1 /*- 2 * Copyright (c) 1989, 1993, 1995 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * Poul-Henning Kamp of the FreeBSD Project. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * @(#)vfs_cache.c 8.5 (Berkeley) 3/22/95 33 */ 34 35 #include <sys/cdefs.h> 36 __FBSDID("$FreeBSD$"); 37 38 #include "opt_ktrace.h" 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/counter.h> 43 #include <sys/filedesc.h> 44 #include <sys/fnv_hash.h> 45 #include <sys/kernel.h> 46 #include <sys/lock.h> 47 #include <sys/malloc.h> 48 #include <sys/fcntl.h> 49 #include <sys/mount.h> 50 #include <sys/namei.h> 51 #include <sys/proc.h> 52 #include <sys/rwlock.h> 53 #include <sys/sdt.h> 54 #include <sys/smp.h> 55 #include <sys/syscallsubr.h> 56 #include <sys/sysctl.h> 57 #include <sys/sysproto.h> 58 #include <sys/vnode.h> 59 #ifdef KTRACE 60 #include <sys/ktrace.h> 61 #endif 62 63 #include <vm/uma.h> 64 65 SDT_PROVIDER_DECLARE(vfs); 66 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *", 67 "struct vnode *"); 68 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *", 69 "char *"); 70 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *"); 71 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *", 72 "char *", "struct vnode *"); 73 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *"); 74 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int", 75 "struct vnode *", "char *"); 76 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *", 77 "struct vnode *"); 78 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative, 79 "struct vnode *", "char *"); 80 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *", 81 "char *"); 82 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *"); 83 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *"); 84 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *"); 85 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *", 86 "struct vnode *"); 87 SDT_PROBE_DEFINE3(vfs, namecache, zap_negative, done, "struct vnode *", 88 "char *", "int"); 89 SDT_PROBE_DEFINE3(vfs, namecache, shrink_negative, done, "struct vnode *", 90 "char *", "int"); 91 92 /* 93 * This structure describes the elements in the cache of recent 94 * names looked up by namei. 95 */ 96 97 struct namecache { 98 LIST_ENTRY(namecache) nc_hash; /* hash chain */ 99 LIST_ENTRY(namecache) nc_src; /* source vnode list */ 100 TAILQ_ENTRY(namecache) nc_dst; /* destination vnode list */ 101 struct vnode *nc_dvp; /* vnode of parent of name */ 102 union { 103 struct vnode *nu_vp; /* vnode the name refers to */ 104 u_int nu_neghits; /* negative entry hits */ 105 } n_un; 106 u_char nc_flag; /* flag bits */ 107 u_char nc_nlen; /* length of name */ 108 char nc_name[0]; /* segment name + nul */ 109 }; 110 111 /* 112 * struct namecache_ts repeats struct namecache layout up to the 113 * nc_nlen member. 114 * struct namecache_ts is used in place of struct namecache when time(s) need 115 * to be stored. The nc_dotdottime field is used when a cache entry is mapping 116 * both a non-dotdot directory name plus dotdot for the directory's 117 * parent. 118 */ 119 struct namecache_ts { 120 LIST_ENTRY(namecache) nc_hash; /* hash chain */ 121 LIST_ENTRY(namecache) nc_src; /* source vnode list */ 122 TAILQ_ENTRY(namecache) nc_dst; /* destination vnode list */ 123 struct vnode *nc_dvp; /* vnode of parent of name */ 124 union { 125 struct vnode *nu_vp; /* vnode the name refers to */ 126 u_int nu_neghits; /* negative entry hits */ 127 } n_un; 128 u_char nc_flag; /* flag bits */ 129 u_char nc_nlen; /* length of name */ 130 struct timespec nc_time; /* timespec provided by fs */ 131 struct timespec nc_dotdottime; /* dotdot timespec provided by fs */ 132 int nc_ticks; /* ticks value when entry was added */ 133 char nc_name[0]; /* segment name + nul */ 134 }; 135 136 #define nc_vp n_un.nu_vp 137 #define nc_neghits n_un.nu_neghits 138 139 /* 140 * Flags in namecache.nc_flag 141 */ 142 #define NCF_WHITE 0x01 143 #define NCF_ISDOTDOT 0x02 144 #define NCF_TS 0x04 145 #define NCF_DTS 0x08 146 #define NCF_DVDROP 0x10 147 #define NCF_NEGATIVE 0x20 148 #define NCF_HOTNEGATIVE 0x40 149 150 /* 151 * Name caching works as follows: 152 * 153 * Names found by directory scans are retained in a cache 154 * for future reference. It is managed LRU, so frequently 155 * used names will hang around. Cache is indexed by hash value 156 * obtained from (vp, name) where vp refers to the directory 157 * containing name. 158 * 159 * If it is a "negative" entry, (i.e. for a name that is known NOT to 160 * exist) the vnode pointer will be NULL. 161 * 162 * Upon reaching the last segment of a path, if the reference 163 * is for DELETE, or NOCACHE is set (rewrite), and the 164 * name is located in the cache, it will be dropped. 165 * 166 * These locks are used (in the order in which they can be taken): 167 * NAME TYPE ROLE 168 * vnodelock mtx vnode lists and v_cache_dd field protection 169 * bucketlock rwlock for access to given set of hash buckets 170 * neglist mtx negative entry LRU management 171 * 172 * Additionally, ncneg_shrink_lock mtx is used to have at most one thread 173 * shrinking the LRU list. 174 * 175 * It is legal to take multiple vnodelock and bucketlock locks. The locking 176 * order is lower address first. Both are recursive. 177 * 178 * "." lookups are lockless. 179 * 180 * ".." and vnode -> name lookups require vnodelock. 181 * 182 * name -> vnode lookup requires the relevant bucketlock to be held for reading. 183 * 184 * Insertions and removals of entries require involved vnodes and bucketlocks 185 * to be write-locked to prevent other threads from seeing the entry. 186 * 187 * Some lookups result in removal of the found entry (e.g. getting rid of a 188 * negative entry with the intent to create a positive one), which poses a 189 * problem when multiple threads reach the state. Similarly, two different 190 * threads can purge two different vnodes and try to remove the same name. 191 * 192 * If the already held vnode lock is lower than the second required lock, we 193 * can just take the other lock. However, in the opposite case, this could 194 * deadlock. As such, this is resolved by trylocking and if that fails unlocking 195 * the first node, locking everything in order and revalidating the state. 196 */ 197 198 /* 199 * Structures associated with name caching. 200 */ 201 #define NCHHASH(hash) \ 202 (&nchashtbl[(hash) & nchash]) 203 static LIST_HEAD(nchashhead, namecache) *nchashtbl; /* Hash Table */ 204 static u_long nchash; /* size of hash table */ 205 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, 206 "Size of namecache hash table"); 207 static u_long ncnegfactor = 16; /* ratio of negative entries */ 208 SYSCTL_ULONG(_vfs, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0, 209 "Ratio of negative namecache entries"); 210 static u_long numneg; /* number of negative entries allocated */ 211 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0, 212 "Number of negative entries in namecache"); 213 static u_long numcache; /* number of cache entries allocated */ 214 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0, 215 "Number of namecache entries"); 216 static u_long numcachehv; /* number of cache entries with vnodes held */ 217 SYSCTL_ULONG(_debug, OID_AUTO, numcachehv, CTLFLAG_RD, &numcachehv, 0, 218 "Number of namecache entries with vnodes held"); 219 u_int ncsizefactor = 2; 220 SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0, 221 "Size factor for namecache"); 222 static u_int ncpurgeminvnodes; 223 SYSCTL_UINT(_vfs, OID_AUTO, ncpurgeminvnodes, CTLFLAG_RW, &ncpurgeminvnodes, 0, 224 "Number of vnodes below which purgevfs ignores the request"); 225 static u_int ncneghitsrequeue = 8; 226 SYSCTL_UINT(_vfs, OID_AUTO, ncneghitsrequeue, CTLFLAG_RW, &ncneghitsrequeue, 0, 227 "Number of hits to requeue a negative entry in the LRU list"); 228 229 struct nchstats nchstats; /* cache effectiveness statistics */ 230 231 static struct mtx ncneg_shrink_lock; 232 MTX_SYSINIT(vfscache_shrink_neg, &ncneg_shrink_lock, "Name Cache shrink neg", 233 MTX_DEF); 234 235 struct neglist { 236 struct mtx nl_lock; 237 TAILQ_HEAD(, namecache) nl_list; 238 } __aligned(CACHE_LINE_SIZE); 239 240 static struct neglist *neglists; 241 static struct neglist ncneg_hot; 242 243 static int shrink_list_turn; 244 245 static u_int numneglists; 246 static inline struct neglist * 247 NCP2NEGLIST(struct namecache *ncp) 248 { 249 250 return (&neglists[(((uintptr_t)(ncp) >> 8) % numneglists)]); 251 } 252 253 static u_int numbucketlocks; 254 static struct rwlock_padalign *bucketlocks; 255 #define HASH2BUCKETLOCK(hash) \ 256 ((struct rwlock *)(&bucketlocks[((hash) % numbucketlocks)])) 257 258 static u_int numvnodelocks; 259 static struct mtx *vnodelocks; 260 static inline struct mtx * 261 VP2VNODELOCK(struct vnode *vp) 262 { 263 struct mtx *vlp; 264 265 if (vp == NULL) 266 return (NULL); 267 vlp = &vnodelocks[(((uintptr_t)(vp) >> 8) % numvnodelocks)]; 268 return (vlp); 269 } 270 271 /* 272 * UMA zones for the VFS cache. 