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 static char __read_frequently cache_fast_lookup_enabled = true; 257 258 /* 259 * This structure describes the elements in the cache of recent 260 * names looked up by namei. 261 */ 262 struct negstate { 263 u_char neg_flag; 264 u_char neg_hit; 265 }; 266 _Static_assert(sizeof(struct negstate) <= sizeof(struct vnode *), 267 "the state must fit in a union with a pointer without growing it"); 268 269 struct namecache { 270 LIST_ENTRY(namecache) nc_src; /* source vnode list */ 271 TAILQ_ENTRY(namecache) nc_dst; /* destination vnode list */ 272 CK_SLIST_ENTRY(namecache) nc_hash;/* hash chain */ 273 struct vnode *nc_dvp; /* vnode of parent of name */ 274 union { 275 struct vnode *nu_vp; /* vnode the name refers to */ 276 struct negstate nu_neg;/* negative entry state */ 277 } n_un; 278 u_char nc_flag; /* flag bits */ 279 u_char nc_nlen; /* length of name */ 280 char nc_name[]; /* segment name + nul */ 281 }; 282 283 /* 284 * struct namecache_ts repeats struct namecache layout up to the 285 * nc_nlen member. 286 * struct namecache_ts is used in place of struct namecache when time(s) need 287 * to be stored. The nc_dotdottime field is used when a cache entry is mapping 288 * both a non-dotdot directory name plus dotdot for the directory's 289 * parent. 290 * 291 * See below for alignment requirement. 292 */ 293 struct namecache_ts { 294 struct timespec nc_time; /* timespec provided by fs */ 295 struct timespec nc_dotdottime; /* dotdot timespec provided by fs */ 296 int nc_ticks; /* ticks value when entry was added */ 297 int nc_pad; 298 struct namecache nc_nc; 299 }; 300 301 TAILQ_HEAD(cache_freebatch, namecache); 302 303 /* 304 * At least mips n32 performs 64-bit accesses to timespec as found 305 * in namecache_ts and requires them to be aligned. Since others 306 * may be in the same spot suffer a little bit and enforce the 307 * alignment for everyone. Note this is a nop for 64-bit platforms. 308 */ 309 #define CACHE_ZONE_ALIGNMENT UMA_ALIGNOF(time_t) 310 311 /* 312 * TODO: the initial value of CACHE_PATH_CUTOFF was inherited from the 313 * 4.4 BSD codebase. Later on struct namecache was tweaked to become 314 * smaller and the value was bumped to retain the total size, but it 315 * was never re-evaluated for suitability. A simple test counting 316 * lengths during package building shows that the value of 45 covers 317 * about 86% of all added entries, reaching 99% at 65. 318 * 319 * Regardless of the above, use of dedicated zones instead of malloc may be 320 * inducing additional waste. This may be hard to address as said zones are 321 * tied to VFS SMR. Even if retaining them, the current split should be 322 * re-evaluated. 323 */ 324 #ifdef __LP64__ 325 #define CACHE_PATH_CUTOFF 45 326 #define CACHE_LARGE_PAD 6 327 #else 328 #define CACHE_PATH_CUTOFF 41 329 #define CACHE_LARGE_PAD 2 330 #endif 331 332 #define CACHE_ZONE_SMALL_SIZE (offsetof(struct namecache, nc_name) + CACHE_PATH_CUTOFF + 1) 333 #define CACHE_ZONE_SMALL_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_SMALL_SIZE) 334 #define CACHE_ZONE_LARGE_SIZE (offsetof(struct namecache, nc_name) + NAME_MAX + 1 + CACHE_LARGE_PAD) 335 #define CACHE_ZONE_LARGE_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_LARGE_SIZE) 336 337 _Static_assert((CACHE_ZONE_SMALL_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size"); 338 _Static_assert((CACHE_ZONE_SMALL_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size"); 339 _Static_assert((CACHE_ZONE_LARGE_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size"); 340 _Static_assert((CACHE_ZONE_LARGE_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size"); 341 342 #define nc_vp n_un.nu_vp 343 #define nc_neg n_un.nu_neg 344 345 /* 346 * Flags in namecache.nc_flag 347 */ 348 #define NCF_WHITE 0x01 349 #define NCF_ISDOTDOT 0x02 350 #define NCF_TS 0x04 351 #define NCF_DTS 0x08 352 #define NCF_DVDROP 0x10 353 #define NCF_NEGATIVE 0x20 354 #define NCF_INVALID 0x40 355 #define NCF_WIP 0x80 356 357 /* 358 * Flags in negstate.neg_flag 359 */ 360 #define NEG_HOT 0x01 361 362 static bool cache_neg_evict_cond(u_long lnumcache); 363 364 /* 365 * Mark an entry as invalid. 366 * 367 * This is called before it starts getting deconstructed. 368 */ 369 static void 370 cache_ncp_invalidate(struct namecache *ncp) 371 { 372 373 KASSERT((ncp->nc_flag & NCF_INVALID) == 0, 374 ("%s: entry %p already invalid", __func__, ncp)); 375 atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_INVALID); 376 atomic_thread_fence_rel(); 377 } 378 379 /* 380 * Check whether the entry can be safely used. 381 * 382 * All places which elide locks are supposed to call this after they are 383 * done with reading from an entry. 384 */ 385 #define cache_ncp_canuse(ncp) ({ \ 386 struct namecache *_ncp = (ncp); \ 387 u_char _nc_flag; \ 388 \ 389 atomic_thread_fence_acq(); \ 390 _nc_flag = atomic_load_char(&_ncp->nc_flag); \ 391 __predict_true((_nc_flag & (NCF_INVALID | NCF_WIP)) == 0); \ 392 }) 393 394 /* 395 * Like the above but also checks NCF_WHITE. 396 */ 397 #define cache_fpl_neg_ncp_canuse(ncp) ({ \ 398 struct namecache *_ncp = (ncp); \ 399 u_char _nc_flag; \ 400 \ 401 atomic_thread_fence_acq(); \ 402 _nc_flag = atomic_load_char(&_ncp->nc_flag); \ 403 __predict_true((_nc_flag & (NCF_INVALID | NCF_WIP | NCF_WHITE)) == 0); \ 404 }) 405 406 VFS_SMR_DECLARE; 407 408 static SYSCTL_NODE(_vfs_cache, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 409 "Name cache parameters"); 410 411 static u_int __read_mostly ncsize; /* the size as computed on creation or resizing */ 412 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, size, CTLFLAG_RW, &ncsize, 0, 413 "Total namecache capacity"); 414 415 u_int ncsizefactor = 2; 416 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, sizefactor, CTLFLAG_RW, &ncsizefactor, 0, 417 "Size factor for namecache"); 418 419 static u_long __read_mostly ncnegfactor = 5; /* ratio of negative entries */ 420 SYSCTL_ULONG(_vfs_cache_param, OID_AUTO, negfactor, CTLFLAG_RW, &ncnegfactor, 0, 421 "Ratio of negative namecache entries"); 422 423 /* 424 * Negative entry % of namecache capacity above which automatic eviction is allowed. 425 * 426 * Check cache_neg_evict_cond for details. 427 */ 428 static u_int ncnegminpct = 3; 429 430 static u_int __read_mostly neg_min; /* the above recomputed against ncsize */ 431 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, negmin, CTLFLAG_RD, &neg_min, 0, 432 "Negative entry count above which automatic eviction is allowed"); 433 434 /* 435 * Structures associated with name caching. 436 */ 437 #define NCHHASH(hash) \ 438 (&nchashtbl[(hash) & nchash]) 439 static __read_mostly CK_SLIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */ 440 static u_long __read_mostly nchash; /* size of hash table */ 441 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, 442 "Size of namecache hash table"); 443 static u_long __exclusive_cache_line numneg; /* number of negative entries allocated */ 444 static u_long __exclusive_cache_line numcache;/* number of cache entries allocated */ 445 446 struct nchstats nchstats; /* cache effectiveness statistics */ 447 448 static u_int __exclusive_cache_line neg_cycle; 449 450 #define ncneghash 3 451 #define numneglists (ncneghash + 1) 452 453 struct neglist { 454 struct mtx nl_evict_lock; 455 struct mtx nl_lock __aligned(CACHE_LINE_SIZE); 456 TAILQ_HEAD(, namecache) nl_list; 457 TAILQ_HEAD(, namecache) nl_hotlist; 458 u_long nl_hotnum; 459 } __aligned(CACHE_LINE_SIZE); 460 461 static struct neglist neglists[numneglists]; 462 463 static inline struct neglist * 464 NCP2NEGLIST(struct namecache *ncp) 465 { 466 467 return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]); 468 } 469 470 static inline struct negstate * 471 NCP2NEGSTATE(struct namecache *ncp) 472 { 473 474 MPASS(atomic_load_char(&ncp->nc_flag) & NCF_NEGATIVE); 475 return (&ncp->nc_neg); 476 } 477 478 #define numbucketlocks (ncbuckethash + 1) 479 static u_int __read_mostly ncbuckethash; 480 static struct mtx_padalign __read_mostly *bucketlocks; 481 #define HASH2BUCKETLOCK(hash) \ 482 ((struct mtx *)(&bucketlocks[((hash) & ncbuckethash)])) 483 484 #define numvnodelocks (ncvnodehash + 1) 485 static u_int __read_mostly ncvnodehash; 486 static struct mtx __read_mostly *vnodelocks; 487 static inline struct mtx * 488 VP2VNODELOCK(struct vnode *vp) 489 { 490 491 return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]); 492 } 493 494 static void 495 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp) 496 { 497 struct namecache_ts *ncp_ts; 498 499 KASSERT((ncp->nc_flag & NCF_TS) != 0 || 500 (tsp == NULL && ticksp == NULL), 501 ("No NCF_TS")); 502 503 if (tsp == NULL) 504 return; 505 506 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc); 507 *tsp = ncp_ts->nc_time; 508 *ticksp = ncp_ts->nc_ticks; 509 } 510 511 #ifdef DEBUG_CACHE 512 static int __read_mostly doingcache = 1; /* 1 => enable the cache */ 513 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0, 514 "VFS namecache enabled"); 515 #endif 516 517 /* Export size information to userland */ 518 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR, 519 sizeof(struct namecache), "sizeof(struct namecache)"); 520 521 /* 522 * The new name cache statistics 523 */ 524 static SYSCTL_NODE(_vfs_cache, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 525 "Name cache statistics"); 526 527 #define STATNODE_ULONG(name, varname, descr) \ 528 SYSCTL_ULONG(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr); 529 #define STATNODE_COUNTER(name, varname, descr) \ 530 static COUNTER_U64_DEFINE_EARLY(varname); \ 531 SYSCTL_COUNTER_U64(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, \ 532 descr); 533 STATNODE_ULONG(neg, numneg, "Number of negative cache entries"); 534 STATNODE_ULONG(count, numcache, "Number of cache entries"); 535 STATNODE_COUNTER(heldvnodes, numcachehv, "Number of namecache entries with vnodes held"); 536 STATNODE_COUNTER(drops, numdrops, "Number of dropped entries due to reaching the limit"); 537 STATNODE_COUNTER(dothits, dothits, "Number of '.' hits"); 538 STATNODE_COUNTER(dotdothis, dotdothits, "Number of '..' hits"); 539 STATNODE_COUNTER(miss, nummiss, "Number of cache misses"); 540 STATNODE_COUNTER(misszap, nummisszap, "Number of cache misses we do not want to cache"); 541 STATNODE_COUNTER(posszaps, numposzaps, 542 "Number of cache hits (positive) we do not want to cache"); 543 STATNODE_COUNTER(poshits, numposhits, "Number of cache hits (positive)"); 544 STATNODE_COUNTER(negzaps, numnegzaps, 545 "Number of cache hits (negative) we do not want to cache"); 546 STATNODE_COUNTER(neghits, numneghits, "Number of cache hits (negative)"); 547 /* These count for vn_getcwd(), too. */ 548 STATNODE_COUNTER(fullpathcalls, numfullpathcalls, "Number of fullpath search calls"); 549 STATNODE_COUNTER(fullpathfail1, numfullpathfail1, "Number of fullpath search errors (ENOTDIR)"); 550 STATNODE_COUNTER(fullpathfail2, numfullpathfail2, 551 "Number of fullpath search errors (VOP_VPTOCNP failures)"); 552 STATNODE_COUNTER(fullpathfail4, numfullpathfail4, "Number of fullpath search errors (ENOMEM)"); 553 STATNODE_COUNTER(fullpathfound, numfullpathfound, "Number of successful fullpath calls"); 554 STATNODE_COUNTER(symlinktoobig, symlinktoobig, "Number of times symlink did not fit the cache"); 555 556 /* 557 * Debug or developer statistics. 558 */ 559 static SYSCTL_NODE(_vfs_cache, OID_AUTO, debug, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 560 "Name cache debugging"); 561 #define DEBUGNODE_ULONG(name, varname, descr) \ 562 SYSCTL_ULONG(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr); 563 #define DEBUGNODE_COUNTER(name, varname, descr) \ 564 static COUNTER_U64_DEFINE_EARLY(varname); \ 565 SYSCTL_COUNTER_U64(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, \ 566 descr); 567 DEBUGNODE_COUNTER(zap_bucket_relock_success, zap_bucket_relock_success, 568 "Number of successful removals after relocking"); 569 static long zap_bucket_fail; 570 DEBUGNODE_ULONG(zap_bucket_fail, zap_bucket_fail, ""); 571 static long zap_bucket_fail2; 572 DEBUGNODE_ULONG(zap_bucket_fail2, zap_bucket_fail2, ""); 573 static long cache_lock_vnodes_cel_3_failures; 574 DEBUGNODE_ULONG(vnodes_cel_3_failures, cache_lock_vnodes_cel_3_failures, 575 "Number of times 3-way vnode locking failed"); 576 577 static void cache_zap_locked(struct namecache *ncp); 578 static int vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf, 579 char **retbuf, size_t *buflen, size_t addend); 580 static int vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, 581 char **retbuf, size_t *buflen); 582 static int vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, 583 char **retbuf, size_t *len, size_t addend); 584 585 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries"); 586 587 static inline void 588 cache_assert_vlp_locked(struct mtx *vlp) 589 { 590 591 if (vlp != NULL) 592 mtx_assert(vlp, MA_OWNED); 593 } 594 595 static inline void 596 cache_assert_vnode_locked(struct vnode *vp) 597 { 598 struct mtx *vlp; 599 600 vlp = VP2VNODELOCK(vp); 601 cache_assert_vlp_locked(vlp); 602 } 603 604 /* 605 * Directory vnodes with entries are held for two reasons: 606 * 1. make them less of a target for reclamation in vnlru 607 * 2. suffer smaller performance penalty in locked lookup as requeieing is avoided 608 * 609 * It will be feasible to stop doing it altogether if all filesystems start 610 * supporting lockless lookup. 611 */ 612 static void 613 cache_hold_vnode(struct vnode *vp) 614 { 615 616 cache_assert_vnode_locked(vp); 617 VNPASS(LIST_EMPTY(&vp->v_cache_src), vp); 618 vhold(vp); 619 counter_u64_add(numcachehv, 1); 620 } 621 622 static void 623 cache_drop_vnode(struct vnode *vp) 624 { 625 626 /* 627 * Called after all locks are dropped, meaning we can't assert 628 * on the state of v_cache_src. 629 */ 630 vdrop(vp); 631 counter_u64_add(numcachehv, -1); 632 } 633 634 /* 635 * UMA zones. 636 */ 637 static uma_zone_t __read_mostly cache_zone_small; 638 static uma_zone_t __read_mostly cache_zone_small_ts; 639 static uma_zone_t __read_mostly cache_zone_large; 640 static uma_zone_t __read_mostly cache_zone_large_ts; 641 642 char * 643 cache_symlink_alloc(size_t size, int flags) 644 { 645 646 if (size < CACHE_ZONE_SMALL_SIZE) { 647 return (uma_zalloc_smr(cache_zone_small, flags)); 648 } 649 if (size < CACHE_ZONE_LARGE_SIZE) { 650 return (uma_zalloc_smr(cache_zone_large, flags)); 651 } 652 counter_u64_add(symlinktoobig, 1); 653 SDT_PROBE1(vfs, namecache, symlink, alloc__fail, size); 654 return (NULL); 655 } 656 657 void 658 cache_symlink_free(char *string, size_t size) 659 { 660 661 MPASS(string != NULL); 662 KASSERT(size < CACHE_ZONE_LARGE_SIZE, 663 ("%s: size %zu too big", __func__, size)); 664 665 if (size < CACHE_ZONE_SMALL_SIZE) { 666 uma_zfree_smr(cache_zone_small, string); 667 return; 668 } 669 if (size < CACHE_ZONE_LARGE_SIZE) { 670 uma_zfree_smr(cache_zone_large, string); 671 return; 672 } 673 __assert_unreachable(); 674 } 675 676 static struct namecache * 677 cache_alloc_uma(int len, bool ts) 678 { 679 struct namecache_ts *ncp_ts; 680 struct namecache *ncp; 681 682 if (__predict_false(ts)) { 683 if (len <= CACHE_PATH_CUTOFF) 684 ncp_ts = uma_zalloc_smr(cache_zone_small_ts, M_WAITOK); 685 else 686 ncp_ts = uma_zalloc_smr(cache_zone_large_ts, M_WAITOK); 687 ncp = &ncp_ts->nc_nc; 688 } else { 689 if (len <= CACHE_PATH_CUTOFF) 690 ncp = uma_zalloc_smr(cache_zone_small, M_WAITOK); 691 else 692 ncp = uma_zalloc_smr(cache_zone_large, M_WAITOK); 693 } 694 return (ncp); 695 } 696 697 static void 698 cache_free_uma(struct namecache *ncp) 699 { 700 struct namecache_ts *ncp_ts; 701 702 if (__predict_false(ncp->nc_flag & NCF_TS)) { 703 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc); 704 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF) 705 uma_zfree_smr(cache_zone_small_ts, ncp_ts); 706 else 707 uma_zfree_smr(cache_zone_large_ts, ncp_ts); 708 } else { 709 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF) 710 uma_zfree_smr(cache_zone_small, ncp); 711 else 712 uma_zfree_smr(cache_zone_large, ncp); 713 } 714 } 715 716 static struct namecache * 717 cache_alloc(int len, bool ts) 718 { 719 u_long lnumcache; 720 721 /* 722 * Avoid blowout in namecache entries. 723 * 724 * Bugs: 725 * 1. filesystems may end up trying to add an already existing entry 726 * (for example this can happen after a cache miss during concurrent 727 * lookup), in which case we will call cache_neg_evict despite not 728 * adding anything. 729 * 2. the routine may fail to free anything and no provisions are made 730 * to make it try harder (see the inside for failure modes) 731 * 3. it only ever looks at negative entries. 732 */ 733 lnumcache = atomic_fetchadd_long(&numcache, 1) + 1; 734 if (cache_neg_evict_cond(lnumcache)) { 735 lnumcache = atomic_load_long(&numcache); 736 } 737 if (__predict_false(lnumcache >= ncsize)) { 738 atomic_subtract_long(&numcache, 1); 739 counter_u64_add(numdrops, 1); 740 return (NULL); 741 } 742 return (cache_alloc_uma(len, ts)); 743 } 744 745 static void 746 cache_free(struct namecache *ncp) 747 { 748 749 MPASS(ncp != NULL); 750 if ((ncp->nc_flag & NCF_DVDROP) != 0) { 751 cache_drop_vnode(ncp->nc_dvp); 752 } 753 cache_free_uma(ncp); 754 atomic_subtract_long(&numcache, 1); 755 } 756 757 static void 758 cache_free_batch(struct cache_freebatch *batch) 759 { 760 struct namecache *ncp, *nnp; 761 int i; 762 763 i = 0; 764 if (TAILQ_EMPTY(batch)) 765 goto out; 766 TAILQ_FOREACH_SAFE(ncp, batch, nc_dst, nnp) { 767 if ((ncp->nc_flag & NCF_DVDROP) != 0) { 768 cache_drop_vnode(ncp->nc_dvp); 769 } 770 cache_free_uma(ncp); 771 i++; 772 } 773 atomic_subtract_long(&numcache, i); 774 out: 775 SDT_PROBE1(vfs, namecache, purge, batch, i); 776 } 777 778 /* 779 * Hashing. 780 * 781 * The code was made to use FNV in 2001 and this choice needs to be revisited. 782 * 783 * Short summary of the difficulty: 784 * The longest name which can be inserted is NAME_MAX characters in length (or 785 * 255 at the time of writing this comment), while majority of names used in 786 * practice are significantly shorter (mostly below 10). More importantly 787 * majority of lookups performed find names are even shorter than that. 788 * 789 * This poses a problem where hashes which do better than FNV past word size 790 * (or so) tend to come with additional overhead when finalizing the result, 791 * making them noticeably slower for the most commonly used range. 792 * 793 * Consider a path like: /usr/obj/usr/src/sys/amd64/GENERIC/vnode_if.c 794 * 795 * When looking it up the most time consuming part by a large margin (at least 796 * on amd64) is hashing. Replacing FNV with something which pessimizes short 797 * input would make the slowest part stand out even more. 798 */ 799 800 /* 801 * TODO: With the value stored we can do better than computing the hash based 802 * on the address. 