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 = 0; 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 begticks; 302 int error; 303 u_int count; 304 305 KKASSERT(ncp->nc_refs != 0); 306 didwarn = 0; 307 begticks = 0; 308 td = curthread; 309 310 for (;;) { 311 count = ncp->nc_lockstatus; 312 cpu_ccfence(); 313 314 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 0) { 315 if (atomic_cmpset_int(&ncp->nc_lockstatus, 316 count, count + 1)) { 317 /* 318 * The vp associated with a locked ncp must 319 * be held to prevent it from being recycled. 320 * 321 * WARNING! If VRECLAIMED is set the vnode 322 * could already be in the middle of a recycle. 323 * Callers must use cache_vref() or 324 * cache_vget() on the locked ncp to 325 * validate the vp or set the cache entry 326 * to unresolved. 327 * 328 * NOTE! vhold() is allowed if we hold a 329 * lock on the ncp (which we do). 330 */ 331 ncp->nc_locktd = td; 332 if (ncp->nc_vp) 333 vhold(ncp->nc_vp); 334 break; 335 } 336 /* cmpset failed */ 337 continue; 338 } 339 if (ncp->nc_locktd == td) { 340 KKASSERT((count & NC_SHLOCK_FLAG) == 0); 341 if (atomic_cmpset_int(&ncp->nc_lockstatus, 342 count, count + 1)) { 343 break; 344 } 345 /* cmpset failed */ 346 continue; 347 } 348 tsleep_interlock(&ncp->nc_locktd, 0); 349 if (atomic_cmpset_int(&ncp->nc_lockstatus, count, 350 count | NC_EXLOCK_REQ) == 0) { 351 /* cmpset failed */ 352 continue; 353 } 354 if (begticks == 0) 355 begticks = ticks; 356 error = tsleep(&ncp->nc_locktd, PINTERLOCKED, 357 "clock", nclockwarn); 358 if (error == EWOULDBLOCK) { 359 if (didwarn == 0) { 360 didwarn = ticks; 361 kprintf("[diagnostic] cache_lock: " 362 "blocked on %p %08x", 363 ncp, count); 364 kprintf(" \"%*.*s\"\n", 365 ncp->nc_nlen, ncp->nc_nlen, 366 ncp->nc_name); 367 } 368 } 369 /* loop */ 370 } 371 if (didwarn) { 372 kprintf("[diagnostic] cache_lock: unblocked %*.*s after " 373 "%d secs\n", 374 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name, 375 (int)(ticks + (hz / 2) - begticks) / hz); 376 } 377 } 378 379 /* 380 * The shared lock works similarly to the exclusive lock except 381 * nc_locktd is left NULL and we need an interlock (VHOLD) to 382 * prevent vhold() races, since the moment our cmpset_int succeeds 383 * another cpu can come in and get its own shared lock. 384 * 385 * A critical section is needed to prevent interruption during the 386 * VHOLD interlock. 387 */ 388 static 389 void 390 _cache_lock_shared(struct namecache *ncp) 391 { 392 int didwarn; 393 int error; 394 u_int count; 395 u_int optreq = NC_EXLOCK_REQ; 396 397 KKASSERT(ncp->nc_refs != 0); 398 didwarn = 0; 399 400 for (;;) { 401 count = ncp->nc_lockstatus; 402 cpu_ccfence(); 403 404 if ((count & ~NC_SHLOCK_REQ) == 0) { 405 crit_enter(); 406 if (atomic_cmpset_int(&ncp->nc_lockstatus, 407 count, 408 (count + 1) | NC_SHLOCK_FLAG | 409 NC_SHLOCK_VHOLD)) { 410 /* 411 * The vp associated with a locked ncp must 412 * be held to prevent it from being recycled. 413 * 414 * WARNING! If VRECLAIMED is set the vnode 415 * could already be in the middle of a recycle. 416 * Callers must use cache_vref() or 417 * cache_vget() on the locked ncp to 418 * validate the vp or set the cache entry 419 * to unresolved. 420 * 421 * NOTE! vhold() is allowed if we hold a 422 * lock on the ncp (which we do). 423 */ 424 if (ncp->nc_vp) 425 vhold(ncp->nc_vp); 426 atomic_clear_int(&ncp->nc_lockstatus, 427 NC_SHLOCK_VHOLD); 428 crit_exit(); 429 break; 430 } 431 /* cmpset failed */ 432 crit_exit(); 433 continue; 434 } 435 436 /* 437 * If already held shared we can just bump the count, but 438 * only allow this if nobody is trying to get the lock 439 * exclusively. If we are blocking too long ignore excl 440 * requests (which can race/deadlock us). 441 * 442 * VHOLD is a bit of a hack. Even though we successfully 443 * added another shared ref, the cpu that got the first 444 * shared ref might not yet have held the vnode. 445 */ 446 if ((count & (optreq|NC_SHLOCK_FLAG)) == NC_SHLOCK_FLAG) { 447 KKASSERT((count & ~(NC_EXLOCK_REQ | 448 NC_SHLOCK_REQ | 449 NC_SHLOCK_FLAG)) > 0); 450 if (atomic_cmpset_int(&ncp->nc_lockstatus, 451 count, count + 1)) { 452 while (ncp->nc_lockstatus & NC_SHLOCK_VHOLD) 453 cpu_pause(); 454 break; 455 } 456 continue; 457 } 458 tsleep_interlock(ncp, 0); 459 if (atomic_cmpset_int(&ncp->nc_lockstatus, count, 460 count | NC_SHLOCK_REQ) == 0) { 461 /* cmpset failed */ 462 continue; 463 } 464 error = tsleep(ncp, PINTERLOCKED, "clocksh", nclockwarn); 465 if (error == EWOULDBLOCK) { 466 optreq = 0; 467 if (didwarn == 0) { 468 didwarn = ticks; 469 kprintf("[diagnostic] cache_lock_shared: " 470 "blocked on %p %08x", 471 ncp, count); 472 kprintf(" \"%*.*s\"\n", 473 ncp->nc_nlen, ncp->nc_nlen, 474 ncp->nc_name); 475 } 476 } 477 /* loop */ 478 } 479 if (didwarn) { 480 kprintf("[diagnostic] cache_lock_shared: " 481 "unblocked %*.*s after %d secs\n", 482 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name, 483 (int)(ticks - didwarn) / hz); 484 } 485 } 486 487 /* 488 * NOTE: nc_refs may be zero if the ncp is interlocked by circumstance, 489 * such as the case where one of its children is locked. 490 */ 491 static 492 int 493 _cache_lock_nonblock(struct namecache *ncp) 494 { 495 thread_t td; 496 u_int count; 497 498 td = curthread; 499 500 for (;;) { 501 count = ncp->nc_lockstatus; 502 503 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 0) { 504 if (atomic_cmpset_int(&ncp->nc_lockstatus, 505 count, count + 1)) { 506 /* 507 * The vp associated with a locked ncp must 508 * be held to prevent it from being recycled. 509 * 510 * WARNING! If VRECLAIMED is set the vnode 511 * could already be in the middle of a recycle. 512 * Callers must use cache_vref() or 513 * cache_vget() on the locked ncp to 514 * validate the vp or set the cache entry 515 * to unresolved. 516 * 517 * NOTE! vhold() is allowed if we hold a 518 * lock on the ncp (which we do). 519 */ 520 ncp->nc_locktd = td; 521 if (ncp->nc_vp) 522 vhold(ncp->nc_vp); 523 break; 524 } 525 /* cmpset failed */ 526 continue; 527 } 528 if (ncp->nc_locktd == td) { 529 if (atomic_cmpset_int(&ncp->nc_lockstatus, 530 count, count + 1)) { 531 break; 532 } 533 /* cmpset failed */ 534 continue; 535 } 536 return(EWOULDBLOCK); 537 } 538 return(0); 539 } 540 541 /* 542 * The shared lock works similarly to the exclusive lock except 543 * nc_locktd is left NULL and we need an interlock (VHOLD) to 544 * prevent vhold() races, since the moment our cmpset_int succeeds 545 * another cpu can come in and get its own shared lock. 546 * 547 * A critical section is needed to prevent interruption during the 548 * VHOLD interlock. 549 */ 550 static 551 int 552 _cache_lock_shared_nonblock(struct namecache *ncp) 553 { 554 u_int count; 555 556 for (;;) { 557 count = ncp->nc_lockstatus; 558 559 if ((count & ~NC_SHLOCK_REQ) == 0) { 560 crit_enter(); 561 if (atomic_cmpset_int(&ncp->nc_lockstatus, 562 count, 563 (count + 1) | NC_SHLOCK_FLAG | 564 NC_SHLOCK_VHOLD)) { 565 /* 566 * The vp associated with a locked ncp must 567 * be held to prevent it from being recycled. 568 * 569 * WARNING! If VRECLAIMED is set the vnode 570 * could already be in the middle of a recycle. 571 * Callers must use cache_vref() or 572 * cache_vget() on the locked ncp to 573 * validate the vp or set the cache entry 574 * to unresolved. 575 * 576 * NOTE! vhold() is allowed if we hold a 577 * lock on the ncp (which we do). 578 */ 579 if (ncp->nc_vp) 580 vhold(ncp->nc_vp); 581 atomic_clear_int(&ncp->nc_lockstatus, 582 NC_SHLOCK_VHOLD); 583 crit_exit(); 584 break; 585 } 586 /* cmpset failed */ 587 crit_exit(); 588 continue; 589 } 590 591 /* 592 * If already held shared we can just bump the count, but 593 * only allow this if nobody is trying to get the lock 594 * exclusively. 595 * 596 * VHOLD is a bit of a hack. Even though we successfully 597 * added another shared ref, the cpu that got the first 598 * shared ref might not yet have held the vnode. 599 */ 600 if ((count & (NC_EXLOCK_REQ|NC_SHLOCK_FLAG)) == 601 NC_SHLOCK_FLAG) { 602 KKASSERT((count & ~(NC_EXLOCK_REQ | 603 NC_SHLOCK_REQ | 604 NC_SHLOCK_FLAG)) > 0); 605 if (atomic_cmpset_int(&ncp->nc_lockstatus, 606 count, count + 1)) { 607 while (ncp->nc_lockstatus & NC_SHLOCK_VHOLD) 608 cpu_pause(); 609 break; 610 } 611 continue; 612 } 613 return(EWOULDBLOCK); 614 } 615 return(0); 616 } 617 618 /* 619 * Helper function 620 * 621 * NOTE: nc_refs can be 0 (degenerate case during _cache_drop). 622 * 623 * nc_locktd must be NULLed out prior to nc_lockstatus getting cleared. 624 */ 625 static 626 void 627 _cache_unlock(struct namecache *ncp) 628 { 629 thread_t td __debugvar = curthread; 630 u_int count; 631 u_int ncount; 632 struct vnode *dropvp; 633 634 KKASSERT(ncp->nc_refs >= 0); 635 KKASSERT((ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) > 0); 636 KKASSERT((ncp->nc_lockstatus & NC_SHLOCK_FLAG) || ncp->nc_locktd == td); 637 638 count = ncp->nc_lockstatus; 639 cpu_ccfence(); 640 641 /* 642 * Clear nc_locktd prior to the atomic op (excl lock only) 643 */ 644 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 1) 645 ncp->nc_locktd = NULL; 646 dropvp = NULL; 647 648 for (;;) { 649 if ((count & 650 ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ|NC_SHLOCK_FLAG)) == 1) { 651 dropvp = ncp->nc_vp; 652 if (count & NC_EXLOCK_REQ) 653 ncount = count & NC_SHLOCK_REQ; /* cnt->0 */ 654 else 655 ncount = 0; 656 657 if (atomic_cmpset_int(&ncp->nc_lockstatus, 658 count, ncount)) { 659 if (count & NC_EXLOCK_REQ) 660 wakeup(&ncp->nc_locktd); 661 else if (count & NC_SHLOCK_REQ) 662 wakeup(ncp); 663 break; 664 } 665 dropvp = NULL; 666 } else { 667 KKASSERT((count & NC_SHLOCK_VHOLD) == 0); 668 KKASSERT((count & ~(NC_EXLOCK_REQ | 669 NC_SHLOCK_REQ | 670 NC_SHLOCK_FLAG)) > 1); 671 if (atomic_cmpset_int(&ncp->nc_lockstatus, 672 count, count - 1)) { 673 break; 674 } 675 } 676 count = ncp->nc_lockstatus; 677 cpu_ccfence(); 678 } 679 680 /* 681 * Don't actually drop the vp until we successfully clean out 682 * the lock, otherwise we may race another shared lock. 683 */ 684 if (dropvp) 685 vdrop(dropvp); 686 } 687 688 static 689 int 690 _cache_lockstatus(struct namecache *ncp) 691 { 692 if (ncp->nc_locktd == curthread) 693 return(LK_EXCLUSIVE); 694 if (ncp->nc_lockstatus & NC_SHLOCK_FLAG) 695 return(LK_SHARED); 696 return(-1); 697 } 698 699 /* 700 * cache_hold() and cache_drop() prevent the premature deletion of a 701 * namecache entry but do not prevent operations (such as zapping) on 702 * that namecache entry. 703 * 704 * This routine may only be called from outside this source module if 705 * nc_refs is already at least 1. 706 * 707 * This is a rare case where callers are allowed to hold a spinlock, 708 * so we can't ourselves. 709 */ 710 static __inline 711 struct namecache * 712 _cache_hold(struct namecache *ncp) 713 { 714 atomic_add_int(&ncp->nc_refs, 1); 715 return(ncp); 716 } 717 718 /* 719 * Drop a cache entry, taking care to deal with races. 720 * 721 * For potential 1->0 transitions we must hold the ncp lock to safely 722 * test its flags. An unresolved entry with no children must be zapped 723 * to avoid leaks. 724 * 725 * The call to cache_zap() itself will handle all remaining races and 726 * will decrement the ncp's refs regardless. If we are resolved or 727 * have children nc_refs can safely be dropped to 0 without having to 728 * zap the entry. 729 * 730 * NOTE: cache_zap() will re-check nc_refs and nc_list in a MPSAFE fashion. 731 * 732 * NOTE: cache_zap() may return a non-NULL referenced parent which must 733 * be dropped in a loop. 734 */ 735 static __inline 736 void 737 _cache_drop(struct namecache *ncp) 738 { 739 int refs; 740 741 while (ncp) { 742 KKASSERT(ncp->nc_refs > 0); 743 refs = ncp->nc_refs; 744 745 if (refs == 1) { 746 if (_cache_lock_nonblock(ncp) == 0) { 747 ncp->nc_flag &= ~NCF_DEFEREDZAP; 748 if ((ncp->nc_flag & NCF_UNRESOLVED) && 749 TAILQ_EMPTY(&ncp->nc_list)) { 750 ncp = cache_zap(ncp, 1); 751 continue; 752 } 753 if (atomic_cmpset_int(&ncp->nc_refs, 1, 0)) { 754 _cache_unlock(ncp); 755 break; 756 } 757 _cache_unlock(ncp); 758 } 759 } else { 760 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) 761 break; 762 } 763 cpu_pause(); 764 } 765 } 766 767 /* 768 * Link a new namecache entry to its parent and to the hash table. Be 769 * careful to avoid races if vhold() blocks in the future. 770 * 771 * Both ncp and par must be referenced and locked. 772 * 773 * NOTE: The hash table spinlock is held during this call, we can't do 774 * anything fancy. 775 */ 776 static void 777 _cache_link_parent(struct namecache *ncp, struct namecache *par, 778 struct nchash_head *nchpp) 779 { 780 KKASSERT(ncp->nc_parent == NULL); 781 ncp->nc_parent = par; 782 ncp->nc_head = nchpp; 783 784 /* 785 * Set inheritance flags. Note that the parent flags may be 786 * stale due to getattr potentially not having been run yet 787 * (it gets run during nlookup()'s). 788 */ 789 ncp->nc_flag &= ~(NCF_SF_PNOCACHE | NCF_UF_PCACHE); 790 if (par->nc_flag & (NCF_SF_NOCACHE | NCF_SF_PNOCACHE)) 791 ncp->nc_flag |= NCF_SF_PNOCACHE; 792 if (par->nc_flag & (NCF_UF_CACHE | NCF_UF_PCACHE)) 793 ncp->nc_flag |= NCF_UF_PCACHE; 794 795 LIST_INSERT_HEAD(&nchpp->list, ncp, nc_hash); 796 797 if (TAILQ_EMPTY(&par->nc_list)) { 798 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry); 799 /* 800 * Any vp associated with an ncp which has children must 801 * be held to prevent it from being recycled. 