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. All advertising materials mentioning features or use of this software 49 * must display the following acknowledgement: 50 * This product includes software developed by the University of 51 * California, Berkeley and its contributors. 52 * 4. Neither the name of the University nor the names of its contributors 53 * may be used to endorse or promote products derived from this software 54 * without specific prior written permission. 55 * 56 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 57 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 58 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 59 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 60 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 61 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 62 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 63 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 64 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 65 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 66 * SUCH DAMAGE. 67 */ 68 69 #include <sys/param.h> 70 #include <sys/systm.h> 71 #include <sys/kernel.h> 72 #include <sys/sysctl.h> 73 #include <sys/mount.h> 74 #include <sys/vnode.h> 75 #include <sys/malloc.h> 76 #include <sys/sysproto.h> 77 #include <sys/spinlock.h> 78 #include <sys/proc.h> 79 #include <sys/namei.h> 80 #include <sys/nlookup.h> 81 #include <sys/filedesc.h> 82 #include <sys/fnv_hash.h> 83 #include <sys/globaldata.h> 84 #include <sys/kern_syscall.h> 85 #include <sys/dirent.h> 86 #include <ddb/ddb.h> 87 88 #include <sys/sysref2.h> 89 #include <sys/spinlock2.h> 90 #include <sys/mplock2.h> 91 92 #define MAX_RECURSION_DEPTH 64 93 94 /* 95 * Random lookups in the cache are accomplished with a hash table using 96 * a hash key of (nc_src_vp, name). Each hash chain has its own spin lock. 97 * 98 * Negative entries may exist and correspond to resolved namecache 99 * structures where nc_vp is NULL. In a negative entry, NCF_WHITEOUT 100 * will be set if the entry corresponds to a whited-out directory entry 101 * (verses simply not finding the entry at all). ncneglist is locked 102 * with a global spinlock (ncspin). 103 * 104 * MPSAFE RULES: 105 * 106 * (1) A ncp must be referenced before it can be locked. 107 * 108 * (2) A ncp must be locked in order to modify it. 109 * 110 * (3) ncp locks are always ordered child -> parent. That may seem 111 * backwards but forward scans use the hash table and thus can hold 112 * the parent unlocked when traversing downward. 113 * 114 * This allows insert/rename/delete/dot-dot and other operations 115 * to use ncp->nc_parent links. 116 * 117 * This also prevents a locked up e.g. NFS node from creating a 118 * chain reaction all the way back to the root vnode / namecache. 119 * 120 * (4) parent linkages require both the parent and child to be locked. 121 */ 122 123 /* 124 * Structures associated with name cacheing. 125 */ 126 #define NCHHASH(hash) (&nchashtbl[(hash) & nchash]) 127 #define MINNEG 1024 128 129 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries"); 130 131 LIST_HEAD(nchash_list, namecache); 132 133 struct nchash_head { 134 struct nchash_list list; 135 struct spinlock spin; 136 }; 137 138 static struct nchash_head *nchashtbl; 139 static struct namecache_list ncneglist; 140 static struct spinlock ncspin; 141 142 /* 143 * ncvp_debug - debug cache_fromvp(). This is used by the NFS server 144 * to create the namecache infrastructure leading to a dangling vnode. 145 * 146 * 0 Only errors are reported 147 * 1 Successes are reported 148 * 2 Successes + the whole directory scan is reported 149 * 3 Force the directory scan code run as if the parent vnode did not 150 * have a namecache record, even if it does have one. 151 */ 152 static int ncvp_debug; 153 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0, ""); 154 155 static u_long nchash; /* size of hash table */ 156 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, ""); 157 158 static int ncnegfactor = 16; /* ratio of negative entries */ 159 SYSCTL_INT(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0, ""); 160 161 static int nclockwarn; /* warn on locked entries in ticks */ 162 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0, ""); 163 164 static int numdefered; /* number of cache entries allocated */ 165 SYSCTL_INT(_debug, OID_AUTO, numdefered, CTLFLAG_RD, &numdefered, 0, ""); 166 167 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode), ""); 168 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache), ""); 169 170 int cache_mpsafe = 1; 171 SYSCTL_INT(_vfs, OID_AUTO, cache_mpsafe, CTLFLAG_RW, &cache_mpsafe, 0, ""); 172 173 static int cache_resolve_mp(struct mount *mp); 174 static struct vnode *cache_dvpref(struct namecache *ncp); 175 static void _cache_lock(struct namecache *ncp); 176 static void _cache_setunresolved(struct namecache *ncp); 177 static void _cache_cleanneg(int count); 178 static void _cache_cleandefered(void); 179 180 /* 181 * The new name cache statistics 182 */ 183 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics"); 184 #define STATNODE(mode, name, var) \ 185 SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, mode, var, 0, ""); 186 #define STATNODE_INT(mode, name, var) \ 187 SYSCTL_UINT(_vfs_cache, OID_AUTO, name, mode, var, 0, ""); 188 static int numneg; STATNODE_INT(CTLFLAG_RD, numneg, &numneg); 189 static int numcache; STATNODE_INT(CTLFLAG_RD, numcache, &numcache); 190 static u_long numcalls; STATNODE(CTLFLAG_RD, numcalls, &numcalls); 191 static u_long dothits; STATNODE(CTLFLAG_RD, dothits, &dothits); 192 static u_long dotdothits; STATNODE(CTLFLAG_RD, dotdothits, &dotdothits); 193 static u_long numchecks; STATNODE(CTLFLAG_RD, numchecks, &numchecks); 194 static u_long nummiss; STATNODE(CTLFLAG_RD, nummiss, &nummiss); 195 static u_long nummisszap; STATNODE(CTLFLAG_RD, nummisszap, &nummisszap); 196 static u_long numposzaps; STATNODE(CTLFLAG_RD, numposzaps, &numposzaps); 197 static u_long numposhits; STATNODE(CTLFLAG_RD, numposhits, &numposhits); 198 static u_long numnegzaps; STATNODE(CTLFLAG_RD, numnegzaps, &numnegzaps); 199 static u_long numneghits; STATNODE(CTLFLAG_RD, numneghits, &numneghits); 200 201 struct nchstats nchstats[SMP_MAXCPU]; 202 /* 203 * Export VFS cache effectiveness statistics to user-land. 204 * 205 * The statistics are left for aggregation to user-land so 206 * neat things can be achieved, like observing per-CPU cache 207 * distribution. 208 */ 209 static int 210 sysctl_nchstats(SYSCTL_HANDLER_ARGS) 211 { 212 struct globaldata *gd; 213 int i, error; 214 215 error = 0; 216 for (i = 0; i < ncpus; ++i) { 217 gd = globaldata_find(i); 218 if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats), 219 sizeof(struct nchstats)))) 220 break; 221 } 222 223 return (error); 224 } 225 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD, 226 0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics"); 227 228 static struct namecache *cache_zap(struct namecache *ncp, int nonblock); 229 230 /* 231 * Namespace locking. The caller must already hold a reference to the 232 * namecache structure in order to lock/unlock it. This function prevents 233 * the namespace from being created or destroyed by accessors other then 234 * the lock holder. 235 * 236 * Note that holding a locked namecache structure prevents other threads 237 * from making namespace changes (e.g. deleting or creating), prevents 238 * vnode association state changes by other threads, and prevents the 239 * namecache entry from being resolved or unresolved by other threads. 240 * 241 * The lock owner has full authority to associate/disassociate vnodes 242 * and resolve/unresolve the locked ncp. 243 * 244 * The primary lock field is nc_exlocks. nc_locktd is set after the 245 * fact (when locking) or cleared prior to unlocking. 246 * 247 * WARNING! Holding a locked ncp will prevent a vnode from being destroyed 248 * or recycled, but it does NOT help you if the vnode had already 249 * initiated a recyclement. If this is important, use cache_get() 250 * rather then cache_lock() (and deal with the differences in the 251 * way the refs counter is handled). Or, alternatively, make an 252 * unconditional call to cache_validate() or cache_resolve() 253 * after cache_lock() returns. 254 * 255 * MPSAFE 256 */ 257 static 258 void 259 _cache_lock(struct namecache *ncp) 260 { 261 thread_t td; 262 int didwarn; 263 int error; 264 u_int count; 265 266 KKASSERT(ncp->nc_refs != 0); 267 didwarn = 0; 268 td = curthread; 269 270 for (;;) { 271 count = ncp->nc_exlocks; 272 273 if (count == 0) { 274 if (atomic_cmpset_int(&ncp->nc_exlocks, 0, 1)) { 275 /* 276 * The vp associated with a locked ncp must 277 * be held to prevent it from being recycled. 278 * 279 * WARNING! If VRECLAIMED is set the vnode 280 * could already be in the middle of a recycle. 281 * Callers must use cache_vref() or 282 * cache_vget() on the locked ncp to 283 * validate the vp or set the cache entry 284 * to unresolved. 285 * 286 * NOTE! vhold() is allowed if we hold a 287 * lock on the ncp (which we do). 288 */ 289 ncp->nc_locktd = td; 290 if (ncp->nc_vp) 291 vhold(ncp->nc_vp); /* MPSAFE */ 292 break; 293 } 294 /* cmpset failed */ 295 continue; 296 } 297 if (ncp->nc_locktd == td) { 298 if (atomic_cmpset_int(&ncp->nc_exlocks, count, 299 count + 1)) { 300 break; 301 } 302 /* cmpset failed */ 303 continue; 304 } 305 tsleep_interlock(ncp, 0); 306 if (atomic_cmpset_int(&ncp->nc_exlocks, count, 307 count | NC_EXLOCK_REQ) == 0) { 308 /* cmpset failed */ 309 continue; 310 } 311 error = tsleep(ncp, PINTERLOCKED, "clock", nclockwarn); 312 if (error == EWOULDBLOCK) { 313 if (didwarn == 0) { 314 didwarn = ticks; 315 kprintf("[diagnostic] cache_lock: blocked " 316 "on %p", 317 ncp); 318 kprintf(" \"%*.*s\"\n", 319 ncp->nc_nlen, ncp->nc_nlen, 320 ncp->nc_name); 321 } 322 } 323 } 324 if (didwarn) { 325 kprintf("[diagnostic] cache_lock: unblocked %*.*s after " 326 "%d secs\n", 327 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name, 328 (int)(ticks - didwarn) / hz); 329 } 330 } 331 332 /* 333 * NOTE: nc_refs may be zero if the ncp is interlocked by circumstance, 334 * such as the case where one of its children is locked. 335 * 336 * MPSAFE 337 */ 338 static 339 int 340 _cache_lock_nonblock(struct namecache *ncp) 341 { 342 thread_t td; 343 u_int count; 344 345 td = curthread; 346 347 for (;;) { 348 count = ncp->nc_exlocks; 349 350 if (count == 0) { 351 if (atomic_cmpset_int(&ncp->nc_exlocks, 0, 1)) { 352 /* 353 * The vp associated with a locked ncp must 354 * be held to prevent it from being recycled. 355 * 356 * WARNING! If VRECLAIMED is set the vnode 357 * could already be in the middle of a recycle. 358 * Callers must use cache_vref() or 359 * cache_vget() on the locked ncp to 360 * validate the vp or set the cache entry 361 * to unresolved. 362 * 363 * NOTE! vhold() is allowed if we hold a 364 * lock on the ncp (which we do). 365 */ 366 ncp->nc_locktd = td; 367 if (ncp->nc_vp) 368 vhold(ncp->nc_vp); /* MPSAFE */ 369 break; 370 } 371 /* cmpset failed */ 372 continue; 373 } 374 if (ncp->nc_locktd == td) { 375 if (atomic_cmpset_int(&ncp->nc_exlocks, count, 376 count + 1)) { 377 break; 378 } 379 /* cmpset failed */ 380 continue; 381 } 382 return(EWOULDBLOCK); 383 } 384 return(0); 385 } 386 387 /* 388 * Helper function 389 * 390 * NOTE: nc_refs can be 0 (degenerate case during _cache_drop). 391 * 392 * nc_locktd must be NULLed out prior to nc_exlocks getting cleared. 