803 */ 804 static void 805 cache_prehash(struct vnode *vp) 806 { 807 808 vp->v_nchash = fnv_32_buf(&vp, sizeof(vp), FNV1_32_INIT); 809 } 810 811 static uint32_t 812 cache_get_hash(char *name, u_char len, struct vnode *dvp) 813 { 814 815 return (fnv_32_buf(name, len, dvp->v_nchash)); 816 } 817 818 static uint32_t 819 cache_get_hash_iter_start(struct vnode *dvp) 820 { 821 822 return (dvp->v_nchash); 823 } 824 825 static uint32_t 826 cache_get_hash_iter(char c, uint32_t hash) 827 { 828 829 return (fnv_32_buf(&c, 1, hash)); 830 } 831 832 static uint32_t 833 cache_get_hash_iter_finish(uint32_t hash) 834 { 835 836 return (hash); 837 } 838 839 static inline struct nchashhead * 840 NCP2BUCKET(struct namecache *ncp) 841 { 842 uint32_t hash; 843 844 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp); 845 return (NCHHASH(hash)); 846 } 847 848 static inline struct mtx * 849 NCP2BUCKETLOCK(struct namecache *ncp) 850 { 851 uint32_t hash; 852 853 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp); 854 return (HASH2BUCKETLOCK(hash)); 855 } 856 857 #ifdef INVARIANTS 858 static void 859 cache_assert_bucket_locked(struct namecache *ncp) 860 { 861 struct mtx *blp; 862 863 blp = NCP2BUCKETLOCK(ncp); 864 mtx_assert(blp, MA_OWNED); 865 } 866 867 static void 868 cache_assert_bucket_unlocked(struct namecache *ncp) 869 { 870 struct mtx *blp; 871 872 blp = NCP2BUCKETLOCK(ncp); 873 mtx_assert(blp, MA_NOTOWNED); 874 } 875 #else 876 #define cache_assert_bucket_locked(x) do { } while (0) 877 #define cache_assert_bucket_unlocked(x) do { } while (0) 878 #endif 879 880 #define cache_sort_vnodes(x, y) _cache_sort_vnodes((void **)(x), (void **)(y)) 881 static void 882 _cache_sort_vnodes(void **p1, void **p2) 883 { 884 void *tmp; 885 886 MPASS(*p1 != NULL || *p2 != NULL); 887 888 if (*p1 > *p2) { 889 tmp = *p2; 890 *p2 = *p1; 891 *p1 = tmp; 892 } 893 } 894 895 static void 896 cache_lock_all_buckets(void) 897 { 898 u_int i; 899 900 for (i = 0; i < numbucketlocks; i++) 901 mtx_lock(&bucketlocks[i]); 902 } 903 904 static void 905 cache_unlock_all_buckets(void) 906 { 907 u_int i; 908 909 for (i = 0; i < numbucketlocks; i++) 910 mtx_unlock(&bucketlocks[i]); 911 } 912 913 static void 914 cache_lock_all_vnodes(void) 915 { 916 u_int i; 917 918 for (i = 0; i < numvnodelocks; i++) 919 mtx_lock(&vnodelocks[i]); 920 } 921 922 static void 923 cache_unlock_all_vnodes(void) 924 { 925 u_int i; 926 927 for (i = 0; i < numvnodelocks; i++) 928 mtx_unlock(&vnodelocks[i]); 929 } 930 931 static int 932 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2) 933 { 934 935 cache_sort_vnodes(&vlp1, &vlp2); 936 937 if (vlp1 != NULL) { 938 if (!mtx_trylock(vlp1)) 939 return (EAGAIN); 940 } 941 if (!mtx_trylock(vlp2)) { 942 if (vlp1 != NULL) 943 mtx_unlock(vlp1); 944 return (EAGAIN); 945 } 946 947 return (0); 948 } 949 950 static void 951 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2) 952 { 953 954 MPASS(vlp1 != NULL || vlp2 != NULL); 955 MPASS(vlp1 <= vlp2); 956 957 if (vlp1 != NULL) 958 mtx_lock(vlp1); 959 if (vlp2 != NULL) 960 mtx_lock(vlp2); 961 } 962 963 static void 964 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2) 965 { 966 967 MPASS(vlp1 != NULL || vlp2 != NULL); 968 969 if (vlp1 != NULL) 970 mtx_unlock(vlp1); 971 if (vlp2 != NULL) 972 mtx_unlock(vlp2); 973 } 974 975 static int 976 sysctl_nchstats(SYSCTL_HANDLER_ARGS) 977 { 978 struct nchstats snap; 979 980 if (req->oldptr == NULL) 981 return (SYSCTL_OUT(req, 0, sizeof(snap))); 982 983 snap = nchstats; 984 snap.ncs_goodhits = counter_u64_fetch(numposhits); 985 snap.ncs_neghits = counter_u64_fetch(numneghits); 986 snap.ncs_badhits = counter_u64_fetch(numposzaps) + 987 counter_u64_fetch(numnegzaps); 988 snap.ncs_miss = counter_u64_fetch(nummisszap) + 989 counter_u64_fetch(nummiss); 990 991 return (SYSCTL_OUT(req, &snap, sizeof(snap))); 992 } 993 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD | 994 CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU", 995 "VFS cache effectiveness statistics"); 996 997 static void 998 cache_recalc_neg_min(u_int val) 999 { 1000 1001 neg_min = (ncsize * val) / 100; 1002 } 1003 1004 static int 1005 sysctl_negminpct(SYSCTL_HANDLER_ARGS) 1006 { 1007 u_int val; 1008 int error; 1009 1010 val = ncnegminpct; 1011 error = sysctl_handle_int(oidp, &val, 0, req); 1012 if (error != 0 || req->newptr == NULL) 1013 return (error); 1014 1015 if (val == ncnegminpct) 1016 return (0); 1017 if (val < 0 || val > 99) 1018 return (EINVAL); 1019 ncnegminpct = val; 1020 cache_recalc_neg_min(val); 1021 return (0); 1022 } 1023 1024 SYSCTL_PROC(_vfs_cache_param, OID_AUTO, negminpct, 1025 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_negminpct, 1026 "I", "Negative entry \% of namecache capacity above which automatic eviction is allowed"); 1027 1028 #ifdef DEBUG_CACHE 1029 /* 1030 * Grab an atomic snapshot of the name cache hash chain lengths 1031 */ 1032 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, 1033 CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 1034 "hash table stats"); 1035 1036 static int 1037 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS) 1038 { 1039 struct nchashhead *ncpp; 1040 struct namecache *ncp; 1041 int i, error, n_nchash, *cntbuf; 1042 1043 retry: 1044 n_nchash = nchash + 1; /* nchash is max index, not count */ 1045 if (req->oldptr == NULL) 1046 return SYSCTL_OUT(req, 0, n_nchash * sizeof(int)); 1047 cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK); 1048 cache_lock_all_buckets(); 1049 if (n_nchash != nchash + 1) { 1050 cache_unlock_all_buckets(); 1051 free(cntbuf, M_TEMP); 1052 goto retry; 1053 } 1054 /* Scan hash tables counting entries */ 1055 for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++) 1056 CK_SLIST_FOREACH(ncp, ncpp, nc_hash) 1057 cntbuf[i]++; 1058 cache_unlock_all_buckets(); 1059 for (error = 0, i = 0; i < n_nchash; i++) 1060 if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0) 1061 break; 1062 free(cntbuf, M_TEMP); 1063 return (error); 1064 } 1065 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD| 1066 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int", 1067 "nchash chain lengths"); 1068 1069 static int 1070 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS) 1071 { 1072 int error; 1073 struct nchashhead *ncpp; 1074 struct namecache *ncp; 1075 int n_nchash; 1076 int count, maxlength, used, pct; 1077 1078 if (!req->oldptr) 1079 return SYSCTL_OUT(req, 0, 4 * sizeof(int)); 1080 1081 cache_lock_all_buckets(); 1082 n_nchash = nchash + 1; /* nchash is max index, not count */ 1083 used = 0; 1084 maxlength = 0; 1085 1086 /* Scan hash tables for applicable entries */ 1087 for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) { 1088 count = 0; 1089 CK_SLIST_FOREACH(ncp, ncpp, nc_hash) { 1090 count++; 1091 } 1092 if (count) 1093 used++; 1094 if (maxlength < count) 1095 maxlength = count; 1096 } 1097 n_nchash = nchash + 1; 1098 cache_unlock_all_buckets(); 1099 pct = (used * 100) / (n_nchash / 100); 1100 error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash)); 1101 if (error) 1102 return (error); 1103 error = SYSCTL_OUT(req, &used, sizeof(used)); 1104 if (error) 1105 return (error); 1106 error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength)); 1107 if (error) 1108 return (error); 1109 error = SYSCTL_OUT(req, &pct, sizeof(pct)); 1110 if (error) 1111 return (error); 1112 return (0); 1113 } 1114 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD| 1115 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I", 1116 "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)"); 1117 #endif 1118 1119 /* 1120 * Negative entries management 1121 * 1122 * Various workloads create plenty of negative entries and barely use them 1123 * afterwards. Moreover malicious users can keep performing bogus lookups 1124 * adding even more entries. For example "make tinderbox" as of writing this 1125 * comment ends up with 2.6M namecache entries in total, 1.2M of which are 1126 * negative. 1127 * 1128 * As such, a rather aggressive eviction method is needed. The currently 1129 * employed method is a placeholder. 1130 * 1131 * Entries are split over numneglists separate lists, each of which is further 1132 * split into hot and cold entries. Entries get promoted after getting a hit. 1133 * Eviction happens on addition of new entry. 1134 */ 1135 static SYSCTL_NODE(_vfs_cache, OID_AUTO, neg, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1136 "Name cache negative entry statistics"); 1137 1138 SYSCTL_ULONG(_vfs_cache_neg, OID_AUTO, count, CTLFLAG_RD, &numneg, 0, 1139 "Number of negative cache entries"); 1140 1141 static COUNTER_U64_DEFINE_EARLY(neg_created); 1142 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, created, CTLFLAG_RD, &neg_created, 1143 "Number of created negative entries"); 1144 1145 static COUNTER_U64_DEFINE_EARLY(neg_evicted); 1146 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evicted, CTLFLAG_RD, &neg_evicted, 1147 "Number of evicted negative entries"); 1148 1149 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_empty); 1150 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_empty, CTLFLAG_RD, 1151 &neg_evict_skipped_empty, 1152 "Number of times evicting failed due to lack of entries"); 1153 1154 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_missed); 1155 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_missed, CTLFLAG_RD, 1156 &neg_evict_skipped_missed, 1157 "Number of times evicting failed due to target entry disappearing"); 1158 1159 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_contended); 1160 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_contended, CTLFLAG_RD, 1161 &neg_evict_skipped_contended, 1162 "Number of times evicting failed due to contention"); 1163 1164 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, hits, CTLFLAG_RD, &numneghits, 1165 "Number of cache hits (negative)"); 1166 1167 static int 1168 sysctl_neg_hot(SYSCTL_HANDLER_ARGS) 1169 { 1170 int i, out; 1171 1172 out = 0; 1173 for (i = 0; i < numneglists; i++) 1174 out += neglists[i].nl_hotnum; 1175 1176 return (SYSCTL_OUT(req, &out, sizeof(out))); 1177 } 1178 SYSCTL_PROC(_vfs_cache_neg, OID_AUTO, hot, CTLTYPE_INT | CTLFLAG_RD | 1179 CTLFLAG_MPSAFE, 0, 0, sysctl_neg_hot, "I", 1180 "Number of hot negative entries"); 1181 1182 static void 1183 cache_neg_init(struct namecache *ncp) 1184 { 1185 struct negstate *ns; 1186 1187 ncp->nc_flag |= NCF_NEGATIVE; 1188 ns = NCP2NEGSTATE(ncp); 1189 ns->neg_flag = 0; 1190 ns->neg_hit = 0; 1191 counter_u64_add(neg_created, 1); 1192 } 1193 1194 #define CACHE_NEG_PROMOTION_THRESH 2 1195 1196 static bool 1197 cache_neg_hit_prep(struct namecache *ncp) 1198 { 1199 struct negstate *ns; 1200 u_char n; 1201 1202 ns = NCP2NEGSTATE(ncp); 1203 n = atomic_load_char(&ns->neg_hit); 1204 for (;;) { 1205 if (n >= CACHE_NEG_PROMOTION_THRESH) 1206 return (false); 1207 if (atomic_fcmpset_8(&ns->neg_hit, &n, n + 1)) 1208 break; 1209 } 1210 return (n + 1 == CACHE_NEG_PROMOTION_THRESH); 1211 } 1212 1213 /* 1214 * Nothing to do here but it is provided for completeness as some 1215 * cache_neg_hit_prep callers may end up returning without even 1216 * trying to promote. 1217 */ 1218 #define cache_neg_hit_abort(ncp) do { } while (0) 1219 1220 static void 1221 cache_neg_hit_finish(struct namecache *ncp) 1222 { 1223 1224 SDT_PROBE2(vfs, namecache, lookup, hit__negative, ncp->nc_dvp, ncp->nc_name); 1225 counter_u64_add(numneghits, 1); 1226 } 1227 1228 /* 1229 * Move a negative entry to the hot list. 1230 */ 1231 static void 1232 cache_neg_promote_locked(struct namecache *ncp) 1233 { 1234 struct neglist *nl; 1235 struct negstate *ns; 1236 1237 ns = NCP2NEGSTATE(ncp); 1238 nl = NCP2NEGLIST(ncp); 1239 mtx_assert(&nl->nl_lock, MA_OWNED); 1240 if ((ns->neg_flag & NEG_HOT) == 0) { 1241 TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst); 1242 TAILQ_INSERT_TAIL(&nl->nl_hotlist, ncp, nc_dst); 1243 nl->nl_hotnum++; 1244 ns->neg_flag |= NEG_HOT; 1245 } 1246 } 1247 1248 /* 1249 * Move a hot negative entry to the cold list. 1250 */ 1251 static void 1252 cache_neg_demote_locked(struct namecache *ncp) 1253 { 1254 struct neglist *nl; 1255 struct negstate *ns; 1256 1257 ns = NCP2NEGSTATE(ncp); 1258 nl = NCP2NEGLIST(ncp); 1259 mtx_assert(&nl->nl_lock, MA_OWNED); 1260 MPASS(ns->neg_flag & NEG_HOT); 1261 TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst); 1262 TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst); 1263 nl->nl_hotnum--; 1264 ns->neg_flag &= ~NEG_HOT; 1265 atomic_store_char(&ns->neg_hit, 0); 1266 } 1267 1268 /* 1269 * Move a negative entry to the hot list if it matches the lookup. 1270 * 1271 * We have to take locks, but they may be contended and in the worst 1272 * case we may need to go off CPU. We don't want to spin within the 1273 * smr section and we can't block with it. Exiting the section means 1274 * the found entry could have been evicted. We are going to look it 1275 * up again. 1276 */ 1277 static bool 1278 cache_neg_promote_cond(struct vnode *dvp, struct componentname *cnp, 1279 struct namecache *oncp, uint32_t hash) 1280 { 1281 struct namecache *ncp; 1282 struct neglist *nl; 1283 u_char nc_flag; 1284 1285 nl = NCP2NEGLIST(oncp); 1286 1287 mtx_lock(&nl->nl_lock); 1288 /* 1289 * For hash iteration. 1290 */ 1291 vfs_smr_enter(); 1292 1293 /* 1294 * Avoid all surprises by only succeeding if we got the same entry and 1295 * bailing completely otherwise. 1296 * XXX There are no provisions to keep the vnode around, meaning we may 1297 * end up promoting a negative entry for a *new* vnode and returning 1298 * ENOENT on its account. This is the error we want to return anyway 1299 * and promotion is harmless. 1300 * 1301 * In particular at this point there can be a new ncp which matches the 1302 * search but hashes to a different neglist. 1303 */ 1304 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 1305 if (ncp == oncp) 1306 break; 1307 } 1308 1309 /* 1310 * No match to begin with. 1311 */ 1312 if (__predict_false(ncp == NULL)) { 1313 goto out_abort; 1314 } 1315 1316 /* 1317 * The newly found entry may be something different... 1318 */ 1319 if (!(ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && 1320 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))) { 1321 goto out_abort; 1322 } 1323 1324 /* 1325 * ... and not even negative. 1326 */ 1327 nc_flag = atomic_load_char(&ncp->nc_flag); 1328 if ((nc_flag & NCF_NEGATIVE) == 0) { 1329 goto out_abort; 1330 } 1331 1332 if (!cache_ncp_canuse(ncp)) { 1333 goto out_abort; 1334 } 1335 1336 cache_neg_promote_locked(ncp); 1337 cache_neg_hit_finish(ncp); 1338 vfs_smr_exit(); 1339 mtx_unlock(&nl->nl_lock); 1340 return (true); 1341 out_abort: 1342 vfs_smr_exit(); 1343 mtx_unlock(&nl->nl_lock); 1344 return (false); 1345 } 1346 1347 static void 1348 cache_neg_promote(struct namecache *ncp) 1349 { 1350 struct neglist *nl; 1351 1352 nl = NCP2NEGLIST(ncp); 1353 mtx_lock(&nl->nl_lock); 1354 cache_neg_promote_locked(ncp); 1355 mtx_unlock(&nl->nl_lock); 1356 } 1357 1358 static void 1359 cache_neg_insert(struct namecache *ncp) 1360 { 1361 struct neglist *nl; 1362 1363 MPASS(ncp->nc_flag & NCF_NEGATIVE); 1364 cache_assert_bucket_locked(ncp); 1365 nl = NCP2NEGLIST(ncp); 1366 mtx_lock(&nl->nl_lock); 1367 TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst); 1368 mtx_unlock(&nl->nl_lock); 1369 atomic_add_long(&numneg, 1); 1370 } 1371 1372 static void 1373 cache_neg_remove(struct namecache *ncp) 1374 { 1375 struct neglist *nl; 1376 struct negstate *ns; 1377 1378 cache_assert_bucket_locked(ncp); 1379 nl = NCP2NEGLIST(ncp); 1380 ns = NCP2NEGSTATE(ncp); 1381 mtx_lock(&nl->nl_lock); 1382 if ((ns->neg_flag & NEG_HOT) != 0) { 1383 TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst); 1384 nl->nl_hotnum--; 1385 } else { 1386 TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst); 1387 } 1388 mtx_unlock(&nl->nl_lock); 1389 atomic_subtract_long(&numneg, 1); 1390 } 1391 1392 static struct neglist * 1393 cache_neg_evict_select_list(void) 1394 { 1395 struct neglist *nl; 1396 u_int c; 1397 1398 c = atomic_fetchadd_int(&neg_cycle, 1) + 1; 1399 nl = &neglists[c % numneglists]; 1400 if (!mtx_trylock(&nl->nl_evict_lock)) { 1401 counter_u64_add(neg_evict_skipped_contended, 1); 1402 return (NULL); 1403 } 1404 return (nl); 1405 } 1406 1407 static struct namecache * 1408 cache_neg_evict_select_entry(struct neglist *nl) 1409 { 1410 struct namecache *ncp, *lncp; 1411 struct negstate *ns, *lns; 1412 int i; 1413 1414 mtx_assert(&nl->nl_evict_lock, MA_OWNED); 1415 mtx_assert(&nl->nl_lock, MA_OWNED); 1416 ncp = TAILQ_FIRST(&nl->nl_list); 1417 if (ncp == NULL) 1418 return (NULL); 1419 lncp = ncp; 1420 lns = NCP2NEGSTATE(lncp); 1421 for (i = 1; i < 4; i++) { 1422 ncp = TAILQ_NEXT(ncp, nc_dst); 1423 if (ncp == NULL) 1424 break; 1425 ns = NCP2NEGSTATE(ncp); 1426 if (ns->neg_hit < lns->neg_hit) { 1427 lncp = ncp; 1428 lns = ns; 1429 } 1430 } 1431 return (lncp); 1432 } 1433 1434 static bool 1435 cache_neg_evict(void) 1436 { 1437 struct namecache *ncp, *ncp2; 1438 struct neglist *nl; 1439 struct vnode *dvp; 1440 struct mtx *dvlp; 1441 struct mtx *blp; 1442 uint32_t hash; 1443 u_char nlen; 1444 bool evicted; 1445 1446 nl = cache_neg_evict_select_list(); 1447 if (nl == NULL) { 1448 return (false); 1449 } 1450 1451 mtx_lock(&nl->nl_lock); 1452 ncp = TAILQ_FIRST(&nl->nl_hotlist); 1453 if (ncp != NULL) { 1454 cache_neg_demote_locked(ncp); 1455 } 1456 ncp = cache_neg_evict_select_entry(nl); 1457 if (ncp == NULL) { 1458 counter_u64_add(neg_evict_skipped_empty, 1); 1459 mtx_unlock(&nl->nl_lock); 1460 mtx_unlock(&nl->nl_evict_lock); 1461 return (false); 1462 } 1463 nlen = ncp->nc_nlen; 1464 dvp = ncp->nc_dvp; 1465 hash = cache_get_hash(ncp->nc_name, nlen, dvp); 1466 dvlp = VP2VNODELOCK(dvp); 1467 blp = HASH2BUCKETLOCK(hash); 1468 mtx_unlock(&nl->nl_lock); 1469 mtx_unlock(&nl->nl_evict_lock); 1470 mtx_lock(dvlp); 1471 mtx_lock(blp); 1472 /* 1473 * Note that since all locks were dropped above, the entry may be 1474 * gone or reallocated to be something else. 1475 */ 1476 CK_SLIST_FOREACH(ncp2, (NCHHASH(hash)), nc_hash) { 1477 if (ncp2 == ncp && ncp2->nc_dvp == dvp && 1478 ncp2->nc_nlen == nlen && (ncp2->nc_flag & NCF_NEGATIVE) != 0) 1479 break; 1480 } 1481 if (ncp2 == NULL) { 1482 counter_u64_add(neg_evict_skipped_missed, 1); 1483 ncp = NULL; 1484 evicted = false; 1485 } else { 1486 MPASS(dvlp == VP2VNODELOCK(ncp->nc_dvp)); 1487 MPASS(blp == NCP2BUCKETLOCK(ncp)); 1488 SDT_PROBE2(vfs, namecache, evict_negative, done, ncp->nc_dvp, 1489 ncp->nc_name); 1490 cache_zap_locked(ncp); 1491 counter_u64_add(neg_evicted, 1); 1492 evicted = true; 1493 } 1494 mtx_unlock(blp); 1495 mtx_unlock(dvlp); 1496 if (ncp != NULL) 1497 cache_free(ncp); 1498 return (evicted); 1499 } 1500 1501 /* 1502 * Maybe evict a negative entry to create more room. 1503 * 1504 * The ncnegfactor parameter limits what fraction of the total count 1505 * can comprise of negative entries. However, if the cache is just 1506 * warming up this leads to excessive evictions. As such, ncnegminpct 1507 * (recomputed to neg_min) dictates whether the above should be 1508 * applied. 1509 * 1510 * Try evicting if the cache is close to full capacity regardless of 1511 * other considerations. 1512 */ 1513 static bool 1514 cache_neg_evict_cond(u_long lnumcache) 1515 { 1516 u_long lnumneg; 1517 1518 if (ncsize - 1000 < lnumcache) 1519 goto out_evict; 1520 lnumneg = atomic_load_long(&numneg); 1521 if (lnumneg < neg_min) 1522 return (false); 1523 if (lnumneg * ncnegfactor < lnumcache) 1524 return (false); 1525 out_evict: 1526 return (cache_neg_evict()); 1527 } 1528 1529 /* 1530 * cache_zap_locked(): 1531 * 1532 * Removes a namecache entry from cache, whether it contains an actual 1533 * pointer to a vnode or if it is just a negative cache entry. 1534 */ 1535 static void 1536 cache_zap_locked(struct namecache *ncp) 1537 { 1538 struct nchashhead *ncpp; 1539 struct vnode *dvp, *vp; 1540 1541 dvp = ncp->nc_dvp; 1542 vp = ncp->nc_vp; 1543 1544 if (!(ncp->nc_flag & NCF_NEGATIVE)) 1545 cache_assert_vnode_locked(vp); 1546 cache_assert_vnode_locked(dvp); 1547 cache_assert_bucket_locked(ncp); 1548 1549 cache_ncp_invalidate(ncp); 1550 1551 ncpp = NCP2BUCKET(ncp); 1552 CK_SLIST_REMOVE(ncpp, ncp, namecache, nc_hash); 1553 if (!(ncp->nc_flag & NCF_NEGATIVE)) { 1554 SDT_PROBE3(vfs, namecache, zap, done, dvp, ncp->nc_name, vp); 1555 TAILQ_REMOVE(&vp->v_cache_dst, ncp, nc_dst); 1556 if (ncp == vp->v_cache_dd) { 1557 atomic_store_ptr(&vp->v_cache_dd, NULL); 1558 } 1559 } else { 1560 SDT_PROBE2(vfs, namecache, zap_negative, done, dvp, ncp->nc_name); 1561 cache_neg_remove(ncp); 1562 } 1563 if (ncp->nc_flag & NCF_ISDOTDOT) { 1564 if (ncp == dvp->v_cache_dd) { 1565 atomic_store_ptr(&dvp->v_cache_dd, NULL); 1566 } 1567 } else { 1568 LIST_REMOVE(ncp, nc_src); 1569 if (LIST_EMPTY(&dvp->v_cache_src)) { 1570 ncp->nc_flag |= NCF_DVDROP; 1571 } 1572 } 1573 } 1574 1575 static void 1576 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp) 1577 { 1578 struct mtx *blp; 1579 1580 MPASS(ncp->nc_dvp == vp); 1581 MPASS(ncp->nc_flag & NCF_NEGATIVE); 1582 cache_assert_vnode_locked(vp); 1583 1584 blp = NCP2BUCKETLOCK(ncp); 1585 mtx_lock(blp); 1586 cache_zap_locked(ncp); 1587 mtx_unlock(blp); 1588 } 1589 1590 static bool 1591 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp, 1592 struct mtx **vlpp) 1593 { 1594 struct mtx *pvlp, *vlp1, *vlp2, *to_unlock; 1595 struct mtx *blp; 1596 1597 MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp); 1598 cache_assert_vnode_locked(vp); 1599 1600 if (ncp->nc_flag & NCF_NEGATIVE) { 1601 if (*vlpp != NULL) { 1602 mtx_unlock(*vlpp); 1603 *vlpp = NULL; 1604 } 1605 cache_zap_negative_locked_vnode_kl(ncp, vp); 1606 return (true); 1607 } 1608 1609 pvlp = VP2VNODELOCK(vp); 1610 blp = NCP2BUCKETLOCK(ncp); 1611 vlp1 = VP2VNODELOCK(ncp->nc_dvp); 1612 vlp2 = VP2VNODELOCK(ncp->nc_vp); 1613 1614 if (*vlpp == vlp1 || *vlpp == vlp2) { 1615 to_unlock = *vlpp; 1616 *vlpp = NULL; 1617 } else { 1618 if (*vlpp != NULL) { 1619 mtx_unlock(*vlpp); 1620 *vlpp = NULL; 1621 } 1622 cache_sort_vnodes(&vlp1, &vlp2); 1623 if (vlp1 == pvlp) { 1624 mtx_lock(vlp2); 1625 to_unlock = vlp2; 1626 } else { 1627 if (!mtx_trylock(vlp1)) 1628 goto out_relock; 1629 to_unlock = vlp1; 1630 } 1631 } 1632 mtx_lock(blp); 1633 cache_zap_locked(ncp); 1634 mtx_unlock(blp); 1635 if (to_unlock != NULL) 1636 mtx_unlock(to_unlock); 1637 return (true); 1638 1639 out_relock: 1640 mtx_unlock(vlp2); 1641 mtx_lock(vlp1); 1642 mtx_lock(vlp2); 1643 MPASS(*vlpp == NULL); 1644 *vlpp = vlp1; 1645 return (false); 1646 } 1647 1648 /* 1649 * If trylocking failed we can get here. We know enough to take all needed locks 1650 * in the right order and re-lookup the entry. 