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 1.61 2006/03/21 18:14:43 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 ncp->nc_fsmid = 1; 298 TAILQ_INIT(&ncp->nc_list); 299 cache_lock(ncp); 300 return(ncp); 301 } 302 303 static void 304 cache_free(struct namecache *ncp) 305 { 306 KKASSERT(ncp->nc_refs == 1 && ncp->nc_exlocks == 1); 307 if (ncp->nc_name) 308 free(ncp->nc_name, M_VFSCACHE); 309 free(ncp, M_VFSCACHE); 310 } 311 312 /* 313 * Ref and deref a namecache structure. 314 */ 315 struct namecache * 316 cache_hold(struct namecache *ncp) 317 { 318 return(_cache_hold(ncp)); 319 } 320 321 void 322 cache_drop(struct namecache *ncp) 323 { 324 _cache_drop(ncp); 325 } 326 327 /* 328 * Namespace locking. The caller must already hold a reference to the 329 * namecache structure in order to lock/unlock it. This function prevents 330 * the namespace from being created or destroyed by accessors other then 331 * the lock holder. 332 * 333 * Note that holding a locked namecache structure prevents other threads 334 * from making namespace changes (e.g. deleting or creating), prevents 335 * vnode association state changes by other threads, and prevents the 336 * namecache entry from being resolved or unresolved by other threads. 337 * 338 * The lock owner has full authority to associate/disassociate vnodes 339 * and resolve/unresolve the locked ncp. 340 * 341 * In particular, if a vnode is associated with a locked cache entry 342 * that vnode will *NOT* be recycled. We accomplish this by vhold()ing the 343 * vnode. XXX we should find a more efficient way to prevent the vnode 344 * from being recycled, but remember that any given vnode may have multiple 345 * namecache associations (think hardlinks). 346 */ 347 void 348 cache_lock(struct namecache *ncp) 349 { 350 thread_t td; 351 int didwarn; 352 353 KKASSERT(ncp->nc_refs != 0); 354 didwarn = 0; 355 td = curthread; 356 357 for (;;) { 358 if (ncp->nc_exlocks == 0) { 359 ncp->nc_exlocks = 1; 360 ncp->nc_locktd = td; 361 /* 362 * The vp associated with a locked ncp must be held 363 * to prevent it from being recycled (which would 364 * cause the ncp to become unresolved). 365 * 366 * XXX loop on race for later MPSAFE work. 367 */ 368 if (ncp->nc_vp) 369 vhold(ncp->nc_vp); 370 break; 371 } 372 if (ncp->nc_locktd == td) { 373 ++ncp->nc_exlocks; 374 break; 375 } 376 ncp->nc_flag |= NCF_LOCKREQ; 377 if (tsleep(ncp, 0, "clock", nclockwarn) == EWOULDBLOCK) { 378 if (didwarn) 379 continue; 380 didwarn = 1; 381 printf("[diagnostic] cache_lock: blocked on %p", ncp); 382 if ((ncp->nc_flag & NCF_MOUNTPT) && ncp->nc_mount) 383 printf(" [MOUNTFROM %s]\n", ncp->nc_mount->mnt_stat.f_mntfromname); 384 else 385 printf(" \"%*.*s\"\n", 386 ncp->nc_nlen, ncp->nc_nlen, 387 ncp->nc_name); 388 } 389 } 390 391 if (didwarn == 1) { 392 printf("[diagnostic] cache_lock: unblocked %*.*s\n", 393 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 394 } 395 } 396 397 int 398 cache_lock_nonblock(struct namecache *ncp) 399 { 400 thread_t td; 401 402 KKASSERT(ncp->nc_refs != 0); 403 td = curthread; 404 if (ncp->nc_exlocks == 0) { 405 ncp->nc_exlocks = 1; 406 ncp->nc_locktd = td; 407 /* 408 * The vp associated with a locked ncp must be held 409 * to prevent it from being recycled (which would 410 * cause the ncp to become unresolved). 411 * 412 * XXX loop on race for later MPSAFE work. 413 */ 414 if (ncp->nc_vp) 415 vhold(ncp->nc_vp); 416 return(0); 417 } else { 418 return(EWOULDBLOCK); 419 } 420 } 421 422 void 423 cache_unlock(struct namecache *ncp) 424 { 425 thread_t td = curthread; 426 427 KKASSERT(ncp->nc_refs > 0); 428 KKASSERT(ncp->nc_exlocks > 0); 429 KKASSERT(ncp->nc_locktd == td); 430 if (--ncp->nc_exlocks == 0) { 431 if (ncp->nc_vp) 432 vdrop(ncp->nc_vp); 433 ncp->nc_locktd = NULL; 434 if (ncp->nc_flag & NCF_LOCKREQ) { 435 ncp->nc_flag &= ~NCF_LOCKREQ; 436 wakeup(ncp); 437 } 438 } 439 } 440 441 /* 442 * ref-and-lock, unlock-and-deref functions. 443 */ 444 struct namecache * 445 cache_get(struct namecache *ncp) 446 { 447 _cache_hold(ncp); 448 cache_lock(ncp); 449 return(ncp); 450 } 451 452 int 453 cache_get_nonblock(struct namecache *ncp) 454 { 455 /* XXX MP */ 456 if (ncp->nc_exlocks == 0 || ncp->nc_locktd == curthread) { 457 _cache_hold(ncp); 458 cache_lock(ncp); 459 return(0); 460 } 461 return(EWOULDBLOCK); 462 } 463 464 void 465 cache_put(struct namecache *ncp) 466 { 467 cache_unlock(ncp); 468 _cache_drop(ncp); 469 } 470 471 /* 472 * Resolve an unresolved ncp by associating a vnode with it. If the 473 * vnode is NULL, a negative cache entry is created. 474 * 475 * The ncp should be locked on entry and will remain locked on return. 476 */ 477 void 478 cache_setvp(struct namecache *ncp, struct vnode *vp) 479 { 480 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED); 481 ncp->nc_vp = vp; 482 if (vp != NULL) { 483 /* 484 * Any vp associated with an ncp which has children must 485 * be held. Any vp associated with a locked ncp must be held. 486 */ 487 if (!TAILQ_EMPTY(&ncp->nc_list)) 488 vhold(vp); 489 TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode); 490 if (ncp->nc_exlocks) 491 vhold(vp); 492 493 /* 494 * Set auxillary flags 495 */ 496 switch(vp->v_type) { 497 case VDIR: 498 ncp->nc_flag |= NCF_ISDIR; 499 break; 500 case VLNK: 501 ncp->nc_flag |= NCF_ISSYMLINK; 502 /* XXX cache the contents of the symlink */ 503 break; 504 default: 505 break; 506 } 507 ++numcache; 508 ncp->nc_error = 0; 509 } else { 510 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 511 ++numneg; 512 ncp->nc_error = ENOENT; 513 } 514 ncp->nc_flag &= ~NCF_UNRESOLVED; 515 } 516 517 void 518 cache_settimeout(struct namecache *ncp, int nticks) 519 { 520 if ((ncp->nc_timeout = ticks + nticks) == 0) 521 ncp->nc_timeout = 1; 522 } 523 524 /* 525 * Disassociate the vnode or negative-cache association and mark a 526 * namecache entry as unresolved again. Note that the ncp is still 527 * left in the hash table and still linked to its parent. 528 * 529 * The ncp should be locked and refd on entry and will remain locked and refd 530 * on return. 531 * 532 * This routine is normally never called on a directory containing children. 533 * However, NFS often does just that in its rename() code as a cop-out to 534 * avoid complex namespace operations. This disconnects a directory vnode 535 * from its namecache and can cause the OLDAPI and NEWAPI to get out of 536 * sync. 537 * 538 * NOTE: NCF_FSMID must be cleared so a refurbishment of the ncp, such as 539 * in a create, properly propogates flag up the chain. 