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 35 #include <sys/cdefs.h> 36 #include "opt_ddb.h" 37 #include "opt_ktrace.h" 38 39 #include <sys/param.h> 40 #include <sys/systm.h> 41 #include <sys/capsicum.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/ktr.h> 47 #include <sys/lock.h> 48 #include <sys/malloc.h> 49 #include <sys/fcntl.h> 50 #include <sys/jail.h> 51 #include <sys/mount.h> 52 #include <sys/namei.h> 53 #include <sys/proc.h> 54 #include <sys/seqc.h> 55 #include <sys/sdt.h> 56 #include <sys/smr.h> 57 #include <sys/smp.h> 58 #include <sys/syscallsubr.h> 59 #include <sys/sysctl.h> 60 #include <sys/sysproto.h> 61 #include <sys/vnode.h> 62 #include <ck_queue.h> 63 #ifdef KTRACE 64 #include <sys/ktrace.h> 65 #endif 66 #ifdef INVARIANTS 67 #include <machine/_inttypes.h> 68 #endif 69 70 #include <security/audit/audit.h> 71 #include <security/mac/mac_framework.h> 72 73 #ifdef DDB 74 #include <ddb/ddb.h> 75 #endif 76 77 #include <vm/uma.h> 78 79 /* 80 * High level overview of name caching in the VFS layer. 81 * 82 * Originally caching was implemented as part of UFS, later extracted to allow 83 * use by other filesystems. A decision was made to make it optional and 84 * completely detached from the rest of the kernel, which comes with limitations 85 * outlined near the end of this comment block. 86 * 87 * This fundamental choice needs to be revisited. In the meantime, the current 88 * state is described below. Significance of all notable routines is explained 89 * in comments placed above their implementation. Scattered thoroughout the 90 * file are TODO comments indicating shortcomings which can be fixed without 91 * reworking everything (most of the fixes will likely be reusable). Various 92 * details are omitted from this explanation to not clutter the overview, they 93 * have to be checked by reading the code and associated commentary. 94 * 95 * Keep in mind that it's individual path components which are cached, not full 96 * paths. That is, for a fully cached path "foo/bar/baz" there are 3 entries, 97 * one for each name. 98 * 99 * I. Data organization 100 * 101 * Entries are described by "struct namecache" objects and stored in a hash 102 * table. See cache_get_hash for more information. 103 * 104 * "struct vnode" contains pointers to source entries (names which can be found 105 * when traversing through said vnode), destination entries (names of that 106 * vnode (see "Limitations" for a breakdown on the subject) and a pointer to 107 * the parent vnode. 108 * 109 * The (directory vnode; name) tuple reliably determines the target entry if 110 * it exists. 111 * 112 * Since there are no small locks at this time (all are 32 bytes in size on 113 * LP64), the code works around the problem by introducing lock arrays to 114 * protect hash buckets and vnode lists. 115 * 116 * II. Filesystem integration 117 * 118 * Filesystems participating in name caching do the following: 119 * - set vop_lookup routine to vfs_cache_lookup 120 * - set vop_cachedlookup to whatever can perform the lookup if the above fails 121 * - if they support lockless lookup (see below), vop_fplookup_vexec and 122 * vop_fplookup_symlink are set along with the MNTK_FPLOOKUP flag on the 123 * mount point 124 * - call cache_purge or cache_vop_* routines to eliminate stale entries as 125 * applicable 126 * - call cache_enter to add entries depending on the MAKEENTRY flag 127 * 128 * With the above in mind, there are 2 entry points when doing lookups: 129 * - ... -> namei -> cache_fplookup -- this is the default 130 * - ... -> VOP_LOOKUP -> vfs_cache_lookup -- normally only called by namei 131 * should the above fail 132 * 133 * Example code flow how an entry is added: 134 * ... -> namei -> cache_fplookup -> cache_fplookup_noentry -> VOP_LOOKUP -> 135 * vfs_cache_lookup -> VOP_CACHEDLOOKUP -> ufs_lookup_ino -> cache_enter 136 * 137 * III. Performance considerations 138 * 139 * For lockless case forward lookup avoids any writes to shared areas apart 140 * from the terminal path component. In other words non-modifying lookups of 141 * different files don't suffer any scalability problems in the namecache. 142 * Looking up the same file is limited by VFS and goes beyond the scope of this 143 * file. 144 * 145 * At least on amd64 the single-threaded bottleneck for long paths is hashing 146 * (see cache_get_hash). There are cases where the code issues acquire fence 147 * multiple times, they can be combined on architectures which suffer from it. 148 * 149 * For locked case each encountered vnode has to be referenced and locked in 150 * order to be handed out to the caller (normally that's namei). This 151 * introduces significant hit single-threaded and serialization multi-threaded. 152 * 153 * Reverse lookup (e.g., "getcwd") fully scales provided it is fully cached -- 154 * avoids any writes to shared areas to any components. 155 * 156 * Unrelated insertions are partially serialized on updating the global entry 157 * counter and possibly serialized on colliding bucket or vnode locks. 158 * 159 * IV. Observability 160 * 161 * Note not everything has an explicit dtrace probe nor it should have, thus 162 * some of the one-liners below depend on implementation details. 163 * 164 * Examples: 165 * 166 * # Check what lookups failed to be handled in a lockless manner. Column 1 is 167 * # line number, column 2 is status code (see cache_fpl_status) 168 * dtrace -n 'vfs:fplookup:lookup:done { @[arg1, arg2] = count(); }' 169 * 170 * # Lengths of names added by binary name 171 * dtrace -n 'fbt::cache_enter_time:entry { @[execname] = quantize(args[2]->cn_namelen); }' 172 * 173 * # Same as above but only those which exceed 64 characters 174 * dtrace -n 'fbt::cache_enter_time:entry /args[2]->cn_namelen > 64/ { @[execname] = quantize(args[2]->cn_namelen); }' 175 * 176 * # Who is performing lookups with spurious slashes (e.g., "foo//bar") and what 177 * # path is it 178 * dtrace -n 'fbt::cache_fplookup_skip_slashes:entry { @[execname, stringof(args[0]->cnp->cn_pnbuf)] = count(); }' 179 * 180 * V. Limitations and implementation defects 181 * 182 * - since it is possible there is no entry for an open file, tools like 183 * "procstat" may fail to resolve fd -> vnode -> path to anything 184 * - even if a filesystem adds an entry, it may get purged (e.g., due to memory 185 * shortage) in which case the above problem applies 186 * - hardlinks are not tracked, thus if a vnode is reachable in more than one 187 * way, resolving a name may return a different path than the one used to 188 * open it (even if said path is still valid) 189 * - by default entries are not added for newly created files 190 * - adding an entry may need to evict negative entry first, which happens in 2 191 * distinct places (evicting on lookup, adding in a later VOP) making it 192 * impossible to simply reuse it 193 * - there is a simple scheme to evict negative entries as the cache is approaching 194 * its capacity, but it is very unclear if doing so is a good idea to begin with 195 * - vnodes are subject to being recycled even if target inode is left in memory, 196 * which loses the name cache entries when it perhaps should not. in case of tmpfs 197 * names get duplicated -- kept by filesystem itself and namecache separately 198 * - struct namecache has a fixed size and comes in 2 variants, often wasting 199 * space. now hard to replace with malloc due to dependence on SMR, which 200 * requires UMA zones to opt in 201 * - lack of better integration with the kernel also turns nullfs into a layered 202 * filesystem instead of something which can take advantage of caching 203 * 204 * Appendix A: where is the time lost, expanding on paragraph III 205 * 206 * While some care went into optimizing lookups, there is still plenty of 207 * performance left on the table, most notably from single-threaded standpoint. 208 * Below is a woefully incomplete list of changes which can help. Ideas are 209 * mostly sketched out, no claim is made all kinks or prerequisites are laid 210 * out. 211 * 212 * Note there is performance lost all over VFS. 213 * 214 * === SMR-only lookup 215 * 216 * For commonly used ops like stat(2), when the terminal vnode *is* cached, 217 * lockless lookup could refrain from refing/locking the found vnode and 218 * instead return while within the SMR section. Then a call to, say, 219 * vop_stat_smr could do the work (or fail with EAGAIN), finally the result 220 * would be validated with seqc not changing. This would be faster 221 * single-threaded as it dodges atomics and would provide full scalability for 222 * multicore uses. This would *not* work for open(2) or other calls which need 223 * the vnode to hang around for the long haul, but would work for aforementioned 224 * stat(2) but also access(2), readlink(2), realpathat(2) and probably more. 225 * 226 * === hotpatching for sdt probes 227 * 228 * They result in *tons* of branches all over with rather regrettable codegen 229 * at times. Removing sdt probes altogether gives over 2% boost in lookup rate. 230 * Reworking the code to patch itself at runtime with asm goto would solve it. 231 * asm goto is fully supported by gcc and clang. 232 * 233 * === copyinstr 234 * 235 * On all architectures it operates one byte at a time, while it could be 236 * word-sized instead thanks to the Mycroft trick. 237 * 238 * API itself is rather pessimal for path lookup, accepting arbitrary sizes and 239 * *optionally* filling in the length parameter. 240 * 241 * Instead a new routine (copyinpath?) could be introduced, demanding a buffer 242 * size which is a multiply of the word (and never zero), with the length 243 * always returned. On top of it the routine could be allowed to transform the 244 * buffer in arbitrary ways, most notably writing past the found length (not to 245 * be confused with writing past buffer size) -- this would allow word-sized 246 * movs while checking for '\0' later. 247 * 248 * === detour through namei 249 * 250 * Currently one suffers being called from namei, which then has to check if 251 * things worked out locklessly. Instead the lockless lookup could be the 252 * actual entry point which calls what is currently namei as a fallback. 253 * 254 * === avoidable branches in cache_can_fplookup 255 * 256 * The cache_fast_lookup_enabled flag check could be hotpatchable (in fact if 257 * this is off, none of fplookup code should execute). 258 * 259 * Both audit and capsicum branches can be combined into one, but it requires 260 * paying off a lot of tech debt first. 261 * 262 * ni_startdir could be indicated with a flag in cn_flags, eliminating the 263 * branch. 264 * 265 * === mount stacks 266 * 267 * Crossing a mount requires checking if perhaps something is mounted on top. 268 * Instead, an additional entry could be added to struct mount with a pointer 269 * to the final mount on the stack. This would be recalculated on each 270 * mount/unmount. 271 * 272 * === root vnodes 273 * 274 * It could become part of the API contract to *always* have a rootvnode set in 275 * mnt_rootvnode. Such vnodes are annotated with VV_ROOT and vnlru would have 276 * to be modified to always skip them. 277 * 278 * === inactive on v_usecount reaching 0 279 * 280 * VOP_NEED_INACTIVE should not exist. Filesystems would indicate need for such 281 * processing with a bit in usecount. 282 * 283 * === v_holdcnt 284 * 285 * Hold count should probably get eliminated, but one can argue it is a useful 286 * feature. Even if so, handling of v_usecount could be decoupled from it -- 287 * vnlru et al would consider the vnode not-freeable if has either hold or 288 * usecount on it. 289 * 290 * This would eliminate 2 atomics. 291 */ 292 293 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 294 "Name cache"); 295 296 SDT_PROVIDER_DECLARE(vfs); 297 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *", 298 "struct vnode *"); 299 SDT_PROBE_DEFINE3(vfs, namecache, enter, duplicate, "struct vnode *", "char *", 300 "struct vnode *"); 301 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *", 302 "char *"); 303 SDT_PROBE_DEFINE2(vfs, namecache, fullpath_smr, hit, "struct vnode *", 304 "const char *"); 305 SDT_PROBE_DEFINE4(vfs, namecache, fullpath_smr, miss, "struct vnode *", 306 "struct namecache *", "int", "int"); 307 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *"); 308 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *", 309 "char *", "struct vnode *"); 310 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *"); 311 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int", 312 "struct vnode *", "char *"); 313 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *", 314 "struct vnode *"); 315 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative, 316 "struct vnode *", "char *"); 317 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *", 318 "char *"); 319 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, hit, "struct vnode *", 320 "struct componentname *"); 321 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, miss, "struct vnode *", 322 "struct componentname *"); 323 SDT_PROBE_DEFINE3(vfs, namecache, purge, done, "struct vnode *", "size_t", "size_t"); 324 SDT_PROBE_DEFINE1(vfs, namecache, purge, batch, "int"); 325 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *"); 326 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *"); 327 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *", 328 "struct vnode *"); 329 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *", 330 "char *"); 331 SDT_PROBE_DEFINE2(vfs, namecache, evict_negative, done, "struct vnode *", 332 "char *"); 333 SDT_PROBE_DEFINE1(vfs, namecache, symlink, alloc__fail, "size_t"); 334 335 SDT_PROBE_DEFINE3(vfs, fplookup, lookup, done, "struct nameidata", "int", "bool"); 336 SDT_PROBE_DECLARE(vfs, namei, lookup, entry); 337 SDT_PROBE_DECLARE(vfs, namei, lookup, return); 338 339 static char __read_frequently cache_fast_lookup_enabled = true; 340 341 /* 342 * This structure describes the elements in the cache of recent 343 * names looked up by namei. 344 */ 345 struct negstate { 346 u_char neg_flag; 347 u_char neg_hit; 348 }; 349 _Static_assert(sizeof(struct negstate) <= sizeof(struct vnode *), 350 "the state must fit in a union with a pointer without growing it"); 351 352 struct namecache { 353 LIST_ENTRY(namecache) nc_src; /* source vnode list */ 354 TAILQ_ENTRY(namecache) nc_dst; /* destination vnode list */ 355 CK_SLIST_ENTRY(namecache) nc_hash;/* hash chain */ 356 struct vnode *nc_dvp; /* vnode of parent of name */ 357 union { 358 struct vnode *nu_vp; /* vnode the name refers to */ 359 struct negstate nu_neg;/* negative entry state */ 360 } n_un; 361 u_char nc_flag; /* flag bits */ 362 u_char nc_nlen; /* length of name */ 363 char nc_name[]; /* segment name + nul */ 364 }; 365 366 /* 367 * struct namecache_ts repeats struct namecache layout up to the 368 * nc_nlen member. 369 * struct namecache_ts is used in place of struct namecache when time(s) need 370 * to be stored. The nc_dotdottime field is used when a cache entry is mapping 371 * both a non-dotdot directory name plus dotdot for the directory's 372 * parent. 373 * 374 * See below for alignment requirement. 375 */ 376 struct namecache_ts { 377 struct timespec nc_time; /* timespec provided by fs */ 378 struct timespec nc_dotdottime; /* dotdot timespec provided by fs */ 379 int nc_ticks; /* ticks value when entry was added */ 380 int nc_pad; 381 struct namecache nc_nc; 382 }; 383 384 TAILQ_HEAD(cache_freebatch, namecache); 385 386 /* 387 * At least mips n32 performs 64-bit accesses to timespec as found 388 * in namecache_ts and requires them to be aligned. Since others 389 * may be in the same spot suffer a little bit and enforce the 390 * alignment for everyone. Note this is a nop for 64-bit platforms. 391 */ 392 #define CACHE_ZONE_ALIGNMENT UMA_ALIGNOF(time_t) 393 394 /* 395 * TODO: the initial value of CACHE_PATH_CUTOFF was inherited from the 396 * 4.4 BSD codebase. Later on struct namecache was tweaked to become 397 * smaller and the value was bumped to retain the total size, but it 398 * was never re-evaluated for suitability. A simple test counting 399 * lengths during package building shows that the value of 45 covers 400 * about 86% of all added entries, reaching 99% at 65. 401 * 402 * Regardless of the above, use of dedicated zones instead of malloc may be 403 * inducing additional waste. This may be hard to address as said zones are 404 * tied to VFS SMR. Even if retaining them, the current split should be 405 * re-evaluated. 406 */ 407 #ifdef __LP64__ 408 #define CACHE_PATH_CUTOFF 45 409 #define CACHE_LARGE_PAD 6 410 #else 411 #define CACHE_PATH_CUTOFF 41 412 #define CACHE_LARGE_PAD 2 413 #endif 414 415 #define CACHE_ZONE_SMALL_SIZE (offsetof(struct namecache, nc_name) + CACHE_PATH_CUTOFF + 1) 416 #define CACHE_ZONE_SMALL_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_SMALL_SIZE) 417 #define CACHE_ZONE_LARGE_SIZE (offsetof(struct namecache, nc_name) + NAME_MAX + 1 + CACHE_LARGE_PAD) 418 #define CACHE_ZONE_LARGE_TS_SIZE (offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_LARGE_SIZE) 419 420 _Static_assert((CACHE_ZONE_SMALL_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size"); 421 _Static_assert((CACHE_ZONE_SMALL_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size"); 422 _Static_assert((CACHE_ZONE_LARGE_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size"); 423 _Static_assert((CACHE_ZONE_LARGE_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size"); 424 425 #define nc_vp n_un.nu_vp 426 #define nc_neg n_un.nu_neg 427 428 /* 429 * Flags in namecache.nc_flag 430 */ 431 #define NCF_WHITE 0x01 432 #define NCF_ISDOTDOT 0x02 433 #define NCF_TS 0x04 434 #define NCF_DTS 0x08 435 #define NCF_DVDROP 0x10 436 #define NCF_NEGATIVE 0x20 437 #define NCF_INVALID 0x40 438 #define NCF_WIP 0x80 439 440 /* 441 * Flags in negstate.neg_flag 442 */ 443 #define NEG_HOT 0x01 444 445 static bool cache_neg_evict_cond(u_long lnumcache); 446 447 /* 448 * Mark an entry as invalid. 449 * 450 * This is called before it starts getting deconstructed. 451 */ 452 static void 453 cache_ncp_invalidate(struct namecache *ncp) 454 { 455 456 KASSERT((ncp->nc_flag & NCF_INVALID) == 0, 457 ("%s: entry %p already invalid", __func__, ncp)); 458 atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_INVALID); 459 atomic_thread_fence_rel(); 460 } 461 462 /* 463 * Check whether the entry can be safely used. 464 * 465 * All places which elide locks are supposed to call this after they are 466 * done with reading from an entry. 467 */ 468 #define cache_ncp_canuse(ncp) ({ \ 469 struct namecache *_ncp = (ncp); \ 470 u_char _nc_flag; \ 471 \ 472 atomic_thread_fence_acq(); \ 473 _nc_flag = atomic_load_char(&_ncp->nc_flag); \ 474 __predict_true((_nc_flag & (NCF_INVALID | NCF_WIP)) == 0); \ 475 }) 476 477 /* 478 * Like the above but also checks NCF_WHITE. 479 */ 480 #define cache_fpl_neg_ncp_canuse(ncp) ({ \ 481 struct namecache *_ncp = (ncp); \ 482 u_char _nc_flag; \ 483 \ 484 atomic_thread_fence_acq(); \ 485 _nc_flag = atomic_load_char(&_ncp->nc_flag); \ 486 __predict_true((_nc_flag & (NCF_INVALID | NCF_WIP | NCF_WHITE)) == 0); \ 487 }) 488 489 VFS_SMR_DECLARE; 490 491 static SYSCTL_NODE(_vfs_cache, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 492 "Name cache parameters"); 493 494 static u_int __read_mostly ncsize; /* the size as computed on creation or resizing */ 495 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, size, CTLFLAG_RD, &ncsize, 0, 496 "Total namecache capacity"); 497 498 u_int ncsizefactor = 2; 499 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, sizefactor, CTLFLAG_RW, &ncsizefactor, 0, 500 "Size factor for namecache"); 501 502 static u_long __read_mostly ncnegfactor = 5; /* ratio of negative entries */ 503 SYSCTL_ULONG(_vfs_cache_param, OID_AUTO, negfactor, CTLFLAG_RW, &ncnegfactor, 0, 504 "Ratio of negative namecache entries"); 505 506 /* 507 * Negative entry % of namecache capacity above which automatic eviction is allowed. 508 * 509 * Check cache_neg_evict_cond for details. 510 */ 511 static u_int ncnegminpct = 3; 512 513 static u_int __read_mostly neg_min; /* the above recomputed against ncsize */ 514 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, negmin, CTLFLAG_RD, &neg_min, 0, 515 "Negative entry count above which automatic eviction is allowed"); 516 517 /* 518 * Structures associated with name caching. 519 */ 520 #define NCHHASH(hash) \ 521 (&nchashtbl[(hash) & nchash]) 522 static __read_mostly CK_SLIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */ 523 static u_long __read_mostly nchash; /* size of hash table */ 524 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, 525 "Size of namecache hash table"); 526 static u_long __exclusive_cache_line numneg; /* number of negative entries allocated */ 527 static u_long __exclusive_cache_line numcache;/* number of cache entries allocated */ 528 529 struct nchstats nchstats; /* cache effectiveness statistics */ 530 531 static u_int __exclusive_cache_line neg_cycle; 532 533 #define ncneghash 3 534 #define numneglists (ncneghash + 1) 535 536 struct neglist { 537 struct mtx nl_evict_lock; 538 struct mtx nl_lock __aligned(CACHE_LINE_SIZE); 539 TAILQ_HEAD(, namecache) nl_list; 540 TAILQ_HEAD(, namecache) nl_hotlist; 541 u_long nl_hotnum; 542 } __aligned(CACHE_LINE_SIZE); 543 544 static struct neglist neglists[numneglists]; 545 546 static inline struct neglist * 547 NCP2NEGLIST(struct namecache *ncp) 548 { 549 550 return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]); 551 } 552 553 static inline struct negstate * 554 NCP2NEGSTATE(struct namecache *ncp) 555 { 556 557 MPASS(atomic_load_char(&ncp->nc_flag) & NCF_NEGATIVE); 558 return (&ncp->nc_neg); 559 } 560 561 #define numbucketlocks (ncbuckethash + 1) 562 static u_int __read_mostly ncbuckethash; 563 static struct mtx_padalign __read_mostly *bucketlocks; 564 #define HASH2BUCKETLOCK(hash) \ 565 ((struct mtx *)(&bucketlocks[((hash) & ncbuckethash)])) 566 567 #define numvnodelocks (ncvnodehash + 1) 568 static u_int __read_mostly ncvnodehash; 569 static struct mtx __read_mostly *vnodelocks; 570 static inline struct mtx * 571 VP2VNODELOCK(struct vnode *vp) 572 { 573 574 return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]); 575 } 576 577 static void 578 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp) 579 { 580 struct namecache_ts *ncp_ts; 581 582 KASSERT((ncp->nc_flag & NCF_TS) != 0 || 583 (tsp == NULL && ticksp == NULL), 584 ("No NCF_TS")); 585 586 if (tsp == NULL) 587 return; 588 589 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc); 590 *tsp = ncp_ts->nc_time; 591 *ticksp = ncp_ts->nc_ticks; 592 } 593 594 #ifdef DEBUG_CACHE 595 static int __read_mostly doingcache = 1; /* 1 => enable the cache */ 596 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0, 597 "VFS namecache enabled"); 598 #endif 599 600 /* Export size information to userland */ 601 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR, 602 sizeof(struct namecache), "sizeof(struct namecache)"); 603 604 /* 605 * The new name cache statistics 606 */ 607 static SYSCTL_NODE(_vfs_cache, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 608 "Name cache statistics"); 609 610 #define STATNODE_ULONG(name, varname, descr) \ 611 SYSCTL_ULONG(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr); 612 #define STATNODE_COUNTER(name, varname, descr) \ 613 static COUNTER_U64_DEFINE_EARLY(varname); \ 614 SYSCTL_COUNTER_U64(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, \ 615 descr); 616 STATNODE_ULONG(neg, numneg, "Number of negative cache entries"); 617 STATNODE_ULONG(count, numcache, "Number of cache entries"); 618 STATNODE_COUNTER(heldvnodes, numcachehv, "Number of namecache entries with vnodes held"); 619 STATNODE_COUNTER(drops, numdrops, "Number of dropped entries due to reaching the limit"); 620 STATNODE_COUNTER(miss, nummiss, "Number of cache misses"); 621 STATNODE_COUNTER(misszap, nummisszap, "Number of cache misses we do not want to cache"); 622 STATNODE_COUNTER(poszaps, numposzaps, 623 "Number of cache hits (positive) we do not want to cache"); 624 STATNODE_COUNTER(poshits, numposhits, "Number of cache hits (positive)"); 625 STATNODE_COUNTER(negzaps, numnegzaps, 626 "Number of cache hits (negative) we do not want to cache"); 627 STATNODE_COUNTER(neghits, numneghits, "Number of cache hits (negative)"); 628 /* These count for vn_getcwd(), too. */ 629 STATNODE_COUNTER(fullpathcalls, numfullpathcalls, "Number of fullpath search calls"); 630 STATNODE_COUNTER(fullpathfail2, numfullpathfail2, 631 "Number of fullpath search errors (VOP_VPTOCNP failures)"); 632 STATNODE_COUNTER(fullpathfail4, numfullpathfail4, "Number of fullpath search errors (ENOMEM)"); 633 STATNODE_COUNTER(fullpathfound, numfullpathfound, "Number of successful fullpath calls"); 634 STATNODE_COUNTER(symlinktoobig, symlinktoobig, "Number of times symlink did not fit the cache"); 635 636 /* 637 * Debug or developer statistics. 