1651 */ 1652 static int 1653 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp, 1654 struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash, 1655 struct mtx *blp) 1656 { 1657 struct namecache *rncp; 1658 1659 cache_assert_bucket_unlocked(ncp); 1660 1661 cache_sort_vnodes(&dvlp, &vlp); 1662 cache_lock_vnodes(dvlp, vlp); 1663 mtx_lock(blp); 1664 CK_SLIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) { 1665 if (rncp == ncp && rncp->nc_dvp == dvp && 1666 rncp->nc_nlen == cnp->cn_namelen && 1667 !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen)) 1668 break; 1669 } 1670 if (rncp != NULL) { 1671 cache_zap_locked(rncp); 1672 mtx_unlock(blp); 1673 cache_unlock_vnodes(dvlp, vlp); 1674 counter_u64_add(zap_bucket_relock_success, 1); 1675 return (0); 1676 } 1677 1678 mtx_unlock(blp); 1679 cache_unlock_vnodes(dvlp, vlp); 1680 return (EAGAIN); 1681 } 1682 1683 static int __noinline 1684 cache_zap_locked_bucket(struct namecache *ncp, struct componentname *cnp, 1685 uint32_t hash, struct mtx *blp) 1686 { 1687 struct mtx *dvlp, *vlp; 1688 struct vnode *dvp; 1689 1690 cache_assert_bucket_locked(ncp); 1691 1692 dvlp = VP2VNODELOCK(ncp->nc_dvp); 1693 vlp = NULL; 1694 if (!(ncp->nc_flag & NCF_NEGATIVE)) 1695 vlp = VP2VNODELOCK(ncp->nc_vp); 1696 if (cache_trylock_vnodes(dvlp, vlp) == 0) { 1697 cache_zap_locked(ncp); 1698 mtx_unlock(blp); 1699 cache_unlock_vnodes(dvlp, vlp); 1700 return (0); 1701 } 1702 1703 dvp = ncp->nc_dvp; 1704 mtx_unlock(blp); 1705 return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp)); 1706 } 1707 1708 static __noinline int 1709 cache_remove_cnp(struct vnode *dvp, struct componentname *cnp) 1710 { 1711 struct namecache *ncp; 1712 struct mtx *blp; 1713 struct mtx *dvlp, *dvlp2; 1714 uint32_t hash; 1715 int error; 1716 1717 if (cnp->cn_namelen == 2 && 1718 cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') { 1719 dvlp = VP2VNODELOCK(dvp); 1720 dvlp2 = NULL; 1721 mtx_lock(dvlp); 1722 retry_dotdot: 1723 ncp = dvp->v_cache_dd; 1724 if (ncp == NULL) { 1725 mtx_unlock(dvlp); 1726 if (dvlp2 != NULL) 1727 mtx_unlock(dvlp2); 1728 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp); 1729 return (0); 1730 } 1731 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) { 1732 if (!cache_zap_locked_vnode_kl2(ncp, dvp, &dvlp2)) 1733 goto retry_dotdot; 1734 MPASS(dvp->v_cache_dd == NULL); 1735 mtx_unlock(dvlp); 1736 if (dvlp2 != NULL) 1737 mtx_unlock(dvlp2); 1738 cache_free(ncp); 1739 } else { 1740 atomic_store_ptr(&dvp->v_cache_dd, NULL); 1741 mtx_unlock(dvlp); 1742 if (dvlp2 != NULL) 1743 mtx_unlock(dvlp2); 1744 } 1745 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp); 1746 return (1); 1747 } 1748 1749 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); 1750 blp = HASH2BUCKETLOCK(hash); 1751 retry: 1752 if (CK_SLIST_EMPTY(NCHHASH(hash))) 1753 goto out_no_entry; 1754 1755 mtx_lock(blp); 1756 1757 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 1758 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && 1759 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) 1760 break; 1761 } 1762 1763 if (ncp == NULL) { 1764 mtx_unlock(blp); 1765 goto out_no_entry; 1766 } 1767 1768 error = cache_zap_locked_bucket(ncp, cnp, hash, blp); 1769 if (__predict_false(error != 0)) { 1770 zap_bucket_fail++; 1771 goto retry; 1772 } 1773 counter_u64_add(numposzaps, 1); 1774 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp); 1775 cache_free(ncp); 1776 return (1); 1777 out_no_entry: 1778 counter_u64_add(nummisszap, 1); 1779 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp); 1780 return (0); 1781 } 1782 1783 static int __noinline 1784 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, 1785 struct timespec *tsp, int *ticksp) 1786 { 1787 int ltype; 1788 1789 *vpp = dvp; 1790 counter_u64_add(dothits, 1); 1791 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp); 1792 if (tsp != NULL) 1793 timespecclear(tsp); 1794 if (ticksp != NULL) 1795 *ticksp = ticks; 1796 vrefact(*vpp); 1797 /* 1798 * When we lookup "." we still can be asked to lock it 1799 * differently... 1800 */ 1801 ltype = cnp->cn_lkflags & LK_TYPE_MASK; 1802 if (ltype != VOP_ISLOCKED(*vpp)) { 1803 if (ltype == LK_EXCLUSIVE) { 1804 vn_lock(*vpp, LK_UPGRADE | LK_RETRY); 1805 if (VN_IS_DOOMED((*vpp))) { 1806 /* forced unmount */ 1807 vrele(*vpp); 1808 *vpp = NULL; 1809 return (ENOENT); 1810 } 1811 } else 1812 vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY); 1813 } 1814 return (-1); 1815 } 1816 1817 static int __noinline 1818 cache_lookup_dotdot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, 1819 struct timespec *tsp, int *ticksp) 1820 { 1821 struct namecache_ts *ncp_ts; 1822 struct namecache *ncp; 1823 struct mtx *dvlp; 1824 enum vgetstate vs; 1825 int error, ltype; 1826 bool whiteout; 1827 1828 MPASS((cnp->cn_flags & ISDOTDOT) != 0); 1829 1830 if ((cnp->cn_flags & MAKEENTRY) == 0) { 1831 cache_remove_cnp(dvp, cnp); 1832 return (0); 1833 } 1834 1835 counter_u64_add(dotdothits, 1); 1836 retry: 1837 dvlp = VP2VNODELOCK(dvp); 1838 mtx_lock(dvlp); 1839 ncp = dvp->v_cache_dd; 1840 if (ncp == NULL) { 1841 SDT_PROBE2(vfs, namecache, lookup, miss, dvp, ".."); 1842 mtx_unlock(dvlp); 1843 return (0); 1844 } 1845 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) { 1846 if (ncp->nc_flag & NCF_NEGATIVE) 1847 *vpp = NULL; 1848 else 1849 *vpp = ncp->nc_vp; 1850 } else 1851 *vpp = ncp->nc_dvp; 1852 if (*vpp == NULL) 1853 goto negative_success; 1854 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", *vpp); 1855 cache_out_ts(ncp, tsp, ticksp); 1856 if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) == 1857 NCF_DTS && tsp != NULL) { 1858 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc); 1859 *tsp = ncp_ts->nc_dotdottime; 1860 } 1861 1862 MPASS(dvp != *vpp); 1863 ltype = VOP_ISLOCKED(dvp); 1864 VOP_UNLOCK(dvp); 1865 vs = vget_prep(*vpp); 1866 mtx_unlock(dvlp); 1867 error = vget_finish(*vpp, cnp->cn_lkflags, vs); 1868 vn_lock(dvp, ltype | LK_RETRY); 1869 if (VN_IS_DOOMED(dvp)) { 1870 if (error == 0) 1871 vput(*vpp); 1872 *vpp = NULL; 1873 return (ENOENT); 1874 } 1875 if (error) { 1876 *vpp = NULL; 1877 goto retry; 1878 } 1879 return (-1); 1880 negative_success: 1881 if (__predict_false(cnp->cn_nameiop == CREATE)) { 1882 if (cnp->cn_flags & ISLASTCN) { 1883 counter_u64_add(numnegzaps, 1); 1884 cache_zap_negative_locked_vnode_kl(ncp, dvp); 1885 mtx_unlock(dvlp); 1886 cache_free(ncp); 1887 return (0); 1888 } 1889 } 1890 1891 whiteout = (ncp->nc_flag & NCF_WHITE); 1892 cache_out_ts(ncp, tsp, ticksp); 1893 if (cache_neg_hit_prep(ncp)) 1894 cache_neg_promote(ncp); 1895 else 1896 cache_neg_hit_finish(ncp); 1897 mtx_unlock(dvlp); 1898 if (whiteout) 1899 cnp->cn_flags |= ISWHITEOUT; 1900 return (ENOENT); 1901 } 1902 1903 /** 1904 * Lookup a name in the name cache 1905 * 1906 * # Arguments 1907 * 1908 * - dvp: Parent directory in which to search. 1909 * - vpp: Return argument. Will contain desired vnode on cache hit. 1910 * - cnp: Parameters of the name search. The most interesting bits of 1911 * the cn_flags field have the following meanings: 1912 * - MAKEENTRY: If clear, free an entry from the cache rather than look 1913 * it up. 1914 * - ISDOTDOT: Must be set if and only if cn_nameptr == ".." 1915 * - tsp: Return storage for cache timestamp. On a successful (positive 1916 * or negative) lookup, tsp will be filled with any timespec that 1917 * was stored when this cache entry was created. However, it will 1918 * be clear for "." entries. 1919 * - ticks: Return storage for alternate cache timestamp. On a successful 1920 * (positive or negative) lookup, it will contain the ticks value 1921 * that was current when the cache entry was created, unless cnp 1922 * was ".". 1923 * 1924 * Either both tsp and ticks have to be provided or neither of them. 1925 * 1926 * # Returns 1927 * 1928 * - -1: A positive cache hit. vpp will contain the desired vnode. 1929 * - ENOENT: A negative cache hit, or dvp was recycled out from under us due 1930 * to a forced unmount. vpp will not be modified. If the entry 1931 * is a whiteout, then the ISWHITEOUT flag will be set in 1932 * cnp->cn_flags. 1933 * - 0: A cache miss. vpp will not be modified. 1934 * 1935 * # Locking 1936 * 1937 * On a cache hit, vpp will be returned locked and ref'd. If we're looking up 1938 * .., dvp is unlocked. If we're looking up . an extra ref is taken, but the 1939 * lock is not recursively acquired. 1940 */ 1941 static int __noinline 1942 cache_lookup_fallback(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, 1943 struct timespec *tsp, int *ticksp) 1944 { 1945 struct namecache *ncp; 1946 struct mtx *blp; 1947 uint32_t hash; 1948 enum vgetstate vs; 1949 int error; 1950 bool whiteout; 1951 1952 MPASS((cnp->cn_flags & ISDOTDOT) == 0); 1953 MPASS((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) != 0); 1954 1955 retry: 1956 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); 1957 blp = HASH2BUCKETLOCK(hash); 1958 mtx_lock(blp); 1959 1960 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 1961 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && 1962 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) 1963 break; 1964 } 1965 1966 if (__predict_false(ncp == NULL)) { 1967 mtx_unlock(blp); 1968 SDT_PROBE2(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr); 1969 counter_u64_add(nummiss, 1); 1970 return (0); 1971 } 1972 1973 if (ncp->nc_flag & NCF_NEGATIVE) 1974 goto negative_success; 1975 1976 counter_u64_add(numposhits, 1); 1977 *vpp = ncp->nc_vp; 1978 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp); 1979 cache_out_ts(ncp, tsp, ticksp); 1980 MPASS(dvp != *vpp); 1981 vs = vget_prep(*vpp); 1982 mtx_unlock(blp); 1983 error = vget_finish(*vpp, cnp->cn_lkflags, vs); 1984 if (error) { 1985 *vpp = NULL; 1986 goto retry; 1987 } 1988 return (-1); 1989 negative_success: 1990 /* 1991 * We don't get here with regular lookup apart from corner cases. 1992 */ 1993 if (__predict_true(cnp->cn_nameiop == CREATE)) { 1994 if (cnp->cn_flags & ISLASTCN) { 1995 counter_u64_add(numnegzaps, 1); 1996 error = cache_zap_locked_bucket(ncp, cnp, hash, blp); 1997 if (__predict_false(error != 0)) { 1998 zap_bucket_fail2++; 1999 goto retry; 2000 } 2001 cache_free(ncp); 2002 return (0); 2003 } 2004 } 2005 2006 whiteout = (ncp->nc_flag & NCF_WHITE); 2007 cache_out_ts(ncp, tsp, ticksp); 2008 if (cache_neg_hit_prep(ncp)) 2009 cache_neg_promote(ncp); 2010 else 2011 cache_neg_hit_finish(ncp); 2012 mtx_unlock(blp); 2013 if (whiteout) 2014 cnp->cn_flags |= ISWHITEOUT; 2015 return (ENOENT); 2016 } 2017 2018 int 2019 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, 2020 struct timespec *tsp, int *ticksp) 2021 { 2022 struct namecache *ncp; 2023 uint32_t hash; 2024 enum vgetstate vs; 2025 int error; 2026 bool whiteout, neg_promote; 2027 u_short nc_flag; 2028 2029 MPASS((tsp == NULL && ticksp == NULL) || (tsp != NULL && ticksp != NULL)); 2030 2031 #ifdef DEBUG_CACHE 2032 if (__predict_false(!doingcache)) { 2033 cnp->cn_flags &= ~MAKEENTRY; 2034 return (0); 2035 } 2036 #endif 2037 2038 if (__predict_false(cnp->cn_nameptr[0] == '.')) { 2039 if (cnp->cn_namelen == 1) 2040 return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp)); 2041 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') 2042 return (cache_lookup_dotdot(dvp, vpp, cnp, tsp, ticksp)); 2043 } 2044 2045 MPASS((cnp->cn_flags & ISDOTDOT) == 0); 2046 2047 if ((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) == 0) { 2048 cache_remove_cnp(dvp, cnp); 2049 return (0); 2050 } 2051 2052 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); 2053 vfs_smr_enter(); 2054 2055 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 2056 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && 2057 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) 2058 break; 2059 } 2060 2061 if (__predict_false(ncp == NULL)) { 2062 vfs_smr_exit(); 2063 SDT_PROBE2(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr); 2064 counter_u64_add(nummiss, 1); 2065 return (0); 2066 } 2067 2068 nc_flag = atomic_load_char(&ncp->nc_flag); 2069 if (nc_flag & NCF_NEGATIVE) 2070 goto negative_success; 2071 2072 counter_u64_add(numposhits, 1); 2073 *vpp = ncp->nc_vp; 2074 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp); 2075 cache_out_ts(ncp, tsp, ticksp); 2076 MPASS(dvp != *vpp); 2077 if (!cache_ncp_canuse(ncp)) { 2078 vfs_smr_exit(); 2079 *vpp = NULL; 2080 goto out_fallback; 2081 } 2082 vs = vget_prep_smr(*vpp); 2083 vfs_smr_exit(); 2084 if (__predict_false(vs == VGET_NONE)) { 2085 *vpp = NULL; 2086 goto out_fallback; 2087 } 2088 error = vget_finish(*vpp, cnp->cn_lkflags, vs); 2089 if (error) { 2090 *vpp = NULL; 2091 goto out_fallback; 2092 } 2093 return (-1); 2094 negative_success: 2095 if (cnp->cn_nameiop == CREATE) { 2096 if (cnp->cn_flags & ISLASTCN) { 2097 vfs_smr_exit(); 2098 goto out_fallback; 2099 } 2100 } 2101 2102 cache_out_ts(ncp, tsp, ticksp); 2103 whiteout = (atomic_load_char(&ncp->nc_flag) & NCF_WHITE); 2104 neg_promote = cache_neg_hit_prep(ncp); 2105 if (!cache_ncp_canuse(ncp)) { 2106 cache_neg_hit_abort(ncp); 2107 vfs_smr_exit(); 2108 goto out_fallback; 2109 } 2110 if (neg_promote) { 2111 vfs_smr_exit(); 2112 if (!cache_neg_promote_cond(dvp, cnp, ncp, hash)) 2113 goto out_fallback; 2114 } else { 2115 cache_neg_hit_finish(ncp); 2116 vfs_smr_exit(); 2117 } 2118 if (whiteout) 2119 cnp->cn_flags |= ISWHITEOUT; 2120 return (ENOENT); 2121 out_fallback: 2122 return (cache_lookup_fallback(dvp, vpp, cnp, tsp, ticksp)); 2123 } 2124 2125 struct celockstate { 2126 struct mtx *vlp[3]; 2127 struct mtx *blp[2]; 2128 }; 2129 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3)); 2130 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2)); 2131 2132 static inline void 2133 cache_celockstate_init(struct celockstate *cel) 2134 { 2135 2136 bzero(cel, sizeof(*cel)); 2137 } 2138 2139 static void 2140 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp, 2141 struct vnode *dvp) 2142 { 2143 struct mtx *vlp1, *vlp2; 2144 2145 MPASS(cel->vlp[0] == NULL); 2146 MPASS(cel->vlp[1] == NULL); 2147 MPASS(cel->vlp[2] == NULL); 2148 2149 MPASS(vp != NULL || dvp != NULL); 2150 2151 vlp1 = VP2VNODELOCK(vp); 2152 vlp2 = VP2VNODELOCK(dvp); 2153 cache_sort_vnodes(&vlp1, &vlp2); 2154 2155 if (vlp1 != NULL) { 2156 mtx_lock(vlp1); 2157 cel->vlp[0] = vlp1; 2158 } 2159 mtx_lock(vlp2); 2160 cel->vlp[1] = vlp2; 2161 } 2162 2163 static void 2164 cache_unlock_vnodes_cel(struct celockstate *cel) 2165 { 2166 2167 MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL); 2168 2169 if (cel->vlp[0] != NULL) 2170 mtx_unlock(cel->vlp[0]); 2171 if (cel->vlp[1] != NULL) 2172 mtx_unlock(cel->vlp[1]); 2173 if (cel->vlp[2] != NULL) 2174 mtx_unlock(cel->vlp[2]); 2175 } 2176 2177 static bool 2178 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp) 2179 { 2180 struct mtx *vlp; 2181 bool ret; 2182 2183 cache_assert_vlp_locked(cel->vlp[0]); 2184 cache_assert_vlp_locked(cel->vlp[1]); 2185 MPASS(cel->vlp[2] == NULL); 2186 2187 MPASS(vp != NULL); 2188 vlp = VP2VNODELOCK(vp); 2189 2190 ret = true; 2191 if (vlp >= cel->vlp[1]) { 2192 mtx_lock(vlp); 2193 } else { 2194 if (mtx_trylock(vlp)) 2195 goto out; 2196 cache_lock_vnodes_cel_3_failures++; 2197 cache_unlock_vnodes_cel(cel); 2198 if (vlp < cel->vlp[0]) { 2199 mtx_lock(vlp); 2200 mtx_lock(cel->vlp[0]); 2201 mtx_lock(cel->vlp[1]); 2202 } else { 2203 if (cel->vlp[0] != NULL) 2204 mtx_lock(cel->vlp[0]); 2205 mtx_lock(vlp); 2206 mtx_lock(cel->vlp[1]); 2207 } 2208 ret = false; 2209 } 2210 out: 2211 cel->vlp[2] = vlp; 2212 return (ret); 2213 } 2214 2215 static void 2216 cache_lock_buckets_cel(struct celockstate *cel, struct mtx *blp1, 2217 struct mtx *blp2) 2218 { 2219 2220 MPASS(cel->blp[0] == NULL); 2221 MPASS(cel->blp[1] == NULL); 2222 2223 cache_sort_vnodes(&blp1, &blp2); 2224 2225 if (blp1 != NULL) { 2226 mtx_lock(blp1); 2227 cel->blp[0] = blp1; 2228 } 2229 mtx_lock(blp2); 2230 cel->blp[1] = blp2; 2231 } 2232 2233 static void 2234 cache_unlock_buckets_cel(struct celockstate *cel) 2235 { 2236 2237 if (cel->blp[0] != NULL) 2238 mtx_unlock(cel->blp[0]); 2239 mtx_unlock(cel->blp[1]); 2240 } 2241 2242 /* 2243 * Lock part of the cache affected by the insertion. 2244 * 2245 * This means vnodelocks for dvp, vp and the relevant bucketlock. 2246 * However, insertion can result in removal of an old entry. In this 2247 * case we have an additional vnode and bucketlock pair to lock. 2248 * 2249 * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while 2250 * preserving the locking order (smaller address first). 2251 */ 2252 static void 2253 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp, 2254 uint32_t hash) 2255 { 2256 struct namecache *ncp; 2257 struct mtx *blps[2]; 2258 u_char nc_flag; 2259 2260 blps[0] = HASH2BUCKETLOCK(hash); 2261 for (;;) { 2262 blps[1] = NULL; 2263 cache_lock_vnodes_cel(cel, dvp, vp); 2264 if (vp == NULL || vp->v_type != VDIR) 2265 break; 2266 ncp = atomic_load_consume_ptr(&vp->v_cache_dd); 2267 if (ncp == NULL) 2268 break; 2269 nc_flag = atomic_load_char(&ncp->nc_flag); 2270 if ((nc_flag & NCF_ISDOTDOT) == 0) 2271 break; 2272 MPASS(ncp->nc_dvp == vp); 2273 blps[1] = NCP2BUCKETLOCK(ncp); 2274 if ((nc_flag & NCF_NEGATIVE) != 0) 2275 break; 2276 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp)) 2277 break; 2278 /* 2279 * All vnodes got re-locked. Re-validate the state and if 2280 * nothing changed we are done. Otherwise restart. 2281 */ 2282 if (ncp == vp->v_cache_dd && 2283 (ncp->nc_flag & NCF_ISDOTDOT) != 0 && 2284 blps[1] == NCP2BUCKETLOCK(ncp) && 2285 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2]) 2286 break; 2287 cache_unlock_vnodes_cel(cel); 2288 cel->vlp[0] = NULL; 2289 cel->vlp[1] = NULL; 2290 cel->vlp[2] = NULL; 2291 } 2292 cache_lock_buckets_cel(cel, blps[0], blps[1]); 2293 } 2294 2295 static void 2296 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp, 2297 uint32_t hash) 2298 { 2299 struct namecache *ncp; 2300 struct mtx *blps[2]; 2301 u_char nc_flag; 2302 2303 blps[0] = HASH2BUCKETLOCK(hash); 2304 for (;;) { 2305 blps[1] = NULL; 2306 cache_lock_vnodes_cel(cel, dvp, vp); 2307 ncp = atomic_load_consume_ptr(&dvp->v_cache_dd); 2308 if (ncp == NULL) 2309 break; 2310 nc_flag = atomic_load_char(&ncp->nc_flag); 2311 if ((nc_flag & NCF_ISDOTDOT) == 0) 2312 break; 2313 MPASS(ncp->nc_dvp == dvp); 2314 blps[1] = NCP2BUCKETLOCK(ncp); 2315 if ((nc_flag & NCF_NEGATIVE) != 0) 2316 break; 2317 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp)) 2318 break; 2319 if (ncp == dvp->v_cache_dd && 2320 (ncp->nc_flag & NCF_ISDOTDOT) != 0 && 2321 blps[1] == NCP2BUCKETLOCK(ncp) && 2322 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2]) 2323 break; 2324 cache_unlock_vnodes_cel(cel); 2325 cel->vlp[0] = NULL; 2326 cel->vlp[1] = NULL; 2327 cel->vlp[2] = NULL; 2328 } 2329 cache_lock_buckets_cel(cel, blps[0], blps[1]); 2330 } 2331 2332 static void 2333 cache_enter_unlock(struct celockstate *cel) 2334 { 2335 2336 cache_unlock_buckets_cel(cel); 2337 cache_unlock_vnodes_cel(cel); 2338 } 2339 2340 static void __noinline 2341 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp, 2342 struct componentname *cnp) 2343 { 2344 struct celockstate cel; 2345 struct namecache *ncp; 2346 uint32_t hash; 2347 int len; 2348 2349 if (atomic_load_ptr(&dvp->v_cache_dd) == NULL) 2350 return; 2351 len = cnp->cn_namelen; 2352 cache_celockstate_init(&cel); 2353 hash = cache_get_hash(cnp->cn_nameptr, len, dvp); 2354 cache_enter_lock_dd(&cel, dvp, vp, hash); 2355 ncp = dvp->v_cache_dd; 2356 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) { 2357 KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent")); 2358 cache_zap_locked(ncp); 2359 } else { 2360 ncp = NULL; 2361 } 2362 atomic_store_ptr(&dvp->v_cache_dd, NULL); 2363 cache_enter_unlock(&cel); 2364 if (ncp != NULL) 2365 cache_free(ncp); 2366 } 2367 2368 /* 2369 * Add an entry to the cache. 2370 */ 2371 void 2372 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, 2373 struct timespec *tsp, struct timespec *dtsp) 2374 { 2375 struct celockstate cel; 2376 struct namecache *ncp, *n2, *ndd; 2377 struct namecache_ts *ncp_ts; 2378 struct nchashhead *ncpp; 2379 uint32_t hash; 2380 int flag; 2381 int len; 2382 2383 KASSERT(cnp->cn_namelen <= NAME_MAX, 2384 ("%s: passed len %ld exceeds NAME_MAX (%d)", __func__, cnp->cn_namelen, 2385 NAME_MAX)); 2386 VNPASS(!VN_IS_DOOMED(dvp), dvp); 2387 VNPASS(dvp->v_type != VNON, dvp); 2388 if (vp != NULL) { 2389 VNPASS(!VN_IS_DOOMED(vp), vp); 2390 VNPASS(vp->v_type != VNON, vp); 2391 } 2392 if (cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.') { 2393 KASSERT(dvp == vp, 2394 ("%s: different vnodes for dot entry (%p; %p)\n", __func__, 2395 dvp, vp)); 2396 } else { 2397 KASSERT(dvp != vp, 2398 ("%s: same vnode for non-dot entry [%s] (%p)\n", __func__, 2399 cnp->cn_nameptr, dvp)); 2400 } 2401 2402 #ifdef DEBUG_CACHE 2403 if (__predict_false(!doingcache)) 2404 return; 2405 #endif 2406 2407 flag = 0; 2408 if (__predict_false(cnp->cn_nameptr[0] == '.')) { 2409 if (cnp->cn_namelen == 1) 2410 return; 2411 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') { 2412 cache_enter_dotdot_prep(dvp, vp, cnp); 2413 flag = NCF_ISDOTDOT; 2414 } 2415 } 2416 2417 ncp = cache_alloc(cnp->cn_namelen, tsp != NULL); 2418 if (ncp == NULL) 2419 return; 2420 2421 cache_celockstate_init(&cel); 2422 ndd = NULL; 2423 ncp_ts = NULL; 2424 2425 /* 2426 * Calculate the hash key and setup as much of the new 2427 * namecache entry as possible before acquiring the lock. 2428 */ 2429 ncp->nc_flag = flag | NCF_WIP; 2430 ncp->nc_vp = vp; 2431 if (vp == NULL) 2432 cache_neg_init(ncp); 2433 ncp->nc_dvp = dvp; 2434 if (tsp != NULL) { 2435 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc); 2436 ncp_ts->nc_time = *tsp; 2437 ncp_ts->nc_ticks = ticks; 2438 ncp_ts->nc_nc.nc_flag |= NCF_TS; 2439 if (dtsp != NULL) { 2440 ncp_ts->nc_dotdottime = *dtsp; 2441 ncp_ts->nc_nc.nc_flag |= NCF_DTS; 2442 } 2443 } 2444 len = ncp->nc_nlen = cnp->cn_namelen; 2445 hash = cache_get_hash(cnp->cn_nameptr, len, dvp); 2446 memcpy(ncp->nc_name, cnp->cn_nameptr, len); 2447 ncp->nc_name[len] = '\0'; 2448 cache_enter_lock(&cel, dvp, vp, hash); 2449 2450 /* 2451 * See if this vnode or negative entry is already in the cache 2452 * with this name. This can happen with concurrent lookups of 2453 * the same path name. 2454 */ 2455 ncpp = NCHHASH(hash); 2456 CK_SLIST_FOREACH(n2, ncpp, nc_hash) { 2457 if (n2->nc_dvp == dvp && 2458 n2->nc_nlen == cnp->cn_namelen && 2459 !