540 */ 541 void 542 cache_setunresolved(struct namecache *ncp) 543 { 544 struct vnode *vp; 545 546 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 547 ncp->nc_flag |= NCF_UNRESOLVED; 548 ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK| 549 NCF_FSMID); 550 ncp->nc_timeout = 0; 551 ncp->nc_error = ENOTCONN; 552 ++numunres; 553 if ((vp = ncp->nc_vp) != NULL) { 554 --numcache; 555 ncp->nc_vp = NULL; 556 TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode); 557 558 /* 559 * Any vp associated with an ncp with children is 560 * held by that ncp. Any vp associated with a locked 561 * ncp is held by that ncp. These conditions must be 562 * undone when the vp is cleared out from the ncp. 563 */ 564 if (!TAILQ_EMPTY(&ncp->nc_list)) 565 vdrop(vp); 566 if (ncp->nc_exlocks) 567 vdrop(vp); 568 } else { 569 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 570 --numneg; 571 } 572 } 573 } 574 575 /* 576 * Invalidate portions of the namecache topology given a starting entry. 577 * The passed ncp is set to an unresolved state and: 578 * 579 * The passed ncp must be locked. 580 * 581 * CINV_DESTROY - Set a flag in the passed ncp entry indicating 582 * that the physical underlying nodes have been 583 * destroyed... as in deleted. For example, when 584 * a directory is removed. This will cause record 585 * lookups on the name to no longer be able to find 586 * the record and tells the resolver to return failure 587 * rather then trying to resolve through the parent. 588 * 589 * The topology itself, including ncp->nc_name, 590 * remains intact. 591 * 592 * This only applies to the passed ncp, if CINV_CHILDREN 593 * is specified the children are not flagged. 594 * 595 * CINV_CHILDREN - Set all children (recursively) to an unresolved 596 * state as well. 597 * 598 * Note that this will also have the side effect of 599 * cleaning out any unreferenced nodes in the topology 600 * from the leaves up as the recursion backs out. 601 * 602 * Note that the topology for any referenced nodes remains intact. 603 * 604 * It is possible for cache_inval() to race a cache_resolve(), meaning that 605 * the namecache entry may not actually be invalidated on return if it was 606 * revalidated while recursing down into its children. This code guarentees 607 * that the node(s) will go through an invalidation cycle, but does not 608 * guarentee that they will remain in an invalidated state. 609 * 610 * Returns non-zero if a revalidation was detected during the invalidation 611 * recursion, zero otherwise. Note that since only the original ncp is 612 * locked the revalidation ultimately can only indicate that the original ncp 613 * *MIGHT* no have been reresolved. 614 */ 615 int 616 cache_inval(struct namecache *ncp, int flags) 617 { 618 struct namecache *kid; 619 struct namecache *nextkid; 620 int rcnt = 0; 621 622 KKASSERT(ncp->nc_exlocks); 623 624 cache_setunresolved(ncp); 625 if (flags & CINV_DESTROY) 626 ncp->nc_flag |= NCF_DESTROYED; 627 628 if ((flags & CINV_CHILDREN) && 629 (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL 630 ) { 631 cache_hold(kid); 632 cache_unlock(ncp); 633 while (kid) { 634 if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL) 635 cache_hold(nextkid); 636 if ((kid->nc_flag & NCF_UNRESOLVED) == 0 || 637 TAILQ_FIRST(&kid->nc_list) 638 ) { 639 cache_lock(kid); 640 rcnt += cache_inval(kid, flags & ~CINV_DESTROY); 641 cache_unlock(kid); 642 } 643 cache_drop(kid); 644 kid = nextkid; 645 } 646 cache_lock(ncp); 647 } 648 649 /* 650 * Someone could have gotten in there while ncp was unlocked, 651 * retry if so. 652 */ 653 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 654 ++rcnt; 655 return (rcnt); 656 } 657 658 /* 659 * Invalidate a vnode's namecache associations. To avoid races against 660 * the resolver we do not invalidate a node which we previously invalidated 661 * but which was then re-resolved while we were in the invalidation loop. 662 * 663 * Returns non-zero if any namecache entries remain after the invalidation 664 * loop completed. 665 * 666 * NOTE: unlike the namecache topology which guarentees that ncp's will not 667 * be ripped out of the topology while held, the vnode's v_namecache list 668 * has no such restriction. NCP's can be ripped out of the list at virtually 669 * any time if not locked, even if held. 670 */ 671 int 672 cache_inval_vp(struct vnode *vp, int flags, int *retflags) 673 { 674 struct namecache *ncp; 675 struct namecache *next; 676 677 restart: 678 ncp = TAILQ_FIRST(&vp->v_namecache); 679 if (ncp) 680 cache_hold(ncp); 681 while (ncp) { 682 /* loop entered with ncp held */ 683 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL) 684 cache_hold(next); 685 cache_lock(ncp); 686 if (ncp->nc_vp != vp) { 687 printf("Warning: cache_inval_vp: race-A detected on " 688 "%s\n", ncp->nc_name); 689 cache_put(ncp); 690 if (next) 691 cache_drop(next); 692 goto restart; 693 } 694 *retflags |= ncp->nc_flag & NCF_FSMID; 695 cache_inval(ncp, flags); 696 cache_put(ncp); /* also releases reference */ 697 ncp = next; 698 if (ncp && ncp->nc_vp != vp) { 699 printf("Warning: cache_inval_vp: race-B detected on " 700 "%s\n", ncp->nc_name); 701 cache_drop(ncp); 702 goto restart; 703 } 704 } 705 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 706 } 707 708 /* 709 * The source ncp has been renamed to the target ncp. Both fncp and tncp 710 * must be locked. Both will be set to unresolved, any children of tncp 711 * will be disconnected (the prior contents of the target is assumed to be 712 * destroyed by the rename operation, e.g. renaming over an empty directory), 713 * and all children of fncp will be moved to tncp. 714 * 715 * XXX the disconnection could pose a problem, check code paths to make 716 * sure any code that blocks can handle the parent being changed out from 717 * under it. Maybe we should lock the children (watch out for deadlocks) ? 718 * 719 * After we return the caller has the option of calling cache_setvp() if 720 * the vnode of the new target ncp is known. 721 * 722 * Any process CD'd into any of the children will no longer be able to ".." 723 * back out. An rm -rf can cause this situation to occur. 