2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1989, 1993, 1995
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * Poul-Henning Kamp of the FreeBSD Project.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * @(#)vfs_cache.c 8.5 (Berkeley) 3/22/95
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
41 #include "opt_ktrace.h"
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/capsicum.h>
46 #include <sys/counter.h>
47 #include <sys/filedesc.h>
48 #include <sys/fnv_hash.h>
49 #include <sys/kernel.h>
52 #include <sys/malloc.h>
53 #include <sys/fcntl.h>
55 #include <sys/mount.h>
56 #include <sys/namei.h>
62 #include <sys/syscallsubr.h>
63 #include <sys/sysctl.h>
64 #include <sys/sysproto.h>
65 #include <sys/vnode.h>
68 #include <sys/ktrace.h>
71 #include <machine/_inttypes.h>
74 #include <sys/capsicum.h>
76 #include <security/audit/audit.h>
77 #include <security/mac/mac_framework.h>
85 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
88 SDT_PROVIDER_DECLARE(vfs);
89 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
91 SDT_PROBE_DEFINE3(vfs, namecache, enter, duplicate, "struct vnode *", "char *",
93 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
95 SDT_PROBE_DEFINE2(vfs, namecache, fullpath_smr, hit, "struct vnode *",
97 SDT_PROBE_DEFINE4(vfs, namecache, fullpath_smr, miss, "struct vnode *",
98 "struct namecache *", "int", "int");
99 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
100 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
101 "char *", "struct vnode *");
102 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
103 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
104 "struct vnode *", "char *");
105 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
107 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
108 "struct vnode *", "char *");
109 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
111 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, hit, "struct vnode *",
112 "struct componentname *");
113 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, miss, "struct vnode *",
114 "struct componentname *");
115 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
116 SDT_PROBE_DEFINE1(vfs, namecache, purge, batch, "int");
117 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
118 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
119 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
121 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
123 SDT_PROBE_DEFINE2(vfs, namecache, evict_negative, done, "struct vnode *",
125 SDT_PROBE_DEFINE1(vfs, namecache, symlink, alloc__fail, "size_t");
127 SDT_PROBE_DEFINE3(vfs, fplookup, lookup, done, "struct nameidata", "int", "bool");
128 SDT_PROBE_DECLARE(vfs, namei, lookup, entry);
129 SDT_PROBE_DECLARE(vfs, namei, lookup, return);
132 * This structure describes the elements in the cache of recent
133 * names looked up by namei.
139 _Static_assert(sizeof(struct negstate) <= sizeof(struct vnode *),
140 "the state must fit in a union with a pointer without growing it");
143 LIST_ENTRY(namecache) nc_src; /* source vnode list */
144 TAILQ_ENTRY(namecache) nc_dst; /* destination vnode list */
145 CK_SLIST_ENTRY(namecache) nc_hash;/* hash chain */
146 struct vnode *nc_dvp; /* vnode of parent of name */
148 struct vnode *nu_vp; /* vnode the name refers to */
149 struct negstate nu_neg;/* negative entry state */
151 u_char nc_flag; /* flag bits */
152 u_char nc_nlen; /* length of name */
153 char nc_name[0]; /* segment name + nul */
157 * struct namecache_ts repeats struct namecache layout up to the
159 * struct namecache_ts is used in place of struct namecache when time(s) need
160 * to be stored. The nc_dotdottime field is used when a cache entry is mapping
161 * both a non-dotdot directory name plus dotdot for the directory's
164 * See below for alignment requirement.
166 struct namecache_ts {
167 struct timespec nc_time; /* timespec provided by fs */
168 struct timespec nc_dotdottime; /* dotdot timespec provided by fs */
169 int nc_ticks; /* ticks value when entry was added */
171 struct namecache nc_nc;
174 TAILQ_HEAD(cache_freebatch, namecache);
177 * At least mips n32 performs 64-bit accesses to timespec as found
178 * in namecache_ts and requires them to be aligned. Since others
179 * may be in the same spot suffer a little bit and enforce the
180 * alignment for everyone. Note this is a nop for 64-bit platforms.
182 #define CACHE_ZONE_ALIGNMENT UMA_ALIGNOF(time_t)
185 * TODO: the initial value of CACHE_PATH_CUTOFF was inherited from the
186 * 4.4 BSD codebase. Later on struct namecache was tweaked to become
187 * smaller and the value was bumped to retain the total size, but it
188 * was never re-evaluated for suitability. A simple test counting
189 * lengths during package building shows that the value of 45 covers
190 * about 86% of all added entries, reaching 99% at 65.
192 * Regardless of the above, use of dedicated zones instead of malloc may be
193 * inducing additional waste. This may be hard to address as said zones are
194 * tied to VFS SMR. Even if retaining them, the current split should be
198 #define CACHE_PATH_CUTOFF 45
199 #define CACHE_LARGE_PAD 6
201 #define CACHE_PATH_CUTOFF 41
202 #define CACHE_LARGE_PAD 2
205 #define CACHE_ZONE_SMALL_SIZE (offsetof(struct namecache, nc_name) + CACHE_PATH_CUTOFF + 1)
206 #define CACHE_ZONE_SMALL_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_SMALL_SIZE)
207 #define CACHE_ZONE_LARGE_SIZE (offsetof(struct namecache, nc_name) + NAME_MAX + 1 + CACHE_LARGE_PAD)
208 #define CACHE_ZONE_LARGE_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_LARGE_SIZE)
210 _Static_assert((CACHE_ZONE_SMALL_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
211 _Static_assert((CACHE_ZONE_SMALL_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
212 _Static_assert((CACHE_ZONE_LARGE_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
213 _Static_assert((CACHE_ZONE_LARGE_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
215 #define nc_vp n_un.nu_vp
216 #define nc_neg n_un.nu_neg
219 * Flags in namecache.nc_flag
221 #define NCF_WHITE 0x01
222 #define NCF_ISDOTDOT 0x02
225 #define NCF_DVDROP 0x10
226 #define NCF_NEGATIVE 0x20
227 #define NCF_INVALID 0x40
231 * Flags in negstate.neg_flag
235 static bool cache_neg_evict_cond(u_long lnumcache);
238 * Mark an entry as invalid.
240 * This is called before it starts getting deconstructed.
243 cache_ncp_invalidate(struct namecache *ncp)
246 KASSERT((ncp->nc_flag & NCF_INVALID) == 0,
247 ("%s: entry %p already invalid", __func__, ncp));
248 atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_INVALID);
249 atomic_thread_fence_rel();
253 * Check whether the entry can be safely used.
255 * All places which elide locks are supposed to call this after they are
256 * done with reading from an entry.
258 #define cache_ncp_canuse(ncp) ({ \
259 struct namecache *_ncp = (ncp); \
262 atomic_thread_fence_acq(); \
263 _nc_flag = atomic_load_char(&_ncp->nc_flag); \
264 __predict_true((_nc_flag & (NCF_INVALID | NCF_WIP)) == 0); \
268 * Like the above but also checks NCF_WHITE.
270 #define cache_fpl_neg_ncp_canuse(ncp) ({ \
271 struct namecache *_ncp = (ncp); \
274 atomic_thread_fence_acq(); \
275 _nc_flag = atomic_load_char(&_ncp->nc_flag); \
276 __predict_true((_nc_flag & (NCF_INVALID | NCF_WIP | NCF_WHITE)) == 0); \
280 * Name caching works as follows:
282 * Names found by directory scans are retained in a cache
283 * for future reference. It is managed LRU, so frequently
284 * used names will hang around. Cache is indexed by hash value
285 * obtained from (dvp, name) where dvp refers to the directory
288 * If it is a "negative" entry, (i.e. for a name that is known NOT to
289 * exist) the vnode pointer will be NULL.
291 * Upon reaching the last segment of a path, if the reference
292 * is for DELETE, or NOCACHE is set (rewrite), and the
293 * name is located in the cache, it will be dropped.
295 * These locks are used (in the order in which they can be taken):
297 * vnodelock mtx vnode lists and v_cache_dd field protection
298 * bucketlock mtx for access to given set of hash buckets
299 * neglist mtx negative entry LRU management
301 * It is legal to take multiple vnodelock and bucketlock locks. The locking
302 * order is lower address first. Both are recursive.
304 * "." lookups are lockless.
306 * ".." and vnode -> name lookups require vnodelock.
308 * name -> vnode lookup requires the relevant bucketlock to be held for reading.
310 * Insertions and removals of entries require involved vnodes and bucketlocks
311 * to be locked to provide safe operation against other threads modifying the
314 * Some lookups result in removal of the found entry (e.g. getting rid of a
315 * negative entry with the intent to create a positive one), which poses a
316 * problem when multiple threads reach the state. Similarly, two different
317 * threads can purge two different vnodes and try to remove the same name.
319 * If the already held vnode lock is lower than the second required lock, we
320 * can just take the other lock. However, in the opposite case, this could
321 * deadlock. As such, this is resolved by trylocking and if that fails unlocking
322 * the first node, locking everything in order and revalidating the state.
327 static SYSCTL_NODE(_vfs_cache, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
328 "Name cache parameters");
330 static u_int __read_mostly ncsize; /* the size as computed on creation or resizing */
331 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, size, CTLFLAG_RW, &ncsize, 0,
332 "Total namecache capacity");
334 u_int ncsizefactor = 2;
335 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, sizefactor, CTLFLAG_RW, &ncsizefactor, 0,
336 "Size factor for namecache");
338 static u_long __read_mostly ncnegfactor = 5; /* ratio of negative entries */
339 SYSCTL_ULONG(_vfs_cache_param, OID_AUTO, negfactor, CTLFLAG_RW, &ncnegfactor, 0,
340 "Ratio of negative namecache entries");
343 * Negative entry % of namecache capacity above which automatic eviction is allowed.
345 * Check cache_neg_evict_cond for details.
347 static u_int ncnegminpct = 3;
349 static u_int __read_mostly neg_min; /* the above recomputed against ncsize */
350 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, negmin, CTLFLAG_RD, &neg_min, 0,
351 "Negative entry count above which automatic eviction is allowed");
354 * Structures associated with name caching.
356 #define NCHHASH(hash) \
357 (&nchashtbl[(hash) & nchash])
358 static __read_mostly CK_SLIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
359 static u_long __read_mostly nchash; /* size of hash table */
360 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
361 "Size of namecache hash table");
362 static u_long __exclusive_cache_line numneg; /* number of negative entries allocated */
363 static u_long __exclusive_cache_line numcache;/* number of cache entries allocated */
365 struct nchstats nchstats; /* cache effectiveness statistics */
367 static bool __read_frequently cache_fast_revlookup = true;
368 SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_revlookup, CTLFLAG_RW,
369 &cache_fast_revlookup, 0, "");
371 static bool __read_mostly cache_rename_add = true;
372 SYSCTL_BOOL(_vfs, OID_AUTO, cache_rename_add, CTLFLAG_RW,
373 &cache_rename_add, 0, "");
375 static u_int __exclusive_cache_line neg_cycle;
378 #define numneglists (ncneghash + 1)
381 struct mtx nl_evict_lock;
382 struct mtx nl_lock __aligned(CACHE_LINE_SIZE);
383 TAILQ_HEAD(, namecache) nl_list;
384 TAILQ_HEAD(, namecache) nl_hotlist;
386 } __aligned(CACHE_LINE_SIZE);
388 static struct neglist neglists[numneglists];
390 static inline struct neglist *
391 NCP2NEGLIST(struct namecache *ncp)
394 return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
397 static inline struct negstate *
398 NCP2NEGSTATE(struct namecache *ncp)
401 MPASS(ncp->nc_flag & NCF_NEGATIVE);
402 return (&ncp->nc_neg);
405 #define numbucketlocks (ncbuckethash + 1)
406 static u_int __read_mostly ncbuckethash;
407 static struct mtx_padalign __read_mostly *bucketlocks;
408 #define HASH2BUCKETLOCK(hash) \
409 ((struct mtx *)(&bucketlocks[((hash) & ncbuckethash)]))
411 #define numvnodelocks (ncvnodehash + 1)
412 static u_int __read_mostly ncvnodehash;
413 static struct mtx __read_mostly *vnodelocks;
414 static inline struct mtx *
415 VP2VNODELOCK(struct vnode *vp)
418 return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
422 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
424 struct namecache_ts *ncp_ts;
426 KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
427 (tsp == NULL && ticksp == NULL),
433 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
434 *tsp = ncp_ts->nc_time;
435 *ticksp = ncp_ts->nc_ticks;
439 static int __read_mostly doingcache = 1; /* 1 => enable the cache */
440 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
441 "VFS namecache enabled");
444 /* Export size information to userland */
445 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
446 sizeof(struct namecache), "sizeof(struct namecache)");
449 * The new name cache statistics
451 static SYSCTL_NODE(_vfs_cache, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
452 "Name cache statistics");
454 #define STATNODE_ULONG(name, varname, descr) \
455 SYSCTL_ULONG(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
456 #define STATNODE_COUNTER(name, varname, descr) \
457 static COUNTER_U64_DEFINE_EARLY(varname); \
458 SYSCTL_COUNTER_U64(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, \
460 STATNODE_ULONG(neg, numneg, "Number of negative cache entries");
461 STATNODE_ULONG(count, numcache, "Number of cache entries");
462 STATNODE_COUNTER(heldvnodes, numcachehv, "Number of namecache entries with vnodes held");
463 STATNODE_COUNTER(drops, numdrops, "Number of dropped entries due to reaching the limit");
464 STATNODE_COUNTER(dothits, dothits, "Number of '.' hits");
465 STATNODE_COUNTER(dotdothis, dotdothits, "Number of '..' hits");
466 STATNODE_COUNTER(miss, nummiss, "Number of cache misses");
467 STATNODE_COUNTER(misszap, nummisszap, "Number of cache misses we do not want to cache");
468 STATNODE_COUNTER(posszaps, numposzaps,
469 "Number of cache hits (positive) we do not want to cache");
470 STATNODE_COUNTER(poshits, numposhits, "Number of cache hits (positive)");
471 STATNODE_COUNTER(negzaps, numnegzaps,
472 "Number of cache hits (negative) we do not want to cache");
473 STATNODE_COUNTER(neghits, numneghits, "Number of cache hits (negative)");
474 /* These count for vn_getcwd(), too. */
475 STATNODE_COUNTER(fullpathcalls, numfullpathcalls, "Number of fullpath search calls");
476 STATNODE_COUNTER(fullpathfail1, numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
477 STATNODE_COUNTER(fullpathfail2, numfullpathfail2,
478 "Number of fullpath search errors (VOP_VPTOCNP failures)");
479 STATNODE_COUNTER(fullpathfail4, numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
480 STATNODE_COUNTER(fullpathfound, numfullpathfound, "Number of successful fullpath calls");
481 STATNODE_COUNTER(symlinktoobig, symlinktoobig, "Number of times symlink did not fit the cache");
484 * Debug or developer statistics.
486 static SYSCTL_NODE(_vfs_cache, OID_AUTO, debug, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
487 "Name cache debugging");
488 #define DEBUGNODE_ULONG(name, varname, descr) \
489 SYSCTL_ULONG(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
490 #define DEBUGNODE_COUNTER(name, varname, descr) \
491 static COUNTER_U64_DEFINE_EARLY(varname); \
492 SYSCTL_COUNTER_U64(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, \
494 DEBUGNODE_COUNTER(zap_bucket_relock_success, zap_bucket_relock_success,
495 "Number of successful removals after relocking");
496 static long zap_bucket_fail;
497 DEBUGNODE_ULONG(zap_bucket_fail, zap_bucket_fail, "");
498 static long zap_bucket_fail2;
499 DEBUGNODE_ULONG(zap_bucket_fail2, zap_bucket_fail2, "");
500 static long cache_lock_vnodes_cel_3_failures;
501 DEBUGNODE_ULONG(vnodes_cel_3_failures, cache_lock_vnodes_cel_3_failures,
502 "Number of times 3-way vnode locking failed");
504 static void cache_zap_locked(struct namecache *ncp);
505 static int vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf,
506 char **freebuf, size_t *buflen);
507 static int vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
508 char **retbuf, size_t *buflen, size_t addend);
509 static int vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf,
510 char **retbuf, size_t *buflen);
511 static int vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf,
512 char **retbuf, size_t *len, size_t addend);
514 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
517 cache_assert_vlp_locked(struct mtx *vlp)
521 mtx_assert(vlp, MA_OWNED);
525 cache_assert_vnode_locked(struct vnode *vp)
529 vlp = VP2VNODELOCK(vp);
530 cache_assert_vlp_locked(vlp);
534 * Directory vnodes with entries are held for two reasons:
535 * 1. make them less of a target for reclamation in vnlru
536 * 2. suffer smaller performance penalty in locked lookup as requeieing is avoided
538 * It will be feasible to stop doing it altogether if all filesystems start
539 * supporting lockless lookup.
542 cache_hold_vnode(struct vnode *vp)
545 cache_assert_vnode_locked(vp);
546 VNPASS(LIST_EMPTY(&vp->v_cache_src), vp);
548 counter_u64_add(numcachehv, 1);
552 cache_drop_vnode(struct vnode *vp)
556 * Called after all locks are dropped, meaning we can't assert
557 * on the state of v_cache_src.