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) { 2460 MPASS(cache_ncp_canuse(n2)); 2461 if ((n2->nc_flag & NCF_NEGATIVE) != 0) 2462 KASSERT(vp == NULL, 2463 ("%s: found entry pointing to a different vnode (%p != %p) ; name [%s]", 2464 __func__, NULL, vp, cnp->cn_nameptr)); 2465 else 2466 KASSERT(n2->nc_vp == vp, 2467 ("%s: found entry pointing to a different vnode (%p != %p) ; name [%s]", 2468 __func__, n2->nc_vp, vp, cnp->cn_nameptr)); 2469 /* 2470 * Entries are supposed to be immutable unless in the 2471 * process of getting destroyed. Accommodating for 2472 * changing timestamps is possible but not worth it. 2473 * This should be harmless in terms of correctness, in 2474 * the worst case resulting in an earlier expiration. 2475 * Alternatively, the found entry can be replaced 2476 * altogether. 2477 */ 2478 MPASS((n2->nc_flag & (NCF_TS | NCF_DTS)) == (ncp->nc_flag & (NCF_TS | NCF_DTS))); 2479 #if 0 2480 if (tsp != NULL) { 2481 KASSERT((n2->nc_flag & NCF_TS) != 0, 2482 ("no NCF_TS")); 2483 n2_ts = __containerof(n2, struct namecache_ts, nc_nc); 2484 n2_ts->nc_time = ncp_ts->nc_time; 2485 n2_ts->nc_ticks = ncp_ts->nc_ticks; 2486 if (dtsp != NULL) { 2487 n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime; 2488 n2_ts->nc_nc.nc_flag |= NCF_DTS; 2489 } 2490 } 2491 #endif 2492 SDT_PROBE3(vfs, namecache, enter, duplicate, dvp, ncp->nc_name, 2493 vp); 2494 goto out_unlock_free; 2495 } 2496 } 2497 2498 if (flag == NCF_ISDOTDOT) { 2499 /* 2500 * See if we are trying to add .. entry, but some other lookup 2501 * has populated v_cache_dd pointer already. 2502 */ 2503 if (dvp->v_cache_dd != NULL) 2504 goto out_unlock_free; 2505 KASSERT(vp == NULL || vp->v_type == VDIR, 2506 ("wrong vnode type %p", vp)); 2507 atomic_thread_fence_rel(); 2508 atomic_store_ptr(&dvp->v_cache_dd, ncp); 2509 } 2510 2511 if (vp != NULL) { 2512 if (flag != NCF_ISDOTDOT) { 2513 /* 2514 * For this case, the cache entry maps both the 2515 * directory name in it and the name ".." for the 2516 * directory's parent. 2517 */ 2518 if ((ndd = vp->v_cache_dd) != NULL) { 2519 if ((ndd->nc_flag & NCF_ISDOTDOT) != 0) 2520 cache_zap_locked(ndd); 2521 else 2522 ndd = NULL; 2523 } 2524 atomic_thread_fence_rel(); 2525 atomic_store_ptr(&vp->v_cache_dd, ncp); 2526 } else if (vp->v_type != VDIR) { 2527 if (vp->v_cache_dd != NULL) { 2528 atomic_store_ptr(&vp->v_cache_dd, NULL); 2529 } 2530 } 2531 } 2532 2533 if (flag != NCF_ISDOTDOT) { 2534 if (LIST_EMPTY(&dvp->v_cache_src)) { 2535 cache_hold_vnode(dvp); 2536 } 2537 LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src); 2538 } 2539 2540 /* 2541 * If the entry is "negative", we place it into the 2542 * "negative" cache queue, otherwise, we place it into the 2543 * destination vnode's cache entries queue. 2544 */ 2545 if (vp != NULL) { 2546 TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst); 2547 SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name, 2548 vp); 2549 } else { 2550 if (cnp->cn_flags & ISWHITEOUT) 2551 atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_WHITE); 2552 cache_neg_insert(ncp); 2553 SDT_PROBE2(vfs, namecache, enter_negative, done, dvp, 2554 ncp->nc_name); 2555 } 2556 2557 /* 2558 * Insert the new namecache entry into the appropriate chain 2559 * within the cache entries table. 2560 */ 2561 CK_SLIST_INSERT_HEAD(ncpp, ncp, nc_hash); 2562 2563 atomic_thread_fence_rel(); 2564 /* 2565 * Mark the entry as fully constructed. 2566 * It is immutable past this point until its removal. 2567 */ 2568 atomic_store_char(&ncp->nc_flag, ncp->nc_flag & ~NCF_WIP); 2569 2570 cache_enter_unlock(&cel); 2571 if (ndd != NULL) 2572 cache_free(ndd); 2573 return; 2574 out_unlock_free: 2575 cache_enter_unlock(&cel); 2576 cache_free(ncp); 2577 return; 2578 } 2579 2580 /* 2581 * A variant of the above accepting flags. 2582 * 2583 * - VFS_CACHE_DROPOLD -- if a conflicting entry is found, drop it. 2584 * 2585 * TODO: this routine is a hack. It blindly removes the old entry, even if it 2586 * happens to match and it is doing it in an inefficient manner. It was added 2587 * to accommodate NFS which runs into a case where the target for a given name 2588 * may change from under it. Note this does nothing to solve the following 2589 * race: 2 callers of cache_enter_time_flags pass a different target vnode for 2590 * the same [dvp, cnp]. It may be argued that code doing this is broken. 2591 */ 2592 void 2593 cache_enter_time_flags(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, 2594 struct timespec *tsp, struct timespec *dtsp, int flags) 2595 { 2596 2597 MPASS((flags & ~(VFS_CACHE_DROPOLD)) == 0); 2598 2599 if (flags & VFS_CACHE_DROPOLD) 2600 cache_remove_cnp(dvp, cnp); 2601 cache_enter_time(dvp, vp, cnp, tsp, dtsp); 2602 } 2603 2604 static u_int 2605 cache_roundup_2(u_int val) 2606 { 2607 u_int res; 2608 2609 for (res = 1; res <= val; res <<= 1) 2610 continue; 2611 2612 return (res); 2613 } 2614 2615 static struct nchashhead * 2616 nchinittbl(u_long elements, u_long *hashmask) 2617 { 2618 struct nchashhead *hashtbl; 2619 u_long hashsize, i; 2620 2621 hashsize = cache_roundup_2(elements) / 2; 2622 2623 hashtbl = malloc((u_long)hashsize * sizeof(*hashtbl), M_VFSCACHE, M_WAITOK); 2624 for (i = 0; i < hashsize; i++) 2625 CK_SLIST_INIT(&hashtbl[i]); 2626 *hashmask = hashsize - 1; 2627 return (hashtbl); 2628 } 2629 2630 static void 2631 ncfreetbl(struct nchashhead *hashtbl) 2632 { 2633 2634 free(hashtbl, M_VFSCACHE); 2635 } 2636 2637 /* 2638 * Name cache initialization, from vfs_init() when we are booting 2639 */ 2640 static void 2641 nchinit(void *dummy __unused) 2642 { 2643 u_int i; 2644 2645 cache_zone_small = uma_zcreate("S VFS Cache", CACHE_ZONE_SMALL_SIZE, 2646 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT); 2647 cache_zone_small_ts = uma_zcreate("STS VFS Cache", CACHE_ZONE_SMALL_TS_SIZE, 2648 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT); 2649 cache_zone_large = uma_zcreate("L VFS Cache", CACHE_ZONE_LARGE_SIZE, 2650 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT); 2651 cache_zone_large_ts = uma_zcreate("LTS VFS Cache", CACHE_ZONE_LARGE_TS_SIZE, 2652 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT); 2653 2654 VFS_SMR_ZONE_SET(cache_zone_small); 2655 VFS_SMR_ZONE_SET(cache_zone_small_ts); 2656 VFS_SMR_ZONE_SET(cache_zone_large); 2657 VFS_SMR_ZONE_SET(cache_zone_large_ts); 2658 2659 ncsize = desiredvnodes * ncsizefactor; 2660 cache_recalc_neg_min(ncnegminpct); 2661 nchashtbl = nchinittbl(desiredvnodes * 2, &nchash); 2662 ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1; 2663 if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */ 2664 ncbuckethash = 7; 2665 if (ncbuckethash > nchash) 2666 ncbuckethash = nchash; 2667 bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE, 2668 M_WAITOK | M_ZERO); 2669 for (i = 0; i < numbucketlocks; i++) 2670 mtx_init(&bucketlocks[i], "ncbuc", NULL, MTX_DUPOK | MTX_RECURSE); 2671 ncvnodehash = ncbuckethash; 2672 vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE, 2673 M_WAITOK | M_ZERO); 2674 for (i = 0; i < numvnodelocks; i++) 2675 mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE); 2676 2677 for (i = 0; i < numneglists; i++) { 2678 mtx_init(&neglists[i].nl_evict_lock, "ncnege", NULL, MTX_DEF); 2679 mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF); 2680 TAILQ_INIT(&neglists[i].nl_list); 2681 TAILQ_INIT(&neglists[i].nl_hotlist); 2682 } 2683 } 2684 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL); 2685 2686 void 2687 cache_vnode_init(struct vnode *vp) 2688 { 2689 2690 LIST_INIT(&vp->v_cache_src); 2691 TAILQ_INIT(&vp->v_cache_dst); 2692 vp->v_cache_dd = NULL; 2693 cache_prehash(vp); 2694 } 2695 2696 /* 2697 * Induce transient cache misses for lockless operation in cache_lookup() by 2698 * using a temporary hash table. 2699 * 2700 * This will force a fs lookup. 2701 * 2702 * Synchronisation is done in 2 steps, calling vfs_smr_synchronize each time 2703 * to observe all CPUs not performing the lookup. 2704 */ 2705 static void 2706 cache_changesize_set_temp(struct nchashhead *temptbl, u_long temphash) 2707 { 2708 2709 MPASS(temphash < nchash); 2710 /* 2711 * Change the size. The new size is smaller and can safely be used 2712 * against the existing table. All lookups which now hash wrong will 2713 * result in a cache miss, which all callers are supposed to know how 2714 * to handle. 2715 */ 2716 atomic_store_long(&nchash, temphash); 2717 atomic_thread_fence_rel(); 2718 vfs_smr_synchronize(); 2719 /* 2720 * At this point everyone sees the updated hash value, but they still 2721 * see the old table. 2722 */ 2723 atomic_store_ptr(&nchashtbl, temptbl); 2724 atomic_thread_fence_rel(); 2725 vfs_smr_synchronize(); 2726 /* 2727 * At this point everyone sees the updated table pointer and size pair. 2728 */ 2729 } 2730 2731 /* 2732 * Set the new hash table. 2733 * 2734 * Similarly to cache_changesize_set_temp(), this has to synchronize against 2735 * lockless operation in cache_lookup(). 2736 */ 2737 static void 2738 cache_changesize_set_new(struct nchashhead *new_tbl, u_long new_hash) 2739 { 2740 2741 MPASS(nchash < new_hash); 2742 /* 2743 * Change the pointer first. This wont result in out of bounds access 2744 * since the temporary table is guaranteed to be smaller. 2745 */ 2746 atomic_store_ptr(&nchashtbl, new_tbl); 2747 atomic_thread_fence_rel(); 2748 vfs_smr_synchronize(); 2749 /* 2750 * At this point everyone sees the updated pointer value, but they 2751 * still see the old size. 2752 */ 2753 atomic_store_long(&nchash, new_hash); 2754 atomic_thread_fence_rel(); 2755 vfs_smr_synchronize(); 2756 /* 2757 * At this point everyone sees the updated table pointer and size pair. 2758 */ 2759 } 2760 2761 void 2762 cache_changesize(u_long newmaxvnodes) 2763 { 2764 struct nchashhead *new_nchashtbl, *old_nchashtbl, *temptbl; 2765 u_long new_nchash, old_nchash, temphash; 2766 struct namecache *ncp; 2767 uint32_t hash; 2768 u_long newncsize; 2769 int i; 2770 2771 newncsize = newmaxvnodes * ncsizefactor; 2772 newmaxvnodes = cache_roundup_2(newmaxvnodes * 2); 2773 if (newmaxvnodes < numbucketlocks) 2774 newmaxvnodes = numbucketlocks; 2775 2776 new_nchashtbl = nchinittbl(newmaxvnodes, &new_nchash); 2777 /* If same hash table size, nothing to do */ 2778 if (nchash == new_nchash) { 2779 ncfreetbl(new_nchashtbl); 2780 return; 2781 } 2782 2783 temptbl = nchinittbl(1, &temphash); 2784 2785 /* 2786 * Move everything from the old hash table to the new table. 2787 * None of the namecache entries in the table can be removed 2788 * because to do so, they have to be removed from the hash table. 2789 */ 2790 cache_lock_all_vnodes(); 2791 cache_lock_all_buckets(); 2792 old_nchashtbl = nchashtbl; 2793 old_nchash = nchash; 2794 cache_changesize_set_temp(temptbl, temphash); 2795 for (i = 0; i <= old_nchash; i++) { 2796 while ((ncp = CK_SLIST_FIRST(&old_nchashtbl[i])) != NULL) { 2797 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, 2798 ncp->nc_dvp); 2799 CK_SLIST_REMOVE(&old_nchashtbl[i], ncp, namecache, nc_hash); 2800 CK_SLIST_INSERT_HEAD(&new_nchashtbl[hash & new_nchash], ncp, nc_hash); 2801 } 2802 } 2803 ncsize = newncsize; 2804 cache_recalc_neg_min(ncnegminpct); 2805 cache_changesize_set_new(new_nchashtbl, new_nchash); 2806 cache_unlock_all_buckets(); 2807 cache_unlock_all_vnodes(); 2808 ncfreetbl(old_nchashtbl); 2809 ncfreetbl(temptbl); 2810 } 2811 2812 /* 2813 * Remove all entries from and to a particular vnode. 2814 */ 2815 static void 2816 cache_purge_impl(struct vnode *vp) 2817 { 2818 struct cache_freebatch batch; 2819 struct namecache *ncp; 2820 struct mtx *vlp, *vlp2; 2821 2822 TAILQ_INIT(&batch); 2823 vlp = VP2VNODELOCK(vp); 2824 vlp2 = NULL; 2825 mtx_lock(vlp); 2826 retry: 2827 while (!LIST_EMPTY(&vp->v_cache_src)) { 2828 ncp = LIST_FIRST(&vp->v_cache_src); 2829 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2)) 2830 goto retry; 2831 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst); 2832 } 2833 while (!TAILQ_EMPTY(&vp->v_cache_dst)) { 2834 ncp = TAILQ_FIRST(&vp->v_cache_dst); 2835 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2)) 2836 goto retry; 2837 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst); 2838 } 2839 ncp = vp->v_cache_dd; 2840 if (ncp != NULL) { 2841 KASSERT(ncp->nc_flag & NCF_ISDOTDOT, 2842 ("lost dotdot link")); 2843 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2)) 2844 goto retry; 2845 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst); 2846 } 2847 KASSERT(vp->v_cache_dd == NULL, ("incomplete purge")); 2848 mtx_unlock(vlp); 2849 if (vlp2 != NULL) 2850 mtx_unlock(vlp2); 2851 cache_free_batch(&batch); 2852 } 2853 2854 /* 2855 * Opportunistic check to see if there is anything to do. 2856 */ 2857 static bool 2858 cache_has_entries(struct vnode *vp) 2859 { 2860 2861 if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) && 2862 atomic_load_ptr(&vp->v_cache_dd) == NULL) 2863 return (false); 2864 return (true); 2865 } 2866 2867 void 2868 cache_purge(struct vnode *vp) 2869 { 2870 2871 SDT_PROBE1(vfs, namecache, purge, done, vp); 2872 if (!cache_has_entries(vp)) 2873 return; 2874 cache_purge_impl(vp); 2875 } 2876 2877 /* 2878 * Only to be used by vgone. 2879 */ 2880 void 2881 cache_purge_vgone(struct vnode *vp) 2882 { 2883 struct mtx *vlp; 2884 2885 VNPASS(VN_IS_DOOMED(vp), vp); 2886 if (cache_has_entries(vp)) { 2887 cache_purge_impl(vp); 2888 return; 2889 } 2890 2891 /* 2892 * Serialize against a potential thread doing cache_purge. 2893 */ 2894 vlp = VP2VNODELOCK(vp); 2895 mtx_wait_unlocked(vlp); 2896 if (cache_has_entries(vp)) { 2897 cache_purge_impl(vp); 2898 return; 2899 } 2900 return; 2901 } 2902 2903 /* 2904 * Remove all negative entries for a particular directory vnode. 2905 */ 2906 void 2907 cache_purge_negative(struct vnode *vp) 2908 { 2909 struct cache_freebatch batch; 2910 struct namecache *ncp, *nnp; 2911 struct mtx *vlp; 2912 2913 SDT_PROBE1(vfs, namecache, purge_negative, done, vp); 2914 if (LIST_EMPTY(&vp->v_cache_src)) 2915 return; 2916 TAILQ_INIT(&batch); 2917 vlp = VP2VNODELOCK(vp); 2918 mtx_lock(vlp); 2919 LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) { 2920 if (!(ncp->nc_flag & NCF_NEGATIVE)) 2921 continue; 2922 cache_zap_negative_locked_vnode_kl(ncp, vp); 2923 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst); 2924 } 2925 mtx_unlock(vlp); 2926 cache_free_batch(&batch); 2927 } 2928 2929 /* 2930 * Entry points for modifying VOP operations. 2931 */ 2932 void 2933 cache_vop_rename(struct vnode *fdvp, struct vnode *fvp, struct vnode *tdvp, 2934 struct vnode *tvp, struct componentname *fcnp, struct componentname *tcnp) 2935 { 2936 2937 ASSERT_VOP_IN_SEQC(fdvp); 2938 ASSERT_VOP_IN_SEQC(fvp); 2939 ASSERT_VOP_IN_SEQC(tdvp); 2940 if (tvp != NULL) 2941 ASSERT_VOP_IN_SEQC(tvp); 2942 2943 cache_purge(fvp); 2944 if (tvp != NULL) { 2945 cache_purge(tvp); 2946 KASSERT(!cache_remove_cnp(tdvp, tcnp), 2947 ("%s: lingering negative entry", __func__)); 2948 } else { 2949 cache_remove_cnp(tdvp, tcnp); 2950 } 2951 2952 /* 2953 * TODO 2954 * 2955 * Historically renaming was always purging all revelang entries, 2956 * but that's quite wasteful. In particular turns out that in many cases 2957 * the target file is immediately accessed after rename, inducing a cache 2958 * miss. 2959 * 2960 * Recode this to reduce relocking and reuse the existing entry (if any) 2961 * instead of just removing it above and allocating a new one here. 2962 */ 2963 cache_enter(tdvp, fvp, tcnp); 2964 } 2965 2966 void 2967 cache_vop_rmdir(struct vnode *dvp, struct vnode *vp) 2968 { 2969 2970 ASSERT_VOP_IN_SEQC(dvp); 2971 ASSERT_VOP_IN_SEQC(vp); 2972 cache_purge(vp); 2973 } 2974 2975 #ifdef INVARIANTS 2976 /* 2977 * Validate that if an entry exists it matches. 2978 */ 2979 void 2980 cache_validate(struct vnode *dvp, struct vnode *vp, struct componentname *cnp) 2981 { 2982 struct namecache *ncp; 2983 struct mtx *blp; 2984 uint32_t hash; 2985 2986 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); 2987 if (CK_SLIST_EMPTY(NCHHASH(hash))) 2988 return; 2989 blp = HASH2BUCKETLOCK(hash); 2990 mtx_lock(blp); 2991 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 2992 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && 2993 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) { 2994 if (ncp->nc_vp != vp) 2995 panic("%s: mismatch (%p != %p); ncp %p [%s] dvp %p\n", 2996 __func__, vp, ncp->nc_vp, ncp, ncp->nc_name, ncp->nc_dvp); 2997 } 2998 } 2999 mtx_unlock(blp); 3000 } 3001 #endif 3002 3003 /* 3004 * Flush all entries referencing a particular filesystem. 3005 */ 3006 void 3007 cache_purgevfs(struct mount *mp) 3008 { 3009 struct vnode *vp, *mvp; 3010 size_t visited, purged; 3011 3012 visited = purged = 0; 3013 /* 3014 * Somewhat wasteful iteration over all vnodes. Would be better to 3015 * support filtering and avoid the interlock to begin with. 3016 */ 3017 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 3018 visited++; 3019 if (!cache_has_entries(vp)) { 3020 VI_UNLOCK(vp); 3021 continue; 3022 } 3023 vholdl(vp); 3024 VI_UNLOCK(vp); 3025 cache_purge(vp); 3026 purged++; 3027 vdrop(vp); 3028 } 3029 3030 SDT_PROBE3(vfs, namecache, purgevfs, done, mp, visited, purged); 3031 } 3032 3033 /* 3034 * Perform canonical checks and cache lookup and pass on to filesystem 3035 * through the vop_cachedlookup only if needed. 3036 */ 3037 3038 int 3039 vfs_cache_lookup(struct vop_lookup_args *ap) 3040 { 3041 struct vnode *dvp; 3042 int error; 3043 struct vnode **vpp = ap->a_vpp; 3044 struct componentname *cnp = ap->a_cnp; 3045 int flags = cnp->cn_flags; 3046 3047 *vpp = NULL; 3048 dvp = ap->a_dvp; 3049 3050 if (dvp->v_type != VDIR) 3051 return (ENOTDIR); 3052 3053 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) && 3054 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) 3055 return (EROFS); 3056 3057 error = vn_dir_check_exec(dvp, cnp); 3058 if (error != 0) 3059 return (error); 3060 3061 error = cache_lookup(dvp, vpp, cnp, NULL, NULL); 3062 if (error == 0) 3063 return (VOP_CACHEDLOOKUP(dvp, vpp, cnp)); 3064 if (error == -1) 3065 return (0); 3066 return (error); 3067 } 3068 3069 /* Implementation of the getcwd syscall. */ 3070 int 3071 sys___getcwd(struct thread *td, struct __getcwd_args *uap) 3072 { 3073 char *buf, *retbuf; 3074 size_t buflen; 3075 int error; 3076 3077 buflen = uap->buflen; 3078 if (__predict_false(buflen < 2)) 3079 return (EINVAL); 3080 if (buflen > MAXPATHLEN) 3081 buflen = MAXPATHLEN; 3082 3083 buf = uma_zalloc(namei_zone, M_WAITOK); 3084 error = vn_getcwd(buf, &retbuf, &buflen); 3085 if (error == 0) 3086 error = copyout(retbuf, uap->buf, buflen); 3087 uma_zfree(namei_zone, buf); 3088 return (error); 3089 } 3090 3091 int 3092 vn_getcwd(char *buf, char **retbuf, size_t *buflen) 3093 { 3094 struct pwd *pwd; 3095 int error; 3096 3097 vfs_smr_enter(); 3098 pwd = pwd_get_smr(); 3099 error = vn_fullpath_any_smr(pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf, 3100 buflen, 0); 3101 VFS_SMR_ASSERT_NOT_ENTERED(); 3102 if (error < 0) { 3103 pwd = pwd_hold(curthread); 3104 error = vn_fullpath_any(pwd->pwd_cdir, pwd->pwd_rdir, buf, 3105 retbuf, buflen); 3106 pwd_drop(pwd); 3107 } 3108 3109 #ifdef KTRACE 3110 if (KTRPOINT(curthread, KTR_NAMEI) && error == 0) 3111 ktrnamei(*retbuf); 3112 #endif 3113 return (error); 3114 } 3115 3116 /* 3117 * Canonicalize a path by walking it forward and back. 3118 * 3119 * BUGS: 3120 * - Nothing guarantees the integrity of the entire chain. Consider the case 3121 * where the path "foo/bar/baz/qux" is passed, but "bar" is moved out of 3122 * "foo" into "quux" during the backwards walk. The result will be 3123 * "quux/bar/baz/qux", which could not have been obtained by an incremental 3124 * walk in userspace. Moreover, the path we return is inaccessible if the 3125 * calling thread lacks permission to traverse "quux". 