638 */ 639 static SYSCTL_NODE(_vfs_cache, OID_AUTO, debug, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 640 "Name cache debugging"); 641 #define DEBUGNODE_ULONG(name, varname, descr) \ 642 SYSCTL_ULONG(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr); 643 static u_long zap_bucket_relock_success; 644 DEBUGNODE_ULONG(zap_bucket_relock_success, zap_bucket_relock_success, 645 "Number of successful removals after relocking"); 646 static u_long zap_bucket_fail; 647 DEBUGNODE_ULONG(zap_bucket_fail, zap_bucket_fail, ""); 648 static u_long zap_bucket_fail2; 649 DEBUGNODE_ULONG(zap_bucket_fail2, zap_bucket_fail2, ""); 650 static u_long cache_lock_vnodes_cel_3_failures; 651 DEBUGNODE_ULONG(vnodes_cel_3_failures, cache_lock_vnodes_cel_3_failures, 652 "Number of times 3-way vnode locking failed"); 653 654 static void cache_zap_locked(struct namecache *ncp); 655 static int vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf, 656 char **retbuf, size_t *buflen, size_t addend); 657 static int vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, 658 char **retbuf, size_t *buflen); 659 static int vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, 660 char **retbuf, size_t *len, size_t addend); 661 662 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries"); 663 664 static inline void 665 cache_assert_vlp_locked(struct mtx *vlp) 666 { 667 668 if (vlp != NULL) 669 mtx_assert(vlp, MA_OWNED); 670 } 671 672 static inline void 673 cache_assert_vnode_locked(struct vnode *vp) 674 { 675 struct mtx *vlp; 676 677 vlp = VP2VNODELOCK(vp); 678 cache_assert_vlp_locked(vlp); 679 } 680 681 /* 682 * Directory vnodes with entries are held for two reasons: 683 * 1. make them less of a target for reclamation in vnlru 684 * 2. suffer smaller performance penalty in locked lookup as requeieing is avoided 685 * 686 * It will be feasible to stop doing it altogether if all filesystems start 687 * supporting lockless lookup. 688 */ 689 static void 690 cache_hold_vnode(struct vnode *vp) 691 { 692 693 cache_assert_vnode_locked(vp); 694 VNPASS(LIST_EMPTY(&vp->v_cache_src), vp); 695 vhold(vp); 696 counter_u64_add(numcachehv, 1); 697 } 698 699 static void 700 cache_drop_vnode(struct vnode *vp) 701 { 702 703 /* 704 * Called after all locks are dropped, meaning we can't assert 705 * on the state of v_cache_src. 706 */ 707 vdrop(vp); 708 counter_u64_add(numcachehv, -1); 709 } 710 711 /* 712 * UMA zones. 713 */ 714 static uma_zone_t __read_mostly cache_zone_small; 715 static uma_zone_t __read_mostly cache_zone_small_ts; 716 static uma_zone_t __read_mostly cache_zone_large; 717 static uma_zone_t __read_mostly cache_zone_large_ts; 718 719 char * 720 cache_symlink_alloc(size_t size, int flags) 721 { 722 723 if (size < CACHE_ZONE_SMALL_SIZE) { 724 return (uma_zalloc_smr(cache_zone_small, flags)); 725 } 726 if (size < CACHE_ZONE_LARGE_SIZE) { 727 return (uma_zalloc_smr(cache_zone_large, flags)); 728 } 729 counter_u64_add(symlinktoobig, 1); 730 SDT_PROBE1(vfs, namecache, symlink, alloc__fail, size); 731 return (NULL); 732 } 733 734 void 735 cache_symlink_free(char *string, size_t size) 736 { 737 738 MPASS(string != NULL); 739 KASSERT(size < CACHE_ZONE_LARGE_SIZE, 740 ("%s: size %zu too big", __func__, size)); 741 742 if (size < CACHE_ZONE_SMALL_SIZE) { 743 uma_zfree_smr(cache_zone_small, string); 744 return; 745 } 746 if (size < CACHE_ZONE_LARGE_SIZE) { 747 uma_zfree_smr(cache_zone_large, string); 748 return; 749 } 750 __assert_unreachable(); 751 } 752 753 static struct namecache * 754 cache_alloc_uma(int len, bool ts) 755 { 756 struct namecache_ts *ncp_ts; 757 struct namecache *ncp; 758 759 if (__predict_false(ts)) { 760 if (len <= CACHE_PATH_CUTOFF) 761 ncp_ts = uma_zalloc_smr(cache_zone_small_ts, M_WAITOK); 762 else 763 ncp_ts = uma_zalloc_smr(cache_zone_large_ts, M_WAITOK); 764 ncp = &ncp_ts->nc_nc; 765 } else { 766 if (len <= CACHE_PATH_CUTOFF) 767 ncp = uma_zalloc_smr(cache_zone_small, M_WAITOK); 768 else 769 ncp = uma_zalloc_smr(cache_zone_large, M_WAITOK); 770 } 771 return (ncp); 772 } 773 774 static void 775 cache_free_uma(struct namecache *ncp) 776 { 777 struct namecache_ts *ncp_ts; 778 779 if (__predict_false(ncp->nc_flag & NCF_TS)) { 780 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc); 781 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF) 782 uma_zfree_smr(cache_zone_small_ts, ncp_ts); 783 else 784 uma_zfree_smr(cache_zone_large_ts, ncp_ts); 785 } else { 786 if (ncp->nc_nlen <= CACHE_PATH_CUTOFF) 787 uma_zfree_smr(cache_zone_small, ncp); 788 else 789 uma_zfree_smr(cache_zone_large, ncp); 790 } 791 } 792 793 static struct namecache * 794 cache_alloc(int len, bool ts) 795 { 796 u_long lnumcache; 797 798 /* 799 * Avoid blowout in namecache entries. 800 * 801 * Bugs: 802 * 1. filesystems may end up trying to add an already existing entry 803 * (for example this can happen after a cache miss during concurrent 804 * lookup), in which case we will call cache_neg_evict despite not 805 * adding anything. 806 * 2. the routine may fail to free anything and no provisions are made 807 * to make it try harder (see the inside for failure modes) 808 * 3. it only ever looks at negative entries. 809 */ 810 lnumcache = atomic_fetchadd_long(&numcache, 1) + 1; 811 if (cache_neg_evict_cond(lnumcache)) { 812 lnumcache = atomic_load_long(&numcache); 813 } 814 if (__predict_false(lnumcache >= ncsize)) { 815 atomic_subtract_long(&numcache, 1); 816 counter_u64_add(numdrops, 1); 817 return (NULL); 818 } 819 return (cache_alloc_uma(len, ts)); 820 } 821 822 static void 823 cache_free(struct namecache *ncp) 824 { 825 826 MPASS(ncp != NULL); 827 if ((ncp->nc_flag & NCF_DVDROP) != 0) { 828 cache_drop_vnode(ncp->nc_dvp); 829 } 830 cache_free_uma(ncp); 831 atomic_subtract_long(&numcache, 1); 832 } 833 834 static void 835 cache_free_batch(struct cache_freebatch *batch) 836 { 837 struct namecache *ncp, *nnp; 838 int i; 839 840 i = 0; 841 if (TAILQ_EMPTY(batch)) 842 goto out; 843 TAILQ_FOREACH_SAFE(ncp, batch, nc_dst, nnp) { 844 if ((ncp->nc_flag & NCF_DVDROP) != 0) { 845 cache_drop_vnode(ncp->nc_dvp); 846 } 847 cache_free_uma(ncp); 848 i++; 849 } 850 atomic_subtract_long(&numcache, i); 851 out: 852 SDT_PROBE1(vfs, namecache, purge, batch, i); 853 } 854 855 /* 856 * Hashing. 857 * 858 * The code was made to use FNV in 2001 and this choice needs to be revisited. 859 * 860 * Short summary of the difficulty: 861 * The longest name which can be inserted is NAME_MAX characters in length (or 862 * 255 at the time of writing this comment), while majority of names used in 863 * practice are significantly shorter (mostly below 10). More importantly 864 * majority of lookups performed find names are even shorter than that. 865 * 866 * This poses a problem where hashes which do better than FNV past word size 867 * (or so) tend to come with additional overhead when finalizing the result, 868 * making them noticeably slower for the most commonly used range. 869 * 870 * Consider a path like: /usr/obj/usr/src/sys/amd64/GENERIC/vnode_if.c 871 * 872 * When looking it up the most time consuming part by a large margin (at least 873 * on amd64) is hashing. Replacing FNV with something which pessimizes short 874 * input would make the slowest part stand out even more. 875 */ 876 877 /* 878 * TODO: With the value stored we can do better than computing the hash based 879 * on the address. 880 */ 881 static void 882 cache_prehash(struct vnode *vp) 883 { 884 885 vp->v_nchash = fnv_32_buf(&vp, sizeof(vp), FNV1_32_INIT); 886 } 887 888 static uint32_t 889 cache_get_hash(char *name, u_char len, struct vnode *dvp) 890 { 891 892 return (fnv_32_buf(name, len, dvp->v_nchash)); 893 } 894 895 static uint32_t 896 cache_get_hash_iter_start(struct vnode *dvp) 897 { 898 899 return (dvp->v_nchash); 900 } 901 902 static uint32_t 903 cache_get_hash_iter(char c, uint32_t hash) 904 { 905 906 return (fnv_32_buf(&c, 1, hash)); 907 } 908 909 static uint32_t 910 cache_get_hash_iter_finish(uint32_t hash) 911 { 912 913 return (hash); 914 } 915 916 static inline struct nchashhead * 917 NCP2BUCKET(struct namecache *ncp) 918 { 919 uint32_t hash; 920 921 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp); 922 return (NCHHASH(hash)); 923 } 924 925 static inline struct mtx * 926 NCP2BUCKETLOCK(struct namecache *ncp) 927 { 928 uint32_t hash; 929 930 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp); 931 return (HASH2BUCKETLOCK(hash)); 932 } 933 934 #ifdef INVARIANTS 935 static void 936 cache_assert_bucket_locked(struct namecache *ncp) 937 { 938 struct mtx *blp; 939 940 blp = NCP2BUCKETLOCK(ncp); 941 mtx_assert(blp, MA_OWNED); 942 } 943 944 static void 945 cache_assert_bucket_unlocked(struct namecache *ncp) 946 { 947 struct mtx *blp; 948 949 blp = NCP2BUCKETLOCK(ncp); 950 mtx_assert(blp, MA_NOTOWNED); 951 } 952 #else 953 #define cache_assert_bucket_locked(x) do { } while (0) 954 #define cache_assert_bucket_unlocked(x) do { } while (0) 955 #endif 956 957 #define cache_sort_vnodes(x, y) _cache_sort_vnodes((void **)(x), (void **)(y)) 958 static void 959 _cache_sort_vnodes(void **p1, void **p2) 960 { 961 void *tmp; 962 963 MPASS(*p1 != NULL || *p2 != NULL); 964 965 if (*p1 > *p2) { 966 tmp = *p2; 967 *p2 = *p1; 968 *p1 = tmp; 969 } 970 } 971 972 static void 973 cache_lock_all_buckets(void) 974 { 975 u_int i; 976 977 for (i = 0; i < numbucketlocks; i++) 978 mtx_lock(&bucketlocks[i]); 979 } 980 981 static void 982 cache_unlock_all_buckets(void) 983 { 984 u_int i; 985 986 for (i = 0; i < numbucketlocks; i++) 987 mtx_unlock(&bucketlocks[i]); 988 } 989 990 static void 991 cache_lock_all_vnodes(void) 992 { 993 u_int i; 994 995 for (i = 0; i < numvnodelocks; i++) 996 mtx_lock(&vnodelocks[i]); 997 } 998 999 static void 1000 cache_unlock_all_vnodes(void) 1001 { 1002 u_int i; 1003 1004 for (i = 0; i < numvnodelocks; i++) 1005 mtx_unlock(&vnodelocks[i]); 1006 } 1007 1008 static int 1009 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2) 1010 { 1011 1012 cache_sort_vnodes(&vlp1, &vlp2); 1013 1014 if (vlp1 != NULL) { 1015 if (!mtx_trylock(vlp1)) 1016 return (EAGAIN); 1017 } 1018 if (!mtx_trylock(vlp2)) { 1019 if (vlp1 != NULL) 1020 mtx_unlock(vlp1); 1021 return (EAGAIN); 1022 } 1023 1024 return (0); 1025 } 1026 1027 static void 1028 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2) 1029 { 1030 1031 MPASS(vlp1 != NULL || vlp2 != NULL); 1032 MPASS(vlp1 <= vlp2); 1033 1034 if (vlp1 != NULL) 1035 mtx_lock(vlp1); 1036 if (vlp2 != NULL) 1037 mtx_lock(vlp2); 1038 } 1039 1040 static void 1041 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2) 1042 { 1043 1044 MPASS(vlp1 != NULL || vlp2 != NULL); 1045 1046 if (vlp1 != NULL) 1047 mtx_unlock(vlp1); 1048 if (vlp2 != NULL) 1049 mtx_unlock(vlp2); 1050 } 1051 1052 static int 1053 sysctl_nchstats(SYSCTL_HANDLER_ARGS) 1054 { 1055 struct nchstats snap; 1056 1057 if (req->oldptr == NULL) 1058 return (SYSCTL_OUT(req, 0, sizeof(snap))); 1059 1060 snap = nchstats; 1061 snap.ncs_goodhits = counter_u64_fetch(numposhits); 1062 snap.ncs_neghits = counter_u64_fetch(numneghits); 1063 snap.ncs_badhits = counter_u64_fetch(numposzaps) + 1064 counter_u64_fetch(numnegzaps); 1065 snap.ncs_miss = counter_u64_fetch(nummisszap) + 1066 counter_u64_fetch(nummiss); 1067 1068 return (SYSCTL_OUT(req, &snap, sizeof(snap))); 1069 } 1070 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD | 1071 CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU", 1072 "VFS cache effectiveness statistics"); 1073 1074 static void 1075 cache_recalc_neg_min(void) 1076 { 1077 1078 neg_min = (ncsize * ncnegminpct) / 100; 1079 } 1080 1081 static int 1082 sysctl_negminpct(SYSCTL_HANDLER_ARGS) 1083 { 1084 u_int val; 1085 int error; 1086 1087 val = ncnegminpct; 1088 error = sysctl_handle_int(oidp, &val, 0, req); 1089 if (error != 0 || req->newptr == NULL) 1090 return (error); 1091 1092 if (val == ncnegminpct) 1093 return (0); 1094 if (val < 0 || val > 99) 1095 return (EINVAL); 1096 ncnegminpct = val; 1097 cache_recalc_neg_min(); 1098 return (0); 1099 } 1100 1101 SYSCTL_PROC(_vfs_cache_param, OID_AUTO, negminpct, 1102 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_negminpct, 1103 "I", "Negative entry \% of namecache capacity above which automatic eviction is allowed"); 1104 1105 #ifdef DEBUG_CACHE 1106 /* 1107 * Grab an atomic snapshot of the name cache hash chain lengths 1108 */ 1109 static SYSCTL_NODE(_debug, OID_AUTO, hashstat, 1110 CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 1111 "hash table stats"); 1112 1113 static int 1114 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS) 1115 { 1116 struct nchashhead *ncpp; 1117 struct namecache *ncp; 1118 int i, error, n_nchash, *cntbuf; 1119 1120 retry: 1121 n_nchash = nchash + 1; /* nchash is max index, not count */ 1122 if (req->oldptr == NULL) 1123 return SYSCTL_OUT(req, 0, n_nchash * sizeof(int)); 1124 cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK); 1125 cache_lock_all_buckets(); 1126 if (n_nchash != nchash + 1) { 1127 cache_unlock_all_buckets(); 1128 free(cntbuf, M_TEMP); 1129 goto retry; 1130 } 1131 /* Scan hash tables counting entries */ 1132 for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++) 1133 CK_SLIST_FOREACH(ncp, ncpp, nc_hash) 1134 cntbuf[i]++; 1135 cache_unlock_all_buckets(); 1136 for (error = 0, i = 0; i < n_nchash; i++) 1137 if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0) 1138 break; 1139 free(cntbuf, M_TEMP); 1140 return (error); 1141 } 1142 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD| 1143 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int", 1144 "nchash chain lengths"); 1145 1146 static int 1147 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS) 1148 { 1149 int error; 1150 struct nchashhead *ncpp; 1151 struct namecache *ncp; 1152 int n_nchash; 1153 int count, maxlength, used, pct; 1154 1155 if (!req->oldptr) 1156 return SYSCTL_OUT(req, 0, 4 * sizeof(int)); 1157 1158 cache_lock_all_buckets(); 1159 n_nchash = nchash + 1; /* nchash is max index, not count */ 1160 used = 0; 1161 maxlength = 0; 1162 1163 /* Scan hash tables for applicable entries */ 1164 for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) { 1165 count = 0; 1166 CK_SLIST_FOREACH(ncp, ncpp, nc_hash) { 1167 count++; 1168 } 1169 if (count) 1170 used++; 1171 if (maxlength < count) 1172 maxlength = count; 1173 } 1174 n_nchash = nchash + 1; 1175 cache_unlock_all_buckets(); 1176 pct = (used * 100) / (n_nchash / 100); 1177 error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash)); 1178 if (error) 1179 return (error); 1180 error = SYSCTL_OUT(req, &used, sizeof(used)); 1181 if (error) 1182 return (error); 1183 error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength)); 1184 if (error) 1185 return (error); 1186 error = SYSCTL_OUT(req, &pct, sizeof(pct)); 1187 if (error) 1188 return (error); 1189 return (0); 1190 } 1191 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD| 1192 CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I", 1193 "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)"); 1194 #endif 1195 1196 /* 1197 * Negative entries management 1198 * 1199 * Various workloads create plenty of negative entries and barely use them 1200 * afterwards. Moreover malicious users can keep performing bogus lookups 1201 * adding even more entries. For example "make tinderbox" as of writing this 1202 * comment ends up with 2.6M namecache entries in total, 1.2M of which are 1203 * negative. 1204 * 1205 * As such, a rather aggressive eviction method is needed. The currently 1206 * employed method is a placeholder. 1207 * 1208 * Entries are split over numneglists separate lists, each of which is further 1209 * split into hot and cold entries. Entries get promoted after getting a hit. 1210 * Eviction happens on addition of new entry. 1211 */ 1212 static SYSCTL_NODE(_vfs_cache, OID_AUTO, neg, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1213 "Name cache negative entry statistics"); 1214 1215 SYSCTL_ULONG(_vfs_cache_neg, OID_AUTO, count, CTLFLAG_RD, &numneg, 0, 1216 "Number of negative cache entries"); 1217 1218 static COUNTER_U64_DEFINE_EARLY(neg_created); 1219 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, created, CTLFLAG_RD, &neg_created, 1220 "Number of created negative entries"); 1221 1222 static COUNTER_U64_DEFINE_EARLY(neg_evicted); 1223 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evicted, CTLFLAG_RD, &neg_evicted, 1224 "Number of evicted negative entries"); 1225 1226 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_empty); 1227 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_empty, CTLFLAG_RD, 1228 &neg_evict_skipped_empty, 1229 "Number of times evicting failed due to lack of entries"); 1230 1231 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_missed); 1232 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_missed, CTLFLAG_RD, 1233 &neg_evict_skipped_missed, 1234 "Number of times evicting failed due to target entry disappearing"); 1235 1236 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_contended); 1237 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_contended, CTLFLAG_RD, 1238 &neg_evict_skipped_contended, 1239 "Number of times evicting failed due to contention"); 1240 1241 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, hits, CTLFLAG_RD, &numneghits, 1242 "Number of cache hits (negative)"); 1243 1244 static int 1245 sysctl_neg_hot(SYSCTL_HANDLER_ARGS) 1246 { 1247 int i, out; 1248 1249 out = 0; 1250 for (i = 0; i < numneglists; i++) 1251 out += neglists[i].nl_hotnum; 1252 1253 return (SYSCTL_OUT(req, &out, sizeof(out))); 1254 } 1255 SYSCTL_PROC(_vfs_cache_neg, OID_AUTO, hot, CTLTYPE_INT | CTLFLAG_RD | 1256 CTLFLAG_MPSAFE, 0, 0, sysctl_neg_hot, "I", 1257 "Number of hot negative entries"); 1258 1259 static void 1260 cache_neg_init(struct namecache *ncp) 1261 { 1262 struct negstate *ns; 1263 1264 ncp->nc_flag |= NCF_NEGATIVE; 1265 ns = NCP2NEGSTATE(ncp); 1266 ns->neg_flag = 0; 1267 ns->neg_hit = 0; 1268 counter_u64_add(neg_created, 1); 1269 } 1270 1271 #define CACHE_NEG_PROMOTION_THRESH 2 1272 1273 static bool 1274 cache_neg_hit_prep(struct namecache *ncp) 1275 { 1276 struct negstate *ns; 1277 u_char n; 1278 1279 ns = NCP2NEGSTATE(ncp); 1280 n = atomic_load_char(&ns->neg_hit); 1281 for (;;) { 1282 if (n >= CACHE_NEG_PROMOTION_THRESH) 1283 return (false); 1284 if (atomic_fcmpset_8(&ns->neg_hit, &n, n + 1)) 1285 break; 1286 } 1287 return (n + 1 == CACHE_NEG_PROMOTION_THRESH); 1288 } 1289 1290 /* 1291 * Nothing to do here but it is provided for completeness as some 1292 * cache_neg_hit_prep callers may end up returning without even 1293 * trying to promote. 1294 */ 1295 #define cache_neg_hit_abort(ncp) do { } while (0) 1296 1297 static void 1298 cache_neg_hit_finish(struct namecache *ncp) 1299 { 1300 1301 SDT_PROBE2(vfs, namecache, lookup, hit__negative, ncp->nc_dvp, ncp->nc_name); 1302 counter_u64_add(numneghits, 1); 1303 } 1304 1305 /* 1306 * Move a negative entry to the hot list. 1307 */ 1308 static void 1309 cache_neg_promote_locked(struct namecache *ncp) 1310 { 1311 struct neglist *nl; 1312 struct negstate *ns; 1313 1314 ns = NCP2NEGSTATE(ncp); 1315 nl = NCP2NEGLIST(ncp); 1316 mtx_assert(&nl->nl_lock, MA_OWNED); 1317 if ((ns->neg_flag & NEG_HOT) == 0) { 1318 TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst); 1319 TAILQ_INSERT_TAIL(&nl->nl_hotlist, ncp, nc_dst); 1320 nl->nl_hotnum++; 1321 ns->neg_flag |= NEG_HOT; 1322 } 1323 } 1324 1325 /* 1326 * Move a hot negative entry to the cold list. 1327 */ 1328 static void 1329 cache_neg_demote_locked(struct namecache *ncp) 1330 { 1331 struct neglist *nl; 1332 struct negstate *ns; 1333 1334 ns = NCP2NEGSTATE(ncp); 1335 nl = NCP2NEGLIST(ncp); 1336 mtx_assert(&nl->nl_lock, MA_OWNED); 1337 MPASS(ns->neg_flag & NEG_HOT); 1338 TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst); 1339 TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst); 1340 nl->nl_hotnum--; 1341 ns->neg_flag &= ~NEG_HOT; 1342 atomic_store_char(&ns->neg_hit, 0); 1343 } 1344 1345 /* 1346 * Move a negative entry to the hot list if it matches the lookup. 1347 * 1348 * We have to take locks, but they may be contended and in the worst 1349 * case we may need to go off CPU. We don't want to spin within the 1350 * smr section and we can't block with it. Exiting the section means 1351 * the found entry could have been evicted. We are going to look it 1352 * up again. 1353 */ 1354 static bool 1355 cache_neg_promote_cond(struct vnode *dvp, struct componentname *cnp, 1356 struct namecache *oncp, uint32_t hash) 1357 { 1358 struct namecache *ncp; 1359 struct neglist *nl; 1360 u_char nc_flag; 1361 1362 nl = NCP2NEGLIST(oncp); 1363 1364 mtx_lock(&nl->nl_lock); 1365 /* 1366 * For hash iteration. 1367 */ 1368 vfs_smr_enter(); 1369 1370 /* 1371 * Avoid all surprises by only succeeding if we got the same entry and 1372 * bailing completely otherwise. 1373 * XXX There are no provisions to keep the vnode around, meaning we may 1374 * end up promoting a negative entry for a *new* vnode and returning 1375 * ENOENT on its account. This is the error we want to return anyway 1376 * and promotion is harmless. 1377 * 1378 * In particular at this point there can be a new ncp which matches the 1379 * search but hashes to a different neglist. 