FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_cache.c
1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1989, 1993, 1995
5 * The Regents of the University of California. All rights reserved.
6 *
7 * This code is derived from software contributed to Berkeley by
8 * Poul-Henning Kamp of the FreeBSD Project.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
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.
21 *
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
32 * SUCH DAMAGE.
33 *
34 * @(#)vfs_cache.c 8.5 (Berkeley) 3/22/95
35 */
36
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39
40 #include "opt_ddb.h"
41 #include "opt_ktrace.h"
42
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>
50 #include <sys/ktr.h>
51 #include <sys/lock.h>
52 #include <sys/malloc.h>
53 #include <sys/fcntl.h>
54 #include <sys/jail.h>
55 #include <sys/mount.h>
56 #include <sys/namei.h>
57 #include <sys/proc.h>
58 #include <sys/seqc.h>
59 #include <sys/sdt.h>
60 #include <sys/smr.h>
61 #include <sys/smp.h>
62 #include <sys/syscallsubr.h>
63 #include <sys/sysctl.h>
64 #include <sys/sysproto.h>
65 #include <sys/vnode.h>
66 #include <ck_queue.h>
67 #ifdef KTRACE
68 #include <sys/ktrace.h>
69 #endif
70 #ifdef INVARIANTS
71 #include <machine/_inttypes.h>
72 #endif
73
74 #include <sys/capsicum.h>
75
76 #include <security/audit/audit.h>
77 #include <security/mac/mac_framework.h>
78
79 #ifdef DDB
80 #include <ddb/ddb.h>
81 #endif
82
83 #include <vm/uma.h>
84
85 /*
86 * High level overview of name caching in the VFS layer.
87 *
88 * Originally caching was implemented as part of UFS, later extracted to allow
89 * use by other filesystems. A decision was made to make it optional and
90 * completely detached from the rest of the kernel, which comes with limitations
91 * outlined near the end of this comment block.
92 *
93 * This fundamental choice needs to be revisited. In the meantime, the current
94 * state is described below. Significance of all notable routines is explained
95 * in comments placed above their implementation. Scattered thoroughout the
96 * file are TODO comments indicating shortcomings which can be fixed without
97 * reworking everything (most of the fixes will likely be reusable). Various
98 * details are omitted from this explanation to not clutter the overview, they
99 * have to be checked by reading the code and associated commentary.
100 *
101 * Keep in mind that it's individual path components which are cached, not full
102 * paths. That is, for a fully cached path "foo/bar/baz" there are 3 entries,
103 * one for each name.
104 *
105 * I. Data organization
106 *
107 * Entries are described by "struct namecache" objects and stored in a hash
108 * table. See cache_get_hash for more information.
109 *
110 * "struct vnode" contains pointers to source entries (names which can be found
111 * when traversing through said vnode), destination entries (names of that
112 * vnode (see "Limitations" for a breakdown on the subject) and a pointer to
113 * the parent vnode.
114 *
115 * The (directory vnode; name) tuple reliably determines the target entry if
116 * it exists.
117 *
118 * Since there are no small locks at this time (all are 32 bytes in size on
119 * LP64), the code works around the problem by introducing lock arrays to
120 * protect hash buckets and vnode lists.
121 *
122 * II. Filesystem integration
123 *
124 * Filesystems participating in name caching do the following:
125 * - set vop_lookup routine to vfs_cache_lookup
126 * - set vop_cachedlookup to whatever can perform the lookup if the above fails
127 * - if they support lockless lookup (see below), vop_fplookup_vexec and
128 * vop_fplookup_symlink are set along with the MNTK_FPLOOKUP flag on the
129 * mount point
130 * - call cache_purge or cache_vop_* routines to eliminate stale entries as
131 * applicable
132 * - call cache_enter to add entries depending on the MAKEENTRY flag
133 *
134 * With the above in mind, there are 2 entry points when doing lookups:
135 * - ... -> namei -> cache_fplookup -- this is the default
136 * - ... -> VOP_LOOKUP -> vfs_cache_lookup -- normally only called by namei
137 * should the above fail
138 *
139 * Example code flow how an entry is added:
140 * ... -> namei -> cache_fplookup -> cache_fplookup_noentry -> VOP_LOOKUP ->
141 * vfs_cache_lookup -> VOP_CACHEDLOOKUP -> ufs_lookup_ino -> cache_enter
142 *
143 * III. Performance considerations
144 *
145 * For lockless case forward lookup avoids any writes to shared areas apart
146 * from the terminal path component. In other words non-modifying lookups of
147 * different files don't suffer any scalability problems in the namecache.
148 * Looking up the same file is limited by VFS and goes beyond the scope of this
149 * file.
150 *
151 * At least on amd64 the single-threaded bottleneck for long paths is hashing
152 * (see cache_get_hash). There are cases where the code issues acquire fence
153 * multiple times, they can be combined on architectures which suffer from it.
154 *
155 * For locked case each encountered vnode has to be referenced and locked in
156 * order to be handed out to the caller (normally that's namei). This
157 * introduces significant hit single-threaded and serialization multi-threaded.
158 *
159 * Reverse lookup (e.g., "getcwd") fully scales provided it is fully cached --
160 * avoids any writes to shared areas to any components.
161 *
162 * Unrelated insertions are partially serialized on updating the global entry
163 * counter and possibly serialized on colliding bucket or vnode locks.
164 *
165 * IV. Observability
166 *
167 * Note not everything has an explicit dtrace probe nor it should have, thus
168 * some of the one-liners below depend on implementation details.
169 *
170 * Examples:
171 *
172 * # Check what lookups failed to be handled in a lockless manner. Column 1 is
173 * # line number, column 2 is status code (see cache_fpl_status)
174 * dtrace -n 'vfs:fplookup:lookup:done { @[arg1, arg2] = count(); }'
175 *
176 * # Lengths of names added by binary name
177 * dtrace -n 'fbt::cache_enter_time:entry { @[execname] = quantize(args[2]->cn_namelen); }'
178 *
179 * # Same as above but only those which exceed 64 characters
180 * dtrace -n 'fbt::cache_enter_time:entry /args[2]->cn_namelen > 64/ { @[execname] = quantize(args[2]->cn_namelen); }'
181 *
182 * # Who is performing lookups with spurious slashes (e.g., "foo//bar") and what
183 * # path is it
184 * dtrace -n 'fbt::cache_fplookup_skip_slashes:entry { @[execname, stringof(args[0]->cnp->cn_pnbuf)] = count(); }'
185 *
186 * V. Limitations and implementation defects
187 *
188 * - since it is possible there is no entry for an open file, tools like
189 * "procstat" may fail to resolve fd -> vnode -> path to anything
190 * - even if a filesystem adds an entry, it may get purged (e.g., due to memory
191 * shortage) in which case the above problem applies
192 * - hardlinks are not tracked, thus if a vnode is reachable in more than one
193 * way, resolving a name may return a different path than the one used to
194 * open it (even if said path is still valid)
195 * - by default entries are not added for newly created files
196 * - adding an entry may need to evict negative entry first, which happens in 2
197 * distinct places (evicting on lookup, adding in a later VOP) making it
198 * impossible to simply reuse it
199 * - there is a simple scheme to evict negative entries as the cache is approaching
200 * its capacity, but it is very unclear if doing so is a good idea to begin with
201 * - vnodes are subject to being recycled even if target inode is left in memory,
202 * which loses the name cache entries when it perhaps should not. in case of tmpfs
203 * names get duplicated -- kept by filesystem itself and namecache separately
204 * - struct namecache has a fixed size and comes in 2 variants, often wasting space.
205 * now hard to replace with malloc due to dependence on SMR.
206 * - lack of better integration with the kernel also turns nullfs into a layered
207 * filesystem instead of something which can take advantage of caching
208 */
209
210 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
211 "Name cache");
212
213 SDT_PROVIDER_DECLARE(vfs);
214 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
215 "struct vnode *");
216 SDT_PROBE_DEFINE3(vfs, namecache, enter, duplicate, "struct vnode *", "char *",
217 "struct vnode *");
218 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
219 "char *");
220 SDT_PROBE_DEFINE2(vfs, namecache, fullpath_smr, hit, "struct vnode *",
221 "const char *");
222 SDT_PROBE_DEFINE4(vfs, namecache, fullpath_smr, miss, "struct vnode *",
223 "struct namecache *", "int", "int");
224 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
225 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
226 "char *", "struct vnode *");
227 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
228 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
229 "struct vnode *", "char *");
230 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
231 "struct vnode *");
232 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
233 "struct vnode *", "char *");
234 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
235 "char *");
236 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, hit, "struct vnode *",
237 "struct componentname *");
238 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, miss, "struct vnode *",
239 "struct componentname *");
240 SDT_PROBE_DEFINE3(vfs, namecache, purge, done, "struct vnode *", "size_t", "size_t");
241 SDT_PROBE_DEFINE1(vfs, namecache, purge, batch, "int");
242 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
243 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
244 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
245 "struct vnode *");
246 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
247 "char *");
248 SDT_PROBE_DEFINE2(vfs, namecache, evict_negative, done, "struct vnode *",
249 "char *");
250 SDT_PROBE_DEFINE1(vfs, namecache, symlink, alloc__fail, "size_t");
251
252 SDT_PROBE_DEFINE3(vfs, fplookup, lookup, done, "struct nameidata", "int", "bool");
253 SDT_PROBE_DECLARE(vfs, namei, lookup, entry);
254 SDT_PROBE_DECLARE(vfs, namei, lookup, return);
255
256 /*
257 * This structure describes the elements in the cache of recent
258 * names looked up by namei.
259 */
260 struct negstate {
261 u_char neg_flag;
262 u_char neg_hit;
263 };
264 _Static_assert(sizeof(struct negstate) <= sizeof(struct vnode *),
265 "the state must fit in a union with a pointer without growing it");
266
267 struct namecache {
268 LIST_ENTRY(namecache) nc_src; /* source vnode list */
269 TAILQ_ENTRY(namecache) nc_dst; /* destination vnode list */
270 CK_SLIST_ENTRY(namecache) nc_hash;/* hash chain */
271 struct vnode *nc_dvp; /* vnode of parent of name */
272 union {
273 struct vnode *nu_vp; /* vnode the name refers to */
274 struct negstate nu_neg;/* negative entry state */
275 } n_un;
276 u_char nc_flag; /* flag bits */
277 u_char nc_nlen; /* length of name */
278 char nc_name[0]; /* segment name + nul */
279 };
280
281 /*
282 * struct namecache_ts repeats struct namecache layout up to the
283 * nc_nlen member.
284 * struct namecache_ts is used in place of struct namecache when time(s) need
285 * to be stored. The nc_dotdottime field is used when a cache entry is mapping
286 * both a non-dotdot directory name plus dotdot for the directory's
287 * parent.
288 *
289 * See below for alignment requirement.
290 */
291 struct namecache_ts {
292 struct timespec nc_time; /* timespec provided by fs */
293 struct timespec nc_dotdottime; /* dotdot timespec provided by fs */
294 int nc_ticks; /* ticks value when entry was added */
295 int nc_pad;
296 struct namecache nc_nc;
297 };
298
299 TAILQ_HEAD(cache_freebatch, namecache);
300
301 /*
302 * At least mips n32 performs 64-bit accesses to timespec as found
303 * in namecache_ts and requires them to be aligned. Since others
304 * may be in the same spot suffer a little bit and enforce the
305 * alignment for everyone. Note this is a nop for 64-bit platforms.
306 */
307 #define CACHE_ZONE_ALIGNMENT UMA_ALIGNOF(time_t)
308
309 /*
310 * TODO: the initial value of CACHE_PATH_CUTOFF was inherited from the
311 * 4.4 BSD codebase. Later on struct namecache was tweaked to become
312 * smaller and the value was bumped to retain the total size, but it
313 * was never re-evaluated for suitability. A simple test counting
314 * lengths during package building shows that the value of 45 covers
315 * about 86% of all added entries, reaching 99% at 65.
316 *
317 * Regardless of the above, use of dedicated zones instead of malloc may be
318 * inducing additional waste. This may be hard to address as said zones are
319 * tied to VFS SMR. Even if retaining them, the current split should be
320 * re-evaluated.
321 */
322 #ifdef __LP64__
323 #define CACHE_PATH_CUTOFF 45
324 #define CACHE_LARGE_PAD 6
325 #else
326 #define CACHE_PATH_CUTOFF 41
327 #define CACHE_LARGE_PAD 2
328 #endif
329
330 #define CACHE_ZONE_SMALL_SIZE (offsetof(struct namecache, nc_name) + CACHE_PATH_CUTOFF + 1)
331 #define CACHE_ZONE_SMALL_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_SMALL_SIZE)
332 #define CACHE_ZONE_LARGE_SIZE (offsetof(struct namecache, nc_name) + NAME_MAX + 1 + CACHE_LARGE_PAD)
333 #define CACHE_ZONE_LARGE_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_LARGE_SIZE)
334
335 _Static_assert((CACHE_ZONE_SMALL_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
336 _Static_assert((CACHE_ZONE_SMALL_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
337 _Static_assert((CACHE_ZONE_LARGE_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
338 _Static_assert((CACHE_ZONE_LARGE_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
339
340 #define nc_vp n_un.nu_vp
341 #define nc_neg n_un.nu_neg
342
343 /*
344 * Flags in namecache.nc_flag
345 */
346 #define NCF_WHITE 0x01
347 #define NCF_ISDOTDOT 0x02
348 #define NCF_TS 0x04
349 #define NCF_DTS 0x08
350 #define NCF_DVDROP 0x10
351 #define NCF_NEGATIVE 0x20
352 #define NCF_INVALID 0x40
353 #define NCF_WIP 0x80
354
355 /*
356 * Flags in negstate.neg_flag
357 */
358 #define NEG_HOT 0x01
359
360 static bool cache_neg_evict_cond(u_long lnumcache);
361
362 /*
363 * Mark an entry as invalid.
364 *
365 * This is called before it starts getting deconstructed.
366 */
367 static void
368 cache_ncp_invalidate(struct namecache *ncp)
369 {
370
371 KASSERT((ncp->nc_flag & NCF_INVALID) == 0,
372 ("%s: entry %p already invalid", __func__, ncp));
373 atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_INVALID);
374 atomic_thread_fence_rel();
375 }
376
377 /*
378 * Check whether the entry can be safely used.
379 *
380 * All places which elide locks are supposed to call this after they are
381 * done with reading from an entry.
382 */
383 #define cache_ncp_canuse(ncp) ({ \
384 struct namecache *_ncp = (ncp); \
385 u_char _nc_flag; \
386 \
387 atomic_thread_fence_acq(); \
388 _nc_flag = atomic_load_char(&_ncp->nc_flag); \
389 __predict_true((_nc_flag & (NCF_INVALID | NCF_WIP)) == 0); \
390 })
391
392 /*
393 * Like the above but also checks NCF_WHITE.
394 */
395 #define cache_fpl_neg_ncp_canuse(ncp) ({ \
396 struct namecache *_ncp = (ncp); \
397 u_char _nc_flag; \
398 \
399 atomic_thread_fence_acq(); \
400 _nc_flag = atomic_load_char(&_ncp->nc_flag); \
401 __predict_true((_nc_flag & (NCF_INVALID | NCF_WIP | NCF_WHITE)) == 0); \
402 })
403
404 VFS_SMR_DECLARE;
405
406 static SYSCTL_NODE(_vfs_cache, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
407 "Name cache parameters");
408
409 static u_int __read_mostly ncsize; /* the size as computed on creation or resizing */
410 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, size, CTLFLAG_RW, &ncsize, 0,
411 "Total namecache capacity");
412
413 u_int ncsizefactor = 2;
414 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, sizefactor, CTLFLAG_RW, &ncsizefactor, 0,
415 "Size factor for namecache");
416
417 static u_long __read_mostly ncnegfactor = 5; /* ratio of negative entries */
418 SYSCTL_ULONG(_vfs_cache_param, OID_AUTO, negfactor, CTLFLAG_RW, &ncnegfactor, 0,
419 "Ratio of negative namecache entries");
420
421 /*
422 * Negative entry % of namecache capacity above which automatic eviction is allowed.
423 *
424 * Check cache_neg_evict_cond for details.
425 */
426 static u_int ncnegminpct = 3;
427
428 static u_int __read_mostly neg_min; /* the above recomputed against ncsize */
429 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, negmin, CTLFLAG_RD, &neg_min, 0,
430 "Negative entry count above which automatic eviction is allowed");
431
432 /*
433 * Structures associated with name caching.
434 */
435 #define NCHHASH(hash) \
436 (&nchashtbl[(hash) & nchash])
437 static __read_mostly CK_SLIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
438 static u_long __read_mostly nchash; /* size of hash table */
439 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
440 "Size of namecache hash table");
441 static u_long __exclusive_cache_line numneg; /* number of negative entries allocated */
442 static u_long __exclusive_cache_line numcache;/* number of cache entries allocated */
443
444 struct nchstats nchstats; /* cache effectiveness statistics */
445
446 static bool __read_frequently cache_fast_revlookup = true;
447 SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_revlookup, CTLFLAG_RW,
448 &cache_fast_revlookup, 0, "");
449
450 static bool __read_mostly cache_rename_add = true;
451 SYSCTL_BOOL(_vfs, OID_AUTO, cache_rename_add, CTLFLAG_RW,
452 &cache_rename_add, 0, "");
453
454 static u_int __exclusive_cache_line neg_cycle;
455
456 #define ncneghash 3
457 #define numneglists (ncneghash + 1)
458
459 struct neglist {
460 struct mtx nl_evict_lock;
461 struct mtx nl_lock __aligned(CACHE_LINE_SIZE);
462 TAILQ_HEAD(, namecache) nl_list;
463 TAILQ_HEAD(, namecache) nl_hotlist;
464 u_long nl_hotnum;
465 } __aligned(CACHE_LINE_SIZE);
466
467 static struct neglist neglists[numneglists];
468
469 static inline struct neglist *
470 NCP2NEGLIST(struct namecache *ncp)
471 {
472
473 return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
474 }
475
476 static inline struct negstate *
477 NCP2NEGSTATE(struct namecache *ncp)
478 {
479
480 MPASS(atomic_load_char(&ncp->nc_flag) & NCF_NEGATIVE);
481 return (&ncp->nc_neg);
482 }
483
484 #define numbucketlocks (ncbuckethash + 1)
485 static u_int __read_mostly ncbuckethash;
486 static struct mtx_padalign __read_mostly *bucketlocks;
487 #define HASH2BUCKETLOCK(hash) \
488 ((struct mtx *)(&bucketlocks[((hash) & ncbuckethash)]))
489
490 #define numvnodelocks (ncvnodehash + 1)
491 static u_int __read_mostly ncvnodehash;
492 static struct mtx __read_mostly *vnodelocks;
493 static inline struct mtx *
494 VP2VNODELOCK(struct vnode *vp)
495 {
496
497 return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
498 }
499
500 static void
501 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
502 {
503 struct namecache_ts *ncp_ts;
504
505 KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
506 (tsp == NULL && ticksp == NULL),
507 ("No NCF_TS"));
508
509 if (tsp == NULL)
510 return;
511
512 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
513 *tsp = ncp_ts->nc_time;
514 *ticksp = ncp_ts->nc_ticks;
515 }
516
517 #ifdef DEBUG_CACHE
518 static int __read_mostly doingcache = 1; /* 1 => enable the cache */
519 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
520 "VFS namecache enabled");
521 #endif
522
523 /* Export size information to userland */
524 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
525 sizeof(struct namecache), "sizeof(struct namecache)");
526
527 /*
528 * The new name cache statistics
529 */
530 static SYSCTL_NODE(_vfs_cache, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
531 "Name cache statistics");
532
533 #define STATNODE_ULONG(name, varname, descr) \
534 SYSCTL_ULONG(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
535 #define STATNODE_COUNTER(name, varname, descr) \
536 static COUNTER_U64_DEFINE_EARLY(varname); \
537 SYSCTL_COUNTER_U64(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, \
538 descr);
539 STATNODE_ULONG(neg, numneg, "Number of negative cache entries");
540 STATNODE_ULONG(count, numcache, "Number of cache entries");
541 STATNODE_COUNTER(heldvnodes, numcachehv, "Number of namecache entries with vnodes held");
542 STATNODE_COUNTER(drops, numdrops, "Number of dropped entries due to reaching the limit");
543 STATNODE_COUNTER(dothits, dothits, "Number of '.' hits");
544 STATNODE_COUNTER(dotdothis, dotdothits, "Number of '..' hits");
545 STATNODE_COUNTER(miss, nummiss, "Number of cache misses");
546 STATNODE_COUNTER(misszap, nummisszap, "Number of cache misses we do not want to cache");
547 STATNODE_COUNTER(posszaps, numposzaps,
548 "Number of cache hits (positive) we do not want to cache");
549 STATNODE_COUNTER(poshits, numposhits, "Number of cache hits (positive)");
550 STATNODE_COUNTER(negzaps, numnegzaps,
551 "Number of cache hits (negative) we do not want to cache");
552 STATNODE_COUNTER(neghits, numneghits, "Number of cache hits (negative)");
553 /* These count for vn_getcwd(), too. */
554 STATNODE_COUNTER(fullpathcalls, numfullpathcalls, "Number of fullpath search calls");
555 STATNODE_COUNTER(fullpathfail1, numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
556 STATNODE_COUNTER(fullpathfail2, numfullpathfail2,
557 "Number of fullpath search errors (VOP_VPTOCNP failures)");
558 STATNODE_COUNTER(fullpathfail4, numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
559 STATNODE_COUNTER(fullpathfound, numfullpathfound, "Number of successful fullpath calls");
560 STATNODE_COUNTER(symlinktoobig, symlinktoobig, "Number of times symlink did not fit the cache");
561
562 /*
563 * Debug or developer statistics.
564 */
565 static SYSCTL_NODE(_vfs_cache, OID_AUTO, debug, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
566 "Name cache debugging");
567 #define DEBUGNODE_ULONG(name, varname, descr) \
568 SYSCTL_ULONG(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
569 #define DEBUGNODE_COUNTER(name, varname, descr) \
570 static COUNTER_U64_DEFINE_EARLY(varname); \
571 SYSCTL_COUNTER_U64(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, \
572 descr);
573 DEBUGNODE_COUNTER(zap_bucket_relock_success, zap_bucket_relock_success,
574 "Number of successful removals after relocking");
575 static long zap_bucket_fail;
576 DEBUGNODE_ULONG(zap_bucket_fail, zap_bucket_fail, "");
577 static long zap_bucket_fail2;
578 DEBUGNODE_ULONG(zap_bucket_fail2, zap_bucket_fail2, "");
579 static long cache_lock_vnodes_cel_3_failures;
580 DEBUGNODE_ULONG(vnodes_cel_3_failures, cache_lock_vnodes_cel_3_failures,
581 "Number of times 3-way vnode locking failed");
582
583 static void cache_zap_locked(struct namecache *ncp);
584 static int vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
585 char **retbuf, size_t *buflen, size_t addend);
586 static int vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf,
587 char **retbuf, size_t *buflen);
588 static int vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf,
589 char **retbuf, size_t *len, size_t addend);
590
591 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
592
593 static inline void
594 cache_assert_vlp_locked(struct mtx *vlp)
595 {
596
597 if (vlp != NULL)
598 mtx_assert(vlp, MA_OWNED);
599 }
600
601 static inline void
602 cache_assert_vnode_locked(struct vnode *vp)
603 {
604 struct mtx *vlp;
605
606 vlp = VP2VNODELOCK(vp);
607 cache_assert_vlp_locked(vlp);
608 }
609
610 /*
611 * Directory vnodes with entries are held for two reasons:
612 * 1. make them less of a target for reclamation in vnlru
613 * 2. suffer smaller performance penalty in locked lookup as requeieing is avoided
614 *
615 * It will be feasible to stop doing it altogether if all filesystems start
616 * supporting lockless lookup.
617 */
618 static void
619 cache_hold_vnode(struct vnode *vp)
620 {
621
622 cache_assert_vnode_locked(vp);
623 VNPASS(LIST_EMPTY(&vp->v_cache_src), vp);
624 vhold(vp);
625 counter_u64_add(numcachehv, 1);
626 }
627
628 static void
629 cache_drop_vnode(struct vnode *vp)
630 {
631
632 /*
633 * Called after all locks are dropped, meaning we can't assert
634 * on the state of v_cache_src.
635 */
636 vdrop(vp);
637 counter_u64_add(numcachehv, -1);
638 }
639
640 /*
641 * UMA zones.
642 */
643 static uma_zone_t __read_mostly cache_zone_small;
644 static uma_zone_t __read_mostly cache_zone_small_ts;
645 static uma_zone_t __read_mostly cache_zone_large;
646 static uma_zone_t __read_mostly cache_zone_large_ts;
647
648 char *
649 cache_symlink_alloc(size_t size, int flags)
650 {
651
652 if (size < CACHE_ZONE_SMALL_SIZE) {
653 return (uma_zalloc_smr(cache_zone_small, flags));
654 }
655 if (size < CACHE_ZONE_LARGE_SIZE) {
656 return (uma_zalloc_smr(cache_zone_large, flags));
657 }
658 counter_u64_add(symlinktoobig, 1);
659 SDT_PROBE1(vfs, namecache, symlink, alloc__fail, size);
660 return (NULL);
661 }
662
663 void
664 cache_symlink_free(char *string, size_t size)
665 {
666
667 MPASS(string != NULL);
668 KASSERT(size < CACHE_ZONE_LARGE_SIZE,
669 ("%s: size %zu too big", __func__, size));
670
671 if (size < CACHE_ZONE_SMALL_SIZE) {
672 uma_zfree_smr(cache_zone_small, string);
673 return;
674 }
675 if (size < CACHE_ZONE_LARGE_SIZE) {
676 uma_zfree_smr(cache_zone_large, string);
677 return;
678 }
679 __assert_unreachable();
680 }
681
682 static struct namecache *
683 cache_alloc_uma(int len, bool ts)
684 {
685 struct namecache_ts *ncp_ts;
686 struct namecache *ncp;
687
688 if (__predict_false(ts)) {
689 if (len <= CACHE_PATH_CUTOFF)
690 ncp_ts = uma_zalloc_smr(cache_zone_small_ts, M_WAITOK);
691 else
692 ncp_ts = uma_zalloc_smr(cache_zone_large_ts, M_WAITOK);
693 ncp = &ncp_ts->nc_nc;
694 } else {
695 if (len <= CACHE_PATH_CUTOFF)
696 ncp = uma_zalloc_smr(cache_zone_small, M_WAITOK);
697 else
698 ncp = uma_zalloc_smr(cache_zone_large, M_WAITOK);
699 }
700 return (ncp);
701 }
702
703 static void
704 cache_free_uma(struct namecache *ncp)
705 {
706 struct namecache_ts *ncp_ts;
707
708 if (__predict_false(ncp->nc_flag & NCF_TS)) {
709 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
710 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
711 uma_zfree_smr(cache_zone_small_ts, ncp_ts);
712 else
713 uma_zfree_smr(cache_zone_large_ts, ncp_ts);
714 } else {
715 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
716 uma_zfree_smr(cache_zone_small, ncp);
717 else
718 uma_zfree_smr(cache_zone_large, ncp);
719 }
720 }
721
722 static struct namecache *
723 cache_alloc(int len, bool ts)
724 {
725 u_long lnumcache;
726
727 /*
728 * Avoid blowout in namecache entries.
