1 /* 2 * Copyright (c) 2003,2004,2009 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * Copyright (c) 1989, 1993, 1995 35 * The Regents of the University of California. All rights reserved. 36 * 37 * This code is derived from software contributed to Berkeley by 38 * Poul-Henning Kamp of the FreeBSD Project. 39 * 40 * Redistribution and use in source and binary forms, with or without 41 * modification, are permitted provided that the following conditions 42 * are met: 43 * 1. Redistributions of source code must retain the above copyright 44 * notice, this list of conditions and the following disclaimer. 45 * 2. Redistributions in binary form must reproduce the above copyright 46 * notice, this list of conditions and the following disclaimer in the 47 * documentation and/or other materials provided with the distribution. 48 * 3. Neither the name of the University nor the names of its contributors 49 * may be used to endorse or promote products derived from this software 50 * without specific prior written permission. 51 * 52 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 53 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 54 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 55 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 56 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 57 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 58 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 59 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 60 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 61 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 62 * SUCH DAMAGE. 63 */ 64 65 #include <sys/param.h> 66 #include <sys/systm.h> 67 #include <sys/kernel.h> 68 #include <sys/sysctl.h> 69 #include <sys/mount.h> 70 #include <sys/vnode.h> 71 #include <sys/malloc.h> 72 #include <sys/sysproto.h> 73 #include <sys/spinlock.h> 74 #include <sys/proc.h> 75 #include <sys/namei.h> 76 #include <sys/nlookup.h> 77 #include <sys/filedesc.h> 78 #include <sys/fnv_hash.h> 79 #include <sys/globaldata.h> 80 #include <sys/kern_syscall.h> 81 #include <sys/dirent.h> 82 #include <ddb/ddb.h> 83 84 #include <sys/sysref2.h> 85 #include <sys/spinlock2.h> 86 #include <sys/mplock2.h> 87 88 #define MAX_RECURSION_DEPTH 64 89 90 /* 91 * Random lookups in the cache are accomplished with a hash table using 92 * a hash key of (nc_src_vp, name). Each hash chain has its own spin lock. 93 * 94 * Negative entries may exist and correspond to resolved namecache 95 * structures where nc_vp is NULL. In a negative entry, NCF_WHITEOUT 96 * will be set if the entry corresponds to a whited-out directory entry 97 * (verses simply not finding the entry at all). ncneglist is locked 98 * with a global spinlock (ncspin). 99 * 100 * MPSAFE RULES: 101 * 102 * (1) A ncp must be referenced before it can be locked. 103 * 104 * (2) A ncp must be locked in order to modify it. 105 * 106 * (3) ncp locks are always ordered child -> parent. That may seem 107 * backwards but forward scans use the hash table and thus can hold 108 * the parent unlocked when traversing downward. 109 * 110 * This allows insert/rename/delete/dot-dot and other operations 111 * to use ncp->nc_parent links. 112 * 113 * This also prevents a locked up e.g. NFS node from creating a 114 * chain reaction all the way back to the root vnode / namecache. 115 * 116 * (4) parent linkages require both the parent and child to be locked. 117 */ 118 119 /* 120 * Structures associated with name cacheing. 121 */ 122 #define NCHHASH(hash) (&nchashtbl[(hash) & nchash]) 123 #define MINNEG 1024 124 #define MINPOS 1024 125 #define NCMOUNT_NUMCACHE 1009 /* prime number */ 126 127 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries"); 128 129 LIST_HEAD(nchash_list, namecache); 130 131 struct nchash_head { 132 struct nchash_list list; 133 struct spinlock spin; 134 }; 135 136 struct ncmount_cache { 137 struct spinlock spin; 138 struct namecache *ncp; 139 struct mount *mp; 140 int isneg; /* if != 0 mp is originator and not target */ 141 }; 142 143 static struct nchash_head *nchashtbl; 144 static struct namecache_list ncneglist; 145 static struct spinlock ncspin; 146 static struct ncmount_cache ncmount_cache[NCMOUNT_NUMCACHE]; 147 148 /* 149 * ncvp_debug - debug cache_fromvp(). This is used by the NFS server 150 * to create the namecache infrastructure leading to a dangling vnode. 151 * 152 * 0 Only errors are reported 153 * 1 Successes are reported 154 * 2 Successes + the whole directory scan is reported 155 * 3 Force the directory scan code run as if the parent vnode did not 156 * have a namecache record, even if it does have one. 157 */ 158 static int ncvp_debug; 159 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0, 160 "Namecache debug level (0-3)"); 161 162 static u_long nchash; /* size of hash table */ 163 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, 164 "Size of namecache hash table"); 165 166 static int ncnegflush = 10; /* burst for negative flush */ 167 SYSCTL_INT(_debug, OID_AUTO, ncnegflush, CTLFLAG_RW, &ncnegflush, 0, 168 "Batch flush negative entries"); 169 170 static int ncposflush = 10; /* burst for positive flush */ 171 SYSCTL_INT(_debug, OID_AUTO, ncposflush, CTLFLAG_RW, &ncposflush, 0, 172 "Batch flush positive entries"); 173 174 static int ncnegfactor = 16; /* ratio of negative entries */ 175 SYSCTL_INT(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0, 176 "Ratio of namecache negative entries"); 177 178 static int nclockwarn; /* warn on locked entries in ticks */ 179 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0, 180 "Warn on locked namecache entries in ticks"); 181 182 static int numdefered; /* number of cache entries allocated */ 183 SYSCTL_INT(_debug, OID_AUTO, numdefered, CTLFLAG_RD, &numdefered, 0, 184 "Number of cache entries allocated"); 185 186 static int ncposlimit; /* number of cache entries allocated */ 187 SYSCTL_INT(_debug, OID_AUTO, ncposlimit, CTLFLAG_RW, &ncposlimit, 0, 188 "Number of cache entries allocated"); 189 190 static int ncp_shared_lock_disable = 1; 191 SYSCTL_INT(_debug, OID_AUTO, ncp_shared_lock_disable, CTLFLAG_RW, 192 &ncp_shared_lock_disable, 0, "Disable shared namecache locks"); 193 194 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode), 195 "sizeof(struct vnode)"); 196 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache), 197 "sizeof(struct namecache)"); 198 199 static int ncmount_cache_enable = 1; 200 SYSCTL_INT(_debug, OID_AUTO, ncmount_cache_enable, CTLFLAG_RW, 201 &ncmount_cache_enable, 0, "mount point cache"); 202 static long ncmount_cache_hit; 203 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_hit, CTLFLAG_RW, 204 &ncmount_cache_hit, 0, "mpcache hits"); 205 static long ncmount_cache_miss; 206 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_miss, CTLFLAG_RW, 207 &ncmount_cache_miss, 0, "mpcache misses"); 208 static long ncmount_cache_overwrite; 209 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_overwrite, CTLFLAG_RW, 210 &ncmount_cache_overwrite, 0, "mpcache entry overwrites"); 211 212 static int cache_resolve_mp(struct mount *mp); 213 static struct vnode *cache_dvpref(struct namecache *ncp); 214 static void _cache_lock(struct namecache *ncp); 215 static void _cache_setunresolved(struct namecache *ncp); 216 static void _cache_cleanneg(int count); 217 static void _cache_cleanpos(int count); 218 static void _cache_cleandefered(void); 219 static void _cache_unlink(struct namecache *ncp); 220 221 /* 222 * The new name cache statistics 223 */ 224 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics"); 225 static int numneg; 226 SYSCTL_INT(_vfs_cache, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0, 227 "Number of negative namecache entries"); 228 static int numcache; 229 SYSCTL_INT(_vfs_cache, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0, 230 "Number of namecaches entries"); 231 static u_long numcalls; 232 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcalls, CTLFLAG_RD, &numcalls, 0, 233 "Number of namecache lookups"); 234 static u_long numchecks; 235 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numchecks, CTLFLAG_RD, &numchecks, 0, 236 "Number of checked entries in namecache lookups"); 237 238 struct nchstats nchstats[SMP_MAXCPU]; 239 /* 240 * Export VFS cache effectiveness statistics to user-land. 241 * 242 * The statistics are left for aggregation to user-land so 243 * neat things can be achieved, like observing per-CPU cache 244 * distribution. 245 */ 246 static int 247 sysctl_nchstats(SYSCTL_HANDLER_ARGS) 248 { 249 struct globaldata *gd; 250 int i, error; 251 252 error = 0; 253 for (i = 0; i < ncpus; ++i) { 254 gd = globaldata_find(i); 255 if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats), 256 sizeof(struct nchstats)))) 257 break; 258 } 259 260 return (error); 261 } 262 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD, 263 0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics"); 264 265 static struct namecache *cache_zap(struct namecache *ncp, int nonblock); 266 267 /* 268 * Namespace locking. The caller must already hold a reference to the 269 * namecache structure in order to lock/unlock it. This function prevents 270 * the namespace from being created or destroyed by accessors other then 271 * the lock holder. 272 * 273 * Note that holding a locked namecache structure prevents other threads 274 * from making namespace changes (e.g. deleting or creating), prevents 275 * vnode association state changes by other threads, and prevents the 276 * namecache entry from being resolved or unresolved by other threads. 277 * 278 * An exclusive lock owner has full authority to associate/disassociate 279 * vnodes and resolve/unresolve the locked ncp. 280 * 281 * A shared lock owner only has authority to acquire the underlying vnode, 282 * if any. 283 * 284 * The primary lock field is nc_lockstatus. nc_locktd is set after the 285 * fact (when locking) or cleared prior to unlocking. 286 * 287 * WARNING! Holding a locked ncp will prevent a vnode from being destroyed 288 * or recycled, but it does NOT help you if the vnode had already 289 * initiated a recyclement. If this is important, use cache_get() 290 * rather then cache_lock() (and deal with the differences in the 291 * way the refs counter is handled). Or, alternatively, make an 292 * unconditional call to cache_validate() or cache_resolve() 293 * after cache_lock() returns. 294 */ 295 static 296 void 297 _cache_lock(struct namecache *ncp) 298 { 299 thread_t td; 300 int didwarn; 301 int error; 302 u_int count; 303 304 KKASSERT(ncp->nc_refs != 0); 305 didwarn = 0; 306 td = curthread; 307 308 for (;;) { 309 count = ncp->nc_lockstatus; 310 cpu_ccfence(); 311 312 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 0) { 313 if (atomic_cmpset_int(&ncp->nc_lockstatus, 314 count, count + 1)) { 315 /* 316 * The vp associated with a locked ncp must 317 * be held to prevent it from being recycled. 318 * 319 * WARNING! If VRECLAIMED is set the vnode 320 * could already be in the middle of a recycle. 321 * Callers must use cache_vref() or 322 * cache_vget() on the locked ncp to 323 * validate the vp or set the cache entry 324 * to unresolved. 325 * 326 * NOTE! vhold() is allowed if we hold a 327 * lock on the ncp (which we do). 328 */ 329 ncp->nc_locktd = td; 330 if (ncp->nc_vp) 331 vhold(ncp->nc_vp); 332 break; 333 } 334 /* cmpset failed */ 335 continue; 336 } 337 if (ncp->nc_locktd == td) { 338 KKASSERT((count & NC_SHLOCK_FLAG) == 0); 339 if (atomic_cmpset_int(&ncp->nc_lockstatus, 340 count, count + 1)) { 341 break; 342 } 343 /* cmpset failed */ 344 continue; 345 } 346 tsleep_interlock(&ncp->nc_locktd, 0); 347 if (atomic_cmpset_int(&ncp->nc_lockstatus, count, 348 count | NC_EXLOCK_REQ) == 0) { 349 /* cmpset failed */ 350 continue; 351 } 352 error = tsleep(&ncp->nc_locktd, PINTERLOCKED, 353 "clock", nclockwarn); 354 if (error == EWOULDBLOCK) { 355 if (didwarn == 0) { 356 didwarn = ticks; 357 kprintf("[diagnostic] cache_lock: " 358 "blocked on %p %08x", 359 ncp, count); 360 kprintf(" \"%*.*s\"\n", 361 ncp->nc_nlen, ncp->nc_nlen, 362 ncp->nc_name); 363 } 364 } 365 /* loop */ 366 } 367 if (didwarn) { 368 kprintf("[diagnostic] cache_lock: unblocked %*.*s after " 369 "%d secs\n", 370 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name, 371 (int)(ticks - didwarn) / hz); 372 } 373 } 374 375 /* 376 * The shared lock works similarly to the exclusive lock except 377 * nc_locktd is left NULL and we need an interlock (VHOLD) to 378 * prevent vhold() races, since the moment our cmpset_int succeeds 379 * another cpu can come in and get its own shared lock. 380 * 381 * A critical section is needed to prevent interruption during the 382 * VHOLD interlock. 383 */ 384 static 385 void 386 _cache_lock_shared(struct namecache *ncp) 387 { 388 int didwarn; 389 int error; 390 u_int count; 391 392 KKASSERT(ncp->nc_refs != 0); 393 didwarn = 0; 394 395 for (;;) { 396 count = ncp->nc_lockstatus; 397 cpu_ccfence(); 398 399 if ((count & ~NC_SHLOCK_REQ) == 0) { 400 crit_enter(); 401 if (atomic_cmpset_int(&ncp->nc_lockstatus, 402 count, 403 (count + 1) | NC_SHLOCK_FLAG | 404 NC_SHLOCK_VHOLD)) { 405 /* 406 * The vp associated with a locked ncp must 407 * be held to prevent it from being recycled. 408 * 409 * WARNING! If VRECLAIMED is set the vnode 410 * could already be in the middle of a recycle. 411 * Callers must use cache_vref() or 412 * cache_vget() on the locked ncp to 413 * validate the vp or set the cache entry 414 * to unresolved. 415 * 416 * NOTE! vhold() is allowed if we hold a 417 * lock on the ncp (which we do). 418 */ 419 if (ncp->nc_vp) 420 vhold(ncp->nc_vp); 421 atomic_clear_int(&ncp->nc_lockstatus, 422 NC_SHLOCK_VHOLD); 423 crit_exit(); 424 break; 425 } 426 /* cmpset failed */ 427 crit_exit(); 428 continue; 429 } 430 431 /* 432 * If already held shared we can just bump the count, but 433 * only allow this if nobody is trying to get the lock 434 * exclusively. 435 * 436 * VHOLD is a bit of a hack. Even though we successfully 437 * added another shared ref, the cpu that got the first 438 * shared ref might not yet have held the vnode. 