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 1.51 2005/02/12 18:56:46 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 int 662 cache_inval_vp(struct vnode *vp, int flags) 663 { 664 struct namecache *ncp; 665 struct namecache *next; 666 667 ncp = TAILQ_FIRST(&vp->v_namecache); 668 if (ncp) 669 cache_hold(ncp); 670 while (ncp) { 671 /* loop entered with ncp held */ 672 if ((next = TAILQ_NEXT(ncp, nc_entry)) != NULL) 673 cache_hold(next); 674 cache_lock(ncp); 675 cache_inval(ncp, flags); 676 cache_put(ncp); /* also releases reference */ 677 ncp = next; 678 } 679 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 680 } 681 682 /* 683 * The source ncp has been renamed to the target ncp. Both fncp and tncp 684 * must be locked. Both will be set to unresolved, any children of tncp 685 * will be disconnected (the prior contents of the target is assumed to be 686 * destroyed by the rename operation, e.g. renaming over an empty directory), 687 * and all children of fncp will be moved to tncp. 688 * 689 * XXX the disconnection could pose a problem, check code paths to make 690 * sure any code that blocks can handle the parent being changed out from 691 * under it. Maybe we should lock the children (watch out for deadlocks) ? 692 * 693 * After we return the caller has the option of calling cache_setvp() if 694 * the vnode of the new target ncp is known. 695 * 696 * Any process CD'd into any of the children will no longer be able to ".." 697 * back out. An rm -rf can cause this situation to occur. 698 */ 699 void 700 cache_rename(struct namecache *fncp, struct namecache *tncp) 701 { 702 struct namecache *scan; 703 int didwarn = 0; 704 705 cache_setunresolved(fncp); 706 cache_setunresolved(tncp); 707 while (cache_inval(tncp, CINV_CHILDREN) != 0) { 708 if (didwarn++ % 10 == 0) { 709 printf("Warning: cache_rename: race during " 710 "rename %s->%s\n", 711 fncp->nc_name, tncp->nc_name); 712 } 713 tsleep(tncp, 0, "mvrace", hz / 10); 714 cache_setunresolved(tncp); 715 } 716 while ((scan = TAILQ_FIRST(&fncp->nc_list)) != NULL) { 717 cache_hold(scan); 718 cache_unlink_parent(scan); 719 cache_link_parent(scan, tncp); 720 if (scan->nc_flag & NCF_HASHED) 721 cache_rehash(scan); 722 cache_drop(scan); 723 } 724 } 725 726 /* 727 * vget the vnode associated with the namecache entry. Resolve the namecache 728 * entry if necessary and deal with namecache/vp races. The passed ncp must 729 * be referenced and may be locked. The ncp's ref/locking state is not 730 * effected by this call. 731 * 732 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked 733 * (depending on the passed lk_type) will be returned in *vpp with an error 734 * of 0, or NULL will be returned in *vpp with a non-0 error code. The 735 * most typical error is ENOENT, meaning that the ncp represents a negative 736 * cache hit and there is no vnode to retrieve, but other errors can occur 737 * too. 738 * 739 * The main race we have to deal with are namecache zaps. The ncp itself 740 * will not disappear since it is referenced, and it turns out that the 741 * validity of the vp pointer can be checked simply by rechecking the 742 * contents of ncp->nc_vp. 743 */ 744 int 745 cache_vget(struct namecache *ncp, struct ucred *cred, 746 int lk_type, struct vnode **vpp) 747 { 748 struct vnode *vp; 749 int error; 750 751 again: 752 vp = NULL; 753 if (ncp->nc_flag & NCF_UNRESOLVED) { 754 cache_lock(ncp); 755 error = cache_resolve(ncp, cred); 756 cache_unlock(ncp); 757 } else { 758 error = 0; 759 } 760 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 761 error = vget(vp, lk_type, curthread); 762 if (error) { 763 if (vp != ncp->nc_vp) /* handle cache_zap race */ 764 goto again; 765 vp = NULL; 766 } else if (vp != ncp->nc_vp) { /* handle cache_zap race */ 767 vput(vp); 768 goto again; 769 } 770 } 771 if (error == 0 && vp == NULL) 772 error = ENOENT; 773 *vpp = vp; 774 return(error); 775 } 776 777 int 778 cache_vref(struct namecache *ncp, struct ucred *cred, struct vnode **vpp) 779 { 780 struct vnode *vp; 781 int error; 782 783 again: 784 vp = NULL; 785 if (ncp->nc_flag & NCF_UNRESOLVED) { 786 cache_lock(ncp); 787 error = cache_resolve(ncp, cred); 788 cache_unlock(ncp); 789 } else { 790 error = 0; 791 } 792 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 793 vref(vp); 794 if (vp != ncp->nc_vp) { /* handle cache_zap race */ 795 vrele(vp); 796 goto again; 797 } 798 } 799 if (error == 0 && vp == NULL) 800 error = ENOENT; 801 *vpp = vp; 802 return(error); 803 } 804 805 /* 806 * Convert a directory vnode to a namecache record without any other 807 * knowledge of the topology. This ONLY works with directory vnodes and 808 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the 809 * returned ncp (if not NULL) will be held and unlocked. 