1 /* 2 * Copyright (c) 2003,2004 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 * @(#)vfs_cache.c 8.5 (Berkeley) 3/22/95 69 * $FreeBSD: src/sys/kern/vfs_cache.c,v 1.42.2.6 2001/10/05 20:07:03 dillon Exp $ 70 * $DragonFly: src/sys/kern/vfs_cache.c,v 1.54 2005/04/19 17:54:42 dillon Exp $ 71 */ 72 73 #include <sys/param.h> 74 #include <sys/systm.h> 75 #include <sys/kernel.h> 76 #include <sys/sysctl.h> 77 #include <sys/mount.h> 78 #include <sys/vnode.h> 79 #include <sys/malloc.h> 80 #include <sys/sysproto.h> 81 #include <sys/proc.h> 82 #include <sys/namei.h> 83 #include <sys/nlookup.h> 84 #include <sys/filedesc.h> 85 #include <sys/fnv_hash.h> 86 #include <sys/globaldata.h> 87 #include <sys/kern_syscall.h> 88 #include <sys/dirent.h> 89 #include <ddb/ddb.h> 90 91 /* 92 * Random lookups in the cache are accomplished with a hash table using 93 * a hash key of (nc_src_vp, name). 94 * 95 * Negative entries may exist and correspond to structures where nc_vp 96 * is NULL. In a negative entry, NCF_WHITEOUT will be set if the entry 97 * corresponds to a whited-out directory entry (verses simply not finding the 98 * entry at all). 99 * 100 * Upon reaching the last segment of a path, if the reference is for DELETE, 101 * or NOCACHE is set (rewrite), and the name is located in the cache, it 102 * will be dropped. 103 */ 104 105 /* 106 * Structures associated with name cacheing. 107 */ 108 #define NCHHASH(hash) (&nchashtbl[(hash) & nchash]) 109 #define MINNEG 1024 110 111 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries"); 112 113 static LIST_HEAD(nchashhead, namecache) *nchashtbl; /* Hash Table */ 114 static struct namecache_list ncneglist; /* instead of vnode */ 115 116 /* 117 * ncvp_debug - debug cache_fromvp(). This is used by the NFS server 118 * to create the namecache infrastructure leading to a dangling vnode. 119 * 120 * 0 Only errors are reported 121 * 1 Successes are reported 122 * 2 Successes + the whole directory scan is reported 123 * 3 Force the directory scan code run as if the parent vnode did not 124 * have a namecache record, even if it does have one. 125 */ 126 static int ncvp_debug; 127 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0, ""); 128 129 static u_long nchash; /* size of hash table */ 130 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, ""); 131 132 static u_long ncnegfactor = 16; /* ratio of negative entries */ 133 SYSCTL_ULONG(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0, ""); 134 135 static int nclockwarn; /* warn on locked entries in ticks */ 136 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0, ""); 137 138 static u_long numneg; /* number of cache entries allocated */ 139 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0, ""); 140 141 static u_long numcache; /* number of cache entries allocated */ 142 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0, ""); 143 144 static u_long numunres; /* number of unresolved entries */ 145 SYSCTL_ULONG(_debug, OID_AUTO, numunres, CTLFLAG_RD, &numunres, 0, ""); 146 147 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode), ""); 148 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache), ""); 149 150 static int cache_resolve_mp(struct namecache *ncp); 151 static void cache_rehash(struct namecache *ncp); 152 153 /* 154 * The new name cache statistics 155 */ 156 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics"); 157 #define STATNODE(mode, name, var) \ 158 SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, mode, var, 0, ""); 159 STATNODE(CTLFLAG_RD, numneg, &numneg); 160 STATNODE(CTLFLAG_RD, numcache, &numcache); 161 static u_long numcalls; STATNODE(CTLFLAG_RD, numcalls, &numcalls); 162 static u_long dothits; STATNODE(CTLFLAG_RD, dothits, &dothits); 163 static u_long dotdothits; STATNODE(CTLFLAG_RD, dotdothits, &dotdothits); 164 static u_long numchecks; STATNODE(CTLFLAG_RD, numchecks, &numchecks); 165 static u_long nummiss; STATNODE(CTLFLAG_RD, nummiss, &nummiss); 166 static u_long nummisszap; STATNODE(CTLFLAG_RD, nummisszap, &nummisszap); 167 static u_long numposzaps; STATNODE(CTLFLAG_RD, numposzaps, &numposzaps); 168 static u_long numposhits; STATNODE(CTLFLAG_RD, numposhits, &numposhits); 169 static u_long numnegzaps; STATNODE(CTLFLAG_RD, numnegzaps, &numnegzaps); 170 static u_long numneghits; STATNODE(CTLFLAG_RD, numneghits, &numneghits); 171 172 struct nchstats nchstats[SMP_MAXCPU]; 173 /* 174 * Export VFS cache effectiveness statistics to user-land. 175 * 176 * The statistics are left for aggregation to user-land so 177 * neat things can be achieved, like observing per-CPU cache 178 * distribution. 179 */ 180 static int 181 sysctl_nchstats(SYSCTL_HANDLER_ARGS) 182 { 183 struct globaldata *gd; 184 int i, error; 185 186 error = 0; 187 for (i = 0; i < ncpus; ++i) { 188 gd = globaldata_find(i); 189 if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats), 190 sizeof(struct nchstats)))) 191 break; 192 } 193 194 return (error); 195 } 196 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD, 197 0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics"); 198 199 static void cache_zap(struct namecache *ncp); 200 201 /* 202 * cache_hold() and cache_drop() prevent the premature deletion of a 203 * namecache entry but do not prevent operations (such as zapping) on 204 * that namecache entry. 205 */ 206 static __inline 207 struct namecache * 208 _cache_hold(struct namecache *ncp) 209 { 210 ++ncp->nc_refs; 211 return(ncp); 212 } 213 214 /* 215 * When dropping an entry, if only one ref remains and the entry has not 216 * been resolved, zap it. Since the one reference is being dropped the 217 * entry had better not be locked. 218 */ 219 static __inline 220 void 221 _cache_drop(struct namecache *ncp) 222 { 223 KKASSERT(ncp->nc_refs > 0); 224 if (ncp->nc_refs == 1 && 225 (ncp->nc_flag & NCF_UNRESOLVED) && 226 TAILQ_EMPTY(&ncp->nc_list) 227 ) { 228 KKASSERT(ncp->nc_exlocks == 0); 229 cache_lock(ncp); 230 cache_zap(ncp); 231 } else { 232 --ncp->nc_refs; 233 } 234 } 235 236 /* 237 * Link a new namecache entry to its parent. Be careful to avoid races 238 * if vhold() blocks in the future. 239 * 240 * If we are creating a child under an oldapi parent we must mark the 241 * child as being an oldapi entry as well. 242 */ 243 static void 244 cache_link_parent(struct namecache *ncp, struct namecache *par) 245 { 246 KKASSERT(ncp->nc_parent == NULL); 247 ncp->nc_parent = par; 248 if (TAILQ_EMPTY(&par->nc_list)) { 249 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry); 250 /* 251 * Any vp associated with an ncp which has children must 252 * be held to prevent it from being recycled. 253 */ 254 if (par->nc_vp) 255 vhold(par->nc_vp); 256 } else { 257 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry); 258 } 259 } 260 261 /* 262 * Remove the parent association from a namecache structure. If this is 263 * the last child of the parent the cache_drop(par) will attempt to 264 * recursively zap the parent. 