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