802 */ 803 if (par->nc_vp) 804 vhold(par->nc_vp); 805 } else { 806 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry); 807 } 808 } 809 810 /* 811 * Remove the parent and hash associations from a namecache structure. 812 * If this is the last child of the parent the cache_drop(par) will 813 * attempt to recursively zap the parent. 814 * 815 * ncp must be locked. This routine will acquire a temporary lock on 816 * the parent as wlel as the appropriate hash chain. 817 */ 818 static void 819 _cache_unlink_parent(struct namecache *ncp) 820 { 821 struct namecache *par; 822 struct vnode *dropvp; 823 824 if ((par = ncp->nc_parent) != NULL) { 825 KKASSERT(ncp->nc_parent == par); 826 _cache_hold(par); 827 _cache_lock(par); 828 spin_lock(&ncp->nc_head->spin); 829 LIST_REMOVE(ncp, nc_hash); 830 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 831 dropvp = NULL; 832 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 833 dropvp = par->nc_vp; 834 spin_unlock(&ncp->nc_head->spin); 835 ncp->nc_parent = NULL; 836 ncp->nc_head = NULL; 837 _cache_unlock(par); 838 _cache_drop(par); 839 840 /* 841 * We can only safely vdrop with no spinlocks held. 842 */ 843 if (dropvp) 844 vdrop(dropvp); 845 } 846 } 847 848 /* 849 * Allocate a new namecache structure. Most of the code does not require 850 * zero-termination of the string but it makes vop_compat_ncreate() easier. 851 */ 852 static struct namecache * 853 cache_alloc(int nlen) 854 { 855 struct namecache *ncp; 856 857 ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO); 858 if (nlen) 859 ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK); 860 ncp->nc_nlen = nlen; 861 ncp->nc_flag = NCF_UNRESOLVED; 862 ncp->nc_error = ENOTCONN; /* needs to be resolved */ 863 ncp->nc_refs = 1; 864 865 TAILQ_INIT(&ncp->nc_list); 866 _cache_lock(ncp); 867 return(ncp); 868 } 869 870 /* 871 * Can only be called for the case where the ncp has never been 872 * associated with anything (so no spinlocks are needed). 873 */ 874 static void 875 _cache_free(struct namecache *ncp) 876 { 877 KKASSERT(ncp->nc_refs == 1 && ncp->nc_lockstatus == 1); 878 if (ncp->nc_name) 879 kfree(ncp->nc_name, M_VFSCACHE); 880 kfree(ncp, M_VFSCACHE); 881 } 882 883 /* 884 * [re]initialize a nchandle. 885 */ 886 void 887 cache_zero(struct nchandle *nch) 888 { 889 nch->ncp = NULL; 890 nch->mount = NULL; 891 } 892 893 /* 894 * Ref and deref a namecache structure. 895 * 896 * The caller must specify a stable ncp pointer, typically meaning the 897 * ncp is already referenced but this can also occur indirectly through 898 * e.g. holding a lock on a direct child. 899 * 900 * WARNING: Caller may hold an unrelated read spinlock, which means we can't 901 * use read spinlocks here. 902 * 903 * MPSAFE if nch is 904 */ 905 struct nchandle * 906 cache_hold(struct nchandle *nch) 907 { 908 _cache_hold(nch->ncp); 909 atomic_add_int(&nch->mount->mnt_refs, 1); 910 return(nch); 911 } 912 913 /* 914 * Create a copy of a namecache handle for an already-referenced 915 * entry. 916 * 917 * MPSAFE if nch is 918 */ 919 void 920 cache_copy(struct nchandle *nch, struct nchandle *target) 921 { 922 *target = *nch; 923 if (target->ncp) 924 _cache_hold(target->ncp); 925 atomic_add_int(&nch->mount->mnt_refs, 1); 926 } 927 928 /* 929 * MPSAFE if nch is 930 */ 931 void 932 cache_changemount(struct nchandle *nch, struct mount *mp) 933 { 934 atomic_add_int(&nch->mount->mnt_refs, -1); 935 nch->mount = mp; 936 atomic_add_int(&nch->mount->mnt_refs, 1); 937 } 938 939 void 940 cache_drop(struct nchandle *nch) 941 { 942 atomic_add_int(&nch->mount->mnt_refs, -1); 943 _cache_drop(nch->ncp); 944 nch->ncp = NULL; 945 nch->mount = NULL; 946 } 947 948 int 949 cache_lockstatus(struct nchandle *nch) 950 { 951 return(_cache_lockstatus(nch->ncp)); 952 } 953 954 void 955 cache_lock(struct nchandle *nch) 956 { 957 _cache_lock(nch->ncp); 958 } 959 960 void 961 cache_lock_maybe_shared(struct nchandle *nch, int excl) 962 { 963 struct namecache *ncp = nch->ncp; 964 965 if (ncp_shared_lock_disable || excl || 966 (ncp->nc_flag & NCF_UNRESOLVED)) { 967 _cache_lock(ncp); 968 } else { 969 _cache_lock_shared(ncp); 970 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 971 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) { 972 _cache_unlock(ncp); 973 _cache_lock(ncp); 974 } 975 } else { 976 _cache_unlock(ncp); 977 _cache_lock(ncp); 978 } 979 } 980 } 981 982 /* 983 * Relock nch1 given an unlocked nch1 and a locked nch2. The caller 984 * is responsible for checking both for validity on return as they 985 * may have become invalid. 986 * 987 * We have to deal with potential deadlocks here, just ping pong 988 * the lock until we get it (we will always block somewhere when 989 * looping so this is not cpu-intensive). 990 * 991 * which = 0 nch1 not locked, nch2 is locked 992 * which = 1 nch1 is locked, nch2 is not locked 993 */ 994 void 995 cache_relock(struct nchandle *nch1, struct ucred *cred1, 996 struct nchandle *nch2, struct ucred *cred2) 997 { 998 int which; 999 1000 which = 0; 1001 1002 for (;;) { 1003 if (which == 0) { 1004 if (cache_lock_nonblock(nch1) == 0) { 1005 cache_resolve(nch1, cred1); 1006 break; 1007 } 1008 cache_unlock(nch2); 1009 cache_lock(nch1); 1010 cache_resolve(nch1, cred1); 1011 which = 1; 1012 } else { 1013 if (cache_lock_nonblock(nch2) == 0) { 1014 cache_resolve(nch2, cred2); 1015 break; 1016 } 1017 cache_unlock(nch1); 1018 cache_lock(nch2); 1019 cache_resolve(nch2, cred2); 1020 which = 0; 1021 } 1022 } 1023 } 1024 1025 int 1026 cache_lock_nonblock(struct nchandle *nch) 1027 { 1028 return(_cache_lock_nonblock(nch->ncp)); 1029 } 1030 1031 void 1032 cache_unlock(struct nchandle *nch) 1033 { 1034 _cache_unlock(nch->ncp); 1035 } 1036 1037 /* 1038 * ref-and-lock, unlock-and-deref functions. 1039 * 1040 * This function is primarily used by nlookup. Even though cache_lock 1041 * holds the vnode, it is possible that the vnode may have already 1042 * initiated a recyclement. 1043 * 1044 * We want cache_get() to return a definitively usable vnode or a 1045 * definitively unresolved ncp. 1046 */ 1047 static 1048 struct namecache * 1049 _cache_get(struct namecache *ncp) 1050 { 1051 _cache_hold(ncp); 1052 _cache_lock(ncp); 1053 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 1054 _cache_setunresolved(ncp); 1055 return(ncp); 1056 } 1057 1058 /* 1059 * Attempt to obtain a shared lock on the ncp. A shared lock will only 1060 * be obtained if the ncp is resolved and the vnode (if not ENOENT) is 1061 * valid. Otherwise an exclusive lock will be acquired instead. 1062 */ 1063 static 1064 struct namecache * 1065 _cache_get_maybe_shared(struct namecache *ncp, int excl) 1066 { 1067 if (ncp_shared_lock_disable || excl || 1068 (ncp->nc_flag & NCF_UNRESOLVED)) { 1069 return(_cache_get(ncp)); 1070 } 1071 _cache_hold(ncp); 1072 _cache_lock_shared(ncp); 1073 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 1074 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) { 1075 _cache_unlock(ncp); 1076 ncp = _cache_get(ncp); 1077 _cache_drop(ncp); 1078 } 1079 } else { 1080 _cache_unlock(ncp); 1081 ncp = _cache_get(ncp); 1082 _cache_drop(ncp); 1083 } 1084 return(ncp); 1085 } 1086 1087 /* 1088 * This is a special form of _cache_lock() which only succeeds if 1089 * it can get a pristine, non-recursive lock. The caller must have 1090 * already ref'd the ncp. 1091 * 1092 * On success the ncp will be locked, on failure it will not. The 1093 * ref count does not change either way. 1094 * 1095 * We want _cache_lock_special() (on success) to return a definitively 1096 * usable vnode or a definitively unresolved ncp. 1097 */ 1098 static int 1099 _cache_lock_special(struct namecache *ncp) 1100 { 1101 if (_cache_lock_nonblock(ncp) == 0) { 1102 if ((ncp->nc_lockstatus & 1103 ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 1) { 1104 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 1105 _cache_setunresolved(ncp); 1106 return(0); 1107 } 1108 _cache_unlock(ncp); 1109 } 1110 return(EWOULDBLOCK); 1111 } 1112 1113 /* 1114 * This function tries to get a shared lock but will back-off to an exclusive 1115 * lock if: 1116 * 1117 * (1) Some other thread is trying to obtain an exclusive lock 1118 * (to prevent the exclusive requester from getting livelocked out 1119 * by many shared locks). 1120 * 1121 * (2) The current thread already owns an exclusive lock (to avoid 1122 * deadlocking). 1123 * 1124 * WARNING! On machines with lots of cores we really want to try hard to 1125 * get a shared lock or concurrent path lookups can chain-react 1126 * into a very high-latency exclusive lock. 1127 */ 1128 static int 1129 _cache_lock_shared_special(struct namecache *ncp) 1130 { 1131 /* 1132 * Only honor a successful shared lock (returning 0) if there is 1133 * no exclusive request pending and the vnode, if present, is not 1134 * in a reclaimed state. 1135 */ 1136 if (_cache_lock_shared_nonblock(ncp) == 0) { 1137 if ((ncp->nc_lockstatus & NC_EXLOCK_REQ) == 0) { 1138 if (ncp->nc_vp == NULL || 1139 (ncp->nc_vp->v_flag & VRECLAIMED) == 0) { 1140 return(0); 1141 } 1142 } 1143 _cache_unlock(ncp); 1144 return(EWOULDBLOCK); 1145 } 1146 1147 /* 1148 * Non-blocking shared lock failed. If we already own the exclusive 1149 * lock just acquire another exclusive lock (instead of deadlocking). 1150 * Otherwise acquire a shared lock. 1151 */ 1152 if (ncp->nc_locktd == curthread) { 1153 _cache_lock(ncp); 1154 return(0); 1155 } 1156 _cache_lock_shared(ncp); 1157 return(0); 1158 } 1159 1160 1161 /* 1162 * NOTE: The same nchandle can be passed for both arguments. 1163 */ 1164 void 1165 cache_get(struct nchandle *nch, struct nchandle *target) 1166 { 1167 KKASSERT(nch->ncp->nc_refs > 0); 1168 target->mount = nch->mount; 1169 target->ncp = _cache_get(nch->ncp); 1170 atomic_add_int(&target->mount->mnt_refs, 1); 1171 } 1172 1173 void 1174 cache_get_maybe_shared(struct nchandle *nch, struct nchandle *target, int excl) 1175 { 1176 KKASSERT(nch->ncp->nc_refs > 0); 1177 target->mount = nch->mount; 1178 target->ncp = _cache_get_maybe_shared(nch->ncp, excl); 1179 atomic_add_int(&target->mount->mnt_refs, 1); 1180 } 1181 1182 /* 1183 * 1184 */ 1185 static __inline 1186 void 1187 _cache_put(struct namecache *ncp) 1188 { 1189 _cache_unlock(ncp); 1190 _cache_drop(ncp); 1191 } 1192 1193 /* 1194 * 1195 */ 1196 void 1197 cache_put(struct nchandle *nch) 1198 { 1199 atomic_add_int(&nch->mount->mnt_refs, -1); 1200 _cache_put(nch->ncp); 1201 nch->ncp = NULL; 1202 nch->mount = NULL; 1203 } 1204 1205 /* 1206 * Resolve an unresolved ncp by associating a vnode with it. If the 1207 * vnode is NULL, a negative cache entry is created. 1208 * 1209 * The ncp should be locked on entry and will remain locked on return. 1210 */ 1211 static 1212 void 1213 _cache_setvp(struct mount *mp, struct namecache *ncp, struct vnode *vp) 1214 { 1215 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED); 1216 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE); 1217 1218 if (vp != NULL) { 1219 /* 1220 * Any vp associated with an ncp which has children must 1221 * be held. Any vp associated with a locked ncp must be held. 1222 */ 1223 if (!TAILQ_EMPTY(&ncp->nc_list)) 1224 vhold(vp); 1225 spin_lock(&vp->v_spin); 1226 ncp->nc_vp = vp; 1227 TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode); 1228 spin_unlock(&vp->v_spin); 1229 if (ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) 1230 vhold(vp); 1231 1232 /* 1233 * Set auxiliary flags 1234 */ 1235 switch(vp->v_type) { 1236 case VDIR: 1237 ncp->nc_flag |= NCF_ISDIR; 1238 break; 1239 case VLNK: 1240 ncp->nc_flag |= NCF_ISSYMLINK; 1241 /* XXX cache the contents of the symlink */ 1242 break; 1243 default: 1244 break; 1245 } 1246 atomic_add_int(&numcache, 1); 1247 ncp->nc_error = 0; 1248 /* XXX: this is a hack to work-around the lack of a real pfs vfs 1249 * implementation*/ 1250 if (mp != NULL) 1251 if (strncmp(mp->mnt_stat.f_fstypename, "null", 5) == 0) 1252 vp->v_pfsmp = mp; 1253 } else { 1254 /* 1255 * When creating a negative cache hit we set the 1256 * namecache_gen. A later resolve will clean out the 1257 * negative cache hit if the mount point's namecache_gen 1258 * has changed. Used by devfs, could also be used by 1259 * other remote FSs. 1260 */ 1261 ncp->nc_vp = NULL; 1262 spin_lock(&ncspin); 1263 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 1264 ++numneg; 1265 spin_unlock(&ncspin); 1266 ncp->nc_error = ENOENT; 1267 if (mp) 1268 VFS_NCPGEN_SET(mp, ncp); 1269 } 1270 ncp->nc_flag &= ~(NCF_UNRESOLVED | NCF_DEFEREDZAP); 1271 } 1272 1273 /* 1274 * 1275 */ 1276 void 1277 cache_setvp(struct nchandle *nch, struct vnode *vp) 1278 { 1279 _cache_setvp(nch->mount, nch->ncp, vp); 1280 } 1281 1282 /* 1283 * 1284 */ 1285 void 1286 cache_settimeout(struct nchandle *nch, int nticks) 1287 { 1288 struct namecache *ncp = nch->ncp; 1289 1290 if ((ncp->nc_timeout = ticks + nticks) == 0) 1291 ncp->nc_timeout = 1; 1292 } 1293 1294 /* 1295 * Disassociate the vnode or negative-cache association and mark a 1296 * namecache entry as unresolved again. Note that the ncp is still 1297 * left in the hash table and still linked to its parent. 1298 * 1299 * The ncp should be locked and refd on entry and will remain locked and refd 1300 * on return. 1301 * 1302 * This routine is normally never called on a directory containing children. 