3126 */ 3127 static int 3128 kern___realpathat(struct thread *td, int fd, const char *path, char *buf, 3129 size_t size, int flags, enum uio_seg pathseg) 3130 { 3131 struct nameidata nd; 3132 char *retbuf, *freebuf; 3133 int error; 3134 3135 if (flags != 0) 3136 return (EINVAL); 3137 NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | SAVENAME | WANTPARENT | AUDITVNODE1, 3138 pathseg, path, fd, &cap_fstat_rights); 3139 if ((error = namei(&nd)) != 0) 3140 return (error); 3141 error = vn_fullpath_hardlink(nd.ni_vp, nd.ni_dvp, nd.ni_cnd.cn_nameptr, 3142 nd.ni_cnd.cn_namelen, &retbuf, &freebuf, &size); 3143 if (error == 0) { 3144 error = copyout(retbuf, buf, size); 3145 free(freebuf, M_TEMP); 3146 } 3147 NDFREE(&nd, 0); 3148 return (error); 3149 } 3150 3151 int 3152 sys___realpathat(struct thread *td, struct __realpathat_args *uap) 3153 { 3154 3155 return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size, 3156 uap->flags, UIO_USERSPACE)); 3157 } 3158 3159 /* 3160 * Retrieve the full filesystem path that correspond to a vnode from the name 3161 * cache (if available) 3162 */ 3163 int 3164 vn_fullpath(struct vnode *vp, char **retbuf, char **freebuf) 3165 { 3166 struct pwd *pwd; 3167 char *buf; 3168 size_t buflen; 3169 int error; 3170 3171 if (__predict_false(vp == NULL)) 3172 return (EINVAL); 3173 3174 buflen = MAXPATHLEN; 3175 buf = malloc(buflen, M_TEMP, M_WAITOK); 3176 vfs_smr_enter(); 3177 pwd = pwd_get_smr(); 3178 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, &buflen, 0); 3179 VFS_SMR_ASSERT_NOT_ENTERED(); 3180 if (error < 0) { 3181 pwd = pwd_hold(curthread); 3182 error = vn_fullpath_any(vp, pwd->pwd_rdir, buf, retbuf, &buflen); 3183 pwd_drop(pwd); 3184 } 3185 if (error == 0) 3186 *freebuf = buf; 3187 else 3188 free(buf, M_TEMP); 3189 return (error); 3190 } 3191 3192 /* 3193 * This function is similar to vn_fullpath, but it attempts to lookup the 3194 * pathname relative to the global root mount point. This is required for the 3195 * auditing sub-system, as audited pathnames must be absolute, relative to the 3196 * global root mount point. 3197 */ 3198 int 3199 vn_fullpath_global(struct vnode *vp, char **retbuf, char **freebuf) 3200 { 3201 char *buf; 3202 size_t buflen; 3203 int error; 3204 3205 if (__predict_false(vp == NULL)) 3206 return (EINVAL); 3207 buflen = MAXPATHLEN; 3208 buf = malloc(buflen, M_TEMP, M_WAITOK); 3209 vfs_smr_enter(); 3210 error = vn_fullpath_any_smr(vp, rootvnode, buf, retbuf, &buflen, 0); 3211 VFS_SMR_ASSERT_NOT_ENTERED(); 3212 if (error < 0) { 3213 error = vn_fullpath_any(vp, rootvnode, buf, retbuf, &buflen); 3214 } 3215 if (error == 0) 3216 *freebuf = buf; 3217 else 3218 free(buf, M_TEMP); 3219 return (error); 3220 } 3221 3222 static struct namecache * 3223 vn_dd_from_dst(struct vnode *vp) 3224 { 3225 struct namecache *ncp; 3226 3227 cache_assert_vnode_locked(vp); 3228 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) { 3229 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) 3230 return (ncp); 3231 } 3232 return (NULL); 3233 } 3234 3235 int 3236 vn_vptocnp(struct vnode **vp, char *buf, size_t *buflen) 3237 { 3238 struct vnode *dvp; 3239 struct namecache *ncp; 3240 struct mtx *vlp; 3241 int error; 3242 3243 vlp = VP2VNODELOCK(*vp); 3244 mtx_lock(vlp); 3245 ncp = (*vp)->v_cache_dd; 3246 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT) == 0) { 3247 KASSERT(ncp == vn_dd_from_dst(*vp), 3248 ("%s: mismatch for dd entry (%p != %p)", __func__, 3249 ncp, vn_dd_from_dst(*vp))); 3250 } else { 3251 ncp = vn_dd_from_dst(*vp); 3252 } 3253 if (ncp != NULL) { 3254 if (*buflen < ncp->nc_nlen) { 3255 mtx_unlock(vlp); 3256 vrele(*vp); 3257 counter_u64_add(numfullpathfail4, 1); 3258 error = ENOMEM; 3259 SDT_PROBE3(vfs, namecache, fullpath, return, error, 3260 vp, NULL); 3261 return (error); 3262 } 3263 *buflen -= ncp->nc_nlen; 3264 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen); 3265 SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp, 3266 ncp->nc_name, vp); 3267 dvp = *vp; 3268 *vp = ncp->nc_dvp; 3269 vref(*vp); 3270 mtx_unlock(vlp); 3271 vrele(dvp); 3272 return (0); 3273 } 3274 SDT_PROBE1(vfs, namecache, fullpath, miss, vp); 3275 3276 mtx_unlock(vlp); 3277 vn_lock(*vp, LK_SHARED | LK_RETRY); 3278 error = VOP_VPTOCNP(*vp, &dvp, buf, buflen); 3279 vput(*vp); 3280 if (error) { 3281 counter_u64_add(numfullpathfail2, 1); 3282 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); 3283 return (error); 3284 } 3285 3286 *vp = dvp; 3287 if (VN_IS_DOOMED(dvp)) { 3288 /* forced unmount */ 3289 vrele(dvp); 3290 error = ENOENT; 3291 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); 3292 return (error); 3293 } 3294 /* 3295 * *vp has its use count incremented still. 3296 */ 3297 3298 return (0); 3299 } 3300 3301 /* 3302 * Resolve a directory to a pathname. 3303 * 3304 * The name of the directory can always be found in the namecache or fetched 3305 * from the filesystem. There is also guaranteed to be only one parent, meaning 3306 * we can just follow vnodes up until we find the root. 3307 * 3308 * The vnode must be referenced. 3309 */ 3310 static int 3311 vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf, 3312 size_t *len, size_t addend) 3313 { 3314 #ifdef KDTRACE_HOOKS 3315 struct vnode *startvp = vp; 3316 #endif 3317 struct vnode *vp1; 3318 size_t buflen; 3319 int error; 3320 bool slash_prefixed; 3321 3322 VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp); 3323 VNPASS(vp->v_usecount > 0, vp); 3324 3325 buflen = *len; 3326 3327 slash_prefixed = true; 3328 if (addend == 0) { 3329 MPASS(*len >= 2); 3330 buflen--; 3331 buf[buflen] = '\0'; 3332 slash_prefixed = false; 3333 } 3334 3335 error = 0; 3336 3337 SDT_PROBE1(vfs, namecache, fullpath, entry, vp); 3338 counter_u64_add(numfullpathcalls, 1); 3339 while (vp != rdir && vp != rootvnode) { 3340 /* 3341 * The vp vnode must be already fully constructed, 3342 * since it is either found in namecache or obtained 3343 * from VOP_VPTOCNP(). We may test for VV_ROOT safely 3344 * without obtaining the vnode lock. 3345 */ 3346 if ((vp->v_vflag & VV_ROOT) != 0) { 3347 vn_lock(vp, LK_RETRY | LK_SHARED); 3348 3349 /* 3350 * With the vnode locked, check for races with 3351 * unmount, forced or not. Note that we 3352 * already verified that vp is not equal to 3353 * the root vnode, which means that 3354 * mnt_vnodecovered can be NULL only for the 3355 * case of unmount. 3356 */ 3357 if (VN_IS_DOOMED(vp) || 3358 (vp1 = vp->v_mount->mnt_vnodecovered) == NULL || 3359 vp1->v_mountedhere != vp->v_mount) { 3360 vput(vp); 3361 error = ENOENT; 3362 SDT_PROBE3(vfs, namecache, fullpath, return, 3363 error, vp, NULL); 3364 break; 3365 } 3366 3367 vref(vp1); 3368 vput(vp); 3369 vp = vp1; 3370 continue; 3371 } 3372 if (vp->v_type != VDIR) { 3373 vrele(vp); 3374 counter_u64_add(numfullpathfail1, 1); 3375 error = ENOTDIR; 3376 SDT_PROBE3(vfs, namecache, fullpath, return, 3377 error, vp, NULL); 3378 break; 3379 } 3380 error = vn_vptocnp(&vp, buf, &buflen); 3381 if (error) 3382 break; 3383 if (buflen == 0) { 3384 vrele(vp); 3385 error = ENOMEM; 3386 SDT_PROBE3(vfs, namecache, fullpath, return, error, 3387 startvp, NULL); 3388 break; 3389 } 3390 buf[--buflen] = '/'; 3391 slash_prefixed = true; 3392 } 3393 if (error) 3394 return (error); 3395 if (!slash_prefixed) { 3396 if (buflen == 0) { 3397 vrele(vp); 3398 counter_u64_add(numfullpathfail4, 1); 3399 SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM, 3400 startvp, NULL); 3401 return (ENOMEM); 3402 } 3403 buf[--buflen] = '/'; 3404 } 3405 counter_u64_add(numfullpathfound, 1); 3406 vrele(vp); 3407 3408 *retbuf = buf + buflen; 3409 SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf); 3410 *len -= buflen; 3411 *len += addend; 3412 return (0); 3413 } 3414 3415 /* 3416 * Resolve an arbitrary vnode to a pathname. 3417 * 3418 * Note 2 caveats: 3419 * - hardlinks are not tracked, thus if the vnode is not a directory this can 3420 * resolve to a different path than the one used to find it 3421 * - namecache is not mandatory, meaning names are not guaranteed to be added 3422 * (in which case resolving fails) 3423 */ 3424 static void __inline 3425 cache_rev_failed_impl(int *reason, int line) 3426 { 3427 3428 *reason = line; 3429 } 3430 #define cache_rev_failed(var) cache_rev_failed_impl((var), __LINE__) 3431 3432 static int 3433 vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf, 3434 char **retbuf, size_t *buflen, size_t addend) 3435 { 3436 #ifdef KDTRACE_HOOKS 3437 struct vnode *startvp = vp; 3438 #endif 3439 struct vnode *tvp; 3440 struct mount *mp; 3441 struct namecache *ncp; 3442 size_t orig_buflen; 3443 int reason; 3444 int error; 3445 #ifdef KDTRACE_HOOKS 3446 int i; 3447 #endif 3448 seqc_t vp_seqc, tvp_seqc; 3449 u_char nc_flag; 3450 3451 VFS_SMR_ASSERT_ENTERED(); 3452 3453 if (!atomic_load_char(&cache_fast_lookup_enabled)) { 3454 vfs_smr_exit(); 3455 return (-1); 3456 } 3457 3458 orig_buflen = *buflen; 3459 3460 if (addend == 0) { 3461 MPASS(*buflen >= 2); 3462 *buflen -= 1; 3463 buf[*buflen] = '\0'; 3464 } 3465 3466 if (vp == rdir || vp == rootvnode) { 3467 if (addend == 0) { 3468 *buflen -= 1; 3469 buf[*buflen] = '/'; 3470 } 3471 goto out_ok; 3472 } 3473 3474 #ifdef KDTRACE_HOOKS 3475 i = 0; 3476 #endif 3477 error = -1; 3478 ncp = NULL; /* for sdt probe down below */ 3479 vp_seqc = vn_seqc_read_any(vp); 3480 if (seqc_in_modify(vp_seqc)) { 3481 cache_rev_failed(&reason); 3482 goto out_abort; 3483 } 3484 3485 for (;;) { 3486 #ifdef KDTRACE_HOOKS 3487 i++; 3488 #endif 3489 if ((vp->v_vflag & VV_ROOT) != 0) { 3490 mp = atomic_load_ptr(&vp->v_mount); 3491 if (mp == NULL) { 3492 cache_rev_failed(&reason); 3493 goto out_abort; 3494 } 3495 tvp = atomic_load_ptr(&mp->mnt_vnodecovered); 3496 tvp_seqc = vn_seqc_read_any(tvp); 3497 if (seqc_in_modify(tvp_seqc)) { 3498 cache_rev_failed(&reason); 3499 goto out_abort; 3500 } 3501 if (!vn_seqc_consistent(vp, vp_seqc)) { 3502 cache_rev_failed(&reason); 3503 goto out_abort; 3504 } 3505 vp = tvp; 3506 vp_seqc = tvp_seqc; 3507 continue; 3508 } 3509 ncp = atomic_load_consume_ptr(&vp->v_cache_dd); 3510 if (ncp == NULL) { 3511 cache_rev_failed(&reason); 3512 goto out_abort; 3513 } 3514 nc_flag = atomic_load_char(&ncp->nc_flag); 3515 if ((nc_flag & NCF_ISDOTDOT) != 0) { 3516 cache_rev_failed(&reason); 3517 goto out_abort; 3518 } 3519 if (ncp->nc_nlen >= *buflen) { 3520 cache_rev_failed(&reason); 3521 error = ENOMEM; 3522 goto out_abort; 3523 } 3524 *buflen -= ncp->nc_nlen; 3525 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen); 3526 *buflen -= 1; 3527 buf[*buflen] = '/'; 3528 tvp = ncp->nc_dvp; 3529 tvp_seqc = vn_seqc_read_any(tvp); 3530 if (seqc_in_modify(tvp_seqc)) { 3531 cache_rev_failed(&reason); 3532 goto out_abort; 3533 } 3534 if (!vn_seqc_consistent(vp, vp_seqc)) { 3535 cache_rev_failed(&reason); 3536 goto out_abort; 3537 } 3538 /* 3539 * Acquire fence provided by vn_seqc_read_any above. 3540 */ 3541 if (__predict_false(atomic_load_ptr(&vp->v_cache_dd) != ncp)) { 3542 cache_rev_failed(&reason); 3543 goto out_abort; 3544 } 3545 if (!cache_ncp_canuse(ncp)) { 3546 cache_rev_failed(&reason); 3547 goto out_abort; 3548 } 3549 vp = tvp; 3550 vp_seqc = tvp_seqc; 3551 if (vp == rdir || vp == rootvnode) 3552 break; 3553 } 3554 out_ok: 3555 vfs_smr_exit(); 3556 *retbuf = buf + *buflen; 3557 *buflen = orig_buflen - *buflen + addend; 3558 SDT_PROBE2(vfs, namecache, fullpath_smr, hit, startvp, *retbuf); 3559 return (0); 3560 3561 out_abort: 3562 *buflen = orig_buflen; 3563 SDT_PROBE4(vfs, namecache, fullpath_smr, miss, startvp, ncp, reason, i); 3564 vfs_smr_exit(); 3565 return (error); 3566 } 3567 3568 static int 3569 vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf, 3570 size_t *buflen) 3571 { 3572 size_t orig_buflen, addend; 3573 int error; 3574 3575 if (*buflen < 2) 3576 return (EINVAL); 3577 3578 orig_buflen = *buflen; 3579 3580 vref(vp); 3581 addend = 0; 3582 if (vp->v_type != VDIR) { 3583 *buflen -= 1; 3584 buf[*buflen] = '\0'; 3585 error = vn_vptocnp(&vp, buf, buflen); 3586 if (error) 3587 return (error); 3588 if (*buflen == 0) { 3589 vrele(vp); 3590 return (ENOMEM); 3591 } 3592 *buflen -= 1; 3593 buf[*buflen] = '/'; 3594 addend = orig_buflen - *buflen; 3595 } 3596 3597 return (vn_fullpath_dir(vp, rdir, buf, retbuf, buflen, addend)); 3598 } 3599 3600 /* 3601 * Resolve an arbitrary vnode to a pathname (taking care of hardlinks). 3602 * 3603 * Since the namecache does not track hardlinks, the caller is 3604 * expected to first look up the target vnode with SAVENAME | 3605 * WANTPARENT flags passed to namei to get dvp and vp. 3606 * 3607 * Then we have 2 cases: 3608 * - if the found vnode is a directory, the path can be constructed just by 3609 * following names up the chain 3610 * - otherwise we populate the buffer with the saved name and start resolving 3611 * from the parent 3612 */ 3613 int 3614 vn_fullpath_hardlink(struct vnode *vp, struct vnode *dvp, 3615 const char *hrdl_name, size_t hrdl_name_length, 3616 char **retbuf, char **freebuf, size_t *buflen) 3617 { 3618 char *buf, *tmpbuf; 3619 struct pwd *pwd; 3620 size_t addend; 3621 int error; 3622 enum vtype type; 3623 3624 if (*buflen < 2) 3625 return (EINVAL); 3626 if (*buflen > MAXPATHLEN) 3627 *buflen = MAXPATHLEN; 3628 3629 buf = malloc(*buflen, M_TEMP, M_WAITOK); 3630 3631 addend = 0; 3632 3633 /* 3634 * Check for VBAD to work around the vp_crossmp bug in lookup(). 3635 * 3636 * For example consider tmpfs on /tmp and realpath /tmp. ni_vp will be 3637 * set to mount point's root vnode while ni_dvp will be vp_crossmp. 3638 * If the type is VDIR (like in this very case) we can skip looking 3639 * at ni_dvp in the first place. However, since vnodes get passed here 3640 * unlocked the target may transition to doomed state (type == VBAD) 3641 * before we get to evaluate the condition. If this happens, we will 3642 * populate part of the buffer and descend to vn_fullpath_dir with 3643 * vp == vp_crossmp. Prevent the problem by checking for VBAD. 3644 * 3645 * This should be atomic_load(&vp->v_type) but it is illegal to take 3646 * an address of a bit field, even if said field is sized to char. 3647 * Work around the problem by reading the value into a full-sized enum 3648 * and then re-reading it with atomic_load which will still prevent 3649 * the compiler from re-reading down the road. 3650 */ 3651 type = vp->v_type; 3652 type = atomic_load_int(&type); 3653 if (type == VBAD) { 3654 error = ENOENT; 3655 goto out_bad; 3656 } 3657 if (type != VDIR) { 3658 addend = hrdl_name_length + 2; 3659 if (*buflen < addend) { 3660 error = ENOMEM; 3661 goto out_bad; 3662 } 3663 *buflen -= addend; 3664 tmpbuf = buf + *buflen; 3665 tmpbuf[0] = '/'; 3666 memcpy(&tmpbuf[1], hrdl_name, hrdl_name_length); 3667 tmpbuf[addend - 1] = '\0'; 3668 vp = dvp; 3669 } 3670 3671 vfs_smr_enter(); 3672 pwd = pwd_get_smr(); 3673 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, buflen, 3674 addend); 3675 VFS_SMR_ASSERT_NOT_ENTERED(); 3676 if (error < 0) { 3677 pwd = pwd_hold(curthread); 3678 vref(vp); 3679 error = vn_fullpath_dir(vp, pwd->pwd_rdir, buf, retbuf, buflen, 3680 addend); 3681 pwd_drop(pwd); 3682 } 3683 if (error != 0) 3684 goto out_bad; 3685 3686 *freebuf = buf; 3687 3688 return (0); 3689 out_bad: 3690 free(buf, M_TEMP); 3691 return (error); 3692 } 3693 3694 struct vnode * 3695 vn_dir_dd_ino(struct vnode *vp) 3696 { 3697 struct namecache *ncp; 3698 struct vnode *ddvp; 3699 struct mtx *vlp; 3700 enum vgetstate vs; 3701 3702 ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino"); 3703 vlp = VP2VNODELOCK(vp); 3704 mtx_lock(vlp); 3705 TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) { 3706 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) 3707 continue; 3708 ddvp = ncp->nc_dvp; 3709 vs = vget_prep(ddvp); 3710 mtx_unlock(vlp); 3711 if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs)) 3712 return (NULL); 3713 return (ddvp); 3714 } 3715 mtx_unlock(vlp); 3716 return (NULL); 3717 } 3718 3719 int 3720 vn_commname(struct vnode *vp, char *buf, u_int buflen) 3721 { 3722 struct namecache *ncp; 3723 struct mtx *vlp; 3724 int l; 3725 3726 vlp = VP2VNODELOCK(vp); 3727 mtx_lock(vlp); 3728 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) 3729 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) 3730 break; 3731 if (ncp == NULL) { 3732 mtx_unlock(vlp); 3733 return (ENOENT); 3734 } 3735 l = min(ncp->nc_nlen, buflen - 1); 3736 memcpy(buf, ncp->nc_name, l); 3737 mtx_unlock(vlp); 3738 buf[l] = '\0'; 3739 return (0); 3740 } 3741 3742 /* 3743 * This function updates path string to vnode's full global path 3744 * and checks the size of the new path string against the pathlen argument. 3745 * 3746 * Requires a locked, referenced vnode. 3747 * Vnode is re-locked on success or ENODEV, otherwise unlocked. 3748 * 3749 * If vp is a directory, the call to vn_fullpath_global() always succeeds 3750 * because it falls back to the ".." lookup if the namecache lookup fails. 3751 */ 3752 int 3753 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path, 3754 u_int pathlen) 3755 { 3756 struct nameidata nd; 3757 struct vnode *vp1; 3758 char *rpath, *fbuf; 3759 int error; 3760 3761 ASSERT_VOP_ELOCKED(vp, __func__); 3762 3763 /* Construct global filesystem path from vp. */ 3764 VOP_UNLOCK(vp); 3765 error = vn_fullpath_global(vp, &rpath, &fbuf); 3766 3767 if (error != 0) { 3768 vrele(vp); 3769 return (error); 3770 } 3771 3772 if (strlen(rpath) >= pathlen) { 3773 vrele(vp); 3774 error = ENAMETOOLONG; 3775 goto out; 3776 } 3777 3778 /* 3779 * Re-lookup the vnode by path to detect a possible rename. 3780 * As a side effect, the vnode is relocked. 3781 * If vnode was renamed, return ENOENT. 3782 */ 3783 NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1, UIO_SYSSPACE, path); 3784 error = namei(&nd); 3785 if (error != 0) { 3786 vrele(vp); 3787 goto out; 3788 } 3789 NDFREE_PNBUF(&nd); 3790 vp1 = nd.ni_vp; 3791 vrele(vp); 3792 if (vp1 == vp) 3793 strcpy(path, rpath); 3794 else { 3795 vput(vp1); 3796 error = ENOENT; 3797 } 3798 3799 out: 3800 free(fbuf, M_TEMP); 3801 return (error); 3802 } 3803 3804 #ifdef DDB 3805 static void 3806 db_print_vpath(struct vnode *vp) 3807 { 3808 3809 while (vp != NULL) { 3810 db_printf("%p: ", vp); 3811 if (vp == rootvnode) { 3812 db_printf("/"); 3813 vp = NULL; 3814 } else { 3815 if (vp->v_vflag & VV_ROOT) { 3816 db_printf("<mount point>"); 3817 vp = vp->v_mount->mnt_vnodecovered; 3818 } else { 3819 struct namecache *ncp; 3820 char *ncn; 3821 int i; 3822 3823 ncp = TAILQ_FIRST(&vp->v_cache_dst); 3824 if (ncp != NULL) { 3825 ncn = ncp->nc_name; 3826 for (i = 0; i < ncp->nc_nlen; i++) 3827 db_printf("%c", *ncn++); 3828 vp = ncp->nc_dvp; 3829 } else { 3830 vp = NULL; 3831 } 3832 } 3833 } 3834 db_printf("\n"); 3835 } 3836 3837 return; 3838 } 3839 3840 DB_SHOW_COMMAND(vpath, db_show_vpath) 3841 { 3842 struct vnode *vp; 3843 3844 if (!have_addr) { 3845 db_printf("usage: show vpath <struct vnode *>\n"); 3846 return; 3847 } 3848 3849 vp = (struct vnode *)addr; 3850 db_print_vpath(vp); 3851 } 3852 3853 #endif 3854 3855 static int cache_fast_lookup = 1; 3856 3857 #define CACHE_FPL_FAILED -2020 3858 3859 void 3860 cache_fast_lookup_enabled_recalc(void) 3861 { 3862 int lookup_flag; 3863 int mac_on; 3864 3865 #ifdef MAC 3866 mac_on = mac_vnode_check_lookup_enabled(); 3867 mac_on |= mac_vnode_check_readlink_enabled(); 3868 #else 3869 mac_on = 0; 3870 #endif 3871 3872 lookup_flag = atomic_load_int(&cache_fast_lookup); 3873 if (lookup_flag && !mac_on) { 3874 atomic_store_char(&cache_fast_lookup_enabled, true); 3875 } else { 3876 atomic_store_char(&cache_fast_lookup_enabled, false); 3877 } 3878 } 3879 3880 static int 3881 syscal_vfs_cache_fast_lookup(SYSCTL_HANDLER_ARGS) 3882 { 3883 int error, old; 3884 3885 old = atomic_load_int(&cache_fast_lookup); 3886 error = sysctl_handle_int(oidp, arg1, arg2, req); 3887 if (error == 0 && req->newptr && old != atomic_load_int(&cache_fast_lookup)) 3888 cache_fast_lookup_enabled_recalc(); 3889 return (error); 3890 } 3891 SYSCTL_PROC(_vfs, OID_AUTO, cache_fast_lookup, CTLTYPE_INT|CTLFLAG_RW|CTLFLAG_MPSAFE, 3892 &cache_fast_lookup, 0, syscal_vfs_cache_fast_lookup, "IU", ""); 3893 3894 /* 3895 * Components of nameidata (or objects it can point to) which may 3896 * need restoring in case fast path lookup fails. 