393 * 394 * MPSAFE 395 */ 396 static 397 void 398 _cache_unlock(struct namecache *ncp) 399 { 400 thread_t td __debugvar = curthread; 401 u_int count; 402 403 KKASSERT(ncp->nc_refs >= 0); 404 KKASSERT(ncp->nc_exlocks > 0); 405 KKASSERT(ncp->nc_locktd == td); 406 407 count = ncp->nc_exlocks; 408 if ((count & ~NC_EXLOCK_REQ) == 1) { 409 ncp->nc_locktd = NULL; 410 if (ncp->nc_vp) 411 vdrop(ncp->nc_vp); 412 } 413 for (;;) { 414 if ((count & ~NC_EXLOCK_REQ) == 1) { 415 if (atomic_cmpset_int(&ncp->nc_exlocks, count, 0)) { 416 if (count & NC_EXLOCK_REQ) 417 wakeup(ncp); 418 break; 419 } 420 } else { 421 if (atomic_cmpset_int(&ncp->nc_exlocks, count, 422 count - 1)) { 423 break; 424 } 425 } 426 count = ncp->nc_exlocks; 427 } 428 } 429 430 431 /* 432 * cache_hold() and cache_drop() prevent the premature deletion of a 433 * namecache entry but do not prevent operations (such as zapping) on 434 * that namecache entry. 435 * 436 * This routine may only be called from outside this source module if 437 * nc_refs is already at least 1. 438 * 439 * This is a rare case where callers are allowed to hold a spinlock, 440 * so we can't ourselves. 441 * 442 * MPSAFE 443 */ 444 static __inline 445 struct namecache * 446 _cache_hold(struct namecache *ncp) 447 { 448 atomic_add_int(&ncp->nc_refs, 1); 449 return(ncp); 450 } 451 452 /* 453 * Drop a cache entry, taking care to deal with races. 454 * 455 * For potential 1->0 transitions we must hold the ncp lock to safely 456 * test its flags. An unresolved entry with no children must be zapped 457 * to avoid leaks. 458 * 459 * The call to cache_zap() itself will handle all remaining races and 460 * will decrement the ncp's refs regardless. If we are resolved or 461 * have children nc_refs can safely be dropped to 0 without having to 462 * zap the entry. 463 * 464 * NOTE: cache_zap() will re-check nc_refs and nc_list in a MPSAFE fashion. 465 * 466 * NOTE: cache_zap() may return a non-NULL referenced parent which must 467 * be dropped in a loop. 468 * 469 * MPSAFE 470 */ 471 static __inline 472 void 473 _cache_drop(struct namecache *ncp) 474 { 475 int refs; 476 477 while (ncp) { 478 KKASSERT(ncp->nc_refs > 0); 479 refs = ncp->nc_refs; 480 481 if (refs == 1) { 482 if (_cache_lock_nonblock(ncp) == 0) { 483 ncp->nc_flag &= ~NCF_DEFEREDZAP; 484 if ((ncp->nc_flag & NCF_UNRESOLVED) && 485 TAILQ_EMPTY(&ncp->nc_list)) { 486 ncp = cache_zap(ncp, 1); 487 continue; 488 } 489 if (atomic_cmpset_int(&ncp->nc_refs, 1, 0)) { 490 _cache_unlock(ncp); 491 break; 492 } 493 _cache_unlock(ncp); 494 } 495 } else { 496 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) 497 break; 498 } 499 cpu_pause(); 500 } 501 } 502 503 /* 504 * Link a new namecache entry to its parent and to the hash table. Be 505 * careful to avoid races if vhold() blocks in the future. 506 * 507 * Both ncp and par must be referenced and locked. 508 * 509 * NOTE: The hash table spinlock is likely held during this call, we 510 * can't do anything fancy. 511 * 512 * MPSAFE 513 */ 514 static void 515 _cache_link_parent(struct namecache *ncp, struct namecache *par, 516 struct nchash_head *nchpp) 517 { 518 KKASSERT(ncp->nc_parent == NULL); 519 ncp->nc_parent = par; 520 ncp->nc_head = nchpp; 521 522 /* 523 * Set inheritance flags. Note that the parent flags may be 524 * stale due to getattr potentially not having been run yet 525 * (it gets run during nlookup()'s). 526 */ 527 ncp->nc_flag &= ~(NCF_SF_PNOCACHE | NCF_UF_PCACHE); 528 if (par->nc_flag & (NCF_SF_NOCACHE | NCF_SF_PNOCACHE)) 529 ncp->nc_flag |= NCF_SF_PNOCACHE; 530 if (par->nc_flag & (NCF_UF_CACHE | NCF_UF_PCACHE)) 531 ncp->nc_flag |= NCF_UF_PCACHE; 532 533 LIST_INSERT_HEAD(&nchpp->list, ncp, nc_hash); 534 535 if (TAILQ_EMPTY(&par->nc_list)) { 536 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry); 537 /* 538 * Any vp associated with an ncp which has children must 539 * be held to prevent it from being recycled. 540 */ 541 if (par->nc_vp) 542 vhold(par->nc_vp); 543 } else { 544 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry); 545 } 546 } 547 548 /* 549 * Remove the parent and hash associations from a namecache structure. 550 * If this is the last child of the parent the cache_drop(par) will 551 * attempt to recursively zap the parent. 552 * 553 * ncp must be locked. This routine will acquire a temporary lock on 554 * the parent as wlel as the appropriate hash chain. 555 * 556 * MPSAFE 557 */ 558 static void 559 _cache_unlink_parent(struct namecache *ncp) 560 { 561 struct namecache *par; 562 struct vnode *dropvp; 563 564 if ((par = ncp->nc_parent) != NULL) { 565 KKASSERT(ncp->nc_parent == par); 566 _cache_hold(par); 567 _cache_lock(par); 568 spin_lock(&ncp->nc_head->spin); 569 LIST_REMOVE(ncp, nc_hash); 570 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 571 dropvp = NULL; 572 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 573 dropvp = par->nc_vp; 574 spin_unlock(&ncp->nc_head->spin); 575 ncp->nc_parent = NULL; 576 ncp->nc_head = NULL; 577 _cache_unlock(par); 578 _cache_drop(par); 579 580 /* 581 * We can only safely vdrop with no spinlocks held. 582 */ 583 if (dropvp) 584 vdrop(dropvp); 585 } 586 } 587 588 /* 589 * Allocate a new namecache structure. Most of the code does not require 590 * zero-termination of the string but it makes vop_compat_ncreate() easier. 591 * 592 * MPSAFE 593 */ 594 static struct namecache * 595 cache_alloc(int nlen) 596 { 597 struct namecache *ncp; 598 599 ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO); 600 if (nlen) 601 ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK); 602 ncp->nc_nlen = nlen; 603 ncp->nc_flag = NCF_UNRESOLVED; 604 ncp->nc_error = ENOTCONN; /* needs to be resolved */ 605 ncp->nc_refs = 1; 606 607 TAILQ_INIT(&ncp->nc_list); 608 _cache_lock(ncp); 609 return(ncp); 610 } 611 612 /* 613 * Can only be called for the case where the ncp has never been 614 * associated with anything (so no spinlocks are needed). 615 * 616 * MPSAFE 617 */ 618 static void 619 _cache_free(struct namecache *ncp) 620 { 621 KKASSERT(ncp->nc_refs == 1 && ncp->nc_exlocks == 1); 622 if (ncp->nc_name) 623 kfree(ncp->nc_name, M_VFSCACHE); 624 kfree(ncp, M_VFSCACHE); 625 } 626 627 /* 628 * MPSAFE 629 */ 630 void 631 cache_zero(struct nchandle *nch) 632 { 633 nch->ncp = NULL; 634 nch->mount = NULL; 635 } 636 637 /* 638 * Ref and deref a namecache structure. 639 * 640 * The caller must specify a stable ncp pointer, typically meaning the 641 * ncp is already referenced but this can also occur indirectly through 642 * e.g. holding a lock on a direct child. 643 * 644 * WARNING: Caller may hold an unrelated read spinlock, which means we can't 645 * use read spinlocks here. 646 * 647 * MPSAFE if nch is 648 */ 649 struct nchandle * 650 cache_hold(struct nchandle *nch) 651 { 652 _cache_hold(nch->ncp); 653 atomic_add_int(&nch->mount->mnt_refs, 1); 654 return(nch); 655 } 656 657 /* 658 * Create a copy of a namecache handle for an already-referenced 659 * entry. 660 * 661 * MPSAFE if nch is 662 */ 663 void 664 cache_copy(struct nchandle *nch, struct nchandle *target) 665 { 666 *target = *nch; 667 if (target->ncp) 668 _cache_hold(target->ncp); 669 atomic_add_int(&nch->mount->mnt_refs, 1); 670 } 671 672 /* 673 * MPSAFE if nch is 674 */ 675 void 676 cache_changemount(struct nchandle *nch, struct mount *mp) 677 { 678 atomic_add_int(&nch->mount->mnt_refs, -1); 679 nch->mount = mp; 680 atomic_add_int(&nch->mount->mnt_refs, 1); 681 } 682 683 /* 684 * MPSAFE 685 */ 686 void 687 cache_drop(struct nchandle *nch) 688 { 689 atomic_add_int(&nch->mount->mnt_refs, -1); 690 _cache_drop(nch->ncp); 691 nch->ncp = NULL; 692 nch->mount = NULL; 693 } 694 695 /* 696 * MPSAFE 697 */ 698 void 699 cache_lock(struct nchandle *nch) 700 { 701 _cache_lock(nch->ncp); 702 } 703 704 /* 705 * Relock nch1 given an unlocked nch1 and a locked nch2. The caller 706 * is responsible for checking both for validity on return as they 707 * may have become invalid. 708 * 709 * We have to deal with potential deadlocks here, just ping pong 710 * the lock until we get it (we will always block somewhere when 711 * looping so this is not cpu-intensive). 712 * 713 * which = 0 nch1 not locked, nch2 is locked 714 * which = 1 nch1 is locked, nch2 is not locked 715 */ 716 void 717 cache_relock(struct nchandle *nch1, struct ucred *cred1, 718 struct nchandle *nch2, struct ucred *cred2) 719 { 720 int which; 721 722 which = 0; 723 724 for (;;) { 725 if (which == 0) { 726 if (cache_lock_nonblock(nch1) == 0) { 727 cache_resolve(nch1, cred1); 728 break; 729 } 730 cache_unlock(nch2); 731 cache_lock(nch1); 732 cache_resolve(nch1, cred1); 733 which = 1; 734 } else { 735 if (cache_lock_nonblock(nch2) == 0) { 736 cache_resolve(nch2, cred2); 737 break; 738 } 739 cache_unlock(nch1); 740 cache_lock(nch2); 741 cache_resolve(nch2, cred2); 742 which = 0; 743 } 744 } 745 } 746 747 /* 748 * MPSAFE 749 */ 750 int 751 cache_lock_nonblock(struct nchandle *nch) 752 { 753 return(_cache_lock_nonblock(nch->ncp)); 754 } 755 756 757 /* 758 * MPSAFE 759 */ 760 void 761 cache_unlock(struct nchandle *nch) 762 { 763 _cache_unlock(nch->ncp); 764 } 765 766 /* 767 * ref-and-lock, unlock-and-deref functions. 768 * 769 * This function is primarily used by nlookup. Even though cache_lock 770 * holds the vnode, it is possible that the vnode may have already 771 * initiated a recyclement. 772 * 773 * We want cache_get() to return a definitively usable vnode or a 774 * definitively unresolved ncp. 775 * 776 * MPSAFE 777 */ 778 static 779 struct namecache * 780 _cache_get(struct namecache *ncp) 781 { 782 _cache_hold(ncp); 783 _cache_lock(ncp); 784 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 785 _cache_setunresolved(ncp); 786 return(ncp); 787 } 788 789 /* 790 * This is a special form of _cache_lock() which only succeeds if 791 * it can get a pristine, non-recursive lock. The caller must have 792 * already ref'd the ncp. 793 * 794 * On success the ncp will be locked, on failure it will not. The 795 * ref count does not change either way. 796 * 797 * We want _cache_lock_special() (on success) to return a definitively 798 * usable vnode or a definitively unresolved ncp. 799 * 800 * MPSAFE 801 */ 802 static int 803 _cache_lock_special(struct namecache *ncp) 804 { 805 if (_cache_lock_nonblock(ncp) == 0) { 806 if ((ncp->nc_exlocks & ~NC_EXLOCK_REQ) == 1) { 807 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 808 _cache_setunresolved(ncp); 809 return(0); 810 } 811 _cache_unlock(ncp); 812 } 813 return(EWOULDBLOCK); 814 } 815 816 817 /* 818 * NOTE: The same nchandle can be passed for both arguments. 819 * 820 * MPSAFE 821 */ 822 void 823 cache_get(struct nchandle *nch, struct nchandle *target) 824 { 825 KKASSERT(nch->ncp->nc_refs > 0); 826 target->mount = nch->mount; 827 target->ncp = _cache_get(nch->ncp); 828 atomic_add_int(&target->mount->mnt_refs, 1); 829 } 830 831 /* 832 * MPSAFE 833 */ 834 static __inline 835 void 836 _cache_put(struct namecache *ncp) 837 { 838 _cache_unlock(ncp); 839 _cache_drop(ncp); 840 } 841 842 /* 843 * MPSAFE 844 */ 845 void 846 cache_put(struct nchandle *nch) 847 { 848 atomic_add_int(&nch->mount->mnt_refs, -1); 849 _cache_put(nch->ncp); 850 nch->ncp = NULL; 851 nch->mount = NULL; 852 } 853 854 /* 855 * Resolve an unresolved ncp by associating a vnode with it. If the 856 * vnode is NULL, a negative cache entry is created. 857 * 858 * The ncp should be locked on entry and will remain locked on return. 