439 */ 440 if ((count & (NC_EXLOCK_REQ|NC_SHLOCK_FLAG)) == 441 NC_SHLOCK_FLAG) { 442 KKASSERT((count & ~(NC_EXLOCK_REQ | 443 NC_SHLOCK_REQ | 444 NC_SHLOCK_FLAG)) > 0); 445 if (atomic_cmpset_int(&ncp->nc_lockstatus, 446 count, count + 1)) { 447 while (ncp->nc_lockstatus & NC_SHLOCK_VHOLD) 448 cpu_pause(); 449 break; 450 } 451 continue; 452 } 453 tsleep_interlock(ncp, 0); 454 if (atomic_cmpset_int(&ncp->nc_lockstatus, count, 455 count | NC_SHLOCK_REQ) == 0) { 456 /* cmpset failed */ 457 continue; 458 } 459 error = tsleep(ncp, PINTERLOCKED, "clocksh", nclockwarn); 460 if (error == EWOULDBLOCK) { 461 if (didwarn == 0) { 462 didwarn = ticks; 463 kprintf("[diagnostic] cache_lock_shared: " 464 "blocked on %p %08x", 465 ncp, count); 466 kprintf(" \"%*.*s\"\n", 467 ncp->nc_nlen, ncp->nc_nlen, 468 ncp->nc_name); 469 } 470 } 471 /* loop */ 472 } 473 if (didwarn) { 474 kprintf("[diagnostic] cache_lock_shared: " 475 "unblocked %*.*s after %d secs\n", 476 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name, 477 (int)(ticks - didwarn) / hz); 478 } 479 } 480 481 /* 482 * NOTE: nc_refs may be zero if the ncp is interlocked by circumstance, 483 * such as the case where one of its children is locked. 484 */ 485 static 486 int 487 _cache_lock_nonblock(struct namecache *ncp) 488 { 489 thread_t td; 490 u_int count; 491 492 td = curthread; 493 494 for (;;) { 495 count = ncp->nc_lockstatus; 496 497 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 0) { 498 if (atomic_cmpset_int(&ncp->nc_lockstatus, 499 count, count + 1)) { 500 /* 501 * The vp associated with a locked ncp must 502 * be held to prevent it from being recycled. 503 * 504 * WARNING! If VRECLAIMED is set the vnode 505 * could already be in the middle of a recycle. 506 * Callers must use cache_vref() or 507 * cache_vget() on the locked ncp to 508 * validate the vp or set the cache entry 509 * to unresolved. 510 * 511 * NOTE! vhold() is allowed if we hold a 512 * lock on the ncp (which we do). 513 */ 514 ncp->nc_locktd = td; 515 if (ncp->nc_vp) 516 vhold(ncp->nc_vp); 517 break; 518 } 519 /* cmpset failed */ 520 continue; 521 } 522 if (ncp->nc_locktd == td) { 523 if (atomic_cmpset_int(&ncp->nc_lockstatus, 524 count, count + 1)) { 525 break; 526 } 527 /* cmpset failed */ 528 continue; 529 } 530 return(EWOULDBLOCK); 531 } 532 return(0); 533 } 534 535 /* 536 * The shared lock works similarly to the exclusive lock except 537 * nc_locktd is left NULL and we need an interlock (VHOLD) to 538 * prevent vhold() races, since the moment our cmpset_int succeeds 539 * another cpu can come in and get its own shared lock. 540 * 541 * A critical section is needed to prevent interruption during the 542 * VHOLD interlock. 543 */ 544 static 545 int 546 _cache_lock_shared_nonblock(struct namecache *ncp) 547 { 548 u_int count; 549 550 for (;;) { 551 count = ncp->nc_lockstatus; 552 553 if ((count & ~NC_SHLOCK_REQ) == 0) { 554 crit_enter(); 555 if (atomic_cmpset_int(&ncp->nc_lockstatus, 556 count, 557 (count + 1) | NC_SHLOCK_FLAG | 558 NC_SHLOCK_VHOLD)) { 559 /* 560 * The vp associated with a locked ncp must 561 * be held to prevent it from being recycled. 562 * 563 * WARNING! If VRECLAIMED is set the vnode 564 * could already be in the middle of a recycle. 565 * Callers must use cache_vref() or 566 * cache_vget() on the locked ncp to 567 * validate the vp or set the cache entry 568 * to unresolved. 569 * 570 * NOTE! vhold() is allowed if we hold a 571 * lock on the ncp (which we do). 572 */ 573 if (ncp->nc_vp) 574 vhold(ncp->nc_vp); 575 atomic_clear_int(&ncp->nc_lockstatus, 576 NC_SHLOCK_VHOLD); 577 crit_exit(); 578 break; 579 } 580 /* cmpset failed */ 581 crit_exit(); 582 continue; 583 } 584 585 /* 586 * If already held shared we can just bump the count, but 587 * only allow this if nobody is trying to get the lock 588 * exclusively. 589 * 590 * VHOLD is a bit of a hack. Even though we successfully 591 * added another shared ref, the cpu that got the first 592 * shared ref might not yet have held the vnode. 593 */ 594 if ((count & (NC_EXLOCK_REQ|NC_SHLOCK_FLAG)) == 595 NC_SHLOCK_FLAG) { 596 KKASSERT((count & ~(NC_EXLOCK_REQ | 597 NC_SHLOCK_REQ | 598 NC_SHLOCK_FLAG)) > 0); 599 if (atomic_cmpset_int(&ncp->nc_lockstatus, 600 count, count + 1)) { 601 while (ncp->nc_lockstatus & NC_SHLOCK_VHOLD) 602 cpu_pause(); 603 break; 604 } 605 continue; 606 } 607 return(EWOULDBLOCK); 608 } 609 return(0); 610 } 611 612 /* 613 * Helper function 614 * 615 * NOTE: nc_refs can be 0 (degenerate case during _cache_drop). 616 * 617 * nc_locktd must be NULLed out prior to nc_lockstatus getting cleared. 618 */ 619 static 620 void 621 _cache_unlock(struct namecache *ncp) 622 { 623 thread_t td __debugvar = curthread; 624 u_int count; 625 u_int ncount; 626 struct vnode *dropvp; 627 628 KKASSERT(ncp->nc_refs >= 0); 629 KKASSERT((ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) > 0); 630 KKASSERT((ncp->nc_lockstatus & NC_SHLOCK_FLAG) || ncp->nc_locktd == td); 631 632 count = ncp->nc_lockstatus; 633 cpu_ccfence(); 634 635 /* 636 * Clear nc_locktd prior to the atomic op (excl lock only) 637 */ 638 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 1) 639 ncp->nc_locktd = NULL; 640 dropvp = NULL; 641 642 for (;;) { 643 if ((count & 644 ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ|NC_SHLOCK_FLAG)) == 1) { 645 dropvp = ncp->nc_vp; 646 if (count & NC_EXLOCK_REQ) 647 ncount = count & NC_SHLOCK_REQ; /* cnt->0 */ 648 else 649 ncount = 0; 650 651 if (atomic_cmpset_int(&ncp->nc_lockstatus, 652 count, ncount)) { 653 if (count & NC_EXLOCK_REQ) 654 wakeup(&ncp->nc_locktd); 655 else if (count & NC_SHLOCK_REQ) 656 wakeup(ncp); 657 break; 658 } 659 dropvp = NULL; 660 } else { 661 KKASSERT((count & NC_SHLOCK_VHOLD) == 0); 662 KKASSERT((count & ~(NC_EXLOCK_REQ | 663 NC_SHLOCK_REQ | 664 NC_SHLOCK_FLAG)) > 1); 665 if (atomic_cmpset_int(&ncp->nc_lockstatus, 666 count, count - 1)) { 667 break; 668 } 669 } 670 count = ncp->nc_lockstatus; 671 cpu_ccfence(); 672 } 673 674 /* 675 * Don't actually drop the vp until we successfully clean out 676 * the lock, otherwise we may race another shared lock. 677 */ 678 if (dropvp) 679 vdrop(dropvp); 680 } 681 682 static 683 int 684 _cache_lockstatus(struct namecache *ncp) 685 { 686 if (ncp->nc_locktd == curthread) 687 return(LK_EXCLUSIVE); 688 if (ncp->nc_lockstatus & NC_SHLOCK_FLAG) 689 return(LK_SHARED); 690 return(-1); 691 } 692 693 /* 694 * cache_hold() and cache_drop() prevent the premature deletion of a 695 * namecache entry but do not prevent operations (such as zapping) on 696 * that namecache entry. 697 * 698 * This routine may only be called from outside this source module if 699 * nc_refs is already at least 1. 700 * 701 * This is a rare case where callers are allowed to hold a spinlock, 702 * so we can't ourselves. 703 */ 704 static __inline 705 struct namecache * 706 _cache_hold(struct namecache *ncp) 707 { 708 atomic_add_int(&ncp->nc_refs, 1); 709 return(ncp); 710 } 711 712 /* 713 * Drop a cache entry, taking care to deal with races. 714 * 715 * For potential 1->0 transitions we must hold the ncp lock to safely 716 * test its flags. An unresolved entry with no children must be zapped 717 * to avoid leaks. 718 * 719 * The call to cache_zap() itself will handle all remaining races and 720 * will decrement the ncp's refs regardless. If we are resolved or 721 * have children nc_refs can safely be dropped to 0 without having to 722 * zap the entry. 723 * 724 * NOTE: cache_zap() will re-check nc_refs and nc_list in a MPSAFE fashion. 725 * 726 * NOTE: cache_zap() may return a non-NULL referenced parent which must 727 * be dropped in a loop. 728 */ 729 static __inline 730 void 731 _cache_drop(struct namecache *ncp) 732 { 733 int refs; 734 735 while (ncp) { 736 KKASSERT(ncp->nc_refs > 0); 737 refs = ncp->nc_refs; 738 739 if (refs == 1) { 740 if (_cache_lock_nonblock(ncp) == 0) { 741 ncp->nc_flag &= ~NCF_DEFEREDZAP; 742 if ((ncp->nc_flag & NCF_UNRESOLVED) && 743 TAILQ_EMPTY(&ncp->nc_list)) { 744 ncp = cache_zap(ncp, 1); 745 continue; 746 } 747 if (atomic_cmpset_int(&ncp->nc_refs, 1, 0)) { 748 _cache_unlock(ncp); 749 break; 750 } 751 _cache_unlock(ncp); 752 } 753 } else { 754 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) 755 break; 756 } 757 cpu_pause(); 758 } 759 } 760 761 /* 762 * Link a new namecache entry to its parent and to the hash table. Be 763 * careful to avoid races if vhold() blocks in the future. 764 * 765 * Both ncp and par must be referenced and locked. 766 * 767 * NOTE: The hash table spinlock is held during this call, we can't do 768 * anything fancy. 769 */ 770 static void 771 _cache_link_parent(struct namecache *ncp, struct namecache *par, 772 struct nchash_head *nchpp) 773 { 774 KKASSERT(ncp->nc_parent == NULL); 775 ncp->nc_parent = par; 776 ncp->nc_head = nchpp; 777 778 /* 779 * Set inheritance flags. Note that the parent flags may be 780 * stale due to getattr potentially not having been run yet 781 * (it gets run during nlookup()'s). 782 */ 783 ncp->nc_flag &= ~(NCF_SF_PNOCACHE | NCF_UF_PCACHE); 784 if (par->nc_flag & (NCF_SF_NOCACHE | NCF_SF_PNOCACHE)) 785 ncp->nc_flag |= NCF_SF_PNOCACHE; 786 if (par->nc_flag & (NCF_UF_CACHE | NCF_UF_PCACHE)) 787 ncp->nc_flag |= NCF_UF_PCACHE; 788 789 LIST_INSERT_HEAD(&nchpp->list, ncp, nc_hash); 790 791 if (TAILQ_EMPTY(&par->nc_list)) { 792 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry); 793 /* 794 * Any vp associated with an ncp which has children must 795 * be held to prevent it from being recycled. 796 */ 797 if (par->nc_vp) 798 vhold(par->nc_vp); 799 } else { 800 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry); 801 } 802 } 803 804 /* 805 * Remove the parent and hash associations from a namecache structure. 806 * If this is the last child of the parent the cache_drop(par) will 807 * attempt to recursively zap the parent. 808 * 809 * ncp must be locked. This routine will acquire a temporary lock on 810 * the parent as wlel as the appropriate hash chain. 811 */ 812 static void 813 _cache_unlink_parent(struct namecache *ncp) 814 { 815 struct namecache *par; 816 struct vnode *dropvp; 817 818 if ((par = ncp->nc_parent) != NULL) { 819 KKASSERT(ncp->nc_parent == par); 820 _cache_hold(par); 821 _cache_lock(par); 822 spin_lock(&ncp->nc_head->spin); 823 LIST_REMOVE(ncp, nc_hash); 824 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 825 dropvp = NULL; 826 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 827 dropvp = par->nc_vp; 828 spin_unlock(&ncp->nc_head->spin); 829 ncp->nc_parent = NULL; 830 ncp->nc_head = NULL; 831 _cache_unlock(par); 832 _cache_drop(par); 833 834 /* 835 * We can only safely vdrop with no spinlocks held. 836 */ 837 if (dropvp) 838 vdrop(dropvp); 839 } 840 } 841 842 /* 843 * Allocate a new namecache structure. Most of the code does not require 844 * zero-termination of the string but it makes vop_compat_ncreate() easier. 845 */ 846 static struct namecache * 847 cache_alloc(int nlen) 848 { 849 struct namecache *ncp; 850 851 ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO); 852 if (nlen) 853 ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK); 854 ncp->nc_nlen = nlen; 855 ncp->nc_flag = NCF_UNRESOLVED; 856 ncp->nc_error = ENOTCONN; /* needs to be resolved */ 857 ncp->nc_refs = 1; 858 859 TAILQ_INIT(&ncp->nc_list); 860 _cache_lock(ncp); 861 return(ncp); 862 } 863 864 /* 865 * Can only be called for the case where the ncp has never been 866 * associated with anything (so no spinlocks are needed). 867 */ 868 static void 869 _cache_free(struct namecache *ncp) 870 { 871 KKASSERT(ncp->nc_refs == 1 && ncp->nc_lockstatus == 1); 872 if (ncp->nc_name) 873 kfree(ncp->nc_name, M_VFSCACHE); 874 kfree(ncp, M_VFSCACHE); 875 } 876 877 /* 878 * [re]initialize a nchandle. 879 */ 880 void 881 cache_zero(struct nchandle *nch) 882 { 883 nch->ncp = NULL; 884 nch->mount = NULL; 885 } 886 887 /* 888 * Ref and deref a namecache structure. 889 * 890 * The caller must specify a stable ncp pointer, typically meaning the 891 * ncp is already referenced but this can also occur indirectly through 892 * e.g. holding a lock on a direct child. 893 * 894 * WARNING: Caller may hold an unrelated read spinlock, which means we can't 895 * use read spinlocks here. 896 * 897 * MPSAFE if nch is 898 */ 899 struct nchandle * 900 cache_hold(struct nchandle *nch) 901 { 902 _cache_hold(nch->ncp); 903 atomic_add_int(&nch->mount->mnt_refs, 1); 904 return(nch); 905 } 906 907 /* 908 * Create a copy of a namecache handle for an already-referenced 909 * entry. 910 * 911 * MPSAFE if nch is 912 */ 913 void 914 cache_copy(struct nchandle *nch, struct nchandle *target) 915 { 916 *target = *nch; 917 if (target->ncp) 918 _cache_hold(target->ncp); 919 atomic_add_int(&nch->mount->mnt_refs, 1); 920 } 921 922 /* 923 * MPSAFE if nch is 924 */ 925 void 926 cache_changemount(struct nchandle *nch, struct mount *mp) 927 { 928 atomic_add_int(&nch->mount->mnt_refs, -1); 929 nch->mount = mp; 930 atomic_add_int(&nch->mount->mnt_refs, 1); 931 } 932 933 void 934 cache_drop(struct nchandle *nch) 935 { 936 atomic_add_int(&nch->mount->mnt_refs, -1); 937 _cache_drop(nch->ncp); 938 nch->ncp = NULL; 939 nch->mount = NULL; 940 } 941 942 int 943 cache_lockstatus(struct nchandle *nch) 944 { 945 return(_cache_lockstatus(nch->ncp)); 946 } 947 948 void 949 cache_lock(struct nchandle *nch) 950 { 951 _cache_lock(nch->ncp); 952 } 953 954 void 955 cache_lock_maybe_shared(struct nchandle *nch, int excl) 956 { 957 struct namecache *ncp = nch->ncp; 958 959 if (ncp_shared_lock_disable || excl || 960 (ncp->nc_flag & NCF_UNRESOLVED)) { 961 _cache_lock(ncp); 962 } else { 963 _cache_lock_shared(ncp); 964 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 965 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) { 966 _cache_unlock(ncp); 967 _cache_lock(ncp); 968 } 969 } else { 970 _cache_unlock(ncp); 971 _cache_lock(ncp); 972 } 973 } 974 } 975 976 /* 977 * Relock nch1 given an unlocked nch1 and a locked nch2. The caller 978 * is responsible for checking both for validity on return as they 979 * may have become invalid. 980 * 981 * We have to deal with potential deadlocks here, just ping pong 982 * the lock until we get it (we will always block somewhere when 983 * looping so this is not cpu-intensive). 