810 * 811 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned. 812 * If 'makeit' is 1 we attempt to track-down and create the namecache topology 813 * for dvp. This will fail only if the directory has been deleted out from 814 * under the caller. 815 * 816 * Callers must always check for a NULL return no matter the value of 'makeit'. 817 */ 818 819 static int cache_inefficient_scan(struct namecache *ncp, struct ucred *cred, 820 struct vnode *dvp); 821 822 struct namecache * 823 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit) 824 { 825 struct namecache *ncp; 826 struct vnode *pvp; 827 int error; 828 829 /* 830 * Temporary debugging code to force the directory scanning code 831 * to be exercised. 832 */ 833 ncp = NULL; 834 if (ncvp_debug >= 3 && makeit && TAILQ_FIRST(&dvp->v_namecache)) { 835 ncp = TAILQ_FIRST(&dvp->v_namecache); 836 printf("cache_fromdvp: forcing %s\n", ncp->nc_name); 837 goto force; 838 } 839 840 /* 841 * Loop until resolution, inside code will break out on error. 842 */ 843 while ((ncp = TAILQ_FIRST(&dvp->v_namecache)) == NULL && makeit) { 844 force: 845 /* 846 * If dvp is the root of its filesystem it should already 847 * have a namecache pointer associated with it as a side 848 * effect of the mount, but it may have been disassociated. 849 */ 850 if (dvp->v_flag & VROOT) { 851 ncp = cache_get(dvp->v_mount->mnt_ncp); 852 error = cache_resolve_mp(ncp); 853 cache_put(ncp); 854 if (ncvp_debug) { 855 printf("cache_fromdvp: resolve root of mount %p error %d", 856 dvp->v_mount, error); 857 } 858 if (error) { 859 if (ncvp_debug) 860 printf(" failed\n"); 861 ncp = NULL; 862 break; 863 } 864 if (ncvp_debug) 865 printf(" succeeded\n"); 866 continue; 867 } 868 869 /* 870 * Get the parent directory and resolve its ncp. 871 */ 872 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred); 873 if (error) { 874 printf("lookupdotdot failed %d %p\n", error, pvp); 875 break; 876 } 877 VOP_UNLOCK(pvp, 0, curthread); 878 879 /* 880 * XXX this recursion could run the kernel out of stack, 881 * change to a less efficient algorithm if we get too deep 882 * (use 'makeit' for a depth counter?) 883 */ 884 ncp = cache_fromdvp(pvp, cred, makeit); 885 vrele(pvp); 886 if (ncp == NULL) 887 break; 888 889 /* 890 * Do an inefficient scan of pvp (embodied by ncp) to look 891 * for dvp. This will create a namecache record for dvp on 892 * success. We loop up to recheck on success. 893 * 894 * ncp and dvp are both held but not locked. 895 */ 896 error = cache_inefficient_scan(ncp, cred, dvp); 897 cache_drop(ncp); 898 if (error) { 899 printf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n", 900 pvp, ncp->nc_name, dvp); 901 ncp = NULL; 902 break; 903 } 904 if (ncvp_debug) { 905 printf("cache_fromdvp: scan %p (%s) succeeded\n", 906 pvp, ncp->nc_name); 907 } 908 } 909 if (ncp) 910 cache_hold(ncp); 911 return (ncp); 912 } 913 914 /* 915 * Do an inefficient scan of the directory represented by ncp looking for 916 * the directory vnode dvp. ncp must be held but not locked on entry and 917 * will be held on return. dvp must be refd but not locked on entry and 918 * will remain refd on return. 919 * 920 * Why do this at all? Well, due to its stateless nature the NFS server 921 * converts file handles directly to vnodes without necessarily going through 922 * the namecache ops that would otherwise create the namecache topology 923 * leading to the vnode. We could either (1) Change the namecache algorithms 924 * to allow disconnect namecache records that are re-merged opportunistically, 925 * or (2) Make the NFS server backtrack and scan to recover a connected 926 * namecache topology in order to then be able to issue new API lookups. 927 * 928 * It turns out that (1) is a huge mess. It takes a nice clean set of 929 * namecache algorithms and introduces a lot of complication in every subsystem 930 * that calls into the namecache to deal with the re-merge case, especially 931 * since we are using the namecache to placehold negative lookups and the 932 * vnode might not be immediately assigned. (2) is certainly far less 933 * efficient then (1), but since we are only talking about directories here 934 * (which are likely to remain cached), the case does not actually run all 935 * that often and has the supreme advantage of not polluting the namecache 936 * algorithms. 