859 * 860 * MPSAFE 861 */ 862 static 863 void 864 _cache_setvp(struct mount *mp, struct namecache *ncp, struct vnode *vp) 865 { 866 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED); 867 868 if (vp != NULL) { 869 /* 870 * Any vp associated with an ncp which has children must 871 * be held. Any vp associated with a locked ncp must be held. 872 */ 873 if (!TAILQ_EMPTY(&ncp->nc_list)) 874 vhold(vp); 875 spin_lock(&vp->v_spinlock); 876 ncp->nc_vp = vp; 877 TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode); 878 spin_unlock(&vp->v_spinlock); 879 if (ncp->nc_exlocks) 880 vhold(vp); 881 882 /* 883 * Set auxiliary flags 884 */ 885 switch(vp->v_type) { 886 case VDIR: 887 ncp->nc_flag |= NCF_ISDIR; 888 break; 889 case VLNK: 890 ncp->nc_flag |= NCF_ISSYMLINK; 891 /* XXX cache the contents of the symlink */ 892 break; 893 default: 894 break; 895 } 896 atomic_add_int(&numcache, 1); 897 ncp->nc_error = 0; 898 } else { 899 /* 900 * When creating a negative cache hit we set the 901 * namecache_gen. A later resolve will clean out the 902 * negative cache hit if the mount point's namecache_gen 903 * has changed. Used by devfs, could also be used by 904 * other remote FSs. 905 */ 906 ncp->nc_vp = NULL; 907 spin_lock(&ncspin); 908 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 909 ++numneg; 910 spin_unlock(&ncspin); 911 ncp->nc_error = ENOENT; 912 if (mp) 913 ncp->nc_namecache_gen = mp->mnt_namecache_gen; 914 } 915 ncp->nc_flag &= ~(NCF_UNRESOLVED | NCF_DEFEREDZAP); 916 } 917 918 /* 919 * MPSAFE 920 */ 921 void 922 cache_setvp(struct nchandle *nch, struct vnode *vp) 923 { 924 _cache_setvp(nch->mount, nch->ncp, vp); 925 } 926 927 /* 928 * MPSAFE 929 */ 930 void 931 cache_settimeout(struct nchandle *nch, int nticks) 932 { 933 struct namecache *ncp = nch->ncp; 934 935 if ((ncp->nc_timeout = ticks + nticks) == 0) 936 ncp->nc_timeout = 1; 937 } 938 939 /* 940 * Disassociate the vnode or negative-cache association and mark a 941 * namecache entry as unresolved again. Note that the ncp is still 942 * left in the hash table and still linked to its parent. 943 * 944 * The ncp should be locked and refd on entry and will remain locked and refd 945 * on return. 946 * 947 * This routine is normally never called on a directory containing children. 948 * However, NFS often does just that in its rename() code as a cop-out to 949 * avoid complex namespace operations. This disconnects a directory vnode 950 * from its namecache and can cause the OLDAPI and NEWAPI to get out of 951 * sync. 952 * 953 * MPSAFE 954 */ 955 static 956 void 957 _cache_setunresolved(struct namecache *ncp) 958 { 959 struct vnode *vp; 960 961 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 962 ncp->nc_flag |= NCF_UNRESOLVED; 963 ncp->nc_timeout = 0; 964 ncp->nc_error = ENOTCONN; 965 if ((vp = ncp->nc_vp) != NULL) { 966 atomic_add_int(&numcache, -1); 967 spin_lock(&vp->v_spinlock); 968 ncp->nc_vp = NULL; 969 TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode); 970 spin_unlock(&vp->v_spinlock); 971 972 /* 973 * Any vp associated with an ncp with children is 974 * held by that ncp. Any vp associated with a locked 975 * ncp is held by that ncp. These conditions must be 976 * undone when the vp is cleared out from the ncp. 977 */ 978 if (!TAILQ_EMPTY(&ncp->nc_list)) 979 vdrop(vp); 980 if (ncp->nc_exlocks) 981 vdrop(vp); 982 } else { 983 spin_lock(&ncspin); 984 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 985 --numneg; 986 spin_unlock(&ncspin); 987 } 988 ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK); 989 } 990 } 991 992 /* 993 * The cache_nresolve() code calls this function to automatically 994 * set a resolved cache element to unresolved if it has timed out 995 * or if it is a negative cache hit and the mount point namecache_gen 996 * has changed. 997 * 998 * MPSAFE 999 */ 1000 static __inline void 1001 _cache_auto_unresolve(struct mount *mp, struct namecache *ncp) 1002 { 1003 /* 1004 * Already in an unresolved state, nothing to do. 1005 */ 1006 if (ncp->nc_flag & NCF_UNRESOLVED) 1007 return; 1008 1009 /* 1010 * Try to zap entries that have timed out. We have 1011 * to be careful here because locked leafs may depend 1012 * on the vnode remaining intact in a parent, so only 1013 * do this under very specific conditions. 1014 */ 1015 if (ncp->nc_timeout && (int)(ncp->nc_timeout - ticks) < 0 && 1016 TAILQ_EMPTY(&ncp->nc_list)) { 1017 _cache_setunresolved(ncp); 1018 return; 1019 } 1020 1021 /* 1022 * If a resolved negative cache hit is invalid due to 1023 * the mount's namecache generation being bumped, zap it. 1024 */ 1025 if (ncp->nc_vp == NULL && 1026 ncp->nc_namecache_gen != mp->mnt_namecache_gen) { 1027 _cache_setunresolved(ncp); 1028 return; 1029 } 1030 } 1031 1032 /* 1033 * MPSAFE 1034 */ 1035 void 1036 cache_setunresolved(struct nchandle *nch) 1037 { 1038 _cache_setunresolved(nch->ncp); 1039 } 1040 1041 /* 1042 * Determine if we can clear NCF_ISMOUNTPT by scanning the mountlist 1043 * looking for matches. This flag tells the lookup code when it must 1044 * check for a mount linkage and also prevents the directories in question 1045 * from being deleted or renamed. 1046 * 1047 * MPSAFE 1048 */ 1049 static 1050 int 1051 cache_clrmountpt_callback(struct mount *mp, void *data) 1052 { 1053 struct nchandle *nch = data; 1054 1055 if (mp->mnt_ncmounton.ncp == nch->ncp) 1056 return(1); 1057 if (mp->mnt_ncmountpt.ncp == nch->ncp) 1058 return(1); 1059 return(0); 1060 } 1061 1062 /* 1063 * MPSAFE 1064 */ 1065 void 1066 cache_clrmountpt(struct nchandle *nch) 1067 { 1068 int count; 1069 1070 count = mountlist_scan(cache_clrmountpt_callback, nch, 1071 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 1072 if (count == 0) 1073 nch->ncp->nc_flag &= ~NCF_ISMOUNTPT; 1074 } 1075 1076 /* 1077 * Invalidate portions of the namecache topology given a starting entry. 1078 * The passed ncp is set to an unresolved state and: 1079 * 1080 * The passed ncp must be referencxed and locked. The routine may unlock 1081 * and relock ncp several times, and will recheck the children and loop 1082 * to catch races. When done the passed ncp will be returned with the 1083 * reference and lock intact. 1084 * 1085 * CINV_DESTROY - Set a flag in the passed ncp entry indicating 1086 * that the physical underlying nodes have been 1087 * destroyed... as in deleted. For example, when 1088 * a directory is removed. This will cause record 1089 * lookups on the name to no longer be able to find 1090 * the record and tells the resolver to return failure 1091 * rather then trying to resolve through the parent. 1092 * 1093 * The topology itself, including ncp->nc_name, 1094 * remains intact. 1095 * 1096 * This only applies to the passed ncp, if CINV_CHILDREN 1097 * is specified the children are not flagged. 1098 * 1099 * CINV_CHILDREN - Set all children (recursively) to an unresolved 1100 * state as well. 1101 * 1102 * Note that this will also have the side effect of 1103 * cleaning out any unreferenced nodes in the topology 1104 * from the leaves up as the recursion backs out. 1105 * 1106 * Note that the topology for any referenced nodes remains intact, but 1107 * the nodes will be marked as having been destroyed and will be set 1108 * to an unresolved state. 1109 * 1110 * It is possible for cache_inval() to race a cache_resolve(), meaning that 1111 * the namecache entry may not actually be invalidated on return if it was 1112 * revalidated while recursing down into its children. This code guarentees 1113 * that the node(s) will go through an invalidation cycle, but does not 1114 * guarentee that they will remain in an invalidated state. 1115 * 1116 * Returns non-zero if a revalidation was detected during the invalidation 1117 * recursion, zero otherwise. Note that since only the original ncp is 1118 * locked the revalidation ultimately can only indicate that the original ncp 1119 * *MIGHT* no have been reresolved. 1120 * 1121 * DEEP RECURSION HANDLING - If a recursive invalidation recurses deeply we 1122 * have to avoid blowing out the kernel stack. We do this by saving the 1123 * deep namecache node and aborting the recursion, then re-recursing at that 1124 * node using a depth-first algorithm in order to allow multiple deep 1125 * recursions to chain through each other, then we restart the invalidation 1126 * from scratch. 1127 * 1128 * MPSAFE 1129 */ 1130 1131 struct cinvtrack { 1132 struct namecache *resume_ncp; 1133 int depth; 1134 }; 1135 1136 static int _cache_inval_internal(struct namecache *, int, struct cinvtrack *); 1137 1138 static 1139 int 1140 _cache_inval(struct namecache *ncp, int flags) 1141 { 1142 struct cinvtrack track; 1143 struct namecache *ncp2; 1144 int r; 1145 1146 track.depth = 0; 1147 track.resume_ncp = NULL; 1148 1149 for (;;) { 1150 r = _cache_inval_internal(ncp, flags, &track); 1151 if (track.resume_ncp == NULL) 1152 break; 1153 kprintf("Warning: deep namecache recursion at %s\n", 1154 ncp->nc_name); 1155 _cache_unlock(ncp); 1156 while ((ncp2 = track.resume_ncp) != NULL) { 1157 track.resume_ncp = NULL; 1158 _cache_lock(ncp2); 1159 _cache_inval_internal(ncp2, flags & ~CINV_DESTROY, 1160 &track); 1161 _cache_put(ncp2); 1162 } 1163 _cache_lock(ncp); 1164 } 1165 return(r); 1166 } 1167 1168 int 1169 cache_inval(struct nchandle *nch, int flags) 1170 { 1171 return(_cache_inval(nch->ncp, flags)); 1172 } 1173 1174 /* 1175 * Helper for _cache_inval(). The passed ncp is refd and locked and 1176 * remains that way on return, but may be unlocked/relocked multiple 1177 * times by the routine. 1178 */ 1179 static int 1180 _cache_inval_internal(struct namecache *ncp, int flags, struct cinvtrack *track) 1181 { 1182 struct namecache *kid; 1183 struct namecache *nextkid; 1184 int rcnt = 0; 1185 1186 KKASSERT(ncp->nc_exlocks); 1187 1188 _cache_setunresolved(ncp); 1189 if (flags & CINV_DESTROY) 1190 ncp->nc_flag |= NCF_DESTROYED; 1191 if ((flags & CINV_CHILDREN) && 1192 (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL 1193 ) { 1194 _cache_hold(kid); 1195 if (++track->depth > MAX_RECURSION_DEPTH) { 1196 track->resume_ncp = ncp; 1197 _cache_hold(ncp); 1198 ++rcnt; 1199 } 1200 _cache_unlock(ncp); 1201 while (kid) { 1202 if (track->resume_ncp) { 1203 _cache_drop(kid); 1204 break; 1205 } 1206 if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL) 1207 _cache_hold(nextkid); 1208 if ((kid->nc_flag & NCF_UNRESOLVED) == 0 || 1209 TAILQ_FIRST(&kid->nc_list) 1210 ) { 1211 _cache_lock(kid); 1212 rcnt += _cache_inval_internal(kid, flags & ~CINV_DESTROY, track); 1213 _cache_unlock(kid); 1214 } 1215 _cache_drop(kid); 1216 kid = nextkid; 1217 } 1218 --track->depth; 1219 _cache_lock(ncp); 1220 } 1221 1222 /* 1223 * Someone could have gotten in there while ncp was unlocked, 1224 * retry if so. 1225 */ 1226 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 1227 ++rcnt; 1228 return (rcnt); 1229 } 1230 1231 /* 1232 * Invalidate a vnode's namecache associations. To avoid races against 1233 * the resolver we do not invalidate a node which we previously invalidated 1234 * but which was then re-resolved while we were in the invalidation loop. 1235 * 1236 * Returns non-zero if any namecache entries remain after the invalidation 1237 * loop completed. 1238 * 1239 * NOTE: Unlike the namecache topology which guarentees that ncp's will not 1240 * be ripped out of the topology while held, the vnode's v_namecache 1241 * list has no such restriction. NCP's can be ripped out of the list 1242 * at virtually any time if not locked, even if held. 1243 * 1244 * In addition, the v_namecache list itself must be locked via 1245 * the vnode's spinlock. 