729 *
730 * Bugs:
731 * 1. filesystems may end up trying to add an already existing entry
732 * (for example this can happen after a cache miss during concurrent
733 * lookup), in which case we will call cache_neg_evict despite not
734 * adding anything.
735 * 2. the routine may fail to free anything and no provisions are made
736 * to make it try harder (see the inside for failure modes)
737 * 3. it only ever looks at negative entries.
738 */
739 lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
740 if (cache_neg_evict_cond(lnumcache)) {
741 lnumcache = atomic_load_long(&numcache);
742 }
743 if (__predict_false(lnumcache >= ncsize)) {
744 atomic_subtract_long(&numcache, 1);
745 counter_u64_add(numdrops, 1);
746 return (NULL);
747 }
748 return (cache_alloc_uma(len, ts));
749 }
750
751 static void
752 cache_free(struct namecache *ncp)
753 {
754
755 MPASS(ncp != NULL);
756 if ((ncp->nc_flag & NCF_DVDROP) != 0) {
757 cache_drop_vnode(ncp->nc_dvp);
758 }
759 cache_free_uma(ncp);
760 atomic_subtract_long(&numcache, 1);
761 }
762
763 static void
764 cache_free_batch(struct cache_freebatch *batch)
765 {
766 struct namecache *ncp, *nnp;
767 int i;
768
769 i = 0;
770 if (TAILQ_EMPTY(batch))
771 goto out;
772 TAILQ_FOREACH_SAFE(ncp, batch, nc_dst, nnp) {
773 if ((ncp->nc_flag & NCF_DVDROP) != 0) {
774 cache_drop_vnode(ncp->nc_dvp);
775 }
776 cache_free_uma(ncp);
777 i++;
778 }
779 atomic_subtract_long(&numcache, i);
780 out:
781 SDT_PROBE1(vfs, namecache, purge, batch, i);
782 }
783
784 /*
785 * Hashing.
786 *
787 * The code was made to use FNV in 2001 and this choice needs to be revisited.
788 *
789 * Short summary of the difficulty:
790 * The longest name which can be inserted is NAME_MAX characters in length (or
791 * 255 at the time of writing this comment), while majority of names used in
792 * practice are significantly shorter (mostly below 10). More importantly
793 * majority of lookups performed find names are even shorter than that.
794 *
795 * This poses a problem where hashes which do better than FNV past word size
796 * (or so) tend to come with additional overhead when finalizing the result,
797 * making them noticeably slower for the most commonly used range.
798 *
799 * Consider a path like: /usr/obj/usr/src/sys/amd64/GENERIC/vnode_if.c
800 *
801 * When looking it up the most time consuming part by a large margin (at least
802 * on amd64) is hashing. Replacing FNV with something which pessimizes short
803 * input would make the slowest part stand out even more.
804 */
805
806 /*
807 * TODO: With the value stored we can do better than computing the hash based
808 * on the address.
809 */
810 static void
811 cache_prehash(struct vnode *vp)
812 {
813
814 vp->v_nchash = fnv_32_buf(&vp, sizeof(vp), FNV1_32_INIT);
815 }
816
817 static uint32_t
818 cache_get_hash(char *name, u_char len, struct vnode *dvp)
819 {
820
821 return (fnv_32_buf(name, len, dvp->v_nchash));
822 }
823
824 static uint32_t
825 cache_get_hash_iter_start(struct vnode *dvp)
826 {
827
828 return (dvp->v_nchash);
829 }
830
831 static uint32_t
832 cache_get_hash_iter(char c, uint32_t hash)
833 {
834
835 return (fnv_32_buf(&c, 1, hash));
836 }
837
838 static uint32_t
839 cache_get_hash_iter_finish(uint32_t hash)
840 {
841
842 return (hash);
843 }
844
845 static inline struct nchashhead *
846 NCP2BUCKET(struct namecache *ncp)
847 {
848 uint32_t hash;
849
850 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
851 return (NCHHASH(hash));
852 }
853
854 static inline struct mtx *
855 NCP2BUCKETLOCK(struct namecache *ncp)
856 {
857 uint32_t hash;
858
859 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
860 return (HASH2BUCKETLOCK(hash));
861 }
862
863 #ifdef INVARIANTS
864 static void
865 cache_assert_bucket_locked(struct namecache *ncp)
866 {
867 struct mtx *blp;
868
869 blp = NCP2BUCKETLOCK(ncp);
870 mtx_assert(blp, MA_OWNED);
871 }
872
873 static void
874 cache_assert_bucket_unlocked(struct namecache *ncp)
875 {
876 struct mtx *blp;
877
878 blp = NCP2BUCKETLOCK(ncp);
879 mtx_assert(blp, MA_NOTOWNED);
880 }
881 #else
882 #define cache_assert_bucket_locked(x) do { } while (0)
883 #define cache_assert_bucket_unlocked(x) do { } while (0)
884 #endif
885
886 #define cache_sort_vnodes(x, y) _cache_sort_vnodes((void **)(x), (void **)(y))
887 static void
888 _cache_sort_vnodes(void **p1, void **p2)
889 {
890 void *tmp;
891
892 MPASS(*p1 != NULL || *p2 != NULL);
893
894 if (*p1 > *p2) {
895 tmp = *p2;
896 *p2 = *p1;
897 *p1 = tmp;
898 }
899 }
900
901 static void
902 cache_lock_all_buckets(void)
903 {
904 u_int i;
905
906 for (i = 0; i < numbucketlocks; i++)
907 mtx_lock(&bucketlocks[i]);
908 }
909
910 static void
911 cache_unlock_all_buckets(void)
912 {
913 u_int i;
914
915 for (i = 0; i < numbucketlocks; i++)
916 mtx_unlock(&bucketlocks[i]);
917 }
918
919 static void
920 cache_lock_all_vnodes(void)
921 {
922 u_int i;
923
924 for (i = 0; i < numvnodelocks; i++)
925 mtx_lock(&vnodelocks[i]);
926 }
927
928 static void
929 cache_unlock_all_vnodes(void)
930 {
931 u_int i;
932
933 for (i = 0; i < numvnodelocks; i++)
934 mtx_unlock(&vnodelocks[i]);
935 }
936
937 static int
938 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
939 {
940
941 cache_sort_vnodes(&vlp1, &vlp2);
942
943 if (vlp1 != NULL) {
944 if (!mtx_trylock(vlp1))
945 return (EAGAIN);
946 }
947 if (!mtx_trylock(vlp2)) {
948 if (vlp1 != NULL)
949 mtx_unlock(vlp1);
950 return (EAGAIN);
951 }
952
953 return (0);
954 }
955
956 static void
957 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
958 {
959
960 MPASS(vlp1 != NULL || vlp2 != NULL);
961 MPASS(vlp1 <= vlp2);
962
963 if (vlp1 != NULL)
964 mtx_lock(vlp1);
965 if (vlp2 != NULL)
966 mtx_lock(vlp2);
967 }
968
969 static void
970 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
971 {
972
973 MPASS(vlp1 != NULL || vlp2 != NULL);
974
975 if (vlp1 != NULL)
976 mtx_unlock(vlp1);
977 if (vlp2 != NULL)
978 mtx_unlock(vlp2);
979 }
980
981 static int
982 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
983 {
984 struct nchstats snap;
985
986 if (req->oldptr == NULL)
987 return (SYSCTL_OUT(req, 0, sizeof(snap)));
988
989 snap = nchstats;
990 snap.ncs_goodhits = counter_u64_fetch(numposhits);
991 snap.ncs_neghits = counter_u64_fetch(numneghits);
992 snap.ncs_badhits = counter_u64_fetch(numposzaps) +
993 counter_u64_fetch(numnegzaps);
994 snap.ncs_miss = counter_u64_fetch(nummisszap) +
995 counter_u64_fetch(nummiss);
996
997 return (SYSCTL_OUT(req, &snap, sizeof(snap)));
998 }
999 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
1000 CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
1001 "VFS cache effectiveness statistics");
1002
1003 static void
1004 cache_recalc_neg_min(u_int val)
1005 {
1006
1007 neg_min = (ncsize * val) / 100;
1008 }
1009
1010 static int
1011 sysctl_negminpct(SYSCTL_HANDLER_ARGS)
1012 {
1013 u_int val;
1014 int error;
1015
1016 val = ncnegminpct;
1017 error = sysctl_handle_int(oidp, &val, 0, req);
1018 if (error != 0 || req->newptr == NULL)
1019 return (error);
1020
1021 if (val == ncnegminpct)
1022 return (0);
1023 if (val < 0 || val > 99)
1024 return (EINVAL);
1025 ncnegminpct = val;
1026 cache_recalc_neg_min(val);
1027 return (0);
1028 }
1029
1030 SYSCTL_PROC(_vfs_cache_param, OID_AUTO, negminpct,
1031 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_negminpct,
1032 "I", "Negative entry \% of namecache capacity above which automatic eviction is allowed");
1033
1034 #ifdef DIAGNOSTIC
1035 /*
1036 * Grab an atomic snapshot of the name cache hash chain lengths
1037 */
1038 static SYSCTL_NODE(_debug, OID_AUTO, hashstat,
1039 CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
1040 "hash table stats");
1041
1042 static int
1043 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
1044 {
1045 struct nchashhead *ncpp;
1046 struct namecache *ncp;
1047 int i, error, n_nchash, *cntbuf;
1048
1049 retry:
1050 n_nchash = nchash + 1; /* nchash is max index, not count */
1051 if (req->oldptr == NULL)
1052 return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
1053 cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
1054 cache_lock_all_buckets();
1055 if (n_nchash != nchash + 1) {
1056 cache_unlock_all_buckets();
1057 free(cntbuf, M_TEMP);
1058 goto retry;
1059 }
1060 /* Scan hash tables counting entries */
1061 for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
1062 CK_SLIST_FOREACH(ncp, ncpp, nc_hash)
1063 cntbuf[i]++;
1064 cache_unlock_all_buckets();
1065 for (error = 0, i = 0; i < n_nchash; i++)
1066 if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
1067 break;
1068 free(cntbuf, M_TEMP);
1069 return (error);
1070 }
1071 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
1072 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
1073 "nchash chain lengths");
1074
1075 static int
1076 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
1077 {
1078 int error;
1079 struct nchashhead *ncpp;
1080 struct namecache *ncp;
1081 int n_nchash;
1082 int count, maxlength, used, pct;
1083
1084 if (!req->oldptr)
1085 return SYSCTL_OUT(req, 0, 4 * sizeof(int));
1086
1087 cache_lock_all_buckets();
1088 n_nchash = nchash + 1; /* nchash is max index, not count */
1089 used = 0;
1090 maxlength = 0;
1091
1092 /* Scan hash tables for applicable entries */
1093 for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
1094 count = 0;
1095 CK_SLIST_FOREACH(ncp, ncpp, nc_hash) {
1096 count++;
1097 }
1098 if (count)
1099 used++;
1100 if (maxlength < count)
1101 maxlength = count;
1102 }
1103 n_nchash = nchash + 1;
1104 cache_unlock_all_buckets();
1105 pct = (used * 100) / (n_nchash / 100);
1106 error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
1107 if (error)
1108 return (error);
1109 error = SYSCTL_OUT(req, &used, sizeof(used));
1110 if (error)
1111 return (error);
1112 error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
1113 if (error)
1114 return (error);
1115 error = SYSCTL_OUT(req, &pct, sizeof(pct));
1116 if (error)
1117 return (error);
1118 return (0);
1119 }
1120 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
1121 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
1122 "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
1123 #endif
1124
1125 /*
1126 * Negative entries management
1127 *
1128 * Various workloads create plenty of negative entries and barely use them
1129 * afterwards. Moreover malicious users can keep performing bogus lookups
1130 * adding even more entries. For example "make tinderbox" as of writing this
1131 * comment ends up with 2.6M namecache entries in total, 1.2M of which are
1132 * negative.
1133 *
1134 * As such, a rather aggressive eviction method is needed. The currently
1135 * employed method is a placeholder.
1136 *
1137 * Entries are split over numneglists separate lists, each of which is further
1138 * split into hot and cold entries. Entries get promoted after getting a hit.
1139 * Eviction happens on addition of new entry.
1140 */
1141 static SYSCTL_NODE(_vfs_cache, OID_AUTO, neg, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1142 "Name cache negative entry statistics");
1143
1144 SYSCTL_ULONG(_vfs_cache_neg, OID_AUTO, count, CTLFLAG_RD, &numneg, 0,
1145 "Number of negative cache entries");
1146
1147 static COUNTER_U64_DEFINE_EARLY(neg_created);
1148 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, created, CTLFLAG_RD, &neg_created,
1149 "Number of created negative entries");
1150
1151 static COUNTER_U64_DEFINE_EARLY(neg_evicted);
1152 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evicted, CTLFLAG_RD, &neg_evicted,
1153 "Number of evicted negative entries");
1154
1155 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_empty);
1156 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_empty, CTLFLAG_RD,
1157 &neg_evict_skipped_empty,
1158 "Number of times evicting failed due to lack of entries");
1159
1160 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_missed);
1161 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_missed, CTLFLAG_RD,
1162 &neg_evict_skipped_missed,
1163 "Number of times evicting failed due to target entry disappearing");
1164
1165 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_contended);
1166 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_contended, CTLFLAG_RD,
1167 &neg_evict_skipped_contended,
1168 "Number of times evicting failed due to contention");
1169
1170 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, hits, CTLFLAG_RD, &numneghits,
1171 "Number of cache hits (negative)");
1172
1173 static int
1174 sysctl_neg_hot(SYSCTL_HANDLER_ARGS)
1175 {
1176 int i, out;
1177
1178 out = 0;
1179 for (i = 0; i < numneglists; i++)
1180 out += neglists[i].nl_hotnum;
1181
1182 return (SYSCTL_OUT(req, &out, sizeof(out)));
1183 }
1184 SYSCTL_PROC(_vfs_cache_neg, OID_AUTO, hot, CTLTYPE_INT | CTLFLAG_RD |
1185 CTLFLAG_MPSAFE, 0, 0, sysctl_neg_hot, "I",
1186 "Number of hot negative entries");
1187
1188 static void
1189 cache_neg_init(struct namecache *ncp)
1190 {
1191 struct negstate *ns;
1192
1193 ncp->nc_flag |= NCF_NEGATIVE;
1194 ns = NCP2NEGSTATE(ncp);
1195 ns->neg_flag = 0;
1196 ns->neg_hit = 0;
1197 counter_u64_add(neg_created, 1);
1198 }
1199
1200 #define CACHE_NEG_PROMOTION_THRESH 2
1201
1202 static bool
1203 cache_neg_hit_prep(struct namecache *ncp)
1204 {
1205 struct negstate *ns;
1206 u_char n;
1207
1208 ns = NCP2NEGSTATE(ncp);
1209 n = atomic_load_char(&ns->neg_hit);
1210 for (;;) {
1211 if (n >= CACHE_NEG_PROMOTION_THRESH)
1212 return (false);
1213 if (atomic_fcmpset_8(&ns->neg_hit, &n, n + 1))
1214 break;
1215 }
1216 return (n + 1 == CACHE_NEG_PROMOTION_THRESH);
1217 }
1218
1219 /*
1220 * Nothing to do here but it is provided for completeness as some
1221 * cache_neg_hit_prep callers may end up returning without even
1222 * trying to promote.
1223 */
1224 #define cache_neg_hit_abort(ncp) do { } while (0)
1225
1226 static void
1227 cache_neg_hit_finish(struct namecache *ncp)
1228 {
1229
1230 SDT_PROBE2(vfs, namecache, lookup, hit__negative, ncp->nc_dvp, ncp->nc_name);
1231 counter_u64_add(numneghits, 1);
1232 }
1233
1234 /*
1235 * Move a negative entry to the hot list.
1236 */
1237 static void
1238 cache_neg_promote_locked(struct namecache *ncp)
1239 {
1240 struct neglist *nl;
1241 struct negstate *ns;
1242
1243 ns = NCP2NEGSTATE(ncp);
1244 nl = NCP2NEGLIST(ncp);
1245 mtx_assert(&nl->nl_lock, MA_OWNED);
1246 if ((ns->neg_flag & NEG_HOT) == 0) {
1247 TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1248 TAILQ_INSERT_TAIL(&nl->nl_hotlist, ncp, nc_dst);
1249 nl->nl_hotnum++;
1250 ns->neg_flag |= NEG_HOT;
1251 }
1252 }
1253
1254 /*
1255 * Move a hot negative entry to the cold list.
1256 */
1257 static void
1258 cache_neg_demote_locked(struct namecache *ncp)
1259 {
1260 struct neglist *nl;
1261 struct negstate *ns;
1262
1263 ns = NCP2NEGSTATE(ncp);
1264 nl = NCP2NEGLIST(ncp);
1265 mtx_assert(&nl->nl_lock, MA_OWNED);
1266 MPASS(ns->neg_flag & NEG_HOT);
1267 TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1268 TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1269 nl->nl_hotnum--;
1270 ns->neg_flag &= ~NEG_HOT;
1271 atomic_store_char(&ns->neg_hit, 0);
1272 }
1273
1274 /*
1275 * Move a negative entry to the hot list if it matches the lookup.
1276 *
1277 * We have to take locks, but they may be contended and in the worst
1278 * case we may need to go off CPU. We don't want to spin within the
1279 * smr section and we can't block with it. Exiting the section means
1280 * the found entry could have been evicted. We are going to look it
1281 * up again.
1282 */
1283 static bool
1284 cache_neg_promote_cond(struct vnode *dvp, struct componentname *cnp,
1285 struct namecache *oncp, uint32_t hash)
1286 {
1287 struct namecache *ncp;
1288 struct neglist *nl;
1289 u_char nc_flag;
1290
1291 nl = NCP2NEGLIST(oncp);
1292
1293 mtx_lock(&nl->nl_lock);
1294 /*
1295 * For hash iteration.
1296 */
1297 vfs_smr_enter();
1298
1299 /*
1300 * Avoid all surprises by only succeeding if we got the same entry and
1301 * bailing completely otherwise.
1302 * XXX There are no provisions to keep the vnode around, meaning we may
1303 * end up promoting a negative entry for a *new* vnode and returning
1304 * ENOENT on its account. This is the error we want to return anyway
1305 * and promotion is harmless.
1306 *
1307 * In particular at this point there can be a new ncp which matches the
1308 * search but hashes to a different neglist.
1309 */
1310 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1311 if (ncp == oncp)
1312 break;
1313 }
1314
1315 /*
1316 * No match to begin with.
1317 */
1318 if (__predict_false(ncp == NULL)) {
1319 goto out_abort;
1320 }
1321
1322 /*
1323 * The newly found entry may be something different...
1324 */
1325 if (!(ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1326 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))) {
1327 goto out_abort;
1328 }
1329
1330 /*
1331 * ... and not even negative.
1332 */
1333 nc_flag = atomic_load_char(&ncp->nc_flag);
1334 if ((nc_flag & NCF_NEGATIVE) == 0) {
1335 goto out_abort;
1336 }
1337
1338 if (!cache_ncp_canuse(ncp)) {
1339 goto out_abort;
1340 }
1341
1342 cache_neg_promote_locked(ncp);
1343 cache_neg_hit_finish(ncp);
1344 vfs_smr_exit();
1345 mtx_unlock(&nl->nl_lock);
1346 return (true);
1347 out_abort:
1348 vfs_smr_exit();
1349 mtx_unlock(&nl->nl_lock);
1350 return (false);
1351 }
1352
1353 static void
1354 cache_neg_promote(struct namecache *ncp)
1355 {
1356 struct neglist *nl;
1357
1358 nl = NCP2NEGLIST(ncp);
1359 mtx_lock(&nl->nl_lock);
1360 cache_neg_promote_locked(ncp);
1361 mtx_unlock(&nl->nl_lock);
1362 }
1363
1364 static void
1365 cache_neg_insert(struct namecache *ncp)
1366 {
1367 struct neglist *nl;
1368
1369 MPASS(ncp->nc_flag & NCF_NEGATIVE);
1370 cache_assert_bucket_locked(ncp);
1371 nl = NCP2NEGLIST(ncp);
1372 mtx_lock(&nl->nl_lock);
1373 TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1374 mtx_unlock(&nl->nl_lock);
1375 atomic_add_long(&numneg, 1);
1376 }
1377
1378 static void
1379 cache_neg_remove(struct namecache *ncp)
1380 {
1381 struct neglist *nl;
1382 struct negstate *ns;
1383
1384 cache_assert_bucket_locked(ncp);
1385 nl = NCP2NEGLIST(ncp);
1386 ns = NCP2NEGSTATE(ncp);
1387 mtx_lock(&nl->nl_lock);
1388 if ((ns->neg_flag & NEG_HOT) != 0) {
1389 TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1390 nl->nl_hotnum--;
1391 } else {
1392 TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1393 }
1394 mtx_unlock(&nl->nl_lock);
1395 atomic_subtract_long(&numneg, 1);
1396 }
1397
1398 static struct neglist *
1399 cache_neg_evict_select_list(void)
1400 {
1401 struct neglist *nl;
1402 u_int c;
1403
1404 c = atomic_fetchadd_int(&neg_cycle, 1) + 1;
1405 nl = &neglists[c % numneglists];
1406 if (!mtx_trylock(&nl->nl_evict_lock)) {
1407 counter_u64_add(neg_evict_skipped_contended, 1);
1408 return (NULL);
1409 }
1410 return (nl);
1411 }
1412
1413 static struct namecache *
1414 cache_neg_evict_select_entry(struct neglist *nl)
1415 {
1416 struct namecache *ncp, *lncp;
1417 struct negstate *ns, *lns;
1418 int i;
1419
1420 mtx_assert(&nl->nl_evict_lock, MA_OWNED);
1421 mtx_assert(&nl->nl_lock, MA_OWNED);
1422 ncp = TAILQ_FIRST(&nl->nl_list);
1423 if (ncp == NULL)
1424 return (NULL);
1425 lncp = ncp;
1426 lns = NCP2NEGSTATE(lncp);
1427 for (i = 1; i < 4; i++) {
1428 ncp = TAILQ_NEXT(ncp, nc_dst);
1429 if (ncp == NULL)
1430 break;
1431 ns = NCP2NEGSTATE(ncp);
1432 if (ns->neg_hit < lns->neg_hit) {
1433 lncp = ncp;
1434 lns = ns;
1435 }
1436 }
1437 return (lncp);
1438 }
1439
1440 static bool
1441 cache_neg_evict(void)
1442 {
1443 struct namecache *ncp, *ncp2;
1444 struct neglist *nl;
1445 struct vnode *dvp;
1446 struct mtx *dvlp;
1447 struct mtx *blp;
1448 uint32_t hash;
1449 u_char nlen;
1450 bool evicted;
1451
1452 nl = cache_neg_evict_select_list();
1453 if (nl == NULL) {
1454 return (false);
1455 }
1456
1457 mtx_lock(&nl->nl_lock);
1458 ncp = TAILQ_FIRST(&nl->nl_hotlist);
1459 if (ncp != NULL) {
1460 cache_neg_demote_locked(ncp);
1461 }
1462 ncp = cache_neg_evict_select_entry(nl);
1463 if (ncp == NULL) {
1464 counter_u64_add(neg_evict_skipped_empty, 1);
1465 mtx_unlock(&nl->nl_lock);
1466 mtx_unlock(&nl->nl_evict_lock);
1467 return (false);
1468 }
1469 nlen = ncp->nc_nlen;
1470 dvp = ncp->nc_dvp;
1471 hash = cache_get_hash(ncp->nc_name, nlen, dvp);
1472 dvlp = VP2VNODELOCK(dvp);
1473 blp = HASH2BUCKETLOCK(hash);
1474 mtx_unlock(&nl->nl_lock);
1475 mtx_unlock(&nl->nl_evict_lock);
1476 mtx_lock(dvlp);
1477 mtx_lock(blp);
1478 /*
1479 * Note that since all locks were dropped above, the entry may be
1480 * gone or reallocated to be something else.
1481 */
1482 CK_SLIST_FOREACH(ncp2, (NCHHASH(hash)), nc_hash) {
1483 if (ncp2 == ncp && ncp2->nc_dvp == dvp &&
1484 ncp2->nc_nlen == nlen && (ncp2->nc_flag & NCF_NEGATIVE) != 0)
1485 break;
1486 }
1487 if (ncp2 == NULL) {
1488 counter_u64_add(neg_evict_skipped_missed, 1);
1489 ncp = NULL;
1490 evicted = false;
1491 } else {
1492 MPASS(dvlp == VP2VNODELOCK(ncp->nc_dvp));
1493 MPASS(blp == NCP2BUCKETLOCK(ncp));
1494 SDT_PROBE2(vfs, namecache, evict_negative, done, ncp->nc_dvp,
1495 ncp->nc_name);
1496 cache_zap_locked(ncp);
1497 counter_u64_add(neg_evicted, 1);
1498 evicted = true;
1499 }
1500 mtx_unlock(blp);
1501 mtx_unlock(dvlp);
1502 if (ncp != NULL)
1503 cache_free(ncp);
1504 return (evicted);
1505 }
1506
1507 /*
1508 * Maybe evict a negative entry to create more room.
1509 *
1510 * The ncnegfactor parameter limits what fraction of the total count
1511 * can comprise of negative entries. However, if the cache is just
1512 * warming up this leads to excessive evictions. As such, ncnegminpct
1513 * (recomputed to neg_min) dictates whether the above should be
1514 * applied.
1515 *
1516 * Try evicting if the cache is close to full capacity regardless of
1517 * other considerations.
