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 1.85 2007/11/02 19:52:25 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 #include <sys/sysref2.h> 92 93 #define MAX_RECURSION_DEPTH 64 94 95 /* 96 * Random lookups in the cache are accomplished with a hash table using 97 * a hash key of (nc_src_vp, name). 98 * 99 * Negative entries may exist and correspond to structures where nc_vp 100 * is NULL. In a negative entry, NCF_WHITEOUT will be set if the entry 101 * corresponds to a whited-out directory entry (verses simply not finding the 102 * entry at all). 103 * 104 * Upon reaching the last segment of a path, if the reference is for DELETE, 105 * or NOCACHE is set (rewrite), and the name is located in the cache, it 106 * will be dropped. 107 */ 108 109 /* 110 * Structures associated with name cacheing. 111 */ 112 #define NCHHASH(hash) (&nchashtbl[(hash) & nchash]) 113 #define MINNEG 1024 114 115 MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries"); 116 117 static LIST_HEAD(nchashhead, namecache) *nchashtbl; /* Hash Table */ 118 static struct namecache_list ncneglist; /* instead of vnode */ 119 120 /* 121 * ncvp_debug - debug cache_fromvp(). This is used by the NFS server 122 * to create the namecache infrastructure leading to a dangling vnode. 123 * 124 * 0 Only errors are reported 125 * 1 Successes are reported 126 * 2 Successes + the whole directory scan is reported 127 * 3 Force the directory scan code run as if the parent vnode did not 128 * have a namecache record, even if it does have one. 129 */ 130 static int ncvp_debug; 131 SYSCTL_INT(_debug, OID_AUTO, ncvp_debug, CTLFLAG_RW, &ncvp_debug, 0, ""); 132 133 static u_long nchash; /* size of hash table */ 134 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, ""); 135 136 static u_long ncnegfactor = 16; /* ratio of negative entries */ 137 SYSCTL_ULONG(_debug, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0, ""); 138 139 static int nclockwarn; /* warn on locked entries in ticks */ 140 SYSCTL_INT(_debug, OID_AUTO, nclockwarn, CTLFLAG_RW, &nclockwarn, 0, ""); 141 142 static u_long numneg; /* number of cache entries allocated */ 143 SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0, ""); 144 145 static u_long numcache; /* number of cache entries allocated */ 146 SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0, ""); 147 148 static u_long numunres; /* number of unresolved entries */ 149 SYSCTL_ULONG(_debug, OID_AUTO, numunres, CTLFLAG_RD, &numunres, 0, ""); 150 151 SYSCTL_INT(_debug, OID_AUTO, vnsize, CTLFLAG_RD, 0, sizeof(struct vnode), ""); 152 SYSCTL_INT(_debug, OID_AUTO, ncsize, CTLFLAG_RD, 0, sizeof(struct namecache), ""); 153 154 static int cache_resolve_mp(struct mount *mp); 155 static void _cache_rehash(struct namecache *ncp); 156 static void _cache_lock(struct namecache *ncp); 157 static void _cache_setunresolved(struct namecache *ncp); 158 159 /* 160 * The new name cache statistics 161 */ 162 SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics"); 163 #define STATNODE(mode, name, var) \ 164 SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, mode, var, 0, ""); 165 STATNODE(CTLFLAG_RD, numneg, &numneg); 166 STATNODE(CTLFLAG_RD, numcache, &numcache); 167 static u_long numcalls; STATNODE(CTLFLAG_RD, numcalls, &numcalls); 168 static u_long dothits; STATNODE(CTLFLAG_RD, dothits, &dothits); 169 static u_long dotdothits; STATNODE(CTLFLAG_RD, dotdothits, &dotdothits); 170 static u_long numchecks; STATNODE(CTLFLAG_RD, numchecks, &numchecks); 171 static u_long nummiss; STATNODE(CTLFLAG_RD, nummiss, &nummiss); 172 static u_long nummisszap; STATNODE(CTLFLAG_RD, nummisszap, &nummisszap); 173 static u_long numposzaps; STATNODE(CTLFLAG_RD, numposzaps, &numposzaps); 174 static u_long numposhits; STATNODE(CTLFLAG_RD, numposhits, &numposhits); 175 static u_long numnegzaps; STATNODE(CTLFLAG_RD, numnegzaps, &numnegzaps); 176 static u_long numneghits; STATNODE(CTLFLAG_RD, numneghits, &numneghits); 177 178 struct nchstats nchstats[SMP_MAXCPU]; 179 /* 180 * Export VFS cache effectiveness statistics to user-land. 181 * 182 * The statistics are left for aggregation to user-land so 183 * neat things can be achieved, like observing per-CPU cache 184 * distribution. 185 */ 186 static int 187 sysctl_nchstats(SYSCTL_HANDLER_ARGS) 188 { 189 struct globaldata *gd; 190 int i, error; 191 192 error = 0; 193 for (i = 0; i < ncpus; ++i) { 194 gd = globaldata_find(i); 195 if ((error = SYSCTL_OUT(req, (void *)&(*gd->gd_nchstats), 196 sizeof(struct nchstats)))) 197 break; 198 } 199 200 return (error); 201 } 202 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE|CTLFLAG_RD, 203 0, 0, sysctl_nchstats, "S,nchstats", "VFS cache effectiveness statistics"); 204 205 static void cache_zap(struct namecache *ncp); 206 207 /* 208 * cache_hold() and cache_drop() prevent the premature deletion of a 209 * namecache entry but do not prevent operations (such as zapping) on 210 * that namecache entry. 211 * 212 * This routine may only be called from outside this source module if 213 * nc_refs is already at least 1. 214 * 215 * This is a rare case where callers are allowed to hold a spinlock, 216 * so we can't ourselves. 217 */ 218 static __inline 219 struct namecache * 220 _cache_hold(struct namecache *ncp) 221 { 222 atomic_add_int(&ncp->nc_refs, 1); 223 return(ncp); 224 } 225 226 /* 227 * When dropping an entry, if only one ref remains and the entry has not 228 * been resolved, zap it. Since the one reference is being dropped the 229 * entry had better not be locked. 230 */ 231 static __inline 232 void 233 _cache_drop(struct namecache *ncp) 234 { 235 KKASSERT(ncp->nc_refs > 0); 236 if (ncp->nc_refs == 1 && 237 (ncp->nc_flag & NCF_UNRESOLVED) && 238 TAILQ_EMPTY(&ncp->nc_list) 239 ) { 240 KKASSERT(ncp->nc_exlocks == 0); 241 _cache_lock(ncp); 242 cache_zap(ncp); 243 } else { 244 atomic_subtract_int(&ncp->nc_refs, 1); 245 } 246 } 247 248 /* 249 * Link a new namecache entry to its parent. Be careful to avoid races 250 * if vhold() blocks in the future. 251 */ 252 static void 253 cache_link_parent(struct namecache *ncp, struct namecache *par) 254 { 255 KKASSERT(ncp->nc_parent == NULL); 256 ncp->nc_parent = par; 257 if (TAILQ_EMPTY(&par->nc_list)) { 258 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry); 259 /* 260 * Any vp associated with an ncp which has children must 261 * be held to prevent it from being recycled. 262 */ 263 if (par->nc_vp) 264 vhold(par->nc_vp); 265 } else { 266 TAILQ_INSERT_HEAD(&par->nc_list, ncp, nc_entry); 267 } 268 } 269 270 /* 271 * Remove the parent association from a namecache structure. If this is 272 * the last child of the parent the cache_drop(par) will attempt to 273 * recursively zap the parent. 274 */ 275 static void 276 cache_unlink_parent(struct namecache *ncp) 277 { 278 struct namecache *par; 279 280 if ((par = ncp->nc_parent) != NULL) { 281 ncp->nc_parent = NULL; 282 par = _cache_hold(par); 283 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 284 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 285 vdrop(par->nc_vp); 286 _cache_drop(par); 287 } 288 } 289 290 /* 291 * Allocate a new namecache structure. Most of the code does not require 292 * zero-termination of the string but it makes vop_compat_ncreate() easier. 293 */ 294 static struct namecache * 295 cache_alloc(int nlen) 296 { 297 struct namecache *ncp; 298 299 ncp = kmalloc(sizeof(*ncp), M_VFSCACHE, M_WAITOK|M_ZERO); 300 if (nlen) 301 ncp->nc_name = kmalloc(nlen + 1, M_VFSCACHE, M_WAITOK); 302 ncp->nc_nlen = nlen; 303 ncp->nc_flag = NCF_UNRESOLVED; 304 ncp->nc_error = ENOTCONN; /* needs to be resolved */ 305 ncp->nc_refs = 1; 306 307 /* 308 * Construct a fake FSMID based on the time of day and a 32 bit 309 * roller for uniqueness. This is used to generate a useful 310 * FSMID for filesystems which do not support it. 311 */ 312 ncp->nc_fsmid = cache_getnewfsmid(); 313 TAILQ_INIT(&ncp->nc_list); 314 _cache_lock(ncp); 315 return(ncp); 316 } 317 318 static void 319 _cache_free(struct namecache *ncp) 320 { 321 KKASSERT(ncp->nc_refs == 1 && ncp->nc_exlocks == 1); 322 if (ncp->nc_name) 323 kfree(ncp->nc_name, M_VFSCACHE); 324 kfree(ncp, M_VFSCACHE); 325 } 326 327 void 328 cache_zero(struct nchandle *nch) 329 { 330 nch->ncp = NULL; 331 nch->mount = NULL; 332 } 333 334 /* 335 * Ref and deref a namecache structure. 336 * 337 * Warning: caller may hold an unrelated read spinlock, which means we can't 338 * use read spinlocks here. 339 */ 340 struct nchandle * 341 cache_hold(struct nchandle *nch) 342 { 343 _cache_hold(nch->ncp); 344 ++nch->mount->mnt_refs; 345 return(nch); 346 } 347 348 void 349 cache_copy(struct nchandle *nch, struct nchandle *target) 350 { 351 *target = *nch; 352 _cache_hold(target->ncp); 353 ++nch->mount->mnt_refs; 354 } 355 356 void 357 cache_changemount(struct nchandle *nch, struct mount *mp) 358 { 359 --nch->mount->mnt_refs; 360 nch->mount = mp; 361 ++nch->mount->mnt_refs; 362 } 363 364 void 365 cache_drop(struct nchandle *nch) 366 { 367 --nch->mount->mnt_refs; 368 _cache_drop(nch->ncp); 369 nch->ncp = NULL; 370 nch->mount = NULL; 371 } 372 373 /* 374 * Namespace locking. The caller must already hold a reference to the 375 * namecache structure in order to lock/unlock it. This function prevents 376 * the namespace from being created or destroyed by accessors other then 377 * the lock holder. 