937 */ 938 static int 939 cache_inefficient_scan(struct namecache *ncp, struct ucred *cred, 940 struct vnode *dvp) 941 { 942 struct nlcomponent nlc; 943 struct namecache *rncp; 944 struct dirent *den; 945 struct vnode *pvp; 946 struct vattr vat; 947 struct iovec iov; 948 struct uio uio; 949 u_long *cookies; 950 off_t baseoff; 951 int ncookies; 952 int blksize; 953 int eofflag; 954 char *rbuf; 955 int error; 956 int xoff; 957 int i; 958 959 vat.va_blocksize = 0; 960 if ((error = VOP_GETATTR(dvp, &vat, curthread)) != 0) 961 return (error); 962 if ((error = cache_vget(ncp, cred, LK_SHARED, &pvp)) != 0) 963 return (error); 964 if (ncvp_debug) 965 printf("inefficient_scan: directory iosize %ld vattr fileid = %ld\n", vat.va_blocksize, (long)vat.va_fileid); 966 if ((blksize = vat.va_blocksize) == 0) 967 blksize = DEV_BSIZE; 968 rbuf = malloc(blksize, M_TEMP, M_WAITOK); 969 rncp = NULL; 970 971 eofflag = 0; 972 uio.uio_offset = 0; 973 cookies = NULL; 974 again: 975 baseoff = uio.uio_offset; 976 iov.iov_base = rbuf; 977 iov.iov_len = blksize; 978 uio.uio_iov = &iov; 979 uio.uio_iovcnt = 1; 980 uio.uio_resid = blksize; 981 uio.uio_segflg = UIO_SYSSPACE; 982 uio.uio_rw = UIO_READ; 983 uio.uio_td = curthread; 984 985 if (cookies) { 986 free(cookies, M_TEMP); 987 cookies = NULL; 988 } 989 if (ncvp_debug >= 2) 990 printf("cache_inefficient_scan: readdir @ %08x\n", (int)baseoff); 991 error = VOP_READDIR(pvp, &uio, cred, &eofflag, &ncookies, &cookies); 992 if (error == 0 && cookies == NULL) 993 error = EPERM; 994 if (error == 0) { 995 for (i = 0; i < ncookies; ++i) { 996 xoff = (int)(cookies[i] - (u_long)baseoff); 997 /* 998 * UFS plays a little trick to skip the first entry 999 * in a directory ("."), by assigning the cookie to 1000 * dpoff + dp->d_reclen in the loop. This causes 1001 * the last cookie to be assigned to the data-end of 1002 * the directory. XXX 1003 */ 1004 if (xoff == blksize) 1005 break; 1006 KKASSERT(xoff >= 0 && xoff <= blksize); 1007 den = (struct dirent *)(rbuf + xoff); 1008 if (ncvp_debug >= 2) 1009 printf("cache_inefficient_scan: %*.*s\n", 1010 den->d_namlen, den->d_namlen, den->d_name); 1011 if (den->d_type != DT_WHT && 1012 den->d_fileno == vat.va_fileid) { 1013 if (ncvp_debug) 1014 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); 1015 nlc.nlc_nameptr = den->d_name; 1016 nlc.nlc_namelen = den->d_namlen; 1017 VOP_UNLOCK(pvp, 0, curthread); 1018 rncp = cache_nlookup(ncp, &nlc); 1019 KKASSERT(rncp != NULL); 1020 break; 1021 } 1022 } 1023 if (rncp == NULL && eofflag == 0 && uio.uio_resid != blksize) 1024 goto again; 1025 } 1026 if (cookies) { 1027 free(cookies, M_TEMP); 1028 cookies = NULL; 1029 } 1030 if (rncp) { 1031 vrele(pvp); 1032 if (rncp->nc_flag & NCF_UNRESOLVED) { 1033 cache_setvp(rncp, dvp); 1034 if (ncvp_debug >= 2) { 1035 printf("cache_inefficient_scan: setvp %s/%s = %p\n", 1036 ncp->nc_name, rncp->nc_name, dvp); 1037 } 1038 } else { 1039 if (ncvp_debug >= 2) { 1040 printf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n", 1041 ncp->nc_name, rncp->nc_name, dvp, 1042 rncp->nc_vp); 1043 } 1044 } 1045 if (rncp->nc_vp == NULL) 1046 error = rncp->nc_error; 1047 cache_put(rncp); 1048 } else { 1049 printf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n", 1050 dvp, ncp->nc_name); 1051 vput(pvp); 1052 error = ENOENT; 1053 } 1054 free(rbuf, M_TEMP); 1055 return (error); 1056 } 1057 1058 /* 1059 * Zap a namecache entry. The ncp is unconditionally set to an unresolved 1060 * state, which disassociates it from its vnode or ncneglist. 1061 * 1062 * Then, if there are no additional references to the ncp and no children, 1063 * the ncp is removed from the topology and destroyed. This function will 1064 * also run through the nc_parent chain and destroy parent ncps if possible. 1065 * As a side benefit, it turns out the only conditions that allow running 1066 * up the chain are also the conditions to ensure no deadlock will occur. 1067 * 1068 * References and/or children may exist if the ncp is in the middle of the 1069 * topology, preventing the ncp from being destroyed. 1070 * 1071 * This function must be called with the ncp held and locked and will unlock 1072 * and drop it during zapping. 1073 */ 1074 static void 1075 cache_zap(struct namecache *ncp) 1076 { 1077 struct namecache *par; 1078 1079 /* 1080 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED. 1081 */ 1082 cache_setunresolved(ncp); 1083 1084 /* 1085 * Try to scrap the entry and possibly tail-recurse on its parent. 1086 * We only scrap unref'd (other then our ref) unresolved entries, 1087 * we do not scrap 'live' entries. 1088 */ 1089 while (ncp->nc_flag & NCF_UNRESOLVED) { 1090 /* 1091 * Someone other then us has a ref, stop. 1092 */ 1093 if (ncp->nc_refs > 1) 1094 goto done; 1095 1096 /* 1097 * We have children, stop. 1098 */ 1099 if (!TAILQ_EMPTY(&ncp->nc_list)) 1100 goto done; 1101 1102 /* 1103 * Remove ncp from the topology: hash table and parent linkage. 