1246 * 1247 * MPSAFE 1248 */ 1249 int 1250 cache_inval_vp(struct vnode *vp, int flags) 1251 { 1252 struct namecache *ncp; 1253 struct namecache *next; 1254 1255 restart: 1256 spin_lock(&vp->v_spinlock); 1257 ncp = TAILQ_FIRST(&vp->v_namecache); 1258 if (ncp) 1259 _cache_hold(ncp); 1260 while (ncp) { 1261 /* loop entered with ncp held and vp spin-locked */ 1262 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL) 1263 _cache_hold(next); 1264 spin_unlock(&vp->v_spinlock); 1265 _cache_lock(ncp); 1266 if (ncp->nc_vp != vp) { 1267 kprintf("Warning: cache_inval_vp: race-A detected on " 1268 "%s\n", ncp->nc_name); 1269 _cache_put(ncp); 1270 if (next) 1271 _cache_drop(next); 1272 goto restart; 1273 } 1274 _cache_inval(ncp, flags); 1275 _cache_put(ncp); /* also releases reference */ 1276 ncp = next; 1277 spin_lock(&vp->v_spinlock); 1278 if (ncp && ncp->nc_vp != vp) { 1279 spin_unlock(&vp->v_spinlock); 1280 kprintf("Warning: cache_inval_vp: race-B detected on " 1281 "%s\n", ncp->nc_name); 1282 _cache_drop(ncp); 1283 goto restart; 1284 } 1285 } 1286 spin_unlock(&vp->v_spinlock); 1287 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 1288 } 1289 1290 /* 1291 * This routine is used instead of the normal cache_inval_vp() when we 1292 * are trying to recycle otherwise good vnodes. 1293 * 1294 * Return 0 on success, non-zero if not all namecache records could be 1295 * disassociated from the vnode (for various reasons). 1296 * 1297 * MPSAFE 1298 */ 1299 int 1300 cache_inval_vp_nonblock(struct vnode *vp) 1301 { 1302 struct namecache *ncp; 1303 struct namecache *next; 1304 1305 spin_lock(&vp->v_spinlock); 1306 ncp = TAILQ_FIRST(&vp->v_namecache); 1307 if (ncp) 1308 _cache_hold(ncp); 1309 while (ncp) { 1310 /* loop entered with ncp held */ 1311 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL) 1312 _cache_hold(next); 1313 spin_unlock(&vp->v_spinlock); 1314 if (_cache_lock_nonblock(ncp)) { 1315 _cache_drop(ncp); 1316 if (next) 1317 _cache_drop(next); 1318 goto done; 1319 } 1320 if (ncp->nc_vp != vp) { 1321 kprintf("Warning: cache_inval_vp: race-A detected on " 1322 "%s\n", ncp->nc_name); 1323 _cache_put(ncp); 1324 if (next) 1325 _cache_drop(next); 1326 goto done; 1327 } 1328 _cache_inval(ncp, 0); 1329 _cache_put(ncp); /* also releases reference */ 1330 ncp = next; 1331 spin_lock(&vp->v_spinlock); 1332 if (ncp && ncp->nc_vp != vp) { 1333 spin_unlock(&vp->v_spinlock); 1334 kprintf("Warning: cache_inval_vp: race-B detected on " 1335 "%s\n", ncp->nc_name); 1336 _cache_drop(ncp); 1337 goto done; 1338 } 1339 } 1340 spin_unlock(&vp->v_spinlock); 1341 done: 1342 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 1343 } 1344 1345 /* 1346 * The source ncp has been renamed to the target ncp. Both fncp and tncp 1347 * must be locked. The target ncp is destroyed (as a normal rename-over 1348 * would destroy the target file or directory). 1349 * 1350 * Because there may be references to the source ncp we cannot copy its 1351 * contents to the target. Instead the source ncp is relinked as the target 1352 * and the target ncp is removed from the namecache topology. 1353 * 1354 * MPSAFE 1355 */ 1356 void 1357 cache_rename(struct nchandle *fnch, struct nchandle *tnch) 1358 { 1359 struct namecache *fncp = fnch->ncp; 1360 struct namecache *tncp = tnch->ncp; 1361 struct namecache *tncp_par; 1362 struct nchash_head *nchpp; 1363 u_int32_t hash; 1364 char *oname; 1365 1366 /* 1367 * Rename fncp (unlink) 1368 */ 1369 _cache_unlink_parent(fncp); 1370 oname = fncp->nc_name; 1371 fncp->nc_name = tncp->nc_name; 1372 fncp->nc_nlen = tncp->nc_nlen; 1373 tncp_par = tncp->nc_parent; 1374 _cache_hold(tncp_par); 1375 _cache_lock(tncp_par); 1376 1377 /* 1378 * Rename fncp (relink) 1379 */ 1380 hash = fnv_32_buf(fncp->nc_name, fncp->nc_nlen, FNV1_32_INIT); 1381 hash = fnv_32_buf(&tncp_par, sizeof(tncp_par), hash); 1382 nchpp = NCHHASH(hash); 1383 1384 spin_lock(&nchpp->spin); 1385 _cache_link_parent(fncp, tncp_par, nchpp); 1386 spin_unlock(&nchpp->spin); 1387 1388 _cache_put(tncp_par); 1389 1390 /* 1391 * Get rid of the overwritten tncp (unlink) 1392 */ 1393 _cache_setunresolved(tncp); 1394 _cache_unlink_parent(tncp); 1395 tncp->nc_name = NULL; 1396 tncp->nc_nlen = 0; 1397 1398 if (oname) 1399 kfree(oname, M_VFSCACHE); 1400 } 1401 1402 /* 1403 * vget the vnode associated with the namecache entry. Resolve the namecache 1404 * entry if necessary. The passed ncp must be referenced and locked. 1405 * 1406 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked 1407 * (depending on the passed lk_type) will be returned in *vpp with an error 1408 * of 0, or NULL will be returned in *vpp with a non-0 error code. The 1409 * most typical error is ENOENT, meaning that the ncp represents a negative 1410 * cache hit and there is no vnode to retrieve, but other errors can occur 1411 * too. 1412 * 1413 * The vget() can race a reclaim. If this occurs we re-resolve the 1414 * namecache entry. 1415 * 1416 * There are numerous places in the kernel where vget() is called on a 1417 * vnode while one or more of its namecache entries is locked. Releasing 1418 * a vnode never deadlocks against locked namecache entries (the vnode 1419 * will not get recycled while referenced ncp's exist). This means we 1420 * can safely acquire the vnode. In fact, we MUST NOT release the ncp 1421 * lock when acquiring the vp lock or we might cause a deadlock. 1422 * 1423 * MPSAFE 1424 */ 1425 int 1426 cache_vget(struct nchandle *nch, struct ucred *cred, 1427 int lk_type, struct vnode **vpp) 1428 { 1429 struct namecache *ncp; 1430 struct vnode *vp; 1431 int error; 1432 1433 ncp = nch->ncp; 1434 KKASSERT(ncp->nc_locktd == curthread); 1435 again: 1436 vp = NULL; 1437 if (ncp->nc_flag & NCF_UNRESOLVED) 1438 error = cache_resolve(nch, cred); 1439 else 1440 error = 0; 1441 1442 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 1443 error = vget(vp, lk_type); 1444 if (error) { 1445 /* 1446 * VRECLAIM race 1447 */ 1448 if (error == ENOENT) { 1449 kprintf("Warning: vnode reclaim race detected " 1450 "in cache_vget on %p (%s)\n", 1451 vp, ncp->nc_name); 1452 _cache_setunresolved(ncp); 1453 goto again; 1454 } 1455 1456 /* 1457 * Not a reclaim race, some other error. 1458 */ 1459 KKASSERT(ncp->nc_vp == vp); 1460 vp = NULL; 1461 } else { 1462 KKASSERT(ncp->nc_vp == vp); 1463 KKASSERT((vp->v_flag & VRECLAIMED) == 0); 1464 } 1465 } 1466 if (error == 0 && vp == NULL) 1467 error = ENOENT; 1468 *vpp = vp; 1469 return(error); 1470 } 1471 1472 int 1473 cache_vref(struct nchandle *nch, struct ucred *cred, struct vnode **vpp) 1474 { 1475 struct namecache *ncp; 1476 struct vnode *vp; 1477 int error; 1478 1479 ncp = nch->ncp; 1480 KKASSERT(ncp->nc_locktd == curthread); 1481 again: 1482 vp = NULL; 1483 if (ncp->nc_flag & NCF_UNRESOLVED) 1484 error = cache_resolve(nch, cred); 1485 else 1486 error = 0; 1487 1488 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 1489 error = vget(vp, LK_SHARED); 1490 if (error) { 1491 /* 1492 * VRECLAIM race 1493 */ 1494 if (error == ENOENT) { 1495 kprintf("Warning: vnode reclaim race detected " 1496 "in cache_vget on %p (%s)\n", 1497 vp, ncp->nc_name); 1498 _cache_setunresolved(ncp); 1499 goto again; 1500 } 1501 1502 /* 1503 * Not a reclaim race, some other error. 1504 */ 1505 KKASSERT(ncp->nc_vp == vp); 1506 vp = NULL; 1507 } else { 1508 KKASSERT(ncp->nc_vp == vp); 1509 KKASSERT((vp->v_flag & VRECLAIMED) == 0); 1510 /* caller does not want a lock */ 1511 vn_unlock(vp); 1512 } 1513 } 1514 if (error == 0 && vp == NULL) 1515 error = ENOENT; 1516 *vpp = vp; 1517 return(error); 1518 } 1519 1520 /* 1521 * Return a referenced vnode representing the parent directory of 1522 * ncp. 1523 * 1524 * Because the caller has locked the ncp it should not be possible for 1525 * the parent ncp to go away. However, the parent can unresolve its 1526 * dvp at any time so we must be able to acquire a lock on the parent 1527 * to safely access nc_vp. 1528 * 1529 * We have to leave par unlocked when vget()ing dvp to avoid a deadlock, 1530 * so use vhold()/vdrop() while holding the lock to prevent dvp from 1531 * getting destroyed. 1532 * 1533 * MPSAFE - Note vhold() is allowed when dvp has 0 refs if we hold a 1534 * lock on the ncp in question.. 1535 */ 1536 static struct vnode * 1537 cache_dvpref(struct namecache *ncp) 1538 { 1539 struct namecache *par; 1540 struct vnode *dvp; 1541 1542 dvp = NULL; 1543 if ((par = ncp->nc_parent) != NULL) { 1544 _cache_hold(par); 1545 _cache_lock(par); 1546 if ((par->nc_flag & NCF_UNRESOLVED) == 0) { 1547 if ((dvp = par->nc_vp) != NULL) 1548 vhold(dvp); 1549 } 1550 _cache_unlock(par); 1551 if (dvp) { 1552 if (vget(dvp, LK_SHARED) == 0) { 1553 vn_unlock(dvp); 1554 vdrop(dvp); 1555 /* return refd, unlocked dvp */ 1556 } else { 1557 vdrop(dvp); 1558 dvp = NULL; 1559 } 1560 } 1561 _cache_drop(par); 1562 } 1563 return(dvp); 1564 } 1565 1566 /* 1567 * Convert a directory vnode to a namecache record without any other 1568 * knowledge of the topology. This ONLY works with directory vnodes and 1569 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the 1570 * returned ncp (if not NULL) will be held and unlocked. 1571 * 1572 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned. 1573 * If 'makeit' is 1 we attempt to track-down and create the namecache topology 1574 * for dvp. This will fail only if the directory has been deleted out from 1575 * under the caller. 1576 * 1577 * Callers must always check for a NULL return no matter the value of 'makeit'. 1578 * 1579 * To avoid underflowing the kernel stack each recursive call increments 1580 * the makeit variable. 1581 */ 1582 1583 static int cache_inefficient_scan(struct nchandle *nch, struct ucred *cred, 1584 struct vnode *dvp, char *fakename); 1585 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred, 1586 struct vnode **saved_dvp); 1587 1588 int 1589 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit, 1590 struct nchandle *nch) 1591 { 1592 struct vnode *saved_dvp; 1593 struct vnode *pvp; 1594 char *fakename; 1595 int error; 1596 1597 nch->ncp = NULL; 1598 nch->mount = dvp->v_mount; 1599 saved_dvp = NULL; 1600 fakename = NULL; 1601 1602 /* 1603 * Handle the makeit == 0 degenerate case 1604 */ 1605 if (makeit == 0) { 1606 spin_lock(&dvp->v_spinlock); 1607 nch->ncp = TAILQ_FIRST(&dvp->v_namecache); 1608 if (nch->ncp) 1609 cache_hold(nch); 1610 spin_unlock(&dvp->v_spinlock); 1611 } 1612 1613 /* 1614 * Loop until resolution, inside code will break out on error. 1615 */ 1616 while (makeit) { 1617 /* 1618 * Break out if we successfully acquire a working ncp. 1619 */ 1620 spin_lock(&dvp->v_spinlock); 1621 nch->ncp = TAILQ_FIRST(&dvp->v_namecache); 1622 if (nch->ncp) { 1623 cache_hold(nch); 1624 spin_unlock(&dvp->v_spinlock); 1625 break; 1626 } 1627 spin_unlock(&dvp->v_spinlock); 1628 1629 /* 1630 * If dvp is the root of its filesystem it should already 1631 * have a namecache pointer associated with it as a side 1632 * effect of the mount, but it may have been disassociated. 