560 counter_u64_add(numcachehv, -1);
566 static uma_zone_t __read_mostly cache_zone_small;
567 static uma_zone_t __read_mostly cache_zone_small_ts;
568 static uma_zone_t __read_mostly cache_zone_large;
569 static uma_zone_t __read_mostly cache_zone_large_ts;
572 cache_symlink_alloc(size_t size, int flags)
575 if (size < CACHE_ZONE_SMALL_SIZE) {
576 return (uma_zalloc_smr(cache_zone_small, flags));
578 if (size < CACHE_ZONE_LARGE_SIZE) {
579 return (uma_zalloc_smr(cache_zone_large, flags));
581 counter_u64_add(symlinktoobig, 1);
582 SDT_PROBE1(vfs, namecache, symlink, alloc__fail, size);
587 cache_symlink_free(char *string, size_t size)
590 MPASS(string != NULL);
591 KASSERT(size < CACHE_ZONE_LARGE_SIZE,
592 ("%s: size %zu too big", __func__, size));
594 if (size < CACHE_ZONE_SMALL_SIZE) {
595 uma_zfree_smr(cache_zone_small, string);
598 if (size < CACHE_ZONE_LARGE_SIZE) {
599 uma_zfree_smr(cache_zone_large, string);
602 __assert_unreachable();
605 static struct namecache *
606 cache_alloc_uma(int len, bool ts)
608 struct namecache_ts *ncp_ts;
609 struct namecache *ncp;
611 if (__predict_false(ts)) {
612 if (len <= CACHE_PATH_CUTOFF)
613 ncp_ts = uma_zalloc_smr(cache_zone_small_ts, M_WAITOK);
615 ncp_ts = uma_zalloc_smr(cache_zone_large_ts, M_WAITOK);
616 ncp = &ncp_ts->nc_nc;
618 if (len <= CACHE_PATH_CUTOFF)
619 ncp = uma_zalloc_smr(cache_zone_small, M_WAITOK);
621 ncp = uma_zalloc_smr(cache_zone_large, M_WAITOK);
627 cache_free_uma(struct namecache *ncp)
629 struct namecache_ts *ncp_ts;
631 if (__predict_false(ncp->nc_flag & NCF_TS)) {
632 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
633 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
634 uma_zfree_smr(cache_zone_small_ts, ncp_ts);
636 uma_zfree_smr(cache_zone_large_ts, ncp_ts);
638 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
639 uma_zfree_smr(cache_zone_small, ncp);
641 uma_zfree_smr(cache_zone_large, ncp);
645 static struct namecache *
646 cache_alloc(int len, bool ts)
651 * Avoid blowout in namecache entries.
654 * 1. filesystems may end up trying to add an already existing entry
655 * (for example this can happen after a cache miss during concurrent
656 * lookup), in which case we will call cache_neg_evict despite not
658 * 2. the routine may fail to free anything and no provisions are made
659 * to make it try harder (see the inside for failure modes)
660 * 3. it only ever looks at negative entries.
662 lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
663 if (cache_neg_evict_cond(lnumcache)) {
664 lnumcache = atomic_load_long(&numcache);
666 if (__predict_false(lnumcache >= ncsize)) {
667 atomic_subtract_long(&numcache, 1);
668 counter_u64_add(numdrops, 1);
671 return (cache_alloc_uma(len, ts));
675 cache_free(struct namecache *ncp)
679 if ((ncp->nc_flag & NCF_DVDROP) != 0) {
680 cache_drop_vnode(ncp->nc_dvp);
683 atomic_subtract_long(&numcache, 1);
687 cache_free_batch(struct cache_freebatch *batch)
689 struct namecache *ncp, *nnp;
693 if (TAILQ_EMPTY(batch))
695 TAILQ_FOREACH_SAFE(ncp, batch, nc_dst, nnp) {
696 if ((ncp->nc_flag & NCF_DVDROP) != 0) {
697 cache_drop_vnode(ncp->nc_dvp);
702 atomic_subtract_long(&numcache, i);
704 SDT_PROBE1(vfs, namecache, purge, batch, i);
708 * TODO: With the value stored we can do better than computing the hash based
709 * on the address. The choice of FNV should also be revisited.
712 cache_prehash(struct vnode *vp)
715 vp->v_nchash = fnv_32_buf(&vp, sizeof(vp), FNV1_32_INIT);
719 cache_get_hash(char *name, u_char len, struct vnode *dvp)
722 return (fnv_32_buf(name, len, dvp->v_nchash));
725 static inline struct nchashhead *
726 NCP2BUCKET(struct namecache *ncp)
730 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
731 return (NCHHASH(hash));
734 static inline struct mtx *
735 NCP2BUCKETLOCK(struct namecache *ncp)
739 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
740 return (HASH2BUCKETLOCK(hash));
745 cache_assert_bucket_locked(struct namecache *ncp)
749 blp = NCP2BUCKETLOCK(ncp);
750 mtx_assert(blp, MA_OWNED);
754 cache_assert_bucket_unlocked(struct namecache *ncp)
758 blp = NCP2BUCKETLOCK(ncp);
759 mtx_assert(blp, MA_NOTOWNED);
762 #define cache_assert_bucket_locked(x) do { } while (0)
763 #define cache_assert_bucket_unlocked(x) do { } while (0)
766 #define cache_sort_vnodes(x, y) _cache_sort_vnodes((void **)(x), (void **)(y))
768 _cache_sort_vnodes(void **p1, void **p2)
772 MPASS(*p1 != NULL || *p2 != NULL);
782 cache_lock_all_buckets(void)
786 for (i = 0; i < numbucketlocks; i++)
787 mtx_lock(&bucketlocks[i]);
791 cache_unlock_all_buckets(void)
795 for (i = 0; i < numbucketlocks; i++)
796 mtx_unlock(&bucketlocks[i]);
800 cache_lock_all_vnodes(void)
804 for (i = 0; i < numvnodelocks; i++)
805 mtx_lock(&vnodelocks[i]);
809 cache_unlock_all_vnodes(void)
813 for (i = 0; i < numvnodelocks; i++)
814 mtx_unlock(&vnodelocks[i]);
818 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
821 cache_sort_vnodes(&vlp1, &vlp2);
824 if (!mtx_trylock(vlp1))
827 if (!mtx_trylock(vlp2)) {
837 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
840 MPASS(vlp1 != NULL || vlp2 != NULL);
850 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
853 MPASS(vlp1 != NULL || vlp2 != NULL);
862 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
864 struct nchstats snap;
866 if (req->oldptr == NULL)
867 return (SYSCTL_OUT(req, 0, sizeof(snap)));
870 snap.ncs_goodhits = counter_u64_fetch(numposhits);
871 snap.ncs_neghits = counter_u64_fetch(numneghits);
872 snap.ncs_badhits = counter_u64_fetch(numposzaps) +
873 counter_u64_fetch(numnegzaps);
874 snap.ncs_miss = counter_u64_fetch(nummisszap) +
875 counter_u64_fetch(nummiss);
877 return (SYSCTL_OUT(req, &snap, sizeof(snap)));
879 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
880 CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
881 "VFS cache effectiveness statistics");
884 cache_recalc_neg_min(u_int val)
887 neg_min = (ncsize * val) / 100;
891 sysctl_negminpct(SYSCTL_HANDLER_ARGS)
897 error = sysctl_handle_int(oidp, &val, 0, req);
898 if (error != 0 || req->newptr == NULL)
901 if (val == ncnegminpct)
903 if (val < 0 || val > 99)
906 cache_recalc_neg_min(val);
910 SYSCTL_PROC(_vfs_cache_param, OID_AUTO, negminpct,
911 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_negminpct,
912 "I", "Negative entry \% of namecache capacity above which automatic eviction is allowed");
916 * Grab an atomic snapshot of the name cache hash chain lengths
918 static SYSCTL_NODE(_debug, OID_AUTO, hashstat,
919 CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
923 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
925 struct nchashhead *ncpp;
926 struct namecache *ncp;
927 int i, error, n_nchash, *cntbuf;
930 n_nchash = nchash + 1; /* nchash is max index, not count */
931 if (req->oldptr == NULL)
932 return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
933 cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
934 cache_lock_all_buckets();
935 if (n_nchash != nchash + 1) {
936 cache_unlock_all_buckets();
937 free(cntbuf, M_TEMP);
940 /* Scan hash tables counting entries */
941 for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
942 CK_SLIST_FOREACH(ncp, ncpp, nc_hash)
944 cache_unlock_all_buckets();
945 for (error = 0, i = 0; i < n_nchash; i++)
946 if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
948 free(cntbuf, M_TEMP);
951 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
952 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
953 "nchash chain lengths");
956 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
959 struct nchashhead *ncpp;
960 struct namecache *ncp;
962 int count, maxlength, used, pct;
965 return SYSCTL_OUT(req, 0, 4 * sizeof(int));
967 cache_lock_all_buckets();
968 n_nchash = nchash + 1; /* nchash is max index, not count */
972 /* Scan hash tables for applicable entries */
973 for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
975 CK_SLIST_FOREACH(ncp, ncpp, nc_hash) {
980 if (maxlength < count)
983 n_nchash = nchash + 1;
984 cache_unlock_all_buckets();
985 pct = (used * 100) / (n_nchash / 100);
986 error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
989 error = SYSCTL_OUT(req, &used, sizeof(used));
992 error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
995 error = SYSCTL_OUT(req, &pct, sizeof(pct));
1000 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
1001 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
1002 "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
1006 * Negative entries management
1008 * Various workloads create plenty of negative entries and barely use them
1009 * afterwards. Moreover malicious users can keep performing bogus lookups
1010 * adding even more entries. For example "make tinderbox" as of writing this
1011 * comment ends up with 2.6M namecache entries in total, 1.2M of which are
1014 * As such, a rather aggressive eviction method is needed. The currently
1015 * employed method is a placeholder.
1017 * Entries are split over numneglists separate lists, each of which is further
1018 * split into hot and cold entries. Entries get promoted after getting a hit.
1019 * Eviction happens on addition of new entry.
1021 static SYSCTL_NODE(_vfs_cache, OID_AUTO, neg, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1022 "Name cache negative entry statistics");
1024 SYSCTL_ULONG(_vfs_cache_neg, OID_AUTO, count, CTLFLAG_RD, &numneg, 0,
1025 "Number of negative cache entries");
1027 static COUNTER_U64_DEFINE_EARLY(neg_created);
1028 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, created, CTLFLAG_RD, &neg_created,
1029 "Number of created negative entries");
1031 static COUNTER_U64_DEFINE_EARLY(neg_evicted);
1032 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evicted, CTLFLAG_RD, &neg_evicted,
1033 "Number of evicted negative entries");
1035 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_empty);
1036 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_empty, CTLFLAG_RD,
1037 &neg_evict_skipped_empty,
1038 "Number of times evicting failed due to lack of entries");
1040 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_missed);
1041 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_missed, CTLFLAG_RD,
1042 &neg_evict_skipped_missed,
1043 "Number of times evicting failed due to target entry disappearing");
1045 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_contended);
1046 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_contended, CTLFLAG_RD,
1047 &neg_evict_skipped_contended,
1048 "Number of times evicting failed due to contention");
1050 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, hits, CTLFLAG_RD, &numneghits,
1051 "Number of cache hits (negative)");
1054 sysctl_neg_hot(SYSCTL_HANDLER_ARGS)
1059 for (i = 0; i < numneglists; i++)
1060 out += neglists[i].nl_hotnum;
1062 return (SYSCTL_OUT(req, &out, sizeof(out)));
1064 SYSCTL_PROC(_vfs_cache_neg, OID_AUTO, hot, CTLTYPE_INT | CTLFLAG_RD |
1065 CTLFLAG_MPSAFE, 0, 0, sysctl_neg_hot, "I",
1066 "Number of hot negative entries");
1069 cache_neg_init(struct namecache *ncp)
1071 struct negstate *ns;
1073 ncp->nc_flag |= NCF_NEGATIVE;
1074 ns = NCP2NEGSTATE(ncp);
1077 counter_u64_add(neg_created, 1);
1080 #define CACHE_NEG_PROMOTION_THRESH 2
1083 cache_neg_hit_prep(struct namecache *ncp)
1085 struct negstate *ns;
1088 ns = NCP2NEGSTATE(ncp);
1089 n = atomic_load_char(&ns->neg_hit);
1091 if (n >= CACHE_NEG_PROMOTION_THRESH)
1093 if (atomic_fcmpset_8(&ns->neg_hit, &n, n + 1))
1096 return (n + 1 == CACHE_NEG_PROMOTION_THRESH);
1100 * Nothing to do here but it is provided for completeness as some
1101 * cache_neg_hit_prep callers may end up returning without even
1102 * trying to promote.
1104 #define cache_neg_hit_abort(ncp) do { } while (0)
1107 cache_neg_hit_finish(struct namecache *ncp)
1110 SDT_PROBE2(vfs, namecache, lookup, hit__negative, ncp->nc_dvp, ncp->nc_name);
1111 counter_u64_add(numneghits, 1);
1115 * Move a negative entry to the hot list.
1118 cache_neg_promote_locked(struct namecache *ncp)
1121 struct negstate *ns;
1123 ns = NCP2NEGSTATE(ncp);
1124 nl = NCP2NEGLIST(ncp);
1125 mtx_assert(&nl->nl_lock, MA_OWNED);
1126 if ((ns->neg_flag & NEG_HOT) == 0) {
1127 TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1128 TAILQ_INSERT_TAIL(&nl->nl_hotlist, ncp, nc_dst);
1130 ns->neg_flag |= NEG_HOT;
1135 * Move a hot negative entry to the cold list.
1138 cache_neg_demote_locked(struct namecache *ncp)
1141 struct negstate *ns;
1143 ns = NCP2NEGSTATE(ncp);
1144 nl = NCP2NEGLIST(ncp);
1145 mtx_assert(&nl->nl_lock, MA_OWNED);
1146 MPASS(ns->neg_flag & NEG_HOT);
1147 TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1148 TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1150 ns->neg_flag &= ~NEG_HOT;
1151 atomic_store_char(&ns->neg_hit, 0);
1155 * Move a negative entry to the hot list if it matches the lookup.
1157 * We have to take locks, but they may be contended and in the worst
1158 * case we may need to go off CPU. We don't want to spin within the
1159 * smr section and we can't block with it. Exiting the section means
1160 * the found entry could have been evicted. We are going to look it
1164 cache_neg_promote_cond(struct vnode *dvp, struct componentname *cnp,
1165 struct namecache *oncp, uint32_t hash)
1167 struct namecache *ncp;
1171 nl = NCP2NEGLIST(oncp);
1173 mtx_lock(&nl->nl_lock);
1175 * For hash iteration.
1180 * Avoid all surprises by only succeeding if we got the same entry and
1181 * bailing completely otherwise.
1182 * XXX There are no provisions to keep the vnode around, meaning we may
1183 * end up promoting a negative entry for a *new* vnode and returning
1184 * ENOENT on its account. This is the error we want to return anyway
1185 * and promotion is harmless.
1187 * In particular at this point there can be a new ncp which matches the
1188 * search but hashes to a different neglist.
1190 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1196 * No match to begin with.
1198 if (__predict_false(ncp == NULL)) {
1203 * The newly found entry may be something different...
1205 if (!(ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1206 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))) {
1211 * ... and not even negative.
1213 nc_flag = atomic_load_char(&ncp->nc_flag);
1214 if ((nc_flag & NCF_NEGATIVE) == 0) {
1218 if (!cache_ncp_canuse(ncp)) {
1222 cache_neg_promote_locked(ncp);
1223 cache_neg_hit_finish(ncp);
1225 mtx_unlock(&nl->nl_lock);
1229 mtx_unlock(&nl->nl_lock);
1234 cache_neg_promote(struct namecache *ncp)
1238 nl = NCP2NEGLIST(ncp);
1239 mtx_lock(&nl->nl_lock);
1240 cache_neg_promote_locked(ncp);
1241 mtx_unlock(&nl->nl_lock);
1245 cache_neg_insert(struct namecache *ncp)
1249 MPASS(ncp->nc_flag & NCF_NEGATIVE);
1250 cache_assert_bucket_locked(ncp);
1251 nl = NCP2NEGLIST(ncp);
1252 mtx_lock(&nl->nl_lock);
1253 TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1254 mtx_unlock(&nl->nl_lock);
1255 atomic_add_long(&numneg, 1);
1259 cache_neg_remove(struct namecache *ncp)
1262 struct negstate *ns;
1264 cache_assert_bucket_locked(ncp);
1265 nl = NCP2NEGLIST(ncp);
1266 ns = NCP2NEGSTATE(ncp);
1267 mtx_lock(&nl->nl_lock);
1268 if ((ns->neg_flag & NEG_HOT) != 0) {
1269 TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1272 TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1274 mtx_unlock(&nl->nl_lock);
1275 atomic_subtract_long(&numneg, 1);
1278 static struct neglist *
1279 cache_neg_evict_select_list(void)
1284 c = atomic_fetchadd_int(&neg_cycle, 1) + 1;
1285 nl = &neglists[c % numneglists];
1286 if (!mtx_trylock(&nl->nl_evict_lock)) {
1287 counter_u64_add(neg_evict_skipped_contended, 1);
1293 static struct namecache *
1294 cache_neg_evict_select_entry(struct neglist *nl)
1296 struct namecache *ncp, *lncp;
1297 struct negstate *ns, *lns;
1300 mtx_assert(&nl->nl_evict_lock, MA_OWNED);
1301 mtx_assert(&nl->nl_lock, MA_OWNED);
1302 ncp = TAILQ_FIRST(&nl->nl_list);
1306 lns = NCP2NEGSTATE(lncp);
1307 for (i = 1; i < 4; i++) {
1308 ncp = TAILQ_NEXT(ncp, nc_dst);
1311 ns = NCP2NEGSTATE(ncp);
1312 if (ns->neg_hit < lns->neg_hit) {
1321 cache_neg_evict(void)
1323 struct namecache *ncp, *ncp2;
1332 nl = cache_neg_evict_select_list();
1337 mtx_lock(&nl->nl_lock);
1338 ncp = TAILQ_FIRST(&nl->nl_hotlist);
1340 cache_neg_demote_locked(ncp);
1342 ncp = cache_neg_evict_select_entry(nl);
1344 counter_u64_add(neg_evict_skipped_empty, 1);
1345 mtx_unlock(&nl->nl_lock);
1346 mtx_unlock(&nl->nl_evict_lock);
1349 nlen = ncp->nc_nlen;
1351 hash = cache_get_hash(ncp->nc_name, nlen, dvp);
1352 dvlp = VP2VNODELOCK(dvp);
1353 blp = HASH2BUCKETLOCK(hash);
1354 mtx_unlock(&nl->nl_lock);
1355 mtx_unlock(&nl->nl_evict_lock);
1359 * Note that since all locks were dropped above, the entry may be
1360 * gone or reallocated to be something else.
