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 87 #define MAX_RECURSION_DEPTH 64 88 89 /* 90 * Random lookups in the cache are accomplished with a hash table using 91 * a hash key of (nc_src_vp, name). Each hash chain has its own spin lock. 92 * 93 * Negative entries may exist and correspond to resolved namecache 94 * structures where nc_vp is NULL. In a negative entry, NCF_WHITEOUT 95 * will be set if the entry corresponds to a whited-out directory entry 96 * (verses simply not finding the entry at all). ncneglist is locked 97 * with a global spinlock (ncspin). 98 * 99 * MPSAFE RULES: 100 * 101 * (1) A ncp must be referenced before it can be locked. 102 * 103 * (2) A ncp must be locked in order to modify it. 104 * 105 * (3) ncp locks are always ordered child -> parent. That may seem 106 * backwards but forward scans use the hash table and thus can hold 107 * the parent unlocked when traversing downward. 108 * 109 * This allows insert/rename/delete/dot-dot and other operations 110 * to use ncp->nc_parent links. 111 * 112 * This also prevents a locked up e.g. NFS node from creating a 113 * chain reaction all the way back to the root vnode / namecache. 114 * 115 * (4) parent linkages require both the parent and child to be locked. 116 */ 117 118 /* 119 * Structures associated with name cacheing. 120 */ 121 #define NCHHASH(hash) (&nchashtbl[(hash) & nchash]) 122 #define MINNEG 1024 123 #define MINPOS 1024 124 #define NCMOUNT_NUMCACHE 1009 /* prime number */ 125 126 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries"); 127 128 LIST_HEAD(nchash_list, namecache); 129 130 /* 131 * Don't cachealign, but at least pad to 32 bytes so entries 132 * don't cross a cache line. 133 */ 134 struct nchash_head { 135 struct nchash_list list; /* 16 bytes */ 136 struct spinlock spin; /* 8 bytes */ 137 long pad01; /* 8 bytes */ 138 }; 139 140 struct ncmount_cache { 141 struct spinlock spin; 142 struct namecache *ncp; 143 struct mount *mp; 144 int isneg; /* if != 0 mp is originator and not target */ 145 }; 146 147 static struct nchash_head *nchashtbl; 148 static struct namecache_list ncneglist; 149 static struct spinlock ncspin; 150 static struct ncmount_cache ncmount_cache[NCMOUNT_NUMCACHE]; 151 152 /* 153 * ncvp_debug - debug cache_fromvp(). This is used by the NFS server 154 * to create the namecache infrastructure leading to a dangling vnode. 155 * 156 * 0 Only errors are reported 157 * 1 Successes are reported 158 * 2 Successes + the whole directory scan is reported 159 * 3 Force the directory scan code run as if the parent vnode did not 160 * have a namecache record, even if it does have one. 161 */ 162 static int ncvp_debug; 163 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0, 164 "Namecache debug level (0-3)"); 165 166 static u_long nchash; /* size of hash table */ 167 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, 168 "Size of namecache hash table"); 169 170 static int ncnegflush = 10; /* burst for negative flush */ 171 SYSCTL_INT(_debug, OID_AUTO, ncnegflush, CTLFLAG_RW, &ncnegflush, 0, 172 "Batch flush negative entries"); 173 174 static int ncposflush = 10; /* burst for positive flush */ 175 SYSCTL_INT(_debug, OID_AUTO, ncposflush, CTLFLAG_RW, &ncposflush, 0, 176 "Batch flush positive entries"); 177 178 static int ncnegfactor = 16; /* ratio of negative entries */ 179 SYSCTL_INT(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0, 180 "Ratio of namecache negative entries"); 181 182 static int nclockwarn; /* warn on locked entries in ticks */ 183 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0, 184 "Warn on locked namecache entries in ticks"); 185 186 static int numdefered; /* number of cache entries allocated */ 187 SYSCTL_INT(_debug, OID_AUTO, numdefered, CTLFLAG_RD, &numdefered, 0, 188 "Number of cache entries allocated"); 189 190 static int ncposlimit; /* number of cache entries allocated */ 191 SYSCTL_INT(_debug, OID_AUTO, ncposlimit, CTLFLAG_RW, &ncposlimit, 0, 192 "Number of cache entries allocated"); 193 194 static int ncp_shared_lock_disable = 0; 195 SYSCTL_INT(_debug, OID_AUTO, ncp_shared_lock_disable, CTLFLAG_RW, 196 &ncp_shared_lock_disable, 0, "Disable shared namecache locks"); 197 198 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode), 199 "sizeof(struct vnode)"); 200 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache), 201 "sizeof(struct namecache)"); 202 203 static int ncmount_cache_enable = 1; 204 SYSCTL_INT(_debug, OID_AUTO, ncmount_cache_enable, CTLFLAG_RW, 205 &ncmount_cache_enable, 0, "mount point cache"); 206 static long ncmount_cache_hit; 207 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_hit, CTLFLAG_RW, 208 &ncmount_cache_hit, 0, "mpcache hits"); 209 static long ncmount_cache_miss; 210 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_miss, CTLFLAG_RW, 211 &ncmount_cache_miss, 0, "mpcache misses"); 212 static long ncmount_cache_overwrite; 213 SYSCTL_LONG(_debug, OID_AUTO, ncmount_cache_overwrite, CTLFLAG_RW, 214 &ncmount_cache_overwrite, 0, "mpcache entry overwrites"); 215 216 static __inline void _cache_drop(struct namecache *ncp); 217 static int cache_resolve_mp(struct mount *mp); 218 static struct vnode *cache_dvpref(struct namecache *ncp); 219 static void _cache_lock(struct namecache *ncp); 220 static void _cache_setunresolved(struct namecache *ncp); 221 static void _cache_cleanneg(int count); 222 static void _cache_cleanpos(int count); 223 static void _cache_cleandefered(void); 224 static void _cache_unlink(struct namecache *ncp); 225 226 /* 227 * The new name cache statistics 228 */ 229 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics"); 230 static int numneg; 231 SYSCTL_INT(_vfs_cache, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0, 232 "Number of negative namecache entries"); 233 static int numcache; 234 SYSCTL_INT(_vfs_cache, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0, 235 "Number of namecaches entries"); 236 237 struct nchstats nchstats[SMP_MAXCPU]; 238 /* 239 * Export VFS cache effectiveness statistics to user-land. 240 * 241 * The statistics are left for aggregation to user-land so 242 * neat things can be achieved, like observing per-CPU cache 243 * distribution. 244 */ 245 static int 246 sysctl_nchstats(SYSCTL_HANDLER_ARGS) 247 { 248 struct globaldata *gd; 249 int i, error; 250 251 error = 0; 252 for (i = 0; i < ncpus; ++i) { 253 gd = globaldata_find(i); 254 if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats), 255 sizeof(struct nchstats)))) 256 break; 257 } 258 259 return (error); 260 } 261 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD, 262 0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics"); 263 264 static struct namecache *cache_zap(struct namecache *ncp, int nonblock); 265 266 /* 267 * Cache mount points and namecache records in order to avoid unnecessary 268 * atomic ops on mnt_refs and ncp->refs. This improves concurrent SMP 269 * performance and is particularly important on multi-socket systems to 270 * reduce cache-line ping-ponging. 271 * 272 * Try to keep the pcpu structure within one cache line (~64 bytes). 273 */ 274 #define MNTCACHE_COUNT 5 275 276 struct mntcache { 277 struct mount *mntary[MNTCACHE_COUNT]; 278 struct namecache *ncp1; 279 struct namecache *ncp2; 280 struct nchandle ncdir; 281 int iter; 282 int unused01; 283 } __cachealign; 284 285 static struct mntcache pcpu_mntcache[MAXCPU]; 286 287 static 288 void 289 _cache_mntref(struct mount *mp) 290 { 291 struct mntcache *cache = &pcpu_mntcache[mycpu->gd_cpuid]; 292 int i; 293 294 for (i = 0; i < MNTCACHE_COUNT; ++i) { 295 if (cache->mntary[i] != mp) 296 continue; 297 if (atomic_cmpset_ptr((void *)&cache->mntary[i], mp, NULL)) 298 return; 299 } 300 atomic_add_int(&mp->mnt_refs, 1); 301 } 302 303 static 304 void 305 _cache_mntrel(struct mount *mp) 306 { 307 struct mntcache *cache = &pcpu_mntcache[mycpu->gd_cpuid]; 308 int i; 309 310 for (i = 0; i < MNTCACHE_COUNT; ++i) { 311 if (cache->mntary[i] == NULL) { 312 mp = atomic_swap_ptr((void *)&cache->mntary[i], mp); 313 if (mp == NULL) 314 return; 315 } 316 } 317 i = (int)((uint32_t)++cache->iter % (uint32_t)MNTCACHE_COUNT); 318 mp = atomic_swap_ptr((void *)&cache->mntary[i], mp); 319 if (mp) 320 atomic_add_int(&mp->mnt_refs, -1); 321 } 322 323 /* 324 * Clears all cached mount points on all cpus. This routine should only 325 * be called when we are waiting for a mount to clear, e.g. so we can 326 * unmount. 327 */ 328 void 329 cache_clearmntcache(void) 330 { 331 int n; 332 333 for (n = 0; n < ncpus; ++n) { 334 struct mntcache *cache = &pcpu_mntcache[n]; 335 struct namecache *ncp; 336 struct mount *mp; 337 int i; 338 339 for (i = 0; i < MNTCACHE_COUNT; ++i) { 340 if (cache->mntary[i]) { 341 mp = atomic_swap_ptr( 342 (void *)&cache->mntary[i], NULL); 343 if (mp) 344 atomic_add_int(&mp->mnt_refs, -1); 345 } 346 } 347 if (cache->ncp1) { 348 ncp = atomic_swap_ptr((void *)&cache->ncp1, NULL); 349 if (ncp) 350 _cache_drop(ncp); 351 } 352 if (cache->ncp2) { 353 ncp = atomic_swap_ptr((void *)&cache->ncp2, NULL); 354 if (ncp) 355 _cache_drop(ncp); 356 } 357 if (cache->ncdir.ncp) { 358 ncp = atomic_swap_ptr((void *)&cache->ncdir.ncp, NULL); 359 if (ncp) 360 _cache_drop(ncp); 361 } 362 if (cache->ncdir.mount) { 363 mp = atomic_swap_ptr((void *)&cache->ncdir.mount, NULL); 364 if (mp) 365 atomic_add_int(&mp->mnt_refs, -1); 366 } 367 } 368 } 369 370 371 /* 372 * Namespace locking. The caller must already hold a reference to the 373 * namecache structure in order to lock/unlock it. This function prevents 374 * the namespace from being created or destroyed by accessors other then 375 * the lock holder. 376 * 377 * Note that holding a locked namecache structure prevents other threads 378 * from making namespace changes (e.g. deleting or creating), prevents 379 * vnode association state changes by other threads, and prevents the 380 * namecache entry from being resolved or unresolved by other threads. 381 * 382 * An exclusive lock owner has full authority to associate/disassociate 383 * vnodes and resolve/unresolve the locked ncp. 384 * 385 * A shared lock owner only has authority to acquire the underlying vnode, 386 * if any. 387 * 388 * The primary lock field is nc_lockstatus. nc_locktd is set after the 389 * fact (when locking) or cleared prior to unlocking. 390 * 391 * WARNING! Holding a locked ncp will prevent a vnode from being destroyed 392 * or recycled, but it does NOT help you if the vnode had already 393 * initiated a recyclement. If this is important, use cache_get() 394 * rather then cache_lock() (and deal with the differences in the 395 * way the refs counter is handled). Or, alternatively, make an 396 * unconditional call to cache_validate() or cache_resolve() 397 * after cache_lock() returns. 398 */ 399 static 400 void 401 _cache_lock(struct namecache *ncp) 402 { 403 thread_t td; 404 int didwarn; 405 int begticks; 406 int error; 407 u_int count; 408 409 KKASSERT(ncp->nc_refs != 0); 410 didwarn = 0; 411 begticks = 0; 412 td = curthread; 413 414 for (;;) { 415 count = ncp->nc_lockstatus; 416 cpu_ccfence(); 417 418 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 0) { 419 if (atomic_cmpset_int(&ncp->nc_lockstatus, 420 count, count + 1)) { 421 /* 422 * The vp associated with a locked ncp must 423 * be held to prevent it from being recycled. 424 * 425 * WARNING! If VRECLAIMED is set the vnode 426 * could already be in the middle of a recycle. 427 * Callers must use cache_vref() or 428 * cache_vget() on the locked ncp to 429 * validate the vp or set the cache entry 430 * to unresolved. 431 * 432 * NOTE! vhold() is allowed if we hold a 433 * lock on the ncp (which we do). 434 */ 435 ncp->nc_locktd = td; 436 if (ncp->nc_vp) 437 vhold(ncp->nc_vp); 438 break; 439 } 440 /* cmpset failed */ 441 continue; 442 } 443 if (ncp->nc_locktd == td) { 444 KKASSERT((count & NC_SHLOCK_FLAG) == 0); 445 if (atomic_cmpset_int(&ncp->nc_lockstatus, 446 count, count + 1)) { 447 break; 448 } 449 /* cmpset failed */ 450 continue; 451 } 452 tsleep_interlock(&ncp->nc_locktd, 0); 453 if (atomic_cmpset_int(&ncp->nc_lockstatus, count, 454 count | NC_EXLOCK_REQ) == 0) { 455 /* cmpset failed */ 456 continue; 457 } 458 if (begticks == 0) 459 begticks = ticks; 460 error = tsleep(&ncp->nc_locktd, PINTERLOCKED, 461 "clock", nclockwarn); 462 if (error == EWOULDBLOCK) { 463 if (didwarn == 0) { 464 didwarn = ticks; 465 kprintf("[diagnostic] cache_lock: " 466 "%s blocked on %p %08x", 467 td->td_comm, ncp, count); 468 kprintf(" \"%*.*s\"\n", 469 ncp->nc_nlen, ncp->nc_nlen, 470 ncp->nc_name); 471 } 472 } 473 /* loop */ 474 } 475 if (didwarn) { 476 kprintf("[diagnostic] cache_lock: %s unblocked %*.*s after " 477 "%d secs\n", 478 td->td_comm, 479 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name, 480 (int)(ticks + (hz / 2) - begticks) / hz); 481 } 482 } 483 484 /* 485 * The shared lock works similarly to the exclusive lock except 486 * nc_locktd is left NULL and we need an interlock (VHOLD) to 487 * prevent vhold() races, since the moment our cmpset_int succeeds 488 * another cpu can come in and get its own shared lock. 489 * 490 * A critical section is needed to prevent interruption during the 491 * VHOLD interlock. 492 */ 493 static 494 void 495 _cache_lock_shared(struct namecache *ncp) 496 { 497 int didwarn; 498 int error; 499 u_int count; 500 u_int optreq = NC_EXLOCK_REQ; 501 502 KKASSERT(ncp->nc_refs != 0); 503 didwarn = 0; 504 505 for (;;) { 506 count = ncp->nc_lockstatus; 507 cpu_ccfence(); 508 509 if ((count & ~NC_SHLOCK_REQ) == 0) { 510 crit_enter(); 511 if (atomic_cmpset_int(&ncp->nc_lockstatus, 512 count, 513 (count + 1) | NC_SHLOCK_FLAG | 514 NC_SHLOCK_VHOLD)) { 515 /* 516 * The vp associated with a locked ncp must 517 * be held to prevent it from being recycled. 518 * 519 * WARNING! If VRECLAIMED is set the vnode 520 * could already be in the middle of a recycle. 521 * Callers must use cache_vref() or 522 * cache_vget() on the locked ncp to 523 * validate the vp or set the cache entry 524 * to unresolved. 525 * 526 * NOTE! vhold() is allowed if we hold a 527 * lock on the ncp (which we do). 528 */ 529 if (ncp->nc_vp) 530 vhold(ncp->nc_vp); 531 atomic_clear_int(&ncp->nc_lockstatus, 532 NC_SHLOCK_VHOLD); 533 crit_exit(); 534 break; 535 } 536 /* cmpset failed */ 537 crit_exit(); 538 continue; 539 } 540 541 /* 542 * If already held shared we can just bump the count, but 543 * only allow this if nobody is trying to get the lock 544 * exclusively. If we are blocking too long ignore excl 545 * requests (which can race/deadlock us). 546 * 547 * VHOLD is a bit of a hack. Even though we successfully 548 * added another shared ref, the cpu that got the first 549 * shared ref might not yet have held the vnode. 550 */ 551 if ((count & (optreq|NC_SHLOCK_FLAG)) == NC_SHLOCK_FLAG) { 552 KKASSERT((count & ~(NC_EXLOCK_REQ | 553 NC_SHLOCK_REQ | 554 NC_SHLOCK_FLAG)) > 0); 555 if (atomic_cmpset_int(&ncp->nc_lockstatus, 556 count, count + 1)) { 557 while (ncp->nc_lockstatus & NC_SHLOCK_VHOLD) 558 cpu_pause(); 559 break; 560 } 561 continue; 562 } 563 tsleep_interlock(ncp, 0); 564 if (atomic_cmpset_int(&ncp->nc_lockstatus, count, 565 count | NC_SHLOCK_REQ) == 0) { 566 /* cmpset failed */ 567 continue; 568 } 569 error = tsleep(ncp, PINTERLOCKED, "clocksh", nclockwarn); 570 if (error == EWOULDBLOCK) { 571 optreq = 0; 572 if (didwarn == 0) { 573 didwarn = ticks - nclockwarn; 574 kprintf("[diagnostic] cache_lock_shared: " 575 "%s blocked on %p %08x", 576 curthread->td_comm, ncp, count); 577 kprintf(" \"%*.