1380 */ 1381 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 1382 if (ncp == oncp) 1383 break; 1384 } 1385 1386 /* 1387 * No match to begin with. 1388 */ 1389 if (__predict_false(ncp == NULL)) { 1390 goto out_abort; 1391 } 1392 1393 /* 1394 * The newly found entry may be something different... 1395 */ 1396 if (!(ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && 1397 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))) { 1398 goto out_abort; 1399 } 1400 1401 /* 1402 * ... and not even negative. 1403 */ 1404 nc_flag = atomic_load_char(&ncp->nc_flag); 1405 if ((nc_flag & NCF_NEGATIVE) == 0) { 1406 goto out_abort; 1407 } 1408 1409 if (!cache_ncp_canuse(ncp)) { 1410 goto out_abort; 1411 } 1412 1413 cache_neg_promote_locked(ncp); 1414 cache_neg_hit_finish(ncp); 1415 vfs_smr_exit(); 1416 mtx_unlock(&nl->nl_lock); 1417 return (true); 1418 out_abort: 1419 vfs_smr_exit(); 1420 mtx_unlock(&nl->nl_lock); 1421 return (false); 1422 } 1423 1424 static void 1425 cache_neg_promote(struct namecache *ncp) 1426 { 1427 struct neglist *nl; 1428 1429 nl = NCP2NEGLIST(ncp); 1430 mtx_lock(&nl->nl_lock); 1431 cache_neg_promote_locked(ncp); 1432 mtx_unlock(&nl->nl_lock); 1433 } 1434 1435 static void 1436 cache_neg_insert(struct namecache *ncp) 1437 { 1438 struct neglist *nl; 1439 1440 MPASS(ncp->nc_flag & NCF_NEGATIVE); 1441 cache_assert_bucket_locked(ncp); 1442 nl = NCP2NEGLIST(ncp); 1443 mtx_lock(&nl->nl_lock); 1444 TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst); 1445 mtx_unlock(&nl->nl_lock); 1446 atomic_add_long(&numneg, 1); 1447 } 1448 1449 static void 1450 cache_neg_remove(struct namecache *ncp) 1451 { 1452 struct neglist *nl; 1453 struct negstate *ns; 1454 1455 cache_assert_bucket_locked(ncp); 1456 nl = NCP2NEGLIST(ncp); 1457 ns = NCP2NEGSTATE(ncp); 1458 mtx_lock(&nl->nl_lock); 1459 if ((ns->neg_flag & NEG_HOT) != 0) { 1460 TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst); 1461 nl->nl_hotnum--; 1462 } else { 1463 TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst); 1464 } 1465 mtx_unlock(&nl->nl_lock); 1466 atomic_subtract_long(&numneg, 1); 1467 } 1468 1469 static struct neglist * 1470 cache_neg_evict_select_list(void) 1471 { 1472 struct neglist *nl; 1473 u_int c; 1474 1475 c = atomic_fetchadd_int(&neg_cycle, 1) + 1; 1476 nl = &neglists[c % numneglists]; 1477 if (!mtx_trylock(&nl->nl_evict_lock)) { 1478 counter_u64_add(neg_evict_skipped_contended, 1); 1479 return (NULL); 1480 } 1481 return (nl); 1482 } 1483 1484 static struct namecache * 1485 cache_neg_evict_select_entry(struct neglist *nl) 1486 { 1487 struct namecache *ncp, *lncp; 1488 struct negstate *ns, *lns; 1489 int i; 1490 1491 mtx_assert(&nl->nl_evict_lock, MA_OWNED); 1492 mtx_assert(&nl->nl_lock, MA_OWNED); 1493 ncp = TAILQ_FIRST(&nl->nl_list); 1494 if (ncp == NULL) 1495 return (NULL); 1496 lncp = ncp; 1497 lns = NCP2NEGSTATE(lncp); 1498 for (i = 1; i < 4; i++) { 1499 ncp = TAILQ_NEXT(ncp, nc_dst); 1500 if (ncp == NULL) 1501 break; 1502 ns = NCP2NEGSTATE(ncp); 1503 if (ns->neg_hit < lns->neg_hit) { 1504 lncp = ncp; 1505 lns = ns; 1506 } 1507 } 1508 return (lncp); 1509 } 1510 1511 static bool 1512 cache_neg_evict(void) 1513 { 1514 struct namecache *ncp, *ncp2; 1515 struct neglist *nl; 1516 struct vnode *dvp; 1517 struct mtx *dvlp; 1518 struct mtx *blp; 1519 uint32_t hash; 1520 u_char nlen; 1521 bool evicted; 1522 1523 nl = cache_neg_evict_select_list(); 1524 if (nl == NULL) { 1525 return (false); 1526 } 1527 1528 mtx_lock(&nl->nl_lock); 1529 ncp = TAILQ_FIRST(&nl->nl_hotlist); 1530 if (ncp != NULL) { 1531 cache_neg_demote_locked(ncp); 1532 } 1533 ncp = cache_neg_evict_select_entry(nl); 1534 if (ncp == NULL) { 1535 counter_u64_add(neg_evict_skipped_empty, 1); 1536 mtx_unlock(&nl->nl_lock); 1537 mtx_unlock(&nl->nl_evict_lock); 1538 return (false); 1539 } 1540 nlen = ncp->nc_nlen; 1541 dvp = ncp->nc_dvp; 1542 hash = cache_get_hash(ncp->nc_name, nlen, dvp); 1543 dvlp = VP2VNODELOCK(dvp); 1544 blp = HASH2BUCKETLOCK(hash); 1545 mtx_unlock(&nl->nl_lock); 1546 mtx_unlock(&nl->nl_evict_lock); 1547 mtx_lock(dvlp); 1548 mtx_lock(blp); 1549 /* 1550 * Note that since all locks were dropped above, the entry may be 1551 * gone or reallocated to be something else. 1552 */ 1553 CK_SLIST_FOREACH(ncp2, (NCHHASH(hash)), nc_hash) { 1554 if (ncp2 == ncp && ncp2->nc_dvp == dvp && 1555 ncp2->nc_nlen == nlen && (ncp2->nc_flag & NCF_NEGATIVE) != 0) 1556 break; 1557 } 1558 if (ncp2 == NULL) { 1559 counter_u64_add(neg_evict_skipped_missed, 1); 1560 ncp = NULL; 1561 evicted = false; 1562 } else { 1563 MPASS(dvlp == VP2VNODELOCK(ncp->nc_dvp)); 1564 MPASS(blp == NCP2BUCKETLOCK(ncp)); 1565 SDT_PROBE2(vfs, namecache, evict_negative, done, ncp->nc_dvp, 1566 ncp->nc_name); 1567 cache_zap_locked(ncp); 1568 counter_u64_add(neg_evicted, 1); 1569 evicted = true; 1570 } 1571 mtx_unlock(blp); 1572 mtx_unlock(dvlp); 1573 if (ncp != NULL) 1574 cache_free(ncp); 1575 return (evicted); 1576 } 1577 1578 /* 1579 * Maybe evict a negative entry to create more room. 1580 * 1581 * The ncnegfactor parameter limits what fraction of the total count 1582 * can comprise of negative entries. However, if the cache is just 1583 * warming up this leads to excessive evictions. As such, ncnegminpct 1584 * (recomputed to neg_min) dictates whether the above should be 1585 * applied. 1586 * 1587 * Try evicting if the cache is close to full capacity regardless of 1588 * other considerations. 1589 */ 1590 static bool 1591 cache_neg_evict_cond(u_long lnumcache) 1592 { 1593 u_long lnumneg; 1594 1595 if (ncsize - 1000 < lnumcache) 1596 goto out_evict; 1597 lnumneg = atomic_load_long(&numneg); 1598 if (lnumneg < neg_min) 1599 return (false); 1600 if (lnumneg * ncnegfactor < lnumcache) 1601 return (false); 1602 out_evict: 1603 return (cache_neg_evict()); 1604 } 1605 1606 /* 1607 * cache_zap_locked(): 1608 * 1609 * Removes a namecache entry from cache, whether it contains an actual 1610 * pointer to a vnode or if it is just a negative cache entry. 1611 */ 1612 static void 1613 cache_zap_locked(struct namecache *ncp) 1614 { 1615 struct nchashhead *ncpp; 1616 struct vnode *dvp, *vp; 1617 1618 dvp = ncp->nc_dvp; 1619 vp = ncp->nc_vp; 1620 1621 if (!(ncp->nc_flag & NCF_NEGATIVE)) 1622 cache_assert_vnode_locked(vp); 1623 cache_assert_vnode_locked(dvp); 1624 cache_assert_bucket_locked(ncp); 1625 1626 cache_ncp_invalidate(ncp); 1627 1628 ncpp = NCP2BUCKET(ncp); 1629 CK_SLIST_REMOVE(ncpp, ncp, namecache, nc_hash); 1630 if (!(ncp->nc_flag & NCF_NEGATIVE)) { 1631 SDT_PROBE3(vfs, namecache, zap, done, dvp, ncp->nc_name, vp); 1632 TAILQ_REMOVE(&vp->v_cache_dst, ncp, nc_dst); 1633 if (ncp == vp->v_cache_dd) { 1634 atomic_store_ptr(&vp->v_cache_dd, NULL); 1635 } 1636 } else { 1637 SDT_PROBE2(vfs, namecache, zap_negative, done, dvp, ncp->nc_name); 1638 cache_neg_remove(ncp); 1639 } 1640 if (ncp->nc_flag & NCF_ISDOTDOT) { 1641 if (ncp == dvp->v_cache_dd) { 1642 atomic_store_ptr(&dvp->v_cache_dd, NULL); 1643 } 1644 } else { 1645 LIST_REMOVE(ncp, nc_src); 1646 if (LIST_EMPTY(&dvp->v_cache_src)) { 1647 ncp->nc_flag |= NCF_DVDROP; 1648 } 1649 } 1650 } 1651 1652 static void 1653 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp) 1654 { 1655 struct mtx *blp; 1656 1657 MPASS(ncp->nc_dvp == vp); 1658 MPASS(ncp->nc_flag & NCF_NEGATIVE); 1659 cache_assert_vnode_locked(vp); 1660 1661 blp = NCP2BUCKETLOCK(ncp); 1662 mtx_lock(blp); 1663 cache_zap_locked(ncp); 1664 mtx_unlock(blp); 1665 } 1666 1667 static bool 1668 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp, 1669 struct mtx **vlpp) 1670 { 1671 struct mtx *pvlp, *vlp1, *vlp2, *to_unlock; 1672 struct mtx *blp; 1673 1674 MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp); 1675 cache_assert_vnode_locked(vp); 1676 1677 if (ncp->nc_flag & NCF_NEGATIVE) { 1678 if (*vlpp != NULL) { 1679 mtx_unlock(*vlpp); 1680 *vlpp = NULL; 1681 } 1682 cache_zap_negative_locked_vnode_kl(ncp, vp); 1683 return (true); 1684 } 1685 1686 pvlp = VP2VNODELOCK(vp); 1687 blp = NCP2BUCKETLOCK(ncp); 1688 vlp1 = VP2VNODELOCK(ncp->nc_dvp); 1689 vlp2 = VP2VNODELOCK(ncp->nc_vp); 1690 1691 if (*vlpp == vlp1 || *vlpp == vlp2) { 1692 to_unlock = *vlpp; 1693 *vlpp = NULL; 1694 } else { 1695 if (*vlpp != NULL) { 1696 mtx_unlock(*vlpp); 1697 *vlpp = NULL; 1698 } 1699 cache_sort_vnodes(&vlp1, &vlp2); 1700 if (vlp1 == pvlp) { 1701 mtx_lock(vlp2); 1702 to_unlock = vlp2; 1703 } else { 1704 if (!mtx_trylock(vlp1)) 1705 goto out_relock; 1706 to_unlock = vlp1; 1707 } 1708 } 1709 mtx_lock(blp); 1710 cache_zap_locked(ncp); 1711 mtx_unlock(blp); 1712 if (to_unlock != NULL) 1713 mtx_unlock(to_unlock); 1714 return (true); 1715 1716 out_relock: 1717 mtx_unlock(vlp2); 1718 mtx_lock(vlp1); 1719 mtx_lock(vlp2); 1720 MPASS(*vlpp == NULL); 1721 *vlpp = vlp1; 1722 return (false); 1723 } 1724 1725 /* 1726 * If trylocking failed we can get here. We know enough to take all needed locks 1727 * in the right order and re-lookup the entry. 1728 */ 1729 static int 1730 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp, 1731 struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash, 1732 struct mtx *blp) 1733 { 1734 struct namecache *rncp; 1735 struct mtx *rvlp; 1736 1737 cache_assert_bucket_unlocked(ncp); 1738 1739 cache_sort_vnodes(&dvlp, &vlp); 1740 cache_lock_vnodes(dvlp, vlp); 1741 mtx_lock(blp); 1742 CK_SLIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) { 1743 if (rncp == ncp && rncp->nc_dvp == dvp && 1744 rncp->nc_nlen == cnp->cn_namelen && 1745 !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen)) 1746 break; 1747 } 1748 1749 if (rncp == NULL) 1750 goto out_mismatch; 1751 1752 if (!(ncp->nc_flag & NCF_NEGATIVE)) 1753 rvlp = VP2VNODELOCK(rncp->nc_vp); 1754 else 1755 rvlp = NULL; 1756 if (rvlp != vlp) 1757 goto out_mismatch; 1758 1759 cache_zap_locked(rncp); 1760 mtx_unlock(blp); 1761 cache_unlock_vnodes(dvlp, vlp); 1762 atomic_add_long(&zap_bucket_relock_success, 1); 1763 return (0); 1764 1765 out_mismatch: 1766 mtx_unlock(blp); 1767 cache_unlock_vnodes(dvlp, vlp); 1768 return (EAGAIN); 1769 } 1770 1771 static int __noinline 1772 cache_zap_locked_bucket(struct namecache *ncp, struct componentname *cnp, 1773 uint32_t hash, struct mtx *blp) 1774 { 1775 struct mtx *dvlp, *vlp; 1776 struct vnode *dvp; 1777 1778 cache_assert_bucket_locked(ncp); 1779 1780 dvlp = VP2VNODELOCK(ncp->nc_dvp); 1781 vlp = NULL; 1782 if (!(ncp->nc_flag & NCF_NEGATIVE)) 1783 vlp = VP2VNODELOCK(ncp->nc_vp); 1784 if (cache_trylock_vnodes(dvlp, vlp) == 0) { 1785 cache_zap_locked(ncp); 1786 mtx_unlock(blp); 1787 cache_unlock_vnodes(dvlp, vlp); 1788 return (0); 1789 } 1790 1791 dvp = ncp->nc_dvp; 1792 mtx_unlock(blp); 1793 return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp)); 1794 } 1795 1796 static __noinline int 1797 cache_remove_cnp(struct vnode *dvp, struct componentname *cnp) 1798 { 1799 struct namecache *ncp; 1800 struct mtx *blp; 1801 struct mtx *dvlp, *dvlp2; 1802 uint32_t hash; 1803 int error; 1804 1805 if (cnp->cn_namelen == 2 && 1806 cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') { 1807 dvlp = VP2VNODELOCK(dvp); 1808 dvlp2 = NULL; 1809 mtx_lock(dvlp); 1810 retry_dotdot: 1811 ncp = dvp->v_cache_dd; 1812 if (ncp == NULL) { 1813 mtx_unlock(dvlp); 1814 if (dvlp2 != NULL) 1815 mtx_unlock(dvlp2); 1816 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp); 1817 return (0); 1818 } 1819 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) { 1820 if (!cache_zap_locked_vnode_kl2(ncp, dvp, &dvlp2)) 1821 goto retry_dotdot; 1822 MPASS(dvp->v_cache_dd == NULL); 1823 mtx_unlock(dvlp); 1824 if (dvlp2 != NULL) 1825 mtx_unlock(dvlp2); 1826 cache_free(ncp); 1827 } else { 1828 atomic_store_ptr(&dvp->v_cache_dd, NULL); 1829 mtx_unlock(dvlp); 1830 if (dvlp2 != NULL) 1831 mtx_unlock(dvlp2); 1832 } 1833 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp); 1834 return (1); 1835 } 1836 1837 /* 1838 * XXX note that access here is completely unlocked with no provisions 1839 * to keep the hash allocated. If one is sufficiently unlucky a 1840 * parallel cache resize can reallocate the hash, unmap backing pages 1841 * and cause the empty check below to fault. 1842 * 1843 * Fixing this has epsilon priority, but can be done with no overhead 1844 * for this codepath with sufficient effort. 1845 */ 1846 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); 1847 blp = HASH2BUCKETLOCK(hash); 1848 retry: 1849 if (CK_SLIST_EMPTY(NCHHASH(hash))) 1850 goto out_no_entry; 1851 1852 mtx_lock(blp); 1853 1854 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 1855 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && 1856 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) 1857 break; 1858 } 1859 1860 if (ncp == NULL) { 1861 mtx_unlock(blp); 1862 goto out_no_entry; 1863 } 1864 1865 error = cache_zap_locked_bucket(ncp, cnp, hash, blp); 1866 if (__predict_false(error != 0)) { 1867 atomic_add_long(&zap_bucket_fail, 1); 1868 goto retry; 1869 } 1870 counter_u64_add(numposzaps, 1); 1871 SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp); 1872 cache_free(ncp); 1873 return (1); 1874 out_no_entry: 1875 counter_u64_add(nummisszap, 1); 1876 SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp); 1877 return (0); 1878 } 1879 1880 static int __noinline 1881 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, 1882 struct timespec *tsp, int *ticksp) 1883 { 1884 int ltype; 1885 1886 *vpp = dvp; 1887 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp); 1888 if (tsp != NULL) 1889 timespecclear(tsp); 1890 if (ticksp != NULL) 1891 *ticksp = ticks; 1892 vrefact(*vpp); 1893 /* 1894 * When we lookup "." we still can be asked to lock it 1895 * differently... 1896 */ 1897 ltype = cnp->cn_lkflags & LK_TYPE_MASK; 1898 if (ltype != VOP_ISLOCKED(*vpp)) { 1899 if (ltype == LK_EXCLUSIVE) { 1900 vn_lock(*vpp, LK_UPGRADE | LK_RETRY); 1901 if (VN_IS_DOOMED((*vpp))) { 1902 /* forced unmount */ 1903 vrele(*vpp); 1904 *vpp = NULL; 1905 return (ENOENT); 1906 } 1907 } else 1908 vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY); 1909 } 1910 return (-1); 1911 } 1912 1913 static int __noinline 1914 cache_lookup_dotdot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, 1915 struct timespec *tsp, int *ticksp) 1916 { 1917 struct namecache_ts *ncp_ts; 1918 struct namecache *ncp; 1919 struct mtx *dvlp; 1920 enum vgetstate vs; 1921 int error, ltype; 1922 bool whiteout; 1923 1924 MPASS((cnp->cn_flags & ISDOTDOT) != 0); 1925 1926 if ((cnp->cn_flags & MAKEENTRY) == 0) { 1927 cache_remove_cnp(dvp, cnp); 1928 return (0); 1929 } 1930 1931 retry: 1932 dvlp = VP2VNODELOCK(dvp); 1933 mtx_lock(dvlp); 1934 ncp = dvp->v_cache_dd; 1935 if (ncp == NULL) { 1936 SDT_PROBE2(vfs, namecache, lookup, miss, dvp, ".."); 1937 mtx_unlock(dvlp); 1938 return (0); 1939 } 1940 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) { 1941 if (ncp->nc_flag & NCF_NEGATIVE) 1942 *vpp = NULL; 1943 else 1944 *vpp = ncp->nc_vp; 1945 } else 1946 *vpp = ncp->nc_dvp; 1947 if (*vpp == NULL) 1948 goto negative_success; 1949 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", *vpp); 1950 cache_out_ts(ncp, tsp, ticksp); 1951 if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) == 1952 NCF_DTS && tsp != NULL) { 1953 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc); 1954 *tsp = ncp_ts->nc_dotdottime; 1955 } 1956 1957 MPASS(dvp != *vpp); 1958 ltype = VOP_ISLOCKED(dvp); 1959 VOP_UNLOCK(dvp); 1960 vs = vget_prep(*vpp); 1961 mtx_unlock(dvlp); 1962 error = vget_finish(*vpp, cnp->cn_lkflags, vs); 1963 vn_lock(dvp, ltype | LK_RETRY); 1964 if (VN_IS_DOOMED(dvp)) { 1965 if (error == 0) 1966 vput(*vpp); 1967 *vpp = NULL; 1968 return (ENOENT); 1969 } 1970 if (error) { 1971 *vpp = NULL; 1972 goto retry; 1973 } 1974 return (-1); 1975 negative_success: 1976 if (__predict_false(cnp->cn_nameiop == CREATE)) { 1977 if (cnp->cn_flags & ISLASTCN) { 1978 counter_u64_add(numnegzaps, 1); 1979 cache_zap_negative_locked_vnode_kl(ncp, dvp); 1980 mtx_unlock(dvlp); 1981 cache_free(ncp); 1982 return (0); 1983 } 1984 } 1985 1986 whiteout = (ncp->nc_flag & NCF_WHITE); 1987 cache_out_ts(ncp, tsp, ticksp); 1988 if (cache_neg_hit_prep(ncp)) 1989 cache_neg_promote(ncp); 1990 else 1991 cache_neg_hit_finish(ncp); 1992 mtx_unlock(dvlp); 1993 if (whiteout) 1994 cnp->cn_flags |= ISWHITEOUT; 1995 return (ENOENT); 1996 } 1997 1998 /** 1999 * Lookup a name in the name cache 2000 * 2001 * # Arguments 2002 * 2003 * - dvp: Parent directory in which to search. 2004 * - vpp: Return argument. Will contain desired vnode on cache hit. 2005 * - cnp: Parameters of the name search. The most interesting bits of 2006 * the cn_flags field have the following meanings: 2007 * - MAKEENTRY: If clear, free an entry from the cache rather than look 2008 * it up. 2009 * - ISDOTDOT: Must be set if and only if cn_nameptr == ".." 2010 * - tsp: Return storage for cache timestamp. On a successful (positive 2011 * or negative) lookup, tsp will be filled with any timespec that 2012 * was stored when this cache entry was created. However, it will 2013 * be clear for "." entries. 2014 * - ticks: Return storage for alternate cache timestamp. On a successful 2015 * (positive or negative) lookup, it will contain the ticks value 2016 * that was current when the cache entry was created, unless cnp 2017 * was ".". 2018 * 2019 * Either both tsp and ticks have to be provided or neither of them. 2020 * 2021 * # Returns 2022 * 2023 * - -1: A positive cache hit. vpp will contain the desired vnode. 2024 * - ENOENT: A negative cache hit, or dvp was recycled out from under us due 2025 * to a forced unmount. vpp will not be modified. If the entry 2026 * is a whiteout, then the ISWHITEOUT flag will be set in 2027 * cnp->cn_flags. 2028 * - 0: A cache miss. vpp will not be modified. 2029 * 2030 * # Locking 2031 * 2032 * On a cache hit, vpp will be returned locked and ref'd. If we're looking up 2033 * .., dvp is unlocked. If we're looking up . an extra ref is taken, but the 2034 * lock is not recursively acquired. 2035 */ 2036 static int __noinline 2037 cache_lookup_fallback(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, 2038 struct timespec *tsp, int *ticksp) 2039 { 2040 struct namecache *ncp; 2041 struct mtx *blp; 2042 uint32_t hash; 2043 enum vgetstate vs; 2044 int error; 2045 bool whiteout; 2046 2047 MPASS((cnp->cn_flags & ISDOTDOT) == 0); 2048 MPASS((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) != 0); 2049 2050 retry: 2051 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); 2052 blp = HASH2BUCKETLOCK(hash); 2053 mtx_lock(blp); 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 mtx_unlock(blp); 2063 SDT_PROBE2(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr); 2064 counter_u64_add(nummiss, 1); 2065 return (0); 2066 } 2067 2068 if (ncp->nc_flag & NCF_NEGATIVE) 2069 goto negative_success; 2070 2071 counter_u64_add(numposhits, 1); 2072 *vpp = ncp->nc_vp; 2073 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp); 2074 cache_out_ts(ncp, tsp, ticksp); 2075 MPASS(dvp != *vpp); 2076 vs = vget_prep(*vpp); 2077 mtx_unlock(blp); 2078 error = vget_finish(*vpp, cnp->cn_lkflags, vs); 2079 if (error) { 2080 *vpp = NULL; 2081 goto retry; 2082 } 2083 return (-1); 2084 negative_success: 2085 /* 2086 * We don't get here with regular lookup apart from corner cases. 2087 */ 2088 if (__predict_true(cnp->cn_nameiop == CREATE)) { 2089 if (cnp->cn_flags & ISLASTCN) { 2090 counter_u64_add(numnegzaps, 1); 2091 error = cache_zap_locked_bucket(ncp, cnp, hash, blp); 2092 if (__predict_false(error != 0)) { 2093 atomic_add_long(&zap_bucket_fail2, 1); 2094 goto retry; 2095 } 2096 cache_free(ncp); 2097 return (0); 2098 } 2099 } 2100 2101 whiteout = (ncp->nc_flag & NCF_WHITE); 2102 cache_out_ts(ncp, tsp, ticksp); 2103 if (cache_neg_hit_prep(ncp)) 2104 cache_neg_promote(ncp); 2105 else 2106 cache_neg_hit_finish(ncp); 2107 mtx_unlock(blp); 2108 if (whiteout) 2109 cnp->cn_flags |= ISWHITEOUT; 2110 return (ENOENT); 2111 } 2112 2113 int 2114 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, 2115 struct timespec *tsp, int *ticksp) 2116 { 2117 struct namecache *ncp; 2118 uint32_t hash; 2119 enum vgetstate vs; 2120 int error; 2121 bool whiteout, neg_promote; 2122 u_short nc_flag; 2123 2124 MPASS((tsp == NULL && ticksp == NULL) || (tsp != NULL && ticksp != NULL)); 2125 2126 #ifdef DEBUG_CACHE 2127 if (__predict_false(!doingcache)) { 2128 cnp->cn_flags &= ~MAKEENTRY; 2129 return (0); 2130 } 2131 #endif 2132 2133 if (__predict_false(cnp->cn_nameptr[0] == '.')) { 2134 if (cnp->cn_namelen == 1) 2135 return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp)); 2136 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') 2137 return (cache_lookup_dotdot(dvp, vpp, cnp, tsp, ticksp)); 2138 } 2139 2140 MPASS((cnp->cn_flags & ISDOTDOT) == 0); 2141 2142 if ((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) == 0) { 2143 cache_remove_cnp(dvp, cnp); 2144 return (0); 2145 } 2146 2147 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); 2148 vfs_smr_enter(); 2149 2150 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 2151 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && 2152 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) 2153 break; 2154 } 2155 2156 if (__predict_false(ncp == NULL)) { 2157 vfs_smr_exit(); 2158 SDT_PROBE2(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr); 2159 counter_u64_add(nummiss, 1); 2160 return (0); 2161 } 2162 2163 nc_flag = atomic_load_char(&ncp->nc_flag); 2164 if (nc_flag & NCF_NEGATIVE) 2165 goto negative_success; 2166 2167 counter_u64_add(numposhits, 1); 2168 *vpp = ncp->nc_vp; 2169 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp); 2170 cache_out_ts(ncp, tsp, ticksp); 2171 MPASS(dvp != *vpp); 2172 if (!cache_ncp_canuse(ncp)) { 2173 vfs_smr_exit(); 2174 *vpp = NULL; 2175 goto out_fallback; 2176 } 2177 vs = vget_prep_smr(*vpp); 2178 vfs_smr_exit(); 2179 if (__predict_false(vs == VGET_NONE)) { 2180 *vpp = NULL; 2181 goto out_fallback; 2182 } 2183 error = vget_finish(*vpp, cnp->cn_lkflags, vs); 2184 if (error) { 2185 *vpp = NULL; 2186 goto out_fallback; 2187 } 2188 return (-1); 2189 negative_success: 2190 if (cnp->cn_nameiop == CREATE) { 2191 if (cnp->cn_flags & ISLASTCN) { 2192 vfs_smr_exit(); 2193 goto out_fallback; 2194 } 2195 } 2196 2197 cache_out_ts(ncp, tsp, ticksp); 2198 whiteout = (atomic_load_char(&ncp->nc_flag) & NCF_WHITE); 2199 neg_promote = cache_neg_hit_prep(ncp); 2200 if (!cache_ncp_canuse(ncp)) { 2201 cache_neg_hit_abort(ncp); 2202 vfs_smr_exit(); 2203 goto out_fallback; 2204 } 2205 if (neg_promote) { 2206 vfs_smr_exit(); 2207 if (!cache_neg_promote_cond(dvp, cnp, ncp, hash)) 2208 goto out_fallback; 2209 } else { 2210 cache_neg_hit_finish(ncp); 2211 vfs_smr_exit(); 2212 } 2213 if (whiteout) 2214 cnp->cn_flags |= ISWHITEOUT; 2215 return (ENOENT); 2216 out_fallback: 2217 return (cache_lookup_fallback(dvp, vpp, cnp, tsp, ticksp)); 2218 } 2219 2220 struct celockstate { 2221 struct mtx *vlp[3]; 2222 struct mtx *blp[2]; 2223 }; 2224 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3)); 2225 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2)); 2226 2227 static inline void 2228 cache_celockstate_init(struct celockstate *cel) 2229 { 2230 2231 bzero(cel, sizeof(*cel)); 2232 } 2233 2234 static void 2235 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp, 2236 struct vnode *dvp) 2237 { 2238 struct mtx *vlp1, *vlp2; 2239 2240 MPASS(cel->vlp[0] == NULL); 2241 MPASS(cel->vlp[1] == NULL); 2242 MPASS(cel->vlp[2] == NULL); 2243 2244 MPASS(vp != NULL || dvp != NULL); 2245 2246 vlp1 = VP2VNODELOCK(vp); 2247 vlp2 = VP2VNODELOCK(dvp); 2248 cache_sort_vnodes(&vlp1, &vlp2); 2249 2250 if (vlp1 != NULL) { 2251 mtx_lock(vlp1); 2252 cel->vlp[0] = vlp1; 2253 } 2254 mtx_lock(vlp2); 2255 cel->vlp[1] = vlp2; 2256 } 2257 2258 static void 2259 cache_unlock_vnodes_cel(struct celockstate *cel) 2260 { 2261 2262 MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL); 2263 2264 if (cel->vlp[0] != NULL) 2265 mtx_unlock(cel->vlp[0]); 2266 if (cel->vlp[1] != NULL) 2267 mtx_unlock(cel->vlp[1]); 2268 if (cel->vlp[2] != NULL) 2269 mtx_unlock(cel->vlp[2]); 2270 } 2271 2272 static bool 2273 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp) 2274 { 2275 struct mtx *vlp; 2276 bool ret; 2277 2278 cache_assert_vlp_locked(cel->vlp[0]); 2279 cache_assert_vlp_locked(cel->vlp[1]); 2280 MPASS(cel->vlp[2] == NULL); 2281 2282 MPASS(vp != NULL); 2283 vlp = VP2VNODELOCK(vp); 2284 2285 ret = true; 2286 if (vlp >= cel->vlp[1]) { 2287 mtx_lock(vlp); 2288 } else { 2289 if (mtx_trylock(vlp)) 2290 goto out; 2291 cache_unlock_vnodes_cel(cel); 2292 atomic_add_long(&cache_lock_vnodes_cel_3_failures, 1); 2293 if (vlp < cel->vlp[0]) { 2294 mtx_lock(vlp); 2295 mtx_lock(cel->vlp[0]); 2296 mtx_lock(cel->vlp[1]); 2297 } else { 2298 if (cel->vlp[0] != NULL) 2299 mtx_lock(cel->vlp[0]); 2300 mtx_lock(vlp); 2301 mtx_lock(cel->vlp[1]); 2302 } 2303 ret = false; 2304 } 2305 out: 2306 cel->vlp[2] = vlp; 2307 return (ret); 2308 } 2309 2310 static void 2311 cache_lock_buckets_cel(struct celockstate *cel, struct mtx *blp1, 2312 struct mtx *blp2) 2313 { 2314 2315 MPASS(cel->blp[0] == NULL); 2316 MPASS(cel->blp[1] == NULL); 2317 2318 cache_sort_vnodes(&blp1, &blp2); 2319 2320 if (blp1 != NULL) { 2321 mtx_lock(blp1); 2322 cel->blp[0] = blp1; 2323 } 2324 mtx_lock(blp2); 2325 cel->blp[1] = blp2; 2326 } 2327 2328 static void 2329 cache_unlock_buckets_cel(struct celockstate *cel) 2330 { 2331 2332 if (cel->blp[0] != NULL) 2333 mtx_unlock(cel->blp[0]); 2334 mtx_unlock(cel->blp[1]); 2335 } 2336 2337 /* 2338 * Lock part of the cache affected by the insertion. 