3897 */ 3898 struct nameidata_outer { 3899 size_t ni_pathlen; 3900 int cn_flags; 3901 }; 3902 3903 struct nameidata_saved { 3904 #ifdef INVARIANTS 3905 char *cn_nameptr; 3906 size_t ni_pathlen; 3907 #endif 3908 }; 3909 3910 #ifdef INVARIANTS 3911 struct cache_fpl_debug { 3912 size_t ni_pathlen; 3913 }; 3914 #endif 3915 3916 struct cache_fpl { 3917 struct nameidata *ndp; 3918 struct componentname *cnp; 3919 char *nulchar; 3920 struct vnode *dvp; 3921 struct vnode *tvp; 3922 seqc_t dvp_seqc; 3923 seqc_t tvp_seqc; 3924 uint32_t hash; 3925 struct nameidata_saved snd; 3926 struct nameidata_outer snd_outer; 3927 int line; 3928 enum cache_fpl_status status:8; 3929 bool in_smr; 3930 bool fsearch; 3931 bool savename; 3932 struct pwd **pwd; 3933 #ifdef INVARIANTS 3934 struct cache_fpl_debug debug; 3935 #endif 3936 }; 3937 3938 static bool cache_fplookup_mp_supported(struct mount *mp); 3939 static bool cache_fplookup_is_mp(struct cache_fpl *fpl); 3940 static int cache_fplookup_cross_mount(struct cache_fpl *fpl); 3941 static int cache_fplookup_partial_setup(struct cache_fpl *fpl); 3942 static int cache_fplookup_skip_slashes(struct cache_fpl *fpl); 3943 static int cache_fplookup_trailingslash(struct cache_fpl *fpl); 3944 static void cache_fpl_pathlen_dec(struct cache_fpl *fpl); 3945 static void cache_fpl_pathlen_inc(struct cache_fpl *fpl); 3946 static void cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n); 3947 static void cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n); 3948 3949 static void 3950 cache_fpl_cleanup_cnp(struct componentname *cnp) 3951 { 3952 3953 uma_zfree(namei_zone, cnp->cn_pnbuf); 3954 #ifdef DIAGNOSTIC 3955 cnp->cn_pnbuf = NULL; 3956 cnp->cn_nameptr = NULL; 3957 #endif 3958 } 3959 3960 static struct vnode * 3961 cache_fpl_handle_root(struct cache_fpl *fpl) 3962 { 3963 struct nameidata *ndp; 3964 struct componentname *cnp; 3965 3966 ndp = fpl->ndp; 3967 cnp = fpl->cnp; 3968 3969 MPASS(*(cnp->cn_nameptr) == '/'); 3970 cnp->cn_nameptr++; 3971 cache_fpl_pathlen_dec(fpl); 3972 3973 if (__predict_false(*(cnp->cn_nameptr) == '/')) { 3974 do { 3975 cnp->cn_nameptr++; 3976 cache_fpl_pathlen_dec(fpl); 3977 } while (*(cnp->cn_nameptr) == '/'); 3978 } 3979 3980 return (ndp->ni_rootdir); 3981 } 3982 3983 static void 3984 cache_fpl_checkpoint_outer(struct cache_fpl *fpl) 3985 { 3986 3987 fpl->snd_outer.ni_pathlen = fpl->ndp->ni_pathlen; 3988 fpl->snd_outer.cn_flags = fpl->ndp->ni_cnd.cn_flags; 3989 } 3990 3991 static void 3992 cache_fpl_checkpoint(struct cache_fpl *fpl) 3993 { 3994 3995 #ifdef INVARIANTS 3996 fpl->snd.cn_nameptr = fpl->ndp->ni_cnd.cn_nameptr; 3997 fpl->snd.ni_pathlen = fpl->debug.ni_pathlen; 3998 #endif 3999 } 4000 4001 static void 4002 cache_fpl_restore_partial(struct cache_fpl *fpl) 4003 { 4004 4005 fpl->ndp->ni_cnd.cn_flags = fpl->snd_outer.cn_flags; 4006 #ifdef INVARIANTS 4007 fpl->debug.ni_pathlen = fpl->snd.ni_pathlen; 4008 #endif 4009 } 4010 4011 static void 4012 cache_fpl_restore_abort(struct cache_fpl *fpl) 4013 { 4014 4015 cache_fpl_restore_partial(fpl); 4016 /* 4017 * It is 0 on entry by API contract. 4018 */ 4019 fpl->ndp->ni_resflags = 0; 4020 fpl->ndp->ni_cnd.cn_nameptr = fpl->ndp->ni_cnd.cn_pnbuf; 4021 fpl->ndp->ni_pathlen = fpl->snd_outer.ni_pathlen; 4022 } 4023 4024 #ifdef INVARIANTS 4025 #define cache_fpl_smr_assert_entered(fpl) ({ \ 4026 struct cache_fpl *_fpl = (fpl); \ 4027 MPASS(_fpl->in_smr == true); \ 4028 VFS_SMR_ASSERT_ENTERED(); \ 4029 }) 4030 #define cache_fpl_smr_assert_not_entered(fpl) ({ \ 4031 struct cache_fpl *_fpl = (fpl); \ 4032 MPASS(_fpl->in_smr == false); \ 4033 VFS_SMR_ASSERT_NOT_ENTERED(); \ 4034 }) 4035 static void 4036 cache_fpl_assert_status(struct cache_fpl *fpl) 4037 { 4038 4039 switch (fpl->status) { 4040 case CACHE_FPL_STATUS_UNSET: 4041 __assert_unreachable(); 4042 break; 4043 case CACHE_FPL_STATUS_DESTROYED: 4044 case CACHE_FPL_STATUS_ABORTED: 4045 case CACHE_FPL_STATUS_PARTIAL: 4046 case CACHE_FPL_STATUS_HANDLED: 4047 break; 4048 } 4049 } 4050 #else 4051 #define cache_fpl_smr_assert_entered(fpl) do { } while (0) 4052 #define cache_fpl_smr_assert_not_entered(fpl) do { } while (0) 4053 #define cache_fpl_assert_status(fpl) do { } while (0) 4054 #endif 4055 4056 #define cache_fpl_smr_enter_initial(fpl) ({ \ 4057 struct cache_fpl *_fpl = (fpl); \ 4058 vfs_smr_enter(); \ 4059 _fpl->in_smr = true; \ 4060 }) 4061 4062 #define cache_fpl_smr_enter(fpl) ({ \ 4063 struct cache_fpl *_fpl = (fpl); \ 4064 MPASS(_fpl->in_smr == false); \ 4065 vfs_smr_enter(); \ 4066 _fpl->in_smr = true; \ 4067 }) 4068 4069 #define cache_fpl_smr_exit(fpl) ({ \ 4070 struct cache_fpl *_fpl = (fpl); \ 4071 MPASS(_fpl->in_smr == true); \ 4072 vfs_smr_exit(); \ 4073 _fpl->in_smr = false; \ 4074 }) 4075 4076 static int 4077 cache_fpl_aborted_early_impl(struct cache_fpl *fpl, int line) 4078 { 4079 4080 if (fpl->status != CACHE_FPL_STATUS_UNSET) { 4081 KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL, 4082 ("%s: converting to abort from %d at %d, set at %d\n", 4083 __func__, fpl->status, line, fpl->line)); 4084 } 4085 cache_fpl_smr_assert_not_entered(fpl); 4086 fpl->status = CACHE_FPL_STATUS_ABORTED; 4087 fpl->line = line; 4088 return (CACHE_FPL_FAILED); 4089 } 4090 4091 #define cache_fpl_aborted_early(x) cache_fpl_aborted_early_impl((x), __LINE__) 4092 4093 static int __noinline 4094 cache_fpl_aborted_impl(struct cache_fpl *fpl, int line) 4095 { 4096 struct nameidata *ndp; 4097 struct componentname *cnp; 4098 4099 ndp = fpl->ndp; 4100 cnp = fpl->cnp; 4101 4102 if (fpl->status != CACHE_FPL_STATUS_UNSET) { 4103 KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL, 4104 ("%s: converting to abort from %d at %d, set at %d\n", 4105 __func__, fpl->status, line, fpl->line)); 4106 } 4107 fpl->status = CACHE_FPL_STATUS_ABORTED; 4108 fpl->line = line; 4109 if (fpl->in_smr) 4110 cache_fpl_smr_exit(fpl); 4111 cache_fpl_restore_abort(fpl); 4112 /* 4113 * Resolving symlinks overwrites data passed by the caller. 4114 * Let namei know. 4115 */ 4116 if (ndp->ni_loopcnt > 0) { 4117 fpl->status = CACHE_FPL_STATUS_DESTROYED; 4118 cache_fpl_cleanup_cnp(cnp); 4119 } 4120 return (CACHE_FPL_FAILED); 4121 } 4122 4123 #define cache_fpl_aborted(x) cache_fpl_aborted_impl((x), __LINE__) 4124 4125 static int __noinline 4126 cache_fpl_partial_impl(struct cache_fpl *fpl, int line) 4127 { 4128 4129 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET, 4130 ("%s: setting to partial at %d, but already set to %d at %d\n", 4131 __func__, line, fpl->status, fpl->line)); 4132 cache_fpl_smr_assert_entered(fpl); 4133 fpl->status = CACHE_FPL_STATUS_PARTIAL; 4134 fpl->line = line; 4135 return (cache_fplookup_partial_setup(fpl)); 4136 } 4137 4138 #define cache_fpl_partial(x) cache_fpl_partial_impl((x), __LINE__) 4139 4140 static int 4141 cache_fpl_handled_impl(struct cache_fpl *fpl, int line) 4142 { 4143 4144 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET, 4145 ("%s: setting to handled at %d, but already set to %d at %d\n", 4146 __func__, line, fpl->status, fpl->line)); 4147 cache_fpl_smr_assert_not_entered(fpl); 4148 fpl->status = CACHE_FPL_STATUS_HANDLED; 4149 fpl->line = line; 4150 return (0); 4151 } 4152 4153 #define cache_fpl_handled(x) cache_fpl_handled_impl((x), __LINE__) 4154 4155 static int 4156 cache_fpl_handled_error_impl(struct cache_fpl *fpl, int error, int line) 4157 { 4158 4159 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET, 4160 ("%s: setting to handled at %d, but already set to %d at %d\n", 4161 __func__, line, fpl->status, fpl->line)); 4162 MPASS(error != 0); 4163 MPASS(error != CACHE_FPL_FAILED); 4164 cache_fpl_smr_assert_not_entered(fpl); 4165 fpl->status = CACHE_FPL_STATUS_HANDLED; 4166 fpl->line = line; 4167 fpl->dvp = NULL; 4168 fpl->tvp = NULL; 4169 fpl->savename = false; 4170 return (error); 4171 } 4172 4173 #define cache_fpl_handled_error(x, e) cache_fpl_handled_error_impl((x), (e), __LINE__) 4174 4175 static bool 4176 cache_fpl_terminated(struct cache_fpl *fpl) 4177 { 4178 4179 return (fpl->status != CACHE_FPL_STATUS_UNSET); 4180 } 4181 4182 #define CACHE_FPL_SUPPORTED_CN_FLAGS \ 4183 (NC_NOMAKEENTRY | NC_KEEPPOSENTRY | LOCKLEAF | LOCKPARENT | WANTPARENT | \ 4184 FAILIFEXISTS | FOLLOW | EMPTYPATH | LOCKSHARED | SAVENAME | SAVESTART | \ 4185 WILLBEDIR | ISOPEN | NOMACCHECK | AUDITVNODE1 | AUDITVNODE2 | NOCAPCHECK | \ 4186 OPENREAD | OPENWRITE | WANTIOCTLCAPS) 4187 4188 #define CACHE_FPL_INTERNAL_CN_FLAGS \ 4189 (ISDOTDOT | MAKEENTRY | ISLASTCN) 4190 4191 _Static_assert((CACHE_FPL_SUPPORTED_CN_FLAGS & CACHE_FPL_INTERNAL_CN_FLAGS) == 0, 4192 "supported and internal flags overlap"); 4193 4194 static bool 4195 cache_fpl_islastcn(struct nameidata *ndp) 4196 { 4197 4198 return (*ndp->ni_next == 0); 4199 } 4200 4201 static bool 4202 cache_fpl_istrailingslash(struct cache_fpl *fpl) 4203 { 4204 4205 MPASS(fpl->nulchar > fpl->cnp->cn_pnbuf); 4206 return (*(fpl->nulchar - 1) == '/'); 4207 } 4208 4209 static bool 4210 cache_fpl_isdotdot(struct componentname *cnp) 4211 { 4212 4213 if (cnp->cn_namelen == 2 && 4214 cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.') 4215 return (true); 4216 return (false); 4217 } 4218 4219 static bool 4220 cache_can_fplookup(struct cache_fpl *fpl) 4221 { 4222 struct nameidata *ndp; 4223 struct componentname *cnp; 4224 struct thread *td; 4225 4226 ndp = fpl->ndp; 4227 cnp = fpl->cnp; 4228 td = curthread; 4229 4230 if (!atomic_load_char(&cache_fast_lookup_enabled)) { 4231 cache_fpl_aborted_early(fpl); 4232 return (false); 4233 } 4234 if ((cnp->cn_flags & ~CACHE_FPL_SUPPORTED_CN_FLAGS) != 0) { 4235 cache_fpl_aborted_early(fpl); 4236 return (false); 4237 } 4238 if (IN_CAPABILITY_MODE(td)) { 4239 cache_fpl_aborted_early(fpl); 4240 return (false); 4241 } 4242 if (AUDITING_TD(td)) { 4243 cache_fpl_aborted_early(fpl); 4244 return (false); 4245 } 4246 if (ndp->ni_startdir != NULL) { 4247 cache_fpl_aborted_early(fpl); 4248 return (false); 4249 } 4250 return (true); 4251 } 4252 4253 static int __noinline 4254 cache_fplookup_dirfd(struct cache_fpl *fpl, struct vnode **vpp) 4255 { 4256 struct nameidata *ndp; 4257 struct componentname *cnp; 4258 int error; 4259 bool fsearch; 4260 4261 ndp = fpl->ndp; 4262 cnp = fpl->cnp; 4263 4264 error = fgetvp_lookup_smr(ndp->ni_dirfd, ndp, vpp, &fsearch); 4265 if (__predict_false(error != 0)) { 4266 return (cache_fpl_aborted(fpl)); 4267 } 4268 fpl->fsearch = fsearch; 4269 if ((*vpp)->v_type != VDIR) { 4270 if (!((cnp->cn_flags & EMPTYPATH) != 0 && cnp->cn_pnbuf[0] == '\0')) { 4271 cache_fpl_smr_exit(fpl); 4272 return (cache_fpl_handled_error(fpl, ENOTDIR)); 4273 } 4274 } 4275 return (0); 4276 } 4277 4278 static int __noinline 4279 cache_fplookup_negative_promote(struct cache_fpl *fpl, struct namecache *oncp, 4280 uint32_t hash) 4281 { 4282 struct componentname *cnp; 4283 struct vnode *dvp; 4284 4285 cnp = fpl->cnp; 4286 dvp = fpl->dvp; 4287 4288 cache_fpl_smr_exit(fpl); 4289 if (cache_neg_promote_cond(dvp, cnp, oncp, hash)) 4290 return (cache_fpl_handled_error(fpl, ENOENT)); 4291 else 4292 return (cache_fpl_aborted(fpl)); 4293 } 4294 4295 /* 4296 * The target vnode is not supported, prepare for the slow path to take over. 4297 */ 4298 static int __noinline 4299 cache_fplookup_partial_setup(struct cache_fpl *fpl) 4300 { 4301 struct nameidata *ndp; 4302 struct componentname *cnp; 4303 enum vgetstate dvs; 4304 struct vnode *dvp; 4305 struct pwd *pwd; 4306 seqc_t dvp_seqc; 4307 4308 ndp = fpl->ndp; 4309 cnp = fpl->cnp; 4310 pwd = *(fpl->pwd); 4311 dvp = fpl->dvp; 4312 dvp_seqc = fpl->dvp_seqc; 4313 4314 if (!pwd_hold_smr(pwd)) { 4315 return (cache_fpl_aborted(fpl)); 4316 } 4317 4318 /* 4319 * Note that seqc is checked before the vnode is locked, so by 4320 * the time regular lookup gets to it it may have moved. 4321 * 4322 * Ultimately this does not affect correctness, any lookup errors 4323 * are userspace racing with itself. It is guaranteed that any 4324 * path which ultimately gets found could also have been found 4325 * by regular lookup going all the way in absence of concurrent 4326 * modifications. 4327 */ 4328 dvs = vget_prep_smr(dvp); 4329 cache_fpl_smr_exit(fpl); 4330 if (__predict_false(dvs == VGET_NONE)) { 4331 pwd_drop(pwd); 4332 return (cache_fpl_aborted(fpl)); 4333 } 4334 4335 vget_finish_ref(dvp, dvs); 4336 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 4337 vrele(dvp); 4338 pwd_drop(pwd); 4339 return (cache_fpl_aborted(fpl)); 4340 } 4341 4342 cache_fpl_restore_partial(fpl); 4343 #ifdef INVARIANTS 4344 if (cnp->cn_nameptr != fpl->snd.cn_nameptr) { 4345 panic("%s: cn_nameptr mismatch (%p != %p) full [%s]\n", __func__, 4346 cnp->cn_nameptr, fpl->snd.cn_nameptr, cnp->cn_pnbuf); 4347 } 4348 #endif 4349 4350 ndp->ni_startdir = dvp; 4351 cnp->cn_flags |= MAKEENTRY; 4352 if (cache_fpl_islastcn(ndp)) 4353 cnp->cn_flags |= ISLASTCN; 4354 if (cache_fpl_isdotdot(cnp)) 4355 cnp->cn_flags |= ISDOTDOT; 4356 4357 /* 4358 * Skip potential extra slashes parsing did not take care of. 4359 * cache_fplookup_skip_slashes explains the mechanism. 4360 */ 4361 if (__predict_false(*(cnp->cn_nameptr) == '/')) { 4362 do { 4363 cnp->cn_nameptr++; 4364 cache_fpl_pathlen_dec(fpl); 4365 } while (*(cnp->cn_nameptr) == '/'); 4366 } 4367 4368 ndp->ni_pathlen = fpl->nulchar - cnp->cn_nameptr + 1; 4369 #ifdef INVARIANTS 4370 if (ndp->ni_pathlen != fpl->debug.ni_pathlen) { 4371 panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n", 4372 __func__, ndp->ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar, 4373 cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf); 4374 } 4375 #endif 4376 return (0); 4377 } 4378 4379 static int 4380 cache_fplookup_final_child(struct cache_fpl *fpl, enum vgetstate tvs) 4381 { 4382 struct componentname *cnp; 4383 struct vnode *tvp; 4384 seqc_t tvp_seqc; 4385 int error, lkflags; 4386 4387 cnp = fpl->cnp; 4388 tvp = fpl->tvp; 4389 tvp_seqc = fpl->tvp_seqc; 4390 4391 if ((cnp->cn_flags & LOCKLEAF) != 0) { 4392 lkflags = LK_SHARED; 4393 if ((cnp->cn_flags & LOCKSHARED) == 0) 4394 lkflags = LK_EXCLUSIVE; 4395 error = vget_finish(tvp, lkflags, tvs); 4396 if (__predict_false(error != 0)) { 4397 return (cache_fpl_aborted(fpl)); 4398 } 4399 } else { 4400 vget_finish_ref(tvp, tvs); 4401 } 4402 4403 if (!vn_seqc_consistent(tvp, tvp_seqc)) { 4404 if ((cnp->cn_flags & LOCKLEAF) != 0) 4405 vput(tvp); 4406 else 4407 vrele(tvp); 4408 return (cache_fpl_aborted(fpl)); 4409 } 4410 4411 return (cache_fpl_handled(fpl)); 4412 } 4413 4414 /* 4415 * They want to possibly modify the state of the namecache. 4416 */ 4417 static int __noinline 4418 cache_fplookup_final_modifying(struct cache_fpl *fpl) 4419 { 4420 struct nameidata *ndp; 4421 struct componentname *cnp; 4422 enum vgetstate dvs; 4423 struct vnode *dvp, *tvp; 4424 struct mount *mp; 4425 seqc_t dvp_seqc; 4426 int error; 4427 bool docache; 4428 4429 ndp = fpl->ndp; 4430 cnp = fpl->cnp; 4431 dvp = fpl->dvp; 4432 dvp_seqc = fpl->dvp_seqc; 4433 4434 MPASS(*(cnp->cn_nameptr) != '/'); 4435 MPASS(cache_fpl_islastcn(ndp)); 4436 if ((cnp->cn_flags & LOCKPARENT) == 0) 4437 MPASS((cnp->cn_flags & WANTPARENT) != 0); 4438 MPASS((cnp->cn_flags & TRAILINGSLASH) == 0); 4439 MPASS(cnp->cn_nameiop == CREATE || cnp->cn_nameiop == DELETE || 4440 cnp->cn_nameiop == RENAME); 4441 MPASS((cnp->cn_flags & MAKEENTRY) == 0); 4442 MPASS((cnp->cn_flags & ISDOTDOT) == 0); 4443 4444 docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE; 4445 if (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME) 4446 docache = false; 4447 4448 /* 4449 * Regular lookup nulifies the slash, which we don't do here. 4450 * Don't take chances with filesystem routines seeing it for 4451 * the last entry. 4452 */ 4453 if (cache_fpl_istrailingslash(fpl)) { 4454 return (cache_fpl_partial(fpl)); 4455 } 4456 4457 mp = atomic_load_ptr(&dvp->v_mount); 4458 if (__predict_false(mp == NULL)) { 4459 return (cache_fpl_aborted(fpl)); 4460 } 4461 4462 if (__predict_false(mp->mnt_flag & MNT_RDONLY)) { 4463 cache_fpl_smr_exit(fpl); 4464 /* 4465 * Original code keeps not checking for CREATE which 4466 * might be a bug. For now let the old lookup decide. 4467 */ 4468 if (cnp->cn_nameiop == CREATE) { 4469 return (cache_fpl_aborted(fpl)); 4470 } 4471 return (cache_fpl_handled_error(fpl, EROFS)); 4472 } 4473 4474 if (fpl->tvp != NULL && (cnp->cn_flags & FAILIFEXISTS) != 0) { 4475 cache_fpl_smr_exit(fpl); 4476 return (cache_fpl_handled_error(fpl, EEXIST)); 4477 } 4478 4479 /* 4480 * Secure access to dvp; check cache_fplookup_partial_setup for 4481 * reasoning. 4482 * 4483 * XXX At least UFS requires its lookup routine to be called for 4484 * the last path component, which leads to some level of complication 4485 * and inefficiency: 4486 * - the target routine always locks the target vnode, but our caller 4487 * may not need it locked 4488 * - some of the VOP machinery asserts that the parent is locked, which 4489 * once more may be not required 4490 * 4491 * TODO: add a flag for filesystems which don't need this. 4492 */ 4493 dvs = vget_prep_smr(dvp); 4494 cache_fpl_smr_exit(fpl); 4495 if (__predict_false(dvs == VGET_NONE)) { 4496 return (cache_fpl_aborted(fpl)); 4497 } 4498 4499 vget_finish_ref(dvp, dvs); 4500 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 4501 vrele(dvp); 4502 return (cache_fpl_aborted(fpl)); 4503 } 4504 4505 error = vn_lock(dvp, LK_EXCLUSIVE); 4506 if (__predict_false(error != 0)) { 4507 vrele(dvp); 4508 return (cache_fpl_aborted(fpl)); 4509 } 4510 4511 tvp = NULL; 4512 cnp->cn_flags |= ISLASTCN; 4513 if (docache) 4514 cnp->cn_flags |= MAKEENTRY; 4515 if (cache_fpl_isdotdot(cnp)) 4516 cnp->cn_flags |= ISDOTDOT; 4517 cnp->cn_lkflags = LK_EXCLUSIVE; 4518 error = VOP_LOOKUP(dvp, &tvp, cnp); 4519 switch (error) { 4520 case EJUSTRETURN: 4521 case 0: 4522 break; 4523 case ENOTDIR: 4524 case ENOENT: 4525 vput(dvp); 4526 return (cache_fpl_handled_error(fpl, error)); 4527 default: 4528 vput(dvp); 4529 return (cache_fpl_aborted(fpl)); 4530 } 4531 4532 fpl->tvp = tvp; 4533 fpl->savename = (cnp->cn_flags & SAVENAME) != 0; 4534 4535 if (tvp == NULL) { 4536 if ((cnp->cn_flags & SAVESTART) != 0) { 4537 ndp->ni_startdir = dvp; 4538 vrefact(ndp->ni_startdir); 4539 cnp->cn_flags |= SAVENAME; 4540 fpl->savename = true; 4541 } 4542 MPASS(error == EJUSTRETURN); 4543 if ((cnp->cn_flags & LOCKPARENT) == 0) { 4544 VOP_UNLOCK(dvp); 4545 } 4546 return (cache_fpl_handled(fpl)); 4547 } 4548 4549 /* 4550 * There are very hairy corner cases concerning various flag combinations 4551 * and locking state. In particular here we only hold one lock instead of 4552 * two. 4553 * 4554 * Skip the complexity as it is of no significance for normal workloads. 4555 */ 4556 if (__predict_false(tvp == dvp)) { 4557 vput(dvp); 4558 vrele(tvp); 4559 return (cache_fpl_aborted(fpl)); 4560 } 4561 4562 /* 4563 * If they want the symlink itself we are fine, but if they want to 4564 * follow it regular lookup has to be engaged. 4565 */ 4566 if (tvp->v_type == VLNK) { 4567 if ((cnp->cn_flags & FOLLOW) != 0) { 4568 vput(dvp); 4569 vput(tvp); 4570 return (cache_fpl_aborted(fpl)); 4571 } 4572 } 4573 4574 /* 4575 * Since we expect this to be the terminal vnode it should almost never 4576 * be a mount point. 4577 */ 4578 if (__predict_false(cache_fplookup_is_mp(fpl))) { 4579 vput(dvp); 4580 vput(tvp); 4581 return (cache_fpl_aborted(fpl)); 4582 } 4583 4584 if ((cnp->cn_flags & FAILIFEXISTS) != 0) { 4585 vput(dvp); 4586 vput(tvp); 4587 return (cache_fpl_handled_error(fpl, EEXIST)); 4588 } 4589 4590 if ((cnp->cn_flags & LOCKLEAF) == 0) { 4591 VOP_UNLOCK(tvp); 4592 } 4593 4594 if ((cnp->cn_flags & LOCKPARENT) == 0) { 4595 VOP_UNLOCK(dvp); 4596 } 4597 4598 if ((cnp->cn_flags & SAVESTART) != 0) { 4599 ndp->ni_startdir = dvp; 4600 vrefact(ndp->ni_startdir); 4601 cnp->cn_flags |= SAVENAME; 4602 fpl->savename = true; 4603 } 4604 4605 return (cache_fpl_handled(fpl)); 4606 } 4607 4608 static int __noinline 4609 cache_fplookup_modifying(struct cache_fpl *fpl) 4610 { 4611 struct nameidata *ndp; 4612 4613 ndp = fpl->ndp; 4614 4615 if (!cache_fpl_islastcn(ndp)) { 4616 return (cache_fpl_partial(fpl)); 4617 } 4618 return (cache_fplookup_final_modifying(fpl)); 4619 } 4620 4621 static int __noinline 4622 cache_fplookup_final_withparent(struct cache_fpl *fpl) 4623 { 4624 struct componentname *cnp; 4625 enum vgetstate dvs, tvs; 4626 struct vnode *dvp, *tvp; 4627 seqc_t dvp_seqc; 4628 int error; 4629 4630 cnp = fpl->cnp; 4631 dvp = fpl->dvp; 4632 dvp_seqc = fpl->dvp_seqc; 4633 tvp = fpl->tvp; 4634 4635 MPASS((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0); 4636 4637 /* 4638 * This is less efficient than it can be for simplicity. 