724 */ 725 void 726 cache_rename(struct namecache *fncp, struct namecache *tncp) 727 { 728 struct namecache *scan; 729 int didwarn = 0; 730 731 cache_setunresolved(fncp); 732 cache_setunresolved(tncp); 733 while (cache_inval(tncp, CINV_CHILDREN) != 0) { 734 if (didwarn++ % 10 == 0) { 735 printf("Warning: cache_rename: race during " 736 "rename %s->%s\n", 737 fncp->nc_name, tncp->nc_name); 738 } 739 tsleep(tncp, 0, "mvrace", hz / 10); 740 cache_setunresolved(tncp); 741 } 742 while ((scan = TAILQ_FIRST(&fncp->nc_list)) != NULL) { 743 cache_hold(scan); 744 cache_unlink_parent(scan); 745 cache_link_parent(scan, tncp); 746 if (scan->nc_flag & NCF_HASHED) 747 cache_rehash(scan); 748 cache_drop(scan); 749 } 750 } 751 752 /* 753 * vget the vnode associated with the namecache entry. Resolve the namecache 754 * entry if necessary and deal with namecache/vp races. The passed ncp must 755 * be referenced and may be locked. The ncp's ref/locking state is not 756 * effected by this call. 757 * 758 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked 759 * (depending on the passed lk_type) will be returned in *vpp with an error 760 * of 0, or NULL will be returned in *vpp with a non-0 error code. The 761 * most typical error is ENOENT, meaning that the ncp represents a negative 762 * cache hit and there is no vnode to retrieve, but other errors can occur 763 * too. 764 * 765 * The main race we have to deal with are namecache zaps. The ncp itself 766 * will not disappear since it is referenced, and it turns out that the 767 * validity of the vp pointer can be checked simply by rechecking the 768 * contents of ncp->nc_vp. 769 */ 770 int 771 cache_vget(struct namecache *ncp, struct ucred *cred, 772 int lk_type, struct vnode **vpp) 773 { 774 struct vnode *vp; 775 int error; 776 777 again: 778 vp = NULL; 779 if (ncp->nc_flag & NCF_UNRESOLVED) { 780 cache_lock(ncp); 781 error = cache_resolve(ncp, cred); 782 cache_unlock(ncp); 783 } else { 784 error = 0; 785 } 786 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 787 error = vget(vp, lk_type, curthread); 788 if (error) { 789 if (vp != ncp->nc_vp) /* handle cache_zap race */ 790 goto again; 791 vp = NULL; 792 } else if (vp != ncp->nc_vp) { /* handle cache_zap race */ 793 vput(vp); 794 goto again; 795 } 796 } 797 if (error == 0 && vp == NULL) 798 error = ENOENT; 799 *vpp = vp; 800 return(error); 801 } 802 803 int 804 cache_vref(struct namecache *ncp, struct ucred *cred, struct vnode **vpp) 805 { 806 struct vnode *vp; 807 int error; 808 809 again: 810 vp = NULL; 811 if (ncp->nc_flag & NCF_UNRESOLVED) { 812 cache_lock(ncp); 813 error = cache_resolve(ncp, cred); 814 cache_unlock(ncp); 815 } else { 816 error = 0; 817 } 818 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 819 vref(vp); 820 if (vp != ncp->nc_vp) { /* handle cache_zap race */ 821 vrele(vp); 822 goto again; 823 } 824 } 825 if (error == 0 && vp == NULL) 826 error = ENOENT; 827 *vpp = vp; 828 return(error); 829 } 830 831 /* 832 * Recursively set the FSMID update flag for namecache nodes leading 833 * to root. This will cause the next getattr or reclaim to increment the 834 * fsmid and mark the inode for lazy updating. 835 * 836 * Stop recursing when we hit a node whos NCF_FSMID flag is already set. 837 * This makes FSMIDs work in an Einsteinian fashion - where the observation 838 * effects the result. In this case a program monitoring a higher level 839 * node will have detected some prior change and started its scan (clearing 840 * NCF_FSMID in higher level nodes), but since it has not yet observed the 841 * node where we find NCF_FSMID still set, we can safely make the related 842 * modification without interfering with the theorized program. 843 * 844 * This also means that FSMIDs cannot represent time-domain quantities 845 * in a hierarchical sense. But the main reason for doing it this way 846 * is to reduce the amount of recursion that occurs in the critical path 847 * when e.g. a program is writing to a file that sits deep in a directory 848 * hierarchy. 849 */ 850 void 851 cache_update_fsmid(struct namecache *ncp) 852 { 853 struct vnode *vp; 854 struct namecache *scan; 855 856 if ((vp = ncp->nc_vp) != NULL) { 857 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 858 for (scan = ncp; scan; scan = scan->nc_parent) { 859 if (scan->nc_flag & NCF_FSMID) 860 break; 861 scan->nc_flag |= NCF_FSMID; 862 } 863 } 864 } else { 865 while (ncp && (ncp->nc_flag & NCF_FSMID) == 0) { 866 ncp->nc_flag |= NCF_FSMID; 867 ncp = ncp->nc_parent; 868 } 869 } 870 } 871 872 void 873 cache_update_fsmid_vp(struct vnode *vp) 874 { 875 struct namecache *ncp; 876 struct namecache *scan; 877 878 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 879 for (scan = ncp; scan; scan = scan->nc_parent) { 880 if (scan->nc_flag & NCF_FSMID) 881 break; 882 scan->nc_flag |= NCF_FSMID; 883 } 884 } 885 } 886 887 /* 888 * If getattr is called on a vnode (e.g. a stat call), the filesystem 889 * may call this routine to determine if the namecache has the hierarchical 890 * change flag set, requiring the fsmid to be updated. 891 * 892 * Since 0 indicates no support, make sure the filesystem fsmid is at least 893 * 1. 894 */ 895 int 896 cache_check_fsmid_vp(struct vnode *vp, int64_t *fsmid) 897 { 898 struct namecache *ncp; 899 int changed = 0; 900 901 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 902 if (ncp->nc_flag & NCF_FSMID) { 903 ncp->nc_flag &= ~NCF_FSMID; 904 changed = 1; 905 } 906 } 907 if (*fsmid == 0) 908 ++*fsmid; 909 if (changed) 910 ++*fsmid; 911 return(changed); 912 } 913 914 /* 915 * Convert a directory vnode to a namecache record without any other 916 * knowledge of the topology. This ONLY works with directory vnodes and 917 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the 918 * returned ncp (if not NULL) will be held and unlocked. 919 * 920 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned. 921 * If 'makeit' is 1 we attempt to track-down and create the namecache topology 922 * for dvp. This will fail only if the directory has been deleted out from 923 * under the caller. 924 * 925 * Callers must always check for a NULL return no matter the value of 'makeit'. 926 * 927 * To avoid underflowing the kernel stack each recursive call increments 928 * the makeit variable. 929 */ 930 931 static int cache_inefficient_scan(struct namecache *ncp, struct ucred *cred, 932 struct vnode *dvp); 933 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred, 934 struct vnode **saved_dvp); 935 936 struct namecache * 937 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit) 938 { 939 struct namecache *ncp; 940 struct vnode *saved_dvp; 941 struct vnode *pvp; 942 int error; 943 944 ncp = NULL; 945 saved_dvp = NULL; 946 947 /* 948 * Temporary debugging code to force the directory scanning code 949 * to be exercised. 950 */ 951 if (ncvp_debug >= 3 && makeit && TAILQ_FIRST(&dvp->v_namecache)) { 952 ncp = TAILQ_FIRST(&dvp->v_namecache); 953 printf("cache_fromdvp: forcing %s\n", ncp->nc_name); 954 goto force; 955 } 956 957 /* 958 * Loop until resolution, inside code will break out on error. 