2339 * 2340 * This means vnodelocks for dvp, vp and the relevant bucketlock. 2341 * However, insertion can result in removal of an old entry. In this 2342 * case we have an additional vnode and bucketlock pair to lock. 2343 * 2344 * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while 2345 * preserving the locking order (smaller address first). 2346 */ 2347 static void 2348 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp, 2349 uint32_t hash) 2350 { 2351 struct namecache *ncp; 2352 struct mtx *blps[2]; 2353 u_char nc_flag; 2354 2355 blps[0] = HASH2BUCKETLOCK(hash); 2356 for (;;) { 2357 blps[1] = NULL; 2358 cache_lock_vnodes_cel(cel, dvp, vp); 2359 if (vp == NULL || vp->v_type != VDIR) 2360 break; 2361 ncp = atomic_load_consume_ptr(&vp->v_cache_dd); 2362 if (ncp == NULL) 2363 break; 2364 nc_flag = atomic_load_char(&ncp->nc_flag); 2365 if ((nc_flag & NCF_ISDOTDOT) == 0) 2366 break; 2367 MPASS(ncp->nc_dvp == vp); 2368 blps[1] = NCP2BUCKETLOCK(ncp); 2369 if ((nc_flag & NCF_NEGATIVE) != 0) 2370 break; 2371 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp)) 2372 break; 2373 /* 2374 * All vnodes got re-locked. Re-validate the state and if 2375 * nothing changed we are done. Otherwise restart. 2376 */ 2377 if (ncp == vp->v_cache_dd && 2378 (ncp->nc_flag & NCF_ISDOTDOT) != 0 && 2379 blps[1] == NCP2BUCKETLOCK(ncp) && 2380 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2]) 2381 break; 2382 cache_unlock_vnodes_cel(cel); 2383 cel->vlp[0] = NULL; 2384 cel->vlp[1] = NULL; 2385 cel->vlp[2] = NULL; 2386 } 2387 cache_lock_buckets_cel(cel, blps[0], blps[1]); 2388 } 2389 2390 static void 2391 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp, 2392 uint32_t hash) 2393 { 2394 struct namecache *ncp; 2395 struct mtx *blps[2]; 2396 u_char nc_flag; 2397 2398 blps[0] = HASH2BUCKETLOCK(hash); 2399 for (;;) { 2400 blps[1] = NULL; 2401 cache_lock_vnodes_cel(cel, dvp, vp); 2402 ncp = atomic_load_consume_ptr(&dvp->v_cache_dd); 2403 if (ncp == NULL) 2404 break; 2405 nc_flag = atomic_load_char(&ncp->nc_flag); 2406 if ((nc_flag & NCF_ISDOTDOT) == 0) 2407 break; 2408 MPASS(ncp->nc_dvp == dvp); 2409 blps[1] = NCP2BUCKETLOCK(ncp); 2410 if ((nc_flag & NCF_NEGATIVE) != 0) 2411 break; 2412 if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp)) 2413 break; 2414 if (ncp == dvp->v_cache_dd && 2415 (ncp->nc_flag & NCF_ISDOTDOT) != 0 && 2416 blps[1] == NCP2BUCKETLOCK(ncp) && 2417 VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2]) 2418 break; 2419 cache_unlock_vnodes_cel(cel); 2420 cel->vlp[0] = NULL; 2421 cel->vlp[1] = NULL; 2422 cel->vlp[2] = NULL; 2423 } 2424 cache_lock_buckets_cel(cel, blps[0], blps[1]); 2425 } 2426 2427 static void 2428 cache_enter_unlock(struct celockstate *cel) 2429 { 2430 2431 cache_unlock_buckets_cel(cel); 2432 cache_unlock_vnodes_cel(cel); 2433 } 2434 2435 static void __noinline 2436 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp, 2437 struct componentname *cnp) 2438 { 2439 struct celockstate cel; 2440 struct namecache *ncp; 2441 uint32_t hash; 2442 int len; 2443 2444 if (atomic_load_ptr(&dvp->v_cache_dd) == NULL) 2445 return; 2446 len = cnp->cn_namelen; 2447 cache_celockstate_init(&cel); 2448 hash = cache_get_hash(cnp->cn_nameptr, len, dvp); 2449 cache_enter_lock_dd(&cel, dvp, vp, hash); 2450 ncp = dvp->v_cache_dd; 2451 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) { 2452 KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent")); 2453 cache_zap_locked(ncp); 2454 } else { 2455 ncp = NULL; 2456 } 2457 atomic_store_ptr(&dvp->v_cache_dd, NULL); 2458 cache_enter_unlock(&cel); 2459 if (ncp != NULL) 2460 cache_free(ncp); 2461 } 2462 2463 /* 2464 * Add an entry to the cache. 2465 */ 2466 void 2467 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, 2468 struct timespec *tsp, struct timespec *dtsp) 2469 { 2470 struct celockstate cel; 2471 struct namecache *ncp, *n2, *ndd; 2472 struct namecache_ts *ncp_ts; 2473 struct nchashhead *ncpp; 2474 uint32_t hash; 2475 int flag; 2476 int len; 2477 2478 KASSERT(cnp->cn_namelen <= NAME_MAX, 2479 ("%s: passed len %ld exceeds NAME_MAX (%d)", __func__, cnp->cn_namelen, 2480 NAME_MAX)); 2481 VNPASS(!VN_IS_DOOMED(dvp), dvp); 2482 VNPASS(dvp->v_type != VNON, dvp); 2483 if (vp != NULL) { 2484 VNPASS(!VN_IS_DOOMED(vp), vp); 2485 VNPASS(vp->v_type != VNON, vp); 2486 } 2487 if (cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.') { 2488 KASSERT(dvp == vp, 2489 ("%s: different vnodes for dot entry (%p; %p)\n", __func__, 2490 dvp, vp)); 2491 } else { 2492 KASSERT(dvp != vp, 2493 ("%s: same vnode for non-dot entry [%s] (%p)\n", __func__, 2494 cnp->cn_nameptr, dvp)); 2495 } 2496 2497 #ifdef DEBUG_CACHE 2498 if (__predict_false(!doingcache)) 2499 return; 2500 #endif 2501 2502 flag = 0; 2503 if (__predict_false(cnp->cn_nameptr[0] == '.')) { 2504 if (cnp->cn_namelen == 1) 2505 return; 2506 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') { 2507 cache_enter_dotdot_prep(dvp, vp, cnp); 2508 flag = NCF_ISDOTDOT; 2509 } 2510 } 2511 2512 ncp = cache_alloc(cnp->cn_namelen, tsp != NULL); 2513 if (ncp == NULL) 2514 return; 2515 2516 cache_celockstate_init(&cel); 2517 ndd = NULL; 2518 ncp_ts = NULL; 2519 2520 /* 2521 * Calculate the hash key and setup as much of the new 2522 * namecache entry as possible before acquiring the lock. 2523 */ 2524 ncp->nc_flag = flag | NCF_WIP; 2525 ncp->nc_vp = vp; 2526 if (vp == NULL) 2527 cache_neg_init(ncp); 2528 ncp->nc_dvp = dvp; 2529 if (tsp != NULL) { 2530 ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc); 2531 ncp_ts->nc_time = *tsp; 2532 ncp_ts->nc_ticks = ticks; 2533 ncp_ts->nc_nc.nc_flag |= NCF_TS; 2534 if (dtsp != NULL) { 2535 ncp_ts->nc_dotdottime = *dtsp; 2536 ncp_ts->nc_nc.nc_flag |= NCF_DTS; 2537 } 2538 } 2539 len = ncp->nc_nlen = cnp->cn_namelen; 2540 hash = cache_get_hash(cnp->cn_nameptr, len, dvp); 2541 memcpy(ncp->nc_name, cnp->cn_nameptr, len); 2542 ncp->nc_name[len] = '\0'; 2543 cache_enter_lock(&cel, dvp, vp, hash); 2544 2545 /* 2546 * See if this vnode or negative entry is already in the cache 2547 * with this name. This can happen with concurrent lookups of 2548 * the same path name. 2549 */ 2550 ncpp = NCHHASH(hash); 2551 CK_SLIST_FOREACH(n2, ncpp, nc_hash) { 2552 if (n2->nc_dvp == dvp && 2553 n2->nc_nlen == cnp->cn_namelen && 2554 !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) { 2555 MPASS(cache_ncp_canuse(n2)); 2556 if ((n2->nc_flag & NCF_NEGATIVE) != 0) 2557 KASSERT(vp == NULL, 2558 ("%s: found entry pointing to a different vnode (%p != %p) ; name [%s]", 2559 __func__, NULL, vp, cnp->cn_nameptr)); 2560 else 2561 KASSERT(n2->nc_vp == vp, 2562 ("%s: found entry pointing to a different vnode (%p != %p) ; name [%s]", 2563 __func__, n2->nc_vp, vp, cnp->cn_nameptr)); 2564 /* 2565 * Entries are supposed to be immutable unless in the 2566 * process of getting destroyed. Accommodating for 2567 * changing timestamps is possible but not worth it. 2568 * This should be harmless in terms of correctness, in 2569 * the worst case resulting in an earlier expiration. 2570 * Alternatively, the found entry can be replaced 2571 * altogether. 2572 */ 2573 MPASS((n2->nc_flag & (NCF_TS | NCF_DTS)) == (ncp->nc_flag & (NCF_TS | NCF_DTS))); 2574 #if 0 2575 if (tsp != NULL) { 2576 KASSERT((n2->nc_flag & NCF_TS) != 0, 2577 ("no NCF_TS")); 2578 n2_ts = __containerof(n2, struct namecache_ts, nc_nc); 2579 n2_ts->nc_time = ncp_ts->nc_time; 2580 n2_ts->nc_ticks = ncp_ts->nc_ticks; 2581 if (dtsp != NULL) { 2582 n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime; 2583 n2_ts->nc_nc.nc_flag |= NCF_DTS; 2584 } 2585 } 2586 #endif 2587 SDT_PROBE3(vfs, namecache, enter, duplicate, dvp, ncp->nc_name, 2588 vp); 2589 goto out_unlock_free; 2590 } 2591 } 2592 2593 if (flag == NCF_ISDOTDOT) { 2594 /* 2595 * See if we are trying to add .. entry, but some other lookup 2596 * has populated v_cache_dd pointer already. 2597 */ 2598 if (dvp->v_cache_dd != NULL) 2599 goto out_unlock_free; 2600 KASSERT(vp == NULL || vp->v_type == VDIR, 2601 ("wrong vnode type %p", vp)); 2602 atomic_thread_fence_rel(); 2603 atomic_store_ptr(&dvp->v_cache_dd, ncp); 2604 } 2605 2606 if (vp != NULL) { 2607 if (flag != NCF_ISDOTDOT) { 2608 /* 2609 * For this case, the cache entry maps both the 2610 * directory name in it and the name ".." for the 2611 * directory's parent. 2612 */ 2613 if ((ndd = vp->v_cache_dd) != NULL) { 2614 if ((ndd->nc_flag & NCF_ISDOTDOT) != 0) 2615 cache_zap_locked(ndd); 2616 else 2617 ndd = NULL; 2618 } 2619 atomic_thread_fence_rel(); 2620 atomic_store_ptr(&vp->v_cache_dd, ncp); 2621 } else if (vp->v_type != VDIR) { 2622 if (vp->v_cache_dd != NULL) { 2623 atomic_store_ptr(&vp->v_cache_dd, NULL); 2624 } 2625 } 2626 } 2627 2628 if (flag != NCF_ISDOTDOT) { 2629 if (LIST_EMPTY(&dvp->v_cache_src)) { 2630 cache_hold_vnode(dvp); 2631 } 2632 LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src); 2633 } 2634 2635 /* 2636 * If the entry is "negative", we place it into the 2637 * "negative" cache queue, otherwise, we place it into the 2638 * destination vnode's cache entries queue. 2639 */ 2640 if (vp != NULL) { 2641 TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst); 2642 SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name, 2643 vp); 2644 } else { 2645 if (cnp->cn_flags & ISWHITEOUT) 2646 atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_WHITE); 2647 cache_neg_insert(ncp); 2648 SDT_PROBE2(vfs, namecache, enter_negative, done, dvp, 2649 ncp->nc_name); 2650 } 2651 2652 /* 2653 * Insert the new namecache entry into the appropriate chain 2654 * within the cache entries table. 2655 */ 2656 CK_SLIST_INSERT_HEAD(ncpp, ncp, nc_hash); 2657 2658 atomic_thread_fence_rel(); 2659 /* 2660 * Mark the entry as fully constructed. 2661 * It is immutable past this point until its removal. 2662 */ 2663 atomic_store_char(&ncp->nc_flag, ncp->nc_flag & ~NCF_WIP); 2664 2665 cache_enter_unlock(&cel); 2666 if (ndd != NULL) 2667 cache_free(ndd); 2668 return; 2669 out_unlock_free: 2670 cache_enter_unlock(&cel); 2671 cache_free(ncp); 2672 return; 2673 } 2674 2675 /* 2676 * A variant of the above accepting flags. 2677 * 2678 * - VFS_CACHE_DROPOLD -- if a conflicting entry is found, drop it. 2679 * 2680 * TODO: this routine is a hack. It blindly removes the old entry, even if it 2681 * happens to match and it is doing it in an inefficient manner. It was added 2682 * to accommodate NFS which runs into a case where the target for a given name 2683 * may change from under it. Note this does nothing to solve the following 2684 * race: 2 callers of cache_enter_time_flags pass a different target vnode for 2685 * the same [dvp, cnp]. It may be argued that code doing this is broken. 2686 */ 2687 void 2688 cache_enter_time_flags(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, 2689 struct timespec *tsp, struct timespec *dtsp, int flags) 2690 { 2691 2692 MPASS((flags & ~(VFS_CACHE_DROPOLD)) == 0); 2693 2694 if (flags & VFS_CACHE_DROPOLD) 2695 cache_remove_cnp(dvp, cnp); 2696 cache_enter_time(dvp, vp, cnp, tsp, dtsp); 2697 } 2698 2699 static u_long 2700 cache_roundup_2(u_long val) 2701 { 2702 u_long res; 2703 2704 for (res = 1; res <= val; res <<= 1) 2705 continue; 2706 2707 return (res); 2708 } 2709 2710 static struct nchashhead * 2711 nchinittbl(u_long elements, u_long *hashmask) 2712 { 2713 struct nchashhead *hashtbl; 2714 u_long hashsize, i; 2715 2716 hashsize = cache_roundup_2(elements) / 2; 2717 2718 hashtbl = malloc(hashsize * sizeof(*hashtbl), M_VFSCACHE, M_WAITOK); 2719 for (i = 0; i < hashsize; i++) 2720 CK_SLIST_INIT(&hashtbl[i]); 2721 *hashmask = hashsize - 1; 2722 return (hashtbl); 2723 } 2724 2725 static void 2726 ncfreetbl(struct nchashhead *hashtbl) 2727 { 2728 2729 free(hashtbl, M_VFSCACHE); 2730 } 2731 2732 /* 2733 * Name cache initialization, from vfs_init() when we are booting 2734 */ 2735 static void 2736 nchinit(void *dummy __unused) 2737 { 2738 u_int i; 2739 2740 cache_zone_small = uma_zcreate("S VFS Cache", CACHE_ZONE_SMALL_SIZE, 2741 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT); 2742 cache_zone_small_ts = uma_zcreate("STS VFS Cache", CACHE_ZONE_SMALL_TS_SIZE, 2743 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT); 2744 cache_zone_large = uma_zcreate("L VFS Cache", CACHE_ZONE_LARGE_SIZE, 2745 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT); 2746 cache_zone_large_ts = uma_zcreate("LTS VFS Cache", CACHE_ZONE_LARGE_TS_SIZE, 2747 NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT); 2748 2749 VFS_SMR_ZONE_SET(cache_zone_small); 2750 VFS_SMR_ZONE_SET(cache_zone_small_ts); 2751 VFS_SMR_ZONE_SET(cache_zone_large); 2752 VFS_SMR_ZONE_SET(cache_zone_large_ts); 2753 2754 ncsize = desiredvnodes * ncsizefactor; 2755 cache_recalc_neg_min(); 2756 nchashtbl = nchinittbl(desiredvnodes * 2, &nchash); 2757 ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1; 2758 if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */ 2759 ncbuckethash = 7; 2760 if (ncbuckethash > nchash) 2761 ncbuckethash = nchash; 2762 bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE, 2763 M_WAITOK | M_ZERO); 2764 for (i = 0; i < numbucketlocks; i++) 2765 mtx_init(&bucketlocks[i], "ncbuc", NULL, MTX_DUPOK | MTX_RECURSE); 2766 ncvnodehash = ncbuckethash; 2767 vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE, 2768 M_WAITOK | M_ZERO); 2769 for (i = 0; i < numvnodelocks; i++) 2770 mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE); 2771 2772 for (i = 0; i < numneglists; i++) { 2773 mtx_init(&neglists[i].nl_evict_lock, "ncnege", NULL, MTX_DEF); 2774 mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF); 2775 TAILQ_INIT(&neglists[i].nl_list); 2776 TAILQ_INIT(&neglists[i].nl_hotlist); 2777 } 2778 } 2779 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL); 2780 2781 void 2782 cache_vnode_init(struct vnode *vp) 2783 { 2784 2785 LIST_INIT(&vp->v_cache_src); 2786 TAILQ_INIT(&vp->v_cache_dst); 2787 vp->v_cache_dd = NULL; 2788 cache_prehash(vp); 2789 } 2790 2791 /* 2792 * Induce transient cache misses for lockless operation in cache_lookup() by 2793 * using a temporary hash table. 2794 * 2795 * This will force a fs lookup. 2796 * 2797 * Synchronisation is done in 2 steps, calling vfs_smr_synchronize each time 2798 * to observe all CPUs not performing the lookup. 2799 */ 2800 static void 2801 cache_changesize_set_temp(struct nchashhead *temptbl, u_long temphash) 2802 { 2803 2804 MPASS(temphash < nchash); 2805 /* 2806 * Change the size. The new size is smaller and can safely be used 2807 * against the existing table. All lookups which now hash wrong will 2808 * result in a cache miss, which all callers are supposed to know how 2809 * to handle. 2810 */ 2811 atomic_store_long(&nchash, temphash); 2812 atomic_thread_fence_rel(); 2813 vfs_smr_synchronize(); 2814 /* 2815 * At this point everyone sees the updated hash value, but they still 2816 * see the old table. 2817 */ 2818 atomic_store_ptr(&nchashtbl, temptbl); 2819 atomic_thread_fence_rel(); 2820 vfs_smr_synchronize(); 2821 /* 2822 * At this point everyone sees the updated table pointer and size pair. 2823 */ 2824 } 2825 2826 /* 2827 * Set the new hash table. 2828 * 2829 * Similarly to cache_changesize_set_temp(), this has to synchronize against 2830 * lockless operation in cache_lookup(). 2831 */ 2832 static void 2833 cache_changesize_set_new(struct nchashhead *new_tbl, u_long new_hash) 2834 { 2835 2836 MPASS(nchash < new_hash); 2837 /* 2838 * Change the pointer first. This wont result in out of bounds access 2839 * since the temporary table is guaranteed to be smaller. 2840 */ 2841 atomic_store_ptr(&nchashtbl, new_tbl); 2842 atomic_thread_fence_rel(); 2843 vfs_smr_synchronize(); 2844 /* 2845 * At this point everyone sees the updated pointer value, but they 2846 * still see the old size. 2847 */ 2848 atomic_store_long(&nchash, new_hash); 2849 atomic_thread_fence_rel(); 2850 vfs_smr_synchronize(); 2851 /* 2852 * At this point everyone sees the updated table pointer and size pair. 2853 */ 2854 } 2855 2856 void 2857 cache_changesize(u_long newmaxvnodes) 2858 { 2859 struct nchashhead *new_nchashtbl, *old_nchashtbl, *temptbl; 2860 u_long new_nchash, old_nchash, temphash; 2861 struct namecache *ncp; 2862 uint32_t hash; 2863 u_long newncsize; 2864 u_long i; 2865 2866 newncsize = newmaxvnodes * ncsizefactor; 2867 newmaxvnodes = cache_roundup_2(newmaxvnodes * 2); 2868 if (newmaxvnodes < numbucketlocks) 2869 newmaxvnodes = numbucketlocks; 2870 2871 new_nchashtbl = nchinittbl(newmaxvnodes, &new_nchash); 2872 /* If same hash table size, nothing to do */ 2873 if (nchash == new_nchash) { 2874 ncfreetbl(new_nchashtbl); 2875 return; 2876 } 2877 2878 temptbl = nchinittbl(1, &temphash); 2879 2880 /* 2881 * Move everything from the old hash table to the new table. 2882 * None of the namecache entries in the table can be removed 2883 * because to do so, they have to be removed from the hash table. 2884 */ 2885 cache_lock_all_vnodes(); 2886 cache_lock_all_buckets(); 2887 old_nchashtbl = nchashtbl; 2888 old_nchash = nchash; 2889 cache_changesize_set_temp(temptbl, temphash); 2890 for (i = 0; i <= old_nchash; i++) { 2891 while ((ncp = CK_SLIST_FIRST(&old_nchashtbl[i])) != NULL) { 2892 hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, 2893 ncp->nc_dvp); 2894 CK_SLIST_REMOVE(&old_nchashtbl[i], ncp, namecache, nc_hash); 2895 CK_SLIST_INSERT_HEAD(&new_nchashtbl[hash & new_nchash], ncp, nc_hash); 2896 } 2897 } 2898 ncsize = newncsize; 2899 cache_recalc_neg_min(); 2900 cache_changesize_set_new(new_nchashtbl, new_nchash); 2901 cache_unlock_all_buckets(); 2902 cache_unlock_all_vnodes(); 2903 ncfreetbl(old_nchashtbl); 2904 ncfreetbl(temptbl); 2905 } 2906 2907 /* 2908 * Remove all entries from and to a particular vnode. 2909 */ 2910 static void 2911 cache_purge_impl(struct vnode *vp) 2912 { 2913 struct cache_freebatch batch; 2914 struct namecache *ncp; 2915 struct mtx *vlp, *vlp2; 2916 2917 TAILQ_INIT(&batch); 2918 vlp = VP2VNODELOCK(vp); 2919 vlp2 = NULL; 2920 mtx_lock(vlp); 2921 retry: 2922 while (!LIST_EMPTY(&vp->v_cache_src)) { 2923 ncp = LIST_FIRST(&vp->v_cache_src); 2924 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2)) 2925 goto retry; 2926 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst); 2927 } 2928 while (!TAILQ_EMPTY(&vp->v_cache_dst)) { 2929 ncp = TAILQ_FIRST(&vp->v_cache_dst); 2930 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2)) 2931 goto retry; 2932 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst); 2933 } 2934 ncp = vp->v_cache_dd; 2935 if (ncp != NULL) { 2936 KASSERT(ncp->nc_flag & NCF_ISDOTDOT, 2937 ("lost dotdot link")); 2938 if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2)) 2939 goto retry; 2940 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst); 2941 } 2942 KASSERT(vp->v_cache_dd == NULL, ("incomplete purge")); 2943 mtx_unlock(vlp); 2944 if (vlp2 != NULL) 2945 mtx_unlock(vlp2); 2946 cache_free_batch(&batch); 2947 } 2948 2949 /* 2950 * Opportunistic check to see if there is anything to do. 2951 */ 2952 static bool 2953 cache_has_entries(struct vnode *vp) 2954 { 2955 2956 if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) && 2957 atomic_load_ptr(&vp->v_cache_dd) == NULL) 2958 return (false); 2959 return (true); 2960 } 2961 2962 void 2963 cache_purge(struct vnode *vp) 2964 { 2965 2966 SDT_PROBE1(vfs, namecache, purge, done, vp); 2967 if (!cache_has_entries(vp)) 2968 return; 2969 cache_purge_impl(vp); 2970 } 2971 2972 /* 2973 * Only to be used by vgone. 2974 */ 2975 void 2976 cache_purge_vgone(struct vnode *vp) 2977 { 2978 struct mtx *vlp; 2979 2980 VNPASS(VN_IS_DOOMED(vp), vp); 2981 if (cache_has_entries(vp)) { 2982 cache_purge_impl(vp); 2983 return; 2984 } 2985 2986 /* 2987 * Serialize against a potential thread doing cache_purge. 2988 */ 2989 vlp = VP2VNODELOCK(vp); 2990 mtx_wait_unlocked(vlp); 2991 if (cache_has_entries(vp)) { 2992 cache_purge_impl(vp); 2993 return; 2994 } 2995 return; 2996 } 2997 2998 /* 2999 * Remove all negative entries for a particular directory vnode. 3000 */ 3001 void 3002 cache_purge_negative(struct vnode *vp) 3003 { 3004 struct cache_freebatch batch; 3005 struct namecache *ncp, *nnp; 3006 struct mtx *vlp; 3007 3008 SDT_PROBE1(vfs, namecache, purge_negative, done, vp); 3009 if (LIST_EMPTY(&vp->v_cache_src)) 3010 return; 3011 TAILQ_INIT(&batch); 3012 vlp = VP2VNODELOCK(vp); 3013 mtx_lock(vlp); 3014 LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) { 3015 if (!(ncp->nc_flag & NCF_NEGATIVE)) 3016 continue; 3017 cache_zap_negative_locked_vnode_kl(ncp, vp); 3018 TAILQ_INSERT_TAIL(&batch, ncp, nc_dst); 3019 } 3020 mtx_unlock(vlp); 3021 cache_free_batch(&batch); 3022 } 3023 3024 /* 3025 * Entry points for modifying VOP operations. 