378 * 379 * Note that holding a locked namecache structure prevents other threads 380 * from making namespace changes (e.g. deleting or creating), prevents 381 * vnode association state changes by other threads, and prevents the 382 * namecache entry from being resolved or unresolved by other threads. 383 * 384 * The lock owner has full authority to associate/disassociate vnodes 385 * and resolve/unresolve the locked ncp. 386 * 387 * WARNING! Holding a locked ncp will prevent a vnode from being destroyed 388 * or recycled, but it does NOT help you if the vnode had already initiated 389 * a recyclement. If this is important, use cache_get() rather then 390 * cache_lock() (and deal with the differences in the way the refs counter 391 * is handled). Or, alternatively, make an unconditional call to 392 * cache_validate() or cache_resolve() after cache_lock() returns. 393 */ 394 static 395 void 396 _cache_lock(struct namecache *ncp) 397 { 398 thread_t td; 399 int didwarn; 400 401 KKASSERT(ncp->nc_refs != 0); 402 didwarn = 0; 403 td = curthread; 404 405 for (;;) { 406 if (ncp->nc_exlocks == 0) { 407 ncp->nc_exlocks = 1; 408 ncp->nc_locktd = td; 409 /* 410 * The vp associated with a locked ncp must be held 411 * to prevent it from being recycled (which would 412 * cause the ncp to become unresolved). 413 * 414 * WARNING! If VRECLAIMED is set the vnode could 415 * already be in the middle of a recycle. Callers 416 * should not assume that nc_vp is usable when 417 * not NULL. cache_vref() or cache_vget() must be 418 * called. 419 * 420 * XXX loop on race for later MPSAFE work. 421 */ 422 if (ncp->nc_vp) 423 vhold(ncp->nc_vp); 424 break; 425 } 426 if (ncp->nc_locktd == td) { 427 ++ncp->nc_exlocks; 428 break; 429 } 430 ncp->nc_flag |= NCF_LOCKREQ; 431 if (tsleep(ncp, 0, "clock", nclockwarn) == EWOULDBLOCK) { 432 if (didwarn) 433 continue; 434 didwarn = 1; 435 kprintf("[diagnostic] cache_lock: blocked on %p", ncp); 436 kprintf(" \"%*.*s\"\n", 437 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 438 } 439 } 440 441 if (didwarn == 1) { 442 kprintf("[diagnostic] cache_lock: unblocked %*.*s\n", 443 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 444 } 445 } 446 447 void 448 cache_lock(struct nchandle *nch) 449 { 450 _cache_lock(nch->ncp); 451 } 452 453 static 454 int 455 _cache_lock_nonblock(struct namecache *ncp) 456 { 457 thread_t td; 458 459 KKASSERT(ncp->nc_refs != 0); 460 td = curthread; 461 if (ncp->nc_exlocks == 0) { 462 ncp->nc_exlocks = 1; 463 ncp->nc_locktd = td; 464 /* 465 * The vp associated with a locked ncp must be held 466 * to prevent it from being recycled (which would 467 * cause the ncp to become unresolved). 468 * 469 * WARNING! If VRECLAIMED is set the vnode could 470 * already be in the middle of a recycle. Callers 471 * should not assume that nc_vp is usable when 472 * not NULL. cache_vref() or cache_vget() must be 473 * called. 474 * 475 * XXX loop on race for later MPSAFE work. 476 */ 477 if (ncp->nc_vp) 478 vhold(ncp->nc_vp); 479 return(0); 480 } else { 481 return(EWOULDBLOCK); 482 } 483 } 484 485 int 486 cache_lock_nonblock(struct nchandle *nch) 487 { 488 return(_cache_lock_nonblock(nch->ncp)); 489 } 490 491 static 492 void 493 _cache_unlock(struct namecache *ncp) 494 { 495 thread_t td = curthread; 496 497 KKASSERT(ncp->nc_refs > 0); 498 KKASSERT(ncp->nc_exlocks > 0); 499 KKASSERT(ncp->nc_locktd == td); 500 if (--ncp->nc_exlocks == 0) { 501 if (ncp->nc_vp) 502 vdrop(ncp->nc_vp); 503 ncp->nc_locktd = NULL; 504 if (ncp->nc_flag & NCF_LOCKREQ) { 505 ncp->nc_flag &= ~NCF_LOCKREQ; 506 wakeup(ncp); 507 } 508 } 509 } 510 511 void 512 cache_unlock(struct nchandle *nch) 513 { 514 _cache_unlock(nch->ncp); 515 } 516 517 /* 518 * ref-and-lock, unlock-and-deref functions. 519 * 520 * This function is primarily used by nlookup. Even though cache_lock 521 * holds the vnode, it is possible that the vnode may have already 522 * initiated a recyclement. We want cache_get() to return a definitively 523 * usable vnode or a definitively unresolved ncp. 524 */ 525 static 526 struct namecache * 527 _cache_get(struct namecache *ncp) 528 { 529 _cache_hold(ncp); 530 _cache_lock(ncp); 531 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 532 _cache_setunresolved(ncp); 533 return(ncp); 534 } 535 536 /* 537 * note: the same nchandle can be passed for both arguments. 538 */ 539 void 540 cache_get(struct nchandle *nch, struct nchandle *target) 541 { 542 target->mount = nch->mount; 543 target->ncp = _cache_get(nch->ncp); 544 ++target->mount->mnt_refs; 545 } 546 547 static int 548 _cache_get_nonblock(struct namecache *ncp) 549 { 550 /* XXX MP */ 551 if (ncp->nc_exlocks == 0 || ncp->nc_locktd == curthread) { 552 _cache_hold(ncp); 553 _cache_lock(ncp); 554 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 555 _cache_setunresolved(ncp); 556 return(0); 557 } 558 return(EWOULDBLOCK); 559 } 560 561 int 562 cache_get_nonblock(struct nchandle *nch) 563 { 564 return(_cache_get_nonblock(nch->ncp)); 565 } 566 567 static __inline 568 void 569 _cache_put(struct namecache *ncp) 570 { 571 _cache_unlock(ncp); 572 _cache_drop(ncp); 573 } 574 575 void 576 cache_put(struct nchandle *nch) 577 { 578 --nch->mount->mnt_refs; 579 _cache_put(nch->ncp); 580 nch->ncp = NULL; 581 nch->mount = NULL; 582 } 583 584 /* 585 * Resolve an unresolved ncp by associating a vnode with it. If the 586 * vnode is NULL, a negative cache entry is created. 587 * 588 * The ncp should be locked on entry and will remain locked on return. 589 */ 590 static 591 void 592 _cache_setvp(struct namecache *ncp, struct vnode *vp) 593 { 594 KKASSERT(ncp->nc_flag & NCF_UNRESOLVED); 595 ncp->nc_vp = vp; 596 if (vp != NULL) { 597 /* 598 * Any vp associated with an ncp which has children must 599 * be held. Any vp associated with a locked ncp must be held. 600 */ 601 if (!TAILQ_EMPTY(&ncp->nc_list)) 602 vhold(vp); 603 TAILQ_INSERT_HEAD(&vp->v_namecache, ncp, nc_vnode); 604 if (ncp->nc_exlocks) 605 vhold(vp); 606 607 /* 608 * Set auxiliary flags 609 */ 610 switch(vp->v_type) { 611 case VDIR: 612 ncp->nc_flag |= NCF_ISDIR; 613 break; 614 case VLNK: 615 ncp->nc_flag |= NCF_ISSYMLINK; 616 /* XXX cache the contents of the symlink */ 617 break; 618 default: 619 break; 620 } 621 ++numcache; 622 ncp->nc_error = 0; 623 } else { 624 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 625 ++numneg; 626 ncp->nc_error = ENOENT; 627 } 628 ncp->nc_flag &= ~NCF_UNRESOLVED; 629 } 630 631 void 632 cache_setvp(struct nchandle *nch, struct vnode *vp) 633 { 634 _cache_setvp(nch->ncp, vp); 635 } 636 637 void 638 cache_settimeout(struct nchandle *nch, int nticks) 639 { 640 struct namecache *ncp = nch->ncp; 641 642 if ((ncp->nc_timeout = ticks + nticks) == 0) 643 ncp->nc_timeout = 1; 644 } 645 646 /* 647 * Disassociate the vnode or negative-cache association and mark a 648 * namecache entry as unresolved again. Note that the ncp is still 649 * left in the hash table and still linked to its parent. 650 * 651 * The ncp should be locked and refd on entry and will remain locked and refd 652 * on return. 653 * 654 * This routine is normally never called on a directory containing children. 655 * However, NFS often does just that in its rename() code as a cop-out to 656 * avoid complex namespace operations. This disconnects a directory vnode 657 * from its namecache and can cause the OLDAPI and NEWAPI to get out of 658 * sync. 659 * 660 * NOTE: NCF_FSMID must be cleared so a refurbishment of the ncp, such as 661 * in a create, properly propogates flag up the chain. 662 */ 663 static 664 void 665 _cache_setunresolved(struct namecache *ncp) 666 { 667 struct vnode *vp; 668 669 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 670 ncp->nc_flag |= NCF_UNRESOLVED; 671 ncp->nc_timeout = 0; 672 ncp->nc_error = ENOTCONN; 673 ++numunres; 674 if ((vp = ncp->nc_vp) != NULL) { 675 --numcache; 676 ncp->nc_vp = NULL; 677 TAILQ_REMOVE(&vp->v_namecache, ncp, nc_vnode); 678 679 /* 680 * Any vp associated with an ncp with children is 681 * held by that ncp. Any vp associated with a locked 682 * ncp is held by that ncp. These conditions must be 683 * undone when the vp is cleared out from the ncp. 684 */ 685 if (ncp->nc_flag & NCF_FSMID) 686 vupdatefsmid(vp); 687 if (!TAILQ_EMPTY(&ncp->nc_list)) 688 vdrop(vp); 689 if (ncp->nc_exlocks) 690 vdrop(vp); 691 } else { 692 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 693 --numneg; 694 } 695 ncp->nc_flag &= ~(NCF_WHITEOUT|NCF_ISDIR|NCF_ISSYMLINK| 696 NCF_FSMID); 697 } 698 } 699 700 void 701 cache_setunresolved(struct nchandle *nch) 702 { 703 _cache_setunresolved(nch->ncp); 704 } 705 706 /* 707 * Determine if we can clear NCF_ISMOUNTPT by scanning the mountlist 708 * looking for matches. This flag tells the lookup code when it must 709 * check for a mount linkage and also prevents the directories in question 710 * from being deleted or renamed. 