959 */ 960 while ((ncp = TAILQ_FIRST(&dvp->v_namecache)) == NULL && makeit) { 961 force: 962 /* 963 * If dvp is the root of its filesystem it should already 964 * have a namecache pointer associated with it as a side 965 * effect of the mount, but it may have been disassociated. 966 */ 967 if (dvp->v_flag & VROOT) { 968 ncp = cache_get(dvp->v_mount->mnt_ncp); 969 error = cache_resolve_mp(ncp); 970 cache_put(ncp); 971 if (ncvp_debug) { 972 printf("cache_fromdvp: resolve root of mount %p error %d", 973 dvp->v_mount, error); 974 } 975 if (error) { 976 if (ncvp_debug) 977 printf(" failed\n"); 978 ncp = NULL; 979 break; 980 } 981 if (ncvp_debug) 982 printf(" succeeded\n"); 983 continue; 984 } 985 986 /* 987 * If we are recursed too deeply resort to an O(n^2) 988 * algorithm to resolve the namecache topology. The 989 * resolved pvp is left referenced in saved_dvp to 990 * prevent the tree from being destroyed while we loop. 991 */ 992 if (makeit > 20) { 993 error = cache_fromdvp_try(dvp, cred, &saved_dvp); 994 if (error) { 995 printf("lookupdotdot(longpath) failed %d " 996 "dvp %p\n", error, dvp); 997 break; 998 } 999 continue; 1000 } 1001 1002 /* 1003 * Get the parent directory and resolve its ncp. 1004 */ 1005 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred); 1006 if (error) { 1007 printf("lookupdotdot failed %d dvp %p\n", error, dvp); 1008 break; 1009 } 1010 VOP_UNLOCK(pvp, 0, curthread); 1011 1012 /* 1013 * Reuse makeit as a recursion depth counter. 1014 */ 1015 ncp = cache_fromdvp(pvp, cred, makeit + 1); 1016 vrele(pvp); 1017 if (ncp == NULL) 1018 break; 1019 1020 /* 1021 * Do an inefficient scan of pvp (embodied by ncp) to look 1022 * for dvp. This will create a namecache record for dvp on 1023 * success. We loop up to recheck on success. 1024 * 1025 * ncp and dvp are both held but not locked. 1026 */ 1027 error = cache_inefficient_scan(ncp, cred, dvp); 1028 cache_drop(ncp); 1029 if (error) { 1030 printf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n", 1031 pvp, ncp->nc_name, dvp); 1032 ncp = NULL; 1033 break; 1034 } 1035 if (ncvp_debug) { 1036 printf("cache_fromdvp: scan %p (%s) succeeded\n", 1037 pvp, ncp->nc_name); 1038 } 1039 } 1040 if (ncp) 1041 cache_hold(ncp); 1042 if (saved_dvp) 1043 vrele(saved_dvp); 1044 return (ncp); 1045 } 1046 1047 /* 1048 * Go up the chain of parent directories until we find something 1049 * we can resolve into the namecache. This is very inefficient. 1050 */ 1051 static 1052 int 1053 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred, 1054 struct vnode **saved_dvp) 1055 { 1056 struct namecache *ncp; 1057 struct vnode *pvp; 1058 int error; 1059 static time_t last_fromdvp_report; 1060 1061 /* 1062 * Loop getting the parent directory vnode until we get something we 1063 * can resolve in the namecache. 1064 */ 1065 vref(dvp); 1066 for (;;) { 1067 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred); 1068 if (error) { 1069 vrele(dvp); 1070 return (error); 1071 } 1072 VOP_UNLOCK(pvp, 0, curthread); 1073 if ((ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) { 1074 cache_hold(ncp); 1075 vrele(pvp); 1076 break; 1077 } 1078 if (pvp->v_flag & VROOT) { 1079 ncp = cache_get(pvp->v_mount->mnt_ncp); 1080 error = cache_resolve_mp(ncp); 1081 cache_unlock(ncp); 1082 vrele(pvp); 1083 if (error) { 1084 cache_drop(ncp); 1085 vrele(dvp); 1086 return (error); 1087 } 1088 break; 1089 } 1090 vrele(dvp); 1091 dvp = pvp; 1092 } 1093 if (last_fromdvp_report != time_second) { 1094 last_fromdvp_report = time_second; 1095 printf("Warning: extremely inefficient path resolution on %s\n", 1096 ncp->nc_name); 1097 } 1098 error = cache_inefficient_scan(ncp, cred, dvp); 1099 1100 /* 1101 * Hopefully dvp now has a namecache record associated with it. 1102 * Leave it referenced to prevent the kernel from recycling the 1103 * vnode. Otherwise extremely long directory paths could result 1104 * in endless recycling. 1105 */ 1106 if (*saved_dvp) 1107 vrele(*saved_dvp); 1108 *saved_dvp = dvp; 1109 return (error); 1110 } 1111 1112 1113 /* 1114 * Do an inefficient scan of the directory represented by ncp looking for 1115 * the directory vnode dvp. ncp must be held but not locked on entry and 1116 * will be held on return. dvp must be refd but not locked on entry and 1117 * will remain refd on return. 1118 * 1119 * Why do this at all? Well, due to its stateless nature the NFS server 1120 * converts file handles directly to vnodes without necessarily going through 1121 * the namecache ops that would otherwise create the namecache topology 1122 * leading to the vnode. We could either (1) Change the namecache algorithms 1123 * to allow disconnect namecache records that are re-merged opportunistically, 1124 * or (2) Make the NFS server backtrack and scan to recover a connected 1125 * namecache topology in order to then be able to issue new API lookups. 1126 * 1127 * It turns out that (1) is a huge mess. It takes a nice clean set of 1128 * namecache algorithms and introduces a lot of complication in every subsystem 1129 * that calls into the namecache to deal with the re-merge case, especially 1130 * since we are using the namecache to placehold negative lookups and the 1131 * vnode might not be immediately assigned. (2) is certainly far less 1132 * efficient then (1), but since we are only talking about directories here 1133 * (which are likely to remain cached), the case does not actually run all 1134 * that often and has the supreme advantage of not polluting the namecache 1135 * algorithms. 1136 */ 1137 static int 1138 cache_inefficient_scan(struct namecache *ncp, struct ucred *cred, 1139 struct vnode *dvp) 1140 { 1141 struct nlcomponent nlc; 1142 struct namecache *rncp; 1143 struct dirent *den; 1144 struct vnode *pvp; 1145 struct vattr vat; 1146 struct iovec iov; 1147 struct uio uio; 1148 int blksize; 1149 int eofflag; 1150 int bytes; 1151 char *rbuf; 1152 int error; 1153 1154 vat.va_blocksize = 0; 1155 if ((error = VOP_GETATTR(dvp, &vat, curthread)) != 0) 1156 return (error); 1157 if ((error = cache_vget(ncp, cred, LK_SHARED, &pvp)) != 0) 1158 return (error); 1159 if (ncvp_debug) 1160 printf("inefficient_scan: directory iosize %ld vattr fileid = %ld\n", vat.va_blocksize, (long)vat.va_fileid); 1161 if ((blksize = vat.va_blocksize) == 0) 1162 blksize = DEV_BSIZE; 1163 rbuf = malloc(blksize, M_TEMP, M_WAITOK); 1164 rncp = NULL; 1165 1166 eofflag = 0; 1167 uio.uio_offset = 0; 1168 again: 1169 iov.iov_base = rbuf; 1170 iov.iov_len = blksize; 1171 uio.uio_iov = &iov; 1172 uio.uio_iovcnt = 1; 1173 uio.uio_resid = blksize; 1174 uio.uio_segflg = UIO_SYSSPACE; 1175 uio.uio_rw = UIO_READ; 1176 uio.uio_td = curthread; 1177 1178 if (ncvp_debug >= 2) 1179 printf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset); 1180 error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL); 1181 if (error == 0) { 1182 den = (struct dirent *)rbuf; 1183 bytes = blksize - uio.uio_resid; 1184 1185 while (bytes > 0) { 1186 if (ncvp_debug >= 2) { 1187 printf("cache_inefficient_scan: %*.