265 */ 266 static void 267 cache_unlink_parent(struct namecache *ncp) 268 { 269 struct namecache *par; 270 271 if ((par = ncp->nc_parent) != NULL) { 272 ncp->nc_parent = NULL; 273 par = cache_hold(par); 274 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 275 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 276 vdrop(par->nc_vp); 277 cache_drop(par); 278 } 279 } 280 281 /* 282 * Allocate a new namecache structure. Most of the code does not require 283 * zero-termination of the string but it makes vop_compat_ncreate() easier. 284 */ 285 static struct namecache * 286 cache_alloc(int nlen) 287 { 288 struct namecache *ncp; 289 290 ncp = malloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO); 291 if (nlen) 292 ncp->nc_name = malloc(nlen + 1, M_VFSCACHE, M_WAITOK); 293 ncp->nc_nlen = nlen; 294 ncp->nc_flag = NCF_UNRESOLVED; 295 ncp->nc_error = ENOTCONN; /* needs to be resolved */ 296 ncp->nc_refs = 1; 297 TAILQ_INIT(&ncp->nc_list); 298 cache_lock(ncp); 299 return(ncp); 300 } 301 302 static void 303 cache_free(struct namecache *ncp) 304 { 305 KKASSERT(ncp->nc_refs == 1 && ncp->nc_exlocks == 1); 306 if (ncp->nc_name) 307 free(ncp->nc_name, M_VFSCACHE); 308 free(ncp, M_VFSCACHE); 309 } 310 311 /* 312 * Ref and deref a namecache structure. 313 */ 314 struct namecache * 315 cache_hold(struct namecache *ncp) 316 { 317 return(_cache_hold(ncp)); 318 } 319 320 void 321 cache_drop(struct namecache *ncp) 322 { 323 _cache_drop(ncp); 324 } 325 326 /* 327 * Namespace locking. The caller must already hold a reference to the 328 * namecache structure in order to lock/unlock it. This function prevents 329 * the namespace from being created or destroyed by accessors other then 330 * the lock holder. 331 * 332 * Note that holding a locked namecache structure prevents other threads 333 * from making namespace changes (e.g. deleting or creating), prevents 334 * vnode association state changes by other threads, and prevents the 335 * namecache entry from being resolved or unresolved by other threads. 336 * 337 * The lock owner has full authority to associate/disassociate vnodes 338 * and resolve/unresolve the locked ncp. 339 * 340 * In particular, if a vnode is associated with a locked cache entry 341 * that vnode will *NOT* be recycled. We accomplish this by vhold()ing the 342 * vnode. XXX we should find a more efficient way to prevent the vnode 343 * from being recycled, but remember that any given vnode may have multiple 344 * namecache associations (think hardlinks). 345 */ 346 void 347 cache_lock(struct namecache *ncp) 348 { 349 thread_t td; 350 int didwarn; 351 352 KKASSERT(ncp->nc_refs != 0); 353 didwarn = 0; 354 td = curthread; 355 356 for (;;) { 357 if (ncp->nc_exlocks == 0) { 358 ncp->nc_exlocks = 1; 359 ncp->nc_locktd = td; 360 /* 361 * The vp associated with a locked ncp must be held 362 * to prevent it from being recycled (which would 363 * cause the ncp to become unresolved). 364 * 365 * XXX loop on race for later MPSAFE work. 366 */ 367 if (ncp->nc_vp) 368 vhold(ncp->nc_vp); 369 break; 370 } 371 if (ncp->nc_locktd == td) { 372 ++ncp->nc_exlocks; 373 break; 374 } 375 ncp->nc_flag |= NCF_LOCKREQ; 376 if (tsleep(ncp, 0, "clock", nclockwarn) == EWOULDBLOCK) { 377 if (didwarn) 378 continue; 379 didwarn = 1; 380 printf("[diagnostic] cache_lock: blocked on %p", ncp); 381 if ((ncp->nc_flag & NCF_MOUNTPT) && ncp->nc_mount) 382 printf(" [MOUNTFROM %s]\n", ncp->nc_mount->mnt_stat.f_mntfromname); 383 else 384 printf(" \"%*.*s\"\n", 385 ncp->nc_nlen, ncp->nc_nlen, 386 ncp->nc_name); 387 } 388 } 389 390 if (didwarn == 1) { 391 printf("[diagnostic] cache_lock: unblocked %*.*s\n", 392 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 393 } 394 } 395 396 int 397 cache_lock_nonblock(struct namecache *ncp) 398 { 399 thread_t td; 400 401 KKASSERT(ncp->nc_refs != 0); 402 td = curthread; 403 if (ncp->nc_exlocks == 0) { 404 ncp->nc_exlocks = 1; 405 ncp->nc_locktd = td; 406 /* 407 * The vp associated with a locked ncp must be held 408 * to prevent it from being recycled (which would 409 * cause the ncp to become unresolved). 410 * 411 * XXX loop on race for later MPSAFE work. 412 */ 413 if (ncp->nc_vp) 414 vhold(ncp->nc_vp); 415 return(0); 416 } else { 417 return(EWOULDBLOCK); 418 } 419 } 420 421 void 422 cache_unlock(struct namecache *ncp) 423 { 424 thread_t td = curthread; 425 426 KKASSERT(ncp->nc_refs > 0); 427 KKASSERT(ncp->nc_exlocks > 0); 428 KKASSERT(ncp->nc_locktd == td); 429 if (--ncp->nc_exlocks == 0) { 430 if (ncp->nc_vp) 431 vdrop(ncp->nc_vp); 432 ncp->nc_locktd = NULL; 433 if (ncp->nc_flag & NCF_LOCKREQ) { 434 ncp->nc_flag &= ~NCF_LOCKREQ; 435 wakeup(ncp); 436 } 437 } 438 } 439 440 /* 441 * ref-and-lock, unlock-and-deref functions. 442 */ 443 struct namecache * 444 cache_get(struct namecache *ncp) 445 { 446 _cache_hold(ncp); 447 cache_lock(ncp); 448 return(ncp); 449 } 450 451 int 452 cache_get_nonblock(struct namecache *ncp) 453 { 454 /* XXX MP */ 455 if (ncp->nc_exlocks == 0 || ncp->nc_locktd == curthread) { 456 _cache_hold(ncp); 457 cache_lock(ncp); 458 return(0); 459 } 460 return(EWOULDBLOCK); 461 } 462 463 void 464 cache_put(struct namecache *ncp) 465 { 466 cache_unlock(ncp); 467 _cache_drop(ncp); 468 } 469 470 /* 471 * Resolve an unresolved ncp by associating a vnode with it. If the 472 * vnode is NULL, a negative cache entry is created. 473 * 474 * The ncp should be locked on entry and will remain locked on return. 475 */ 476 void 477 cache_setvp(struct namecache *ncp, struct vnode *vp) 478 { 479 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED); 480 ncp->nc_vp = vp; 481 if (vp != NULL) { 482 /* 483 * Any vp associated with an ncp which has children must 484 * be held. Any vp associated with a locked ncp must be held. 485 */ 486 if (!TAILQ_EMPTY(&ncp->nc_list)) 487 vhold(vp); 488 TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode); 489 if (ncp->nc_exlocks) 490 vhold(vp); 491 492 /* 493 * Set auxillary flags 494 */ 495 switch(vp->v_type) { 496 case VDIR: 497 ncp->nc_flag |= NCF_ISDIR; 498 break; 499 case VLNK: 500 ncp->nc_flag |= NCF_ISSYMLINK; 501 /* XXX cache the contents of the symlink */ 502 break; 503 default: 504 break; 505 } 506 ++numcache; 507 ncp->nc_error = 0; 508 } else { 509 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 510 ++numneg; 511 ncp->nc_error = ENOENT; 512 } 513 ncp->nc_flag &= ~NCF_UNRESOLVED; 514 } 515 516 void 517 cache_settimeout(struct namecache *ncp, int nticks) 518 { 519 if ((ncp->nc_timeout = ticks + nticks) == 0) 520 ncp->nc_timeout = 1; 521 } 522 523 /* 524 * Disassociate the vnode or negative-cache association and mark a 525 * namecache entry as unresolved again. Note that the ncp is still 526 * left in the hash table and still linked to its parent. 527 * 528 * The ncp should be locked and refd on entry and will remain locked and refd 529 * on return. 530 * 531 * This routine is normally never called on a directory containing children. 