1303 * However, NFS often does just that in its rename() code as a cop-out to 1304 * avoid complex namespace operations. This disconnects a directory vnode 1305 * from its namecache and can cause the OLDAPI and NEWAPI to get out of 1306 * sync. 1307 * 1308 */ 1309 static 1310 void 1311 _cache_setunresolved(struct namecache *ncp) 1312 { 1313 struct vnode *vp; 1314 1315 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 1316 ncp->nc_flag |= NCF_UNRESOLVED; 1317 ncp->nc_timeout = 0; 1318 ncp->nc_error = ENOTCONN; 1319 if ((vp = ncp->nc_vp) != NULL) { 1320 atomic_add_int(&numcache, -1); 1321 spin_lock(&vp->v_spin); 1322 ncp->nc_vp = NULL; 1323 TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode); 1324 spin_unlock(&vp->v_spin); 1325 1326 /* 1327 * Any vp associated with an ncp with children is 1328 * held by that ncp. Any vp associated with a locked 1329 * ncp is held by that ncp. These conditions must be 1330 * undone when the vp is cleared out from the ncp. 1331 */ 1332 if (!TAILQ_EMPTY(&ncp->nc_list)) 1333 vdrop(vp); 1334 if (ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) 1335 vdrop(vp); 1336 } else { 1337 spin_lock(&ncspin); 1338 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 1339 --numneg; 1340 spin_unlock(&ncspin); 1341 } 1342 ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK); 1343 } 1344 } 1345 1346 /* 1347 * The cache_nresolve() code calls this function to automatically 1348 * set a resolved cache element to unresolved if it has timed out 1349 * or if it is a negative cache hit and the mount point namecache_gen 1350 * has changed. 1351 */ 1352 static __inline int 1353 _cache_auto_unresolve_test(struct mount *mp, struct namecache *ncp) 1354 { 1355 /* 1356 * Try to zap entries that have timed out. We have 1357 * to be careful here because locked leafs may depend 1358 * on the vnode remaining intact in a parent, so only 1359 * do this under very specific conditions. 1360 */ 1361 if (ncp->nc_timeout && (int)(ncp->nc_timeout - ticks) < 0 && 1362 TAILQ_EMPTY(&ncp->nc_list)) { 1363 return 1; 1364 } 1365 1366 /* 1367 * If a resolved negative cache hit is invalid due to 1368 * the mount's namecache generation being bumped, zap it. 1369 */ 1370 if (ncp->nc_vp == NULL && VFS_NCPGEN_TEST(mp, ncp)) { 1371 return 1; 1372 } 1373 1374 /* 1375 * Otherwise we are good 1376 */ 1377 return 0; 1378 } 1379 1380 static __inline void 1381 _cache_auto_unresolve(struct mount *mp, struct namecache *ncp) 1382 { 1383 /* 1384 * Already in an unresolved state, nothing to do. 1385 */ 1386 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 1387 if (_cache_auto_unresolve_test(mp, ncp)) 1388 _cache_setunresolved(ncp); 1389 } 1390 } 1391 1392 /* 1393 * 1394 */ 1395 void 1396 cache_setunresolved(struct nchandle *nch) 1397 { 1398 _cache_setunresolved(nch->ncp); 1399 } 1400 1401 /* 1402 * Determine if we can clear NCF_ISMOUNTPT by scanning the mountlist 1403 * looking for matches. This flag tells the lookup code when it must 1404 * check for a mount linkage and also prevents the directories in question 1405 * from being deleted or renamed. 1406 */ 1407 static 1408 int 1409 cache_clrmountpt_callback(struct mount *mp, void *data) 1410 { 1411 struct nchandle *nch = data; 1412 1413 if (mp->mnt_ncmounton.ncp == nch->ncp) 1414 return(1); 1415 if (mp->mnt_ncmountpt.ncp == nch->ncp) 1416 return(1); 1417 return(0); 1418 } 1419 1420 /* 1421 * 1422 */ 1423 void 1424 cache_clrmountpt(struct nchandle *nch) 1425 { 1426 int count; 1427 1428 count = mountlist_scan(cache_clrmountpt_callback, nch, 1429 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 1430 if (count == 0) 1431 nch->ncp->nc_flag &= ~NCF_ISMOUNTPT; 1432 } 1433 1434 /* 1435 * Invalidate portions of the namecache topology given a starting entry. 1436 * The passed ncp is set to an unresolved state and: 1437 * 1438 * The passed ncp must be referencxed and locked. The routine may unlock 1439 * and relock ncp several times, and will recheck the children and loop 1440 * to catch races. When done the passed ncp will be returned with the 1441 * reference and lock intact. 1442 * 1443 * CINV_DESTROY - Set a flag in the passed ncp entry indicating 1444 * that the physical underlying nodes have been 1445 * destroyed... as in deleted. For example, when 1446 * a directory is removed. This will cause record 1447 * lookups on the name to no longer be able to find 1448 * the record and tells the resolver to return failure 1449 * rather then trying to resolve through the parent. 1450 * 1451 * The topology itself, including ncp->nc_name, 1452 * remains intact. 1453 * 1454 * This only applies to the passed ncp, if CINV_CHILDREN 1455 * is specified the children are not flagged. 1456 * 1457 * CINV_CHILDREN - Set all children (recursively) to an unresolved 1458 * state as well. 1459 * 1460 * Note that this will also have the side effect of 1461 * cleaning out any unreferenced nodes in the topology 1462 * from the leaves up as the recursion backs out. 1463 * 1464 * Note that the topology for any referenced nodes remains intact, but 1465 * the nodes will be marked as having been destroyed and will be set 1466 * to an unresolved state. 1467 * 1468 * It is possible for cache_inval() to race a cache_resolve(), meaning that 1469 * the namecache entry may not actually be invalidated on return if it was 1470 * revalidated while recursing down into its children. This code guarentees 1471 * that the node(s) will go through an invalidation cycle, but does not 1472 * guarentee that they will remain in an invalidated state. 1473 * 1474 * Returns non-zero if a revalidation was detected during the invalidation 1475 * recursion, zero otherwise. Note that since only the original ncp is 1476 * locked the revalidation ultimately can only indicate that the original ncp 1477 * *MIGHT* no have been reresolved. 1478 * 1479 * DEEP RECURSION HANDLING - If a recursive invalidation recurses deeply we 1480 * have to avoid blowing out the kernel stack. We do this by saving the 1481 * deep namecache node and aborting the recursion, then re-recursing at that 1482 * node using a depth-first algorithm in order to allow multiple deep 1483 * recursions to chain through each other, then we restart the invalidation 1484 * from scratch. 1485 */ 1486 1487 struct cinvtrack { 1488 struct namecache *resume_ncp; 1489 int depth; 1490 }; 1491 1492 static int _cache_inval_internal(struct namecache *, int, struct cinvtrack *); 1493 1494 static 1495 int 1496 _cache_inval(struct namecache *ncp, int flags) 1497 { 1498 struct cinvtrack track; 1499 struct namecache *ncp2; 1500 int r; 1501 1502 track.depth = 0; 1503 track.resume_ncp = NULL; 1504 1505 for (;;) { 1506 r = _cache_inval_internal(ncp, flags, &track); 1507 if (track.resume_ncp == NULL) 1508 break; 1509 kprintf("Warning: deep namecache recursion at %s\n", 1510 ncp->nc_name); 1511 _cache_unlock(ncp); 1512 while ((ncp2 = track.resume_ncp) != NULL) { 1513 track.resume_ncp = NULL; 1514 _cache_lock(ncp2); 1515 _cache_inval_internal(ncp2, flags & ~CINV_DESTROY, 1516 &track); 1517 _cache_put(ncp2); 1518 } 1519 _cache_lock(ncp); 1520 } 1521 return(r); 1522 } 1523 1524 int 1525 cache_inval(struct nchandle *nch, int flags) 1526 { 1527 return(_cache_inval(nch->ncp, flags)); 1528 } 1529 1530 /* 1531 * Helper for _cache_inval(). The passed ncp is refd and locked and 1532 * remains that way on return, but may be unlocked/relocked multiple 1533 * times by the routine. 1534 */ 1535 static int 1536 _cache_inval_internal(struct namecache *ncp, int flags, struct cinvtrack *track) 1537 { 1538 struct namecache *kid; 1539 struct namecache *nextkid; 1540 int rcnt = 0; 1541 1542 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE); 1543 1544 _cache_setunresolved(ncp); 1545 if (flags & CINV_DESTROY) 1546 ncp->nc_flag |= NCF_DESTROYED; 1547 if ((flags & CINV_CHILDREN) && 1548 (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL 1549 ) { 1550 _cache_hold(kid); 1551 if (++track->depth > MAX_RECURSION_DEPTH) { 1552 track->resume_ncp = ncp; 1553 _cache_hold(ncp); 1554 ++rcnt; 1555 } 1556 _cache_unlock(ncp); 1557 while (kid) { 1558 if (track->resume_ncp) { 1559 _cache_drop(kid); 1560 break; 1561 } 1562 if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL) 1563 _cache_hold(nextkid); 1564 if ((kid->nc_flag & NCF_UNRESOLVED) == 0 || 1565 TAILQ_FIRST(&kid->nc_list) 1566 ) { 1567 _cache_lock(kid); 1568 rcnt += _cache_inval_internal(kid, flags & ~CINV_DESTROY, track); 1569 _cache_unlock(kid); 1570 } 1571 _cache_drop(kid); 1572 kid = nextkid; 1573 } 1574 --track->depth; 1575 _cache_lock(ncp); 1576 } 1577 1578 /* 1579 * Someone could have gotten in there while ncp was unlocked, 1580 * retry if so. 1581 */ 1582 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 1583 ++rcnt; 1584 return (rcnt); 1585 } 1586 1587 /* 1588 * Invalidate a vnode's namecache associations. To avoid races against 1589 * the resolver we do not invalidate a node which we previously invalidated 1590 * but which was then re-resolved while we were in the invalidation loop. 1591 * 1592 * Returns non-zero if any namecache entries remain after the invalidation 1593 * loop completed. 1594 * 1595 * NOTE: Unlike the namecache topology which guarentees that ncp's will not 1596 * be ripped out of the topology while held, the vnode's v_namecache 1597 * list has no such restriction. NCP's can be ripped out of the list 1598 * at virtually any time if not locked, even if held. 1599 * 1600 * In addition, the v_namecache list itself must be locked via 1601 * the vnode's spinlock. 1602 */ 1603 int 1604 cache_inval_vp(struct vnode *vp, int flags) 1605 { 1606 struct namecache *ncp; 1607 struct namecache *next; 1608 1609 restart: 1610 spin_lock(&vp->v_spin); 1611 ncp = TAILQ_FIRST(&vp->v_namecache); 1612 if (ncp) 1613 _cache_hold(ncp); 1614 while (ncp) { 1615 /* loop entered with ncp held and vp spin-locked */ 1616 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL) 1617 _cache_hold(next); 1618 spin_unlock(&vp->v_spin); 1619 _cache_lock(ncp); 1620 if (ncp->nc_vp != vp) { 1621 kprintf("Warning: cache_inval_vp: race-A detected on " 1622 "%s\n", ncp->nc_name); 1623 _cache_put(ncp); 1624 if (next) 1625 _cache_drop(next); 1626 goto restart; 1627 } 1628 _cache_inval(ncp, flags); 1629 _cache_put(ncp); /* also releases reference */ 1630 ncp = next; 1631 spin_lock(&vp->v_spin); 1632 if (ncp && ncp->nc_vp != vp) { 1633 spin_unlock(&vp->v_spin); 1634 kprintf("Warning: cache_inval_vp: race-B detected on " 1635 "%s\n", ncp->nc_name); 1636 _cache_drop(ncp); 1637 goto restart; 1638 } 1639 } 1640 spin_unlock(&vp->v_spin); 1641 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 1642 } 1643 1644 /* 1645 * This routine is used instead of the normal cache_inval_vp() when we 1646 * are trying to recycle otherwise good vnodes. 1647 * 1648 * Return 0 on success, non-zero if not all namecache records could be 1649 * disassociated from the vnode (for various reasons). 1650 */ 1651 int 1652 cache_inval_vp_nonblock(struct vnode *vp) 1653 { 1654 struct namecache *ncp; 1655 struct namecache *next; 1656 1657 spin_lock(&vp->v_spin); 1658 ncp = TAILQ_FIRST(&vp->v_namecache); 1659 if (ncp) 1660 _cache_hold(ncp); 1661 while (ncp) { 1662 /* loop entered with ncp held */ 1663 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL) 1664 _cache_hold(next); 1665 spin_unlock(&vp->v_spin); 1666 if (_cache_lock_nonblock(ncp)) { 1667 _cache_drop(ncp); 1668 if (next) 1669 _cache_drop(next); 1670 goto done; 1671 } 1672 if (ncp->nc_vp != vp) { 1673 kprintf("Warning: cache_inval_vp: race-A detected on " 1674 "%s\n", ncp->nc_name); 1675 _cache_put(ncp); 1676 if (next) 1677 _cache_drop(next); 1678 goto done; 1679 } 1680 _cache_inval(ncp, 0); 1681 _cache_put(ncp); /* also releases reference */ 1682 ncp = next; 1683 spin_lock(&vp->v_spin); 1684 if (ncp && ncp->nc_vp != vp) { 1685 spin_unlock(&vp->v_spin); 1686 kprintf("Warning: cache_inval_vp: race-B detected on " 1687 "%s\n", ncp->nc_name); 1688 _cache_drop(ncp); 1689 goto done; 1690 } 1691 } 1692 spin_unlock(&vp->v_spin); 1693 done: 1694 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 1695 } 1696 1697 /* 1698 * The source ncp has been renamed to the target ncp. Both fncp and tncp 1699 * must be locked. The target ncp is destroyed (as a normal rename-over 1700 * would destroy the target file or directory). 1701 * 1702 * Because there may be references to the source ncp we cannot copy its 1703 * contents to the target. Instead the source ncp is relinked as the target 1704 * and the target ncp is removed from the namecache topology. 1705 */ 1706 void 1707 cache_rename(struct nchandle *fnch, struct nchandle *tnch) 1708 { 1709 struct namecache *fncp = fnch->ncp; 1710 struct namecache *tncp = tnch->ncp; 1711 struct namecache *tncp_par; 1712 struct nchash_head *nchpp; 1713 u_int32_t hash; 1714 char *oname; 1715 char *nname; 1716 1717 if (tncp->nc_nlen) { 1718 nname = kmalloc(tncp->nc_nlen + 1, M_VFSCACHE, M_WAITOK); 1719 bcopy(tncp->nc_name, nname, tncp->nc_nlen); 1720 nname[tncp->nc_nlen] = 0; 1721 } else { 1722 nname = NULL; 1723 } 1724 1725 /* 1726 * Rename fncp (unlink) 1727 */ 1728 _cache_unlink_parent(fncp); 1729 oname = fncp->nc_name; 1730 fncp->nc_name = nname; 1731 fncp->nc_nlen = tncp->nc_nlen; 1732 if (oname) 1733 kfree(oname, M_VFSCACHE); 1734 1735 tncp_par = tncp->nc_parent; 1736 _cache_hold(tncp_par); 1737 _cache_lock(tncp_par); 1738 1739 /* 1740 * Rename fncp (relink) 1741 */ 1742 hash = fnv_32_buf(fncp->nc_name, fncp->nc_nlen, FNV1_32_INIT); 1743 hash = fnv_32_buf(&tncp_par, sizeof(tncp_par), hash); 1744 nchpp = NCHHASH(hash); 1745 1746 spin_lock(&nchpp->spin); 1747 _cache_link_parent(fncp, tncp_par, nchpp); 1748 spin_unlock(&nchpp->spin); 1749 1750 _cache_put(tncp_par); 1751 1752 /* 1753 * Get rid of the overwritten tncp (unlink) 1754 */ 1755 _cache_unlink(tncp); 1756 } 1757 1758 /* 1759 * Perform actions consistent with unlinking a file. The passed-in ncp 1760 * must be locked. 