1362 CK_SLIST_FOREACH(ncp2, (NCHHASH(hash)), nc_hash) {
1363 if (ncp2 == ncp && ncp2->nc_dvp == dvp &&
1364 ncp2->nc_nlen == nlen && (ncp2->nc_flag & NCF_NEGATIVE) != 0)
1368 counter_u64_add(neg_evict_skipped_missed, 1);
1372 MPASS(dvlp == VP2VNODELOCK(ncp->nc_dvp));
1373 MPASS(blp == NCP2BUCKETLOCK(ncp));
1374 SDT_PROBE2(vfs, namecache, evict_negative, done, ncp->nc_dvp,
1376 cache_zap_locked(ncp);
1377 counter_u64_add(neg_evicted, 1);
1388 * Maybe evict a negative entry to create more room.
1390 * The ncnegfactor parameter limits what fraction of the total count
1391 * can comprise of negative entries. However, if the cache is just
1392 * warming up this leads to excessive evictions. As such, ncnegminpct
1393 * (recomputed to neg_min) dictates whether the above should be
1396 * Try evicting if the cache is close to full capacity regardless of
1397 * other considerations.
1400 cache_neg_evict_cond(u_long lnumcache)
1404 if (ncsize - 1000 < lnumcache)
1406 lnumneg = atomic_load_long(&numneg);
1407 if (lnumneg < neg_min)
1409 if (lnumneg * ncnegfactor < lnumcache)
1412 return (cache_neg_evict());
1416 * cache_zap_locked():
1418 * Removes a namecache entry from cache, whether it contains an actual
1419 * pointer to a vnode or if it is just a negative cache entry.
1422 cache_zap_locked(struct namecache *ncp)
1424 struct nchashhead *ncpp;
1425 struct vnode *dvp, *vp;
1430 if (!(ncp->nc_flag & NCF_NEGATIVE))
1431 cache_assert_vnode_locked(vp);
1432 cache_assert_vnode_locked(dvp);
1433 cache_assert_bucket_locked(ncp);
1435 cache_ncp_invalidate(ncp);
1437 ncpp = NCP2BUCKET(ncp);
1438 CK_SLIST_REMOVE(ncpp, ncp, namecache, nc_hash);
1439 if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1440 SDT_PROBE3(vfs, namecache, zap, done, dvp, ncp->nc_name, vp);
1441 TAILQ_REMOVE(&vp->v_cache_dst, ncp, nc_dst);
1442 if (ncp == vp->v_cache_dd) {
1443 vn_seqc_write_begin_unheld(vp);
1444 vp->v_cache_dd = NULL;
1445 vn_seqc_write_end(vp);
1448 SDT_PROBE2(vfs, namecache, zap_negative, done, dvp, ncp->nc_name);
1449 cache_neg_remove(ncp);
1451 if (ncp->nc_flag & NCF_ISDOTDOT) {
1452 if (ncp == dvp->v_cache_dd) {
1453 vn_seqc_write_begin_unheld(dvp);
1454 dvp->v_cache_dd = NULL;
1455 vn_seqc_write_end(dvp);
1458 LIST_REMOVE(ncp, nc_src);
1459 if (LIST_EMPTY(&dvp->v_cache_src)) {
1460 ncp->nc_flag |= NCF_DVDROP;
1466 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
1470 MPASS(ncp->nc_dvp == vp);
1471 MPASS(ncp->nc_flag & NCF_NEGATIVE);
1472 cache_assert_vnode_locked(vp);
1474 blp = NCP2BUCKETLOCK(ncp);
1476 cache_zap_locked(ncp);
1481 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
1484 struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
1487 MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
1488 cache_assert_vnode_locked(vp);
1490 if (ncp->nc_flag & NCF_NEGATIVE) {
1491 if (*vlpp != NULL) {
1495 cache_zap_negative_locked_vnode_kl(ncp, vp);
1499 pvlp = VP2VNODELOCK(vp);
1500 blp = NCP2BUCKETLOCK(ncp);
1501 vlp1 = VP2VNODELOCK(ncp->nc_dvp);
1502 vlp2 = VP2VNODELOCK(ncp->nc_vp);
1504 if (*vlpp == vlp1 || *vlpp == vlp2) {
1508 if (*vlpp != NULL) {
1512 cache_sort_vnodes(&vlp1, &vlp2);
1517 if (!mtx_trylock(vlp1))
1523 cache_zap_locked(ncp);
1525 if (to_unlock != NULL)
1526 mtx_unlock(to_unlock);
1533 MPASS(*vlpp == NULL);
1539 * If trylocking failed we can get here. We know enough to take all needed locks
1540 * in the right order and re-lookup the entry.
1543 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1544 struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1547 struct namecache *rncp;
1549 cache_assert_bucket_unlocked(ncp);
1551 cache_sort_vnodes(&dvlp, &vlp);
1552 cache_lock_vnodes(dvlp, vlp);
1554 CK_SLIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1555 if (rncp == ncp && rncp->nc_dvp == dvp &&
1556 rncp->nc_nlen == cnp->cn_namelen &&
1557 !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1561 cache_zap_locked(rncp);
1563 cache_unlock_vnodes(dvlp, vlp);
1564 counter_u64_add(zap_bucket_relock_success, 1);
1569 cache_unlock_vnodes(dvlp, vlp);
1573 static int __noinline
1574 cache_zap_locked_bucket(struct namecache *ncp, struct componentname *cnp,
1575 uint32_t hash, struct mtx *blp)
1577 struct mtx *dvlp, *vlp;
1580 cache_assert_bucket_locked(ncp);
1582 dvlp = VP2VNODELOCK(ncp->nc_dvp);
1584 if (!(ncp->nc_flag & NCF_NEGATIVE))
1585 vlp = VP2VNODELOCK(ncp->nc_vp);
1586 if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1587 cache_zap_locked(ncp);
1589 cache_unlock_vnodes(dvlp, vlp);
1595 return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1598 static __noinline int
1599 cache_remove_cnp(struct vnode *dvp, struct componentname *cnp)
1601 struct namecache *ncp;
1603 struct mtx *dvlp, *dvlp2;
1607 if (cnp->cn_namelen == 2 &&
1608 cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1609 dvlp = VP2VNODELOCK(dvp);
1613 ncp = dvp->v_cache_dd;
1618 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1621 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1622 if (!cache_zap_locked_vnode_kl2(ncp, dvp, &dvlp2))
1624 MPASS(dvp->v_cache_dd == NULL);
1630 vn_seqc_write_begin(dvp);
1631 dvp->v_cache_dd = NULL;
1632 vn_seqc_write_end(dvp);
1637 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1641 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1642 blp = HASH2BUCKETLOCK(hash);
1644 if (CK_SLIST_EMPTY(NCHHASH(hash)))
1649 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1650 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1651 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1660 error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1661 if (__predict_false(error != 0)) {
1665 counter_u64_add(numposzaps, 1);
1666 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1670 counter_u64_add(nummisszap, 1);
1671 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1675 static int __noinline
1676 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1677 struct timespec *tsp, int *ticksp)
1682 counter_u64_add(dothits, 1);
1683 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1690 * When we lookup "." we still can be asked to lock it
1693 ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1694 if (ltype != VOP_ISLOCKED(*vpp)) {
1695 if (ltype == LK_EXCLUSIVE) {
1696 vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1697 if (VN_IS_DOOMED((*vpp))) {
1698 /* forced unmount */
1704 vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1709 static int __noinline
1710 cache_lookup_dotdot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1711 struct timespec *tsp, int *ticksp)
1713 struct namecache_ts *ncp_ts;
1714 struct namecache *ncp;
1720 MPASS((cnp->cn_flags & ISDOTDOT) != 0);
1722 if ((cnp->cn_flags & MAKEENTRY) == 0) {
1723 cache_remove_cnp(dvp, cnp);
1727 counter_u64_add(dotdothits, 1);
1729 dvlp = VP2VNODELOCK(dvp);
1731 ncp = dvp->v_cache_dd;
1733 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, "..", NULL);
1737 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1738 if (ncp->nc_flag & NCF_NEGATIVE)
1745 goto negative_success;
1746 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", *vpp);
1747 cache_out_ts(ncp, tsp, ticksp);
1748 if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1749 NCF_DTS && tsp != NULL) {
1750 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1751 *tsp = ncp_ts->nc_dotdottime;
1755 ltype = VOP_ISLOCKED(dvp);
1757 vs = vget_prep(*vpp);
1759 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1760 vn_lock(dvp, ltype | LK_RETRY);
1761 if (VN_IS_DOOMED(dvp)) {
1773 if (__predict_false(cnp->cn_nameiop == CREATE)) {
1774 if (cnp->cn_flags & ISLASTCN) {
1775 counter_u64_add(numnegzaps, 1);
1776 cache_zap_negative_locked_vnode_kl(ncp, dvp);
1783 whiteout = (ncp->nc_flag & NCF_WHITE);
1784 cache_out_ts(ncp, tsp, ticksp);
1785 if (cache_neg_hit_prep(ncp))
1786 cache_neg_promote(ncp);
1788 cache_neg_hit_finish(ncp);
1791 cnp->cn_flags |= ISWHITEOUT;
1796 * Lookup a name in the name cache
1800 * - dvp: Parent directory in which to search.
1801 * - vpp: Return argument. Will contain desired vnode on cache hit.
1802 * - cnp: Parameters of the name search. The most interesting bits of
1803 * the cn_flags field have the following meanings:
1804 * - MAKEENTRY: If clear, free an entry from the cache rather than look
1806 * - ISDOTDOT: Must be set if and only if cn_nameptr == ".."
1807 * - tsp: Return storage for cache timestamp. On a successful (positive
1808 * or negative) lookup, tsp will be filled with any timespec that
1809 * was stored when this cache entry was created. However, it will
1810 * be clear for "." entries.
1811 * - ticks: Return storage for alternate cache timestamp. On a successful
1812 * (positive or negative) lookup, it will contain the ticks value
1813 * that was current when the cache entry was created, unless cnp
1816 * Either both tsp and ticks have to be provided or neither of them.
1820 * - -1: A positive cache hit. vpp will contain the desired vnode.
1821 * - ENOENT: A negative cache hit, or dvp was recycled out from under us due
1822 * to a forced unmount. vpp will not be modified. If the entry
1823 * is a whiteout, then the ISWHITEOUT flag will be set in
1825 * - 0: A cache miss. vpp will not be modified.
1829 * On a cache hit, vpp will be returned locked and ref'd. If we're looking up
1830 * .., dvp is unlocked. If we're looking up . an extra ref is taken, but the
1831 * lock is not recursively acquired.
1833 static int __noinline
1834 cache_lookup_fallback(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1835 struct timespec *tsp, int *ticksp)
1837 struct namecache *ncp;
1844 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
1845 MPASS((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) != 0);
1848 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1849 blp = HASH2BUCKETLOCK(hash);
1852 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1853 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1854 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1858 if (__predict_false(ncp == NULL)) {
1860 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1862 counter_u64_add(nummiss, 1);
1866 if (ncp->nc_flag & NCF_NEGATIVE)
1867 goto negative_success;
1869 counter_u64_add(numposhits, 1);
1871 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1872 cache_out_ts(ncp, tsp, ticksp);
1874 vs = vget_prep(*vpp);
1876 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1884 * We don't get here with regular lookup apart from corner cases.
1886 if (__predict_true(cnp->cn_nameiop == CREATE)) {
1887 if (cnp->cn_flags & ISLASTCN) {
1888 counter_u64_add(numnegzaps, 1);
1889 error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1890 if (__predict_false(error != 0)) {
1899 whiteout = (ncp->nc_flag & NCF_WHITE);
1900 cache_out_ts(ncp, tsp, ticksp);
1901 if (cache_neg_hit_prep(ncp))
1902 cache_neg_promote(ncp);
1904 cache_neg_hit_finish(ncp);
1907 cnp->cn_flags |= ISWHITEOUT;
1912 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1913 struct timespec *tsp, int *ticksp)
1915 struct namecache *ncp;
1919 bool whiteout, neg_promote;
1922 MPASS((tsp == NULL && ticksp == NULL) || (tsp != NULL && ticksp != NULL));
1925 if (__predict_false(!doingcache)) {
1926 cnp->cn_flags &= ~MAKEENTRY;
1931 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
1932 if (cnp->cn_namelen == 1)
1933 return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1934 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.')
1935 return (cache_lookup_dotdot(dvp, vpp, cnp, tsp, ticksp));
1938 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
1940 if ((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) == 0) {
1941 cache_remove_cnp(dvp, cnp);
1945 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1948 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1949 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1950 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1954 if (__predict_false(ncp == NULL)) {
1956 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1958 counter_u64_add(nummiss, 1);
1962 nc_flag = atomic_load_char(&ncp->nc_flag);
1963 if (nc_flag & NCF_NEGATIVE)
1964 goto negative_success;
1966 counter_u64_add(numposhits, 1);
1968 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1969 cache_out_ts(ncp, tsp, ticksp);
1971 if (!cache_ncp_canuse(ncp)) {
1976 vs = vget_prep_smr(*vpp);
1978 if (__predict_false(vs == VGET_NONE)) {
1982 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1989 if (cnp->cn_nameiop == CREATE) {
1990 if (cnp->cn_flags & ISLASTCN) {
1996 cache_out_ts(ncp, tsp, ticksp);
1997 whiteout = (ncp->nc_flag & NCF_WHITE);
1998 neg_promote = cache_neg_hit_prep(ncp);
1999 if (!cache_ncp_canuse(ncp)) {
2000 cache_neg_hit_abort(ncp);
2006 if (!cache_neg_promote_cond(dvp, cnp, ncp, hash))
2009 cache_neg_hit_finish(ncp);
2013 cnp->cn_flags |= ISWHITEOUT;
2016 return (cache_lookup_fallback(dvp, vpp, cnp, tsp, ticksp));
2019 struct celockstate {
2023 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
2024 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
2027 cache_celockstate_init(struct celockstate *cel)
2030 bzero(cel, sizeof(*cel));
2034 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
2037 struct mtx *vlp1, *vlp2;
2039 MPASS(cel->vlp[0] == NULL);
2040 MPASS(cel->vlp[1] == NULL);
2041 MPASS(cel->vlp[2] == NULL);
2043 MPASS(vp != NULL || dvp != NULL);
2045 vlp1 = VP2VNODELOCK(vp);
2046 vlp2 = VP2VNODELOCK(dvp);
2047 cache_sort_vnodes(&vlp1, &vlp2);
2058 cache_unlock_vnodes_cel(struct celockstate *cel)
2061 MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
2063 if (cel->vlp[0] != NULL)
2064 mtx_unlock(cel->vlp[0]);
2065 if (cel->vlp[1] != NULL)
2066 mtx_unlock(cel->vlp[1]);
2067 if (cel->vlp[2] != NULL)
2068 mtx_unlock(cel->vlp[2]);
2072 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
2077 cache_assert_vlp_locked(cel->vlp[0]);
2078 cache_assert_vlp_locked(cel->vlp[1]);
2079 MPASS(cel->vlp[2] == NULL);
2082 vlp = VP2VNODELOCK(vp);
2085 if (vlp >= cel->vlp[1]) {
2088 if (mtx_trylock(vlp))
2090 cache_lock_vnodes_cel_3_failures++;
2091 cache_unlock_vnodes_cel(cel);
2092 if (vlp < cel->vlp[0]) {
2094 mtx_lock(cel->vlp[0]);
2095 mtx_lock(cel->vlp[1]);
2097 if (cel->vlp[0] != NULL)
2098 mtx_lock(cel->vlp[0]);
2100 mtx_lock(cel->vlp[1]);
2110 cache_lock_buckets_cel(struct celockstate *cel, struct mtx *blp1,
2114 MPASS(cel->blp[0] == NULL);
2115 MPASS(cel->blp[1] == NULL);
2117 cache_sort_vnodes(&blp1, &blp2);
2128 cache_unlock_buckets_cel(struct celockstate *cel)
2131 if (cel->blp[0] != NULL)
2132 mtx_unlock(cel->blp[0]);
2133 mtx_unlock(cel->blp[1]);
2137 * Lock part of the cache affected by the insertion.
2139 * This means vnodelocks for dvp, vp and the relevant bucketlock.
2140 * However, insertion can result in removal of an old entry. In this
2141 * case we have an additional vnode and bucketlock pair to lock.
2143 * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
2144 * preserving the locking order (smaller address first).