532 * However, NFS often does just that in its rename() code as a cop-out to 533 * avoid complex namespace operations. This disconnects a directory vnode 534 * from its namecache and can cause the OLDAPI and NEWAPI to get out of 535 * sync. 536 */ 537 void 538 cache_setunresolved(struct namecache *ncp) 539 { 540 struct vnode *vp; 541 542 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 543 ncp->nc_flag |= NCF_UNRESOLVED; 544 ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK); 545 ncp->nc_timeout = 0; 546 ncp->nc_error = ENOTCONN; 547 ++numunres; 548 if ((vp = ncp->nc_vp) != NULL) { 549 --numcache; 550 ncp->nc_vp = NULL; 551 TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode); 552 553 /* 554 * Any vp associated with an ncp with children is 555 * held by that ncp. Any vp associated with a locked 556 * ncp is held by that ncp. These conditions must be 557 * undone when the vp is cleared out from the ncp. 558 */ 559 if (!TAILQ_EMPTY(&ncp->nc_list)) 560 vdrop(vp); 561 if (ncp->nc_exlocks) 562 vdrop(vp); 563 } else { 564 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 565 --numneg; 566 } 567 } 568 } 569 570 /* 571 * Invalidate portions of the namecache topology given a starting entry. 572 * The passed ncp is set to an unresolved state and: 573 * 574 * The passed ncp must be locked. 575 * 576 * CINV_DESTROY - Set a flag in the passed ncp entry indicating 577 * that the physical underlying nodes have been 578 * destroyed... as in deleted. For example, when 579 * a directory is removed. This will cause record 580 * lookups on the name to no longer be able to find 581 * the record and tells the resolver to return failure 582 * rather then trying to resolve through the parent. 583 * 584 * The topology itself, including ncp->nc_name, 585 * remains intact. 586 * 587 * This only applies to the passed ncp, if CINV_CHILDREN 588 * is specified the children are not flagged. 589 * 590 * CINV_CHILDREN - Set all children (recursively) to an unresolved 591 * state as well. 592 * 593 * Note that this will also have the side effect of 594 * cleaning out any unreferenced nodes in the topology 595 * from the leaves up as the recursion backs out. 596 * 597 * Note that the topology for any referenced nodes remains intact. 598 * 599 * It is possible for cache_inval() to race a cache_resolve(), meaning that 600 * the namecache entry may not actually be invalidated on return if it was 601 * revalidated while recursing down into its children. This code guarentees 602 * that the node(s) will go through an invalidation cycle, but does not 603 * guarentee that they will remain in an invalidated state. 604 * 605 * Returns non-zero if a revalidation was detected during the invalidation 606 * recursion, zero otherwise. Note that since only the original ncp is 607 * locked the revalidation ultimately can only indicate that the original ncp 608 * *MIGHT* no have been reresolved. 609 */ 610 int 611 cache_inval(struct namecache *ncp, int flags) 612 { 613 struct namecache *kid; 614 struct namecache *nextkid; 615 int rcnt = 0; 616 617 KKASSERT(ncp->nc_exlocks); 618 619 cache_setunresolved(ncp); 620 if (flags & CINV_DESTROY) 621 ncp->nc_flag |= NCF_DESTROYED; 622 623 if ((flags & CINV_CHILDREN) && 624 (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL 625 ) { 626 cache_hold(kid); 627 cache_unlock(ncp); 628 while (kid) { 629 if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL) 630 cache_hold(nextkid); 631 if ((kid->nc_flag & NCF_UNRESOLVED) == 0 || 632 TAILQ_FIRST(&kid->nc_list) 633 ) { 634 cache_lock(kid); 635 rcnt += cache_inval(kid, flags & ~CINV_DESTROY); 636 cache_unlock(kid); 637 } 638 cache_drop(kid); 639 kid = nextkid; 640 } 641 cache_lock(ncp); 642 } 643 644 /* 645 * Someone could have gotten in there while ncp was unlocked, 646 * retry if so. 647 */ 648 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 649 ++rcnt; 650 return (rcnt); 651 } 652 653 /* 654 * Invalidate a vnode's namecache associations. To avoid races against 655 * the resolver we do not invalidate a node which we previously invalidated 656 * but which was then re-resolved while we were in the invalidation loop. 657 * 658 * Returns non-zero if any namecache entries remain after the invalidation 659 * loop completed. 660 * 661 * NOTE: unlike the namecache topology which guarentees that ncp's will not 662 * be ripped out of the topology while held, the vnode's v_namecache list 663 * has no such restriction. NCP's can be ripped out of the list at virtually 664 * any time if not locked, even if held. 665 */ 666 int 667 cache_inval_vp(struct vnode *vp, int flags) 668 { 669 struct namecache *ncp; 670 struct namecache *next; 671 672 restart: 673 ncp = TAILQ_FIRST(&vp->v_namecache); 674 if (ncp) 675 cache_hold(ncp); 676 while (ncp) { 677 /* loop entered with ncp held */ 678 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL) 679 cache_hold(next); 680 cache_lock(ncp); 681 if (ncp->nc_vp != vp) { 682 printf("Warning: cache_inval_vp: race-A detected on " 683 "%s\n", ncp->nc_name); 684 cache_put(ncp); 685 if (next) 686 cache_drop(next); 687 goto restart; 688 } 689 cache_inval(ncp, flags); 690 cache_put(ncp); /* also releases reference */ 691 ncp = next; 692 if (ncp && ncp->nc_vp != vp) { 693 printf("Warning: cache_inval_vp: race-B detected on " 694 "%s\n", ncp->nc_name); 695 cache_drop(ncp); 696 goto restart; 697 } 698 } 699 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 700 } 701 702 /* 703 * The source ncp has been renamed to the target ncp. Both fncp and tncp 704 * must be locked. Both will be set to unresolved, any children of tncp 705 * will be disconnected (the prior contents of the target is assumed to be 706 * destroyed by the rename operation, e.g. renaming over an empty directory), 707 * and all children of fncp will be moved to tncp. 708 * 709 * XXX the disconnection could pose a problem, check code paths to make 710 * sure any code that blocks can handle the parent being changed out from 711 * under it. Maybe we should lock the children (watch out for deadlocks) ? 712 * 713 * After we return the caller has the option of calling cache_setvp() if 714 * the vnode of the new target ncp is known. 715 * 716 * Any process CD'd into any of the children will no longer be able to ".." 717 * back out. An rm -rf can cause this situation to occur. 