*s\"\n", 578 ncp->nc_nlen, ncp->nc_nlen, 579 ncp->nc_name); 580 } 581 } 582 /* loop */ 583 } 584 if (didwarn) { 585 kprintf("[diagnostic] cache_lock_shared: " 586 "%s unblocked %*.*s after %d secs\n", 587 curthread->td_comm, 588 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name, 589 (int)(ticks - didwarn) / hz); 590 } 591 } 592 593 /* 594 * Lock ncp exclusively, return 0 on success. 595 * 596 * NOTE: nc_refs may be zero if the ncp is interlocked by circumstance, 597 * such as the case where one of its children is locked. 598 */ 599 static 600 int 601 _cache_lock_nonblock(struct namecache *ncp) 602 { 603 thread_t td; 604 u_int count; 605 606 td = curthread; 607 608 for (;;) { 609 count = ncp->nc_lockstatus; 610 611 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 0) { 612 if (atomic_cmpset_int(&ncp->nc_lockstatus, 613 count, count + 1)) { 614 /* 615 * The vp associated with a locked ncp must 616 * be held to prevent it from being recycled. 617 * 618 * WARNING! If VRECLAIMED is set the vnode 619 * could already be in the middle of a recycle. 620 * Callers must use cache_vref() or 621 * cache_vget() on the locked ncp to 622 * validate the vp or set the cache entry 623 * to unresolved. 624 * 625 * NOTE! vhold() is allowed if we hold a 626 * lock on the ncp (which we do). 627 */ 628 ncp->nc_locktd = td; 629 if (ncp->nc_vp) 630 vhold(ncp->nc_vp); 631 break; 632 } 633 /* cmpset failed */ 634 continue; 635 } 636 if (ncp->nc_locktd == td) { 637 if (atomic_cmpset_int(&ncp->nc_lockstatus, 638 count, count + 1)) { 639 break; 640 } 641 /* cmpset failed */ 642 continue; 643 } 644 return(EWOULDBLOCK); 645 } 646 return(0); 647 } 648 649 /* 650 * The shared lock works similarly to the exclusive lock except 651 * nc_locktd is left NULL and we need an interlock (VHOLD) to 652 * prevent vhold() races, since the moment our cmpset_int succeeds 653 * another cpu can come in and get its own shared lock. 654 * 655 * A critical section is needed to prevent interruption during the 656 * VHOLD interlock. 657 */ 658 static 659 int 660 _cache_lock_shared_nonblock(struct namecache *ncp) 661 { 662 u_int count; 663 664 for (;;) { 665 count = ncp->nc_lockstatus; 666 667 if ((count & ~NC_SHLOCK_REQ) == 0) { 668 crit_enter(); 669 if (atomic_cmpset_int(&ncp->nc_lockstatus, 670 count, 671 (count + 1) | NC_SHLOCK_FLAG | 672 NC_SHLOCK_VHOLD)) { 673 /* 674 * The vp associated with a locked ncp must 675 * be held to prevent it from being recycled. 676 * 677 * WARNING! If VRECLAIMED is set the vnode 678 * could already be in the middle of a recycle. 679 * Callers must use cache_vref() or 680 * cache_vget() on the locked ncp to 681 * validate the vp or set the cache entry 682 * to unresolved. 683 * 684 * NOTE! vhold() is allowed if we hold a 685 * lock on the ncp (which we do). 686 */ 687 if (ncp->nc_vp) 688 vhold(ncp->nc_vp); 689 atomic_clear_int(&ncp->nc_lockstatus, 690 NC_SHLOCK_VHOLD); 691 crit_exit(); 692 break; 693 } 694 /* cmpset failed */ 695 crit_exit(); 696 continue; 697 } 698 699 /* 700 * If already held shared we can just bump the count, but 701 * only allow this if nobody is trying to get the lock 702 * exclusively. 703 * 704 * VHOLD is a bit of a hack. Even though we successfully 705 * added another shared ref, the cpu that got the first 706 * shared ref might not yet have held the vnode. 707 */ 708 if ((count & (NC_EXLOCK_REQ|NC_SHLOCK_FLAG)) == 709 NC_SHLOCK_FLAG) { 710 KKASSERT((count & ~(NC_EXLOCK_REQ | 711 NC_SHLOCK_REQ | 712 NC_SHLOCK_FLAG)) > 0); 713 if (atomic_cmpset_int(&ncp->nc_lockstatus, 714 count, count + 1)) { 715 while (ncp->nc_lockstatus & NC_SHLOCK_VHOLD) 716 cpu_pause(); 717 break; 718 } 719 continue; 720 } 721 return(EWOULDBLOCK); 722 } 723 return(0); 724 } 725 726 /* 727 * Helper function 728 * 729 * NOTE: nc_refs can be 0 (degenerate case during _cache_drop). 730 * 731 * nc_locktd must be NULLed out prior to nc_lockstatus getting cleared. 732 */ 733 static 734 void 735 _cache_unlock(struct namecache *ncp) 736 { 737 thread_t td __debugvar = curthread; 738 u_int count; 739 u_int ncount; 740 struct vnode *dropvp; 741 742 KKASSERT(ncp->nc_refs >= 0); 743 KKASSERT((ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) > 0); 744 KKASSERT((ncp->nc_lockstatus & NC_SHLOCK_FLAG) || ncp->nc_locktd == td); 745 746 count = ncp->nc_lockstatus; 747 cpu_ccfence(); 748 749 /* 750 * Clear nc_locktd prior to the atomic op (excl lock only) 751 */ 752 if ((count & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 1) 753 ncp->nc_locktd = NULL; 754 dropvp = NULL; 755 756 for (;;) { 757 if ((count & 758 ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ|NC_SHLOCK_FLAG)) == 1) { 759 dropvp = ncp->nc_vp; 760 if (count & NC_EXLOCK_REQ) 761 ncount = count & NC_SHLOCK_REQ; /* cnt->0 */ 762 else 763 ncount = 0; 764 765 if (atomic_cmpset_int(&ncp->nc_lockstatus, 766 count, ncount)) { 767 if (count & NC_EXLOCK_REQ) 768 wakeup(&ncp->nc_locktd); 769 else if (count & NC_SHLOCK_REQ) 770 wakeup(ncp); 771 break; 772 } 773 dropvp = NULL; 774 } else { 775 KKASSERT((count & NC_SHLOCK_VHOLD) == 0); 776 KKASSERT((count & ~(NC_EXLOCK_REQ | 777 NC_SHLOCK_REQ | 778 NC_SHLOCK_FLAG)) > 1); 779 if (atomic_cmpset_int(&ncp->nc_lockstatus, 780 count, count - 1)) { 781 break; 782 } 783 } 784 count = ncp->nc_lockstatus; 785 cpu_ccfence(); 786 } 787 788 /* 789 * Don't actually drop the vp until we successfully clean out 790 * the lock, otherwise we may race another shared lock. 791 */ 792 if (dropvp) 793 vdrop(dropvp); 794 } 795 796 static 797 int 798 _cache_lockstatus(struct namecache *ncp) 799 { 800 if (ncp->nc_locktd == curthread) 801 return(LK_EXCLUSIVE); 802 if (ncp->nc_lockstatus & NC_SHLOCK_FLAG) 803 return(LK_SHARED); 804 return(-1); 805 } 806 807 /* 808 * cache_hold() and cache_drop() prevent the premature deletion of a 809 * namecache entry but do not prevent operations (such as zapping) on 810 * that namecache entry. 811 * 812 * This routine may only be called from outside this source module if 813 * nc_refs is already at least 1. 814 * 815 * This is a rare case where callers are allowed to hold a spinlock, 816 * so we can't ourselves. 817 */ 818 static __inline 819 struct namecache * 820 _cache_hold(struct namecache *ncp) 821 { 822 atomic_add_int(&ncp->nc_refs, 1); 823 return(ncp); 824 } 825 826 /* 827 * Drop a cache entry, taking care to deal with races. 828 * 829 * For potential 1->0 transitions we must hold the ncp lock to safely 830 * test its flags. An unresolved entry with no children must be zapped 831 * to avoid leaks. 832 * 833 * The call to cache_zap() itself will handle all remaining races and 834 * will decrement the ncp's refs regardless. If we are resolved or 835 * have children nc_refs can safely be dropped to 0 without having to 836 * zap the entry. 837 * 838 * NOTE: cache_zap() will re-check nc_refs and nc_list in a MPSAFE fashion. 839 * 840 * NOTE: cache_zap() may return a non-NULL referenced parent which must 841 * be dropped in a loop. 842 */ 843 static __inline 844 void 845 _cache_drop(struct namecache *ncp) 846 { 847 int refs; 848 849 while (ncp) { 850 KKASSERT(ncp->nc_refs > 0); 851 refs = ncp->nc_refs; 852 853 if (refs == 1) { 854 if (_cache_lock_nonblock(ncp) == 0) { 855 ncp->nc_flag &= ~NCF_DEFEREDZAP; 856 if ((ncp->nc_flag & NCF_UNRESOLVED) && 857 TAILQ_EMPTY(&ncp->nc_list)) { 858 ncp = cache_zap(ncp, 1); 859 continue; 860 } 861 if (atomic_cmpset_int(&ncp->nc_refs, 1, 0)) { 862 _cache_unlock(ncp); 863 break; 864 } 865 _cache_unlock(ncp); 866 } 867 } else { 868 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) 869 break; 870 } 871 cpu_pause(); 872 } 873 } 874 875 /* 876 * Link a new namecache entry to its parent and to the hash table. Be 877 * careful to avoid races if vhold() blocks in the future. 878 * 879 * Both ncp and par must be referenced and locked. 880 * 881 * NOTE: The hash table spinlock is held during this call, we can't do 882 * anything fancy. 883 */ 884 static void 885 _cache_link_parent(struct namecache *ncp, struct namecache *par, 886 struct nchash_head *nchpp) 887 { 888 KKASSERT(ncp->nc_parent == NULL); 889 ncp->nc_parent = par; 890 ncp->nc_head = nchpp; 891 892 /* 893 * Set inheritance flags. Note that the parent flags may be 894 * stale due to getattr potentially not having been run yet 895 * (it gets run during nlookup()'s). 896 */ 897 ncp->nc_flag &= ~(NCF_SF_PNOCACHE | NCF_UF_PCACHE); 898 if (par->nc_flag & (NCF_SF_NOCACHE | NCF_SF_PNOCACHE)) 899 ncp->nc_flag |= NCF_SF_PNOCACHE; 900 if (par->nc_flag & (NCF_UF_CACHE | NCF_UF_PCACHE)) 901 ncp->nc_flag |= NCF_UF_PCACHE; 902 903 LIST_INSERT_HEAD(&nchpp->list, ncp, nc_hash); 904 905 if (TAILQ_EMPTY(&par->nc_list)) { 906 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry); 907 /* 908 * Any vp associated with an ncp which has children must 909 * be held to prevent it from being recycled. 910 */ 911 if (par->nc_vp) 912 vhold(par->nc_vp); 913 } else { 914 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry); 915 } 916 } 917 918 /* 919 * Remove the parent and hash associations from a namecache structure. 920 * If this is the last child of the parent the cache_drop(par) will 921 * attempt to recursively zap the parent. 922 * 923 * ncp must be locked. This routine will acquire a temporary lock on 924 * the parent as wlel as the appropriate hash chain. 925 */ 926 static void 927 _cache_unlink_parent(struct namecache *ncp) 928 { 929 struct namecache *par; 930 struct vnode *dropvp; 931 932 if ((par = ncp->nc_parent) != NULL) { 933 KKASSERT(ncp->nc_parent == par); 934 _cache_hold(par); 935 _cache_lock(par); 936 spin_lock(&ncp->nc_head->spin); 937 LIST_REMOVE(ncp, nc_hash); 938 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 939 dropvp = NULL; 940 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 941 dropvp = par->nc_vp; 942 spin_unlock(&ncp->nc_head->spin); 943 ncp->nc_parent = NULL; 944 ncp->nc_head = NULL; 945 _cache_unlock(par); 946 _cache_drop(par); 947 948 /* 949 * We can only safely vdrop with no spinlocks held. 950 */ 951 if (dropvp) 952 vdrop(dropvp); 953 } 954 } 955 956 /* 957 * Allocate a new namecache structure. Most of the code does not require 958 * zero-termination of the string but it makes vop_compat_ncreate() easier. 959 */ 960 static struct namecache * 961 cache_alloc(int nlen) 962 { 963 struct namecache *ncp; 964 965 ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO); 966 if (nlen) 967 ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK); 968 ncp->nc_nlen = nlen; 969 ncp->nc_flag = NCF_UNRESOLVED; 970 ncp->nc_error = ENOTCONN; /* needs to be resolved */ 971 ncp->nc_refs = 1; 972 973 TAILQ_INIT(&ncp->nc_list); 974 _cache_lock(ncp); 975 return(ncp); 976 } 977 978 /* 979 * Can only be called for the case where the ncp has never been 980 * associated with anything (so no spinlocks are needed). 981 */ 982 static void 983 _cache_free(struct namecache *ncp) 984 { 985 KKASSERT(ncp->nc_refs == 1 && ncp->nc_lockstatus == 1); 986 if (ncp->nc_name) 987 kfree(ncp->nc_name, M_VFSCACHE); 988 kfree(ncp, M_VFSCACHE); 989 } 990 991 /* 992 * [re]initialize a nchandle. 993 */ 994 void 995 cache_zero(struct nchandle *nch) 996 { 997 nch->ncp = NULL; 998 nch->mount = NULL; 999 } 1000 1001 /* 1002 * Ref and deref a namecache structure. 1003 * 1004 * The caller must specify a stable ncp pointer, typically meaning the 1005 * ncp is already referenced but this can also occur indirectly through 1006 * e.g. holding a lock on a direct child. 1007 * 1008 * WARNING: Caller may hold an unrelated read spinlock, which means we can't 1009 * use read spinlocks here. 1010 */ 1011 struct nchandle * 1012 cache_hold(struct nchandle *nch) 1013 { 1014 _cache_hold(nch->ncp); 1015 _cache_mntref(nch->mount); 1016 return(nch); 1017 } 1018 1019 /* 1020 * Create a copy of a namecache handle for an already-referenced 1021 * entry. 1022 */ 1023 void 1024 cache_copy(struct nchandle *nch, struct nchandle *target) 1025 { 1026 struct mntcache *cache = &pcpu_mntcache[mycpu->gd_cpuid]; 1027 struct namecache *ncp; 1028 1029 *target = *nch; 1030 _cache_mntref(target->mount); 1031 ncp = target->ncp; 1032 if (ncp) { 1033 if (ncp == cache->ncp1) { 1034 if (atomic_cmpset_ptr((void *)&cache->ncp1, ncp, NULL)) 1035 return; 1036 } 1037 if (ncp == cache->ncp2) { 1038 if (atomic_cmpset_ptr((void *)&cache->ncp2, ncp, NULL)) 1039 return; 1040 } 1041 _cache_hold(ncp); 1042 } 1043 } 1044 1045 /* 1046 * Caller wants to copy the current directory, copy it out from our 1047 * pcpu cache if possible (the entire critical path is just two localized 1048 * cmpset ops). If the pcpu cache has a snapshot at all it will be a 1049 * valid one, so we don't have to lock p->p_fd even though we are loading 1050 * two fields. 1051 * 1052 * This has a limited effect since nlookup must still ref and shlock the 1053 * vnode to check perms. We do avoid the per-proc spin-lock though, which 1054 * can aid threaded programs. 1055 */ 1056 void 1057 cache_copy_ncdir(struct proc *p, struct nchandle *target) 1058 { 1059 struct mntcache *cache = &pcpu_mntcache[mycpu->gd_cpuid]; 1060 1061 *target = p->p_fd->fd_ncdir; 1062 if (target->ncp == cache->ncdir.ncp && 1063 target->mount == cache->ncdir.mount) { 1064 if (atomic_cmpset_ptr((void *)&cache->ncdir.ncp, 1065 target->ncp, NULL)) { 1066 if (atomic_cmpset_ptr((void *)&cache->ncdir.mount, 1067 target->mount, NULL)) { 1068 /* CRITICAL PATH */ 1069 return; 1070 } 1071 _cache_drop(target->ncp); 1072 } 1073 } 1074 spin_lock_shared(&p->p_fd->fd_spin); 1075 cache_copy(&p->p_fd->fd_ncdir, target); 1076 spin_unlock_shared(&p->p_fd->fd_spin); 1077 } 1078 1079 void 1080 cache_changemount(struct nchandle *nch, struct mount *mp) 1081 { 1082 _cache_mntref(mp); 1083 _cache_mntrel(nch->mount); 1084 nch->mount = mp; 1085 } 1086 1087 void 1088 cache_drop(struct nchandle *nch) 1089 { 1090 _cache_mntrel(nch->mount); 1091 _cache_drop(nch->ncp); 1092 nch->ncp = NULL; 1093 nch->mount = NULL; 1094 } 1095 1096 /* 1097 * Drop the nchandle, but try to cache the ref to avoid global atomic 1098 * ops. This is typically done on the system root and jail root nchandles. 1099 */ 1100 void 1101 cache_drop_and_cache(struct nchandle *nch) 1102 { 1103 struct mntcache *cache = &pcpu_mntcache[mycpu->gd_cpuid]; 1104 struct namecache *ncp; 1105 1106 _cache_mntrel(nch->mount); 1107 ncp = nch->ncp; 1108 if (cache->ncp1 == NULL) { 1109 ncp = atomic_swap_ptr((void *)&cache->ncp1, ncp); 1110 if (ncp == NULL) 1111 goto done; 1112 } 1113 if (cache->ncp2 == NULL) { 1114 ncp = atomic_swap_ptr((void *)&cache->ncp2, ncp); 1115 if (ncp == NULL) 1116 goto done; 1117 } 1118 if (++cache->iter & 1) 1119 ncp = atomic_swap_ptr((void *)&cache->ncp2, ncp); 1120 else 1121 ncp = atomic_swap_ptr((void *)&cache->ncp1, ncp); 1122 if (ncp) 1123 _cache_drop(ncp); 1124 done: 1125 nch->ncp = NULL; 1126 nch->mount = NULL; 1127 } 1128 1129 /* 1130 * We are dropping what the caller believes is the current directory, 1131 * unconditionally store it in our pcpu cache. Anything already in 1132 * the cache will be discarded. 