1761 * 1762 * The ncp is marked DESTROYED so it no longer shows up in searches, 1763 * and will be physically deleted when the vnode goes away. 1764 * 1765 * If the related vnode has no refs then we cycle it through vget()/vput() 1766 * to (possibly if we don't have a ref race) trigger a deactivation, 1767 * allowing the VFS to trivially detect and recycle the deleted vnode 1768 * via VOP_INACTIVE(). 1769 * 1770 * NOTE: _cache_rename() will automatically call _cache_unlink() on the 1771 * target ncp. 1772 */ 1773 void 1774 cache_unlink(struct nchandle *nch) 1775 { 1776 _cache_unlink(nch->ncp); 1777 } 1778 1779 static void 1780 _cache_unlink(struct namecache *ncp) 1781 { 1782 struct vnode *vp; 1783 1784 /* 1785 * Causes lookups to fail and allows another ncp with the same 1786 * name to be created under ncp->nc_parent. 1787 */ 1788 ncp->nc_flag |= NCF_DESTROYED; 1789 1790 /* 1791 * Attempt to trigger a deactivation. Set VREF_FINALIZE to 1792 * force action on the 1->0 transition. 1793 */ 1794 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 && 1795 (vp = ncp->nc_vp) != NULL) { 1796 atomic_set_int(&vp->v_refcnt, VREF_FINALIZE); 1797 if (VREFCNT(vp) <= 0) { 1798 if (vget(vp, LK_SHARED) == 0) 1799 vput(vp); 1800 } 1801 } 1802 } 1803 1804 /* 1805 * Return non-zero if the nch might be associated with an open and/or mmap()'d 1806 * file. The easy solution is to just return non-zero if the vnode has refs. 1807 * Used to interlock hammer2 reclaims (VREF_FINALIZE should already be set to 1808 * force the reclaim). 1809 */ 1810 int 1811 cache_isopen(struct nchandle *nch) 1812 { 1813 struct vnode *vp; 1814 struct namecache *ncp = nch->ncp; 1815 1816 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 && 1817 (vp = ncp->nc_vp) != NULL && 1818 VREFCNT(vp)) { 1819 return 1; 1820 } 1821 return 0; 1822 } 1823 1824 1825 /* 1826 * vget the vnode associated with the namecache entry. Resolve the namecache 1827 * entry if necessary. The passed ncp must be referenced and locked. If 1828 * the ncp is resolved it might be locked shared. 1829 * 1830 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked 1831 * (depending on the passed lk_type) will be returned in *vpp with an error 1832 * of 0, or NULL will be returned in *vpp with a non-0 error code. The 1833 * most typical error is ENOENT, meaning that the ncp represents a negative 1834 * cache hit and there is no vnode to retrieve, but other errors can occur 1835 * too. 1836 * 1837 * The vget() can race a reclaim. If this occurs we re-resolve the 1838 * namecache entry. 1839 * 1840 * There are numerous places in the kernel where vget() is called on a 1841 * vnode while one or more of its namecache entries is locked. Releasing 1842 * a vnode never deadlocks against locked namecache entries (the vnode 1843 * will not get recycled while referenced ncp's exist). This means we 1844 * can safely acquire the vnode. In fact, we MUST NOT release the ncp 1845 * lock when acquiring the vp lock or we might cause a deadlock. 1846 * 1847 * NOTE: The passed-in ncp must be locked exclusively if it is initially 1848 * unresolved. If a reclaim race occurs the passed-in ncp will be 1849 * relocked exclusively before being re-resolved. 1850 */ 1851 int 1852 cache_vget(struct nchandle *nch, struct ucred *cred, 1853 int lk_type, struct vnode **vpp) 1854 { 1855 struct namecache *ncp; 1856 struct vnode *vp; 1857 int error; 1858 1859 ncp = nch->ncp; 1860 again: 1861 vp = NULL; 1862 if (ncp->nc_flag & NCF_UNRESOLVED) 1863 error = cache_resolve(nch, cred); 1864 else 1865 error = 0; 1866 1867 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 1868 error = vget(vp, lk_type); 1869 if (error) { 1870 /* 1871 * VRECLAIM race 1872 * 1873 * The ncp may have been locked shared, we must relock 1874 * it exclusively before we can set it to unresolved. 1875 */ 1876 if (error == ENOENT) { 1877 kprintf("Warning: vnode reclaim race detected " 1878 "in cache_vget on %p (%s)\n", 1879 vp, ncp->nc_name); 1880 _cache_unlock(ncp); 1881 _cache_lock(ncp); 1882 _cache_setunresolved(ncp); 1883 goto again; 1884 } 1885 1886 /* 1887 * Not a reclaim race, some other error. 1888 */ 1889 KKASSERT(ncp->nc_vp == vp); 1890 vp = NULL; 1891 } else { 1892 KKASSERT(ncp->nc_vp == vp); 1893 KKASSERT((vp->v_flag & VRECLAIMED) == 0); 1894 } 1895 } 1896 if (error == 0 && vp == NULL) 1897 error = ENOENT; 1898 *vpp = vp; 1899 return(error); 1900 } 1901 1902 /* 1903 * Similar to cache_vget() but only acquires a ref on the vnode. 1904 * 1905 * NOTE: The passed-in ncp must be locked exclusively if it is initially 1906 * unresolved. If a reclaim race occurs the passed-in ncp will be 1907 * relocked exclusively before being re-resolved. 1908 */ 1909 int 1910 cache_vref(struct nchandle *nch, struct ucred *cred, struct vnode **vpp) 1911 { 1912 struct namecache *ncp; 1913 struct vnode *vp; 1914 int error; 1915 1916 ncp = nch->ncp; 1917 again: 1918 vp = NULL; 1919 if (ncp->nc_flag & NCF_UNRESOLVED) 1920 error = cache_resolve(nch, cred); 1921 else 1922 error = 0; 1923 1924 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 1925 error = vget(vp, LK_SHARED); 1926 if (error) { 1927 /* 1928 * VRECLAIM race 1929 */ 1930 if (error == ENOENT) { 1931 kprintf("Warning: vnode reclaim race detected " 1932 "in cache_vget on %p (%s)\n", 1933 vp, ncp->nc_name); 1934 _cache_unlock(ncp); 1935 _cache_lock(ncp); 1936 _cache_setunresolved(ncp); 1937 goto again; 1938 } 1939 1940 /* 1941 * Not a reclaim race, some other error. 1942 */ 1943 KKASSERT(ncp->nc_vp == vp); 1944 vp = NULL; 1945 } else { 1946 KKASSERT(ncp->nc_vp == vp); 1947 KKASSERT((vp->v_flag & VRECLAIMED) == 0); 1948 /* caller does not want a lock */ 1949 vn_unlock(vp); 1950 } 1951 } 1952 if (error == 0 && vp == NULL) 1953 error = ENOENT; 1954 *vpp = vp; 1955 return(error); 1956 } 1957 1958 /* 1959 * Return a referenced vnode representing the parent directory of 1960 * ncp. 1961 * 1962 * Because the caller has locked the ncp it should not be possible for 1963 * the parent ncp to go away. However, the parent can unresolve its 1964 * dvp at any time so we must be able to acquire a lock on the parent 1965 * to safely access nc_vp. 1966 * 1967 * We have to leave par unlocked when vget()ing dvp to avoid a deadlock, 1968 * so use vhold()/vdrop() while holding the lock to prevent dvp from 1969 * getting destroyed. 1970 * 1971 * NOTE: vhold() is allowed when dvp has 0 refs if we hold a 1972 * lock on the ncp in question.. 1973 */ 1974 static struct vnode * 1975 cache_dvpref(struct namecache *ncp) 1976 { 1977 struct namecache *par; 1978 struct vnode *dvp; 1979 1980 dvp = NULL; 1981 if ((par = ncp->nc_parent) != NULL) { 1982 _cache_hold(par); 1983 _cache_lock(par); 1984 if ((par->nc_flag & NCF_UNRESOLVED) == 0) { 1985 if ((dvp = par->nc_vp) != NULL) 1986 vhold(dvp); 1987 } 1988 _cache_unlock(par); 1989 if (dvp) { 1990 if (vget(dvp, LK_SHARED) == 0) { 1991 vn_unlock(dvp); 1992 vdrop(dvp); 1993 /* return refd, unlocked dvp */ 1994 } else { 1995 vdrop(dvp); 1996 dvp = NULL; 1997 } 1998 } 1999 _cache_drop(par); 2000 } 2001 return(dvp); 2002 } 2003 2004 /* 2005 * Convert a directory vnode to a namecache record without any other 2006 * knowledge of the topology. This ONLY works with directory vnodes and 2007 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the 2008 * returned ncp (if not NULL) will be held and unlocked. 2009 * 2010 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned. 2011 * If 'makeit' is 1 we attempt to track-down and create the namecache topology 2012 * for dvp. This will fail only if the directory has been deleted out from 2013 * under the caller. 2014 * 2015 * Callers must always check for a NULL return no matter the value of 'makeit'. 2016 * 2017 * To avoid underflowing the kernel stack each recursive call increments 2018 * the makeit variable. 2019 */ 2020 2021 static int cache_inefficient_scan(struct nchandle *nch, struct ucred *cred, 2022 struct vnode *dvp, char *fakename); 2023 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred, 2024 struct vnode **saved_dvp); 2025 2026 int 2027 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit, 2028 struct nchandle *nch) 2029 { 2030 struct vnode *saved_dvp; 2031 struct vnode *pvp; 2032 char *fakename; 2033 int error; 2034 2035 nch->ncp = NULL; 2036 nch->mount = dvp->v_mount; 2037 saved_dvp = NULL; 2038 fakename = NULL; 2039 2040 /* 2041 * Handle the makeit == 0 degenerate case 2042 */ 2043 if (makeit == 0) { 2044 spin_lock_shared(&dvp->v_spin); 2045 nch->ncp = TAILQ_FIRST(&dvp->v_namecache); 2046 if (nch->ncp) 2047 cache_hold(nch); 2048 spin_unlock_shared(&dvp->v_spin); 2049 } 2050 2051 /* 2052 * Loop until resolution, inside code will break out on error. 2053 */ 2054 while (makeit) { 2055 /* 2056 * Break out if we successfully acquire a working ncp. 2057 */ 2058 spin_lock_shared(&dvp->v_spin); 2059 nch->ncp = TAILQ_FIRST(&dvp->v_namecache); 2060 if (nch->ncp) { 2061 cache_hold(nch); 2062 spin_unlock_shared(&dvp->v_spin); 2063 break; 2064 } 2065 spin_unlock_shared(&dvp->v_spin); 2066 2067 /* 2068 * If dvp is the root of its filesystem it should already 2069 * have a namecache pointer associated with it as a side 2070 * effect of the mount, but it may have been disassociated. 2071 */ 2072 if (dvp->v_flag & VROOT) { 2073 nch->ncp = _cache_get(nch->mount->mnt_ncmountpt.ncp); 2074 error = cache_resolve_mp(nch->mount); 2075 _cache_put(nch->ncp); 2076 if (ncvp_debug) { 2077 kprintf("cache_fromdvp: resolve root of mount %p error %d", 2078 dvp->v_mount, error); 2079 } 2080 if (error) { 2081 if (ncvp_debug) 2082 kprintf(" failed\n"); 2083 nch->ncp = NULL; 2084 break; 2085 } 2086 if (ncvp_debug) 2087 kprintf(" succeeded\n"); 2088 continue; 2089 } 2090 2091 /* 2092 * If we are recursed too deeply resort to an O(n^2) 2093 * algorithm to resolve the namecache topology. The 2094 * resolved pvp is left referenced in saved_dvp to 2095 * prevent the tree from being destroyed while we loop. 2096 */ 2097 if (makeit > 20) { 2098 error = cache_fromdvp_try(dvp, cred, &saved_dvp); 2099 if (error) { 2100 kprintf("lookupdotdot(longpath) failed %d " 2101 "dvp %p\n", error, dvp); 2102 nch->ncp = NULL; 2103 break; 2104 } 2105 continue; 2106 } 2107 2108 /* 2109 * Get the parent directory and resolve its ncp. 2110 */ 2111 if (fakename) { 2112 kfree(fakename, M_TEMP); 2113 fakename = NULL; 2114 } 2115 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred, 2116 &fakename); 2117 if (error) { 2118 kprintf("lookupdotdot failed %d dvp %p\n", error, dvp); 2119 break; 2120 } 2121 vn_unlock(pvp); 2122 2123 /* 2124 * Reuse makeit as a recursion depth counter. On success 2125 * nch will be fully referenced. 2126 */ 2127 cache_fromdvp(pvp, cred, makeit + 1, nch); 2128 vrele(pvp); 2129 if (nch->ncp == NULL) 2130 break; 2131 2132 /* 2133 * Do an inefficient scan of pvp (embodied by ncp) to look 2134 * for dvp. This will create a namecache record for dvp on 2135 * success. We loop up to recheck on success. 2136 * 2137 * ncp and dvp are both held but not locked. 2138 */ 2139 error = cache_inefficient_scan(nch, cred, dvp, fakename); 2140 if (error) { 2141 kprintf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n", 2142 pvp, nch->ncp->nc_name, dvp); 2143 cache_drop(nch); 2144 /* nch was NULLed out, reload mount */ 2145 nch->mount = dvp->v_mount; 2146 break; 2147 } 2148 if (ncvp_debug) { 2149 kprintf("cache_fromdvp: scan %p (%s) succeeded\n", 2150 pvp, nch->ncp->nc_name); 2151 } 2152 cache_drop(nch); 2153 /* nch was NULLed out, reload mount */ 2154 nch->mount = dvp->v_mount; 2155 } 2156 2157 /* 2158 * If nch->ncp is non-NULL it will have been held already. 2159 */ 2160 if (fakename) 2161 kfree(fakename, M_TEMP); 2162 if (saved_dvp) 2163 vrele(saved_dvp); 2164 if (nch->ncp) 2165 return (0); 2166 return (EINVAL); 2167 } 2168 2169 /* 2170 * Go up the chain of parent directories until we find something 2171 * we can resolve into the namecache. This is very inefficient. 2172 */ 2173 static 2174 int 2175 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred, 2176 struct vnode **saved_dvp) 2177 { 2178 struct nchandle nch; 2179 struct vnode *pvp; 2180 int error; 2181 static time_t last_fromdvp_report; 2182 char *fakename; 2183 2184 /* 2185 * Loop getting the parent directory vnode until we get something we 2186 * can resolve in the namecache. 2187 */ 2188 vref(dvp); 2189 nch.mount = dvp->v_mount; 2190 nch.ncp = NULL; 2191 fakename = NULL; 2192 2193 for (;;) { 2194 if (fakename) { 2195 kfree(fakename, M_TEMP); 2196 fakename = NULL; 2197 } 2198 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred, 2199 &fakename); 2200 if (error) { 2201 vrele(dvp); 2202 break; 2203 } 2204 vn_unlock(pvp); 2205 spin_lock_shared(&pvp->v_spin); 2206 if ((nch.ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) { 2207 _cache_hold(nch.ncp); 2208 spin_unlock_shared(&pvp->v_spin); 2209 vrele(pvp); 2210 break; 2211 } 2212 spin_unlock_shared(&pvp->v_spin); 2213 if (pvp->v_flag & VROOT) { 2214 nch.ncp = _cache_get(pvp->v_mount->mnt_ncmountpt.ncp); 2215 error = cache_resolve_mp(nch.mount); 2216 _cache_unlock(nch.ncp); 2217 vrele(pvp); 2218 if (error) { 2219 _cache_drop(nch.ncp); 2220 nch.ncp = NULL; 2221 vrele(dvp); 2222 } 2223 break; 2224 } 2225 vrele(dvp); 2226 dvp = pvp; 2227 } 2228 if (error == 0) { 2229 if (last_fromdvp_report != time_uptime) { 2230 last_fromdvp_report = time_uptime; 2231 kprintf("Warning: extremely inefficient path " 2232 "resolution on %s\n", 2233 nch.ncp->nc_name); 2234 } 2235 error = cache_inefficient_scan(&nch, cred, dvp, fakename); 2236 2237 /* 2238 * Hopefully dvp now has a namecache record associated with 2239 * it. Leave it referenced to prevent the kernel from 2240 * recycling the vnode. Otherwise extremely long directory 2241 * paths could result in endless recycling. 