*s\n", 1188 den->d_namlen, den->d_namlen, 1189 den->d_name); 1190 } 1191 if (den->d_type != DT_WHT && 1192 den->d_ino == vat.va_fileid) { 1193 if (ncvp_debug) { 1194 printf("cache_inefficient_scan: " 1195 "MATCHED inode %ld path %s/%*.*s\n", 1196 vat.va_fileid, ncp->nc_name, 1197 den->d_namlen, den->d_namlen, 1198 den->d_name); 1199 } 1200 nlc.nlc_nameptr = den->d_name; 1201 nlc.nlc_namelen = den->d_namlen; 1202 VOP_UNLOCK(pvp, 0, curthread); 1203 rncp = cache_nlookup(ncp, &nlc); 1204 KKASSERT(rncp != NULL); 1205 break; 1206 } 1207 bytes -= _DIRENT_DIRSIZ(den); 1208 den = _DIRENT_NEXT(den); 1209 } 1210 if (rncp == NULL && eofflag == 0 && uio.uio_resid != blksize) 1211 goto again; 1212 } 1213 if (rncp) { 1214 vrele(pvp); 1215 if (rncp->nc_flag & NCF_UNRESOLVED) { 1216 cache_setvp(rncp, dvp); 1217 if (ncvp_debug >= 2) { 1218 printf("cache_inefficient_scan: setvp %s/%s = %p\n", 1219 ncp->nc_name, rncp->nc_name, dvp); 1220 } 1221 } else { 1222 if (ncvp_debug >= 2) { 1223 printf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n", 1224 ncp->nc_name, rncp->nc_name, dvp, 1225 rncp->nc_vp); 1226 } 1227 } 1228 if (rncp->nc_vp == NULL) 1229 error = rncp->nc_error; 1230 cache_put(rncp); 1231 } else { 1232 printf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n", 1233 dvp, ncp->nc_name); 1234 vput(pvp); 1235 error = ENOENT; 1236 } 1237 free(rbuf, M_TEMP); 1238 return (error); 1239 } 1240 1241 /* 1242 * Zap a namecache entry. The ncp is unconditionally set to an unresolved 1243 * state, which disassociates it from its vnode or ncneglist. 1244 * 1245 * Then, if there are no additional references to the ncp and no children, 1246 * the ncp is removed from the topology and destroyed. This function will 1247 * also run through the nc_parent chain and destroy parent ncps if possible. 1248 * As a side benefit, it turns out the only conditions that allow running 1249 * up the chain are also the conditions to ensure no deadlock will occur. 1250 * 1251 * References and/or children may exist if the ncp is in the middle of the 1252 * topology, preventing the ncp from being destroyed. 1253 * 1254 * This function must be called with the ncp held and locked and will unlock 1255 * and drop it during zapping. 1256 */ 1257 static void 1258 cache_zap(struct namecache *ncp) 1259 { 1260 struct namecache *par; 1261 1262 /* 1263 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED. 1264 */ 1265 cache_setunresolved(ncp); 1266 1267 /* 1268 * Try to scrap the entry and possibly tail-recurse on its parent. 1269 * We only scrap unref'd (other then our ref) unresolved entries, 1270 * we do not scrap 'live' entries. 1271 */ 1272 while (ncp->nc_flag & NCF_UNRESOLVED) { 1273 /* 1274 * Someone other then us has a ref, stop. 1275 */ 1276 if (ncp->nc_refs > 1) 1277 goto done; 1278 1279 /* 1280 * We have children, stop. 1281 */ 1282 if (!TAILQ_EMPTY(&ncp->nc_list)) 1283 goto done; 1284 1285 /* 1286 * Remove ncp from the topology: hash table and parent linkage. 1287 */ 1288 if (ncp->nc_flag & NCF_HASHED) { 1289 ncp->nc_flag &= ~NCF_HASHED; 1290 LIST_REMOVE(ncp, nc_hash); 1291 } 1292 if ((par = ncp->nc_parent) != NULL) { 1293 par = cache_hold(par); 1294 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 1295 ncp->nc_parent = NULL; 1296 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 1297 vdrop(par->nc_vp); 1298 } 1299 1300 /* 1301 * ncp should not have picked up any refs. Physically 1302 * destroy the ncp. 1303 */ 1304 KKASSERT(ncp->nc_refs == 1); 1305 --numunres; 1306 /* cache_unlock(ncp) not required */ 1307 ncp->nc_refs = -1; /* safety */ 1308 if (ncp->nc_name) 1309 free(ncp->nc_name, M_VFSCACHE); 1310 free(ncp, M_VFSCACHE); 1311 1312 /* 1313 * Loop on the parent (it may be NULL). Only bother looping 1314 * if the parent has a single ref (ours), which also means 1315 * we can lock it trivially. 1316 */ 1317 ncp = par; 1318 if (ncp == NULL) 1319 return; 1320 if (ncp->nc_refs != 1) { 1321 cache_drop(ncp); 1322 return; 1323 } 1324 KKASSERT(par->nc_exlocks == 0); 1325 cache_lock(ncp); 1326 } 1327 done: 1328 cache_unlock(ncp); 1329 --ncp->nc_refs; 1330 } 1331 1332 static enum { CHI_LOW, CHI_HIGH } cache_hysteresis_state = CHI_LOW; 1333 1334 static __inline 1335 void 1336 cache_hysteresis(void) 1337 { 1338 /* 1339 * Don't cache too many negative hits. We use hysteresis to reduce 1340 * the impact on the critical path. 1341 */ 1342 switch(cache_hysteresis_state) { 1343 case CHI_LOW: 1344 if (numneg > MINNEG && numneg * ncnegfactor > numcache) { 1345 cache_cleanneg(10); 1346 cache_hysteresis_state = CHI_HIGH; 1347 } 1348 break; 1349 case CHI_HIGH: 1350 if (numneg > MINNEG * 9 / 10 && 1351 numneg * ncnegfactor * 9 / 10 > numcache 1352 ) { 1353 cache_cleanneg(10); 1354 } else { 1355 cache_hysteresis_state = CHI_LOW; 1356 } 1357 break; 1358 } 1359 } 1360 1361 /* 1362 * NEW NAMECACHE LOOKUP API 1363 * 1364 * Lookup an entry in the cache. A locked, referenced, non-NULL 1365 * entry is *always* returned, even if the supplied component is illegal. 1366 * The resulting namecache entry should be returned to the system with 1367 * cache_put() or cache_unlock() + cache_drop(). 1368 * 1369 * namecache locks are recursive but care must be taken to avoid lock order 1370 * reversals. 1371 * 1372 * Nobody else will be able to manipulate the associated namespace (e.g. 1373 * create, delete, rename, rename-target) until the caller unlocks the 1374 * entry. 1375 * 1376 * The returned entry will be in one of three states: positive hit (non-null 1377 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set). 1378 * Unresolved entries must be resolved through the filesystem to associate the 1379 * vnode and/or determine whether a positive or negative hit has occured. 1380 * 1381 * It is not necessary to lock a directory in order to lock namespace under 1382 * that directory. In fact, it is explicitly not allowed to do that. A 1383 * directory is typically only locked when being created, renamed, or 1384 * destroyed. 1385 * 1386 * The directory (par) may be unresolved, in which case any returned child 1387 * will likely also be marked unresolved. Likely but not guarenteed. Since 1388 * the filesystem lookup requires a resolved directory vnode the caller is 1389 * responsible for resolving the namecache chain top-down. This API 1390 * specifically allows whole chains to be created in an unresolved state. 