3026 */ 3027 void 3028 cache_vop_rename(struct vnode *fdvp, struct vnode *fvp, struct vnode *tdvp, 3029 struct vnode *tvp, struct componentname *fcnp, struct componentname *tcnp) 3030 { 3031 3032 ASSERT_VOP_IN_SEQC(fdvp); 3033 ASSERT_VOP_IN_SEQC(fvp); 3034 ASSERT_VOP_IN_SEQC(tdvp); 3035 if (tvp != NULL) 3036 ASSERT_VOP_IN_SEQC(tvp); 3037 3038 cache_purge(fvp); 3039 if (tvp != NULL) { 3040 cache_purge(tvp); 3041 KASSERT(!cache_remove_cnp(tdvp, tcnp), 3042 ("%s: lingering negative entry", __func__)); 3043 } else { 3044 cache_remove_cnp(tdvp, tcnp); 3045 } 3046 3047 /* 3048 * TODO 3049 * 3050 * Historically renaming was always purging all revelang entries, 3051 * but that's quite wasteful. In particular turns out that in many cases 3052 * the target file is immediately accessed after rename, inducing a cache 3053 * miss. 3054 * 3055 * Recode this to reduce relocking and reuse the existing entry (if any) 3056 * instead of just removing it above and allocating a new one here. 3057 */ 3058 cache_enter(tdvp, fvp, tcnp); 3059 } 3060 3061 void 3062 cache_vop_rmdir(struct vnode *dvp, struct vnode *vp) 3063 { 3064 3065 ASSERT_VOP_IN_SEQC(dvp); 3066 ASSERT_VOP_IN_SEQC(vp); 3067 cache_purge(vp); 3068 } 3069 3070 #ifdef INVARIANTS 3071 /* 3072 * Validate that if an entry exists it matches. 3073 */ 3074 void 3075 cache_validate(struct vnode *dvp, struct vnode *vp, struct componentname *cnp) 3076 { 3077 struct namecache *ncp; 3078 struct mtx *blp; 3079 uint32_t hash; 3080 3081 hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); 3082 if (CK_SLIST_EMPTY(NCHHASH(hash))) 3083 return; 3084 blp = HASH2BUCKETLOCK(hash); 3085 mtx_lock(blp); 3086 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 3087 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && 3088 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) { 3089 if (ncp->nc_vp != vp) 3090 panic("%s: mismatch (%p != %p); ncp %p [%s] dvp %p\n", 3091 __func__, vp, ncp->nc_vp, ncp, ncp->nc_name, ncp->nc_dvp); 3092 } 3093 } 3094 mtx_unlock(blp); 3095 } 3096 3097 void 3098 cache_assert_no_entries(struct vnode *vp) 3099 { 3100 3101 VNPASS(TAILQ_EMPTY(&vp->v_cache_dst), vp); 3102 VNPASS(LIST_EMPTY(&vp->v_cache_src), vp); 3103 VNPASS(vp->v_cache_dd == NULL, vp); 3104 } 3105 #endif 3106 3107 /* 3108 * Flush all entries referencing a particular filesystem. 3109 */ 3110 void 3111 cache_purgevfs(struct mount *mp) 3112 { 3113 struct vnode *vp, *mvp; 3114 size_t visited __sdt_used, purged __sdt_used; 3115 3116 visited = purged = 0; 3117 /* 3118 * Somewhat wasteful iteration over all vnodes. Would be better to 3119 * support filtering and avoid the interlock to begin with. 3120 */ 3121 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { 3122 visited++; 3123 if (!cache_has_entries(vp)) { 3124 VI_UNLOCK(vp); 3125 continue; 3126 } 3127 vholdl(vp); 3128 VI_UNLOCK(vp); 3129 cache_purge(vp); 3130 purged++; 3131 vdrop(vp); 3132 } 3133 3134 SDT_PROBE3(vfs, namecache, purgevfs, done, mp, visited, purged); 3135 } 3136 3137 /* 3138 * Perform canonical checks and cache lookup and pass on to filesystem 3139 * through the vop_cachedlookup only if needed. 3140 */ 3141 3142 int 3143 vfs_cache_lookup(struct vop_lookup_args *ap) 3144 { 3145 struct vnode *dvp; 3146 int error; 3147 struct vnode **vpp = ap->a_vpp; 3148 struct componentname *cnp = ap->a_cnp; 3149 int flags = cnp->cn_flags; 3150 3151 *vpp = NULL; 3152 dvp = ap->a_dvp; 3153 3154 if (dvp->v_type != VDIR) 3155 return (ENOTDIR); 3156 3157 if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) && 3158 (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) 3159 return (EROFS); 3160 3161 error = vn_dir_check_exec(dvp, cnp); 3162 if (error != 0) 3163 return (error); 3164 3165 error = cache_lookup(dvp, vpp, cnp, NULL, NULL); 3166 if (error == 0) 3167 return (VOP_CACHEDLOOKUP(dvp, vpp, cnp)); 3168 if (error == -1) 3169 return (0); 3170 return (error); 3171 } 3172 3173 /* Implementation of the getcwd syscall. */ 3174 int 3175 sys___getcwd(struct thread *td, struct __getcwd_args *uap) 3176 { 3177 char *buf, *retbuf; 3178 size_t buflen; 3179 int error; 3180 3181 buflen = uap->buflen; 3182 if (__predict_false(buflen < 2)) 3183 return (EINVAL); 3184 if (buflen > MAXPATHLEN) 3185 buflen = MAXPATHLEN; 3186 3187 buf = uma_zalloc(namei_zone, M_WAITOK); 3188 error = vn_getcwd(buf, &retbuf, &buflen); 3189 if (error == 0) 3190 error = copyout(retbuf, uap->buf, buflen); 3191 uma_zfree(namei_zone, buf); 3192 return (error); 3193 } 3194 3195 int 3196 vn_getcwd(char *buf, char **retbuf, size_t *buflen) 3197 { 3198 struct pwd *pwd; 3199 int error; 3200 3201 vfs_smr_enter(); 3202 pwd = pwd_get_smr(); 3203 error = vn_fullpath_any_smr(pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf, 3204 buflen, 0); 3205 VFS_SMR_ASSERT_NOT_ENTERED(); 3206 if (error < 0) { 3207 pwd = pwd_hold(curthread); 3208 error = vn_fullpath_any(pwd->pwd_cdir, pwd->pwd_rdir, buf, 3209 retbuf, buflen); 3210 pwd_drop(pwd); 3211 } 3212 3213 #ifdef KTRACE 3214 if (KTRPOINT(curthread, KTR_NAMEI) && error == 0) 3215 ktrnamei(*retbuf); 3216 #endif 3217 return (error); 3218 } 3219 3220 /* 3221 * Canonicalize a path by walking it forward and back. 3222 * 3223 * BUGS: 3224 * - Nothing guarantees the integrity of the entire chain. Consider the case 3225 * where the path "foo/bar/baz/qux" is passed, but "bar" is moved out of 3226 * "foo" into "quux" during the backwards walk. The result will be 3227 * "quux/bar/baz/qux", which could not have been obtained by an incremental 3228 * walk in userspace. Moreover, the path we return is inaccessible if the 3229 * calling thread lacks permission to traverse "quux". 3230 */ 3231 static int 3232 kern___realpathat(struct thread *td, int fd, const char *path, char *buf, 3233 size_t size, int flags, enum uio_seg pathseg) 3234 { 3235 struct nameidata nd; 3236 char *retbuf, *freebuf; 3237 int error; 3238 3239 if (flags != 0) 3240 return (EINVAL); 3241 NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | WANTPARENT | AUDITVNODE1, 3242 pathseg, path, fd, &cap_fstat_rights); 3243 if ((error = namei(&nd)) != 0) 3244 return (error); 3245 3246 if (nd.ni_vp->v_type == VREG && nd.ni_dvp->v_type != VDIR && 3247 (nd.ni_vp->v_vflag & VV_ROOT) != 0) { 3248 /* 3249 * This happens if vp is a file mount. The call to 3250 * vn_fullpath_hardlink can panic if path resolution can't be 3251 * handled without the directory. 3252 * 3253 * To resolve this, we find the vnode which was mounted on - 3254 * this should have a unique global path since we disallow 3255 * mounting on linked files. 3256 */ 3257 struct vnode *covered_vp; 3258 error = vn_lock(nd.ni_vp, LK_SHARED); 3259 if (error != 0) 3260 goto out; 3261 covered_vp = nd.ni_vp->v_mount->mnt_vnodecovered; 3262 vref(covered_vp); 3263 VOP_UNLOCK(nd.ni_vp); 3264 error = vn_fullpath(covered_vp, &retbuf, &freebuf); 3265 vrele(covered_vp); 3266 } else { 3267 error = vn_fullpath_hardlink(nd.ni_vp, nd.ni_dvp, nd.ni_cnd.cn_nameptr, 3268 nd.ni_cnd.cn_namelen, &retbuf, &freebuf, &size); 3269 } 3270 if (error == 0) { 3271 error = copyout(retbuf, buf, size); 3272 free(freebuf, M_TEMP); 3273 } 3274 out: 3275 vrele(nd.ni_vp); 3276 vrele(nd.ni_dvp); 3277 NDFREE_PNBUF(&nd); 3278 return (error); 3279 } 3280 3281 int 3282 sys___realpathat(struct thread *td, struct __realpathat_args *uap) 3283 { 3284 3285 return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size, 3286 uap->flags, UIO_USERSPACE)); 3287 } 3288 3289 /* 3290 * Retrieve the full filesystem path that correspond to a vnode from the name 3291 * cache (if available) 3292 */ 3293 int 3294 vn_fullpath(struct vnode *vp, char **retbuf, char **freebuf) 3295 { 3296 struct pwd *pwd; 3297 char *buf; 3298 size_t buflen; 3299 int error; 3300 3301 if (__predict_false(vp == NULL)) 3302 return (EINVAL); 3303 3304 buflen = MAXPATHLEN; 3305 buf = malloc(buflen, M_TEMP, M_WAITOK); 3306 vfs_smr_enter(); 3307 pwd = pwd_get_smr(); 3308 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, &buflen, 0); 3309 VFS_SMR_ASSERT_NOT_ENTERED(); 3310 if (error < 0) { 3311 pwd = pwd_hold(curthread); 3312 error = vn_fullpath_any(vp, pwd->pwd_rdir, buf, retbuf, &buflen); 3313 pwd_drop(pwd); 3314 } 3315 if (error == 0) 3316 *freebuf = buf; 3317 else 3318 free(buf, M_TEMP); 3319 return (error); 3320 } 3321 3322 /* 3323 * This function is similar to vn_fullpath, but it attempts to lookup the 3324 * pathname relative to the global root mount point. This is required for the 3325 * auditing sub-system, as audited pathnames must be absolute, relative to the 3326 * global root mount point. 3327 */ 3328 int 3329 vn_fullpath_global(struct vnode *vp, char **retbuf, char **freebuf) 3330 { 3331 char *buf; 3332 size_t buflen; 3333 int error; 3334 3335 if (__predict_false(vp == NULL)) 3336 return (EINVAL); 3337 buflen = MAXPATHLEN; 3338 buf = malloc(buflen, M_TEMP, M_WAITOK); 3339 vfs_smr_enter(); 3340 error = vn_fullpath_any_smr(vp, rootvnode, buf, retbuf, &buflen, 0); 3341 VFS_SMR_ASSERT_NOT_ENTERED(); 3342 if (error < 0) { 3343 error = vn_fullpath_any(vp, rootvnode, buf, retbuf, &buflen); 3344 } 3345 if (error == 0) 3346 *freebuf = buf; 3347 else 3348 free(buf, M_TEMP); 3349 return (error); 3350 } 3351 3352 static struct namecache * 3353 vn_dd_from_dst(struct vnode *vp) 3354 { 3355 struct namecache *ncp; 3356 3357 cache_assert_vnode_locked(vp); 3358 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) { 3359 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) 3360 return (ncp); 3361 } 3362 return (NULL); 3363 } 3364 3365 int 3366 vn_vptocnp(struct vnode **vp, char *buf, size_t *buflen) 3367 { 3368 struct vnode *dvp; 3369 struct namecache *ncp; 3370 struct mtx *vlp; 3371 int error; 3372 3373 vlp = VP2VNODELOCK(*vp); 3374 mtx_lock(vlp); 3375 ncp = (*vp)->v_cache_dd; 3376 if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT) == 0) { 3377 KASSERT(ncp == vn_dd_from_dst(*vp), 3378 ("%s: mismatch for dd entry (%p != %p)", __func__, 3379 ncp, vn_dd_from_dst(*vp))); 3380 } else { 3381 ncp = vn_dd_from_dst(*vp); 3382 } 3383 if (ncp != NULL) { 3384 if (*buflen < ncp->nc_nlen) { 3385 mtx_unlock(vlp); 3386 vrele(*vp); 3387 counter_u64_add(numfullpathfail4, 1); 3388 error = ENOMEM; 3389 SDT_PROBE3(vfs, namecache, fullpath, return, error, 3390 vp, NULL); 3391 return (error); 3392 } 3393 *buflen -= ncp->nc_nlen; 3394 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen); 3395 SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp, 3396 ncp->nc_name, vp); 3397 dvp = *vp; 3398 *vp = ncp->nc_dvp; 3399 vref(*vp); 3400 mtx_unlock(vlp); 3401 vrele(dvp); 3402 return (0); 3403 } 3404 SDT_PROBE1(vfs, namecache, fullpath, miss, vp); 3405 3406 mtx_unlock(vlp); 3407 vn_lock(*vp, LK_SHARED | LK_RETRY); 3408 error = VOP_VPTOCNP(*vp, &dvp, buf, buflen); 3409 vput(*vp); 3410 if (error) { 3411 counter_u64_add(numfullpathfail2, 1); 3412 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); 3413 return (error); 3414 } 3415 3416 *vp = dvp; 3417 if (VN_IS_DOOMED(dvp)) { 3418 /* forced unmount */ 3419 vrele(dvp); 3420 error = ENOENT; 3421 SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); 3422 return (error); 3423 } 3424 /* 3425 * *vp has its use count incremented still. 3426 */ 3427 3428 return (0); 3429 } 3430 3431 /* 3432 * Resolve a directory to a pathname. 3433 * 3434 * The name of the directory can always be found in the namecache or fetched 3435 * from the filesystem. There is also guaranteed to be only one parent, meaning 3436 * we can just follow vnodes up until we find the root. 3437 * 3438 * The vnode must be referenced. 3439 */ 3440 static int 3441 vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf, 3442 size_t *len, size_t addend) 3443 { 3444 #ifdef KDTRACE_HOOKS 3445 struct vnode *startvp = vp; 3446 #endif 3447 struct vnode *vp1; 3448 size_t buflen; 3449 int error; 3450 bool slash_prefixed; 3451 3452 VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp); 3453 VNPASS(vp->v_usecount > 0, vp); 3454 3455 buflen = *len; 3456 3457 slash_prefixed = true; 3458 if (addend == 0) { 3459 MPASS(*len >= 2); 3460 buflen--; 3461 buf[buflen] = '\0'; 3462 slash_prefixed = false; 3463 } 3464 3465 error = 0; 3466 3467 SDT_PROBE1(vfs, namecache, fullpath, entry, vp); 3468 counter_u64_add(numfullpathcalls, 1); 3469 while (vp != rdir && vp != rootvnode) { 3470 /* 3471 * The vp vnode must be already fully constructed, 3472 * since it is either found in namecache or obtained 3473 * from VOP_VPTOCNP(). We may test for VV_ROOT safely 3474 * without obtaining the vnode lock. 3475 */ 3476 if ((vp->v_vflag & VV_ROOT) != 0) { 3477 vn_lock(vp, LK_RETRY | LK_SHARED); 3478 3479 /* 3480 * With the vnode locked, check for races with 3481 * unmount, forced or not. Note that we 3482 * already verified that vp is not equal to 3483 * the root vnode, which means that 3484 * mnt_vnodecovered can be NULL only for the 3485 * case of unmount. 3486 */ 3487 if (VN_IS_DOOMED(vp) || 3488 (vp1 = vp->v_mount->mnt_vnodecovered) == NULL || 3489 vp1->v_mountedhere != vp->v_mount) { 3490 vput(vp); 3491 error = ENOENT; 3492 SDT_PROBE3(vfs, namecache, fullpath, return, 3493 error, vp, NULL); 3494 break; 3495 } 3496 3497 vref(vp1); 3498 vput(vp); 3499 vp = vp1; 3500 continue; 3501 } 3502 VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp); 3503 error = vn_vptocnp(&vp, buf, &buflen); 3504 if (error) 3505 break; 3506 if (buflen == 0) { 3507 vrele(vp); 3508 error = ENOMEM; 3509 SDT_PROBE3(vfs, namecache, fullpath, return, error, 3510 startvp, NULL); 3511 break; 3512 } 3513 buf[--buflen] = '/'; 3514 slash_prefixed = true; 3515 } 3516 if (error) 3517 return (error); 3518 if (!slash_prefixed) { 3519 if (buflen == 0) { 3520 vrele(vp); 3521 counter_u64_add(numfullpathfail4, 1); 3522 SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM, 3523 startvp, NULL); 3524 return (ENOMEM); 3525 } 3526 buf[--buflen] = '/'; 3527 } 3528 counter_u64_add(numfullpathfound, 1); 3529 vrele(vp); 3530 3531 *retbuf = buf + buflen; 3532 SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf); 3533 *len -= buflen; 3534 *len += addend; 3535 return (0); 3536 } 3537 3538 /* 3539 * Resolve an arbitrary vnode to a pathname. 3540 * 3541 * Note 2 caveats: 3542 * - hardlinks are not tracked, thus if the vnode is not a directory this can 3543 * resolve to a different path than the one used to find it 3544 * - namecache is not mandatory, meaning names are not guaranteed to be added 3545 * (in which case resolving fails) 3546 */ 3547 static void __inline 3548 cache_rev_failed_impl(int *reason, int line) 3549 { 3550 3551 *reason = line; 3552 } 3553 #define cache_rev_failed(var) cache_rev_failed_impl((var), __LINE__) 3554 3555 static int 3556 vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf, 3557 char **retbuf, size_t *buflen, size_t addend) 3558 { 3559 #ifdef KDTRACE_HOOKS 3560 struct vnode *startvp = vp; 3561 #endif 3562 struct vnode *tvp; 3563 struct mount *mp; 3564 struct namecache *ncp; 3565 size_t orig_buflen; 3566 int reason; 3567 int error; 3568 #ifdef KDTRACE_HOOKS 3569 int i; 3570 #endif 3571 seqc_t vp_seqc, tvp_seqc; 3572 u_char nc_flag; 3573 3574 VFS_SMR_ASSERT_ENTERED(); 3575 3576 if (!atomic_load_char(&cache_fast_lookup_enabled)) { 3577 vfs_smr_exit(); 3578 return (-1); 3579 } 3580 3581 orig_buflen = *buflen; 3582 3583 if (addend == 0) { 3584 MPASS(*buflen >= 2); 3585 *buflen -= 1; 3586 buf[*buflen] = '\0'; 3587 } 3588 3589 if (vp == rdir || vp == rootvnode) { 3590 if (addend == 0) { 3591 *buflen -= 1; 3592 buf[*buflen] = '/'; 3593 } 3594 goto out_ok; 3595 } 3596 3597 #ifdef KDTRACE_HOOKS 3598 i = 0; 3599 #endif 3600 error = -1; 3601 ncp = NULL; /* for sdt probe down below */ 3602 vp_seqc = vn_seqc_read_any(vp); 3603 if (seqc_in_modify(vp_seqc)) { 3604 cache_rev_failed(&reason); 3605 goto out_abort; 3606 } 3607 3608 for (;;) { 3609 #ifdef KDTRACE_HOOKS 3610 i++; 3611 #endif 3612 if ((vp->v_vflag & VV_ROOT) != 0) { 3613 mp = atomic_load_ptr(&vp->v_mount); 3614 if (mp == NULL) { 3615 cache_rev_failed(&reason); 3616 goto out_abort; 3617 } 3618 tvp = atomic_load_ptr(&mp->mnt_vnodecovered); 3619 tvp_seqc = vn_seqc_read_any(tvp); 3620 if (seqc_in_modify(tvp_seqc)) { 3621 cache_rev_failed(&reason); 3622 goto out_abort; 3623 } 3624 if (!vn_seqc_consistent(vp, vp_seqc)) { 3625 cache_rev_failed(&reason); 3626 goto out_abort; 3627 } 3628 vp = tvp; 3629 vp_seqc = tvp_seqc; 3630 continue; 3631 } 3632 ncp = atomic_load_consume_ptr(&vp->v_cache_dd); 3633 if (ncp == NULL) { 3634 cache_rev_failed(&reason); 3635 goto out_abort; 3636 } 3637 nc_flag = atomic_load_char(&ncp->nc_flag); 3638 if ((nc_flag & NCF_ISDOTDOT) != 0) { 3639 cache_rev_failed(&reason); 3640 goto out_abort; 3641 } 3642 if (ncp->nc_nlen >= *buflen) { 3643 cache_rev_failed(&reason); 3644 error = ENOMEM; 3645 goto out_abort; 3646 } 3647 *buflen -= ncp->nc_nlen; 3648 memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen); 3649 *buflen -= 1; 3650 buf[*buflen] = '/'; 3651 tvp = ncp->nc_dvp; 3652 tvp_seqc = vn_seqc_read_any(tvp); 3653 if (seqc_in_modify(tvp_seqc)) { 3654 cache_rev_failed(&reason); 3655 goto out_abort; 3656 } 3657 if (!vn_seqc_consistent(vp, vp_seqc)) { 3658 cache_rev_failed(&reason); 3659 goto out_abort; 3660 } 3661 /* 3662 * Acquire fence provided by vn_seqc_read_any above. 3663 */ 3664 if (__predict_false(atomic_load_ptr(&vp->v_cache_dd) != ncp)) { 3665 cache_rev_failed(&reason); 3666 goto out_abort; 3667 } 3668 if (!cache_ncp_canuse(ncp)) { 3669 cache_rev_failed(&reason); 3670 goto out_abort; 3671 } 3672 vp = tvp; 3673 vp_seqc = tvp_seqc; 3674 if (vp == rdir || vp == rootvnode) 3675 break; 3676 } 3677 out_ok: 3678 vfs_smr_exit(); 3679 *retbuf = buf + *buflen; 3680 *buflen = orig_buflen - *buflen + addend; 3681 SDT_PROBE2(vfs, namecache, fullpath_smr, hit, startvp, *retbuf); 3682 return (0); 3683 3684 out_abort: 3685 *buflen = orig_buflen; 3686 SDT_PROBE4(vfs, namecache, fullpath_smr, miss, startvp, ncp, reason, i); 3687 vfs_smr_exit(); 3688 return (error); 3689 } 3690 3691 static int 3692 vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf, 3693 size_t *buflen) 3694 { 3695 size_t orig_buflen, addend; 3696 int error; 3697 3698 if (*buflen < 2) 3699 return (EINVAL); 3700 3701 orig_buflen = *buflen; 3702 3703 vref(vp); 3704 addend = 0; 3705 if (vp->v_type != VDIR) { 3706 *buflen -= 1; 3707 buf[*buflen] = '\0'; 3708 error = vn_vptocnp(&vp, buf, buflen); 3709 if (error) 3710 return (error); 3711 if (*buflen == 0) { 3712 vrele(vp); 3713 return (ENOMEM); 3714 } 3715 *buflen -= 1; 3716 buf[*buflen] = '/'; 3717 addend = orig_buflen - *buflen; 3718 } 3719 3720 return (vn_fullpath_dir(vp, rdir, buf, retbuf, buflen, addend)); 3721 } 3722 3723 /* 3724 * Resolve an arbitrary vnode to a pathname (taking care of hardlinks). 3725 * 3726 * Since the namecache does not track hardlinks, the caller is expected to 3727 * first look up the target vnode with WANTPARENT flag passed to namei to get 3728 * dvp and vp. 3729 * 3730 * Then we have 2 cases: 3731 * - if the found vnode is a directory, the path can be constructed just by 3732 * following names up the chain 3733 * - otherwise we populate the buffer with the saved name and start resolving 3734 * from the parent 3735 */ 3736 int 3737 vn_fullpath_hardlink(struct vnode *vp, struct vnode *dvp, 3738 const char *hrdl_name, size_t hrdl_name_length, 3739 char **retbuf, char **freebuf, size_t *buflen) 3740 { 3741 char *buf, *tmpbuf; 3742 struct pwd *pwd; 3743 size_t addend; 3744 int error; 3745 __enum_uint8(vtype) type; 3746 3747 if (*buflen < 2) 3748 return (EINVAL); 3749 if (*buflen > MAXPATHLEN) 3750 *buflen = MAXPATHLEN; 3751 3752 buf = malloc(*buflen, M_TEMP, M_WAITOK); 3753 3754 addend = 0; 3755 3756 /* 3757 * Check for VBAD to work around the vp_crossmp bug in lookup(). 3758 * 3759 * For example consider tmpfs on /tmp and realpath /tmp. ni_vp will be 3760 * set to mount point's root vnode while ni_dvp will be vp_crossmp. 3761 * If the type is VDIR (like in this very case) we can skip looking 3762 * at ni_dvp in the first place. However, since vnodes get passed here 3763 * unlocked the target may transition to doomed state (type == VBAD) 3764 * before we get to evaluate the condition. If this happens, we will 3765 * populate part of the buffer and descend to vn_fullpath_dir with 3766 * vp == vp_crossmp. Prevent the problem by checking for VBAD. 3767 */ 3768 type = atomic_load_8(&vp->v_type); 3769 if (type == VBAD) { 3770 error = ENOENT; 3771 goto out_bad; 3772 } 3773 if (type != VDIR) { 3774 addend = hrdl_name_length + 2; 3775 if (*buflen < addend) { 3776 error = ENOMEM; 3777 goto out_bad; 3778 } 3779 *buflen -= addend; 3780 tmpbuf = buf + *buflen; 3781 tmpbuf[0] = '/'; 3782 memcpy(&tmpbuf[1], hrdl_name, hrdl_name_length); 3783 tmpbuf[addend - 1] = '\0'; 3784 vp = dvp; 3785 } 3786 3787 vfs_smr_enter(); 3788 pwd = pwd_get_smr(); 3789 error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, buflen, 3790 addend); 3791 VFS_SMR_ASSERT_NOT_ENTERED(); 3792 if (error < 0) { 3793 pwd = pwd_hold(curthread); 3794 vref(vp); 3795 error = vn_fullpath_dir(vp, pwd->pwd_rdir, buf, retbuf, buflen, 3796 addend); 3797 pwd_drop(pwd); 3798 } 3799 if (error != 0) 3800 goto out_bad; 3801 3802 *freebuf = buf; 3803 3804 return (0); 3805 out_bad: 3806 free(buf, M_TEMP); 3807 return (error); 3808 } 3809 3810 struct vnode * 3811 vn_dir_dd_ino(struct vnode *vp) 3812 { 3813 struct namecache *ncp; 3814 struct vnode *ddvp; 3815 struct mtx *vlp; 3816 enum vgetstate vs; 3817 3818 ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino"); 3819 vlp = VP2VNODELOCK(vp); 3820 mtx_lock(vlp); 3821 TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) { 3822 if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) 3823 continue; 3824 ddvp = ncp->nc_dvp; 3825 vs = vget_prep(ddvp); 3826 mtx_unlock(vlp); 3827 if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs)) 3828 return (NULL); 3829 return (ddvp); 3830 } 3831 mtx_unlock(vlp); 3832 return (NULL); 3833 } 3834 3835 int 3836 vn_commname(struct vnode *vp, char *buf, u_int buflen) 3837 { 3838 struct namecache *ncp; 3839 struct mtx *vlp; 3840 int l; 3841 3842 vlp = VP2VNODELOCK(vp); 3843 mtx_lock(vlp); 3844 TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) 3845 if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) 3846 break; 3847 if (ncp == NULL) { 3848 mtx_unlock(vlp); 3849 return (ENOENT); 3850 } 3851 l = min(ncp->nc_nlen, buflen - 1); 3852 memcpy(buf, ncp->nc_name, l); 3853 mtx_unlock(vlp); 3854 buf[l] = '\0'; 3855 return (0); 3856 } 3857 3858 /* 3859 * This function updates path string to vnode's full global path 3860 * and checks the size of the new path string against the pathlen argument. 3861 * 3862 * Requires a locked, referenced vnode. 3863 * Vnode is re-locked on success or ENODEV, otherwise unlocked. 3864 * 3865 * If vp is a directory, the call to vn_fullpath_global() always succeeds 3866 * because it falls back to the ".." lookup if the namecache lookup fails. 3867 */ 3868 int 3869 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path, 3870 u_int pathlen) 3871 { 3872 struct nameidata nd; 3873 struct vnode *vp1; 3874 char *rpath, *fbuf; 3875 int error; 3876 3877 ASSERT_VOP_ELOCKED(vp, __func__); 3878 3879 /* Construct global filesystem path from vp. */ 3880 VOP_UNLOCK(vp); 3881 error = vn_fullpath_global(vp, &rpath, &fbuf); 3882 3883 if (error != 0) { 3884 vrele(vp); 3885 return (error); 3886 } 3887 3888 if (strlen(rpath) >= pathlen) { 3889 vrele(vp); 3890 error = ENAMETOOLONG; 3891 goto out; 3892 } 3893 3894 /* 3895 * Re-lookup the vnode by path to detect a possible rename. 