2242 */ 2243 if (*saved_dvp) 2244 vrele(*saved_dvp); 2245 *saved_dvp = dvp; 2246 _cache_drop(nch.ncp); 2247 } 2248 if (fakename) 2249 kfree(fakename, M_TEMP); 2250 return (error); 2251 } 2252 2253 /* 2254 * Do an inefficient scan of the directory represented by ncp looking for 2255 * the directory vnode dvp. ncp must be held but not locked on entry and 2256 * will be held on return. dvp must be refd but not locked on entry and 2257 * will remain refd on return. 2258 * 2259 * Why do this at all? Well, due to its stateless nature the NFS server 2260 * converts file handles directly to vnodes without necessarily going through 2261 * the namecache ops that would otherwise create the namecache topology 2262 * leading to the vnode. We could either (1) Change the namecache algorithms 2263 * to allow disconnect namecache records that are re-merged opportunistically, 2264 * or (2) Make the NFS server backtrack and scan to recover a connected 2265 * namecache topology in order to then be able to issue new API lookups. 2266 * 2267 * It turns out that (1) is a huge mess. It takes a nice clean set of 2268 * namecache algorithms and introduces a lot of complication in every subsystem 2269 * that calls into the namecache to deal with the re-merge case, especially 2270 * since we are using the namecache to placehold negative lookups and the 2271 * vnode might not be immediately assigned. (2) is certainly far less 2272 * efficient then (1), but since we are only talking about directories here 2273 * (which are likely to remain cached), the case does not actually run all 2274 * that often and has the supreme advantage of not polluting the namecache 2275 * algorithms. 2276 * 2277 * If a fakename is supplied just construct a namecache entry using the 2278 * fake name. 2279 */ 2280 static int 2281 cache_inefficient_scan(struct nchandle *nch, struct ucred *cred, 2282 struct vnode *dvp, char *fakename) 2283 { 2284 struct nlcomponent nlc; 2285 struct nchandle rncp; 2286 struct dirent *den; 2287 struct vnode *pvp; 2288 struct vattr vat; 2289 struct iovec iov; 2290 struct uio uio; 2291 int blksize; 2292 int eofflag; 2293 int bytes; 2294 char *rbuf; 2295 int error; 2296 2297 vat.va_blocksize = 0; 2298 if ((error = VOP_GETATTR(dvp, &vat)) != 0) 2299 return (error); 2300 cache_lock(nch); 2301 error = cache_vref(nch, cred, &pvp); 2302 cache_unlock(nch); 2303 if (error) 2304 return (error); 2305 if (ncvp_debug) { 2306 kprintf("inefficient_scan: directory iosize %ld " 2307 "vattr fileid = %lld\n", 2308 vat.va_blocksize, 2309 (long long)vat.va_fileid); 2310 } 2311 2312 /* 2313 * Use the supplied fakename if not NULL. Fake names are typically 2314 * not in the actual filesystem hierarchy. This is used by HAMMER 2315 * to glue @@timestamp recursions together. 2316 */ 2317 if (fakename) { 2318 nlc.nlc_nameptr = fakename; 2319 nlc.nlc_namelen = strlen(fakename); 2320 rncp = cache_nlookup(nch, &nlc); 2321 goto done; 2322 } 2323 2324 if ((blksize = vat.va_blocksize) == 0) 2325 blksize = DEV_BSIZE; 2326 rbuf = kmalloc(blksize, M_TEMP, M_WAITOK); 2327 rncp.ncp = NULL; 2328 2329 eofflag = 0; 2330 uio.uio_offset = 0; 2331 again: 2332 iov.iov_base = rbuf; 2333 iov.iov_len = blksize; 2334 uio.uio_iov = &iov; 2335 uio.uio_iovcnt = 1; 2336 uio.uio_resid = blksize; 2337 uio.uio_segflg = UIO_SYSSPACE; 2338 uio.uio_rw = UIO_READ; 2339 uio.uio_td = curthread; 2340 2341 if (ncvp_debug >= 2) 2342 kprintf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset); 2343 error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL); 2344 if (error == 0) { 2345 den = (struct dirent *)rbuf; 2346 bytes = blksize - uio.uio_resid; 2347 2348 while (bytes > 0) { 2349 if (ncvp_debug >= 2) { 2350 kprintf("cache_inefficient_scan: %*.*s\n", 2351 den->d_namlen, den->d_namlen, 2352 den->d_name); 2353 } 2354 if (den->d_type != DT_WHT && 2355 den->d_ino == vat.va_fileid) { 2356 if (ncvp_debug) { 2357 kprintf("cache_inefficient_scan: " 2358 "MATCHED inode %lld path %s/%*.*s\n", 2359 (long long)vat.va_fileid, 2360 nch->ncp->nc_name, 2361 den->d_namlen, den->d_namlen, 2362 den->d_name); 2363 } 2364 nlc.nlc_nameptr = den->d_name; 2365 nlc.nlc_namelen = den->d_namlen; 2366 rncp = cache_nlookup(nch, &nlc); 2367 KKASSERT(rncp.ncp != NULL); 2368 break; 2369 } 2370 bytes -= _DIRENT_DIRSIZ(den); 2371 den = _DIRENT_NEXT(den); 2372 } 2373 if (rncp.ncp == NULL && eofflag == 0 && uio.uio_resid != blksize) 2374 goto again; 2375 } 2376 kfree(rbuf, M_TEMP); 2377 done: 2378 vrele(pvp); 2379 if (rncp.ncp) { 2380 if (rncp.ncp->nc_flag & NCF_UNRESOLVED) { 2381 _cache_setvp(rncp.mount, rncp.ncp, dvp); 2382 if (ncvp_debug >= 2) { 2383 kprintf("cache_inefficient_scan: setvp %s/%s = %p\n", 2384 nch->ncp->nc_name, rncp.ncp->nc_name, dvp); 2385 } 2386 } else { 2387 if (ncvp_debug >= 2) { 2388 kprintf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n", 2389 nch->ncp->nc_name, rncp.ncp->nc_name, dvp, 2390 rncp.ncp->nc_vp); 2391 } 2392 } 2393 if (rncp.ncp->nc_vp == NULL) 2394 error = rncp.ncp->nc_error; 2395 /* 2396 * Release rncp after a successful nlookup. rncp was fully 2397 * referenced. 2398 */ 2399 cache_put(&rncp); 2400 } else { 2401 kprintf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n", 2402 dvp, nch->ncp->nc_name); 2403 error = ENOENT; 2404 } 2405 return (error); 2406 } 2407 2408 /* 2409 * Zap a namecache entry. The ncp is unconditionally set to an unresolved 2410 * state, which disassociates it from its vnode or ncneglist. 2411 * 2412 * Then, if there are no additional references to the ncp and no children, 2413 * the ncp is removed from the topology and destroyed. 2414 * 2415 * References and/or children may exist if the ncp is in the middle of the 2416 * topology, preventing the ncp from being destroyed. 2417 * 2418 * This function must be called with the ncp held and locked and will unlock 2419 * and drop it during zapping. 2420 * 2421 * If nonblock is non-zero and the parent ncp cannot be locked we give up. 2422 * This case can occur in the cache_drop() path. 2423 * 2424 * This function may returned a held (but NOT locked) parent node which the 2425 * caller must drop. We do this so _cache_drop() can loop, to avoid 2426 * blowing out the kernel stack. 2427 * 2428 * WARNING! For MPSAFE operation this routine must acquire up to three 2429 * spin locks to be able to safely test nc_refs. Lock order is 2430 * very important. 2431 * 2432 * hash spinlock if on hash list 2433 * parent spinlock if child of parent 2434 * (the ncp is unresolved so there is no vnode association) 2435 */ 2436 static struct namecache * 2437 cache_zap(struct namecache *ncp, int nonblock) 2438 { 2439 struct namecache *par; 2440 struct vnode *dropvp; 2441 int refs; 2442 2443 /* 2444 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED. 2445 */ 2446 _cache_setunresolved(ncp); 2447 2448 /* 2449 * Try to scrap the entry and possibly tail-recurse on its parent. 2450 * We only scrap unref'd (other then our ref) unresolved entries, 2451 * we do not scrap 'live' entries. 2452 * 2453 * Note that once the spinlocks are acquired if nc_refs == 1 no 2454 * other references are possible. If it isn't, however, we have 2455 * to decrement but also be sure to avoid a 1->0 transition. 2456 */ 2457 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED); 2458 KKASSERT(ncp->nc_refs > 0); 2459 2460 /* 2461 * Acquire locks. Note that the parent can't go away while we hold 2462 * a child locked. 2463 */ 2464 if ((par = ncp->nc_parent) != NULL) { 2465 if (nonblock) { 2466 for (;;) { 2467 if (_cache_lock_nonblock(par) == 0) 2468 break; 2469 refs = ncp->nc_refs; 2470 ncp->nc_flag |= NCF_DEFEREDZAP; 2471 ++numdefered; /* MP race ok */ 2472 if (atomic_cmpset_int(&ncp->nc_refs, 2473 refs, refs - 1)) { 2474 _cache_unlock(ncp); 2475 return(NULL); 2476 } 2477 cpu_pause(); 2478 } 2479 _cache_hold(par); 2480 } else { 2481 _cache_hold(par); 2482 _cache_lock(par); 2483 } 2484 spin_lock(&ncp->nc_head->spin); 2485 } 2486 2487 /* 2488 * If someone other then us has a ref or we have children 2489 * we cannot zap the entry. The 1->0 transition and any 2490 * further list operation is protected by the spinlocks 2491 * we have acquired but other transitions are not. 2492 */ 2493 for (;;) { 2494 refs = ncp->nc_refs; 2495 if (refs == 1 && TAILQ_EMPTY(&ncp->nc_list)) 2496 break; 2497 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) { 2498 if (par) { 2499 spin_unlock(&ncp->nc_head->spin); 2500 _cache_put(par); 2501 } 2502 _cache_unlock(ncp); 2503 return(NULL); 2504 } 2505 cpu_pause(); 2506 } 2507 2508 /* 2509 * We are the only ref and with the spinlocks held no further 2510 * refs can be acquired by others. 2511 * 2512 * Remove us from the hash list and parent list. We have to 2513 * drop a ref on the parent's vp if the parent's list becomes 2514 * empty. 2515 */ 2516 dropvp = NULL; 2517 if (par) { 2518 struct nchash_head *nchpp = ncp->nc_head; 2519 2520 KKASSERT(nchpp != NULL); 2521 LIST_REMOVE(ncp, nc_hash); 2522 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 2523 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 2524 dropvp = par->nc_vp; 2525 ncp->nc_head = NULL; 2526 ncp->nc_parent = NULL; 2527 spin_unlock(&nchpp->spin); 2528 _cache_unlock(par); 2529 } else { 2530 KKASSERT(ncp->nc_head == NULL); 2531 } 2532 2533 /* 2534 * ncp should not have picked up any refs. Physically 2535 * destroy the ncp. 2536 */ 2537 KKASSERT(ncp->nc_refs == 1); 2538 /* _cache_unlock(ncp) not required */ 2539 ncp->nc_refs = -1; /* safety */ 2540 if (ncp->nc_name) 2541 kfree(ncp->nc_name, M_VFSCACHE); 2542 kfree(ncp, M_VFSCACHE); 2543 2544 /* 2545 * Delayed drop (we had to release our spinlocks) 2546 * 2547 * The refed parent (if not NULL) must be dropped. The 2548 * caller is responsible for looping. 2549 */ 2550 if (dropvp) 2551 vdrop(dropvp); 2552 return(par); 2553 } 2554 2555 /* 2556 * Clean up dangling negative cache and defered-drop entries in the 2557 * namecache. 2558 * 2559 * This routine is called in the critical path and also called from 2560 * vnlru(). When called from vnlru we use a lower limit to try to 2561 * deal with the negative cache before the critical path has to start 2562 * dealing with it. 2563 */ 2564 typedef enum { CHI_LOW, CHI_HIGH } cache_hs_t; 2565 2566 static cache_hs_t neg_cache_hysteresis_state[2] = { CHI_LOW, CHI_LOW }; 2567 static cache_hs_t pos_cache_hysteresis_state[2] = { CHI_LOW, CHI_LOW }; 2568 2569 void 2570 cache_hysteresis(int critpath) 2571 { 2572 int poslimit; 2573 int neglimit = desiredvnodes / ncnegfactor; 2574 int xnumcache = numcache; 2575 2576 if (critpath == 0) 2577 neglimit = neglimit * 8 / 10; 2578 2579 /* 2580 * Don't cache too many negative hits. We use hysteresis to reduce 2581 * the impact on the critical path. 2582 */ 2583 switch(neg_cache_hysteresis_state[critpath]) { 2584 case CHI_LOW: 2585 if (numneg > MINNEG && numneg > neglimit) { 2586 if (critpath) 2587 _cache_cleanneg(ncnegflush); 2588 else 2589 _cache_cleanneg(ncnegflush + 2590 numneg - neglimit); 2591 neg_cache_hysteresis_state[critpath] = CHI_HIGH; 2592 } 2593 break; 2594 case CHI_HIGH: 2595 if (numneg > MINNEG * 9 / 10 && 2596 numneg * 9 / 10 > neglimit 2597 ) { 2598 if (critpath) 2599 _cache_cleanneg(ncnegflush); 2600 else 2601 _cache_cleanneg(ncnegflush + 2602 numneg * 9 / 10 - neglimit); 2603 } else { 2604 neg_cache_hysteresis_state[critpath] = CHI_LOW; 2605 } 2606 break; 2607 } 2608 2609 /* 2610 * Don't cache too many positive hits. We use hysteresis to reduce 2611 * the impact on the critical path. 2612 * 2613 * Excessive positive hits can accumulate due to large numbers of 2614 * hardlinks (the vnode cache will not prevent hl ncps from growing 2615 * into infinity). 2616 */ 2617 if ((poslimit = ncposlimit) == 0) 2618 poslimit = desiredvnodes * 2; 2619 if (critpath == 0) 2620 poslimit = poslimit * 8 / 10; 2621 2622 switch(pos_cache_hysteresis_state[critpath]) { 2623 case CHI_LOW: 2624 if (xnumcache > poslimit && xnumcache > MINPOS) { 2625 if (critpath) 2626 _cache_cleanpos(ncposflush); 2627 else 2628 _cache_cleanpos(ncposflush + 2629 xnumcache - poslimit); 2630 pos_cache_hysteresis_state[critpath] = CHI_HIGH; 2631 } 2632 break; 2633 case CHI_HIGH: 2634 if (xnumcache > poslimit * 5 / 6 && xnumcache > MINPOS) { 2635 if (critpath) 2636 _cache_cleanpos(ncposflush); 2637 else 2638 _cache_cleanpos(ncposflush + 2639 xnumcache - poslimit * 5 / 6); 2640 } else { 2641 pos_cache_hysteresis_state[critpath] = CHI_LOW; 2642 } 2643 break; 2644 } 2645 2646 /* 2647 * Clean out dangling defered-zap ncps which could not 2648 * be cleanly dropped if too many build up. Note 2649 * that numdefered is not an exact number as such ncps 2650 * can be reused and the counter is not handled in a MP 2651 * safe manner by design. 2652 */ 2653 if (numdefered > neglimit) { 2654 _cache_cleandefered(); 2655 } 2656 } 2657 2658 /* 2659 * NEW NAMECACHE LOOKUP API 2660 * 2661 * Lookup an entry in the namecache. The passed par_nch must be referenced 2662 * and unlocked. A referenced and locked nchandle with a non-NULL nch.ncp 2663 * is ALWAYS returned, eve if the supplied component is illegal. 