2147 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2150 struct namecache *ncp;
2151 struct mtx *blps[2];
2153 blps[0] = HASH2BUCKETLOCK(hash);
2156 cache_lock_vnodes_cel(cel, dvp, vp);
2157 if (vp == NULL || vp->v_type != VDIR)
2159 ncp = vp->v_cache_dd;
2162 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2164 MPASS(ncp->nc_dvp == vp);
2165 blps[1] = NCP2BUCKETLOCK(ncp);
2166 if (ncp->nc_flag & NCF_NEGATIVE)
2168 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2171 * All vnodes got re-locked. Re-validate the state and if
2172 * nothing changed we are done. Otherwise restart.
2174 if (ncp == vp->v_cache_dd &&
2175 (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2176 blps[1] == NCP2BUCKETLOCK(ncp) &&
2177 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2179 cache_unlock_vnodes_cel(cel);
2184 cache_lock_buckets_cel(cel, blps[0], blps[1]);
2188 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2191 struct namecache *ncp;
2192 struct mtx *blps[2];
2194 blps[0] = HASH2BUCKETLOCK(hash);
2197 cache_lock_vnodes_cel(cel, dvp, vp);
2198 ncp = dvp->v_cache_dd;
2201 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2203 MPASS(ncp->nc_dvp == dvp);
2204 blps[1] = NCP2BUCKETLOCK(ncp);
2205 if (ncp->nc_flag & NCF_NEGATIVE)
2207 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2209 if (ncp == dvp->v_cache_dd &&
2210 (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2211 blps[1] == NCP2BUCKETLOCK(ncp) &&
2212 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2214 cache_unlock_vnodes_cel(cel);
2219 cache_lock_buckets_cel(cel, blps[0], blps[1]);
2223 cache_enter_unlock(struct celockstate *cel)
2226 cache_unlock_buckets_cel(cel);
2227 cache_unlock_vnodes_cel(cel);
2230 static void __noinline
2231 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
2232 struct componentname *cnp)
2234 struct celockstate cel;
2235 struct namecache *ncp;
2239 if (dvp->v_cache_dd == NULL)
2241 len = cnp->cn_namelen;
2242 cache_celockstate_init(&cel);
2243 hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2244 cache_enter_lock_dd(&cel, dvp, vp, hash);
2245 vn_seqc_write_begin(dvp);
2246 ncp = dvp->v_cache_dd;
2247 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
2248 KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
2249 cache_zap_locked(ncp);
2253 dvp->v_cache_dd = NULL;
2254 vn_seqc_write_end(dvp);
2255 cache_enter_unlock(&cel);
2261 * Add an entry to the cache.
2264 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
2265 struct timespec *tsp, struct timespec *dtsp)
2267 struct celockstate cel;
2268 struct namecache *ncp, *n2, *ndd;
2269 struct namecache_ts *ncp_ts;
2270 struct nchashhead *ncpp;
2275 KASSERT(cnp->cn_namelen <= NAME_MAX,
2276 ("%s: passed len %ld exceeds NAME_MAX (%d)", __func__, cnp->cn_namelen,
2278 VNPASS(dvp != vp, dvp);
2279 VNPASS(!VN_IS_DOOMED(dvp), dvp);
2280 VNPASS(dvp->v_type != VNON, dvp);
2282 VNPASS(!VN_IS_DOOMED(vp), vp);
2283 VNPASS(vp->v_type != VNON, vp);
2287 if (__predict_false(!doingcache))
2292 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
2293 if (cnp->cn_namelen == 1)
2295 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
2296 cache_enter_dotdot_prep(dvp, vp, cnp);
2297 flag = NCF_ISDOTDOT;
2301 ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
2305 cache_celockstate_init(&cel);
2310 * Calculate the hash key and setup as much of the new
2311 * namecache entry as possible before acquiring the lock.
2313 ncp->nc_flag = flag | NCF_WIP;
2316 cache_neg_init(ncp);
2319 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
2320 ncp_ts->nc_time = *tsp;
2321 ncp_ts->nc_ticks = ticks;
2322 ncp_ts->nc_nc.nc_flag |= NCF_TS;
2324 ncp_ts->nc_dotdottime = *dtsp;
2325 ncp_ts->nc_nc.nc_flag |= NCF_DTS;
2328 len = ncp->nc_nlen = cnp->cn_namelen;
2329 hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2330 memcpy(ncp->nc_name, cnp->cn_nameptr, len);
2331 ncp->nc_name[len] = '\0';
2332 cache_enter_lock(&cel, dvp, vp, hash);
2335 * See if this vnode or negative entry is already in the cache
2336 * with this name. This can happen with concurrent lookups of
2337 * the same path name.
2339 ncpp = NCHHASH(hash);
2340 CK_SLIST_FOREACH(n2, ncpp, nc_hash) {
2341 if (n2->nc_dvp == dvp &&
2342 n2->nc_nlen == cnp->cn_namelen &&
2343 !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
2344 MPASS(cache_ncp_canuse(n2));
2345 if ((n2->nc_flag & NCF_NEGATIVE) != 0)
2347 ("%s: found entry pointing to a different vnode (%p != %p)",
2348 __func__, NULL, vp));
2350 KASSERT(n2->nc_vp == vp,
2351 ("%s: found entry pointing to a different vnode (%p != %p)",
2352 __func__, n2->nc_vp, vp));
2354 * Entries are supposed to be immutable unless in the
2355 * process of getting destroyed. Accommodating for
2356 * changing timestamps is possible but not worth it.
2357 * This should be harmless in terms of correctness, in
2358 * the worst case resulting in an earlier expiration.
2359 * Alternatively, the found entry can be replaced
2362 MPASS((n2->nc_flag & (NCF_TS | NCF_DTS)) == (ncp->nc_flag & (NCF_TS | NCF_DTS)));
2365 KASSERT((n2->nc_flag & NCF_TS) != 0,
2367 n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
2368 n2_ts->nc_time = ncp_ts->nc_time;
2369 n2_ts->nc_ticks = ncp_ts->nc_ticks;
2371 n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
2372 n2_ts->nc_nc.nc_flag |= NCF_DTS;
2376 SDT_PROBE3(vfs, namecache, enter, duplicate, dvp, ncp->nc_name,
2378 goto out_unlock_free;
2382 if (flag == NCF_ISDOTDOT) {
2384 * See if we are trying to add .. entry, but some other lookup
2385 * has populated v_cache_dd pointer already.
2387 if (dvp->v_cache_dd != NULL)
2388 goto out_unlock_free;
2389 KASSERT(vp == NULL || vp->v_type == VDIR,
2390 ("wrong vnode type %p", vp));
2391 vn_seqc_write_begin(dvp);
2392 dvp->v_cache_dd = ncp;
2393 vn_seqc_write_end(dvp);
2397 if (flag != NCF_ISDOTDOT) {
2399 * For this case, the cache entry maps both the
2400 * directory name in it and the name ".." for the
2401 * directory's parent.
2403 vn_seqc_write_begin(vp);
2404 if ((ndd = vp->v_cache_dd) != NULL) {
2405 if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
2406 cache_zap_locked(ndd);
2410 vp->v_cache_dd = ncp;
2411 vn_seqc_write_end(vp);
2412 } else if (vp->v_type != VDIR) {
2413 if (vp->v_cache_dd != NULL) {
2414 vn_seqc_write_begin(vp);
2415 vp->v_cache_dd = NULL;
2416 vn_seqc_write_end(vp);
2421 if (flag != NCF_ISDOTDOT) {
2422 if (LIST_EMPTY(&dvp->v_cache_src)) {
2423 cache_hold_vnode(dvp);
2425 LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
2429 * If the entry is "negative", we place it into the
2430 * "negative" cache queue, otherwise, we place it into the
2431 * destination vnode's cache entries queue.
2434 TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
2435 SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
2438 if (cnp->cn_flags & ISWHITEOUT)
2439 ncp->nc_flag |= NCF_WHITE;
2440 cache_neg_insert(ncp);
2441 SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
2446 * Insert the new namecache entry into the appropriate chain
2447 * within the cache entries table.
2449 CK_SLIST_INSERT_HEAD(ncpp, ncp, nc_hash);
2451 atomic_thread_fence_rel();
2453 * Mark the entry as fully constructed.
2454 * It is immutable past this point until its removal.
2456 atomic_store_char(&ncp->nc_flag, ncp->nc_flag & ~NCF_WIP);
2458 cache_enter_unlock(&cel);
2463 cache_enter_unlock(&cel);
2469 cache_roundup_2(u_int val)
2473 for (res = 1; res <= val; res <<= 1)
2479 static struct nchashhead *
2480 nchinittbl(u_long elements, u_long *hashmask)
2482 struct nchashhead *hashtbl;
2485 hashsize = cache_roundup_2(elements) / 2;
2487 hashtbl = malloc((u_long)hashsize * sizeof(*hashtbl), M_VFSCACHE, M_WAITOK);
2488 for (i = 0; i < hashsize; i++)
2489 CK_SLIST_INIT(&hashtbl[i]);
2490 *hashmask = hashsize - 1;
2495 ncfreetbl(struct nchashhead *hashtbl)
2498 free(hashtbl, M_VFSCACHE);
2502 * Name cache initialization, from vfs_init() when we are booting
2505 nchinit(void *dummy __unused)
2509 cache_zone_small = uma_zcreate("S VFS Cache", CACHE_ZONE_SMALL_SIZE,
2510 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2511 cache_zone_small_ts = uma_zcreate("STS VFS Cache", CACHE_ZONE_SMALL_TS_SIZE,
2512 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2513 cache_zone_large = uma_zcreate("L VFS Cache", CACHE_ZONE_LARGE_SIZE,
2514 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2515 cache_zone_large_ts = uma_zcreate("LTS VFS Cache", CACHE_ZONE_LARGE_TS_SIZE,
2516 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2518 VFS_SMR_ZONE_SET(cache_zone_small);
2519 VFS_SMR_ZONE_SET(cache_zone_small_ts);
2520 VFS_SMR_ZONE_SET(cache_zone_large);
2521 VFS_SMR_ZONE_SET(cache_zone_large_ts);
2523 ncsize = desiredvnodes * ncsizefactor;
2524 cache_recalc_neg_min(ncnegminpct);
2525 nchashtbl = nchinittbl(desiredvnodes * 2, &nchash);
2526 ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
2527 if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
2529 if (ncbuckethash > nchash)
2530 ncbuckethash = nchash;
2531 bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
2533 for (i = 0; i < numbucketlocks; i++)
2534 mtx_init(&bucketlocks[i], "ncbuc", NULL, MTX_DUPOK | MTX_RECURSE);
2535 ncvnodehash = ncbuckethash;
2536 vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
2538 for (i = 0; i < numvnodelocks; i++)
2539 mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
2541 for (i = 0; i < numneglists; i++) {
2542 mtx_init(&neglists[i].nl_evict_lock, "ncnege", NULL, MTX_DEF);
2543 mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
2544 TAILQ_INIT(&neglists[i].nl_list);
2545 TAILQ_INIT(&neglists[i].nl_hotlist);
2548 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
2551 cache_vnode_init(struct vnode *vp)
2554 LIST_INIT(&vp->v_cache_src);
2555 TAILQ_INIT(&vp->v_cache_dst);
2556 vp->v_cache_dd = NULL;
2561 cache_changesize(u_long newmaxvnodes)
2563 struct nchashhead *new_nchashtbl, *old_nchashtbl;
2564 u_long new_nchash, old_nchash;
2565 struct namecache *ncp;
2570 newncsize = newmaxvnodes * ncsizefactor;
2571 newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
2572 if (newmaxvnodes < numbucketlocks)
2573 newmaxvnodes = numbucketlocks;
2575 new_nchashtbl = nchinittbl(newmaxvnodes, &new_nchash);
2576 /* If same hash table size, nothing to do */
2577 if (nchash == new_nchash) {
2578 ncfreetbl(new_nchashtbl);
2582 * Move everything from the old hash table to the new table.
2583 * None of the namecache entries in the table can be removed
2584 * because to do so, they have to be removed from the hash table.
2586 cache_lock_all_vnodes();
2587 cache_lock_all_buckets();
2588 old_nchashtbl = nchashtbl;
2589 old_nchash = nchash;
2590 nchashtbl = new_nchashtbl;
2591 nchash = new_nchash;
2592 for (i = 0; i <= old_nchash; i++) {
2593 while ((ncp = CK_SLIST_FIRST(&old_nchashtbl[i])) != NULL) {
2594 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
2596 CK_SLIST_REMOVE(&old_nchashtbl[i], ncp, namecache, nc_hash);
2597 CK_SLIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
2601 cache_recalc_neg_min(ncnegminpct);
2602 cache_unlock_all_buckets();
2603 cache_unlock_all_vnodes();
2604 ncfreetbl(old_nchashtbl);
2608 * Invalidate all entries from and to a particular vnode.
2611 cache_purge_impl(struct vnode *vp)
2613 struct cache_freebatch batch;
2614 struct namecache *ncp;
2615 struct mtx *vlp, *vlp2;
2618 vlp = VP2VNODELOCK(vp);
2622 while (!LIST_EMPTY(&vp->v_cache_src)) {
2623 ncp = LIST_FIRST(&vp->v_cache_src);
2624 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2626 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2628 while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2629 ncp = TAILQ_FIRST(&vp->v_cache_dst);
2630 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2632 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2634 ncp = vp->v_cache_dd;
2636 KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2637 ("lost dotdot link"));
2638 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2640 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2642 KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2646 cache_free_batch(&batch);
2650 * Opportunistic check to see if there is anything to do.
2653 cache_has_entries(struct vnode *vp)
2656 if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2657 vp->v_cache_dd == NULL)
2663 cache_purge(struct vnode *vp)
2666 SDT_PROBE1(vfs, namecache, purge, done, vp);
2667 if (!cache_has_entries(vp))
2669 cache_purge_impl(vp);
2673 * Only to be used by vgone.
2676 cache_purge_vgone(struct vnode *vp)
2680 VNPASS(VN_IS_DOOMED(vp), vp);
2681 if (cache_has_entries(vp)) {
2682 cache_purge_impl(vp);
2687 * Serialize against a potential thread doing cache_purge.
2689 vlp = VP2VNODELOCK(vp);
2690 mtx_wait_unlocked(vlp);
2691 if (cache_has_entries(vp)) {
2692 cache_purge_impl(vp);
2699 * Invalidate all negative entries for a particular directory vnode.
2702 cache_purge_negative(struct vnode *vp)
2704 struct cache_freebatch batch;
2705 struct namecache *ncp, *nnp;
2708 SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2709 if (LIST_EMPTY(&vp->v_cache_src))
2712 vlp = VP2VNODELOCK(vp);
2714 LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2715 if (!(ncp->nc_flag & NCF_NEGATIVE))
2717 cache_zap_negative_locked_vnode_kl(ncp, vp);
2718 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2721 cache_free_batch(&batch);
2725 * Entry points for modifying VOP operations.
2728 cache_vop_rename(struct vnode *fdvp, struct vnode *fvp, struct vnode *tdvp,
2729 struct vnode *tvp, struct componentname *fcnp, struct componentname *tcnp)
2732 ASSERT_VOP_IN_SEQC(fdvp);
2733 ASSERT_VOP_IN_SEQC(fvp);
2734 ASSERT_VOP_IN_SEQC(tdvp);
2736 ASSERT_VOP_IN_SEQC(tvp);
2741 KASSERT(!cache_remove_cnp(tdvp, tcnp),
2742 ("%s: lingering negative entry", __func__));
2744 cache_remove_cnp(tdvp, tcnp);
2750 * Historically renaming was always purging all revelang entries,
2751 * but that's quite wasteful. In particular turns out that in many cases
2752 * the target file is immediately accessed after rename, inducing a cache
2755 * Recode this to reduce relocking and reuse the existing entry (if any)
2756 * instead of just removing it above and allocating a new one here.
2758 if (cache_rename_add) {
2759 cache_enter(tdvp, fvp, tcnp);
2764 cache_vop_rmdir(struct vnode *dvp, struct vnode *vp)
2767 ASSERT_VOP_IN_SEQC(dvp);
2768 ASSERT_VOP_IN_SEQC(vp);
2774 * Validate that if an entry exists it matches.
2777 cache_validate(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
2779 struct namecache *ncp;
2783 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
2784 if (CK_SLIST_EMPTY(NCHHASH(hash)))
2786 blp = HASH2BUCKETLOCK(hash);
2788 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
2789 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
2790 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) {
2791 if (ncp->nc_vp != vp)
2792 panic("%s: mismatch (%p != %p); ncp %p [%s] dvp %p vp %p\n",
2793 __func__, vp, ncp->nc_vp, ncp, ncp->nc_name, ncp->nc_dvp,
2802 * Flush all entries referencing a particular filesystem.
2805 cache_purgevfs(struct mount *mp)
2807 struct vnode *vp, *mvp;
2809 SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2811 * Somewhat wasteful iteration over all vnodes. Would be better to
2812 * support filtering and avoid the interlock to begin with.
2814 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2815 if (!cache_has_entries(vp)) {
2827 * Perform canonical checks and cache lookup and pass on to filesystem
2828 * through the vop_cachedlookup only if needed.