1133 */ 1134 void 1135 cache_drop_ncdir(struct nchandle *nch) 1136 { 1137 struct mntcache *cache = &pcpu_mntcache[mycpu->gd_cpuid]; 1138 1139 nch->ncp = atomic_swap_ptr((void *)&cache->ncdir.ncp, nch->ncp); 1140 nch->mount = atomic_swap_ptr((void *)&cache->ncdir.mount, nch->mount); 1141 if (nch->ncp) 1142 _cache_drop(nch->ncp); 1143 if (nch->mount) 1144 _cache_mntrel(nch->mount); 1145 nch->ncp = NULL; 1146 nch->mount = NULL; 1147 } 1148 1149 int 1150 cache_lockstatus(struct nchandle *nch) 1151 { 1152 return(_cache_lockstatus(nch->ncp)); 1153 } 1154 1155 void 1156 cache_lock(struct nchandle *nch) 1157 { 1158 _cache_lock(nch->ncp); 1159 } 1160 1161 void 1162 cache_lock_maybe_shared(struct nchandle *nch, int excl) 1163 { 1164 struct namecache *ncp = nch->ncp; 1165 1166 if (ncp_shared_lock_disable || excl || 1167 (ncp->nc_flag & NCF_UNRESOLVED)) { 1168 _cache_lock(ncp); 1169 } else { 1170 _cache_lock_shared(ncp); 1171 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 1172 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) { 1173 _cache_unlock(ncp); 1174 _cache_lock(ncp); 1175 } 1176 } else { 1177 _cache_unlock(ncp); 1178 _cache_lock(ncp); 1179 } 1180 } 1181 } 1182 1183 /* 1184 * Relock nch1 given an unlocked nch1 and a locked nch2. The caller 1185 * is responsible for checking both for validity on return as they 1186 * may have become invalid. 1187 * 1188 * We have to deal with potential deadlocks here, just ping pong 1189 * the lock until we get it (we will always block somewhere when 1190 * looping so this is not cpu-intensive). 1191 * 1192 * which = 0 nch1 not locked, nch2 is locked 1193 * which = 1 nch1 is locked, nch2 is not locked 1194 */ 1195 void 1196 cache_relock(struct nchandle *nch1, struct ucred *cred1, 1197 struct nchandle *nch2, struct ucred *cred2) 1198 { 1199 int which; 1200 1201 which = 0; 1202 1203 for (;;) { 1204 if (which == 0) { 1205 if (cache_lock_nonblock(nch1) == 0) { 1206 cache_resolve(nch1, cred1); 1207 break; 1208 } 1209 cache_unlock(nch2); 1210 cache_lock(nch1); 1211 cache_resolve(nch1, cred1); 1212 which = 1; 1213 } else { 1214 if (cache_lock_nonblock(nch2) == 0) { 1215 cache_resolve(nch2, cred2); 1216 break; 1217 } 1218 cache_unlock(nch1); 1219 cache_lock(nch2); 1220 cache_resolve(nch2, cred2); 1221 which = 0; 1222 } 1223 } 1224 } 1225 1226 int 1227 cache_lock_nonblock(struct nchandle *nch) 1228 { 1229 return(_cache_lock_nonblock(nch->ncp)); 1230 } 1231 1232 void 1233 cache_unlock(struct nchandle *nch) 1234 { 1235 _cache_unlock(nch->ncp); 1236 } 1237 1238 /* 1239 * ref-and-lock, unlock-and-deref functions. 1240 * 1241 * This function is primarily used by nlookup. Even though cache_lock 1242 * holds the vnode, it is possible that the vnode may have already 1243 * initiated a recyclement. 1244 * 1245 * We want cache_get() to return a definitively usable vnode or a 1246 * definitively unresolved ncp. 1247 */ 1248 static 1249 struct namecache * 1250 _cache_get(struct namecache *ncp) 1251 { 1252 _cache_hold(ncp); 1253 _cache_lock(ncp); 1254 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 1255 _cache_setunresolved(ncp); 1256 return(ncp); 1257 } 1258 1259 /* 1260 * Attempt to obtain a shared lock on the ncp. A shared lock will only 1261 * be obtained if the ncp is resolved and the vnode (if not ENOENT) is 1262 * valid. Otherwise an exclusive lock will be acquired instead. 1263 */ 1264 static 1265 struct namecache * 1266 _cache_get_maybe_shared(struct namecache *ncp, int excl) 1267 { 1268 if (ncp_shared_lock_disable || excl || 1269 (ncp->nc_flag & NCF_UNRESOLVED)) { 1270 return(_cache_get(ncp)); 1271 } 1272 _cache_hold(ncp); 1273 _cache_lock_shared(ncp); 1274 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 1275 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) { 1276 _cache_unlock(ncp); 1277 ncp = _cache_get(ncp); 1278 _cache_drop(ncp); 1279 } 1280 } else { 1281 _cache_unlock(ncp); 1282 ncp = _cache_get(ncp); 1283 _cache_drop(ncp); 1284 } 1285 return(ncp); 1286 } 1287 1288 /* 1289 * This is a special form of _cache_lock() which only succeeds if 1290 * it can get a pristine, non-recursive lock. The caller must have 1291 * already ref'd the ncp. 1292 * 1293 * On success the ncp will be locked, on failure it will not. The 1294 * ref count does not change either way. 1295 * 1296 * We want _cache_lock_special() (on success) to return a definitively 1297 * usable vnode or a definitively unresolved ncp. 1298 */ 1299 static int 1300 _cache_lock_special(struct namecache *ncp) 1301 { 1302 if (_cache_lock_nonblock(ncp) == 0) { 1303 if ((ncp->nc_lockstatus & 1304 ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) == 1) { 1305 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 1306 _cache_setunresolved(ncp); 1307 return(0); 1308 } 1309 _cache_unlock(ncp); 1310 } 1311 return(EWOULDBLOCK); 1312 } 1313 1314 /* 1315 * This function tries to get a shared lock but will back-off to an exclusive 1316 * lock if: 1317 * 1318 * (1) Some other thread is trying to obtain an exclusive lock 1319 * (to prevent the exclusive requester from getting livelocked out 1320 * by many shared locks). 1321 * 1322 * (2) The current thread already owns an exclusive lock (to avoid 1323 * deadlocking). 1324 * 1325 * WARNING! On machines with lots of cores we really want to try hard to 1326 * get a shared lock or concurrent path lookups can chain-react 1327 * into a very high-latency exclusive lock. 1328 */ 1329 static int 1330 _cache_lock_shared_special(struct namecache *ncp) 1331 { 1332 /* 1333 * Only honor a successful shared lock (returning 0) if there is 1334 * no exclusive request pending and the vnode, if present, is not 1335 * in a reclaimed state. 1336 */ 1337 if (_cache_lock_shared_nonblock(ncp) == 0) { 1338 if ((ncp->nc_lockstatus & NC_EXLOCK_REQ) == 0) { 1339 if (ncp->nc_vp == NULL || 1340 (ncp->nc_vp->v_flag & VRECLAIMED) == 0) { 1341 return(0); 1342 } 1343 } 1344 _cache_unlock(ncp); 1345 return(EWOULDBLOCK); 1346 } 1347 1348 /* 1349 * Non-blocking shared lock failed. If we already own the exclusive 1350 * lock just acquire another exclusive lock (instead of deadlocking). 1351 * Otherwise acquire a shared lock. 1352 */ 1353 if (ncp->nc_locktd == curthread) { 1354 _cache_lock(ncp); 1355 return(0); 1356 } 1357 _cache_lock_shared(ncp); 1358 return(0); 1359 } 1360 1361 1362 /* 1363 * NOTE: The same nchandle can be passed for both arguments. 1364 */ 1365 void 1366 cache_get(struct nchandle *nch, struct nchandle *target) 1367 { 1368 KKASSERT(nch->ncp->nc_refs > 0); 1369 target->mount = nch->mount; 1370 target->ncp = _cache_get(nch->ncp); 1371 _cache_mntref(target->mount); 1372 } 1373 1374 void 1375 cache_get_maybe_shared(struct nchandle *nch, struct nchandle *target, int excl) 1376 { 1377 KKASSERT(nch->ncp->nc_refs > 0); 1378 target->mount = nch->mount; 1379 target->ncp = _cache_get_maybe_shared(nch->ncp, excl); 1380 _cache_mntref(target->mount); 1381 } 1382 1383 /* 1384 * 1385 */ 1386 static __inline 1387 void 1388 _cache_put(struct namecache *ncp) 1389 { 1390 _cache_unlock(ncp); 1391 _cache_drop(ncp); 1392 } 1393 1394 /* 1395 * 1396 */ 1397 void 1398 cache_put(struct nchandle *nch) 1399 { 1400 _cache_mntrel(nch->mount); 1401 _cache_put(nch->ncp); 1402 nch->ncp = NULL; 1403 nch->mount = NULL; 1404 } 1405 1406 /* 1407 * Resolve an unresolved ncp by associating a vnode with it. If the 1408 * vnode is NULL, a negative cache entry is created. 1409 * 1410 * The ncp should be locked on entry and will remain locked on return. 1411 */ 1412 static 1413 void 1414 _cache_setvp(struct mount *mp, struct namecache *ncp, struct vnode *vp) 1415 { 1416 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED); 1417 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE); 1418 1419 if (vp != NULL) { 1420 /* 1421 * Any vp associated with an ncp which has children must 1422 * be held. Any vp associated with a locked ncp must be held. 1423 */ 1424 if (!TAILQ_EMPTY(&ncp->nc_list)) 1425 vhold(vp); 1426 spin_lock(&vp->v_spin); 1427 ncp->nc_vp = vp; 1428 TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode); 1429 spin_unlock(&vp->v_spin); 1430 if (ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) 1431 vhold(vp); 1432 1433 /* 1434 * Set auxiliary flags 1435 */ 1436 switch(vp->v_type) { 1437 case VDIR: 1438 ncp->nc_flag |= NCF_ISDIR; 1439 break; 1440 case VLNK: 1441 ncp->nc_flag |= NCF_ISSYMLINK; 1442 /* XXX cache the contents of the symlink */ 1443 break; 1444 default: 1445 break; 1446 } 1447 atomic_add_int(&numcache, 1); 1448 ncp->nc_error = 0; 1449 /* XXX: this is a hack to work-around the lack of a real pfs vfs 1450 * implementation*/ 1451 if (mp != NULL) 1452 if (strncmp(mp->mnt_stat.f_fstypename, "null", 5) == 0) 1453 vp->v_pfsmp = mp; 1454 } else { 1455 /* 1456 * When creating a negative cache hit we set the 1457 * namecache_gen. A later resolve will clean out the 1458 * negative cache hit if the mount point's namecache_gen 1459 * has changed. Used by devfs, could also be used by 1460 * other remote FSs. 1461 */ 1462 ncp->nc_vp = NULL; 1463 spin_lock(&ncspin); 1464 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 1465 ++numneg; 1466 spin_unlock(&ncspin); 1467 ncp->nc_error = ENOENT; 1468 if (mp) 1469 VFS_NCPGEN_SET(mp, ncp); 1470 } 1471 ncp->nc_flag &= ~(NCF_UNRESOLVED | NCF_DEFEREDZAP); 1472 } 1473 1474 /* 1475 * 1476 */ 1477 void 1478 cache_setvp(struct nchandle *nch, struct vnode *vp) 1479 { 1480 _cache_setvp(nch->mount, nch->ncp, vp); 1481 } 1482 1483 /* 1484 * 1485 */ 1486 void 1487 cache_settimeout(struct nchandle *nch, int nticks) 1488 { 1489 struct namecache *ncp = nch->ncp; 1490 1491 if ((ncp->nc_timeout = ticks + nticks) == 0) 1492 ncp->nc_timeout = 1; 1493 } 1494 1495 /* 1496 * Disassociate the vnode or negative-cache association and mark a 1497 * namecache entry as unresolved again. Note that the ncp is still 1498 * left in the hash table and still linked to its parent. 1499 * 1500 * The ncp should be locked and refd on entry and will remain locked and refd 1501 * on return. 1502 * 1503 * This routine is normally never called on a directory containing children. 1504 * However, NFS often does just that in its rename() code as a cop-out to 1505 * avoid complex namespace operations. This disconnects a directory vnode 1506 * from its namecache and can cause the OLDAPI and NEWAPI to get out of 1507 * sync. 1508 * 1509 */ 1510 static 1511 void 1512 _cache_setunresolved(struct namecache *ncp) 1513 { 1514 struct vnode *vp; 1515 1516 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 1517 ncp->nc_flag |= NCF_UNRESOLVED; 1518 ncp->nc_timeout = 0; 1519 ncp->nc_error = ENOTCONN; 1520 if ((vp = ncp->nc_vp) != NULL) { 1521 atomic_add_int(&numcache, -1); 1522 spin_lock(&vp->v_spin); 1523 ncp->nc_vp = NULL; 1524 TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode); 1525 spin_unlock(&vp->v_spin); 1526 1527 /* 1528 * Any vp associated with an ncp with children is 1529 * held by that ncp. Any vp associated with a locked 1530 * ncp is held by that ncp. These conditions must be 1531 * undone when the vp is cleared out from the ncp. 1532 */ 1533 if (!TAILQ_EMPTY(&ncp->nc_list)) 1534 vdrop(vp); 1535 if (ncp->nc_lockstatus & ~(NC_EXLOCK_REQ|NC_SHLOCK_REQ)) 1536 vdrop(vp); 1537 } else { 1538 spin_lock(&ncspin); 1539 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 1540 --numneg; 1541 spin_unlock(&ncspin); 1542 } 1543 ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK); 1544 } 1545 } 1546 1547 /* 1548 * The cache_nresolve() code calls this function to automatically 1549 * set a resolved cache element to unresolved if it has timed out 1550 * or if it is a negative cache hit and the mount point namecache_gen 1551 * has changed. 1552 */ 1553 static __inline int 1554 _cache_auto_unresolve_test(struct mount *mp, struct namecache *ncp) 1555 { 1556 /* 1557 * Try to zap entries that have timed out. We have 1558 * to be careful here because locked leafs may depend 1559 * on the vnode remaining intact in a parent, so only 1560 * do this under very specific conditions. 1561 */ 1562 if (ncp->nc_timeout && (int)(ncp->nc_timeout - ticks) < 0 && 1563 TAILQ_EMPTY(&ncp->nc_list)) { 1564 return 1; 1565 } 1566 1567 /* 1568 * If a resolved negative cache hit is invalid due to 1569 * the mount's namecache generation being bumped, zap it. 1570 */ 1571 if (ncp->nc_vp == NULL && VFS_NCPGEN_TEST(mp, ncp)) { 1572 return 1; 1573 } 1574 1575 /* 1576 * Otherwise we are good 1577 */ 1578 return 0; 1579 } 1580 1581 static __inline void 1582 _cache_auto_unresolve(struct mount *mp, struct namecache *ncp) 1583 { 1584 /* 1585 * Already in an unresolved state, nothing to do. 1586 */ 1587 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 1588 if (_cache_auto_unresolve_test(mp, ncp)) 1589 _cache_setunresolved(ncp); 1590 } 1591 } 1592 1593 /* 1594 * 1595 */ 1596 void 1597 cache_setunresolved(struct nchandle *nch) 1598 { 1599 _cache_setunresolved(nch->ncp); 1600 } 1601 1602 /* 1603 * Determine if we can clear NCF_ISMOUNTPT by scanning the mountlist 1604 * looking for matches. This flag tells the lookup code when it must 1605 * check for a mount linkage and also prevents the directories in question 1606 * from being deleted or renamed. 1607 */ 1608 static 1609 int 1610 cache_clrmountpt_callback(struct mount *mp, void *data) 1611 { 1612 struct nchandle *nch = data; 1613 1614 if (mp->mnt_ncmounton.ncp == nch->ncp) 1615 return(1); 1616 if (mp->mnt_ncmountpt.ncp == nch->ncp) 1617 return(1); 1618 return(0); 1619 } 1620 1621 /* 1622 * 1623 */ 1624 void 1625 cache_clrmountpt(struct nchandle *nch) 1626 { 1627 int count; 1628 1629 count = mountlist_scan(cache_clrmountpt_callback, nch, 1630 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 1631 if (count == 0) 1632 nch->ncp->nc_flag &= ~NCF_ISMOUNTPT; 1633 } 1634 1635 /* 1636 * Invalidate portions of the namecache topology given a starting entry. 1637 * The passed ncp is set to an unresolved state and: 1638 * 1639 * The passed ncp must be referencxed and locked. The routine may unlock 1640 * and relock ncp several times, and will recheck the children and loop 1641 * to catch races. When done the passed ncp will be returned with the 1642 * reference and lock intact. 1643 * 1644 * CINV_DESTROY - Set a flag in the passed ncp entry indicating 1645 * that the physical underlying nodes have been 1646 * destroyed... as in deleted. For example, when 1647 * a directory is removed. This will cause record 1648 * lookups on the name to no longer be able to find 1649 * the record and tells the resolver to return failure 1650 * rather then trying to resolve through the parent. 1651 * 1652 * The topology itself, including ncp->nc_name, 1653 * remains intact. 1654 * 1655 * This only applies to the passed ncp, if CINV_CHILDREN 1656 * is specified the children are not flagged. 1657 * 1658 * CINV_CHILDREN - Set all children (recursively) to an unresolved 1659 * state as well. 1660 * 1661 * Note that this will also have the side effect of 1662 * cleaning out any unreferenced nodes in the topology 1663 * from the leaves up as the recursion backs out. 1664 * 1665 * Note that the topology for any referenced nodes remains intact, but 1666 * the nodes will be marked as having been destroyed and will be set 1667 * to an unresolved state. 1668 * 1669 * It is possible for cache_inval() to race a cache_resolve(), meaning that 1670 * the namecache entry may not actually be invalidated on return if it was 1671 * revalidated while recursing down into its children. This code guarentees 1672 * that the node(s) will go through an invalidation cycle, but does not 1673 * guarentee that they will remain in an invalidated state. 