273 * 274 * The small cache is used for entries with short names, which are the 275 * most common. The large cache is used for entries which are too big to 276 * fit in the small cache. 277 */ 278 static uma_zone_t cache_zone_small; 279 static uma_zone_t cache_zone_small_ts; 280 static uma_zone_t cache_zone_large; 281 static uma_zone_t cache_zone_large_ts; 282 283 #define CACHE_PATH_CUTOFF 35 284 285 static struct namecache * 286 cache_alloc(int len, int ts) 287 { 288 289 if (len > CACHE_PATH_CUTOFF) { 290 if (ts) 291 return (uma_zalloc(cache_zone_large_ts, M_WAITOK)); 292 else 293 return (uma_zalloc(cache_zone_large, M_WAITOK)); 294 } 295 if (ts) 296 return (uma_zalloc(cache_zone_small_ts, M_WAITOK)); 297 else 298 return (uma_zalloc(cache_zone_small, M_WAITOK)); 299 } 300 301 static void 302 cache_free(struct namecache *ncp) 303 { 304 int ts; 305 306 if (ncp == NULL) 307 return; 308 ts = ncp->nc_flag & NCF_TS; 309 if ((ncp->nc_flag & NCF_DVDROP) != 0) 310 vdrop(ncp->nc_dvp); 311 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF) { 312 if (ts) 313 uma_zfree(cache_zone_small_ts, ncp); 314 else 315 uma_zfree(cache_zone_small, ncp); 316 } else if (ts) 317 uma_zfree(cache_zone_large_ts, ncp); 318 else 319 uma_zfree(cache_zone_large, ncp); 320 } 321 322 static char * 323 nc_get_name(struct namecache *ncp) 324 { 325 struct namecache_ts *ncp_ts; 326 327 if ((ncp->nc_flag & NCF_TS) == 0) 328 return (ncp->nc_name); 329 ncp_ts = (struct namecache_ts *)ncp; 330 return (ncp_ts->nc_name); 331 } 332 333 static void 334 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp) 335 { 336 337 KASSERT((ncp->nc_flag & NCF_TS) != 0 || 338 (tsp == NULL && ticksp == NULL), 339 ("No NCF_TS")); 340 341 if (tsp != NULL) 342 *tsp = ((struct namecache_ts *)ncp)->nc_time; 343 if (ticksp != NULL) 344 *ticksp = ((struct namecache_ts *)ncp)->nc_ticks; 345 } 346 347 static int doingcache = 1; /* 1 => enable the cache */ 348 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0, 349 "VFS namecache enabled"); 350 351 /* Export size information to userland */ 352 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR, 353 sizeof(struct namecache), "sizeof(struct namecache)"); 354 355 /* 356 * The new name cache statistics 357 */ 358 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, 359 "Name cache statistics"); 360 #define STATNODE_ULONG(name, descr) \ 361 SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr); 362 #define STATNODE_COUNTER(name, descr) \ 363 static counter_u64_t name; \ 364 SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, descr); 365 STATNODE_ULONG(numneg, "Number of negative cache entries"); 366 STATNODE_ULONG(numcache, "Number of cache entries"); 367 STATNODE_COUNTER(numcalls, "Number of cache lookups"); 368 STATNODE_COUNTER(dothits, "Number of '.' hits"); 369 STATNODE_COUNTER(dotdothits, "Number of '..' hits"); 370 STATNODE_COUNTER(numchecks, "Number of checks in lookup"); 371 STATNODE_COUNTER(nummiss, "Number of cache misses"); 372 STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache"); 373 STATNODE_COUNTER(numposzaps, 374 "Number of cache hits (positive) we do not want to cache"); 375 STATNODE_COUNTER(numposhits, "Number of cache hits (positive)"); 376 STATNODE_COUNTER(numnegzaps, 377 "Number of cache hits (negative) we do not want to cache"); 378 STATNODE_COUNTER(numneghits, "Number of cache hits (negative)"); 379 /* These count for kern___getcwd(), too. */ 380 STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls"); 381 STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)"); 382 STATNODE_COUNTER(numfullpathfail2, 383 "Number of fullpath search errors (VOP_VPTOCNP failures)"); 384 STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)"); 385 STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls"); 386 static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail, 387 "Number of times zap_and_exit failed to lock"); 388 static long cache_lock_vnodes_cel_3_failures; 389 STATNODE_ULONG(cache_lock_vnodes_cel_3_failures, 390 "Number of times 3-way vnode locking failed"); 391 392 static void cache_zap_locked(struct namecache *ncp, bool neg_locked); 393 static int vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir, 394 char *buf, char **retbuf, u_int buflen); 395 396 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries"); 397 398 static int cache_yield; 399 SYSCTL_INT(_vfs_cache, OID_AUTO, yield, CTLFLAG_RD, &cache_yield, 0, 400 "Number of times cache called yield"); 401 402 static void 403 cache_maybe_yield(void) 404 { 405 406 if (should_yield()) { 407 cache_yield++; 408 kern_yield(PRI_USER); 409 } 410 } 411 412 static inline void 413 cache_assert_vlp_locked(struct mtx *vlp) 414 { 415 416 if (vlp != NULL) 417 mtx_assert(vlp, MA_OWNED); 418 } 419 420 static inline void 421 cache_assert_vnode_locked(struct vnode *vp) 422 { 423 struct mtx *vlp; 424 425 vlp = VP2VNODELOCK(vp); 426 cache_assert_vlp_locked(vlp); 427 } 428 429 static uint32_t 430 cache_get_hash(char *name, u_char len, struct vnode *dvp) 431 { 432 uint32_t hash; 433 434 hash = fnv_32_buf(name, len, FNV1_32_INIT); 435 hash = fnv_32_buf(&dvp, sizeof(dvp), hash); 436 return (hash); 437 } 438 439 static inline struct rwlock * 440 NCP2BUCKETLOCK(struct namecache *ncp) 441 { 442 uint32_t hash; 443 444 hash = cache_get_hash(nc_get_name(ncp), ncp->nc_nlen, ncp->nc_dvp); 445 return (HASH2BUCKETLOCK(hash)); 446 } 447 448 #ifdef INVARIANTS 449 static void 450 cache_assert_bucket_locked(struct namecache *ncp, int mode) 451 { 452 struct rwlock *blp; 453 454 blp = NCP2BUCKETLOCK(ncp); 455 rw_assert(blp, mode); 456 } 457 #else 458 #define cache_assert_bucket_locked(x, y) do { } while (0) 459 #endif 460 461 #define cache_sort(x, y) _cache_sort((void **)(x), (void **)(y)) 462 static void 463 _cache_sort(void **p1, void **p2) 464 { 465 void *tmp; 466 467 if (*p1 > *p2) { 468 tmp = *p2; 469 *p2 = *p1; 470 *p1 = tmp; 471 } 472 } 473 474 static void 475 cache_lock_all_buckets(void) 476 { 477 u_int i; 478 479 for (i = 0; i < numbucketlocks; i++) 480 rw_wlock(&bucketlocks[i]); 481 } 482 483 static void 484 cache_unlock_all_buckets(void) 485 { 486 u_int i; 487 488 for (i = 0; i < numbucketlocks; i++) 489 rw_wunlock(&bucketlocks[i]); 490 } 491 492 static void 493 cache_lock_all_vnodes(void) 494 { 495 u_int i; 496 497 for (i = 0; i < numvnodelocks; i++) 498 mtx_lock(&vnodelocks[i]); 499 } 500 501 static void 502 cache_unlock_all_vnodes(void) 503 { 504 u_int i; 505 506 for (i = 0; i < numvnodelocks; i++) 507 mtx_unlock(&vnodelocks[i]); 508 } 509 510 static int 511 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2) 512 { 513 514 cache_sort(&vlp1, &vlp2); 515 MPASS(vlp2 != NULL); 516 517 if (vlp1 != NULL) { 518 if (!mtx_trylock(vlp1)) 519 return (EAGAIN); 520 } 521 if (!mtx_trylock(vlp2)) { 522 if (vlp1 != NULL) 523 mtx_unlock(vlp1); 524 return (EAGAIN); 525 } 526 527 return (0); 528 } 529 530 static void 531 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2) 532 { 533 534 MPASS(vlp1 != NULL || vlp2 != NULL); 535 536 if (vlp1 != NULL) 537 mtx_unlock(vlp1); 538 if (vlp2 != NULL) 539 mtx_unlock(vlp2); 540 } 541 542 static int 543 sysctl_nchstats(SYSCTL_HANDLER_ARGS) 544 { 545 struct nchstats snap; 546 547 if (req->oldptr == NULL) 548 return (SYSCTL_OUT(req, 0, sizeof(snap))); 549 550 snap = nchstats; 551 snap.ncs_goodhits = counter_u64_fetch(numposhits); 552 snap.ncs_neghits = counter_u64_fetch(numneghits); 553 snap.ncs_badhits = counter_u64_fetch(numposzaps) + 554 counter_u64_fetch(numnegzaps); 555 snap.ncs_miss = counter_u64_fetch(nummisszap) + 556 counter_u64_fetch(nummiss); 557 558 return (SYSCTL_OUT(req, &snap, sizeof(snap))); 559 } 560 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD | 561 CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU", 562 "VFS cache effectiveness statistics"); 563 564 #ifdef DIAGNOSTIC 565 /* 566 * Grab an atomic snapshot of the name cache hash chain lengths 567 */ 568 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW, NULL, 569 "hash table stats"); 570 571 static int 572 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS) 