2832 vfs_cache_lookup(struct vop_lookup_args *ap)
2836 struct vnode **vpp = ap->a_vpp;
2837 struct componentname *cnp = ap->a_cnp;
2838 int flags = cnp->cn_flags;
2843 if (dvp->v_type != VDIR)
2846 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2847 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2850 error = vn_dir_check_exec(dvp, cnp);
2854 error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2856 return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2862 /* Implementation of the getcwd syscall. */
2864 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2870 buflen = uap->buflen;
2871 if (__predict_false(buflen < 2))
2873 if (buflen > MAXPATHLEN)
2874 buflen = MAXPATHLEN;
2876 buf = uma_zalloc(namei_zone, M_WAITOK);
2877 error = vn_getcwd(buf, &retbuf, &buflen);
2879 error = copyout(retbuf, uap->buf, buflen);
2880 uma_zfree(namei_zone, buf);
2885 vn_getcwd(char *buf, char **retbuf, size_t *buflen)
2891 pwd = pwd_get_smr();
2892 error = vn_fullpath_any_smr(pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf,
2894 VFS_SMR_ASSERT_NOT_ENTERED();
2896 pwd = pwd_hold(curthread);
2897 error = vn_fullpath_any(pwd->pwd_cdir, pwd->pwd_rdir, buf,
2903 if (KTRPOINT(curthread, KTR_NAMEI) && error == 0)
2910 kern___realpathat(struct thread *td, int fd, const char *path, char *buf,
2911 size_t size, int flags, enum uio_seg pathseg)
2913 struct nameidata nd;
2914 char *retbuf, *freebuf;
2919 NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | SAVENAME | WANTPARENT | AUDITVNODE1,
2920 pathseg, path, fd, &cap_fstat_rights, td);
2921 if ((error = namei(&nd)) != 0)
2923 error = vn_fullpath_hardlink(&nd, &retbuf, &freebuf, &size);
2925 error = copyout(retbuf, buf, size);
2926 free(freebuf, M_TEMP);
2933 sys___realpathat(struct thread *td, struct __realpathat_args *uap)
2936 return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size,
2937 uap->flags, UIO_USERSPACE));
2941 * Retrieve the full filesystem path that correspond to a vnode from the name
2942 * cache (if available)
2945 vn_fullpath(struct vnode *vp, char **retbuf, char **freebuf)
2952 if (__predict_false(vp == NULL))
2955 buflen = MAXPATHLEN;
2956 buf = malloc(buflen, M_TEMP, M_WAITOK);
2958 pwd = pwd_get_smr();
2959 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, &buflen, 0);
2960 VFS_SMR_ASSERT_NOT_ENTERED();
2962 pwd = pwd_hold(curthread);
2963 error = vn_fullpath_any(vp, pwd->pwd_rdir, buf, retbuf, &buflen);
2974 * This function is similar to vn_fullpath, but it attempts to lookup the
2975 * pathname relative to the global root mount point. This is required for the
2976 * auditing sub-system, as audited pathnames must be absolute, relative to the
2977 * global root mount point.
2980 vn_fullpath_global(struct vnode *vp, char **retbuf, char **freebuf)
2986 if (__predict_false(vp == NULL))
2988 buflen = MAXPATHLEN;
2989 buf = malloc(buflen, M_TEMP, M_WAITOK);
2991 error = vn_fullpath_any_smr(vp, rootvnode, buf, retbuf, &buflen, 0);
2992 VFS_SMR_ASSERT_NOT_ENTERED();
2994 error = vn_fullpath_any(vp, rootvnode, buf, retbuf, &buflen);
3003 static struct namecache *
3004 vn_dd_from_dst(struct vnode *vp)
3006 struct namecache *ncp;
3008 cache_assert_vnode_locked(vp);
3009 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) {
3010 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
3017 vn_vptocnp(struct vnode **vp, char *buf, size_t *buflen)
3020 struct namecache *ncp;
3024 vlp = VP2VNODELOCK(*vp);
3026 ncp = (*vp)->v_cache_dd;
3027 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT) == 0) {
3028 KASSERT(ncp == vn_dd_from_dst(*vp),
3029 ("%s: mismatch for dd entry (%p != %p)", __func__,
3030 ncp, vn_dd_from_dst(*vp)));
3032 ncp = vn_dd_from_dst(*vp);
3035 if (*buflen < ncp->nc_nlen) {
3038 counter_u64_add(numfullpathfail4, 1);
3040 SDT_PROBE3(vfs, namecache, fullpath, return, error,
3044 *buflen -= ncp->nc_nlen;
3045 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
3046 SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
3055 SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
3058 vn_lock(*vp, LK_SHARED | LK_RETRY);
3059 error = VOP_VPTOCNP(*vp, &dvp, buf, buflen);
3062 counter_u64_add(numfullpathfail2, 1);
3063 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
3068 if (VN_IS_DOOMED(dvp)) {
3069 /* forced unmount */
3072 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
3076 * *vp has its use count incremented still.
3083 * Resolve a directory to a pathname.
3085 * The name of the directory can always be found in the namecache or fetched
3086 * from the filesystem. There is also guaranteed to be only one parent, meaning
3087 * we can just follow vnodes up until we find the root.
3089 * The vnode must be referenced.
3092 vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3093 size_t *len, size_t addend)
3095 #ifdef KDTRACE_HOOKS
3096 struct vnode *startvp = vp;
3101 bool slash_prefixed;
3103 VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp);
3104 VNPASS(vp->v_usecount > 0, vp);
3108 slash_prefixed = true;
3113 slash_prefixed = false;
3118 SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
3119 counter_u64_add(numfullpathcalls, 1);
3120 while (vp != rdir && vp != rootvnode) {
3122 * The vp vnode must be already fully constructed,
3123 * since it is either found in namecache or obtained
3124 * from VOP_VPTOCNP(). We may test for VV_ROOT safely
3125 * without obtaining the vnode lock.
3127 if ((vp->v_vflag & VV_ROOT) != 0) {
3128 vn_lock(vp, LK_RETRY | LK_SHARED);
3131 * With the vnode locked, check for races with
3132 * unmount, forced or not. Note that we
3133 * already verified that vp is not equal to
3134 * the root vnode, which means that
3135 * mnt_vnodecovered can be NULL only for the
3138 if (VN_IS_DOOMED(vp) ||
3139 (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
3140 vp1->v_mountedhere != vp->v_mount) {
3143 SDT_PROBE3(vfs, namecache, fullpath, return,
3153 if (vp->v_type != VDIR) {
3155 counter_u64_add(numfullpathfail1, 1);
3157 SDT_PROBE3(vfs, namecache, fullpath, return,
3161 error = vn_vptocnp(&vp, buf, &buflen);
3167 SDT_PROBE3(vfs, namecache, fullpath, return, error,
3171 buf[--buflen] = '/';
3172 slash_prefixed = true;
3176 if (!slash_prefixed) {
3179 counter_u64_add(numfullpathfail4, 1);
3180 SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
3184 buf[--buflen] = '/';
3186 counter_u64_add(numfullpathfound, 1);
3189 *retbuf = buf + buflen;
3190 SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf);
3197 * Resolve an arbitrary vnode to a pathname.
3200 * - hardlinks are not tracked, thus if the vnode is not a directory this can
3201 * resolve to a different path than the one used to find it
3202 * - namecache is not mandatory, meaning names are not guaranteed to be added
3203 * (in which case resolving fails)
3205 static void __inline
3206 cache_rev_failed_impl(int *reason, int line)
3211 #define cache_rev_failed(var) cache_rev_failed_impl((var), __LINE__)
3214 vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
3215 char **retbuf, size_t *buflen, size_t addend)
3217 #ifdef KDTRACE_HOOKS
3218 struct vnode *startvp = vp;
3222 struct namecache *ncp;
3226 #ifdef KDTRACE_HOOKS
3229 seqc_t vp_seqc, tvp_seqc;
3232 VFS_SMR_ASSERT_ENTERED();
3234 if (!cache_fast_revlookup) {
3239 orig_buflen = *buflen;
3242 MPASS(*buflen >= 2);
3244 buf[*buflen] = '\0';
3247 if (vp == rdir || vp == rootvnode) {
3255 #ifdef KDTRACE_HOOKS
3259 ncp = NULL; /* for sdt probe down below */
3260 vp_seqc = vn_seqc_read_any(vp);
3261 if (seqc_in_modify(vp_seqc)) {
3262 cache_rev_failed(&reason);
3267 #ifdef KDTRACE_HOOKS
3270 if ((vp->v_vflag & VV_ROOT) != 0) {
3271 mp = atomic_load_ptr(&vp->v_mount);
3273 cache_rev_failed(&reason);
3276 tvp = atomic_load_ptr(&mp->mnt_vnodecovered);
3277 tvp_seqc = vn_seqc_read_any(tvp);
3278 if (seqc_in_modify(tvp_seqc)) {
3279 cache_rev_failed(&reason);
3282 if (!vn_seqc_consistent(vp, vp_seqc)) {
3283 cache_rev_failed(&reason);
3290 ncp = atomic_load_ptr(&vp->v_cache_dd);
3292 cache_rev_failed(&reason);
3295 nc_flag = atomic_load_char(&ncp->nc_flag);
3296 if ((nc_flag & NCF_ISDOTDOT) != 0) {
3297 cache_rev_failed(&reason);
3300 if (!cache_ncp_canuse(ncp)) {
3301 cache_rev_failed(&reason);
3304 if (ncp->nc_nlen >= *buflen) {
3305 cache_rev_failed(&reason);
3309 *buflen -= ncp->nc_nlen;
3310 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
3314 tvp_seqc = vn_seqc_read_any(tvp);
3315 if (seqc_in_modify(tvp_seqc)) {
3316 cache_rev_failed(&reason);
3319 if (!vn_seqc_consistent(vp, vp_seqc)) {
3320 cache_rev_failed(&reason);
3325 if (vp == rdir || vp == rootvnode)
3330 *retbuf = buf + *buflen;
3331 *buflen = orig_buflen - *buflen + addend;
3332 SDT_PROBE2(vfs, namecache, fullpath_smr, hit, startvp, *retbuf);
3336 *buflen = orig_buflen;
3337 SDT_PROBE4(vfs, namecache, fullpath_smr, miss, startvp, ncp, reason, i);
3343 vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3346 size_t orig_buflen, addend;
3352 orig_buflen = *buflen;
3356 if (vp->v_type != VDIR) {
3358 buf[*buflen] = '\0';
3359 error = vn_vptocnp(&vp, buf, buflen);
3368 addend = orig_buflen - *buflen;
3371 return (vn_fullpath_dir(vp, rdir, buf, retbuf, buflen, addend));
3375 * Resolve an arbitrary vnode to a pathname (taking care of hardlinks).
3377 * Since the namecache does not track hardlinks, the caller is expected to first
3378 * look up the target vnode with SAVENAME | WANTPARENT flags passed to namei.
3380 * Then we have 2 cases:
3381 * - if the found vnode is a directory, the path can be constructed just by
3382 * following names up the chain
3383 * - otherwise we populate the buffer with the saved name and start resolving
3387 vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf, char **freebuf,
3392 struct componentname *cnp;
3400 if (*buflen > MAXPATHLEN)
3401 *buflen = MAXPATHLEN;
3403 buf = malloc(*buflen, M_TEMP, M_WAITOK);
3408 * Check for VBAD to work around the vp_crossmp bug in lookup().
3410 * For example consider tmpfs on /tmp and realpath /tmp. ni_vp will be
3411 * set to mount point's root vnode while ni_dvp will be vp_crossmp.
3412 * If the type is VDIR (like in this very case) we can skip looking
3413 * at ni_dvp in the first place. However, since vnodes get passed here
3414 * unlocked the target may transition to doomed state (type == VBAD)
3415 * before we get to evaluate the condition. If this happens, we will
3416 * populate part of the buffer and descend to vn_fullpath_dir with
3417 * vp == vp_crossmp. Prevent the problem by checking for VBAD.
3419 * This should be atomic_load(&vp->v_type) but it is illegal to take
3420 * an address of a bit field, even if said field is sized to char.
3421 * Work around the problem by reading the value into a full-sized enum
3422 * and then re-reading it with atomic_load which will still prevent
3423 * the compiler from re-reading down the road.
3426 type = atomic_load_int(&type);
3433 addend = cnp->cn_namelen + 2;
3434 if (*buflen < addend) {
3439 tmpbuf = buf + *buflen;
3441 memcpy(&tmpbuf[1], cnp->cn_nameptr, cnp->cn_namelen);
3442 tmpbuf[addend - 1] = '\0';
3447 pwd = pwd_get_smr();
3448 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3450 VFS_SMR_ASSERT_NOT_ENTERED();
3452 pwd = pwd_hold(curthread);
3454 error = vn_fullpath_dir(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3470 vn_dir_dd_ino(struct vnode *vp)
3472 struct namecache *ncp;
3477 ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
3478 vlp = VP2VNODELOCK(vp);
3480 TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
3481 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
3484 vs = vget_prep(ddvp);
3486 if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
3495 vn_commname(struct vnode *vp, char *buf, u_int buflen)
3497 struct namecache *ncp;
3501 vlp = VP2VNODELOCK(vp);
3503 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
3504 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
3510 l = min(ncp->nc_nlen, buflen - 1);
3511 memcpy(buf, ncp->nc_name, l);
3518 * This function updates path string to vnode's full global path
3519 * and checks the size of the new path string against the pathlen argument.
3521 * Requires a locked, referenced vnode.
3522 * Vnode is re-locked on success or ENODEV, otherwise unlocked.
3524 * If vp is a directory, the call to vn_fullpath_global() always succeeds
3525 * because it falls back to the ".." lookup if the namecache lookup fails.
3528 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
3531 struct nameidata nd;
3536 ASSERT_VOP_ELOCKED(vp, __func__);
3538 /* Construct global filesystem path from vp. */
3540 error = vn_fullpath_global(vp, &rpath, &fbuf);
3547 if (strlen(rpath) >= pathlen) {
3549 error = ENAMETOOLONG;
3554 * Re-lookup the vnode by path to detect a possible rename.
3555 * As a side effect, the vnode is relocked.
3556 * If vnode was renamed, return ENOENT.
3558 NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
3559 UIO_SYSSPACE, path, td);
3565 NDFREE(&nd, NDF_ONLY_PNBUF);
3569 strcpy(path, rpath);
3582 db_print_vpath(struct vnode *vp)
3585 while (vp != NULL) {
3586 db_printf("%p: ", vp);
3587 if (vp == rootvnode) {
3591 if (vp->v_vflag & VV_ROOT) {
3592 db_printf("<mount point>");
3593 vp = vp->v_mount->mnt_vnodecovered;
3595 struct namecache *ncp;
3599 ncp = TAILQ_FIRST(&vp->v_cache_dst);
3602 for (i = 0; i < ncp->nc_nlen; i++)
3603 db_printf("%c", *ncn++);
3616 DB_SHOW_COMMAND(vpath, db_show_vpath)
3621 db_printf("usage: show vpath <struct vnode *>\n");
3625 vp = (struct vnode *)addr;
3631 static int cache_fast_lookup = 1;
3632 static char __read_frequently cache_fast_lookup_enabled = true;
3634 #define CACHE_FPL_FAILED -2020
3637 cache_fast_lookup_enabled_recalc(void)
3643 mac_on = mac_vnode_check_lookup_enabled();
3644 mac_on |= mac_vnode_check_readlink_enabled();
3649 lookup_flag = atomic_load_int(&cache_fast_lookup);
3650 if (lookup_flag && !mac_on) {
3651 atomic_store_char(&cache_fast_lookup_enabled, true);
3653 atomic_store_char(&cache_fast_lookup_enabled, false);
3658 syscal_vfs_cache_fast_lookup(SYSCTL_HANDLER_ARGS)
3662 old = atomic_load_int(&cache_fast_lookup);
3663 error = sysctl_handle_int(oidp, arg1, arg2, req);
3664 if (error == 0 && req->newptr && old != atomic_load_int(&cache_fast_lookup))
3665 cache_fast_lookup_enabled_recalc();
3668 SYSCTL_PROC(_vfs, OID_AUTO, cache_fast_lookup, CTLTYPE_INT|CTLFLAG_RW|CTLFLAG_MPSAFE,
3669 &cache_fast_lookup, 0, syscal_vfs_cache_fast_lookup, "IU", "");
3672 * Components of nameidata (or objects it can point to) which may
3673 * need restoring in case fast path lookup fails.
