1 /* 2 * fs/dcache.c 3 * 4 * Complete reimplementation 5 * (C) 1997 Thomas Schoebel-Theuer, 6 * with heavy changes by Linus Torvalds 7 */ 8 9 /* 10 * Notes on the allocation strategy: 11 * 12 * The dcache is a master of the icache - whenever a dcache entry 13 * exists, the inode will always exist. "iput()" is done either when 14 * the dcache entry is deleted or garbage collected. 15 */ 16 17 #include <linux/syscalls.h> 18 #include <linux/string.h> 19 #include <linux/mm.h> 20 #include <linux/fs.h> 21 #include <linux/fsnotify.h> 22 #include <linux/slab.h> 23 #include <linux/init.h> 24 #include <linux/hash.h> 25 #include <linux/cache.h> 26 #include <linux/export.h> 27 #include <linux/mount.h> 28 #include <linux/file.h> 29 #include <asm/uaccess.h> 30 #include <linux/security.h> 31 #include <linux/seqlock.h> 32 #include <linux/swap.h> 33 #include <linux/bootmem.h> 34 #include <linux/fs_struct.h> 35 #include <linux/hardirq.h> 36 #include <linux/bit_spinlock.h> 37 #include <linux/rculist_bl.h> 38 #include <linux/prefetch.h> 39 #include <linux/ratelimit.h> 40 #include <linux/list_lru.h> 41 #include "internal.h" 42 #include "mount.h" 43 44 /* 45 * Usage: 46 * dcache->d_inode->i_lock protects: 47 * - i_dentry, d_alias, d_inode of aliases 48 * dcache_hash_bucket lock protects: 49 * - the dcache hash table 50 * s_anon bl list spinlock protects: 51 * - the s_anon list (see __d_drop) 52 * dentry->d_sb->s_dentry_lru_lock protects: 53 * - the dcache lru lists and counters 54 * d_lock protects: 55 * - d_flags 56 * - d_name 57 * - d_lru 58 * - d_count 59 * - d_unhashed() 60 * - d_parent and d_subdirs 61 * - childrens' d_child and d_parent 62 * - d_alias, d_inode 63 * 64 * Ordering: 65 * dentry->d_inode->i_lock 66 * dentry->d_lock 67 * dentry->d_sb->s_dentry_lru_lock 68 * dcache_hash_bucket lock 69 * s_anon lock 70 * 71 * If there is an ancestor relationship: 72 * dentry->d_parent->...->d_parent->d_lock 73 * ... 74 * dentry->d_parent->d_lock 75 * dentry->d_lock 76 * 77 * If no ancestor relationship: 78 * if (dentry1 < dentry2) 79 * dentry1->d_lock 80 * dentry2->d_lock 81 */ 82 int sysctl_vfs_cache_pressure __read_mostly = 100; 83 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure); 84 85 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock); 86 87 EXPORT_SYMBOL(rename_lock); 88 89 static struct kmem_cache *dentry_cache __read_mostly; 90 91 /* 92 * This is the single most critical data structure when it comes 93 * to the dcache: the hashtable for lookups. Somebody should try 94 * to make this good - I've just made it work. 95 * 96 * This hash-function tries to avoid losing too many bits of hash 97 * information, yet avoid using a prime hash-size or similar. 98 */ 99 100 static unsigned int d_hash_mask __read_mostly; 101 static unsigned int d_hash_shift __read_mostly; 102 103 static struct hlist_bl_head *dentry_hashtable __read_mostly; 104 105 static inline struct hlist_bl_head *d_hash(const struct dentry *parent, 106 unsigned int hash) 107 { 108 hash += (unsigned long) parent / L1_CACHE_BYTES; 109 hash = hash + (hash >> d_hash_shift); 110 return dentry_hashtable + (hash & d_hash_mask); 111 } 112 113 /* Statistics gathering. */ 114 struct dentry_stat_t dentry_stat = { 115 .age_limit = 45, 116 }; 117 118 static DEFINE_PER_CPU(long, nr_dentry); 119 static DEFINE_PER_CPU(long, nr_dentry_unused); 120 121 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS) 122 123 /* 124 * Here we resort to our own counters instead of using generic per-cpu counters 125 * for consistency with what the vfs inode code does. We are expected to harvest 126 * better code and performance by having our own specialized counters. 