718 */ 719 void 720 cache_rename(struct namecache *fncp, struct namecache *tncp) 721 { 722 struct namecache *scan; 723 int didwarn = 0; 724 725 cache_setunresolved(fncp); 726 cache_setunresolved(tncp); 727 while (cache_inval(tncp, CINV_CHILDREN) != 0) { 728 if (didwarn++ % 10 == 0) { 729 printf("Warning: cache_rename: race during " 730 "rename %s->%s\n", 731 fncp->nc_name, tncp->nc_name); 732 } 733 tsleep(tncp, 0, "mvrace", hz / 10); 734 cache_setunresolved(tncp); 735 } 736 while ((scan = TAILQ_FIRST(&fncp->nc_list)) != NULL) { 737 cache_hold(scan); 738 cache_unlink_parent(scan); 739 cache_link_parent(scan, tncp); 740 if (scan->nc_flag & NCF_HASHED) 741 cache_rehash(scan); 742 cache_drop(scan); 743 } 744 } 745 746 /* 747 * vget the vnode associated with the namecache entry. Resolve the namecache 748 * entry if necessary and deal with namecache/vp races. The passed ncp must 749 * be referenced and may be locked. The ncp's ref/locking state is not 750 * effected by this call. 751 * 752 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked 753 * (depending on the passed lk_type) will be returned in *vpp with an error 754 * of 0, or NULL will be returned in *vpp with a non-0 error code. The 755 * most typical error is ENOENT, meaning that the ncp represents a negative 756 * cache hit and there is no vnode to retrieve, but other errors can occur 757 * too. 758 * 759 * The main race we have to deal with are namecache zaps. The ncp itself 760 * will not disappear since it is referenced, and it turns out that the 761 * validity of the vp pointer can be checked simply by rechecking the 762 * contents of ncp->nc_vp. 763 */ 764 int 765 cache_vget(struct namecache *ncp, struct ucred *cred, 766 int lk_type, struct vnode **vpp) 767 { 768 struct vnode *vp; 769 int error; 770 771 again: 772 vp = NULL; 773 if (ncp->nc_flag & NCF_UNRESOLVED) { 774 cache_lock(ncp); 775 error = cache_resolve(ncp, cred); 776 cache_unlock(ncp); 777 } else { 778 error = 0; 779 } 780 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 781 error = vget(vp, lk_type, curthread); 782 if (error) { 783 if (vp != ncp->nc_vp) /* handle cache_zap race */ 784 goto again; 785 vp = NULL; 786 } else if (vp != ncp->nc_vp) { /* handle cache_zap race */ 787 vput(vp); 788 goto again; 789 } 790 } 791 if (error == 0 && vp == NULL) 792 error = ENOENT; 793 *vpp = vp; 794 return(error); 795 } 796 797 int 798 cache_vref(struct namecache *ncp, struct ucred *cred, struct vnode **vpp) 799 { 800 struct vnode *vp; 801 int error; 802 803 again: 804 vp = NULL; 805 if (ncp->nc_flag & NCF_UNRESOLVED) { 806 cache_lock(ncp); 807 error = cache_resolve(ncp, cred); 808 cache_unlock(ncp); 809 } else { 810 error = 0; 811 } 812 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 813 vref(vp); 814 if (vp != ncp->nc_vp) { /* handle cache_zap race */ 815 vrele(vp); 816 goto again; 817 } 818 } 819 if (error == 0 && vp == NULL) 820 error = ENOENT; 821 *vpp = vp; 822 return(error); 823 } 824 825 /* 826 * Convert a directory vnode to a namecache record without any other 827 * knowledge of the topology. This ONLY works with directory vnodes and 828 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the 829 * returned ncp (if not NULL) will be held and unlocked. 830 * 831 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned. 832 * If 'makeit' is 1 we attempt to track-down and create the namecache topology 833 * for dvp. This will fail only if the directory has been deleted out from 834 * under the caller. 835 * 836 * Callers must always check for a NULL return no matter the value of 'makeit'. 837 */ 838 839 static int cache_inefficient_scan(struct namecache *ncp, struct ucred *cred, 840 struct vnode *dvp); 841 842 struct namecache * 843 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit) 844 { 845 struct namecache *ncp; 846 struct vnode *pvp; 847 int error; 848 849 /* 850 * Temporary debugging code to force the directory scanning code 851 * to be exercised. 852 */ 853 ncp = NULL; 854 if (ncvp_debug >= 3 && makeit && TAILQ_FIRST(&dvp->v_namecache)) { 855 ncp = TAILQ_FIRST(&dvp->v_namecache); 856 printf("cache_fromdvp: forcing %s\n", ncp->nc_name); 857 goto force; 858 } 859 860 /* 861 * Loop until resolution, inside code will break out on error. 862 */ 863 while ((ncp = TAILQ_FIRST(&dvp->v_namecache)) == NULL && makeit) { 864 force: 865 /* 866 * If dvp is the root of its filesystem it should already 867 * have a namecache pointer associated with it as a side 868 * effect of the mount, but it may have been disassociated. 869 */ 870 if (dvp->v_flag & VROOT) { 871 ncp = cache_get(dvp->v_mount->mnt_ncp); 872 error = cache_resolve_mp(ncp); 873 cache_put(ncp); 874 if (ncvp_debug) { 875 printf("cache_fromdvp: resolve root of mount %p error %d", 876 dvp->v_mount, error); 877 } 878 if (error) { 879 if (ncvp_debug) 880 printf(" failed\n"); 881 ncp = NULL; 882 break; 883 } 884 if (ncvp_debug) 885 printf(" succeeded\n"); 886 continue; 887 } 888 889 /* 890 * Get the parent directory and resolve its ncp. 891 */ 892 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred); 893 if (error) { 894 printf("lookupdotdot failed %d %p\n", error, pvp); 895 break; 896 } 897 VOP_UNLOCK(pvp, 0, curthread); 898 899 /* 900 * XXX this recursion could run the kernel out of stack, 901 * change to a less efficient algorithm if we get too deep 902 * (use 'makeit' for a depth counter?) 903 */ 904 ncp = cache_fromdvp(pvp, cred, makeit); 905 vrele(pvp); 906 if (ncp == NULL) 907 break; 908 909 /* 910 * Do an inefficient scan of pvp (embodied by ncp) to look 911 * for dvp. This will create a namecache record for dvp on 912 * success. We loop up to recheck on success. 913 * 914 * ncp and dvp are both held but not locked. 915 */ 916 error = cache_inefficient_scan(ncp, cred, dvp); 917 cache_drop(ncp); 918 if (error) { 919 printf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n", 920 pvp, ncp->nc_name, dvp); 921 ncp = NULL; 922 break; 923 } 924 if (ncvp_debug) { 925 printf("cache_fromdvp: scan %p (%s) succeeded\n", 926 pvp, ncp->nc_name); 927 } 928 } 929 if (ncp) 930 cache_hold(ncp); 931 return (ncp); 932 } 933 934 /* 935 * Do an inefficient scan of the directory represented by ncp looking for 936 * the directory vnode dvp. ncp must be held but not locked on entry and 937 * will be held on return. dvp must be refd but not locked on entry and 938 * will remain refd on return. 939 * 940 * Why do this at all? Well, due to its stateless nature the NFS server 941 * converts file handles directly to vnodes without necessarily going through 942 * the namecache ops that would otherwise create the namecache topology 943 * leading to the vnode. We could either (1) Change the namecache algorithms 944 * to allow disconnect namecache records that are re-merged opportunistically, 945 * or (2) Make the NFS server backtrack and scan to recover a connected 946 * namecache topology in order to then be able to issue new API lookups. 947 * 948 * It turns out that (1) is a huge mess. It takes a nice clean set of 949 * namecache algorithms and introduces a lot of complication in every subsystem 950 * that calls into the namecache to deal with the re-merge case, especially 951 * since we are using the namecache to placehold negative lookups and the 952 * vnode might not be immediately assigned. (2) is certainly far less 953 * efficient then (1), but since we are only talking about directories here 954 * (which are likely to remain cached), the case does not actually run all 955 * that often and has the supreme advantage of not polluting the namecache 956 * algorithms. 