573 { 574 struct nchashhead *ncpp; 575 struct namecache *ncp; 576 int i, error, n_nchash, *cntbuf; 577 578 retry: 579 n_nchash = nchash + 1; /* nchash is max index, not count */ 580 if (req->oldptr == NULL) 581 return SYSCTL_OUT(req, 0, n_nchash * sizeof(int)); 582 cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK); 583 cache_lock_all_buckets(); 584 if (n_nchash != nchash + 1) { 585 cache_unlock_all_buckets(); 586 free(cntbuf, M_TEMP); 587 goto retry; 588 } 589 /* Scan hash tables counting entries */ 590 for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++) 591 LIST_FOREACH(ncp, ncpp, nc_hash) 592 cntbuf[i]++; 593 cache_unlock_all_buckets(); 594 for (error = 0, i = 0; i < n_nchash; i++) 595 if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0) 596 break; 597 free(cntbuf, M_TEMP); 598 return (error); 599 } 600 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD| 601 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int", 602 "nchash chain lengths"); 603 604 static int 605 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS) 606 { 607 int error; 608 struct nchashhead *ncpp; 609 struct namecache *ncp; 610 int n_nchash; 611 int count, maxlength, used, pct; 612 613 if (!req->oldptr) 614 return SYSCTL_OUT(req, 0, 4 * sizeof(int)); 615 616 cache_lock_all_buckets(); 617 n_nchash = nchash + 1; /* nchash is max index, not count */ 618 used = 0; 619 maxlength = 0; 620 621 /* Scan hash tables for applicable entries */ 622 for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) { 623 count = 0; 624 LIST_FOREACH(ncp, ncpp, nc_hash) { 625 count++; 626 } 627 if (count) 628 used++; 629 if (maxlength < count) 630 maxlength = count; 631 } 632 n_nchash = nchash + 1; 633 cache_unlock_all_buckets(); 634 pct = (used * 100) / (n_nchash / 100); 635 error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash)); 636 if (error) 637 return (error); 638 error = SYSCTL_OUT(req, &used, sizeof(used)); 639 if (error) 640 return (error); 641 error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength)); 642 if (error) 643 return (error); 644 error = SYSCTL_OUT(req, &pct, sizeof(pct)); 645 if (error) 646 return (error); 647 return (0); 648 } 649 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD| 650 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I", 651 "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)"); 652 #endif 653 654 /* 655 * Negative entries management 656 * 657 * A variation of LRU scheme is used. New entries are hashed into one of 658 * numneglists cold lists. Entries get promoted to the hot list on first hit. 659 * Partial LRU for the hot list is maintained by requeueing them every 660 * ncneghitsrequeue hits. 661 * 662 * The shrinker will demote hot list head and evict from the cold list in a 663 * round-robin manner. 664 */ 665 static void 666 cache_negative_hit(struct namecache *ncp) 667 { 668 struct neglist *neglist; 669 u_int hits; 670 671 MPASS(ncp->nc_flag & NCF_NEGATIVE); 672 hits = atomic_fetchadd_int(&ncp->nc_neghits, 1); 673 if (ncp->nc_flag & NCF_HOTNEGATIVE) { 674 if ((hits % ncneghitsrequeue) != 0) 675 return; 676 mtx_lock(&ncneg_hot.nl_lock); 677 if (ncp->nc_flag & NCF_HOTNEGATIVE) { 678 TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst); 679 TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst); 680 mtx_unlock(&ncneg_hot.nl_lock); 681 return; 682 } 683 /* 684 * The shrinker cleared the flag and removed the entry from 685 * the hot list. Put it back. 686 */ 687 } else { 688 mtx_lock(&ncneg_hot.nl_lock); 689 } 690 neglist = NCP2NEGLIST(ncp); 691 mtx_lock(&neglist->nl_lock); 692 if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) { 693 TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst); 694 TAILQ_INSERT_TAIL(&ncneg_hot.nl_list, ncp, nc_dst); 695 ncp->nc_flag |= NCF_HOTNEGATIVE; 696 } 697 mtx_unlock(&neglist->nl_lock); 698 mtx_unlock(&ncneg_hot.nl_lock); 699 } 700 701 static void 702 cache_negative_insert(struct namecache *ncp, bool neg_locked) 703 { 704 struct neglist *neglist; 705 706 MPASS(ncp->nc_flag & NCF_NEGATIVE); 707 cache_assert_bucket_locked(ncp, RA_WLOCKED); 708 neglist = NCP2NEGLIST(ncp); 709 if (!neg_locked) { 710 mtx_lock(&neglist->nl_lock); 711 } else { 712 mtx_assert(&neglist->nl_lock, MA_OWNED); 713 } 714 TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst); 715 if (!neg_locked) 716 mtx_unlock(&neglist->nl_lock); 717 atomic_add_rel_long(&numneg, 1); 718 } 719 720 static void 721 cache_negative_remove(struct namecache *ncp, bool neg_locked) 722 { 723 struct neglist *neglist; 724 bool hot_locked = false; 725 bool list_locked = false; 726 727 MPASS(ncp->nc_flag & NCF_NEGATIVE); 728 cache_assert_bucket_locked(ncp, RA_WLOCKED); 729 neglist = NCP2NEGLIST(ncp); 730 if (!neg_locked) { 731 if (ncp->nc_flag & NCF_HOTNEGATIVE) { 732 hot_locked = true; 733 mtx_lock(&ncneg_hot.nl_lock); 734 if (!(ncp->nc_flag & NCF_HOTNEGATIVE)) { 735 list_locked = true; 736 mtx_lock(&neglist->nl_lock); 737 } 738 } else { 739 list_locked = true; 740 mtx_lock(&neglist->nl_lock); 741 } 742 } else { 743 mtx_assert(&neglist->nl_lock, MA_OWNED); 744 mtx_assert(&ncneg_hot.nl_lock, MA_OWNED); 745 } 746 if (ncp->nc_flag & NCF_HOTNEGATIVE) { 747 TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst); 748 } else { 749 TAILQ_REMOVE(&neglist->nl_list, ncp, nc_dst); 750 } 751 if (list_locked) 752 mtx_unlock(&neglist->nl_lock); 753 if (hot_locked) 754 mtx_unlock(&ncneg_hot.nl_lock); 755 atomic_subtract_rel_long(&numneg, 1); 756 } 757 758 static void 759 cache_negative_shrink_select(int start, struct namecache **ncpp, 760 struct neglist **neglistpp) 761 { 762 struct neglist *neglist; 763 struct namecache *ncp; 764 int i; 765 766 *ncpp = ncp = NULL; 767 768 for (i = start; i < numneglists; i++) { 769 neglist = &neglists[i]; 770 if (TAILQ_FIRST(&neglist->nl_list) == NULL) 771 continue; 772 mtx_lock(&neglist->nl_lock); 773 ncp = TAILQ_FIRST(&neglist->nl_list); 774 if (ncp != NULL) 775 break; 776 mtx_unlock(&neglist->nl_lock); 777 } 778 779 *neglistpp = neglist; 780 *ncpp = ncp; 781 } 782 783 static void 784 cache_negative_zap_one(void) 785 { 786 struct namecache *ncp, *ncp2, *ncpc; 787 struct neglist *neglist; 788 struct mtx *dvlp; 789 struct rwlock *blp; 790 791 if (!mtx_trylock(&ncneg_shrink_lock)) 792 return; 793 794 ncpc = NULL; 795 mtx_lock(&ncneg_hot.nl_lock); 796 ncp = TAILQ_FIRST(&ncneg_hot.nl_list); 797 if (ncp != NULL) { 798 neglist = NCP2NEGLIST(ncp); 799 mtx_lock(&neglist->nl_lock); 800 TAILQ_REMOVE(&ncneg_hot.nl_list, ncp, nc_dst); 801 TAILQ_INSERT_TAIL(&neglist->nl_list, ncp, nc_dst); 802 ncp->nc_flag &= ~NCF_HOTNEGATIVE; 803 mtx_unlock(&neglist->nl_lock); 804 } 805 806 cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist); 807 shrink_list_turn++; 808 if (shrink_list_turn == numneglists) 809 shrink_list_turn = 0; 810 if (ncp == NULL && shrink_list_turn == 0) 811 cache_negative_shrink_select(shrink_list_turn, &ncp, &neglist); 812 if (ncp == NULL) { 813 mtx_unlock(&ncneg_hot.nl_lock); 814 goto out; 815 } 816 817 MPASS(ncp->nc_flag & NCF_NEGATIVE); 818 dvlp = VP2VNODELOCK(ncp->nc_dvp); 819 blp = NCP2BUCKETLOCK(ncp); 820 mtx_unlock(&neglist->nl_lock); 821 mtx_unlock(&ncneg_hot.nl_lock); 822 mtx_lock(dvlp); 823 rw_wlock(blp); 824 mtx_lock(&ncneg_hot.nl_lock); 825 mtx_lock(&neglist->nl_lock); 826 ncp2 = TAILQ_FIRST(&neglist->nl_list); 827 if (ncp != ncp2 || dvlp != VP2VNODELOCK(ncp2->nc_dvp) || 828 blp != NCP2BUCKETLOCK(ncp2) || !(ncp2->nc_flag & NCF_NEGATIVE)) { 829 ncp = NULL; 830 goto out_unlock_all; 831 } 832 SDT_PROBE3(vfs, namecache, shrink_negative, done, ncp->nc_dvp, 833 nc_get_name(ncp), ncp->nc_neghits); 834 835 cache_zap_locked(ncp, true); 836 out_unlock_all: 837 mtx_unlock(&neglist->nl_lock); 838 mtx_unlock(&ncneg_hot.nl_lock); 839 rw_wunlock(blp); 840 mtx_unlock(dvlp); 841 out: 842 mtx_unlock(&ncneg_shrink_lock); 843 cache_free(ncp); 844 } 845 846 /* 847 * cache_zap_locked(): 848 * 849 * Removes a namecache entry from cache, whether it contains an actual 850 * pointer to a vnode or if it is just a negative cache entry. 851 */ 852 static void 853 cache_zap_locked(struct namecache *ncp, bool neg_locked) 854 { 855 856 if (!(ncp->nc_flag & NCF_NEGATIVE)) 857 cache_assert_vnode_locked(ncp->nc_vp); 858 cache_assert_vnode_locked(ncp->nc_dvp); 859 cache_assert_bucket_locked(ncp, RA_WLOCKED); 860 861 CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp, 862 (ncp->nc_flag & NCF_NEGATIVE) ? NULL : ncp->nc_vp); 863 if (!