3675 struct nameidata_outer {
3680 struct nameidata_saved {
3688 struct cache_fpl_debug {
3694 struct nameidata *ndp;
3695 struct componentname *cnp;
3702 struct nameidata_saved snd;
3703 struct nameidata_outer snd_outer;
3705 enum cache_fpl_status status:8;
3710 struct cache_fpl_debug debug;
3714 static bool cache_fplookup_is_mp(struct cache_fpl *fpl);
3715 static int cache_fplookup_cross_mount(struct cache_fpl *fpl);
3716 static int cache_fplookup_partial_setup(struct cache_fpl *fpl);
3717 static int cache_fplookup_skip_slashes(struct cache_fpl *fpl);
3718 static int cache_fplookup_preparse(struct cache_fpl *fpl);
3719 static void cache_fpl_pathlen_dec(struct cache_fpl *fpl);
3720 static void cache_fpl_pathlen_inc(struct cache_fpl *fpl);
3721 static void cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n);
3722 static void cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n);
3725 cache_fpl_cleanup_cnp(struct componentname *cnp)
3728 uma_zfree(namei_zone, cnp->cn_pnbuf);
3730 cnp->cn_pnbuf = NULL;
3731 cnp->cn_nameptr = NULL;
3735 static struct vnode *
3736 cache_fpl_handle_root(struct cache_fpl *fpl)
3738 struct nameidata *ndp;
3739 struct componentname *cnp;
3744 MPASS(*(cnp->cn_nameptr) == '/');
3746 cache_fpl_pathlen_dec(fpl);
3748 if (__predict_false(*(cnp->cn_nameptr) == '/')) {
3751 cache_fpl_pathlen_dec(fpl);
3752 } while (*(cnp->cn_nameptr) == '/');
3755 return (ndp->ni_rootdir);
3759 cache_fpl_checkpoint_outer(struct cache_fpl *fpl)
3762 fpl->snd_outer.ni_pathlen = fpl->ndp->ni_pathlen;
3763 fpl->snd_outer.cn_flags = fpl->ndp->ni_cnd.cn_flags;
3767 cache_fpl_checkpoint(struct cache_fpl *fpl)
3771 fpl->snd.cn_nameptr = fpl->ndp->ni_cnd.cn_nameptr;
3772 fpl->snd.ni_pathlen = fpl->debug.ni_pathlen;
3777 cache_fpl_restore_partial(struct cache_fpl *fpl)
3780 fpl->ndp->ni_cnd.cn_flags = fpl->snd_outer.cn_flags;
3782 fpl->debug.ni_pathlen = fpl->snd.ni_pathlen;
3787 cache_fpl_restore_abort(struct cache_fpl *fpl)
3790 cache_fpl_restore_partial(fpl);
3792 * It is 0 on entry by API contract.
3794 fpl->ndp->ni_resflags = 0;
3795 fpl->ndp->ni_cnd.cn_nameptr = fpl->ndp->ni_cnd.cn_pnbuf;
3796 fpl->ndp->ni_pathlen = fpl->snd_outer.ni_pathlen;
3800 #define cache_fpl_smr_assert_entered(fpl) ({ \
3801 struct cache_fpl *_fpl = (fpl); \
3802 MPASS(_fpl->in_smr == true); \
3803 VFS_SMR_ASSERT_ENTERED(); \
3805 #define cache_fpl_smr_assert_not_entered(fpl) ({ \
3806 struct cache_fpl *_fpl = (fpl); \
3807 MPASS(_fpl->in_smr == false); \
3808 VFS_SMR_ASSERT_NOT_ENTERED(); \
3811 cache_fpl_assert_status(struct cache_fpl *fpl)
3814 switch (fpl->status) {
3815 case CACHE_FPL_STATUS_UNSET:
3816 __assert_unreachable();
3818 case CACHE_FPL_STATUS_DESTROYED:
3819 case CACHE_FPL_STATUS_ABORTED:
3820 case CACHE_FPL_STATUS_PARTIAL:
3821 case CACHE_FPL_STATUS_HANDLED:
3826 #define cache_fpl_smr_assert_entered(fpl) do { } while (0)
3827 #define cache_fpl_smr_assert_not_entered(fpl) do { } while (0)
3828 #define cache_fpl_assert_status(fpl) do { } while (0)
3831 #define cache_fpl_smr_enter_initial(fpl) ({ \
3832 struct cache_fpl *_fpl = (fpl); \
3834 _fpl->in_smr = true; \
3837 #define cache_fpl_smr_enter(fpl) ({ \
3838 struct cache_fpl *_fpl = (fpl); \
3839 MPASS(_fpl->in_smr == false); \
3841 _fpl->in_smr = true; \
3844 #define cache_fpl_smr_exit(fpl) ({ \
3845 struct cache_fpl *_fpl = (fpl); \
3846 MPASS(_fpl->in_smr == true); \
3848 _fpl->in_smr = false; \
3852 cache_fpl_aborted_early_impl(struct cache_fpl *fpl, int line)
3855 if (fpl->status != CACHE_FPL_STATUS_UNSET) {
3856 KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL,
3857 ("%s: converting to abort from %d at %d, set at %d\n",
3858 __func__, fpl->status, line, fpl->line));
3860 cache_fpl_smr_assert_not_entered(fpl);
3861 fpl->status = CACHE_FPL_STATUS_ABORTED;
3863 return (CACHE_FPL_FAILED);
3866 #define cache_fpl_aborted_early(x) cache_fpl_aborted_early_impl((x), __LINE__)
3868 static int __noinline
3869 cache_fpl_aborted_impl(struct cache_fpl *fpl, int line)
3871 struct nameidata *ndp;
3872 struct componentname *cnp;
3877 if (fpl->status != CACHE_FPL_STATUS_UNSET) {
3878 KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL,
3879 ("%s: converting to abort from %d at %d, set at %d\n",
3880 __func__, fpl->status, line, fpl->line));
3882 fpl->status = CACHE_FPL_STATUS_ABORTED;
3885 cache_fpl_smr_exit(fpl);
3886 cache_fpl_restore_abort(fpl);
3888 * Resolving symlinks overwrites data passed by the caller.
3891 if (ndp->ni_loopcnt > 0) {
3892 fpl->status = CACHE_FPL_STATUS_DESTROYED;
3893 cache_fpl_cleanup_cnp(cnp);
3895 return (CACHE_FPL_FAILED);
3898 #define cache_fpl_aborted(x) cache_fpl_aborted_impl((x), __LINE__)
3900 static int __noinline
3901 cache_fpl_partial_impl(struct cache_fpl *fpl, int line)
3904 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3905 ("%s: setting to partial at %d, but already set to %d at %d\n",
3906 __func__, line, fpl->status, fpl->line));
3907 cache_fpl_smr_assert_entered(fpl);
3908 fpl->status = CACHE_FPL_STATUS_PARTIAL;
3910 return (cache_fplookup_partial_setup(fpl));
3913 #define cache_fpl_partial(x) cache_fpl_partial_impl((x), __LINE__)
3916 cache_fpl_handled_impl(struct cache_fpl *fpl, int line)
3919 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3920 ("%s: setting to handled at %d, but already set to %d at %d\n",
3921 __func__, line, fpl->status, fpl->line));
3922 cache_fpl_smr_assert_not_entered(fpl);
3923 fpl->status = CACHE_FPL_STATUS_HANDLED;
3928 #define cache_fpl_handled(x) cache_fpl_handled_impl((x), __LINE__)
3931 cache_fpl_handled_error_impl(struct cache_fpl *fpl, int error, int line)
3934 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3935 ("%s: setting to handled at %d, but already set to %d at %d\n",
3936 __func__, line, fpl->status, fpl->line));
3938 MPASS(error != CACHE_FPL_FAILED);
3939 cache_fpl_smr_assert_not_entered(fpl);
3940 fpl->status = CACHE_FPL_STATUS_HANDLED;
3944 fpl->savename = false;
3948 #define cache_fpl_handled_error(x, e) cache_fpl_handled_error_impl((x), (e), __LINE__)
3951 cache_fpl_terminated(struct cache_fpl *fpl)
3954 return (fpl->status != CACHE_FPL_STATUS_UNSET);
3957 #define CACHE_FPL_SUPPORTED_CN_FLAGS \
3958 (NC_NOMAKEENTRY | NC_KEEPPOSENTRY | LOCKLEAF | LOCKPARENT | WANTPARENT | \
3959 FAILIFEXISTS | FOLLOW | LOCKSHARED | SAVENAME | SAVESTART | WILLBEDIR | \
3960 ISOPEN | NOMACCHECK | AUDITVNODE1 | AUDITVNODE2 | NOCAPCHECK)
3962 #define CACHE_FPL_INTERNAL_CN_FLAGS \
3963 (ISDOTDOT | MAKEENTRY | ISLASTCN)
3965 _Static_assert((CACHE_FPL_SUPPORTED_CN_FLAGS & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
3966 "supported and internal flags overlap");
3969 cache_fpl_islastcn(struct nameidata *ndp)
3972 return (*ndp->ni_next == 0);
3976 cache_fpl_isdotdot(struct componentname *cnp)
3979 if (cnp->cn_namelen == 2 &&
3980 cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
3986 cache_can_fplookup(struct cache_fpl *fpl)
3988 struct nameidata *ndp;
3989 struct componentname *cnp;
3994 td = cnp->cn_thread;
3996 if (!atomic_load_char(&cache_fast_lookup_enabled)) {
3997 cache_fpl_aborted_early(fpl);
4000 if ((cnp->cn_flags & ~CACHE_FPL_SUPPORTED_CN_FLAGS) != 0) {
4001 cache_fpl_aborted_early(fpl);
4004 if (IN_CAPABILITY_MODE(td)) {
4005 cache_fpl_aborted_early(fpl);
4008 if (AUDITING_TD(td)) {
4009 cache_fpl_aborted_early(fpl);
4012 if (ndp->ni_startdir != NULL) {
4013 cache_fpl_aborted_early(fpl);
4020 cache_fplookup_dirfd(struct cache_fpl *fpl, struct vnode **vpp)
4022 struct nameidata *ndp;
4027 error = fgetvp_lookup_smr(ndp->ni_dirfd, ndp, vpp, &fsearch);
4028 if (__predict_false(error != 0)) {
4029 return (cache_fpl_aborted(fpl));
4031 fpl->fsearch = fsearch;
4035 static int __noinline
4036 cache_fplookup_negative_promote(struct cache_fpl *fpl, struct namecache *oncp,
4039 struct componentname *cnp;
4045 cache_fpl_smr_exit(fpl);
4046 if (cache_neg_promote_cond(dvp, cnp, oncp, hash))
4047 return (cache_fpl_handled_error(fpl, ENOENT));
4049 return (cache_fpl_aborted(fpl));
4053 * The target vnode is not supported, prepare for the slow path to take over.
4055 static int __noinline
4056 cache_fplookup_partial_setup(struct cache_fpl *fpl)
4058 struct nameidata *ndp;
4059 struct componentname *cnp;
4069 dvp_seqc = fpl->dvp_seqc;
4071 if (!pwd_hold_smr(pwd)) {
4072 return (cache_fpl_aborted(fpl));
4076 * Note that seqc is checked before the vnode is locked, so by
4077 * the time regular lookup gets to it it may have moved.
4079 * Ultimately this does not affect correctness, any lookup errors
4080 * are userspace racing with itself. It is guaranteed that any
4081 * path which ultimately gets found could also have been found
4082 * by regular lookup going all the way in absence of concurrent
4085 dvs = vget_prep_smr(dvp);
4086 cache_fpl_smr_exit(fpl);
4087 if (__predict_false(dvs == VGET_NONE)) {
4089 return (cache_fpl_aborted(fpl));
4092 vget_finish_ref(dvp, dvs);
4093 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4096 return (cache_fpl_aborted(fpl));
4099 cache_fpl_restore_partial(fpl);
4101 if (cnp->cn_nameptr != fpl->snd.cn_nameptr) {
4102 panic("%s: cn_nameptr mismatch (%p != %p) full [%s]\n", __func__,
4103 cnp->cn_nameptr, fpl->snd.cn_nameptr, cnp->cn_pnbuf);
4107 ndp->ni_startdir = dvp;
4108 cnp->cn_flags |= MAKEENTRY;
4109 if (cache_fpl_islastcn(ndp))
4110 cnp->cn_flags |= ISLASTCN;
4111 if (cache_fpl_isdotdot(cnp))
4112 cnp->cn_flags |= ISDOTDOT;
4115 * Skip potential extra slashes parsing did not take care of.
4116 * cache_fplookup_skip_slashes explains the mechanism.
4118 if (__predict_false(*(cnp->cn_nameptr) == '/')) {
4121 cache_fpl_pathlen_dec(fpl);
4122 } while (*(cnp->cn_nameptr) == '/');
4125 ndp->ni_pathlen = fpl->nulchar - cnp->cn_nameptr + 1;
4127 if (ndp->ni_pathlen != fpl->debug.ni_pathlen) {
4128 panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n",
4129 __func__, ndp->ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar,
4130 cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf);
4137 cache_fplookup_final_child(struct cache_fpl *fpl, enum vgetstate tvs)
4139 struct componentname *cnp;
4146 tvp_seqc = fpl->tvp_seqc;
4148 if ((cnp->cn_flags & LOCKLEAF) != 0) {
4149 lkflags = LK_SHARED;
4150 if ((cnp->cn_flags & LOCKSHARED) == 0)
4151 lkflags = LK_EXCLUSIVE;
4152 error = vget_finish(tvp, lkflags, tvs);
4153 if (__predict_false(error != 0)) {
4154 return (cache_fpl_aborted(fpl));
4157 vget_finish_ref(tvp, tvs);
4160 if (!vn_seqc_consistent(tvp, tvp_seqc)) {
4161 if ((cnp->cn_flags & LOCKLEAF) != 0)
4165 return (cache_fpl_aborted(fpl));
4168 return (cache_fpl_handled(fpl));
4172 * They want to possibly modify the state of the namecache.
4174 static int __noinline
4175 cache_fplookup_final_modifying(struct cache_fpl *fpl)
4177 struct nameidata *ndp;
4178 struct componentname *cnp;
4180 struct vnode *dvp, *tvp;
4189 dvp_seqc = fpl->dvp_seqc;
4191 MPASS(*(cnp->cn_nameptr) != '/');
4192 MPASS(cache_fpl_islastcn(ndp));
4193 if ((cnp->cn_flags & LOCKPARENT) == 0)
4194 MPASS((cnp->cn_flags & WANTPARENT) != 0);
4195 MPASS((cnp->cn_flags & TRAILINGSLASH) == 0);
4196 MPASS(cnp->cn_nameiop == CREATE || cnp->cn_nameiop == DELETE ||
4197 cnp->cn_nameiop == RENAME);
4198 MPASS((cnp->cn_flags & MAKEENTRY) == 0);
4199 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
4201 docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
4202 if (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)
4205 mp = atomic_load_ptr(&dvp->v_mount);
4206 if (__predict_false(mp == NULL)) {
4207 return (cache_fpl_aborted(fpl));
4210 if (__predict_false(mp->mnt_flag & MNT_RDONLY)) {
4211 cache_fpl_smr_exit(fpl);
4213 * Original code keeps not checking for CREATE which
4214 * might be a bug. For now let the old lookup decide.
4216 if (cnp->cn_nameiop == CREATE) {
4217 return (cache_fpl_aborted(fpl));
4219 return (cache_fpl_handled_error(fpl, EROFS));
4222 if (fpl->tvp != NULL && (cnp->cn_flags & FAILIFEXISTS) != 0) {
4223 cache_fpl_smr_exit(fpl);
4224 return (cache_fpl_handled_error(fpl, EEXIST));
4228 * Secure access to dvp; check cache_fplookup_partial_setup for
4231 * XXX At least UFS requires its lookup routine to be called for
4232 * the last path component, which leads to some level of complication
4234 * - the target routine always locks the target vnode, but our caller
4235 * may not need it locked
4236 * - some of the VOP machinery asserts that the parent is locked, which
4237 * once more may be not required
4239 * TODO: add a flag for filesystems which don't need this.
4241 dvs = vget_prep_smr(dvp);
4242 cache_fpl_smr_exit(fpl);
4243 if (__predict_false(dvs == VGET_NONE)) {
4244 return (cache_fpl_aborted(fpl));
4247 vget_finish_ref(dvp, dvs);
4248 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4250 return (cache_fpl_aborted(fpl));
4253 error = vn_lock(dvp, LK_EXCLUSIVE);
4254 if (__predict_false(error != 0)) {
4256 return (cache_fpl_aborted(fpl));
4260 cnp->cn_flags |= ISLASTCN;
4262 cnp->cn_flags |= MAKEENTRY;
4263 if (cache_fpl_isdotdot(cnp))
4264 cnp->cn_flags |= ISDOTDOT;
4265 cnp->cn_lkflags = LK_EXCLUSIVE;
4266 error = VOP_LOOKUP(dvp, &tvp, cnp);
4274 return (cache_fpl_handled_error(fpl, error));
4277 return (cache_fpl_aborted(fpl));
4281 fpl->savename = (cnp->cn_flags & SAVENAME) != 0;
4284 if ((cnp->cn_flags & SAVESTART) != 0) {
4285 ndp->ni_startdir = dvp;
4286 vrefact(ndp->ni_startdir);
4287 cnp->cn_flags |= SAVENAME;
4288 fpl->savename = true;
4290 MPASS(error == EJUSTRETURN);
4291 if ((cnp->cn_flags & LOCKPARENT) == 0) {
4294 return (cache_fpl_handled(fpl));
4298 * There are very hairy corner cases concerning various flag combinations
4299 * and locking state. In particular here we only hold one lock instead of
4302 * Skip the complexity as it is of no significance for normal workloads.
4304 if (__predict_false(tvp == dvp)) {
4307 return (cache_fpl_aborted(fpl));
4311 * Check if the target is either a symlink or a mount point.
4312 * Since we expect this to be the terminal vnode it should
4313 * almost never be true.
4315 if (__predict_false(tvp->v_type == VLNK || cache_fplookup_is_mp(fpl))) {
4318 return (cache_fpl_aborted(fpl));
4321 if ((cnp->cn_flags & FAILIFEXISTS) != 0) {
4324 return (cache_fpl_handled_error(fpl, EEXIST));
4327 if ((cnp->cn_flags & LOCKLEAF) == 0) {
4331 if ((cnp->cn_flags & LOCKPARENT) == 0) {
4335 if ((cnp->cn_flags & SAVESTART) != 0) {
4336 ndp->ni_startdir = dvp;
4337 vrefact(ndp->ni_startdir);
4338 cnp->cn_flags |= SAVENAME;
4339 fpl->savename = true;
4342 return (cache_fpl_handled(fpl));
4345 static int __noinline
4346 cache_fplookup_modifying(struct cache_fpl *fpl)
4348 struct nameidata *ndp;
4352 if (!cache_fpl_islastcn(ndp)) {
4353 return (cache_fpl_partial(fpl));
4355 return (cache_fplookup_final_modifying(fpl));
4358 static int __noinline
4359 cache_fplookup_final_withparent(struct cache_fpl *fpl)
4361 struct componentname *cnp;
4362 enum vgetstate dvs, tvs;
4363 struct vnode *dvp, *tvp;
4369 dvp_seqc = fpl->dvp_seqc;
4372 MPASS((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0);
4375 * This is less efficient than it can be for simplicity.