1391 */ 1392 struct namecache * 1393 cache_nlookup(struct namecache *par, struct nlcomponent *nlc) 1394 { 1395 struct namecache *ncp; 1396 struct namecache *new_ncp; 1397 struct nchashhead *nchpp; 1398 u_int32_t hash; 1399 globaldata_t gd; 1400 1401 numcalls++; 1402 gd = mycpu; 1403 1404 /* 1405 * Try to locate an existing entry 1406 */ 1407 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 1408 hash = fnv_32_buf(&par, sizeof(par), hash); 1409 new_ncp = NULL; 1410 restart: 1411 LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 1412 numchecks++; 1413 1414 /* 1415 * Zap entries that have timed out. 1416 */ 1417 if (ncp->nc_timeout && 1418 (int)(ncp->nc_timeout - ticks) < 0 && 1419 (ncp->nc_flag & NCF_UNRESOLVED) == 0 && 1420 ncp->nc_exlocks == 0 1421 ) { 1422 cache_zap(cache_get(ncp)); 1423 goto restart; 1424 } 1425 1426 /* 1427 * Break out if we find a matching entry. Note that 1428 * UNRESOLVED entries may match, but DESTROYED entries 1429 * do not. 1430 */ 1431 if (ncp->nc_parent == par && 1432 ncp->nc_nlen == nlc->nlc_namelen && 1433 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 1434 (ncp->nc_flag & NCF_DESTROYED) == 0 1435 ) { 1436 if (cache_get_nonblock(ncp) == 0) { 1437 if (new_ncp) 1438 cache_free(new_ncp); 1439 goto found; 1440 } 1441 cache_get(ncp); 1442 cache_put(ncp); 1443 goto restart; 1444 } 1445 } 1446 1447 /* 1448 * We failed to locate an entry, create a new entry and add it to 1449 * the cache. We have to relookup after possibly blocking in 1450 * malloc. 1451 */ 1452 if (new_ncp == NULL) { 1453 new_ncp = cache_alloc(nlc->nlc_namelen); 1454 goto restart; 1455 } 1456 1457 ncp = new_ncp; 1458 1459 /* 1460 * Initialize as a new UNRESOLVED entry, lock (non-blocking), 1461 * and link to the parent. The mount point is usually inherited 1462 * from the parent unless this is a special case such as a mount 1463 * point where nlc_namelen is 0. The caller is responsible for 1464 * setting nc_mount in that case. If nlc_namelen is 0 nc_name will 1465 * be NULL. 1466 */ 1467 if (nlc->nlc_namelen) { 1468 bcopy(nlc->nlc_nameptr, ncp->nc_name, nlc->nlc_namelen); 1469 ncp->nc_name[nlc->nlc_namelen] = 0; 1470 ncp->nc_mount = par->nc_mount; 1471 } 1472 nchpp = NCHHASH(hash); 1473 LIST_INSERT_HEAD(nchpp, ncp, nc_hash); 1474 ncp->nc_flag |= NCF_HASHED; 1475 cache_link_parent(ncp, par); 1476 found: 1477 /* 1478 * stats and namecache size management 1479 */ 1480 if (ncp->nc_flag & NCF_UNRESOLVED) 1481 ++gd->gd_nchstats->ncs_miss; 1482 else if (ncp->nc_vp) 1483 ++gd->gd_nchstats->ncs_goodhits; 1484 else 1485 ++gd->gd_nchstats->ncs_neghits; 1486 cache_hysteresis(); 1487 return(ncp); 1488 } 1489 1490 /* 1491 * Resolve an unresolved namecache entry, generally by looking it up. 1492 * The passed ncp must be locked and refd. 1493 * 1494 * Theoretically since a vnode cannot be recycled while held, and since 1495 * the nc_parent chain holds its vnode as long as children exist, the 1496 * direct parent of the cache entry we are trying to resolve should 1497 * have a valid vnode. If not then generate an error that we can 1498 * determine is related to a resolver bug. 1499 * 1500 * Note that successful resolution does not necessarily return an error 1501 * code of 0. If the ncp resolves to a negative cache hit then ENOENT 1502 * will be returned. 1503 */ 1504 int 1505 cache_resolve(struct namecache *ncp, struct ucred *cred) 1506 { 1507 struct namecache *par; 1508 int error; 1509 1510 restart: 1511 /* 1512 * If the ncp is already resolved we have nothing to do. 1513 */ 1514 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 1515 return (ncp->nc_error); 1516 1517 /* 1518 * Mount points need special handling because the parent does not 1519 * belong to the same filesystem as the ncp. 1520 */ 1521 if (ncp->nc_flag & NCF_MOUNTPT) 1522 return (cache_resolve_mp(ncp)); 1523 1524 /* 1525 * We expect an unbroken chain of ncps to at least the mount point, 1526 * and even all the way to root (but this code doesn't have to go 1527 * past the mount point). 1528 */ 1529 if (ncp->nc_parent == NULL) { 1530 printf("EXDEV case 1 %p %*.*s\n", ncp, 1531 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 1532 ncp->nc_error = EXDEV; 1533 return(ncp->nc_error); 1534 } 1535 1536 /* 1537 * The vp's of the parent directories in the chain are held via vhold() 1538 * due to the existance of the child, and should not disappear. 1539 * However, there are cases where they can disappear: 1540 * 1541 * - due to filesystem I/O errors. 1542 * - due to NFS being stupid about tracking the namespace and 1543 * destroys the namespace for entire directories quite often. 1544 * - due to forced unmounts. 1545 * - due to an rmdir (parent will be marked DESTROYED) 1546 * 1547 * When this occurs we have to track the chain backwards and resolve 1548 * it, looping until the resolver catches up to the current node. We 1549 * could recurse here but we might run ourselves out of kernel stack 1550 * so we do it in a more painful manner. This situation really should 1551 * not occur all that often, or if it does not have to go back too 1552 * many nodes to resolve the ncp. 1553 */ 1554 while (ncp->nc_parent->nc_vp == NULL) { 1555 /* 1556 * This case can occur if a process is CD'd into a 1557 * directory which is then rmdir'd. If the parent is marked 1558 * destroyed there is no point trying to resolve it. 1559 */ 1560 if (ncp->nc_parent->nc_flag & NCF_DESTROYED) 1561 return(ENOENT); 1562 1563 par = ncp->nc_parent; 1564 while (par->nc_parent && par->nc_parent->nc_vp == NULL) 1565 par = par->nc_parent; 1566 if (par->nc_parent == NULL) { 1567 printf("EXDEV case 2 %*.*s\n", 1568 par->nc_nlen, par->nc_nlen, par->nc_name); 1569 return (EXDEV); 1570 } 1571 printf("[diagnostic] cache_resolve: had to recurse on %*.*s\n", 1572 par->nc_nlen, par->nc_nlen, par->nc_name); 1573 /* 1574 * The parent is not set in stone, ref and lock it to prevent 1575 * it from disappearing. Also note that due to renames it 1576 * is possible for our ncp to move and for par to no longer 1577 * be one of its parents. We resolve it anyway, the loop 1578 * will handle any moves. 1579 */ 1580 cache_get(par); 1581 if (par->nc_flag & NCF_MOUNTPT) { 1582 cache_resolve_mp(par); 1583 } else if (par->nc_parent->nc_vp == NULL) { 1584 printf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name); 1585 cache_put(par); 1586 continue; 1587 } else if (par->nc_flag & NCF_UNRESOLVED) { 1588 par->nc_error = VOP_NRESOLVE(par, cred); 1589 } 1590 if ((error = par->nc_error) != 0) { 1591 if (par->nc_error != EAGAIN) { 1592 printf("EXDEV case 3 %*.