1104 */ 1105 if (ncp->nc_flag & NCF_HASHED) { 1106 ncp->nc_flag &= ~NCF_HASHED; 1107 LIST_REMOVE(ncp, nc_hash); 1108 } 1109 if ((par = ncp->nc_parent) != NULL) { 1110 par = cache_hold(par); 1111 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 1112 ncp->nc_parent = NULL; 1113 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 1114 vdrop(par->nc_vp); 1115 } 1116 1117 /* 1118 * ncp should not have picked up any refs. Physically 1119 * destroy the ncp. 1120 */ 1121 KKASSERT(ncp->nc_refs == 1); 1122 --numunres; 1123 /* cache_unlock(ncp) not required */ 1124 ncp->nc_refs = -1; /* safety */ 1125 if (ncp->nc_name) 1126 free(ncp->nc_name, M_VFSCACHE); 1127 free(ncp, M_VFSCACHE); 1128 1129 /* 1130 * Loop on the parent (it may be NULL). Only bother looping 1131 * if the parent has a single ref (ours), which also means 1132 * we can lock it trivially. 1133 */ 1134 ncp = par; 1135 if (ncp == NULL) 1136 return; 1137 if (ncp->nc_refs != 1) { 1138 cache_drop(ncp); 1139 return; 1140 } 1141 KKASSERT(par->nc_exlocks == 0); 1142 cache_lock(ncp); 1143 } 1144 done: 1145 cache_unlock(ncp); 1146 --ncp->nc_refs; 1147 } 1148 1149 static enum { CHI_LOW, CHI_HIGH } cache_hysteresis_state = CHI_LOW; 1150 1151 static __inline 1152 void 1153 cache_hysteresis(void) 1154 { 1155 /* 1156 * Don't cache too many negative hits. We use hysteresis to reduce 1157 * the impact on the critical path. 1158 */ 1159 switch(cache_hysteresis_state) { 1160 case CHI_LOW: 1161 if (numneg > MINNEG && numneg * ncnegfactor > numcache) { 1162 cache_cleanneg(10); 1163 cache_hysteresis_state = CHI_HIGH; 1164 } 1165 break; 1166 case CHI_HIGH: 1167 if (numneg > MINNEG * 9 / 10 && 1168 numneg * ncnegfactor * 9 / 10 > numcache 1169 ) { 1170 cache_cleanneg(10); 1171 } else { 1172 cache_hysteresis_state = CHI_LOW; 1173 } 1174 break; 1175 } 1176 } 1177 1178 /* 1179 * NEW NAMECACHE LOOKUP API 1180 * 1181 * Lookup an entry in the cache. A locked, referenced, non-NULL 1182 * entry is *always* returned, even if the supplied component is illegal. 1183 * The resulting namecache entry should be returned to the system with 1184 * cache_put() or cache_unlock() + cache_drop(). 1185 * 1186 * namecache locks are recursive but care must be taken to avoid lock order 1187 * reversals. 1188 * 1189 * Nobody else will be able to manipulate the associated namespace (e.g. 1190 * create, delete, rename, rename-target) until the caller unlocks the 1191 * entry. 1192 * 1193 * The returned entry will be in one of three states: positive hit (non-null 1194 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set). 1195 * Unresolved entries must be resolved through the filesystem to associate the 1196 * vnode and/or determine whether a positive or negative hit has occured. 1197 * 1198 * It is not necessary to lock a directory in order to lock namespace under 1199 * that directory. In fact, it is explicitly not allowed to do that. A 1200 * directory is typically only locked when being created, renamed, or 1201 * destroyed. 1202 * 1203 * The directory (par) may be unresolved, in which case any returned child 1204 * will likely also be marked unresolved. Likely but not guarenteed. Since 1205 * the filesystem lookup requires a resolved directory vnode the caller is 1206 * responsible for resolving the namecache chain top-down. This API 1207 * specifically allows whole chains to be created in an unresolved state. 1208 */ 1209 struct namecache * 1210 cache_nlookup(struct namecache *par, struct nlcomponent *nlc) 1211 { 1212 struct namecache *ncp; 1213 struct namecache *new_ncp; 1214 struct nchashhead *nchpp; 1215 u_int32_t hash; 1216 globaldata_t gd; 1217 1218 numcalls++; 1219 gd = mycpu; 1220 1221 /* 1222 * Try to locate an existing entry 1223 */ 1224 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 1225 hash = fnv_32_buf(&par, sizeof(par), hash); 1226 new_ncp = NULL; 1227 restart: 1228 LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 1229 numchecks++; 1230 1231 /* 1232 * Zap entries that have timed out. 1233 */ 1234 if (ncp->nc_timeout && 1235 (int)(ncp->nc_timeout - ticks) < 0 && 1236 (ncp->nc_flag & NCF_UNRESOLVED) == 0 && 1237 ncp->nc_exlocks == 0 1238 ) { 1239 cache_zap(cache_get(ncp)); 1240 goto restart; 1241 } 1242 1243 /* 1244 * Break out if we find a matching entry. Note that 1245 * UNRESOLVED entries may match, but DESTROYED entries 1246 * do not. 1247 */ 1248 if (ncp->nc_parent == par && 1249 ncp->nc_nlen == nlc->nlc_namelen && 1250 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 1251 (ncp->nc_flag & NCF_DESTROYED) == 0 1252 ) { 1253 if (cache_get_nonblock(ncp) == 0) { 1254 if (new_ncp) 1255 cache_free(new_ncp); 1256 goto found; 1257 } 1258 cache_get(ncp); 1259 cache_put(ncp); 1260 goto restart; 1261 } 1262 } 1263 1264 /* 1265 * We failed to locate an entry, create a new entry and add it to 1266 * the cache. We have to relookup after possibly blocking in 1267 * malloc. 