(ncp->nc_flag & NCF_NEGATIVE)) { 864 SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp, 865 nc_get_name(ncp), ncp->nc_vp); 866 } else { 867 SDT_PROBE3(vfs, namecache, zap_negative, done, ncp->nc_dvp, 868 nc_get_name(ncp), ncp->nc_neghits); 869 } 870 LIST_REMOVE(ncp, nc_hash); 871 if (ncp->nc_flag & NCF_ISDOTDOT) { 872 if (ncp == ncp->nc_dvp->v_cache_dd) 873 ncp->nc_dvp->v_cache_dd = NULL; 874 } else { 875 LIST_REMOVE(ncp, nc_src); 876 if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) { 877 ncp->nc_flag |= NCF_DVDROP; 878 atomic_subtract_rel_long(&numcachehv, 1); 879 } 880 } 881 if (!(ncp->nc_flag & NCF_NEGATIVE)) { 882 TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst); 883 if (ncp == ncp->nc_vp->v_cache_dd) 884 ncp->nc_vp->v_cache_dd = NULL; 885 } else { 886 cache_negative_remove(ncp, neg_locked); 887 } 888 atomic_subtract_rel_long(&numcache, 1); 889 } 890 891 static void 892 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp) 893 { 894 struct rwlock *blp; 895 896 MPASS(ncp->nc_dvp == vp); 897 MPASS(ncp->nc_flag & NCF_NEGATIVE); 898 cache_assert_vnode_locked(vp); 899 900 blp = NCP2BUCKETLOCK(ncp); 901 rw_wlock(blp); 902 cache_zap_locked(ncp, false); 903 rw_wunlock(blp); 904 } 905 906 static bool 907 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp, 908 struct mtx **vlpp) 909 { 910 struct mtx *pvlp, *vlp1, *vlp2, *to_unlock; 911 struct rwlock *blp; 912 913 MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp); 914 cache_assert_vnode_locked(vp); 915 916 if (ncp->nc_flag & NCF_NEGATIVE) { 917 if (*vlpp != NULL) { 918 mtx_unlock(*vlpp); 919 *vlpp = NULL; 920 } 921 cache_zap_negative_locked_vnode_kl(ncp, vp); 922 return (true); 923 } 924 925 pvlp = VP2VNODELOCK(vp); 926 blp = NCP2BUCKETLOCK(ncp); 927 vlp1 = VP2VNODELOCK(ncp->nc_dvp); 928 vlp2 = VP2VNODELOCK(ncp->nc_vp); 929 930 if (*vlpp == vlp1 || *vlpp == vlp2) { 931 to_unlock = *vlpp; 932 *vlpp = NULL; 933 } else { 934 if (*vlpp != NULL) { 935 mtx_unlock(*vlpp); 936 *vlpp = NULL; 937 } 938 cache_sort(&vlp1, &vlp2); 939 if (vlp1 == pvlp) { 940 mtx_lock(vlp2); 941 to_unlock = vlp2; 942 } else { 943 if (!mtx_trylock(vlp1)) 944 goto out_relock; 945 to_unlock = vlp1; 946 } 947 } 948 rw_wlock(blp); 949 cache_zap_locked(ncp, false); 950 rw_wunlock(blp); 951 if (to_unlock != NULL) 952 mtx_unlock(to_unlock); 953 return (true); 954 955 out_relock: 956 mtx_unlock(vlp2); 957 mtx_lock(vlp1); 958 mtx_lock(vlp2); 959 MPASS(*vlpp == NULL); 960 *vlpp = vlp1; 961 return (false); 962 } 963 964 static int 965 cache_zap_locked_vnode(struct namecache *ncp, struct vnode *vp) 966 { 967 struct mtx *pvlp, *vlp1, *vlp2, *to_unlock; 968 struct rwlock *blp; 969 int error = 0; 970 971 MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp); 972 cache_assert_vnode_locked(vp); 973 974 pvlp = VP2VNODELOCK(vp); 975 if (ncp->nc_flag & NCF_NEGATIVE) { 976 cache_zap_negative_locked_vnode_kl(ncp, vp); 977 goto out; 978 } 979 980 blp = NCP2BUCKETLOCK(ncp); 981 vlp1 = VP2VNODELOCK(ncp->nc_dvp); 982 vlp2 = VP2VNODELOCK(ncp->nc_vp); 983 cache_sort(&vlp1, &vlp2); 984 if (vlp1 == pvlp) { 985 mtx_lock(vlp2); 986 to_unlock = vlp2; 987 } else { 988 if (!mtx_trylock(vlp1)) { 989 error = EAGAIN; 990 goto out; 991 } 992 to_unlock = vlp1; 993 } 994 rw_wlock(blp); 995 cache_zap_locked(ncp, false); 996 rw_wunlock(blp); 997 mtx_unlock(to_unlock); 998 out: 999 mtx_unlock(pvlp); 1000 return (error); 1001 } 1002 1003 static int 1004 cache_zap_rlocked_bucket(struct namecache *ncp, struct rwlock *blp) 1005 { 1006 struct mtx *dvlp, *vlp; 1007 1008 cache_assert_bucket_locked(ncp, RA_RLOCKED); 1009 1010 dvlp = VP2VNODELOCK(ncp->nc_dvp); 1011 vlp = NULL; 1012 if (!(ncp->nc_flag & NCF_NEGATIVE)) 1013 vlp = VP2VNODELOCK(ncp->nc_vp); 1014 if (cache_trylock_vnodes(dvlp, vlp) == 0) { 1015 rw_runlock(blp); 1016 rw_wlock(blp); 1017 cache_zap_locked(ncp, false); 1018 rw_wunlock(blp); 1019 cache_unlock_vnodes(dvlp, vlp); 1020 return (0); 1021 } 1022 1023 rw_runlock(blp); 1024 return (EAGAIN); 1025 } 1026 1027 static int 1028 cache_zap_wlocked_bucket_kl(struct namecache *ncp, struct rwlock *blp, 1029 struct mtx **vlpp1, struct mtx **vlpp2) 1030 { 1031 struct mtx *dvlp, *vlp; 1032 1033 cache_assert_bucket_locked(ncp, RA_WLOCKED); 1034 1035 dvlp = VP2VNODELOCK(ncp->nc_dvp); 1036 vlp = NULL; 1037 if (!(ncp->nc_flag & NCF_NEGATIVE)) 1038 vlp = VP2VNODELOCK(ncp->nc_vp); 1039 cache_sort(&dvlp, &vlp); 1040 1041 if (*vlpp1 == dvlp && *vlpp2 == vlp) { 1042 cache_zap_locked(ncp, false); 1043 cache_unlock_vnodes(dvlp, vlp); 1044 *vlpp1 = NULL; 1045 *vlpp2 = NULL; 1046 return (0); 1047 } 1048 1049 if (*vlpp1 != NULL) 1050 mtx_unlock(*vlpp1); 1051 if (*vlpp2 != NULL) 1052 mtx_unlock(*vlpp2); 1053 *vlpp1 = NULL; 1054 *vlpp2 = NULL; 1055 1056 if (cache_trylock_vnodes(dvlp, vlp) == 0) { 1057 cache_zap_locked(ncp, false); 1058 cache_unlock_vnodes(dvlp, vlp); 1059 return (0); 1060 } 1061 1062 rw_wunlock(blp); 1063 *vlpp1 = dvlp; 1064 *vlpp2 = vlp; 1065 if (*vlpp1 != NULL) 1066 mtx_lock(*vlpp1); 1067 mtx_lock(*vlpp2); 1068 rw_wlock(blp); 1069 return (EAGAIN); 1070 } 1071 1072 static void 1073 cache_lookup_unlock(struct rwlock *blp, struct mtx *vlp) 1074 { 1075 1076 if (blp != NULL) { 1077 rw_runlock(blp); 1078 mtx_assert(vlp, MA_NOTOWNED); 1079 } else { 1080 mtx_unlock(vlp); 1081 } 1082 } 1083 1084 /* 1085 * Lookup an entry in the cache 1086 * 1087 * Lookup is called with dvp pointing to the directory to search, 1088 * cnp pointing to the name of the entry being sought. If the lookup 1089 * succeeds, the vnode is returned in *vpp, and a status of -1 is 1090 * returned. If the lookup determines that the name does not exist 1091 * (negative caching), a status of ENOENT is returned. If the lookup 1092 * fails, a status of zero is returned. If the directory vnode is 1093 * recycled out from under us due to a forced unmount, a status of 1094 * ENOENT is returned. 1095 * 1096 * vpp is locked and ref'd on return. If we're looking up DOTDOT, dvp is 1097 * unlocked. If we're looking up . an extra ref is taken, but the lock is 1098 * not recursively acquired. 1099 */ 1100 1101 int 1102 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, 1103 struct timespec *tsp, int *ticksp) 1104 { 1105 struct namecache *ncp; 1106 struct rwlock *blp; 1107 struct mtx *dvlp, *dvlp2; 1108 uint32_t hash; 1109 int error, ltype; 1110 1111 if (!doingcache) { 1112 cnp->cn_flags &= ~MAKEENTRY; 1113 return (0); 1114 } 1115 retry: 1116 blp = NULL; 1117 dvlp = VP2VNODELOCK(dvp); 1118 error = 0; 1119 counter_u64_add(numcalls, 1); 1120 1121 if (cnp->cn_nameptr[0] == '.') { 1122 if (cnp->cn_namelen == 1) { 1123 *vpp = dvp; 1124 CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .", 1125 dvp, cnp->cn_nameptr); 1126 counter_u64_add(dothits, 1); 1127 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp); 1128 if (tsp != NULL) 1129 timespecclear(tsp); 1130 if (ticksp != NULL) 1131 *ticksp = ticks; 1132 VREF(*vpp); 1133 /* 1134 * When we lookup "." we still can be asked to lock it 1135 * differently... 1136 */ 1137 ltype = cnp->cn_lkflags & LK_TYPE_MASK; 1138 if (ltype != VOP_ISLOCKED(*vpp)) { 1139 if (ltype == LK_EXCLUSIVE) { 1140 vn_lock(*vpp, LK_UPGRADE | LK_RETRY); 1141 if ((*vpp)->v_iflag & VI_DOOMED) { 1142 /* forced unmount */ 1143 vrele(*vpp); 1144 *vpp = NULL; 1145 return (ENOENT); 1146 } 1147 } else 1148 vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY); 1149 } 1150 return (-1); 1151 } 1152 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') { 1153 counter_u64_add(dotdothits, 1); 1154 dvlp2 = NULL; 1155 mtx_lock(dvlp); 1156 retry_dotdot: 1157 ncp = dvp->v_cache_dd; 1158 if (ncp == NULL) { 1159 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, 1160 "..", NULL); 1161 mtx_unlock(dvlp); 1162 return (0); 1163 } 1164 if ((cnp->cn_flags & MAKEENTRY) == 0) { 1165 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) { 1166 if (ncp->nc_dvp != dvp) 1167 panic("dvp %p v_cache_dd %p\n", dvp, ncp); 1168 if (!cache_zap_locked_vnode_kl2(ncp, 1169 dvp, &dvlp2)) 1170 goto retry_dotdot; 1171 MPASS(dvp->v_cache_dd == NULL); 1172 mtx_unlock(dvlp); 1173 if (dvlp2 != NULL) 1174 mtx_unlock(dvlp2); 1175 cache_free(ncp); 1176 } else { 1177 dvp->v_cache_dd = NULL; 1178 mtx_unlock(dvlp); 1179 if (dvlp2 != NULL) 1180 mtx_unlock(dvlp2); 1181 } 1182 return (0); 1183 } 1184 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) { 1185 if (ncp->nc_flag & NCF_NEGATIVE) 1186 *vpp = NULL; 1187 else 1188 *vpp = ncp->nc_vp; 1189 } else 1190 *vpp = ncp->nc_dvp; 1191 /* Return failure if negative entry was found. */ 1192 if (*vpp == NULL) 1193 goto negative_success; 1194 CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..", 1195 dvp, cnp->cn_nameptr, *vpp); 1196 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", 1197 *vpp); 1198 cache_out_ts(ncp, tsp, ticksp); 1199 if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) == 1200 NCF_DTS && tsp != NULL) 1201 *tsp = ((struct namecache_ts *)ncp)-> 1202 nc_dotdottime; 1203 goto success; 1204 } 1205 } 1206 1207 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); 1208 blp = HASH2BUCKETLOCK(hash); 1209 rw_rlock(blp); 1210 1211 LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 1212 counter_u64_add(numchecks, 1); 1213 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && 1214 !bcmp(nc_get_name(ncp), cnp->cn_nameptr, ncp->nc_nlen)) 1215 break; 1216 } 1217 1218 /* We failed to find an entry */ 1219 if (ncp == NULL) { 1220 SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, 1221 NULL); 1222 if ((cnp->cn_flags & MAKEENTRY) == 0) { 1223 counter_u64_add(nummisszap, 1); 1224 } else { 1225 counter_u64_add(nummiss, 1); 1226 } 1227 goto unlock; 1228 } 1229 1230 /* We don't want to have an entry, so dump it */ 1231 if ((cnp->cn_flags & MAKEENTRY) == 0) { 1232 counter_u64_add(numposzaps, 1); 1233 goto zap_and_exit; 1234 } 1235 1236 /* We found a "positive" match, return the vnode */ 1237 if (!