984 * 985 * which = 0 nch1 not locked, nch2 is locked 986 * which = 1 nch1 is locked, nch2 is not locked 987 */ 988 void 989 cache_relock(struct nchandle *nch1, struct ucred *cred1, 990 struct nchandle *nch2, struct ucred *cred2) 991 { 992 int which; 993 994 which = 0; 995 996 for (;;) { 997 if (which == 0) { 998 if (cache_lock_nonblock(nch1) == 0) { 999 cache_resolve(nch1, cred1); 1000 break; 1001 } 1002 cache_unlock(nch2); 1003 cache_lock(nch1); 1004 cache_resolve(nch1, cred1); 1005 which = 1; 1006 } else { 1007 if (cache_lock_nonblock(nch2) == 0) { 1008 cache_resolve(nch2, cred2); 1009 break; 1010 } 1011 cache_unlock(nch1); 1012 cache_lock(nch2); 1013 cache_resolve(nch2, cred2); 1014 which = 0; 1015 } 1016 } 1017 } 1018 1019 int 1020 cache_lock_nonblock(struct nchandle *nch) 1021 { 1022 return(_cache_lock_nonblock(nch->ncp)); 1023 } 1024 1025 void 1026 cache_unlock(struct nchandle *nch) 1027 { 1028 _cache_unlock(nch->ncp); 1029 } 1030 1031 /* 1032 * ref-and-lock, unlock-and-deref functions. 1033 * 1034 * This function is primarily used by nlookup. Even though cache_lock 1035 * holds the vnode, it is possible that the vnode may have already 1036 * initiated a recyclement. 1037 * 1038 * We want cache_get() to return a definitively usable vnode or a 1039 * definitively unresolved ncp. 1040 */ 1041 static 1042 struct namecache * 1043 _cache_get(struct namecache *ncp) 1044 { 1045 _cache_hold(ncp); 1046 _cache_lock(ncp); 1047 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 1048 _cache_setunresolved(ncp); 1049 return(ncp); 1050 } 1051 1052 /* 1053 * Attempt to obtain a shared lock on the ncp. A shared lock will only 1054 * be obtained if the ncp is resolved and the vnode (if not ENOENT) is 1055 * valid. Otherwise an exclusive lock will be acquired instead. 1056 */ 1057 static 1058 struct namecache * 1059 _cache_get_maybe_shared(struct namecache *ncp, int excl) 1060 { 1061 if (ncp_shared_lock_disable || excl || 1062 (ncp->nc_flag & NCF_UNRESOLVED)) { 1063 return(_cache_get(ncp)); 1064 } 1065 _cache_hold(ncp); 1066 _cache_lock_shared(ncp); 1067 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 1068 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) { 1069 _cache_unlock(ncp); 1070 ncp = _cache_get(ncp); 1071 _cache_drop(ncp); 1072 } 1073 } else { 1074 _cache_unlock(ncp); 1075 ncp = _cache_get(ncp); 1076 _cache_drop(ncp); 1077 } 1078 return(ncp); 1079 } 1080 1081 /* 1082 * This is a special form of _cache_lock() which only succeeds if 1083 * it can get a pristine, non-recursive lock. The caller must have 1084 * already ref'd the ncp. 1085 * 1086 * On success the ncp will be locked, on failure it will not. The 1087 * ref count does not change either way. 1088 * 1089 * We want _cache_lock_special() (on success) to return a definitively 1090 * usable vnode or a definitively unresolved ncp. 1091 */ 1092 static int 1093 _cache_lock_special(struct namecache *ncp) 1094 { 1095 if (_cache_lock_nonblock(ncp) == 0) { 1096 if ((ncp->nc_lockstatus & 1097 ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 1) { 1098 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 1099 _cache_setunresolved(ncp); 1100 return(0); 1101 } 1102 _cache_unlock(ncp); 1103 } 1104 return(EWOULDBLOCK); 1105 } 1106 1107 static int 1108 _cache_lock_shared_special(struct namecache *ncp) 1109 { 1110 if (_cache_lock_shared_nonblock(ncp) == 0) { 1111 if ((ncp->nc_lockstatus & 1112 ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == (NC_SHLOCK_FLAG | 1)) { 1113 if (ncp->nc_vp == NULL || 1114 (ncp->nc_vp->v_flag & VRECLAIMED) == 0) { 1115 return(0); 1116 } 1117 } 1118 _cache_unlock(ncp); 1119 } 1120 return(EWOULDBLOCK); 1121 } 1122 1123 1124 /* 1125 * NOTE: The same nchandle can be passed for both arguments. 1126 */ 1127 void 1128 cache_get(struct nchandle *nch, struct nchandle *target) 1129 { 1130 KKASSERT(nch->ncp->nc_refs > 0); 1131 target->mount = nch->mount; 1132 target->ncp = _cache_get(nch->ncp); 1133 atomic_add_int(&target->mount->mnt_refs, 1); 1134 } 1135 1136 void 1137 cache_get_maybe_shared(struct nchandle *nch, struct nchandle *target, int excl) 1138 { 1139 KKASSERT(nch->ncp->nc_refs > 0); 1140 target->mount = nch->mount; 1141 target->ncp = _cache_get_maybe_shared(nch->ncp, excl); 1142 atomic_add_int(&target->mount->mnt_refs, 1); 1143 } 1144 1145 /* 1146 * 1147 */ 1148 static __inline 1149 void 1150 _cache_put(struct namecache *ncp) 1151 { 1152 _cache_unlock(ncp); 1153 _cache_drop(ncp); 1154 } 1155 1156 /* 1157 * 1158 */ 1159 void 1160 cache_put(struct nchandle *nch) 1161 { 1162 atomic_add_int(&nch->mount->mnt_refs, -1); 1163 _cache_put(nch->ncp); 1164 nch->ncp = NULL; 1165 nch->mount = NULL; 1166 } 1167 1168 /* 1169 * Resolve an unresolved ncp by associating a vnode with it. If the 1170 * vnode is NULL, a negative cache entry is created. 1171 * 1172 * The ncp should be locked on entry and will remain locked on return. 1173 */ 1174 static 1175 void 1176 _cache_setvp(struct mount *mp, struct namecache *ncp, struct vnode *vp) 1177 { 1178 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED); 1179 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE); 1180 1181 if (vp != NULL) { 1182 /* 1183 * Any vp associated with an ncp which has children must 1184 * be held. Any vp associated with a locked ncp must be held. 1185 */ 1186 if (!TAILQ_EMPTY(&ncp->nc_list)) 1187 vhold(vp); 1188 spin_lock(&vp->v_spin); 1189 ncp->nc_vp = vp; 1190 TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode); 1191 spin_unlock(&vp->v_spin); 1192 if (ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) 1193 vhold(vp); 1194 1195 /* 1196 * Set auxiliary flags 1197 */ 1198 switch(vp->v_type) { 1199 case VDIR: 1200 ncp->nc_flag |= NCF_ISDIR; 1201 break; 1202 case VLNK: 1203 ncp->nc_flag |= NCF_ISSYMLINK; 1204 /* XXX cache the contents of the symlink */ 1205 break; 1206 default: 1207 break; 1208 } 1209 atomic_add_int(&numcache, 1); 1210 ncp->nc_error = 0; 1211 /* XXX: this is a hack to work-around the lack of a real pfs vfs 1212 * implementation*/ 1213 if (mp != NULL) 1214 if (strncmp(mp->mnt_stat.f_fstypename, "null", 5) == 0) 1215 vp->v_pfsmp = mp; 1216 } else { 1217 /* 1218 * When creating a negative cache hit we set the 1219 * namecache_gen. A later resolve will clean out the 1220 * negative cache hit if the mount point's namecache_gen 1221 * has changed. Used by devfs, could also be used by 1222 * other remote FSs. 1223 */ 1224 ncp->nc_vp = NULL; 1225 spin_lock(&ncspin); 1226 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 1227 ++numneg; 1228 spin_unlock(&ncspin); 1229 ncp->nc_error = ENOENT; 1230 if (mp) 1231 VFS_NCPGEN_SET(mp, ncp); 1232 } 1233 ncp->nc_flag &= ~(NCF_UNRESOLVED | NCF_DEFEREDZAP); 1234 } 1235 1236 /* 1237 * 1238 */ 1239 void 1240 cache_setvp(struct nchandle *nch, struct vnode *vp) 1241 { 1242 _cache_setvp(nch->mount, nch->ncp, vp); 1243 } 1244 1245 /* 1246 * 1247 */ 1248 void 1249 cache_settimeout(struct nchandle *nch, int nticks) 1250 { 1251 struct namecache *ncp = nch->ncp; 1252 1253 if ((ncp->nc_timeout = ticks + nticks) == 0) 1254 ncp->nc_timeout = 1; 1255 } 1256 1257 /* 1258 * Disassociate the vnode or negative-cache association and mark a 1259 * namecache entry as unresolved again. Note that the ncp is still 1260 * left in the hash table and still linked to its parent. 1261 * 1262 * The ncp should be locked and refd on entry and will remain locked and refd 1263 * on return. 1264 * 1265 * This routine is normally never called on a directory containing children. 1266 * However, NFS often does just that in its rename() code as a cop-out to 1267 * avoid complex namespace operations. This disconnects a directory vnode 1268 * from its namecache and can cause the OLDAPI and NEWAPI to get out of 1269 * sync. 1270 * 1271 */ 1272 static 1273 void 1274 _cache_setunresolved(struct namecache *ncp) 1275 { 1276 struct vnode *vp; 1277 1278 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 1279 ncp->nc_flag |= NCF_UNRESOLVED; 1280 ncp->nc_timeout = 0; 1281 ncp->nc_error = ENOTCONN; 1282 if ((vp = ncp->nc_vp) != NULL) { 1283 atomic_add_int(&numcache, -1); 1284 spin_lock(&vp->v_spin); 1285 ncp->nc_vp = NULL; 1286 TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode); 1287 spin_unlock(&vp->v_spin); 1288 1289 /* 1290 * Any vp associated with an ncp with children is 1291 * held by that ncp. Any vp associated with a locked 1292 * ncp is held by that ncp. These conditions must be 1293 * undone when the vp is cleared out from the ncp. 1294 */ 1295 if (!TAILQ_EMPTY(&ncp->nc_list)) 1296 vdrop(vp); 1297 if (ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) 1298 vdrop(vp); 1299 } else { 1300 spin_lock(&ncspin); 1301 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 1302 --numneg; 1303 spin_unlock(&ncspin); 1304 } 1305 ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK); 1306 } 1307 } 1308 1309 /* 1310 * The cache_nresolve() code calls this function to automatically 1311 * set a resolved cache element to unresolved if it has timed out 1312 * or if it is a negative cache hit and the mount point namecache_gen 1313 * has changed. 1314 */ 1315 static __inline int 1316 _cache_auto_unresolve_test(struct mount *mp, struct namecache *ncp) 1317 { 1318 /* 1319 * Try to zap entries that have timed out. We have 1320 * to be careful here because locked leafs may depend 1321 * on the vnode remaining intact in a parent, so only 1322 * do this under very specific conditions. 1323 */ 1324 if (ncp->nc_timeout && (int)(ncp->nc_timeout - ticks) < 0 && 1325 TAILQ_EMPTY(&ncp->nc_list)) { 1326 return 1; 1327 } 1328 1329 /* 1330 * If a resolved negative cache hit is invalid due to 1331 * the mount's namecache generation being bumped, zap it. 1332 */ 1333 if (ncp->nc_vp == NULL && VFS_NCPGEN_TEST(mp, ncp)) { 1334 return 1; 1335 } 1336 1337 /* 1338 * Otherwise we are good 1339 */ 1340 return 0; 1341 } 1342 1343 static __inline void 1344 _cache_auto_unresolve(struct mount *mp, struct namecache *ncp) 1345 { 1346 /* 1347 * Already in an unresolved state, nothing to do. 1348 */ 1349 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 1350 if (_cache_auto_unresolve_test(mp, ncp)) 1351 _cache_setunresolved(ncp); 1352 } 1353 } 1354 1355 /* 1356 * 1357 */ 1358 void 1359 cache_setunresolved(struct nchandle *nch) 1360 { 1361 _cache_setunresolved(nch->ncp); 1362 } 1363 1364 /* 1365 * Determine if we can clear NCF_ISMOUNTPT by scanning the mountlist 1366 * looking for matches. This flag tells the lookup code when it must 1367 * check for a mount linkage and also prevents the directories in question 1368 * from being deleted or renamed. 1369 */ 1370 static 1371 int 1372 cache_clrmountpt_callback(struct mount *mp, void *data) 1373 { 1374 struct nchandle *nch = data; 1375 1376 if (mp->mnt_ncmounton.ncp == nch->ncp) 1377 return(1); 1378 if (mp->mnt_ncmountpt.ncp == nch->ncp) 1379 return(1); 1380 return(0); 1381 } 1382 1383 /* 1384 * 1385 */ 1386 void 1387 cache_clrmountpt(struct nchandle *nch) 1388 { 1389 int count; 1390 1391 count = mountlist_scan(cache_clrmountpt_callback, nch, 1392 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 1393 if (count == 0) 1394 nch->ncp->nc_flag &= ~NCF_ISMOUNTPT; 1395 } 1396 1397 /* 1398 * Invalidate portions of the namecache topology given a starting entry. 1399 * The passed ncp is set to an unresolved state and: 1400 * 1401 * The passed ncp must be referencxed and locked. The routine may unlock 1402 * and relock ncp several times, and will recheck the children and loop 1403 * to catch races. When done the passed ncp will be returned with the 1404 * reference and lock intact. 1405 * 1406 * CINV_DESTROY - Set a flag in the passed ncp entry indicating 1407 * that the physical underlying nodes have been 1408 * destroyed... as in deleted. For example, when 1409 * a directory is removed. This will cause record 1410 * lookups on the name to no longer be able to find 1411 * the record and tells the resolver to return failure 1412 * rather then trying to resolve through the parent. 1413 * 1414 * The topology itself, including ncp->nc_name, 1415 * remains intact. 1416 * 1417 * This only applies to the passed ncp, if CINV_CHILDREN 1418 * is specified the children are not flagged. 1419 * 1420 * CINV_CHILDREN - Set all children (recursively) to an unresolved 1421 * state as well. 1422 * 1423 * Note that this will also have the side effect of 1424 * cleaning out any unreferenced nodes in the topology 1425 * from the leaves up as the recursion backs out. 1426 * 1427 * Note that the topology for any referenced nodes remains intact, but 1428 * the nodes will be marked as having been destroyed and will be set 1429 * to an unresolved state. 1430 * 1431 * It is possible for cache_inval() to race a cache_resolve(), meaning that 1432 * the namecache entry may not actually be invalidated on return if it was 1433 * revalidated while recursing down into its children. This code guarentees 1434 * that the node(s) will go through an invalidation cycle, but does not 1435 * guarentee that they will remain in an invalidated state. 1436 * 1437 * Returns non-zero if a revalidation was detected during the invalidation 1438 * recursion, zero otherwise. Note that since only the original ncp is 1439 * locked the revalidation ultimately can only indicate that the original ncp 1440 * *MIGHT* no have been reresolved. 1441 * 1442 * DEEP RECURSION HANDLING - If a recursive invalidation recurses deeply we 1443 * have to avoid blowing out the kernel stack. We do this by saving the 1444 * deep namecache node and aborting the recursion, then re-recursing at that 1445 * node using a depth-first algorithm in order to allow multiple deep 1446 * recursions to chain through each other, then we restart the invalidation 1447 * from scratch. 1448 */ 1449 1450 struct cinvtrack { 1451 struct namecache *resume_ncp; 1452 int depth; 1453 }; 1454 1455 static int _cache_inval_internal(struct namecache *, int, struct cinvtrack *); 1456 1457 static 1458 int 1459 _cache_inval(struct namecache *ncp, int flags) 1460 { 1461 struct cinvtrack track; 1462 struct namecache *ncp2; 1463 int r; 1464 1465 track.depth = 0; 1466 track.resume_ncp = NULL; 1467 1468 for (;;) { 1469 r = _cache_inval_internal(ncp, flags, &track); 1470 if (track.resume_ncp == NULL) 1471 break; 1472 kprintf("Warning: deep namecache recursion at %s\n", 1473 ncp->nc_name); 1474 _cache_unlock(ncp); 1475 while ((ncp2 = track.resume_ncp) != NULL) { 1476 track.resume_ncp = NULL; 1477 _cache_lock(ncp2); 1478 _cache_inval_internal(ncp2, flags & ~CINV_DESTROY, 1479 &track); 1480 _cache_put(ncp2); 1481 } 1482 _cache_lock(ncp); 1483 } 1484 return(r); 1485 } 1486 1487 int 1488 cache_inval(struct nchandle *nch, int flags) 1489 { 1490 return(_cache_inval(nch->ncp, flags)); 1491 } 1492 1493 /* 1494 * Helper for _cache_inval(). The passed ncp is refd and locked and 1495 * remains that way on return, but may be unlocked/relocked multiple 1496 * times by the routine. 