1518 */
1519 static bool
1520 cache_neg_evict_cond(u_long lnumcache)
1521 {
1522 u_long lnumneg;
1523
1524 if (ncsize - 1000 < lnumcache)
1525 goto out_evict;
1526 lnumneg = atomic_load_long(&numneg);
1527 if (lnumneg < neg_min)
1528 return (false);
1529 if (lnumneg * ncnegfactor < lnumcache)
1530 return (false);
1531 out_evict:
1532 return (cache_neg_evict());
1533 }
1534
1535 /*
1536 * cache_zap_locked():
1537 *
1538 * Removes a namecache entry from cache, whether it contains an actual
1539 * pointer to a vnode or if it is just a negative cache entry.
1540 */
1541 static void
1542 cache_zap_locked(struct namecache *ncp)
1543 {
1544 struct nchashhead *ncpp;
1545 struct vnode *dvp, *vp;
1546
1547 dvp = ncp->nc_dvp;
1548 vp = ncp->nc_vp;
1549
1550 if (!(ncp->nc_flag & NCF_NEGATIVE))
1551 cache_assert_vnode_locked(vp);
1552 cache_assert_vnode_locked(dvp);
1553 cache_assert_bucket_locked(ncp);
1554
1555 cache_ncp_invalidate(ncp);
1556
1557 ncpp = NCP2BUCKET(ncp);
1558 CK_SLIST_REMOVE(ncpp, ncp, namecache, nc_hash);
1559 if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1560 SDT_PROBE3(vfs, namecache, zap, done, dvp, ncp->nc_name, vp);
1561 TAILQ_REMOVE(&vp->v_cache_dst, ncp, nc_dst);
1562 if (ncp == vp->v_cache_dd) {
1563 atomic_store_ptr(&vp->v_cache_dd, NULL);
1564 }
1565 } else {
1566 SDT_PROBE2(vfs, namecache, zap_negative, done, dvp, ncp->nc_name);
1567 cache_neg_remove(ncp);
1568 }
1569 if (ncp->nc_flag & NCF_ISDOTDOT) {
1570 if (ncp == dvp->v_cache_dd) {
1571 atomic_store_ptr(&dvp->v_cache_dd, NULL);
1572 }
1573 } else {
1574 LIST_REMOVE(ncp, nc_src);
1575 if (LIST_EMPTY(&dvp->v_cache_src)) {
1576 ncp->nc_flag |= NCF_DVDROP;
1577 }
1578 }
1579 }
1580
1581 static void
1582 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
1583 {
1584 struct mtx *blp;
1585
1586 MPASS(ncp->nc_dvp == vp);
1587 MPASS(ncp->nc_flag & NCF_NEGATIVE);
1588 cache_assert_vnode_locked(vp);
1589
1590 blp = NCP2BUCKETLOCK(ncp);
1591 mtx_lock(blp);
1592 cache_zap_locked(ncp);
1593 mtx_unlock(blp);
1594 }
1595
1596 static bool
1597 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
1598 struct mtx **vlpp)
1599 {
1600 struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
1601 struct mtx *blp;
1602
1603 MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
1604 cache_assert_vnode_locked(vp);
1605
1606 if (ncp->nc_flag & NCF_NEGATIVE) {
1607 if (*vlpp != NULL) {
1608 mtx_unlock(*vlpp);
1609 *vlpp = NULL;
1610 }
1611 cache_zap_negative_locked_vnode_kl(ncp, vp);
1612 return (true);
1613 }
1614
1615 pvlp = VP2VNODELOCK(vp);
1616 blp = NCP2BUCKETLOCK(ncp);
1617 vlp1 = VP2VNODELOCK(ncp->nc_dvp);
1618 vlp2 = VP2VNODELOCK(ncp->nc_vp);
1619
1620 if (*vlpp == vlp1 || *vlpp == vlp2) {
1621 to_unlock = *vlpp;
1622 *vlpp = NULL;
1623 } else {
1624 if (*vlpp != NULL) {
1625 mtx_unlock(*vlpp);
1626 *vlpp = NULL;
1627 }
1628 cache_sort_vnodes(&vlp1, &vlp2);
1629 if (vlp1 == pvlp) {
1630 mtx_lock(vlp2);
1631 to_unlock = vlp2;
1632 } else {
1633 if (!mtx_trylock(vlp1))
1634 goto out_relock;
1635 to_unlock = vlp1;
1636 }
1637 }
1638 mtx_lock(blp);
1639 cache_zap_locked(ncp);
1640 mtx_unlock(blp);
1641 if (to_unlock != NULL)
1642 mtx_unlock(to_unlock);
1643 return (true);
1644
1645 out_relock:
1646 mtx_unlock(vlp2);
1647 mtx_lock(vlp1);
1648 mtx_lock(vlp2);
1649 MPASS(*vlpp == NULL);
1650 *vlpp = vlp1;
1651 return (false);
1652 }
1653
1654 /*
1655 * If trylocking failed we can get here. We know enough to take all needed locks
1656 * in the right order and re-lookup the entry.
1657 */
1658 static int
1659 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1660 struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1661 struct mtx *blp)
1662 {
1663 struct namecache *rncp;
1664
1665 cache_assert_bucket_unlocked(ncp);
1666
1667 cache_sort_vnodes(&dvlp, &vlp);
1668 cache_lock_vnodes(dvlp, vlp);
1669 mtx_lock(blp);
1670 CK_SLIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1671 if (rncp == ncp && rncp->nc_dvp == dvp &&
1672 rncp->nc_nlen == cnp->cn_namelen &&
1673 !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1674 break;
1675 }
1676 if (rncp != NULL) {
1677 cache_zap_locked(rncp);
1678 mtx_unlock(blp);
1679 cache_unlock_vnodes(dvlp, vlp);
1680 counter_u64_add(zap_bucket_relock_success, 1);
1681 return (0);
1682 }
1683
1684 mtx_unlock(blp);
1685 cache_unlock_vnodes(dvlp, vlp);
1686 return (EAGAIN);
1687 }
1688
1689 static int __noinline
1690 cache_zap_locked_bucket(struct namecache *ncp, struct componentname *cnp,
1691 uint32_t hash, struct mtx *blp)
1692 {
1693 struct mtx *dvlp, *vlp;
1694 struct vnode *dvp;
1695
1696 cache_assert_bucket_locked(ncp);
1697
1698 dvlp = VP2VNODELOCK(ncp->nc_dvp);
1699 vlp = NULL;
1700 if (!(ncp->nc_flag & NCF_NEGATIVE))
1701 vlp = VP2VNODELOCK(ncp->nc_vp);
1702 if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1703 cache_zap_locked(ncp);
1704 mtx_unlock(blp);
1705 cache_unlock_vnodes(dvlp, vlp);
1706 return (0);
1707 }
1708
1709 dvp = ncp->nc_dvp;
1710 mtx_unlock(blp);
1711 return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1712 }
1713
1714 static __noinline int
1715 cache_remove_cnp(struct vnode *dvp, struct componentname *cnp)
1716 {
1717 struct namecache *ncp;
1718 struct mtx *blp;
1719 struct mtx *dvlp, *dvlp2;
1720 uint32_t hash;
1721 int error;
1722
1723 if (cnp->cn_namelen == 2 &&
1724 cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1725 dvlp = VP2VNODELOCK(dvp);
1726 dvlp2 = NULL;
1727 mtx_lock(dvlp);
1728 retry_dotdot:
1729 ncp = dvp->v_cache_dd;
1730 if (ncp == NULL) {
1731 mtx_unlock(dvlp);
1732 if (dvlp2 != NULL)
1733 mtx_unlock(dvlp2);
1734 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1735 return (0);
1736 }
1737 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1738 if (!cache_zap_locked_vnode_kl2(ncp, dvp, &dvlp2))
1739 goto retry_dotdot;
1740 MPASS(dvp->v_cache_dd == NULL);
1741 mtx_unlock(dvlp);
1742 if (dvlp2 != NULL)
1743 mtx_unlock(dvlp2);
1744 cache_free(ncp);
1745 } else {
1746 atomic_store_ptr(&dvp->v_cache_dd, NULL);
1747 mtx_unlock(dvlp);
1748 if (dvlp2 != NULL)
1749 mtx_unlock(dvlp2);
1750 }
1751 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1752 return (1);
1753 }
1754
1755 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1756 blp = HASH2BUCKETLOCK(hash);
1757 retry:
1758 if (CK_SLIST_EMPTY(NCHHASH(hash)))
1759 goto out_no_entry;
1760
1761 mtx_lock(blp);
1762
1763 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1764 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1765 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1766 break;
1767 }
1768
1769 if (ncp == NULL) {
1770 mtx_unlock(blp);
1771 goto out_no_entry;
1772 }
1773
1774 error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1775 if (__predict_false(error != 0)) {
1776 zap_bucket_fail++;
1777 goto retry;
1778 }
1779 counter_u64_add(numposzaps, 1);
1780 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1781 cache_free(ncp);
1782 return (1);
1783 out_no_entry:
1784 counter_u64_add(nummisszap, 1);
1785 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1786 return (0);
1787 }
1788
1789 static int __noinline
1790 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1791 struct timespec *tsp, int *ticksp)
1792 {
1793 int ltype;
1794
1795 *vpp = dvp;
1796 counter_u64_add(dothits, 1);
1797 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1798 if (tsp != NULL)
1799 timespecclear(tsp);
1800 if (ticksp != NULL)
1801 *ticksp = ticks;
1802 vrefact(*vpp);
1803 /*
1804 * When we lookup "." we still can be asked to lock it
1805 * differently...
1806 */
1807 ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1808 if (ltype != VOP_ISLOCKED(*vpp)) {
1809 if (ltype == LK_EXCLUSIVE) {
1810 vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1811 if (VN_IS_DOOMED((*vpp))) {
1812 /* forced unmount */
1813 vrele(*vpp);
1814 *vpp = NULL;
1815 return (ENOENT);
1816 }
1817 } else
1818 vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1819 }
1820 return (-1);
1821 }
1822
1823 static int __noinline
1824 cache_lookup_dotdot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1825 struct timespec *tsp, int *ticksp)
1826 {
1827 struct namecache_ts *ncp_ts;
1828 struct namecache *ncp;
1829 struct mtx *dvlp;
1830 enum vgetstate vs;
1831 int error, ltype;
1832 bool whiteout;
1833
1834 MPASS((cnp->cn_flags & ISDOTDOT) != 0);
1835
1836 if ((cnp->cn_flags & MAKEENTRY) == 0) {
1837 cache_remove_cnp(dvp, cnp);
1838 return (0);
1839 }
1840
1841 counter_u64_add(dotdothits, 1);
1842 retry:
1843 dvlp = VP2VNODELOCK(dvp);
1844 mtx_lock(dvlp);
1845 ncp = dvp->v_cache_dd;
1846 if (ncp == NULL) {
1847 SDT_PROBE2(vfs, namecache, lookup, miss, dvp, "..");
1848 mtx_unlock(dvlp);
1849 return (0);
1850 }
1851 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1852 if (ncp->nc_flag & NCF_NEGATIVE)
1853 *vpp = NULL;
1854 else
1855 *vpp = ncp->nc_vp;
1856 } else
1857 *vpp = ncp->nc_dvp;
1858 if (*vpp == NULL)
1859 goto negative_success;
1860 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", *vpp);
1861 cache_out_ts(ncp, tsp, ticksp);
1862 if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1863 NCF_DTS && tsp != NULL) {
1864 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1865 *tsp = ncp_ts->nc_dotdottime;
1866 }
1867
1868 MPASS(dvp != *vpp);
1869 ltype = VOP_ISLOCKED(dvp);
1870 VOP_UNLOCK(dvp);
1871 vs = vget_prep(*vpp);
1872 mtx_unlock(dvlp);
1873 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1874 vn_lock(dvp, ltype | LK_RETRY);
1875 if (VN_IS_DOOMED(dvp)) {
1876 if (error == 0)
1877 vput(*vpp);
1878 *vpp = NULL;
1879 return (ENOENT);
1880 }
1881 if (error) {
1882 *vpp = NULL;
1883 goto retry;
1884 }
1885 return (-1);
1886 negative_success:
1887 if (__predict_false(cnp->cn_nameiop == CREATE)) {
1888 if (cnp->cn_flags & ISLASTCN) {
1889 counter_u64_add(numnegzaps, 1);
1890 cache_zap_negative_locked_vnode_kl(ncp, dvp);
1891 mtx_unlock(dvlp);
1892 cache_free(ncp);
1893 return (0);
1894 }
1895 }
1896
1897 whiteout = (ncp->nc_flag & NCF_WHITE);
1898 cache_out_ts(ncp, tsp, ticksp);
1899 if (cache_neg_hit_prep(ncp))
1900 cache_neg_promote(ncp);
1901 else
1902 cache_neg_hit_finish(ncp);
1903 mtx_unlock(dvlp);
1904 if (whiteout)
1905 cnp->cn_flags |= ISWHITEOUT;
1906 return (ENOENT);
1907 }
1908
1909 /**
1910 * Lookup a name in the name cache
1911 *
1912 * # Arguments
1913 *
1914 * - dvp: Parent directory in which to search.
1915 * - vpp: Return argument. Will contain desired vnode on cache hit.
1916 * - cnp: Parameters of the name search. The most interesting bits of
1917 * the cn_flags field have the following meanings:
1918 * - MAKEENTRY: If clear, free an entry from the cache rather than look
1919 * it up.
1920 * - ISDOTDOT: Must be set if and only if cn_nameptr == ".."
1921 * - tsp: Return storage for cache timestamp. On a successful (positive
1922 * or negative) lookup, tsp will be filled with any timespec that
1923 * was stored when this cache entry was created. However, it will
1924 * be clear for "." entries.
1925 * - ticks: Return storage for alternate cache timestamp. On a successful
1926 * (positive or negative) lookup, it will contain the ticks value
1927 * that was current when the cache entry was created, unless cnp
1928 * was ".".
1929 *
1930 * Either both tsp and ticks have to be provided or neither of them.
1931 *
1932 * # Returns
1933 *
1934 * - -1: A positive cache hit. vpp will contain the desired vnode.
1935 * - ENOENT: A negative cache hit, or dvp was recycled out from under us due
1936 * to a forced unmount. vpp will not be modified. If the entry
1937 * is a whiteout, then the ISWHITEOUT flag will be set in
1938 * cnp->cn_flags.
1939 * - 0: A cache miss. vpp will not be modified.
1940 *
1941 * # Locking
1942 *
1943 * On a cache hit, vpp will be returned locked and ref'd. If we're looking up
1944 * .., dvp is unlocked. If we're looking up . an extra ref is taken, but the
1945 * lock is not recursively acquired.
1946 */
1947 static int __noinline
1948 cache_lookup_fallback(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1949 struct timespec *tsp, int *ticksp)
1950 {
1951 struct namecache *ncp;
1952 struct mtx *blp;
1953 uint32_t hash;
1954 enum vgetstate vs;
1955 int error;
1956 bool whiteout;
1957
1958 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
1959 MPASS((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) != 0);
1960
1961 retry:
1962 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1963 blp = HASH2BUCKETLOCK(hash);
1964 mtx_lock(blp);
1965
1966 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1967 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1968 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1969 break;
1970 }
1971
1972 if (__predict_false(ncp == NULL)) {
1973 mtx_unlock(blp);
1974 SDT_PROBE2(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr);
1975 counter_u64_add(nummiss, 1);
1976 return (0);
1977 }
1978
1979 if (ncp->nc_flag & NCF_NEGATIVE)
1980 goto negative_success;
1981
1982 counter_u64_add(numposhits, 1);
1983 *vpp = ncp->nc_vp;
1984 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1985 cache_out_ts(ncp, tsp, ticksp);
1986 MPASS(dvp != *vpp);
1987 vs = vget_prep(*vpp);
1988 mtx_unlock(blp);
1989 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1990 if (error) {
1991 *vpp = NULL;
1992 goto retry;
1993 }
1994 return (-1);
1995 negative_success:
1996 /*
1997 * We don't get here with regular lookup apart from corner cases.
1998 */
1999 if (__predict_true(cnp->cn_nameiop == CREATE)) {
2000 if (cnp->cn_flags & ISLASTCN) {
2001 counter_u64_add(numnegzaps, 1);
2002 error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
2003 if (__predict_false(error != 0)) {
2004 zap_bucket_fail2++;
2005 goto retry;
2006 }
2007 cache_free(ncp);
2008 return (0);
2009 }
2010 }
2011
2012 whiteout = (ncp->nc_flag & NCF_WHITE);
2013 cache_out_ts(ncp, tsp, ticksp);
2014 if (cache_neg_hit_prep(ncp))
2015 cache_neg_promote(ncp);
2016 else
2017 cache_neg_hit_finish(ncp);
2018 mtx_unlock(blp);
2019 if (whiteout)
2020 cnp->cn_flags |= ISWHITEOUT;
2021 return (ENOENT);
2022 }
2023
2024 int
2025 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
2026 struct timespec *tsp, int *ticksp)
2027 {
2028 struct namecache *ncp;
2029 uint32_t hash;
2030 enum vgetstate vs;
2031 int error;
2032 bool whiteout, neg_promote;
2033 u_short nc_flag;
2034
2035 MPASS((tsp == NULL && ticksp == NULL) || (tsp != NULL && ticksp != NULL));
2036
2037 #ifdef DEBUG_CACHE
2038 if (__predict_false(!doingcache)) {
2039 cnp->cn_flags &= ~MAKEENTRY;
2040 return (0);
2041 }
2042 #endif
2043
2044 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
2045 if (cnp->cn_namelen == 1)
2046 return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
2047 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.')
2048 return (cache_lookup_dotdot(dvp, vpp, cnp, tsp, ticksp));
2049 }
2050
2051 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
2052
2053 if ((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) == 0) {
2054 cache_remove_cnp(dvp, cnp);
2055 return (0);
2056 }
2057
2058 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
2059 vfs_smr_enter();
2060
2061 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
2062 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
2063 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
2064 break;
2065 }
2066
2067 if (__predict_false(ncp == NULL)) {
2068 vfs_smr_exit();
2069 SDT_PROBE2(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr);
2070 counter_u64_add(nummiss, 1);
2071 return (0);
2072 }
2073
2074 nc_flag = atomic_load_char(&ncp->nc_flag);
2075 if (nc_flag & NCF_NEGATIVE)
2076 goto negative_success;
2077
2078 counter_u64_add(numposhits, 1);
2079 *vpp = ncp->nc_vp;
2080 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
2081 cache_out_ts(ncp, tsp, ticksp);
2082 MPASS(dvp != *vpp);
2083 if (!cache_ncp_canuse(ncp)) {
2084 vfs_smr_exit();
2085 *vpp = NULL;
2086 goto out_fallback;
2087 }
2088 vs = vget_prep_smr(*vpp);
2089 vfs_smr_exit();
2090 if (__predict_false(vs == VGET_NONE)) {
2091 *vpp = NULL;
2092 goto out_fallback;
2093 }
2094 error = vget_finish(*vpp, cnp->cn_lkflags, vs);
2095 if (error) {
2096 *vpp = NULL;
2097 goto out_fallback;
2098 }
2099 return (-1);
2100 negative_success:
2101 if (cnp->cn_nameiop == CREATE) {
2102 if (cnp->cn_flags & ISLASTCN) {
2103 vfs_smr_exit();
2104 goto out_fallback;
2105 }
2106 }
2107
2108 cache_out_ts(ncp, tsp, ticksp);
2109 whiteout = (atomic_load_char(&ncp->nc_flag) & NCF_WHITE);
2110 neg_promote = cache_neg_hit_prep(ncp);
2111 if (!cache_ncp_canuse(ncp)) {
2112 cache_neg_hit_abort(ncp);
2113 vfs_smr_exit();
2114 goto out_fallback;
2115 }
2116 if (neg_promote) {
2117 vfs_smr_exit();
2118 if (!cache_neg_promote_cond(dvp, cnp, ncp, hash))
2119 goto out_fallback;
2120 } else {
2121 cache_neg_hit_finish(ncp);
2122 vfs_smr_exit();
2123 }
2124 if (whiteout)
2125 cnp->cn_flags |= ISWHITEOUT;
2126 return (ENOENT);
2127 out_fallback:
2128 return (cache_lookup_fallback(dvp, vpp, cnp, tsp, ticksp));
2129 }
2130
2131 struct celockstate {
2132 struct mtx *vlp[3];
2133 struct mtx *blp[2];
2134 };
2135 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
2136 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
2137
2138 static inline void
2139 cache_celockstate_init(struct celockstate *cel)
2140 {
2141
2142 bzero(cel, sizeof(*cel));
2143 }
2144
2145 static void
2146 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
2147 struct vnode *dvp)
2148 {
2149 struct mtx *vlp1, *vlp2;
2150
2151 MPASS(cel->vlp[0] == NULL);
2152 MPASS(cel->vlp[1] == NULL);
2153 MPASS(cel->vlp[2] == NULL);
2154
2155 MPASS(vp != NULL || dvp != NULL);
2156
2157 vlp1 = VP2VNODELOCK(vp);
2158 vlp2 = VP2VNODELOCK(dvp);
2159 cache_sort_vnodes(&vlp1, &vlp2);
2160
2161 if (vlp1 != NULL) {
2162 mtx_lock(vlp1);
2163 cel->vlp[0] = vlp1;
2164 }
2165 mtx_lock(vlp2);
2166 cel->vlp[1] = vlp2;
2167 }
2168
2169 static void
2170 cache_unlock_vnodes_cel(struct celockstate *cel)
2171 {
2172
2173 MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
2174
2175 if (cel->vlp[0] != NULL)
2176 mtx_unlock(cel->vlp[0]);
2177 if (cel->vlp[1] != NULL)
2178 mtx_unlock(cel->vlp[1]);
2179 if (cel->vlp[2] != NULL)
2180 mtx_unlock(cel->vlp[2]);
2181 }
2182
2183 static bool
2184 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
2185 {
2186 struct mtx *vlp;
2187 bool ret;
2188
2189 cache_assert_vlp_locked(cel->vlp[0]);
2190 cache_assert_vlp_locked(cel->vlp[1]);
2191 MPASS(cel->vlp[2] == NULL);
2192
2193 MPASS(vp != NULL);
2194 vlp = VP2VNODELOCK(vp);
2195
2196 ret = true;
2197 if (vlp >= cel->vlp[1]) {
2198 mtx_lock(vlp);
2199 } else {
2200 if (mtx_trylock(vlp))
2201 goto out;
2202 cache_lock_vnodes_cel_3_failures++;
2203 cache_unlock_vnodes_cel(cel);
2204 if (vlp < cel->vlp[0]) {
2205 mtx_lock(vlp);
2206 mtx_lock(cel->vlp[0]);
2207 mtx_lock(cel->vlp[1]);
2208 } else {
2209 if (cel->vlp[0] != NULL)
2210 mtx_lock(cel->vlp[0]);
2211 mtx_lock(vlp);
2212 mtx_lock(cel->vlp[1]);
2213 }
2214 ret = false;
2215 }
2216 out:
2217 cel->vlp[2] = vlp;
2218 return (ret);
2219 }
2220
2221 static void
2222 cache_lock_buckets_cel(struct celockstate *cel, struct mtx *blp1,
2223 struct mtx *blp2)
2224 {
2225
2226 MPASS(cel->blp[0] == NULL);
2227 MPASS(cel->blp[1] == NULL);
2228
2229 cache_sort_vnodes(&blp1, &blp2);
2230
2231 if (blp1 != NULL) {
2232 mtx_lock(blp1);
2233 cel->blp[0] = blp1;
2234 }
2235 mtx_lock(blp2);
2236 cel->blp[1] = blp2;
2237 }
2238
2239 static void
2240 cache_unlock_buckets_cel(struct celockstate *cel)
2241 {
2242
2243 if (cel->blp[0] != NULL)
2244 mtx_unlock(cel->blp[0]);
2245 mtx_unlock(cel->blp[1]);
2246 }
2247
2248 /*
2249 * Lock part of the cache affected by the insertion.
2250 *
2251 * This means vnodelocks for dvp, vp and the relevant bucketlock.
2252 * However, insertion can result in removal of an old entry. In this
2253 * case we have an additional vnode and bucketlock pair to lock.
2254 *
2255 * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
2256 * preserving the locking order (smaller address first).
2257 */
2258 static void
2259 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2260 uint32_t hash)
2261 {
2262 struct namecache *ncp;
2263 struct mtx *blps[2];
2264 u_char nc_flag;
2265
2266 blps[0] = HASH2BUCKETLOCK(hash);
2267 for (;;) {
2268 blps[1] = NULL;
2269 cache_lock_vnodes_cel(cel, dvp, vp);
2270 if (vp == NULL || vp->v_type != VDIR)
2271 break;
2272 ncp = atomic_load_consume_ptr(&vp->v_cache_dd);
2273 if (ncp == NULL)
2274 break;
2275 nc_flag = atomic_load_char(&ncp->nc_flag);
2276 if ((nc_flag & NCF_ISDOTDOT) == 0)
2277 break;
2278 MPASS(ncp->nc_dvp == vp);
2279 blps[1] = NCP2BUCKETLOCK(ncp);
2280 if ((nc_flag & NCF_NEGATIVE) != 0)
2281 break;
2282 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2283 break;
2284 /*
2285 * All vnodes got re-locked. Re-validate the state and if
2286 * nothing changed we are done. Otherwise restart.