(ncp->nc_flag & NCF_NEGATIVE)) { 1238 counter_u64_add(numposhits, 1); 1239 *vpp = ncp->nc_vp; 1240 CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p", 1241 dvp, cnp->cn_nameptr, *vpp, ncp); 1242 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, nc_get_name(ncp), 1243 *vpp); 1244 cache_out_ts(ncp, tsp, ticksp); 1245 goto success; 1246 } 1247 1248 negative_success: 1249 /* We found a negative match, and want to create it, so purge */ 1250 if (cnp->cn_nameiop == CREATE) { 1251 counter_u64_add(numnegzaps, 1); 1252 goto zap_and_exit; 1253 } 1254 1255 counter_u64_add(numneghits, 1); 1256 cache_negative_hit(ncp); 1257 if (ncp->nc_flag & NCF_WHITE) 1258 cnp->cn_flags |= ISWHITEOUT; 1259 SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp, 1260 nc_get_name(ncp)); 1261 cache_out_ts(ncp, tsp, ticksp); 1262 cache_lookup_unlock(blp, dvlp); 1263 return (ENOENT); 1264 1265 success: 1266 /* 1267 * On success we return a locked and ref'd vnode as per the lookup 1268 * protocol. 1269 */ 1270 MPASS(dvp != *vpp); 1271 ltype = 0; /* silence gcc warning */ 1272 if (cnp->cn_flags & ISDOTDOT) { 1273 ltype = VOP_ISLOCKED(dvp); 1274 VOP_UNLOCK(dvp, 0); 1275 } 1276 vhold(*vpp); 1277 cache_lookup_unlock(blp, dvlp); 1278 error = vget(*vpp, cnp->cn_lkflags | LK_VNHELD, cnp->cn_thread); 1279 if (cnp->cn_flags & ISDOTDOT) { 1280 vn_lock(dvp, ltype | LK_RETRY); 1281 if (dvp->v_iflag & VI_DOOMED) { 1282 if (error == 0) 1283 vput(*vpp); 1284 *vpp = NULL; 1285 return (ENOENT); 1286 } 1287 } 1288 if (error) { 1289 *vpp = NULL; 1290 goto retry; 1291 } 1292 if ((cnp->cn_flags & ISLASTCN) && 1293 (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) { 1294 ASSERT_VOP_ELOCKED(*vpp, "cache_lookup"); 1295 } 1296 return (-1); 1297 1298 unlock: 1299 cache_lookup_unlock(blp, dvlp); 1300 return (0); 1301 1302 zap_and_exit: 1303 if (blp != NULL) 1304 error = cache_zap_rlocked_bucket(ncp, blp); 1305 else 1306 error = cache_zap_locked_vnode(ncp, dvp); 1307 if (error != 0) { 1308 zap_and_exit_bucket_fail++; 1309 cache_maybe_yield(); 1310 goto retry; 1311 } 1312 cache_free(ncp); 1313 return (0); 1314 } 1315 1316 struct celockstate { 1317 struct mtx *vlp[3]; 1318 struct rwlock *blp[2]; 1319 }; 1320 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3)); 1321 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2)); 1322 1323 static inline void 1324 cache_celockstate_init(struct celockstate *cel) 1325 { 1326 1327 bzero(cel, sizeof(*cel)); 1328 } 1329 1330 static void 1331 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp, 1332 struct vnode *dvp) 1333 { 1334 struct mtx *vlp1, *vlp2; 1335 1336 MPASS(cel->vlp[0] == NULL); 1337 MPASS(cel->vlp[1] == NULL); 1338 MPASS(cel->vlp[2] == NULL); 1339 1340 MPASS(vp != NULL || dvp != NULL); 1341 1342 vlp1 = VP2VNODELOCK(vp); 1343 vlp2 = VP2VNODELOCK(dvp); 1344 cache_sort(&vlp1, &vlp2); 1345 1346 if (vlp1 != NULL) { 1347 mtx_lock(vlp1); 1348 cel->vlp[0] = vlp1; 1349 } 1350 mtx_lock(vlp2); 1351 cel->vlp[1] = vlp2; 1352 } 1353 1354 static void 1355 cache_unlock_vnodes_cel(struct celockstate *cel) 1356 { 1357 1358 MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL); 1359 1360 if (cel->vlp[0] != NULL) 1361 mtx_unlock(cel->vlp[0]); 1362 if (cel->vlp[1] != NULL) 1363 mtx_unlock(cel->vlp[1]); 1364 if (cel->vlp[2] != NULL) 1365 mtx_unlock(cel->vlp[2]); 1366 } 1367 1368 static bool 1369 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp) 1370 { 1371 struct mtx *vlp; 1372 bool ret; 1373 1374 cache_assert_vlp_locked(cel->vlp[0]); 1375 cache_assert_vlp_locked(cel->vlp[1]); 1376 MPASS(cel->vlp[2] == NULL); 1377 1378 vlp = VP2VNODELOCK(vp); 1379 MPASS(vlp != NULL); 1380 1381 ret = true; 1382 if (vlp >= cel->vlp[1]) { 1383 mtx_lock(vlp); 1384 } else { 1385 if (mtx_trylock(vlp)) 1386 goto out; 1387 cache_lock_vnodes_cel_3_failures++; 1388 cache_unlock_vnodes_cel(cel); 1389 if (vlp < cel->vlp[0]) { 1390 mtx_lock(vlp); 1391 mtx_lock(cel->vlp[0]); 1392 mtx_lock(cel->vlp[1]); 1393 } else { 1394 if (cel->vlp[0] != NULL) 1395 mtx_lock(cel->vlp[0]); 1396 mtx_lock(vlp); 1397 mtx_lock(cel->vlp[1]); 1398 } 1399 ret = false; 1400 } 1401 out: 1402 cel->vlp[2] = vlp; 1403 return (ret); 1404 } 1405 1406 static void 1407 cache_lock_buckets_cel(struct celockstate *cel, struct rwlock *blp1, 1408 struct rwlock *blp2) 1409 { 1410 1411 MPASS(cel->blp[0] == NULL); 1412 MPASS(cel->blp[1] == NULL); 1413 1414 cache_sort(&blp1, &blp2); 1415 1416 if (blp1 != NULL) { 1417 rw_wlock(blp1); 1418 cel->blp[0] = blp1; 1419 } 1420 rw_wlock(blp2); 1421 cel->blp[1] = blp2; 1422 } 1423 1424 static void 1425 cache_unlock_buckets_cel(struct celockstate *cel) 1426 { 1427 1428 if (cel->blp[0] != NULL) 1429 rw_wunlock(cel->blp[0]); 1430 rw_wunlock(cel->blp[1]); 1431 } 1432 1433 /* 1434 * Lock part of the cache affected by the insertion. 1435 * 1436 * This means vnodelocks for dvp, vp and the relevant bucketlock. 1437 * However, insertion can result in removal of an old entry. In this 1438 * case we have an additional vnode and bucketlock pair to lock. If the 1439 * entry is negative, ncelock is locked instead of the vnode. 1440 * 1441 * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while 1442 * preserving the locking order (smaller address first). 1443 */ 1444 static void 1445 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp, 1446 uint32_t hash) 1447 { 1448 struct namecache *ncp; 1449 struct rwlock *blps[2]; 1450 1451 blps[0] = HASH2BUCKETLOCK(hash); 1452 for (;;) { 1453 blps[1] = NULL; 1454 cache_lock_vnodes_cel(cel, dvp, vp); 1455 if (vp == NULL || vp->v_type != VDIR) 1456 break; 1457 ncp = vp->v_cache_dd; 1458 if (ncp == NULL) 1459 break; 1460 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) 1461 break; 1462 MPASS(ncp->nc_dvp == vp); 1463 blps[1] = NCP2BUCKETLOCK(ncp); 1464 if (ncp->nc_flag & NCF_NEGATIVE) 1465 break; 1466 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp)) 1467 break; 1468 /* 1469 * All vnodes got re-locked. Re-validate the state and if 1470 * nothing changed we are done. Otherwise restart. 1471 */ 1472 if (ncp == vp->v_cache_dd && 1473 (ncp->nc_flag & NCF_ISDOTDOT) != 0 && 1474 blps[1] == NCP2BUCKETLOCK(ncp) && 1475 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2]) 1476 break; 1477 cache_unlock_vnodes_cel(cel); 1478 cel->vlp[0] = NULL; 1479 cel->vlp[1] = NULL; 1480 cel->vlp[2] = NULL; 1481 } 1482 cache_lock_buckets_cel(cel, blps[0], blps[1]); 1483 } 1484 1485 static void 1486 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp, 1487 uint32_t hash) 1488 { 1489 struct namecache *ncp; 1490 struct rwlock *blps[2]; 1491 1492 blps[0] = HASH2BUCKETLOCK(hash); 1493 for (;;) { 1494 blps[1] = NULL; 1495 cache_lock_vnodes_cel(cel, dvp, vp); 1496 ncp = dvp->v_cache_dd; 1497 if (ncp == NULL) 1498 break; 1499 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) 1500 break; 1501 MPASS(ncp->nc_dvp == dvp); 1502 blps[1] = NCP2BUCKETLOCK(ncp); 1503 if (ncp->nc_flag & NCF_NEGATIVE) 1504 break; 1505 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp)) 1506 break; 1507 if (ncp == dvp->v_cache_dd && 1508 (ncp->nc_flag & NCF_ISDOTDOT) != 0 && 1509 blps[1] == NCP2BUCKETLOCK(ncp) && 1510 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2]) 1511 break; 1512 cache_unlock_vnodes_cel(cel); 1513 cel->vlp[0] = NULL; 1514 cel->vlp[1] = NULL; 1515 cel->vlp[2] = NULL; 1516 } 1517 cache_lock_buckets_cel(cel, blps[0], blps[1]); 1518 } 1519 1520 static void 1521 cache_enter_unlock(struct celockstate *cel) 1522 { 1523 1524 cache_unlock_buckets_cel(cel); 1525 cache_unlock_vnodes_cel(cel); 1526 } 1527 1528 /* 1529 * Add an entry to the cache. 