2664 * 2665 * The resulting namecache entry should be returned to the system with 2666 * cache_put() or cache_unlock() + cache_drop(). 2667 * 2668 * namecache locks are recursive but care must be taken to avoid lock order 2669 * reversals (hence why the passed par_nch must be unlocked). Locking 2670 * rules are to order for parent traversals, not for child traversals. 2671 * 2672 * Nobody else will be able to manipulate the associated namespace (e.g. 2673 * create, delete, rename, rename-target) until the caller unlocks the 2674 * entry. 2675 * 2676 * The returned entry will be in one of three states: positive hit (non-null 2677 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set). 2678 * Unresolved entries must be resolved through the filesystem to associate the 2679 * vnode and/or determine whether a positive or negative hit has occured. 2680 * 2681 * It is not necessary to lock a directory in order to lock namespace under 2682 * that directory. In fact, it is explicitly not allowed to do that. A 2683 * directory is typically only locked when being created, renamed, or 2684 * destroyed. 2685 * 2686 * The directory (par) may be unresolved, in which case any returned child 2687 * will likely also be marked unresolved. Likely but not guarenteed. Since 2688 * the filesystem lookup requires a resolved directory vnode the caller is 2689 * responsible for resolving the namecache chain top-down. This API 2690 * specifically allows whole chains to be created in an unresolved state. 2691 */ 2692 struct nchandle 2693 cache_nlookup(struct nchandle *par_nch, struct nlcomponent *nlc) 2694 { 2695 struct nchandle nch; 2696 struct namecache *ncp; 2697 struct namecache *new_ncp; 2698 struct nchash_head *nchpp; 2699 struct mount *mp; 2700 u_int32_t hash; 2701 globaldata_t gd; 2702 int par_locked; 2703 2704 numcalls++; 2705 gd = mycpu; 2706 mp = par_nch->mount; 2707 par_locked = 0; 2708 2709 /* 2710 * This is a good time to call it, no ncp's are locked by 2711 * the caller or us. 2712 */ 2713 cache_hysteresis(1); 2714 2715 /* 2716 * Try to locate an existing entry 2717 */ 2718 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 2719 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash); 2720 new_ncp = NULL; 2721 nchpp = NCHHASH(hash); 2722 restart: 2723 if (new_ncp) 2724 spin_lock(&nchpp->spin); 2725 else 2726 spin_lock_shared(&nchpp->spin); 2727 2728 LIST_FOREACH(ncp, &nchpp->list, nc_hash) { 2729 numchecks++; 2730 2731 /* 2732 * Break out if we find a matching entry. Note that 2733 * UNRESOLVED entries may match, but DESTROYED entries 2734 * do not. 2735 */ 2736 if (ncp->nc_parent == par_nch->ncp && 2737 ncp->nc_nlen == nlc->nlc_namelen && 2738 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 2739 (ncp->nc_flag & NCF_DESTROYED) == 0 2740 ) { 2741 _cache_hold(ncp); 2742 if (new_ncp) 2743 spin_unlock(&nchpp->spin); 2744 else 2745 spin_unlock_shared(&nchpp->spin); 2746 if (par_locked) { 2747 _cache_unlock(par_nch->ncp); 2748 par_locked = 0; 2749 } 2750 if (_cache_lock_special(ncp) == 0) { 2751 /* 2752 * Successfully locked but we must re-test 2753 * conditions that might have changed since 2754 * we did not have the lock before. 2755 */ 2756 if ((ncp->nc_flag & NCF_DESTROYED) || 2757 ncp->nc_parent != par_nch->ncp) { 2758 _cache_put(ncp); 2759 goto restart; 2760 } 2761 _cache_auto_unresolve(mp, ncp); 2762 if (new_ncp) 2763 _cache_free(new_ncp); 2764 goto found; 2765 } 2766 _cache_get(ncp); /* cycle the lock to block */ 2767 _cache_put(ncp); 2768 _cache_drop(ncp); 2769 goto restart; 2770 } 2771 } 2772 2773 /* 2774 * We failed to locate an entry, create a new entry and add it to 2775 * the cache. The parent ncp must also be locked so we 2776 * can link into it. 2777 * 2778 * We have to relookup after possibly blocking in kmalloc or 2779 * when locking par_nch. 2780 * 2781 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special 2782 * mount case, in which case nc_name will be NULL. 2783 */ 2784 if (new_ncp == NULL) { 2785 spin_unlock_shared(&nchpp->spin); 2786 new_ncp = cache_alloc(nlc->nlc_namelen); 2787 if (nlc->nlc_namelen) { 2788 bcopy(nlc->nlc_nameptr, new_ncp->nc_name, 2789 nlc->nlc_namelen); 2790 new_ncp->nc_name[nlc->nlc_namelen] = 0; 2791 } 2792 goto restart; 2793 } 2794 2795 /* 2796 * NOTE! The spinlock is held exclusively here because new_ncp 2797 * is non-NULL. 2798 */ 2799 if (par_locked == 0) { 2800 spin_unlock(&nchpp->spin); 2801 _cache_lock(par_nch->ncp); 2802 par_locked = 1; 2803 goto restart; 2804 } 2805 2806 /* 2807 * WARNING! We still hold the spinlock. We have to set the hash 2808 * table entry atomically. 2809 */ 2810 ncp = new_ncp; 2811 _cache_link_parent(ncp, par_nch->ncp, nchpp); 2812 spin_unlock(&nchpp->spin); 2813 _cache_unlock(par_nch->ncp); 2814 /* par_locked = 0 - not used */ 2815 found: 2816 /* 2817 * stats and namecache size management 2818 */ 2819 if (ncp->nc_flag & NCF_UNRESOLVED) 2820 ++gd->gd_nchstats->ncs_miss; 2821 else if (ncp->nc_vp) 2822 ++gd->gd_nchstats->ncs_goodhits; 2823 else 2824 ++gd->gd_nchstats->ncs_neghits; 2825 nch.mount = mp; 2826 nch.ncp = ncp; 2827 atomic_add_int(&nch.mount->mnt_refs, 1); 2828 return(nch); 2829 } 2830 2831 /* 2832 * Attempt to lookup a namecache entry and return with a shared namecache 2833 * lock. 2834 */ 2835 int 2836 cache_nlookup_maybe_shared(struct nchandle *par_nch, struct nlcomponent *nlc, 2837 int excl, struct nchandle *res_nch) 2838 { 2839 struct namecache *ncp; 2840 struct nchash_head *nchpp; 2841 struct mount *mp; 2842 u_int32_t hash; 2843 globaldata_t gd; 2844 2845 /* 2846 * If exclusive requested or shared namecache locks are disabled, 2847 * return failure. 2848 */ 2849 if (ncp_shared_lock_disable || excl) 2850 return(EWOULDBLOCK); 2851 2852 numcalls++; 2853 gd = mycpu; 2854 mp = par_nch->mount; 2855 2856 /* 2857 * This is a good time to call it, no ncp's are locked by 2858 * the caller or us. 2859 */ 2860 cache_hysteresis(1); 2861 2862 /* 2863 * Try to locate an existing entry 2864 */ 2865 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 2866 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash); 2867 nchpp = NCHHASH(hash); 2868 2869 spin_lock_shared(&nchpp->spin); 2870 2871 LIST_FOREACH(ncp, &nchpp->list, nc_hash) { 2872 numchecks++; 2873 2874 /* 2875 * Break out if we find a matching entry. Note that 2876 * UNRESOLVED entries may match, but DESTROYED entries 2877 * do not. 2878 */ 2879 if (ncp->nc_parent == par_nch->ncp && 2880 ncp->nc_nlen == nlc->nlc_namelen && 2881 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 2882 (ncp->nc_flag & NCF_DESTROYED) == 0 2883 ) { 2884 _cache_hold(ncp); 2885 spin_unlock_shared(&nchpp->spin); 2886 if (_cache_lock_shared_special(ncp) == 0) { 2887 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 && 2888 (ncp->nc_flag & NCF_DESTROYED) == 0 && 2889 _cache_auto_unresolve_test(mp, ncp) == 0) { 2890 goto found; 2891 } 2892 _cache_unlock(ncp); 2893 } 2894 _cache_drop(ncp); 2895 spin_lock_shared(&nchpp->spin); 2896 break; 2897 } 2898 } 2899 2900 /* 2901 * Failure 2902 */ 2903 spin_unlock_shared(&nchpp->spin); 2904 return(EWOULDBLOCK); 2905 2906 /* 2907 * Success 2908 * 2909 * Note that nc_error might be non-zero (e.g ENOENT). 2910 */ 2911 found: 2912 res_nch->mount = mp; 2913 res_nch->ncp = ncp; 2914 ++gd->gd_nchstats->ncs_goodhits; 2915 atomic_add_int(&res_nch->mount->mnt_refs, 1); 2916 2917 KKASSERT(ncp->nc_error != EWOULDBLOCK); 2918 return(ncp->nc_error); 2919 } 2920 2921 /* 2922 * This is a non-blocking verison of cache_nlookup() used by 2923 * nfs_readdirplusrpc_uio(). It can fail for any reason and 2924 * will return nch.ncp == NULL in that case. 2925 */ 2926 struct nchandle 2927 cache_nlookup_nonblock(struct nchandle *par_nch, struct nlcomponent *nlc) 2928 { 2929 struct nchandle nch; 2930 struct namecache *ncp; 2931 struct namecache *new_ncp; 2932 struct nchash_head *nchpp; 2933 struct mount *mp; 2934 u_int32_t hash; 2935 globaldata_t gd; 2936 int par_locked; 2937 2938 numcalls++; 2939 gd = mycpu; 2940 mp = par_nch->mount; 2941 par_locked = 0; 2942 2943 /* 2944 * Try to locate an existing entry 2945 */ 2946 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 2947 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash); 2948 new_ncp = NULL; 2949 nchpp = NCHHASH(hash); 2950 restart: 2951 spin_lock(&nchpp->spin); 2952 LIST_FOREACH(ncp, &nchpp->list, nc_hash) { 2953 numchecks++; 2954 2955 /* 2956 * Break out if we find a matching entry. Note that 2957 * UNRESOLVED entries may match, but DESTROYED entries 2958 * do not. 2959 */ 2960 if (ncp->nc_parent == par_nch->ncp && 2961 ncp->nc_nlen == nlc->nlc_namelen && 2962 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 2963 (ncp->nc_flag & NCF_DESTROYED) == 0 2964 ) { 2965 _cache_hold(ncp); 2966 spin_unlock(&nchpp->spin); 2967 if (par_locked) { 2968 _cache_unlock(par_nch->ncp); 2969 par_locked = 0; 2970 } 2971 if (_cache_lock_special(ncp) == 0) { 2972 _cache_auto_unresolve(mp, ncp); 2973 if (new_ncp) { 2974 _cache_free(new_ncp); 2975 new_ncp = NULL; 2976 } 2977 goto found; 2978 } 2979 _cache_drop(ncp); 2980 goto failed; 2981 } 2982 } 2983 2984 /* 2985 * We failed to locate an entry, create a new entry and add it to 2986 * the cache. The parent ncp must also be locked so we 2987 * can link into it. 2988 * 2989 * We have to relookup after possibly blocking in kmalloc or 2990 * when locking par_nch. 2991 * 2992 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special 2993 * mount case, in which case nc_name will be NULL. 2994 */ 2995 if (new_ncp == NULL) { 2996 spin_unlock(&nchpp->spin); 2997 new_ncp = cache_alloc(nlc->nlc_namelen); 2998 if (nlc->nlc_namelen) { 2999 bcopy(nlc->nlc_nameptr, new_ncp->nc_name, 3000 nlc->nlc_namelen); 3001 new_ncp->nc_name[nlc->nlc_namelen] = 0; 3002 } 3003 goto restart; 3004 } 3005 if (par_locked == 0) { 3006 spin_unlock(&nchpp->spin); 3007 if (_cache_lock_nonblock(par_nch->ncp) == 0) { 3008 par_locked = 1; 3009 goto restart; 3010 } 3011 goto failed; 3012 } 3013 3014 /* 3015 * WARNING! We still hold the spinlock. We have to set the hash 3016 * table entry atomically. 3017 */ 3018 ncp = new_ncp; 3019 _cache_link_parent(ncp, par_nch->ncp, nchpp); 3020 spin_unlock(&nchpp->spin); 3021 _cache_unlock(par_nch->ncp); 3022 /* par_locked = 0 - not used */ 3023 found: 3024 /* 3025 * stats and namecache size management 3026 */ 3027 if (ncp->nc_flag & NCF_UNRESOLVED) 3028 ++gd->gd_nchstats->ncs_miss; 3029 else if (ncp->nc_vp) 3030 ++gd->gd_nchstats->ncs_goodhits; 3031 else 3032 ++gd->gd_nchstats->ncs_neghits; 3033 nch.mount = mp; 3034 nch.ncp = ncp; 3035 atomic_add_int(&nch.mount->mnt_refs, 1); 3036 return(nch); 3037 failed: 3038 if (new_ncp) { 3039 _cache_free(new_ncp); 3040 new_ncp = NULL; 3041 } 3042 nch.mount = NULL; 3043 nch.ncp = NULL; 3044 return(nch); 3045 } 3046 3047 /* 3048 * The namecache entry is marked as being used as a mount point. 3049 * Locate the mount if it is visible to the caller. The DragonFly 3050 * mount system allows arbitrary loops in the topology and disentangles 3051 * those loops by matching against (mp, ncp) rather than just (ncp). 3052 * This means any given ncp can dive any number of mounts, depending 3053 * on the relative mount (e.g. nullfs) the caller is at in the topology. 3054 * 3055 * We use a very simple frontend cache to reduce SMP conflicts, 3056 * which we have to do because the mountlist scan needs an exclusive 3057 * lock around its ripout info list. Not to mention that there might 3058 * be a lot of mounts. 3059 */ 3060 struct findmount_info { 3061 struct mount *result; 3062 struct mount *nch_mount; 3063 struct namecache *nch_ncp; 3064 }; 3065 3066 static 3067 struct ncmount_cache * 3068 ncmount_cache_lookup(struct mount *mp, struct namecache *ncp) 3069 { 3070 int hash; 3071 3072 hash = ((int)(intptr_t)mp / sizeof(*mp)) ^ 3073 ((int)(intptr_t)ncp / sizeof(*ncp)); 3074 hash = (hash & 0x7FFFFFFF) % NCMOUNT_NUMCACHE; 3075 return (&ncmount_cache[hash]); 3076 } 3077 3078 static 3079 int 3080 cache_findmount_callback(struct mount *mp, void *data) 3081 { 3082 struct findmount_info *info = data; 3083 3084 /* 3085 * Check the mount's mounted-on point against the passed nch. 3086 */ 3087 if (mp->mnt_ncmounton.mount == info->nch_mount && 3088 mp->mnt_ncmounton.ncp == info->nch_ncp 3089 ) { 3090 info->result = mp; 3091 atomic_add_int(&mp->mnt_refs, 1); 3092 return(-1); 3093 } 3094 return(0); 3095 } 3096 3097 struct mount * 3098 cache_findmount(struct nchandle *nch) 3099 { 3100 struct findmount_info info; 3101 struct ncmount_cache *ncc; 3102 struct mount *mp; 3103 3104 /* 3105 * Fast 3106 */ 3107 if (ncmount_cache_enable == 0) { 3108 ncc = NULL; 3109 goto skip; 3110 } 3111 ncc = ncmount_cache_lookup(nch->mount, nch->ncp); 3112 if (ncc->ncp == nch->ncp) { 3113 spin_lock_shared(&ncc->spin); 3114 if (ncc->isneg == 0 && 3115 ncc->ncp == nch->ncp && (mp = ncc->mp) != NULL) { 3116 if (mp->mnt_ncmounton.mount == nch->mount && 3117 mp->mnt_ncmounton.ncp == nch->ncp) { 3118 /* 3119 * Cache hit (positive) 3120 */ 3121 atomic_add_int(&mp->mnt_refs, 1); 3122 spin_unlock_shared(&ncc->spin); 3123 ++ncmount_cache_hit; 3124 return(mp); 3125 } 3126 /* else cache miss */ 3127 } 3128 if (ncc->isneg && 3129 ncc->ncp == nch->ncp && ncc->mp == nch->mount) { 3130 /* 3131 * Cache hit (negative) 3132 */ 3133 spin_unlock_shared(&ncc->spin); 3134 ++ncmount_cache_hit; 3135 return(NULL); 3136 } 3137 spin_unlock_shared(&ncc->spin); 3138 } 3139 skip: 3140 3141 /* 3142 * Slow 3143 */ 3144 info.result = NULL; 3145 info.nch_mount = nch->mount; 3146 info.nch_ncp = nch->ncp; 3147 mountlist_scan(cache_findmount_callback, &info, 3148 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 3149 3150 /* 3151 * Cache the result. 