1674 * 1675 * Returns non-zero if a revalidation was detected during the invalidation 1676 * recursion, zero otherwise. Note that since only the original ncp is 1677 * locked the revalidation ultimately can only indicate that the original ncp 1678 * *MIGHT* no have been reresolved. 1679 * 1680 * DEEP RECURSION HANDLING - If a recursive invalidation recurses deeply we 1681 * have to avoid blowing out the kernel stack. We do this by saving the 1682 * deep namecache node and aborting the recursion, then re-recursing at that 1683 * node using a depth-first algorithm in order to allow multiple deep 1684 * recursions to chain through each other, then we restart the invalidation 1685 * from scratch. 1686 */ 1687 1688 struct cinvtrack { 1689 struct namecache *resume_ncp; 1690 int depth; 1691 }; 1692 1693 static int _cache_inval_internal(struct namecache *, int, struct cinvtrack *); 1694 1695 static 1696 int 1697 _cache_inval(struct namecache *ncp, int flags) 1698 { 1699 struct cinvtrack track; 1700 struct namecache *ncp2; 1701 int r; 1702 1703 track.depth = 0; 1704 track.resume_ncp = NULL; 1705 1706 for (;;) { 1707 r = _cache_inval_internal(ncp, flags, &track); 1708 if (track.resume_ncp == NULL) 1709 break; 1710 _cache_unlock(ncp); 1711 while ((ncp2 = track.resume_ncp) != NULL) { 1712 track.resume_ncp = NULL; 1713 _cache_lock(ncp2); 1714 _cache_inval_internal(ncp2, flags & ~CINV_DESTROY, 1715 &track); 1716 _cache_put(ncp2); 1717 } 1718 _cache_lock(ncp); 1719 } 1720 return(r); 1721 } 1722 1723 int 1724 cache_inval(struct nchandle *nch, int flags) 1725 { 1726 return(_cache_inval(nch->ncp, flags)); 1727 } 1728 1729 /* 1730 * Helper for _cache_inval(). The passed ncp is refd and locked and 1731 * remains that way on return, but may be unlocked/relocked multiple 1732 * times by the routine. 1733 */ 1734 static int 1735 _cache_inval_internal(struct namecache *ncp, int flags, struct cinvtrack *track) 1736 { 1737 struct namecache *nextkid; 1738 int rcnt = 0; 1739 1740 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE); 1741 1742 _cache_setunresolved(ncp); 1743 if (flags & CINV_DESTROY) { 1744 ncp->nc_flag |= NCF_DESTROYED; 1745 ++ncp->nc_generation; 1746 } 1747 while ((flags & CINV_CHILDREN) && 1748 (nextkid = TAILQ_FIRST(&ncp->nc_list)) != NULL 1749 ) { 1750 struct namecache *kid; 1751 int restart; 1752 1753 restart = 0; 1754 _cache_hold(nextkid); 1755 if (++track->depth > MAX_RECURSION_DEPTH) { 1756 track->resume_ncp = ncp; 1757 _cache_hold(ncp); 1758 ++rcnt; 1759 } 1760 while ((kid = nextkid) != NULL) { 1761 /* 1762 * Parent (ncp) must be locked for the iteration. 1763 */ 1764 nextkid = NULL; 1765 if (kid->nc_parent != ncp) { 1766 _cache_drop(kid); 1767 kprintf("cache_inval_internal restartA %s\n", 1768 ncp->nc_name); 1769 restart = 1; 1770 break; 1771 } 1772 if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL) 1773 _cache_hold(nextkid); 1774 1775 /* 1776 * Parent unlocked for this section to avoid 1777 * deadlocks. 1778 */ 1779 _cache_unlock(ncp); 1780 if (track->resume_ncp) { 1781 _cache_drop(kid); 1782 _cache_lock(ncp); 1783 break; 1784 } 1785 if ((kid->nc_flag & NCF_UNRESOLVED) == 0 || 1786 TAILQ_FIRST(&kid->nc_list) 1787 ) { 1788 _cache_lock(kid); 1789 if (kid->nc_parent != ncp) { 1790 kprintf("cache_inval_internal " 1791 "restartB %s\n", 1792 ncp->nc_name); 1793 restart = 1; 1794 _cache_unlock(kid); 1795 _cache_drop(kid); 1796 _cache_lock(ncp); 1797 break; 1798 } 1799 1800 rcnt += _cache_inval_internal(kid, flags & ~CINV_DESTROY, track); 1801 _cache_unlock(kid); 1802 } 1803 _cache_drop(kid); 1804 _cache_lock(ncp); 1805 } 1806 if (nextkid) 1807 _cache_drop(nextkid); 1808 --track->depth; 1809 if (restart == 0) 1810 break; 1811 } 1812 1813 /* 1814 * Someone could have gotten in there while ncp was unlocked, 1815 * retry if so. 1816 */ 1817 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 1818 ++rcnt; 1819 return (rcnt); 1820 } 1821 1822 /* 1823 * Invalidate a vnode's namecache associations. To avoid races against 1824 * the resolver we do not invalidate a node which we previously invalidated 1825 * but which was then re-resolved while we were in the invalidation loop. 1826 * 1827 * Returns non-zero if any namecache entries remain after the invalidation 1828 * loop completed. 1829 * 1830 * NOTE: Unlike the namecache topology which guarentees that ncp's will not 1831 * be ripped out of the topology while held, the vnode's v_namecache 1832 * list has no such restriction. NCP's can be ripped out of the list 1833 * at virtually any time if not locked, even if held. 1834 * 1835 * In addition, the v_namecache list itself must be locked via 1836 * the vnode's spinlock. 1837 */ 1838 int 1839 cache_inval_vp(struct vnode *vp, int flags) 1840 { 1841 struct namecache *ncp; 1842 struct namecache *next; 1843 1844 restart: 1845 spin_lock(&vp->v_spin); 1846 ncp = TAILQ_FIRST(&vp->v_namecache); 1847 if (ncp) 1848 _cache_hold(ncp); 1849 while (ncp) { 1850 /* loop entered with ncp held and vp spin-locked */ 1851 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL) 1852 _cache_hold(next); 1853 spin_unlock(&vp->v_spin); 1854 _cache_lock(ncp); 1855 if (ncp->nc_vp != vp) { 1856 kprintf("Warning: cache_inval_vp: race-A detected on " 1857 "%s\n", ncp->nc_name); 1858 _cache_put(ncp); 1859 if (next) 1860 _cache_drop(next); 1861 goto restart; 1862 } 1863 _cache_inval(ncp, flags); 1864 _cache_put(ncp); /* also releases reference */ 1865 ncp = next; 1866 spin_lock(&vp->v_spin); 1867 if (ncp && ncp->nc_vp != vp) { 1868 spin_unlock(&vp->v_spin); 1869 kprintf("Warning: cache_inval_vp: race-B detected on " 1870 "%s\n", ncp->nc_name); 1871 _cache_drop(ncp); 1872 goto restart; 1873 } 1874 } 1875 spin_unlock(&vp->v_spin); 1876 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 1877 } 1878 1879 /* 1880 * This routine is used instead of the normal cache_inval_vp() when we 1881 * are trying to recycle otherwise good vnodes. 1882 * 1883 * Return 0 on success, non-zero if not all namecache records could be 1884 * disassociated from the vnode (for various reasons). 1885 */ 1886 int 1887 cache_inval_vp_nonblock(struct vnode *vp) 1888 { 1889 struct namecache *ncp; 1890 struct namecache *next; 1891 1892 spin_lock(&vp->v_spin); 1893 ncp = TAILQ_FIRST(&vp->v_namecache); 1894 if (ncp) 1895 _cache_hold(ncp); 1896 while (ncp) { 1897 /* loop entered with ncp held */ 1898 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL) 1899 _cache_hold(next); 1900 spin_unlock(&vp->v_spin); 1901 if (_cache_lock_nonblock(ncp)) { 1902 _cache_drop(ncp); 1903 if (next) 1904 _cache_drop(next); 1905 goto done; 1906 } 1907 if (ncp->nc_vp != vp) { 1908 kprintf("Warning: cache_inval_vp: race-A detected on " 1909 "%s\n", ncp->nc_name); 1910 _cache_put(ncp); 1911 if (next) 1912 _cache_drop(next); 1913 goto done; 1914 } 1915 _cache_inval(ncp, 0); 1916 _cache_put(ncp); /* also releases reference */ 1917 ncp = next; 1918 spin_lock(&vp->v_spin); 1919 if (ncp && ncp->nc_vp != vp) { 1920 spin_unlock(&vp->v_spin); 1921 kprintf("Warning: cache_inval_vp: race-B detected on " 1922 "%s\n", ncp->nc_name); 1923 _cache_drop(ncp); 1924 goto done; 1925 } 1926 } 1927 spin_unlock(&vp->v_spin); 1928 done: 1929 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 1930 } 1931 1932 /* 1933 * Clears the universal directory search 'ok' flag. This flag allows 1934 * nlookup() to bypass normal vnode checks. This flag is a cached flag 1935 * so clearing it simply forces revalidation. 1936 */ 1937 void 1938 cache_inval_wxok(struct vnode *vp) 1939 { 1940 struct namecache *ncp; 1941 1942 spin_lock(&vp->v_spin); 1943 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 1944 if (ncp->nc_flag & NCF_WXOK) 1945 atomic_clear_short(&ncp->nc_flag, NCF_WXOK); 1946 } 1947 spin_unlock(&vp->v_spin); 1948 } 1949 1950 /* 1951 * The source ncp has been renamed to the target ncp. Both fncp and tncp 1952 * must be locked. The target ncp is destroyed (as a normal rename-over 1953 * would destroy the target file or directory). 1954 * 1955 * Because there may be references to the source ncp we cannot copy its 1956 * contents to the target. Instead the source ncp is relinked as the target 1957 * and the target ncp is removed from the namecache topology. 1958 */ 1959 void 1960 cache_rename(struct nchandle *fnch, struct nchandle *tnch) 1961 { 1962 struct namecache *fncp = fnch->ncp; 1963 struct namecache *tncp = tnch->ncp; 1964 struct namecache *tncp_par; 1965 struct nchash_head *nchpp; 1966 u_int32_t hash; 1967 char *oname; 1968 char *nname; 1969 1970 ++fncp->nc_generation; 1971 ++tncp->nc_generation; 1972 if (tncp->nc_nlen) { 1973 nname = kmalloc(tncp->nc_nlen + 1, M_VFSCACHE, M_WAITOK); 1974 bcopy(tncp->nc_name, nname, tncp->nc_nlen); 1975 nname[tncp->nc_nlen] = 0; 1976 } else { 1977 nname = NULL; 1978 } 1979 1980 /* 1981 * Rename fncp (unlink) 1982 */ 1983 _cache_unlink_parent(fncp); 1984 oname = fncp->nc_name; 1985 fncp->nc_name = nname; 1986 fncp->nc_nlen = tncp->nc_nlen; 1987 if (oname) 1988 kfree(oname, M_VFSCACHE); 1989 1990 tncp_par = tncp->nc_parent; 1991 _cache_hold(tncp_par); 1992 _cache_lock(tncp_par); 1993 1994 /* 1995 * Rename fncp (relink) 1996 */ 1997 hash = fnv_32_buf(fncp->nc_name, fncp->nc_nlen, FNV1_32_INIT); 1998 hash = fnv_32_buf(&tncp_par, sizeof(tncp_par), hash); 1999 nchpp = NCHHASH(hash); 2000 2001 spin_lock(&nchpp->spin); 2002 _cache_link_parent(fncp, tncp_par, nchpp); 2003 spin_unlock(&nchpp->spin); 2004 2005 _cache_put(tncp_par); 2006 2007 /* 2008 * Get rid of the overwritten tncp (unlink) 2009 */ 2010 _cache_unlink(tncp); 2011 } 2012 2013 /* 2014 * Perform actions consistent with unlinking a file. The passed-in ncp 2015 * must be locked. 2016 * 2017 * The ncp is marked DESTROYED so it no longer shows up in searches, 2018 * and will be physically deleted when the vnode goes away. 2019 * 2020 * If the related vnode has no refs then we cycle it through vget()/vput() 2021 * to (possibly if we don't have a ref race) trigger a deactivation, 2022 * allowing the VFS to trivially detect and recycle the deleted vnode 2023 * via VOP_INACTIVE(). 2024 * 2025 * NOTE: _cache_rename() will automatically call _cache_unlink() on the 2026 * target ncp. 2027 */ 2028 void 2029 cache_unlink(struct nchandle *nch) 2030 { 2031 _cache_unlink(nch->ncp); 2032 } 2033 2034 static void 2035 _cache_unlink(struct namecache *ncp) 2036 { 2037 struct vnode *vp; 2038 2039 /* 2040 * Causes lookups to fail and allows another ncp with the same 2041 * name to be created under ncp->nc_parent. 2042 */ 2043 ncp->nc_flag |= NCF_DESTROYED; 2044 ++ncp->nc_generation; 2045 2046 /* 2047 * Attempt to trigger a deactivation. Set VREF_FINALIZE to 2048 * force action on the 1->0 transition. 2049 */ 2050 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 && 2051 (vp = ncp->nc_vp) != NULL) { 2052 atomic_set_int(&vp->v_refcnt, VREF_FINALIZE); 2053 if (VREFCNT(vp) <= 0) { 2054 if (vget(vp, LK_SHARED) == 0) 2055 vput(vp); 2056 } 2057 } 2058 } 2059 2060 /* 2061 * Return non-zero if the nch might be associated with an open and/or mmap()'d 2062 * file. The easy solution is to just return non-zero if the vnode has refs. 2063 * Used to interlock hammer2 reclaims (VREF_FINALIZE should already be set to 2064 * force the reclaim). 2065 */ 2066 int 2067 cache_isopen(struct nchandle *nch) 2068 { 2069 struct vnode *vp; 2070 struct namecache *ncp = nch->ncp; 2071 2072 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0 && 2073 (vp = ncp->nc_vp) != NULL && 2074 VREFCNT(vp)) { 2075 return 1; 2076 } 2077 return 0; 2078 } 2079 2080 2081 /* 2082 * vget the vnode associated with the namecache entry. Resolve the namecache 2083 * entry if necessary. The passed ncp must be referenced and locked. If 2084 * the ncp is resolved it might be locked shared. 2085 * 2086 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked 2087 * (depending on the passed lk_type) will be returned in *vpp with an error 2088 * of 0, or NULL will be returned in *vpp with a non-0 error code. The 2089 * most typical error is ENOENT, meaning that the ncp represents a negative 2090 * cache hit and there is no vnode to retrieve, but other errors can occur 2091 * too. 2092 * 2093 * The vget() can race a reclaim. If this occurs we re-resolve the 2094 * namecache entry. 2095 * 2096 * There are numerous places in the kernel where vget() is called on a 2097 * vnode while one or more of its namecache entries is locked. Releasing 2098 * a vnode never deadlocks against locked namecache entries (the vnode 2099 * will not get recycled while referenced ncp's exist). This means we 2100 * can safely acquire the vnode. In fact, we MUST NOT release the ncp 2101 * lock when acquiring the vp lock or we might cause a deadlock. 2102 * 2103 * NOTE: The passed-in ncp must be locked exclusively if it is initially 2104 * unresolved. If a reclaim race occurs the passed-in ncp will be 2105 * relocked exclusively before being re-resolved. 2106 */ 2107 int 2108 cache_vget(struct nchandle *nch, struct ucred *cred, 2109 int lk_type, struct vnode **vpp) 2110 { 2111 struct namecache *ncp; 2112 struct vnode *vp; 2113 int error; 2114 2115 ncp = nch->ncp; 2116 again: 2117 vp = NULL; 2118 if (ncp->nc_flag & NCF_UNRESOLVED) 2119 error = cache_resolve(nch, cred); 2120 else 2121 error = 0; 2122 2123 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 2124 error = vget(vp, lk_type); 2125 if (error) { 2126 /* 2127 * VRECLAIM race 2128 * 2129 * The ncp may have been locked shared, we must relock 2130 * it exclusively before we can set it to unresolved. 2131 */ 2132 if (error == ENOENT) { 2133 kprintf("Warning: vnode reclaim race detected " 2134 "in cache_vget on %p (%s)\n", 2135 vp, ncp->nc_name); 2136 _cache_unlock(ncp); 2137 _cache_lock(ncp); 2138 _cache_setunresolved(ncp); 2139 goto again; 2140 } 2141 2142 /* 2143 * Not a reclaim race, some other error. 2144 */ 2145 KKASSERT(ncp->nc_vp == vp); 2146 vp = NULL; 2147 } else { 2148 KKASSERT(ncp->nc_vp == vp); 2149 KKASSERT((vp->v_flag & VRECLAIMED) == 0); 2150 } 2151 } 2152 if (error == 0 && vp == NULL) 2153 error = ENOENT; 2154 *vpp = vp; 2155 return(error); 2156 } 2157 2158 /* 2159 * Similar to cache_vget() but only acquires a ref on the vnode. 2160 * 2161 * NOTE: The passed-in ncp must be locked exclusively if it is initially 2162 * unresolved. If a reclaim race occurs the passed-in ncp will be 2163 * relocked exclusively before being re-resolved. 2164 */ 2165 int 2166 cache_vref(struct nchandle *nch, struct ucred *cred, struct vnode **vpp) 2167 { 2168 struct namecache *ncp; 2169 struct vnode *vp; 2170 int error; 2171 2172 ncp = nch->ncp; 2173 again: 2174 vp = NULL; 2175 if (ncp->nc_flag & NCF_UNRESOLVED) 2176 error = cache_resolve(nch, cred); 2177 else 2178 error = 0; 2179 2180 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 2181 error = vget(vp, LK_SHARED); 2182 if (error) { 2183 /* 2184 * VRECLAIM race 2185 */ 2186 if (error == ENOENT) { 2187 kprintf("Warning: vnode reclaim race detected " 2188 "in cache_vget on %p (%s)\n", 2189 vp, ncp->nc_name); 2190 _cache_unlock(ncp); 2191 _cache_lock(ncp); 2192 _cache_setunresolved(ncp); 2193 goto again; 2194 } 2195 2196 /* 2197 * Not a reclaim race, some other error. 