1268 */ 1269 if (new_ncp == NULL) { 1270 new_ncp = cache_alloc(nlc->nlc_namelen); 1271 goto restart; 1272 } 1273 1274 ncp = new_ncp; 1275 1276 /* 1277 * Initialize as a new UNRESOLVED entry, lock (non-blocking), 1278 * and link to the parent. The mount point is usually inherited 1279 * from the parent unless this is a special case such as a mount 1280 * point where nlc_namelen is 0. The caller is responsible for 1281 * setting nc_mount in that case. If nlc_namelen is 0 nc_name will 1282 * be NULL. 1283 */ 1284 if (nlc->nlc_namelen) { 1285 bcopy(nlc->nlc_nameptr, ncp->nc_name, nlc->nlc_namelen); 1286 ncp->nc_name[nlc->nlc_namelen] = 0; 1287 ncp->nc_mount = par->nc_mount; 1288 } 1289 nchpp = NCHHASH(hash); 1290 LIST_INSERT_HEAD(nchpp, ncp, nc_hash); 1291 ncp->nc_flag |= NCF_HASHED; 1292 cache_link_parent(ncp, par); 1293 found: 1294 /* 1295 * stats and namecache size management 1296 */ 1297 if (ncp->nc_flag & NCF_UNRESOLVED) 1298 ++gd->gd_nchstats->ncs_miss; 1299 else if (ncp->nc_vp) 1300 ++gd->gd_nchstats->ncs_goodhits; 1301 else 1302 ++gd->gd_nchstats->ncs_neghits; 1303 cache_hysteresis(); 1304 return(ncp); 1305 } 1306 1307 /* 1308 * Resolve an unresolved namecache entry, generally by looking it up. 1309 * The passed ncp must be locked and refd. 1310 * 1311 * Theoretically since a vnode cannot be recycled while held, and since 1312 * the nc_parent chain holds its vnode as long as children exist, the 1313 * direct parent of the cache entry we are trying to resolve should 1314 * have a valid vnode. If not then generate an error that we can 1315 * determine is related to a resolver bug. 1316 * 1317 * Note that successful resolution does not necessarily return an error 1318 * code of 0. If the ncp resolves to a negative cache hit then ENOENT 1319 * will be returned. 1320 */ 1321 int 1322 cache_resolve(struct namecache *ncp, struct ucred *cred) 1323 { 1324 struct namecache *par; 1325 int error; 1326 1327 restart: 1328 /* 1329 * If the ncp is already resolved we have nothing to do. 1330 */ 1331 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 1332 return (ncp->nc_error); 1333 1334 /* 1335 * Mount points need special handling because the parent does not 1336 * belong to the same filesystem as the ncp. 1337 */ 1338 if (ncp->nc_flag & NCF_MOUNTPT) 1339 return (cache_resolve_mp(ncp)); 1340 1341 /* 1342 * We expect an unbroken chain of ncps to at least the mount point, 1343 * and even all the way to root (but this code doesn't have to go 1344 * past the mount point). 1345 */ 1346 if (ncp->nc_parent == NULL) { 1347 printf("EXDEV case 1 %p %*.*s\n", ncp, 1348 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 1349 ncp->nc_error = EXDEV; 1350 return(ncp->nc_error); 1351 } 1352 1353 /* 1354 * The vp's of the parent directories in the chain are held via vhold() 1355 * due to the existance of the child, and should not disappear. 1356 * However, there are cases where they can disappear: 1357 * 1358 * - due to filesystem I/O errors. 1359 * - due to NFS being stupid about tracking the namespace and 1360 * destroys the namespace for entire directories quite often. 1361 * - due to forced unmounts. 1362 * - due to an rmdir (parent will be marked DESTROYED) 1363 * 1364 * When this occurs we have to track the chain backwards and resolve 1365 * it, looping until the resolver catches up to the current node. We 1366 * could recurse here but we might run ourselves out of kernel stack 1367 * so we do it in a more painful manner. This situation really should 1368 * not occur all that often, or if it does not have to go back too 1369 * many nodes to resolve the ncp. 1370 */ 1371 while (ncp->nc_parent->nc_vp == NULL) { 1372 /* 1373 * This case can occur if a process is CD'd into a 1374 * directory which is then rmdir'd. If the parent is marked 1375 * destroyed there is no point trying to resolve it. 1376 */ 1377 if (ncp->nc_parent->nc_flag & NCF_DESTROYED) 1378 return(ENOENT); 1379 1380 par = ncp->nc_parent; 1381 while (par->nc_parent && par->nc_parent->nc_vp == NULL) 1382 par = par->nc_parent; 1383 if (par->nc_parent == NULL) { 1384 printf("EXDEV case 2 %*.*s\n", 1385 par->nc_nlen, par->nc_nlen, par->nc_name); 1386 return (EXDEV); 1387 } 1388 printf("[diagnostic] cache_resolve: had to recurse on %*.