711 */ 712 static 713 int 714 cache_clrmountpt_callback(struct mount *mp, void *data) 715 { 716 struct nchandle *nch = data; 717 718 if (mp->mnt_ncmounton.ncp == nch->ncp) 719 return(1); 720 if (mp->mnt_ncmountpt.ncp == nch->ncp) 721 return(1); 722 return(0); 723 } 724 725 void 726 cache_clrmountpt(struct nchandle *nch) 727 { 728 int count; 729 730 count = mountlist_scan(cache_clrmountpt_callback, nch, 731 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 732 if (count == 0) 733 nch->ncp->nc_flag &= ~NCF_ISMOUNTPT; 734 } 735 736 /* 737 * Invalidate portions of the namecache topology given a starting entry. 738 * The passed ncp is set to an unresolved state and: 739 * 740 * The passed ncp must be locked. 741 * 742 * CINV_DESTROY - Set a flag in the passed ncp entry indicating 743 * that the physical underlying nodes have been 744 * destroyed... as in deleted. For example, when 745 * a directory is removed. This will cause record 746 * lookups on the name to no longer be able to find 747 * the record and tells the resolver to return failure 748 * rather then trying to resolve through the parent. 749 * 750 * The topology itself, including ncp->nc_name, 751 * remains intact. 752 * 753 * This only applies to the passed ncp, if CINV_CHILDREN 754 * is specified the children are not flagged. 755 * 756 * CINV_CHILDREN - Set all children (recursively) to an unresolved 757 * state as well. 758 * 759 * Note that this will also have the side effect of 760 * cleaning out any unreferenced nodes in the topology 761 * from the leaves up as the recursion backs out. 762 * 763 * Note that the topology for any referenced nodes remains intact. 764 * 765 * It is possible for cache_inval() to race a cache_resolve(), meaning that 766 * the namecache entry may not actually be invalidated on return if it was 767 * revalidated while recursing down into its children. This code guarentees 768 * that the node(s) will go through an invalidation cycle, but does not 769 * guarentee that they will remain in an invalidated state. 770 * 771 * Returns non-zero if a revalidation was detected during the invalidation 772 * recursion, zero otherwise. Note that since only the original ncp is 773 * locked the revalidation ultimately can only indicate that the original ncp 774 * *MIGHT* no have been reresolved. 775 * 776 * DEEP RECURSION HANDLING - If a recursive invalidation recurses deeply we 777 * have to avoid blowing out the kernel stack. We do this by saving the 778 * deep namecache node and aborting the recursion, then re-recursing at that 779 * node using a depth-first algorithm in order to allow multiple deep 780 * recursions to chain through each other, then we restart the invalidation 781 * from scratch. 782 */ 783 784 struct cinvtrack { 785 struct namecache *resume_ncp; 786 int depth; 787 }; 788 789 static int _cache_inval_internal(struct namecache *, int, struct cinvtrack *); 790 791 static 792 int 793 _cache_inval(struct namecache *ncp, int flags) 794 { 795 struct cinvtrack track; 796 struct namecache *ncp2; 797 int r; 798 799 track.depth = 0; 800 track.resume_ncp = NULL; 801 802 for (;;) { 803 r = _cache_inval_internal(ncp, flags, &track); 804 if (track.resume_ncp == NULL) 805 break; 806 kprintf("Warning: deep namecache recursion at %s\n", 807 ncp->nc_name); 808 _cache_unlock(ncp); 809 while ((ncp2 = track.resume_ncp) != NULL) { 810 track.resume_ncp = NULL; 811 _cache_lock(ncp2); 812 _cache_inval_internal(ncp2, flags & ~CINV_DESTROY, 813 &track); 814 _cache_put(ncp2); 815 } 816 _cache_lock(ncp); 817 } 818 return(r); 819 } 820 821 int 822 cache_inval(struct nchandle *nch, int flags) 823 { 824 return(_cache_inval(nch->ncp, flags)); 825 } 826 827 static int 828 _cache_inval_internal(struct namecache *ncp, int flags, struct cinvtrack *track) 829 { 830 struct namecache *kid; 831 struct namecache *nextkid; 832 int rcnt = 0; 833 834 KKASSERT(ncp->nc_exlocks); 835 836 _cache_setunresolved(ncp); 837 if (flags & CINV_DESTROY) 838 ncp->nc_flag |= NCF_DESTROYED; 839 840 if ((flags & CINV_CHILDREN) && 841 (kid = TAILQ_FIRST(&ncp->nc_list)) != NULL 842 ) { 843 if (++track->depth > MAX_RECURSION_DEPTH) { 844 track->resume_ncp = ncp; 845 _cache_hold(ncp); 846 ++rcnt; 847 } 848 _cache_hold(kid); 849 _cache_unlock(ncp); 850 while (kid) { 851 if (track->resume_ncp) { 852 _cache_drop(kid); 853 break; 854 } 855 if ((nextkid = TAILQ_NEXT(kid, nc_entry)) != NULL) 856 _cache_hold(nextkid); 857 if ((kid->nc_flag & NCF_UNRESOLVED) == 0 || 858 TAILQ_FIRST(&kid->nc_list) 859 ) { 860 _cache_lock(kid); 861 rcnt += _cache_inval_internal(kid, flags & ~CINV_DESTROY, track); 862 _cache_unlock(kid); 863 } 864 _cache_drop(kid); 865 kid = nextkid; 866 } 867 --track->depth; 868 _cache_lock(ncp); 869 } 870 871 /* 872 * Someone could have gotten in there while ncp was unlocked, 873 * retry if so. 874 */ 875 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 876 ++rcnt; 877 return (rcnt); 878 } 879 880 /* 881 * Invalidate a vnode's namecache associations. To avoid races against 882 * the resolver we do not invalidate a node which we previously invalidated 883 * but which was then re-resolved while we were in the invalidation loop. 884 * 885 * Returns non-zero if any namecache entries remain after the invalidation 886 * loop completed. 887 * 888 * NOTE: unlike the namecache topology which guarentees that ncp's will not 889 * be ripped out of the topology while held, the vnode's v_namecache list 890 * has no such restriction. NCP's can be ripped out of the list at virtually 891 * any time if not locked, even if held. 892 */ 893 int 894 cache_inval_vp(struct vnode *vp, int flags) 895 { 896 struct namecache *ncp; 897 struct namecache *next; 898 899 restart: 900 ncp = TAILQ_FIRST(&vp->v_namecache); 901 if (ncp) 902 _cache_hold(ncp); 903 while (ncp) { 904 /* loop entered with ncp held */ 905 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL) 906 _cache_hold(next); 907 _cache_lock(ncp); 908 if (ncp->nc_vp != vp) { 909 kprintf("Warning: cache_inval_vp: race-A detected on " 910 "%s\n", ncp->nc_name); 911 _cache_put(ncp); 912 if (next) 913 _cache_drop(next); 914 goto restart; 915 } 916 _cache_inval(ncp, flags); 917 _cache_put(ncp); /* also releases reference */ 918 ncp = next; 919 if (ncp && ncp->nc_vp != vp) { 920 kprintf("Warning: cache_inval_vp: race-B detected on " 921 "%s\n", ncp->nc_name); 922 _cache_drop(ncp); 923 goto restart; 924 } 925 } 926 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 927 } 928 929 /* 930 * This routine is used instead of the normal cache_inval_vp() when we 931 * are trying to recycle otherwise good vnodes. 932 * 933 * Return 0 on success, non-zero if not all namecache records could be 934 * disassociated from the vnode (for various reasons). 935 */ 936 int 937 cache_inval_vp_nonblock(struct vnode *vp) 938 { 939 struct namecache *ncp; 940 struct namecache *next; 941 942 ncp = TAILQ_FIRST(&vp->v_namecache); 943 if (ncp) 944 _cache_hold(ncp); 945 while (ncp) { 946 /* loop entered with ncp held */ 947 if ((next = TAILQ_NEXT(ncp, nc_vnode)) != NULL) 948 _cache_hold(next); 949 if (_cache_lock_nonblock(ncp)) { 950 _cache_drop(ncp); 951 if (next) 952 _cache_drop(next); 953 break; 954 } 955 if (ncp->nc_vp != vp) { 956 kprintf("Warning: cache_inval_vp: race-A detected on " 957 "%s\n", ncp->nc_name); 958 _cache_put(ncp); 959 if (next) 960 _cache_drop(next); 961 break; 962 } 963 _cache_inval(ncp, 0); 964 _cache_put(ncp); /* also releases reference */ 965 ncp = next; 966 if (ncp && ncp->nc_vp != vp) { 967 kprintf("Warning: cache_inval_vp: race-B detected on " 968 "%s\n", ncp->nc_name); 969 _cache_drop(ncp); 970 break; 971 } 972 } 973 return(TAILQ_FIRST(&vp->v_namecache) != NULL); 974 } 975 976 /* 977 * The source ncp has been renamed to the target ncp. Both fncp and tncp 978 * must be locked. Both will be set to unresolved, any children of tncp 979 * will be disconnected (the prior contents of the target is assumed to be 980 * destroyed by the rename operation, e.g. renaming over an empty directory), 981 * and all children of fncp will be moved to tncp. 982 * 983 * XXX the disconnection could pose a problem, check code paths to make 984 * sure any code that blocks can handle the parent being changed out from 985 * under it. Maybe we should lock the children (watch out for deadlocks) ? 986 * 987 * After we return the caller has the option of calling cache_setvp() if 988 * the vnode of the new target ncp is known. 989 * 990 * Any process CD'd into any of the children will no longer be able to ".." 991 * back out. An rm -rf can cause this situation to occur. 