1530 */ 1531 void 1532 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, 1533 struct timespec *tsp, struct timespec *dtsp) 1534 { 1535 struct celockstate cel; 1536 struct namecache *ncp, *n2, *ndd; 1537 struct namecache_ts *n3; 1538 struct nchashhead *ncpp; 1539 struct neglist *neglist; 1540 uint32_t hash; 1541 int flag; 1542 int len; 1543 bool neg_locked; 1544 1545 CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr); 1546 VNASSERT(vp == NULL || (vp->v_iflag & VI_DOOMED) == 0, vp, 1547 ("cache_enter: Adding a doomed vnode")); 1548 VNASSERT(dvp == NULL || (dvp->v_iflag & VI_DOOMED) == 0, dvp, 1549 ("cache_enter: Doomed vnode used as src")); 1550 1551 if (!doingcache) 1552 return; 1553 1554 /* 1555 * Avoid blowout in namecache entries. 1556 */ 1557 if (numcache >= desiredvnodes * ncsizefactor) 1558 return; 1559 1560 cache_celockstate_init(&cel); 1561 ndd = NULL; 1562 flag = 0; 1563 if (cnp->cn_nameptr[0] == '.') { 1564 if (cnp->cn_namelen == 1) 1565 return; 1566 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') { 1567 len = cnp->cn_namelen; 1568 hash = cache_get_hash(cnp->cn_nameptr, len, dvp); 1569 cache_enter_lock_dd(&cel, dvp, vp, hash); 1570 /* 1571 * If dotdot entry already exists, just retarget it 1572 * to new parent vnode, otherwise continue with new 1573 * namecache entry allocation. 1574 */ 1575 if ((ncp = dvp->v_cache_dd) != NULL && 1576 ncp->nc_flag & NCF_ISDOTDOT) { 1577 KASSERT(ncp->nc_dvp == dvp, 1578 ("wrong isdotdot parent")); 1579 neg_locked = false; 1580 if (ncp->nc_flag & NCF_NEGATIVE || vp == NULL) { 1581 neglist = NCP2NEGLIST(ncp); 1582 mtx_lock(&ncneg_hot.nl_lock); 1583 mtx_lock(&neglist->nl_lock); 1584 neg_locked = true; 1585 } 1586 if (!(ncp->nc_flag & NCF_NEGATIVE)) { 1587 TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, 1588 ncp, nc_dst); 1589 } else { 1590 cache_negative_remove(ncp, true); 1591 } 1592 if (vp != NULL) { 1593 TAILQ_INSERT_HEAD(&vp->v_cache_dst, 1594 ncp, nc_dst); 1595 ncp->nc_flag &= ~(NCF_NEGATIVE|NCF_HOTNEGATIVE); 1596 } else { 1597 ncp->nc_flag &= ~(NCF_HOTNEGATIVE); 1598 ncp->nc_flag |= NCF_NEGATIVE; 1599 cache_negative_insert(ncp, true); 1600 } 1601 if (neg_locked) { 1602 mtx_unlock(&neglist->nl_lock); 1603 mtx_unlock(&ncneg_hot.nl_lock); 1604 } 1605 ncp->nc_vp = vp; 1606 cache_enter_unlock(&cel); 1607 return; 1608 } 1609 dvp->v_cache_dd = NULL; 1610 cache_enter_unlock(&cel); 1611 cache_celockstate_init(&cel); 1612 SDT_PROBE3(vfs, namecache, enter, done, dvp, "..", vp); 1613 flag = NCF_ISDOTDOT; 1614 } 1615 } 1616 1617 /* 1618 * Calculate the hash key and setup as much of the new 1619 * namecache entry as possible before acquiring the lock. 1620 */ 1621 ncp = cache_alloc(cnp->cn_namelen, tsp != NULL); 1622 ncp->nc_flag = flag; 1623 ncp->nc_vp = vp; 1624 if (vp == NULL) 1625 ncp->nc_flag |= NCF_NEGATIVE; 1626 ncp->nc_dvp = dvp; 1627 if (tsp != NULL) { 1628 n3 = (struct namecache_ts *)ncp; 1629 n3->nc_time = *tsp; 1630 n3->nc_ticks = ticks; 1631 n3->nc_flag |= NCF_TS; 1632 if (dtsp != NULL) { 1633 n3->nc_dotdottime = *dtsp; 1634 n3->nc_flag |= NCF_DTS; 1635 } 1636 } 1637 len = ncp->nc_nlen = cnp->cn_namelen; 1638 hash = cache_get_hash(cnp->cn_nameptr, len, dvp); 1639 strlcpy(nc_get_name(ncp), cnp->cn_nameptr, len + 1); 1640 cache_enter_lock(&cel, dvp, vp, hash); 1641 1642 /* 1643 * See if this vnode or negative entry is already in the cache 1644 * with this name. This can happen with concurrent lookups of 1645 * the same path name. 1646 */ 1647 ncpp = NCHHASH(hash); 1648 LIST_FOREACH(n2, ncpp, nc_hash) { 1649 if (n2->nc_dvp == dvp && 1650 n2->nc_nlen == cnp->cn_namelen && 1651 !bcmp(nc_get_name(n2), cnp->cn_nameptr, n2->nc_nlen)) { 1652 if (tsp != NULL) { 1653 KASSERT((n2->nc_flag & NCF_TS) != 0, 1654 ("no NCF_TS")); 1655 n3 = (struct namecache_ts *)n2; 1656 n3->nc_time = 1657 ((struct namecache_ts *)ncp)->nc_time; 1658 n3->nc_ticks = 1659 ((struct namecache_ts *)ncp)->nc_ticks; 1660 if (dtsp != NULL) { 1661 n3->nc_dotdottime = 1662 ((struct namecache_ts *)ncp)-> 1663 nc_dotdottime; 1664 if (ncp->nc_flag & NCF_NEGATIVE) 1665 mtx_lock(&ncneg_hot.nl_lock); 1666 n3->nc_flag |= NCF_DTS; 1667 if (ncp->nc_flag & NCF_NEGATIVE) 1668 mtx_unlock(&ncneg_hot.nl_lock); 1669 } 1670 } 1671 goto out_unlock_free; 1672 } 1673 } 1674 1675 if (flag == NCF_ISDOTDOT) { 1676 /* 1677 * See if we are trying to add .. entry, but some other lookup 1678 * has populated v_cache_dd pointer already. 1679 */ 1680 if (dvp->v_cache_dd != NULL) 1681 goto out_unlock_free; 1682 KASSERT(vp == NULL || vp->v_type == VDIR, 1683 ("wrong vnode type %p", vp)); 1684 dvp->v_cache_dd = ncp; 1685 } 1686 1687 atomic_add_rel_long(&numcache, 1); 1688 if (vp != NULL) { 1689 if (vp->v_type == VDIR) { 1690 if (flag != NCF_ISDOTDOT) { 1691 /* 1692 * For this case, the cache entry maps both the 1693 * directory name in it and the name ".." for the 1694 * directory's parent. 1695 */ 1696 if ((ndd = vp->v_cache_dd) != NULL) { 1697 if ((ndd->nc_flag & NCF_ISDOTDOT) != 0) 1698 cache_zap_locked(ndd, false); 1699 else 1700 ndd = NULL; 1701 } 1702 vp->v_cache_dd = ncp; 1703 } 1704 } else { 1705 vp->v_cache_dd = NULL; 1706 } 1707 } 1708 1709 if (flag != NCF_ISDOTDOT) { 1710 if (LIST_EMPTY(&dvp->v_cache_src)) { 1711 vhold(dvp); 1712 atomic_add_rel_long(&numcachehv, 1); 1713 } 1714 LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src); 1715 } 1716 1717 /* 1718 * Insert the new namecache entry into the appropriate chain 1719 * within the cache entries table. 1720 */ 1721 LIST_INSERT_HEAD(ncpp, ncp, nc_hash); 1722 1723 /* 1724 * If the entry is "negative", we place it into the 1725 * "negative" cache queue, otherwise, we place it into the 1726 * destination vnode's cache entries queue. 1727 */ 1728 if (vp != NULL) { 1729 TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst); 1730 SDT_PROBE3(vfs, namecache, enter, done, dvp, nc_get_name(ncp), 1731 vp); 1732 } else { 1733 if (cnp->cn_flags & ISWHITEOUT) 1734 ncp->nc_flag |= NCF_WHITE; 1735 cache_negative_insert(ncp, false); 1736 SDT_PROBE2(vfs, namecache, enter_negative, done, dvp, 1737 nc_get_name(ncp)); 1738 } 1739 cache_enter_unlock(&cel); 1740 if (numneg * ncnegfactor > numcache) 1741 cache_negative_zap_one(); 1742 cache_free(ndd); 1743 return; 1744 out_unlock_free: 1745 cache_enter_unlock(&cel); 1746 cache_free(ncp); 1747 return; 1748 } 1749 1750 static u_int 1751 cache_roundup_2(u_int val) 1752 { 1753 u_int res; 1754 1755 for (res = 1; res <= val; res <<= 1) 1756 continue; 1757 1758 return (res); 1759 } 1760 1761 /* 1762 * Name cache initialization, from vfs_init() when we are booting 1763 */ 1764 static void 1765 nchinit(void *dummy __unused) 1766 { 1767 u_int i; 1768 1769 cache_zone_small = uma_zcreate("S VFS Cache", 1770 sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1, 1771 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT); 1772 cache_zone_small_ts = uma_zcreate("STS VFS Cache", 1773 sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1, 1774 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT); 1775 cache_zone_large = uma_zcreate("L VFS Cache", 1776 sizeof(struct namecache) + NAME_MAX + 1, 1777 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT); 1778 cache_zone_large_ts = uma_zcreate("LTS VFS Cache", 1779 sizeof(struct namecache_ts) + NAME_MAX + 1, 1780 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT); 1781 1782 nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash); 1783 numbucketlocks = cache_roundup_2(mp_ncpus * 64); 1784 bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE, 1785 M_WAITOK | M_ZERO); 1786 for (i = 0; i < numbucketlocks; i++) 1787 rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK | RW_RECURSE); 1788 numvnodelocks = cache_roundup_2(mp_ncpus * 64); 1789 vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE, 1790 M_WAITOK | M_ZERO); 1791 for (i = 0; i < numvnodelocks; i++) 1792 mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE); 1793 ncpurgeminvnodes = numbucketlocks; 1794 1795 numneglists = 4; 1796 neglists = malloc(sizeof(*neglists) * numneglists, M_VFSCACHE, 1797 M_WAITOK | M_ZERO); 1798 for (i = 0; i < numneglists; i++) { 1799 mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF); 1800 TAILQ_INIT(&neglists[i].nl_list); 1801 } 1802 mtx_init(&ncneg_hot.nl_lock, "ncneglh", NULL, MTX_DEF); 1803 