127 * 128 * Please note that the loop is done over all possible CPUs, not over all online 129 * CPUs. The reason for this is that we don't want to play games with CPUs going 130 * on and off. If one of them goes off, we will just keep their counters. 131 * 132 * glommer: See cffbc8a for details, and if you ever intend to change this, 133 * please update all vfs counters to match. 134 */ 135 static long get_nr_dentry(void) 136 { 137 int i; 138 long sum = 0; 139 for_each_possible_cpu(i) 140 sum += per_cpu(nr_dentry, i); 141 return sum < 0 ? 0 : sum; 142 } 143 144 static long get_nr_dentry_unused(void) 145 { 146 int i; 147 long sum = 0; 148 for_each_possible_cpu(i) 149 sum += per_cpu(nr_dentry_unused, i); 150 return sum < 0 ? 0 : sum; 151 } 152 153 int proc_nr_dentry(ctl_table *table, int write, void __user *buffer, 154 size_t *lenp, loff_t *ppos) 155 { 156 dentry_stat.nr_dentry = get_nr_dentry(); 157 dentry_stat.nr_unused = get_nr_dentry_unused(); 158 return proc_doulongvec_minmax(table, write, buffer, lenp, ppos); 159 } 160 #endif 161 162 /* 163 * Compare 2 name strings, return 0 if they match, otherwise non-zero. 164 * The strings are both count bytes long, and count is non-zero. 165 */ 166 #ifdef CONFIG_DCACHE_WORD_ACCESS 167 168 #include <asm/word-at-a-time.h> 169 /* 170 * NOTE! 'cs' and 'scount' come from a dentry, so it has a 171 * aligned allocation for this particular component. We don't 172 * strictly need the load_unaligned_zeropad() safety, but it 173 * doesn't hurt either. 174 * 175 * In contrast, 'ct' and 'tcount' can be from a pathname, and do 176 * need the careful unaligned handling. 177 */ 178 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount) 179 { 180 unsigned long a,b,mask; 181 182 for (;;) { 183 a = *(unsigned long *)cs; 184 b = load_unaligned_zeropad(ct); 185 if (tcount < sizeof(unsigned long)) 186 break; 187 if (unlikely(a != b)) 188 return 1; 189 cs += sizeof(unsigned long); 190 ct += sizeof(unsigned long); 191 tcount -= sizeof(unsigned long); 192 if (!tcount) 193 return 0; 194 } 195 mask = bytemask_from_count(tcount); 196 return unlikely(!!((a ^ b) & mask)); 197 } 198 199 #else 200 201 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount) 202 { 203 do { 204 if (*cs != *ct) 205 return 1; 206 cs++; 207 ct++; 208 tcount--; 209 } while (tcount); 210 return 0; 211 } 212 213 #endif 214 215 static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount) 216 { 217 const unsigned char *cs; 218 /* 219 * Be careful about RCU walk racing with rename: 220 * use ACCESS_ONCE to fetch the name pointer. 221 * 222 * NOTE! Even if a rename will mean that the length 223 * was not loaded atomically, we don't care. The 224 * RCU walk will check the sequence count eventually, 225 * and catch it. And we won't overrun the buffer, 226 * because we're reading the name pointer atomically, 227 * and a dentry name is guaranteed to be properly 228 * terminated with a NUL byte. 229 * 230 * End result: even if 'len' is wrong, we'll exit 231 * early because the data cannot match (there can 232 * be no NUL in the ct/tcount data) 233 */ 234 cs = ACCESS_ONCE(dentry->d_name.name); 235 smp_read_barrier_depends(); 236 return dentry_string_cmp(cs, ct, tcount); 237 } 238 239 static void __d_free(struct rcu_head *head) 240 { 241 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu); 242 243 WARN_ON(!hlist_unhashed(&dentry->d_alias)); 244 if (dname_external(dentry)) 245 kfree(dentry->d_name.name); 246 kmem_cache_free(dentry_cache, dentry); 247 } 248 249 static void dentry_free(struct dentry *dentry) 250 { 251 /* if dentry was never visible to RCU, immediate free is OK */ 252 if (!