992 */ 993 void 994 cache_rename(struct nchandle *fnch, struct nchandle *tnch) 995 { 996 struct namecache *fncp = fnch->ncp; 997 struct namecache *tncp = tnch->ncp; 998 struct namecache *scan; 999 int didwarn = 0; 1000 1001 _cache_setunresolved(fncp); 1002 _cache_setunresolved(tncp); 1003 while (_cache_inval(tncp, CINV_CHILDREN) != 0) { 1004 if (didwarn++ % 10 == 0) { 1005 kprintf("Warning: cache_rename: race during " 1006 "rename %s->%s\n", 1007 fncp->nc_name, tncp->nc_name); 1008 } 1009 tsleep(tncp, 0, "mvrace", hz / 10); 1010 _cache_setunresolved(tncp); 1011 } 1012 while ((scan = TAILQ_FIRST(&fncp->nc_list)) != NULL) { 1013 _cache_hold(scan); 1014 cache_unlink_parent(scan); 1015 cache_link_parent(scan, tncp); 1016 if (scan->nc_flag & NCF_HASHED) 1017 _cache_rehash(scan); 1018 _cache_drop(scan); 1019 } 1020 } 1021 1022 /* 1023 * vget the vnode associated with the namecache entry. Resolve the namecache 1024 * entry if necessary and deal with namecache/vp races. The passed ncp must 1025 * be referenced and may be locked. The ncp's ref/locking state is not 1026 * effected by this call. 1027 * 1028 * lk_type may be LK_SHARED, LK_EXCLUSIVE. A ref'd, possibly locked 1029 * (depending on the passed lk_type) will be returned in *vpp with an error 1030 * of 0, or NULL will be returned in *vpp with a non-0 error code. The 1031 * most typical error is ENOENT, meaning that the ncp represents a negative 1032 * cache hit and there is no vnode to retrieve, but other errors can occur 1033 * too. 1034 * 1035 * The main race we have to deal with are namecache zaps. The ncp itself 1036 * will not disappear since it is referenced, and it turns out that the 1037 * validity of the vp pointer can be checked simply by rechecking the 1038 * contents of ncp->nc_vp. 1039 */ 1040 int 1041 cache_vget(struct nchandle *nch, struct ucred *cred, 1042 int lk_type, struct vnode **vpp) 1043 { 1044 struct namecache *ncp; 1045 struct vnode *vp; 1046 int error; 1047 1048 ncp = nch->ncp; 1049 again: 1050 vp = NULL; 1051 if (ncp->nc_flag & NCF_UNRESOLVED) { 1052 _cache_lock(ncp); 1053 error = cache_resolve(nch, cred); 1054 _cache_unlock(ncp); 1055 } else { 1056 error = 0; 1057 } 1058 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 1059 /* 1060 * Accessing the vnode from the namecache is a bit 1061 * dangerous. Because there are no refs on the vnode, it 1062 * could be in the middle of a reclaim. 1063 */ 1064 if (vp->v_flag & VRECLAIMED) { 1065 kprintf("Warning: vnode reclaim race detected in cache_vget on %p (%s)\n", vp, ncp->nc_name); 1066 _cache_lock(ncp); 1067 _cache_setunresolved(ncp); 1068 _cache_unlock(ncp); 1069 goto again; 1070 } 1071 error = vget(vp, lk_type); 1072 if (error) { 1073 if (vp != ncp->nc_vp) 1074 goto again; 1075 vp = NULL; 1076 } else if (vp != ncp->nc_vp) { 1077 vput(vp); 1078 goto again; 1079 } else if (vp->v_flag & VRECLAIMED) { 1080 panic("vget succeeded on a VRECLAIMED node! vp %p", vp); 1081 } 1082 } 1083 if (error == 0 && vp == NULL) 1084 error = ENOENT; 1085 *vpp = vp; 1086 return(error); 1087 } 1088 1089 int 1090 cache_vref(struct nchandle *nch, struct ucred *cred, struct vnode **vpp) 1091 { 1092 struct namecache *ncp; 1093 struct vnode *vp; 1094 int error; 1095 1096 ncp = nch->ncp; 1097 1098 again: 1099 vp = NULL; 1100 if (ncp->nc_flag & NCF_UNRESOLVED) { 1101 _cache_lock(ncp); 1102 error = cache_resolve(nch, cred); 1103 _cache_unlock(ncp); 1104 } else { 1105 error = 0; 1106 } 1107 if (error == 0 && (vp = ncp->nc_vp) != NULL) { 1108 /* 1109 * Since we did not obtain any locks, a cache zap 1110 * race can occur here if the vnode is in the middle 1111 * of being reclaimed and has not yet been able to 1112 * clean out its cache node. If that case occurs, 1113 * we must lock and unresolve the cache, then loop 1114 * to retry. 1115 */ 1116 if ((error = vget(vp, LK_SHARED)) != 0) { 1117 if (error == ENOENT) { 1118 kprintf("Warning: vnode reclaim race detected on cache_vref %p (%s)\n", vp, ncp->nc_name); 1119 _cache_lock(ncp); 1120 _cache_setunresolved(ncp); 1121 _cache_unlock(ncp); 1122 goto again; 1123 } 1124 /* fatal error */ 1125 } else { 1126 /* caller does not want a lock */ 1127 vn_unlock(vp); 1128 } 1129 } 1130 if (error == 0 && vp == NULL) 1131 error = ENOENT; 1132 *vpp = vp; 1133 return(error); 1134 } 1135 1136 /* 1137 * Recursively set the FSMID update flag for namecache nodes leading 1138 * to root. This will cause the next getattr or reclaim to increment the 1139 * fsmid and mark the inode for lazy updating. 1140 * 1141 * Stop recursing when we hit a node whos NCF_FSMID flag is already set. 1142 * This makes FSMIDs work in an Einsteinian fashion - where the observation 1143 * effects the result. In this case a program monitoring a higher level 1144 * node will have detected some prior change and started its scan (clearing 1145 * NCF_FSMID in higher level nodes), but since it has not yet observed the 1146 * node where we find NCF_FSMID still set, we can safely make the related 1147 * modification without interfering with the theorized program. 1148 * 1149 * This also means that FSMIDs cannot represent time-domain quantities 1150 * in a hierarchical sense. But the main reason for doing it this way 1151 * is to reduce the amount of recursion that occurs in the critical path 1152 * when e.g. a program is writing to a file that sits deep in a directory 1153 * hierarchy. 1154 */ 1155 void 1156 cache_update_fsmid(struct nchandle *nch) 1157 { 1158 struct namecache *ncp; 1159 struct namecache *scan; 1160 struct vnode *vp; 1161 1162 ncp = nch->ncp; 1163 1164 /* 1165 * Warning: even if we get a non-NULL vp it could still be in the 1166 * middle of a recyclement. Don't do anything fancy, just set 1167 * NCF_FSMID. 1168 */ 1169 if ((vp = ncp->nc_vp) != NULL) { 1170 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 1171 for (scan = ncp; scan; scan = scan->nc_parent) { 1172 if (scan->nc_flag & NCF_FSMID) 1173 break; 1174 scan->nc_flag |= NCF_FSMID; 1175 } 1176 } 1177 } else { 1178 while (ncp && (ncp->nc_flag & NCF_FSMID) == 0) { 1179 ncp->nc_flag |= NCF_FSMID; 1180 ncp = ncp->nc_parent; 1181 } 1182 } 1183 } 1184 1185 void 1186 cache_update_fsmid_vp(struct vnode *vp) 1187 { 1188 struct namecache *ncp; 1189 struct namecache *scan; 1190 1191 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 1192 for (scan = ncp; scan; scan = scan->nc_parent) { 1193 if (scan->nc_flag & NCF_FSMID) 1194 break; 1195 scan->nc_flag |= NCF_FSMID; 1196 } 1197 } 1198 } 1199 1200 /* 1201 * If getattr is called on a vnode (e.g. a stat call), the filesystem 1202 * may call this routine to determine if the namecache has the hierarchical 1203 * change flag set, requiring the fsmid to be updated. 1204 * 1205 * Since 0 indicates no support, make sure the filesystem fsmid is at least 1206 * 1. 1207 */ 1208 int 1209 cache_check_fsmid_vp(struct vnode *vp, int64_t *fsmid) 1210 { 1211 struct namecache *ncp; 1212 int changed = 0; 1213 1214 TAILQ_FOREACH(ncp, &vp->v_namecache, nc_vnode) { 1215 if (ncp->nc_flag & NCF_FSMID) { 1216 ncp->nc_flag &= ~NCF_FSMID; 1217 changed = 1; 1218 } 1219 } 1220 if (*fsmid == 0) 1221 ++*fsmid; 1222 if (changed) 1223 ++*fsmid; 1224 return(changed); 1225 } 1226 1227 /* 1228 * Obtain the FSMID for a vnode for filesystems which do not support 1229 * a built-in FSMID. 1230 */ 1231 int64_t 1232 cache_sync_fsmid_vp(struct vnode *vp) 1233 { 1234 struct namecache *ncp; 1235 1236 if ((ncp = TAILQ_FIRST(&vp->v_namecache)) != NULL) { 1237 if (ncp->nc_flag & NCF_FSMID) { 1238 ncp->nc_flag &= ~NCF_FSMID; 1239 ++ncp->nc_fsmid; 1240 } 1241 return(ncp->nc_fsmid); 1242 } 1243 return(VNOVAL); 1244 } 1245 1246 /* 1247 * Convert a directory vnode to a namecache record without any other 1248 * knowledge of the topology. This ONLY works with directory vnodes and 1249 * is ONLY used by the NFS server. dvp must be refd but unlocked, and the 1250 * returned ncp (if not NULL) will be held and unlocked. 1251 * 1252 * If 'makeit' is 0 and dvp has no existing namecache record, NULL is returned. 1253 * If 'makeit' is 1 we attempt to track-down and create the namecache topology 1254 * for dvp. This will fail only if the directory has been deleted out from 1255 * under the caller. 1256 * 1257 * Callers must always check for a NULL return no matter the value of 'makeit'. 1258 * 1259 * To avoid underflowing the kernel stack each recursive call increments 1260 * the makeit variable. 1261 */ 1262 1263 static int cache_inefficient_scan(struct nchandle *nch, struct ucred *cred, 1264 struct vnode *dvp); 1265 static int cache_fromdvp_try(struct vnode *dvp, struct ucred *cred, 1266 struct vnode **saved_dvp); 1267 1268 int 1269 cache_fromdvp(struct vnode *dvp, struct ucred *cred, int makeit, 1270 struct nchandle *nch) 1271 { 1272 struct vnode *saved_dvp; 1273 struct vnode *pvp; 1274 int error; 1275 1276 nch->ncp = NULL; 1277 nch->mount = dvp->v_mount; 1278 saved_dvp = NULL; 1279 1280 /* 1281 * Temporary debugging code to force the directory scanning code 1282 * to be exercised. 1283 */ 1284 if (ncvp_debug >= 3 && makeit && TAILQ_FIRST(&dvp->v_namecache)) { 1285 nch->ncp = TAILQ_FIRST(&dvp->v_namecache); 1286 kprintf("cache_fromdvp: forcing %s\n", nch->ncp->nc_name); 1287 goto force; 1288 } 1289 1290 /* 1291 * Loop until resolution, inside code will break out on error. 