1497 */ 1498 static int 1499 _cache_inval_internal(struct namecache *ncp, int flags, struct cinvtrack *track) 1500 { 1501 struct namecache *kid; 1502 struct namecache *nextkid; 1503 int rcnt = 0; 1504 1505 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE); 1506 1507 _cache_setunresolved(ncp); 1508 if (flags & CINV_DESTROY) 1509 ncp->nc_flag |= NCF_DESTROYED; 1510 if ((flags & CINV_CHILDREN) && 1511 (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL 1512 ) { 1513 _cache_hold(kid); 1514 if (++track->depth > MAX_RECURSION_DEPTH) { 1515 track->resume_ncp = ncp; 1516 _cache_hold(ncp); 1517 ++rcnt; 1518 } 1519 _cache_unlock(ncp); 1520 while (kid) { 1521 if (track->resume_ncp) { 1522 _cache_drop(kid); 1523 break; 1524 } 1525 if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL) 1526 _cache_hold(nextkid); 1527 if ((kid->nc_flag & NCF_UNRESOLVED) == 0 || 1528 TAILQ_FIRST(&kid->nc_list) 1529 ) { 1530 _cache_lock(kid); 1531 rcnt += _cache_inval_internal(kid, flags & ~CINV_DESTROY, track); 1532 _cache_unlock(kid); 1533 } 1534 _cache_drop(kid); 1535 kid = nextkid; 1536 } 1537 --track->depth; 1538 _cache_lock(ncp); 1539 } 1540 1541 /* 1542 * Someone could have gotten in there while ncp was unlocked, 1543 * retry if so. 1544 */ 1545 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 1546 ++rcnt; 1547 return (rcnt); 1548 } 1549 1550 /* 1551 * Invalidate a vnode's namecache associations. To avoid races against 1552 * the resolver we do not invalidate a node which we previously invalidated 1553 * but which was then re-resolved while we were in the invalidation loop. 1554 * 1555 * Returns non-zero if any namecache entries remain after the invalidation 1556 * loop completed. 1557 * 1558 * NOTE: Unlike the namecache topology which guarentees that ncp's will not 1559 * be ripped out of the topology while held, the vnode's v_namecache 1560 * list has no such restriction. NCP's can be ripped out of the list 1561 * at virtually any time if not locked, even if held. 1562 * 1563 * In addition, the v_namecache list itself must be locked via 1564 * the vnode's spinlock. 1565 */ 1566 int 1567 cache_inval_vp(struct vnode *vp, int flags) 1568 { 1569 struct namecache *ncp; 1570 struct namecache *next; 1571 1572 restart: 1573 spin_lock(&vp->v_spin); 1574 ncp = TAILQ_FIRST(&vp->v_namecache); 1575 if (ncp) 1576 _cache_hold(ncp); 1577 while (ncp) { 1578 /* loop entered with ncp held and vp spin-locked */ 1579 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL) 1580 _cache_hold(next); 1581 spin_unlock(&vp->v_spin); 1582 _cache_lock(ncp); 1583 if (ncp->nc_vp != vp) { 1584 kprintf("Warning: cache_inval_vp: race-A detected on " 1585 "%s\n", ncp->nc_name); 1586 _cache_put(ncp); 1587 if (next) 1588 _cache_drop(next); 1589 goto restart; 1590 } 1591 _cache_inval(ncp, flags); 1592 _cache_put(ncp); /* also releases reference */ 1593 ncp = next; 1594 spin_lock(&vp->v_spin); 1595 if (ncp && ncp->nc_vp != vp) { 1596 spin_unlock(&vp->v_spin); 1597 kprintf("Warning: cache_inval_vp: race-B detected on " 1598 "%s\n", ncp->nc_name); 1599 _cache_drop(ncp); 1600 goto restart; 1601 } 1602 } 1603 spin_unlock(&vp->v_spin); 1604 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 1605 } 1606 1607 /* 1608 * This routine is used instead of the normal cache_inval_vp() when we 1609 * are trying to recycle otherwise good vnodes. 1610 * 1611 * Return 0 on success, non-zero if not all namecache records could be 1612 * disassociated from the vnode (for various reasons). 1613 */ 1614 int 1615 cache_inval_vp_nonblock(struct vnode *vp) 1616 { 1617 struct namecache *ncp; 1618 struct namecache *next; 1619 1620 spin_lock(&vp->v_spin); 1621 ncp = TAILQ_FIRST(&vp->v_namecache); 1622 if (ncp) 1623 _cache_hold(ncp); 1624 while (ncp) { 1625 /* loop entered with ncp held */ 1626 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL) 1627 _cache_hold(next); 1628 spin_unlock(&vp->v_spin); 1629 if (_cache_lock_nonblock(ncp)) { 1630 _cache_drop(ncp); 1631 if (next) 1632 _cache_drop(next); 1633 goto done; 1634 } 1635 if (ncp->nc_vp != vp) { 1636 kprintf("Warning: cache_inval_vp: race-A detected on " 1637 "%s\n", ncp->nc_name); 1638 _cache_put(ncp); 1639 if (next) 1640 _cache_drop(next); 1641 goto done; 1642 } 1643 _cache_inval(ncp, 0); 1644 _cache_put(ncp); /* also releases reference */ 1645 ncp = next; 1646 spin_lock(&vp->v_spin); 1647 if (ncp && ncp->nc_vp != vp) { 1648 spin_unlock(&vp->v_spin); 1649 kprintf("Warning: cache_inval_vp: race-B detected on " 1650 "%s\n", ncp->nc_name); 1651 _cache_drop(ncp); 1652 goto done; 1653 } 1654 } 1655 spin_unlock(&vp->v_spin); 1656 done: 1657 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 1658 } 1659 1660 /* 1661 * The source ncp has been renamed to the target ncp. Both fncp and tncp 1662 * must be locked. The target ncp is destroyed (as a normal rename-over 1663 * would destroy the target file or directory). 1664 * 1665 * Because there may be references to the source ncp we cannot copy its 1666 * contents to the target. Instead the source ncp is relinked as the target 1667 * and the target ncp is removed from the namecache topology. 1668 */ 1669 void 1670 cache_rename(struct nchandle *fnch, struct nchandle *tnch) 1671 { 1672 struct namecache *fncp = fnch->ncp; 1673 struct namecache *tncp = tnch->ncp; 1674 struct namecache *tncp_par; 1675 struct nchash_head *nchpp; 1676 u_int32_t hash; 1677 char *oname; 1678 char *nname; 1679 1680 if (tncp->nc_nlen) { 1681 nname = kmalloc(tncp->nc_nlen + 1, M_VFSCACHE, M_WAITOK); 1682 bcopy(tncp->nc_name, nname, tncp->nc_nlen); 1683 nname[tncp->nc_nlen] = 0; 1684 } else { 1685 nname = NULL; 1686 } 1687 1688 /* 1689 * Rename fncp (unlink) 1690 */ 1691 _cache_unlink_parent(fncp); 1692 oname = fncp->nc_name; 1693 fncp->nc_name = nname; 1694 fncp->nc_nlen = tncp->nc_nlen; 1695 if (oname) 1696 kfree(oname, M_VFSCACHE); 1697 1698 tncp_par = tncp->nc_parent; 1699 _cache_hold(tncp_par); 1700 _cache_lock(tncp_par); 1701 1702 /* 1703 * Rename fncp (relink) 1704 */ 1705 hash = fnv_32_buf(fncp->nc_name, fncp->nc_nlen, FNV1_32_INIT); 1706 hash = fnv_32_buf(&tncp_par, sizeof(tncp_par), hash); 1707 nchpp = NCHHASH(hash); 1708 1709 spin_lock(&nchpp->spin); 1710 _cache_link_parent(fncp, tncp_par, nchpp); 1711 spin_unlock(&nchpp->spin); 1712 1713 _cache_put(tncp_par); 1714 1715 /* 1716 * Get rid of the overwritten tncp (unlink) 1717 */ 1718 _cache_unlink(tncp); 1719 } 1720 1721 /* 1722 * Perform actions consistent with unlinking a file. The passed-in ncp 1723 * must be locked. 1724 * 1725 * The ncp is marked DESTROYED so it no longer shows up in searches, 1726 * and will be physically deleted when the vnode goes away. 1727 * 1728 * If the related vnode has no refs then we cycle it through vget()/vput() 1729 * to (possibly if we don't have a ref race) trigger a deactivation, 1730 * allowing the VFS to trivially detect and recycle the deleted vnode 1731 * via VOP_INACTIVE(). 1732 * 1733 * NOTE: _cache_rename() will automatically call _cache_unlink() on the 1734 * target ncp. 1735 */ 1736 void 1737 cache_unlink(struct nchandle *nch) 1738 { 1739 _cache_unlink(nch->ncp); 1740 } 1741 1742 static void 1743 _cache_unlink(struct namecache *ncp) 1744 { 1745 struct vnode *vp; 1746 1747 /* 1748 * Causes lookups to fail and allows another ncp with the same 1749 * name to be created under ncp->nc_parent. 1750 */ 1751 ncp->nc_flag |= NCF_DESTROYED; 1752 1753 /* 1754 * Attempt to trigger a deactivation. 1755 */ 1756 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 && 1757 (vp = ncp->nc_vp) != NULL && 1758 !sysref_isactive(&vp->v_sysref)) { 1759 if (vget(vp, LK_SHARED) == 0) 1760 vput(vp); 1761 } 1762 } 1763 1764 /* 1765 * vget the vnode associated with the namecache entry. Resolve the namecache 1766 * entry if necessary. The passed ncp must be referenced and locked. 1767 * 1768 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked 1769 * (depending on the passed lk_type) will be returned in *vpp with an error 1770 * of 0, or NULL will be returned in *vpp with a non-0 error code. The 1771 * most typical error is ENOENT, meaning that the ncp represents a negative 1772 * cache hit and there is no vnode to retrieve, but other errors can occur 1773 * too. 1774 * 1775 * The vget() can race a reclaim. If this occurs we re-resolve the 1776 * namecache entry. 1777 * 1778 * There are numerous places in the kernel where vget() is called on a 1779 * vnode while one or more of its namecache entries is locked. Releasing 1780 * a vnode never deadlocks against locked namecache entries (the vnode 1781 * will not get recycled while referenced ncp's exist). This means we 1782 * can safely acquire the vnode. In fact, we MUST NOT release the ncp 1783 * lock when acquiring the vp lock or we might cause a deadlock. 1784 * 1785 * NOTE: The passed-in ncp must be locked exclusively if it is initially 1786 * unresolved. If a reclaim race occurs the passed-in ncp will be 1787 * relocked exclusively before being re-resolved. 1788 */ 1789 int 1790 cache_vget(struct nchandle *nch, struct ucred *cred, 1791 int lk_type, struct vnode **vpp) 1792 { 1793 struct namecache *ncp; 1794 struct vnode *vp; 1795 int error; 1796 1797 ncp = nch->ncp; 1798 again: 1799 vp = NULL; 1800 if (ncp->nc_flag & NCF_UNRESOLVED) 1801 error = cache_resolve(nch, cred); 1802 else 1803 error = 0; 1804 1805 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 1806 error = vget(vp, lk_type); 1807 if (error) { 1808 /* 1809 * VRECLAIM race 1810 */ 1811 if (error == ENOENT) { 1812 kprintf("Warning: vnode reclaim race detected " 1813 "in cache_vget on %p (%s)\n", 1814 vp, ncp->nc_name); 1815 _cache_unlock(ncp); 1816 _cache_lock(ncp); 1817 _cache_setunresolved(ncp); 1818 goto again; 1819 } 1820 1821 /* 1822 * Not a reclaim race, some other error. 1823 */ 1824 KKASSERT(ncp->nc_vp == vp); 1825 vp = NULL; 1826 } else { 1827 KKASSERT(ncp->nc_vp == vp); 1828 KKASSERT((vp->v_flag & VRECLAIMED) == 0); 1829 } 1830 } 1831 if (error == 0 && vp == NULL) 1832 error = ENOENT; 1833 *vpp = vp; 1834 return(error); 1835 } 1836 1837 /* 1838 * Similar to cache_vget() but only acquires a ref on the vnode. 1839 * 1840 * NOTE: The passed-in ncp must be locked exclusively if it is initially 1841 * unresolved. If a reclaim race occurs the passed-in ncp will be 1842 * relocked exclusively before being re-resolved. 1843 */ 1844 int 1845 cache_vref(struct nchandle *nch, struct ucred *cred, struct vnode **vpp) 1846 { 1847 struct namecache *ncp; 1848 struct vnode *vp; 1849 int error; 1850 1851 ncp = nch->ncp; 1852 again: 1853 vp = NULL; 1854 if (ncp->nc_flag & NCF_UNRESOLVED) 1855 error = cache_resolve(nch, cred); 1856 else 1857 error = 0; 1858 1859 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 1860 error = vget(vp, LK_SHARED); 1861 if (error) { 1862 /* 1863 * VRECLAIM race 1864 */ 1865 if (error == ENOENT) { 1866 kprintf("Warning: vnode reclaim race detected " 1867 "in cache_vget on %p (%s)\n", 1868 vp, ncp->nc_name); 1869 _cache_unlock(ncp); 1870 _cache_lock(ncp); 1871 _cache_setunresolved(ncp); 1872 goto again; 1873 } 1874 1875 /* 1876 * Not a reclaim race, some other error. 1877 */ 1878 KKASSERT(ncp->nc_vp == vp); 1879 vp = NULL; 1880 } else { 1881 KKASSERT(ncp->nc_vp == vp); 1882 KKASSERT((vp->v_flag & VRECLAIMED) == 0); 1883 /* caller does not want a lock */ 1884 vn_unlock(vp); 1885 } 1886 } 1887 if (error == 0 && vp == NULL) 1888 error = ENOENT; 1889 *vpp = vp; 1890 return(error); 1891 } 1892 1893 /* 1894 * Return a referenced vnode representing the parent directory of 1895 * ncp. 1896 * 1897 * Because the caller has locked the ncp it should not be possible for 1898 * the parent ncp to go away. However, the parent can unresolve its 1899 * dvp at any time so we must be able to acquire a lock on the parent 1900 * to safely access nc_vp. 1901 * 1902 * We have to leave par unlocked when vget()ing dvp to avoid a deadlock, 1903 * so use vhold()/vdrop() while holding the lock to prevent dvp from 1904 * getting destroyed. 1905 * 1906 * NOTE: vhold() is allowed when dvp has 0 refs if we hold a 1907 * lock on the ncp in question.. 1908 */ 1909 static struct vnode * 1910 cache_dvpref(struct namecache *ncp) 1911 { 1912 struct namecache *par; 1913 struct vnode *dvp; 1914 1915 dvp = NULL; 1916 if ((par = ncp->nc_parent) != NULL) { 1917 _cache_hold(par); 1918 _cache_lock(par); 1919 if ((par->nc_flag & NCF_UNRESOLVED) == 0) { 1920 if ((dvp = par->nc_vp) != NULL) 1921 vhold(dvp); 1922 } 1923 _cache_unlock(par); 1924 if (dvp) { 1925 if (vget(dvp, LK_SHARED) == 0) { 1926 vn_unlock(dvp); 1927 vdrop(dvp); 1928 /* return refd, unlocked dvp */ 1929 } else { 1930 vdrop(dvp); 1931 dvp = NULL; 1932 } 1933 } 1934 _cache_drop(par); 1935 } 1936 return(dvp); 1937 } 1938 1939 /* 1940 * Convert a directory vnode to a namecache record without any other 1941 * knowledge of the topology. This ONLY works with directory vnodes and 1942 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the 1943 * returned ncp (if not NULL) will be held and unlocked. 1944 * 1945 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned. 1946 * If 'makeit' is 1 we attempt to track-down and create the namecache topology 1947 * for dvp. This will fail only if the directory has been deleted out from 1948 * under the caller. 1949 * 1950 * Callers must always check for a NULL return no matter the value of 'makeit'. 1951 * 1952 * To avoid underflowing the kernel stack each recursive call increments 1953 * the makeit variable. 1954 */ 1955 1956 static int cache_inefficient_scan(struct nchandle *nch, struct ucred *cred, 1957 struct vnode *dvp, char *fakename); 1958 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred, 1959 struct vnode **saved_dvp); 1960 1961 int 1962 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit, 1963 struct nchandle *nch) 1964 { 1965 struct vnode *saved_dvp; 1966 struct vnode *pvp; 1967 char *fakename; 1968 int error; 1969 1970 nch->ncp = NULL; 1971 nch->mount = dvp->v_mount; 1972 saved_dvp = NULL; 1973 fakename = NULL; 1974 1975 /* 1976 * Handle the makeit == 0 degenerate case 1977 */ 1978 if (makeit == 0) { 1979 spin_lock(&dvp->v_spin); 1980 nch->ncp = TAILQ_FIRST(&dvp->v_namecache); 1981 if (nch->ncp) 1982 cache_hold(nch); 1983 spin_unlock(&dvp->v_spin); 1984 } 1985 1986 /* 1987 * Loop until resolution, inside code will break out on error. 