2198 */ 2199 KKASSERT(ncp->nc_vp == vp); 2200 vp = NULL; 2201 } else { 2202 KKASSERT(ncp->nc_vp == vp); 2203 KKASSERT((vp->v_flag & VRECLAIMED) == 0); 2204 /* caller does not want a lock */ 2205 vn_unlock(vp); 2206 } 2207 } 2208 if (error == 0 && vp == NULL) 2209 error = ENOENT; 2210 *vpp = vp; 2211 return(error); 2212 } 2213 2214 /* 2215 * Return a referenced vnode representing the parent directory of 2216 * ncp. 2217 * 2218 * Because the caller has locked the ncp it should not be possible for 2219 * the parent ncp to go away. However, the parent can unresolve its 2220 * dvp at any time so we must be able to acquire a lock on the parent 2221 * to safely access nc_vp. 2222 * 2223 * We have to leave par unlocked when vget()ing dvp to avoid a deadlock, 2224 * so use vhold()/vdrop() while holding the lock to prevent dvp from 2225 * getting destroyed. 2226 * 2227 * NOTE: vhold() is allowed when dvp has 0 refs if we hold a 2228 * lock on the ncp in question.. 2229 */ 2230 static struct vnode * 2231 cache_dvpref(struct namecache *ncp) 2232 { 2233 struct namecache *par; 2234 struct vnode *dvp; 2235 2236 dvp = NULL; 2237 if ((par = ncp->nc_parent) != NULL) { 2238 _cache_hold(par); 2239 _cache_lock(par); 2240 if ((par->nc_flag & NCF_UNRESOLVED) == 0) { 2241 if ((dvp = par->nc_vp) != NULL) 2242 vhold(dvp); 2243 } 2244 _cache_unlock(par); 2245 if (dvp) { 2246 if (vget(dvp, LK_SHARED) == 0) { 2247 vn_unlock(dvp); 2248 vdrop(dvp); 2249 /* return refd, unlocked dvp */ 2250 } else { 2251 vdrop(dvp); 2252 dvp = NULL; 2253 } 2254 } 2255 _cache_drop(par); 2256 } 2257 return(dvp); 2258 } 2259 2260 /* 2261 * Convert a directory vnode to a namecache record without any other 2262 * knowledge of the topology. This ONLY works with directory vnodes and 2263 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the 2264 * returned ncp (if not NULL) will be held and unlocked. 2265 * 2266 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned. 2267 * If 'makeit' is 1 we attempt to track-down and create the namecache topology 2268 * for dvp. This will fail only if the directory has been deleted out from 2269 * under the caller. 2270 * 2271 * Callers must always check for a NULL return no matter the value of 'makeit'. 2272 * 2273 * To avoid underflowing the kernel stack each recursive call increments 2274 * the makeit variable. 2275 */ 2276 2277 static int cache_inefficient_scan(struct nchandle *nch, struct ucred *cred, 2278 struct vnode *dvp, char *fakename); 2279 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred, 2280 struct vnode **saved_dvp); 2281 2282 int 2283 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit, 2284 struct nchandle *nch) 2285 { 2286 struct vnode *saved_dvp; 2287 struct vnode *pvp; 2288 char *fakename; 2289 int error; 2290 2291 nch->ncp = NULL; 2292 nch->mount = dvp->v_mount; 2293 saved_dvp = NULL; 2294 fakename = NULL; 2295 2296 /* 2297 * Handle the makeit == 0 degenerate case 2298 */ 2299 if (makeit == 0) { 2300 spin_lock_shared(&dvp->v_spin); 2301 nch->ncp = TAILQ_FIRST(&dvp->v_namecache); 2302 if (nch->ncp) 2303 cache_hold(nch); 2304 spin_unlock_shared(&dvp->v_spin); 2305 } 2306 2307 /* 2308 * Loop until resolution, inside code will break out on error. 2309 */ 2310 while (makeit) { 2311 /* 2312 * Break out if we successfully acquire a working ncp. 2313 */ 2314 spin_lock_shared(&dvp->v_spin); 2315 nch->ncp = TAILQ_FIRST(&dvp->v_namecache); 2316 if (nch->ncp) { 2317 cache_hold(nch); 2318 spin_unlock_shared(&dvp->v_spin); 2319 break; 2320 } 2321 spin_unlock_shared(&dvp->v_spin); 2322 2323 /* 2324 * If dvp is the root of its filesystem it should already 2325 * have a namecache pointer associated with it as a side 2326 * effect of the mount, but it may have been disassociated. 2327 */ 2328 if (dvp->v_flag & VROOT) { 2329 nch->ncp = _cache_get(nch->mount->mnt_ncmountpt.ncp); 2330 error = cache_resolve_mp(nch->mount); 2331 _cache_put(nch->ncp); 2332 if (ncvp_debug) { 2333 kprintf("cache_fromdvp: resolve root of mount %p error %d", 2334 dvp->v_mount, error); 2335 } 2336 if (error) { 2337 if (ncvp_debug) 2338 kprintf(" failed\n"); 2339 nch->ncp = NULL; 2340 break; 2341 } 2342 if (ncvp_debug) 2343 kprintf(" succeeded\n"); 2344 continue; 2345 } 2346 2347 /* 2348 * If we are recursed too deeply resort to an O(n^2) 2349 * algorithm to resolve the namecache topology. The 2350 * resolved pvp is left referenced in saved_dvp to 2351 * prevent the tree from being destroyed while we loop. 2352 */ 2353 if (makeit > 20) { 2354 error = cache_fromdvp_try(dvp, cred, &saved_dvp); 2355 if (error) { 2356 kprintf("lookupdotdot(longpath) failed %d " 2357 "dvp %p\n", error, dvp); 2358 nch->ncp = NULL; 2359 break; 2360 } 2361 continue; 2362 } 2363 2364 /* 2365 * Get the parent directory and resolve its ncp. 2366 */ 2367 if (fakename) { 2368 kfree(fakename, M_TEMP); 2369 fakename = NULL; 2370 } 2371 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred, 2372 &fakename); 2373 if (error) { 2374 kprintf("lookupdotdot failed %d dvp %p\n", error, dvp); 2375 break; 2376 } 2377 vn_unlock(pvp); 2378 2379 /* 2380 * Reuse makeit as a recursion depth counter. On success 2381 * nch will be fully referenced. 2382 */ 2383 cache_fromdvp(pvp, cred, makeit + 1, nch); 2384 vrele(pvp); 2385 if (nch->ncp == NULL) 2386 break; 2387 2388 /* 2389 * Do an inefficient scan of pvp (embodied by ncp) to look 2390 * for dvp. This will create a namecache record for dvp on 2391 * success. We loop up to recheck on success. 2392 * 2393 * ncp and dvp are both held but not locked. 2394 */ 2395 error = cache_inefficient_scan(nch, cred, dvp, fakename); 2396 if (error) { 2397 kprintf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n", 2398 pvp, nch->ncp->nc_name, dvp); 2399 cache_drop(nch); 2400 /* nch was NULLed out, reload mount */ 2401 nch->mount = dvp->v_mount; 2402 break; 2403 } 2404 if (ncvp_debug) { 2405 kprintf("cache_fromdvp: scan %p (%s) succeeded\n", 2406 pvp, nch->ncp->nc_name); 2407 } 2408 cache_drop(nch); 2409 /* nch was NULLed out, reload mount */ 2410 nch->mount = dvp->v_mount; 2411 } 2412 2413 /* 2414 * If nch->ncp is non-NULL it will have been held already. 2415 */ 2416 if (fakename) 2417 kfree(fakename, M_TEMP); 2418 if (saved_dvp) 2419 vrele(saved_dvp); 2420 if (nch->ncp) 2421 return (0); 2422 return (EINVAL); 2423 } 2424 2425 /* 2426 * Go up the chain of parent directories until we find something 2427 * we can resolve into the namecache. This is very inefficient. 2428 */ 2429 static 2430 int 2431 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred, 2432 struct vnode **saved_dvp) 2433 { 2434 struct nchandle nch; 2435 struct vnode *pvp; 2436 int error; 2437 static time_t last_fromdvp_report; 2438 char *fakename; 2439 2440 /* 2441 * Loop getting the parent directory vnode until we get something we 2442 * can resolve in the namecache. 2443 */ 2444 vref(dvp); 2445 nch.mount = dvp->v_mount; 2446 nch.ncp = NULL; 2447 fakename = NULL; 2448 2449 for (;;) { 2450 if (fakename) { 2451 kfree(fakename, M_TEMP); 2452 fakename = NULL; 2453 } 2454 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred, 2455 &fakename); 2456 if (error) { 2457 vrele(dvp); 2458 break; 2459 } 2460 vn_unlock(pvp); 2461 spin_lock_shared(&pvp->v_spin); 2462 if ((nch.ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) { 2463 _cache_hold(nch.ncp); 2464 spin_unlock_shared(&pvp->v_spin); 2465 vrele(pvp); 2466 break; 2467 } 2468 spin_unlock_shared(&pvp->v_spin); 2469 if (pvp->v_flag & VROOT) { 2470 nch.ncp = _cache_get(pvp->v_mount->mnt_ncmountpt.ncp); 2471 error = cache_resolve_mp(nch.mount); 2472 _cache_unlock(nch.ncp); 2473 vrele(pvp); 2474 if (error) { 2475 _cache_drop(nch.ncp); 2476 nch.ncp = NULL; 2477 vrele(dvp); 2478 } 2479 break; 2480 } 2481 vrele(dvp); 2482 dvp = pvp; 2483 } 2484 if (error == 0) { 2485 if (last_fromdvp_report != time_uptime) { 2486 last_fromdvp_report = time_uptime; 2487 kprintf("Warning: extremely inefficient path " 2488 "resolution on %s\n", 2489 nch.ncp->nc_name); 2490 } 2491 error = cache_inefficient_scan(&nch, cred, dvp, fakename); 2492 2493 /* 2494 * Hopefully dvp now has a namecache record associated with 2495 * it. Leave it referenced to prevent the kernel from 2496 * recycling the vnode. Otherwise extremely long directory 2497 * paths could result in endless recycling. 2498 */ 2499 if (*saved_dvp) 2500 vrele(*saved_dvp); 2501 *saved_dvp = dvp; 2502 _cache_drop(nch.ncp); 2503 } 2504 if (fakename) 2505 kfree(fakename, M_TEMP); 2506 return (error); 2507 } 2508 2509 /* 2510 * Do an inefficient scan of the directory represented by ncp looking for 2511 * the directory vnode dvp. ncp must be held but not locked on entry and 2512 * will be held on return. dvp must be refd but not locked on entry and 2513 * will remain refd on return. 2514 * 2515 * Why do this at all? Well, due to its stateless nature the NFS server 2516 * converts file handles directly to vnodes without necessarily going through 2517 * the namecache ops that would otherwise create the namecache topology 2518 * leading to the vnode. We could either (1) Change the namecache algorithms 2519 * to allow disconnect namecache records that are re-merged opportunistically, 2520 * or (2) Make the NFS server backtrack and scan to recover a connected 2521 * namecache topology in order to then be able to issue new API lookups. 2522 * 2523 * It turns out that (1) is a huge mess. It takes a nice clean set of 2524 * namecache algorithms and introduces a lot of complication in every subsystem 2525 * that calls into the namecache to deal with the re-merge case, especially 2526 * since we are using the namecache to placehold negative lookups and the 2527 * vnode might not be immediately assigned. (2) is certainly far less 2528 * efficient then (1), but since we are only talking about directories here 2529 * (which are likely to remain cached), the case does not actually run all 2530 * that often and has the supreme advantage of not polluting the namecache 2531 * algorithms. 2532 * 2533 * If a fakename is supplied just construct a namecache entry using the 2534 * fake name. 2535 */ 2536 static int 2537 cache_inefficient_scan(struct nchandle *nch, struct ucred *cred, 2538 struct vnode *dvp, char *fakename) 2539 { 2540 struct nlcomponent nlc; 2541 struct nchandle rncp; 2542 struct dirent *den; 2543 struct vnode *pvp; 2544 struct vattr vat; 2545 struct iovec iov; 2546 struct uio uio; 2547 int blksize; 2548 int eofflag; 2549 int bytes; 2550 char *rbuf; 2551 int error; 2552 2553 vat.va_blocksize = 0; 2554 if ((error = VOP_GETATTR(dvp, &vat)) != 0) 2555 return (error); 2556 cache_lock(nch); 2557 error = cache_vref(nch, cred, &pvp); 2558 cache_unlock(nch); 2559 if (error) 2560 return (error); 2561 if (ncvp_debug) { 2562 kprintf("inefficient_scan of (%p,%s): directory iosize %ld " 2563 "vattr fileid = %lld\n", 2564 nch->ncp, nch->ncp->nc_name, 2565 vat.va_blocksize, 2566 (long long)vat.va_fileid); 2567 } 2568 2569 /* 2570 * Use the supplied fakename if not NULL. Fake names are typically 2571 * not in the actual filesystem hierarchy. This is used by HAMMER 2572 * to glue @@timestamp recursions together. 2573 */ 2574 if (fakename) { 2575 nlc.nlc_nameptr = fakename; 2576 nlc.nlc_namelen = strlen(fakename); 2577 rncp = cache_nlookup(nch, &nlc); 2578 goto done; 2579 } 2580 2581 if ((blksize = vat.va_blocksize) == 0) 2582 blksize = DEV_BSIZE; 2583 rbuf = kmalloc(blksize, M_TEMP, M_WAITOK); 2584 rncp.ncp = NULL; 2585 2586 eofflag = 0; 2587 uio.uio_offset = 0; 2588 again: 2589 iov.iov_base = rbuf; 2590 iov.iov_len = blksize; 2591 uio.uio_iov = &iov; 2592 uio.uio_iovcnt = 1; 2593 uio.uio_resid = blksize; 2594 uio.uio_segflg = UIO_SYSSPACE; 2595 uio.uio_rw = UIO_READ; 2596 uio.uio_td = curthread; 2597 2598 if (ncvp_debug >= 2) 2599 kprintf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset); 2600 error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL); 2601 if (error == 0) { 2602 den = (struct dirent *)rbuf; 2603 bytes = blksize - uio.uio_resid; 2604 2605 while (bytes > 0) { 2606 if (ncvp_debug >= 2) { 2607 kprintf("cache_inefficient_scan: %*.*s\n", 2608 den->d_namlen, den->d_namlen, 2609 den->d_name); 2610 } 2611 if (den->d_type != DT_WHT && 2612 den->d_ino == vat.va_fileid) { 2613 if (ncvp_debug) { 2614 kprintf("cache_inefficient_scan: " 2615 "MATCHED inode %lld path %s/%*.*s\n", 2616 (long long)vat.va_fileid, 2617 nch->ncp->nc_name, 2618 den->d_namlen, den->d_namlen, 2619 den->d_name); 2620 } 2621 nlc.nlc_nameptr = den->d_name; 2622 nlc.nlc_namelen = den->d_namlen; 2623 rncp = cache_nlookup(nch, &nlc); 2624 KKASSERT(rncp.ncp != NULL); 2625 break; 2626 } 2627 bytes -= _DIRENT_DIRSIZ(den); 2628 den = _DIRENT_NEXT(den); 2629 } 2630 if (rncp.ncp == NULL && eofflag == 0 && uio.uio_resid != blksize) 2631 goto again; 2632 } 2633 kfree(rbuf, M_TEMP); 2634 done: 2635 vrele(pvp); 2636 if (rncp.ncp) { 2637 if (rncp.ncp->nc_flag & NCF_UNRESOLVED) { 2638 _cache_setvp(rncp.mount, rncp.ncp, dvp); 2639 if (ncvp_debug >= 2) { 2640 kprintf("cache_inefficient_scan: setvp %s/%s = %p\n", 2641 nch->ncp->nc_name, rncp.ncp->nc_name, dvp); 2642 } 2643 } else { 2644 if (ncvp_debug >= 2) { 2645 kprintf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n", 2646 nch->ncp->nc_name, rncp.ncp->nc_name, dvp, 2647 rncp.ncp->nc_vp); 2648 } 2649 } 2650 if (rncp.ncp->nc_vp == NULL) 2651 error = rncp.ncp->nc_error; 2652 /* 2653 * Release rncp after a successful nlookup. rncp was fully 2654 * referenced. 2655 */ 2656 cache_put(&rncp); 2657 } else { 2658 kprintf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n", 2659 dvp, nch->ncp->nc_name); 2660 error = ENOENT; 2661 } 2662 return (error); 2663 } 2664 2665 /* 2666 * Zap a namecache entry. The ncp is unconditionally set to an unresolved 2667 * state, which disassociates it from its vnode or ncneglist. 2668 * 2669 * Then, if there are no additional references to the ncp and no children, 2670 * the ncp is removed from the topology and destroyed. 2671 * 2672 * References and/or children may exist if the ncp is in the middle of the 2673 * topology, preventing the ncp from being destroyed. 2674 * 2675 * This function must be called with the ncp held and locked and will unlock 2676 * and drop it during zapping. 2677 * 2678 * If nonblock is non-zero and the parent ncp cannot be locked we give up. 2679 * This case can occur in the cache_drop() path. 2680 * 2681 * This function may returned a held (but NOT locked) parent node which the 2682 * caller must drop. We do this so _cache_drop() can loop, to avoid 2683 * blowing out the kernel stack. 2684 * 2685 * WARNING! For MPSAFE operation this routine must acquire up to three 2686 * spin locks to be able to safely test nc_refs. Lock order is 2687 * very important. 2688 * 2689 * hash spinlock if on hash list 2690 * parent spinlock if child of parent 2691 * (the ncp is unresolved so there is no vnode association) 2692 */ 2693 static struct namecache * 2694 cache_zap(struct namecache *ncp, int nonblock) 2695 { 2696 struct namecache *par; 2697 struct vnode *dropvp; 2698 struct nchash_head *nchpp; 2699 int refs; 2700 2701 /* 2702 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED. 2703 */ 2704 _cache_setunresolved(ncp); 2705 2706 /* 2707 * Try to scrap the entry and possibly tail-recurse on its parent. 2708 * We only scrap unref'd (other then our ref) unresolved entries, 2709 * we do not scrap 'live' entries. 2710 * 2711 * Note that once the spinlocks are acquired if nc_refs == 1 no 2712 * other references are possible. If it isn't, however, we have 2713 * to decrement but also be sure to avoid a 1->0 transition. 