1292 */ 1293 while ((nch->ncp = TAILQ_FIRST(&dvp->v_namecache)) == NULL && makeit) { 1294 force: 1295 /* 1296 * If dvp is the root of its filesystem it should already 1297 * have a namecache pointer associated with it as a side 1298 * effect of the mount, but it may have been disassociated. 1299 */ 1300 if (dvp->v_flag & VROOT) { 1301 nch->ncp = _cache_get(nch->mount->mnt_ncmountpt.ncp); 1302 error = cache_resolve_mp(nch->mount); 1303 _cache_put(nch->ncp); 1304 if (ncvp_debug) { 1305 kprintf("cache_fromdvp: resolve root of mount %p error %d", 1306 dvp->v_mount, error); 1307 } 1308 if (error) { 1309 if (ncvp_debug) 1310 kprintf(" failed\n"); 1311 nch->ncp = NULL; 1312 break; 1313 } 1314 if (ncvp_debug) 1315 kprintf(" succeeded\n"); 1316 continue; 1317 } 1318 1319 /* 1320 * If we are recursed too deeply resort to an O(n^2) 1321 * algorithm to resolve the namecache topology. The 1322 * resolved pvp is left referenced in saved_dvp to 1323 * prevent the tree from being destroyed while we loop. 1324 */ 1325 if (makeit > 20) { 1326 error = cache_fromdvp_try(dvp, cred, &saved_dvp); 1327 if (error) { 1328 kprintf("lookupdotdot(longpath) failed %d " 1329 "dvp %p\n", error, dvp); 1330 break; 1331 } 1332 continue; 1333 } 1334 1335 /* 1336 * Get the parent directory and resolve its ncp. 1337 */ 1338 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred); 1339 if (error) { 1340 kprintf("lookupdotdot failed %d dvp %p\n", error, dvp); 1341 break; 1342 } 1343 vn_unlock(pvp); 1344 1345 /* 1346 * Reuse makeit as a recursion depth counter. 1347 */ 1348 cache_fromdvp(pvp, cred, makeit + 1, nch); 1349 vrele(pvp); 1350 if (nch->ncp == NULL) 1351 break; 1352 1353 /* 1354 * Do an inefficient scan of pvp (embodied by ncp) to look 1355 * for dvp. This will create a namecache record for dvp on 1356 * success. We loop up to recheck on success. 1357 * 1358 * ncp and dvp are both held but not locked. 1359 */ 1360 error = cache_inefficient_scan(nch, cred, dvp); 1361 _cache_drop(nch->ncp); 1362 if (error) { 1363 kprintf("cache_fromdvp: scan %p (%s) failed on dvp=%p\n", 1364 pvp, nch->ncp->nc_name, dvp); 1365 nch->ncp = NULL; 1366 break; 1367 } 1368 if (ncvp_debug) { 1369 kprintf("cache_fromdvp: scan %p (%s) succeeded\n", 1370 pvp, nch->ncp->nc_name); 1371 } 1372 } 1373 1374 /* 1375 * hold it for real so the mount gets a ref 1376 */ 1377 if (nch->ncp) 1378 cache_hold(nch); 1379 if (saved_dvp) 1380 vrele(saved_dvp); 1381 if (nch->ncp) 1382 return (0); 1383 return (EINVAL); 1384 } 1385 1386 /* 1387 * Go up the chain of parent directories until we find something 1388 * we can resolve into the namecache. This is very inefficient. 1389 */ 1390 static 1391 int 1392 cache_fromdvp_try(struct vnode *dvp, struct ucred *cred, 1393 struct vnode **saved_dvp) 1394 { 1395 struct nchandle nch; 1396 struct vnode *pvp; 1397 int error; 1398 static time_t last_fromdvp_report; 1399 1400 /* 1401 * Loop getting the parent directory vnode until we get something we 1402 * can resolve in the namecache. 1403 */ 1404 vref(dvp); 1405 nch.mount = dvp->v_mount; 1406 1407 for (;;) { 1408 error = vop_nlookupdotdot(*dvp->v_ops, dvp, &pvp, cred); 1409 if (error) { 1410 vrele(dvp); 1411 return (error); 1412 } 1413 vn_unlock(pvp); 1414 if ((nch.ncp = TAILQ_FIRST(&pvp->v_namecache)) != NULL) { 1415 _cache_hold(nch.ncp); 1416 vrele(pvp); 1417 break; 1418 } 1419 if (pvp->v_flag & VROOT) { 1420 nch.ncp = _cache_get(pvp->v_mount->mnt_ncmountpt.ncp); 1421 error = cache_resolve_mp(nch.mount); 1422 _cache_unlock(nch.ncp); 1423 vrele(pvp); 1424 if (error) { 1425 _cache_drop(nch.ncp); 1426 vrele(dvp); 1427 return (error); 1428 } 1429 break; 1430 } 1431 vrele(dvp); 1432 dvp = pvp; 1433 } 1434 if (last_fromdvp_report != time_second) { 1435 last_fromdvp_report = time_second; 1436 kprintf("Warning: extremely inefficient path resolution on %s\n", 1437 nch.ncp->nc_name); 1438 } 1439 error = cache_inefficient_scan(&nch, cred, dvp); 1440 1441 /* 1442 * Hopefully dvp now has a namecache record associated with it. 1443 * Leave it referenced to prevent the kernel from recycling the 1444 * vnode. Otherwise extremely long directory paths could result 1445 * in endless recycling. 1446 */ 1447 if (*saved_dvp) 1448 vrele(*saved_dvp); 1449 *saved_dvp = dvp; 1450 return (error); 1451 } 1452 1453 1454 /* 1455 * Do an inefficient scan of the directory represented by ncp looking for 1456 * the directory vnode dvp. ncp must be held but not locked on entry and 1457 * will be held on return. dvp must be refd but not locked on entry and 1458 * will remain refd on return. 1459 * 1460 * Why do this at all? Well, due to its stateless nature the NFS server 1461 * converts file handles directly to vnodes without necessarily going through 1462 * the namecache ops that would otherwise create the namecache topology 1463 * leading to the vnode. We could either (1) Change the namecache algorithms 1464 * to allow disconnect namecache records that are re-merged opportunistically, 1465 * or (2) Make the NFS server backtrack and scan to recover a connected 1466 * namecache topology in order to then be able to issue new API lookups. 1467 * 1468 * It turns out that (1) is a huge mess. It takes a nice clean set of 1469 * namecache algorithms and introduces a lot of complication in every subsystem 1470 * that calls into the namecache to deal with the re-merge case, especially 1471 * since we are using the namecache to placehold negative lookups and the 1472 * vnode might not be immediately assigned. (2) is certainly far less 1473 * efficient then (1), but since we are only talking about directories here 1474 * (which are likely to remain cached), the case does not actually run all 1475 * that often and has the supreme advantage of not polluting the namecache 1476 * algorithms. 1477 */ 1478 static int 1479 cache_inefficient_scan(struct nchandle *nch, struct ucred *cred, 1480 struct vnode *dvp) 1481 { 1482 struct nlcomponent nlc; 1483 struct nchandle rncp; 1484 struct dirent *den; 1485 struct vnode *pvp; 1486 struct vattr vat; 1487 struct iovec iov; 1488 struct uio uio; 1489 int blksize; 1490 int eofflag; 1491 int bytes; 1492 char *rbuf; 1493 int error; 1494 1495 vat.va_blocksize = 0; 1496 if ((error = VOP_GETATTR(dvp, &vat)) != 0) 1497 return (error); 1498 if ((error = cache_vref(nch, cred, &pvp)) != 0) 1499 return (error); 1500 if (ncvp_debug) 1501 kprintf("inefficient_scan: directory iosize %ld vattr fileid = %lld\n", vat.va_blocksize, vat.va_fileid); 1502 if ((blksize = vat.va_blocksize) == 0) 1503 blksize = DEV_BSIZE; 1504 rbuf = kmalloc(blksize, M_TEMP, M_WAITOK); 1505 rncp.ncp = NULL; 1506 1507 eofflag = 0; 1508 uio.uio_offset = 0; 1509 again: 1510 iov.iov_base = rbuf; 1511 iov.iov_len = blksize; 1512 uio.uio_iov = &iov; 1513 uio.uio_iovcnt = 1; 1514 uio.uio_resid = blksize; 1515 uio.uio_segflg = UIO_SYSSPACE; 1516 uio.uio_rw = UIO_READ; 1517 uio.uio_td = curthread; 1518 1519 if (ncvp_debug >= 2) 1520 kprintf("cache_inefficient_scan: readdir @ %08x\n", (int)uio.uio_offset); 1521 error = VOP_READDIR(pvp, &uio, cred, &eofflag, NULL, NULL); 1522 if (error == 0) { 1523 den = (struct dirent *)rbuf; 1524 bytes = blksize - uio.uio_resid; 1525 1526 while (bytes > 0) { 1527 if (ncvp_debug >= 2) { 1528 kprintf("cache_inefficient_scan: %*.*s\n", 1529 den->d_namlen, den->d_namlen, 1530 den->d_name); 1531 } 1532 if (den->d_type != DT_WHT && 1533 den->d_ino == vat.va_fileid) { 1534 if (ncvp_debug) { 1535 kprintf("cache_inefficient_scan: " 1536 "MATCHED inode %lld path %s/%*.*s\n", 1537 vat.va_fileid, nch->ncp->nc_name, 1538 den->d_namlen, den->d_namlen, 1539 den->d_name); 1540 } 1541 nlc.nlc_nameptr = den->d_name; 1542 nlc.nlc_namelen = den->d_namlen; 1543 rncp = cache_nlookup(nch, &nlc); 1544 KKASSERT(rncp.ncp != NULL); 1545 break; 1546 } 1547 bytes -= _DIRENT_DIRSIZ(den); 1548 den = _DIRENT_NEXT(den); 1549 } 1550 if (rncp.ncp == NULL && eofflag == 0 && uio.uio_resid != blksize) 1551 goto again; 1552 } 1553 vrele(pvp); 1554 if (rncp.ncp) { 1555 if (rncp.ncp->nc_flag & NCF_UNRESOLVED) { 1556 _cache_setvp(rncp.ncp, dvp); 1557 if (ncvp_debug >= 2) { 1558 kprintf("cache_inefficient_scan: setvp %s/%s = %p\n", 1559 nch->ncp->nc_name, rncp.ncp->nc_name, dvp); 1560 } 1561 } else { 1562 if (ncvp_debug >= 2) { 1563 kprintf("cache_inefficient_scan: setvp %s/%s already set %p/%p\n", 1564 nch->ncp->nc_name, rncp.ncp->nc_name, dvp, 1565 rncp.ncp->nc_vp); 1566 } 1567 } 1568 if (rncp.ncp->nc_vp == NULL) 1569 error = rncp.ncp->nc_error; 1570 _cache_put(rncp.ncp); 1571 } else { 1572 kprintf("cache_inefficient_scan: dvp %p NOT FOUND in %s\n", 1573 dvp, nch->ncp->nc_name); 1574 error = ENOENT; 1575 } 1576 kfree(rbuf, M_TEMP); 1577 return (error); 1578 } 1579 1580 /* 1581 * Zap a namecache entry. The ncp is unconditionally set to an unresolved 1582 * state, which disassociates it from its vnode or ncneglist. 1583 * 1584 * Then, if there are no additional references to the ncp and no children, 1585 * the ncp is removed from the topology and destroyed. This function will 1586 * also run through the nc_parent chain and destroy parent ncps if possible. 