TAILQ_INIT(&ncneg_hot.nl_list); 1804 1805 numcalls = counter_u64_alloc(M_WAITOK); 1806 dothits = counter_u64_alloc(M_WAITOK); 1807 dotdothits = counter_u64_alloc(M_WAITOK); 1808 numchecks = counter_u64_alloc(M_WAITOK); 1809 nummiss = counter_u64_alloc(M_WAITOK); 1810 nummisszap = counter_u64_alloc(M_WAITOK); 1811 numposzaps = counter_u64_alloc(M_WAITOK); 1812 numposhits = counter_u64_alloc(M_WAITOK); 1813 numnegzaps = counter_u64_alloc(M_WAITOK); 1814 numneghits = counter_u64_alloc(M_WAITOK); 1815 numfullpathcalls = counter_u64_alloc(M_WAITOK); 1816 numfullpathfail1 = counter_u64_alloc(M_WAITOK); 1817 numfullpathfail2 = counter_u64_alloc(M_WAITOK); 1818 numfullpathfail4 = counter_u64_alloc(M_WAITOK); 1819 numfullpathfound = counter_u64_alloc(M_WAITOK); 1820 } 1821 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL); 1822 1823 void 1824 cache_changesize(int newmaxvnodes) 1825 { 1826 struct nchashhead *new_nchashtbl, *old_nchashtbl; 1827 u_long new_nchash, old_nchash; 1828 struct namecache *ncp; 1829 uint32_t hash; 1830 int i; 1831 1832 new_nchashtbl = hashinit(newmaxvnodes * 2, M_VFSCACHE, &new_nchash); 1833 /* If same hash table size, nothing to do */ 1834 if (nchash == new_nchash) { 1835 free(new_nchashtbl, M_VFSCACHE); 1836 return; 1837 } 1838 /* 1839 * Move everything from the old hash table to the new table. 1840 * None of the namecache entries in the table can be removed 1841 * because to do so, they have to be removed from the hash table. 1842 */ 1843 cache_lock_all_vnodes(); 1844 cache_lock_all_buckets(); 1845 old_nchashtbl = nchashtbl; 1846 old_nchash = nchash; 1847 nchashtbl = new_nchashtbl; 1848 nchash = new_nchash; 1849 for (i = 0; i <= old_nchash; i++) { 1850 while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) { 1851 hash = cache_get_hash(nc_get_name(ncp), ncp->nc_nlen, 1852 ncp->nc_dvp); 1853 LIST_REMOVE(ncp, nc_hash); 1854 LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash); 1855 } 1856 } 1857 cache_unlock_all_buckets(); 1858 cache_unlock_all_vnodes(); 1859 free(old_nchashtbl, M_VFSCACHE); 1860 } 1861 1862 /* 1863 * Invalidate all entries to a particular vnode. 1864 */ 1865 void 1866 cache_purge(struct vnode *vp) 1867 { 1868 TAILQ_HEAD(, namecache) ncps; 1869 struct namecache *ncp, *nnp; 1870 struct mtx *vlp, *vlp2; 1871 1872 CTR1(KTR_VFS, "cache_purge(%p)", vp); 1873 SDT_PROBE1(vfs, namecache, purge, done, vp); 1874 if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) && 1875 vp->v_cache_dd == NULL) 1876 return; 1877 TAILQ_INIT(&ncps); 1878 vlp = VP2VNODELOCK(vp); 1879 vlp2 = NULL; 1880 mtx_lock(vlp); 1881 retry: 1882 while (!LIST_EMPTY(&vp->v_cache_src)) { 1883 ncp = LIST_FIRST(&vp->v_cache_src); 1884 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2)) 1885 goto retry; 1886 TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst); 1887 } 1888 while (!TAILQ_EMPTY(&vp->v_cache_dst)) { 1889 ncp = TAILQ_FIRST(&vp->v_cache_dst); 1890 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2)) 1891 goto retry; 1892 TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst); 1893 } 1894 ncp = vp->v_cache_dd; 1895 if (ncp != NULL) { 1896 KASSERT(ncp->nc_flag & NCF_ISDOTDOT, 1897 ("lost dotdot link")); 1898 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2)) 1899 goto retry; 1900 TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst); 1901 } 1902 KASSERT(vp->v_cache_dd == NULL, ("incomplete purge")); 1903 mtx_unlock(vlp); 1904 if (vlp2 != NULL) 1905 mtx_unlock(vlp2); 1906 TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) { 1907 cache_free(ncp); 1908 } 1909 } 1910 1911 /* 1912 * Invalidate all negative entries for a particular directory vnode. 1913 */ 1914 void 1915 cache_purge_negative(struct vnode *vp) 1916 { 1917 TAILQ_HEAD(, namecache) ncps; 1918 struct namecache *ncp, *nnp; 1919 struct mtx *vlp; 1920 1921 CTR1(KTR_VFS, "cache_purge_negative(%p)", vp); 1922 SDT_PROBE1(vfs, namecache, purge_negative, done, vp); 1923 TAILQ_INIT(&ncps); 1924 vlp = VP2VNODELOCK(vp); 1925 mtx_lock(vlp); 1926 LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) { 1927 if (!(ncp->nc_flag & NCF_NEGATIVE)) 1928 continue; 1929 cache_zap_negative_locked_vnode_kl(ncp, vp); 1930 TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst); 1931 } 1932 mtx_unlock(vlp); 1933 TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) { 1934 cache_free(ncp); 1935 } 1936 } 1937 1938 /* 1939 * Flush all entries referencing a particular filesystem. 1940 */ 1941 void 1942 cache_purgevfs(struct mount *mp, bool force) 1943 { 1944 TAILQ_HEAD(, namecache) ncps; 1945 struct mtx *vlp1, *vlp2; 1946 struct rwlock *blp; 1947 struct nchashhead *bucket; 1948 struct namecache *ncp, *nnp; 1949 u_long i, j, n_nchash; 1950 int error; 1951 1952 /* Scan hash tables for applicable entries */ 1953 SDT_PROBE1(vfs, namecache, purgevfs, done, mp); 1954 if (!force && mp->mnt_nvnodelistsize <= ncpurgeminvnodes) 1955 return; 1956 TAILQ_INIT(&ncps); 1957 n_nchash = nchash + 1; 1958 vlp1 = vlp2 = NULL; 1959 for (i = 0; i < numbucketlocks; i++) { 1960 blp = (struct rwlock *)&bucketlocks[i]; 1961 rw_wlock(blp); 1962 for (j = i; j < n_nchash; j += numbucketlocks) { 1963 retry: 1964 bucket = &nchashtbl[j]; 1965 LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) { 1966 cache_assert_bucket_locked(ncp, RA_WLOCKED); 1967 if (ncp->nc_dvp->v_mount != mp) 1968 continue; 1969 error = cache_zap_wlocked_bucket_kl(ncp, blp, 1970 &vlp1, &vlp2); 1971 if (error != 0) 1972 goto retry; 1973 TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst); 1974 } 1975 } 1976 rw_wunlock(blp); 1977 if (vlp1 == NULL && vlp2 == NULL) 1978 cache_maybe_yield(); 1979 } 1980 if (vlp1 != NULL) 1981 mtx_unlock(vlp1); 1982 if (vlp2 != NULL) 1983 mtx_unlock(vlp2); 1984 1985 TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) { 1986 cache_free(ncp); 1987 } 1988 } 1989 1990 /* 1991 * Perform canonical checks and cache lookup and pass on to filesystem 1992 * through the vop_cachedlookup only if needed. 1993 */ 1994 1995 int 1996 vfs_cache_lookup(struct vop_lookup_args *ap) 1997 { 1998 struct vnode *dvp; 1999 int error; 2000 struct vnode **vpp = ap->a_vpp; 2001 struct componentname *cnp = ap->a_cnp; 2002 struct ucred *cred = cnp->cn_cred; 2003 int flags = cnp->cn_flags; 2004 struct thread *td = cnp->cn_thread; 2005 2006 *vpp = NULL; 2007 dvp = ap->a_dvp; 2008 2009 if (dvp->v_type != VDIR) 2010 return (ENOTDIR); 2011 2012 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) && 2013 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) 2014 return (EROFS); 2015 2016 error = VOP_ACCESS(dvp, VEXEC, cred, td); 2017 if (error) 2018 return (error); 2019 2020 error = cache_lookup(dvp, vpp, cnp, NULL, NULL); 2021 if (error == 0) 2022 return (VOP_CACHEDLOOKUP(dvp, vpp, cnp)); 2023 if (error == -1) 2024 return (0); 2025 return (error); 2026 } 2027 2028 /* 2029 * XXX All of these sysctls would probably be more productive dead. 2030 */ 2031 static int disablecwd; 2032 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, 2033 "Disable the getcwd syscall"); 2034 2035 /* Implementation of the getcwd syscall. */ 2036 int 2037 sys___getcwd(struct thread *td, struct __getcwd_args *uap) 2038 { 2039 2040 return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen, 2041 MAXPATHLEN)); 2042 } 2043 2044 int 2045 kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, u_int buflen, 2046 u_int path_max) 2047 { 2048 char *bp, *tmpbuf; 2049 struct filedesc *fdp; 2050 struct vnode *cdir, *rdir; 2051 int error; 2052 2053 if (disablecwd) 2054 return (ENODEV); 2055 if (buflen < 2) 2056 return (EINVAL); 2057 if (buflen > path_max) 2058 buflen = path_max; 2059 2060 tmpbuf = malloc(buflen, M_TEMP, M_WAITOK); 2061 fdp = td->td_proc->p_fd; 2062 FILEDESC_SLOCK(fdp); 2063 cdir = fdp->fd_cdir; 2064 VREF(cdir); 2065 rdir = fdp->fd_rdir; 2066 VREF(rdir); 2067 FILEDESC_SUNLOCK(fdp); 2068 error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen); 2069 vrele(rdir); 2070 vrele(cdir); 2071 2072 if (!error) { 2073 if (bufseg == UIO_SYSSPACE) 2074 bcopy(bp, buf, strlen(bp) + 1); 2075 else 2076 error = copyout(bp, buf, strlen(bp) + 1); 2077 #ifdef KTRACE 2078 if (KTRPOINT(curthread, KTR_NAMEI)) 2079 ktrnamei(bp); 2080 #endif 2081 } 2082 free(tmpbuf, M_TEMP); 2083 return (error); 2084 } 2085 2086 /* 2087 * Thus begins the fullpath magic. 