1633 */ 1634 if (dvp->v_flag & VROOT) { 1635 nch->ncp = _cache_get(nch->mount->mnt_ncmountpt.ncp); 1636 error = cache_resolve_mp(nch->mount); 1637 _cache_put(nch->ncp); 1638 if (ncvp_debug) { 1639 kprintf("cache_fromdvp: resolve root of mount %p error %d", 1640 dvp->v_mount, error); 1641 } 1642 if (error) { 1643 if (ncvp_debug) 1644 kprintf(" failed\n"); 1645 nch->ncp = NULL; 1646 break; 1647 } 1648 if (ncvp_debug) 1649 kprintf(" succeeded\n"); 1650 continue; 1651 } 1652 1653 /* 1654 * If we are recursed too deeply resort to an O(n^2) 1655 * algorithm to resolve the namecache topology. The 1656 * resolved pvp is left referenced in saved_dvp to 1657 * prevent the tree from being destroyed while we loop. 1658 */ 1659 if (makeit > 20) { 1660 error = cache_fromdvp_try(dvp, cred, &saved_dvp); 1661 if (error) { 1662 kprintf("lookupdotdot(longpath) failed %d " 1663 "dvp %p\n", error, dvp); 1664 nch->ncp = NULL; 1665 break; 1666 } 1667 continue; 1668 } 1669 1670 /* 1671 * Get the parent directory and resolve its ncp. 1672 */ 1673 if (fakename) { 1674 kfree(fakename, M_TEMP); 1675 fakename = NULL; 1676 } 1677 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred, 1678 &fakename); 1679 if (error) { 1680 kprintf("lookupdotdot failed %d dvp %p\n", error, dvp); 1681 break; 1682 } 1683 vn_unlock(pvp); 1684 1685 /* 1686 * Reuse makeit as a recursion depth counter. On success 1687 * nch will be fully referenced. 1688 */ 1689 cache_fromdvp(pvp, cred, makeit + 1, nch); 1690 vrele(pvp); 1691 if (nch->ncp == NULL) 1692 break; 1693 1694 /* 1695 * Do an inefficient scan of pvp (embodied by ncp) to look 1696 * for dvp. This will create a namecache record for dvp on 1697 * success. We loop up to recheck on success. 1698 * 1699 * ncp and dvp are both held but not locked. 1700 */ 1701 error = cache_inefficient_scan(nch, cred, dvp, fakename); 1702 if (error) { 1703 kprintf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n", 1704 pvp, nch->ncp->nc_name, dvp); 1705 cache_drop(nch); 1706 /* nch was NULLed out, reload mount */ 1707 nch->mount = dvp->v_mount; 1708 break; 1709 } 1710 if (ncvp_debug) { 1711 kprintf("cache_fromdvp: scan %p (%s) succeeded\n", 1712 pvp, nch->ncp->nc_name); 1713 } 1714 cache_drop(nch); 1715 /* nch was NULLed out, reload mount */ 1716 nch->mount = dvp->v_mount; 1717 } 1718 1719 /* 1720 * If nch->ncp is non-NULL it will have been held already. 1721 */ 1722 if (fakename) 1723 kfree(fakename, M_TEMP); 1724 if (saved_dvp) 1725 vrele(saved_dvp); 1726 if (nch->ncp) 1727 return (0); 1728 return (EINVAL); 1729 } 1730 1731 /* 1732 * Go up the chain of parent directories until we find something 1733 * we can resolve into the namecache. This is very inefficient. 1734 */ 1735 static 1736 int 1737 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred, 1738 struct vnode **saved_dvp) 1739 { 1740 struct nchandle nch; 1741 struct vnode *pvp; 1742 int error; 1743 static time_t last_fromdvp_report; 1744 char *fakename; 1745 1746 /* 1747 * Loop getting the parent directory vnode until we get something we 1748 * can resolve in the namecache. 1749 */ 1750 vref(dvp); 1751 nch.mount = dvp->v_mount; 1752 nch.ncp = NULL; 1753 fakename = NULL; 1754 1755 for (;;) { 1756 if (fakename) { 1757 kfree(fakename, M_TEMP); 1758 fakename = NULL; 1759 } 1760 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred, 1761 &fakename); 1762 if (error) { 1763 vrele(dvp); 1764 break; 1765 } 1766 vn_unlock(pvp); 1767 spin_lock(&pvp->v_spinlock); 1768 if ((nch.ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) { 1769 _cache_hold(nch.ncp); 1770 spin_unlock(&pvp->v_spinlock); 1771 vrele(pvp); 1772 break; 1773 } 1774 spin_unlock(&pvp->v_spinlock); 1775 if (pvp->v_flag & VROOT) { 1776 nch.ncp = _cache_get(pvp->v_mount->mnt_ncmountpt.ncp); 1777 error = cache_resolve_mp(nch.mount); 1778 _cache_unlock(nch.ncp); 1779 vrele(pvp); 1780 if (error) { 1781 _cache_drop(nch.ncp); 1782 nch.ncp = NULL; 1783 vrele(dvp); 1784 } 1785 break; 1786 } 1787 vrele(dvp); 1788 dvp = pvp; 1789 } 1790 if (error == 0) { 1791 if (last_fromdvp_report != time_second) { 1792 last_fromdvp_report = time_second; 1793 kprintf("Warning: extremely inefficient path " 1794 "resolution on %s\n", 1795 nch.ncp->nc_name); 1796 } 1797 error = cache_inefficient_scan(&nch, cred, dvp, fakename); 1798 1799 /* 1800 * Hopefully dvp now has a namecache record associated with 1801 * it. Leave it referenced to prevent the kernel from 1802 * recycling the vnode. Otherwise extremely long directory 1803 * paths could result in endless recycling. 1804 */ 1805 if (*saved_dvp) 1806 vrele(*saved_dvp); 1807 *saved_dvp = dvp; 1808 _cache_drop(nch.ncp); 1809 } 1810 if (fakename) 1811 kfree(fakename, M_TEMP); 1812 return (error); 1813 } 1814 1815 /* 1816 * Do an inefficient scan of the directory represented by ncp looking for 1817 * the directory vnode dvp. ncp must be held but not locked on entry and 1818 * will be held on return. dvp must be refd but not locked on entry and 1819 * will remain refd on return. 1820 * 1821 * Why do this at all? Well, due to its stateless nature the NFS server 1822 * converts file handles directly to vnodes without necessarily going through 1823 * the namecache ops that would otherwise create the namecache topology 1824 * leading to the vnode. We could either (1) Change the namecache algorithms 1825 * to allow disconnect namecache records that are re-merged opportunistically, 1826 * or (2) Make the NFS server backtrack and scan to recover a connected 1827 * namecache topology in order to then be able to issue new API lookups. 1828 * 1829 * It turns out that (1) is a huge mess. It takes a nice clean set of 1830 * namecache algorithms and introduces a lot of complication in every subsystem 1831 * that calls into the namecache to deal with the re-merge case, especially 1832 * since we are using the namecache to placehold negative lookups and the 1833 * vnode might not be immediately assigned. (2) is certainly far less 1834 * efficient then (1), but since we are only talking about directories here 1835 * (which are likely to remain cached), the case does not actually run all 1836 * that often and has the supreme advantage of not polluting the namecache 1837 * algorithms. 1838 * 1839 * If a fakename is supplied just construct a namecache entry using the 1840 * fake name. 1841 */ 1842 static int 1843 cache_inefficient_scan(struct nchandle *nch, struct ucred *cred, 1844 struct vnode *dvp, char *fakename) 1845 { 1846 struct nlcomponent nlc; 1847 struct nchandle rncp; 1848 struct dirent *den; 1849 struct vnode *pvp; 1850 struct vattr vat; 1851 struct iovec iov; 1852 struct uio uio; 1853 int blksize; 1854 int eofflag; 1855 int bytes; 1856 char *rbuf; 1857 int error; 1858 1859 vat.va_blocksize = 0; 1860 if ((error = VOP_GETATTR(dvp, &vat)) != 0) 1861 return (error); 1862 cache_lock(nch); 1863 error = cache_vref(nch, cred, &pvp); 1864 cache_unlock(nch); 1865 if (error) 1866 return (error); 1867 if (ncvp_debug) { 1868 kprintf("inefficient_scan: directory iosize %ld " 1869 "vattr fileid = %lld\n", 1870 vat.va_blocksize, 1871 (long long)vat.va_fileid); 1872 } 1873 1874 /* 1875 * Use the supplied fakename if not NULL. Fake names are typically 1876 * not in the actual filesystem hierarchy. This is used by HAMMER 1877 * to glue @@timestamp recursions together. 1878 */ 1879 if (fakename) { 1880 nlc.nlc_nameptr = fakename; 1881 nlc.nlc_namelen = strlen(fakename); 1882 rncp = cache_nlookup(nch, &nlc); 1883 goto done; 1884 } 1885 1886 if ((blksize = vat.va_blocksize) == 0) 1887 blksize = DEV_BSIZE; 1888 rbuf = kmalloc(blksize, M_TEMP, M_WAITOK); 1889 rncp.ncp = NULL; 1890 1891 eofflag = 0; 1892 uio.uio_offset = 0; 1893 again: 1894 iov.iov_base = rbuf; 1895 iov.iov_len = blksize; 1896 uio.uio_iov = &iov; 1897 uio.uio_iovcnt = 1; 1898 uio.uio_resid = blksize; 1899 uio.uio_segflg = UIO_SYSSPACE; 1900 uio.uio_rw = UIO_READ; 1901 uio.uio_td = curthread; 1902 1903 if (ncvp_debug >= 2) 1904 kprintf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset); 1905 error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL); 1906 if (error == 0) { 1907 den = (struct dirent *)rbuf; 1908 bytes = blksize - uio.uio_resid; 1909 1910 while (bytes > 0) { 1911 if (ncvp_debug >= 2) { 1912 kprintf("cache_inefficient_scan: %*.*s\n", 1913 den->d_namlen, den->d_namlen, 1914 den->d_name); 1915 } 1916 if (den->d_type != DT_WHT && 1917 den->d_ino == vat.va_fileid) { 1918 if (ncvp_debug) { 1919 kprintf("cache_inefficient_scan: " 1920 "MATCHED inode %lld path %s/%*.*s\n", 1921 (long long)vat.va_fileid, 1922 nch->ncp->nc_name, 1923 den->d_namlen, den->d_namlen, 1924 den->d_name); 1925 } 1926 nlc.nlc_nameptr = den->d_name; 1927 nlc.nlc_namelen = den->d_namlen; 1928 rncp = cache_nlookup(nch, &nlc); 1929 KKASSERT(rncp.ncp != NULL); 1930 break; 1931 } 1932 bytes -= _DIRENT_DIRSIZ(den); 1933 den = _DIRENT_NEXT(den); 1934 } 1935 if (rncp.ncp == NULL && eofflag == 0 && uio.uio_resid != blksize) 1936 goto again; 1937 } 1938 kfree(rbuf, M_TEMP); 1939 done: 1940 vrele(pvp); 1941 if (rncp.ncp) { 1942 if (rncp.ncp->nc_flag & NCF_UNRESOLVED) { 1943 _cache_setvp(rncp.mount, rncp.ncp, dvp); 1944 if (ncvp_debug >= 2) { 1945 kprintf("cache_inefficient_scan: setvp %s/%s = %p\n", 1946 nch->ncp->nc_name, rncp.ncp->nc_name, dvp); 1947 } 1948 } else { 1949 if (ncvp_debug >= 2) { 1950 kprintf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n", 1951 nch->ncp->nc_name, rncp.ncp->nc_name, dvp, 1952 rncp.ncp->nc_vp); 1953 } 1954 } 1955 if (rncp.ncp->nc_vp == NULL) 1956 error = rncp.ncp->nc_error; 1957 /* 1958 * Release rncp after a successful nlookup. rncp was fully 1959 * referenced. 1960 */ 1961 cache_put(&rncp); 1962 } else { 1963 kprintf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n", 1964 dvp, nch->ncp->nc_name); 1965 error = ENOENT; 1966 } 1967 return (error); 1968 } 1969 1970 /* 1971 * Zap a namecache entry. The ncp is unconditionally set to an unresolved 1972 * state, which disassociates it from its vnode or ncneglist. 1973 * 1974 * Then, if there are no additional references to the ncp and no children, 1975 * the ncp is removed from the topology and destroyed. 1976 * 1977 * References and/or children may exist if the ncp is in the middle of the 1978 * topology, preventing the ncp from being destroyed. 1979 * 1980 * This function must be called with the ncp held and locked and will unlock 1981 * and drop it during zapping. 1982 * 1983 * If nonblock is non-zero and the parent ncp cannot be locked we give up. 1984 * This case can occur in the cache_drop() path. 1985 * 1986 * This function may returned a held (but NOT locked) parent node which the 1987 * caller must drop. We do this so _cache_drop() can loop, to avoid 1988 * blowing out the kernel stack. 1989 * 1990 * WARNING! For MPSAFE operation this routine must acquire up to three 1991 * spin locks to be able to safely test nc_refs. Lock order is 1992 * very important. 1993 * 1994 * hash spinlock if on hash list 1995 * parent spinlock if child of parent 1996 * (the ncp is unresolved so there is no vnode association) 1997 */ 1998 static struct namecache * 1999 cache_zap(struct namecache *ncp, int nonblock) 2000 { 2001 struct namecache *par; 2002 struct vnode *dropvp; 2003 int refs; 2004 2005 /* 2006 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED. 2007 */ 2008 _cache_setunresolved(ncp); 2009 2010 /* 2011 * Try to scrap the entry and possibly tail-recurse on its parent. 2012 * We only scrap unref'd (other then our ref) unresolved entries, 2013 * we do not scrap 'live' entries. 2014 * 2015 * Note that once the spinlocks are acquired if nc_refs == 1 no 2016 * other references are possible. If it isn't, however, we have 2017 * to decrement but also be sure to avoid a 1->0 transition. 2018 */ 2019 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED); 2020 KKASSERT(ncp->nc_refs > 0); 2021 2022 /* 2023 * Acquire locks. Note that the parent can't go away while we hold 2024 * a child locked. 