957 */ 958 static int 959 cache_inefficient_scan(struct namecache *ncp, struct ucred *cred, 960 struct vnode *dvp) 961 { 962 struct nlcomponent nlc; 963 struct namecache *rncp; 964 struct dirent *den; 965 struct vnode *pvp; 966 struct vattr vat; 967 struct iovec iov; 968 struct uio uio; 969 u_long *cookies; 970 off_t baseoff; 971 int ncookies; 972 int blksize; 973 int eofflag; 974 char *rbuf; 975 int error; 976 int xoff; 977 int i; 978 979 vat.va_blocksize = 0; 980 if ((error = VOP_GETATTR(dvp, &vat, curthread)) != 0) 981 return (error); 982 if ((error = cache_vget(ncp, cred, LK_SHARED, &pvp)) != 0) 983 return (error); 984 if (ncvp_debug) 985 printf("inefficient_scan: directory iosize %ld vattr fileid = %ld\n", vat.va_blocksize, (long)vat.va_fileid); 986 if ((blksize = vat.va_blocksize) == 0) 987 blksize = DEV_BSIZE; 988 rbuf = malloc(blksize, M_TEMP, M_WAITOK); 989 rncp = NULL; 990 991 eofflag = 0; 992 uio.uio_offset = 0; 993 cookies = NULL; 994 again: 995 baseoff = uio.uio_offset; 996 iov.iov_base = rbuf; 997 iov.iov_len = blksize; 998 uio.uio_iov = &iov; 999 uio.uio_iovcnt = 1; 1000 uio.uio_resid = blksize; 1001 uio.uio_segflg = UIO_SYSSPACE; 1002 uio.uio_rw = UIO_READ; 1003 uio.uio_td = curthread; 1004 1005 if (cookies) { 1006 free(cookies, M_TEMP); 1007 cookies = NULL; 1008 } 1009 if (ncvp_debug >= 2) 1010 printf("cache_inefficient_scan: readdir @ %08x\n", (int)baseoff); 1011 error = VOP_READDIR(pvp, &uio, cred, &eofflag, &ncookies, &cookies); 1012 if (error == 0 && cookies == NULL) 1013 error = EPERM; 1014 if (error == 0) { 1015 for (i = 0; i < ncookies; ++i) { 1016 xoff = (int)(cookies[i] - (u_long)baseoff); 1017 /* 1018 * UFS plays a little trick to skip the first entry 1019 * in a directory ("."), by assigning the cookie to 1020 * dpoff + dp->d_reclen in the loop. This causes 1021 * the last cookie to be assigned to the data-end of 1022 * the directory. XXX 1023 */ 1024 if (xoff == blksize) 1025 break; 1026 KKASSERT(xoff >= 0 && xoff <= blksize); 1027 den = (struct dirent *)(rbuf + xoff); 1028 if (ncvp_debug >= 2) 1029 printf("cache_inefficient_scan: %*.*s\n", 1030 den->d_namlen, den->d_namlen, den->d_name); 1031 if (den->d_type != DT_WHT && 1032 den->d_fileno == vat.va_fileid) { 1033 if (ncvp_debug) 1034 printf("cache_inefficient_scan: MATCHED inode %ld path %s/%*.*s\n", vat.va_fileid, ncp->nc_name, den->d_namlen, den->d_namlen, den->d_name); 1035 nlc.nlc_nameptr = den->d_name; 1036 nlc.nlc_namelen = den->d_namlen; 1037 VOP_UNLOCK(pvp, 0, curthread); 1038 rncp = cache_nlookup(ncp, &nlc); 1039 KKASSERT(rncp != NULL); 1040 break; 1041 } 1042 } 1043 if (rncp == NULL && eofflag == 0 && uio.uio_resid != blksize) 1044 goto again; 1045 } 1046 if (cookies) { 1047 free(cookies, M_TEMP); 1048 cookies = NULL; 1049 } 1050 if (rncp) { 1051 vrele(pvp); 1052 if (rncp->nc_flag & NCF_UNRESOLVED) { 1053 cache_setvp(rncp, dvp); 1054 if (ncvp_debug >= 2) { 1055 printf("cache_inefficient_scan: setvp %s/%s = %p\n", 1056 ncp->nc_name, rncp->nc_name, dvp); 1057 } 1058 } else { 1059 if (ncvp_debug >= 2) { 1060 printf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n", 1061 ncp->nc_name, rncp->nc_name, dvp, 1062 rncp->nc_vp); 1063 } 1064 } 1065 if (rncp->nc_vp == NULL) 1066 error = rncp->nc_error; 1067 cache_put(rncp); 1068 } else { 1069 printf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n", 1070 dvp, ncp->nc_name); 1071 vput(pvp); 1072 error = ENOENT; 1073 } 1074 free(rbuf, M_TEMP); 1075 return (error); 1076 } 1077 1078 /* 1079 * Zap a namecache entry. The ncp is unconditionally set to an unresolved 1080 * state, which disassociates it from its vnode or ncneglist. 1081 * 1082 * Then, if there are no additional references to the ncp and no children, 1083 * the ncp is removed from the topology and destroyed. This function will 1084 * also run through the nc_parent chain and destroy parent ncps if possible. 1085 * As a side benefit, it turns out the only conditions that allow running 1086 * up the chain are also the conditions to ensure no deadlock will occur. 1087 * 1088 * References and/or children may exist if the ncp is in the middle of the 1089 * topology, preventing the ncp from being destroyed. 1090 * 1091 * This function must be called with the ncp held and locked and will unlock 1092 * and drop it during zapping. 1093 */ 1094 static void 1095 cache_zap(struct namecache *ncp) 1096 { 1097 struct namecache *par; 1098 1099 /* 1100 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED. 1101 */ 1102 cache_setunresolved(ncp); 1103 1104 /* 1105 * Try to scrap the entry and possibly tail-recurse on its parent. 1106 * We only scrap unref'd (other then our ref) unresolved entries, 1107 * we do not scrap 'live' entries. 1108 */ 1109 while (ncp->nc_flag & NCF_UNRESOLVED) { 1110 /* 1111 * Someone other then us has a ref, stop. 1112 */ 1113 if (ncp->nc_refs > 1) 1114 goto done; 1115 1116 /* 1117 * We have children, stop. 1118 */ 1119 if (!TAILQ_EMPTY(&ncp->nc_list)) 1120 goto done; 1121 1122 /* 1123 * Remove ncp from the topology: hash table and parent linkage. 1124 */ 1125 if (ncp->nc_flag & NCF_HASHED) { 1126 ncp->nc_flag &= ~NCF_HASHED; 1127 LIST_REMOVE(ncp, nc_hash); 1128 } 1129 if ((par = ncp->nc_parent) != NULL) { 1130 par = cache_hold(par); 1131 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 1132 ncp->nc_parent = NULL; 1133 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 1134 vdrop(par->nc_vp); 1135 } 1136 1137 /* 1138 * ncp should not have picked up any refs. Physically 1139 * destroy the ncp. 1140 */ 1141 KKASSERT(ncp->nc_refs == 1); 1142 --numunres; 1143 /* cache_unlock(ncp) not required */ 1144 ncp->nc_refs = -1; /* safety */ 1145 if (ncp->nc_name) 1146 free(ncp->nc_name, M_VFSCACHE); 1147 free(ncp, M_VFSCACHE); 1148 1149 /* 1150 * Loop on the parent (it may be NULL). Only bother looping 1151 * if the parent has a single ref (ours), which also means 1152 * we can lock it trivially. 1153 */ 1154 ncp = par; 1155 if (ncp == NULL) 1156 return; 1157 if (ncp->nc_refs != 1) { 1158 cache_drop(ncp); 1159 return; 1160 } 1161 KKASSERT(par->nc_exlocks == 0); 1162 cache_lock(ncp); 1163 } 1164 done: 1165 cache_unlock(ncp); 1166 --ncp->nc_refs; 1167 } 1168 1169 static enum { CHI_LOW, CHI_HIGH } cache_hysteresis_state = CHI_LOW; 1170 1171 static __inline 1172 void 1173 cache_hysteresis(void) 1174 { 1175 /* 1176 * Don't cache too many negative hits. We use hysteresis to reduce 1177 * the impact on the critical path. 1178 */ 1179 switch(cache_hysteresis_state) { 1180 case CHI_LOW: 1181 if (numneg > MINNEG && numneg * ncnegfactor > numcache) { 1182 cache_cleanneg(10); 1183 cache_hysteresis_state = CHI_HIGH; 1184 } 1185 break; 1186 case CHI_HIGH: 1187 if (numneg > MINNEG * 9 / 10 && 1188 numneg * ncnegfactor * 9 / 10 > numcache 1189 ) { 1190 cache_cleanneg(10); 1191 } else { 1192 cache_hysteresis_state = CHI_LOW; 1193 } 1194 break; 1195 } 1196 } 1197 1198 /* 1199 * NEW NAMECACHE LOOKUP API 1200 * 1201 * Lookup an entry in the cache. A locked, referenced, non-NULL 1202 * entry is *always* returned, even if the supplied component is illegal. 1203 * The resulting namecache entry should be returned to the system with 1204 * cache_put() or cache_unlock() + cache_drop(). 