*s\n", 1389 par->nc_nlen, par->nc_nlen, par->nc_name); 1390 /* 1391 * The parent is not set in stone, ref and lock it to prevent 1392 * it from disappearing. Also note that due to renames it 1393 * is possible for our ncp to move and for par to no longer 1394 * be one of its parents. We resolve it anyway, the loop 1395 * will handle any moves. 1396 */ 1397 cache_get(par); 1398 if (par->nc_flag & NCF_MOUNTPT) { 1399 cache_resolve_mp(par); 1400 } else if (par->nc_parent->nc_vp == NULL) { 1401 printf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name); 1402 cache_put(par); 1403 continue; 1404 } else if (par->nc_flag & NCF_UNRESOLVED) { 1405 par->nc_error = VOP_NRESOLVE(par, cred); 1406 } 1407 if ((error = par->nc_error) != 0) { 1408 if (par->nc_error != EAGAIN) { 1409 printf("EXDEV case 3 %*.*s error %d\n", 1410 par->nc_nlen, par->nc_nlen, par->nc_name, 1411 par->nc_error); 1412 cache_put(par); 1413 return(error); 1414 } 1415 printf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n", 1416 par, par->nc_nlen, par->nc_nlen, par->nc_name); 1417 } 1418 cache_put(par); 1419 /* loop */ 1420 } 1421 1422 /* 1423 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected 1424 * ncp's and reattach them. If this occurs the original ncp is marked 1425 * EAGAIN to force a relookup. 1426 * 1427 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed 1428 * ncp must already be resolved. 1429 */ 1430 KKASSERT((ncp->nc_flag & NCF_MOUNTPT) == 0); 1431 ncp->nc_error = VOP_NRESOLVE(ncp, cred); 1432 /*vop_nresolve(*ncp->nc_parent->nc_vp->v_ops, ncp, cred);*/ 1433 if (ncp->nc_error == EAGAIN) { 1434 printf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n", 1435 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 1436 goto restart; 1437 } 1438 return(ncp->nc_error); 1439 } 1440 1441 /* 1442 * Resolve the ncp associated with a mount point. Such ncp's almost always 1443 * remain resolved and this routine is rarely called. NFS MPs tends to force 1444 * re-resolution more often due to its mac-truck-smash-the-namecache 1445 * method of tracking namespace changes. 1446 * 1447 * The semantics for this call is that the passed ncp must be locked on 1448 * entry and will be locked on return. However, if we actually have to 1449 * resolve the mount point we temporarily unlock the entry in order to 1450 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of 1451 * the unlock we have to recheck the flags after we relock. 1452 */ 1453 static int 1454 cache_resolve_mp(struct namecache *ncp) 1455 { 1456 struct vnode *vp; 1457 struct mount *mp = ncp->nc_mount; 1458 int error; 1459 1460 KKASSERT(mp != NULL); 1461 if (ncp->nc_flag & NCF_UNRESOLVED) { 1462 cache_unlock(ncp); 1463 while (vfs_busy(mp, 0, NULL, curthread)) 1464 ; 1465 error = VFS_ROOT(mp, &vp); 1466 cache_lock(ncp); 1467 1468 /* 1469 * recheck the ncp state after relocking. 1470 */ 1471 if (ncp->nc_flag & NCF_UNRESOLVED) { 1472 ncp->nc_error = error; 1473 if (error == 0) { 1474 cache_setvp(ncp, vp); 1475 vput(vp); 1476 } else { 1477 printf("[diagnostic] cache_resolve_mp: failed to resolve mount %p\n", mp); 1478 cache_setvp(ncp, NULL); 1479 } 1480 } else if (error == 0) { 1481 vput(vp); 1482 } 1483 vfs_unbusy(mp, curthread); 1484 } 1485 return(ncp->nc_error); 1486 } 1487 1488 void 1489 cache_cleanneg(int count) 1490 { 1491 struct namecache *ncp; 1492 1493 /* 1494 * Automode from the vnlru proc - clean out 10% of the negative cache 1495 * entries. 1496 */ 1497 if (count == 0) 1498 count = numneg / 10 + 1; 1499 1500 /* 1501 * Attempt to clean out the specified number of negative cache 1502 * entries. 1503 */ 1504 while (count) { 1505 ncp = TAILQ_FIRST(&ncneglist); 1506 if (ncp == NULL) { 1507 KKASSERT(numneg == 0); 1508 break; 1509 } 1510 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 1511 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 1512 if (cache_get_nonblock(ncp) == 0) 1513 cache_zap(ncp); 1514 --count; 1515 } 1516 } 1517 1518 /* 1519 * Rehash a ncp. Rehashing is typically required if the name changes (should 1520 * not generally occur) or the parent link changes. This function will 1521 * unhash the ncp if the ncp is no longer hashable. 