2287 */
2288 if (ncp == vp->v_cache_dd &&
2289 (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2290 blps[1] == NCP2BUCKETLOCK(ncp) &&
2291 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2292 break;
2293 cache_unlock_vnodes_cel(cel);
2294 cel->vlp[0] = NULL;
2295 cel->vlp[1] = NULL;
2296 cel->vlp[2] = NULL;
2297 }
2298 cache_lock_buckets_cel(cel, blps[0], blps[1]);
2299 }
2300
2301 static void
2302 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2303 uint32_t hash)
2304 {
2305 struct namecache *ncp;
2306 struct mtx *blps[2];
2307 u_char nc_flag;
2308
2309 blps[0] = HASH2BUCKETLOCK(hash);
2310 for (;;) {
2311 blps[1] = NULL;
2312 cache_lock_vnodes_cel(cel, dvp, vp);
2313 ncp = atomic_load_consume_ptr(&dvp->v_cache_dd);
2314 if (ncp == NULL)
2315 break;
2316 nc_flag = atomic_load_char(&ncp->nc_flag);
2317 if ((nc_flag & NCF_ISDOTDOT) == 0)
2318 break;
2319 MPASS(ncp->nc_dvp == dvp);
2320 blps[1] = NCP2BUCKETLOCK(ncp);
2321 if ((nc_flag & NCF_NEGATIVE) != 0)
2322 break;
2323 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2324 break;
2325 if (ncp == dvp->v_cache_dd &&
2326 (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2327 blps[1] == NCP2BUCKETLOCK(ncp) &&
2328 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2329 break;
2330 cache_unlock_vnodes_cel(cel);
2331 cel->vlp[0] = NULL;
2332 cel->vlp[1] = NULL;
2333 cel->vlp[2] = NULL;
2334 }
2335 cache_lock_buckets_cel(cel, blps[0], blps[1]);
2336 }
2337
2338 static void
2339 cache_enter_unlock(struct celockstate *cel)
2340 {
2341
2342 cache_unlock_buckets_cel(cel);
2343 cache_unlock_vnodes_cel(cel);
2344 }
2345
2346 static void __noinline
2347 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
2348 struct componentname *cnp)
2349 {
2350 struct celockstate cel;
2351 struct namecache *ncp;
2352 uint32_t hash;
2353 int len;
2354
2355 if (atomic_load_ptr(&dvp->v_cache_dd) == NULL)
2356 return;
2357 len = cnp->cn_namelen;
2358 cache_celockstate_init(&cel);
2359 hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2360 cache_enter_lock_dd(&cel, dvp, vp, hash);
2361 ncp = dvp->v_cache_dd;
2362 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
2363 KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
2364 cache_zap_locked(ncp);
2365 } else {
2366 ncp = NULL;
2367 }
2368 atomic_store_ptr(&dvp->v_cache_dd, NULL);
2369 cache_enter_unlock(&cel);
2370 if (ncp != NULL)
2371 cache_free(ncp);
2372 }
2373
2374 /*
2375 * Add an entry to the cache.
2376 */
2377 void
2378 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
2379 struct timespec *tsp, struct timespec *dtsp)
2380 {
2381 struct celockstate cel;
2382 struct namecache *ncp, *n2, *ndd;
2383 struct namecache_ts *ncp_ts;
2384 struct nchashhead *ncpp;
2385 uint32_t hash;
2386 int flag;
2387 int len;
2388
2389 KASSERT(cnp->cn_namelen <= NAME_MAX,
2390 ("%s: passed len %ld exceeds NAME_MAX (%d)", __func__, cnp->cn_namelen,
2391 NAME_MAX));
2392 #ifdef notyet
2393 /*
2394 * Not everything doing this is weeded out yet.
2395 */
2396 VNPASS(dvp != vp, dvp);
2397 #endif
2398 VNPASS(!VN_IS_DOOMED(dvp), dvp);
2399 VNPASS(dvp->v_type != VNON, dvp);
2400 if (vp != NULL) {
2401 VNPASS(!VN_IS_DOOMED(vp), vp);
2402 VNPASS(vp->v_type != VNON, vp);
2403 }
2404
2405 #ifdef DEBUG_CACHE
2406 if (__predict_false(!doingcache))
2407 return;
2408 #endif
2409
2410 flag = 0;
2411 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
2412 if (cnp->cn_namelen == 1)
2413 return;
2414 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
2415 cache_enter_dotdot_prep(dvp, vp, cnp);
2416 flag = NCF_ISDOTDOT;
2417 }
2418 }
2419
2420 ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
2421 if (ncp == NULL)
2422 return;
2423
2424 cache_celockstate_init(&cel);
2425 ndd = NULL;
2426 ncp_ts = NULL;
2427
2428 /*
2429 * Calculate the hash key and setup as much of the new
2430 * namecache entry as possible before acquiring the lock.
2431 */
2432 ncp->nc_flag = flag | NCF_WIP;
2433 ncp->nc_vp = vp;
2434 if (vp == NULL)
2435 cache_neg_init(ncp);
2436 ncp->nc_dvp = dvp;
2437 if (tsp != NULL) {
2438 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
2439 ncp_ts->nc_time = *tsp;
2440 ncp_ts->nc_ticks = ticks;
2441 ncp_ts->nc_nc.nc_flag |= NCF_TS;
2442 if (dtsp != NULL) {
2443 ncp_ts->nc_dotdottime = *dtsp;
2444 ncp_ts->nc_nc.nc_flag |= NCF_DTS;
2445 }
2446 }
2447 len = ncp->nc_nlen = cnp->cn_namelen;
2448 hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2449 memcpy(ncp->nc_name, cnp->cn_nameptr, len);
2450 ncp->nc_name[len] = '\0';
2451 cache_enter_lock(&cel, dvp, vp, hash);
2452
2453 /*
2454 * See if this vnode or negative entry is already in the cache
2455 * with this name. This can happen with concurrent lookups of
2456 * the same path name.
2457 */
2458 ncpp = NCHHASH(hash);
2459 CK_SLIST_FOREACH(n2, ncpp, nc_hash) {
2460 if (n2->nc_dvp == dvp &&
2461 n2->nc_nlen == cnp->cn_namelen &&
2462 !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
2463 MPASS(cache_ncp_canuse(n2));
2464 if ((n2->nc_flag & NCF_NEGATIVE) != 0)
2465 KASSERT(vp == NULL,
2466 ("%s: found entry pointing to a different vnode (%p != %p) ; name [%s]",
2467 __func__, NULL, vp, cnp->cn_nameptr));
2468 else
2469 KASSERT(n2->nc_vp == vp,
2470 ("%s: found entry pointing to a different vnode (%p != %p) ; name [%s]",
2471 __func__, n2->nc_vp, vp, cnp->cn_nameptr));
2472 /*
2473 * Entries are supposed to be immutable unless in the
2474 * process of getting destroyed. Accommodating for
2475 * changing timestamps is possible but not worth it.
2476 * This should be harmless in terms of correctness, in
2477 * the worst case resulting in an earlier expiration.
2478 * Alternatively, the found entry can be replaced
2479 * altogether.
2480 */
2481 MPASS((n2->nc_flag & (NCF_TS | NCF_DTS)) == (ncp->nc_flag & (NCF_TS | NCF_DTS)));
2482 #if 0
2483 if (tsp != NULL) {
2484 KASSERT((n2->nc_flag & NCF_TS) != 0,
2485 ("no NCF_TS"));
2486 n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
2487 n2_ts->nc_time = ncp_ts->nc_time;
2488 n2_ts->nc_ticks = ncp_ts->nc_ticks;
2489 if (dtsp != NULL) {
2490 n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
2491 n2_ts->nc_nc.nc_flag |= NCF_DTS;
2492 }
2493 }
2494 #endif
2495 SDT_PROBE3(vfs, namecache, enter, duplicate, dvp, ncp->nc_name,
2496 vp);
2497 goto out_unlock_free;
2498 }
2499 }
2500
2501 if (flag == NCF_ISDOTDOT) {
2502 /*
2503 * See if we are trying to add .. entry, but some other lookup
2504 * has populated v_cache_dd pointer already.
2505 */
2506 if (dvp->v_cache_dd != NULL)
2507 goto out_unlock_free;
2508 KASSERT(vp == NULL || vp->v_type == VDIR,
2509 ("wrong vnode type %p", vp));
2510 atomic_thread_fence_rel();
2511 atomic_store_ptr(&dvp->v_cache_dd, ncp);
2512 }
2513
2514 if (vp != NULL) {
2515 if (flag != NCF_ISDOTDOT) {
2516 /*
2517 * For this case, the cache entry maps both the
2518 * directory name in it and the name ".." for the
2519 * directory's parent.
2520 */
2521 if ((ndd = vp->v_cache_dd) != NULL) {
2522 if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
2523 cache_zap_locked(ndd);
2524 else
2525 ndd = NULL;
2526 }
2527 atomic_thread_fence_rel();
2528 atomic_store_ptr(&vp->v_cache_dd, ncp);
2529 } else if (vp->v_type != VDIR) {
2530 if (vp->v_cache_dd != NULL) {
2531 atomic_store_ptr(&vp->v_cache_dd, NULL);
2532 }
2533 }
2534 }
2535
2536 if (flag != NCF_ISDOTDOT) {
2537 if (LIST_EMPTY(&dvp->v_cache_src)) {
2538 cache_hold_vnode(dvp);
2539 }
2540 LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
2541 }
2542
2543 /*
2544 * If the entry is "negative", we place it into the
2545 * "negative" cache queue, otherwise, we place it into the
2546 * destination vnode's cache entries queue.
2547 */
2548 if (vp != NULL) {
2549 TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
2550 SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
2551 vp);
2552 } else {
2553 if (cnp->cn_flags & ISWHITEOUT)
2554 atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_WHITE);
2555 cache_neg_insert(ncp);
2556 SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
2557 ncp->nc_name);
2558 }
2559
2560 /*
2561 * Insert the new namecache entry into the appropriate chain
2562 * within the cache entries table.
2563 */
2564 CK_SLIST_INSERT_HEAD(ncpp, ncp, nc_hash);
2565
2566 atomic_thread_fence_rel();
2567 /*
2568 * Mark the entry as fully constructed.
2569 * It is immutable past this point until its removal.
2570 */
2571 atomic_store_char(&ncp->nc_flag, ncp->nc_flag & ~NCF_WIP);
2572
2573 cache_enter_unlock(&cel);
2574 if (ndd != NULL)
2575 cache_free(ndd);
2576 return;
2577 out_unlock_free:
2578 cache_enter_unlock(&cel);
2579 cache_free(ncp);
2580 return;
2581 }
2582
2583 /*
2584 * A variant of the above accepting flags.
2585 *
2586 * - VFS_CACHE_DROPOLD -- if a conflicting entry is found, drop it.
2587 *
2588 * TODO: this routine is a hack. It blindly removes the old entry, even if it
2589 * happens to match and it is doing it in an inefficient manner. It was added
2590 * to accomodate NFS which runs into a case where the target for a given name
2591 * may change from under it. Note this does nothing to solve the following
2592 * race: 2 callers of cache_enter_time_flags pass a different target vnode for
2593 * the same [dvp, cnp]. It may be argued that code doing this is broken.
2594 */
2595 void
2596 cache_enter_time_flags(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
2597 struct timespec *tsp, struct timespec *dtsp, int flags)
2598 {
2599
2600 MPASS((flags & ~(VFS_CACHE_DROPOLD)) == 0);
2601
2602 if (flags & VFS_CACHE_DROPOLD)
2603 cache_remove_cnp(dvp, cnp);
2604 cache_enter_time(dvp, vp, cnp, tsp, dtsp);
2605 }
2606
2607 static u_int
2608 cache_roundup_2(u_int val)
2609 {
2610 u_int res;
2611
2612 for (res = 1; res <= val; res <<= 1)
2613 continue;
2614
2615 return (res);
2616 }
2617
2618 static struct nchashhead *
2619 nchinittbl(u_long elements, u_long *hashmask)
2620 {
2621 struct nchashhead *hashtbl;
2622 u_long hashsize, i;
2623
2624 hashsize = cache_roundup_2(elements) / 2;
2625
2626 hashtbl = malloc((u_long)hashsize * sizeof(*hashtbl), M_VFSCACHE, M_WAITOK);
2627 for (i = 0; i < hashsize; i++)
2628 CK_SLIST_INIT(&hashtbl[i]);
2629 *hashmask = hashsize - 1;
2630 return (hashtbl);
2631 }
2632
2633 static void
2634 ncfreetbl(struct nchashhead *hashtbl)
2635 {
2636
2637 free(hashtbl, M_VFSCACHE);
2638 }
2639
2640 /*
2641 * Name cache initialization, from vfs_init() when we are booting
2642 */
2643 static void
2644 nchinit(void *dummy __unused)
2645 {
2646 u_int i;
2647
2648 cache_zone_small = uma_zcreate("S VFS Cache", CACHE_ZONE_SMALL_SIZE,
2649 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2650 cache_zone_small_ts = uma_zcreate("STS VFS Cache", CACHE_ZONE_SMALL_TS_SIZE,
2651 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2652 cache_zone_large = uma_zcreate("L VFS Cache", CACHE_ZONE_LARGE_SIZE,
2653 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2654 cache_zone_large_ts = uma_zcreate("LTS VFS Cache", CACHE_ZONE_LARGE_TS_SIZE,
2655 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2656
2657 VFS_SMR_ZONE_SET(cache_zone_small);
2658 VFS_SMR_ZONE_SET(cache_zone_small_ts);
2659 VFS_SMR_ZONE_SET(cache_zone_large);
2660 VFS_SMR_ZONE_SET(cache_zone_large_ts);
2661
2662 ncsize = desiredvnodes * ncsizefactor;
2663 cache_recalc_neg_min(ncnegminpct);
2664 nchashtbl = nchinittbl(desiredvnodes * 2, &nchash);
2665 ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
2666 if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
2667 ncbuckethash = 7;
2668 if (ncbuckethash > nchash)
2669 ncbuckethash = nchash;
2670 bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
2671 M_WAITOK | M_ZERO);
2672 for (i = 0; i < numbucketlocks; i++)
2673 mtx_init(&bucketlocks[i], "ncbuc", NULL, MTX_DUPOK | MTX_RECURSE);
2674 ncvnodehash = ncbuckethash;
2675 vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
2676 M_WAITOK | M_ZERO);
2677 for (i = 0; i < numvnodelocks; i++)
2678 mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
2679
2680 for (i = 0; i < numneglists; i++) {
2681 mtx_init(&neglists[i].nl_evict_lock, "ncnege", NULL, MTX_DEF);
2682 mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
2683 TAILQ_INIT(&neglists[i].nl_list);
2684 TAILQ_INIT(&neglists[i].nl_hotlist);
2685 }
2686 }
2687 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
2688
2689 void
2690 cache_vnode_init(struct vnode *vp)
2691 {
2692
2693 LIST_INIT(&vp->v_cache_src);
2694 TAILQ_INIT(&vp->v_cache_dst);
2695 vp->v_cache_dd = NULL;
2696 cache_prehash(vp);
2697 }
2698
2699 /*
2700 * Induce transient cache misses for lockless operation in cache_lookup() by
2701 * using a temporary hash table.
2702 *
2703 * This will force a fs lookup.
2704 *
2705 * Synchronisation is done in 2 steps, calling vfs_smr_synchronize each time
2706 * to observe all CPUs not performing the lookup.
2707 */
2708 static void
2709 cache_changesize_set_temp(struct nchashhead *temptbl, u_long temphash)
2710 {
2711
2712 MPASS(temphash < nchash);
2713 /*
2714 * Change the size. The new size is smaller and can safely be used
2715 * against the existing table. All lookups which now hash wrong will
2716 * result in a cache miss, which all callers are supposed to know how
2717 * to handle.
2718 */
2719 atomic_store_long(&nchash, temphash);
2720 atomic_thread_fence_rel();
2721 vfs_smr_synchronize();
2722 /*
2723 * At this point everyone sees the updated hash value, but they still
2724 * see the old table.
2725 */
2726 atomic_store_ptr(&nchashtbl, temptbl);
2727 atomic_thread_fence_rel();
2728 vfs_smr_synchronize();
2729 /*
2730 * At this point everyone sees the updated table pointer and size pair.
2731 */
2732 }
2733
2734 /*
2735 * Set the new hash table.
2736 *
2737 * Similarly to cache_changesize_set_temp(), this has to synchronize against
2738 * lockless operation in cache_lookup().
2739 */
2740 static void
2741 cache_changesize_set_new(struct nchashhead *new_tbl, u_long new_hash)
2742 {
2743
2744 MPASS(nchash < new_hash);
2745 /*
2746 * Change the pointer first. This wont result in out of bounds access
2747 * since the temporary table is guaranteed to be smaller.
2748 */
2749 atomic_store_ptr(&nchashtbl, new_tbl);
2750 atomic_thread_fence_rel();
2751 vfs_smr_synchronize();
2752 /*
2753 * At this point everyone sees the updated pointer value, but they
2754 * still see the old size.
2755 */
2756 atomic_store_long(&nchash, new_hash);
2757 atomic_thread_fence_rel();
2758 vfs_smr_synchronize();
2759 /*
2760 * At this point everyone sees the updated table pointer and size pair.
2761 */
2762 }
2763
2764 void
2765 cache_changesize(u_long newmaxvnodes)
2766 {
2767 struct nchashhead *new_nchashtbl, *old_nchashtbl, *temptbl;
2768 u_long new_nchash, old_nchash, temphash;
2769 struct namecache *ncp;
2770 uint32_t hash;
2771 u_long newncsize;
2772 int i;
2773
2774 newncsize = newmaxvnodes * ncsizefactor;
2775 newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
2776 if (newmaxvnodes < numbucketlocks)
2777 newmaxvnodes = numbucketlocks;
2778
2779 new_nchashtbl = nchinittbl(newmaxvnodes, &new_nchash);
2780 /* If same hash table size, nothing to do */
2781 if (nchash == new_nchash) {
2782 ncfreetbl(new_nchashtbl);
2783 return;
2784 }
2785
2786 temptbl = nchinittbl(1, &temphash);
2787
2788 /*
2789 * Move everything from the old hash table to the new table.
2790 * None of the namecache entries in the table can be removed
2791 * because to do so, they have to be removed from the hash table.
2792 */
2793 cache_lock_all_vnodes();
2794 cache_lock_all_buckets();
2795 old_nchashtbl = nchashtbl;
2796 old_nchash = nchash;
2797 cache_changesize_set_temp(temptbl, temphash);
2798 for (i = 0; i <= old_nchash; i++) {
2799 while ((ncp = CK_SLIST_FIRST(&old_nchashtbl[i])) != NULL) {
2800 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
2801 ncp->nc_dvp);
2802 CK_SLIST_REMOVE(&old_nchashtbl[i], ncp, namecache, nc_hash);
2803 CK_SLIST_INSERT_HEAD(&new_nchashtbl[hash & new_nchash], ncp, nc_hash);
2804 }
2805 }
2806 ncsize = newncsize;
2807 cache_recalc_neg_min(ncnegminpct);
2808 cache_changesize_set_new(new_nchashtbl, new_nchash);
2809 cache_unlock_all_buckets();
2810 cache_unlock_all_vnodes();
2811 ncfreetbl(old_nchashtbl);
2812 ncfreetbl(temptbl);
2813 }
2814
2815 /*
2816 * Remove all entries from and to a particular vnode.
2817 */
2818 static void
2819 cache_purge_impl(struct vnode *vp)
2820 {
2821 struct cache_freebatch batch;
2822 struct namecache *ncp;
2823 struct mtx *vlp, *vlp2;
2824
2825 TAILQ_INIT(&batch);
2826 vlp = VP2VNODELOCK(vp);
2827 vlp2 = NULL;
2828 mtx_lock(vlp);
2829 retry:
2830 while (!LIST_EMPTY(&vp->v_cache_src)) {
2831 ncp = LIST_FIRST(&vp->v_cache_src);
2832 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2833 goto retry;
2834 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2835 }
2836 while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2837 ncp = TAILQ_FIRST(&vp->v_cache_dst);
2838 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2839 goto retry;
2840 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2841 }
2842 ncp = vp->v_cache_dd;
2843 if (ncp != NULL) {
2844 KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2845 ("lost dotdot link"));
2846 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2847 goto retry;
2848 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2849 }
2850 KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2851 mtx_unlock(vlp);
2852 if (vlp2 != NULL)
2853 mtx_unlock(vlp2);
2854 cache_free_batch(&batch);
2855 }
2856
2857 /*
2858 * Opportunistic check to see if there is anything to do.
2859 */
2860 static bool
2861 cache_has_entries(struct vnode *vp)
2862 {
2863
2864 if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2865 atomic_load_ptr(&vp->v_cache_dd) == NULL)
2866 return (false);
2867 return (true);
2868 }
2869
2870 void
2871 cache_purge(struct vnode *vp)
2872 {
2873
2874 SDT_PROBE1(vfs, namecache, purge, done, vp);
2875 if (!cache_has_entries(vp))
2876 return;
2877 cache_purge_impl(vp);
2878 }
2879
2880 /*
2881 * Only to be used by vgone.
2882 */
2883 void
2884 cache_purge_vgone(struct vnode *vp)
2885 {
2886 struct mtx *vlp;
2887
2888 VNPASS(VN_IS_DOOMED(vp), vp);
2889 if (cache_has_entries(vp)) {
2890 cache_purge_impl(vp);
2891 return;
2892 }
2893
2894 /*
2895 * Serialize against a potential thread doing cache_purge.
2896 */
2897 vlp = VP2VNODELOCK(vp);
2898 mtx_wait_unlocked(vlp);
2899 if (cache_has_entries(vp)) {
2900 cache_purge_impl(vp);
2901 return;
2902 }
2903 return;
2904 }
2905
2906 /*
2907 * Remove all negative entries for a particular directory vnode.
2908 */
2909 void
2910 cache_purge_negative(struct vnode *vp)
2911 {
2912 struct cache_freebatch batch;
2913 struct namecache *ncp, *nnp;
2914 struct mtx *vlp;
2915
2916 SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2917 if (LIST_EMPTY(&vp->v_cache_src))
2918 return;
2919 TAILQ_INIT(&batch);
2920 vlp = VP2VNODELOCK(vp);
2921 mtx_lock(vlp);
2922 LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2923 if (!(ncp->nc_flag & NCF_NEGATIVE))
2924 continue;
2925 cache_zap_negative_locked_vnode_kl(ncp, vp);
2926 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2927 }
2928 mtx_unlock(vlp);
2929 cache_free_batch(&batch);
2930 }
2931
2932 /*
2933 * Entry points for modifying VOP operations.
2934 */
2935 void
2936 cache_vop_rename(struct vnode *fdvp, struct vnode *fvp, struct vnode *tdvp,
2937 struct vnode *tvp, struct componentname *fcnp, struct componentname *tcnp)
2938 {
2939
2940 ASSERT_VOP_IN_SEQC(fdvp);
2941 ASSERT_VOP_IN_SEQC(fvp);
2942 ASSERT_VOP_IN_SEQC(tdvp);
2943 if (tvp != NULL)
2944 ASSERT_VOP_IN_SEQC(tvp);
2945
2946 cache_purge(fvp);
2947 if (tvp != NULL) {
2948 cache_purge(tvp);
2949 KASSERT(!cache_remove_cnp(tdvp, tcnp),
2950 ("%s: lingering negative entry", __func__));
2951 } else {
2952 cache_remove_cnp(tdvp, tcnp);
2953 }
2954
2955 /*
2956 * TODO
2957 *
2958 * Historically renaming was always purging all revelang entries,
2959 * but that's quite wasteful. In particular turns out that in many cases
2960 * the target file is immediately accessed after rename, inducing a cache
2961 * miss.
2962 *
2963 * Recode this to reduce relocking and reuse the existing entry (if any)
2964 * instead of just removing it above and allocating a new one here.
2965 */
2966 if (cache_rename_add) {
2967 cache_enter(tdvp, fvp, tcnp);
2968 }
2969 }
2970
2971 void
2972 cache_vop_rmdir(struct vnode *dvp, struct vnode *vp)
2973 {
2974
2975 ASSERT_VOP_IN_SEQC(dvp);
2976 ASSERT_VOP_IN_SEQC(vp);
2977 cache_purge(vp);
2978 }
2979
2980 #ifdef INVARIANTS
2981 /*
2982 * Validate that if an entry exists it matches.
2983 */
2984 void
2985 cache_validate(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
2986 {
2987 struct namecache *ncp;
2988 struct mtx *blp;
2989 uint32_t hash;
2990
2991 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
2992 if (CK_SLIST_EMPTY(NCHHASH(hash)))
2993 return;
2994 blp = HASH2BUCKETLOCK(hash);
2995 mtx_lock(blp);
2996 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
2997 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
2998 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) {
2999 if (ncp->nc_vp != vp)
3000 panic("%s: mismatch (%p != %p); ncp %p [%s] dvp %p\n",
3001 __func__, vp, ncp->nc_vp, ncp, ncp->nc_name, ncp->nc_dvp);
3002 }
3003 }
3004 mtx_unlock(blp);
3005 }
3006 #endif
3007
3008 /*
3009 * Flush all entries referencing a particular filesystem.
3010 */
3011 void
3012 cache_purgevfs(struct mount *mp)
3013 {
3014 struct vnode *vp, *mvp;
3015 size_t visited, purged;
3016
3017 visited = purged = 0;
3018 /*
3019 * Somewhat wasteful iteration over all vnodes. Would be better to
3020 * support filtering and avoid the interlock to begin with.
3021 */
3022 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3023 visited++;
3024 if (!cache_has_entries(vp)) {
3025 VI_UNLOCK(vp);
3026 continue;
3027 }
3028 vholdl(vp);
3029 VI_UNLOCK(vp);
3030 cache_purge(vp);
3031 purged++;
3032 vdrop(vp);
3033 }
3034
3035 SDT_PROBE3(vfs, namecache, purgevfs, done, mp, visited, purged);
3036 }
3037
3038 /*
3039 * Perform canonical checks and cache lookup and pass on to filesystem
3040 * through the vop_cachedlookup only if needed.