1205 * 1206 * namecache locks are recursive but care must be taken to avoid lock order 1207 * reversals. 1208 * 1209 * Nobody else will be able to manipulate the associated namespace (e.g. 1210 * create, delete, rename, rename-target) until the caller unlocks the 1211 * entry. 1212 * 1213 * The returned entry will be in one of three states: positive hit (non-null 1214 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set). 1215 * Unresolved entries must be resolved through the filesystem to associate the 1216 * vnode and/or determine whether a positive or negative hit has occured. 1217 * 1218 * It is not necessary to lock a directory in order to lock namespace under 1219 * that directory. In fact, it is explicitly not allowed to do that. A 1220 * directory is typically only locked when being created, renamed, or 1221 * destroyed. 1222 * 1223 * The directory (par) may be unresolved, in which case any returned child 1224 * will likely also be marked unresolved. Likely but not guarenteed. Since 1225 * the filesystem lookup requires a resolved directory vnode the caller is 1226 * responsible for resolving the namecache chain top-down. This API 1227 * specifically allows whole chains to be created in an unresolved state. 1228 */ 1229 struct namecache * 1230 cache_nlookup(struct namecache *par, struct nlcomponent *nlc) 1231 { 1232 struct namecache *ncp; 1233 struct namecache *new_ncp; 1234 struct nchashhead *nchpp; 1235 u_int32_t hash; 1236 globaldata_t gd; 1237 1238 numcalls++; 1239 gd = mycpu; 1240 1241 /* 1242 * Try to locate an existing entry 1243 */ 1244 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 1245 hash = fnv_32_buf(&par, sizeof(par), hash); 1246 new_ncp = NULL; 1247 restart: 1248 LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 1249 numchecks++; 1250 1251 /* 1252 * Zap entries that have timed out. 1253 */ 1254 if (ncp->nc_timeout && 1255 (int)(ncp->nc_timeout - ticks) < 0 && 1256 (ncp->nc_flag & NCF_UNRESOLVED) == 0 && 1257 ncp->nc_exlocks == 0 1258 ) { 1259 cache_zap(cache_get(ncp)); 1260 goto restart; 1261 } 1262 1263 /* 1264 * Break out if we find a matching entry. Note that 1265 * UNRESOLVED entries may match, but DESTROYED entries 1266 * do not. 1267 */ 1268 if (ncp->nc_parent == par && 1269 ncp->nc_nlen == nlc->nlc_namelen && 1270 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 1271 (ncp->nc_flag & NCF_DESTROYED) == 0 1272 ) { 1273 if (cache_get_nonblock(ncp) == 0) { 1274 if (new_ncp) 1275 cache_free(new_ncp); 1276 goto found; 1277 } 1278 cache_get(ncp); 1279 cache_put(ncp); 1280 goto restart; 1281 } 1282 } 1283 1284 /* 1285 * We failed to locate an entry, create a new entry and add it to 1286 * the cache. We have to relookup after possibly blocking in 1287 * malloc. 1288 */ 1289 if (new_ncp == NULL) { 1290 new_ncp = cache_alloc(nlc->nlc_namelen); 1291 goto restart; 1292 } 1293 1294 ncp = new_ncp; 1295 1296 /* 1297 * Initialize as a new UNRESOLVED entry, lock (non-blocking), 1298 * and link to the parent. The mount point is usually inherited 1299 * from the parent unless this is a special case such as a mount 1300 * point where nlc_namelen is 0. The caller is responsible for 1301 * setting nc_mount in that case. If nlc_namelen is 0 nc_name will 1302 * be NULL. 1303 */ 1304 if (nlc->nlc_namelen) { 1305 bcopy(nlc->nlc_nameptr, ncp->nc_name, nlc->nlc_namelen); 1306 ncp->nc_name[nlc->nlc_namelen] = 0; 1307 ncp->nc_mount = par->nc_mount; 1308 } 1309 nchpp = NCHHASH(hash); 1310 LIST_INSERT_HEAD(nchpp, ncp, nc_hash); 1311 ncp->nc_flag |= NCF_HASHED; 1312 cache_link_parent(ncp, par); 1313 found: 1314 /* 1315 * stats and namecache size management 1316 */ 1317 if (ncp->nc_flag & NCF_UNRESOLVED) 1318 ++gd->gd_nchstats->ncs_miss; 1319 else if (ncp->nc_vp) 1320 ++gd->gd_nchstats->ncs_goodhits; 1321 else 1322 ++gd->gd_nchstats->ncs_neghits; 1323 cache_hysteresis(); 1324 return(ncp); 1325 } 1326 1327 /* 1328 * Resolve an unresolved namecache entry, generally by looking it up. 1329 * The passed ncp must be locked and refd. 1330 * 1331 * Theoretically since a vnode cannot be recycled while held, and since 1332 * the nc_parent chain holds its vnode as long as children exist, the 1333 * direct parent of the cache entry we are trying to resolve should 1334 * have a valid vnode. If not then generate an error that we can 1335 * determine is related to a resolver bug. 1336 * 1337 * Note that successful resolution does not necessarily return an error 1338 * code of 0. If the ncp resolves to a negative cache hit then ENOENT 1339 * will be returned. 1340 */ 1341 int 1342 cache_resolve(struct namecache *ncp, struct ucred *cred) 1343 { 1344 struct namecache *par; 1345 int error; 1346 1347 restart: 1348 /* 1349 * If the ncp is already resolved we have nothing to do. 1350 */ 1351 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 1352 return (ncp->nc_error); 1353 1354 /* 1355 * Mount points need special handling because the parent does not 1356 * belong to the same filesystem as the ncp. 1357 */ 1358 if (ncp->nc_flag & NCF_MOUNTPT) 1359 return (cache_resolve_mp(ncp)); 1360 1361 /* 1362 * We expect an unbroken chain of ncps to at least the mount point, 1363 * and even all the way to root (but this code doesn't have to go 1364 * past the mount point). 1365 */ 1366 if (ncp->nc_parent == NULL) { 1367 printf("EXDEV case 1 %p %*.*s\n", ncp, 1368 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 1369 ncp->nc_error = EXDEV; 1370 return(ncp->nc_error); 1371 } 1372 1373 /* 1374 * The vp's of the parent directories in the chain are held via vhold() 1375 * due to the existance of the child, and should not disappear. 1376 * However, there are cases where they can disappear: 1377 * 1378 * - due to filesystem I/O errors. 1379 * - due to NFS being stupid about tracking the namespace and 1380 * destroys the namespace for entire directories quite often. 1381 * - due to forced unmounts. 1382 * - due to an rmdir (parent will be marked DESTROYED) 1383 * 1384 * When this occurs we have to track the chain backwards and resolve 1385 * it, looping until the resolver catches up to the current node. We 1386 * could recurse here but we might run ourselves out of kernel stack 1387 * so we do it in a more painful manner. This situation really should 1388 * not occur all that often, or if it does not have to go back too 1389 * many nodes to resolve the ncp. 1390 */ 1391 while (ncp->nc_parent->nc_vp == NULL) { 1392 /* 1393 * This case can occur if a process is CD'd into a 1394 * directory which is then rmdir'd. If the parent is marked 1395 * destroyed there is no point trying to resolve it. 1396 */ 1397 if (ncp->nc_parent->nc_flag & NCF_DESTROYED) 1398 return(ENOENT); 1399 1400 par = ncp->nc_parent; 1401 while (par->nc_parent && par->nc_parent->nc_vp == NULL) 1402 par = par->nc_parent; 1403 if (par->nc_parent == NULL) { 1404 printf("EXDEV case 2 %*.