1522 */ 1523 static void 1524 cache_rehash(struct namecache *ncp) 1525 { 1526 struct nchashhead *nchpp; 1527 u_int32_t hash; 1528 1529 if (ncp->nc_flag & NCF_HASHED) { 1530 ncp->nc_flag &= ~NCF_HASHED; 1531 LIST_REMOVE(ncp, nc_hash); 1532 } 1533 if (ncp->nc_nlen && ncp->nc_parent) { 1534 hash = fnv_32_buf(ncp->nc_name, ncp->nc_nlen, FNV1_32_INIT); 1535 hash = fnv_32_buf(&ncp->nc_parent, 1536 sizeof(ncp->nc_parent), hash); 1537 nchpp = NCHHASH(hash); 1538 LIST_INSERT_HEAD(nchpp, ncp, nc_hash); 1539 ncp->nc_flag |= NCF_HASHED; 1540 } 1541 } 1542 1543 /* 1544 * Name cache initialization, from vfsinit() when we are booting 1545 */ 1546 void 1547 nchinit(void) 1548 { 1549 int i; 1550 globaldata_t gd; 1551 1552 /* initialise per-cpu namecache effectiveness statistics. */ 1553 for (i = 0; i < ncpus; ++i) { 1554 gd = globaldata_find(i); 1555 gd->gd_nchstats = &nchstats[i]; 1556 } 1557 TAILQ_INIT(&ncneglist); 1558 nchashtbl = hashinit(desiredvnodes*2, M_VFSCACHE, &nchash); 1559 nclockwarn = 1 * hz; 1560 } 1561 1562 /* 1563 * Called from start_init() to bootstrap the root filesystem. Returns 1564 * a referenced, unlocked namecache record. 1565 */ 1566 struct namecache * 1567 cache_allocroot(struct mount *mp, struct vnode *vp) 1568 { 1569 struct namecache *ncp = cache_alloc(0); 1570 1571 ncp->nc_flag |= NCF_MOUNTPT | NCF_ROOT; 1572 ncp->nc_mount = mp; 1573 cache_setvp(ncp, vp); 1574 return(ncp); 1575 } 1576 1577 /* 1578 * vfs_cache_setroot() 1579 * 1580 * Create an association between the root of our namecache and 1581 * the root vnode. This routine may be called several times during 1582 * booting. 1583 * 1584 * If the caller intends to save the returned namecache pointer somewhere 1585 * it must cache_hold() it. 1586 */ 1587 void 1588 vfs_cache_setroot(struct vnode *nvp, struct namecache *ncp) 1589 { 1590 struct vnode *ovp; 1591 struct namecache *oncp; 1592 1593 ovp = rootvnode; 1594 oncp = rootncp; 1595 rootvnode = nvp; 1596 rootncp = ncp; 1597 1598 if (ovp) 1599 vrele(ovp); 1600 if (oncp) 1601 cache_drop(oncp); 1602 } 1603 1604 /* 1605 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache 1606 * topology and is being removed as quickly as possible. The new VOP_N*() 1607 * API calls are required to make specific adjustments using the supplied 1608 * ncp pointers rather then just bogusly purging random vnodes. 1609 * 1610 * Invalidate all namecache entries to a particular vnode as well as 1611 * any direct children of that vnode in the namecache. This is a 1612 * 'catch all' purge used by filesystems that do not know any better. 1613 * 1614 * A new vnode v_id is generated. Note that no vnode will ever have a 1615 * v_id of 0. 1616 * 1617 * Note that the linkage between the vnode and its namecache entries will 1618 * be removed, but the namecache entries themselves might stay put due to 1619 * active references from elsewhere in the system or due to the existance of 1620 * the children. The namecache topology is left intact even if we do not 1621 * know what the vnode association is. Such entries will be marked 1622 * NCF_UNRESOLVED. 1623 * 1624 * XXX: Only time and the size of v_id prevents this from failing: 1625 * XXX: In theory we should hunt down all (struct vnode*, v_id) 1626 * XXX: soft references and nuke them, at least on the global 1627 * XXX: v_id wraparound. The period of resistance can be extended 1628 * XXX: by incrementing each vnodes v_id individually instead of 1629 * XXX: using the global v_id. 1630 */ 1631 void 1632 cache_purge(struct vnode *vp) 1633 { 1634 static u_long nextid; 1635 1636 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN); 1637 1638 /* 1639 * Calculate a new unique id for ".." handling 1640 */ 1641 do { 1642 nextid++; 1643 } while (nextid == vp->v_id || nextid == 0); 1644 vp->v_id = nextid; 1645 } 1646 1647 /* 1648 * Flush all entries referencing a particular filesystem. 1649 * 1650 * Since we need to check it anyway, we will flush all the invalid 1651 * entries at the same time. 1652 */ 1653 void 1654 cache_purgevfs(struct mount *mp) 1655 { 1656 struct nchashhead *nchpp; 1657 struct namecache *ncp, *nnp; 1658 1659 /* 1660 * Scan hash tables for applicable entries. 1661 */ 1662 for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) { 1663 ncp = LIST_FIRST(nchpp); 1664 if (ncp) 1665 cache_hold(ncp); 1666 while (ncp) { 1667 nnp = LIST_NEXT(ncp, nc_hash); 1668 if (nnp) 1669 cache_hold(nnp); 1670 if (ncp->nc_mount == mp) { 1671 cache_lock(ncp); 1672 cache_zap(ncp); 1673 } else { 1674 cache_drop(ncp); 1675 } 1676 ncp = nnp; 1677 } 1678 } 1679 } 1680 1681 static int disablecwd; 1682 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, ""); 1683 1684 static u_long numcwdcalls; STATNODE(CTLFLAG_RD, numcwdcalls, &numcwdcalls); 1685 static u_long numcwdfail1; STATNODE(CTLFLAG_RD, numcwdfail1, &numcwdfail1); 1686 static u_long numcwdfail2; STATNODE(CTLFLAG_RD, numcwdfail2, &numcwdfail2); 1687 static u_long numcwdfail3; STATNODE(CTLFLAG_RD, numcwdfail3, &numcwdfail3); 1688 static u_long numcwdfail4; STATNODE(CTLFLAG_RD, numcwdfail4, &numcwdfail4); 1689 static u_long