4639 */ 4640 dvs = vget_prep_smr(dvp); 4641 if (__predict_false(dvs == VGET_NONE)) { 4642 return (cache_fpl_aborted(fpl)); 4643 } 4644 tvs = vget_prep_smr(tvp); 4645 if (__predict_false(tvs == VGET_NONE)) { 4646 cache_fpl_smr_exit(fpl); 4647 vget_abort(dvp, dvs); 4648 return (cache_fpl_aborted(fpl)); 4649 } 4650 4651 cache_fpl_smr_exit(fpl); 4652 4653 if ((cnp->cn_flags & LOCKPARENT) != 0) { 4654 error = vget_finish(dvp, LK_EXCLUSIVE, dvs); 4655 if (__predict_false(error != 0)) { 4656 vget_abort(tvp, tvs); 4657 return (cache_fpl_aborted(fpl)); 4658 } 4659 } else { 4660 vget_finish_ref(dvp, dvs); 4661 } 4662 4663 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 4664 vget_abort(tvp, tvs); 4665 if ((cnp->cn_flags & LOCKPARENT) != 0) 4666 vput(dvp); 4667 else 4668 vrele(dvp); 4669 return (cache_fpl_aborted(fpl)); 4670 } 4671 4672 error = cache_fplookup_final_child(fpl, tvs); 4673 if (__predict_false(error != 0)) { 4674 MPASS(fpl->status == CACHE_FPL_STATUS_ABORTED || 4675 fpl->status == CACHE_FPL_STATUS_DESTROYED); 4676 if ((cnp->cn_flags & LOCKPARENT) != 0) 4677 vput(dvp); 4678 else 4679 vrele(dvp); 4680 return (error); 4681 } 4682 4683 MPASS(fpl->status == CACHE_FPL_STATUS_HANDLED); 4684 return (0); 4685 } 4686 4687 static int 4688 cache_fplookup_final(struct cache_fpl *fpl) 4689 { 4690 struct componentname *cnp; 4691 enum vgetstate tvs; 4692 struct vnode *dvp, *tvp; 4693 seqc_t dvp_seqc; 4694 4695 cnp = fpl->cnp; 4696 dvp = fpl->dvp; 4697 dvp_seqc = fpl->dvp_seqc; 4698 tvp = fpl->tvp; 4699 4700 MPASS(*(cnp->cn_nameptr) != '/'); 4701 4702 if (cnp->cn_nameiop != LOOKUP) { 4703 return (cache_fplookup_final_modifying(fpl)); 4704 } 4705 4706 if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0) 4707 return (cache_fplookup_final_withparent(fpl)); 4708 4709 tvs = vget_prep_smr(tvp); 4710 if (__predict_false(tvs == VGET_NONE)) { 4711 return (cache_fpl_partial(fpl)); 4712 } 4713 4714 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 4715 cache_fpl_smr_exit(fpl); 4716 vget_abort(tvp, tvs); 4717 return (cache_fpl_aborted(fpl)); 4718 } 4719 4720 cache_fpl_smr_exit(fpl); 4721 return (cache_fplookup_final_child(fpl, tvs)); 4722 } 4723 4724 /* 4725 * Comment from locked lookup: 4726 * Check for degenerate name (e.g. / or "") which is a way of talking about a 4727 * directory, e.g. like "/." or ".". 4728 */ 4729 static int __noinline 4730 cache_fplookup_degenerate(struct cache_fpl *fpl) 4731 { 4732 struct componentname *cnp; 4733 struct vnode *dvp; 4734 enum vgetstate dvs; 4735 int error, lkflags; 4736 #ifdef INVARIANTS 4737 char *cp; 4738 #endif 4739 4740 fpl->tvp = fpl->dvp; 4741 fpl->tvp_seqc = fpl->dvp_seqc; 4742 4743 cnp = fpl->cnp; 4744 dvp = fpl->dvp; 4745 4746 #ifdef INVARIANTS 4747 for (cp = cnp->cn_pnbuf; *cp != '\0'; cp++) { 4748 KASSERT(*cp == '/', 4749 ("%s: encountered non-slash; string [%s]\n", __func__, 4750 cnp->cn_pnbuf)); 4751 } 4752 #endif 4753 4754 if (__predict_false(cnp->cn_nameiop != LOOKUP)) { 4755 cache_fpl_smr_exit(fpl); 4756 return (cache_fpl_handled_error(fpl, EISDIR)); 4757 } 4758 4759 MPASS((cnp->cn_flags & SAVESTART) == 0); 4760 4761 if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0) { 4762 return (cache_fplookup_final_withparent(fpl)); 4763 } 4764 4765 dvs = vget_prep_smr(dvp); 4766 cache_fpl_smr_exit(fpl); 4767 if (__predict_false(dvs == VGET_NONE)) { 4768 return (cache_fpl_aborted(fpl)); 4769 } 4770 4771 if ((cnp->cn_flags & LOCKLEAF) != 0) { 4772 lkflags = LK_SHARED; 4773 if ((cnp->cn_flags & LOCKSHARED) == 0) 4774 lkflags = LK_EXCLUSIVE; 4775 error = vget_finish(dvp, lkflags, dvs); 4776 if (__predict_false(error != 0)) { 4777 return (cache_fpl_aborted(fpl)); 4778 } 4779 } else { 4780 vget_finish_ref(dvp, dvs); 4781 } 4782 return (cache_fpl_handled(fpl)); 4783 } 4784 4785 static int __noinline 4786 cache_fplookup_emptypath(struct cache_fpl *fpl) 4787 { 4788 struct nameidata *ndp; 4789 struct componentname *cnp; 4790 enum vgetstate tvs; 4791 struct vnode *tvp; 4792 int error, lkflags; 4793 4794 fpl->tvp = fpl->dvp; 4795 fpl->tvp_seqc = fpl->dvp_seqc; 4796 4797 ndp = fpl->ndp; 4798 cnp = fpl->cnp; 4799 tvp = fpl->tvp; 4800 4801 MPASS(*cnp->cn_pnbuf == '\0'); 4802 4803 if (__predict_false((cnp->cn_flags & EMPTYPATH) == 0)) { 4804 cache_fpl_smr_exit(fpl); 4805 return (cache_fpl_handled_error(fpl, ENOENT)); 4806 } 4807 4808 MPASS((cnp->cn_flags & (LOCKPARENT | WANTPARENT)) == 0); 4809 4810 tvs = vget_prep_smr(tvp); 4811 cache_fpl_smr_exit(fpl); 4812 if (__predict_false(tvs == VGET_NONE)) { 4813 return (cache_fpl_aborted(fpl)); 4814 } 4815 4816 if ((cnp->cn_flags & LOCKLEAF) != 0) { 4817 lkflags = LK_SHARED; 4818 if ((cnp->cn_flags & LOCKSHARED) == 0) 4819 lkflags = LK_EXCLUSIVE; 4820 error = vget_finish(tvp, lkflags, tvs); 4821 if (__predict_false(error != 0)) { 4822 return (cache_fpl_aborted(fpl)); 4823 } 4824 } else { 4825 vget_finish_ref(tvp, tvs); 4826 } 4827 4828 ndp->ni_resflags |= NIRES_EMPTYPATH; 4829 return (cache_fpl_handled(fpl)); 4830 } 4831 4832 static int __noinline 4833 cache_fplookup_noentry(struct cache_fpl *fpl) 4834 { 4835 struct nameidata *ndp; 4836 struct componentname *cnp; 4837 enum vgetstate dvs; 4838 struct vnode *dvp, *tvp; 4839 seqc_t dvp_seqc; 4840 int error; 4841 4842 ndp = fpl->ndp; 4843 cnp = fpl->cnp; 4844 dvp = fpl->dvp; 4845 dvp_seqc = fpl->dvp_seqc; 4846 4847 MPASS((cnp->cn_flags & MAKEENTRY) == 0); 4848 MPASS((cnp->cn_flags & ISDOTDOT) == 0); 4849 if (cnp->cn_nameiop == LOOKUP) 4850 MPASS((cnp->cn_flags & NOCACHE) == 0); 4851 MPASS(!cache_fpl_isdotdot(cnp)); 4852 4853 /* 4854 * Hack: delayed name len checking. 4855 */ 4856 if (__predict_false(cnp->cn_namelen > NAME_MAX)) { 4857 cache_fpl_smr_exit(fpl); 4858 return (cache_fpl_handled_error(fpl, ENAMETOOLONG)); 4859 } 4860 4861 if (cnp->cn_nameptr[0] == '/') { 4862 return (cache_fplookup_skip_slashes(fpl)); 4863 } 4864 4865 if (cnp->cn_pnbuf[0] == '\0') { 4866 return (cache_fplookup_emptypath(fpl)); 4867 } 4868 4869 if (cnp->cn_nameptr[0] == '\0') { 4870 if (fpl->tvp == NULL) { 4871 return (cache_fplookup_degenerate(fpl)); 4872 } 4873 return (cache_fplookup_trailingslash(fpl)); 4874 } 4875 4876 if (cnp->cn_nameiop != LOOKUP) { 4877 fpl->tvp = NULL; 4878 return (cache_fplookup_modifying(fpl)); 4879 } 4880 4881 MPASS((cnp->cn_flags & SAVESTART) == 0); 4882 4883 /* 4884 * Only try to fill in the component if it is the last one, 4885 * otherwise not only there may be several to handle but the 4886 * walk may be complicated. 4887 */ 4888 if (!cache_fpl_islastcn(ndp)) { 4889 return (cache_fpl_partial(fpl)); 4890 } 4891 4892 /* 4893 * Regular lookup nulifies the slash, which we don't do here. 4894 * Don't take chances with filesystem routines seeing it for 4895 * the last entry. 4896 */ 4897 if (cache_fpl_istrailingslash(fpl)) { 4898 return (cache_fpl_partial(fpl)); 4899 } 4900 4901 /* 4902 * Secure access to dvp; check cache_fplookup_partial_setup for 4903 * reasoning. 4904 */ 4905 dvs = vget_prep_smr(dvp); 4906 cache_fpl_smr_exit(fpl); 4907 if (__predict_false(dvs == VGET_NONE)) { 4908 return (cache_fpl_aborted(fpl)); 4909 } 4910 4911 vget_finish_ref(dvp, dvs); 4912 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 4913 vrele(dvp); 4914 return (cache_fpl_aborted(fpl)); 4915 } 4916 4917 error = vn_lock(dvp, LK_SHARED); 4918 if (__predict_false(error != 0)) { 4919 vrele(dvp); 4920 return (cache_fpl_aborted(fpl)); 4921 } 4922 4923 tvp = NULL; 4924 /* 4925 * TODO: provide variants which don't require locking either vnode. 4926 */ 4927 cnp->cn_flags |= ISLASTCN | MAKEENTRY; 4928 cnp->cn_lkflags = LK_SHARED; 4929 if ((cnp->cn_flags & LOCKSHARED) == 0) { 4930 cnp->cn_lkflags = LK_EXCLUSIVE; 4931 } 4932 error = VOP_LOOKUP(dvp, &tvp, cnp); 4933 switch (error) { 4934 case EJUSTRETURN: 4935 case 0: 4936 break; 4937 case ENOTDIR: 4938 case ENOENT: 4939 vput(dvp); 4940 return (cache_fpl_handled_error(fpl, error)); 4941 default: 4942 vput(dvp); 4943 return (cache_fpl_aborted(fpl)); 4944 } 4945 4946 fpl->tvp = tvp; 4947 if (!fpl->savename) { 4948 MPASS((cnp->cn_flags & SAVENAME) == 0); 4949 } 4950 4951 if (tvp == NULL) { 4952 MPASS(error == EJUSTRETURN); 4953 if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) { 4954 vput(dvp); 4955 } else if ((cnp->cn_flags & LOCKPARENT) == 0) { 4956 VOP_UNLOCK(dvp); 4957 } 4958 return (cache_fpl_handled(fpl)); 4959 } 4960 4961 if (tvp->v_type == VLNK) { 4962 if ((cnp->cn_flags & FOLLOW) != 0) { 4963 vput(dvp); 4964 vput(tvp); 4965 return (cache_fpl_aborted(fpl)); 4966 } 4967 } 4968 4969 if (__predict_false(cache_fplookup_is_mp(fpl))) { 4970 vput(dvp); 4971 vput(tvp); 4972 return (cache_fpl_aborted(fpl)); 4973 } 4974 4975 if ((cnp->cn_flags & LOCKLEAF) == 0) { 4976 VOP_UNLOCK(tvp); 4977 } 4978 4979 if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) { 4980 vput(dvp); 4981 } else if ((cnp->cn_flags & LOCKPARENT) == 0) { 4982 VOP_UNLOCK(dvp); 4983 } 4984 return (cache_fpl_handled(fpl)); 4985 } 4986 4987 static int __noinline 4988 cache_fplookup_dot(struct cache_fpl *fpl) 4989 { 4990 int error; 4991 4992 MPASS(!seqc_in_modify(fpl->dvp_seqc)); 4993 /* 4994 * Just re-assign the value. seqc will be checked later for the first 4995 * non-dot path component in line and/or before deciding to return the 4996 * vnode. 4997 */ 4998 fpl->tvp = fpl->dvp; 4999 fpl->tvp_seqc = fpl->dvp_seqc; 5000 5001 counter_u64_add(dothits, 1); 5002 SDT_PROBE3(vfs, namecache, lookup, hit, fpl->dvp, ".", fpl->dvp); 5003 5004 error = 0; 5005 if (cache_fplookup_is_mp(fpl)) { 5006 error = cache_fplookup_cross_mount(fpl); 5007 } 5008 return (error); 5009 } 5010 5011 static int __noinline 5012 cache_fplookup_dotdot(struct cache_fpl *fpl) 5013 { 5014 struct nameidata *ndp; 5015 struct componentname *cnp; 5016 struct namecache *ncp; 5017 struct vnode *dvp; 5018 struct prison *pr; 5019 u_char nc_flag; 5020 5021 ndp = fpl->ndp; 5022 cnp = fpl->cnp; 5023 dvp = fpl->dvp; 5024 5025 MPASS(cache_fpl_isdotdot(cnp)); 5026 5027 /* 5028 * XXX this is racy the same way regular lookup is 5029 */ 5030 for (pr = cnp->cn_cred->cr_prison; pr != NULL; 5031 pr = pr->pr_parent) 5032 if (dvp == pr->pr_root) 5033 break; 5034 5035 if (dvp == ndp->ni_rootdir || 5036 dvp == ndp->ni_topdir || 5037 dvp == rootvnode || 5038 pr != NULL) { 5039 fpl->tvp = dvp; 5040 fpl->tvp_seqc = vn_seqc_read_any(dvp); 5041 if (seqc_in_modify(fpl->tvp_seqc)) { 5042 return (cache_fpl_aborted(fpl)); 5043 } 5044 return (0); 5045 } 5046 5047 if ((dvp->v_vflag & VV_ROOT) != 0) { 5048 /* 5049 * TODO 5050 * The opposite of climb mount is needed here. 5051 */ 5052 return (cache_fpl_partial(fpl)); 5053 } 5054 5055 ncp = atomic_load_consume_ptr(&dvp->v_cache_dd); 5056 if (ncp == NULL) { 5057 return (cache_fpl_aborted(fpl)); 5058 } 5059 5060 nc_flag = atomic_load_char(&ncp->nc_flag); 5061 if ((nc_flag & NCF_ISDOTDOT) != 0) { 5062 if ((nc_flag & NCF_NEGATIVE) != 0) 5063 return (cache_fpl_aborted(fpl)); 5064 fpl->tvp = ncp->nc_vp; 5065 } else { 5066 fpl->tvp = ncp->nc_dvp; 5067 } 5068 5069 fpl->tvp_seqc = vn_seqc_read_any(fpl->tvp); 5070 if (seqc_in_modify(fpl->tvp_seqc)) { 5071 return (cache_fpl_partial(fpl)); 5072 } 5073 5074 /* 5075 * Acquire fence provided by vn_seqc_read_any above. 5076 */ 5077 if (__predict_false(atomic_load_ptr(&dvp->v_cache_dd) != ncp)) { 5078 return (cache_fpl_aborted(fpl)); 5079 } 5080 5081 if (!cache_ncp_canuse(ncp)) { 5082 return (cache_fpl_aborted(fpl)); 5083 } 5084 5085 counter_u64_add(dotdothits, 1); 5086 return (0); 5087 } 5088 5089 static int __noinline 5090 cache_fplookup_neg(struct cache_fpl *fpl, struct namecache *ncp, uint32_t hash) 5091 { 5092 u_char nc_flag __diagused; 5093 bool neg_promote; 5094 5095 #ifdef INVARIANTS 5096 nc_flag = atomic_load_char(&ncp->nc_flag); 5097 MPASS((nc_flag & NCF_NEGATIVE) != 0); 5098 #endif 5099 /* 5100 * If they want to create an entry we need to replace this one. 5101 */ 5102 if (__predict_false(fpl->cnp->cn_nameiop != LOOKUP)) { 5103 fpl->tvp = NULL; 5104 return (cache_fplookup_modifying(fpl)); 5105 } 5106 neg_promote = cache_neg_hit_prep(ncp); 5107 if (!cache_fpl_neg_ncp_canuse(ncp)) { 5108 cache_neg_hit_abort(ncp); 5109 return (cache_fpl_partial(fpl)); 5110 } 5111 if (neg_promote) { 5112 return (cache_fplookup_negative_promote(fpl, ncp, hash)); 5113 } 5114 cache_neg_hit_finish(ncp); 5115 cache_fpl_smr_exit(fpl); 5116 return (cache_fpl_handled_error(fpl, ENOENT)); 5117 } 5118 5119 /* 5120 * Resolve a symlink. Called by filesystem-specific routines. 5121 * 5122 * Code flow is: 5123 * ... -> cache_fplookup_symlink -> VOP_FPLOOKUP_SYMLINK -> cache_symlink_resolve 5124 */ 5125 int 5126 cache_symlink_resolve(struct cache_fpl *fpl, const char *string, size_t len) 5127 { 5128 struct nameidata *ndp; 5129 struct componentname *cnp; 5130 size_t adjust; 5131 5132 ndp = fpl->ndp; 5133 cnp = fpl->cnp; 5134 5135 if (__predict_false(len == 0)) { 5136 return (ENOENT); 5137 } 5138 5139 if (__predict_false(len > MAXPATHLEN - 2)) { 5140 if (cache_fpl_istrailingslash(fpl)) { 5141 return (EAGAIN); 5142 } 5143 } 5144 5145 ndp->ni_pathlen = fpl->nulchar - cnp->cn_nameptr - cnp->cn_namelen + 1; 5146 #ifdef INVARIANTS 5147 if (ndp->ni_pathlen != fpl->debug.ni_pathlen) { 5148 panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n", 5149 __func__, ndp->ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar, 5150 cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf); 5151 } 5152 #endif 5153 5154 if (__predict_false(len + ndp->ni_pathlen > MAXPATHLEN)) { 5155 return (ENAMETOOLONG); 5156 } 5157 5158 if (__predict_false(ndp->ni_loopcnt++ >= MAXSYMLINKS)) { 5159 return (ELOOP); 5160 } 5161 5162 adjust = len; 5163 if (ndp->ni_pathlen > 1) { 5164 bcopy(ndp->ni_next, cnp->cn_pnbuf + len, ndp->ni_pathlen); 5165 } else { 5166 if (cache_fpl_istrailingslash(fpl)) { 5167 adjust = len + 1; 5168 cnp->cn_pnbuf[len] = '/'; 5169 cnp->cn_pnbuf[len + 1] = '\0'; 5170 } else { 5171 cnp->cn_pnbuf[len] = '\0'; 5172 } 5173 } 5174 bcopy(string, cnp->cn_pnbuf, len); 5175 5176 ndp->ni_pathlen += adjust; 5177 cache_fpl_pathlen_add(fpl, adjust); 5178 cnp->cn_nameptr = cnp->cn_pnbuf; 5179 fpl->nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1]; 5180 fpl->tvp = NULL; 5181 return (0); 5182 } 5183 5184 static int __noinline 5185 cache_fplookup_symlink(struct cache_fpl *fpl) 5186 { 5187 struct mount *mp; 5188 struct nameidata *ndp; 5189 struct componentname *cnp; 5190 struct vnode *dvp, *tvp; 5191 int error; 5192 5193 ndp = fpl->ndp; 5194 cnp = fpl->cnp; 5195 dvp = fpl->dvp; 5196 tvp = fpl->tvp; 5197 5198 if (cache_fpl_islastcn(ndp)) { 5199 if ((cnp->cn_flags & FOLLOW) == 0) { 5200 return (cache_fplookup_final(fpl)); 5201 } 5202 } 5203 5204 mp = atomic_load_ptr(&dvp->v_mount); 5205 if (__predict_false(mp == NULL)) { 5206 return (cache_fpl_aborted(fpl)); 5207 } 5208 5209 /* 5210 * Note this check races against setting the flag just like regular 5211 * lookup. 5212 */ 5213 if (__predict_false((mp->mnt_flag & MNT_NOSYMFOLLOW) != 0)) { 5214 cache_fpl_smr_exit(fpl); 5215 return (cache_fpl_handled_error(fpl, EACCES)); 5216 } 5217 5218 error = VOP_FPLOOKUP_SYMLINK(tvp, fpl); 5219 if (__predict_false(error != 0)) { 5220 switch (error) { 5221 case EAGAIN: 5222 return (cache_fpl_partial(fpl)); 5223 case ENOENT: 5224 case ENAMETOOLONG: 5225 case ELOOP: 5226 cache_fpl_smr_exit(fpl); 5227 return (cache_fpl_handled_error(fpl, error)); 5228 default: 5229 return (cache_fpl_aborted(fpl)); 5230 } 5231 } 5232 5233 if (*(cnp->cn_nameptr) == '/') { 5234 fpl->dvp = cache_fpl_handle_root(fpl); 5235 fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp); 5236 if (seqc_in_modify(fpl->dvp_seqc)) { 5237 return (cache_fpl_aborted(fpl)); 5238 } 5239 /* 5240 * The main loop assumes that ->dvp points to a vnode belonging 5241 * to a filesystem which can do lockless lookup, but the absolute 5242 * symlink can be wandering off to one which does not. 5243 */ 5244 mp = atomic_load_ptr(&fpl->dvp->v_mount); 5245 if (__predict_false(mp == NULL)) { 5246 return (cache_fpl_aborted(fpl)); 5247 } 5248 if (!cache_fplookup_mp_supported(mp)) { 5249 cache_fpl_checkpoint(fpl); 5250 return (cache_fpl_partial(fpl)); 5251 } 5252 } 5253 return (0); 5254 } 5255 5256 static int 5257 cache_fplookup_next(struct cache_fpl *fpl) 5258 { 5259 struct componentname *cnp; 5260 struct namecache *ncp; 5261 struct vnode *dvp, *tvp; 5262 u_char nc_flag; 5263 uint32_t hash; 5264 int error; 5265 5266 cnp = fpl->cnp; 5267 dvp = fpl->dvp; 5268 hash = fpl->hash; 5269 5270 if (__predict_false(cnp->cn_nameptr[0] == '.')) { 5271 if (cnp->cn_namelen == 1) { 5272 return (cache_fplookup_dot(fpl)); 5273 } 5274 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') { 5275 return (cache_fplookup_dotdot(fpl)); 5276 } 5277 } 5278 5279 MPASS(!cache_fpl_isdotdot(cnp)); 5280 5281 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 5282 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && 5283 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) 5284 break; 5285 } 5286 5287 if (__predict_false(ncp == NULL)) { 5288 return (cache_fplookup_noentry(fpl)); 5289 } 5290 5291 tvp = atomic_load_ptr(&ncp->nc_vp); 5292 nc_flag = atomic_load_char(&ncp->nc_flag); 5293 if ((nc_flag & NCF_NEGATIVE) != 0) { 5294 return (cache_fplookup_neg(fpl, ncp, hash)); 5295 } 5296 5297 if (!cache_ncp_canuse(ncp)) { 5298 return (cache_fpl_partial(fpl)); 5299 } 5300 5301 fpl->tvp = tvp; 5302 fpl->tvp_seqc = vn_seqc_read_any(tvp); 5303 if (seqc_in_modify(fpl->tvp_seqc)) { 5304 return (cache_fpl_partial(fpl)); 5305 } 5306 5307 counter_u64_add(numposhits, 1); 5308 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, tvp); 5309 5310 error = 0; 5311 if (cache_fplookup_is_mp(fpl)) { 5312 error = cache_fplookup_cross_mount(fpl); 5313 } 5314 return (error); 5315 } 5316 5317 static bool 5318 cache_fplookup_mp_supported(struct mount *mp) 5319 { 5320 5321 MPASS(mp != NULL); 5322 if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0) 5323 return (false); 5324 return (true); 5325 } 5326 5327 /* 5328 * Walk up the mount stack (if any). 5329 * 5330 * Correctness is provided in the following ways: 5331 * - all vnodes are protected from freeing with SMR 5332 * - struct mount objects are type stable making them always safe to access 5333 * - stability of the particular mount is provided by busying it 5334 * - relationship between the vnode which is mounted on and the mount is 5335 * verified with the vnode sequence counter after busying 5336 * - association between root vnode of the mount and the mount is protected 5337 * by busy 5338 * 5339 * From that point on we can read the sequence counter of the root vnode 5340 * and get the next mount on the stack (if any) using the same protection. 5341 * 5342 * By the end of successful walk we are guaranteed the reached state was 5343 * indeed present at least at some point which matches the regular lookup. 5344 */ 5345 static int __noinline 5346 cache_fplookup_climb_mount(struct cache_fpl *fpl) 5347 { 5348 struct mount *mp, *prev_mp; 5349 struct mount_pcpu *mpcpu, *prev_mpcpu; 5350 struct vnode *vp; 5351 seqc_t vp_seqc; 5352 5353 vp = fpl->tvp; 5354 vp_seqc = fpl->tvp_seqc; 5355 5356 VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp); 5357 mp = atomic_load_ptr(&vp->v_mountedhere); 5358 if (__predict_false(mp == NULL)) { 5359 return (0); 5360 } 5361 5362 prev_mp = NULL; 5363 for (;;) { 5364 if (!vfs_op_thread_enter_crit(mp, mpcpu)) { 5365 if (prev_mp != NULL) 5366 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu); 5367 return (cache_fpl_partial(fpl)); 5368 } 5369 if (prev_mp != NULL) 5370 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu); 5371 if (!vn_seqc_consistent(vp, vp_seqc)) { 5372 vfs_op_thread_exit_crit(mp, mpcpu); 5373 return (cache_fpl_partial(fpl)); 5374 } 5375 if (!cache_fplookup_mp_supported(mp)) { 5376 vfs_op_thread_exit_crit(mp, mpcpu); 5377 return (cache_fpl_partial(fpl)); 5378 } 5379 vp = atomic_load_ptr(&mp->mnt_rootvnode); 5380 if (vp == NULL) { 5381 vfs_op_thread_exit_crit(mp, mpcpu); 5382 return (cache_fpl_partial(fpl)); 5383 } 5384 vp_seqc = vn_seqc_read_any(vp); 5385 if (seqc_in_modify(vp_seqc)) { 5386 vfs_op_thread_exit_crit(mp, mpcpu); 5387 return (cache_fpl_partial(fpl)); 5388 } 5389 prev_mp = mp; 5390 prev_mpcpu = mpcpu; 5391 mp = atomic_load_ptr(&vp->v_mountedhere); 5392 if (mp == NULL) 5393 break; 5394 } 5395 5396 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu); 5397 fpl->tvp = vp; 5398 fpl->tvp_seqc = vp_seqc; 5399 return (0); 5400 } 5401 5402 static int __noinline 5403 cache_fplookup_cross_mount(struct cache_fpl *fpl) 5404 { 5405 struct mount *mp; 5406 struct mount_pcpu *mpcpu; 5407 struct vnode *vp; 5408 seqc_t vp_seqc; 5409 5410 vp = fpl->tvp; 5411 vp_seqc = fpl->tvp_seqc; 5412 5413 VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp); 5414 mp = atomic_load_ptr(&vp->v_mountedhere); 5415 if (__predict_false(mp == NULL)) { 5416 return (0); 5417 } 5418 5419 if (!vfs_op_thread_enter_crit(mp, mpcpu)) { 5420 return (cache_fpl_partial(fpl)); 5421 } 5422 if (!vn_seqc_consistent(vp, vp_seqc)) { 5423 vfs_op_thread_exit_crit(mp, mpcpu); 5424 return (cache_fpl_partial(fpl)); 5425 } 5426 if (!cache_fplookup_mp_supported(mp)) { 5427 vfs_op_thread_exit_crit(mp, mpcpu); 5428 return (cache_fpl_partial(fpl)); 5429 } 5430 vp = atomic_load_ptr(&mp->mnt_rootvnode); 5431 if (__predict_false(vp == NULL)) { 5432 vfs_op_thread_exit_crit(mp, mpcpu); 5433 return (cache_fpl_partial(fpl)); 5434 } 5435 vp_seqc = vn_seqc_read_any(vp); 5436 vfs_op_thread_exit_crit(mp, mpcpu); 5437 if (seqc_in_modify(vp_seqc)) { 5438 return (cache_fpl_partial(fpl)); 5439 } 5440 mp = atomic_load_ptr(&vp->v_mountedhere); 5441 if (__predict_false(mp != NULL)) { 5442 /* 5443 * There are possibly more mount points on top. 5444 * Normally this does not happen so for simplicity just start 5445 * over. 5446 */ 5447 return (cache_fplookup_climb_mount(fpl)); 5448 } 5449 5450 fpl->tvp = vp; 5451 fpl->tvp_seqc = vp_seqc; 5452 return (0); 5453 } 5454 5455 /* 5456 * Check if a vnode is mounted on. 5457 */ 5458 static bool 5459 cache_fplookup_is_mp(struct cache_fpl *fpl) 5460 { 5461 struct vnode *vp; 5462 5463 vp = fpl->tvp; 5464 return ((vn_irflag_read(vp) & VIRF_MOUNTPOINT) != 0); 5465 } 5466 5467 /* 5468 * Parse the path. 5469 * 5470 * The code was originally copy-pasted from regular lookup and despite 5471 * clean ups leaves performance on the table. Any modifications here 5472 * must take into account that in case off fallback the resulting 5473 * nameidata state has to be compatible with the original. 