*s\n", 1405 par->nc_nlen, par->nc_nlen, par->nc_name); 1406 return (EXDEV); 1407 } 1408 printf("[diagnostic] cache_resolve: had to recurse on %*.*s\n", 1409 par->nc_nlen, par->nc_nlen, par->nc_name); 1410 /* 1411 * The parent is not set in stone, ref and lock it to prevent 1412 * it from disappearing. Also note that due to renames it 1413 * is possible for our ncp to move and for par to no longer 1414 * be one of its parents. We resolve it anyway, the loop 1415 * will handle any moves. 1416 */ 1417 cache_get(par); 1418 if (par->nc_flag & NCF_MOUNTPT) { 1419 cache_resolve_mp(par); 1420 } else if (par->nc_parent->nc_vp == NULL) { 1421 printf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name); 1422 cache_put(par); 1423 continue; 1424 } else if (par->nc_flag & NCF_UNRESOLVED) { 1425 par->nc_error = VOP_NRESOLVE(par, cred); 1426 } 1427 if ((error = par->nc_error) != 0) { 1428 if (par->nc_error != EAGAIN) { 1429 printf("EXDEV case 3 %*.*s error %d\n", 1430 par->nc_nlen, par->nc_nlen, par->nc_name, 1431 par->nc_error); 1432 cache_put(par); 1433 return(error); 1434 } 1435 printf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n", 1436 par, par->nc_nlen, par->nc_nlen, par->nc_name); 1437 } 1438 cache_put(par); 1439 /* loop */ 1440 } 1441 1442 /* 1443 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected 1444 * ncp's and reattach them. If this occurs the original ncp is marked 1445 * EAGAIN to force a relookup. 1446 * 1447 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed 1448 * ncp must already be resolved. 1449 */ 1450 KKASSERT((ncp->nc_flag & NCF_MOUNTPT) == 0); 1451 ncp->nc_error = VOP_NRESOLVE(ncp, cred); 1452 /*vop_nresolve(*ncp->nc_parent->nc_vp->v_ops, ncp, cred);*/ 1453 if (ncp->nc_error == EAGAIN) { 1454 printf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n", 1455 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 1456 goto restart; 1457 } 1458 return(ncp->nc_error); 1459 } 1460 1461 /* 1462 * Resolve the ncp associated with a mount point. Such ncp's almost always 1463 * remain resolved and this routine is rarely called. NFS MPs tends to force 1464 * re-resolution more often due to its mac-truck-smash-the-namecache 1465 * method of tracking namespace changes. 1466 * 1467 * The semantics for this call is that the passed ncp must be locked on 1468 * entry and will be locked on return. However, if we actually have to 1469 * resolve the mount point we temporarily unlock the entry in order to 1470 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of 1471 * the unlock we have to recheck the flags after we relock. 1472 */ 1473 static int 1474 cache_resolve_mp(struct namecache *ncp) 1475 { 1476 struct vnode *vp; 1477 struct mount *mp = ncp->nc_mount; 1478 int error; 1479 1480 KKASSERT(mp != NULL); 1481 if (ncp->nc_flag & NCF_UNRESOLVED) { 1482 cache_unlock(ncp); 1483 while (vfs_busy(mp, 0, curthread)) 1484 ; 1485 error = VFS_ROOT(mp, &vp); 1486 cache_lock(ncp); 1487 1488 /* 1489 * recheck the ncp state after relocking. 1490 */ 1491 if (ncp->nc_flag & NCF_UNRESOLVED) { 1492 ncp->nc_error = error; 1493 if (error == 0) { 1494 cache_setvp(ncp, vp); 1495 vput(vp); 1496 } else { 1497 printf("[diagnostic] cache_resolve_mp: failed to resolve mount %p\n", mp); 1498 cache_setvp(ncp, NULL); 1499 } 1500 } else if (error == 0) { 1501 vput(vp); 1502 } 1503 vfs_unbusy(mp, curthread); 1504 } 1505 return(ncp->nc_error); 1506 } 1507 1508 void 1509 cache_cleanneg(int count) 1510 { 1511 struct namecache *ncp; 1512 1513 /* 1514 * Automode from the vnlru proc - clean out 10% of the negative cache 1515 * entries. 1516 */ 1517 if (count == 0) 1518 count = numneg / 10 + 1; 1519 1520 /* 1521 * Attempt to clean out the specified number of negative cache 1522 * entries. 1523 */ 1524 while (count) { 1525 ncp = TAILQ_FIRST(&ncneglist); 1526 if (ncp == NULL) { 1527 KKASSERT(numneg == 0); 1528 break; 1529 } 1530 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 1531 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 1532 if (cache_get_nonblock(ncp) == 0) 1533 cache_zap(ncp); 1534 --count; 1535 } 1536 } 1537 1538 /* 1539 * Rehash a ncp. Rehashing is typically required if the name changes (should 1540 * not generally occur) or the parent link changes. This function will 1541 * unhash the ncp if the ncp is no longer hashable. 1542 */ 1543 static void 1544 cache_rehash(struct namecache *ncp) 1545 { 1546 struct nchashhead *nchpp; 1547 u_int32_t hash; 1548 1549 if (ncp->nc_flag & NCF_HASHED) { 1550 ncp->nc_flag &= ~NCF_HASHED; 1551 LIST_REMOVE(ncp, nc_hash); 1552 } 1553 if (ncp->nc_nlen && ncp->nc_parent) { 1554 hash = fnv_32_buf(ncp->nc_name, ncp->nc_nlen, FNV1_32_INIT); 1555 hash = fnv_32_buf(&ncp->nc_parent, 1556 sizeof(ncp->nc_parent), hash); 1557 nchpp = NCHHASH(hash); 1558 LIST_INSERT_HEAD(nchpp, ncp, nc_hash); 1559 ncp->nc_flag |= NCF_HASHED; 1560 } 1561 } 1562 1563 /* 1564 * Name cache initialization, from vfsinit() when we are booting 1565 */ 1566 void 1567 nchinit(void) 1568 { 1569 int i; 1570 globaldata_t gd; 1571 1572 /* initialise per-cpu namecache effectiveness statistics. */ 1573 for (i = 0; i < ncpus; ++i) { 1574 gd = globaldata_find(i); 1575 gd->gd_nchstats = &nchstats[i]; 1576 } 1577 TAILQ_INIT(&ncneglist); 1578 nchashtbl = hashinit(desiredvnodes*2, M_VFSCACHE, &nchash); 1579 nclockwarn = 1 * hz; 1580 } 1581 1582 /* 1583 * Called from start_init() to bootstrap the root filesystem. Returns 1584 * a referenced, unlocked namecache record. 1585 */ 1586 struct namecache * 1587 cache_allocroot(struct mount *mp, struct vnode *vp) 1588 { 1589 struct namecache *ncp = cache_alloc(0); 1590 1591 ncp->nc_flag |= NCF_MOUNTPT | NCF_ROOT; 1592 ncp->nc_mount = mp; 1593 cache_setvp(ncp, vp); 1594 return(ncp); 1595 } 1596 1597 /* 1598 * vfs_cache_setroot() 1599 * 1600 * Create an association between the root of our namecache and 1601 * the root vnode. This routine may be called several times during 1602 * booting. 1603 * 1604 * If the caller intends to save the returned namecache pointer somewhere 1605 * it must cache_hold() it. 1606 */ 1607 void 1608 vfs_cache_setroot(struct vnode *nvp, struct namecache *ncp) 1609 { 1610 struct vnode *ovp; 1611 struct namecache *oncp; 1612 1613 ovp = rootvnode; 1614 oncp = rootncp; 1615 rootvnode = nvp; 1616 rootncp = ncp; 1617 1618 if (ovp) 1619 vrele(ovp); 1620 if (oncp) 1621 cache_drop(oncp); 1622 } 1623 1624 /* 1625 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache 1626 * topology and is being removed as quickly as possible. The new VOP_N*() 1627 * API calls are required to make specific adjustments using the supplied 1628 * ncp pointers rather then just bogusly purging random vnodes. 1629 * 1630 * Invalidate all namecache entries to a particular vnode as well as 1631 * any direct children of that vnode in the namecache. This is a 1632 * 'catch all' purge used by filesystems that do not know any better. 1633 * 1634 * A new vnode v_id is generated. Note that no vnode will ever have a 1635 * v_id of 0. 