1587 * As a side benefit, it turns out the only conditions that allow running 1588 * up the chain are also the conditions to ensure no deadlock will occur. 1589 * 1590 * References and/or children may exist if the ncp is in the middle of the 1591 * topology, preventing the ncp from being destroyed. 1592 * 1593 * This function must be called with the ncp held and locked and will unlock 1594 * and drop it during zapping. 1595 */ 1596 static void 1597 cache_zap(struct namecache *ncp) 1598 { 1599 struct namecache *par; 1600 1601 /* 1602 * Disassociate the vnode or negative cache ref and set NCF_UNRESOLVED. 1603 */ 1604 _cache_setunresolved(ncp); 1605 1606 /* 1607 * Try to scrap the entry and possibly tail-recurse on its parent. 1608 * We only scrap unref'd (other then our ref) unresolved entries, 1609 * we do not scrap 'live' entries. 1610 */ 1611 while (ncp->nc_flag & NCF_UNRESOLVED) { 1612 /* 1613 * Someone other then us has a ref, stop. 1614 */ 1615 if (ncp->nc_refs > 1) 1616 goto done; 1617 1618 /* 1619 * We have children, stop. 1620 */ 1621 if (!TAILQ_EMPTY(&ncp->nc_list)) 1622 goto done; 1623 1624 /* 1625 * Remove ncp from the topology: hash table and parent linkage. 1626 */ 1627 if (ncp->nc_flag & NCF_HASHED) { 1628 ncp->nc_flag &= ~NCF_HASHED; 1629 LIST_REMOVE(ncp, nc_hash); 1630 } 1631 if ((par = ncp->nc_parent) != NULL) { 1632 par = _cache_hold(par); 1633 TAILQ_REMOVE(&par->nc_list, ncp, nc_entry); 1634 ncp->nc_parent = NULL; 1635 if (par->nc_vp && TAILQ_EMPTY(&par->nc_list)) 1636 vdrop(par->nc_vp); 1637 } 1638 1639 /* 1640 * ncp should not have picked up any refs. Physically 1641 * destroy the ncp. 1642 */ 1643 KKASSERT(ncp->nc_refs == 1); 1644 --numunres; 1645 /* _cache_unlock(ncp) not required */ 1646 ncp->nc_refs = -1; /* safety */ 1647 if (ncp->nc_name) 1648 kfree(ncp->nc_name, M_VFSCACHE); 1649 kfree(ncp, M_VFSCACHE); 1650 1651 /* 1652 * Loop on the parent (it may be NULL). Only bother looping 1653 * if the parent has a single ref (ours), which also means 1654 * we can lock it trivially. 1655 */ 1656 ncp = par; 1657 if (ncp == NULL) 1658 return; 1659 if (ncp->nc_refs != 1) { 1660 _cache_drop(ncp); 1661 return; 1662 } 1663 KKASSERT(par->nc_exlocks == 0); 1664 _cache_lock(ncp); 1665 } 1666 done: 1667 _cache_unlock(ncp); 1668 atomic_subtract_int(&ncp->nc_refs, 1); 1669 } 1670 1671 static enum { CHI_LOW, CHI_HIGH } cache_hysteresis_state = CHI_LOW; 1672 1673 static __inline 1674 void 1675 cache_hysteresis(void) 1676 { 1677 /* 1678 * Don't cache too many negative hits. We use hysteresis to reduce 1679 * the impact on the critical path. 1680 */ 1681 switch(cache_hysteresis_state) { 1682 case CHI_LOW: 1683 if (numneg > MINNEG && numneg * ncnegfactor > numcache) { 1684 cache_cleanneg(10); 1685 cache_hysteresis_state = CHI_HIGH; 1686 } 1687 break; 1688 case CHI_HIGH: 1689 if (numneg > MINNEG * 9 / 10 && 1690 numneg * ncnegfactor * 9 / 10 > numcache 1691 ) { 1692 cache_cleanneg(10); 1693 } else { 1694 cache_hysteresis_state = CHI_LOW; 1695 } 1696 break; 1697 } 1698 } 1699 1700 /* 1701 * NEW NAMECACHE LOOKUP API 1702 * 1703 * Lookup an entry in the cache. A locked, referenced, non-NULL 1704 * entry is *always* returned, even if the supplied component is illegal. 1705 * The resulting namecache entry should be returned to the system with 1706 * cache_put() or _cache_unlock() + cache_drop(). 1707 * 1708 * namecache locks are recursive but care must be taken to avoid lock order 1709 * reversals. 1710 * 1711 * Nobody else will be able to manipulate the associated namespace (e.g. 1712 * create, delete, rename, rename-target) until the caller unlocks the 1713 * entry. 1714 * 1715 * The returned entry will be in one of three states: positive hit (non-null 1716 * vnode), negative hit (null vnode), or unresolved (NCF_UNRESOLVED is set). 1717 * Unresolved entries must be resolved through the filesystem to associate the 1718 * vnode and/or determine whether a positive or negative hit has occured. 1719 * 1720 * It is not necessary to lock a directory in order to lock namespace under 1721 * that directory. In fact, it is explicitly not allowed to do that. A 1722 * directory is typically only locked when being created, renamed, or 1723 * destroyed. 1724 * 1725 * The directory (par) may be unresolved, in which case any returned child 1726 * will likely also be marked unresolved. Likely but not guarenteed. Since 1727 * the filesystem lookup requires a resolved directory vnode the caller is 1728 * responsible for resolving the namecache chain top-down. This API 1729 * specifically allows whole chains to be created in an unresolved state. 1730 */ 1731 struct nchandle 1732 cache_nlookup(struct nchandle *par_nch, struct nlcomponent *nlc) 1733 { 1734 struct nchandle nch; 1735 struct namecache *ncp; 1736 struct namecache *new_ncp; 1737 struct nchashhead *nchpp; 1738 u_int32_t hash; 1739 globaldata_t gd; 1740 1741 numcalls++; 1742 gd = mycpu; 1743 1744 /* 1745 * Try to locate an existing entry 1746 */ 1747 hash = fnv_32_buf(nlc->nlc_nameptr, nlc->nlc_namelen, FNV1_32_INIT); 1748 hash = fnv_32_buf(&par_nch->ncp, sizeof(par_nch->ncp), hash); 1749 new_ncp = NULL; 1750 restart: 1751 LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { 1752 numchecks++; 1753 1754 /* 1755 * Try to zap entries that have timed out. We have 1756 * to be careful here because locked leafs may depend 1757 * on the vnode remaining intact in a parent, so only 1758 * do this under very specific conditions. 1759 */ 1760 if (ncp->nc_timeout && 1761 (int)(ncp->nc_timeout - ticks) < 0 && 1762 (ncp->nc_flag & NCF_UNRESOLVED) == 0 && 1763 ncp->nc_exlocks == 0 && 1764 TAILQ_EMPTY(&ncp->nc_list) 1765 ) { 1766 cache_zap(_cache_get(ncp)); 1767 goto restart; 1768 } 1769 1770 /* 1771 * Break out if we find a matching entry. Note that 1772 * UNRESOLVED entries may match, but DESTROYED entries 1773 * do not. 1774 */ 1775 if (ncp->nc_parent == par_nch->ncp && 1776 ncp->nc_nlen == nlc->nlc_namelen && 1777 bcmp(ncp->nc_name, nlc->nlc_nameptr, ncp->nc_nlen) == 0 && 1778 (ncp->nc_flag & NCF_DESTROYED) == 0 1779 ) { 1780 if (_cache_get_nonblock(ncp) == 0) { 1781 if (new_ncp) 1782 _cache_free(new_ncp); 1783 goto found; 1784 } 1785 _cache_get(ncp); 1786 _cache_put(ncp); 1787 goto restart; 1788 } 1789 } 1790 1791 /* 1792 * We failed to locate an entry, create a new entry and add it to 1793 * the cache. We have to relookup after possibly blocking in 1794 * malloc. 1795 */ 1796 if (new_ncp == NULL) { 1797 new_ncp = cache_alloc(nlc->nlc_namelen); 1798 goto restart; 1799 } 1800 1801 ncp = new_ncp; 1802 1803 /* 1804 * Initialize as a new UNRESOLVED entry, lock (non-blocking), 1805 * and link to the parent. The mount point is usually inherited 1806 * from the parent unless this is a special case such as a mount 1807 * point where nlc_namelen is 0. If nlc_namelen is 0 nc_name will 1808 * be NULL. 1809 */ 1810 if (nlc->nlc_namelen) { 1811 bcopy(nlc->nlc_nameptr, ncp->nc_name, nlc->nlc_namelen); 1812 ncp->nc_name[nlc->nlc_namelen] = 0; 1813 } 1814 nchpp = NCHHASH(hash); 1815 LIST_INSERT_HEAD(nchpp, ncp, nc_hash); 1816 ncp->nc_flag |= NCF_HASHED; 1817 cache_link_parent(ncp, par_nch->ncp); 1818 found: 1819 /* 1820 * stats and namecache size management 1821 */ 1822 if (ncp->nc_flag & NCF_UNRESOLVED) 1823 ++gd->gd_nchstats->ncs_miss; 1824 else if (ncp->nc_vp) 1825 ++gd->gd_nchstats->ncs_goodhits; 1826 else 1827 ++gd->gd_nchstats->ncs_neghits; 1828 cache_hysteresis(); 1829 nch.mount = par_nch->mount; 1830 nch.ncp = ncp; 1831 ++nch.mount->mnt_refs; 1832 return(nch); 1833 } 1834 1835 /* 1836 * The namecache entry is marked as being used as a mount point. 1837 * Locate the mount if it is visible to the caller. 1838 */ 1839 struct findmount_info { 1840 struct mount *result; 1841 struct mount *nch_mount; 1842 struct namecache *nch_ncp; 1843 }; 1844 1845 static 1846 int 1847 cache_findmount_callback(struct mount *mp, void *data) 1848 { 1849 struct findmount_info *info = data; 1850 1851 /* 1852 * Check the mount's mounted-on point against the passed nch. 1853 */ 1854 if (mp->mnt_ncmounton.mount == info->nch_mount && 1855 mp->mnt_ncmounton.ncp == info->nch_ncp 1856 ) { 1857 info->result = mp; 1858 return(-1); 1859 } 1860 return(0); 1861 } 1862 1863 struct mount * 1864 cache_findmount(struct nchandle *nch) 1865 { 1866 struct findmount_info info; 1867 1868 info.result = NULL; 1869 info.nch_mount = nch->mount; 1870 info.nch_ncp = nch->ncp; 1871 mountlist_scan(cache_findmount_callback, &info, 1872 MNTSCAN_FORWARD|MNTSCAN_NOBUSY); 1873 return(info.result); 1874 } 1875 1876 /* 1877 * Resolve an unresolved namecache entry, generally by looking it up. 1878 * The passed ncp must be locked and refd. 1879 * 1880 * Theoretically since a vnode cannot be recycled while held, and since 1881 * the nc_parent chain holds its vnode as long as children exist, the 1882 * direct parent of the cache entry we are trying to resolve should 1883 * have a valid vnode. If not then generate an error that we can 1884 * determine is related to a resolver bug. 1885 * 1886 * However, if a vnode was in the middle of a recyclement when the NCP 1887 * got locked, ncp->nc_vp might point to a vnode that is about to become 1888 * invalid. cache_resolve() handles this case by unresolving the entry 1889 * and then re-resolving it. 1890 * 1891 * Note that successful resolution does not necessarily return an error 1892 * code of 0. If the ncp resolves to a negative cache hit then ENOENT 1893 * will be returned. 