4377 dvs = vget_prep_smr(dvp);
4378 if (__predict_false(dvs == VGET_NONE)) {
4379 return (cache_fpl_aborted(fpl));
4381 tvs = vget_prep_smr(tvp);
4382 if (__predict_false(tvs == VGET_NONE)) {
4383 cache_fpl_smr_exit(fpl);
4384 vget_abort(dvp, dvs);
4385 return (cache_fpl_aborted(fpl));
4388 cache_fpl_smr_exit(fpl);
4390 if ((cnp->cn_flags & LOCKPARENT) != 0) {
4391 error = vget_finish(dvp, LK_EXCLUSIVE, dvs);
4392 if (__predict_false(error != 0)) {
4393 vget_abort(tvp, tvs);
4394 return (cache_fpl_aborted(fpl));
4397 vget_finish_ref(dvp, dvs);
4400 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4401 vget_abort(tvp, tvs);
4402 if ((cnp->cn_flags & LOCKPARENT) != 0)
4406 return (cache_fpl_aborted(fpl));
4409 error = cache_fplookup_final_child(fpl, tvs);
4410 if (__predict_false(error != 0)) {
4411 MPASS(fpl->status == CACHE_FPL_STATUS_ABORTED);
4412 if ((cnp->cn_flags & LOCKPARENT) != 0)
4419 MPASS(fpl->status == CACHE_FPL_STATUS_HANDLED);
4424 cache_fplookup_final(struct cache_fpl *fpl)
4426 struct componentname *cnp;
4428 struct vnode *dvp, *tvp;
4433 dvp_seqc = fpl->dvp_seqc;
4436 MPASS(*(cnp->cn_nameptr) != '/');
4438 if (cnp->cn_nameiop != LOOKUP) {
4439 return (cache_fplookup_final_modifying(fpl));
4442 if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0)
4443 return (cache_fplookup_final_withparent(fpl));
4445 tvs = vget_prep_smr(tvp);
4446 if (__predict_false(tvs == VGET_NONE)) {
4447 return (cache_fpl_partial(fpl));
4450 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4451 cache_fpl_smr_exit(fpl);
4452 vget_abort(tvp, tvs);
4453 return (cache_fpl_aborted(fpl));
4456 cache_fpl_smr_exit(fpl);
4457 return (cache_fplookup_final_child(fpl, tvs));
4461 * Comment from locked lookup:
4462 * Check for degenerate name (e.g. / or "") which is a way of talking about a
4463 * directory, e.g. like "/." or ".".
4465 static int __noinline
4466 cache_fplookup_degenerate(struct cache_fpl *fpl)
4468 struct componentname *cnp;
4473 fpl->tvp = fpl->dvp;
4474 fpl->tvp_seqc = fpl->dvp_seqc;
4479 if (__predict_false(cnp->cn_nameiop != LOOKUP)) {
4480 cache_fpl_smr_exit(fpl);
4481 return (cache_fpl_handled_error(fpl, EISDIR));
4484 MPASS((cnp->cn_flags & SAVESTART) == 0);
4486 if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0) {
4487 return (cache_fplookup_final_withparent(fpl));
4490 dvs = vget_prep_smr(dvp);
4491 cache_fpl_smr_exit(fpl);
4492 if (__predict_false(dvs == VGET_NONE)) {
4493 return (cache_fpl_aborted(fpl));
4496 if ((cnp->cn_flags & LOCKLEAF) != 0) {
4497 lkflags = LK_SHARED;
4498 if ((cnp->cn_flags & LOCKSHARED) == 0)
4499 lkflags = LK_EXCLUSIVE;
4500 error = vget_finish(dvp, lkflags, dvs);
4501 if (__predict_false(error != 0)) {
4502 return (cache_fpl_aborted(fpl));
4505 vget_finish_ref(dvp, dvs);
4507 return (cache_fpl_handled(fpl));
4510 static int __noinline
4511 cache_fplookup_noentry(struct cache_fpl *fpl)
4513 struct nameidata *ndp;
4514 struct componentname *cnp;
4516 struct vnode *dvp, *tvp;
4524 dvp_seqc = fpl->dvp_seqc;
4526 MPASS(*(cnp->cn_nameptr) != '/');
4527 MPASS((cnp->cn_flags & MAKEENTRY) == 0);
4528 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
4529 MPASS(!cache_fpl_isdotdot(cnp));
4532 * Hack: delayed name len checking.
4534 if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
4535 cache_fpl_smr_exit(fpl);
4536 return (cache_fpl_handled_error(fpl, ENAMETOOLONG));
4539 if (cnp->cn_nameiop != LOOKUP) {
4541 return (cache_fplookup_modifying(fpl));
4544 MPASS((cnp->cn_flags & SAVESTART) == 0);
4547 * Only try to fill in the component if it is the last one,
4548 * otherwise not only there may be several to handle but the
4549 * walk may be complicated.
4551 if (!cache_fpl_islastcn(ndp)) {
4552 return (cache_fpl_partial(fpl));
4556 * Secure access to dvp; check cache_fplookup_partial_setup for
4559 dvs = vget_prep_smr(dvp);
4560 cache_fpl_smr_exit(fpl);
4561 if (__predict_false(dvs == VGET_NONE)) {
4562 return (cache_fpl_aborted(fpl));
4565 vget_finish_ref(dvp, dvs);
4566 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4568 return (cache_fpl_aborted(fpl));
4571 error = vn_lock(dvp, LK_SHARED);
4572 if (__predict_false(error != 0)) {
4574 return (cache_fpl_aborted(fpl));
4579 * TODO: provide variants which don't require locking either vnode.
4581 cnp->cn_flags |= ISLASTCN;
4582 docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
4584 cnp->cn_flags |= MAKEENTRY;
4585 cnp->cn_lkflags = LK_SHARED;
4586 if ((cnp->cn_flags & LOCKSHARED) == 0) {
4587 cnp->cn_lkflags = LK_EXCLUSIVE;
4589 error = VOP_LOOKUP(dvp, &tvp, cnp);
4597 return (cache_fpl_handled_error(fpl, error));
4600 return (cache_fpl_aborted(fpl));
4604 if (!fpl->savename) {
4605 MPASS((cnp->cn_flags & SAVENAME) == 0);
4609 MPASS(error == EJUSTRETURN);
4610 if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) {
4612 } else if ((cnp->cn_flags & LOCKPARENT) == 0) {
4615 return (cache_fpl_handled(fpl));
4618 if (__predict_false(tvp->v_type == VLNK || cache_fplookup_is_mp(fpl))) {
4621 return (cache_fpl_aborted(fpl));
4624 if ((cnp->cn_flags & LOCKLEAF) == 0) {
4628 if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) {
4630 } else if ((cnp->cn_flags & LOCKPARENT) == 0) {
4633 return (cache_fpl_handled(fpl));
4636 static int __noinline
4637 cache_fplookup_dot(struct cache_fpl *fpl)
4641 MPASS(!seqc_in_modify(fpl->dvp_seqc));
4643 * Just re-assign the value. seqc will be checked later for the first
4644 * non-dot path component in line and/or before deciding to return the
4647 fpl->tvp = fpl->dvp;
4648 fpl->tvp_seqc = fpl->dvp_seqc;
4650 counter_u64_add(dothits, 1);
4651 SDT_PROBE3(vfs, namecache, lookup, hit, fpl->dvp, ".", fpl->dvp);
4654 if (cache_fplookup_is_mp(fpl)) {
4655 error = cache_fplookup_cross_mount(fpl);
4660 static int __noinline
4661 cache_fplookup_dotdot(struct cache_fpl *fpl)
4663 struct nameidata *ndp;
4664 struct componentname *cnp;
4665 struct namecache *ncp;
4674 MPASS(cache_fpl_isdotdot(cnp));
4677 * XXX this is racy the same way regular lookup is
4679 for (pr = cnp->cn_cred->cr_prison; pr != NULL;
4681 if (dvp == pr->pr_root)
4684 if (dvp == ndp->ni_rootdir ||
4685 dvp == ndp->ni_topdir ||
4689 fpl->tvp_seqc = vn_seqc_read_any(dvp);
4690 if (seqc_in_modify(fpl->tvp_seqc)) {
4691 return (cache_fpl_aborted(fpl));
4696 if ((dvp->v_vflag & VV_ROOT) != 0) {
4699 * The opposite of climb mount is needed here.
4701 return (cache_fpl_aborted(fpl));
4704 ncp = atomic_load_ptr(&dvp->v_cache_dd);
4706 return (cache_fpl_aborted(fpl));
4709 nc_flag = atomic_load_char(&ncp->nc_flag);
4710 if ((nc_flag & NCF_ISDOTDOT) != 0) {
4711 if ((nc_flag & NCF_NEGATIVE) != 0)
4712 return (cache_fpl_aborted(fpl));
4713 fpl->tvp = ncp->nc_vp;
4715 fpl->tvp = ncp->nc_dvp;
4718 if (!cache_ncp_canuse(ncp)) {
4719 return (cache_fpl_aborted(fpl));
4722 fpl->tvp_seqc = vn_seqc_read_any(fpl->tvp);
4723 if (seqc_in_modify(fpl->tvp_seqc)) {
4724 return (cache_fpl_partial(fpl));
4727 counter_u64_add(dotdothits, 1);
4731 static int __noinline
4732 cache_fplookup_neg(struct cache_fpl *fpl, struct namecache *ncp, uint32_t hash)
4737 nc_flag = atomic_load_char(&ncp->nc_flag);
4738 MPASS((nc_flag & NCF_NEGATIVE) != 0);
4740 * If they want to create an entry we need to replace this one.
4742 if (__predict_false(fpl->cnp->cn_nameiop != LOOKUP)) {
4744 return (cache_fplookup_modifying(fpl));
4746 neg_promote = cache_neg_hit_prep(ncp);
4747 if (!cache_fpl_neg_ncp_canuse(ncp)) {
4748 cache_neg_hit_abort(ncp);
4749 return (cache_fpl_partial(fpl));
4752 return (cache_fplookup_negative_promote(fpl, ncp, hash));
4754 cache_neg_hit_finish(ncp);
4755 cache_fpl_smr_exit(fpl);
4756 return (cache_fpl_handled_error(fpl, ENOENT));
4760 * Resolve a symlink. Called by filesystem-specific routines.
4763 * ... -> cache_fplookup_symlink -> VOP_FPLOOKUP_SYMLINK -> cache_symlink_resolve
4766 cache_symlink_resolve(struct cache_fpl *fpl, const char *string, size_t len)
4768 struct nameidata *ndp;
4769 struct componentname *cnp;
4774 if (__predict_false(len == 0)) {
4778 ndp->ni_pathlen = fpl->nulchar - cnp->cn_nameptr - cnp->cn_namelen + 1;
4780 if (ndp->ni_pathlen != fpl->debug.ni_pathlen) {
4781 panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n",
4782 __func__, ndp->ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar,
4783 cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf);
4787 if (__predict_false(len + ndp->ni_pathlen > MAXPATHLEN)) {
4788 return (ENAMETOOLONG);
4791 if (__predict_false(ndp->ni_loopcnt++ >= MAXSYMLINKS)) {
4795 if (ndp->ni_pathlen > 1) {
4796 bcopy(ndp->ni_next, cnp->cn_pnbuf + len, ndp->ni_pathlen);
4798 cnp->cn_pnbuf[len] = '\0';
4800 bcopy(string, cnp->cn_pnbuf, len);
4802 ndp->ni_pathlen += len;
4803 cache_fpl_pathlen_add(fpl, len);
4804 cnp->cn_nameptr = cnp->cn_pnbuf;
4805 fpl->nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1];
4810 static int __noinline
4811 cache_fplookup_symlink(struct cache_fpl *fpl)
4813 struct nameidata *ndp;
4814 struct componentname *cnp;
4815 struct vnode *dvp, *tvp;
4823 if (cache_fpl_islastcn(ndp)) {
4824 if ((cnp->cn_flags & FOLLOW) == 0) {
4825 return (cache_fplookup_final(fpl));
4829 error = VOP_FPLOOKUP_SYMLINK(tvp, fpl);
4830 if (__predict_false(error != 0)) {
4833 return (cache_fpl_partial(fpl));
4837 cache_fpl_smr_exit(fpl);
4838 return (cache_fpl_handled_error(fpl, error));
4840 return (cache_fpl_aborted(fpl));
4844 if (*(cnp->cn_nameptr) == '/') {
4845 fpl->dvp = cache_fpl_handle_root(fpl);
4846 fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp);
4847 if (seqc_in_modify(fpl->dvp_seqc)) {
4848 return (cache_fpl_aborted(fpl));
4852 return (cache_fplookup_preparse(fpl));
4856 cache_fplookup_next(struct cache_fpl *fpl)
4858 struct componentname *cnp;
4859 struct namecache *ncp;
4860 struct vnode *dvp, *tvp;
4868 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
4869 if (cnp->cn_namelen == 1) {
4870 return (cache_fplookup_dot(fpl));
4872 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
4873 return (cache_fplookup_dotdot(fpl));
4877 MPASS(!cache_fpl_isdotdot(cnp));
4879 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
4881 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
4882 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
4883 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
4887 if (__predict_false(ncp == NULL)) {
4888 if (cnp->cn_nameptr[0] == '/') {
4889 return (cache_fplookup_skip_slashes(fpl));
4891 return (cache_fplookup_noentry(fpl));
4894 tvp = atomic_load_ptr(&ncp->nc_vp);
4895 nc_flag = atomic_load_char(&ncp->nc_flag);
4896 if ((nc_flag & NCF_NEGATIVE) != 0) {
4897 return (cache_fplookup_neg(fpl, ncp, hash));
4900 if (!cache_ncp_canuse(ncp)) {
4901 return (cache_fpl_partial(fpl));
4905 fpl->tvp_seqc = vn_seqc_read_any(tvp);
4906 if (seqc_in_modify(fpl->tvp_seqc)) {
4907 return (cache_fpl_partial(fpl));
4910 counter_u64_add(numposhits, 1);
4911 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, tvp);
4914 if (cache_fplookup_is_mp(fpl)) {
4915 error = cache_fplookup_cross_mount(fpl);
4921 cache_fplookup_mp_supported(struct mount *mp)
4925 if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0)
4931 * Walk up the mount stack (if any).
4933 * Correctness is provided in the following ways:
4934 * - all vnodes are protected from freeing with SMR
4935 * - struct mount objects are type stable making them always safe to access
4936 * - stability of the particular mount is provided by busying it
4937 * - relationship between the vnode which is mounted on and the mount is
4938 * verified with the vnode sequence counter after busying
4939 * - association between root vnode of the mount and the mount is protected
4942 * From that point on we can read the sequence counter of the root vnode
4943 * and get the next mount on the stack (if any) using the same protection.
4945 * By the end of successful walk we are guaranteed the reached state was
4946 * indeed present at least at some point which matches the regular lookup.
4948 static int __noinline
4949 cache_fplookup_climb_mount(struct cache_fpl *fpl)
4951 struct mount *mp, *prev_mp;
4952 struct mount_pcpu *mpcpu, *prev_mpcpu;
4957 vp_seqc = fpl->tvp_seqc;
4959 VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp);
4960 mp = atomic_load_ptr(&vp->v_mountedhere);
4961 if (__predict_false(mp == NULL)) {
4967 if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
4968 if (prev_mp != NULL)
4969 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
4970 return (cache_fpl_partial(fpl));
4972 if (prev_mp != NULL)
4973 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
4974 if (!vn_seqc_consistent(vp, vp_seqc)) {
4975 vfs_op_thread_exit_crit(mp, mpcpu);
4976 return (cache_fpl_partial(fpl));
4978 if (!cache_fplookup_mp_supported(mp)) {
4979 vfs_op_thread_exit_crit(mp, mpcpu);
4980 return (cache_fpl_partial(fpl));
4982 vp = atomic_load_ptr(&mp->mnt_rootvnode);
4984 vfs_op_thread_exit_crit(mp, mpcpu);
4985 return (cache_fpl_partial(fpl));
4987 vp_seqc = vn_seqc_read_any(vp);
4988 if (seqc_in_modify(vp_seqc)) {
4989 vfs_op_thread_exit_crit(mp, mpcpu);
4990 return (cache_fpl_partial(fpl));
4994 mp = atomic_load_ptr(&vp->v_mountedhere);
4999 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
5001 fpl->tvp_seqc = vp_seqc;
5005 static int __noinline
5006 cache_fplookup_cross_mount(struct cache_fpl *fpl)
5009 struct mount_pcpu *mpcpu;
5014 vp_seqc = fpl->tvp_seqc;
5016 VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp);
5017 mp = atomic_load_ptr(&vp->v_mountedhere);
5018 if (__predict_false(mp == NULL)) {
5022 if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
5023 return (cache_fpl_partial(fpl));
5025 if (!vn_seqc_consistent(vp, vp_seqc)) {
5026 vfs_op_thread_exit_crit(mp, mpcpu);
5027 return (cache_fpl_partial(fpl));
5029 if (!cache_fplookup_mp_supported(mp)) {
5030 vfs_op_thread_exit_crit(mp, mpcpu);
5031 return (cache_fpl_partial(fpl));
5033 vp = atomic_load_ptr(&mp->mnt_rootvnode);
5034 if (__predict_false(vp == NULL)) {
5035 vfs_op_thread_exit_crit(mp, mpcpu);
5036 return (cache_fpl_partial(fpl));
5038 vp_seqc = vn_seqc_read_any(vp);
5039 vfs_op_thread_exit_crit(mp, mpcpu);
5040 if (seqc_in_modify(vp_seqc)) {
5041 return (cache_fpl_partial(fpl));
5043 mp = atomic_load_ptr(&vp->v_mountedhere);
5044 if (__predict_false(mp != NULL)) {
5046 * There are possibly more mount points on top.