3041 */
3042
3043 int
3044 vfs_cache_lookup(struct vop_lookup_args *ap)
3045 {
3046 struct vnode *dvp;
3047 int error;
3048 struct vnode **vpp = ap->a_vpp;
3049 struct componentname *cnp = ap->a_cnp;
3050 int flags = cnp->cn_flags;
3051
3052 *vpp = NULL;
3053 dvp = ap->a_dvp;
3054
3055 if (dvp->v_type != VDIR)
3056 return (ENOTDIR);
3057
3058 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
3059 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
3060 return (EROFS);
3061
3062 error = vn_dir_check_exec(dvp, cnp);
3063 if (error != 0)
3064 return (error);
3065
3066 error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
3067 if (error == 0)
3068 return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
3069 if (error == -1)
3070 return (0);
3071 return (error);
3072 }
3073
3074 /* Implementation of the getcwd syscall. */
3075 int
3076 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
3077 {
3078 char *buf, *retbuf;
3079 size_t buflen;
3080 int error;
3081
3082 buflen = uap->buflen;
3083 if (__predict_false(buflen < 2))
3084 return (EINVAL);
3085 if (buflen > MAXPATHLEN)
3086 buflen = MAXPATHLEN;
3087
3088 buf = uma_zalloc(namei_zone, M_WAITOK);
3089 error = vn_getcwd(buf, &retbuf, &buflen);
3090 if (error == 0)
3091 error = copyout(retbuf, uap->buf, buflen);
3092 uma_zfree(namei_zone, buf);
3093 return (error);
3094 }
3095
3096 int
3097 vn_getcwd(char *buf, char **retbuf, size_t *buflen)
3098 {
3099 struct pwd *pwd;
3100 int error;
3101
3102 vfs_smr_enter();
3103 pwd = pwd_get_smr();
3104 error = vn_fullpath_any_smr(pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf,
3105 buflen, 0);
3106 VFS_SMR_ASSERT_NOT_ENTERED();
3107 if (error < 0) {
3108 pwd = pwd_hold(curthread);
3109 error = vn_fullpath_any(pwd->pwd_cdir, pwd->pwd_rdir, buf,
3110 retbuf, buflen);
3111 pwd_drop(pwd);
3112 }
3113
3114 #ifdef KTRACE
3115 if (KTRPOINT(curthread, KTR_NAMEI) && error == 0)
3116 ktrnamei(*retbuf);
3117 #endif
3118 return (error);
3119 }
3120
3121 static int
3122 kern___realpathat(struct thread *td, int fd, const char *path, char *buf,
3123 size_t size, int flags, enum uio_seg pathseg)
3124 {
3125 struct nameidata nd;
3126 char *retbuf, *freebuf;
3127 int error;
3128
3129 if (flags != 0)
3130 return (EINVAL);
3131 NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | SAVENAME | WANTPARENT | AUDITVNODE1,
3132 pathseg, path, fd, &cap_fstat_rights, td);
3133 if ((error = namei(&nd)) != 0)
3134 return (error);
3135 error = vn_fullpath_hardlink(nd.ni_vp, nd.ni_dvp, nd.ni_cnd.cn_nameptr,
3136 nd.ni_cnd.cn_namelen, &retbuf, &freebuf, &size);
3137 if (error == 0) {
3138 error = copyout(retbuf, buf, size);
3139 free(freebuf, M_TEMP);
3140 }
3141 NDFREE(&nd, 0);
3142 return (error);
3143 }
3144
3145 int
3146 sys___realpathat(struct thread *td, struct __realpathat_args *uap)
3147 {
3148
3149 return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size,
3150 uap->flags, UIO_USERSPACE));
3151 }
3152
3153 /*
3154 * Retrieve the full filesystem path that correspond to a vnode from the name
3155 * cache (if available)
3156 */
3157 int
3158 vn_fullpath(struct vnode *vp, char **retbuf, char **freebuf)
3159 {
3160 struct pwd *pwd;
3161 char *buf;
3162 size_t buflen;
3163 int error;
3164
3165 if (__predict_false(vp == NULL))
3166 return (EINVAL);
3167
3168 buflen = MAXPATHLEN;
3169 buf = malloc(buflen, M_TEMP, M_WAITOK);
3170 vfs_smr_enter();
3171 pwd = pwd_get_smr();
3172 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, &buflen, 0);
3173 VFS_SMR_ASSERT_NOT_ENTERED();
3174 if (error < 0) {
3175 pwd = pwd_hold(curthread);
3176 error = vn_fullpath_any(vp, pwd->pwd_rdir, buf, retbuf, &buflen);
3177 pwd_drop(pwd);
3178 }
3179 if (error == 0)
3180 *freebuf = buf;
3181 else
3182 free(buf, M_TEMP);
3183 return (error);
3184 }
3185
3186 /*
3187 * This function is similar to vn_fullpath, but it attempts to lookup the
3188 * pathname relative to the global root mount point. This is required for the
3189 * auditing sub-system, as audited pathnames must be absolute, relative to the
3190 * global root mount point.
3191 */
3192 int
3193 vn_fullpath_global(struct vnode *vp, char **retbuf, char **freebuf)
3194 {
3195 char *buf;
3196 size_t buflen;
3197 int error;
3198
3199 if (__predict_false(vp == NULL))
3200 return (EINVAL);
3201 buflen = MAXPATHLEN;
3202 buf = malloc(buflen, M_TEMP, M_WAITOK);
3203 vfs_smr_enter();
3204 error = vn_fullpath_any_smr(vp, rootvnode, buf, retbuf, &buflen, 0);
3205 VFS_SMR_ASSERT_NOT_ENTERED();
3206 if (error < 0) {
3207 error = vn_fullpath_any(vp, rootvnode, buf, retbuf, &buflen);
3208 }
3209 if (error == 0)
3210 *freebuf = buf;
3211 else
3212 free(buf, M_TEMP);
3213 return (error);
3214 }
3215
3216 static struct namecache *
3217 vn_dd_from_dst(struct vnode *vp)
3218 {
3219 struct namecache *ncp;
3220
3221 cache_assert_vnode_locked(vp);
3222 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) {
3223 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
3224 return (ncp);
3225 }
3226 return (NULL);
3227 }
3228
3229 int
3230 vn_vptocnp(struct vnode **vp, char *buf, size_t *buflen)
3231 {
3232 struct vnode *dvp;
3233 struct namecache *ncp;
3234 struct mtx *vlp;
3235 int error;
3236
3237 vlp = VP2VNODELOCK(*vp);
3238 mtx_lock(vlp);
3239 ncp = (*vp)->v_cache_dd;
3240 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT) == 0) {
3241 KASSERT(ncp == vn_dd_from_dst(*vp),
3242 ("%s: mismatch for dd entry (%p != %p)", __func__,
3243 ncp, vn_dd_from_dst(*vp)));
3244 } else {
3245 ncp = vn_dd_from_dst(*vp);
3246 }
3247 if (ncp != NULL) {
3248 if (*buflen < ncp->nc_nlen) {
3249 mtx_unlock(vlp);
3250 vrele(*vp);
3251 counter_u64_add(numfullpathfail4, 1);
3252 error = ENOMEM;
3253 SDT_PROBE3(vfs, namecache, fullpath, return, error,
3254 vp, NULL);
3255 return (error);
3256 }
3257 *buflen -= ncp->nc_nlen;
3258 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
3259 SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
3260 ncp->nc_name, vp);
3261 dvp = *vp;
3262 *vp = ncp->nc_dvp;
3263 vref(*vp);
3264 mtx_unlock(vlp);
3265 vrele(dvp);
3266 return (0);
3267 }
3268 SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
3269
3270 mtx_unlock(vlp);
3271 vn_lock(*vp, LK_SHARED | LK_RETRY);
3272 error = VOP_VPTOCNP(*vp, &dvp, buf, buflen);
3273 vput(*vp);
3274 if (error) {
3275 counter_u64_add(numfullpathfail2, 1);
3276 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
3277 return (error);
3278 }
3279
3280 *vp = dvp;
3281 if (VN_IS_DOOMED(dvp)) {
3282 /* forced unmount */
3283 vrele(dvp);
3284 error = ENOENT;
3285 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
3286 return (error);
3287 }
3288 /*
3289 * *vp has its use count incremented still.
3290 */
3291
3292 return (0);
3293 }
3294
3295 /*
3296 * Resolve a directory to a pathname.
3297 *
3298 * The name of the directory can always be found in the namecache or fetched
3299 * from the filesystem. There is also guaranteed to be only one parent, meaning
3300 * we can just follow vnodes up until we find the root.
3301 *
3302 * The vnode must be referenced.
3303 */
3304 static int
3305 vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3306 size_t *len, size_t addend)
3307 {
3308 #ifdef KDTRACE_HOOKS
3309 struct vnode *startvp = vp;
3310 #endif
3311 struct vnode *vp1;
3312 size_t buflen;
3313 int error;
3314 bool slash_prefixed;
3315
3316 VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp);
3317 VNPASS(vp->v_usecount > 0, vp);
3318
3319 buflen = *len;
3320
3321 slash_prefixed = true;
3322 if (addend == 0) {
3323 MPASS(*len >= 2);
3324 buflen--;
3325 buf[buflen] = '\0';
3326 slash_prefixed = false;
3327 }
3328
3329 error = 0;
3330
3331 SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
3332 counter_u64_add(numfullpathcalls, 1);
3333 while (vp != rdir && vp != rootvnode) {
3334 /*
3335 * The vp vnode must be already fully constructed,
3336 * since it is either found in namecache or obtained
3337 * from VOP_VPTOCNP(). We may test for VV_ROOT safely
3338 * without obtaining the vnode lock.
3339 */
3340 if ((vp->v_vflag & VV_ROOT) != 0) {
3341 vn_lock(vp, LK_RETRY | LK_SHARED);
3342
3343 /*
3344 * With the vnode locked, check for races with
3345 * unmount, forced or not. Note that we
3346 * already verified that vp is not equal to
3347 * the root vnode, which means that
3348 * mnt_vnodecovered can be NULL only for the
3349 * case of unmount.
3350 */
3351 if (VN_IS_DOOMED(vp) ||
3352 (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
3353 vp1->v_mountedhere != vp->v_mount) {
3354 vput(vp);
3355 error = ENOENT;
3356 SDT_PROBE3(vfs, namecache, fullpath, return,
3357 error, vp, NULL);
3358 break;
3359 }
3360
3361 vref(vp1);
3362 vput(vp);
3363 vp = vp1;
3364 continue;
3365 }
3366 if (vp->v_type != VDIR) {
3367 vrele(vp);
3368 counter_u64_add(numfullpathfail1, 1);
3369 error = ENOTDIR;
3370 SDT_PROBE3(vfs, namecache, fullpath, return,
3371 error, vp, NULL);
3372 break;
3373 }
3374 error = vn_vptocnp(&vp, buf, &buflen);
3375 if (error)
3376 break;
3377 if (buflen == 0) {
3378 vrele(vp);
3379 error = ENOMEM;
3380 SDT_PROBE3(vfs, namecache, fullpath, return, error,
3381 startvp, NULL);
3382 break;
3383 }
3384 buf[--buflen] = '/';
3385 slash_prefixed = true;
3386 }
3387 if (error)
3388 return (error);
3389 if (!slash_prefixed) {
3390 if (buflen == 0) {
3391 vrele(vp);
3392 counter_u64_add(numfullpathfail4, 1);
3393 SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
3394 startvp, NULL);
3395 return (ENOMEM);
3396 }
3397 buf[--buflen] = '/';
3398 }
3399 counter_u64_add(numfullpathfound, 1);
3400 vrele(vp);
3401
3402 *retbuf = buf + buflen;
3403 SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf);
3404 *len -= buflen;
3405 *len += addend;
3406 return (0);
3407 }
3408
3409 /*
3410 * Resolve an arbitrary vnode to a pathname.
3411 *
3412 * Note 2 caveats:
3413 * - hardlinks are not tracked, thus if the vnode is not a directory this can
3414 * resolve to a different path than the one used to find it
3415 * - namecache is not mandatory, meaning names are not guaranteed to be added
3416 * (in which case resolving fails)
3417 */
3418 static void __inline
3419 cache_rev_failed_impl(int *reason, int line)
3420 {
3421
3422 *reason = line;
3423 }
3424 #define cache_rev_failed(var) cache_rev_failed_impl((var), __LINE__)
3425
3426 static int
3427 vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
3428 char **retbuf, size_t *buflen, size_t addend)
3429 {
3430 #ifdef KDTRACE_HOOKS
3431 struct vnode *startvp = vp;
3432 #endif
3433 struct vnode *tvp;
3434 struct mount *mp;
3435 struct namecache *ncp;
3436 size_t orig_buflen;
3437 int reason;
3438 int error;
3439 #ifdef KDTRACE_HOOKS
3440 int i;
3441 #endif
3442 seqc_t vp_seqc, tvp_seqc;
3443 u_char nc_flag;
3444
3445 VFS_SMR_ASSERT_ENTERED();
3446
3447 if (!cache_fast_revlookup) {
3448 vfs_smr_exit();
3449 return (-1);
3450 }
3451
3452 orig_buflen = *buflen;
3453
3454 if (addend == 0) {
3455 MPASS(*buflen >= 2);
3456 *buflen -= 1;
3457 buf[*buflen] = '\0';
3458 }
3459
3460 if (vp == rdir || vp == rootvnode) {
3461 if (addend == 0) {
3462 *buflen -= 1;
3463 buf[*buflen] = '/';
3464 }
3465 goto out_ok;
3466 }
3467
3468 #ifdef KDTRACE_HOOKS
3469 i = 0;
3470 #endif
3471 error = -1;
3472 ncp = NULL; /* for sdt probe down below */
3473 vp_seqc = vn_seqc_read_any(vp);
3474 if (seqc_in_modify(vp_seqc)) {
3475 cache_rev_failed(&reason);
3476 goto out_abort;
3477 }
3478
3479 for (;;) {
3480 #ifdef KDTRACE_HOOKS
3481 i++;
3482 #endif
3483 if ((vp->v_vflag & VV_ROOT) != 0) {
3484 mp = atomic_load_ptr(&vp->v_mount);
3485 if (mp == NULL) {
3486 cache_rev_failed(&reason);
3487 goto out_abort;
3488 }
3489 tvp = atomic_load_ptr(&mp->mnt_vnodecovered);
3490 tvp_seqc = vn_seqc_read_any(tvp);
3491 if (seqc_in_modify(tvp_seqc)) {
3492 cache_rev_failed(&reason);
3493 goto out_abort;
3494 }
3495 if (!vn_seqc_consistent(vp, vp_seqc)) {
3496 cache_rev_failed(&reason);
3497 goto out_abort;
3498 }
3499 vp = tvp;
3500 vp_seqc = tvp_seqc;
3501 continue;
3502 }
3503 ncp = atomic_load_consume_ptr(&vp->v_cache_dd);
3504 if (ncp == NULL) {
3505 cache_rev_failed(&reason);
3506 goto out_abort;
3507 }
3508 nc_flag = atomic_load_char(&ncp->nc_flag);
3509 if ((nc_flag & NCF_ISDOTDOT) != 0) {
3510 cache_rev_failed(&reason);
3511 goto out_abort;
3512 }
3513 if (ncp->nc_nlen >= *buflen) {
3514 cache_rev_failed(&reason);
3515 error = ENOMEM;
3516 goto out_abort;
3517 }
3518 *buflen -= ncp->nc_nlen;
3519 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
3520 *buflen -= 1;
3521 buf[*buflen] = '/';
3522 tvp = ncp->nc_dvp;
3523 tvp_seqc = vn_seqc_read_any(tvp);
3524 if (seqc_in_modify(tvp_seqc)) {
3525 cache_rev_failed(&reason);
3526 goto out_abort;
3527 }
3528 if (!vn_seqc_consistent(vp, vp_seqc)) {
3529 cache_rev_failed(&reason);
3530 goto out_abort;
3531 }
3532 /*
3533 * Acquire fence provided by vn_seqc_read_any above.
3534 */
3535 if (__predict_false(atomic_load_ptr(&vp->v_cache_dd) != ncp)) {
3536 cache_rev_failed(&reason);
3537 goto out_abort;
3538 }
3539 if (!cache_ncp_canuse(ncp)) {
3540 cache_rev_failed(&reason);
3541 goto out_abort;
3542 }
3543 vp = tvp;
3544 vp_seqc = tvp_seqc;
3545 if (vp == rdir || vp == rootvnode)
3546 break;
3547 }
3548 out_ok:
3549 vfs_smr_exit();
3550 *retbuf = buf + *buflen;
3551 *buflen = orig_buflen - *buflen + addend;
3552 SDT_PROBE2(vfs, namecache, fullpath_smr, hit, startvp, *retbuf);
3553 return (0);
3554
3555 out_abort:
3556 *buflen = orig_buflen;
3557 SDT_PROBE4(vfs, namecache, fullpath_smr, miss, startvp, ncp, reason, i);
3558 vfs_smr_exit();
3559 return (error);
3560 }
3561
3562 static int
3563 vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3564 size_t *buflen)
3565 {
3566 size_t orig_buflen, addend;
3567 int error;
3568
3569 if (*buflen < 2)
3570 return (EINVAL);
3571
3572 orig_buflen = *buflen;
3573
3574 vref(vp);
3575 addend = 0;
3576 if (vp->v_type != VDIR) {
3577 *buflen -= 1;
3578 buf[*buflen] = '\0';
3579 error = vn_vptocnp(&vp, buf, buflen);
3580 if (error)
3581 return (error);
3582 if (*buflen == 0) {
3583 vrele(vp);
3584 return (ENOMEM);
3585 }
3586 *buflen -= 1;
3587 buf[*buflen] = '/';
3588 addend = orig_buflen - *buflen;
3589 }
3590
3591 return (vn_fullpath_dir(vp, rdir, buf, retbuf, buflen, addend));
3592 }
3593
3594 /*
3595 * Resolve an arbitrary vnode to a pathname (taking care of hardlinks).
3596 *
3597 * Since the namecache does not track hardlinks, the caller is
3598 * expected to first look up the target vnode with SAVENAME |
3599 * WANTPARENT flags passed to namei to get dvp and vp.
3600 *
3601 * Then we have 2 cases:
3602 * - if the found vnode is a directory, the path can be constructed just by
3603 * following names up the chain
3604 * - otherwise we populate the buffer with the saved name and start resolving
3605 * from the parent
3606 */
3607 int
3608 vn_fullpath_hardlink(struct vnode *vp, struct vnode *dvp,
3609 const char *hrdl_name, size_t hrdl_name_length,
3610 char **retbuf, char **freebuf, size_t *buflen)
3611 {
3612 char *buf, *tmpbuf;
3613 struct pwd *pwd;
3614 size_t addend;
3615 int error;
3616 enum vtype type;
3617
3618 if (*buflen < 2)
3619 return (EINVAL);
3620 if (*buflen > MAXPATHLEN)
3621 *buflen = MAXPATHLEN;
3622
3623 buf = malloc(*buflen, M_TEMP, M_WAITOK);
3624
3625 addend = 0;
3626
3627 /*
3628 * Check for VBAD to work around the vp_crossmp bug in lookup().
3629 *
3630 * For example consider tmpfs on /tmp and realpath /tmp. ni_vp will be
3631 * set to mount point's root vnode while ni_dvp will be vp_crossmp.
3632 * If the type is VDIR (like in this very case) we can skip looking
3633 * at ni_dvp in the first place. However, since vnodes get passed here
3634 * unlocked the target may transition to doomed state (type == VBAD)
3635 * before we get to evaluate the condition. If this happens, we will
3636 * populate part of the buffer and descend to vn_fullpath_dir with
3637 * vp == vp_crossmp. Prevent the problem by checking for VBAD.
3638 *
3639 * This should be atomic_load(&vp->v_type) but it is illegal to take
3640 * an address of a bit field, even if said field is sized to char.
3641 * Work around the problem by reading the value into a full-sized enum
3642 * and then re-reading it with atomic_load which will still prevent
3643 * the compiler from re-reading down the road.
3644 */
3645 type = vp->v_type;
3646 type = atomic_load_int(&type);
3647 if (type == VBAD) {
3648 error = ENOENT;
3649 goto out_bad;
3650 }
3651 if (type != VDIR) {
3652 addend = hrdl_name_length + 2;
3653 if (*buflen < addend) {
3654 error = ENOMEM;
3655 goto out_bad;
3656 }
3657 *buflen -= addend;
3658 tmpbuf = buf + *buflen;
3659 tmpbuf[0] = '/';
3660 memcpy(&tmpbuf[1], hrdl_name, hrdl_name_length);
3661 tmpbuf[addend - 1] = '\0';
3662 vp = dvp;
3663 }
3664
3665 vfs_smr_enter();
3666 pwd = pwd_get_smr();
3667 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3668 addend);
3669 VFS_SMR_ASSERT_NOT_ENTERED();
3670 if (error < 0) {
3671 pwd = pwd_hold(curthread);
3672 vref(vp);
3673 error = vn_fullpath_dir(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3674 addend);
3675 pwd_drop(pwd);
3676 }
3677 if (error != 0)
3678 goto out_bad;
3679
3680 *freebuf = buf;
3681
3682 return (0);
3683 out_bad:
3684 free(buf, M_TEMP);
3685 return (error);
3686 }
3687
3688 struct vnode *
3689 vn_dir_dd_ino(struct vnode *vp)
3690 {
3691 struct namecache *ncp;
3692 struct vnode *ddvp;
3693 struct mtx *vlp;
3694 enum vgetstate vs;
3695
3696 ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
3697 vlp = VP2VNODELOCK(vp);
3698 mtx_lock(vlp);
3699 TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
3700 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
3701 continue;
3702 ddvp = ncp->nc_dvp;
3703 vs = vget_prep(ddvp);
3704 mtx_unlock(vlp);
3705 if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
3706 return (NULL);
3707 return (ddvp);
3708 }
3709 mtx_unlock(vlp);
3710 return (NULL);
3711 }
3712
3713 int
3714 vn_commname(struct vnode *vp, char *buf, u_int buflen)
3715 {
3716 struct namecache *ncp;
3717 struct mtx *vlp;
3718 int l;
3719
3720 vlp = VP2VNODELOCK(vp);
3721 mtx_lock(vlp);
3722 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
3723 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
3724 break;
3725 if (ncp == NULL) {
3726 mtx_unlock(vlp);
3727 return (ENOENT);
3728 }
3729 l = min(ncp->nc_nlen, buflen - 1);
3730 memcpy(buf, ncp->nc_name, l);
3731 mtx_unlock(vlp);
3732 buf[l] = '\0';
3733 return (0);
3734 }
3735
3736 /*
3737 * This function updates path string to vnode's full global path
3738 * and checks the size of the new path string against the pathlen argument.
3739 *
3740 * Requires a locked, referenced vnode.
3741 * Vnode is re-locked on success or ENODEV, otherwise unlocked.
3742 *
3743 * If vp is a directory, the call to vn_fullpath_global() always succeeds
3744 * because it falls back to the ".." lookup if the namecache lookup fails.
3745 */
3746 int
3747 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
3748 u_int pathlen)
3749 {
3750 struct nameidata nd;
3751 struct vnode *vp1;
3752 char *rpath, *fbuf;
3753 int error;
3754
3755 ASSERT_VOP_ELOCKED(vp, __func__);
3756
3757 /* Construct global filesystem path from vp. */
3758 VOP_UNLOCK(vp);
3759 error = vn_fullpath_global(vp, &rpath, &fbuf);
3760
3761 if (error != 0) {
3762 vrele(vp);
3763 return (error);
3764 }
3765
3766 if (strlen(rpath) >= pathlen) {
3767 vrele(vp);
3768 error = ENAMETOOLONG;
3769 goto out;
3770 }
3771
3772 /*
3773 * Re-lookup the vnode by path to detect a possible rename.
3774 * As a side effect, the vnode is relocked.
3775 * If vnode was renamed, return ENOENT.
3776 */
3777 NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
3778 UIO_SYSSPACE, path, td);
3779 error = namei(&nd);
3780 if (error != 0) {
3781 vrele(vp);
3782 goto out;
3783 }
3784 NDFREE(&nd, NDF_ONLY_PNBUF);
3785 vp1 = nd.ni_vp;
3786 vrele(vp);
3787 if (vp1 == vp)
3788 strcpy(path, rpath);
3789 else {
3790 vput(vp1);
3791 error = ENOENT;
3792 }
3793
3794 out:
3795 free(fbuf, M_TEMP);
3796 return (error);
3797 }
3798
3799 #ifdef DDB
3800 static void
3801 db_print_vpath(struct vnode *vp)
3802 {
3803
3804 while (vp != NULL) {
3805 db_printf("%p: ", vp);
3806 if (vp == rootvnode) {
3807 db_printf("/");
3808 vp = NULL;
3809 } else {
3810 if (vp->v_vflag & VV_ROOT) {
3811 db_printf("<mount point>");
3812 vp = vp->v_mount->mnt_vnodecovered;
3813 } else {
3814 struct namecache *ncp;
3815 char *ncn;
3816 int i;
3817
3818 ncp = TAILQ_FIRST(&vp->v_cache_dst);
3819 if (ncp != NULL) {
3820 ncn = ncp->nc_name;
3821 for (i = 0; i < ncp->nc_nlen; i++)
3822 db_printf("%c", *ncn++);
3823 vp = ncp->nc_dvp;
3824 } else {
3825 vp = NULL;
3826 }
3827 }
3828 }
3829 db_printf("\n");
3830 }
3831
3832 return;
3833 }
3834
3835 DB_SHOW_COMMAND(vpath, db_show_vpath)
3836 {
3837 struct vnode *vp;
3838
3839 if (!have_addr) {
3840 db_printf("usage: show vpath <struct vnode *>\n");
3841 return;
3842 }
3843
3844 vp = (struct vnode *)addr;
3845 db_print_vpath(vp);
3846 }
3847
3848 #endif
3849
3850 static int cache_fast_lookup = 1;
3851 static char __read_frequently cache_fast_lookup_enabled = true;
3852
3853 #define CACHE_FPL_FAILED -2020
3854
3855 void
3856 cache_fast_lookup_enabled_recalc(void)
3857 {
3858 int lookup_flag;
3859 int mac_on;
3860
3861 #ifdef MAC
3862 mac_on = mac_vnode_check_lookup_enabled();
3863 mac_on |= mac_vnode_check_readlink_enabled();
3864 #else
3865 mac_on = 0;
3866 #endif
3867
3868 lookup_flag = atomic_load_int(&cache_fast_lookup);
3869 if (lookup_flag && !mac_on) {
3870 atomic_store_char(&cache_fast_lookup_enabled, true);
3871 } else {
3872 atomic_store_char(&cache_fast_lookup_enabled, false);
3873 }
3874 }
3875
3876 static int
3877 syscal_vfs_cache_fast_lookup(SYSCTL_HANDLER_ARGS)
3878 {
3879 int error, old;
3880
3881 old = atomic_load_int(&cache_fast_lookup);
3882 error = sysctl_handle_int(oidp, arg1, arg2, req);
3883 if (error == 0 && req->newptr && old != atomic_load_int(&cache_fast_lookup))
3884 cache_fast_lookup_enabled_recalc();
3885 return (error);
3886 }
3887 SYSCTL_PROC(_vfs, OID_AUTO, cache_fast_lookup, CTLTYPE_INT|CTLFLAG_RW|CTLFLAG_MPSAFE,
3888 &cache_fast_lookup, 0, syscal_vfs_cache_fast_lookup, "IU", "");
3889
3890 /*
3891 * Components of nameidata (or objects it can point to) which may
3892 * need restoring in case fast path lookup fails.