2088 */ 2089 2090 static int disablefullpath; 2091 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0, 2092 "Disable the vn_fullpath function"); 2093 2094 /* 2095 * Retrieve the full filesystem path that correspond to a vnode from the name 2096 * cache (if available) 2097 */ 2098 int 2099 vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf) 2100 { 2101 char *buf; 2102 struct filedesc *fdp; 2103 struct vnode *rdir; 2104 int error; 2105 2106 if (disablefullpath) 2107 return (ENODEV); 2108 if (vn == NULL) 2109 return (EINVAL); 2110 2111 buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK); 2112 fdp = td->td_proc->p_fd; 2113 FILEDESC_SLOCK(fdp); 2114 rdir = fdp->fd_rdir; 2115 VREF(rdir); 2116 FILEDESC_SUNLOCK(fdp); 2117 error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN); 2118 vrele(rdir); 2119 2120 if (!error) 2121 *freebuf = buf; 2122 else 2123 free(buf, M_TEMP); 2124 return (error); 2125 } 2126 2127 /* 2128 * This function is similar to vn_fullpath, but it attempts to lookup the 2129 * pathname relative to the global root mount point. This is required for the 2130 * auditing sub-system, as audited pathnames must be absolute, relative to the 2131 * global root mount point. 2132 */ 2133 int 2134 vn_fullpath_global(struct thread *td, struct vnode *vn, 2135 char **retbuf, char **freebuf) 2136 { 2137 char *buf; 2138 int error; 2139 2140 if (disablefullpath) 2141 return (ENODEV); 2142 if (vn == NULL) 2143 return (EINVAL); 2144 buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK); 2145 error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN); 2146 if (!error) 2147 *freebuf = buf; 2148 else 2149 free(buf, M_TEMP); 2150 return (error); 2151 } 2152 2153 int 2154 vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen) 2155 { 2156 struct vnode *dvp; 2157 struct namecache *ncp; 2158 struct mtx *vlp; 2159 int error; 2160 2161 vlp = VP2VNODELOCK(*vp); 2162 mtx_lock(vlp); 2163 TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) { 2164 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) 2165 break; 2166 } 2167 if (ncp != NULL) { 2168 if (*buflen < ncp->nc_nlen) { 2169 mtx_unlock(vlp); 2170 vrele(*vp); 2171 counter_u64_add(numfullpathfail4, 1); 2172 error = ENOMEM; 2173 SDT_PROBE3(vfs, namecache, fullpath, return, error, 2174 vp, NULL); 2175 return (error); 2176 } 2177 *buflen -= ncp->nc_nlen; 2178 memcpy(buf + *buflen, nc_get_name(ncp), ncp->nc_nlen); 2179 SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp, 2180 nc_get_name(ncp), vp); 2181 dvp = *vp; 2182 *vp = ncp->nc_dvp; 2183 vref(*vp); 2184 mtx_unlock(vlp); 2185 vrele(dvp); 2186 return (0); 2187 } 2188 SDT_PROBE1(vfs, namecache, fullpath, miss, vp); 2189 2190 mtx_unlock(vlp); 2191 vn_lock(*vp, LK_SHARED | LK_RETRY); 2192 error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen); 2193 vput(*vp); 2194 if (error) { 2195 counter_u64_add(numfullpathfail2, 1); 2196 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); 2197 return (error); 2198 } 2199 2200 *vp = dvp; 2201 if (dvp->v_iflag & VI_DOOMED) { 2202 /* forced unmount */ 2203 vrele(dvp); 2204 error = ENOENT; 2205 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); 2206 return (error); 2207 } 2208 /* 2209 * *vp has its use count incremented still. 2210 */ 2211 2212 return (0); 2213 } 2214 2215 /* 2216 * The magic behind kern___getcwd() and vn_fullpath(). 2217 */ 2218 static int 2219 vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir, 2220 char *buf, char **retbuf, u_int buflen) 2221 { 2222 int error, slash_prefixed; 2223 #ifdef KDTRACE_HOOKS 2224 struct vnode *startvp = vp; 2225 #endif 2226 struct vnode *vp1; 2227 2228 buflen--; 2229 buf[buflen] = '\0'; 2230 error = 0; 2231 slash_prefixed = 0; 2232 2233 SDT_PROBE1(vfs, namecache, fullpath, entry, vp); 2234 counter_u64_add(numfullpathcalls, 1); 2235 vref(vp); 2236 if (vp->v_type != VDIR) { 2237 error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen); 2238 if (error) 2239 return (error); 2240 if (buflen == 0) { 2241 vrele(vp); 2242 return (ENOMEM); 2243 } 2244 buf[--buflen] = '/'; 2245 slash_prefixed = 1; 2246 } 2247 while (vp != rdir && vp != rootvnode) { 2248 if (vp->v_vflag & VV_ROOT) { 2249 if (vp->v_iflag & VI_DOOMED) { /* forced unmount */ 2250 vrele(vp); 2251 error = ENOENT; 2252 SDT_PROBE3(vfs, namecache, fullpath, return, 2253 error, vp, NULL); 2254 break; 2255 } 2256 vp1 = vp->v_mount->mnt_vnodecovered; 2257 vref(vp1); 2258 vrele(vp); 2259 vp = vp1; 2260 continue; 2261 } 2262 if (vp->v_type != VDIR) { 2263 vrele(vp); 2264 counter_u64_add(numfullpathfail1, 1); 2265 error = ENOTDIR; 2266 SDT_PROBE3(vfs, namecache, fullpath, return, 2267 error, vp, NULL); 2268 break; 2269 } 2270 error = vn_vptocnp(&vp, td->td_ucred, buf, &buflen); 2271 if (error) 2272 break; 2273 if (buflen == 0) { 2274 vrele(vp); 2275 error = ENOMEM; 2276 SDT_PROBE3(vfs, namecache, fullpath, return, error, 2277 startvp, NULL); 2278 break; 2279 } 2280 buf[--buflen] = '/'; 2281 slash_prefixed = 1; 2282 } 2283 if (error) 2284 return (error); 2285 if (!slash_prefixed) { 2286 if (buflen == 0) { 2287 vrele(vp); 2288 counter_u64_add(numfullpathfail4, 1); 2289 SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM, 2290 startvp, NULL); 2291 return (ENOMEM); 2292 } 2293 buf[--buflen] = '/'; 2294 } 2295 counter_u64_add(numfullpathfound, 1); 2296 vrele(vp); 2297 2298 SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen); 2299 *retbuf = buf + buflen; 2300 return (0); 2301 } 2302 2303 struct vnode * 2304 vn_dir_dd_ino(struct vnode *vp) 2305 { 2306 struct namecache *ncp; 2307 struct vnode *ddvp; 2308 struct mtx *vlp; 2309 2310 ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino"); 2311 vlp = VP2VNODELOCK(vp); 2312 mtx_lock(vlp); 2313 TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) { 2314 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) 2315 continue; 2316 ddvp = ncp->nc_dvp; 2317 vhold(ddvp); 2318 mtx_unlock(vlp); 2319 if (vget(ddvp, LK_SHARED | LK_NOWAIT | LK_VNHELD, curthread)) 2320 return (NULL); 2321 return (ddvp); 2322 } 2323 mtx_unlock(vlp); 2324 return (NULL); 2325 } 2326 2327 int 2328 vn_commname(struct vnode *vp, char *buf, u_int buflen) 2329 { 2330 struct namecache *ncp; 2331 struct mtx *vlp; 2332 int l; 2333 2334 vlp = VP2VNODELOCK(vp); 2335 mtx_lock(vlp); 2336 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) 2337 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) 2338 break; 2339 if (ncp == NULL) { 2340 mtx_unlock(vlp); 2341 return (ENOENT); 2342 } 2343 l = min(ncp->nc_nlen, buflen - 1); 2344 memcpy(buf, nc_get_name(ncp), l); 2345 mtx_unlock(vlp); 2346 buf[l] = '\0'; 2347 return (0); 2348 } 2349 2350 /* ABI compat shims for old kernel modules. */ 2351 #undef cache_enter 2352 2353 void cache_enter(struct vnode *dvp, struct vnode *vp, 2354 struct componentname *cnp); 2355 2356 void 2357 cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp) 2358 { 2359 2360 cache_enter_time(dvp, vp, cnp, NULL, NULL); 2361 } 2362 2363 /* 2364 * This function updates path string to vnode's full global path 2365 * and checks the size of the new path string against the pathlen argument. 2366 * 2367 * Requires a locked, referenced vnode. 2368 * Vnode is re-locked on success or ENODEV, otherwise unlocked. 2369 * 2370 * If sysctl debug.disablefullpath is set, ENODEV is returned, 2371 * vnode is left locked and path remain untouched. 2372 * 2373 * If vp is a directory, the call to vn_fullpath_global() always succeeds 2374 * because it falls back to the ".." lookup if the namecache lookup fails. 2375 */ 2376 int 2377 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path, 2378 u_int pathlen) 2379 { 2380 struct nameidata nd; 2381 struct vnode *vp1; 2382 char *rpath, *fbuf; 2383 int error; 2384 2385 ASSERT_VOP_ELOCKED(vp, __func__); 2386 2387 /* Return ENODEV if sysctl debug.disablefullpath==1 */ 2388 if (disablefullpath) 2389 return (ENODEV); 2390 2391 /* Construct global filesystem path from vp. */ 2392 VOP_UNLOCK(vp, 0); 2393 error = vn_fullpath_global(td, vp, &rpath, &fbuf); 2394 2395 if (error != 0) { 2396 vrele(vp); 2397 return (error); 2398 } 2399 2400 if (strlen(rpath) >= pathlen) { 2401 vrele(vp); 2402 error = ENAMETOOLONG; 2403 goto out; 2404 } 2405 2406 /* 2407 * Re-lookup the vnode by path to detect a possible rename. 2408 * As a side effect, the vnode is relocked. 2409 * If vnode was renamed, return ENOENT. 2410 */ 2411 NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1, 2412 UIO_SYSSPACE, path, td); 2413 error = namei(&nd); 2414 if (error != 0) { 2415 vrele(vp); 2416 goto out; 2417 } 2418 NDFREE(&nd, NDF_ONLY_PNBUF); 2419 vp1 = nd.ni_vp; 2420 vrele(vp); 2421 if (vp1 == vp) 2422 strcpy(path, rpath); 2423 else { 2424 vput(vp1); 2425 error = ENOENT; 2426 } 2427 2428 out: 2429 free(fbuf, M_TEMP); 2430 return (error); 2431 } 2432