3896 * As a side effect, the vnode is relocked. 3897 * If vnode was renamed, return ENOENT. 3898 */ 3899 NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1, UIO_SYSSPACE, path); 3900 error = namei(&nd); 3901 if (error != 0) { 3902 vrele(vp); 3903 goto out; 3904 } 3905 NDFREE_PNBUF(&nd); 3906 vp1 = nd.ni_vp; 3907 vrele(vp); 3908 if (vp1 == vp) 3909 strcpy(path, rpath); 3910 else { 3911 vput(vp1); 3912 error = ENOENT; 3913 } 3914 3915 out: 3916 free(fbuf, M_TEMP); 3917 return (error); 3918 } 3919 3920 /* 3921 * This is similar to vn_path_to_global_path but allows for regular 3922 * files which may not be present in the cache. 3923 * 3924 * Requires a locked, referenced vnode. 3925 * Vnode is re-locked on success or ENODEV, otherwise unlocked. 3926 */ 3927 int 3928 vn_path_to_global_path_hardlink(struct thread *td, struct vnode *vp, 3929 struct vnode *dvp, char *path, u_int pathlen, const char *leaf_name, 3930 size_t leaf_length) 3931 { 3932 struct nameidata nd; 3933 struct vnode *vp1; 3934 char *rpath, *fbuf; 3935 size_t len; 3936 int error; 3937 3938 ASSERT_VOP_ELOCKED(vp, __func__); 3939 3940 /* 3941 * Construct global filesystem path from dvp, vp and leaf 3942 * name. 3943 */ 3944 VOP_UNLOCK(vp); 3945 len = pathlen; 3946 error = vn_fullpath_hardlink(vp, dvp, leaf_name, leaf_length, 3947 &rpath, &fbuf, &len); 3948 3949 if (error != 0) { 3950 vrele(vp); 3951 return (error); 3952 } 3953 3954 if (strlen(rpath) >= pathlen) { 3955 vrele(vp); 3956 error = ENAMETOOLONG; 3957 goto out; 3958 } 3959 3960 /* 3961 * Re-lookup the vnode by path to detect a possible rename. 3962 * As a side effect, the vnode is relocked. 3963 * If vnode was renamed, return ENOENT. 3964 */ 3965 NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1, UIO_SYSSPACE, path); 3966 error = namei(&nd); 3967 if (error != 0) { 3968 vrele(vp); 3969 goto out; 3970 } 3971 NDFREE_PNBUF(&nd); 3972 vp1 = nd.ni_vp; 3973 vrele(vp); 3974 if (vp1 == vp) 3975 strcpy(path, rpath); 3976 else { 3977 vput(vp1); 3978 error = ENOENT; 3979 } 3980 3981 out: 3982 free(fbuf, M_TEMP); 3983 return (error); 3984 } 3985 3986 #ifdef DDB 3987 static void 3988 db_print_vpath(struct vnode *vp) 3989 { 3990 3991 while (vp != NULL) { 3992 db_printf("%p: ", vp); 3993 if (vp == rootvnode) { 3994 db_printf("/"); 3995 vp = NULL; 3996 } else { 3997 if (vp->v_vflag & VV_ROOT) { 3998 db_printf("<mount point>"); 3999 vp = vp->v_mount->mnt_vnodecovered; 4000 } else { 4001 struct namecache *ncp; 4002 char *ncn; 4003 int i; 4004 4005 ncp = TAILQ_FIRST(&vp->v_cache_dst); 4006 if (ncp != NULL) { 4007 ncn = ncp->nc_name; 4008 for (i = 0; i < ncp->nc_nlen; i++) 4009 db_printf("%c", *ncn++); 4010 vp = ncp->nc_dvp; 4011 } else { 4012 vp = NULL; 4013 } 4014 } 4015 } 4016 db_printf("\n"); 4017 } 4018 4019 return; 4020 } 4021 4022 DB_SHOW_COMMAND(vpath, db_show_vpath) 4023 { 4024 struct vnode *vp; 4025 4026 if (!have_addr) { 4027 db_printf("usage: show vpath <struct vnode *>\n"); 4028 return; 4029 } 4030 4031 vp = (struct vnode *)addr; 4032 db_print_vpath(vp); 4033 } 4034 4035 #endif 4036 4037 static int cache_fast_lookup = 1; 4038 4039 #define CACHE_FPL_FAILED -2020 4040 4041 static int 4042 cache_vop_bad_vexec(struct vop_fplookup_vexec_args *v) 4043 { 4044 vn_printf(v->a_vp, "no proper vop_fplookup_vexec\n"); 4045 panic("no proper vop_fplookup_vexec"); 4046 } 4047 4048 static int 4049 cache_vop_bad_symlink(struct vop_fplookup_symlink_args *v) 4050 { 4051 vn_printf(v->a_vp, "no proper vop_fplookup_symlink\n"); 4052 panic("no proper vop_fplookup_symlink"); 4053 } 4054 4055 void 4056 cache_vop_vector_register(struct vop_vector *v) 4057 { 4058 size_t ops; 4059 4060 ops = 0; 4061 if (v->vop_fplookup_vexec != NULL) { 4062 ops++; 4063 } 4064 if (v->vop_fplookup_symlink != NULL) { 4065 ops++; 4066 } 4067 4068 if (ops == 2) { 4069 return; 4070 } 4071 4072 if (ops == 0) { 4073 v->vop_fplookup_vexec = cache_vop_bad_vexec; 4074 v->vop_fplookup_symlink = cache_vop_bad_symlink; 4075 return; 4076 } 4077 4078 printf("%s: invalid vop vector %p -- either all or none fplookup vops " 4079 "need to be provided", __func__, v); 4080 if (v->vop_fplookup_vexec == NULL) { 4081 printf("%s: missing vop_fplookup_vexec\n", __func__); 4082 } 4083 if (v->vop_fplookup_symlink == NULL) { 4084 printf("%s: missing vop_fplookup_symlink\n", __func__); 4085 } 4086 panic("bad vop vector %p", v); 4087 } 4088 4089 #ifdef INVARIANTS 4090 void 4091 cache_validate_vop_vector(struct mount *mp, struct vop_vector *vops) 4092 { 4093 if (mp == NULL) 4094 return; 4095 4096 if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0) 4097 return; 4098 4099 if (vops->vop_fplookup_vexec == NULL || 4100 vops->vop_fplookup_vexec == cache_vop_bad_vexec) 4101 panic("bad vop_fplookup_vexec on vector %p for filesystem %s", 4102 vops, mp->mnt_vfc->vfc_name); 4103 4104 if (vops->vop_fplookup_symlink == NULL || 4105 vops->vop_fplookup_symlink == cache_vop_bad_symlink) 4106 panic("bad vop_fplookup_symlink on vector %p for filesystem %s", 4107 vops, mp->mnt_vfc->vfc_name); 4108 } 4109 #endif 4110 4111 void 4112 cache_fast_lookup_enabled_recalc(void) 4113 { 4114 int lookup_flag; 4115 int mac_on; 4116 4117 #ifdef MAC 4118 mac_on = mac_vnode_check_lookup_enabled(); 4119 mac_on |= mac_vnode_check_readlink_enabled(); 4120 #else 4121 mac_on = 0; 4122 #endif 4123 4124 lookup_flag = atomic_load_int(&cache_fast_lookup); 4125 if (lookup_flag && !mac_on) { 4126 atomic_store_char(&cache_fast_lookup_enabled, true); 4127 } else { 4128 atomic_store_char(&cache_fast_lookup_enabled, false); 4129 } 4130 } 4131 4132 static int 4133 syscal_vfs_cache_fast_lookup(SYSCTL_HANDLER_ARGS) 4134 { 4135 int error, old; 4136 4137 old = atomic_load_int(&cache_fast_lookup); 4138 error = sysctl_handle_int(oidp, arg1, arg2, req); 4139 if (error == 0 && req->newptr && old != atomic_load_int(&cache_fast_lookup)) 4140 cache_fast_lookup_enabled_recalc(); 4141 return (error); 4142 } 4143 SYSCTL_PROC(_vfs_cache_param, OID_AUTO, fast_lookup, CTLTYPE_INT|CTLFLAG_RW|CTLFLAG_MPSAFE, 4144 &cache_fast_lookup, 0, syscal_vfs_cache_fast_lookup, "IU", ""); 4145 4146 /* 4147 * Components of nameidata (or objects it can point to) which may 4148 * need restoring in case fast path lookup fails. 4149 */ 4150 struct nameidata_outer { 4151 size_t ni_pathlen; 4152 int cn_flags; 4153 }; 4154 4155 struct nameidata_saved { 4156 #ifdef INVARIANTS 4157 char *cn_nameptr; 4158 size_t ni_pathlen; 4159 #endif 4160 }; 4161 4162 #ifdef INVARIANTS 4163 struct cache_fpl_debug { 4164 size_t ni_pathlen; 4165 }; 4166 #endif 4167 4168 struct cache_fpl { 4169 struct nameidata *ndp; 4170 struct componentname *cnp; 4171 char *nulchar; 4172 struct vnode *dvp; 4173 struct vnode *tvp; 4174 seqc_t dvp_seqc; 4175 seqc_t tvp_seqc; 4176 uint32_t hash; 4177 struct nameidata_saved snd; 4178 struct nameidata_outer snd_outer; 4179 int line; 4180 enum cache_fpl_status status:8; 4181 bool in_smr; 4182 bool fsearch; 4183 struct pwd **pwd; 4184 #ifdef INVARIANTS 4185 struct cache_fpl_debug debug; 4186 #endif 4187 }; 4188 4189 static bool cache_fplookup_mp_supported(struct mount *mp); 4190 static bool cache_fplookup_is_mp(struct cache_fpl *fpl); 4191 static int cache_fplookup_cross_mount(struct cache_fpl *fpl); 4192 static int cache_fplookup_partial_setup(struct cache_fpl *fpl); 4193 static int cache_fplookup_skip_slashes(struct cache_fpl *fpl); 4194 static int cache_fplookup_trailingslash(struct cache_fpl *fpl); 4195 static void cache_fpl_pathlen_dec(struct cache_fpl *fpl); 4196 static void cache_fpl_pathlen_inc(struct cache_fpl *fpl); 4197 static void cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n); 4198 static void cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n); 4199 4200 static void 4201 cache_fpl_cleanup_cnp(struct componentname *cnp) 4202 { 4203 4204 uma_zfree(namei_zone, cnp->cn_pnbuf); 4205 cnp->cn_pnbuf = NULL; 4206 cnp->cn_nameptr = NULL; 4207 } 4208 4209 static struct vnode * 4210 cache_fpl_handle_root(struct cache_fpl *fpl) 4211 { 4212 struct nameidata *ndp; 4213 struct componentname *cnp; 4214 4215 ndp = fpl->ndp; 4216 cnp = fpl->cnp; 4217 4218 MPASS(*(cnp->cn_nameptr) == '/'); 4219 cnp->cn_nameptr++; 4220 cache_fpl_pathlen_dec(fpl); 4221 4222 if (__predict_false(*(cnp->cn_nameptr) == '/')) { 4223 do { 4224 cnp->cn_nameptr++; 4225 cache_fpl_pathlen_dec(fpl); 4226 } while (*(cnp->cn_nameptr) == '/'); 4227 } 4228 4229 return (ndp->ni_rootdir); 4230 } 4231 4232 static void 4233 cache_fpl_checkpoint_outer(struct cache_fpl *fpl) 4234 { 4235 4236 fpl->snd_outer.ni_pathlen = fpl->ndp->ni_pathlen; 4237 fpl->snd_outer.cn_flags = fpl->ndp->ni_cnd.cn_flags; 4238 } 4239 4240 static void 4241 cache_fpl_checkpoint(struct cache_fpl *fpl) 4242 { 4243 4244 #ifdef INVARIANTS 4245 fpl->snd.cn_nameptr = fpl->ndp->ni_cnd.cn_nameptr; 4246 fpl->snd.ni_pathlen = fpl->debug.ni_pathlen; 4247 #endif 4248 } 4249 4250 static void 4251 cache_fpl_restore_partial(struct cache_fpl *fpl) 4252 { 4253 4254 fpl->ndp->ni_cnd.cn_flags = fpl->snd_outer.cn_flags; 4255 #ifdef INVARIANTS 4256 fpl->debug.ni_pathlen = fpl->snd.ni_pathlen; 4257 #endif 4258 } 4259 4260 static void 4261 cache_fpl_restore_abort(struct cache_fpl *fpl) 4262 { 4263 4264 cache_fpl_restore_partial(fpl); 4265 /* 4266 * It is 0 on entry by API contract. 4267 */ 4268 fpl->ndp->ni_resflags = 0; 4269 fpl->ndp->ni_cnd.cn_nameptr = fpl->ndp->ni_cnd.cn_pnbuf; 4270 fpl->ndp->ni_pathlen = fpl->snd_outer.ni_pathlen; 4271 } 4272 4273 #ifdef INVARIANTS 4274 #define cache_fpl_smr_assert_entered(fpl) ({ \ 4275 struct cache_fpl *_fpl = (fpl); \ 4276 MPASS(_fpl->in_smr == true); \ 4277 VFS_SMR_ASSERT_ENTERED(); \ 4278 }) 4279 #define cache_fpl_smr_assert_not_entered(fpl) ({ \ 4280 struct cache_fpl *_fpl = (fpl); \ 4281 MPASS(_fpl->in_smr == false); \ 4282 VFS_SMR_ASSERT_NOT_ENTERED(); \ 4283 }) 4284 static void 4285 cache_fpl_assert_status(struct cache_fpl *fpl) 4286 { 4287 4288 switch (fpl->status) { 4289 case CACHE_FPL_STATUS_UNSET: 4290 __assert_unreachable(); 4291 break; 4292 case CACHE_FPL_STATUS_DESTROYED: 4293 case CACHE_FPL_STATUS_ABORTED: 4294 case CACHE_FPL_STATUS_PARTIAL: 4295 case CACHE_FPL_STATUS_HANDLED: 4296 break; 4297 } 4298 } 4299 #else 4300 #define cache_fpl_smr_assert_entered(fpl) do { } while (0) 4301 #define cache_fpl_smr_assert_not_entered(fpl) do { } while (0) 4302 #define cache_fpl_assert_status(fpl) do { } while (0) 4303 #endif 4304 4305 #define cache_fpl_smr_enter_initial(fpl) ({ \ 4306 struct cache_fpl *_fpl = (fpl); \ 4307 vfs_smr_enter(); \ 4308 _fpl->in_smr = true; \ 4309 }) 4310 4311 #define cache_fpl_smr_enter(fpl) ({ \ 4312 struct cache_fpl *_fpl = (fpl); \ 4313 MPASS(_fpl->in_smr == false); \ 4314 vfs_smr_enter(); \ 4315 _fpl->in_smr = true; \ 4316 }) 4317 4318 #define cache_fpl_smr_exit(fpl) ({ \ 4319 struct cache_fpl *_fpl = (fpl); \ 4320 MPASS(_fpl->in_smr == true); \ 4321 vfs_smr_exit(); \ 4322 _fpl->in_smr = false; \ 4323 }) 4324 4325 static int 4326 cache_fpl_aborted_early_impl(struct cache_fpl *fpl, int line) 4327 { 4328 4329 if (fpl->status != CACHE_FPL_STATUS_UNSET) { 4330 KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL, 4331 ("%s: converting to abort from %d at %d, set at %d\n", 4332 __func__, fpl->status, line, fpl->line)); 4333 } 4334 cache_fpl_smr_assert_not_entered(fpl); 4335 fpl->status = CACHE_FPL_STATUS_ABORTED; 4336 fpl->line = line; 4337 return (CACHE_FPL_FAILED); 4338 } 4339 4340 #define cache_fpl_aborted_early(x) cache_fpl_aborted_early_impl((x), __LINE__) 4341 4342 static int __noinline 4343 cache_fpl_aborted_impl(struct cache_fpl *fpl, int line) 4344 { 4345 struct nameidata *ndp; 4346 struct componentname *cnp; 4347 4348 ndp = fpl->ndp; 4349 cnp = fpl->cnp; 4350 4351 if (fpl->status != CACHE_FPL_STATUS_UNSET) { 4352 KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL, 4353 ("%s: converting to abort from %d at %d, set at %d\n", 4354 __func__, fpl->status, line, fpl->line)); 4355 } 4356 fpl->status = CACHE_FPL_STATUS_ABORTED; 4357 fpl->line = line; 4358 if (fpl->in_smr) 4359 cache_fpl_smr_exit(fpl); 4360 cache_fpl_restore_abort(fpl); 4361 /* 4362 * Resolving symlinks overwrites data passed by the caller. 4363 * Let namei know. 4364 */ 4365 if (ndp->ni_loopcnt > 0) { 4366 fpl->status = CACHE_FPL_STATUS_DESTROYED; 4367 cache_fpl_cleanup_cnp(cnp); 4368 } 4369 return (CACHE_FPL_FAILED); 4370 } 4371 4372 #define cache_fpl_aborted(x) cache_fpl_aborted_impl((x), __LINE__) 4373 4374 static int __noinline 4375 cache_fpl_partial_impl(struct cache_fpl *fpl, int line) 4376 { 4377 4378 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET, 4379 ("%s: setting to partial at %d, but already set to %d at %d\n", 4380 __func__, line, fpl->status, fpl->line)); 4381 cache_fpl_smr_assert_entered(fpl); 4382 fpl->status = CACHE_FPL_STATUS_PARTIAL; 4383 fpl->line = line; 4384 return (cache_fplookup_partial_setup(fpl)); 4385 } 4386 4387 #define cache_fpl_partial(x) cache_fpl_partial_impl((x), __LINE__) 4388 4389 static int 4390 cache_fpl_handled_impl(struct cache_fpl *fpl, int line) 4391 { 4392 4393 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET, 4394 ("%s: setting to handled at %d, but already set to %d at %d\n", 4395 __func__, line, fpl->status, fpl->line)); 4396 cache_fpl_smr_assert_not_entered(fpl); 4397 fpl->status = CACHE_FPL_STATUS_HANDLED; 4398 fpl->line = line; 4399 return (0); 4400 } 4401 4402 #define cache_fpl_handled(x) cache_fpl_handled_impl((x), __LINE__) 4403 4404 static int 4405 cache_fpl_handled_error_impl(struct cache_fpl *fpl, int error, int line) 4406 { 4407 4408 KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET, 4409 ("%s: setting to handled at %d, but already set to %d at %d\n", 4410 __func__, line, fpl->status, fpl->line)); 4411 MPASS(error != 0); 4412 MPASS(error != CACHE_FPL_FAILED); 4413 cache_fpl_smr_assert_not_entered(fpl); 4414 fpl->status = CACHE_FPL_STATUS_HANDLED; 4415 fpl->line = line; 4416 fpl->dvp = NULL; 4417 fpl->tvp = NULL; 4418 return (error); 4419 } 4420 4421 #define cache_fpl_handled_error(x, e) cache_fpl_handled_error_impl((x), (e), __LINE__) 4422 4423 static bool 4424 cache_fpl_terminated(struct cache_fpl *fpl) 4425 { 4426 4427 return (fpl->status != CACHE_FPL_STATUS_UNSET); 4428 } 4429 4430 #define CACHE_FPL_SUPPORTED_CN_FLAGS \ 4431 (NC_NOMAKEENTRY | NC_KEEPPOSENTRY | LOCKLEAF | LOCKPARENT | WANTPARENT | \ 4432 FAILIFEXISTS | FOLLOW | EMPTYPATH | LOCKSHARED | ISRESTARTED | WILLBEDIR | \ 4433 ISOPEN | NOMACCHECK | AUDITVNODE1 | AUDITVNODE2 | NOCAPCHECK | OPENREAD | \ 4434 OPENWRITE | WANTIOCTLCAPS) 4435 4436 #define CACHE_FPL_INTERNAL_CN_FLAGS \ 4437 (ISDOTDOT | MAKEENTRY | ISLASTCN) 4438 4439 _Static_assert((CACHE_FPL_SUPPORTED_CN_FLAGS & CACHE_FPL_INTERNAL_CN_FLAGS) == 0, 4440 "supported and internal flags overlap"); 4441 4442 static bool 4443 cache_fpl_islastcn(struct nameidata *ndp) 4444 { 4445 4446 return (*ndp->ni_next == 0); 4447 } 4448 4449 static bool 4450 cache_fpl_istrailingslash(struct cache_fpl *fpl) 4451 { 4452 4453 MPASS(fpl->nulchar > fpl->cnp->cn_pnbuf); 4454 return (*(fpl->nulchar - 1) == '/'); 4455 } 4456 4457 static bool 4458 cache_fpl_isdotdot(struct componentname *cnp) 4459 { 4460 4461 if (cnp->cn_namelen == 2 && 4462 cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.') 4463 return (true); 4464 return (false); 4465 } 4466 4467 static bool 4468 cache_can_fplookup(struct cache_fpl *fpl) 4469 { 4470 struct nameidata *ndp; 4471 struct componentname *cnp; 4472 struct thread *td; 4473 4474 ndp = fpl->ndp; 4475 cnp = fpl->cnp; 4476 td = curthread; 4477 4478 if (!atomic_load_char(&cache_fast_lookup_enabled)) { 4479 cache_fpl_aborted_early(fpl); 4480 return (false); 4481 } 4482 if ((cnp->cn_flags & ~CACHE_FPL_SUPPORTED_CN_FLAGS) != 0) { 4483 cache_fpl_aborted_early(fpl); 4484 return (false); 4485 } 4486 if (IN_CAPABILITY_MODE(td) || CAP_TRACING(td)) { 4487 cache_fpl_aborted_early(fpl); 4488 return (false); 4489 } 4490 if (AUDITING_TD(td)) { 4491 cache_fpl_aborted_early(fpl); 4492 return (false); 4493 } 4494 if (ndp->ni_startdir != NULL) { 4495 cache_fpl_aborted_early(fpl); 4496 return (false); 4497 } 4498 return (true); 4499 } 4500 4501 static int __noinline 4502 cache_fplookup_dirfd(struct cache_fpl *fpl, struct vnode **vpp) 4503 { 4504 struct nameidata *ndp; 4505 struct componentname *cnp; 4506 int error; 4507 bool fsearch; 4508 4509 ndp = fpl->ndp; 4510 cnp = fpl->cnp; 4511 4512 error = fgetvp_lookup_smr(ndp, vpp, &fsearch); 4513 if (__predict_false(error != 0)) { 4514 return (cache_fpl_aborted(fpl)); 4515 } 4516 fpl->fsearch = fsearch; 4517 if ((*vpp)->v_type != VDIR) { 4518 if (!((cnp->cn_flags & EMPTYPATH) != 0 && cnp->cn_pnbuf[0] == '\0')) { 4519 cache_fpl_smr_exit(fpl); 4520 return (cache_fpl_handled_error(fpl, ENOTDIR)); 4521 } 4522 } 4523 return (0); 4524 } 4525 4526 static int __noinline 4527 cache_fplookup_negative_promote(struct cache_fpl *fpl, struct namecache *oncp, 4528 uint32_t hash) 4529 { 4530 struct componentname *cnp; 4531 struct vnode *dvp; 4532 4533 cnp = fpl->cnp; 4534 dvp = fpl->dvp; 4535 4536 cache_fpl_smr_exit(fpl); 4537 if (cache_neg_promote_cond(dvp, cnp, oncp, hash)) 4538 return (cache_fpl_handled_error(fpl, ENOENT)); 4539 else 4540 return (cache_fpl_aborted(fpl)); 4541 } 4542 4543 /* 4544 * The target vnode is not supported, prepare for the slow path to take over. 4545 */ 4546 static int __noinline 4547 cache_fplookup_partial_setup(struct cache_fpl *fpl) 4548 { 4549 struct nameidata *ndp; 4550 struct componentname *cnp; 4551 enum vgetstate dvs; 4552 struct vnode *dvp; 4553 struct pwd *pwd; 4554 seqc_t dvp_seqc; 4555 4556 ndp = fpl->ndp; 4557 cnp = fpl->cnp; 4558 pwd = *(fpl->pwd); 4559 dvp = fpl->dvp; 4560 dvp_seqc = fpl->dvp_seqc; 4561 4562 if (!pwd_hold_smr(pwd)) { 4563 return (cache_fpl_aborted(fpl)); 4564 } 4565 4566 /* 4567 * Note that seqc is checked before the vnode is locked, so by 4568 * the time regular lookup gets to it it may have moved. 4569 * 4570 * Ultimately this does not affect correctness, any lookup errors 4571 * are userspace racing with itself. It is guaranteed that any 4572 * path which ultimately gets found could also have been found 4573 * by regular lookup going all the way in absence of concurrent 4574 * modifications. 4575 */ 4576 dvs = vget_prep_smr(dvp); 4577 cache_fpl_smr_exit(fpl); 4578 if (__predict_false(dvs == VGET_NONE)) { 4579 pwd_drop(pwd); 4580 return (cache_fpl_aborted(fpl)); 4581 } 4582 4583 vget_finish_ref(dvp, dvs); 4584 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 4585 vrele(dvp); 4586 pwd_drop(pwd); 4587 return (cache_fpl_aborted(fpl)); 4588 } 4589 4590 cache_fpl_restore_partial(fpl); 4591 #ifdef INVARIANTS 4592 if (cnp->cn_nameptr != fpl->snd.cn_nameptr) { 4593 panic("%s: cn_nameptr mismatch (%p != %p) full [%s]\n", __func__, 4594 cnp->cn_nameptr, fpl->snd.cn_nameptr, cnp->cn_pnbuf); 4595 } 4596 #endif 4597 4598 ndp->ni_startdir = dvp; 4599 cnp->cn_flags |= MAKEENTRY; 4600 if (cache_fpl_islastcn(ndp)) 4601 cnp->cn_flags |= ISLASTCN; 4602 if (cache_fpl_isdotdot(cnp)) 4603 cnp->cn_flags |= ISDOTDOT; 4604 4605 /* 4606 * Skip potential extra slashes parsing did not take care of. 4607 * cache_fplookup_skip_slashes explains the mechanism. 4608 */ 4609 if (__predict_false(*(cnp->cn_nameptr) == '/')) { 4610 do { 4611 cnp->cn_nameptr++; 4612 cache_fpl_pathlen_dec(fpl); 4613 } while (*(cnp->cn_nameptr) == '/'); 4614 } 4615 4616 ndp->ni_pathlen = fpl->nulchar - cnp->cn_nameptr + 1; 4617 #ifdef INVARIANTS 4618 if (ndp->ni_pathlen != fpl->debug.ni_pathlen) { 4619 panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n", 4620 __func__, ndp->ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar, 4621 cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf); 4622 } 4623 #endif 4624 return (0); 4625 } 4626 4627 static int 4628 cache_fplookup_final_child(struct cache_fpl *fpl, enum vgetstate tvs) 4629 { 4630 struct componentname *cnp; 4631 struct vnode *tvp; 4632 seqc_t tvp_seqc; 4633 int error, lkflags; 4634 4635 cnp = fpl->cnp; 4636 tvp = fpl->tvp; 4637 tvp_seqc = fpl->tvp_seqc; 4638 4639 if ((cnp->cn_flags & LOCKLEAF) != 0) { 4640 lkflags = LK_SHARED; 4641 if ((cnp->cn_flags & LOCKSHARED) == 0) 4642 lkflags = LK_EXCLUSIVE; 4643 error = vget_finish(tvp, lkflags, tvs); 4644 if (__predict_false(error != 0)) { 4645 return (cache_fpl_aborted(fpl)); 4646 } 4647 } else { 4648 vget_finish_ref(tvp, tvs); 4649 } 4650 4651 if (!vn_seqc_consistent(tvp, tvp_seqc)) { 4652 if ((cnp->cn_flags & LOCKLEAF) != 0) 4653 vput(tvp); 4654 else 4655 vrele(tvp); 4656 return (cache_fpl_aborted(fpl)); 4657 } 4658 4659 return (cache_fpl_handled(fpl)); 4660 } 4661 4662 /* 4663 * They want to possibly modify the state of the namecache. 4664 */ 4665 static int __noinline 4666 cache_fplookup_final_modifying(struct cache_fpl *fpl) 4667 { 4668 struct nameidata *ndp __diagused; 4669 struct componentname *cnp; 4670 enum vgetstate dvs; 4671 struct vnode *dvp, *tvp; 4672 struct mount *mp; 4673 seqc_t dvp_seqc; 4674 int error; 4675 bool docache; 4676 4677 ndp = fpl->ndp; 4678 cnp = fpl->cnp; 4679 dvp = fpl->dvp; 4680 dvp_seqc = fpl->dvp_seqc; 4681 4682 MPASS(*(cnp->cn_nameptr) != '/'); 4683 MPASS(cache_fpl_islastcn(ndp)); 4684 if ((cnp->cn_flags & LOCKPARENT) == 0) 4685 MPASS((cnp->cn_flags & WANTPARENT) != 0); 4686 MPASS((cnp->cn_flags & TRAILINGSLASH) == 0); 4687 MPASS(cnp->cn_nameiop == CREATE || cnp->cn_nameiop == DELETE || 4688 cnp->cn_nameiop == RENAME); 4689 MPASS((cnp->cn_flags & MAKEENTRY) == 0); 4690 MPASS((cnp->cn_flags & ISDOTDOT) == 0); 4691 4692 docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE; 4693 if (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME) 4694 docache = false; 4695 4696 /* 4697 * Regular lookup nulifies the slash, which we don't do here. 4698 * Don't take chances with filesystem routines seeing it for 4699 * the last entry. 4700 */ 4701 if (cache_fpl_istrailingslash(fpl)) { 4702 return (cache_fpl_partial(fpl)); 4703 } 4704 4705 mp = atomic_load_ptr(&dvp->v_mount); 4706 if (__predict_false(mp == NULL)) { 4707 return (cache_fpl_aborted(fpl)); 4708 } 4709 4710 if (__predict_false(mp->mnt_flag & MNT_RDONLY)) { 4711 cache_fpl_smr_exit(fpl); 4712 /* 4713 * Original code keeps not checking for CREATE which 4714 * might be a bug. For now let the old lookup decide. 4715 */ 4716 if (cnp->cn_nameiop == CREATE) { 4717 return (cache_fpl_aborted(fpl)); 4718 } 4719 return (cache_fpl_handled_error(fpl, EROFS)); 4720 } 4721 4722 if (fpl->tvp != NULL && (cnp->cn_flags & FAILIFEXISTS) != 0) { 4723 cache_fpl_smr_exit(fpl); 4724 return (cache_fpl_handled_error(fpl, EEXIST)); 4725 } 4726 4727 /* 4728 * Secure access to dvp; check cache_fplookup_partial_setup for 4729 * reasoning. 