*s error %d\n", 1593 par->nc_nlen, par->nc_nlen, par->nc_name, 1594 par->nc_error); 1595 cache_put(par); 1596 return(error); 1597 } 1598 printf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n", 1599 par, par->nc_nlen, par->nc_nlen, par->nc_name); 1600 } 1601 cache_put(par); 1602 /* loop */ 1603 } 1604 1605 /* 1606 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected 1607 * ncp's and reattach them. If this occurs the original ncp is marked 1608 * EAGAIN to force a relookup. 1609 * 1610 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed 1611 * ncp must already be resolved. 1612 */ 1613 KKASSERT((ncp->nc_flag & NCF_MOUNTPT) == 0); 1614 ncp->nc_error = VOP_NRESOLVE(ncp, cred); 1615 /*vop_nresolve(*ncp->nc_parent->nc_vp->v_ops, ncp, cred);*/ 1616 if (ncp->nc_error == EAGAIN) { 1617 printf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n", 1618 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 1619 goto restart; 1620 } 1621 return(ncp->nc_error); 1622 } 1623 1624 /* 1625 * Resolve the ncp associated with a mount point. Such ncp's almost always 1626 * remain resolved and this routine is rarely called. NFS MPs tends to force 1627 * re-resolution more often due to its mac-truck-smash-the-namecache 1628 * method of tracking namespace changes. 1629 * 1630 * The semantics for this call is that the passed ncp must be locked on 1631 * entry and will be locked on return. However, if we actually have to 1632 * resolve the mount point we temporarily unlock the entry in order to 1633 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of 1634 * the unlock we have to recheck the flags after we relock. 1635 */ 1636 static int 1637 cache_resolve_mp(struct namecache *ncp) 1638 { 1639 struct vnode *vp; 1640 struct mount *mp = ncp->nc_mount; 1641 int error; 1642 1643 KKASSERT(mp != NULL); 1644 if (ncp->nc_flag & NCF_UNRESOLVED) { 1645 cache_unlock(ncp); 1646 while (vfs_busy(mp, 0, curthread)) 1647 ; 1648 error = VFS_ROOT(mp, &vp); 1649 cache_lock(ncp); 1650 1651 /* 1652 * recheck the ncp state after relocking. 1653 */ 1654 if (ncp->nc_flag & NCF_UNRESOLVED) { 1655 ncp->nc_error = error; 1656 if (error == 0) { 1657 cache_setvp(ncp, vp); 1658 vput(vp); 1659 } else { 1660 printf("[diagnostic] cache_resolve_mp: failed to resolve mount %p\n", mp); 1661 cache_setvp(ncp, NULL); 1662 } 1663 } else if (error == 0) { 1664 vput(vp); 1665 } 1666 vfs_unbusy(mp, curthread); 1667 } 1668 return(ncp->nc_error); 1669 } 1670 1671 void 1672 cache_cleanneg(int count) 1673 { 1674 struct namecache *ncp; 1675 1676 /* 1677 * Automode from the vnlru proc - clean out 10% of the negative cache 1678 * entries. 1679 */ 1680 if (count == 0) 1681 count = numneg / 10 + 1; 1682 1683 /* 1684 * Attempt to clean out the specified number of negative cache 1685 * entries. 1686 */ 1687 while (count) { 1688 ncp = TAILQ_FIRST(&ncneglist); 1689 if (ncp == NULL) { 1690 KKASSERT(numneg == 0); 1691 break; 1692 } 1693 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 1694 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 1695 if (cache_get_nonblock(ncp) == 0) 1696 cache_zap(ncp); 1697 --count; 1698 } 1699 } 1700 1701 /* 1702 * Rehash a ncp. Rehashing is typically required if the name changes (should 1703 * not generally occur) or the parent link changes. This function will 1704 * unhash the ncp if the ncp is no longer hashable. 1705 */ 1706 static void 1707 cache_rehash(struct namecache *ncp) 1708 { 1709 struct nchashhead *nchpp; 1710 u_int32_t hash; 1711 1712 if (ncp->nc_flag & NCF_HASHED) { 1713 ncp->nc_flag &= ~NCF_HASHED; 1714 LIST_REMOVE(ncp, nc_hash); 1715 } 1716 if (ncp->nc_nlen && ncp->nc_parent) { 1717 hash = fnv_32_buf(ncp->nc_name, ncp->nc_nlen, FNV1_32_INIT); 1718 hash = fnv_32_buf(&ncp->nc_parent, 1719 sizeof(ncp->nc_parent), hash); 1720 nchpp = NCHHASH(hash); 1721 LIST_INSERT_HEAD(nchpp, ncp, nc_hash); 1722 ncp->nc_flag |= NCF_HASHED; 1723 } 1724 } 1725 1726 /* 1727 * Name cache initialization, from vfsinit() when we are booting 1728 */ 1729 void 1730 nchinit(void) 1731 { 1732 int i; 1733 globaldata_t gd; 1734 1735 /* initialise per-cpu namecache effectiveness statistics. */ 1736 for (i = 0; i < ncpus; ++i) { 1737 gd = globaldata_find(i); 1738 gd->gd_nchstats = &nchstats[i]; 1739 } 1740 TAILQ_INIT(&ncneglist); 1741 nchashtbl = hashinit(desiredvnodes*2, M_VFSCACHE, &nchash); 1742 nclockwarn = 1 * hz; 1743 } 1744 1745 /* 1746 * Called from start_init() to bootstrap the root filesystem. Returns 1747 * a referenced, unlocked namecache record. 1748 */ 1749 struct namecache * 1750 cache_allocroot(struct mount *mp, struct vnode *vp) 1751 { 1752 struct namecache *ncp = cache_alloc(0); 1753 1754 ncp->nc_flag |= NCF_MOUNTPT | NCF_ROOT; 1755 ncp->nc_mount = mp; 1756 cache_setvp(ncp, vp); 1757 return(ncp); 1758 } 1759 1760 /* 1761 * vfs_cache_setroot() 1762 * 1763 * Create an association between the root of our namecache and 1764 * the root vnode. This routine may be called several times during 1765 * booting. 1766 * 1767 * If the caller intends to save the returned namecache pointer somewhere 1768 * it must cache_hold() it. 1769 */ 1770 void 1771 vfs_cache_setroot(struct vnode *nvp, struct namecache *ncp) 1772 { 1773 struct vnode *ovp; 1774 struct namecache *oncp; 1775 1776 ovp = rootvnode; 1777 oncp = rootncp; 1778 rootvnode = nvp; 1779 rootncp = ncp; 1780 1781 if (ovp) 1782 vrele(ovp); 1783 if (oncp) 1784 cache_drop(oncp); 1785 } 1786 1787 /* 1788 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache 1789 * topology and is being removed as quickly as possible. The new VOP_N*() 1790 * API calls are required to make specific adjustments using the supplied 1791 * ncp pointers rather then just bogusly purging random vnodes. 1792 * 1793 * Invalidate all namecache entries to a particular vnode as well as 1794 * any direct children of that vnode in the namecache. This is a 1795 * 'catch all' purge used by filesystems that do not know any better. 1796 * 1797 * A new vnode v_id is generated. Note that no vnode will ever have a 1798 * v_id of 0. 1799 * 1800 * Note that the linkage between the vnode and its namecache entries will 1801 * be removed, but the namecache entries themselves might stay put due to 1802 * active references from elsewhere in the system or due to the existance of 1803 * the children. The namecache topology is left intact even if we do not 1804 * know what the vnode association is. Such entries will be marked 1805 * NCF_UNRESOLVED. 