numcwdfound; STATNODE(CTLFLAG_RD, numcwdfound, &numcwdfound); 1690 1691 int 1692 __getcwd(struct __getcwd_args *uap) 1693 { 1694 int buflen; 1695 int error; 1696 char *buf; 1697 char *bp; 1698 1699 if (disablecwd) 1700 return (ENODEV); 1701 1702 buflen = uap->buflen; 1703 if (buflen < 2) 1704 return (EINVAL); 1705 if (buflen > MAXPATHLEN) 1706 buflen = MAXPATHLEN; 1707 1708 buf = malloc(buflen, M_TEMP, M_WAITOK); 1709 bp = kern_getcwd(buf, buflen, &error); 1710 if (error == 0) 1711 error = copyout(bp, uap->buf, strlen(bp) + 1); 1712 free(buf, M_TEMP); 1713 return (error); 1714 } 1715 1716 char * 1717 kern_getcwd(char *buf, size_t buflen, int *error) 1718 { 1719 struct proc *p = curproc; 1720 char *bp; 1721 int i, slash_prefixed; 1722 struct filedesc *fdp; 1723 struct namecache *ncp; 1724 1725 numcwdcalls++; 1726 bp = buf; 1727 bp += buflen - 1; 1728 *bp = '\0'; 1729 fdp = p->p_fd; 1730 slash_prefixed = 0; 1731 1732 ncp = fdp->fd_ncdir; 1733 while (ncp && ncp != fdp->fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) { 1734 if (ncp->nc_flag & NCF_MOUNTPT) { 1735 if (ncp->nc_mount == NULL) { 1736 *error = EBADF; /* forced unmount? */ 1737 return(NULL); 1738 } 1739 ncp = ncp->nc_parent; 1740 continue; 1741 } 1742 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 1743 if (bp == buf) { 1744 numcwdfail4++; 1745 *error = ENOMEM; 1746 return(NULL); 1747 } 1748 *--bp = ncp->nc_name[i]; 1749 } 1750 if (bp == buf) { 1751 numcwdfail4++; 1752 *error = ENOMEM; 1753 return(NULL); 1754 } 1755 *--bp = '/'; 1756 slash_prefixed = 1; 1757 ncp = ncp->nc_parent; 1758 } 1759 if (ncp == NULL) { 1760 numcwdfail2++; 1761 *error = ENOENT; 1762 return(NULL); 1763 } 1764 if (!slash_prefixed) { 1765 if (bp == buf) { 1766 numcwdfail4++; 1767 *error = ENOMEM; 1768 return(NULL); 1769 } 1770 *--bp = '/'; 1771 } 1772 numcwdfound++; 1773 *error = 0; 1774 return (bp); 1775 } 1776 1777 /* 1778 * Thus begins the fullpath magic. 1779 */ 1780 1781 #undef STATNODE 1782 #define STATNODE(name) \ 1783 static u_int name; \ 1784 SYSCTL_UINT(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, "") 1785 1786 static int disablefullpath; 1787 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, 1788 &disablefullpath, 0, ""); 1789 1790 STATNODE(numfullpathcalls); 1791 STATNODE(numfullpathfail1); 1792 STATNODE(numfullpathfail2); 1793 STATNODE(numfullpathfail3); 1794 STATNODE(numfullpathfail4); 1795 STATNODE(numfullpathfound); 1796 1797 int 1798 cache_fullpath(struct proc *p, struct namecache *ncp, char **retbuf, char **freebuf) 1799 { 1800 char *bp, *buf; 1801 int i, slash_prefixed; 1802 struct namecache *fd_nrdir; 1803 1804 numfullpathcalls--; 1805 1806 buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK); 1807 bp = buf + MAXPATHLEN - 1; 1808 *bp = '\0'; 1809 if (p != NULL) 1810 fd_nrdir = p->p_fd->fd_nrdir; 1811 else 1812 fd_nrdir = NULL; 1813 slash_prefixed = 0; 1814 while (ncp && ncp != fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) { 1815 if (ncp->nc_flag & NCF_MOUNTPT) { 1816 if (ncp->nc_mount == NULL) { 1817 free(buf, M_TEMP); 1818 return(EBADF); 1819 } 1820 ncp = ncp->nc_parent; 1821 continue; 1822 } 1823 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 1824 if (bp == buf) { 1825 numfullpathfail4++; 1826 free(buf, M_TEMP); 1827 return(ENOMEM); 1828 } 1829 *--bp = ncp->nc_name[i]; 1830 } 1831 if (bp == buf) { 1832 numfullpathfail4++; 1833 free(buf, M_TEMP); 1834 return(ENOMEM); 1835 } 1836 *--bp = '/'; 1837 slash_prefixed = 1; 1838 ncp = ncp->nc_parent; 1839 } 1840 if (ncp == NULL) { 1841 numfullpathfail2++; 1842 free(buf, M_TEMP); 1843 return(ENOENT); 1844 } 1845 if (p != NULL && (ncp->nc_flag & NCF_ROOT) && ncp != fd_nrdir) { 1846 bp = buf + MAXPATHLEN - 1; 1847 *bp = '\0'; 1848 slash_prefixed = 0; 1849 } 1850 if (!slash_prefixed) { 1851 if (bp == buf) { 1852 numfullpathfail4++; 1853 free(buf, M_TEMP); 1854 return(ENOMEM); 1855 } 1856 *--bp = '/'; 1857 } 1858 numfullpathfound++; 1859 *retbuf = bp; 1860 *freebuf = buf; 1861 1862 return(0); 1863 } 1864 1865 int 1866 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf) 1867 { 1868 struct namecache *ncp; 1869 1870 numfullpathcalls++; 1871 if (disablefullpath) 1872 return (ENODEV); 1873 1874 if (p == NULL) 1875 return (EINVAL); 1876 1877 /* vn is NULL, client wants us to use p->p_textvp */ 1878 if (vn == NULL) { 1879 if ((vn = p->p_textvp) == NULL) 1880 return (EINVAL); 1881 } 1882 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) { 1883 if (ncp->nc_nlen) 1884 break; 1885 } 1886 if (ncp == NULL) 1887 return (EINVAL); 1888 1889 numfullpathcalls--; 1890 return(cache_fullpath(p, ncp, retbuf, freebuf)); 1891 } 1892