1894 */ 1895 int 1896 cache_resolve(struct nchandle *nch, struct ucred *cred) 1897 { 1898 struct namecache *par; 1899 struct namecache *ncp; 1900 struct nchandle nctmp; 1901 struct mount *mp; 1902 struct vnode *dvp; 1903 int error; 1904 1905 ncp = nch->ncp; 1906 mp = nch->mount; 1907 restart: 1908 /* 1909 * If the ncp is already resolved we have nothing to do. However, 1910 * we do want to guarentee that a usable vnode is returned when 1911 * a vnode is present, so make sure it hasn't been reclaimed. 1912 */ 1913 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 1914 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 1915 _cache_setunresolved(ncp); 1916 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) 1917 return (ncp->nc_error); 1918 } 1919 1920 /* 1921 * Mount points need special handling because the parent does not 1922 * belong to the same filesystem as the ncp. 1923 */ 1924 if (ncp == mp->mnt_ncmountpt.ncp) 1925 return (cache_resolve_mp(mp)); 1926 1927 /* 1928 * We expect an unbroken chain of ncps to at least the mount point, 1929 * and even all the way to root (but this code doesn't have to go 1930 * past the mount point). 1931 */ 1932 if (ncp->nc_parent == NULL) { 1933 kprintf("EXDEV case 1 %p %*.*s\n", ncp, 1934 ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 1935 ncp->nc_error = EXDEV; 1936 return(ncp->nc_error); 1937 } 1938 1939 /* 1940 * The vp's of the parent directories in the chain are held via vhold() 1941 * due to the existance of the child, and should not disappear. 1942 * However, there are cases where they can disappear: 1943 * 1944 * - due to filesystem I/O errors. 1945 * - due to NFS being stupid about tracking the namespace and 1946 * destroys the namespace for entire directories quite often. 1947 * - due to forced unmounts. 1948 * - due to an rmdir (parent will be marked DESTROYED) 1949 * 1950 * When this occurs we have to track the chain backwards and resolve 1951 * it, looping until the resolver catches up to the current node. We 1952 * could recurse here but we might run ourselves out of kernel stack 1953 * so we do it in a more painful manner. This situation really should 1954 * not occur all that often, or if it does not have to go back too 1955 * many nodes to resolve the ncp. 1956 */ 1957 while (ncp->nc_parent->nc_vp == NULL) { 1958 /* 1959 * This case can occur if a process is CD'd into a 1960 * directory which is then rmdir'd. If the parent is marked 1961 * destroyed there is no point trying to resolve it. 1962 */ 1963 if (ncp->nc_parent->nc_flag & NCF_DESTROYED) 1964 return(ENOENT); 1965 1966 par = ncp->nc_parent; 1967 while (par->nc_parent && par->nc_parent->nc_vp == NULL) 1968 par = par->nc_parent; 1969 if (par->nc_parent == NULL) { 1970 kprintf("EXDEV case 2 %*.*s\n", 1971 par->nc_nlen, par->nc_nlen, par->nc_name); 1972 return (EXDEV); 1973 } 1974 kprintf("[diagnostic] cache_resolve: had to recurse on %*.*s\n", 1975 par->nc_nlen, par->nc_nlen, par->nc_name); 1976 /* 1977 * The parent is not set in stone, ref and lock it to prevent 1978 * it from disappearing. Also note that due to renames it 1979 * is possible for our ncp to move and for par to no longer 1980 * be one of its parents. We resolve it anyway, the loop 1981 * will handle any moves. 1982 */ 1983 _cache_get(par); 1984 if (par == nch->mount->mnt_ncmountpt.ncp) { 1985 cache_resolve_mp(nch->mount); 1986 } else if ((dvp = par->nc_parent->nc_vp) == NULL) { 1987 kprintf("[diagnostic] cache_resolve: raced on %*.*s\n", par->nc_nlen, par->nc_nlen, par->nc_name); 1988 _cache_put(par); 1989 continue; 1990 } else if (par->nc_flag & NCF_UNRESOLVED) { 1991 /* vhold(dvp); - DVP can't go away */ 1992 nctmp.mount = mp; 1993 nctmp.ncp = par; 1994 par->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred); 1995 /* vdrop(dvp); */ 1996 } 1997 if ((error = par->nc_error) != 0) { 1998 if (par->nc_error != EAGAIN) { 1999 kprintf("EXDEV case 3 %*.*s error %d\n", 2000 par->nc_nlen, par->nc_nlen, par->nc_name, 2001 par->nc_error); 2002 _cache_put(par); 2003 return(error); 2004 } 2005 kprintf("[diagnostic] cache_resolve: EAGAIN par %p %*.*s\n", 2006 par, par->nc_nlen, par->nc_nlen, par->nc_name); 2007 } 2008 _cache_put(par); 2009 /* loop */ 2010 } 2011 2012 /* 2013 * Call VOP_NRESOLVE() to get the vp, then scan for any disconnected 2014 * ncp's and reattach them. If this occurs the original ncp is marked 2015 * EAGAIN to force a relookup. 2016 * 2017 * NOTE: in order to call VOP_NRESOLVE(), the parent of the passed 2018 * ncp must already be resolved. 2019 */ 2020 dvp = ncp->nc_parent->nc_vp; 2021 /* vhold(dvp); - dvp can't go away */ 2022 nctmp.mount = mp; 2023 nctmp.ncp = ncp; 2024 ncp->nc_error = VOP_NRESOLVE(&nctmp, dvp, cred); 2025 /* vdrop(dvp); */ 2026 if (ncp->nc_error == EAGAIN) { 2027 kprintf("[diagnostic] cache_resolve: EAGAIN ncp %p %*.*s\n", 2028 ncp, ncp->nc_nlen, ncp->nc_nlen, ncp->nc_name); 2029 goto restart; 2030 } 2031 return(ncp->nc_error); 2032 } 2033 2034 /* 2035 * Resolve the ncp associated with a mount point. Such ncp's almost always 2036 * remain resolved and this routine is rarely called. NFS MPs tends to force 2037 * re-resolution more often due to its mac-truck-smash-the-namecache 2038 * method of tracking namespace changes. 2039 * 2040 * The semantics for this call is that the passed ncp must be locked on 2041 * entry and will be locked on return. However, if we actually have to 2042 * resolve the mount point we temporarily unlock the entry in order to 2043 * avoid race-to-root deadlocks due to e.g. dead NFS mounts. Because of 2044 * the unlock we have to recheck the flags after we relock. 2045 */ 2046 static int 2047 cache_resolve_mp(struct mount *mp) 2048 { 2049 struct namecache *ncp = mp->mnt_ncmountpt.ncp; 2050 struct vnode *vp; 2051 int error; 2052 2053 KKASSERT(mp != NULL); 2054 2055 /* 2056 * If the ncp is already resolved we have nothing to do. However, 2057 * we do want to guarentee that a usable vnode is returned when 2058 * a vnode is present, so make sure it hasn't been reclaimed. 2059 */ 2060 if ((ncp->nc_flag & NCF_UNRESOLVED) == 0) { 2061 if (ncp->nc_vp && (ncp->nc_vp->v_flag & VRECLAIMED)) 2062 _cache_setunresolved(ncp); 2063 } 2064 2065 if (ncp->nc_flag & NCF_UNRESOLVED) { 2066 _cache_unlock(ncp); 2067 while (vfs_busy(mp, 0)) 2068 ; 2069 error = VFS_ROOT(mp, &vp); 2070 _cache_lock(ncp); 2071 2072 /* 2073 * recheck the ncp state after relocking. 2074 */ 2075 if (ncp->nc_flag & NCF_UNRESOLVED) { 2076 ncp->nc_error = error; 2077 if (error == 0) { 2078 _cache_setvp(ncp, vp); 2079 vput(vp); 2080 } else { 2081 kprintf("[diagnostic] cache_resolve_mp: failed to resolve mount %p\n", mp); 2082 _cache_setvp(ncp, NULL); 2083 } 2084 } else if (error == 0) { 2085 vput(vp); 2086 } 2087 vfs_unbusy(mp); 2088 } 2089 return(ncp->nc_error); 2090 } 2091 2092 void 2093 cache_cleanneg(int count) 2094 { 2095 struct namecache *ncp; 2096 2097 /* 2098 * Automode from the vnlru proc - clean out 10% of the negative cache 2099 * entries. 2100 */ 2101 if (count == 0) 2102 count = numneg / 10 + 1; 2103 2104 /* 2105 * Attempt to clean out the specified number of negative cache 2106 * entries. 2107 */ 2108 while (count) { 2109 ncp = TAILQ_FIRST(&ncneglist); 2110 if (ncp == NULL) { 2111 KKASSERT(numneg == 0); 2112 break; 2113 } 2114 TAILQ_REMOVE(&ncneglist, ncp, nc_vnode); 2115 TAILQ_INSERT_TAIL(&ncneglist, ncp, nc_vnode); 2116 if (_cache_get_nonblock(ncp) == 0) 2117 cache_zap(ncp); 2118 --count; 2119 } 2120 } 2121 2122 /* 2123 * Rehash a ncp. Rehashing is typically required if the name changes (should 2124 * not generally occur) or the parent link changes. This function will 2125 * unhash the ncp if the ncp is no longer hashable. 2126 */ 2127 static void 2128 _cache_rehash(struct namecache *ncp) 2129 { 2130 struct nchashhead *nchpp; 2131 u_int32_t hash; 2132 2133 if (ncp->nc_flag & NCF_HASHED) { 2134 ncp->nc_flag &= ~NCF_HASHED; 2135 LIST_REMOVE(ncp, nc_hash); 2136 } 2137 if (ncp->nc_nlen && ncp->nc_parent) { 2138 hash = fnv_32_buf(ncp->nc_name, ncp->nc_nlen, FNV1_32_INIT); 2139 hash = fnv_32_buf(&ncp->nc_parent, 2140 sizeof(ncp->nc_parent), hash); 2141 nchpp = NCHHASH(hash); 2142 LIST_INSERT_HEAD(nchpp, ncp, nc_hash); 2143 ncp->nc_flag |= NCF_HASHED; 2144 } 2145 } 2146 2147 /* 2148 * Name cache initialization, from vfsinit() when we are booting 2149 */ 2150 void 2151 nchinit(void) 2152 { 2153 int i; 2154 globaldata_t gd; 2155 2156 /* initialise per-cpu namecache effectiveness statistics. */ 2157 for (i = 0; i < ncpus; ++i) { 2158 gd = globaldata_find(i); 2159 gd->gd_nchstats = &nchstats[i]; 2160 } 2161 TAILQ_INIT(&ncneglist); 2162 nchashtbl = hashinit(desiredvnodes*2, M_VFSCACHE, &nchash); 2163 nclockwarn = 1 * hz; 2164 } 2165 2166 /* 2167 * Called from start_init() to bootstrap the root filesystem. Returns 2168 * a referenced, unlocked namecache record. 