1636 * 1637 * Note that the linkage between the vnode and its namecache entries will 1638 * be removed, but the namecache entries themselves might stay put due to 1639 * active references from elsewhere in the system or due to the existance of 1640 * the children. The namecache topology is left intact even if we do not 1641 * know what the vnode association is. Such entries will be marked 1642 * NCF_UNRESOLVED. 1643 * 1644 * XXX: Only time and the size of v_id prevents this from failing: 1645 * XXX: In theory we should hunt down all (struct vnode*, v_id) 1646 * XXX: soft references and nuke them, at least on the global 1647 * XXX: v_id wraparound. The period of resistance can be extended 1648 * XXX: by incrementing each vnodes v_id individually instead of 1649 * XXX: using the global v_id. 1650 */ 1651 void 1652 cache_purge(struct vnode *vp) 1653 { 1654 static u_long nextid; 1655 1656 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN); 1657 1658 /* 1659 * Calculate a new unique id for ".." handling 1660 */ 1661 do { 1662 nextid++; 1663 } while (nextid == vp->v_id || nextid == 0); 1664 vp->v_id = nextid; 1665 } 1666 1667 /* 1668 * Flush all entries referencing a particular filesystem. 1669 * 1670 * Since we need to check it anyway, we will flush all the invalid 1671 * entries at the same time. 1672 */ 1673 void 1674 cache_purgevfs(struct mount *mp) 1675 { 1676 struct nchashhead *nchpp; 1677 struct namecache *ncp, *nnp; 1678 1679 /* 1680 * Scan hash tables for applicable entries. 1681 */ 1682 for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) { 1683 ncp = LIST_FIRST(nchpp); 1684 if (ncp) 1685 cache_hold(ncp); 1686 while (ncp) { 1687 nnp = LIST_NEXT(ncp, nc_hash); 1688 if (nnp) 1689 cache_hold(nnp); 1690 if (ncp->nc_mount == mp) { 1691 cache_lock(ncp); 1692 cache_zap(ncp); 1693 } else { 1694 cache_drop(ncp); 1695 } 1696 ncp = nnp; 1697 } 1698 } 1699 } 1700 1701 static int disablecwd; 1702 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, ""); 1703 1704 static u_long numcwdcalls; STATNODE(CTLFLAG_RD, numcwdcalls, &numcwdcalls); 1705 static u_long numcwdfail1; STATNODE(CTLFLAG_RD, numcwdfail1, &numcwdfail1); 1706 static u_long numcwdfail2; STATNODE(CTLFLAG_RD, numcwdfail2, &numcwdfail2); 1707 static u_long numcwdfail3; STATNODE(CTLFLAG_RD, numcwdfail3, &numcwdfail3); 1708 static u_long numcwdfail4; STATNODE(CTLFLAG_RD, numcwdfail4, &numcwdfail4); 1709 static u_long numcwdfound; STATNODE(CTLFLAG_RD, numcwdfound, &numcwdfound); 1710 1711 int 1712 __getcwd(struct __getcwd_args *uap) 1713 { 1714 int buflen; 1715 int error; 1716 char *buf; 1717 char *bp; 1718 1719 if (disablecwd) 1720 return (ENODEV); 1721 1722 buflen = uap->buflen; 1723 if (buflen < 2) 1724 return (EINVAL); 1725 if (buflen > MAXPATHLEN) 1726 buflen = MAXPATHLEN; 1727 1728 buf = malloc(buflen, M_TEMP, M_WAITOK); 1729 bp = kern_getcwd(buf, buflen, &error); 1730 if (error == 0) 1731 error = copyout(bp, uap->buf, strlen(bp) + 1); 1732 free(buf, M_TEMP); 1733 return (error); 1734 } 1735 1736 char * 1737 kern_getcwd(char *buf, size_t buflen, int *error) 1738 { 1739 struct proc *p = curproc; 1740 char *bp; 1741 int i, slash_prefixed; 1742 struct filedesc *fdp; 1743 struct namecache *ncp; 1744 1745 numcwdcalls++; 1746 bp = buf; 1747 bp += buflen - 1; 1748 *bp = '\0'; 1749 fdp = p->p_fd; 1750 slash_prefixed = 0; 1751 1752 ncp = fdp->fd_ncdir; 1753 while (ncp && ncp != fdp->fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) { 1754 if (ncp->nc_flag & NCF_MOUNTPT) { 1755 if (ncp->nc_mount == NULL) { 1756 *error = EBADF; /* forced unmount? */ 1757 return(NULL); 1758 } 1759 ncp = ncp->nc_parent; 1760 continue; 1761 } 1762 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 1763 if (bp == buf) { 1764 numcwdfail4++; 1765 *error = ENOMEM; 1766 return(NULL); 1767 } 1768 *--bp = ncp->nc_name[i]; 1769 } 1770 if (bp == buf) { 1771 numcwdfail4++; 1772 *error = ENOMEM; 1773 return(NULL); 1774 } 1775 *--bp = '/'; 1776 slash_prefixed = 1; 1777 ncp = ncp->nc_parent; 1778 } 1779 if (ncp == NULL) { 1780 numcwdfail2++; 1781 *error = ENOENT; 1782 return(NULL); 1783 } 1784 if (!slash_prefixed) { 1785 if (bp == buf) { 1786 numcwdfail4++; 1787 *error = ENOMEM; 1788 return(NULL); 1789 } 1790 *--bp = '/'; 1791 } 1792 numcwdfound++; 1793 *error = 0; 1794 return (bp); 1795 } 1796 1797 /* 1798 * Thus begins the fullpath magic. 1799 */ 1800 1801 #undef STATNODE 1802 #define STATNODE(name) \ 1803 static u_int name; \ 1804 SYSCTL_UINT(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, "") 1805 1806 static int disablefullpath; 1807 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, 1808 &disablefullpath, 0, ""); 1809 1810 STATNODE(numfullpathcalls); 1811 STATNODE(numfullpathfail1); 1812 STATNODE(numfullpathfail2); 1813 STATNODE(numfullpathfail3); 1814 STATNODE(numfullpathfail4); 1815 STATNODE(numfullpathfound); 1816 1817 int 1818 cache_fullpath(struct proc *p, struct namecache *ncp, char **retbuf, char **freebuf) 1819 { 1820 char *bp, *buf; 1821 int i, slash_prefixed; 1822 struct namecache *fd_nrdir; 1823 1824 numfullpathcalls--; 1825 1826 buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK); 1827 bp = buf + MAXPATHLEN - 1; 1828 *bp = '\0'; 1829 if (p != NULL) 1830 fd_nrdir = p->p_fd->fd_nrdir; 1831 else 1832 fd_nrdir = NULL; 1833 slash_prefixed = 0; 1834 while (ncp && ncp != fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) { 1835 if (ncp->nc_flag & NCF_MOUNTPT) { 1836 if (ncp->nc_mount == NULL) { 1837 free(buf, M_TEMP); 1838 return(EBADF); 1839 } 1840 ncp = ncp->nc_parent; 1841 continue; 1842 } 1843 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 1844 if (bp == buf) { 1845 numfullpathfail4++; 1846 free(buf, M_TEMP); 1847 return(ENOMEM); 1848 } 1849 *--bp = ncp->nc_name[i]; 1850 } 1851 if (bp == buf) { 1852 numfullpathfail4++; 1853 free(buf, M_TEMP); 1854 return(ENOMEM); 1855 } 1856 *--bp = '/'; 1857 slash_prefixed = 1; 1858 ncp = ncp->nc_parent; 1859 } 1860 if (ncp == NULL) { 1861 numfullpathfail2++; 1862 free(buf, M_TEMP); 1863 return(ENOENT); 1864 } 1865 if (p != NULL && (ncp->nc_flag & NCF_ROOT) && ncp != fd_nrdir) { 1866 bp = buf + MAXPATHLEN - 1; 1867 *bp = '\0'; 1868 slash_prefixed = 0; 1869 } 1870 if (!slash_prefixed) { 1871 if (bp == buf) { 1872 numfullpathfail4++; 1873 free(buf, M_TEMP); 1874 return(ENOMEM); 1875 } 1876 *--bp = '/'; 1877 } 1878 numfullpathfound++; 1879 *retbuf = bp; 1880 *freebuf = buf; 1881 1882 return(0); 1883 } 1884 1885 int 1886 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf) 1887 { 1888 struct namecache *ncp; 1889 1890 numfullpathcalls++; 1891 if (disablefullpath) 1892 return (ENODEV); 1893 1894 if (p == NULL) 1895 return (EINVAL); 1896 1897 /* vn is NULL, client wants us to use p->p_textvp */ 1898 if (vn == NULL) { 1899 if ((vn = p->p_textvp) == NULL) 1900 return (EINVAL); 1901 } 1902 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) { 1903 if (ncp->nc_nlen) 1904 break; 1905 } 1906 if (ncp == NULL) 1907 return (EINVAL); 1908 1909 numfullpathcalls--; 1910 return(cache_fullpath(p, ncp, retbuf, freebuf)); 1911 } 1912