1806 * 1807 * XXX: Only time and the size of v_id prevents this from failing: 1808 * XXX: In theory we should hunt down all (struct vnode*, v_id) 1809 * XXX: soft references and nuke them, at least on the global 1810 * XXX: v_id wraparound. The period of resistance can be extended 1811 * XXX: by incrementing each vnodes v_id individually instead of 1812 * XXX: using the global v_id. 1813 * 1814 * Does not support NCP_FSMID accumulation on invalidation (retflags is 1815 * not used). 1816 */ 1817 void 1818 cache_purge(struct vnode *vp) 1819 { 1820 static u_long nextid; 1821 int retflags = 0; 1822 1823 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN, &retflags); 1824 1825 /* 1826 * Calculate a new unique id for ".." handling 1827 */ 1828 do { 1829 nextid++; 1830 } while (nextid == vp->v_id || nextid == 0); 1831 vp->v_id = nextid; 1832 } 1833 1834 /* 1835 * Flush all entries referencing a particular filesystem. 1836 * 1837 * Since we need to check it anyway, we will flush all the invalid 1838 * entries at the same time. 1839 */ 1840 void 1841 cache_purgevfs(struct mount *mp) 1842 { 1843 struct nchashhead *nchpp; 1844 struct namecache *ncp, *nnp; 1845 1846 /* 1847 * Scan hash tables for applicable entries. 1848 */ 1849 for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) { 1850 ncp = LIST_FIRST(nchpp); 1851 if (ncp) 1852 cache_hold(ncp); 1853 while (ncp) { 1854 nnp = LIST_NEXT(ncp, nc_hash); 1855 if (nnp) 1856 cache_hold(nnp); 1857 if (ncp->nc_mount == mp) { 1858 cache_lock(ncp); 1859 cache_zap(ncp); 1860 } else { 1861 cache_drop(ncp); 1862 } 1863 ncp = nnp; 1864 } 1865 } 1866 } 1867 1868 static int disablecwd; 1869 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, ""); 1870 1871 static u_long numcwdcalls; STATNODE(CTLFLAG_RD, numcwdcalls, &numcwdcalls); 1872 static u_long numcwdfail1; STATNODE(CTLFLAG_RD, numcwdfail1, &numcwdfail1); 1873 static u_long numcwdfail2; STATNODE(CTLFLAG_RD, numcwdfail2, &numcwdfail2); 1874 static u_long numcwdfail3; STATNODE(CTLFLAG_RD, numcwdfail3, &numcwdfail3); 1875 static u_long numcwdfail4; STATNODE(CTLFLAG_RD, numcwdfail4, &numcwdfail4); 1876 static u_long numcwdfound; STATNODE(CTLFLAG_RD, numcwdfound, &numcwdfound); 1877 1878 int 1879 __getcwd(struct __getcwd_args *uap) 1880 { 1881 int buflen; 1882 int error; 1883 char *buf; 1884 char *bp; 1885 1886 if (disablecwd) 1887 return (ENODEV); 1888 1889 buflen = uap->buflen; 1890 if (buflen < 2) 1891 return (EINVAL); 1892 if (buflen > MAXPATHLEN) 1893 buflen = MAXPATHLEN; 1894 1895 buf = malloc(buflen, M_TEMP, M_WAITOK); 1896 bp = kern_getcwd(buf, buflen, &error); 1897 if (error == 0) 1898 error = copyout(bp, uap->buf, strlen(bp) + 1); 1899 free(buf, M_TEMP); 1900 return (error); 1901 } 1902 1903 char * 1904 kern_getcwd(char *buf, size_t buflen, int *error) 1905 { 1906 struct proc *p = curproc; 1907 char *bp; 1908 int i, slash_prefixed; 1909 struct filedesc *fdp; 1910 struct namecache *ncp; 1911 1912 numcwdcalls++; 1913 bp = buf; 1914 bp += buflen - 1; 1915 *bp = '\0'; 1916 fdp = p->p_fd; 1917 slash_prefixed = 0; 1918 1919 ncp = fdp->fd_ncdir; 1920 while (ncp && ncp != fdp->fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) { 1921 if (ncp->nc_flag & NCF_MOUNTPT) { 1922 if (ncp->nc_mount == NULL) { 1923 *error = EBADF; /* forced unmount? */ 1924 return(NULL); 1925 } 1926 ncp = ncp->nc_parent; 1927 continue; 1928 } 1929 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 1930 if (bp == buf) { 1931 numcwdfail4++; 1932 *error = ENOMEM; 1933 return(NULL); 1934 } 1935 *--bp = ncp->nc_name[i]; 1936 } 1937 if (bp == buf) { 1938 numcwdfail4++; 1939 *error = ENOMEM; 1940 return(NULL); 1941 } 1942 *--bp = '/'; 1943 slash_prefixed = 1; 1944 ncp = ncp->nc_parent; 1945 } 1946 if (ncp == NULL) { 1947 numcwdfail2++; 1948 *error = ENOENT; 1949 return(NULL); 1950 } 1951 if (!slash_prefixed) { 1952 if (bp == buf) { 1953 numcwdfail4++; 1954 *error = ENOMEM; 1955 return(NULL); 1956 } 1957 *--bp = '/'; 1958 } 1959 numcwdfound++; 1960 *error = 0; 1961 return (bp); 1962 } 1963 1964 /* 1965 * Thus begins the fullpath magic. 1966 */ 1967 1968 #undef STATNODE 1969 #define STATNODE(name) \ 1970 static u_int name; \ 1971 SYSCTL_UINT(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, "") 1972 1973 static int disablefullpath; 1974 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, 1975 &disablefullpath, 0, ""); 1976 1977 STATNODE(numfullpathcalls); 1978 STATNODE(numfullpathfail1); 1979 STATNODE(numfullpathfail2); 1980 STATNODE(numfullpathfail3); 1981 STATNODE(numfullpathfail4); 1982 STATNODE(numfullpathfound); 1983 1984 int 1985 cache_fullpath(struct proc *p, struct namecache *ncp, char **retbuf, char **freebuf) 1986 { 1987 char *bp, *buf; 1988 int i, slash_prefixed; 1989 struct namecache *fd_nrdir; 1990 1991 numfullpathcalls--; 1992 1993 buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK); 1994 bp = buf + MAXPATHLEN - 1; 1995 *bp = '\0'; 1996 if (p != NULL) 1997 fd_nrdir = p->p_fd->fd_nrdir; 1998 else 1999 fd_nrdir = NULL; 2000 slash_prefixed = 0; 2001 while (ncp && ncp != fd_nrdir && (ncp->nc_flag & NCF_ROOT) == 0) { 2002 if (ncp->nc_flag & NCF_MOUNTPT) { 2003 if (ncp->nc_mount == NULL) { 2004 free(buf, M_TEMP); 2005 return(EBADF); 2006 } 2007 ncp = ncp->nc_parent; 2008 continue; 2009 } 2010 for (i = ncp->nc_nlen - 1; i >= 0; i--) { 2011 if (bp == buf) { 2012 numfullpathfail4++; 2013 free(buf, M_TEMP); 2014 return(ENOMEM); 2015 } 2016 *--bp = ncp->nc_name[i]; 2017 } 2018 if (bp == buf) { 2019 numfullpathfail4++; 2020 free(buf, M_TEMP); 2021 return(ENOMEM); 2022 } 2023 *--bp = '/'; 2024 slash_prefixed = 1; 2025 ncp = ncp->nc_parent; 2026 } 2027 if (ncp == NULL) { 2028 numfullpathfail2++; 2029 free(buf, M_TEMP); 2030 return(ENOENT); 2031 } 2032 if (p != NULL && (ncp->nc_flag & NCF_ROOT) && ncp != fd_nrdir) { 2033 bp = buf + MAXPATHLEN - 1; 2034 *bp = '\0'; 2035 slash_prefixed = 0; 2036 } 2037 if (!slash_prefixed) { 2038 if (bp == buf) { 2039 numfullpathfail4++; 2040 free(buf, M_TEMP); 2041 return(ENOMEM); 2042 } 2043 *--bp = '/'; 2044 } 2045 numfullpathfound++; 2046 *retbuf = bp; 2047 *freebuf = buf; 2048 2049 return(0); 2050 } 2051 2052 int 2053 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf) 2054 { 2055 struct namecache *ncp; 2056 2057 numfullpathcalls++; 2058 if (disablefullpath) 2059 return (ENODEV); 2060 2061 if (p == NULL) 2062 return (EINVAL); 2063 2064 /* vn is NULL, client wants us to use p->p_textvp */ 2065 if (vn == NULL) { 2066 if ((vn = p->p_textvp) == NULL) 2067 return (EINVAL); 2068 } 2069 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) { 2070 if (ncp->nc_nlen) 2071 break; 2072 } 2073 if (ncp == NULL) 2074 return (EINVAL); 2075 2076 numfullpathcalls--; 2077 return(cache_fullpath(p, ncp, retbuf, freebuf)); 2078 } 2079