1988 */ 1989 while (makeit) { 1990 /* 1991 * Break out if we successfully acquire a working ncp. 1992 */ 1993 spin_lock(&dvp->v_spin); 1994 nch->ncp = TAILQ_FIRST(&dvp->v_namecache); 1995 if (nch->ncp) { 1996 cache_hold(nch); 1997 spin_unlock(&dvp->v_spin); 1998 break; 1999 } 2000 spin_unlock(&dvp->v_spin); 2001 2002 /* 2003 * If dvp is the root of its filesystem it should already 2004 * have a namecache pointer associated with it as a side 2005 * effect of the mount, but it may have been disassociated. 2006 */ 2007 if (dvp->v_flag & VROOT) { 2008 nch->ncp = _cache_get(nch->mount->mnt_ncmountpt.ncp); 2009 error = cache_resolve_mp(nch->mount); 2010 _cache_put(nch->ncp); 2011 if (ncvp_debug) { 2012 kprintf("cache_fromdvp: resolve root of mount %p error %d", 2013 dvp->v_mount, error); 2014 } 2015 if (error) { 2016 if (ncvp_debug) 2017 kprintf(" failed\n"); 2018 nch->ncp = NULL; 2019 break; 2020 } 2021 if (ncvp_debug) 2022 kprintf(" succeeded\n"); 2023 continue; 2024 } 2025 2026 /* 2027 * If we are recursed too deeply resort to an O(n^2) 2028 * algorithm to resolve the namecache topology. The 2029 * resolved pvp is left referenced in saved_dvp to 2030 * prevent the tree from being destroyed while we loop. 2031 */ 2032 if (makeit > 20) { 2033 error = cache_fromdvp_try(dvp, cred, &saved_dvp); 2034 if (error) { 2035 kprintf("lookupdotdot(longpath) failed %d " 2036 "dvp %p\n", error, dvp); 2037 nch->ncp = NULL; 2038 break; 2039 } 2040 continue; 2041 } 2042 2043 /* 2044 * Get the parent directory and resolve its ncp. 2045 */ 2046 if (fakename) { 2047 kfree(fakename, M_TEMP); 2048 fakename = NULL; 2049 } 2050 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred, 2051 &fakename); 2052 if (error) { 2053 kprintf("lookupdotdot failed %d dvp %p\n", error, dvp); 2054 break; 2055 } 2056 vn_unlock(pvp); 2057 2058 /* 2059 * Reuse makeit as a recursion depth counter. On success 2060 * nch will be fully referenced. 2061 */ 2062 cache_fromdvp(pvp, cred, makeit + 1, nch); 2063 vrele(pvp); 2064 if (nch->ncp == NULL) 2065 break; 2066 2067 /* 2068 * Do an inefficient scan of pvp (embodied by ncp) to look 2069 * for dvp. This will create a namecache record for dvp on 2070 * success. We loop up to recheck on success. 2071 * 2072 * ncp and dvp are both held but not locked. 2073 */ 2074 error = cache_inefficient_scan(nch, cred, dvp, fakename); 2075 if (error) { 2076 kprintf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n", 2077 pvp, nch->ncp->nc_name, dvp); 2078 cache_drop(nch); 2079 /* nch was NULLed out, reload mount */ 2080 nch->mount = dvp->v_mount; 2081 break; 2082 } 2083 if (ncvp_debug) { 2084 kprintf("cache_fromdvp: scan %p (%s) succeeded\n", 2085 pvp, nch->ncp->nc_name); 2086 } 2087 cache_drop(nch); 2088 /* nch was NULLed out, reload mount */ 2089 nch->mount = dvp->v_mount; 2090 } 2091 2092 /* 2093 * If nch->ncp is non-NULL it will have been held already. 2094 */ 2095 if (fakename) 2096 kfree(fakename, M_TEMP); 2097 if (saved_dvp) 2098 vrele(saved_dvp); 2099 if (nch->ncp) 2100 return (0); 2101 return (EINVAL); 2102 } 2103 2104 /* 2105 * Go up the chain of parent directories until we find something 2106 * we can resolve into the namecache. This is very inefficient. 2107 */ 2108 static 2109 int 2110 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred, 2111 struct vnode **saved_dvp) 2112 { 2113 struct nchandle nch; 2114 struct vnode *pvp; 2115 int error; 2116 static time_t last_fromdvp_report; 2117 char *fakename; 2118 2119 /* 2120 * Loop getting the parent directory vnode until we get something we 2121 * can resolve in the namecache. 2122 */ 2123 vref(dvp); 2124 nch.mount = dvp->v_mount; 2125 nch.ncp = NULL; 2126 fakename = NULL; 2127 2128 for (;;) { 2129 if (fakename) { 2130 kfree(fakename, M_TEMP); 2131 fakename = NULL; 2132 } 2133 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred, 2134 &fakename); 2135 if (error) { 2136 vrele(dvp); 2137 break; 2138 } 2139 vn_unlock(pvp); 2140 spin_lock(&pvp->v_spin); 2141 if ((nch.ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) { 2142 _cache_hold(nch.ncp); 2143 spin_unlock(&pvp->v_spin); 2144 vrele(pvp); 2145 break; 2146 } 2147 spin_unlock(&pvp->v_spin); 2148 if (pvp->v_flag & VROOT) { 2149 nch.ncp = _cache_get(pvp->v_mount->mnt_ncmountpt.ncp); 2150 error = cache_resolve_mp(nch.mount); 2151 _cache_unlock(nch.ncp); 2152 vrele(pvp); 2153 if (error) { 2154 _cache_drop(nch.ncp); 2155 nch.ncp = NULL; 2156 vrele(dvp); 2157 } 2158 break; 2159 } 2160 vrele(dvp); 2161 dvp = pvp; 2162 } 2163 if (error == 0) { 2164 if (last_fromdvp_report != time_second) { 2165 last_fromdvp_report = time_second; 2166 kprintf("Warning: extremely inefficient path " 2167 "resolution on %s\n", 2168 nch.ncp->nc_name); 2169 } 2170 error = cache_inefficient_scan(&nch, cred, dvp, fakename); 2171 2172 /* 2173 * Hopefully dvp now has a namecache record associated with 2174 * it. Leave it referenced to prevent the kernel from 2175 * recycling the vnode. Otherwise extremely long directory 2176 * paths could result in endless recycling. 2177 */ 2178 if (*saved_dvp) 2179 vrele(*saved_dvp); 2180 *saved_dvp = dvp; 2181 _cache_drop(nch.ncp); 2182 } 2183 if (fakename) 2184 kfree(fakename, M_TEMP); 2185 return (error); 2186 } 2187 2188 /* 2189 * Do an inefficient scan of the directory represented by ncp looking for 2190 * the directory vnode dvp. ncp must be held but not locked on entry and 2191 * will be held on return. dvp must be refd but not locked on entry and 2192 * will remain refd on return. 2193 * 2194 * Why do this at all? Well, due to its stateless nature the NFS server 2195 * converts file handles directly to vnodes without necessarily going through 2196 * the namecache ops that would otherwise create the namecache topology 2197 * leading to the vnode. We could either (1) Change the namecache algorithms 2198 * to allow disconnect namecache records that are re-merged opportunistically, 2199 * or (2) Make the NFS server backtrack and scan to recover a connected 2200 * namecache topology in order to then be able to issue new API lookups. 2201 * 2202 * It turns out that (1) is a huge mess. It takes a nice clean set of 2203 * namecache algorithms and introduces a lot of complication in every subsystem 2204 * that calls into the namecache to deal with the re-merge case, especially 2205 * since we are using the namecache to placehold negative lookups and the 2206 * vnode might not be immediately assigned. (2) is certainly far less 2207 * efficient then (1), but since we are only talking about directories here 2208 * (which are likely to remain cached), the case does not actually run all 2209 * that often and has the supreme advantage of not polluting the namecache 2210 * algorithms. 2211 * 2212 * If a fakename is supplied just construct a namecache entry using the 2213 * fake name. 2214 */ 2215 static int 2216 cache_inefficient_scan(struct nchandle *nch, struct ucred *cred, 2217 struct vnode *dvp, char *fakename) 2218 { 2219 struct nlcomponent nlc; 2220 struct nchandle rncp; 2221 struct dirent *den; 2222 struct vnode *pvp; 2223 struct vattr vat; 2224 struct iovec iov; 2225 struct uio uio; 2226 int blksize; 2227 int eofflag; 2228 int bytes; 2229 char *rbuf; 2230 int error; 2231 2232 vat.va_blocksize = 0; 2233 if ((error = VOP_GETATTR(dvp, &vat)) != 0) 2234 return (error); 2235 cache_lock(nch); 2236 error = cache_vref(nch, cred, &pvp); 2237 cache_unlock(nch); 2238 if (error) 2239 return (error); 2240 if (ncvp_debug) { 2241 kprintf("inefficient_scan: directory iosize %ld " 2242 "vattr fileid = %lld\n", 2243 vat.va_blocksize, 2244 (long long)vat.va_fileid); 2245 } 2246 2247 /* 2248 * Use the supplied fakename if not NULL. Fake names are typically 2249 * not in the actual filesystem hierarchy. This is used by HAMMER 2250 * to glue @@timestamp recursions together. 2251 */ 2252 if (fakename) { 2253 nlc.nlc_nameptr = fakename; 2254 nlc.nlc_namelen = strlen(fakename); 2255 rncp = cache_nlookup(nch, &nlc); 2256 goto done; 2257 } 2258 2259 if ((blksize = vat.va_blocksize) == 0) 2260 blksize = DEV_BSIZE; 2261 rbuf = kmalloc(blksize, M_TEMP, M_WAITOK); 2262 rncp.ncp = NULL; 2263 2264 eofflag = 0; 2265 uio.uio_offset = 0; 2266 again: 2267 iov.iov_base = rbuf; 2268 iov.iov_len = blksize; 2269 uio.uio_iov = &iov; 2270 uio.uio_iovcnt = 1; 2271 uio.uio_resid = blksize; 2272 uio.uio_segflg = UIO_SYSSPACE; 2273 uio.uio_rw = UIO_READ; 2274 uio.uio_td = curthread; 2275 2276 if (ncvp_debug >= 2) 2277 kprintf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset); 2278 error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL); 2279 if (error == 0) { 2280 den = (struct dirent *)rbuf; 2281 bytes = blksize - uio.uio_resid; 2282 2283 while (bytes > 0) { 2284 if (ncvp_debug >= 2) { 2285 kprintf("cache_inefficient_scan: %*.*s\n", 2286 den->d_namlen, den->d_namlen, 2287 den->d_name); 2288 } 2289 if (den->d_type != DT_WHT && 2290 den->d_ino == vat.va_fileid) { 2291 if (ncvp_debug) { 2292 kprintf("cache_inefficient_scan: " 2293 "MATCHED inode %lld path %s/%*.*s\n", 2294 (long long)vat.va_fileid, 2295 nch->ncp->nc_name, 2296 den->d_namlen, den->d_namlen, 2297 den->d_name); 2298 } 2299 nlc.nlc_nameptr = den->d_name; 2300 nlc.nlc_namelen = den->d_namlen; 2301 rncp = cache_nlookup(nch, &nlc); 2302 KKASSERT(rncp.ncp != NULL); 2303 break; 2304 } 2305 bytes -= _DIRENT_DIRSIZ(den); 2306 den = _DIRENT_NEXT(den); 2307 } 2308 if (rncp.ncp == NULL && eofflag == 0 && uio.uio_resid != blksize) 2309 goto again; 2310 } 2311 kfree(rbuf, M_TEMP); 2312 done: 2313 vrele(pvp); 2314 if (rncp.ncp) { 2315 if (rncp.ncp->nc_flag & NCF_UNRESOLVED) { 2316 _cache_setvp(rncp.mount, rncp.ncp, dvp); 2317 if (ncvp_debug >= 2) { 2318 kprintf("cache_inefficient_scan: setvp %s/%s = %p\n", 2319 nch->ncp->nc_name, rncp.ncp->nc_name, dvp); 2320 } 2321 } else { 2322 if (ncvp_debug >= 2) { 2323 kprintf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n", 2324 nch->ncp->nc_name, rncp.ncp->nc_name, dvp, 2325 rncp.ncp->nc_vp); 2326 } 2327 } 2328 if (rncp.ncp->nc_vp == NULL) 2329 error = rncp.ncp->nc_error; 2330 /* 2331 * Release rncp after a successful nlookup. rncp was fully 2332 * referenced. 2333 */ 2334 cache_put(&rncp); 2335 } else { 2336 kprintf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n", 2337 dvp, nch->ncp->nc_name); 2338 error = ENOENT; 2339 } 2340 return (error); 2341 } 2342 2343 /* 2344 * Zap a namecache entry. The ncp is unconditionally set to an unresolved 2345 * state, which disassociates it from its vnode or ncneglist. 2346 * 2347 * Then, if there are no additional references to the ncp and no children, 2348 * the ncp is removed from the topology and destroyed. 2349 * 2350 * References and/or children may exist if the ncp is in the middle of the 2351 * topology, preventing the ncp from being destroyed. 2352 * 2353 * This function must be called with the ncp held and locked and will unlock 2354 * and drop it during zapping. 2355 * 2356 * If nonblock is non-zero and the parent ncp cannot be locked we give up. 2357 * This case can occur in the cache_drop() path. 2358 * 2359 * This function may returned a held (but NOT locked) parent node which the 2360 * caller must drop. We do this so _cache_drop() can loop, to avoid 2361 * blowing out the kernel stack. 2362 * 2363 * WARNING! For MPSAFE operation this routine must acquire up to three 2364 * spin locks to be able to safely test nc_refs. Lock order is 2365 * very important. 2366 * 2367 * hash spinlock if on hash list 2368 * parent spinlock if child of parent 2369 * (the ncp is unresolved so there is no vnode association) 2370 */ 2371 static struct namecache * 2372 cache_zap(struct namecache *ncp, int nonblock) 2373 { 2374 struct namecache *par; 2375 struct vnode *dropvp; 2376 int refs; 2377 2378 /* 2379 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED. 2380 */ 2381 _cache_setunresolved(ncp); 2382 2383 /* 2384 * Try to scrap the entry and possibly tail-recurse on its parent. 2385 * We only scrap unref'd (other then our ref) unresolved entries, 2386 * we do not scrap 'live' entries. 2387 * 2388 * Note that once the spinlocks are acquired if nc_refs == 1 no 2389 * other references are possible. If it isn't, however, we have 2390 * to decrement but also be sure to avoid a 1->0 transition. 2391 */ 2392 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED); 2393 KKASSERT(ncp->nc_refs > 0); 2394 2395 /* 2396 * Acquire locks. Note that the parent can't go away while we hold 2397 * a child locked. 2398 */ 2399 if ((par = ncp->nc_parent) != NULL) { 2400 if (nonblock) { 2401 for (;;) { 2402 if (_cache_lock_nonblock(par) == 0) 2403 break; 2404 refs = ncp->nc_refs; 2405 ncp->nc_flag |= NCF_DEFEREDZAP; 2406 ++numdefered; /* MP race ok */ 2407 if (atomic_cmpset_int(&ncp->nc_refs, 2408 refs, refs - 1)) { 2409 _cache_unlock(ncp); 2410 return(NULL); 2411 } 2412 cpu_pause(); 2413 } 2414 _cache_hold(par); 2415 } else { 2416 _cache_hold(par); 2417 _cache_lock(par); 2418 } 2419 spin_lock(&ncp->nc_head->spin); 2420 } 2421 2422 /* 2423 * If someone other then us has a ref or we have children 2424 * we cannot zap the entry. The 1->0 transition and any 2425 * further list operation is protected by the spinlocks 2426 * we have acquired but other transitions are not. 2427 */ 2428 for (;;) { 2429 refs = ncp->nc_refs; 2430 if (refs == 1 && TAILQ_EMPTY(&ncp->nc_list)) 2431 break; 2432 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) { 2433 if (par) { 2434 spin_unlock(&ncp->nc_head->spin); 2435 _cache_put(par); 2436 } 2437 _cache_unlock(ncp); 2438 return(NULL); 2439 } 2440 cpu_pause(); 2441 } 2442 2443 /* 2444 * We are the only ref and with the spinlocks held no further 2445 * refs can be acquired by others. 2446 * 2447 * Remove us from the hash list and parent list. We have to 2448 * drop a ref on the parent's vp if the parent's list becomes 2449 * empty. 2450 */ 2451 dropvp = NULL; 2452 if (par) { 2453 struct nchash_head *nchpp = ncp->nc_head; 2454 2455 KKASSERT(nchpp != NULL); 2456 LIST_REMOVE(ncp, nc_hash); 2457 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 2458 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 2459 dropvp = par->nc_vp; 2460 ncp->nc_head = NULL; 2461 ncp->nc_parent = NULL; 2462 spin_unlock(&nchpp->spin); 2463 _cache_unlock(par); 2464 } else { 2465 KKASSERT(ncp->nc_head == NULL); 2466 } 2467 2468 /* 2469 * ncp should not have picked up any refs. Physically 2470 * destroy the ncp. 2471 */ 2472 KKASSERT(ncp->nc_refs == 1); 2473 /* _cache_unlock(ncp) not required */ 2474 ncp->nc_refs = -1; /* safety */ 2475 if (ncp->nc_name) 2476 kfree(ncp->nc_name, M_VFSCACHE); 2477 kfree(ncp, M_VFSCACHE); 2478 2479 /* 2480 * Delayed drop (we had to release our spinlocks) 2481 * 2482 * The refed parent (if not NULL) must be dropped. The 2483 * caller is responsible for looping. 