3893 */
3894 struct nameidata_outer {
3895 size_t ni_pathlen;
3896 int cn_flags;
3897 };
3898
3899 struct nameidata_saved {
3900 #ifdef INVARIANTS
3901 char *cn_nameptr;
3902 size_t ni_pathlen;
3903 #endif
3904 };
3905
3906 #ifdef INVARIANTS
3907 struct cache_fpl_debug {
3908 size_t ni_pathlen;
3909 };
3910 #endif
3911
3912 struct cache_fpl {
3913 struct nameidata *ndp;
3914 struct componentname *cnp;
3915 char *nulchar;
3916 struct vnode *dvp;
3917 struct vnode *tvp;
3918 seqc_t dvp_seqc;
3919 seqc_t tvp_seqc;
3920 uint32_t hash;
3921 struct nameidata_saved snd;
3922 struct nameidata_outer snd_outer;
3923 int line;
3924 enum cache_fpl_status status:8;
3925 bool in_smr;
3926 bool fsearch;
3927 bool savename;
3928 struct pwd **pwd;
3929 #ifdef INVARIANTS
3930 struct cache_fpl_debug debug;
3931 #endif
3932 };
3933
3934 static bool cache_fplookup_mp_supported(struct mount *mp);
3935 static bool cache_fplookup_is_mp(struct cache_fpl *fpl);
3936 static int cache_fplookup_cross_mount(struct cache_fpl *fpl);
3937 static int cache_fplookup_partial_setup(struct cache_fpl *fpl);
3938 static int cache_fplookup_skip_slashes(struct cache_fpl *fpl);
3939 static int cache_fplookup_trailingslash(struct cache_fpl *fpl);
3940 static void cache_fpl_pathlen_dec(struct cache_fpl *fpl);
3941 static void cache_fpl_pathlen_inc(struct cache_fpl *fpl);
3942 static void cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n);
3943 static void cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n);
3944
3945 static void
3946 cache_fpl_cleanup_cnp(struct componentname *cnp)
3947 {
3948
3949 uma_zfree(namei_zone, cnp->cn_pnbuf);
3950 #ifdef DIAGNOSTIC
3951 cnp->cn_pnbuf = NULL;
3952 cnp->cn_nameptr = NULL;
3953 #endif
3954 }
3955
3956 static struct vnode *
3957 cache_fpl_handle_root(struct cache_fpl *fpl)
3958 {
3959 struct nameidata *ndp;
3960 struct componentname *cnp;
3961
3962 ndp = fpl->ndp;
3963 cnp = fpl->cnp;
3964
3965 MPASS(*(cnp->cn_nameptr) == '/');
3966 cnp->cn_nameptr++;
3967 cache_fpl_pathlen_dec(fpl);
3968
3969 if (__predict_false(*(cnp->cn_nameptr) == '/')) {
3970 do {
3971 cnp->cn_nameptr++;
3972 cache_fpl_pathlen_dec(fpl);
3973 } while (*(cnp->cn_nameptr) == '/');
3974 }
3975
3976 return (ndp->ni_rootdir);
3977 }
3978
3979 static void
3980 cache_fpl_checkpoint_outer(struct cache_fpl *fpl)
3981 {
3982
3983 fpl->snd_outer.ni_pathlen = fpl->ndp->ni_pathlen;
3984 fpl->snd_outer.cn_flags = fpl->ndp->ni_cnd.cn_flags;
3985 }
3986
3987 static void
3988 cache_fpl_checkpoint(struct cache_fpl *fpl)
3989 {
3990
3991 #ifdef INVARIANTS
3992 fpl->snd.cn_nameptr = fpl->ndp->ni_cnd.cn_nameptr;
3993 fpl->snd.ni_pathlen = fpl->debug.ni_pathlen;
3994 #endif
3995 }
3996
3997 static void
3998 cache_fpl_restore_partial(struct cache_fpl *fpl)
3999 {
4000
4001 fpl->ndp->ni_cnd.cn_flags = fpl->snd_outer.cn_flags;
4002 #ifdef INVARIANTS
4003 fpl->debug.ni_pathlen = fpl->snd.ni_pathlen;
4004 #endif
4005 }
4006
4007 static void
4008 cache_fpl_restore_abort(struct cache_fpl *fpl)
4009 {
4010
4011 cache_fpl_restore_partial(fpl);
4012 /*
4013 * It is 0 on entry by API contract.
4014 */
4015 fpl->ndp->ni_resflags = 0;
4016 fpl->ndp->ni_cnd.cn_nameptr = fpl->ndp->ni_cnd.cn_pnbuf;
4017 fpl->ndp->ni_pathlen = fpl->snd_outer.ni_pathlen;
4018 }
4019
4020 #ifdef INVARIANTS
4021 #define cache_fpl_smr_assert_entered(fpl) ({ \
4022 struct cache_fpl *_fpl = (fpl); \
4023 MPASS(_fpl->in_smr == true); \
4024 VFS_SMR_ASSERT_ENTERED(); \
4025 })
4026 #define cache_fpl_smr_assert_not_entered(fpl) ({ \
4027 struct cache_fpl *_fpl = (fpl); \
4028 MPASS(_fpl->in_smr == false); \
4029 VFS_SMR_ASSERT_NOT_ENTERED(); \
4030 })
4031 static void
4032 cache_fpl_assert_status(struct cache_fpl *fpl)
4033 {
4034
4035 switch (fpl->status) {
4036 case CACHE_FPL_STATUS_UNSET:
4037 __assert_unreachable();
4038 break;
4039 case CACHE_FPL_STATUS_DESTROYED:
4040 case CACHE_FPL_STATUS_ABORTED:
4041 case CACHE_FPL_STATUS_PARTIAL:
4042 case CACHE_FPL_STATUS_HANDLED:
4043 break;
4044 }
4045 }
4046 #else
4047 #define cache_fpl_smr_assert_entered(fpl) do { } while (0)
4048 #define cache_fpl_smr_assert_not_entered(fpl) do { } while (0)
4049 #define cache_fpl_assert_status(fpl) do { } while (0)
4050 #endif
4051
4052 #define cache_fpl_smr_enter_initial(fpl) ({ \
4053 struct cache_fpl *_fpl = (fpl); \
4054 vfs_smr_enter(); \
4055 _fpl->in_smr = true; \
4056 })
4057
4058 #define cache_fpl_smr_enter(fpl) ({ \
4059 struct cache_fpl *_fpl = (fpl); \
4060 MPASS(_fpl->in_smr == false); \
4061 vfs_smr_enter(); \
4062 _fpl->in_smr = true; \
4063 })
4064
4065 #define cache_fpl_smr_exit(fpl) ({ \
4066 struct cache_fpl *_fpl = (fpl); \
4067 MPASS(_fpl->in_smr == true); \
4068 vfs_smr_exit(); \
4069 _fpl->in_smr = false; \
4070 })
4071
4072 static int
4073 cache_fpl_aborted_early_impl(struct cache_fpl *fpl, int line)
4074 {
4075
4076 if (fpl->status != CACHE_FPL_STATUS_UNSET) {
4077 KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL,
4078 ("%s: converting to abort from %d at %d, set at %d\n",
4079 __func__, fpl->status, line, fpl->line));
4080 }
4081 cache_fpl_smr_assert_not_entered(fpl);
4082 fpl->status = CACHE_FPL_STATUS_ABORTED;
4083 fpl->line = line;
4084 return (CACHE_FPL_FAILED);
4085 }
4086
4087 #define cache_fpl_aborted_early(x) cache_fpl_aborted_early_impl((x), __LINE__)
4088
4089 static int __noinline
4090 cache_fpl_aborted_impl(struct cache_fpl *fpl, int line)
4091 {
4092 struct nameidata *ndp;
4093 struct componentname *cnp;
4094
4095 ndp = fpl->ndp;
4096 cnp = fpl->cnp;
4097
4098 if (fpl->status != CACHE_FPL_STATUS_UNSET) {
4099 KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL,
4100 ("%s: converting to abort from %d at %d, set at %d\n",
4101 __func__, fpl->status, line, fpl->line));
4102 }
4103 fpl->status = CACHE_FPL_STATUS_ABORTED;
4104 fpl->line = line;
4105 if (fpl->in_smr)
4106 cache_fpl_smr_exit(fpl);
4107 cache_fpl_restore_abort(fpl);
4108 /*
4109 * Resolving symlinks overwrites data passed by the caller.
4110 * Let namei know.
4111 */
4112 if (ndp->ni_loopcnt > 0) {
4113 fpl->status = CACHE_FPL_STATUS_DESTROYED;
4114 cache_fpl_cleanup_cnp(cnp);
4115 }
4116 return (CACHE_FPL_FAILED);
4117 }
4118
4119 #define cache_fpl_aborted(x) cache_fpl_aborted_impl((x), __LINE__)
4120
4121 static int __noinline
4122 cache_fpl_partial_impl(struct cache_fpl *fpl, int line)
4123 {
4124
4125 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
4126 ("%s: setting to partial at %d, but already set to %d at %d\n",
4127 __func__, line, fpl->status, fpl->line));
4128 cache_fpl_smr_assert_entered(fpl);
4129 fpl->status = CACHE_FPL_STATUS_PARTIAL;
4130 fpl->line = line;
4131 return (cache_fplookup_partial_setup(fpl));
4132 }
4133
4134 #define cache_fpl_partial(x) cache_fpl_partial_impl((x), __LINE__)
4135
4136 static int
4137 cache_fpl_handled_impl(struct cache_fpl *fpl, int line)
4138 {
4139
4140 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
4141 ("%s: setting to handled at %d, but already set to %d at %d\n",
4142 __func__, line, fpl->status, fpl->line));
4143 cache_fpl_smr_assert_not_entered(fpl);
4144 fpl->status = CACHE_FPL_STATUS_HANDLED;
4145 fpl->line = line;
4146 return (0);
4147 }
4148
4149 #define cache_fpl_handled(x) cache_fpl_handled_impl((x), __LINE__)
4150
4151 static int
4152 cache_fpl_handled_error_impl(struct cache_fpl *fpl, int error, int line)
4153 {
4154
4155 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
4156 ("%s: setting to handled at %d, but already set to %d at %d\n",
4157 __func__, line, fpl->status, fpl->line));
4158 MPASS(error != 0);
4159 MPASS(error != CACHE_FPL_FAILED);
4160 cache_fpl_smr_assert_not_entered(fpl);
4161 fpl->status = CACHE_FPL_STATUS_HANDLED;
4162 fpl->line = line;
4163 fpl->dvp = NULL;
4164 fpl->tvp = NULL;
4165 fpl->savename = false;
4166 return (error);
4167 }
4168
4169 #define cache_fpl_handled_error(x, e) cache_fpl_handled_error_impl((x), (e), __LINE__)
4170
4171 static bool
4172 cache_fpl_terminated(struct cache_fpl *fpl)
4173 {
4174
4175 return (fpl->status != CACHE_FPL_STATUS_UNSET);
4176 }
4177
4178 #define CACHE_FPL_SUPPORTED_CN_FLAGS \
4179 (NC_NOMAKEENTRY | NC_KEEPPOSENTRY | LOCKLEAF | LOCKPARENT | WANTPARENT | \
4180 FAILIFEXISTS | FOLLOW | LOCKSHARED | SAVENAME | SAVESTART | WILLBEDIR | \
4181 ISOPEN | NOMACCHECK | AUDITVNODE1 | AUDITVNODE2 | NOCAPCHECK)
4182
4183 #define CACHE_FPL_INTERNAL_CN_FLAGS \
4184 (ISDOTDOT | MAKEENTRY | ISLASTCN)
4185
4186 _Static_assert((CACHE_FPL_SUPPORTED_CN_FLAGS & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
4187 "supported and internal flags overlap");
4188
4189 static bool
4190 cache_fpl_islastcn(struct nameidata *ndp)
4191 {
4192
4193 return (*ndp->ni_next == 0);
4194 }
4195
4196 static bool
4197 cache_fpl_istrailingslash(struct cache_fpl *fpl)
4198 {
4199
4200 return (*(fpl->nulchar - 1) == '/');
4201 }
4202
4203 static bool
4204 cache_fpl_isdotdot(struct componentname *cnp)
4205 {
4206
4207 if (cnp->cn_namelen == 2 &&
4208 cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
4209 return (true);
4210 return (false);
4211 }
4212
4213 static bool
4214 cache_can_fplookup(struct cache_fpl *fpl)
4215 {
4216 struct nameidata *ndp;
4217 struct componentname *cnp;
4218 struct thread *td;
4219
4220 ndp = fpl->ndp;
4221 cnp = fpl->cnp;
4222 td = cnp->cn_thread;
4223
4224 if (!atomic_load_char(&cache_fast_lookup_enabled)) {
4225 cache_fpl_aborted_early(fpl);
4226 return (false);
4227 }
4228 if ((cnp->cn_flags & ~CACHE_FPL_SUPPORTED_CN_FLAGS) != 0) {
4229 cache_fpl_aborted_early(fpl);
4230 return (false);
4231 }
4232 if (IN_CAPABILITY_MODE(td)) {
4233 cache_fpl_aborted_early(fpl);
4234 return (false);
4235 }
4236 if (AUDITING_TD(td)) {
4237 cache_fpl_aborted_early(fpl);
4238 return (false);
4239 }
4240 if (ndp->ni_startdir != NULL) {
4241 cache_fpl_aborted_early(fpl);
4242 return (false);
4243 }
4244 return (true);
4245 }
4246
4247 static int
4248 cache_fplookup_dirfd(struct cache_fpl *fpl, struct vnode **vpp)
4249 {
4250 struct nameidata *ndp;
4251 int error;
4252 bool fsearch;
4253
4254 ndp = fpl->ndp;
4255 error = fgetvp_lookup_smr(ndp->ni_dirfd, ndp, vpp, &fsearch);
4256 if (__predict_false(error != 0)) {
4257 return (cache_fpl_aborted(fpl));
4258 }
4259 fpl->fsearch = fsearch;
4260 return (0);
4261 }
4262
4263 static int __noinline
4264 cache_fplookup_negative_promote(struct cache_fpl *fpl, struct namecache *oncp,
4265 uint32_t hash)
4266 {
4267 struct componentname *cnp;
4268 struct vnode *dvp;
4269
4270 cnp = fpl->cnp;
4271 dvp = fpl->dvp;
4272
4273 cache_fpl_smr_exit(fpl);
4274 if (cache_neg_promote_cond(dvp, cnp, oncp, hash))
4275 return (cache_fpl_handled_error(fpl, ENOENT));
4276 else
4277 return (cache_fpl_aborted(fpl));
4278 }
4279
4280 /*
4281 * The target vnode is not supported, prepare for the slow path to take over.
4282 */
4283 static int __noinline
4284 cache_fplookup_partial_setup(struct cache_fpl *fpl)
4285 {
4286 struct nameidata *ndp;
4287 struct componentname *cnp;
4288 enum vgetstate dvs;
4289 struct vnode *dvp;
4290 struct pwd *pwd;
4291 seqc_t dvp_seqc;
4292
4293 ndp = fpl->ndp;
4294 cnp = fpl->cnp;
4295 pwd = *(fpl->pwd);
4296 dvp = fpl->dvp;
4297 dvp_seqc = fpl->dvp_seqc;
4298
4299 if (!pwd_hold_smr(pwd)) {
4300 return (cache_fpl_aborted(fpl));
4301 }
4302
4303 /*
4304 * Note that seqc is checked before the vnode is locked, so by
4305 * the time regular lookup gets to it it may have moved.
4306 *
4307 * Ultimately this does not affect correctness, any lookup errors
4308 * are userspace racing with itself. It is guaranteed that any
4309 * path which ultimately gets found could also have been found
4310 * by regular lookup going all the way in absence of concurrent
4311 * modifications.
4312 */
4313 dvs = vget_prep_smr(dvp);
4314 cache_fpl_smr_exit(fpl);
4315 if (__predict_false(dvs == VGET_NONE)) {
4316 pwd_drop(pwd);
4317 return (cache_fpl_aborted(fpl));
4318 }
4319
4320 vget_finish_ref(dvp, dvs);
4321 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4322 vrele(dvp);
4323 pwd_drop(pwd);
4324 return (cache_fpl_aborted(fpl));
4325 }
4326
4327 cache_fpl_restore_partial(fpl);
4328 #ifdef INVARIANTS
4329 if (cnp->cn_nameptr != fpl->snd.cn_nameptr) {
4330 panic("%s: cn_nameptr mismatch (%p != %p) full [%s]\n", __func__,
4331 cnp->cn_nameptr, fpl->snd.cn_nameptr, cnp->cn_pnbuf);
4332 }
4333 #endif
4334
4335 ndp->ni_startdir = dvp;
4336 cnp->cn_flags |= MAKEENTRY;
4337 if (cache_fpl_islastcn(ndp))
4338 cnp->cn_flags |= ISLASTCN;
4339 if (cache_fpl_isdotdot(cnp))
4340 cnp->cn_flags |= ISDOTDOT;
4341
4342 /*
4343 * Skip potential extra slashes parsing did not take care of.
4344 * cache_fplookup_skip_slashes explains the mechanism.
4345 */
4346 if (__predict_false(*(cnp->cn_nameptr) == '/')) {
4347 do {
4348 cnp->cn_nameptr++;
4349 cache_fpl_pathlen_dec(fpl);
4350 } while (*(cnp->cn_nameptr) == '/');
4351 }
4352
4353 ndp->ni_pathlen = fpl->nulchar - cnp->cn_nameptr + 1;
4354 #ifdef INVARIANTS
4355 if (ndp->ni_pathlen != fpl->debug.ni_pathlen) {
4356 panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n",
4357 __func__, ndp->ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar,
4358 cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf);
4359 }
4360 #endif
4361 return (0);
4362 }
4363
4364 static int
4365 cache_fplookup_final_child(struct cache_fpl *fpl, enum vgetstate tvs)
4366 {
4367 struct componentname *cnp;
4368 struct vnode *tvp;
4369 seqc_t tvp_seqc;
4370 int error, lkflags;
4371
4372 cnp = fpl->cnp;
4373 tvp = fpl->tvp;
4374 tvp_seqc = fpl->tvp_seqc;
4375
4376 if ((cnp->cn_flags & LOCKLEAF) != 0) {
4377 lkflags = LK_SHARED;
4378 if ((cnp->cn_flags & LOCKSHARED) == 0)
4379 lkflags = LK_EXCLUSIVE;
4380 error = vget_finish(tvp, lkflags, tvs);
4381 if (__predict_false(error != 0)) {
4382 return (cache_fpl_aborted(fpl));
4383 }
4384 } else {
4385 vget_finish_ref(tvp, tvs);
4386 }
4387
4388 if (!vn_seqc_consistent(tvp, tvp_seqc)) {
4389 if ((cnp->cn_flags & LOCKLEAF) != 0)
4390 vput(tvp);
4391 else
4392 vrele(tvp);
4393 return (cache_fpl_aborted(fpl));
4394 }
4395
4396 return (cache_fpl_handled(fpl));
4397 }
4398
4399 /*
4400 * They want to possibly modify the state of the namecache.
4401 */
4402 static int __noinline
4403 cache_fplookup_final_modifying(struct cache_fpl *fpl)
4404 {
4405 struct nameidata *ndp;
4406 struct componentname *cnp;
4407 enum vgetstate dvs;
4408 struct vnode *dvp, *tvp;
4409 struct mount *mp;
4410 seqc_t dvp_seqc;
4411 int error;
4412 bool docache;
4413
4414 ndp = fpl->ndp;
4415 cnp = fpl->cnp;
4416 dvp = fpl->dvp;
4417 dvp_seqc = fpl->dvp_seqc;
4418
4419 MPASS(*(cnp->cn_nameptr) != '/');
4420 MPASS(cache_fpl_islastcn(ndp));
4421 if ((cnp->cn_flags & LOCKPARENT) == 0)
4422 MPASS((cnp->cn_flags & WANTPARENT) != 0);
4423 MPASS((cnp->cn_flags & TRAILINGSLASH) == 0);
4424 MPASS(cnp->cn_nameiop == CREATE || cnp->cn_nameiop == DELETE ||
4425 cnp->cn_nameiop == RENAME);
4426 MPASS((cnp->cn_flags & MAKEENTRY) == 0);
4427 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
4428
4429 docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
4430 if (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)
4431 docache = false;
4432
4433 /*
4434 * Regular lookup nulifies the slash, which we don't do here.
4435 * Don't take chances with filesystem routines seeing it for
4436 * the last entry.
4437 */
4438 if (cache_fpl_istrailingslash(fpl)) {
4439 return (cache_fpl_partial(fpl));
4440 }
4441
4442 mp = atomic_load_ptr(&dvp->v_mount);
4443 if (__predict_false(mp == NULL)) {
4444 return (cache_fpl_aborted(fpl));
4445 }
4446
4447 if (__predict_false(mp->mnt_flag & MNT_RDONLY)) {
4448 cache_fpl_smr_exit(fpl);
4449 /*
4450 * Original code keeps not checking for CREATE which
4451 * might be a bug. For now let the old lookup decide.
4452 */
4453 if (cnp->cn_nameiop == CREATE) {
4454 return (cache_fpl_aborted(fpl));
4455 }
4456 return (cache_fpl_handled_error(fpl, EROFS));
4457 }
4458
4459 if (fpl->tvp != NULL && (cnp->cn_flags & FAILIFEXISTS) != 0) {
4460 cache_fpl_smr_exit(fpl);
4461 return (cache_fpl_handled_error(fpl, EEXIST));
4462 }
4463
4464 /*
4465 * Secure access to dvp; check cache_fplookup_partial_setup for
4466 * reasoning.
4467 *
4468 * XXX At least UFS requires its lookup routine to be called for
4469 * the last path component, which leads to some level of complication
4470 * and inefficiency:
4471 * - the target routine always locks the target vnode, but our caller
4472 * may not need it locked
4473 * - some of the VOP machinery asserts that the parent is locked, which
4474 * once more may be not required
4475 *
4476 * TODO: add a flag for filesystems which don't need this.
4477 */
4478 dvs = vget_prep_smr(dvp);
4479 cache_fpl_smr_exit(fpl);
4480 if (__predict_false(dvs == VGET_NONE)) {
4481 return (cache_fpl_aborted(fpl));
4482 }
4483
4484 vget_finish_ref(dvp, dvs);
4485 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4486 vrele(dvp);
4487 return (cache_fpl_aborted(fpl));
4488 }
4489
4490 error = vn_lock(dvp, LK_EXCLUSIVE);
4491 if (__predict_false(error != 0)) {
4492 vrele(dvp);
4493 return (cache_fpl_aborted(fpl));
4494 }
4495
4496 tvp = NULL;
4497 cnp->cn_flags |= ISLASTCN;
4498 if (docache)
4499 cnp->cn_flags |= MAKEENTRY;
4500 if (cache_fpl_isdotdot(cnp))
4501 cnp->cn_flags |= ISDOTDOT;
4502 cnp->cn_lkflags = LK_EXCLUSIVE;
4503 error = VOP_LOOKUP(dvp, &tvp, cnp);
4504 switch (error) {
4505 case EJUSTRETURN:
4506 case 0:
4507 break;
4508 case ENOTDIR:
4509 case ENOENT:
4510 vput(dvp);
4511 return (cache_fpl_handled_error(fpl, error));
4512 default:
4513 vput(dvp);
4514 return (cache_fpl_aborted(fpl));
4515 }
4516
4517 fpl->tvp = tvp;
4518 fpl->savename = (cnp->cn_flags & SAVENAME) != 0;
4519
4520 if (tvp == NULL) {
4521 if ((cnp->cn_flags & SAVESTART) != 0) {
4522 ndp->ni_startdir = dvp;
4523 vrefact(ndp->ni_startdir);
4524 cnp->cn_flags |= SAVENAME;
4525 fpl->savename = true;
4526 }
4527 MPASS(error == EJUSTRETURN);
4528 if ((cnp->cn_flags & LOCKPARENT) == 0) {
4529 VOP_UNLOCK(dvp);
4530 }
4531 return (cache_fpl_handled(fpl));
4532 }
4533
4534 /*
4535 * There are very hairy corner cases concerning various flag combinations
4536 * and locking state. In particular here we only hold one lock instead of
4537 * two.
4538 *
4539 * Skip the complexity as it is of no significance for normal workloads.
4540 */
4541 if (__predict_false(tvp == dvp)) {
4542 vput(dvp);
4543 vrele(tvp);
4544 return (cache_fpl_aborted(fpl));
4545 }
4546
4547 /*
4548 * If they want the symlink itself we are fine, but if they want to
4549 * follow it regular lookup has to be engaged.
4550 */
4551 if (tvp->v_type == VLNK) {
4552 if ((cnp->cn_flags & FOLLOW) != 0) {
4553 vput(dvp);
4554 vput(tvp);
4555 return (cache_fpl_aborted(fpl));
4556 }
4557 }
4558
4559 /*
4560 * Since we expect this to be the terminal vnode it should almost never
4561 * be a mount point.
4562 */
4563 if (__predict_false(cache_fplookup_is_mp(fpl))) {
4564 vput(dvp);
4565 vput(tvp);
4566 return (cache_fpl_aborted(fpl));
4567 }
4568
4569 if ((cnp->cn_flags & FAILIFEXISTS) != 0) {
4570 vput(dvp);
4571 vput(tvp);
4572 return (cache_fpl_handled_error(fpl, EEXIST));
4573 }
4574
4575 if ((cnp->cn_flags & LOCKLEAF) == 0) {
4576 VOP_UNLOCK(tvp);
4577 }
4578
4579 if ((cnp->cn_flags & LOCKPARENT) == 0) {
4580 VOP_UNLOCK(dvp);
4581 }
4582
4583 if ((cnp->cn_flags & SAVESTART) != 0) {
4584 ndp->ni_startdir = dvp;
4585 vrefact(ndp->ni_startdir);
4586 cnp->cn_flags |= SAVENAME;
4587 fpl->savename = true;
4588 }
4589
4590 return (cache_fpl_handled(fpl));
4591 }
4592
4593 static int __noinline
4594 cache_fplookup_modifying(struct cache_fpl *fpl)
4595 {
4596 struct nameidata *ndp;
4597
4598 ndp = fpl->ndp;
4599
4600 if (!cache_fpl_islastcn(ndp)) {
4601 return (cache_fpl_partial(fpl));
4602 }
4603 return (cache_fplookup_final_modifying(fpl));
4604 }
4605
4606 static int __noinline
4607 cache_fplookup_final_withparent(struct cache_fpl *fpl)
4608 {
4609 struct componentname *cnp;
4610 enum vgetstate dvs, tvs;
4611 struct vnode *dvp, *tvp;
4612 seqc_t dvp_seqc;
4613 int error;
4614
4615 cnp = fpl->cnp;
4616 dvp = fpl->dvp;
4617 dvp_seqc = fpl->dvp_seqc;
4618 tvp = fpl->tvp;
4619
4620 MPASS((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0);
4621
4622 /*
4623 * This is less efficient than it can be for simplicity.
