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