2714 */ 2715 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED); 2716 KKASSERT(ncp->nc_refs > 0); 2717 2718 /* 2719 * Acquire locks. Note that the parent can't go away while we hold 2720 * a child locked. 2721 */ 2722 nchpp = NULL; 2723 if ((par = ncp->nc_parent) != NULL) { 2724 if (nonblock) { 2725 for (;;) { 2726 if (_cache_lock_nonblock(par) == 0) 2727 break; 2728 refs = ncp->nc_refs; 2729 ncp->nc_flag |= NCF_DEFEREDZAP; 2730 ++numdefered; /* MP race ok */ 2731 if (atomic_cmpset_int(&ncp->nc_refs, 2732 refs, refs - 1)) { 2733 _cache_unlock(ncp); 2734 return(NULL); 2735 } 2736 cpu_pause(); 2737 } 2738 _cache_hold(par); 2739 } else { 2740 _cache_hold(par); 2741 _cache_lock(par); 2742 } 2743 nchpp = ncp->nc_head; 2744 spin_lock(&nchpp->spin); 2745 } 2746 2747 /* 2748 * At this point if we find refs == 1 it should not be possible for 2749 * anyone else to have access to the ncp. We are holding the only 2750 * possible access point left (nchpp) spin-locked. 2751 * 2752 * If someone other then us has a ref or we have children 2753 * we cannot zap the entry. The 1->0 transition and any 2754 * further list operation is protected by the spinlocks 2755 * we have acquired but other transitions are not. 2756 */ 2757 for (;;) { 2758 refs = ncp->nc_refs; 2759 cpu_ccfence(); 2760 if (refs == 1 && TAILQ_EMPTY(&ncp->nc_list)) 2761 break; 2762 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) { 2763 if (par) { 2764 spin_unlock(&nchpp->spin); 2765 _cache_put(par); 2766 } 2767 _cache_unlock(ncp); 2768 return(NULL); 2769 } 2770 cpu_pause(); 2771 } 2772 2773 /* 2774 * We are the only ref and with the spinlocks held no further 2775 * refs can be acquired by others. 2776 * 2777 * Remove us from the hash list and parent list. We have to 2778 * drop a ref on the parent's vp if the parent's list becomes 2779 * empty. 2780 */ 2781 dropvp = NULL; 2782 if (par) { 2783 KKASSERT(nchpp == ncp->nc_head); 2784 LIST_REMOVE(ncp, nc_hash); 2785 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 2786 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 2787 dropvp = par->nc_vp; 2788 ncp->nc_head = NULL; 2789 ncp->nc_parent = NULL; 2790 spin_unlock(&nchpp->spin); 2791 _cache_unlock(par); 2792 } else { 2793 KKASSERT(ncp->nc_head == NULL); 2794 } 2795 2796 /* 2797 * ncp should not have picked up any refs. Physically 2798 * destroy the ncp. 2799 */ 2800 if (ncp->nc_refs != 1) { 2801 int save_refs = ncp->nc_refs; 2802 cpu_ccfence(); 2803 panic("cache_zap: %p bad refs %d (%d)\n", 2804 ncp, save_refs, atomic_fetchadd_int(&ncp->nc_refs, 0)); 2805 } 2806 KKASSERT(ncp->nc_refs == 1); 2807 /* _cache_unlock(ncp) not required */ 2808 ncp->nc_refs = -1; /* safety */ 2809 if (ncp->nc_name) 2810 kfree(ncp->nc_name, M_VFSCACHE); 2811 kfree(ncp, M_VFSCACHE); 2812 2813 /* 2814 * Delayed drop (we had to release our spinlocks) 2815 * 2816 * The refed parent (if not NULL) must be dropped. The 2817 * caller is responsible for looping. 2818 */ 2819 if (dropvp) 2820 vdrop(dropvp); 2821 return(par); 2822 } 2823 2824 /* 2825 * Clean up dangling negative cache and defered-drop entries in the 2826 * namecache. 2827 * 2828 * This routine is called in the critical path and also called from 2829 * vnlru(). When called from vnlru we use a lower limit to try to 2830 * deal with the negative cache before the critical path has to start 2831 * dealing with it. 2832 */ 2833 typedef enum { CHI_LOW, CHI_HIGH } cache_hs_t; 2834 2835 static cache_hs_t neg_cache_hysteresis_state[2] = { CHI_LOW, CHI_LOW }; 2836 static cache_hs_t pos_cache_hysteresis_state[2] = { CHI_LOW, CHI_LOW }; 2837 2838 void 2839 cache_hysteresis(int critpath) 2840 { 2841 int poslimit; 2842 int neglimit = maxvnodes / ncnegfactor; 2843 int xnumcache = numcache; 2844 2845 if (critpath == 0) 2846 neglimit = neglimit * 8 / 10; 2847 2848 /* 2849 * Don't cache too many negative hits. We use hysteresis to reduce 2850 * the impact on the critical path. 2851 */ 2852 switch(neg_cache_hysteresis_state[critpath]) { 2853 case CHI_LOW: 2854 if (numneg > MINNEG && numneg > neglimit) { 2855 if (critpath) 2856 _cache_cleanneg(ncnegflush); 2857 else 2858 _cache_cleanneg(ncnegflush + 2859 numneg - neglimit); 2860 neg_cache_hysteresis_state[critpath] = CHI_HIGH; 2861 } 2862 break; 2863 case CHI_HIGH: 2864 if (numneg > MINNEG * 9 / 10 && 2865 numneg * 9 / 10 > neglimit 2866 ) { 2867 if (critpath) 2868 _cache_cleanneg(ncnegflush); 2869 else 2870 _cache_cleanneg(ncnegflush + 2871 numneg * 9 / 10 - neglimit); 2872 } else { 2873 neg_cache_hysteresis_state[critpath] = CHI_LOW; 2874 } 2875 break; 2876 } 2877 2878 /* 2879 * Don't cache too many positive hits. We use hysteresis to reduce 2880 * the impact on the critical path. 2881 * 2882 * Excessive positive hits can accumulate due to large numbers of 2883 * hardlinks (the vnode cache will not prevent hl ncps from growing 2884 * into infinity). 2885 */ 2886 if ((poslimit = ncposlimit) == 0) 2887 poslimit = maxvnodes * 2; 2888 if (critpath == 0) 2889 poslimit = poslimit * 8 / 10; 2890 2891 switch(pos_cache_hysteresis_state[critpath]) { 2892 case CHI_LOW: 2893 if (xnumcache > poslimit && xnumcache > MINPOS) { 2894 if (critpath) 2895 _cache_cleanpos(ncposflush); 2896 else 2897 _cache_cleanpos(ncposflush + 2898 xnumcache - poslimit); 2899 pos_cache_hysteresis_state[critpath] = CHI_HIGH; 2900 } 2901 break; 2902 case CHI_HIGH: 2903 if (xnumcache > poslimit * 5 / 6 && xnumcache > MINPOS) { 2904 if (critpath) 2905 _cache_cleanpos(ncposflush); 2906 else 2907 _cache_cleanpos(ncposflush + 2908 xnumcache - poslimit * 5 / 6); 2909 } else { 2910 pos_cache_hysteresis_state[critpath] = CHI_LOW; 2911 } 2912 break; 2913 } 2914 2915 /* 2916 * Clean out dangling defered-zap ncps which could not 2917 * be cleanly dropped if too many build up. Note 2918 * that numdefered is not an exact number as such ncps 2919 * can be reused and the counter is not handled in a MP 2920 * safe manner by design. 2921 */ 2922 if (numdefered > neglimit) { 2923 _cache_cleandefered(); 2924 } 2925 } 2926 2927 /* 2928 * NEW NAMECACHE LOOKUP API 2929 * 2930 * Lookup an entry in the namecache. The passed par_nch must be referenced 2931 * and unlocked. A referenced and locked nchandle with a non-NULL nch.ncp 2932 * is ALWAYS returned, eve if the supplied component is illegal. 2933 * 2934 * The resulting namecache entry should be returned to the system with 2935 * cache_put() or cache_unlock() + cache_drop(). 2936 * 2937 * namecache locks are recursive but care must be taken to avoid lock order 2938 * reversals (hence why the passed par_nch must be unlocked). Locking 2939 * rules are to order for parent traversals, not for child traversals. 2940 * 2941 * Nobody else will be able to manipulate the associated namespace (e.g. 2942 * create, delete, rename, rename-target) until the caller unlocks the 2943 * entry. 2944 * 2945 * The returned entry will be in one of three states: positive hit (non-null 2946 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set). 2947 * Unresolved entries must be resolved through the filesystem to associate the 2948 * vnode and/or determine whether a positive or negative hit has occured. 2949 * 2950 * It is not necessary to lock a directory in order to lock namespace under 2951 * that directory. In fact, it is explicitly not allowed to do that. A 2952 * directory is typically only locked when being created, renamed, or 2953 * destroyed. 2954 * 2955 * The directory (par) may be unresolved, in which case any returned child 2956 * will likely also be marked unresolved. Likely but not guarenteed. Since 2957 * the filesystem lookup requires a resolved directory vnode the caller is 2958 * responsible for resolving the namecache chain top-down. This API 2959 * specifically allows whole chains to be created in an unresolved state. 2960 */ 2961 struct nchandle 2962 cache_nlookup(struct nchandle *par_nch, struct nlcomponent *nlc) 2963 { 2964 struct nchandle nch; 2965 struct namecache *ncp; 2966 struct namecache *new_ncp; 2967 struct nchash_head *nchpp; 2968 struct mount *mp; 2969 u_int32_t hash; 2970 globaldata_t gd; 2971 int par_locked; 2972 2973 gd = mycpu; 2974 mp = par_nch->mount; 2975 par_locked = 0; 2976 2977 /* 2978 * This is a good time to call it, no ncp's are locked by 2979 * the caller or us. 2980 */ 2981 cache_hysteresis(1); 2982 2983 /* 2984 * Try to locate an existing entry 2985 */ 2986 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 2987 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash); 2988 new_ncp = NULL; 2989 nchpp = NCHHASH(hash); 2990 restart: 2991 if (new_ncp) 2992 spin_lock(&nchpp->spin); 2993 else 2994 spin_lock_shared(&nchpp->spin); 2995 2996 LIST_FOREACH(ncp, &nchpp->list, nc_hash) { 2997 /* 2998 * Break out if we find a matching entry. Note that 2999 * UNRESOLVED entries may match, but DESTROYED entries 3000 * do not. 3001 */ 3002 if (ncp->nc_parent == par_nch->ncp && 3003 ncp->nc_nlen == nlc->nlc_namelen && 3004 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 3005 (ncp->nc_flag & NCF_DESTROYED) == 0 3006 ) { 3007 _cache_hold(ncp); 3008 if (new_ncp) 3009 spin_unlock(&nchpp->spin); 3010 else 3011 spin_unlock_shared(&nchpp->spin); 3012 if (par_locked) { 3013 _cache_unlock(par_nch->ncp); 3014 par_locked = 0; 3015 } 3016 if (_cache_lock_special(ncp) == 0) { 3017 /* 3018 * Successfully locked but we must re-test 3019 * conditions that might have changed since 3020 * we did not have the lock before. 3021 */ 3022 if (ncp->nc_parent != par_nch->ncp || 3023 ncp->nc_nlen != nlc->nlc_namelen || 3024 bcmp(ncp->nc_name, nlc->nlc_nameptr, 3025 ncp->nc_nlen) || 3026 (ncp->nc_flag & NCF_DESTROYED)) { 3027 _cache_put(ncp); 3028 goto restart; 3029 } 3030 _cache_auto_unresolve(mp, ncp); 3031 if (new_ncp) 3032 _cache_free(new_ncp); 3033 goto found; 3034 } 3035 _cache_get(ncp); /* cycle the lock to block */ 3036 _cache_put(ncp); 3037 _cache_drop(ncp); 3038 goto restart; 3039 } 3040 } 3041 3042 /* 3043 * We failed to locate an entry, create a new entry and add it to 3044 * the cache. The parent ncp must also be locked so we 3045 * can link into it. 3046 * 3047 * We have to relookup after possibly blocking in kmalloc or 3048 * when locking par_nch. 3049 * 3050 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special 3051 * mount case, in which case nc_name will be NULL. 3052 */ 3053 if (new_ncp == NULL) { 3054 spin_unlock_shared(&nchpp->spin); 3055 new_ncp = cache_alloc(nlc->nlc_namelen); 3056 if (nlc->nlc_namelen) { 3057 bcopy(nlc->nlc_nameptr, new_ncp->nc_name, 3058 nlc->nlc_namelen); 3059 new_ncp->nc_name[nlc->nlc_namelen] = 0; 3060 } 3061 goto restart; 3062 } 3063 3064 /* 3065 * NOTE! The spinlock is held exclusively here because new_ncp 3066 * is non-NULL. 3067 */ 3068 if (par_locked == 0) { 3069 spin_unlock(&nchpp->spin); 3070 _cache_lock(par_nch->ncp); 3071 par_locked = 1; 3072 goto restart; 3073 } 3074 3075 /* 3076 * WARNING! We still hold the spinlock. We have to set the hash 3077 * table entry atomically. 3078 */ 3079 ncp = new_ncp; 3080 _cache_link_parent(ncp, par_nch->ncp, nchpp); 3081 spin_unlock(&nchpp->spin); 3082 _cache_unlock(par_nch->ncp); 3083 /* par_locked = 0 - not used */ 3084 found: 3085 /* 3086 * stats and namecache size management 3087 */ 3088 if (ncp->nc_flag & NCF_UNRESOLVED) 3089 ++gd->gd_nchstats->ncs_miss; 3090 else if (ncp->nc_vp) 3091 ++gd->gd_nchstats->ncs_goodhits; 3092 else 3093 ++gd->gd_nchstats->ncs_neghits; 3094 nch.mount = mp; 3095 nch.ncp = ncp; 3096 _cache_mntref(nch.mount); 3097 3098 return(nch); 3099 } 3100 3101 /* 3102 * Attempt to lookup a namecache entry and return with a shared namecache 3103 * lock. 3104 */ 3105 int 3106 cache_nlookup_maybe_shared(struct nchandle *par_nch, struct nlcomponent *nlc, 3107 int excl, struct nchandle *res_nch) 3108 { 3109 struct namecache *ncp; 3110 struct nchash_head *nchpp; 3111 struct mount *mp; 3112 u_int32_t hash; 3113 globaldata_t gd; 3114 3115 /* 3116 * If exclusive requested or shared namecache locks are disabled, 3117 * return failure. 3118 */ 3119 if (ncp_shared_lock_disable || excl) 3120 return(EWOULDBLOCK); 3121 3122 gd = mycpu; 3123 mp = par_nch->mount; 3124 3125 /* 3126 * This is a good time to call it, no ncp's are locked by 3127 * the caller or us. 3128 */ 3129 cache_hysteresis(1); 3130 3131 /* 3132 * Try to locate an existing entry 3133 */ 3134 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 3135 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash); 3136 nchpp = NCHHASH(hash); 3137 3138 spin_lock_shared(&nchpp->spin); 3139 3140 LIST_FOREACH(ncp, &nchpp->list, nc_hash) { 3141 /* 3142 * Break out if we find a matching entry. Note that 3143 * UNRESOLVED entries may match, but DESTROYED entries 3144 * do not. 3145 */ 3146 if (ncp->nc_parent == par_nch->ncp && 3147 ncp->nc_nlen == nlc->nlc_namelen && 3148 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 3149 (ncp->nc_flag & NCF_DESTROYED) == 0 3150 ) { 3151 _cache_hold(ncp); 3152 spin_unlock_shared(&nchpp->spin); 3153 if (_cache_lock_shared_special(ncp) == 0) { 3154 if (ncp->nc_parent == par_nch->ncp && 3155 ncp->nc_nlen == nlc->nlc_namelen && 3156 bcmp(ncp->nc_name, nlc->nlc_nameptr, 3157 ncp->nc_nlen) == 0 && 3158 (ncp->nc_flag & NCF_DESTROYED) == 0 && 3159 (ncp->nc_flag & NCF_UNRESOLVED) == 0 && 3160 _cache_auto_unresolve_test(mp, ncp) == 0) { 3161 goto found; 3162 } 3163 _cache_unlock(ncp); 3164 } 3165 _cache_drop(ncp); 3166 spin_lock_shared(&nchpp->spin); 3167 break; 3168 } 3169 } 3170 3171 /* 3172 * Failure 3173 */ 3174 spin_unlock_shared(&nchpp->spin); 3175 return(EWOULDBLOCK); 3176 3177 /* 3178 * Success 3179 * 3180 * Note that nc_error might be non-zero (e.g ENOENT). 3181 */ 3182 found: 3183 res_nch->mount = mp; 3184 res_nch->ncp = ncp; 3185 ++gd->gd_nchstats->ncs_goodhits; 3186 _cache_mntref(res_nch->mount); 3187 3188 KKASSERT(ncp->nc_error != EWOULDBLOCK); 3189 return(ncp->nc_error); 3190 } 3191 3192 /* 3193 * This is a non-blocking verison of cache_nlookup() used by 3194 * nfs_readdirplusrpc_uio(). It can fail for any reason and 3195 * will return nch.ncp == NULL in that case. 3196 */ 3197 struct nchandle 3198 cache_nlookup_nonblock(struct nchandle *par_nch, struct nlcomponent *nlc) 3199 { 3200 struct nchandle nch; 3201 struct namecache *ncp; 3202 struct namecache *new_ncp; 3203 struct nchash_head *nchpp; 3204 struct mount *mp; 3205 u_int32_t hash; 3206 globaldata_t gd; 3207 int par_locked; 3208 3209 gd = mycpu; 3210 mp = par_nch->mount; 3211 par_locked = 0; 3212 3213 /* 3214 * Try to locate an existing entry 3215 */ 3216 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 3217 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash); 3218 new_ncp = NULL; 3219 nchpp = NCHHASH(hash); 3220 restart: 3221 spin_lock(&nchpp->spin); 3222 LIST_FOREACH(ncp, &nchpp->list, nc_hash) { 3223 /* 3224 * Break out if we find a matching entry. Note that 3225 * UNRESOLVED entries may match, but DESTROYED entries 3226 * do not. 3227 */ 3228 if (ncp->nc_parent == par_nch->ncp && 3229 ncp->nc_nlen == nlc->nlc_namelen && 3230 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 3231 (ncp->nc_flag & NCF_DESTROYED) == 0 3232 ) { 3233 _cache_hold(ncp); 3234 spin_unlock(&nchpp->spin); 3235 if (par_locked) { 3236 _cache_unlock(par_nch->ncp); 3237 par_locked = 0; 3238 } 3239 if (_cache_lock_special(ncp) == 0) { 3240 if (ncp->nc_parent != par_nch->ncp || 3241 ncp->nc_nlen != nlc->nlc_namelen || 3242 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) || 3243 (ncp->nc_flag & NCF_DESTROYED)) { 3244 kprintf("cache_lookup_nonblock: " 3245 "ncp-race %p %*.