4624 */
4625 dvs = vget_prep_smr(dvp);
4626 if (__predict_false(dvs == VGET_NONE)) {
4627 return (cache_fpl_aborted(fpl));
4628 }
4629 tvs = vget_prep_smr(tvp);
4630 if (__predict_false(tvs == VGET_NONE)) {
4631 cache_fpl_smr_exit(fpl);
4632 vget_abort(dvp, dvs);
4633 return (cache_fpl_aborted(fpl));
4634 }
4635
4636 cache_fpl_smr_exit(fpl);
4637
4638 if ((cnp->cn_flags & LOCKPARENT) != 0) {
4639 error = vget_finish(dvp, LK_EXCLUSIVE, dvs);
4640 if (__predict_false(error != 0)) {
4641 vget_abort(tvp, tvs);
4642 return (cache_fpl_aborted(fpl));
4643 }
4644 } else {
4645 vget_finish_ref(dvp, dvs);
4646 }
4647
4648 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4649 vget_abort(tvp, tvs);
4650 if ((cnp->cn_flags & LOCKPARENT) != 0)
4651 vput(dvp);
4652 else
4653 vrele(dvp);
4654 return (cache_fpl_aborted(fpl));
4655 }
4656
4657 error = cache_fplookup_final_child(fpl, tvs);
4658 if (__predict_false(error != 0)) {
4659 MPASS(fpl->status == CACHE_FPL_STATUS_ABORTED ||
4660 fpl->status == CACHE_FPL_STATUS_DESTROYED);
4661 if ((cnp->cn_flags & LOCKPARENT) != 0)
4662 vput(dvp);
4663 else
4664 vrele(dvp);
4665 return (error);
4666 }
4667
4668 MPASS(fpl->status == CACHE_FPL_STATUS_HANDLED);
4669 return (0);
4670 }
4671
4672 static int
4673 cache_fplookup_final(struct cache_fpl *fpl)
4674 {
4675 struct componentname *cnp;
4676 enum vgetstate tvs;
4677 struct vnode *dvp, *tvp;
4678 seqc_t dvp_seqc;
4679
4680 cnp = fpl->cnp;
4681 dvp = fpl->dvp;
4682 dvp_seqc = fpl->dvp_seqc;
4683 tvp = fpl->tvp;
4684
4685 MPASS(*(cnp->cn_nameptr) != '/');
4686
4687 if (cnp->cn_nameiop != LOOKUP) {
4688 return (cache_fplookup_final_modifying(fpl));
4689 }
4690
4691 if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0)
4692 return (cache_fplookup_final_withparent(fpl));
4693
4694 tvs = vget_prep_smr(tvp);
4695 if (__predict_false(tvs == VGET_NONE)) {
4696 return (cache_fpl_partial(fpl));
4697 }
4698
4699 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4700 cache_fpl_smr_exit(fpl);
4701 vget_abort(tvp, tvs);
4702 return (cache_fpl_aborted(fpl));
4703 }
4704
4705 cache_fpl_smr_exit(fpl);
4706 return (cache_fplookup_final_child(fpl, tvs));
4707 }
4708
4709 /*
4710 * Comment from locked lookup:
4711 * Check for degenerate name (e.g. / or "") which is a way of talking about a
4712 * directory, e.g. like "/." or ".".
4713 */
4714 static int __noinline
4715 cache_fplookup_degenerate(struct cache_fpl *fpl)
4716 {
4717 struct componentname *cnp;
4718 struct vnode *dvp;
4719 enum vgetstate dvs;
4720 int error, lkflags;
4721 #ifdef INVARIANTS
4722 char *cp;
4723 #endif
4724
4725 fpl->tvp = fpl->dvp;
4726 fpl->tvp_seqc = fpl->dvp_seqc;
4727
4728 cnp = fpl->cnp;
4729 dvp = fpl->dvp;
4730
4731 #ifdef INVARIANTS
4732 for (cp = cnp->cn_pnbuf; *cp != '\0'; cp++) {
4733 KASSERT(*cp == '/',
4734 ("%s: encountered non-slash; string [%s]\n", __func__,
4735 cnp->cn_pnbuf));
4736 }
4737 #endif
4738
4739 if (__predict_false(cnp->cn_nameiop != LOOKUP)) {
4740 cache_fpl_smr_exit(fpl);
4741 return (cache_fpl_handled_error(fpl, EISDIR));
4742 }
4743
4744 MPASS((cnp->cn_flags & SAVESTART) == 0);
4745
4746 if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0) {
4747 return (cache_fplookup_final_withparent(fpl));
4748 }
4749
4750 dvs = vget_prep_smr(dvp);
4751 cache_fpl_smr_exit(fpl);
4752 if (__predict_false(dvs == VGET_NONE)) {
4753 return (cache_fpl_aborted(fpl));
4754 }
4755
4756 if ((cnp->cn_flags & LOCKLEAF) != 0) {
4757 lkflags = LK_SHARED;
4758 if ((cnp->cn_flags & LOCKSHARED) == 0)
4759 lkflags = LK_EXCLUSIVE;
4760 error = vget_finish(dvp, lkflags, dvs);
4761 if (__predict_false(error != 0)) {
4762 return (cache_fpl_aborted(fpl));
4763 }
4764 } else {
4765 vget_finish_ref(dvp, dvs);
4766 }
4767 return (cache_fpl_handled(fpl));
4768 }
4769
4770 static int __noinline
4771 cache_fplookup_noentry(struct cache_fpl *fpl)
4772 {
4773 struct nameidata *ndp;
4774 struct componentname *cnp;
4775 enum vgetstate dvs;
4776 struct vnode *dvp, *tvp;
4777 seqc_t dvp_seqc;
4778 int error;
4779 bool docache;
4780
4781 ndp = fpl->ndp;
4782 cnp = fpl->cnp;
4783 dvp = fpl->dvp;
4784 dvp_seqc = fpl->dvp_seqc;
4785
4786 MPASS((cnp->cn_flags & MAKEENTRY) == 0);
4787 MPASS((cnp->cn_flags & ISDOTDOT) == 0);
4788 MPASS(!cache_fpl_isdotdot(cnp));
4789
4790 /*
4791 * Hack: delayed name len checking.
4792 */
4793 if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
4794 cache_fpl_smr_exit(fpl);
4795 return (cache_fpl_handled_error(fpl, ENAMETOOLONG));
4796 }
4797
4798 if (cnp->cn_nameptr[0] == '/') {
4799 return (cache_fplookup_skip_slashes(fpl));
4800 }
4801
4802 if (cnp->cn_nameptr[0] == '\0') {
4803 if (fpl->tvp == NULL) {
4804 return (cache_fplookup_degenerate(fpl));
4805 }
4806 return (cache_fplookup_trailingslash(fpl));
4807 }
4808
4809 if (cnp->cn_nameiop != LOOKUP) {
4810 fpl->tvp = NULL;
4811 return (cache_fplookup_modifying(fpl));
4812 }
4813
4814 MPASS((cnp->cn_flags & SAVESTART) == 0);
4815
4816 /*
4817 * Only try to fill in the component if it is the last one,
4818 * otherwise not only there may be several to handle but the
4819 * walk may be complicated.
4820 */
4821 if (!cache_fpl_islastcn(ndp)) {
4822 return (cache_fpl_partial(fpl));
4823 }
4824
4825 /*
4826 * Regular lookup nulifies the slash, which we don't do here.
4827 * Don't take chances with filesystem routines seeing it for
4828 * the last entry.
4829 */
4830 if (cache_fpl_istrailingslash(fpl)) {
4831 return (cache_fpl_partial(fpl));
4832 }
4833
4834 /*
4835 * Secure access to dvp; check cache_fplookup_partial_setup for
4836 * reasoning.
4837 */
4838 dvs = vget_prep_smr(dvp);
4839 cache_fpl_smr_exit(fpl);
4840 if (__predict_false(dvs == VGET_NONE)) {
4841 return (cache_fpl_aborted(fpl));
4842 }
4843
4844 vget_finish_ref(dvp, dvs);
4845 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4846 vrele(dvp);
4847 return (cache_fpl_aborted(fpl));
4848 }
4849
4850 error = vn_lock(dvp, LK_SHARED);
4851 if (__predict_false(error != 0)) {
4852 vrele(dvp);
4853 return (cache_fpl_aborted(fpl));
4854 }
4855
4856 tvp = NULL;
4857 /*
4858 * TODO: provide variants which don't require locking either vnode.
4859 */
4860 cnp->cn_flags |= ISLASTCN;
4861 docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
4862 if (docache)
4863 cnp->cn_flags |= MAKEENTRY;
4864 cnp->cn_lkflags = LK_SHARED;
4865 if ((cnp->cn_flags & LOCKSHARED) == 0) {
4866 cnp->cn_lkflags = LK_EXCLUSIVE;
4867 }
4868 error = VOP_LOOKUP(dvp, &tvp, cnp);
4869 switch (error) {
4870 case EJUSTRETURN:
4871 case 0:
4872 break;
4873 case ENOTDIR:
4874 case ENOENT:
4875 vput(dvp);
4876 return (cache_fpl_handled_error(fpl, error));
4877 default:
4878 vput(dvp);
4879 return (cache_fpl_aborted(fpl));
4880 }
4881
4882 fpl->tvp = tvp;
4883 if (!fpl->savename) {
4884 MPASS((cnp->cn_flags & SAVENAME) == 0);
4885 }
4886
4887 if (tvp == NULL) {
4888 MPASS(error == EJUSTRETURN);
4889 if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) {
4890 vput(dvp);
4891 } else if ((cnp->cn_flags & LOCKPARENT) == 0) {
4892 VOP_UNLOCK(dvp);
4893 }
4894 return (cache_fpl_handled(fpl));
4895 }
4896
4897 if (tvp->v_type == VLNK) {
4898 if ((cnp->cn_flags & FOLLOW) != 0) {
4899 vput(dvp);
4900 vput(tvp);
4901 return (cache_fpl_aborted(fpl));
4902 }
4903 }
4904
4905 if (__predict_false(cache_fplookup_is_mp(fpl))) {
4906 vput(dvp);
4907 vput(tvp);
4908 return (cache_fpl_aborted(fpl));
4909 }
4910
4911 if ((cnp->cn_flags & LOCKLEAF) == 0) {
4912 VOP_UNLOCK(tvp);
4913 }
4914
4915 if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) {
4916 vput(dvp);
4917 } else if ((cnp->cn_flags & LOCKPARENT) == 0) {
4918 VOP_UNLOCK(dvp);
4919 }
4920 return (cache_fpl_handled(fpl));
4921 }
4922
4923 static int __noinline
4924 cache_fplookup_dot(struct cache_fpl *fpl)
4925 {
4926 int error;
4927
4928 MPASS(!seqc_in_modify(fpl->dvp_seqc));
4929 /*
4930 * Just re-assign the value. seqc will be checked later for the first
4931 * non-dot path component in line and/or before deciding to return the
4932 * vnode.
4933 */
4934 fpl->tvp = fpl->dvp;
4935 fpl->tvp_seqc = fpl->dvp_seqc;
4936
4937 counter_u64_add(dothits, 1);
4938 SDT_PROBE3(vfs, namecache, lookup, hit, fpl->dvp, ".", fpl->dvp);
4939
4940 error = 0;
4941 if (cache_fplookup_is_mp(fpl)) {
4942 error = cache_fplookup_cross_mount(fpl);
4943 }
4944 return (error);
4945 }
4946
4947 static int __noinline
4948 cache_fplookup_dotdot(struct cache_fpl *fpl)
4949 {
4950 struct nameidata *ndp;
4951 struct componentname *cnp;
4952 struct namecache *ncp;
4953 struct vnode *dvp;
4954 struct prison *pr;
4955 u_char nc_flag;
4956
4957 ndp = fpl->ndp;
4958 cnp = fpl->cnp;
4959 dvp = fpl->dvp;
4960
4961 MPASS(cache_fpl_isdotdot(cnp));
4962
4963 /*
4964 * XXX this is racy the same way regular lookup is
4965 */
4966 for (pr = cnp->cn_cred->cr_prison; pr != NULL;
4967 pr = pr->pr_parent)
4968 if (dvp == pr->pr_root)
4969 break;
4970
4971 if (dvp == ndp->ni_rootdir ||
4972 dvp == ndp->ni_topdir ||
4973 dvp == rootvnode ||
4974 pr != NULL) {
4975 fpl->tvp = dvp;
4976 fpl->tvp_seqc = vn_seqc_read_any(dvp);
4977 if (seqc_in_modify(fpl->tvp_seqc)) {
4978 return (cache_fpl_aborted(fpl));
4979 }
4980 return (0);
4981 }
4982
4983 if ((dvp->v_vflag & VV_ROOT) != 0) {
4984 /*
4985 * TODO
4986 * The opposite of climb mount is needed here.
4987 */
4988 return (cache_fpl_partial(fpl));
4989 }
4990
4991 ncp = atomic_load_consume_ptr(&dvp->v_cache_dd);
4992 if (ncp == NULL) {
4993 return (cache_fpl_aborted(fpl));
4994 }
4995
4996 nc_flag = atomic_load_char(&ncp->nc_flag);
4997 if ((nc_flag & NCF_ISDOTDOT) != 0) {
4998 if ((nc_flag & NCF_NEGATIVE) != 0)
4999 return (cache_fpl_aborted(fpl));
5000 fpl->tvp = ncp->nc_vp;
5001 } else {
5002 fpl->tvp = ncp->nc_dvp;
5003 }
5004
5005 fpl->tvp_seqc = vn_seqc_read_any(fpl->tvp);
5006 if (seqc_in_modify(fpl->tvp_seqc)) {
5007 return (cache_fpl_partial(fpl));
5008 }
5009
5010 /*
5011 * Acquire fence provided by vn_seqc_read_any above.
5012 */
5013 if (__predict_false(atomic_load_ptr(&dvp->v_cache_dd) != ncp)) {
5014 return (cache_fpl_aborted(fpl));
5015 }
5016
5017 if (!cache_ncp_canuse(ncp)) {
5018 return (cache_fpl_aborted(fpl));
5019 }
5020
5021 counter_u64_add(dotdothits, 1);
5022 return (0);
5023 }
5024
5025 static int __noinline
5026 cache_fplookup_neg(struct cache_fpl *fpl, struct namecache *ncp, uint32_t hash)
5027 {
5028 u_char nc_flag;
5029 bool neg_promote;
5030
5031 nc_flag = atomic_load_char(&ncp->nc_flag);
5032 MPASS((nc_flag & NCF_NEGATIVE) != 0);
5033 /*
5034 * If they want to create an entry we need to replace this one.
5035 */
5036 if (__predict_false(fpl->cnp->cn_nameiop != LOOKUP)) {
5037 fpl->tvp = NULL;
5038 return (cache_fplookup_modifying(fpl));
5039 }
5040 neg_promote = cache_neg_hit_prep(ncp);
5041 if (!cache_fpl_neg_ncp_canuse(ncp)) {
5042 cache_neg_hit_abort(ncp);
5043 return (cache_fpl_partial(fpl));
5044 }
5045 if (neg_promote) {
5046 return (cache_fplookup_negative_promote(fpl, ncp, hash));
5047 }
5048 cache_neg_hit_finish(ncp);
5049 cache_fpl_smr_exit(fpl);
5050 return (cache_fpl_handled_error(fpl, ENOENT));
5051 }
5052
5053 /*
5054 * Resolve a symlink. Called by filesystem-specific routines.
5055 *
5056 * Code flow is:
5057 * ... -> cache_fplookup_symlink -> VOP_FPLOOKUP_SYMLINK -> cache_symlink_resolve
5058 */
5059 int
5060 cache_symlink_resolve(struct cache_fpl *fpl, const char *string, size_t len)
5061 {
5062 struct nameidata *ndp;
5063 struct componentname *cnp;
5064 size_t adjust;
5065
5066 ndp = fpl->ndp;
5067 cnp = fpl->cnp;
5068
5069 if (__predict_false(len == 0)) {
5070 return (ENOENT);
5071 }
5072
5073 if (__predict_false(len > MAXPATHLEN - 2)) {
5074 if (cache_fpl_istrailingslash(fpl)) {
5075 return (EAGAIN);
5076 }
5077 }
5078
5079 ndp->ni_pathlen = fpl->nulchar - cnp->cn_nameptr - cnp->cn_namelen + 1;
5080 #ifdef INVARIANTS
5081 if (ndp->ni_pathlen != fpl->debug.ni_pathlen) {
5082 panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n",
5083 __func__, ndp->ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar,
5084 cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf);
5085 }
5086 #endif
5087
5088 if (__predict_false(len + ndp->ni_pathlen > MAXPATHLEN)) {
5089 return (ENAMETOOLONG);
5090 }
5091
5092 if (__predict_false(ndp->ni_loopcnt++ >= MAXSYMLINKS)) {
5093 return (ELOOP);
5094 }
5095
5096 adjust = len;
5097 if (ndp->ni_pathlen > 1) {
5098 bcopy(ndp->ni_next, cnp->cn_pnbuf + len, ndp->ni_pathlen);
5099 } else {
5100 if (cache_fpl_istrailingslash(fpl)) {
5101 adjust = len + 1;
5102 cnp->cn_pnbuf[len] = '/';
5103 cnp->cn_pnbuf[len + 1] = '\0';
5104 } else {
5105 cnp->cn_pnbuf[len] = '\0';
5106 }
5107 }
5108 bcopy(string, cnp->cn_pnbuf, len);
5109
5110 ndp->ni_pathlen += adjust;
5111 cache_fpl_pathlen_add(fpl, adjust);
5112 cnp->cn_nameptr = cnp->cn_pnbuf;
5113 fpl->nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1];
5114 fpl->tvp = NULL;
5115 return (0);
5116 }
5117
5118 static int __noinline
5119 cache_fplookup_symlink(struct cache_fpl *fpl)
5120 {
5121 struct mount *mp;
5122 struct nameidata *ndp;
5123 struct componentname *cnp;
5124 struct vnode *dvp, *tvp;
5125 int error;
5126
5127 ndp = fpl->ndp;
5128 cnp = fpl->cnp;
5129 dvp = fpl->dvp;
5130 tvp = fpl->tvp;
5131
5132 if (cache_fpl_islastcn(ndp)) {
5133 if ((cnp->cn_flags & FOLLOW) == 0) {
5134 return (cache_fplookup_final(fpl));
5135 }
5136 }
5137
5138 mp = atomic_load_ptr(&dvp->v_mount);
5139 if (__predict_false(mp == NULL)) {
5140 return (cache_fpl_aborted(fpl));
5141 }
5142
5143 /*
5144 * Note this check races against setting the flag just like regular
5145 * lookup.
5146 */
5147 if (__predict_false((mp->mnt_flag & MNT_NOSYMFOLLOW) != 0)) {
5148 cache_fpl_smr_exit(fpl);
5149 return (cache_fpl_handled_error(fpl, EACCES));
5150 }
5151
5152 error = VOP_FPLOOKUP_SYMLINK(tvp, fpl);
5153 if (__predict_false(error != 0)) {
5154 switch (error) {
5155 case EAGAIN:
5156 return (cache_fpl_partial(fpl));
5157 case ENOENT:
5158 case ENAMETOOLONG:
5159 case ELOOP:
5160 cache_fpl_smr_exit(fpl);
5161 return (cache_fpl_handled_error(fpl, error));
5162 default:
5163 return (cache_fpl_aborted(fpl));
5164 }
5165 }
5166
5167 if (*(cnp->cn_nameptr) == '/') {
5168 fpl->dvp = cache_fpl_handle_root(fpl);
5169 fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp);
5170 if (seqc_in_modify(fpl->dvp_seqc)) {
5171 return (cache_fpl_aborted(fpl));
5172 }
5173 /*
5174 * The main loop assumes that ->dvp points to a vnode belonging
5175 * to a filesystem which can do lockless lookup, but the absolute
5176 * symlink can be wandering off to one which does not.
5177 */
5178 mp = atomic_load_ptr(&fpl->dvp->v_mount);
5179 if (__predict_false(mp == NULL)) {
5180 return (cache_fpl_aborted(fpl));
5181 }
5182 if (!cache_fplookup_mp_supported(mp)) {
5183 cache_fpl_checkpoint(fpl);
5184 return (cache_fpl_partial(fpl));
5185 }
5186 }
5187 return (0);
5188 }
5189
5190 static int
5191 cache_fplookup_next(struct cache_fpl *fpl)
5192 {
5193 struct componentname *cnp;
5194 struct namecache *ncp;
5195 struct vnode *dvp, *tvp;
5196 u_char nc_flag;
5197 uint32_t hash;
5198 int error;
5199
5200 cnp = fpl->cnp;
5201 dvp = fpl->dvp;
5202 hash = fpl->hash;
5203
5204 if (__predict_false(cnp->cn_nameptr[0] == '.')) {
5205 if (cnp->cn_namelen == 1) {
5206 return (cache_fplookup_dot(fpl));
5207 }
5208 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
5209 return (cache_fplookup_dotdot(fpl));
5210 }
5211 }
5212
5213 MPASS(!cache_fpl_isdotdot(cnp));
5214
5215 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
5216 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
5217 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
5218 break;
5219 }
5220
5221 if (__predict_false(ncp == NULL)) {
5222 return (cache_fplookup_noentry(fpl));
5223 }
5224
5225 tvp = atomic_load_ptr(&ncp->nc_vp);
5226 nc_flag = atomic_load_char(&ncp->nc_flag);
5227 if ((nc_flag & NCF_NEGATIVE) != 0) {
5228 return (cache_fplookup_neg(fpl, ncp, hash));
5229 }
5230
5231 if (!cache_ncp_canuse(ncp)) {
5232 return (cache_fpl_partial(fpl));
5233 }
5234
5235 fpl->tvp = tvp;
5236 fpl->tvp_seqc = vn_seqc_read_any(tvp);
5237 if (seqc_in_modify(fpl->tvp_seqc)) {
5238 return (cache_fpl_partial(fpl));
5239 }
5240
5241 counter_u64_add(numposhits, 1);
5242 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, tvp);
5243
5244 error = 0;
5245 if (cache_fplookup_is_mp(fpl)) {
5246 error = cache_fplookup_cross_mount(fpl);
5247 }
5248 return (error);
5249 }
5250
5251 static bool
5252 cache_fplookup_mp_supported(struct mount *mp)
5253 {
5254
5255 MPASS(mp != NULL);
5256 if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0)
5257 return (false);
5258 return (true);
5259 }
5260
5261 /*
5262 * Walk up the mount stack (if any).
5263 *
5264 * Correctness is provided in the following ways:
5265 * - all vnodes are protected from freeing with SMR
5266 * - struct mount objects are type stable making them always safe to access
5267 * - stability of the particular mount is provided by busying it
5268 * - relationship between the vnode which is mounted on and the mount is
5269 * verified with the vnode sequence counter after busying
5270 * - association between root vnode of the mount and the mount is protected
5271 * by busy
5272 *
5273 * From that point on we can read the sequence counter of the root vnode
5274 * and get the next mount on the stack (if any) using the same protection.
5275 *
5276 * By the end of successful walk we are guaranteed the reached state was
5277 * indeed present at least at some point which matches the regular lookup.
5278 */
5279 static int __noinline
5280 cache_fplookup_climb_mount(struct cache_fpl *fpl)
5281 {
5282 struct mount *mp, *prev_mp;
5283 struct mount_pcpu *mpcpu, *prev_mpcpu;
5284 struct vnode *vp;
5285 seqc_t vp_seqc;
5286
5287 vp = fpl->tvp;
5288 vp_seqc = fpl->tvp_seqc;
5289
5290 VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp);
5291 mp = atomic_load_ptr(&vp->v_mountedhere);
5292 if (__predict_false(mp == NULL)) {
5293 return (0);
5294 }
5295
5296 prev_mp = NULL;
5297 for (;;) {
5298 if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
5299 if (prev_mp != NULL)
5300 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
5301 return (cache_fpl_partial(fpl));
5302 }
5303 if (prev_mp != NULL)
5304 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
5305 if (!vn_seqc_consistent(vp, vp_seqc)) {
5306 vfs_op_thread_exit_crit(mp, mpcpu);
5307 return (cache_fpl_partial(fpl));
5308 }
5309 if (!cache_fplookup_mp_supported(mp)) {
5310 vfs_op_thread_exit_crit(mp, mpcpu);
5311 return (cache_fpl_partial(fpl));
5312 }
5313 vp = atomic_load_ptr(&mp->mnt_rootvnode);
5314 if (vp == NULL) {
5315 vfs_op_thread_exit_crit(mp, mpcpu);
5316 return (cache_fpl_partial(fpl));
5317 }
5318 vp_seqc = vn_seqc_read_any(vp);
5319 if (seqc_in_modify(vp_seqc)) {
5320 vfs_op_thread_exit_crit(mp, mpcpu);
5321 return (cache_fpl_partial(fpl));
5322 }
5323 prev_mp = mp;
5324 prev_mpcpu = mpcpu;
5325 mp = atomic_load_ptr(&vp->v_mountedhere);
5326 if (mp == NULL)
5327 break;
5328 }
5329
5330 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
5331 fpl->tvp = vp;
5332 fpl->tvp_seqc = vp_seqc;
5333 return (0);
5334 }
5335
5336 static int __noinline
5337 cache_fplookup_cross_mount(struct cache_fpl *fpl)
5338 {
5339 struct mount *mp;
5340 struct mount_pcpu *mpcpu;
5341 struct vnode *vp;
5342 seqc_t vp_seqc;
5343
5344 vp = fpl->tvp;
5345 vp_seqc = fpl->tvp_seqc;
5346
5347 VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp);
5348 mp = atomic_load_ptr(&vp->v_mountedhere);
5349 if (__predict_false(mp == NULL)) {
5350 return (0);
5351 }
5352
5353 if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
5354 return (cache_fpl_partial(fpl));
5355 }
5356 if (!vn_seqc_consistent(vp, vp_seqc)) {
5357 vfs_op_thread_exit_crit(mp, mpcpu);
5358 return (cache_fpl_partial(fpl));
5359 }
5360 if (!cache_fplookup_mp_supported(mp)) {
5361 vfs_op_thread_exit_crit(mp, mpcpu);
5362 return (cache_fpl_partial(fpl));
5363 }
5364 vp = atomic_load_ptr(&mp->mnt_rootvnode);
5365 if (__predict_false(vp == NULL)) {
5366 vfs_op_thread_exit_crit(mp, mpcpu);
5367 return (cache_fpl_partial(fpl));
5368 }
5369 vp_seqc = vn_seqc_read_any(vp);
5370 vfs_op_thread_exit_crit(mp, mpcpu);
5371 if (seqc_in_modify(vp_seqc)) {
5372 return (cache_fpl_partial(fpl));
5373 }
5374 mp = atomic_load_ptr(&vp->v_mountedhere);
5375 if (__predict_false(mp != NULL)) {
5376 /*
5377 * There are possibly more mount points on top.