2025 */ 2026 if ((par = ncp->nc_parent) != NULL) { 2027 if (nonblock) { 2028 for (;;) { 2029 if (_cache_lock_nonblock(par) == 0) 2030 break; 2031 refs = ncp->nc_refs; 2032 ncp->nc_flag |= NCF_DEFEREDZAP; 2033 ++numdefered; /* MP race ok */ 2034 if (atomic_cmpset_int(&ncp->nc_refs, 2035 refs, refs - 1)) { 2036 _cache_unlock(ncp); 2037 return(NULL); 2038 } 2039 cpu_pause(); 2040 } 2041 _cache_hold(par); 2042 } else { 2043 _cache_hold(par); 2044 _cache_lock(par); 2045 } 2046 spin_lock(&ncp->nc_head->spin); 2047 } 2048 2049 /* 2050 * If someone other then us has a ref or we have children 2051 * we cannot zap the entry. The 1->0 transition and any 2052 * further list operation is protected by the spinlocks 2053 * we have acquired but other transitions are not. 2054 */ 2055 for (;;) { 2056 refs = ncp->nc_refs; 2057 if (refs == 1 && TAILQ_EMPTY(&ncp->nc_list)) 2058 break; 2059 if (atomic_cmpset_int(&ncp->nc_refs, refs, refs - 1)) { 2060 if (par) { 2061 spin_unlock(&ncp->nc_head->spin); 2062 _cache_put(par); 2063 } 2064 _cache_unlock(ncp); 2065 return(NULL); 2066 } 2067 cpu_pause(); 2068 } 2069 2070 /* 2071 * We are the only ref and with the spinlocks held no further 2072 * refs can be acquired by others. 2073 * 2074 * Remove us from the hash list and parent list. We have to 2075 * drop a ref on the parent's vp if the parent's list becomes 2076 * empty. 2077 */ 2078 dropvp = NULL; 2079 if (par) { 2080 struct nchash_head *nchpp = ncp->nc_head; 2081 2082 KKASSERT(nchpp != NULL); 2083 LIST_REMOVE(ncp, nc_hash); 2084 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 2085 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 2086 dropvp = par->nc_vp; 2087 ncp->nc_head = NULL; 2088 ncp->nc_parent = NULL; 2089 spin_unlock(&nchpp->spin); 2090 _cache_unlock(par); 2091 } else { 2092 KKASSERT(ncp->nc_head == NULL); 2093 } 2094 2095 /* 2096 * ncp should not have picked up any refs. Physically 2097 * destroy the ncp. 2098 */ 2099 KKASSERT(ncp->nc_refs == 1); 2100 /* _cache_unlock(ncp) not required */ 2101 ncp->nc_refs = -1; /* safety */ 2102 if (ncp->nc_name) 2103 kfree(ncp->nc_name, M_VFSCACHE); 2104 kfree(ncp, M_VFSCACHE); 2105 2106 /* 2107 * Delayed drop (we had to release our spinlocks) 2108 * 2109 * The refed parent (if not NULL) must be dropped. The 2110 * caller is responsible for looping. 2111 */ 2112 if (dropvp) 2113 vdrop(dropvp); 2114 return(par); 2115 } 2116 2117 /* 2118 * Clean up dangling negative cache and defered-drop entries in the 2119 * namecache. 2120 */ 2121 static enum { CHI_LOW, CHI_HIGH } cache_hysteresis_state = CHI_LOW; 2122 2123 void 2124 cache_hysteresis(void) 2125 { 2126 /* 2127 * Don't cache too many negative hits. We use hysteresis to reduce 2128 * the impact on the critical path. 2129 */ 2130 switch(cache_hysteresis_state) { 2131 case CHI_LOW: 2132 if (numneg > MINNEG && numneg * ncnegfactor > numcache) { 2133 _cache_cleanneg(10); 2134 cache_hysteresis_state = CHI_HIGH; 2135 } 2136 break; 2137 case CHI_HIGH: 2138 if (numneg > MINNEG * 9 / 10 && 2139 numneg * ncnegfactor * 9 / 10 > numcache 2140 ) { 2141 _cache_cleanneg(10); 2142 } else { 2143 cache_hysteresis_state = CHI_LOW; 2144 } 2145 break; 2146 } 2147 2148 /* 2149 * Clean out dangling defered-zap ncps which could not 2150 * be cleanly dropped if too many build up. Note 2151 * that numdefered is not an exact number as such ncps 2152 * can be reused and the counter is not handled in a MP 2153 * safe manner by design. 2154 */ 2155 if (numdefered * ncnegfactor > numcache) { 2156 _cache_cleandefered(); 2157 } 2158 } 2159 2160 /* 2161 * NEW NAMECACHE LOOKUP API 2162 * 2163 * Lookup an entry in the namecache. The passed par_nch must be referenced 2164 * and unlocked. A referenced and locked nchandle with a non-NULL nch.ncp 2165 * is ALWAYS returned, eve if the supplied component is illegal. 2166 * 2167 * The resulting namecache entry should be returned to the system with 2168 * cache_put() or cache_unlock() + cache_drop(). 2169 * 2170 * namecache locks are recursive but care must be taken to avoid lock order 2171 * reversals (hence why the passed par_nch must be unlocked). Locking 2172 * rules are to order for parent traversals, not for child traversals. 2173 * 2174 * Nobody else will be able to manipulate the associated namespace (e.g. 2175 * create, delete, rename, rename-target) until the caller unlocks the 2176 * entry. 2177 * 2178 * The returned entry will be in one of three states: positive hit (non-null 2179 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set). 2180 * Unresolved entries must be resolved through the filesystem to associate the 2181 * vnode and/or determine whether a positive or negative hit has occured. 2182 * 2183 * It is not necessary to lock a directory in order to lock namespace under 2184 * that directory. In fact, it is explicitly not allowed to do that. A 2185 * directory is typically only locked when being created, renamed, or 2186 * destroyed. 2187 * 2188 * The directory (par) may be unresolved, in which case any returned child 2189 * will likely also be marked unresolved. Likely but not guarenteed. Since 2190 * the filesystem lookup requires a resolved directory vnode the caller is 2191 * responsible for resolving the namecache chain top-down. This API 2192 * specifically allows whole chains to be created in an unresolved state. 2193 */ 2194 struct nchandle 2195 cache_nlookup(struct nchandle *par_nch, struct nlcomponent *nlc) 2196 { 2197 struct nchandle nch; 2198 struct namecache *ncp; 2199 struct namecache *new_ncp; 2200 struct nchash_head *nchpp; 2201 struct mount *mp; 2202 u_int32_t hash; 2203 globaldata_t gd; 2204 int par_locked; 2205 2206 numcalls++; 2207 gd = mycpu; 2208 mp = par_nch->mount; 2209 par_locked = 0; 2210 2211 /* 2212 * This is a good time to call it, no ncp's are locked by 2213 * the caller or us. 2214 */ 2215 cache_hysteresis(); 2216 2217 /* 2218 * Try to locate an existing entry 2219 */ 2220 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 2221 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash); 2222 new_ncp = NULL; 2223 nchpp = NCHHASH(hash); 2224 restart: 2225 spin_lock(&nchpp->spin); 2226 LIST_FOREACH(ncp, &nchpp->list, nc_hash) { 2227 numchecks++; 2228 2229 /* 2230 * Break out if we find a matching entry. Note that 2231 * UNRESOLVED entries may match, but DESTROYED entries 2232 * do not. 2233 */ 2234 if (ncp->nc_parent == par_nch->ncp && 2235 ncp->nc_nlen == nlc->nlc_namelen && 2236 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 2237 (ncp->nc_flag & NCF_DESTROYED) == 0 2238 ) { 2239 _cache_hold(ncp); 2240 spin_unlock(&nchpp->spin); 2241 if (par_locked) { 2242 _cache_unlock(par_nch->ncp); 2243 par_locked = 0; 2244 } 2245 if (_cache_lock_special(ncp) == 0) { 2246 _cache_auto_unresolve(mp, ncp); 2247 if (new_ncp) 2248 _cache_free(new_ncp); 2249 goto found; 2250 } 2251 _cache_get(ncp); 2252 _cache_put(ncp); 2253 _cache_drop(ncp); 2254 goto restart; 2255 } 2256 } 2257 2258 /* 2259 * We failed to locate an entry, create a new entry and add it to 2260 * the cache. The parent ncp must also be locked so we 2261 * can link into it. 2262 * 2263 * We have to relookup after possibly blocking in kmalloc or 2264 * when locking par_nch. 2265 * 2266 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special 2267 * mount case, in which case nc_name will be NULL. 2268 */ 2269 if (new_ncp == NULL) { 2270 spin_unlock(&nchpp->spin); 2271 new_ncp = cache_alloc(nlc->nlc_namelen); 2272 if (nlc->nlc_namelen) { 2273 bcopy(nlc->nlc_nameptr, new_ncp->nc_name, 2274 nlc->nlc_namelen); 2275 new_ncp->nc_name[nlc->nlc_namelen] = 0; 2276 } 2277 goto restart; 2278 } 2279 if (par_locked == 0) { 2280 spin_unlock(&nchpp->spin); 2281 _cache_lock(par_nch->ncp); 2282 par_locked = 1; 2283 goto restart; 2284 } 2285 2286 /* 2287 * WARNING! We still hold the spinlock. We have to set the hash 2288 * table entry atomically. 2289 */ 2290 ncp = new_ncp; 2291 _cache_link_parent(ncp, par_nch->ncp, nchpp); 2292 spin_unlock(&nchpp->spin); 2293 _cache_unlock(par_nch->ncp); 2294 /* par_locked = 0 - not used */ 2295 found: 2296 /* 2297 * stats and namecache size management 2298 */ 2299 if (ncp->nc_flag & NCF_UNRESOLVED) 2300 ++gd->gd_nchstats->ncs_miss; 2301 else if (ncp->nc_vp) 2302 ++gd->gd_nchstats->ncs_goodhits; 2303 else 2304 ++gd->gd_nchstats->ncs_neghits; 2305 nch.mount = mp; 2306 nch.ncp = ncp; 2307 atomic_add_int(&nch.mount->mnt_refs, 1); 2308 return(nch); 2309 } 2310 2311 /* 2312 * This is a non-blocking verison of cache_nlookup() used by 2313 * nfs_readdirplusrpc_uio(). It can fail for any reason and 2314 * will return nch.ncp == NULL in that case. 2315 */ 2316 struct nchandle 2317 cache_nlookup_nonblock(struct nchandle *par_nch, struct nlcomponent *nlc) 2318 { 2319 struct nchandle nch; 2320 struct namecache *ncp; 2321 struct namecache *new_ncp; 2322 struct nchash_head *nchpp; 2323 struct mount *mp; 2324 u_int32_t hash; 2325 globaldata_t gd; 2326 int par_locked; 2327 2328 numcalls++; 2329 gd = mycpu; 2330 mp = par_nch->mount; 2331 par_locked = 0; 2332 2333 /* 2334 * Try to locate an existing entry 2335 */ 2336 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 2337 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash); 2338 new_ncp = NULL; 2339 nchpp = NCHHASH(hash); 2340 restart: 2341 spin_lock(&nchpp->spin); 2342 LIST_FOREACH(ncp, &nchpp->list, nc_hash) { 2343 numchecks++; 2344 2345 /* 2346 * Break out if we find a matching entry. Note that 2347 * UNRESOLVED entries may match, but DESTROYED entries 2348 * do not. 2349 */ 2350 if (ncp->nc_parent == par_nch->ncp && 2351 ncp->nc_nlen == nlc->nlc_namelen && 2352 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 2353 (ncp->nc_flag & NCF_DESTROYED) == 0 2354 ) { 2355 _cache_hold(ncp); 2356 spin_unlock(&nchpp->spin); 2357 if (par_locked) { 2358 _cache_unlock(par_nch->ncp); 2359 par_locked = 0; 2360 } 2361 if (_cache_lock_special(ncp) == 0) { 2362 _cache_auto_unresolve(mp, ncp); 2363 if (new_ncp) { 2364 _cache_free(new_ncp); 2365 new_ncp = NULL; 2366 } 2367 goto found; 2368 } 2369 _cache_drop(ncp); 2370 goto failed; 2371 } 2372 } 2373 2374 /* 2375 * We failed to locate an entry, create a new entry and add it to 2376 * the cache. The parent ncp must also be locked so we 2377 * can link into it. 2378 * 2379 * We have to relookup after possibly blocking in kmalloc or 2380 * when locking par_nch. 2381 * 2382 * NOTE: nlc_namelen can be 0 and nlc_nameptr NULL as a special 2383 * mount case, in which case nc_name will be NULL. 2384 */ 2385 if (new_ncp == NULL) { 2386 spin_unlock(&nchpp->spin); 2387 new_ncp = cache_alloc(nlc->nlc_namelen); 2388 if (nlc->nlc_namelen) { 2389 bcopy(nlc->nlc_nameptr, new_ncp->nc_name, 2390 nlc->nlc_namelen); 2391 new_ncp->nc_name[nlc->nlc_namelen] = 0; 2392 } 2393 goto restart; 2394 } 2395 if (par_locked == 0) { 2396 spin_unlock(&nchpp->spin); 2397 if (_cache_lock_nonblock(par_nch->ncp) == 0) { 2398 par_locked = 1; 2399 goto restart; 2400 } 2401 goto failed; 2402 } 2403 2404 /* 2405 * WARNING! We still hold the spinlock. We have to set the hash 2406 * table entry atomically. 