4730 * 4731 * XXX At least UFS requires its lookup routine to be called for 4732 * the last path component, which leads to some level of complication 4733 * and inefficiency: 4734 * - the target routine always locks the target vnode, but our caller 4735 * may not need it locked 4736 * - some of the VOP machinery asserts that the parent is locked, which 4737 * once more may be not required 4738 * 4739 * TODO: add a flag for filesystems which don't need this. 4740 */ 4741 dvs = vget_prep_smr(dvp); 4742 cache_fpl_smr_exit(fpl); 4743 if (__predict_false(dvs == VGET_NONE)) { 4744 return (cache_fpl_aborted(fpl)); 4745 } 4746 4747 vget_finish_ref(dvp, dvs); 4748 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 4749 vrele(dvp); 4750 return (cache_fpl_aborted(fpl)); 4751 } 4752 4753 error = vn_lock(dvp, LK_EXCLUSIVE); 4754 if (__predict_false(error != 0)) { 4755 vrele(dvp); 4756 return (cache_fpl_aborted(fpl)); 4757 } 4758 4759 tvp = NULL; 4760 cnp->cn_flags |= ISLASTCN; 4761 if (docache) 4762 cnp->cn_flags |= MAKEENTRY; 4763 if (cache_fpl_isdotdot(cnp)) 4764 cnp->cn_flags |= ISDOTDOT; 4765 cnp->cn_lkflags = LK_EXCLUSIVE; 4766 error = VOP_LOOKUP(dvp, &tvp, cnp); 4767 switch (error) { 4768 case EJUSTRETURN: 4769 case 0: 4770 break; 4771 case ENOTDIR: 4772 case ENOENT: 4773 vput(dvp); 4774 return (cache_fpl_handled_error(fpl, error)); 4775 default: 4776 vput(dvp); 4777 return (cache_fpl_aborted(fpl)); 4778 } 4779 4780 fpl->tvp = tvp; 4781 4782 if (tvp == NULL) { 4783 MPASS(error == EJUSTRETURN); 4784 if ((cnp->cn_flags & LOCKPARENT) == 0) { 4785 VOP_UNLOCK(dvp); 4786 } 4787 return (cache_fpl_handled(fpl)); 4788 } 4789 4790 /* 4791 * There are very hairy corner cases concerning various flag combinations 4792 * and locking state. In particular here we only hold one lock instead of 4793 * two. 4794 * 4795 * Skip the complexity as it is of no significance for normal workloads. 4796 */ 4797 if (__predict_false(tvp == dvp)) { 4798 vput(dvp); 4799 vrele(tvp); 4800 return (cache_fpl_aborted(fpl)); 4801 } 4802 4803 /* 4804 * If they want the symlink itself we are fine, but if they want to 4805 * follow it regular lookup has to be engaged. 4806 */ 4807 if (tvp->v_type == VLNK) { 4808 if ((cnp->cn_flags & FOLLOW) != 0) { 4809 vput(dvp); 4810 vput(tvp); 4811 return (cache_fpl_aborted(fpl)); 4812 } 4813 } 4814 4815 /* 4816 * Since we expect this to be the terminal vnode it should almost never 4817 * be a mount point. 4818 */ 4819 if (__predict_false(cache_fplookup_is_mp(fpl))) { 4820 vput(dvp); 4821 vput(tvp); 4822 return (cache_fpl_aborted(fpl)); 4823 } 4824 4825 if ((cnp->cn_flags & FAILIFEXISTS) != 0) { 4826 vput(dvp); 4827 vput(tvp); 4828 return (cache_fpl_handled_error(fpl, EEXIST)); 4829 } 4830 4831 if ((cnp->cn_flags & LOCKLEAF) == 0) { 4832 VOP_UNLOCK(tvp); 4833 } 4834 4835 if ((cnp->cn_flags & LOCKPARENT) == 0) { 4836 VOP_UNLOCK(dvp); 4837 } 4838 4839 return (cache_fpl_handled(fpl)); 4840 } 4841 4842 static int __noinline 4843 cache_fplookup_modifying(struct cache_fpl *fpl) 4844 { 4845 struct nameidata *ndp; 4846 4847 ndp = fpl->ndp; 4848 4849 if (!cache_fpl_islastcn(ndp)) { 4850 return (cache_fpl_partial(fpl)); 4851 } 4852 return (cache_fplookup_final_modifying(fpl)); 4853 } 4854 4855 static int __noinline 4856 cache_fplookup_final_withparent(struct cache_fpl *fpl) 4857 { 4858 struct componentname *cnp; 4859 enum vgetstate dvs, tvs; 4860 struct vnode *dvp, *tvp; 4861 seqc_t dvp_seqc; 4862 int error; 4863 4864 cnp = fpl->cnp; 4865 dvp = fpl->dvp; 4866 dvp_seqc = fpl->dvp_seqc; 4867 tvp = fpl->tvp; 4868 4869 MPASS((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0); 4870 4871 /* 4872 * This is less efficient than it can be for simplicity. 4873 */ 4874 dvs = vget_prep_smr(dvp); 4875 if (__predict_false(dvs == VGET_NONE)) { 4876 return (cache_fpl_aborted(fpl)); 4877 } 4878 tvs = vget_prep_smr(tvp); 4879 if (__predict_false(tvs == VGET_NONE)) { 4880 cache_fpl_smr_exit(fpl); 4881 vget_abort(dvp, dvs); 4882 return (cache_fpl_aborted(fpl)); 4883 } 4884 4885 cache_fpl_smr_exit(fpl); 4886 4887 if ((cnp->cn_flags & LOCKPARENT) != 0) { 4888 error = vget_finish(dvp, LK_EXCLUSIVE, dvs); 4889 if (__predict_false(error != 0)) { 4890 vget_abort(tvp, tvs); 4891 return (cache_fpl_aborted(fpl)); 4892 } 4893 } else { 4894 vget_finish_ref(dvp, dvs); 4895 } 4896 4897 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 4898 vget_abort(tvp, tvs); 4899 if ((cnp->cn_flags & LOCKPARENT) != 0) 4900 vput(dvp); 4901 else 4902 vrele(dvp); 4903 return (cache_fpl_aborted(fpl)); 4904 } 4905 4906 error = cache_fplookup_final_child(fpl, tvs); 4907 if (__predict_false(error != 0)) { 4908 MPASS(fpl->status == CACHE_FPL_STATUS_ABORTED || 4909 fpl->status == CACHE_FPL_STATUS_DESTROYED); 4910 if ((cnp->cn_flags & LOCKPARENT) != 0) 4911 vput(dvp); 4912 else 4913 vrele(dvp); 4914 return (error); 4915 } 4916 4917 MPASS(fpl->status == CACHE_FPL_STATUS_HANDLED); 4918 return (0); 4919 } 4920 4921 static int 4922 cache_fplookup_final(struct cache_fpl *fpl) 4923 { 4924 struct componentname *cnp; 4925 enum vgetstate tvs; 4926 struct vnode *dvp, *tvp; 4927 seqc_t dvp_seqc; 4928 4929 cnp = fpl->cnp; 4930 dvp = fpl->dvp; 4931 dvp_seqc = fpl->dvp_seqc; 4932 tvp = fpl->tvp; 4933 4934 MPASS(*(cnp->cn_nameptr) != '/'); 4935 4936 if (cnp->cn_nameiop != LOOKUP) { 4937 return (cache_fplookup_final_modifying(fpl)); 4938 } 4939 4940 if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0) 4941 return (cache_fplookup_final_withparent(fpl)); 4942 4943 tvs = vget_prep_smr(tvp); 4944 if (__predict_false(tvs == VGET_NONE)) { 4945 return (cache_fpl_partial(fpl)); 4946 } 4947 4948 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 4949 cache_fpl_smr_exit(fpl); 4950 vget_abort(tvp, tvs); 4951 return (cache_fpl_aborted(fpl)); 4952 } 4953 4954 cache_fpl_smr_exit(fpl); 4955 return (cache_fplookup_final_child(fpl, tvs)); 4956 } 4957 4958 /* 4959 * Comment from locked lookup: 4960 * Check for degenerate name (e.g. / or "") which is a way of talking about a 4961 * directory, e.g. like "/." or ".". 4962 */ 4963 static int __noinline 4964 cache_fplookup_degenerate(struct cache_fpl *fpl) 4965 { 4966 struct componentname *cnp; 4967 struct vnode *dvp; 4968 enum vgetstate dvs; 4969 int error, lkflags; 4970 #ifdef INVARIANTS 4971 char *cp; 4972 #endif 4973 4974 fpl->tvp = fpl->dvp; 4975 fpl->tvp_seqc = fpl->dvp_seqc; 4976 4977 cnp = fpl->cnp; 4978 dvp = fpl->dvp; 4979 4980 #ifdef INVARIANTS 4981 for (cp = cnp->cn_pnbuf; *cp != '\0'; cp++) { 4982 KASSERT(*cp == '/', 4983 ("%s: encountered non-slash; string [%s]\n", __func__, 4984 cnp->cn_pnbuf)); 4985 } 4986 #endif 4987 4988 if (__predict_false(cnp->cn_nameiop != LOOKUP)) { 4989 cache_fpl_smr_exit(fpl); 4990 return (cache_fpl_handled_error(fpl, EISDIR)); 4991 } 4992 4993 if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0) { 4994 return (cache_fplookup_final_withparent(fpl)); 4995 } 4996 4997 dvs = vget_prep_smr(dvp); 4998 cache_fpl_smr_exit(fpl); 4999 if (__predict_false(dvs == VGET_NONE)) { 5000 return (cache_fpl_aborted(fpl)); 5001 } 5002 5003 if ((cnp->cn_flags & LOCKLEAF) != 0) { 5004 lkflags = LK_SHARED; 5005 if ((cnp->cn_flags & LOCKSHARED) == 0) 5006 lkflags = LK_EXCLUSIVE; 5007 error = vget_finish(dvp, lkflags, dvs); 5008 if (__predict_false(error != 0)) { 5009 return (cache_fpl_aborted(fpl)); 5010 } 5011 } else { 5012 vget_finish_ref(dvp, dvs); 5013 } 5014 return (cache_fpl_handled(fpl)); 5015 } 5016 5017 static int __noinline 5018 cache_fplookup_emptypath(struct cache_fpl *fpl) 5019 { 5020 struct nameidata *ndp; 5021 struct componentname *cnp; 5022 enum vgetstate tvs; 5023 struct vnode *tvp; 5024 int error, lkflags; 5025 5026 fpl->tvp = fpl->dvp; 5027 fpl->tvp_seqc = fpl->dvp_seqc; 5028 5029 ndp = fpl->ndp; 5030 cnp = fpl->cnp; 5031 tvp = fpl->tvp; 5032 5033 MPASS(*cnp->cn_pnbuf == '\0'); 5034 5035 if (__predict_false((cnp->cn_flags & EMPTYPATH) == 0)) { 5036 cache_fpl_smr_exit(fpl); 5037 return (cache_fpl_handled_error(fpl, ENOENT)); 5038 } 5039 5040 MPASS((cnp->cn_flags & (LOCKPARENT | WANTPARENT)) == 0); 5041 5042 tvs = vget_prep_smr(tvp); 5043 cache_fpl_smr_exit(fpl); 5044 if (__predict_false(tvs == VGET_NONE)) { 5045 return (cache_fpl_aborted(fpl)); 5046 } 5047 5048 if ((cnp->cn_flags & LOCKLEAF) != 0) { 5049 lkflags = LK_SHARED; 5050 if ((cnp->cn_flags & LOCKSHARED) == 0) 5051 lkflags = LK_EXCLUSIVE; 5052 error = vget_finish(tvp, lkflags, tvs); 5053 if (__predict_false(error != 0)) { 5054 return (cache_fpl_aborted(fpl)); 5055 } 5056 } else { 5057 vget_finish_ref(tvp, tvs); 5058 } 5059 5060 ndp->ni_resflags |= NIRES_EMPTYPATH; 5061 return (cache_fpl_handled(fpl)); 5062 } 5063 5064 static int __noinline 5065 cache_fplookup_noentry(struct cache_fpl *fpl) 5066 { 5067 struct nameidata *ndp; 5068 struct componentname *cnp; 5069 enum vgetstate dvs; 5070 struct vnode *dvp, *tvp; 5071 seqc_t dvp_seqc; 5072 int error; 5073 5074 ndp = fpl->ndp; 5075 cnp = fpl->cnp; 5076 dvp = fpl->dvp; 5077 dvp_seqc = fpl->dvp_seqc; 5078 5079 MPASS((cnp->cn_flags & MAKEENTRY) == 0); 5080 MPASS((cnp->cn_flags & ISDOTDOT) == 0); 5081 if (cnp->cn_nameiop == LOOKUP) 5082 MPASS((cnp->cn_flags & NOCACHE) == 0); 5083 MPASS(!cache_fpl_isdotdot(cnp)); 5084 5085 /* 5086 * Hack: delayed name len checking. 5087 */ 5088 if (__predict_false(cnp->cn_namelen > NAME_MAX)) { 5089 cache_fpl_smr_exit(fpl); 5090 return (cache_fpl_handled_error(fpl, ENAMETOOLONG)); 5091 } 5092 5093 if (cnp->cn_nameptr[0] == '/') { 5094 return (cache_fplookup_skip_slashes(fpl)); 5095 } 5096 5097 if (cnp->cn_pnbuf[0] == '\0') { 5098 return (cache_fplookup_emptypath(fpl)); 5099 } 5100 5101 if (cnp->cn_nameptr[0] == '\0') { 5102 if (fpl->tvp == NULL) { 5103 return (cache_fplookup_degenerate(fpl)); 5104 } 5105 return (cache_fplookup_trailingslash(fpl)); 5106 } 5107 5108 if (cnp->cn_nameiop != LOOKUP) { 5109 fpl->tvp = NULL; 5110 return (cache_fplookup_modifying(fpl)); 5111 } 5112 5113 /* 5114 * Only try to fill in the component if it is the last one, 5115 * otherwise not only there may be several to handle but the 5116 * walk may be complicated. 5117 */ 5118 if (!cache_fpl_islastcn(ndp)) { 5119 return (cache_fpl_partial(fpl)); 5120 } 5121 5122 /* 5123 * Regular lookup nulifies the slash, which we don't do here. 5124 * Don't take chances with filesystem routines seeing it for 5125 * the last entry. 5126 */ 5127 if (cache_fpl_istrailingslash(fpl)) { 5128 return (cache_fpl_partial(fpl)); 5129 } 5130 5131 /* 5132 * Secure access to dvp; check cache_fplookup_partial_setup for 5133 * reasoning. 5134 */ 5135 dvs = vget_prep_smr(dvp); 5136 cache_fpl_smr_exit(fpl); 5137 if (__predict_false(dvs == VGET_NONE)) { 5138 return (cache_fpl_aborted(fpl)); 5139 } 5140 5141 vget_finish_ref(dvp, dvs); 5142 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 5143 vrele(dvp); 5144 return (cache_fpl_aborted(fpl)); 5145 } 5146 5147 error = vn_lock(dvp, LK_SHARED); 5148 if (__predict_false(error != 0)) { 5149 vrele(dvp); 5150 return (cache_fpl_aborted(fpl)); 5151 } 5152 5153 tvp = NULL; 5154 /* 5155 * TODO: provide variants which don't require locking either vnode. 5156 */ 5157 cnp->cn_flags |= ISLASTCN | MAKEENTRY; 5158 cnp->cn_lkflags = LK_SHARED; 5159 if ((cnp->cn_flags & LOCKSHARED) == 0) { 5160 cnp->cn_lkflags = LK_EXCLUSIVE; 5161 } 5162 error = VOP_LOOKUP(dvp, &tvp, cnp); 5163 switch (error) { 5164 case EJUSTRETURN: 5165 case 0: 5166 break; 5167 case ENOTDIR: 5168 case ENOENT: 5169 vput(dvp); 5170 return (cache_fpl_handled_error(fpl, error)); 5171 default: 5172 vput(dvp); 5173 return (cache_fpl_aborted(fpl)); 5174 } 5175 5176 fpl->tvp = tvp; 5177 5178 if (tvp == NULL) { 5179 MPASS(error == EJUSTRETURN); 5180 if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) { 5181 vput(dvp); 5182 } else if ((cnp->cn_flags & LOCKPARENT) == 0) { 5183 VOP_UNLOCK(dvp); 5184 } 5185 return (cache_fpl_handled(fpl)); 5186 } 5187 5188 if (tvp->v_type == VLNK) { 5189 if ((cnp->cn_flags & FOLLOW) != 0) { 5190 vput(dvp); 5191 vput(tvp); 5192 return (cache_fpl_aborted(fpl)); 5193 } 5194 } 5195 5196 if (__predict_false(cache_fplookup_is_mp(fpl))) { 5197 vput(dvp); 5198 vput(tvp); 5199 return (cache_fpl_aborted(fpl)); 5200 } 5201 5202 if ((cnp->cn_flags & LOCKLEAF) == 0) { 5203 VOP_UNLOCK(tvp); 5204 } 5205 5206 if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) { 5207 vput(dvp); 5208 } else if ((cnp->cn_flags & LOCKPARENT) == 0) { 5209 VOP_UNLOCK(dvp); 5210 } 5211 return (cache_fpl_handled(fpl)); 5212 } 5213 5214 static int __noinline 5215 cache_fplookup_dot(struct cache_fpl *fpl) 5216 { 5217 int error; 5218 5219 MPASS(!seqc_in_modify(fpl->dvp_seqc)); 5220 5221 if (__predict_false(fpl->dvp->v_type != VDIR)) { 5222 cache_fpl_smr_exit(fpl); 5223 return (cache_fpl_handled_error(fpl, ENOTDIR)); 5224 } 5225 5226 /* 5227 * Just re-assign the value. seqc will be checked later for the first 5228 * non-dot path component in line and/or before deciding to return the 5229 * vnode. 5230 */ 5231 fpl->tvp = fpl->dvp; 5232 fpl->tvp_seqc = fpl->dvp_seqc; 5233 5234 SDT_PROBE3(vfs, namecache, lookup, hit, fpl->dvp, ".", fpl->dvp); 5235 5236 error = 0; 5237 if (cache_fplookup_is_mp(fpl)) { 5238 error = cache_fplookup_cross_mount(fpl); 5239 } 5240 return (error); 5241 } 5242 5243 static int __noinline 5244 cache_fplookup_dotdot(struct cache_fpl *fpl) 5245 { 5246 struct nameidata *ndp; 5247 struct componentname *cnp; 5248 struct namecache *ncp; 5249 struct vnode *dvp; 5250 struct prison *pr; 5251 u_char nc_flag; 5252 5253 ndp = fpl->ndp; 5254 cnp = fpl->cnp; 5255 dvp = fpl->dvp; 5256 5257 MPASS(cache_fpl_isdotdot(cnp)); 5258 5259 /* 5260 * XXX this is racy the same way regular lookup is 5261 */ 5262 for (pr = cnp->cn_cred->cr_prison; pr != NULL; 5263 pr = pr->pr_parent) 5264 if (dvp == pr->pr_root) 5265 break; 5266 5267 if (dvp == ndp->ni_rootdir || 5268 dvp == ndp->ni_topdir || 5269 dvp == rootvnode || 5270 pr != NULL) { 5271 fpl->tvp = dvp; 5272 fpl->tvp_seqc = vn_seqc_read_any(dvp); 5273 if (seqc_in_modify(fpl->tvp_seqc)) { 5274 return (cache_fpl_aborted(fpl)); 5275 } 5276 return (0); 5277 } 5278 5279 if ((dvp->v_vflag & VV_ROOT) != 0) { 5280 /* 5281 * TODO 5282 * The opposite of climb mount is needed here. 5283 */ 5284 return (cache_fpl_partial(fpl)); 5285 } 5286 5287 if (__predict_false(dvp->v_type != VDIR)) { 5288 cache_fpl_smr_exit(fpl); 5289 return (cache_fpl_handled_error(fpl, ENOTDIR)); 5290 } 5291 5292 ncp = atomic_load_consume_ptr(&dvp->v_cache_dd); 5293 if (ncp == NULL) { 5294 return (cache_fpl_aborted(fpl)); 5295 } 5296 5297 nc_flag = atomic_load_char(&ncp->nc_flag); 5298 if ((nc_flag & NCF_ISDOTDOT) != 0) { 5299 if ((nc_flag & NCF_NEGATIVE) != 0) 5300 return (cache_fpl_aborted(fpl)); 5301 fpl->tvp = ncp->nc_vp; 5302 } else { 5303 fpl->tvp = ncp->nc_dvp; 5304 } 5305 5306 fpl->tvp_seqc = vn_seqc_read_any(fpl->tvp); 5307 if (seqc_in_modify(fpl->tvp_seqc)) { 5308 return (cache_fpl_partial(fpl)); 5309 } 5310 5311 /* 5312 * Acquire fence provided by vn_seqc_read_any above. 5313 */ 5314 if (__predict_false(atomic_load_ptr(&dvp->v_cache_dd) != ncp)) { 5315 return (cache_fpl_aborted(fpl)); 5316 } 5317 5318 if (!cache_ncp_canuse(ncp)) { 5319 return (cache_fpl_aborted(fpl)); 5320 } 5321 5322 return (0); 5323 } 5324 5325 static int __noinline 5326 cache_fplookup_neg(struct cache_fpl *fpl, struct namecache *ncp, uint32_t hash) 5327 { 5328 u_char nc_flag __diagused; 5329 bool neg_promote; 5330 5331 #ifdef INVARIANTS 5332 nc_flag = atomic_load_char(&ncp->nc_flag); 5333 MPASS((nc_flag & NCF_NEGATIVE) != 0); 5334 #endif 5335 /* 5336 * If they want to create an entry we need to replace this one. 5337 */ 5338 if (__predict_false(fpl->cnp->cn_nameiop != LOOKUP)) { 5339 fpl->tvp = NULL; 5340 return (cache_fplookup_modifying(fpl)); 5341 } 5342 neg_promote = cache_neg_hit_prep(ncp); 5343 if (!cache_fpl_neg_ncp_canuse(ncp)) { 5344 cache_neg_hit_abort(ncp); 5345 return (cache_fpl_partial(fpl)); 5346 } 5347 if (neg_promote) { 5348 return (cache_fplookup_negative_promote(fpl, ncp, hash)); 5349 } 5350 cache_neg_hit_finish(ncp); 5351 cache_fpl_smr_exit(fpl); 5352 return (cache_fpl_handled_error(fpl, ENOENT)); 5353 } 5354 5355 /* 5356 * Resolve a symlink. Called by filesystem-specific routines. 5357 * 5358 * Code flow is: 5359 * ... -> cache_fplookup_symlink -> VOP_FPLOOKUP_SYMLINK -> cache_symlink_resolve 5360 */ 5361 int 5362 cache_symlink_resolve(struct cache_fpl *fpl, const char *string, size_t len) 5363 { 5364 struct nameidata *ndp; 5365 struct componentname *cnp; 5366 size_t adjust; 5367 5368 ndp = fpl->ndp; 5369 cnp = fpl->cnp; 5370 5371 if (__predict_false(len == 0)) { 5372 return (ENOENT); 5373 } 5374 5375 if (__predict_false(len > MAXPATHLEN - 2)) { 5376 if (cache_fpl_istrailingslash(fpl)) { 5377 return (EAGAIN); 5378 } 5379 } 5380 5381 ndp->ni_pathlen = fpl->nulchar - cnp->cn_nameptr - cnp->cn_namelen + 1; 5382 #ifdef INVARIANTS 5383 if (ndp->ni_pathlen != fpl->debug.ni_pathlen) { 5384 panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n", 5385 __func__, ndp->ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar, 5386 cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf); 5387 } 5388 #endif 5389 5390 if (__predict_false(len + ndp->ni_pathlen > MAXPATHLEN)) { 5391 return (ENAMETOOLONG); 5392 } 5393 5394 if (__predict_false(ndp->ni_loopcnt++ >= MAXSYMLINKS)) { 5395 return (ELOOP); 5396 } 5397 5398 adjust = len; 5399 if (ndp->ni_pathlen > 1) { 5400 bcopy(ndp->ni_next, cnp->cn_pnbuf + len, ndp->ni_pathlen); 5401 } else { 5402 if (cache_fpl_istrailingslash(fpl)) { 5403 adjust = len + 1; 5404 cnp->cn_pnbuf[len] = '/'; 5405 cnp->cn_pnbuf[len + 1] = '\0'; 5406 } else { 5407 cnp->cn_pnbuf[len] = '\0'; 5408 } 5409 } 5410 bcopy(string, cnp->cn_pnbuf, len); 5411 5412 ndp->ni_pathlen += adjust; 5413 cache_fpl_pathlen_add(fpl, adjust); 5414 cnp->cn_nameptr = cnp->cn_pnbuf; 5415 fpl->nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1]; 5416 fpl->tvp = NULL; 5417 return (0); 5418 } 5419 5420 static int __noinline 5421 cache_fplookup_symlink(struct cache_fpl *fpl) 5422 { 5423 struct mount *mp; 5424 struct nameidata *ndp; 5425 struct componentname *cnp; 5426 struct vnode *dvp, *tvp; 5427 struct pwd *pwd; 5428 int error; 5429 5430 ndp = fpl->ndp; 5431 cnp = fpl->cnp; 5432 dvp = fpl->dvp; 5433 tvp = fpl->tvp; 5434 pwd = *(fpl->pwd); 5435 5436 if (cache_fpl_islastcn(ndp)) { 5437 if ((cnp->cn_flags & FOLLOW) == 0) { 5438 return (cache_fplookup_final(fpl)); 5439 } 5440 } 5441 5442 mp = atomic_load_ptr(&dvp->v_mount); 5443 if (__predict_false(mp == NULL)) { 5444 return (cache_fpl_aborted(fpl)); 5445 } 5446 5447 /* 5448 * Note this check races against setting the flag just like regular 5449 * lookup. 5450 */ 5451 if (__predict_false((mp->mnt_flag & MNT_NOSYMFOLLOW) != 0)) { 5452 cache_fpl_smr_exit(fpl); 5453 return (cache_fpl_handled_error(fpl, EACCES)); 5454 } 5455 5456 error = VOP_FPLOOKUP_SYMLINK(tvp, fpl); 5457 if (__predict_false(error != 0)) { 5458 switch (error) { 5459 case EAGAIN: 5460 return (cache_fpl_partial(fpl)); 5461 case ENOENT: 5462 case ENAMETOOLONG: 5463 case ELOOP: 5464 cache_fpl_smr_exit(fpl); 5465 return (cache_fpl_handled_error(fpl, error)); 5466 default: 5467 return (cache_fpl_aborted(fpl)); 5468 } 5469 } 5470 5471 if (*(cnp->cn_nameptr) == '/') { 5472 fpl->dvp = cache_fpl_handle_root(fpl); 5473 fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp); 5474 if (seqc_in_modify(fpl->dvp_seqc)) { 5475 return (cache_fpl_aborted(fpl)); 5476 } 5477 /* 5478 * The main loop assumes that ->dvp points to a vnode belonging 5479 * to a filesystem which can do lockless lookup, but the absolute 5480 * symlink can be wandering off to one which does not. 5481 */ 5482 mp = atomic_load_ptr(&fpl->dvp->v_mount); 5483 if (__predict_false(mp == NULL)) { 5484 return (cache_fpl_aborted(fpl)); 5485 } 5486 if (!cache_fplookup_mp_supported(mp)) { 5487 cache_fpl_checkpoint(fpl); 5488 return (cache_fpl_partial(fpl)); 5489 } 5490 if (__predict_false(pwd->pwd_adir != pwd->pwd_rdir)) { 5491 return (cache_fpl_aborted(fpl)); 5492 } 5493 } 5494 return (0); 5495 } 5496 5497 static int 5498 cache_fplookup_next(struct cache_fpl *fpl) 5499 { 5500 struct componentname *cnp; 5501 struct namecache *ncp; 5502 struct vnode *dvp, *tvp; 5503 u_char nc_flag; 5504 uint32_t hash; 5505 int error; 5506 5507 cnp = fpl->cnp; 5508 dvp = fpl->dvp; 5509 hash = fpl->hash; 5510 5511 if (__predict_false(cnp->cn_nameptr[0] == '.')) { 5512 if (cnp->cn_namelen == 1) { 5513 return (cache_fplookup_dot(fpl)); 5514 } 5515 if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') { 5516 return (cache_fplookup_dotdot(fpl)); 5517 } 5518 } 5519 5520 MPASS(!cache_fpl_isdotdot(cnp)); 5521 5522 CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 5523 if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && 5524 !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) 5525 break; 5526 } 5527 5528 if (__predict_false(ncp == NULL)) { 5529 return (cache_fplookup_noentry(fpl)); 5530 } 5531 5532 tvp = atomic_load_ptr(&ncp->nc_vp); 5533 nc_flag = atomic_load_char(&ncp->nc_flag); 5534 if ((nc_flag & NCF_NEGATIVE) != 0) { 5535 return (cache_fplookup_neg(fpl, ncp, hash)); 5536 } 5537 5538 if (!cache_ncp_canuse(ncp)) { 5539 return (cache_fpl_partial(fpl)); 5540 } 5541 5542 fpl->tvp = tvp; 5543 fpl->tvp_seqc = vn_seqc_read_any(tvp); 5544 if (seqc_in_modify(fpl->tvp_seqc)) { 5545 return (cache_fpl_partial(fpl)); 5546 } 5547 5548 counter_u64_add(numposhits, 1); 5549 SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, tvp); 5550 5551 error = 0; 5552 if (cache_fplookup_is_mp(fpl)) { 5553 error = cache_fplookup_cross_mount(fpl); 5554 } 5555 return (error); 5556 } 5557 5558 static bool 5559 cache_fplookup_mp_supported(struct mount *mp) 5560 { 5561 5562 MPASS(mp != NULL); 5563 if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0) 5564 return (false); 5565 return (true); 5566 } 5567 5568 /* 5569 * Walk up the mount stack (if any). 5570 * 5571 * Correctness is provided in the following ways: 5572 * - all vnodes are protected from freeing with SMR 5573 * - struct mount objects are type stable making them always safe to access 5574 * - stability of the particular mount is provided by busying it 5575 * - relationship between the vnode which is mounted on and the mount is 5576 * verified with the vnode sequence counter after busying 5577 * - association between root vnode of the mount and the mount is protected 5578 * by busy 5579 * 5580 * From that point on we can read the sequence counter of the root vnode 5581 * and get the next mount on the stack (if any) using the same protection. 5582 * 5583 * By the end of successful walk we are guaranteed the reached state was 5584 * indeed present at least at some point which matches the regular lookup. 