5378 * Normally this does not happen so for simplicity just start
5379 * over.
5380 */
5381 return (cache_fplookup_climb_mount(fpl));
5382 }
5383
5384 fpl->tvp = vp;
5385 fpl->tvp_seqc = vp_seqc;
5386 return (0);
5387 }
5388
5389 /*
5390 * Check if a vnode is mounted on.
5391 */
5392 static bool
5393 cache_fplookup_is_mp(struct cache_fpl *fpl)
5394 {
5395 struct vnode *vp;
5396
5397 vp = fpl->tvp;
5398 return ((vn_irflag_read(vp) & VIRF_MOUNTPOINT) != 0);
5399 }
5400
5401 /*
5402 * Parse the path.
5403 *
5404 * The code was originally copy-pasted from regular lookup and despite
5405 * clean ups leaves performance on the table. Any modifications here
5406 * must take into account that in case off fallback the resulting
5407 * nameidata state has to be compatible with the original.
5408 */
5409
5410 /*
5411 * Debug ni_pathlen tracking.
5412 */
5413 #ifdef INVARIANTS
5414 static void
5415 cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n)
5416 {
5417
5418 fpl->debug.ni_pathlen += n;
5419 KASSERT(fpl->debug.ni_pathlen <= PATH_MAX,
5420 ("%s: pathlen overflow to %zd\n", __func__, fpl->debug.ni_pathlen));
5421 }
5422
5423 static void
5424 cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n)
5425 {
5426
5427 fpl->debug.ni_pathlen -= n;
5428 KASSERT(fpl->debug.ni_pathlen <= PATH_MAX,
5429 ("%s: pathlen underflow to %zd\n", __func__, fpl->debug.ni_pathlen));
5430 }
5431
5432 static void
5433 cache_fpl_pathlen_inc(struct cache_fpl *fpl)
5434 {
5435
5436 cache_fpl_pathlen_add(fpl, 1);
5437 }
5438
5439 static void
5440 cache_fpl_pathlen_dec(struct cache_fpl *fpl)
5441 {
5442
5443 cache_fpl_pathlen_sub(fpl, 1);
5444 }
5445 #else
5446 static void
5447 cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n)
5448 {
5449 }
5450
5451 static void
5452 cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n)
5453 {
5454 }
5455
5456 static void
5457 cache_fpl_pathlen_inc(struct cache_fpl *fpl)
5458 {
5459 }
5460
5461 static void
5462 cache_fpl_pathlen_dec(struct cache_fpl *fpl)
5463 {
5464 }
5465 #endif
5466
5467 static void
5468 cache_fplookup_parse(struct cache_fpl *fpl)
5469 {
5470 struct nameidata *ndp;
5471 struct componentname *cnp;
5472 struct vnode *dvp;
5473 char *cp;
5474 uint32_t hash;
5475
5476 ndp = fpl->ndp;
5477 cnp = fpl->cnp;
5478 dvp = fpl->dvp;
5479
5480 /*
5481 * Find the end of this path component, it is either / or nul.
5482 *
5483 * Store / as a temporary sentinel so that we only have one character
5484 * to test for. Pathnames tend to be short so this should not be
5485 * resulting in cache misses.
5486 *
5487 * TODO: fix this to be word-sized.
5488 */
5489 KASSERT(&cnp->cn_nameptr[fpl->debug.ni_pathlen - 1] == fpl->nulchar,
5490 ("%s: mismatch between pathlen (%zu) and nulchar (%p != %p), string [%s]\n",
5491 __func__, fpl->debug.ni_pathlen, &cnp->cn_nameptr[fpl->debug.ni_pathlen - 1],
5492 fpl->nulchar, cnp->cn_pnbuf));
5493 KASSERT(*fpl->nulchar == '\0',
5494 ("%s: expected nul at %p; string [%s]\n", __func__, fpl->nulchar,
5495 cnp->cn_pnbuf));
5496 hash = cache_get_hash_iter_start(dvp);
5497 *fpl->nulchar = '/';
5498 for (cp = cnp->cn_nameptr; *cp != '/'; cp++) {
5499 KASSERT(*cp != '\0',
5500 ("%s: encountered unexpected nul; string [%s]\n", __func__,
5501 cnp->cn_nameptr));
5502 hash = cache_get_hash_iter(*cp, hash);
5503 continue;
5504 }
5505 *fpl->nulchar = '\0';
5506 fpl->hash = cache_get_hash_iter_finish(hash);
5507
5508 cnp->cn_namelen = cp - cnp->cn_nameptr;
5509 cache_fpl_pathlen_sub(fpl, cnp->cn_namelen);
5510
5511 #ifdef INVARIANTS
5512 /*
5513 * cache_get_hash only accepts lengths up to NAME_MAX. This is fine since
5514 * we are going to fail this lookup with ENAMETOOLONG (see below).
5515 */
5516 if (cnp->cn_namelen <= NAME_MAX) {
5517 if (fpl->hash != cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp)) {
5518 panic("%s: mismatched hash for [%s] len %ld", __func__,
5519 cnp->cn_nameptr, cnp->cn_namelen);
5520 }
5521 }
5522 #endif
5523
5524 /*
5525 * Hack: we have to check if the found path component's length exceeds
5526 * NAME_MAX. However, the condition is very rarely true and check can
5527 * be elided in the common case -- if an entry was found in the cache,
5528 * then it could not have been too long to begin with.
5529 */
5530 ndp->ni_next = cp;
5531 }
5532
5533 static void
5534 cache_fplookup_parse_advance(struct cache_fpl *fpl)
5535 {
5536 struct nameidata *ndp;
5537 struct componentname *cnp;
5538
5539 ndp = fpl->ndp;
5540 cnp = fpl->cnp;
5541
5542 cnp->cn_nameptr = ndp->ni_next;
5543 KASSERT(*(cnp->cn_nameptr) == '/',
5544 ("%s: should have seen slash at %p ; buf %p [%s]\n", __func__,
5545 cnp->cn_nameptr, cnp->cn_pnbuf, cnp->cn_pnbuf));
5546 cnp->cn_nameptr++;
5547 cache_fpl_pathlen_dec(fpl);
5548 }
5549
5550 /*
5551 * Skip spurious slashes in a pathname (e.g., "foo///bar") and retry.
5552 *
5553 * Lockless lookup tries to elide checking for spurious slashes and should they
5554 * be present is guaranteed to fail to find an entry. In this case the caller
5555 * must check if the name starts with a slash and call this routine. It is
5556 * going to fast forward across the spurious slashes and set the state up for
5557 * retry.
5558 */
5559 static int __noinline
5560 cache_fplookup_skip_slashes(struct cache_fpl *fpl)
5561 {
5562 struct nameidata *ndp;
5563 struct componentname *cnp;
5564
5565 ndp = fpl->ndp;
5566 cnp = fpl->cnp;
5567
5568 MPASS(*(cnp->cn_nameptr) == '/');
5569 do {
5570 cnp->cn_nameptr++;
5571 cache_fpl_pathlen_dec(fpl);
5572 } while (*(cnp->cn_nameptr) == '/');
5573
5574 /*
5575 * Go back to one slash so that cache_fplookup_parse_advance has
5576 * something to skip.
5577 */
5578 cnp->cn_nameptr--;
5579 cache_fpl_pathlen_inc(fpl);
5580
5581 /*
5582 * cache_fplookup_parse_advance starts from ndp->ni_next
5583 */
5584 ndp->ni_next = cnp->cn_nameptr;
5585
5586 /*
5587 * See cache_fplookup_dot.
5588 */
5589 fpl->tvp = fpl->dvp;
5590 fpl->tvp_seqc = fpl->dvp_seqc;
5591
5592 return (0);
5593 }
5594
5595 /*
5596 * Handle trailing slashes (e.g., "foo/").
5597 *
5598 * If a trailing slash is found the terminal vnode must be a directory.
5599 * Regular lookup shortens the path by nulifying the first trailing slash and
5600 * sets the TRAILINGSLASH flag to denote this took place. There are several
5601 * checks on it performed later.
5602 *
5603 * Similarly to spurious slashes, lockless lookup handles this in a speculative
5604 * manner relying on an invariant that a non-directory vnode will get a miss.
5605 * In this case cn_nameptr[0] == '\0' and cn_namelen == 0.
5606 *
5607 * Thus for a path like "foo/bar/" the code unwinds the state back to "bar/"
5608 * and denotes this is the last path component, which avoids looping back.
5609 *
5610 * Only plain lookups are supported for now to restrict corner cases to handle.
5611 */
5612 static int __noinline
5613 cache_fplookup_trailingslash(struct cache_fpl *fpl)
5614 {
5615 #ifdef INVARIANTS
5616 size_t ni_pathlen;
5617 #endif
5618 struct nameidata *ndp;
5619 struct componentname *cnp;
5620 struct namecache *ncp;
5621 struct vnode *tvp;
5622 char *cn_nameptr_orig, *cn_nameptr_slash;
5623 seqc_t tvp_seqc;
5624 u_char nc_flag;
5625
5626 ndp = fpl->ndp;
5627 cnp = fpl->cnp;
5628 tvp = fpl->tvp;
5629 tvp_seqc = fpl->tvp_seqc;
5630
5631 MPASS(fpl->dvp == fpl->tvp);
5632 KASSERT(cache_fpl_istrailingslash(fpl),
5633 ("%s: expected trailing slash at %p; string [%s]\n", __func__, fpl->nulchar - 1,
5634 cnp->cn_pnbuf));
5635 KASSERT(cnp->cn_nameptr[0] == '\0',
5636 ("%s: expected nul char at %p; string [%s]\n", __func__, &cnp->cn_nameptr[0],
5637 cnp->cn_pnbuf));
5638 KASSERT(cnp->cn_namelen == 0,
5639 ("%s: namelen 0 but got %ld; string [%s]\n", __func__, cnp->cn_namelen,
5640 cnp->cn_pnbuf));
5641 MPASS(cnp->cn_nameptr > cnp->cn_pnbuf);
5642
5643 if (cnp->cn_nameiop != LOOKUP) {
5644 return (cache_fpl_aborted(fpl));
5645 }
5646
5647 if (__predict_false(tvp->v_type != VDIR)) {
5648 if (!vn_seqc_consistent(tvp, tvp_seqc)) {
5649 return (cache_fpl_aborted(fpl));
5650 }
5651 cache_fpl_smr_exit(fpl);
5652 return (cache_fpl_handled_error(fpl, ENOTDIR));
5653 }
5654
5655 /*
5656 * Denote the last component.
5657 */
5658 ndp->ni_next = &cnp->cn_nameptr[0];
5659 MPASS(cache_fpl_islastcn(ndp));
5660
5661 /*
5662 * Unwind trailing slashes.
5663 */
5664 cn_nameptr_orig = cnp->cn_nameptr;
5665 while (cnp->cn_nameptr >= cnp->cn_pnbuf) {
5666 cnp->cn_nameptr--;
5667 if (cnp->cn_nameptr[0] != '/') {
5668 break;
5669 }
5670 }
5671
5672 /*
5673 * Unwind to the beginning of the path component.
5674 *
5675 * Note the path may or may not have started with a slash.
5676 */
5677 cn_nameptr_slash = cnp->cn_nameptr;
5678 while (cnp->cn_nameptr > cnp->cn_pnbuf) {
5679 cnp->cn_nameptr--;
5680 if (cnp->cn_nameptr[0] == '/') {
5681 break;
5682 }
5683 }
5684 if (cnp->cn_nameptr[0] == '/') {
5685 cnp->cn_nameptr++;
5686 }
5687
5688 cnp->cn_namelen = cn_nameptr_slash - cnp->cn_nameptr + 1;
5689 cache_fpl_pathlen_add(fpl, cn_nameptr_orig - cnp->cn_nameptr);
5690 cache_fpl_checkpoint(fpl);
5691
5692 #ifdef INVARIANTS
5693 ni_pathlen = fpl->nulchar - cnp->cn_nameptr + 1;
5694 if (ni_pathlen != fpl->debug.ni_pathlen) {
5695 panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n",
5696 __func__, ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar,
5697 cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf);
5698 }
5699 #endif
5700
5701 /*
5702 * If this was a "./" lookup the parent directory is already correct.
5703 */
5704 if (cnp->cn_nameptr[0] == '.' && cnp->cn_namelen == 1) {
5705 return (0);
5706 }
5707
5708 /*
5709 * Otherwise we need to look it up.
5710 */
5711 tvp = fpl->tvp;
5712 ncp = atomic_load_consume_ptr(&tvp->v_cache_dd);
5713 if (__predict_false(ncp == NULL)) {
5714 return (cache_fpl_aborted(fpl));
5715 }
5716 nc_flag = atomic_load_char(&ncp->nc_flag);
5717 if ((nc_flag & NCF_ISDOTDOT) != 0) {
5718 return (cache_fpl_aborted(fpl));
5719 }
5720 fpl->dvp = ncp->nc_dvp;
5721 fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp);
5722 if (seqc_in_modify(fpl->dvp_seqc)) {
5723 return (cache_fpl_aborted(fpl));
5724 }
5725 return (0);
5726 }
5727
5728 /*
5729 * See the API contract for VOP_FPLOOKUP_VEXEC.
5730 */
5731 static int __noinline
5732 cache_fplookup_failed_vexec(struct cache_fpl *fpl, int error)
5733 {
5734 struct componentname *cnp;
5735 struct vnode *dvp;
5736 seqc_t dvp_seqc;
5737
5738 cnp = fpl->cnp;
5739 dvp = fpl->dvp;
5740 dvp_seqc = fpl->dvp_seqc;
5741
5742 /*
5743 * TODO: Due to ignoring trailing slashes lookup will perform a
5744 * permission check on the last dir when it should not be doing it. It
5745 * may fail, but said failure should be ignored. It is possible to fix
5746 * it up fully without resorting to regular lookup, but for now just
5747 * abort.
5748 */
5749 if (cache_fpl_istrailingslash(fpl)) {
5750 return (cache_fpl_aborted(fpl));
5751 }
5752
5753 /*
5754 * Hack: delayed degenerate path checking.
5755 */
5756 if (cnp->cn_nameptr[0] == '\0' && fpl->tvp == NULL) {
5757 return (cache_fplookup_degenerate(fpl));
5758 }
5759
5760 /*
5761 * Hack: delayed name len checking.
5762 */
5763 if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
5764 cache_fpl_smr_exit(fpl);
5765 return (cache_fpl_handled_error(fpl, ENAMETOOLONG));
5766 }
5767
5768 /*
5769 * Hack: they may be looking up foo/bar, where foo is not a directory.
5770 * In such a case we need to return ENOTDIR, but we may happen to get
5771 * here with a different error.
5772 */
5773 if (dvp->v_type != VDIR) {
5774 error = ENOTDIR;
5775 }
5776
5777 /*
5778 * Hack: handle O_SEARCH.
5779 *
5780 * Open Group Base Specifications Issue 7, 2018 edition states:
5781 * <quote>
5782 * If the access mode of the open file description associated with the
5783 * file descriptor is not O_SEARCH, the function shall check whether
5784 * directory searches are permitted using the current permissions of
5785 * the directory underlying the file descriptor. If the access mode is
5786 * O_SEARCH, the function shall not perform the check.
5787 * </quote>
5788 *
5789 * Regular lookup tests for the NOEXECCHECK flag for every path
5790 * component to decide whether to do the permission check. However,
5791 * since most lookups never have the flag (and when they do it is only
5792 * present for the first path component), lockless lookup only acts on
5793 * it if there is a permission problem. Here the flag is represented
5794 * with a boolean so that we don't have to clear it on the way out.
5795 *
5796 * For simplicity this always aborts.
5797 * TODO: check if this is the first lookup and ignore the permission
5798 * problem. Note the flag has to survive fallback (if it happens to be
5799 * performed).
5800 */
5801 if (fpl->fsearch) {
5802 return (cache_fpl_aborted(fpl));
5803 }
5804
5805 switch (error) {
5806 case EAGAIN:
5807 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
5808 error = cache_fpl_aborted(fpl);
5809 } else {
5810 cache_fpl_partial(fpl);
5811 }
5812 break;
5813 default:
5814 if (!vn_seqc_consistent(dvp, dvp_seqc)) {
5815 error = cache_fpl_aborted(fpl);
5816 } else {
5817 cache_fpl_smr_exit(fpl);
5818 cache_fpl_handled_error(fpl, error);
5819 }
5820 break;
5821 }
5822 return (error);
5823 }
5824
5825 static int
5826 cache_fplookup_impl(struct vnode *dvp, struct cache_fpl *fpl)
5827 {
5828 struct nameidata *ndp;
5829 struct componentname *cnp;
5830 struct mount *mp;
5831 int error;
5832
5833 ndp = fpl->ndp;
5834 cnp = fpl->cnp;
5835
5836 cache_fpl_checkpoint(fpl);
5837
5838 /*
5839 * The vnode at hand is almost always stable, skip checking for it.
5840 * Worst case this postpones the check towards the end of the iteration
5841 * of the main loop.
5842 */
5843 fpl->dvp = dvp;
5844 fpl->dvp_seqc = vn_seqc_read_notmodify(fpl->dvp);
5845
5846 mp = atomic_load_ptr(&dvp->v_mount);
5847 if (__predict_false(mp == NULL || !cache_fplookup_mp_supported(mp))) {
5848 return (cache_fpl_aborted(fpl));
5849 }
5850
5851 MPASS(fpl->tvp == NULL);
5852
5853 for (;;) {
5854 cache_fplookup_parse(fpl);
5855
5856 error = VOP_FPLOOKUP_VEXEC(fpl->dvp, cnp->cn_cred);
5857 if (__predict_false(error != 0)) {
5858 error = cache_fplookup_failed_vexec(fpl, error);
5859 break;
5860 }
5861
5862 error = cache_fplookup_next(fpl);
5863 if (__predict_false(cache_fpl_terminated(fpl))) {
5864 break;
5865 }
5866
5867 VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp);
5868
5869 if (fpl->tvp->v_type == VLNK) {
5870 error = cache_fplookup_symlink(fpl);
5871 if (cache_fpl_terminated(fpl)) {
5872 break;
5873 }
5874 } else {
5875 if (cache_fpl_islastcn(ndp)) {
5876 error = cache_fplookup_final(fpl);
5877 break;
5878 }
5879
5880 if (!vn_seqc_consistent(fpl->dvp, fpl->dvp_seqc)) {
5881 error = cache_fpl_aborted(fpl);
5882 break;
5883 }
5884
5885 fpl->dvp = fpl->tvp;
5886 fpl->dvp_seqc = fpl->tvp_seqc;
5887 cache_fplookup_parse_advance(fpl);
5888 }
5889
5890 cache_fpl_checkpoint(fpl);
5891 }
5892
5893 return (error);
5894 }
5895
5896 /*
5897 * Fast path lookup protected with SMR and sequence counters.
5898 *
5899 * Note: all VOP_FPLOOKUP_VEXEC routines have a comment referencing this one.
5900 *
5901 * Filesystems can opt in by setting the MNTK_FPLOOKUP flag and meeting criteria
5902 * outlined below.
5903 *
5904 * Traditional vnode lookup conceptually looks like this:
5905 *
5906 * vn_lock(current);
5907 * for (;;) {
5908 * next = find();
5909 * vn_lock(next);
5910 * vn_unlock(current);
5911 * current = next;
5912 * if (last)
5913 * break;
5914 * }
5915 * return (current);
5916 *
5917 * Each jump to the next vnode is safe memory-wise and atomic with respect to
5918 * any modifications thanks to holding respective locks.
5919 *
5920 * The same guarantee can be provided with a combination of safe memory
5921 * reclamation and sequence counters instead. If all operations which affect
5922 * the relationship between the current vnode and the one we are looking for
5923 * also modify the counter, we can verify whether all the conditions held as
5924 * we made the jump. This includes things like permissions, mount points etc.
5925 * Counter modification is provided by enclosing relevant places in
5926 * vn_seqc_write_begin()/end() calls.
5927 *
5928 * Thus this translates to:
5929 *
5930 * vfs_smr_enter();
5931 * dvp_seqc = seqc_read_any(dvp);
5932 * if (seqc_in_modify(dvp_seqc)) // someone is altering the vnode
5933 * abort();
5934 * for (;;) {
5935 * tvp = find();
5936 * tvp_seqc = seqc_read_any(tvp);
5937 * if (seqc_in_modify(tvp_seqc)) // someone is altering the target vnode
5938 * abort();
5939 * if (!seqc_consistent(dvp, dvp_seqc) // someone is altering the vnode
5940 * abort();
5941 * dvp = tvp; // we know nothing of importance has changed
5942 * dvp_seqc = tvp_seqc; // store the counter for the tvp iteration
5943 * if (last)
5944 * break;
5945 * }
5946 * vget(); // secure the vnode
5947 * if (!seqc_consistent(tvp, tvp_seqc) // final check
5948 * abort();
5949 * // at this point we know nothing has changed for any parent<->child pair
5950 * // as they were crossed during the lookup, meaning we matched the guarantee
5951 * // of the locked variant
5952 * return (tvp);
5953 *
5954 * The API contract for VOP_FPLOOKUP_VEXEC routines is as follows:
5955 * - they are called while within vfs_smr protection which they must never exit
5956 * - EAGAIN can be returned to denote checking could not be performed, it is
5957 * always valid to return it
5958 * - if the sequence counter has not changed the result must be valid
5959 * - if the sequence counter has changed both false positives and false negatives
5960 * are permitted (since the result will be rejected later)
5961 * - for simple cases of unix permission checks vaccess_vexec_smr can be used
5962 *
5963 * Caveats to watch out for:
5964 * - vnodes are passed unlocked and unreferenced with nothing stopping
5965 * VOP_RECLAIM, in turn meaning that ->v_data can become NULL. It is advised
5966 * to use atomic_load_ptr to fetch it.
5967 * - the aforementioned object can also get freed, meaning absent other means it
5968 * should be protected with vfs_smr
5969 * - either safely checking permissions as they are modified or guaranteeing
5970 * their stability is left to the routine
5971 */
5972 int
5973 cache_fplookup(struct nameidata *ndp, enum cache_fpl_status *status,
5974 struct pwd **pwdp)
5975 {
5976 struct cache_fpl fpl;
5977 struct pwd *pwd;
5978 struct vnode *dvp;
5979 struct componentname *cnp;
5980 int error;
5981
5982 fpl.status = CACHE_FPL_STATUS_UNSET;
5983 fpl.in_smr = false;
5984 fpl.ndp = ndp;
5985 fpl.cnp = cnp = &ndp->ni_cnd;
5986 MPASS(ndp->ni_lcf == 0);
5987 MPASS(curthread == cnp->cn_thread);
5988 KASSERT ((cnp->cn_flags & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
5989 ("%s: internal flags found in cn_flags %" PRIx64, __func__,
5990 cnp->cn_flags));
5991 if ((cnp->cn_flags & SAVESTART) != 0) {
5992 MPASS(cnp->cn_nameiop != LOOKUP);
5993 }
5994 MPASS(cnp->cn_nameptr == cnp->cn_pnbuf);
5995
5996 if (__predict_false(!cache_can_fplookup(&fpl))) {
5997 *status = fpl.status;
5998 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
5999 return (EOPNOTSUPP);
6000 }
6001
6002 cache_fpl_checkpoint_outer(&fpl);
6003
6004 cache_fpl_smr_enter_initial(&fpl);
6005 #ifdef INVARIANTS
6006 fpl.debug.ni_pathlen = ndp->ni_pathlen;
6007 #endif
6008 fpl.nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1];
6009 fpl.fsearch = false;
6010 fpl.savename = (cnp->cn_flags & SAVENAME) != 0;
6011 fpl.tvp = NULL; /* for degenerate path handling */
6012 fpl.pwd = pwdp;
6013 pwd = pwd_get_smr();
6014 *(fpl.pwd) = pwd;
6015 ndp->ni_rootdir = pwd->pwd_rdir;
6016 ndp->ni_topdir = pwd->pwd_jdir;
6017
6018 if (cnp->cn_pnbuf[0] == '/') {
6019 dvp = cache_fpl_handle_root(&fpl);
6020 MPASS(ndp->ni_resflags == 0);
6021 ndp->ni_resflags = NIRES_ABS;
6022 } else {
6023 if (ndp->ni_dirfd == AT_FDCWD) {
6024 dvp = pwd->pwd_cdir;
6025 } else {
6026 error = cache_fplookup_dirfd(&fpl, &dvp);
6027 if (__predict_false(error != 0)) {
6028 goto out;
6029 }
6030 }
6031 }
6032
6033 SDT_PROBE4(vfs, namei, lookup, entry, dvp, cnp->cn_pnbuf, cnp->cn_flags, true);
6034 error = cache_fplookup_impl(dvp, &fpl);
6035 out:
6036 cache_fpl_smr_assert_not_entered(&fpl);
6037 cache_fpl_assert_status(&fpl);
6038 *status = fpl.status;
6039 if (SDT_PROBES_ENABLED()) {
6040 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
6041 if (fpl.status == CACHE_FPL_STATUS_HANDLED)
6042 SDT_PROBE4(vfs, namei, lookup, return, error, ndp->ni_vp, true,
6043 ndp);
6044 }
6045
6046 if (__predict_true(fpl.status == CACHE_FPL_STATUS_HANDLED)) {
6047 MPASS(error != CACHE_FPL_FAILED);
6048 if (error != 0) {
6049 MPASS(fpl.dvp == NULL);
6050 MPASS(fpl.tvp == NULL);
6051 MPASS(fpl.savename == false);
6052 }
6053 ndp->ni_dvp = fpl.dvp;
6054 ndp->ni_vp = fpl.tvp;
6055 if (fpl.savename) {
6056 cnp->cn_flags |= HASBUF;
6057 } else {
6058 cache_fpl_cleanup_cnp(cnp);
6059 }
6060 }
6061 return (error);
6062 }
Cache object: e4f584c9061261f7659c150f567d6ea0
|