3152 * 3153 * Negative lookups: We cache the originating {ncp,mp}. (mp) is 3154 * only used for pointer comparisons and is not 3155 * referenced (otherwise there would be dangling 3156 * refs). 3157 * 3158 * Positive lookups: We cache the originating {ncp} and the target 3159 * (mp). (mp) is referenced. 3160 * 3161 * Indeterminant: If the match is undergoing an unmount we do 3162 * not cache it to avoid racing cache_unmounting(), 3163 * but still return the match. 3164 */ 3165 if (ncc) { 3166 spin_lock(&ncc->spin); 3167 if (info.result == NULL) { 3168 if (ncc->isneg == 0 && ncc->mp) 3169 atomic_add_int(&ncc->mp->mnt_refs, -1); 3170 ncc->ncp = nch->ncp; 3171 ncc->mp = nch->mount; 3172 ncc->isneg = 1; 3173 spin_unlock(&ncc->spin); 3174 ++ncmount_cache_overwrite; 3175 } else if ((info.result->mnt_kern_flag & MNTK_UNMOUNT) == 0) { 3176 if (ncc->isneg == 0 && ncc->mp) 3177 atomic_add_int(&ncc->mp->mnt_refs, -1); 3178 atomic_add_int(&info.result->mnt_refs, 1); 3179 ncc->ncp = nch->ncp; 3180 ncc->mp = info.result; 3181 ncc->isneg = 0; 3182 spin_unlock(&ncc->spin); 3183 ++ncmount_cache_overwrite; 3184 } else { 3185 spin_unlock(&ncc->spin); 3186 } 3187 ++ncmount_cache_miss; 3188 } 3189 return(info.result); 3190 } 3191 3192 void 3193 cache_dropmount(struct mount *mp) 3194 { 3195 atomic_add_int(&mp->mnt_refs, -1); 3196 } 3197 3198 void 3199 cache_ismounting(struct mount *mp) 3200 { 3201 struct nchandle *nch = &mp->mnt_ncmounton; 3202 struct ncmount_cache *ncc; 3203 3204 ncc = ncmount_cache_lookup(nch->mount, nch->ncp); 3205 if (ncc->isneg && 3206 ncc->ncp == nch->ncp && ncc->mp == nch->mount) { 3207 spin_lock(&ncc->spin); 3208 if (ncc->isneg && 3209 ncc->ncp == nch->ncp && ncc->mp == nch->mount) { 3210 ncc->ncp = NULL; 3211 ncc->mp = NULL; 3212 } 3213 spin_unlock(&ncc->spin); 3214 } 3215 } 3216 3217 void 3218 cache_unmounting(struct mount *mp) 3219 { 3220 struct nchandle *nch = &mp->mnt_ncmounton; 3221 struct ncmount_cache *ncc; 3222 3223 ncc = ncmount_cache_lookup(nch->mount, nch->ncp); 3224 if (ncc->isneg == 0 && 3225 ncc->ncp == nch->ncp && ncc->mp == mp) { 3226 spin_lock(&ncc->spin); 3227 if (ncc->isneg == 0 && 3228 ncc->ncp == nch->ncp && ncc->mp == mp) { 3229 atomic_add_int(&mp->mnt_refs, -1); 3230 ncc->ncp = NULL; 3231 ncc->mp = NULL; 3232 } 3233 spin_unlock(&ncc->spin); 3234 } 3235 } 3236 3237 /* 3238 * Resolve an unresolved namecache entry, generally by looking it up. 3239 * The passed ncp must be locked and refd. 3240 * 3241 * Theoretically since a vnode cannot be recycled while held, and since 3242 * the nc_parent chain holds its vnode as long as children exist, the 3243 * direct parent of the cache entry we are trying to resolve should 3244 * have a valid vnode. If not then generate an error that we can 3245 * determine is related to a resolver bug. 3246 * 3247 * However, if a vnode was in the middle of a recyclement when the NCP 3248 * got locked, ncp->nc_vp might point to a vnode that is about to become 3249 * invalid. cache_resolve() handles this case by unresolving the entry 3250 * and then re-resolving it. 3251 * 3252 * Note that successful resolution does not necessarily return an error 3253 * code of 0. If the ncp resolves to a negative cache hit then ENOENT 3254 * will be returned. 3255 */ 3256 int 3257 cache_resolve(struct nchandle *nch, struct ucred *cred) 3258 { 3259 struct namecache *par_tmp; 3260 struct namecache *par; 3261 struct namecache *ncp; 3262 struct nchandle nctmp; 3263 struct mount *mp; 3264 struct vnode *dvp; 3265 int error; 3266 3267 ncp = nch->ncp; 3268 mp = nch->mount; 3269 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE); 3270 restart: 3271 /* 3272 * If the ncp is already resolved we have nothing to do. However, 3273 * we do want to guarentee that a usable vnode is returned when 3274 * a vnode is present, so make sure it hasn't been reclaimed. 3275 */ 3276 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 3277 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 3278 _cache_setunresolved(ncp); 3279 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 3280 return (ncp->nc_error); 3281 } 3282 3283 /* 3284 * If the ncp was destroyed it will never resolve again. This 3285 * can basically only happen when someone is chdir'd into an 3286 * empty directory which is then rmdir'd. We want to catch this 3287 * here and not dive the VFS because the VFS might actually 3288 * have a way to re-resolve the disconnected ncp, which will 3289 * result in inconsistencies in the cdir/nch for proc->p_fd. 3290 */ 3291 if (ncp->nc_flag & NCF_DESTROYED) { 3292 kprintf("Warning: cache_resolve: ncp '%s' was unlinked\n", 3293 ncp->nc_name); 3294 return(EINVAL); 3295 } 3296 3297 /* 3298 * Mount points need special handling because the parent does not 3299 * belong to the same filesystem as the ncp. 3300 */ 3301 if (ncp == mp->mnt_ncmountpt.ncp) 3302 return (cache_resolve_mp(mp)); 3303 3304 /* 3305 * We expect an unbroken chain of ncps to at least the mount point, 3306 * and even all the way to root (but this code doesn't have to go 3307 * past the mount point). 3308 */ 3309 if (ncp->nc_parent == NULL) { 3310 kprintf("EXDEV case 1 %p %*.*s\n", ncp, 3311 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 3312 ncp->nc_error = EXDEV; 3313 return(ncp->nc_error); 3314 } 3315 3316 /* 3317 * The vp's of the parent directories in the chain are held via vhold() 3318 * due to the existance of the child, and should not disappear. 3319 * However, there are cases where they can disappear: 3320 * 3321 * - due to filesystem I/O errors. 3322 * - due to NFS being stupid about tracking the namespace and 3323 * destroys the namespace for entire directories quite often. 3324 * - due to forced unmounts. 3325 * - due to an rmdir (parent will be marked DESTROYED) 3326 * 3327 * When this occurs we have to track the chain backwards and resolve 3328 * it, looping until the resolver catches up to the current node. We 3329 * could recurse here but we might run ourselves out of kernel stack 3330 * so we do it in a more painful manner. This situation really should 3331 * not occur all that often, or if it does not have to go back too 3332 * many nodes to resolve the ncp. 3333 */ 3334 while ((dvp = cache_dvpref(ncp)) == NULL) { 3335 /* 3336 * This case can occur if a process is CD'd into a 3337 * directory which is then rmdir'd. If the parent is marked 3338 * destroyed there is no point trying to resolve it. 3339 */ 3340 if (ncp->nc_parent->nc_flag & NCF_DESTROYED) 3341 return(ENOENT); 3342 par = ncp->nc_parent; 3343 _cache_hold(par); 3344 _cache_lock(par); 3345 while ((par_tmp = par->nc_parent) != NULL && 3346 par_tmp->nc_vp == NULL) { 3347 _cache_hold(par_tmp); 3348 _cache_lock(par_tmp); 3349 _cache_put(par); 3350 par = par_tmp; 3351 } 3352 if (par->nc_parent == NULL) { 3353 kprintf("EXDEV case 2 %*.*s\n", 3354 par->nc_nlen, par->nc_nlen, par->nc_name); 3355 _cache_put(par); 3356 return (EXDEV); 3357 } 3358 /* 3359 * The parent is not set in stone, ref and lock it to prevent 3360 * it from disappearing. Also note that due to renames it 3361 * is possible for our ncp to move and for par to no longer 3362 * be one of its parents. We resolve it anyway, the loop 3363 * will handle any moves. 3364 */ 3365 _cache_get(par); /* additional hold/lock */ 3366 _cache_put(par); /* from earlier hold/lock */ 3367 if (par == nch->mount->mnt_ncmountpt.ncp) { 3368 cache_resolve_mp(nch->mount); 3369 } else if ((dvp = cache_dvpref(par)) == NULL) { 3370 kprintf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name); 3371 _cache_put(par); 3372 continue; 3373 } else { 3374 if (par->nc_flag & NCF_UNRESOLVED) { 3375 nctmp.mount = mp; 3376 nctmp.ncp = par; 3377 par->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred); 3378 } 3379 vrele(dvp); 3380 } 3381 if ((error = par->nc_error) != 0) { 3382 if (par->nc_error != EAGAIN) { 3383 kprintf("EXDEV case 3 %*.*s error %d\n", 3384 par->nc_nlen, par->nc_nlen, par->nc_name, 3385 par->nc_error); 3386 _cache_put(par); 3387 return(error); 3388 } 3389 kprintf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n", 3390 par, par->nc_nlen, par->nc_nlen, par->nc_name); 3391 } 3392 _cache_put(par); 3393 /* loop */ 3394 } 3395 3396 /* 3397 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected 3398 * ncp's and reattach them. If this occurs the original ncp is marked 3399 * EAGAIN to force a relookup. 3400 * 3401 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed 3402 * ncp must already be resolved. 3403 */ 3404 if (dvp) { 3405 nctmp.mount = mp; 3406 nctmp.ncp = ncp; 3407 ncp->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred); 3408 vrele(dvp); 3409 } else { 3410 ncp->nc_error = EPERM; 3411 } 3412 if (ncp->nc_error == EAGAIN) { 3413 kprintf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n", 3414 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 3415 goto restart; 3416 } 3417 return(ncp->nc_error); 3418 } 3419 3420 /* 3421 * Resolve the ncp associated with a mount point. Such ncp's almost always 3422 * remain resolved and this routine is rarely called. NFS MPs tends to force 3423 * re-resolution more often due to its mac-truck-smash-the-namecache 3424 * method of tracking namespace changes. 3425 * 3426 * The semantics for this call is that the passed ncp must be locked on 3427 * entry and will be locked on return. However, if we actually have to 3428 * resolve the mount point we temporarily unlock the entry in order to 3429 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of 3430 * the unlock we have to recheck the flags after we relock. 3431 */ 3432 static int 3433 cache_resolve_mp(struct mount *mp) 3434 { 3435 struct namecache *ncp = mp->mnt_ncmountpt.ncp; 3436 struct vnode *vp; 3437 int error; 3438 3439 KKASSERT(mp != NULL); 3440 3441 /* 3442 * If the ncp is already resolved we have nothing to do. However, 3443 * we do want to guarentee that a usable vnode is returned when 3444 * a vnode is present, so make sure it hasn't been reclaimed. 3445 */ 3446 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 3447 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 3448 _cache_setunresolved(ncp); 3449 } 3450 3451 if (ncp->nc_flag & NCF_UNRESOLVED) { 3452 _cache_unlock(ncp); 3453 while (vfs_busy(mp, 0)) 3454 ; 3455 error = VFS_ROOT(mp, &vp); 3456 _cache_lock(ncp); 3457 3458 /* 3459 * recheck the ncp state after relocking. 3460 */ 3461 if (ncp->nc_flag & NCF_UNRESOLVED) { 3462 ncp->nc_error = error; 3463 if (error == 0) { 3464 _cache_setvp(mp, ncp, vp); 3465 vput(vp); 3466 } else { 3467 kprintf("[diagnostic] cache_resolve_mp: failed" 3468 " to resolve mount %p err=%d ncp=%p\n", 3469 mp, error, ncp); 3470 _cache_setvp(mp, ncp, NULL); 3471 } 3472 } else if (error == 0) { 3473 vput(vp); 3474 } 3475 vfs_unbusy(mp); 3476 } 3477 return(ncp->nc_error); 3478 } 3479 3480 /* 3481 * Clean out negative cache entries when too many have accumulated. 3482 */ 3483 static void 3484 _cache_cleanneg(int count) 3485 { 3486 struct namecache *ncp; 3487 3488 /* 3489 * Attempt to clean out the specified number of negative cache 3490 * entries. 3491 */ 3492 while (count) { 3493 spin_lock(&ncspin); 3494 ncp = TAILQ_FIRST(&ncneglist); 3495 if (ncp == NULL) { 3496 spin_unlock(&ncspin); 3497 break; 3498 } 3499 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 3500 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 3501 _cache_hold(ncp); 3502 spin_unlock(&ncspin); 3503 3504 /* 3505 * This can race, so we must re-check that the ncp 3506 * is on the ncneglist after successfully locking it. 3507 */ 3508 if (_cache_lock_special(ncp) == 0) { 3509 if (ncp->nc_vp == NULL && 3510 (ncp->nc_flag & NCF_UNRESOLVED) == 0) { 3511 ncp = cache_zap(ncp, 1); 3512 if (ncp) 3513 _cache_drop(ncp); 3514 } else { 3515 kprintf("cache_cleanneg: race avoided\n"); 3516 _cache_unlock(ncp); 3517 } 3518 } else { 3519 _cache_drop(ncp); 3520 } 3521 --count; 3522 } 3523 } 3524 3525 /* 3526 * Clean out positive cache entries when too many have accumulated. 3527 */ 3528 static void 3529 _cache_cleanpos(int count) 3530 { 3531 static volatile int rover; 3532 struct nchash_head *nchpp; 3533 struct namecache *ncp; 3534 int rover_copy; 3535 3536 /* 3537 * Attempt to clean out the specified number of negative cache 3538 * entries. 3539 */ 3540 while (count) { 3541 rover_copy = ++rover; /* MPSAFEENOUGH */ 3542 cpu_ccfence(); 3543 nchpp = NCHHASH(rover_copy); 3544 3545 spin_lock_shared(&nchpp->spin); 3546 ncp = LIST_FIRST(&nchpp->list); 3547 while (ncp && (ncp->nc_flag & NCF_DESTROYED)) 3548 ncp = LIST_NEXT(ncp, nc_hash); 3549 if (ncp) 3550 _cache_hold(ncp); 3551 spin_unlock_shared(&nchpp->spin); 3552 3553 if (ncp) { 3554 if (_cache_lock_special(ncp) == 0) { 3555 ncp = cache_zap(ncp, 1); 3556 if (ncp) 3557 _cache_drop(ncp); 3558 } else { 3559 _cache_drop(ncp); 3560 } 3561 } 3562 --count; 3563 } 3564 } 3565 3566 /* 3567 * This is a kitchen sink function to clean out ncps which we 3568 * tried to zap from cache_drop() but failed because we were 3569 * unable to acquire the parent lock. 3570 * 3571 * Such entries can also be removed via cache_inval_vp(), such 3572 * as when unmounting. 