2407 */ 2408 ncp = new_ncp; 2409 _cache_link_parent(ncp, par_nch->ncp, nchpp); 2410 spin_unlock(&nchpp->spin); 2411 _cache_unlock(par_nch->ncp); 2412 /* par_locked = 0 - not used */ 2413 found: 2414 /* 2415 * stats and namecache size management 2416 */ 2417 if (ncp->nc_flag & NCF_UNRESOLVED) 2418 ++gd->gd_nchstats->ncs_miss; 2419 else if (ncp->nc_vp) 2420 ++gd->gd_nchstats->ncs_goodhits; 2421 else 2422 ++gd->gd_nchstats->ncs_neghits; 2423 nch.mount = mp; 2424 nch.ncp = ncp; 2425 atomic_add_int(&nch.mount->mnt_refs, 1); 2426 return(nch); 2427 failed: 2428 if (new_ncp) { 2429 _cache_free(new_ncp); 2430 new_ncp = NULL; 2431 } 2432 nch.mount = NULL; 2433 nch.ncp = NULL; 2434 return(nch); 2435 } 2436 2437 /* 2438 * The namecache entry is marked as being used as a mount point. 2439 * Locate the mount if it is visible to the caller. 2440 */ 2441 struct findmount_info { 2442 struct mount *result; 2443 struct mount *nch_mount; 2444 struct namecache *nch_ncp; 2445 }; 2446 2447 static 2448 int 2449 cache_findmount_callback(struct mount *mp, void *data) 2450 { 2451 struct findmount_info *info = data; 2452 2453 /* 2454 * Check the mount's mounted-on point against the passed nch. 2455 */ 2456 if (mp->mnt_ncmounton.mount == info->nch_mount && 2457 mp->mnt_ncmounton.ncp == info->nch_ncp 2458 ) { 2459 info->result = mp; 2460 return(-1); 2461 } 2462 return(0); 2463 } 2464 2465 struct mount * 2466 cache_findmount(struct nchandle *nch) 2467 { 2468 struct findmount_info info; 2469 2470 info.result = NULL; 2471 info.nch_mount = nch->mount; 2472 info.nch_ncp = nch->ncp; 2473 mountlist_scan(cache_findmount_callback, &info, 2474 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 2475 return(info.result); 2476 } 2477 2478 /* 2479 * Resolve an unresolved namecache entry, generally by looking it up. 2480 * The passed ncp must be locked and refd. 2481 * 2482 * Theoretically since a vnode cannot be recycled while held, and since 2483 * the nc_parent chain holds its vnode as long as children exist, the 2484 * direct parent of the cache entry we are trying to resolve should 2485 * have a valid vnode. If not then generate an error that we can 2486 * determine is related to a resolver bug. 2487 * 2488 * However, if a vnode was in the middle of a recyclement when the NCP 2489 * got locked, ncp->nc_vp might point to a vnode that is about to become 2490 * invalid. cache_resolve() handles this case by unresolving the entry 2491 * and then re-resolving it. 2492 * 2493 * Note that successful resolution does not necessarily return an error 2494 * code of 0. If the ncp resolves to a negative cache hit then ENOENT 2495 * will be returned. 2496 * 2497 * MPSAFE 2498 */ 2499 int 2500 cache_resolve(struct nchandle *nch, struct ucred *cred) 2501 { 2502 struct namecache *par_tmp; 2503 struct namecache *par; 2504 struct namecache *ncp; 2505 struct nchandle nctmp; 2506 struct mount *mp; 2507 struct vnode *dvp; 2508 int error; 2509 2510 ncp = nch->ncp; 2511 mp = nch->mount; 2512 restart: 2513 /* 2514 * If the ncp is already resolved we have nothing to do. However, 2515 * we do want to guarentee that a usable vnode is returned when 2516 * a vnode is present, so make sure it hasn't been reclaimed. 2517 */ 2518 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 2519 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 2520 _cache_setunresolved(ncp); 2521 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 2522 return (ncp->nc_error); 2523 } 2524 2525 /* 2526 * Mount points need special handling because the parent does not 2527 * belong to the same filesystem as the ncp. 2528 */ 2529 if (ncp == mp->mnt_ncmountpt.ncp) 2530 return (cache_resolve_mp(mp)); 2531 2532 /* 2533 * We expect an unbroken chain of ncps to at least the mount point, 2534 * and even all the way to root (but this code doesn't have to go 2535 * past the mount point). 2536 */ 2537 if (ncp->nc_parent == NULL) { 2538 kprintf("EXDEV case 1 %p %*.*s\n", ncp, 2539 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 2540 ncp->nc_error = EXDEV; 2541 return(ncp->nc_error); 2542 } 2543 2544 /* 2545 * The vp's of the parent directories in the chain are held via vhold() 2546 * due to the existance of the child, and should not disappear. 2547 * However, there are cases where they can disappear: 2548 * 2549 * - due to filesystem I/O errors. 2550 * - due to NFS being stupid about tracking the namespace and 2551 * destroys the namespace for entire directories quite often. 2552 * - due to forced unmounts. 2553 * - due to an rmdir (parent will be marked DESTROYED) 2554 * 2555 * When this occurs we have to track the chain backwards and resolve 2556 * it, looping until the resolver catches up to the current node. We 2557 * could recurse here but we might run ourselves out of kernel stack 2558 * so we do it in a more painful manner. This situation really should 2559 * not occur all that often, or if it does not have to go back too 2560 * many nodes to resolve the ncp. 2561 */ 2562 while ((dvp = cache_dvpref(ncp)) == NULL) { 2563 /* 2564 * This case can occur if a process is CD'd into a 2565 * directory which is then rmdir'd. If the parent is marked 2566 * destroyed there is no point trying to resolve it. 2567 */ 2568 if (ncp->nc_parent->nc_flag & NCF_DESTROYED) 2569 return(ENOENT); 2570 par = ncp->nc_parent; 2571 _cache_hold(par); 2572 _cache_lock(par); 2573 while ((par_tmp = par->nc_parent) != NULL && 2574 par_tmp->nc_vp == NULL) { 2575 _cache_hold(par_tmp); 2576 _cache_lock(par_tmp); 2577 _cache_put(par); 2578 par = par_tmp; 2579 } 2580 if (par->nc_parent == NULL) { 2581 kprintf("EXDEV case 2 %*.*s\n", 2582 par->nc_nlen, par->nc_nlen, par->nc_name); 2583 _cache_put(par); 2584 return (EXDEV); 2585 } 2586 kprintf("[diagnostic] cache_resolve: had to recurse on %*.*s\n", 2587 par->nc_nlen, par->nc_nlen, par->nc_name); 2588 /* 2589 * The parent is not set in stone, ref and lock it to prevent 2590 * it from disappearing. Also note that due to renames it 2591 * is possible for our ncp to move and for par to no longer 2592 * be one of its parents. We resolve it anyway, the loop 2593 * will handle any moves. 2594 */ 2595 _cache_get(par); /* additional hold/lock */ 2596 _cache_put(par); /* from earlier hold/lock */ 2597 if (par == nch->mount->mnt_ncmountpt.ncp) { 2598 cache_resolve_mp(nch->mount); 2599 } else if ((dvp = cache_dvpref(par)) == NULL) { 2600 kprintf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name); 2601 _cache_put(par); 2602 continue; 2603 } else { 2604 if (par->nc_flag & NCF_UNRESOLVED) { 2605 nctmp.mount = mp; 2606 nctmp.ncp = par; 2607 par->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred); 2608 } 2609 vrele(dvp); 2610 } 2611 if ((error = par->nc_error) != 0) { 2612 if (par->nc_error != EAGAIN) { 2613 kprintf("EXDEV case 3 %*.*s error %d\n", 2614 par->nc_nlen, par->nc_nlen, par->nc_name, 2615 par->nc_error); 2616 _cache_put(par); 2617 return(error); 2618 } 2619 kprintf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n", 2620 par, par->nc_nlen, par->nc_nlen, par->nc_name); 2621 } 2622 _cache_put(par); 2623 /* loop */ 2624 } 2625 2626 /* 2627 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected 2628 * ncp's and reattach them. If this occurs the original ncp is marked 2629 * EAGAIN to force a relookup. 2630 * 2631 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed 2632 * ncp must already be resolved. 2633 */ 2634 if (dvp) { 2635 nctmp.mount = mp; 2636 nctmp.ncp = ncp; 2637 ncp->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred); 2638 vrele(dvp); 2639 } else { 2640 ncp->nc_error = EPERM; 2641 } 2642 if (ncp->nc_error == EAGAIN) { 2643 kprintf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n", 2644 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 2645 goto restart; 2646 } 2647 return(ncp->nc_error); 2648 } 2649 2650 /* 2651 * Resolve the ncp associated with a mount point. Such ncp's almost always 2652 * remain resolved and this routine is rarely called. NFS MPs tends to force 2653 * re-resolution more often due to its mac-truck-smash-the-namecache 2654 * method of tracking namespace changes. 2655 * 2656 * The semantics for this call is that the passed ncp must be locked on 2657 * entry and will be locked on return. However, if we actually have to 2658 * resolve the mount point we temporarily unlock the entry in order to 2659 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of 2660 * the unlock we have to recheck the flags after we relock. 2661 */ 2662 static int 2663 cache_resolve_mp(struct mount *mp) 2664 { 2665 struct namecache *ncp = mp->mnt_ncmountpt.ncp; 2666 struct vnode *vp; 2667 int error; 2668 2669 KKASSERT(mp != NULL); 2670 2671 /* 2672 * If the ncp is already resolved we have nothing to do. However, 2673 * we do want to guarentee that a usable vnode is returned when 2674 * a vnode is present, so make sure it hasn't been reclaimed. 2675 */ 2676 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 2677 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 2678 _cache_setunresolved(ncp); 2679 } 2680 2681 if (ncp->nc_flag & NCF_UNRESOLVED) { 2682 _cache_unlock(ncp); 2683 while (vfs_busy(mp, 0)) 2684 ; 2685 error = VFS_ROOT(mp, &vp); 2686 _cache_lock(ncp); 2687 2688 /* 2689 * recheck the ncp state after relocking. 2690 */ 2691 if (ncp->nc_flag & NCF_UNRESOLVED) { 2692 ncp->nc_error = error; 2693 if (error == 0) { 2694 _cache_setvp(mp, ncp, vp); 2695 vput(vp); 2696 } else { 2697 kprintf("[diagnostic] cache_resolve_mp: failed" 2698 " to resolve mount %p err=%d ncp=%p\n", 2699 mp, error, ncp); 2700 _cache_setvp(mp, ncp, NULL); 2701 } 2702 } else if (error == 0) { 2703 vput(vp); 2704 } 2705 vfs_unbusy(mp); 2706 } 2707 return(ncp->nc_error); 2708 } 2709 2710 /* 2711 * Clean out negative cache entries when too many have accumulated. 2712 * 2713 * MPSAFE 2714 */ 2715 static void 2716 _cache_cleanneg(int count) 2717 { 2718 struct namecache *ncp; 2719 2720 /* 2721 * Automode from the vnlru proc - clean out 10% of the negative cache 2722 * entries. 2723 */ 2724 if (count == 0) 2725 count = numneg / 10 + 1; 2726 2727 /* 2728 * Attempt to clean out the specified number of negative cache 2729 * entries. 2730 */ 2731 while (count) { 2732 spin_lock(&ncspin); 2733 ncp = TAILQ_FIRST(&ncneglist); 2734 if (ncp == NULL) { 2735 spin_unlock(&ncspin); 2736 break; 2737 } 2738 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 2739 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 2740 _cache_hold(ncp); 2741 spin_unlock(&ncspin); 2742 if (_cache_lock_special(ncp) == 0) { 2743 ncp = cache_zap(ncp, 1); 2744 if (ncp) 2745 _cache_drop(ncp); 2746 } else { 2747 _cache_drop(ncp); 2748 } 2749 --count; 2750 } 2751 } 2752 2753 /* 2754 * This is a kitchen sink function to clean out ncps which we 2755 * tried to zap from cache_drop() but failed because we were 2756 * unable to acquire the parent lock. 2757 * 2758 * Such entries can also be removed via cache_inval_vp(), such 2759 * as when unmounting. 2760 * 2761 * MPSAFE 2762 */ 2763 static void 2764 _cache_cleandefered(void) 2765 { 2766 struct nchash_head *nchpp; 2767 struct namecache *ncp; 2768 struct namecache dummy; 2769 int i; 2770 2771 numdefered = 0; 2772 bzero(&dummy, sizeof(dummy)); 2773 dummy.nc_flag = NCF_DESTROYED; 2774 2775 for (i = 0; i <= nchash; ++i) { 2776 nchpp = &nchashtbl[i]; 2777 2778 spin_lock(&nchpp->spin); 2779 LIST_INSERT_HEAD(&nchpp->list, &dummy, nc_hash); 2780 ncp = &dummy; 2781 while ((ncp = LIST_NEXT(ncp, nc_hash)) != NULL) { 2782 if ((ncp->nc_flag & NCF_DEFEREDZAP) == 0) 2783 continue; 2784 LIST_REMOVE(&dummy, nc_hash); 2785 LIST_INSERT_AFTER(ncp, &dummy, nc_hash); 2786 _cache_hold(ncp); 2787 spin_unlock(&nchpp->spin); 2788 if (_cache_lock_nonblock(ncp) == 0) { 2789 ncp->nc_flag &= ~NCF_DEFEREDZAP; 2790 _cache_unlock(ncp); 2791 } 2792 _cache_drop(ncp); 2793 spin_lock(&nchpp->spin); 2794 ncp = &dummy; 2795 } 2796 LIST_REMOVE(&dummy, nc_hash); 2797 spin_unlock(&nchpp->spin); 2798 } 2799 } 2800 2801 /* 2802 * Name cache initialization, from vfsinit() when we are booting 2803 */ 2804 void 2805 nchinit(void) 2806 { 2807 int i; 2808 globaldata_t gd; 2809 2810 /* initialise per-cpu namecache effectiveness statistics. */ 2811 for (i = 0; i < ncpus; ++i) { 2812 gd = globaldata_find(i); 2813 gd->gd_nchstats = &nchstats[i]; 2814 } 2815 TAILQ_INIT(&ncneglist); 2816 spin_init(&ncspin); 2817 nchashtbl = hashinit_ext(desiredvnodes*2, sizeof(struct nchash_head), 2818 M_VFSCACHE, &nchash); 2819 for (i = 0; i <= (int)nchash; ++i) { 2820 LIST_INIT(&nchashtbl[i].list); 2821 spin_init(&nchashtbl[i].spin); 2822 } 2823 nclockwarn = 5 * hz; 2824 } 2825 2826 /* 2827 * Called from start_init() to bootstrap the root filesystem. Returns 2828 * a referenced, unlocked namecache record. 