*s\n", 3246 ncp, 3247 nlc->nlc_namelen, 3248 nlc->nlc_namelen, 3249 nlc->nlc_nameptr); 3250 _cache_unlock(ncp); 3251 _cache_drop(ncp); 3252 goto failed; 3253 } 3254 _cache_auto_unresolve(mp, ncp); 3255 if (new_ncp) { 3256 _cache_free(new_ncp); 3257 new_ncp = NULL; 3258 } 3259 goto found; 3260 } 3261 _cache_drop(ncp); 3262 goto failed; 3263 } 3264 } 3265 3266 /* 3267 * We failed to locate an entry, create a new entry and add it to 3268 * the cache. The parent ncp must also be locked so we 3269 * can link into it. 3270 * 3271 * We have to relookup after possibly blocking in kmalloc or 3272 * when locking par_nch. 3273 * 3274 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special 3275 * mount case, in which case nc_name will be NULL. 3276 */ 3277 if (new_ncp == NULL) { 3278 spin_unlock(&nchpp->spin); 3279 new_ncp = cache_alloc(nlc->nlc_namelen); 3280 if (nlc->nlc_namelen) { 3281 bcopy(nlc->nlc_nameptr, new_ncp->nc_name, 3282 nlc->nlc_namelen); 3283 new_ncp->nc_name[nlc->nlc_namelen] = 0; 3284 } 3285 goto restart; 3286 } 3287 if (par_locked == 0) { 3288 spin_unlock(&nchpp->spin); 3289 if (_cache_lock_nonblock(par_nch->ncp) == 0) { 3290 par_locked = 1; 3291 goto restart; 3292 } 3293 goto failed; 3294 } 3295 3296 /* 3297 * WARNING! We still hold the spinlock. We have to set the hash 3298 * table entry atomically. 3299 */ 3300 ncp = new_ncp; 3301 _cache_link_parent(ncp, par_nch->ncp, nchpp); 3302 spin_unlock(&nchpp->spin); 3303 _cache_unlock(par_nch->ncp); 3304 /* par_locked = 0 - not used */ 3305 found: 3306 /* 3307 * stats and namecache size management 3308 */ 3309 if (ncp->nc_flag & NCF_UNRESOLVED) 3310 ++gd->gd_nchstats->ncs_miss; 3311 else if (ncp->nc_vp) 3312 ++gd->gd_nchstats->ncs_goodhits; 3313 else 3314 ++gd->gd_nchstats->ncs_neghits; 3315 nch.mount = mp; 3316 nch.ncp = ncp; 3317 _cache_mntref(nch.mount); 3318 3319 return(nch); 3320 failed: 3321 if (new_ncp) { 3322 _cache_free(new_ncp); 3323 new_ncp = NULL; 3324 } 3325 nch.mount = NULL; 3326 nch.ncp = NULL; 3327 return(nch); 3328 } 3329 3330 /* 3331 * The namecache entry is marked as being used as a mount point. 3332 * Locate the mount if it is visible to the caller. The DragonFly 3333 * mount system allows arbitrary loops in the topology and disentangles 3334 * those loops by matching against (mp, ncp) rather than just (ncp). 3335 * This means any given ncp can dive any number of mounts, depending 3336 * on the relative mount (e.g. nullfs) the caller is at in the topology. 3337 * 3338 * We use a very simple frontend cache to reduce SMP conflicts, 3339 * which we have to do because the mountlist scan needs an exclusive 3340 * lock around its ripout info list. Not to mention that there might 3341 * be a lot of mounts. 3342 */ 3343 struct findmount_info { 3344 struct mount *result; 3345 struct mount *nch_mount; 3346 struct namecache *nch_ncp; 3347 }; 3348 3349 static 3350 struct ncmount_cache * 3351 ncmount_cache_lookup(struct mount *mp, struct namecache *ncp) 3352 { 3353 int hash; 3354 3355 hash = ((int)(intptr_t)mp / sizeof(*mp)) ^ 3356 ((int)(intptr_t)ncp / sizeof(*ncp)); 3357 hash = (hash & 0x7FFFFFFF) % NCMOUNT_NUMCACHE; 3358 return (&ncmount_cache[hash]); 3359 } 3360 3361 static 3362 int 3363 cache_findmount_callback(struct mount *mp, void *data) 3364 { 3365 struct findmount_info *info = data; 3366 3367 /* 3368 * Check the mount's mounted-on point against the passed nch. 3369 */ 3370 if (mp->mnt_ncmounton.mount == info->nch_mount && 3371 mp->mnt_ncmounton.ncp == info->nch_ncp 3372 ) { 3373 info->result = mp; 3374 _cache_mntref(mp); 3375 return(-1); 3376 } 3377 return(0); 3378 } 3379 3380 struct mount * 3381 cache_findmount(struct nchandle *nch) 3382 { 3383 struct findmount_info info; 3384 struct ncmount_cache *ncc; 3385 struct mount *mp; 3386 3387 /* 3388 * Fast 3389 */ 3390 if (ncmount_cache_enable == 0) { 3391 ncc = NULL; 3392 goto skip; 3393 } 3394 ncc = ncmount_cache_lookup(nch->mount, nch->ncp); 3395 if (ncc->ncp == nch->ncp) { 3396 spin_lock_shared(&ncc->spin); 3397 if (ncc->isneg == 0 && 3398 ncc->ncp == nch->ncp && (mp = ncc->mp) != NULL) { 3399 if (mp->mnt_ncmounton.mount == nch->mount && 3400 mp->mnt_ncmounton.ncp == nch->ncp) { 3401 /* 3402 * Cache hit (positive) 3403 */ 3404 _cache_mntref(mp); 3405 spin_unlock_shared(&ncc->spin); 3406 ++ncmount_cache_hit; 3407 return(mp); 3408 } 3409 /* else cache miss */ 3410 } 3411 if (ncc->isneg && 3412 ncc->ncp == nch->ncp && ncc->mp == nch->mount) { 3413 /* 3414 * Cache hit (negative) 3415 */ 3416 spin_unlock_shared(&ncc->spin); 3417 ++ncmount_cache_hit; 3418 return(NULL); 3419 } 3420 spin_unlock_shared(&ncc->spin); 3421 } 3422 skip: 3423 3424 /* 3425 * Slow 3426 */ 3427 info.result = NULL; 3428 info.nch_mount = nch->mount; 3429 info.nch_ncp = nch->ncp; 3430 mountlist_scan(cache_findmount_callback, &info, 3431 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 3432 3433 /* 3434 * Cache the result. 3435 * 3436 * Negative lookups: We cache the originating {ncp,mp}. (mp) is 3437 * only used for pointer comparisons and is not 3438 * referenced (otherwise there would be dangling 3439 * refs). 3440 * 3441 * Positive lookups: We cache the originating {ncp} and the target 3442 * (mp). (mp) is referenced. 3443 * 3444 * Indeterminant: If the match is undergoing an unmount we do 3445 * not cache it to avoid racing cache_unmounting(), 3446 * but still return the match. 3447 */ 3448 if (ncc) { 3449 spin_lock(&ncc->spin); 3450 if (info.result == NULL) { 3451 if (ncc->isneg == 0 && ncc->mp) 3452 _cache_mntrel(ncc->mp); 3453 ncc->ncp = nch->ncp; 3454 ncc->mp = nch->mount; 3455 ncc->isneg = 1; 3456 spin_unlock(&ncc->spin); 3457 ++ncmount_cache_overwrite; 3458 } else if ((info.result->mnt_kern_flag & MNTK_UNMOUNT) == 0) { 3459 if (ncc->isneg == 0 && ncc->mp) 3460 _cache_mntrel(ncc->mp); 3461 _cache_mntref(info.result); 3462 ncc->ncp = nch->ncp; 3463 ncc->mp = info.result; 3464 ncc->isneg = 0; 3465 spin_unlock(&ncc->spin); 3466 ++ncmount_cache_overwrite; 3467 } else { 3468 spin_unlock(&ncc->spin); 3469 } 3470 ++ncmount_cache_miss; 3471 } 3472 return(info.result); 3473 } 3474 3475 void 3476 cache_dropmount(struct mount *mp) 3477 { 3478 _cache_mntrel(mp); 3479 } 3480 3481 void 3482 cache_ismounting(struct mount *mp) 3483 { 3484 struct nchandle *nch = &mp->mnt_ncmounton; 3485 struct ncmount_cache *ncc; 3486 3487 ncc = ncmount_cache_lookup(nch->mount, nch->ncp); 3488 if (ncc->isneg && 3489 ncc->ncp == nch->ncp && ncc->mp == nch->mount) { 3490 spin_lock(&ncc->spin); 3491 if (ncc->isneg && 3492 ncc->ncp == nch->ncp && ncc->mp == nch->mount) { 3493 ncc->ncp = NULL; 3494 ncc->mp = NULL; 3495 } 3496 spin_unlock(&ncc->spin); 3497 } 3498 } 3499 3500 void 3501 cache_unmounting(struct mount *mp) 3502 { 3503 struct nchandle *nch = &mp->mnt_ncmounton; 3504 struct ncmount_cache *ncc; 3505 3506 ncc = ncmount_cache_lookup(nch->mount, nch->ncp); 3507 if (ncc->isneg == 0 && 3508 ncc->ncp == nch->ncp && ncc->mp == mp) { 3509 spin_lock(&ncc->spin); 3510 if (ncc->isneg == 0 && 3511 ncc->ncp == nch->ncp && ncc->mp == mp) { 3512 _cache_mntrel(mp); 3513 ncc->ncp = NULL; 3514 ncc->mp = NULL; 3515 } 3516 spin_unlock(&ncc->spin); 3517 } 3518 } 3519 3520 /* 3521 * Resolve an unresolved namecache entry, generally by looking it up. 3522 * The passed ncp must be locked and refd. 3523 * 3524 * Theoretically since a vnode cannot be recycled while held, and since 3525 * the nc_parent chain holds its vnode as long as children exist, the 3526 * direct parent of the cache entry we are trying to resolve should 3527 * have a valid vnode. If not then generate an error that we can 3528 * determine is related to a resolver bug. 3529 * 3530 * However, if a vnode was in the middle of a recyclement when the NCP 3531 * got locked, ncp->nc_vp might point to a vnode that is about to become 3532 * invalid. cache_resolve() handles this case by unresolving the entry 3533 * and then re-resolving it. 3534 * 3535 * Note that successful resolution does not necessarily return an error 3536 * code of 0. If the ncp resolves to a negative cache hit then ENOENT 3537 * will be returned. 3538 */ 3539 int 3540 cache_resolve(struct nchandle *nch, struct ucred *cred) 3541 { 3542 struct namecache *par_tmp; 3543 struct namecache *par; 3544 struct namecache *ncp; 3545 struct nchandle nctmp; 3546 struct mount *mp; 3547 struct vnode *dvp; 3548 int error; 3549 3550 ncp = nch->ncp; 3551 mp = nch->mount; 3552 KKASSERT(_cache_lockstatus(ncp) == LK_EXCLUSIVE); 3553 restart: 3554 /* 3555 * If the ncp is already resolved we have nothing to do. However, 3556 * we do want to guarentee that a usable vnode is returned when 3557 * a vnode is present, so make sure it hasn't been reclaimed. 3558 */ 3559 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 3560 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 3561 _cache_setunresolved(ncp); 3562 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 3563 return (ncp->nc_error); 3564 } 3565 3566 /* 3567 * If the ncp was destroyed it will never resolve again. This 3568 * can basically only happen when someone is chdir'd into an 3569 * empty directory which is then rmdir'd. We want to catch this 3570 * here and not dive the VFS because the VFS might actually 3571 * have a way to re-resolve the disconnected ncp, which will 3572 * result in inconsistencies in the cdir/nch for proc->p_fd. 3573 */ 3574 if (ncp->nc_flag & NCF_DESTROYED) 3575 return(EINVAL); 3576 3577 /* 3578 * Mount points need special handling because the parent does not 3579 * belong to the same filesystem as the ncp. 3580 */ 3581 if (ncp == mp->mnt_ncmountpt.ncp) 3582 return (cache_resolve_mp(mp)); 3583 3584 /* 3585 * We expect an unbroken chain of ncps to at least the mount point, 3586 * and even all the way to root (but this code doesn't have to go 3587 * past the mount point). 3588 */ 3589 if (ncp->nc_parent == NULL) { 3590 kprintf("EXDEV case 1 %p %*.*s\n", ncp, 3591 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 3592 ncp->nc_error = EXDEV; 3593 return(ncp->nc_error); 3594 } 3595 3596 /* 3597 * The vp's of the parent directories in the chain are held via vhold() 3598 * due to the existance of the child, and should not disappear. 3599 * However, there are cases where they can disappear: 3600 * 3601 * - due to filesystem I/O errors. 3602 * - due to NFS being stupid about tracking the namespace and 3603 * destroys the namespace for entire directories quite often. 3604 * - due to forced unmounts. 3605 * - due to an rmdir (parent will be marked DESTROYED) 3606 * 3607 * When this occurs we have to track the chain backwards and resolve 3608 * it, looping until the resolver catches up to the current node. We 3609 * could recurse here but we might run ourselves out of kernel stack 3610 * so we do it in a more painful manner. This situation really should 3611 * not occur all that often, or if it does not have to go back too 3612 * many nodes to resolve the ncp. 3613 */ 3614 while ((dvp = cache_dvpref(ncp)) == NULL) { 3615 /* 3616 * This case can occur if a process is CD'd into a 3617 * directory which is then rmdir'd. If the parent is marked 3618 * destroyed there is no point trying to resolve it. 3619 */ 3620 if (ncp->nc_parent->nc_flag & NCF_DESTROYED) 3621 return(ENOENT); 3622 par = ncp->nc_parent; 3623 _cache_hold(par); 3624 _cache_lock(par); 3625 while ((par_tmp = par->nc_parent) != NULL && 3626 par_tmp->nc_vp == NULL) { 3627 _cache_hold(par_tmp); 3628 _cache_lock(par_tmp); 3629 _cache_put(par); 3630 par = par_tmp; 3631 } 3632 if (par->nc_parent == NULL) { 3633 kprintf("EXDEV case 2 %*.*s\n", 3634 par->nc_nlen, par->nc_nlen, par->nc_name); 3635 _cache_put(par); 3636 return (EXDEV); 3637 } 3638 /* 3639 * The parent is not set in stone, ref and lock it to prevent 3640 * it from disappearing. Also note that due to renames it 3641 * is possible for our ncp to move and for par to no longer 3642 * be one of its parents. We resolve it anyway, the loop 3643 * will handle any moves. 3644 */ 3645 _cache_get(par); /* additional hold/lock */ 3646 _cache_put(par); /* from earlier hold/lock */ 3647 if (par == nch->mount->mnt_ncmountpt.ncp) { 3648 cache_resolve_mp(nch->mount); 3649 } else if ((dvp = cache_dvpref(par)) == NULL) { 3650 kprintf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name); 3651 _cache_put(par); 3652 continue; 3653 } else { 3654 if (par->nc_flag & NCF_UNRESOLVED) { 3655 nctmp.mount = mp; 3656 nctmp.ncp = par; 3657 par->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred); 3658 } 3659 vrele(dvp); 3660 } 3661 if ((error = par->nc_error) != 0) { 3662 if (par->nc_error != EAGAIN) { 3663 kprintf("EXDEV case 3 %*.*s error %d\n", 3664 par->nc_nlen, par->nc_nlen, par->nc_name, 3665 par->nc_error); 3666 _cache_put(par); 3667 return(error); 3668 } 3669 kprintf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n", 3670 par, par->nc_nlen, par->nc_nlen, par->nc_name); 3671 } 3672 _cache_put(par); 3673 /* loop */ 3674 } 3675 3676 /* 3677 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected 3678 * ncp's and reattach them. If this occurs the original ncp is marked 3679 * EAGAIN to force a relookup. 3680 * 3681 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed 3682 * ncp must already be resolved. 3683 */ 3684 if (dvp) { 3685 nctmp.mount = mp; 3686 nctmp.ncp = ncp; 3687 ncp->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred); 3688 vrele(dvp); 3689 } else { 3690 ncp->nc_error = EPERM; 3691 } 3692 if (ncp->nc_error == EAGAIN) { 3693 kprintf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n", 3694 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 3695 goto restart; 3696 } 3697 return(ncp->nc_error); 3698 } 3699 3700 /* 3701 * Resolve the ncp associated with a mount point. Such ncp's almost always 3702 * remain resolved and this routine is rarely called. NFS MPs tends to force 3703 * re-resolution more often due to its mac-truck-smash-the-namecache 3704 * method of tracking namespace changes. 3705 * 3706 * The semantics for this call is that the passed ncp must be locked on 3707 * entry and will be locked on return. However, if we actually have to 3708 * resolve the mount point we temporarily unlock the entry in order to 3709 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of 3710 * the unlock we have to recheck the flags after we relock. 3711 */ 3712 static int 3713 cache_resolve_mp(struct mount *mp) 3714 { 3715 struct namecache *ncp = mp->mnt_ncmountpt.ncp; 3716 struct vnode *vp; 3717 int error; 3718 3719 KKASSERT(mp != NULL); 3720 3721 /* 3722 * If the ncp is already resolved we have nothing to do. However, 3723 * we do want to guarentee that a usable vnode is returned when 3724 * a vnode is present, so make sure it hasn't been reclaimed. 3725 */ 3726 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 3727 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 3728 _cache_setunresolved(ncp); 3729 } 3730 3731 if (ncp->nc_flag & NCF_UNRESOLVED) { 3732 _cache_unlock(ncp); 3733 while (vfs_busy(mp, 0)) 3734 ; 3735 error = VFS_ROOT(mp, &vp); 3736 _cache_lock(ncp); 3737 3738 /* 3739 * recheck the ncp state after relocking. 