2484 */ 2485 if (dropvp) 2486 vdrop(dropvp); 2487 return(par); 2488 } 2489 2490 /* 2491 * Clean up dangling negative cache and defered-drop entries in the 2492 * namecache. 2493 * 2494 * This routine is called in the critical path and also called from 2495 * vnlru(). When called from vnlru we use a lower limit to try to 2496 * deal with the negative cache before the critical path has to start 2497 * dealing with it. 2498 */ 2499 typedef enum { CHI_LOW, CHI_HIGH } cache_hs_t; 2500 2501 static cache_hs_t neg_cache_hysteresis_state[2] = { CHI_LOW, CHI_LOW }; 2502 static cache_hs_t pos_cache_hysteresis_state[2] = { CHI_LOW, CHI_LOW }; 2503 2504 void 2505 cache_hysteresis(int critpath) 2506 { 2507 int poslimit; 2508 int neglimit = desiredvnodes / ncnegfactor; 2509 int xnumcache = numcache; 2510 2511 if (critpath == 0) 2512 neglimit = neglimit * 8 / 10; 2513 2514 /* 2515 * Don't cache too many negative hits. We use hysteresis to reduce 2516 * the impact on the critical path. 2517 */ 2518 switch(neg_cache_hysteresis_state[critpath]) { 2519 case CHI_LOW: 2520 if (numneg > MINNEG && numneg > neglimit) { 2521 if (critpath) 2522 _cache_cleanneg(ncnegflush); 2523 else 2524 _cache_cleanneg(ncnegflush + 2525 numneg - neglimit); 2526 neg_cache_hysteresis_state[critpath] = CHI_HIGH; 2527 } 2528 break; 2529 case CHI_HIGH: 2530 if (numneg > MINNEG * 9 / 10 && 2531 numneg * 9 / 10 > neglimit 2532 ) { 2533 if (critpath) 2534 _cache_cleanneg(ncnegflush); 2535 else 2536 _cache_cleanneg(ncnegflush + 2537 numneg * 9 / 10 - neglimit); 2538 } else { 2539 neg_cache_hysteresis_state[critpath] = CHI_LOW; 2540 } 2541 break; 2542 } 2543 2544 /* 2545 * Don't cache too many positive hits. We use hysteresis to reduce 2546 * the impact on the critical path. 2547 * 2548 * Excessive positive hits can accumulate due to large numbers of 2549 * hardlinks (the vnode cache will not prevent hl ncps from growing 2550 * into infinity). 2551 */ 2552 if ((poslimit = ncposlimit) == 0) 2553 poslimit = desiredvnodes * 2; 2554 if (critpath == 0) 2555 poslimit = poslimit * 8 / 10; 2556 2557 switch(pos_cache_hysteresis_state[critpath]) { 2558 case CHI_LOW: 2559 if (xnumcache > poslimit && xnumcache > MINPOS) { 2560 if (critpath) 2561 _cache_cleanpos(ncposflush); 2562 else 2563 _cache_cleanpos(ncposflush + 2564 xnumcache - poslimit); 2565 pos_cache_hysteresis_state[critpath] = CHI_HIGH; 2566 } 2567 break; 2568 case CHI_HIGH: 2569 if (xnumcache > poslimit * 5 / 6 && xnumcache > MINPOS) { 2570 if (critpath) 2571 _cache_cleanpos(ncposflush); 2572 else 2573 _cache_cleanpos(ncposflush + 2574 xnumcache - poslimit * 5 / 6); 2575 } else { 2576 pos_cache_hysteresis_state[critpath] = CHI_LOW; 2577 } 2578 break; 2579 } 2580 2581 /* 2582 * Clean out dangling defered-zap ncps which could not 2583 * be cleanly dropped if too many build up. Note 2584 * that numdefered is not an exact number as such ncps 2585 * can be reused and the counter is not handled in a MP 2586 * safe manner by design. 2587 */ 2588 if (numdefered > neglimit) { 2589 _cache_cleandefered(); 2590 } 2591 } 2592 2593 /* 2594 * NEW NAMECACHE LOOKUP API 2595 * 2596 * Lookup an entry in the namecache. The passed par_nch must be referenced 2597 * and unlocked. A referenced and locked nchandle with a non-NULL nch.ncp 2598 * is ALWAYS returned, eve if the supplied component is illegal. 2599 * 2600 * The resulting namecache entry should be returned to the system with 2601 * cache_put() or cache_unlock() + cache_drop(). 2602 * 2603 * namecache locks are recursive but care must be taken to avoid lock order 2604 * reversals (hence why the passed par_nch must be unlocked). Locking 2605 * rules are to order for parent traversals, not for child traversals. 2606 * 2607 * Nobody else will be able to manipulate the associated namespace (e.g. 2608 * create, delete, rename, rename-target) until the caller unlocks the 2609 * entry. 2610 * 2611 * The returned entry will be in one of three states: positive hit (non-null 2612 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set). 2613 * Unresolved entries must be resolved through the filesystem to associate the 2614 * vnode and/or determine whether a positive or negative hit has occured. 2615 * 2616 * It is not necessary to lock a directory in order to lock namespace under 2617 * that directory. In fact, it is explicitly not allowed to do that. A 2618 * directory is typically only locked when being created, renamed, or 2619 * destroyed. 2620 * 2621 * The directory (par) may be unresolved, in which case any returned child 2622 * will likely also be marked unresolved. Likely but not guarenteed. Since 2623 * the filesystem lookup requires a resolved directory vnode the caller is 2624 * responsible for resolving the namecache chain top-down. This API 2625 * specifically allows whole chains to be created in an unresolved state. 2626 */ 2627 struct nchandle 2628 cache_nlookup(struct nchandle *par_nch, struct nlcomponent *nlc) 2629 { 2630 struct nchandle nch; 2631 struct namecache *ncp; 2632 struct namecache *new_ncp; 2633 struct nchash_head *nchpp; 2634 struct mount *mp; 2635 u_int32_t hash; 2636 globaldata_t gd; 2637 int par_locked; 2638 2639 numcalls++; 2640 gd = mycpu; 2641 mp = par_nch->mount; 2642 par_locked = 0; 2643 2644 /* 2645 * This is a good time to call it, no ncp's are locked by 2646 * the caller or us. 2647 */ 2648 cache_hysteresis(1); 2649 2650 /* 2651 * Try to locate an existing entry 2652 */ 2653 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 2654 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash); 2655 new_ncp = NULL; 2656 nchpp = NCHHASH(hash); 2657 restart: 2658 spin_lock(&nchpp->spin); 2659 LIST_FOREACH(ncp, &nchpp->list, nc_hash) { 2660 numchecks++; 2661 2662 /* 2663 * Break out if we find a matching entry. Note that 2664 * UNRESOLVED entries may match, but DESTROYED entries 2665 * do not. 2666 */ 2667 if (ncp->nc_parent == par_nch->ncp && 2668 ncp->nc_nlen == nlc->nlc_namelen && 2669 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 2670 (ncp->nc_flag & NCF_DESTROYED) == 0 2671 ) { 2672 _cache_hold(ncp); 2673 spin_unlock(&nchpp->spin); 2674 if (par_locked) { 2675 _cache_unlock(par_nch->ncp); 2676 par_locked = 0; 2677 } 2678 if (_cache_lock_special(ncp) == 0) { 2679 _cache_auto_unresolve(mp, ncp); 2680 if (new_ncp) 2681 _cache_free(new_ncp); 2682 goto found; 2683 } 2684 _cache_get(ncp); 2685 _cache_put(ncp); 2686 _cache_drop(ncp); 2687 goto restart; 2688 } 2689 } 2690 2691 /* 2692 * We failed to locate an entry, create a new entry and add it to 2693 * the cache. The parent ncp must also be locked so we 2694 * can link into it. 2695 * 2696 * We have to relookup after possibly blocking in kmalloc or 2697 * when locking par_nch. 2698 * 2699 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special 2700 * mount case, in which case nc_name will be NULL. 2701 */ 2702 if (new_ncp == NULL) { 2703 spin_unlock(&nchpp->spin); 2704 new_ncp = cache_alloc(nlc->nlc_namelen); 2705 if (nlc->nlc_namelen) { 2706 bcopy(nlc->nlc_nameptr, new_ncp->nc_name, 2707 nlc->nlc_namelen); 2708 new_ncp->nc_name[nlc->nlc_namelen] = 0; 2709 } 2710 goto restart; 2711 } 2712 if (par_locked == 0) { 2713 spin_unlock(&nchpp->spin); 2714 _cache_lock(par_nch->ncp); 2715 par_locked = 1; 2716 goto restart; 2717 } 2718 2719 /* 2720 * WARNING! We still hold the spinlock. We have to set the hash 2721 * table entry atomically. 2722 */ 2723 ncp = new_ncp; 2724 _cache_link_parent(ncp, par_nch->ncp, nchpp); 2725 spin_unlock(&nchpp->spin); 2726 _cache_unlock(par_nch->ncp); 2727 /* par_locked = 0 - not used */ 2728 found: 2729 /* 2730 * stats and namecache size management 2731 */ 2732 if (ncp->nc_flag & NCF_UNRESOLVED) 2733 ++gd->gd_nchstats->ncs_miss; 2734 else if (ncp->nc_vp) 2735 ++gd->gd_nchstats->ncs_goodhits; 2736 else 2737 ++gd->gd_nchstats->ncs_neghits; 2738 nch.mount = mp; 2739 nch.ncp = ncp; 2740 atomic_add_int(&nch.mount->mnt_refs, 1); 2741 return(nch); 2742 } 2743 2744 /* 2745 * Attempt to lookup a namecache entry and return with a shared namecache 2746 * lock. 2747 */ 2748 int 2749 cache_nlookup_maybe_shared(struct nchandle *par_nch, struct nlcomponent *nlc, 2750 int excl, struct nchandle *res_nch) 2751 { 2752 struct namecache *ncp; 2753 struct nchash_head *nchpp; 2754 struct mount *mp; 2755 u_int32_t hash; 2756 globaldata_t gd; 2757 2758 /* 2759 * If exclusive requested or shared namecache locks are disabled, 2760 * return failure. 2761 */ 2762 if (ncp_shared_lock_disable || excl) 2763 return(EWOULDBLOCK); 2764 2765 numcalls++; 2766 gd = mycpu; 2767 mp = par_nch->mount; 2768 2769 /* 2770 * This is a good time to call it, no ncp's are locked by 2771 * the caller or us. 2772 */ 2773 cache_hysteresis(1); 2774 2775 /* 2776 * Try to locate an existing entry 2777 */ 2778 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 2779 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash); 2780 nchpp = NCHHASH(hash); 2781 2782 spin_lock(&nchpp->spin); 2783 2784 LIST_FOREACH(ncp, &nchpp->list, nc_hash) { 2785 numchecks++; 2786 2787 /* 2788 * Break out if we find a matching entry. Note that 2789 * UNRESOLVED entries may match, but DESTROYED entries 2790 * do not. 2791 */ 2792 if (ncp->nc_parent == par_nch->ncp && 2793 ncp->nc_nlen == nlc->nlc_namelen && 2794 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 2795 (ncp->nc_flag & NCF_DESTROYED) == 0 2796 ) { 2797 _cache_hold(ncp); 2798 spin_unlock(&nchpp->spin); 2799 if (_cache_lock_shared_special(ncp) == 0) { 2800 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 && 2801 (ncp->nc_flag & NCF_DESTROYED) == 0 && 2802 _cache_auto_unresolve_test(mp, ncp) == 0) { 2803 goto found; 2804 } 2805 _cache_unlock(ncp); 2806 } 2807 _cache_drop(ncp); 2808 spin_lock(&nchpp->spin); 2809 break; 2810 } 2811 } 2812 2813 /* 2814 * Failure 2815 */ 2816 spin_unlock(&nchpp->spin); 2817 return(EWOULDBLOCK); 2818 2819 /* 2820 * Success 2821 * 2822 * Note that nc_error might be non-zero (e.g ENOENT). 2823 */ 2824 found: 2825 res_nch->mount = mp; 2826 res_nch->ncp = ncp; 2827 ++gd->gd_nchstats->ncs_goodhits; 2828 atomic_add_int(&res_nch->mount->mnt_refs, 1); 2829 2830 KKASSERT(ncp->nc_error != EWOULDBLOCK); 2831 return(ncp->nc_error); 2832 } 2833 2834 /* 2835 * This is a non-blocking verison of cache_nlookup() used by 2836 * nfs_readdirplusrpc_uio(). It can fail for any reason and 2837 * will return nch.ncp == NULL in that case. 2838 */ 2839 struct nchandle 2840 cache_nlookup_nonblock(struct nchandle *par_nch, struct nlcomponent *nlc) 2841 { 2842 struct nchandle nch; 2843 struct namecache *ncp; 2844 struct namecache *new_ncp; 2845 struct nchash_head *nchpp; 2846 struct mount *mp; 2847 u_int32_t hash; 2848 globaldata_t gd; 2849 int par_locked; 2850 2851 numcalls++; 2852 gd = mycpu; 2853 mp = par_nch->mount; 2854 par_locked = 0; 2855 2856 /* 2857 * Try to locate an existing entry 2858 */ 2859 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 2860 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash); 2861 new_ncp = NULL; 2862 nchpp = NCHHASH(hash); 2863 restart: 2864 spin_lock(&nchpp->spin); 2865 LIST_FOREACH(ncp, &nchpp->list, nc_hash) { 2866 numchecks++; 2867 2868 /* 2869 * Break out if we find a matching entry. Note that 2870 * UNRESOLVED entries may match, but DESTROYED entries 2871 * do not. 2872 */ 2873 if (ncp->nc_parent == par_nch->ncp && 2874 ncp->nc_nlen == nlc->nlc_namelen && 2875 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 2876 (ncp->nc_flag & NCF_DESTROYED) == 0 2877 ) { 2878 _cache_hold(ncp); 2879 spin_unlock(&nchpp->spin); 2880 if (par_locked) { 2881 _cache_unlock(par_nch->ncp); 2882 par_locked = 0; 2883 } 2884 if (_cache_lock_special(ncp) == 0) { 2885 _cache_auto_unresolve(mp, ncp); 2886 if (new_ncp) { 2887 _cache_free(new_ncp); 2888 new_ncp = NULL; 2889 } 2890 goto found; 2891 } 2892 _cache_drop(ncp); 2893 goto failed; 2894 } 2895 } 2896 2897 /* 2898 * We failed to locate an entry, create a new entry and add it to 2899 * the cache. The parent ncp must also be locked so we 2900 * can link into it. 2901 * 2902 * We have to relookup after possibly blocking in kmalloc or 2903 * when locking par_nch. 2904 * 2905 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special 2906 * mount case, in which case nc_name will be NULL. 2907 */ 2908 if (new_ncp == NULL) { 2909 spin_unlock(&nchpp->spin); 2910 new_ncp = cache_alloc(nlc->nlc_namelen); 2911 if (nlc->nlc_namelen) { 2912 bcopy(nlc->nlc_nameptr, new_ncp->nc_name, 2913 nlc->nlc_namelen); 2914 new_ncp->nc_name[nlc->nlc_namelen] = 0; 2915 } 2916 goto restart; 2917 } 2918 if (par_locked == 0) { 2919 spin_unlock(&nchpp->spin); 2920 if (_cache_lock_nonblock(par_nch->ncp) == 0) { 2921 par_locked = 1; 2922 goto restart; 2923 } 2924 goto failed; 2925 } 2926 2927 /* 2928 * WARNING! We still hold the spinlock. We have to set the hash 2929 * table entry atomically. 2930 */ 2931 ncp = new_ncp; 2932 _cache_link_parent(ncp, par_nch->ncp, nchpp); 2933 spin_unlock(&nchpp->spin); 2934 _cache_unlock(par_nch->ncp); 2935 /* par_locked = 0 - not used */ 2936 found: 2937 /* 2938 * stats and namecache size management 2939 */ 2940 if (ncp->nc_flag & NCF_UNRESOLVED) 2941 ++gd->gd_nchstats->ncs_miss; 2942 else if (ncp->nc_vp) 2943 ++gd->gd_nchstats->ncs_goodhits; 2944 else 2945 ++gd->gd_nchstats->ncs_neghits; 2946 nch.mount = mp; 2947 nch.ncp = ncp; 2948 atomic_add_int(&nch.mount->mnt_refs, 1); 2949 return(nch); 2950 failed: 2951 if (new_ncp) { 2952 _cache_free(new_ncp); 2953 new_ncp = NULL; 2954 } 2955 nch.mount = NULL; 2956 nch.ncp = NULL; 2957 return(nch); 2958 } 2959 2960 /* 2961 * The namecache entry is marked as being used as a mount point. 2962 * Locate the mount if it is visible to the caller. The DragonFly 2963 * mount system allows arbitrary loops in the topology and disentangles 2964 * those loops by matching against (mp, ncp) rather than just (ncp). 2965 * This means any given ncp can dive any number of mounts, depending 2966 * on the relative mount (e.g. nullfs) the caller is at in the topology. 2967 * 2968 * We use a very simple frontend cache to reduce SMP conflicts, 2969 * which we have to do because the mountlist scan needs an exclusive 2970 * lock around its ripout info list. Not to mention that there might 2971 * be a lot of mounts. 