3573 */ 3574 static void 3575 _cache_cleandefered(void) 3576 { 3577 struct nchash_head *nchpp; 3578 struct namecache *ncp; 3579 struct namecache dummy; 3580 int i; 3581 3582 numdefered = 0; 3583 bzero(&dummy, sizeof(dummy)); 3584 dummy.nc_flag = NCF_DESTROYED; 3585 dummy.nc_refs = 1; 3586 3587 for (i = 0; i <= nchash; ++i) { 3588 nchpp = &nchashtbl[i]; 3589 3590 spin_lock(&nchpp->spin); 3591 LIST_INSERT_HEAD(&nchpp->list, &dummy, nc_hash); 3592 ncp = &dummy; 3593 while ((ncp = LIST_NEXT(ncp, nc_hash)) != NULL) { 3594 if ((ncp->nc_flag & NCF_DEFEREDZAP) == 0) 3595 continue; 3596 LIST_REMOVE(&dummy, nc_hash); 3597 LIST_INSERT_AFTER(ncp, &dummy, nc_hash); 3598 _cache_hold(ncp); 3599 spin_unlock(&nchpp->spin); 3600 if (_cache_lock_nonblock(ncp) == 0) { 3601 ncp->nc_flag &= ~NCF_DEFEREDZAP; 3602 _cache_unlock(ncp); 3603 } 3604 _cache_drop(ncp); 3605 spin_lock(&nchpp->spin); 3606 ncp = &dummy; 3607 } 3608 LIST_REMOVE(&dummy, nc_hash); 3609 spin_unlock(&nchpp->spin); 3610 } 3611 } 3612 3613 /* 3614 * Name cache initialization, from vfsinit() when we are booting 3615 */ 3616 void 3617 nchinit(void) 3618 { 3619 int i; 3620 globaldata_t gd; 3621 3622 /* initialise per-cpu namecache effectiveness statistics. */ 3623 for (i = 0; i < ncpus; ++i) { 3624 gd = globaldata_find(i); 3625 gd->gd_nchstats = &nchstats[i]; 3626 } 3627 TAILQ_INIT(&ncneglist); 3628 spin_init(&ncspin, "nchinit"); 3629 nchashtbl = hashinit_ext(desiredvnodes / 2, 3630 sizeof(struct nchash_head), 3631 M_VFSCACHE, &nchash); 3632 for (i = 0; i <= (int)nchash; ++i) { 3633 LIST_INIT(&nchashtbl[i].list); 3634 spin_init(&nchashtbl[i].spin, "nchinit_hash"); 3635 } 3636 for (i = 0; i < NCMOUNT_NUMCACHE; ++i) 3637 spin_init(&ncmount_cache[i].spin, "nchinit_cache"); 3638 nclockwarn = 5 * hz; 3639 } 3640 3641 /* 3642 * Called from start_init() to bootstrap the root filesystem. Returns 3643 * a referenced, unlocked namecache record. 3644 */ 3645 void 3646 cache_allocroot(struct nchandle *nch, struct mount *mp, struct vnode *vp) 3647 { 3648 nch->ncp = cache_alloc(0); 3649 nch->mount = mp; 3650 atomic_add_int(&mp->mnt_refs, 1); 3651 if (vp) 3652 _cache_setvp(nch->mount, nch->ncp, vp); 3653 } 3654 3655 /* 3656 * vfs_cache_setroot() 3657 * 3658 * Create an association between the root of our namecache and 3659 * the root vnode. This routine may be called several times during 3660 * booting. 3661 * 3662 * If the caller intends to save the returned namecache pointer somewhere 3663 * it must cache_hold() it. 3664 */ 3665 void 3666 vfs_cache_setroot(struct vnode *nvp, struct nchandle *nch) 3667 { 3668 struct vnode *ovp; 3669 struct nchandle onch; 3670 3671 ovp = rootvnode; 3672 onch = rootnch; 3673 rootvnode = nvp; 3674 if (nch) 3675 rootnch = *nch; 3676 else 3677 cache_zero(&rootnch); 3678 if (ovp) 3679 vrele(ovp); 3680 if (onch.ncp) 3681 cache_drop(&onch); 3682 } 3683 3684 /* 3685 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache 3686 * topology and is being removed as quickly as possible. The new VOP_N*() 3687 * API calls are required to make specific adjustments using the supplied 3688 * ncp pointers rather then just bogusly purging random vnodes. 3689 * 3690 * Invalidate all namecache entries to a particular vnode as well as 3691 * any direct children of that vnode in the namecache. This is a 3692 * 'catch all' purge used by filesystems that do not know any better. 3693 * 3694 * Note that the linkage between the vnode and its namecache entries will 3695 * be removed, but the namecache entries themselves might stay put due to 3696 * active references from elsewhere in the system or due to the existance of 3697 * the children. The namecache topology is left intact even if we do not 3698 * know what the vnode association is. Such entries will be marked 3699 * NCF_UNRESOLVED. 3700 */ 3701 void 3702 cache_purge(struct vnode *vp) 3703 { 3704 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN); 3705 } 3706 3707 /* 3708 * Flush all entries referencing a particular filesystem. 3709 * 3710 * Since we need to check it anyway, we will flush all the invalid 3711 * entries at the same time. 3712 */ 3713 #if 0 3714 3715 void 3716 cache_purgevfs(struct mount *mp) 3717 { 3718 struct nchash_head *nchpp; 3719 struct namecache *ncp, *nnp; 3720 3721 /* 3722 * Scan hash tables for applicable entries. 3723 */ 3724 for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) { 3725 spin_lock_wr(&nchpp->spin); XXX 3726 ncp = LIST_FIRST(&nchpp->list); 3727 if (ncp) 3728 _cache_hold(ncp); 3729 while (ncp) { 3730 nnp = LIST_NEXT(ncp, nc_hash); 3731 if (nnp) 3732 _cache_hold(nnp); 3733 if (ncp->nc_mount == mp) { 3734 _cache_lock(ncp); 3735 ncp = cache_zap(ncp, 0); 3736 if (ncp) 3737 _cache_drop(ncp); 3738 } else { 3739 _cache_drop(ncp); 3740 } 3741 ncp = nnp; 3742 } 3743 spin_unlock_wr(&nchpp->spin); XXX 3744 } 3745 } 3746 3747 #endif 3748 3749 static int disablecwd; 3750 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, 3751 "Disable getcwd"); 3752 3753 static u_long numcwdcalls; 3754 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdcalls, CTLFLAG_RD, &numcwdcalls, 0, 3755 "Number of current directory resolution calls"); 3756 static u_long numcwdfailnf; 3757 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailnf, CTLFLAG_RD, &numcwdfailnf, 0, 3758 "Number of current directory failures due to lack of file"); 3759 static u_long numcwdfailsz; 3760 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailsz, CTLFLAG_RD, &numcwdfailsz, 0, 3761 "Number of current directory failures due to large result"); 3762 static u_long numcwdfound; 3763 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfound, CTLFLAG_RD, &numcwdfound, 0, 3764 "Number of current directory resolution successes"); 3765 3766 /* 3767 * MPALMOSTSAFE 3768 */ 3769 int 3770 sys___getcwd(struct __getcwd_args *uap) 3771 { 3772 u_int buflen; 3773 int error; 3774 char *buf; 3775 char *bp; 3776 3777 if (disablecwd) 3778 return (ENODEV); 3779 3780 buflen = uap->buflen; 3781 if (buflen == 0) 3782 return (EINVAL); 3783 if (buflen > MAXPATHLEN) 3784 buflen = MAXPATHLEN; 3785 3786 buf = kmalloc(buflen, M_TEMP, M_WAITOK); 3787 bp = kern_getcwd(buf, buflen, &error); 3788 if (error == 0) 3789 error = copyout(bp, uap->buf, strlen(bp) + 1); 3790 kfree(buf, M_TEMP); 3791 return (error); 3792 } 3793 3794 char * 3795 kern_getcwd(char *buf, size_t buflen, int *error) 3796 { 3797 struct proc *p = curproc; 3798 char *bp; 3799 int i, slash_prefixed; 3800 struct filedesc *fdp; 3801 struct nchandle nch; 3802 struct namecache *ncp; 3803 3804 numcwdcalls++; 3805 bp = buf; 3806 bp += buflen - 1; 3807 *bp = '\0'; 3808 fdp = p->p_fd; 3809 slash_prefixed = 0; 3810 3811 nch = fdp->fd_ncdir; 3812 ncp = nch.ncp; 3813 if (ncp) 3814 _cache_hold(ncp); 3815 3816 while (ncp && (ncp != fdp->fd_nrdir.ncp || 3817 nch.mount != fdp->fd_nrdir.mount) 3818 ) { 3819 /* 3820 * While traversing upwards if we encounter the root 3821 * of the current mount we have to skip to the mount point 3822 * in the underlying filesystem. 3823 */ 3824 if (ncp == nch.mount->mnt_ncmountpt.ncp) { 3825 nch = nch.mount->mnt_ncmounton; 3826 _cache_drop(ncp); 3827 ncp = nch.ncp; 3828 if (ncp) 3829 _cache_hold(ncp); 3830 continue; 3831 } 3832 3833 /* 3834 * Prepend the path segment 3835 */ 3836 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 3837 if (bp == buf) { 3838 numcwdfailsz++; 3839 *error = ERANGE; 3840 bp = NULL; 3841 goto done; 3842 } 3843 *--bp = ncp->nc_name[i]; 3844 } 3845 if (bp == buf) { 3846 numcwdfailsz++; 3847 *error = ERANGE; 3848 bp = NULL; 3849 goto done; 3850 } 3851 *--bp = '/'; 3852 slash_prefixed = 1; 3853 3854 /* 3855 * Go up a directory. This isn't a mount point so we don't 3856 * have to check again. 3857 */ 3858 while ((nch.ncp = ncp->nc_parent) != NULL) { 3859 if (ncp_shared_lock_disable) 3860 _cache_lock(ncp); 3861 else 3862 _cache_lock_shared(ncp); 3863 if (nch.ncp != ncp->nc_parent) { 3864 _cache_unlock(ncp); 3865 continue; 3866 } 3867 _cache_hold(nch.ncp); 3868 _cache_unlock(ncp); 3869 break; 3870 } 3871 _cache_drop(ncp); 3872 ncp = nch.ncp; 3873 } 3874 if (ncp == NULL) { 3875 numcwdfailnf++; 3876 *error = ENOENT; 3877 bp = NULL; 3878 goto done; 3879 } 3880 if (!slash_prefixed) { 3881 if (bp == buf) { 3882 numcwdfailsz++; 3883 *error = ERANGE; 3884 bp = NULL; 3885 goto done; 3886 } 3887 *--bp = '/'; 3888 } 3889 numcwdfound++; 3890 *error = 0; 3891 done: 3892 if (ncp) 3893 _cache_drop(ncp); 3894 return (bp); 3895 } 3896 3897 /* 3898 * Thus begins the fullpath magic. 3899 * 3900 * The passed nchp is referenced but not locked. 3901 */ 3902 static int disablefullpath; 3903 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, 3904 &disablefullpath, 0, 3905 "Disable fullpath lookups"); 3906 3907 static u_int numfullpathcalls; 3908 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathcalls, CTLFLAG_RD, 3909 &numfullpathcalls, 0, 3910 "Number of full path resolutions in progress"); 3911 static u_int numfullpathfailnf; 3912 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailnf, CTLFLAG_RD, 3913 &numfullpathfailnf, 0, 3914 "Number of full path resolution failures due to lack of file"); 3915 static u_int numfullpathfailsz; 3916 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailsz, CTLFLAG_RD, 3917 &numfullpathfailsz, 0, 3918 "Number of full path resolution failures due to insufficient memory"); 3919 static u_int numfullpathfound; 3920 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfound, CTLFLAG_RD, 3921 &numfullpathfound, 0, 3922 "Number of full path resolution successes"); 3923 3924 int 3925 cache_fullpath(struct proc *p, struct nchandle *nchp, struct nchandle *nchbase, 3926 char **retbuf, char **freebuf, int guess) 3927 { 3928 struct nchandle fd_nrdir; 3929 struct nchandle nch; 3930 struct namecache *ncp; 3931 struct mount *mp, *new_mp; 3932 char *bp, *buf; 3933 int slash_prefixed; 3934 int error = 0; 3935 int i; 3936 3937 atomic_add_int(&numfullpathcalls, -1); 3938 3939 *retbuf = NULL; 3940 *freebuf = NULL; 3941 3942 buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK); 3943 bp = buf + MAXPATHLEN - 1; 3944 *bp = '\0'; 3945 if (nchbase) 3946 fd_nrdir = *nchbase; 3947 else if (p != NULL) 3948 fd_nrdir = p->p_fd->fd_nrdir; 3949 else 3950 fd_nrdir = rootnch; 3951 slash_prefixed = 0; 3952 nch = *nchp; 3953 ncp = nch.ncp; 3954 if (ncp) 3955 _cache_hold(ncp); 3956 mp = nch.mount; 3957 3958 while (ncp && (ncp != fd_nrdir.ncp || mp != fd_nrdir.mount)) { 3959 new_mp = NULL; 3960 3961 /* 3962 * If we are asked to guess the upwards path, we do so whenever 3963 * we encounter an ncp marked as a mountpoint. We try to find 3964 * the actual mountpoint by finding the mountpoint with this 3965 * ncp. 3966 */ 3967 if (guess && (ncp->nc_flag & NCF_ISMOUNTPT)) { 3968 new_mp = mount_get_by_nc(ncp); 3969 } 3970 /* 3971 * While traversing upwards if we encounter the root 3972 * of the current mount we have to skip to the mount point. 3973 */ 3974 if (ncp == mp->mnt_ncmountpt.ncp) { 3975 new_mp = mp; 3976 } 3977 if (new_mp) { 3978 nch = new_mp->mnt_ncmounton; 3979 _cache_drop(ncp); 3980 ncp = nch.ncp; 3981 if (ncp) 3982 _cache_hold(ncp); 3983 mp = nch.mount; 3984 continue; 3985 } 3986 3987 /* 3988 * Prepend the path segment 3989 */ 3990 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 3991 if (bp == buf) { 3992 numfullpathfailsz++; 3993 kfree(buf, M_TEMP); 3994 error = ENOMEM; 3995 goto done; 3996 } 3997 *--bp = ncp->nc_name[i]; 3998 } 3999 if (bp == buf) { 4000 numfullpathfailsz++; 4001 kfree(buf, M_TEMP); 4002 error = ENOMEM; 4003 goto done; 4004 } 4005 *--bp = '/'; 4006 slash_prefixed = 1; 4007 4008 /* 4009 * Go up a directory. This isn't a mount point so we don't 4010 * have to check again. 4011 * 4012 * We can only safely access nc_parent with ncp held locked. 4013 */ 4014 while ((nch.ncp = ncp->nc_parent) != NULL) { 4015 _cache_lock(ncp); 4016 if (nch.ncp != ncp->nc_parent) { 4017 _cache_unlock(ncp); 4018 continue; 4019 } 4020 _cache_hold(nch.ncp); 4021 _cache_unlock(ncp); 4022 break; 4023 } 4024 _cache_drop(ncp); 4025 ncp = nch.ncp; 4026 } 4027 if (ncp == NULL) { 4028 numfullpathfailnf++; 4029 kfree(buf, M_TEMP); 4030 error = ENOENT; 4031 goto done; 4032 } 4033 4034 if (!slash_prefixed) { 4035 if (bp == buf) { 4036 numfullpathfailsz++; 4037 kfree(buf, M_TEMP); 4038 error = ENOMEM; 4039 goto done; 4040 } 4041 *--bp = '/'; 4042 } 4043 numfullpathfound++; 4044 *retbuf = bp; 4045 *freebuf = buf; 4046 error = 0; 4047 done: 4048 if (ncp) 4049 _cache_drop(ncp); 4050 return(error); 4051 } 4052 4053 int 4054 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, 4055 char **freebuf, int guess) 4056 { 4057 struct namecache *ncp; 4058 struct nchandle nch; 4059 int error; 4060 4061 *freebuf = NULL; 4062 atomic_add_int(&numfullpathcalls, 1); 4063 if (disablefullpath) 4064 return (ENODEV); 4065 4066 if (p == NULL) 4067 return (EINVAL); 4068 4069 /* vn is NULL, client wants us to use p->p_textvp */ 4070 if (vn == NULL) { 4071 if ((vn = p->p_textvp) == NULL) 4072 return (EINVAL); 4073 } 4074 spin_lock_shared(&vn->v_spin); 4075 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) { 4076 if (ncp->nc_nlen) 4077 break; 4078 } 4079 if (ncp == NULL) { 4080 spin_unlock_shared(&vn->v_spin); 4081 return (EINVAL); 4082 } 4083 _cache_hold(ncp); 4084 spin_unlock_shared(&vn->v_spin); 4085 4086 atomic_add_int(&numfullpathcalls, -1); 4087 nch.ncp = ncp; 4088 nch.mount = vn->v_mount; 4089 error = cache_fullpath(p, &nch, NULL, retbuf, freebuf, guess); 4090 _cache_drop(ncp); 4091 return (error); 4092 } 4093