5047 * Normally this does not happen so for simplicity just start
5050 return (cache_fplookup_climb_mount(fpl));
5054 fpl->tvp_seqc = vp_seqc;
5059 * Check if a vnode is mounted on.
5062 cache_fplookup_is_mp(struct cache_fpl *fpl)
5067 return ((vn_irflag_read(vp) & VIRF_MOUNTPOINT) != 0);
5073 * The code was originally copy-pasted from regular lookup and despite
5074 * clean ups leaves performance on the table. Any modifications here
5075 * must take into account that in case off fallback the resulting
5076 * nameidata state has to be compatible with the original.
5080 * Debug ni_pathlen tracking.
5084 cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n)
5087 fpl->debug.ni_pathlen += n;
5088 KASSERT(fpl->debug.ni_pathlen <= PATH_MAX,
5089 ("%s: pathlen overflow to %zd\n", __func__, fpl->debug.ni_pathlen));
5093 cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n)
5096 fpl->debug.ni_pathlen -= n;
5097 KASSERT(fpl->debug.ni_pathlen <= PATH_MAX,
5098 ("%s: pathlen underflow to %zd\n", __func__, fpl->debug.ni_pathlen));
5102 cache_fpl_pathlen_inc(struct cache_fpl *fpl)
5105 cache_fpl_pathlen_add(fpl, 1);
5109 cache_fpl_pathlen_dec(struct cache_fpl *fpl)
5112 cache_fpl_pathlen_sub(fpl, 1);
5116 cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n)
5121 cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n)
5126 cache_fpl_pathlen_inc(struct cache_fpl *fpl)
5131 cache_fpl_pathlen_dec(struct cache_fpl *fpl)
5136 static int __always_inline
5137 cache_fplookup_preparse(struct cache_fpl *fpl)
5139 struct componentname *cnp;
5143 if (__predict_false(cnp->cn_nameptr[0] == '\0')) {
5144 return (cache_fplookup_degenerate(fpl));
5148 * By this point the shortest possible pathname is one character + nul
5149 * terminator, hence 2.
5151 KASSERT(fpl->debug.ni_pathlen >= 2, ("%s: pathlen %zu\n", __func__,
5152 fpl->debug.ni_pathlen));
5153 KASSERT(&cnp->cn_nameptr[fpl->debug.ni_pathlen - 2] == fpl->nulchar - 1,
5154 ("%s: mismatch on string (%p != %p) [%s]\n", __func__,
5155 &cnp->cn_nameptr[fpl->debug.ni_pathlen - 2], fpl->nulchar - 1,
5157 if (__predict_false(*(fpl->nulchar - 1) == '/')) {
5160 * Regular lookup performs the following:
5161 * *ndp->ni_next = '\0';
5162 * cnp->cn_flags |= TRAILINGSLASH;
5164 * Which is problematic since it modifies data read
5165 * from userspace. Then if fast path lookup was to
5166 * abort we would have to either restore it or convey
5167 * the flag. Since this is a corner case just ignore
5168 * it for simplicity.
5170 return (cache_fpl_aborted(fpl));
5176 cache_fplookup_parse(struct cache_fpl *fpl)
5178 struct nameidata *ndp;
5179 struct componentname *cnp;
5186 * Find the end of this path component, it is either / or nul.
5188 * Store / as a temporary sentinel so that we only have one character
5189 * to test for. Pathnames tend to be short so this should not be
5190 * resulting in cache misses.
5192 KASSERT(&cnp->cn_nameptr[fpl->debug.ni_pathlen - 1] == fpl->nulchar,
5193 ("%s: mismatch between pathlen (%zu) and nulchar (%p != %p), string [%s]\n",
5194 __func__, fpl->debug.ni_pathlen, &cnp->cn_nameptr[fpl->debug.ni_pathlen - 1],
5195 fpl->nulchar, cnp->cn_pnbuf));
5196 KASSERT(*fpl->nulchar == '\0',
5197 ("%s: expected nul at %p; string [%s]\n", __func__, fpl->nulchar,
5199 *fpl->nulchar = '/';
5200 for (cp = cnp->cn_nameptr; *cp != '/'; cp++) {
5201 KASSERT(*cp != '\0',
5202 ("%s: encountered unexpected nul; string [%s]\n", __func__,
5206 *fpl->nulchar = '\0';
5208 cnp->cn_namelen = cp - cnp->cn_nameptr;
5209 cache_fpl_pathlen_sub(fpl, cnp->cn_namelen);
5211 * Hack: we have to check if the found path component's length exceeds
5212 * NAME_MAX. However, the condition is very rarely true and check can
5213 * be elided in the common case -- if an entry was found in the cache,
5214 * then it could not have been too long to begin with.
5220 * Code below is only here to assure compatibility with regular lookup.
5221 * It covers handling of trailing slashes and names like "/", both of
5222 * which of can be taken care of upfront which lockless lookup does
5223 * in cache_fplookup_preparse. Regular lookup performs these for each
5226 while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
5229 panic("%s: ran into TRAILINGSLASH handling from [%s]\n",
5230 __func__, cnp->cn_pnbuf);
5234 if (cnp->cn_nameptr[0] == '\0') {
5235 panic("%s: ran into degenerate name from [%s]\n", __func__, cnp->cn_pnbuf);
5242 cache_fplookup_parse_advance(struct cache_fpl *fpl)
5244 struct nameidata *ndp;
5245 struct componentname *cnp;
5250 cnp->cn_nameptr = ndp->ni_next;
5251 KASSERT(*(cnp->cn_nameptr) == '/',
5252 ("%s: should have seen slash at %p ; buf %p [%s]\n", __func__,
5253 cnp->cn_nameptr, cnp->cn_pnbuf, cnp->cn_pnbuf));
5255 cache_fpl_pathlen_dec(fpl);
5259 * Skip spurious slashes in a pathname (e.g., "foo///bar") and retry.
5261 * Lockless lookup tries to elide checking for spurious slashes and should they
5262 * be present is guaranteed to fail to find an entry. In this case the caller
5263 * must check if the name starts with a slash and call this routine. It is
5264 * going to fast forward across the spurious slashes and set the state up for
5267 static int __noinline
5268 cache_fplookup_skip_slashes(struct cache_fpl *fpl)
5270 struct nameidata *ndp;
5271 struct componentname *cnp;
5276 MPASS(*(cnp->cn_nameptr) == '/');
5279 cache_fpl_pathlen_dec(fpl);
5280 } while (*(cnp->cn_nameptr) == '/');
5283 * Go back to one slash so that cache_fplookup_parse_advance has
5284 * something to skip.
5287 cache_fpl_pathlen_inc(fpl);
5290 * cache_fplookup_parse_advance starts from ndp->ni_next
5292 ndp->ni_next = cnp->cn_nameptr;
5295 * See cache_fplookup_dot.
5297 fpl->tvp = fpl->dvp;
5298 fpl->tvp_seqc = fpl->dvp_seqc;
5304 * See the API contract for VOP_FPLOOKUP_VEXEC.
5306 static int __noinline
5307 cache_fplookup_failed_vexec(struct cache_fpl *fpl, int error)
5309 struct componentname *cnp;
5315 dvp_seqc = fpl->dvp_seqc;
5318 * Hack: delayed name len checking.
5320 if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
5321 cache_fpl_smr_exit(fpl);
5322 return (cache_fpl_handled_error(fpl, ENAMETOOLONG));
5326 * Hack: they may be looking up foo/bar, where foo is a
5327 * regular file. In such a case we need to turn ENOTDIR,
5328 * but we may happen to get here with a different error.
5330 if (dvp->v_type != VDIR) {
5332 * The check here is predominantly to catch
5333 * EOPNOTSUPP from dead_vnodeops. If the vnode
5334 * gets doomed past this point it is going to
5335 * fail seqc verification.
5337 if (VN_IS_DOOMED(dvp)) {
5338 return (cache_fpl_aborted(fpl));
5344 * Hack: handle O_SEARCH.
5346 * Open Group Base Specifications Issue 7, 2018 edition states:
5348 * If the access mode of the open file description associated with the
5349 * file descriptor is not O_SEARCH, the function shall check whether
5350 * directory searches are permitted using the current permissions of
5351 * the directory underlying the file descriptor. If the access mode is
5352 * O_SEARCH, the function shall not perform the check.
5355 * Regular lookup tests for the NOEXECCHECK flag for every path
5356 * component to decide whether to do the permission check. However,
5357 * since most lookups never have the flag (and when they do it is only
5358 * present for the first path component), lockless lookup only acts on
5359 * it if there is a permission problem. Here the flag is represented
5360 * with a boolean so that we don't have to clear it on the way out.
5362 * For simplicity this always aborts.
5363 * TODO: check if this is the first lookup and ignore the permission
5364 * problem. Note the flag has to survive fallback (if it happens to be
5368 return (cache_fpl_aborted(fpl));
5373 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
5374 error = cache_fpl_aborted(fpl);
5376 cache_fpl_partial(fpl);
5380 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
5381 error = cache_fpl_aborted(fpl);
5383 cache_fpl_smr_exit(fpl);
5384 cache_fpl_handled_error(fpl, error);
5392 cache_fplookup_impl(struct vnode *dvp, struct cache_fpl *fpl)
5394 struct nameidata *ndp;
5395 struct componentname *cnp;
5402 cache_fpl_checkpoint(fpl);
5405 * The vnode at hand is almost always stable, skip checking for it.
5406 * Worst case this postpones the check towards the end of the iteration
5410 fpl->dvp_seqc = vn_seqc_read_notmodify(fpl->dvp);
5412 mp = atomic_load_ptr(&dvp->v_mount);
5413 if (__predict_false(mp == NULL || !cache_fplookup_mp_supported(mp))) {
5414 return (cache_fpl_aborted(fpl));
5417 error = cache_fplookup_preparse(fpl);
5418 if (__predict_false(cache_fpl_terminated(fpl))) {
5423 error = cache_fplookup_parse(fpl);
5424 if (__predict_false(error != 0)) {
5428 error = VOP_FPLOOKUP_VEXEC(fpl->dvp, cnp->cn_cred);
5429 if (__predict_false(error != 0)) {
5430 error = cache_fplookup_failed_vexec(fpl, error);
5434 error = cache_fplookup_next(fpl);
5435 if (__predict_false(cache_fpl_terminated(fpl))) {
5439 VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp);
5441 if (fpl->tvp->v_type == VLNK) {
5442 error = cache_fplookup_symlink(fpl);
5443 if (cache_fpl_terminated(fpl)) {
5447 if (cache_fpl_islastcn(ndp)) {
5448 error = cache_fplookup_final(fpl);
5452 if (!vn_seqc_consistent(fpl->dvp, fpl->dvp_seqc)) {
5453 error = cache_fpl_aborted(fpl);
5457 fpl->dvp = fpl->tvp;
5458 fpl->dvp_seqc = fpl->tvp_seqc;
5459 cache_fplookup_parse_advance(fpl);
5462 cache_fpl_checkpoint(fpl);
5469 * Fast path lookup protected with SMR and sequence counters.
5471 * Note: all VOP_FPLOOKUP_VEXEC routines have a comment referencing this one.
5473 * Filesystems can opt in by setting the MNTK_FPLOOKUP flag and meeting criteria
5476 * Traditional vnode lookup conceptually looks like this:
5482 * vn_unlock(current);
5489 * Each jump to the next vnode is safe memory-wise and atomic with respect to
5490 * any modifications thanks to holding respective locks.
5492 * The same guarantee can be provided with a combination of safe memory
5493 * reclamation and sequence counters instead. If all operations which affect
5494 * the relationship between the current vnode and the one we are looking for
5495 * also modify the counter, we can verify whether all the conditions held as
5496 * we made the jump. This includes things like permissions, mount points etc.
5497 * Counter modification is provided by enclosing relevant places in
5498 * vn_seqc_write_begin()/end() calls.
5500 * Thus this translates to:
5503 * dvp_seqc = seqc_read_any(dvp);
5504 * if (seqc_in_modify(dvp_seqc)) // someone is altering the vnode
5508 * tvp_seqc = seqc_read_any(tvp);
5509 * if (seqc_in_modify(tvp_seqc)) // someone is altering the target vnode
5511 * if (!seqc_consistent(dvp, dvp_seqc) // someone is altering the vnode
5513 * dvp = tvp; // we know nothing of importance has changed
5514 * dvp_seqc = tvp_seqc; // store the counter for the tvp iteration
5518 * vget(); // secure the vnode
5519 * if (!seqc_consistent(tvp, tvp_seqc) // final check
5521 * // at this point we know nothing has changed for any parent<->child pair
5522 * // as they were crossed during the lookup, meaning we matched the guarantee
5523 * // of the locked variant
5526 * The API contract for VOP_FPLOOKUP_VEXEC routines is as follows:
5527 * - they are called while within vfs_smr protection which they must never exit
5528 * - EAGAIN can be returned to denote checking could not be performed, it is
5529 * always valid to return it
5530 * - if the sequence counter has not changed the result must be valid
5531 * - if the sequence counter has changed both false positives and false negatives
5532 * are permitted (since the result will be rejected later)
5533 * - for simple cases of unix permission checks vaccess_vexec_smr can be used
5535 * Caveats to watch out for:
5536 * - vnodes are passed unlocked and unreferenced with nothing stopping
5537 * VOP_RECLAIM, in turn meaning that ->v_data can become NULL. It is advised
5538 * to use atomic_load_ptr to fetch it.
5539 * - the aforementioned object can also get freed, meaning absent other means it
5540 * should be protected with vfs_smr
5541 * - either safely checking permissions as they are modified or guaranteeing
5542 * their stability is left to the routine
5545 cache_fplookup(struct nameidata *ndp, enum cache_fpl_status *status,
5548 struct cache_fpl fpl;
5551 struct componentname *cnp;
5554 fpl.status = CACHE_FPL_STATUS_UNSET;
5557 fpl.cnp = cnp = &ndp->ni_cnd;
5558 MPASS(ndp->ni_lcf == 0);
5559 MPASS(curthread == cnp->cn_thread);
5560 KASSERT ((cnp->cn_flags & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
5561 ("%s: internal flags found in cn_flags %" PRIx64, __func__,
5563 if ((cnp->cn_flags & SAVESTART) != 0) {
5564 MPASS(cnp->cn_nameiop != LOOKUP);
5566 MPASS(cnp->cn_nameptr == cnp->cn_pnbuf);
5568 if (__predict_false(!cache_can_fplookup(&fpl))) {
5569 *status = fpl.status;
5570 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
5571 return (EOPNOTSUPP);
5574 cache_fpl_checkpoint_outer(&fpl);
5576 cache_fpl_smr_enter_initial(&fpl);
5578 fpl.debug.ni_pathlen = ndp->ni_pathlen;
5580 fpl.nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1];
5581 fpl.fsearch = false;
5582 fpl.savename = (cnp->cn_flags & SAVENAME) != 0;
5584 pwd = pwd_get_smr();
5586 ndp->ni_rootdir = pwd->pwd_rdir;
5587 ndp->ni_topdir = pwd->pwd_jdir;
5589 if (cnp->cn_pnbuf[0] == '/') {
5590 dvp = cache_fpl_handle_root(&fpl);
5591 MPASS(ndp->ni_resflags == 0);
5592 ndp->ni_resflags = NIRES_ABS;
5594 if (ndp->ni_dirfd == AT_FDCWD) {
5595 dvp = pwd->pwd_cdir;
5597 error = cache_fplookup_dirfd(&fpl, &dvp);
5598 if (__predict_false(error != 0)) {
5604 SDT_PROBE4(vfs, namei, lookup, entry, dvp, cnp->cn_pnbuf, cnp->cn_flags, true);
5605 error = cache_fplookup_impl(dvp, &fpl);
5607 cache_fpl_smr_assert_not_entered(&fpl);
5608 cache_fpl_assert_status(&fpl);
5609 *status = fpl.status;
5610 if (SDT_PROBES_ENABLED()) {
5611 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
5612 if (fpl.status == CACHE_FPL_STATUS_HANDLED)
5613 SDT_PROBE4(vfs, namei, lookup, return, error, ndp->ni_vp, true,
5617 if (__predict_true(fpl.status == CACHE_FPL_STATUS_HANDLED)) {
5618 MPASS(error != CACHE_FPL_FAILED);
5620 MPASS(fpl.dvp == NULL);
5621 MPASS(fpl.tvp == NULL);
5622 MPASS(fpl.savename == false);
5624 ndp->ni_dvp = fpl.dvp;
5625 ndp->ni_vp = fpl.tvp;
5627 cnp->cn_flags |= HASBUF;
5629 cache_fpl_cleanup_cnp(cnp);