2972 */ 2973 struct findmount_info { 2974 struct mount *result; 2975 struct mount *nch_mount; 2976 struct namecache *nch_ncp; 2977 }; 2978 2979 static 2980 struct ncmount_cache * 2981 ncmount_cache_lookup(struct mount *mp, struct namecache *ncp) 2982 { 2983 int hash; 2984 2985 hash = ((int)(intptr_t)mp / sizeof(*mp)) ^ 2986 ((int)(intptr_t)ncp / sizeof(*ncp)); 2987 hash = (hash & 0x7FFFFFFF) % NCMOUNT_NUMCACHE; 2988 return (&ncmount_cache[hash]); 2989 } 2990 2991 static 2992 int 2993 cache_findmount_callback(struct mount *mp, void *data) 2994 { 2995 struct findmount_info *info = data; 2996 2997 /* 2998 * Check the mount's mounted-on point against the passed nch. 2999 */ 3000 if (mp->mnt_ncmounton.mount == info->nch_mount && 3001 mp->mnt_ncmounton.ncp == info->nch_ncp 3002 ) { 3003 info->result = mp; 3004 atomic_add_int(&mp->mnt_refs, 1); 3005 return(-1); 3006 } 3007 return(0); 3008 } 3009 3010 struct mount * 3011 cache_findmount(struct nchandle *nch) 3012 { 3013 struct findmount_info info; 3014 struct ncmount_cache *ncc; 3015 struct mount *mp; 3016 3017 /* 3018 * Fast 3019 */ 3020 if (ncmount_cache_enable == 0) { 3021 ncc = NULL; 3022 goto skip; 3023 } 3024 ncc = ncmount_cache_lookup(nch->mount, nch->ncp); 3025 if (ncc->ncp == nch->ncp) { 3026 spin_lock_shared(&ncc->spin); 3027 if (ncc->isneg == 0 && 3028 ncc->ncp == nch->ncp && (mp = ncc->mp) != NULL) { 3029 if (mp->mnt_ncmounton.mount == nch->mount && 3030 mp->mnt_ncmounton.ncp == nch->ncp) { 3031 /* 3032 * Cache hit (positive) 3033 */ 3034 atomic_add_int(&mp->mnt_refs, 1); 3035 spin_unlock_shared(&ncc->spin); 3036 ++ncmount_cache_hit; 3037 return(mp); 3038 } 3039 /* else cache miss */ 3040 } 3041 if (ncc->isneg && 3042 ncc->ncp == nch->ncp && ncc->mp == nch->mount) { 3043 /* 3044 * Cache hit (negative) 3045 */ 3046 spin_unlock_shared(&ncc->spin); 3047 ++ncmount_cache_hit; 3048 return(NULL); 3049 } 3050 spin_unlock_shared(&ncc->spin); 3051 } 3052 skip: 3053 3054 /* 3055 * Slow 3056 */ 3057 info.result = NULL; 3058 info.nch_mount = nch->mount; 3059 info.nch_ncp = nch->ncp; 3060 mountlist_scan(cache_findmount_callback, &info, 3061 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 3062 3063 /* 3064 * Cache the result. 3065 * 3066 * Negative lookups: We cache the originating {ncp,mp}. (mp) is 3067 * only used for pointer comparisons and is not 3068 * referenced (otherwise there would be dangling 3069 * refs). 3070 * 3071 * Positive lookups: We cache the originating {ncp} and the target 3072 * (mp). (mp) is referenced. 3073 * 3074 * Indeterminant: If the match is undergoing an unmount we do 3075 * not cache it to avoid racing cache_unmounting(), 3076 * but still return the match. 3077 */ 3078 if (ncc) { 3079 spin_lock(&ncc->spin); 3080 if (info.result == NULL) { 3081 if (ncc->isneg == 0 && ncc->mp) 3082 atomic_add_int(&ncc->mp->mnt_refs, -1); 3083 ncc->ncp = nch->ncp; 3084 ncc->mp = nch->mount; 3085 ncc->isneg = 1; 3086 spin_unlock(&ncc->spin); 3087 ++ncmount_cache_overwrite; 3088 } else if ((info.result->mnt_kern_flag & MNTK_UNMOUNT) == 0) { 3089 if (ncc->isneg == 0 && ncc->mp) 3090 atomic_add_int(&ncc->mp->mnt_refs, -1); 3091 atomic_add_int(&info.result->mnt_refs, 1); 3092 ncc->ncp = nch->ncp; 3093 ncc->mp = info.result; 3094 ncc->isneg = 0; 3095 spin_unlock(&ncc->spin); 3096 ++ncmount_cache_overwrite; 3097 } else { 3098 spin_unlock(&ncc->spin); 3099 } 3100 ++ncmount_cache_miss; 3101 } 3102 return(info.result); 3103 } 3104 3105 void 3106 cache_dropmount(struct mount *mp) 3107 { 3108 atomic_add_int(&mp->mnt_refs, -1); 3109 } 3110 3111 void 3112 cache_ismounting(struct mount *mp) 3113 { 3114 struct nchandle *nch = &mp->mnt_ncmounton; 3115 struct ncmount_cache *ncc; 3116 3117 ncc = ncmount_cache_lookup(nch->mount, nch->ncp); 3118 if (ncc->isneg && 3119 ncc->ncp == nch->ncp && ncc->mp == nch->mount) { 3120 spin_lock(&ncc->spin); 3121 if (ncc->isneg && 3122 ncc->ncp == nch->ncp && ncc->mp == nch->mount) { 3123 ncc->ncp = NULL; 3124 ncc->mp = NULL; 3125 } 3126 spin_unlock(&ncc->spin); 3127 } 3128 } 3129 3130 void 3131 cache_unmounting(struct mount *mp) 3132 { 3133 struct nchandle *nch = &mp->mnt_ncmounton; 3134 struct ncmount_cache *ncc; 3135 3136 ncc = ncmount_cache_lookup(nch->mount, nch->ncp); 3137 if (ncc->isneg == 0 && 3138 ncc->ncp == nch->ncp && ncc->mp == mp) { 3139 spin_lock(&ncc->spin); 3140 if (ncc->isneg == 0 && 3141 ncc->ncp == nch->ncp && ncc->mp == mp) { 3142 atomic_add_int(&mp->mnt_refs, -1); 3143 ncc->ncp = NULL; 3144 ncc->mp = NULL; 3145 } 3146 spin_unlock(&ncc->spin); 3147 } 3148 } 3149 3150 /* 3151 * Resolve an unresolved namecache entry, generally by looking it up. 3152 * The passed ncp must be locked and refd. 3153 * 3154 * Theoretically since a vnode cannot be recycled while held, and since 3155 * the nc_parent chain holds its vnode as long as children exist, the 3156 * direct parent of the cache entry we are trying to resolve should 3157 * have a valid vnode. If not then generate an error that we can 3158 * determine is related to a resolver bug. 3159 * 3160 * However, if a vnode was in the middle of a recyclement when the NCP 3161 * got locked, ncp->nc_vp might point to a vnode that is about to become 3162 * invalid. cache_resolve() handles this case by unresolving the entry 3163 * and then re-resolving it. 3164 * 3165 * Note that successful resolution does not necessarily return an error 3166 * code of 0. If the ncp resolves to a negative cache hit then ENOENT 3167 * will be returned. 3168 */ 3169 int 3170 cache_resolve(struct nchandle *nch, struct ucred *cred) 3171 { 3172 struct namecache *par_tmp; 3173 struct namecache *par; 3174 struct namecache *ncp; 3175 struct nchandle nctmp; 3176 struct mount *mp; 3177 struct vnode *dvp; 3178 int error; 3179 3180 ncp = nch->ncp; 3181 mp = nch->mount; 3182 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE); 3183 restart: 3184 /* 3185 * If the ncp is already resolved we have nothing to do. However, 3186 * we do want to guarentee that a usable vnode is returned when 3187 * a vnode is present, so make sure it hasn't been reclaimed. 3188 */ 3189 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 3190 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 3191 _cache_setunresolved(ncp); 3192 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 3193 return (ncp->nc_error); 3194 } 3195 3196 /* 3197 * If the ncp was destroyed it will never resolve again. This 3198 * can basically only happen when someone is chdir'd into an 3199 * empty directory which is then rmdir'd. We want to catch this 3200 * here and not dive the VFS because the VFS might actually 3201 * have a way to re-resolve the disconnected ncp, which will 3202 * result in inconsistencies in the cdir/nch for proc->p_fd. 3203 */ 3204 if (ncp->nc_flag & NCF_DESTROYED) { 3205 kprintf("Warning: cache_resolve: ncp '%s' was unlinked\n", 3206 ncp->nc_name); 3207 return(EINVAL); 3208 } 3209 3210 /* 3211 * Mount points need special handling because the parent does not 3212 * belong to the same filesystem as the ncp. 3213 */ 3214 if (ncp == mp->mnt_ncmountpt.ncp) 3215 return (cache_resolve_mp(mp)); 3216 3217 /* 3218 * We expect an unbroken chain of ncps to at least the mount point, 3219 * and even all the way to root (but this code doesn't have to go 3220 * past the mount point). 3221 */ 3222 if (ncp->nc_parent == NULL) { 3223 kprintf("EXDEV case 1 %p %*.*s\n", ncp, 3224 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 3225 ncp->nc_error = EXDEV; 3226 return(ncp->nc_error); 3227 } 3228 3229 /* 3230 * The vp's of the parent directories in the chain are held via vhold() 3231 * due to the existance of the child, and should not disappear. 3232 * However, there are cases where they can disappear: 3233 * 3234 * - due to filesystem I/O errors. 3235 * - due to NFS being stupid about tracking the namespace and 3236 * destroys the namespace for entire directories quite often. 3237 * - due to forced unmounts. 3238 * - due to an rmdir (parent will be marked DESTROYED) 3239 * 3240 * When this occurs we have to track the chain backwards and resolve 3241 * it, looping until the resolver catches up to the current node. We 3242 * could recurse here but we might run ourselves out of kernel stack 3243 * so we do it in a more painful manner. This situation really should 3244 * not occur all that often, or if it does not have to go back too 3245 * many nodes to resolve the ncp. 3246 */ 3247 while ((dvp = cache_dvpref(ncp)) == NULL) { 3248 /* 3249 * This case can occur if a process is CD'd into a 3250 * directory which is then rmdir'd. If the parent is marked 3251 * destroyed there is no point trying to resolve it. 3252 */ 3253 if (ncp->nc_parent->nc_flag & NCF_DESTROYED) 3254 return(ENOENT); 3255 par = ncp->nc_parent; 3256 _cache_hold(par); 3257 _cache_lock(par); 3258 while ((par_tmp = par->nc_parent) != NULL && 3259 par_tmp->nc_vp == NULL) { 3260 _cache_hold(par_tmp); 3261 _cache_lock(par_tmp); 3262 _cache_put(par); 3263 par = par_tmp; 3264 } 3265 if (par->nc_parent == NULL) { 3266 kprintf("EXDEV case 2 %*.*s\n", 3267 par->nc_nlen, par->nc_nlen, par->nc_name); 3268 _cache_put(par); 3269 return (EXDEV); 3270 } 3271 kprintf("[diagnostic] cache_resolve: had to recurse on %*.*s\n", 3272 par->nc_nlen, par->nc_nlen, par->nc_name); 3273 /* 3274 * The parent is not set in stone, ref and lock it to prevent 3275 * it from disappearing. Also note that due to renames it 3276 * is possible for our ncp to move and for par to no longer 3277 * be one of its parents. We resolve it anyway, the loop 3278 * will handle any moves. 3279 */ 3280 _cache_get(par); /* additional hold/lock */ 3281 _cache_put(par); /* from earlier hold/lock */ 3282 if (par == nch->mount->mnt_ncmountpt.ncp) { 3283 cache_resolve_mp(nch->mount); 3284 } else if ((dvp = cache_dvpref(par)) == NULL) { 3285 kprintf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name); 3286 _cache_put(par); 3287 continue; 3288 } else { 3289 if (par->nc_flag & NCF_UNRESOLVED) { 3290 nctmp.mount = mp; 3291 nctmp.ncp = par; 3292 par->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred); 3293 } 3294 vrele(dvp); 3295 } 3296 if ((error = par->nc_error) != 0) { 3297 if (par->nc_error != EAGAIN) { 3298 kprintf("EXDEV case 3 %*.*s error %d\n", 3299 par->nc_nlen, par->nc_nlen, par->nc_name, 3300 par->nc_error); 3301 _cache_put(par); 3302 return(error); 3303 } 3304 kprintf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n", 3305 par, par->nc_nlen, par->nc_nlen, par->nc_name); 3306 } 3307 _cache_put(par); 3308 /* loop */ 3309 } 3310 3311 /* 3312 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected 3313 * ncp's and reattach them. If this occurs the original ncp is marked 3314 * EAGAIN to force a relookup. 3315 * 3316 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed 3317 * ncp must already be resolved. 3318 */ 3319 if (dvp) { 3320 nctmp.mount = mp; 3321 nctmp.ncp = ncp; 3322 ncp->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred); 3323 vrele(dvp); 3324 } else { 3325 ncp->nc_error = EPERM; 3326 } 3327 if (ncp->nc_error == EAGAIN) { 3328 kprintf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n", 3329 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 3330 goto restart; 3331 } 3332 return(ncp->nc_error); 3333 } 3334 3335 /* 3336 * Resolve the ncp associated with a mount point. Such ncp's almost always 3337 * remain resolved and this routine is rarely called. NFS MPs tends to force 3338 * re-resolution more often due to its mac-truck-smash-the-namecache 3339 * method of tracking namespace changes. 3340 * 3341 * The semantics for this call is that the passed ncp must be locked on 3342 * entry and will be locked on return. However, if we actually have to 3343 * resolve the mount point we temporarily unlock the entry in order to 3344 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of 3345 * the unlock we have to recheck the flags after we relock. 3346 */ 3347 static int 3348 cache_resolve_mp(struct mount *mp) 3349 { 3350 struct namecache *ncp = mp->mnt_ncmountpt.ncp; 3351 struct vnode *vp; 3352 int error; 3353 3354 KKASSERT(mp != NULL); 3355 3356 /* 3357 * If the ncp is already resolved we have nothing to do. However, 3358 * we do want to guarentee that a usable vnode is returned when 3359 * a vnode is present, so make sure it hasn't been reclaimed. 3360 */ 3361 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 3362 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 3363 _cache_setunresolved(ncp); 3364 } 3365 3366 if (ncp->nc_flag & NCF_UNRESOLVED) { 3367 _cache_unlock(ncp); 3368 while (vfs_busy(mp, 0)) 3369 ; 3370 error = VFS_ROOT(mp, &vp); 3371 _cache_lock(ncp); 3372 3373 /* 3374 * recheck the ncp state after relocking. 3375 */ 3376 if (ncp->nc_flag & NCF_UNRESOLVED) { 3377 ncp->nc_error = error; 3378 if (error == 0) { 3379 _cache_setvp(mp, ncp, vp); 3380 vput(vp); 3381 } else { 3382 kprintf("[diagnostic] cache_resolve_mp: failed" 3383 " to resolve mount %p err=%d ncp=%p\n", 3384 mp, error, ncp); 3385 _cache_setvp(mp, ncp, NULL); 3386 } 3387 } else if (error == 0) { 3388 vput(vp); 3389 } 3390 vfs_unbusy(mp); 3391 } 3392 return(ncp->nc_error); 3393 } 3394 3395 /* 3396 * Clean out negative cache entries when too many have accumulated. 3397 */ 3398 static void 3399 _cache_cleanneg(int count) 3400 { 3401 struct namecache *ncp; 3402 3403 /* 3404 * Attempt to clean out the specified number of negative cache 3405 * entries. 3406 */ 3407 while (count) { 3408 spin_lock(&ncspin); 3409 ncp = TAILQ_FIRST(&ncneglist); 3410 if (ncp == NULL) { 3411 spin_unlock(&ncspin); 3412 break; 3413 } 3414 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 3415 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 3416 _cache_hold(ncp); 3417 spin_unlock(&ncspin); 3418 3419 /* 3420 * This can race, so we must re-check that the ncp 3421 * is on the ncneglist after successfully locking it. 3422 */ 3423 if (_cache_lock_special(ncp) == 0) { 3424 if (ncp->nc_vp == NULL && 3425 (ncp->nc_flag & NCF_UNRESOLVED) == 0) { 3426 ncp = cache_zap(ncp, 1); 3427 if (ncp) 3428 _cache_drop(ncp); 3429 } else { 3430 kprintf("cache_cleanneg: race avoided\n"); 3431 _cache_unlock(ncp); 3432 } 3433 } else { 3434 _cache_drop(ncp); 3435 } 3436 --count; 3437 } 3438 } 3439 3440 /* 3441 * Clean out positive cache entries when too many have accumulated. 