2829 */ 2830 void 2831 cache_allocroot(struct nchandle *nch, struct mount *mp, struct vnode *vp) 2832 { 2833 nch->ncp = cache_alloc(0); 2834 nch->mount = mp; 2835 atomic_add_int(&mp->mnt_refs, 1); 2836 if (vp) 2837 _cache_setvp(nch->mount, nch->ncp, vp); 2838 } 2839 2840 /* 2841 * vfs_cache_setroot() 2842 * 2843 * Create an association between the root of our namecache and 2844 * the root vnode. This routine may be called several times during 2845 * booting. 2846 * 2847 * If the caller intends to save the returned namecache pointer somewhere 2848 * it must cache_hold() it. 2849 */ 2850 void 2851 vfs_cache_setroot(struct vnode *nvp, struct nchandle *nch) 2852 { 2853 struct vnode *ovp; 2854 struct nchandle onch; 2855 2856 ovp = rootvnode; 2857 onch = rootnch; 2858 rootvnode = nvp; 2859 if (nch) 2860 rootnch = *nch; 2861 else 2862 cache_zero(&rootnch); 2863 if (ovp) 2864 vrele(ovp); 2865 if (onch.ncp) 2866 cache_drop(&onch); 2867 } 2868 2869 /* 2870 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache 2871 * topology and is being removed as quickly as possible. The new VOP_N*() 2872 * API calls are required to make specific adjustments using the supplied 2873 * ncp pointers rather then just bogusly purging random vnodes. 2874 * 2875 * Invalidate all namecache entries to a particular vnode as well as 2876 * any direct children of that vnode in the namecache. This is a 2877 * 'catch all' purge used by filesystems that do not know any better. 2878 * 2879 * Note that the linkage between the vnode and its namecache entries will 2880 * be removed, but the namecache entries themselves might stay put due to 2881 * active references from elsewhere in the system or due to the existance of 2882 * the children. The namecache topology is left intact even if we do not 2883 * know what the vnode association is. Such entries will be marked 2884 * NCF_UNRESOLVED. 2885 */ 2886 void 2887 cache_purge(struct vnode *vp) 2888 { 2889 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN); 2890 } 2891 2892 /* 2893 * Flush all entries referencing a particular filesystem. 2894 * 2895 * Since we need to check it anyway, we will flush all the invalid 2896 * entries at the same time. 2897 */ 2898 #if 0 2899 2900 void 2901 cache_purgevfs(struct mount *mp) 2902 { 2903 struct nchash_head *nchpp; 2904 struct namecache *ncp, *nnp; 2905 2906 /* 2907 * Scan hash tables for applicable entries. 2908 */ 2909 for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) { 2910 spin_lock_wr(&nchpp->spin); XXX 2911 ncp = LIST_FIRST(&nchpp->list); 2912 if (ncp) 2913 _cache_hold(ncp); 2914 while (ncp) { 2915 nnp = LIST_NEXT(ncp, nc_hash); 2916 if (nnp) 2917 _cache_hold(nnp); 2918 if (ncp->nc_mount == mp) { 2919 _cache_lock(ncp); 2920 ncp = cache_zap(ncp, 0); 2921 if (ncp) 2922 _cache_drop(ncp); 2923 } else { 2924 _cache_drop(ncp); 2925 } 2926 ncp = nnp; 2927 } 2928 spin_unlock_wr(&nchpp->spin); XXX 2929 } 2930 } 2931 2932 #endif 2933 2934 static int disablecwd; 2935 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, ""); 2936 2937 static u_long numcwdcalls; STATNODE(CTLFLAG_RD, numcwdcalls, &numcwdcalls); 2938 static u_long numcwdfail1; STATNODE(CTLFLAG_RD, numcwdfail1, &numcwdfail1); 2939 static u_long numcwdfail2; STATNODE(CTLFLAG_RD, numcwdfail2, &numcwdfail2); 2940 static u_long numcwdfail3; STATNODE(CTLFLAG_RD, numcwdfail3, &numcwdfail3); 2941 static u_long numcwdfail4; STATNODE(CTLFLAG_RD, numcwdfail4, &numcwdfail4); 2942 static u_long numcwdfound; STATNODE(CTLFLAG_RD, numcwdfound, &numcwdfound); 2943 2944 /* 2945 * MPALMOSTSAFE 2946 */ 2947 int 2948 sys___getcwd(struct __getcwd_args *uap) 2949 { 2950 u_int buflen; 2951 int error; 2952 char *buf; 2953 char *bp; 2954 2955 if (disablecwd) 2956 return (ENODEV); 2957 2958 buflen = uap->buflen; 2959 if (buflen == 0) 2960 return (EINVAL); 2961 if (buflen > MAXPATHLEN) 2962 buflen = MAXPATHLEN; 2963 2964 buf = kmalloc(buflen, M_TEMP, M_WAITOK); 2965 get_mplock(); 2966 bp = kern_getcwd(buf, buflen, &error); 2967 rel_mplock(); 2968 if (error == 0) 2969 error = copyout(bp, uap->buf, strlen(bp) + 1); 2970 kfree(buf, M_TEMP); 2971 return (error); 2972 } 2973 2974 char * 2975 kern_getcwd(char *buf, size_t buflen, int *error) 2976 { 2977 struct proc *p = curproc; 2978 char *bp; 2979 int i, slash_prefixed; 2980 struct filedesc *fdp; 2981 struct nchandle nch; 2982 struct namecache *ncp; 2983 2984 numcwdcalls++; 2985 bp = buf; 2986 bp += buflen - 1; 2987 *bp = '\0'; 2988 fdp = p->p_fd; 2989 slash_prefixed = 0; 2990 2991 nch = fdp->fd_ncdir; 2992 ncp = nch.ncp; 2993 if (ncp) 2994 _cache_hold(ncp); 2995 2996 while (ncp && (ncp != fdp->fd_nrdir.ncp || 2997 nch.mount != fdp->fd_nrdir.mount) 2998 ) { 2999 /* 3000 * While traversing upwards if we encounter the root 3001 * of the current mount we have to skip to the mount point 3002 * in the underlying filesystem. 3003 */ 3004 if (ncp == nch.mount->mnt_ncmountpt.ncp) { 3005 nch = nch.mount->mnt_ncmounton; 3006 _cache_drop(ncp); 3007 ncp = nch.ncp; 3008 if (ncp) 3009 _cache_hold(ncp); 3010 continue; 3011 } 3012 3013 /* 3014 * Prepend the path segment 3015 */ 3016 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 3017 if (bp == buf) { 3018 numcwdfail4++; 3019 *error = ERANGE; 3020 bp = NULL; 3021 goto done; 3022 } 3023 *--bp = ncp->nc_name[i]; 3024 } 3025 if (bp == buf) { 3026 numcwdfail4++; 3027 *error = ERANGE; 3028 bp = NULL; 3029 goto done; 3030 } 3031 *--bp = '/'; 3032 slash_prefixed = 1; 3033 3034 /* 3035 * Go up a directory. This isn't a mount point so we don't 3036 * have to check again. 3037 */ 3038 while ((nch.ncp = ncp->nc_parent) != NULL) { 3039 _cache_lock(ncp); 3040 if (nch.ncp != ncp->nc_parent) { 3041 _cache_unlock(ncp); 3042 continue; 3043 } 3044 _cache_hold(nch.ncp); 3045 _cache_unlock(ncp); 3046 break; 3047 } 3048 _cache_drop(ncp); 3049 ncp = nch.ncp; 3050 } 3051 if (ncp == NULL) { 3052 numcwdfail2++; 3053 *error = ENOENT; 3054 bp = NULL; 3055 goto done; 3056 } 3057 if (!slash_prefixed) { 3058 if (bp == buf) { 3059 numcwdfail4++; 3060 *error = ERANGE; 3061 bp = NULL; 3062 goto done; 3063 } 3064 *--bp = '/'; 3065 } 3066 numcwdfound++; 3067 *error = 0; 3068 done: 3069 if (ncp) 3070 _cache_drop(ncp); 3071 return (bp); 3072 } 3073 3074 /* 3075 * Thus begins the fullpath magic. 3076 * 3077 * The passed nchp is referenced but not locked. 3078 */ 3079 #undef STATNODE 3080 #define STATNODE(name) \ 3081 static u_int name; \ 3082 SYSCTL_UINT(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, "") 3083 3084 static int disablefullpath; 3085 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, 3086 &disablefullpath, 0, ""); 3087 3088 STATNODE(numfullpathcalls); 3089 STATNODE(numfullpathfail1); 3090 STATNODE(numfullpathfail2); 3091 STATNODE(numfullpathfail3); 3092 STATNODE(numfullpathfail4); 3093 STATNODE(numfullpathfound); 3094 3095 int 3096 cache_fullpath(struct proc *p, struct nchandle *nchp, 3097 char **retbuf, char **freebuf, int guess) 3098 { 3099 struct nchandle fd_nrdir; 3100 struct nchandle nch; 3101 struct namecache *ncp; 3102 struct mount *mp, *new_mp; 3103 char *bp, *buf; 3104 int slash_prefixed; 3105 int error = 0; 3106 int i; 3107 3108 atomic_add_int(&numfullpathcalls, -1); 3109 3110 *retbuf = NULL; 3111 *freebuf = NULL; 3112 3113 buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK); 3114 bp = buf + MAXPATHLEN - 1; 3115 *bp = '\0'; 3116 if (p != NULL) 3117 fd_nrdir = p->p_fd->fd_nrdir; 3118 else 3119 fd_nrdir = rootnch; 3120 slash_prefixed = 0; 3121 nch = *nchp; 3122 ncp = nch.ncp; 3123 if (ncp) 3124 _cache_hold(ncp); 3125 mp = nch.mount; 3126 3127 while (ncp && (ncp != fd_nrdir.ncp || mp != fd_nrdir.mount)) { 3128 new_mp = NULL; 3129 3130 /* 3131 * If we are asked to guess the upwards path, we do so whenever 3132 * we encounter an ncp marked as a mountpoint. We try to find 3133 * the actual mountpoint by finding the mountpoint with this ncp. 3134 */ 3135 if (guess && (ncp->nc_flag & NCF_ISMOUNTPT)) { 3136 new_mp = mount_get_by_nc(ncp); 3137 } 3138 /* 3139 * While traversing upwards if we encounter the root 3140 * of the current mount we have to skip to the mount point. 3141 */ 3142 if (ncp == mp->mnt_ncmountpt.ncp) { 3143 new_mp = mp; 3144 } 3145 if (new_mp) { 3146 nch = new_mp->mnt_ncmounton; 3147 _cache_drop(ncp); 3148 ncp = nch.ncp; 3149 if (ncp) 3150 _cache_hold(ncp); 3151 mp = nch.mount; 3152 continue; 3153 } 3154 3155 /* 3156 * Prepend the path segment 3157 */ 3158 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 3159 if (bp == buf) { 3160 numfullpathfail4++; 3161 kfree(buf, M_TEMP); 3162 error = ENOMEM; 3163 goto done; 3164 } 3165 *--bp = ncp->nc_name[i]; 3166 } 3167 if (bp == buf) { 3168 numfullpathfail4++; 3169 kfree(buf, M_TEMP); 3170 error = ENOMEM; 3171 goto done; 3172 } 3173 *--bp = '/'; 3174 slash_prefixed = 1; 3175 3176 /* 3177 * Go up a directory. This isn't a mount point so we don't 3178 * have to check again. 3179 * 3180 * We can only safely access nc_parent with ncp held locked. 3181 */ 3182 while ((nch.ncp = ncp->nc_parent) != NULL) { 3183 _cache_lock(ncp); 3184 if (nch.ncp != ncp->nc_parent) { 3185 _cache_unlock(ncp); 3186 continue; 3187 } 3188 _cache_hold(nch.ncp); 3189 _cache_unlock(ncp); 3190 break; 3191 } 3192 _cache_drop(ncp); 3193 ncp = nch.ncp; 3194 } 3195 if (ncp == NULL) { 3196 numfullpathfail2++; 3197 kfree(buf, M_TEMP); 3198 error = ENOENT; 3199 goto done; 3200 } 3201 3202 if (!slash_prefixed) { 3203 if (bp == buf) { 3204 numfullpathfail4++; 3205 kfree(buf, M_TEMP); 3206 error = ENOMEM; 3207 goto done; 3208 } 3209 *--bp = '/'; 3210 } 3211 numfullpathfound++; 3212 *retbuf = bp; 3213 *freebuf = buf; 3214 error = 0; 3215 done: 3216 if (ncp) 3217 _cache_drop(ncp); 3218 return(error); 3219 } 3220 3221 int 3222 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf, int guess) 3223 { 3224 struct namecache *ncp; 3225 struct nchandle nch; 3226 int error; 3227 3228 atomic_add_int(&numfullpathcalls, 1); 3229 if (disablefullpath) 3230 return (ENODEV); 3231 3232 if (p == NULL) 3233 return (EINVAL); 3234 3235 /* vn is NULL, client wants us to use p->p_textvp */ 3236 if (vn == NULL) { 3237 if ((vn = p->p_textvp) == NULL) 3238 return (EINVAL); 3239 } 3240 spin_lock(&vn->v_spinlock); 3241 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) { 3242 if (ncp->nc_nlen) 3243 break; 3244 } 3245 if (ncp == NULL) { 3246 spin_unlock(&vn->v_spinlock); 3247 return (EINVAL); 3248 } 3249 _cache_hold(ncp); 3250 spin_unlock(&vn->v_spinlock); 3251 3252 atomic_add_int(&numfullpathcalls, -1); 3253 nch.ncp = ncp;; 3254 nch.mount = vn->v_mount; 3255 error = cache_fullpath(p, &nch, retbuf, freebuf, guess); 3256 _cache_drop(ncp); 3257 return (error); 3258 } 3259