5585 */ 5586 static int __noinline 5587 cache_fplookup_climb_mount(struct cache_fpl *fpl) 5588 { 5589 struct mount *mp, *prev_mp; 5590 struct mount_pcpu *mpcpu, *prev_mpcpu; 5591 struct vnode *vp; 5592 seqc_t vp_seqc; 5593 5594 vp = fpl->tvp; 5595 vp_seqc = fpl->tvp_seqc; 5596 5597 VNPASS(vp->v_type == VDIR || vp->v_type == VREG || vp->v_type == VBAD, vp); 5598 mp = atomic_load_ptr(&vp->v_mountedhere); 5599 if (__predict_false(mp == NULL)) { 5600 return (0); 5601 } 5602 5603 prev_mp = NULL; 5604 for (;;) { 5605 if (!vfs_op_thread_enter_crit(mp, mpcpu)) { 5606 if (prev_mp != NULL) 5607 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu); 5608 return (cache_fpl_partial(fpl)); 5609 } 5610 if (prev_mp != NULL) 5611 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu); 5612 if (!vn_seqc_consistent(vp, vp_seqc)) { 5613 vfs_op_thread_exit_crit(mp, mpcpu); 5614 return (cache_fpl_partial(fpl)); 5615 } 5616 if (!cache_fplookup_mp_supported(mp)) { 5617 vfs_op_thread_exit_crit(mp, mpcpu); 5618 return (cache_fpl_partial(fpl)); 5619 } 5620 vp = atomic_load_ptr(&mp->mnt_rootvnode); 5621 if (vp == NULL) { 5622 vfs_op_thread_exit_crit(mp, mpcpu); 5623 return (cache_fpl_partial(fpl)); 5624 } 5625 vp_seqc = vn_seqc_read_any(vp); 5626 if (seqc_in_modify(vp_seqc)) { 5627 vfs_op_thread_exit_crit(mp, mpcpu); 5628 return (cache_fpl_partial(fpl)); 5629 } 5630 prev_mp = mp; 5631 prev_mpcpu = mpcpu; 5632 mp = atomic_load_ptr(&vp->v_mountedhere); 5633 if (mp == NULL) 5634 break; 5635 } 5636 5637 vfs_op_thread_exit_crit(prev_mp, prev_mpcpu); 5638 fpl->tvp = vp; 5639 fpl->tvp_seqc = vp_seqc; 5640 return (0); 5641 } 5642 5643 static int __noinline 5644 cache_fplookup_cross_mount(struct cache_fpl *fpl) 5645 { 5646 struct mount *mp; 5647 struct mount_pcpu *mpcpu; 5648 struct vnode *vp; 5649 seqc_t vp_seqc; 5650 5651 vp = fpl->tvp; 5652 vp_seqc = fpl->tvp_seqc; 5653 5654 VNPASS(vp->v_type == VDIR || vp->v_type == VREG || vp->v_type == VBAD, vp); 5655 mp = atomic_load_ptr(&vp->v_mountedhere); 5656 if (__predict_false(mp == NULL)) { 5657 return (0); 5658 } 5659 5660 if (!vfs_op_thread_enter_crit(mp, mpcpu)) { 5661 return (cache_fpl_partial(fpl)); 5662 } 5663 if (!vn_seqc_consistent(vp, vp_seqc)) { 5664 vfs_op_thread_exit_crit(mp, mpcpu); 5665 return (cache_fpl_partial(fpl)); 5666 } 5667 if (!cache_fplookup_mp_supported(mp)) { 5668 vfs_op_thread_exit_crit(mp, mpcpu); 5669 return (cache_fpl_partial(fpl)); 5670 } 5671 vp = atomic_load_ptr(&mp->mnt_rootvnode); 5672 if (__predict_false(vp == NULL)) { 5673 vfs_op_thread_exit_crit(mp, mpcpu); 5674 return (cache_fpl_partial(fpl)); 5675 } 5676 vp_seqc = vn_seqc_read_any(vp); 5677 vfs_op_thread_exit_crit(mp, mpcpu); 5678 if (seqc_in_modify(vp_seqc)) { 5679 return (cache_fpl_partial(fpl)); 5680 } 5681 mp = atomic_load_ptr(&vp->v_mountedhere); 5682 if (__predict_false(mp != NULL)) { 5683 /* 5684 * There are possibly more mount points on top. 5685 * Normally this does not happen so for simplicity just start 5686 * over. 5687 */ 5688 return (cache_fplookup_climb_mount(fpl)); 5689 } 5690 5691 fpl->tvp = vp; 5692 fpl->tvp_seqc = vp_seqc; 5693 return (0); 5694 } 5695 5696 /* 5697 * Check if a vnode is mounted on. 5698 */ 5699 static bool 5700 cache_fplookup_is_mp(struct cache_fpl *fpl) 5701 { 5702 struct vnode *vp; 5703 5704 vp = fpl->tvp; 5705 return ((vn_irflag_read(vp) & VIRF_MOUNTPOINT) != 0); 5706 } 5707 5708 /* 5709 * Parse the path. 5710 * 5711 * The code was originally copy-pasted from regular lookup and despite 5712 * clean ups leaves performance on the table. Any modifications here 5713 * must take into account that in case off fallback the resulting 5714 * nameidata state has to be compatible with the original. 5715 */ 5716 5717 /* 5718 * Debug ni_pathlen tracking. 5719 */ 5720 #ifdef INVARIANTS 5721 static void 5722 cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n) 5723 { 5724 5725 fpl->debug.ni_pathlen += n; 5726 KASSERT(fpl->debug.ni_pathlen <= PATH_MAX, 5727 ("%s: pathlen overflow to %zd\n", __func__, fpl->debug.ni_pathlen)); 5728 } 5729 5730 static void 5731 cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n) 5732 { 5733 5734 fpl->debug.ni_pathlen -= n; 5735 KASSERT(fpl->debug.ni_pathlen <= PATH_MAX, 5736 ("%s: pathlen underflow to %zd\n", __func__, fpl->debug.ni_pathlen)); 5737 } 5738 5739 static void 5740 cache_fpl_pathlen_inc(struct cache_fpl *fpl) 5741 { 5742 5743 cache_fpl_pathlen_add(fpl, 1); 5744 } 5745 5746 static void 5747 cache_fpl_pathlen_dec(struct cache_fpl *fpl) 5748 { 5749 5750 cache_fpl_pathlen_sub(fpl, 1); 5751 } 5752 #else 5753 static void 5754 cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n) 5755 { 5756 } 5757 5758 static void 5759 cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n) 5760 { 5761 } 5762 5763 static void 5764 cache_fpl_pathlen_inc(struct cache_fpl *fpl) 5765 { 5766 } 5767 5768 static void 5769 cache_fpl_pathlen_dec(struct cache_fpl *fpl) 5770 { 5771 } 5772 #endif 5773 5774 static void 5775 cache_fplookup_parse(struct cache_fpl *fpl) 5776 { 5777 struct nameidata *ndp; 5778 struct componentname *cnp; 5779 struct vnode *dvp; 5780 char *cp; 5781 uint32_t hash; 5782 5783 ndp = fpl->ndp; 5784 cnp = fpl->cnp; 5785 dvp = fpl->dvp; 5786 5787 /* 5788 * Find the end of this path component, it is either / or nul. 5789 * 5790 * Store / as a temporary sentinel so that we only have one character 5791 * to test for. Pathnames tend to be short so this should not be 5792 * resulting in cache misses. 5793 * 5794 * TODO: fix this to be word-sized. 5795 */ 5796 MPASS(&cnp->cn_nameptr[fpl->debug.ni_pathlen - 1] >= cnp->cn_pnbuf); 5797 KASSERT(&cnp->cn_nameptr[fpl->debug.ni_pathlen - 1] == fpl->nulchar, 5798 ("%s: mismatch between pathlen (%zu) and nulchar (%p != %p), string [%s]\n", 5799 __func__, fpl->debug.ni_pathlen, &cnp->cn_nameptr[fpl->debug.ni_pathlen - 1], 5800 fpl->nulchar, cnp->cn_pnbuf)); 5801 KASSERT(*fpl->nulchar == '\0', 5802 ("%s: expected nul at %p; string [%s]\n", __func__, fpl->nulchar, 5803 cnp->cn_pnbuf)); 5804 hash = cache_get_hash_iter_start(dvp); 5805 *fpl->nulchar = '/'; 5806 for (cp = cnp->cn_nameptr; *cp != '/'; cp++) { 5807 KASSERT(*cp != '\0', 5808 ("%s: encountered unexpected nul; string [%s]\n", __func__, 5809 cnp->cn_nameptr)); 5810 hash = cache_get_hash_iter(*cp, hash); 5811 continue; 5812 } 5813 *fpl->nulchar = '\0'; 5814 fpl->hash = cache_get_hash_iter_finish(hash); 5815 5816 cnp->cn_namelen = cp - cnp->cn_nameptr; 5817 cache_fpl_pathlen_sub(fpl, cnp->cn_namelen); 5818 5819 #ifdef INVARIANTS 5820 /* 5821 * cache_get_hash only accepts lengths up to NAME_MAX. This is fine since 5822 * we are going to fail this lookup with ENAMETOOLONG (see below). 5823 */ 5824 if (cnp->cn_namelen <= NAME_MAX) { 5825 if (fpl->hash != cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp)) { 5826 panic("%s: mismatched hash for [%s] len %ld", __func__, 5827 cnp->cn_nameptr, cnp->cn_namelen); 5828 } 5829 } 5830 #endif 5831 5832 /* 5833 * Hack: we have to check if the found path component's length exceeds 5834 * NAME_MAX. However, the condition is very rarely true and check can 5835 * be elided in the common case -- if an entry was found in the cache, 5836 * then it could not have been too long to begin with. 5837 */ 5838 ndp->ni_next = cp; 5839 } 5840 5841 static void 5842 cache_fplookup_parse_advance(struct cache_fpl *fpl) 5843 { 5844 struct nameidata *ndp; 5845 struct componentname *cnp; 5846 5847 ndp = fpl->ndp; 5848 cnp = fpl->cnp; 5849 5850 cnp->cn_nameptr = ndp->ni_next; 5851 KASSERT(*(cnp->cn_nameptr) == '/', 5852 ("%s: should have seen slash at %p ; buf %p [%s]\n", __func__, 5853 cnp->cn_nameptr, cnp->cn_pnbuf, cnp->cn_pnbuf)); 5854 cnp->cn_nameptr++; 5855 cache_fpl_pathlen_dec(fpl); 5856 } 5857 5858 /* 5859 * Skip spurious slashes in a pathname (e.g., "foo///bar") and retry. 5860 * 5861 * Lockless lookup tries to elide checking for spurious slashes and should they 5862 * be present is guaranteed to fail to find an entry. In this case the caller 5863 * must check if the name starts with a slash and call this routine. It is 5864 * going to fast forward across the spurious slashes and set the state up for 5865 * retry. 5866 */ 5867 static int __noinline 5868 cache_fplookup_skip_slashes(struct cache_fpl *fpl) 5869 { 5870 struct nameidata *ndp; 5871 struct componentname *cnp; 5872 5873 ndp = fpl->ndp; 5874 cnp = fpl->cnp; 5875 5876 MPASS(*(cnp->cn_nameptr) == '/'); 5877 do { 5878 cnp->cn_nameptr++; 5879 cache_fpl_pathlen_dec(fpl); 5880 } while (*(cnp->cn_nameptr) == '/'); 5881 5882 /* 5883 * Go back to one slash so that cache_fplookup_parse_advance has 5884 * something to skip. 5885 */ 5886 cnp->cn_nameptr--; 5887 cache_fpl_pathlen_inc(fpl); 5888 5889 /* 5890 * cache_fplookup_parse_advance starts from ndp->ni_next 5891 */ 5892 ndp->ni_next = cnp->cn_nameptr; 5893 5894 /* 5895 * See cache_fplookup_dot. 5896 */ 5897 fpl->tvp = fpl->dvp; 5898 fpl->tvp_seqc = fpl->dvp_seqc; 5899 5900 return (0); 5901 } 5902 5903 /* 5904 * Handle trailing slashes (e.g., "foo/"). 5905 * 5906 * If a trailing slash is found the terminal vnode must be a directory. 5907 * Regular lookup shortens the path by nulifying the first trailing slash and 5908 * sets the TRAILINGSLASH flag to denote this took place. There are several 5909 * checks on it performed later. 5910 * 5911 * Similarly to spurious slashes, lockless lookup handles this in a speculative 5912 * manner relying on an invariant that a non-directory vnode will get a miss. 5913 * In this case cn_nameptr[0] == '\0' and cn_namelen == 0. 5914 * 5915 * Thus for a path like "foo/bar/" the code unwinds the state back to "bar/" 5916 * and denotes this is the last path component, which avoids looping back. 5917 * 5918 * Only plain lookups are supported for now to restrict corner cases to handle. 5919 */ 5920 static int __noinline 5921 cache_fplookup_trailingslash(struct cache_fpl *fpl) 5922 { 5923 #ifdef INVARIANTS 5924 size_t ni_pathlen; 5925 #endif 5926 struct nameidata *ndp; 5927 struct componentname *cnp; 5928 struct namecache *ncp; 5929 struct vnode *tvp; 5930 char *cn_nameptr_orig, *cn_nameptr_slash; 5931 seqc_t tvp_seqc; 5932 u_char nc_flag; 5933 5934 ndp = fpl->ndp; 5935 cnp = fpl->cnp; 5936 tvp = fpl->tvp; 5937 tvp_seqc = fpl->tvp_seqc; 5938 5939 MPASS(fpl->dvp == fpl->tvp); 5940 KASSERT(cache_fpl_istrailingslash(fpl), 5941 ("%s: expected trailing slash at %p; string [%s]\n", __func__, fpl->nulchar - 1, 5942 cnp->cn_pnbuf)); 5943 KASSERT(cnp->cn_nameptr[0] == '\0', 5944 ("%s: expected nul char at %p; string [%s]\n", __func__, &cnp->cn_nameptr[0], 5945 cnp->cn_pnbuf)); 5946 KASSERT(cnp->cn_namelen == 0, 5947 ("%s: namelen 0 but got %ld; string [%s]\n", __func__, cnp->cn_namelen, 5948 cnp->cn_pnbuf)); 5949 MPASS(cnp->cn_nameptr > cnp->cn_pnbuf); 5950 5951 if (cnp->cn_nameiop != LOOKUP) { 5952 return (cache_fpl_aborted(fpl)); 5953 } 5954 5955 if (__predict_false(tvp->v_type != VDIR)) { 5956 if (!vn_seqc_consistent(tvp, tvp_seqc)) { 5957 return (cache_fpl_aborted(fpl)); 5958 } 5959 cache_fpl_smr_exit(fpl); 5960 return (cache_fpl_handled_error(fpl, ENOTDIR)); 5961 } 5962 5963 /* 5964 * Denote the last component. 5965 */ 5966 ndp->ni_next = &cnp->cn_nameptr[0]; 5967 MPASS(cache_fpl_islastcn(ndp)); 5968 5969 /* 5970 * Unwind trailing slashes. 5971 */ 5972 cn_nameptr_orig = cnp->cn_nameptr; 5973 while (cnp->cn_nameptr >= cnp->cn_pnbuf) { 5974 cnp->cn_nameptr--; 5975 if (cnp->cn_nameptr[0] != '/') { 5976 break; 5977 } 5978 } 5979 5980 /* 5981 * Unwind to the beginning of the path component. 5982 * 5983 * Note the path may or may not have started with a slash. 5984 */ 5985 cn_nameptr_slash = cnp->cn_nameptr; 5986 while (cnp->cn_nameptr > cnp->cn_pnbuf) { 5987 cnp->cn_nameptr--; 5988 if (cnp->cn_nameptr[0] == '/') { 5989 break; 5990 } 5991 } 5992 if (cnp->cn_nameptr[0] == '/') { 5993 cnp->cn_nameptr++; 5994 } 5995 5996 cnp->cn_namelen = cn_nameptr_slash - cnp->cn_nameptr + 1; 5997 cache_fpl_pathlen_add(fpl, cn_nameptr_orig - cnp->cn_nameptr); 5998 cache_fpl_checkpoint(fpl); 5999 6000 #ifdef INVARIANTS 6001 ni_pathlen = fpl->nulchar - cnp->cn_nameptr + 1; 6002 if (ni_pathlen != fpl->debug.ni_pathlen) { 6003 panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n", 6004 __func__, ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar, 6005 cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf); 6006 } 6007 #endif 6008 6009 /* 6010 * If this was a "./" lookup the parent directory is already correct. 6011 */ 6012 if (cnp->cn_nameptr[0] == '.' && cnp->cn_namelen == 1) { 6013 return (0); 6014 } 6015 6016 /* 6017 * Otherwise we need to look it up. 6018 */ 6019 tvp = fpl->tvp; 6020 ncp = atomic_load_consume_ptr(&tvp->v_cache_dd); 6021 if (__predict_false(ncp == NULL)) { 6022 return (cache_fpl_aborted(fpl)); 6023 } 6024 nc_flag = atomic_load_char(&ncp->nc_flag); 6025 if ((nc_flag & NCF_ISDOTDOT) != 0) { 6026 return (cache_fpl_aborted(fpl)); 6027 } 6028 fpl->dvp = ncp->nc_dvp; 6029 fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp); 6030 if (seqc_in_modify(fpl->dvp_seqc)) { 6031 return (cache_fpl_aborted(fpl)); 6032 } 6033 return (0); 6034 } 6035 6036 /* 6037 * See the API contract for VOP_FPLOOKUP_VEXEC. 6038 */ 6039 static int __noinline 6040 cache_fplookup_failed_vexec(struct cache_fpl *fpl, int error) 6041 { 6042 struct componentname *cnp; 6043 struct vnode *dvp; 6044 seqc_t dvp_seqc; 6045 6046 cnp = fpl->cnp; 6047 dvp = fpl->dvp; 6048 dvp_seqc = fpl->dvp_seqc; 6049 6050 /* 6051 * Hack: delayed empty path checking. 6052 */ 6053 if (cnp->cn_pnbuf[0] == '\0') { 6054 return (cache_fplookup_emptypath(fpl)); 6055 } 6056 6057 /* 6058 * TODO: Due to ignoring trailing slashes lookup will perform a 6059 * permission check on the last dir when it should not be doing it. It 6060 * may fail, but said failure should be ignored. It is possible to fix 6061 * it up fully without resorting to regular lookup, but for now just 6062 * abort. 6063 */ 6064 if (cache_fpl_istrailingslash(fpl)) { 6065 return (cache_fpl_aborted(fpl)); 6066 } 6067 6068 /* 6069 * Hack: delayed degenerate path checking. 6070 */ 6071 if (cnp->cn_nameptr[0] == '\0' && fpl->tvp == NULL) { 6072 return (cache_fplookup_degenerate(fpl)); 6073 } 6074 6075 /* 6076 * Hack: delayed name len checking. 6077 */ 6078 if (__predict_false(cnp->cn_namelen > NAME_MAX)) { 6079 cache_fpl_smr_exit(fpl); 6080 return (cache_fpl_handled_error(fpl, ENAMETOOLONG)); 6081 } 6082 6083 /* 6084 * Hack: they may be looking up foo/bar, where foo is not a directory. 6085 * In such a case we need to return ENOTDIR, but we may happen to get 6086 * here with a different error. 6087 */ 6088 if (dvp->v_type != VDIR) { 6089 error = ENOTDIR; 6090 } 6091 6092 /* 6093 * Hack: handle O_SEARCH. 6094 * 6095 * Open Group Base Specifications Issue 7, 2018 edition states: 6096 * <quote> 6097 * If the access mode of the open file description associated with the 6098 * file descriptor is not O_SEARCH, the function shall check whether 6099 * directory searches are permitted using the current permissions of 6100 * the directory underlying the file descriptor. If the access mode is 6101 * O_SEARCH, the function shall not perform the check. 6102 * </quote> 6103 * 6104 * Regular lookup tests for the NOEXECCHECK flag for every path 6105 * component to decide whether to do the permission check. However, 6106 * since most lookups never have the flag (and when they do it is only 6107 * present for the first path component), lockless lookup only acts on 6108 * it if there is a permission problem. Here the flag is represented 6109 * with a boolean so that we don't have to clear it on the way out. 6110 * 6111 * For simplicity this always aborts. 6112 * TODO: check if this is the first lookup and ignore the permission 6113 * problem. Note the flag has to survive fallback (if it happens to be 6114 * performed). 6115 */ 6116 if (fpl->fsearch) { 6117 return (cache_fpl_aborted(fpl)); 6118 } 6119 6120 switch (error) { 6121 case EAGAIN: 6122 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 6123 error = cache_fpl_aborted(fpl); 6124 } else { 6125 cache_fpl_partial(fpl); 6126 } 6127 break; 6128 default: 6129 if (!vn_seqc_consistent(dvp, dvp_seqc)) { 6130 error = cache_fpl_aborted(fpl); 6131 } else { 6132 cache_fpl_smr_exit(fpl); 6133 cache_fpl_handled_error(fpl, error); 6134 } 6135 break; 6136 } 6137 return (error); 6138 } 6139 6140 static int 6141 cache_fplookup_impl(struct vnode *dvp, struct cache_fpl *fpl) 6142 { 6143 struct nameidata *ndp; 6144 struct componentname *cnp; 6145 struct mount *mp; 6146 int error; 6147 6148 ndp = fpl->ndp; 6149 cnp = fpl->cnp; 6150 6151 cache_fpl_checkpoint(fpl); 6152 6153 /* 6154 * The vnode at hand is almost always stable, skip checking for it. 6155 * Worst case this postpones the check towards the end of the iteration 6156 * of the main loop. 6157 */ 6158 fpl->dvp = dvp; 6159 fpl->dvp_seqc = vn_seqc_read_notmodify(fpl->dvp); 6160 6161 mp = atomic_load_ptr(&dvp->v_mount); 6162 if (__predict_false(mp == NULL || !cache_fplookup_mp_supported(mp))) { 6163 return (cache_fpl_aborted(fpl)); 6164 } 6165 6166 MPASS(fpl->tvp == NULL); 6167 6168 for (;;) { 6169 cache_fplookup_parse(fpl); 6170 6171 error = VOP_FPLOOKUP_VEXEC(fpl->dvp, cnp->cn_cred); 6172 if (__predict_false(error != 0)) { 6173 error = cache_fplookup_failed_vexec(fpl, error); 6174 break; 6175 } 6176 6177 error = cache_fplookup_next(fpl); 6178 if (__predict_false(cache_fpl_terminated(fpl))) { 6179 break; 6180 } 6181 6182 VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp); 6183 6184 if (fpl->tvp->v_type == VLNK) { 6185 error = cache_fplookup_symlink(fpl); 6186 if (cache_fpl_terminated(fpl)) { 6187 break; 6188 } 6189 } else { 6190 if (cache_fpl_islastcn(ndp)) { 6191 error = cache_fplookup_final(fpl); 6192 break; 6193 } 6194 6195 if (!vn_seqc_consistent(fpl->dvp, fpl->dvp_seqc)) { 6196 error = cache_fpl_aborted(fpl); 6197 break; 6198 } 6199 6200 fpl->dvp = fpl->tvp; 6201 fpl->dvp_seqc = fpl->tvp_seqc; 6202 cache_fplookup_parse_advance(fpl); 6203 } 6204 6205 cache_fpl_checkpoint(fpl); 6206 } 6207 6208 return (error); 6209 } 6210 6211 /* 6212 * Fast path lookup protected with SMR and sequence counters. 6213 * 6214 * Note: all VOP_FPLOOKUP_VEXEC routines have a comment referencing this one. 6215 * 6216 * Filesystems can opt in by setting the MNTK_FPLOOKUP flag and meeting criteria 6217 * outlined below. 6218 * 6219 * Traditional vnode lookup conceptually looks like this: 6220 * 6221 * vn_lock(current); 6222 * for (;;) { 6223 * next = find(); 6224 * vn_lock(next); 6225 * vn_unlock(current); 6226 * current = next; 6227 * if (last) 6228 * break; 6229 * } 6230 * return (current); 6231 * 6232 * Each jump to the next vnode is safe memory-wise and atomic with respect to 6233 * any modifications thanks to holding respective locks. 6234 * 6235 * The same guarantee can be provided with a combination of safe memory 6236 * reclamation and sequence counters instead. If all operations which affect 6237 * the relationship between the current vnode and the one we are looking for 6238 * also modify the counter, we can verify whether all the conditions held as 6239 * we made the jump. This includes things like permissions, mount points etc. 6240 * Counter modification is provided by enclosing relevant places in 6241 * vn_seqc_write_begin()/end() calls. 6242 * 6243 * Thus this translates to: 6244 * 6245 * vfs_smr_enter(); 6246 * dvp_seqc = seqc_read_any(dvp); 6247 * if (seqc_in_modify(dvp_seqc)) // someone is altering the vnode 6248 * abort(); 6249 * for (;;) { 6250 * tvp = find(); 6251 * tvp_seqc = seqc_read_any(tvp); 6252 * if (seqc_in_modify(tvp_seqc)) // someone is altering the target vnode 6253 * abort(); 6254 * if (!seqc_consistent(dvp, dvp_seqc) // someone is altering the vnode 6255 * abort(); 6256 * dvp = tvp; // we know nothing of importance has changed 6257 * dvp_seqc = tvp_seqc; // store the counter for the tvp iteration 6258 * if (last) 6259 * break; 6260 * } 6261 * vget(); // secure the vnode 6262 * if (!seqc_consistent(tvp, tvp_seqc) // final check 6263 * abort(); 6264 * // at this point we know nothing has changed for any parent<->child pair 6265 * // as they were crossed during the lookup, meaning we matched the guarantee 6266 * // of the locked variant 6267 * return (tvp); 6268 * 6269 * The API contract for VOP_FPLOOKUP_VEXEC routines is as follows: 6270 * - they are called while within vfs_smr protection which they must never exit 6271 * - EAGAIN can be returned to denote checking could not be performed, it is 6272 * always valid to return it 6273 * - if the sequence counter has not changed the result must be valid 6274 * - if the sequence counter has changed both false positives and false negatives 6275 * are permitted (since the result will be rejected later) 6276 * - for simple cases of unix permission checks vaccess_vexec_smr can be used 6277 * 6278 * Caveats to watch out for: 6279 * - vnodes are passed unlocked and unreferenced with nothing stopping 6280 * VOP_RECLAIM, in turn meaning that ->v_data can become NULL. It is advised 6281 * to use atomic_load_ptr to fetch it. 6282 * - the aforementioned object can also get freed, meaning absent other means it 6283 * should be protected with vfs_smr 6284 * - either safely checking permissions as they are modified or guaranteeing 6285 * their stability is left to the routine 6286 */ 6287 int 6288 cache_fplookup(struct nameidata *ndp, enum cache_fpl_status *status, 6289 struct pwd **pwdp) 6290 { 6291 struct cache_fpl fpl; 6292 struct pwd *pwd; 6293 struct vnode *dvp; 6294 struct componentname *cnp; 6295 int error; 6296 6297 fpl.status = CACHE_FPL_STATUS_UNSET; 6298 fpl.in_smr = false; 6299 fpl.ndp = ndp; 6300 fpl.cnp = cnp = &ndp->ni_cnd; 6301 MPASS(ndp->ni_lcf == 0); 6302 KASSERT ((cnp->cn_flags & CACHE_FPL_INTERNAL_CN_FLAGS) == 0, 6303 ("%s: internal flags found in cn_flags %" PRIx64, __func__, 6304 cnp->cn_flags)); 6305 MPASS(cnp->cn_nameptr == cnp->cn_pnbuf); 6306 MPASS(ndp->ni_resflags == 0); 6307 6308 if (__predict_false(!cache_can_fplookup(&fpl))) { 6309 *status = fpl.status; 6310 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status); 6311 return (EOPNOTSUPP); 6312 } 6313 6314 cache_fpl_checkpoint_outer(&fpl); 6315 6316 cache_fpl_smr_enter_initial(&fpl); 6317 #ifdef INVARIANTS 6318 fpl.debug.ni_pathlen = ndp->ni_pathlen; 6319 #endif 6320 fpl.nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1]; 6321 fpl.fsearch = false; 6322 fpl.tvp = NULL; /* for degenerate path handling */ 6323 fpl.pwd = pwdp; 6324 pwd = pwd_get_smr(); 6325 *(fpl.pwd) = pwd; 6326 namei_setup_rootdir(ndp, cnp, pwd); 6327 ndp->ni_topdir = pwd->pwd_jdir; 6328 6329 if (cnp->cn_pnbuf[0] == '/') { 6330 dvp = cache_fpl_handle_root(&fpl); 6331 ndp->ni_resflags = NIRES_ABS; 6332 } else { 6333 if (ndp->ni_dirfd == AT_FDCWD) { 6334 dvp = pwd->pwd_cdir; 6335 } else { 6336 error = cache_fplookup_dirfd(&fpl, &dvp); 6337 if (__predict_false(error != 0)) { 6338 goto out; 6339 } 6340 } 6341 } 6342 6343 SDT_PROBE4(vfs, namei, lookup, entry, dvp, cnp->cn_pnbuf, cnp->cn_flags, true); 6344 error = cache_fplookup_impl(dvp, &fpl); 6345 out: 6346 cache_fpl_smr_assert_not_entered(&fpl); 6347 cache_fpl_assert_status(&fpl); 6348 *status = fpl.status; 6349 if (SDT_PROBES_ENABLED()) { 6350 SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status); 6351 if (fpl.status == CACHE_FPL_STATUS_HANDLED) 6352 SDT_PROBE4(vfs, namei, lookup, return, error, ndp->ni_vp, true, 6353 ndp); 6354 } 6355 6356 if (__predict_true(fpl.status == CACHE_FPL_STATUS_HANDLED)) { 6357 MPASS(error != CACHE_FPL_FAILED); 6358 if (error != 0) { 6359 cache_fpl_cleanup_cnp(fpl.cnp); 6360 MPASS(fpl.dvp == NULL); 6361 MPASS(fpl.tvp == NULL); 6362 } 6363 ndp->ni_dvp = fpl.dvp; 6364 ndp->ni_vp = fpl.tvp; 6365 } 6366 return (error); 6367 } 6368