3442 */ 3443 static void 3444 _cache_cleanpos(int count) 3445 { 3446 static volatile int rover; 3447 struct nchash_head *nchpp; 3448 struct namecache *ncp; 3449 int rover_copy; 3450 3451 /* 3452 * Attempt to clean out the specified number of negative cache 3453 * entries. 3454 */ 3455 while (count) { 3456 rover_copy = ++rover; /* MPSAFEENOUGH */ 3457 cpu_ccfence(); 3458 nchpp = NCHHASH(rover_copy); 3459 3460 spin_lock(&nchpp->spin); 3461 ncp = LIST_FIRST(&nchpp->list); 3462 while (ncp && (ncp->nc_flag & NCF_DESTROYED)) 3463 ncp = LIST_NEXT(ncp, nc_hash); 3464 if (ncp) 3465 _cache_hold(ncp); 3466 spin_unlock(&nchpp->spin); 3467 3468 if (ncp) { 3469 if (_cache_lock_special(ncp) == 0) { 3470 ncp = cache_zap(ncp, 1); 3471 if (ncp) 3472 _cache_drop(ncp); 3473 } else { 3474 _cache_drop(ncp); 3475 } 3476 } 3477 --count; 3478 } 3479 } 3480 3481 /* 3482 * This is a kitchen sink function to clean out ncps which we 3483 * tried to zap from cache_drop() but failed because we were 3484 * unable to acquire the parent lock. 3485 * 3486 * Such entries can also be removed via cache_inval_vp(), such 3487 * as when unmounting. 3488 */ 3489 static void 3490 _cache_cleandefered(void) 3491 { 3492 struct nchash_head *nchpp; 3493 struct namecache *ncp; 3494 struct namecache dummy; 3495 int i; 3496 3497 numdefered = 0; 3498 bzero(&dummy, sizeof(dummy)); 3499 dummy.nc_flag = NCF_DESTROYED; 3500 dummy.nc_refs = 1; 3501 3502 for (i = 0; i <= nchash; ++i) { 3503 nchpp = &nchashtbl[i]; 3504 3505 spin_lock(&nchpp->spin); 3506 LIST_INSERT_HEAD(&nchpp->list, &dummy, nc_hash); 3507 ncp = &dummy; 3508 while ((ncp = LIST_NEXT(ncp, nc_hash)) != NULL) { 3509 if ((ncp->nc_flag & NCF_DEFEREDZAP) == 0) 3510 continue; 3511 LIST_REMOVE(&dummy, nc_hash); 3512 LIST_INSERT_AFTER(ncp, &dummy, nc_hash); 3513 _cache_hold(ncp); 3514 spin_unlock(&nchpp->spin); 3515 if (_cache_lock_nonblock(ncp) == 0) { 3516 ncp->nc_flag &= ~NCF_DEFEREDZAP; 3517 _cache_unlock(ncp); 3518 } 3519 _cache_drop(ncp); 3520 spin_lock(&nchpp->spin); 3521 ncp = &dummy; 3522 } 3523 LIST_REMOVE(&dummy, nc_hash); 3524 spin_unlock(&nchpp->spin); 3525 } 3526 } 3527 3528 /* 3529 * Name cache initialization, from vfsinit() when we are booting 3530 */ 3531 void 3532 nchinit(void) 3533 { 3534 int i; 3535 globaldata_t gd; 3536 3537 /* initialise per-cpu namecache effectiveness statistics. */ 3538 for (i = 0; i < ncpus; ++i) { 3539 gd = globaldata_find(i); 3540 gd->gd_nchstats = &nchstats[i]; 3541 } 3542 TAILQ_INIT(&ncneglist); 3543 spin_init(&ncspin); 3544 nchashtbl = hashinit_ext(desiredvnodes / 2, 3545 sizeof(struct nchash_head), 3546 M_VFSCACHE, &nchash); 3547 for (i = 0; i <= (int)nchash; ++i) { 3548 LIST_INIT(&nchashtbl[i].list); 3549 spin_init(&nchashtbl[i].spin); 3550 } 3551 for (i = 0; i < NCMOUNT_NUMCACHE; ++i) 3552 spin_init(&ncmount_cache[i].spin); 3553 nclockwarn = 5 * hz; 3554 } 3555 3556 /* 3557 * Called from start_init() to bootstrap the root filesystem. Returns 3558 * a referenced, unlocked namecache record. 3559 */ 3560 void 3561 cache_allocroot(struct nchandle *nch, struct mount *mp, struct vnode *vp) 3562 { 3563 nch->ncp = cache_alloc(0); 3564 nch->mount = mp; 3565 atomic_add_int(&mp->mnt_refs, 1); 3566 if (vp) 3567 _cache_setvp(nch->mount, nch->ncp, vp); 3568 } 3569 3570 /* 3571 * vfs_cache_setroot() 3572 * 3573 * Create an association between the root of our namecache and 3574 * the root vnode. This routine may be called several times during 3575 * booting. 3576 * 3577 * If the caller intends to save the returned namecache pointer somewhere 3578 * it must cache_hold() it. 3579 */ 3580 void 3581 vfs_cache_setroot(struct vnode *nvp, struct nchandle *nch) 3582 { 3583 struct vnode *ovp; 3584 struct nchandle onch; 3585 3586 ovp = rootvnode; 3587 onch = rootnch; 3588 rootvnode = nvp; 3589 if (nch) 3590 rootnch = *nch; 3591 else 3592 cache_zero(&rootnch); 3593 if (ovp) 3594 vrele(ovp); 3595 if (onch.ncp) 3596 cache_drop(&onch); 3597 } 3598 3599 /* 3600 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache 3601 * topology and is being removed as quickly as possible. The new VOP_N*() 3602 * API calls are required to make specific adjustments using the supplied 3603 * ncp pointers rather then just bogusly purging random vnodes. 3604 * 3605 * Invalidate all namecache entries to a particular vnode as well as 3606 * any direct children of that vnode in the namecache. This is a 3607 * 'catch all' purge used by filesystems that do not know any better. 3608 * 3609 * Note that the linkage between the vnode and its namecache entries will 3610 * be removed, but the namecache entries themselves might stay put due to 3611 * active references from elsewhere in the system or due to the existance of 3612 * the children. The namecache topology is left intact even if we do not 3613 * know what the vnode association is. Such entries will be marked 3614 * NCF_UNRESOLVED. 3615 */ 3616 void 3617 cache_purge(struct vnode *vp) 3618 { 3619 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN); 3620 } 3621 3622 /* 3623 * Flush all entries referencing a particular filesystem. 3624 * 3625 * Since we need to check it anyway, we will flush all the invalid 3626 * entries at the same time. 3627 */ 3628 #if 0 3629 3630 void 3631 cache_purgevfs(struct mount *mp) 3632 { 3633 struct nchash_head *nchpp; 3634 struct namecache *ncp, *nnp; 3635 3636 /* 3637 * Scan hash tables for applicable entries. 3638 */ 3639 for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) { 3640 spin_lock_wr(&nchpp->spin); XXX 3641 ncp = LIST_FIRST(&nchpp->list); 3642 if (ncp) 3643 _cache_hold(ncp); 3644 while (ncp) { 3645 nnp = LIST_NEXT(ncp, nc_hash); 3646 if (nnp) 3647 _cache_hold(nnp); 3648 if (ncp->nc_mount == mp) { 3649 _cache_lock(ncp); 3650 ncp = cache_zap(ncp, 0); 3651 if (ncp) 3652 _cache_drop(ncp); 3653 } else { 3654 _cache_drop(ncp); 3655 } 3656 ncp = nnp; 3657 } 3658 spin_unlock_wr(&nchpp->spin); XXX 3659 } 3660 } 3661 3662 #endif 3663 3664 static int disablecwd; 3665 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, 3666 "Disable getcwd"); 3667 3668 static u_long numcwdcalls; 3669 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdcalls, CTLFLAG_RD, &numcwdcalls, 0, 3670 "Number of current directory resolution calls"); 3671 static u_long numcwdfailnf; 3672 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailnf, CTLFLAG_RD, &numcwdfailnf, 0, 3673 "Number of current directory failures due to lack of file"); 3674 static u_long numcwdfailsz; 3675 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailsz, CTLFLAG_RD, &numcwdfailsz, 0, 3676 "Number of current directory failures due to large result"); 3677 static u_long numcwdfound; 3678 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfound, CTLFLAG_RD, &numcwdfound, 0, 3679 "Number of current directory resolution successes"); 3680 3681 /* 3682 * MPALMOSTSAFE 3683 */ 3684 int 3685 sys___getcwd(struct __getcwd_args *uap) 3686 { 3687 u_int buflen; 3688 int error; 3689 char *buf; 3690 char *bp; 3691 3692 if (disablecwd) 3693 return (ENODEV); 3694 3695 buflen = uap->buflen; 3696 if (buflen == 0) 3697 return (EINVAL); 3698 if (buflen > MAXPATHLEN) 3699 buflen = MAXPATHLEN; 3700 3701 buf = kmalloc(buflen, M_TEMP, M_WAITOK); 3702 bp = kern_getcwd(buf, buflen, &error); 3703 if (error == 0) 3704 error = copyout(bp, uap->buf, strlen(bp) + 1); 3705 kfree(buf, M_TEMP); 3706 return (error); 3707 } 3708 3709 char * 3710 kern_getcwd(char *buf, size_t buflen, int *error) 3711 { 3712 struct proc *p = curproc; 3713 char *bp; 3714 int i, slash_prefixed; 3715 struct filedesc *fdp; 3716 struct nchandle nch; 3717 struct namecache *ncp; 3718 3719 numcwdcalls++; 3720 bp = buf; 3721 bp += buflen - 1; 3722 *bp = '\0'; 3723 fdp = p->p_fd; 3724 slash_prefixed = 0; 3725 3726 nch = fdp->fd_ncdir; 3727 ncp = nch.ncp; 3728 if (ncp) 3729 _cache_hold(ncp); 3730 3731 while (ncp && (ncp != fdp->fd_nrdir.ncp || 3732 nch.mount != fdp->fd_nrdir.mount) 3733 ) { 3734 /* 3735 * While traversing upwards if we encounter the root 3736 * of the current mount we have to skip to the mount point 3737 * in the underlying filesystem. 3738 */ 3739 if (ncp == nch.mount->mnt_ncmountpt.ncp) { 3740 nch = nch.mount->mnt_ncmounton; 3741 _cache_drop(ncp); 3742 ncp = nch.ncp; 3743 if (ncp) 3744 _cache_hold(ncp); 3745 continue; 3746 } 3747 3748 /* 3749 * Prepend the path segment 3750 */ 3751 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 3752 if (bp == buf) { 3753 numcwdfailsz++; 3754 *error = ERANGE; 3755 bp = NULL; 3756 goto done; 3757 } 3758 *--bp = ncp->nc_name[i]; 3759 } 3760 if (bp == buf) { 3761 numcwdfailsz++; 3762 *error = ERANGE; 3763 bp = NULL; 3764 goto done; 3765 } 3766 *--bp = '/'; 3767 slash_prefixed = 1; 3768 3769 /* 3770 * Go up a directory. This isn't a mount point so we don't 3771 * have to check again. 3772 */ 3773 while ((nch.ncp = ncp->nc_parent) != NULL) { 3774 if (ncp_shared_lock_disable) 3775 _cache_lock(ncp); 3776 else 3777 _cache_lock_shared(ncp); 3778 if (nch.ncp != ncp->nc_parent) { 3779 _cache_unlock(ncp); 3780 continue; 3781 } 3782 _cache_hold(nch.ncp); 3783 _cache_unlock(ncp); 3784 break; 3785 } 3786 _cache_drop(ncp); 3787 ncp = nch.ncp; 3788 } 3789 if (ncp == NULL) { 3790 numcwdfailnf++; 3791 *error = ENOENT; 3792 bp = NULL; 3793 goto done; 3794 } 3795 if (!slash_prefixed) { 3796 if (bp == buf) { 3797 numcwdfailsz++; 3798 *error = ERANGE; 3799 bp = NULL; 3800 goto done; 3801 } 3802 *--bp = '/'; 3803 } 3804 numcwdfound++; 3805 *error = 0; 3806 done: 3807 if (ncp) 3808 _cache_drop(ncp); 3809 return (bp); 3810 } 3811 3812 /* 3813 * Thus begins the fullpath magic. 3814 * 3815 * The passed nchp is referenced but not locked. 3816 */ 3817 static int disablefullpath; 3818 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, 3819 &disablefullpath, 0, 3820 "Disable fullpath lookups"); 3821 3822 static u_int numfullpathcalls; 3823 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathcalls, CTLFLAG_RD, 3824 &numfullpathcalls, 0, 3825 "Number of full path resolutions in progress"); 3826 static u_int numfullpathfailnf; 3827 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailnf, CTLFLAG_RD, 3828 &numfullpathfailnf, 0, 3829 "Number of full path resolution failures due to lack of file"); 3830 static u_int numfullpathfailsz; 3831 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailsz, CTLFLAG_RD, 3832 &numfullpathfailsz, 0, 3833 "Number of full path resolution failures due to insufficient memory"); 3834 static u_int numfullpathfound; 3835 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfound, CTLFLAG_RD, 3836 &numfullpathfound, 0, 3837 "Number of full path resolution successes"); 3838 3839 int 3840 cache_fullpath(struct proc *p, struct nchandle *nchp, struct nchandle *nchbase, 3841 char **retbuf, char **freebuf, int guess) 3842 { 3843 struct nchandle fd_nrdir; 3844 struct nchandle nch; 3845 struct namecache *ncp; 3846 struct mount *mp, *new_mp; 3847 char *bp, *buf; 3848 int slash_prefixed; 3849 int error = 0; 3850 int i; 3851 3852 atomic_add_int(&numfullpathcalls, -1); 3853 3854 *retbuf = NULL; 3855 *freebuf = NULL; 3856 3857 buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK); 3858 bp = buf + MAXPATHLEN - 1; 3859 *bp = '\0'; 3860 if (nchbase) 3861 fd_nrdir = *nchbase; 3862 else if (p != NULL) 3863 fd_nrdir = p->p_fd->fd_nrdir; 3864 else 3865 fd_nrdir = rootnch; 3866 slash_prefixed = 0; 3867 nch = *nchp; 3868 ncp = nch.ncp; 3869 if (ncp) 3870 _cache_hold(ncp); 3871 mp = nch.mount; 3872 3873 while (ncp && (ncp != fd_nrdir.ncp || mp != fd_nrdir.mount)) { 3874 new_mp = NULL; 3875 3876 /* 3877 * If we are asked to guess the upwards path, we do so whenever 3878 * we encounter an ncp marked as a mountpoint. We try to find 3879 * the actual mountpoint by finding the mountpoint with this 3880 * ncp. 3881 */ 3882 if (guess && (ncp->nc_flag & NCF_ISMOUNTPT)) { 3883 new_mp = mount_get_by_nc(ncp); 3884 } 3885 /* 3886 * While traversing upwards if we encounter the root 3887 * of the current mount we have to skip to the mount point. 3888 */ 3889 if (ncp == mp->mnt_ncmountpt.ncp) { 3890 new_mp = mp; 3891 } 3892 if (new_mp) { 3893 nch = new_mp->mnt_ncmounton; 3894 _cache_drop(ncp); 3895 ncp = nch.ncp; 3896 if (ncp) 3897 _cache_hold(ncp); 3898 mp = nch.mount; 3899 continue; 3900 } 3901 3902 /* 3903 * Prepend the path segment 3904 */ 3905 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 3906 if (bp == buf) { 3907 numfullpathfailsz++; 3908 kfree(buf, M_TEMP); 3909 error = ENOMEM; 3910 goto done; 3911 } 3912 *--bp = ncp->nc_name[i]; 3913 } 3914 if (bp == buf) { 3915 numfullpathfailsz++; 3916 kfree(buf, M_TEMP); 3917 error = ENOMEM; 3918 goto done; 3919 } 3920 *--bp = '/'; 3921 slash_prefixed = 1; 3922 3923 /* 3924 * Go up a directory. This isn't a mount point so we don't 3925 * have to check again. 3926 * 3927 * We can only safely access nc_parent with ncp held locked. 3928 */ 3929 while ((nch.ncp = ncp->nc_parent) != NULL) { 3930 _cache_lock(ncp); 3931 if (nch.ncp != ncp->nc_parent) { 3932 _cache_unlock(ncp); 3933 continue; 3934 } 3935 _cache_hold(nch.ncp); 3936 _cache_unlock(ncp); 3937 break; 3938 } 3939 _cache_drop(ncp); 3940 ncp = nch.ncp; 3941 } 3942 if (ncp == NULL) { 3943 numfullpathfailnf++; 3944 kfree(buf, M_TEMP); 3945 error = ENOENT; 3946 goto done; 3947 } 3948 3949 if (!slash_prefixed) { 3950 if (bp == buf) { 3951 numfullpathfailsz++; 3952 kfree(buf, M_TEMP); 3953 error = ENOMEM; 3954 goto done; 3955 } 3956 *--bp = '/'; 3957 } 3958 numfullpathfound++; 3959 *retbuf = bp; 3960 *freebuf = buf; 3961 error = 0; 3962 done: 3963 if (ncp) 3964 _cache_drop(ncp); 3965 return(error); 3966 } 3967 3968 int 3969 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf, 3970 int guess) 3971 { 3972 struct namecache *ncp; 3973 struct nchandle nch; 3974 int error; 3975 3976 *freebuf = NULL; 3977 atomic_add_int(&numfullpathcalls, 1); 3978 if (disablefullpath) 3979 return (ENODEV); 3980 3981 if (p == NULL) 3982 return (EINVAL); 3983 3984 /* vn is NULL, client wants us to use p->p_textvp */ 3985 if (vn == NULL) { 3986 if ((vn = p->p_textvp) == NULL) 3987 return (EINVAL); 3988 } 3989 spin_lock(&vn->v_spin); 3990 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) { 3991 if (ncp->nc_nlen) 3992 break; 3993 } 3994 if (ncp == NULL) { 3995 spin_unlock(&vn->v_spin); 3996 return (EINVAL); 3997 } 3998 _cache_hold(ncp); 3999 spin_unlock(&vn->v_spin); 4000 4001 atomic_add_int(&numfullpathcalls, -1); 4002 nch.ncp = ncp; 4003 nch.mount = vn->v_mount; 4004 error = cache_fullpath(p, &nch, NULL, retbuf, freebuf, guess); 4005 _cache_drop(ncp); 4006 return (error); 4007 } 4008