3740 */ 3741 if (ncp->nc_flag & NCF_UNRESOLVED) { 3742 ncp->nc_error = error; 3743 if (error == 0) { 3744 _cache_setvp(mp, ncp, vp); 3745 vput(vp); 3746 } else { 3747 kprintf("[diagnostic] cache_resolve_mp: failed" 3748 " to resolve mount %p err=%d ncp=%p\n", 3749 mp, error, ncp); 3750 _cache_setvp(mp, ncp, NULL); 3751 } 3752 } else if (error == 0) { 3753 vput(vp); 3754 } 3755 vfs_unbusy(mp); 3756 } 3757 return(ncp->nc_error); 3758 } 3759 3760 /* 3761 * Clean out negative cache entries when too many have accumulated. 3762 */ 3763 static void 3764 _cache_cleanneg(int count) 3765 { 3766 struct namecache *ncp; 3767 3768 /* 3769 * Attempt to clean out the specified number of negative cache 3770 * entries. 3771 */ 3772 while (count) { 3773 spin_lock(&ncspin); 3774 ncp = TAILQ_FIRST(&ncneglist); 3775 if (ncp == NULL) { 3776 spin_unlock(&ncspin); 3777 break; 3778 } 3779 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 3780 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 3781 _cache_hold(ncp); 3782 spin_unlock(&ncspin); 3783 3784 /* 3785 * This can race, so we must re-check that the ncp 3786 * is on the ncneglist after successfully locking it. 3787 */ 3788 if (_cache_lock_special(ncp) == 0) { 3789 if (ncp->nc_vp == NULL && 3790 (ncp->nc_flag & NCF_UNRESOLVED) == 0) { 3791 ncp = cache_zap(ncp, 1); 3792 if (ncp) 3793 _cache_drop(ncp); 3794 } else { 3795 kprintf("cache_cleanneg: race avoided\n"); 3796 _cache_unlock(ncp); 3797 } 3798 } else { 3799 _cache_drop(ncp); 3800 } 3801 --count; 3802 } 3803 } 3804 3805 /* 3806 * Clean out positive cache entries when too many have accumulated. 3807 */ 3808 static void 3809 _cache_cleanpos(int count) 3810 { 3811 static volatile int rover; 3812 struct nchash_head *nchpp; 3813 struct namecache *ncp; 3814 int rover_copy; 3815 3816 /* 3817 * Attempt to clean out the specified number of negative cache 3818 * entries. 3819 */ 3820 while (count) { 3821 rover_copy = ++rover; /* MPSAFEENOUGH */ 3822 cpu_ccfence(); 3823 nchpp = NCHHASH(rover_copy); 3824 3825 spin_lock_shared(&nchpp->spin); 3826 ncp = LIST_FIRST(&nchpp->list); 3827 while (ncp && (ncp->nc_flag & NCF_DESTROYED)) 3828 ncp = LIST_NEXT(ncp, nc_hash); 3829 if (ncp) 3830 _cache_hold(ncp); 3831 spin_unlock_shared(&nchpp->spin); 3832 3833 if (ncp) { 3834 if (_cache_lock_special(ncp) == 0) { 3835 ncp = cache_zap(ncp, 1); 3836 if (ncp) 3837 _cache_drop(ncp); 3838 } else { 3839 _cache_drop(ncp); 3840 } 3841 } 3842 --count; 3843 } 3844 } 3845 3846 /* 3847 * This is a kitchen sink function to clean out ncps which we 3848 * tried to zap from cache_drop() but failed because we were 3849 * unable to acquire the parent lock. 3850 * 3851 * Such entries can also be removed via cache_inval_vp(), such 3852 * as when unmounting. 3853 */ 3854 static void 3855 _cache_cleandefered(void) 3856 { 3857 struct nchash_head *nchpp; 3858 struct namecache *ncp; 3859 struct namecache dummy; 3860 int i; 3861 3862 numdefered = 0; 3863 bzero(&dummy, sizeof(dummy)); 3864 dummy.nc_flag = NCF_DESTROYED; 3865 dummy.nc_refs = 1; 3866 3867 for (i = 0; i <= nchash; ++i) { 3868 nchpp = &nchashtbl[i]; 3869 3870 spin_lock(&nchpp->spin); 3871 LIST_INSERT_HEAD(&nchpp->list, &dummy, nc_hash); 3872 ncp = &dummy; 3873 while ((ncp = LIST_NEXT(ncp, nc_hash)) != NULL) { 3874 if ((ncp->nc_flag & NCF_DEFEREDZAP) == 0) 3875 continue; 3876 LIST_REMOVE(&dummy, nc_hash); 3877 LIST_INSERT_AFTER(ncp, &dummy, nc_hash); 3878 _cache_hold(ncp); 3879 spin_unlock(&nchpp->spin); 3880 if (_cache_lock_nonblock(ncp) == 0) { 3881 ncp->nc_flag &= ~NCF_DEFEREDZAP; 3882 _cache_unlock(ncp); 3883 } 3884 _cache_drop(ncp); 3885 spin_lock(&nchpp->spin); 3886 ncp = &dummy; 3887 } 3888 LIST_REMOVE(&dummy, nc_hash); 3889 spin_unlock(&nchpp->spin); 3890 } 3891 } 3892 3893 /* 3894 * Name cache initialization, from vfsinit() when we are booting 3895 */ 3896 void 3897 nchinit(void) 3898 { 3899 int i; 3900 globaldata_t gd; 3901 3902 /* 3903 * Initialise per-cpu namecache effectiveness statistics. 3904 */ 3905 for (i = 0; i < ncpus; ++i) { 3906 gd = globaldata_find(i); 3907 gd->gd_nchstats = &nchstats[i]; 3908 } 3909 3910 /* 3911 * Create a generous namecache hash table 3912 */ 3913 TAILQ_INIT(&ncneglist); 3914 spin_init(&ncspin, "nchinit"); 3915 nchashtbl = hashinit_ext(vfs_inodehashsize(), 3916 sizeof(struct nchash_head), 3917 M_VFSCACHE, &nchash); 3918 for (i = 0; i <= (int)nchash; ++i) { 3919 LIST_INIT(&nchashtbl[i].list); 3920 spin_init(&nchashtbl[i].spin, "nchinit_hash"); 3921 } 3922 for (i = 0; i < NCMOUNT_NUMCACHE; ++i) 3923 spin_init(&ncmount_cache[i].spin, "nchinit_cache"); 3924 nclockwarn = 5 * hz; 3925 } 3926 3927 /* 3928 * Called from start_init() to bootstrap the root filesystem. Returns 3929 * a referenced, unlocked namecache record. 3930 */ 3931 void 3932 cache_allocroot(struct nchandle *nch, struct mount *mp, struct vnode *vp) 3933 { 3934 nch->ncp = cache_alloc(0); 3935 nch->mount = mp; 3936 _cache_mntref(mp); 3937 if (vp) 3938 _cache_setvp(nch->mount, nch->ncp, vp); 3939 } 3940 3941 /* 3942 * vfs_cache_setroot() 3943 * 3944 * Create an association between the root of our namecache and 3945 * the root vnode. This routine may be called several times during 3946 * booting. 3947 * 3948 * If the caller intends to save the returned namecache pointer somewhere 3949 * it must cache_hold() it. 3950 */ 3951 void 3952 vfs_cache_setroot(struct vnode *nvp, struct nchandle *nch) 3953 { 3954 struct vnode *ovp; 3955 struct nchandle onch; 3956 3957 ovp = rootvnode; 3958 onch = rootnch; 3959 rootvnode = nvp; 3960 if (nch) 3961 rootnch = *nch; 3962 else 3963 cache_zero(&rootnch); 3964 if (ovp) 3965 vrele(ovp); 3966 if (onch.ncp) 3967 cache_drop(&onch); 3968 } 3969 3970 /* 3971 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache 3972 * topology and is being removed as quickly as possible. The new VOP_N*() 3973 * API calls are required to make specific adjustments using the supplied 3974 * ncp pointers rather then just bogusly purging random vnodes. 3975 * 3976 * Invalidate all namecache entries to a particular vnode as well as 3977 * any direct children of that vnode in the namecache. This is a 3978 * 'catch all' purge used by filesystems that do not know any better. 3979 * 3980 * Note that the linkage between the vnode and its namecache entries will 3981 * be removed, but the namecache entries themselves might stay put due to 3982 * active references from elsewhere in the system or due to the existance of 3983 * the children. The namecache topology is left intact even if we do not 3984 * know what the vnode association is. Such entries will be marked 3985 * NCF_UNRESOLVED. 3986 */ 3987 void 3988 cache_purge(struct vnode *vp) 3989 { 3990 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN); 3991 } 3992 3993 static int disablecwd; 3994 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, 3995 "Disable getcwd"); 3996 3997 static u_long numcwdcalls; 3998 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdcalls, CTLFLAG_RD, &numcwdcalls, 0, 3999 "Number of current directory resolution calls"); 4000 static u_long numcwdfailnf; 4001 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailnf, CTLFLAG_RD, &numcwdfailnf, 0, 4002 "Number of current directory failures due to lack of file"); 4003 static u_long numcwdfailsz; 4004 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfailsz, CTLFLAG_RD, &numcwdfailsz, 0, 4005 "Number of current directory failures due to large result"); 4006 static u_long numcwdfound; 4007 SYSCTL_ULONG(_vfs_cache, OID_AUTO, numcwdfound, CTLFLAG_RD, &numcwdfound, 0, 4008 "Number of current directory resolution successes"); 4009 4010 /* 4011 * MPALMOSTSAFE 4012 */ 4013 int 4014 sys___getcwd(struct __getcwd_args *uap) 4015 { 4016 u_int buflen; 4017 int error; 4018 char *buf; 4019 char *bp; 4020 4021 if (disablecwd) 4022 return (ENODEV); 4023 4024 buflen = uap->buflen; 4025 if (buflen == 0) 4026 return (EINVAL); 4027 if (buflen > MAXPATHLEN) 4028 buflen = MAXPATHLEN; 4029 4030 buf = kmalloc(buflen, M_TEMP, M_WAITOK); 4031 bp = kern_getcwd(buf, buflen, &error); 4032 if (error == 0) 4033 error = copyout(bp, uap->buf, strlen(bp) + 1); 4034 kfree(buf, M_TEMP); 4035 return (error); 4036 } 4037 4038 char * 4039 kern_getcwd(char *buf, size_t buflen, int *error) 4040 { 4041 struct proc *p = curproc; 4042 char *bp; 4043 int i, slash_prefixed; 4044 struct filedesc *fdp; 4045 struct nchandle nch; 4046 struct namecache *ncp; 4047 4048 numcwdcalls++; 4049 bp = buf; 4050 bp += buflen - 1; 4051 *bp = '\0'; 4052 fdp = p->p_fd; 4053 slash_prefixed = 0; 4054 4055 nch = fdp->fd_ncdir; 4056 ncp = nch.ncp; 4057 if (ncp) 4058 _cache_hold(ncp); 4059 4060 while (ncp && (ncp != fdp->fd_nrdir.ncp || 4061 nch.mount != fdp->fd_nrdir.mount) 4062 ) { 4063 /* 4064 * While traversing upwards if we encounter the root 4065 * of the current mount we have to skip to the mount point 4066 * in the underlying filesystem. 4067 */ 4068 if (ncp == nch.mount->mnt_ncmountpt.ncp) { 4069 nch = nch.mount->mnt_ncmounton; 4070 _cache_drop(ncp); 4071 ncp = nch.ncp; 4072 if (ncp) 4073 _cache_hold(ncp); 4074 continue; 4075 } 4076 4077 /* 4078 * Prepend the path segment 4079 */ 4080 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 4081 if (bp == buf) { 4082 numcwdfailsz++; 4083 *error = ERANGE; 4084 bp = NULL; 4085 goto done; 4086 } 4087 *--bp = ncp->nc_name[i]; 4088 } 4089 if (bp == buf) { 4090 numcwdfailsz++; 4091 *error = ERANGE; 4092 bp = NULL; 4093 goto done; 4094 } 4095 *--bp = '/'; 4096 slash_prefixed = 1; 4097 4098 /* 4099 * Go up a directory. This isn't a mount point so we don't 4100 * have to check again. 4101 */ 4102 while ((nch.ncp = ncp->nc_parent) != NULL) { 4103 if (ncp_shared_lock_disable) 4104 _cache_lock(ncp); 4105 else 4106 _cache_lock_shared(ncp); 4107 if (nch.ncp != ncp->nc_parent) { 4108 _cache_unlock(ncp); 4109 continue; 4110 } 4111 _cache_hold(nch.ncp); 4112 _cache_unlock(ncp); 4113 break; 4114 } 4115 _cache_drop(ncp); 4116 ncp = nch.ncp; 4117 } 4118 if (ncp == NULL) { 4119 numcwdfailnf++; 4120 *error = ENOENT; 4121 bp = NULL; 4122 goto done; 4123 } 4124 if (!slash_prefixed) { 4125 if (bp == buf) { 4126 numcwdfailsz++; 4127 *error = ERANGE; 4128 bp = NULL; 4129 goto done; 4130 } 4131 *--bp = '/'; 4132 } 4133 numcwdfound++; 4134 *error = 0; 4135 done: 4136 if (ncp) 4137 _cache_drop(ncp); 4138 return (bp); 4139 } 4140 4141 /* 4142 * Thus begins the fullpath magic. 4143 * 4144 * The passed nchp is referenced but not locked. 4145 */ 4146 static int disablefullpath; 4147 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, 4148 &disablefullpath, 0, 4149 "Disable fullpath lookups"); 4150 4151 static u_int numfullpathcalls; 4152 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathcalls, CTLFLAG_RD, 4153 &numfullpathcalls, 0, 4154 "Number of full path resolutions in progress"); 4155 static u_int numfullpathfailnf; 4156 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailnf, CTLFLAG_RD, 4157 &numfullpathfailnf, 0, 4158 "Number of full path resolution failures due to lack of file"); 4159 static u_int numfullpathfailsz; 4160 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfailsz, CTLFLAG_RD, 4161 &numfullpathfailsz, 0, 4162 "Number of full path resolution failures due to insufficient memory"); 4163 static u_int numfullpathfound; 4164 SYSCTL_UINT(_vfs_cache, OID_AUTO, numfullpathfound, CTLFLAG_RD, 4165 &numfullpathfound, 0, 4166 "Number of full path resolution successes"); 4167 4168 int 4169 cache_fullpath(struct proc *p, struct nchandle *nchp, struct nchandle *nchbase, 4170 char **retbuf, char **freebuf, int guess) 4171 { 4172 struct nchandle fd_nrdir; 4173 struct nchandle nch; 4174 struct namecache *ncp; 4175 struct mount *mp, *new_mp; 4176 char *bp, *buf; 4177 int slash_prefixed; 4178 int error = 0; 4179 int i; 4180 4181 atomic_add_int(&numfullpathcalls, -1); 4182 4183 *retbuf = NULL; 4184 *freebuf = NULL; 4185 4186 buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK); 4187 bp = buf + MAXPATHLEN - 1; 4188 *bp = '\0'; 4189 if (nchbase) 4190 fd_nrdir = *nchbase; 4191 else if (p != NULL) 4192 fd_nrdir = p->p_fd->fd_nrdir; 4193 else 4194 fd_nrdir = rootnch; 4195 slash_prefixed = 0; 4196 nch = *nchp; 4197 ncp = nch.ncp; 4198 if (ncp) 4199 _cache_hold(ncp); 4200 mp = nch.mount; 4201 4202 while (ncp && (ncp != fd_nrdir.ncp || mp != fd_nrdir.mount)) { 4203 new_mp = NULL; 4204 4205 /* 4206 * If we are asked to guess the upwards path, we do so whenever 4207 * we encounter an ncp marked as a mountpoint. We try to find 4208 * the actual mountpoint by finding the mountpoint with this 4209 * ncp. 4210 */ 4211 if (guess && (ncp->nc_flag & NCF_ISMOUNTPT)) { 4212 new_mp = mount_get_by_nc(ncp); 4213 } 4214 /* 4215 * While traversing upwards if we encounter the root 4216 * of the current mount we have to skip to the mount point. 4217 */ 4218 if (ncp == mp->mnt_ncmountpt.ncp) { 4219 new_mp = mp; 4220 } 4221 if (new_mp) { 4222 nch = new_mp->mnt_ncmounton; 4223 _cache_drop(ncp); 4224 ncp = nch.ncp; 4225 if (ncp) 4226 _cache_hold(ncp); 4227 mp = nch.mount; 4228 continue; 4229 } 4230 4231 /* 4232 * Prepend the path segment 4233 */ 4234 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 4235 if (bp == buf) { 4236 numfullpathfailsz++; 4237 kfree(buf, M_TEMP); 4238 error = ENOMEM; 4239 goto done; 4240 } 4241 *--bp = ncp->nc_name[i]; 4242 } 4243 if (bp == buf) { 4244 numfullpathfailsz++; 4245 kfree(buf, M_TEMP); 4246 error = ENOMEM; 4247 goto done; 4248 } 4249 *--bp = '/'; 4250 slash_prefixed = 1; 4251 4252 /* 4253 * Go up a directory. This isn't a mount point so we don't 4254 * have to check again. 4255 * 4256 * We can only safely access nc_parent with ncp held locked. 4257 */ 4258 while ((nch.ncp = ncp->nc_parent) != NULL) { 4259 _cache_lock(ncp); 4260 if (nch.ncp != ncp->nc_parent) { 4261 _cache_unlock(ncp); 4262 continue; 4263 } 4264 _cache_hold(nch.ncp); 4265 _cache_unlock(ncp); 4266 break; 4267 } 4268 _cache_drop(ncp); 4269 ncp = nch.ncp; 4270 } 4271 if (ncp == NULL) { 4272 numfullpathfailnf++; 4273 kfree(buf, M_TEMP); 4274 error = ENOENT; 4275 goto done; 4276 } 4277 4278 if (!slash_prefixed) { 4279 if (bp == buf) { 4280 numfullpathfailsz++; 4281 kfree(buf, M_TEMP); 4282 error = ENOMEM; 4283 goto done; 4284 } 4285 *--bp = '/'; 4286 } 4287 numfullpathfound++; 4288 *retbuf = bp; 4289 *freebuf = buf; 4290 error = 0; 4291 done: 4292 if (ncp) 4293 _cache_drop(ncp); 4294 return(error); 4295 } 4296 4297 int 4298 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, 4299 char **freebuf, int guess) 4300 { 4301 struct namecache *ncp; 4302 struct nchandle nch; 4303 int error; 4304 4305 *freebuf = NULL; 4306 atomic_add_int(&numfullpathcalls, 1); 4307 if (disablefullpath) 4308 return (ENODEV); 4309 4310 if (p == NULL) 4311 return (EINVAL); 4312 4313 /* vn is NULL, client wants us to use p->p_textvp */ 4314 if (vn == NULL) { 4315 if ((vn = p->p_textvp) == NULL) 4316 return (EINVAL); 4317 } 4318 spin_lock_shared(&vn->v_spin); 4319 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) { 4320 if (ncp->nc_nlen) 4321 break; 4322 } 4323 if (ncp == NULL) { 4324 spin_unlock_shared(&vn->v_spin); 4325 return (EINVAL); 4326 } 4327 _cache_hold(ncp); 4328 spin_unlock_shared(&vn->v_spin); 4329 4330 atomic_add_int(&numfullpathcalls, -1); 4331 nch.ncp = ncp; 4332 nch.mount = vn->v_mount; 4333 error = cache_fullpath(p, &nch, NULL, retbuf, freebuf, guess); 4334 _cache_drop(ncp); 4335 return (error); 4336 } 4337