2169 */ 2170 void 2171 cache_allocroot(struct nchandle *nch, struct mount *mp, struct vnode *vp) 2172 { 2173 nch->ncp = cache_alloc(0); 2174 nch->mount = mp; 2175 ++mp->mnt_refs; 2176 if (vp) 2177 _cache_setvp(nch->ncp, vp); 2178 } 2179 2180 /* 2181 * vfs_cache_setroot() 2182 * 2183 * Create an association between the root of our namecache and 2184 * the root vnode. This routine may be called several times during 2185 * booting. 2186 * 2187 * If the caller intends to save the returned namecache pointer somewhere 2188 * it must cache_hold() it. 2189 */ 2190 void 2191 vfs_cache_setroot(struct vnode *nvp, struct nchandle *nch) 2192 { 2193 struct vnode *ovp; 2194 struct nchandle onch; 2195 2196 ovp = rootvnode; 2197 onch = rootnch; 2198 rootvnode = nvp; 2199 if (nch) 2200 rootnch = *nch; 2201 else 2202 cache_zero(&rootnch); 2203 if (ovp) 2204 vrele(ovp); 2205 if (onch.ncp) 2206 cache_drop(&onch); 2207 } 2208 2209 /* 2210 * XXX OLD API COMPAT FUNCTION. This really messes up the new namecache 2211 * topology and is being removed as quickly as possible. The new VOP_N*() 2212 * API calls are required to make specific adjustments using the supplied 2213 * ncp pointers rather then just bogusly purging random vnodes. 2214 * 2215 * Invalidate all namecache entries to a particular vnode as well as 2216 * any direct children of that vnode in the namecache. This is a 2217 * 'catch all' purge used by filesystems that do not know any better. 2218 * 2219 * Note that the linkage between the vnode and its namecache entries will 2220 * be removed, but the namecache entries themselves might stay put due to 2221 * active references from elsewhere in the system or due to the existance of 2222 * the children. The namecache topology is left intact even if we do not 2223 * know what the vnode association is. Such entries will be marked 2224 * NCF_UNRESOLVED. 2225 */ 2226 void 2227 cache_purge(struct vnode *vp) 2228 { 2229 cache_inval_vp(vp, CINV_DESTROY | CINV_CHILDREN); 2230 } 2231 2232 /* 2233 * Flush all entries referencing a particular filesystem. 2234 * 2235 * Since we need to check it anyway, we will flush all the invalid 2236 * entries at the same time. 2237 */ 2238 #if 0 2239 2240 void 2241 cache_purgevfs(struct mount *mp) 2242 { 2243 struct nchashhead *nchpp; 2244 struct namecache *ncp, *nnp; 2245 2246 /* 2247 * Scan hash tables for applicable entries. 2248 */ 2249 for (nchpp = &nchashtbl[nchash]; nchpp >= nchashtbl; nchpp--) { 2250 ncp = LIST_FIRST(nchpp); 2251 if (ncp) 2252 _cache_hold(ncp); 2253 while (ncp) { 2254 nnp = LIST_NEXT(ncp, nc_hash); 2255 if (nnp) 2256 _cache_hold(nnp); 2257 if (ncp->nc_mount == mp) { 2258 _cache_lock(ncp); 2259 cache_zap(ncp); 2260 } else { 2261 _cache_drop(ncp); 2262 } 2263 ncp = nnp; 2264 } 2265 } 2266 } 2267 2268 #endif 2269 2270 /* 2271 * Create a new (theoretically) unique fsmid 2272 */ 2273 int64_t 2274 cache_getnewfsmid(void) 2275 { 2276 static int fsmid_roller; 2277 int64_t fsmid; 2278 2279 ++fsmid_roller; 2280 fsmid = ((int64_t)time_second << 32) | 2281 (fsmid_roller & 0x7FFFFFFF); 2282 return (fsmid); 2283 } 2284 2285 2286 static int disablecwd; 2287 SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, ""); 2288 2289 static u_long numcwdcalls; STATNODE(CTLFLAG_RD, numcwdcalls, &numcwdcalls); 2290 static u_long numcwdfail1; STATNODE(CTLFLAG_RD, numcwdfail1, &numcwdfail1); 2291 static u_long numcwdfail2; STATNODE(CTLFLAG_RD, numcwdfail2, &numcwdfail2); 2292 static u_long numcwdfail3; STATNODE(CTLFLAG_RD, numcwdfail3, &numcwdfail3); 2293 static u_long numcwdfail4; STATNODE(CTLFLAG_RD, numcwdfail4, &numcwdfail4); 2294 static u_long numcwdfound; STATNODE(CTLFLAG_RD, numcwdfound, &numcwdfound); 2295 2296 int 2297 sys___getcwd(struct __getcwd_args *uap) 2298 { 2299 int buflen; 2300 int error; 2301 char *buf; 2302 char *bp; 2303 2304 if (disablecwd) 2305 return (ENODEV); 2306 2307 buflen = uap->buflen; 2308 if (buflen < 2) 2309 return (EINVAL); 2310 if (buflen > MAXPATHLEN) 2311 buflen = MAXPATHLEN; 2312 2313 buf = kmalloc(buflen, M_TEMP, M_WAITOK); 2314 bp = kern_getcwd(buf, buflen, &error); 2315 if (error == 0) 2316 error = copyout(bp, uap->buf, strlen(bp) + 1); 2317 kfree(buf, M_TEMP); 2318 return (error); 2319 } 2320 2321 char * 2322 kern_getcwd(char *buf, size_t buflen, int *error) 2323 { 2324 struct proc *p = curproc; 2325 char *bp; 2326 int i, slash_prefixed; 2327 struct filedesc *fdp; 2328 struct nchandle nch; 2329 2330 numcwdcalls++; 2331 bp = buf; 2332 bp += buflen - 1; 2333 *bp = '\0'; 2334 fdp = p->p_fd; 2335 slash_prefixed = 0; 2336 2337 nch = fdp->fd_ncdir; 2338 while (nch.ncp && (nch.ncp != fdp->fd_nrdir.ncp || 2339 nch.mount != fdp->fd_nrdir.mount) 2340 ) { 2341 /* 2342 * While traversing upwards if we encounter the root 2343 * of the current mount we have to skip to the mount point 2344 * in the underlying filesystem. 2345 */ 2346 if (nch.ncp == nch.mount->mnt_ncmountpt.ncp) { 2347 nch = nch.mount->mnt_ncmounton; 2348 continue; 2349 } 2350 2351 /* 2352 * Prepend the path segment 2353 */ 2354 for (i = nch.ncp->nc_nlen - 1; i >= 0; i--) { 2355 if (bp == buf) { 2356 numcwdfail4++; 2357 *error = ENOMEM; 2358 return(NULL); 2359 } 2360 *--bp = nch.ncp->nc_name[i]; 2361 } 2362 if (bp == buf) { 2363 numcwdfail4++; 2364 *error = ENOMEM; 2365 return(NULL); 2366 } 2367 *--bp = '/'; 2368 slash_prefixed = 1; 2369 2370 /* 2371 * Go up a directory. This isn't a mount point so we don't 2372 * have to check again. 2373 */ 2374 nch.ncp = nch.ncp->nc_parent; 2375 } 2376 if (nch.ncp == NULL) { 2377 numcwdfail2++; 2378 *error = ENOENT; 2379 return(NULL); 2380 } 2381 if (!slash_prefixed) { 2382 if (bp == buf) { 2383 numcwdfail4++; 2384 *error = ENOMEM; 2385 return(NULL); 2386 } 2387 *--bp = '/'; 2388 } 2389 numcwdfound++; 2390 *error = 0; 2391 return (bp); 2392 } 2393 2394 /* 2395 * Thus begins the fullpath magic. 2396 */ 2397 2398 #undef STATNODE 2399 #define STATNODE(name) \ 2400 static u_int name; \ 2401 SYSCTL_UINT(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, "") 2402 2403 static int disablefullpath; 2404 SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, 2405 &disablefullpath, 0, ""); 2406 2407 STATNODE(numfullpathcalls); 2408 STATNODE(numfullpathfail1); 2409 STATNODE(numfullpathfail2); 2410 STATNODE(numfullpathfail3); 2411 STATNODE(numfullpathfail4); 2412 STATNODE(numfullpathfound); 2413 2414 int 2415 cache_fullpath(struct proc *p, struct nchandle *nchp, char **retbuf, char **freebuf) 2416 { 2417 char *bp, *buf; 2418 int i, slash_prefixed; 2419 struct nchandle fd_nrdir; 2420 struct nchandle nch; 2421 2422 numfullpathcalls--; 2423 2424 *retbuf = NULL; 2425 *freebuf = NULL; 2426 2427 buf = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK); 2428 bp = buf + MAXPATHLEN - 1; 2429 *bp = '\0'; 2430 if (p != NULL) 2431 fd_nrdir = p->p_fd->fd_nrdir; 2432 else 2433 fd_nrdir = rootnch; 2434 slash_prefixed = 0; 2435 nch = *nchp; 2436 2437 while (nch.ncp && 2438 (nch.ncp != fd_nrdir.ncp || nch.mount != fd_nrdir.mount) 2439 ) { 2440 /* 2441 * While traversing upwards if we encounter the root 2442 * of the current mount we have to skip to the mount point. 2443 */ 2444 if (nch.ncp == nch.mount->mnt_ncmountpt.ncp) { 2445 nch = nch.mount->mnt_ncmounton; 2446 continue; 2447 } 2448 2449 /* 2450 * Prepend the path segment 2451 */ 2452 for (i = nch.ncp->nc_nlen - 1; i >= 0; i--) { 2453 if (bp == buf) { 2454 numfullpathfail4++; 2455 kfree(buf, M_TEMP); 2456 return(ENOMEM); 2457 } 2458 *--bp = nch.ncp->nc_name[i]; 2459 } 2460 if (bp == buf) { 2461 numfullpathfail4++; 2462 kfree(buf, M_TEMP); 2463 return(ENOMEM); 2464 } 2465 *--bp = '/'; 2466 slash_prefixed = 1; 2467 2468 /* 2469 * Go up a directory. This isn't a mount point so we don't 2470 * have to check again. 2471 */ 2472 nch.ncp = nch.ncp->nc_parent; 2473 } 2474 if (nch.ncp == NULL) { 2475 numfullpathfail2++; 2476 kfree(buf, M_TEMP); 2477 return(ENOENT); 2478 } 2479 2480 if (!slash_prefixed) { 2481 if (bp == buf) { 2482 numfullpathfail4++; 2483 kfree(buf, M_TEMP); 2484 return(ENOMEM); 2485 } 2486 *--bp = '/'; 2487 } 2488 numfullpathfound++; 2489 *retbuf = bp; 2490 *freebuf = buf; 2491 2492 return(0); 2493 } 2494 2495 int 2496 vn_fullpath(struct proc *p, struct vnode *vn, char **retbuf, char **freebuf) 2497 { 2498 struct namecache *ncp; 2499 struct nchandle nch; 2500 2501 numfullpathcalls++; 2502 if (disablefullpath) 2503 return (ENODEV); 2504 2505 if (p == NULL) 2506 return (EINVAL); 2507 2508 /* vn is NULL, client wants us to use p->p_textvp */ 2509 if (vn == NULL) { 2510 if ((vn = p->p_textvp) == NULL) 2511 return (EINVAL); 2512 } 2513 TAILQ_FOREACH(ncp, &vn->v_namecache, nc_vnode) { 2514 if (ncp->nc_nlen) 2515 break; 2516 } 2517 if (ncp == NULL) 2518 return (EINVAL); 2519 2520 numfullpathcalls--; 2521 nch.ncp = ncp;; 2522 nch.mount = vn->v_mount; 2523 return(cache_fullpath(p, &nch, retbuf, freebuf)); 2524 } 2525