5474 */ 5475 5476 /* 5477 * Debug ni_pathlen tracking. 5478 */ 5479 #ifdef INVARIANTS 5480 static void 5481 cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n) 5482 { 5483 5484 fpl->debug.ni_pathlen += n; 5485 KASSERT(fpl->debug.ni_pathlen <= PATH_MAX, 5486 ("%s: pathlen overflow to %zd\n", __func__, fpl->debug.ni_pathlen)); 5487 } 5488 5489 static void 5490 cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n) 5491 { 5492 5493 fpl->debug.ni_pathlen -= n; 5494 KASSERT(fpl->debug.ni_pathlen <= PATH_MAX, 5495 ("%s: pathlen underflow to %zd\n", __func__, fpl->debug.ni_pathlen)); 5496 } 5497 5498 static void 5499 cache_fpl_pathlen_inc(struct cache_fpl *fpl) 5500 { 5501 5502 cache_fpl_pathlen_add(fpl, 1); 5503 } 5504 5505 static void 5506 cache_fpl_pathlen_dec(struct cache_fpl *fpl) 5507 { 5508 5509 cache_fpl_pathlen_sub(fpl, 1); 5510 } 5511 #else 5512 static void 5513 cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n) 5514 { 5515 } 5516 5517 static void 5518 cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n) 5519 { 5520 } 5521 5522 static void 5523 cache_fpl_pathlen_inc(struct cache_fpl *fpl) 5524 { 5525 } 5526 5527 static void 5528 cache_fpl_pathlen_dec(struct cache_fpl *fpl) 5529 { 5530 } 5531 #endif 5532 5533 static void 5534 cache_fplookup_parse(struct cache_fpl *fpl) 5535 { 5536 struct nameidata *ndp; 5537 struct componentname *cnp; 5538 struct vnode *dvp; 5539 char *cp; 5540 uint32_t hash; 5541 5542 ndp = fpl->ndp; 5543 cnp = fpl->cnp; 5544 dvp = fpl->dvp; 5545 5546 /* 5547 * Find the end of this path component, it is either / or nul. 5548 * 5549 * Store / as a temporary sentinel so that we only have one character 5550 * to test for. Pathnames tend to be short so this should not be 5551 * resulting in cache misses. 5552 * 5553 * TODO: fix this to be word-sized. 5554 */ 5555 MPASS(&cnp->cn_nameptr[fpl->debug.ni_pathlen - 1] >= cnp->cn_pnbuf); 5556 KASSERT(&cnp->cn_nameptr[fpl->debug.ni_pathlen - 1] == fpl->nulchar, 5557 ("%s: mismatch between pathlen (%zu) and nulchar (%p != %p), string [%s]\n", 5558 __func__, fpl->debug.ni_pathlen, &cnp->cn_nameptr[fpl->debug.ni_pathlen - 1], 5559 fpl->nulchar, cnp->cn_pnbuf)); 5560 KASSERT(*fpl->nulchar == '\0', 5561 ("%s: expected nul at %p; string [%s]\n", __func__, fpl->nulchar, 5562 cnp->cn_pnbuf)); 5563 hash = cache_get_hash_iter_start(dvp); 5564 *fpl->nulchar = '/'; 5565 for (cp = cnp->cn_nameptr; *cp != '/'; cp++) { 5566 KASSERT(*cp != '\0', 5567 ("%s: encountered unexpected nul; string [%s]\n", __func__, 5568 cnp->cn_nameptr)); 5569 hash = cache_get_hash_iter(*cp, hash); 5570 continue; 5571 } 5572 *fpl->nulchar = '\0'; 5573 fpl->hash = cache_get_hash_iter_finish(hash); 5574 5575 cnp->cn_namelen = cp - cnp->cn_nameptr; 5576 cache_fpl_pathlen_sub(fpl, cnp->cn_namelen); 5577 5578 #ifdef INVARIANTS 5579 /* 5580 * cache_get_hash only accepts lengths up to NAME_MAX. This is fine since 5581 * we are going to fail this lookup with ENAMETOOLONG (see below). 5582 */ 5583 if (cnp->cn_namelen <= NAME_MAX) { 5584 if (fpl->hash != cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp)) { 5585 panic("%s: mismatched hash for [%s] len %ld", __func__, 5586 cnp->cn_nameptr, cnp->cn_namelen); 5587 } 5588 } 5589 #endif 5590 5591 /* 5592 * Hack: we have to check if the found path component's length exceeds 5593 * NAME_MAX. However, the condition is very rarely true and check can 5594 * be elided in the common case -- if an entry was found in the cache, 5595 * then it could not have been too long to begin with. 5596 */ 5597 ndp->ni_next = cp; 5598 } 5599 5600 static void 5601 cache_fplookup_parse_advance(struct cache_fpl *fpl) 5602 { 5603 struct nameidata *ndp; 5604 struct componentname *cnp; 5605 5606 ndp = fpl->ndp; 5607 cnp = fpl->cnp; 5608 5609 cnp->cn_nameptr = ndp->ni_next; 5610 KASSERT(*(cnp->cn_nameptr) == '/', 5611 ("%s: should have seen slash at %p ; buf %p [%s]\n", __func__, 5612 cnp->cn_nameptr, cnp->cn_pnbuf, cnp->cn_pnbuf)); 5613 cnp->cn_nameptr++; 5614 cache_fpl_pathlen_dec(fpl); 5615 } 5616 5617 /* 5618 * Skip spurious slashes in a pathname (e.g., "foo///bar") and retry. 5619 * 5620 * Lockless lookup tries to elide checking for spurious slashes and should they 5621 * be present is guaranteed to fail to find an entry. In this case the caller 5622 * must check if the name starts with a slash and call this routine. It is 5623 * going to fast forward across the spurious slashes and set the state up for 5624 * retry. 5625 */ 5626 static int __noinline 5627 cache_fplookup_skip_slashes(struct cache_fpl *fpl) 5628 { 5629 struct nameidata *ndp; 5630 struct componentname *cnp; 5631 5632 ndp = fpl->ndp; 5633 cnp = fpl->cnp; 5634 5635 MPASS(*(cnp->cn_nameptr) == '/'); 5636 do { 5637 cnp->cn_nameptr++; 5638 cache_fpl_pathlen_dec(fpl); 5639 } while (*(cnp->cn_nameptr) == '/'); 5640 5641 /* 5642 * Go back to one slash so that cache_fplookup_parse_advance has 5643 * something to skip. 5644 */ 5645 cnp->cn_nameptr--; 5646 cache_fpl_pathlen_inc(fpl); 5647 5648 /* 5649 * cache_fplookup_parse_advance starts from ndp->ni_next 5650 */ 5651 ndp->ni_next = cnp->cn_nameptr; 5652 5653 /* 5654 * See cache_fplookup_dot. 5655 */ 5656 fpl->tvp = fpl->dvp; 5657 fpl->tvp_seqc = fpl->dvp_seqc; 5658 5659 return (0); 5660 } 5661 5662 /* 5663 * Handle trailing slashes (e.g., "foo/"). 5664 * 5665 * If a trailing slash is found the terminal vnode must be a directory. 5666 * Regular lookup shortens the path by nulifying the first trailing slash and 5667 * sets the TRAILINGSLASH flag to denote this took place. There are several 5668 * checks on it performed later. 5669 * 5670 * Similarly to spurious slashes, lockless lookup handles this in a speculative 5671 * manner relying on an invariant that a non-directory vnode will get a miss. 5672 * In this case cn_nameptr[0] == '\0' and cn_namelen == 0. 5673 * 5674 * Thus for a path like "foo/bar/" the code unwinds the state back to "bar/" 5675 * and denotes this is the last path component, which avoids looping back. 5676 * 5677 * Only plain lookups are supported for now to restrict corner cases to handle. 5678 */ 5679 static int __noinline 5680 cache_fplookup_trailingslash(struct cache_fpl *fpl) 5681 { 5682 #ifdef INVARIANTS 5683 size_t ni_pathlen; 5684 #endif 5685 struct nameidata *ndp; 5686 struct componentname *cnp; 5687 struct namecache *ncp; 5688 struct vnode *tvp; 5689 char *cn_nameptr_orig, *cn_nameptr_slash; 5690 seqc_t tvp_seqc; 5691 u_char nc_flag; 5692 5693 ndp = fpl->ndp; 5694 cnp = fpl->cnp; 5695 tvp = fpl->tvp; 5696 tvp_seqc = fpl->tvp_seqc; 5697 5698 MPASS(fpl->dvp == fpl->tvp); 5699 KASSERT(cache_fpl_istrailingslash(fpl), 5700 ("%s: expected trailing slash at %p; string [%s]\n", __func__, fpl->nulchar - 1, 5701 cnp->cn_pnbuf)); 5702 KASSERT(cnp->cn_nameptr[0] == '\0', 5703 ("%s: expected nul char at %p; string [%s]\n", __func__, &cnp->cn_nameptr[0], 5704 cnp->cn_pnbuf)); 5705 KASSERT(cnp->cn_namelen == 0, 5706 ("%s: namelen 0 but got %ld; string [%s]\n", __func__, cnp->cn_namelen, 5707 cnp->cn_pnbuf)); 5708 MPASS(cnp->cn_nameptr > cnp->cn_pnbuf); 5709 5710 if (cnp->cn_nameiop != LOOKUP) { 5711 return (cache_fpl_aborted(fpl)); 5712 } 5713 5714 if (__predict_false(tvp->v_type != VDIR)) { 5715 if (!vn_seqc_consistent(tvp, tvp_seqc)) { 5716 return (cache_fpl_aborted(fpl)); 5717 } 5718 cache_fpl_smr_exit(fpl); 5719 return (cache_fpl_handled_error(fpl, ENOTDIR)); 5720 } 5721 5722 /* 5723 * Denote the last component. 5724 */ 5725 ndp->ni_next = &cnp->cn_nameptr[0]; 5726 MPASS(cache_fpl_islastcn(ndp)); 5727 5728 /* 5729 * Unwind trailing slashes. 5730 */ 5731 cn_nameptr_orig = cnp->cn_nameptr; 5732 while (cnp->cn_nameptr >= cnp->cn_pnbuf) { 5733 cnp->cn_nameptr--; 5734 if (cnp->cn_nameptr[0] != '/') { 5735 break; 5736 } 5737 } 5738 5739 /* 5740 * Unwind to the beginning of the path component. 5741 * 5742 * Note the path may or may not have started with a slash. 5743 */ 5744 cn_nameptr_slash = cnp->cn_nameptr; 5745 while (cnp->cn_nameptr > cnp->cn_pnbuf) { 5746 cnp->cn_nameptr--; 5747 if (cnp->cn_nameptr[0] == '/') { 5748 break; 5749 } 5750 } 5751 if (cnp->cn_nameptr[0] == '/') { 5752 cnp->cn_nameptr++; 5753 } 5754 5755 cnp->cn_namelen = cn_nameptr_slash - cnp->cn_nameptr + 1; 5756 cache_fpl_pathlen_add(fpl, cn_nameptr_orig - cnp->cn_nameptr); 5757 cache_fpl_checkpoint(fpl); 5758 5759 #ifdef INVARIANTS 5760 ni_pathlen = fpl->nulchar - cnp->cn_nameptr + 1; 5761 if (ni_pathlen != fpl->debug.ni_pathlen) { 5762 panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n", 5763 __func__, ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar, 5764 cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf); 5765 } 5766 #endif 5767 5768 /* 5769 * If this was a "./" lookup the parent directory is already correct. 5770 */ 5771 if (cnp->cn_nameptr[0] == '.' && cnp->cn_namelen == 1) { 5772 return (0); 5773 } 5774 5775 /* 5776 * Otherwise we need to look it up. 5777 */ 5778 tvp = fpl->tvp; 5779 ncp = atomic_load_consume_ptr(&tvp->v_cache_dd); 5780 if (__predict_false(ncp == NULL)) { 5781 return (cache_fpl_aborted(fpl)); 5782 } 5783 nc_flag = atomic_load_char(&ncp->nc_flag); 5784 if ((nc_flag & NCF_ISDOTDOT) != 0) { 5785 return (cache_fpl_aborted(fpl)); 5786 } 5787 fpl->dvp = ncp->nc_dvp; 5788 fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp); 5789 if (seqc_in_modify(fpl->dvp_seqc)) { 5790 return (cache_fpl_aborted(fpl)); 5791 } 5792 return (0); 5793 } 5794 5795 /* 5796 * See the API contract for VOP_FPLOOKUP_VEXEC. 5797 */ 5798 static int __noinline 5799 cache_fplookup_failed_vexec(struct cache_fpl *fpl, int error) 5800 { 5801 struct componentname *cnp; 5802 struct vnode *dvp; 5803 seqc_t dvp_seqc; 5804 5805 cnp = fpl->cnp; 5806 dvp = fpl->dvp; 5807 dvp_seqc = fpl->dvp_seqc; 5808 5809 /* 5810 * Hack: delayed empty path checking. 5811 */ 5812 if (cnp->cn_pnbuf[0] == '\0') { 5813 return (cache_fplookup_emptypath(fpl)); 5814 } 5815 5816 /* 5817 * TODO: Due to ignoring trailing slashes lookup will perform a 5818 * permission check on the last dir when it should not be doing it. It 5819 * may fail, but said failure should be ignored. It is possible to fix 5820 * it up fully without resorting to regular lookup, but for now just 5821 * abort. 5822 */ 5823 if (cache_fpl_istrailingslash(fpl)) { 5824 return (cache_fpl_aborted(fpl)); 5825 } 5826 5827 /* 5828 * Hack: delayed degenerate path checking. 5829 */ 5830 if (cnp->cn_nameptr[0] == '\0' && fpl->tvp == NULL) { 5831 return (cache_fplookup_degenerate(fpl)); 5832 } 5833 5834 /* 5835 * Hack: delayed name len checking. 5836 */ 5837 if (__predict_false(cnp->cn_namelen > NAME_MAX)) { 5838 cache_fpl_smr_exit(fpl); 5839 return (cache_fpl_handled_error(fpl, ENAMETOOLONG)); 5840 } 5841 5842 /* 5843 * Hack: they may be looking up foo/bar, where foo is not a directory. 5844 * In such a case we need to return ENOTDIR, but we may happen to get 5845 * here with a different error. 5846 */ 5847 if (dvp->v_type != VDIR) { 5848 error = ENOTDIR; 5849 } 5850 5851 /* 5852 * Hack: handle O_SEARCH. 5853 * 5854 * Open Group Base Specifications Issue 7, 2018 edition states: 5855 * <quote> 5856 * If the access mode of the open file description associated with the 5857 * file descriptor is not O_SEARCH, the function shall check whether 5858 * directory searches are permitted using the current permissions of 5859 * the directory underlying the file descriptor. If the access mode is 5860 * O_SEARCH, the function shall not perform the check. 5861 * </quote> 5862 * 5863 * Regular lookup tests for the NOEXECCHECK flag for every path 5864 * component to decide whether to do the permission check. However, 5865 * since most lookups never have the flag (and when they do it is only 5866 * present for the first path component), lockless lookup only acts on 5867 * it if there is a permission problem. Here the flag is represented 5868 * with a boolean so that we don't have to clear it on the way out. 5869 * 5870 * For simplicity this always aborts. 5871 * TODO: check if this is the first lookup and ignore the permission 5872 * problem. Note the flag has to survive fallback (if it happens to be 5873 * performed). 5874 */ 5875 if (fpl->fsearch) { 5876 return (cache_fpl_aborted(fpl)); 5877 } 5878 5879 switch (error) { 5880 case EAGAIN: 5881 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 5882 error = cache_fpl_aborted(fpl); 5883 } else { 5884 cache_fpl_partial(fpl); 5885 } 5886 break; 5887 default: 5888 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 5889 error = cache_fpl_aborted(fpl); 5890 } else { 5891 cache_fpl_smr_exit(fpl); 5892 cache_fpl_handled_error(fpl, error); 5893 } 5894 break; 5895 } 5896 return (error); 5897 } 5898 5899 static int 5900 cache_fplookup_impl(struct vnode *dvp, struct cache_fpl *fpl) 5901 { 5902 struct nameidata *ndp; 5903 struct componentname *cnp; 5904 struct mount *mp; 5905 int error; 5906 5907 ndp = fpl->ndp; 5908 cnp = fpl->cnp; 5909 5910 cache_fpl_checkpoint(fpl); 5911 5912 /* 5913 * The vnode at hand is almost always stable, skip checking for it. 5914 * Worst case this postpones the check towards the end of the iteration 5915 * of the main loop. 5916 */ 5917 fpl->dvp = dvp; 5918 fpl->dvp_seqc = vn_seqc_read_notmodify(fpl->dvp); 5919 5920 mp = atomic_load_ptr(&dvp->v_mount); 5921 if (__predict_false(mp == NULL || !cache_fplookup_mp_supported(mp))) { 5922 return (cache_fpl_aborted(fpl)); 5923 } 5924 5925 MPASS(fpl->tvp == NULL); 5926 5927 for (;;) { 5928 cache_fplookup_parse(fpl); 5929 5930 error = VOP_FPLOOKUP_VEXEC(fpl->dvp, cnp->cn_cred); 5931 if (__predict_false(error != 0)) { 5932 error = cache_fplookup_failed_vexec(fpl, error); 5933 break; 5934 } 5935 5936 error = cache_fplookup_next(fpl); 5937 if (__predict_false(cache_fpl_terminated(fpl))) { 5938 break; 5939 } 5940 5941 VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp); 5942 5943 if (fpl->tvp->v_type == VLNK) { 5944 error = cache_fplookup_symlink(fpl); 5945 if (cache_fpl_terminated(fpl)) { 5946 break; 5947 } 5948 } else { 5949 if (cache_fpl_islastcn(ndp)) { 5950 error = cache_fplookup_final(fpl); 5951 break; 5952 } 5953 5954 if (!vn_seqc_consistent(fpl->dvp, fpl->dvp_seqc)) { 5955 error = cache_fpl_aborted(fpl); 5956 break; 5957 } 5958 5959 fpl->dvp = fpl->tvp; 5960 fpl->dvp_seqc = fpl->tvp_seqc; 5961 cache_fplookup_parse_advance(fpl); 5962 } 5963 5964 cache_fpl_checkpoint(fpl); 5965 } 5966 5967 return (error); 5968 } 5969 5970 /* 5971 * Fast path lookup protected with SMR and sequence counters. 5972 * 5973 * Note: all VOP_FPLOOKUP_VEXEC routines have a comment referencing this one. 5974 * 5975 * Filesystems can opt in by setting the MNTK_FPLOOKUP flag and meeting criteria 5976 * outlined below. 5977 * 5978 * Traditional vnode lookup conceptually looks like this: 5979 * 5980 * vn_lock(current); 5981 * for (;;) { 5982 * next = find(); 5983 * vn_lock(next); 5984 * vn_unlock(current); 5985 * current = next; 5986 * if (last) 5987 * break; 5988 * } 5989 * return (current); 5990 * 5991 * Each jump to the next vnode is safe memory-wise and atomic with respect to 5992 * any modifications thanks to holding respective locks. 5993 * 5994 * The same guarantee can be provided with a combination of safe memory 5995 * reclamation and sequence counters instead. If all operations which affect 5996 * the relationship between the current vnode and the one we are looking for 5997 * also modify the counter, we can verify whether all the conditions held as 5998 * we made the jump. This includes things like permissions, mount points etc. 5999 * Counter modification is provided by enclosing relevant places in 6000 * vn_seqc_write_begin()/end() calls. 6001 * 6002 * Thus this translates to: 6003 * 6004 * vfs_smr_enter(); 6005 * dvp_seqc = seqc_read_any(dvp); 6006 * if (seqc_in_modify(dvp_seqc)) // someone is altering the vnode 6007 * abort(); 6008 * for (;;) { 6009 * tvp = find(); 6010 * tvp_seqc = seqc_read_any(tvp); 6011 * if (seqc_in_modify(tvp_seqc)) // someone is altering the target vnode 6012 * abort(); 6013 * if (!seqc_consistent(dvp, dvp_seqc) // someone is altering the vnode 6014 * abort(); 6015 * dvp = tvp; // we know nothing of importance has changed 6016 * dvp_seqc = tvp_seqc; // store the counter for the tvp iteration 6017 * if (last) 6018 * break; 6019 * } 6020 * vget(); // secure the vnode 6021 * if (!seqc_consistent(tvp, tvp_seqc) // final check 6022 * abort(); 6023 * // at this point we know nothing has changed for any parent<->child pair 6024 * // as they were crossed during the lookup, meaning we matched the guarantee 6025 * // of the locked variant 6026 * return (tvp); 6027 * 6028 * The API contract for VOP_FPLOOKUP_VEXEC routines is as follows: 6029 * - they are called while within vfs_smr protection which they must never exit 6030 * - EAGAIN can be returned to denote checking could not be performed, it is 6031 * always valid to return it 6032 * - if the sequence counter has not changed the result must be valid 6033 * - if the sequence counter has changed both false positives and false negatives 6034 * are permitted (since the result will be rejected later) 6035 * - for simple cases of unix permission checks vaccess_vexec_smr can be used 6036 * 6037 * Caveats to watch out for: 6038 * - vnodes are passed unlocked and unreferenced with nothing stopping 6039 * VOP_RECLAIM, in turn meaning that ->v_data can become NULL. It is advised 6040 * to use atomic_load_ptr to fetch it. 6041 * - the aforementioned object can also get freed, meaning absent other means it 6042 * should be protected with vfs_smr 6043 * - either safely checking permissions as they are modified or guaranteeing 6044 * their stability is left to the routine 6045 */ 6046 int 6047 cache_fplookup(struct nameidata *ndp, enum cache_fpl_status *status, 6048 struct pwd **pwdp) 6049 { 6050 struct cache_fpl fpl; 6051 struct pwd *pwd; 6052 struct vnode *dvp; 6053 struct componentname *cnp; 6054 int error; 6055 6056 fpl.status = CACHE_FPL_STATUS_UNSET; 6057 fpl.in_smr = false; 6058 fpl.ndp = ndp; 6059 fpl.cnp = cnp = &ndp->ni_cnd; 6060 MPASS(ndp->ni_lcf == 0); 6061 KASSERT ((cnp->cn_flags & CACHE_FPL_INTERNAL_CN_FLAGS) == 0, 6062 ("%s: internal flags found in cn_flags %" PRIx64, __func__, 6063 cnp->cn_flags)); 6064 if ((cnp->cn_flags & SAVESTART) != 0) { 6065 MPASS(cnp->cn_nameiop != LOOKUP); 6066 } 6067 MPASS(cnp->cn_nameptr == cnp->cn_pnbuf); 6068 6069 if (__predict_false(!cache_can_fplookup(&fpl))) { 6070 *status = fpl.status; 6071 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status); 6072 return (EOPNOTSUPP); 6073 } 6074 6075 cache_fpl_checkpoint_outer(&fpl); 6076 6077 cache_fpl_smr_enter_initial(&fpl); 6078 #ifdef INVARIANTS 6079 fpl.debug.ni_pathlen = ndp->ni_pathlen; 6080 #endif 6081 fpl.nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1]; 6082 fpl.fsearch = false; 6083 fpl.savename = (cnp->cn_flags & SAVENAME) != 0; 6084 fpl.tvp = NULL; /* for degenerate path handling */ 6085 fpl.pwd = pwdp; 6086 pwd = pwd_get_smr(); 6087 *(fpl.pwd) = pwd; 6088 ndp->ni_rootdir = pwd->pwd_rdir; 6089 ndp->ni_topdir = pwd->pwd_jdir; 6090 6091 if (cnp->cn_pnbuf[0] == '/') { 6092 dvp = cache_fpl_handle_root(&fpl); 6093 MPASS(ndp->ni_resflags == 0); 6094 ndp->ni_resflags = NIRES_ABS; 6095 } else { 6096 if (ndp->ni_dirfd == AT_FDCWD) { 6097 dvp = pwd->pwd_cdir; 6098 } else { 6099 error = cache_fplookup_dirfd(&fpl, &dvp); 6100 if (__predict_false(error != 0)) { 6101 goto out; 6102 } 6103 } 6104 } 6105 6106 SDT_PROBE4(vfs, namei, lookup, entry, dvp, cnp->cn_pnbuf, cnp->cn_flags, true); 6107 error = cache_fplookup_impl(dvp, &fpl); 6108 out: 6109 cache_fpl_smr_assert_not_entered(&fpl); 6110 cache_fpl_assert_status(&fpl); 6111 *status = fpl.status; 6112 if (SDT_PROBES_ENABLED()) { 6113 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status); 6114 if (fpl.status == CACHE_FPL_STATUS_HANDLED) 6115 SDT_PROBE4(vfs, namei, lookup, return, error, ndp->ni_vp, true, 6116 ndp); 6117 } 6118 6119 if (__predict_true(fpl.status == CACHE_FPL_STATUS_HANDLED)) { 6120 MPASS(error != CACHE_FPL_FAILED); 6121 if (error != 0) { 6122 MPASS(fpl.dvp == NULL); 6123 MPASS(fpl.tvp == NULL); 6124 MPASS(fpl.savename == false); 6125 } 6126 ndp->ni_dvp = fpl.dvp; 6127 ndp->ni_vp = fpl.tvp; 6128 if (fpl.savename) { 6129 cnp->cn_flags |= HASBUF; 6130 } else { 6131 cache_fpl_cleanup_cnp(cnp); 6132 } 6133 } 6134 return (error); 6135 } 6136