(dentry->d_flags & DCACHE_RCUACCESS)) 253 __d_free(&dentry->d_u.d_rcu); 254 else 255 call_rcu(&dentry->d_u.d_rcu, __d_free); 256 } 257 258 /** 259 * dentry_rcuwalk_barrier - invalidate in-progress rcu-walk lookups 260 * @dentry: the target dentry 261 * After this call, in-progress rcu-walk path lookup will fail. This 262 * should be called after unhashing, and after changing d_inode (if 263 * the dentry has not already been unhashed). 264 */ 265 static inline void dentry_rcuwalk_barrier(struct dentry *dentry) 266 { 267 assert_spin_locked(&dentry->d_lock); 268 /* Go through a barrier */ 269 write_seqcount_barrier(&dentry->d_seq); 270 } 271 272 /* 273 * Release the dentry's inode, using the filesystem 274 * d_iput() operation if defined. Dentry has no refcount 275 * and is unhashed. 276 */ 277 static void dentry_iput(struct dentry * dentry) 278 __releases(dentry->d_lock) 279 __releases(dentry->d_inode->i_lock) 280 { 281 struct inode *inode = dentry->d_inode; 282 if (inode) { 283 dentry->d_inode = NULL; 284 hlist_del_init(&dentry->d_alias); 285 spin_unlock(&dentry->d_lock); 286 spin_unlock(&inode->i_lock); 287 if (!inode->i_nlink) 288 fsnotify_inoderemove(inode); 289 if (dentry->d_op && dentry->d_op->d_iput) 290 dentry->d_op->d_iput(dentry, inode); 291 else 292 iput(inode); 293 } else { 294 spin_unlock(&dentry->d_lock); 295 } 296 } 297 298 /* 299 * Release the dentry's inode, using the filesystem 300 * d_iput() operation if defined. dentry remains in-use. 301 */ 302 static void dentry_unlink_inode(struct dentry * dentry) 303 __releases(dentry->d_lock) 304 __releases(dentry->d_inode->i_lock) 305 { 306 struct inode *inode = dentry->d_inode; 307 __d_clear_type(dentry); 308 dentry->d_inode = NULL; 309 hlist_del_init(&dentry->d_alias); 310 dentry_rcuwalk_barrier(dentry); 311 spin_unlock(&dentry->d_lock); 312 spin_unlock(&inode->i_lock); 313 if (!inode->i_nlink) 314 fsnotify_inoderemove(inode); 315 if (dentry->d_op && dentry->d_op->d_iput) 316 dentry->d_op->d_iput(dentry, inode); 317 else 318 iput(inode); 319 } 320 321 /* 322 * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry 323 * is in use - which includes both the "real" per-superblock 324 * LRU list _and_ the DCACHE_SHRINK_LIST use. 325 * 326 * The DCACHE_SHRINK_LIST bit is set whenever the dentry is 327 * on the shrink list (ie not on the superblock LRU list). 328 * 329 * The per-cpu "nr_dentry_unused" counters are updated with 330 * the DCACHE_LRU_LIST bit. 331 * 332 * These helper functions make sure we always follow the 333 * rules. d_lock must be held by the caller. 334 */ 335 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x)) 336 static void d_lru_add(struct dentry *dentry) 337 { 338 D_FLAG_VERIFY(dentry, 0); 339 dentry->d_flags |= DCACHE_LRU_LIST; 340 this_cpu_inc(nr_dentry_unused); 341 WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru)); 342 } 343 344 static void d_lru_del(struct dentry *dentry) 345 { 346 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST); 347 dentry->d_flags &= ~DCACHE_LRU_LIST; 348 this_cpu_dec(nr_dentry_unused); 349 WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru)); 350 } 351 352 static void d_shrink_del(struct dentry *dentry) 353 { 354 D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST); 355 list_del_init(&dentry->d_lru); 356 dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST); 357 this_cpu_dec(nr_dentry_unused); 358 } 359 360 static void d_shrink_add(struct dentry *dentry, struct list_head *list) 361 { 362 D_FLAG_VERIFY(dentry, 0); 363 list_add(&dentry->d_lru, list); 364 dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST; 365 this_cpu_inc(nr_dentry_unused); 366 } 367 368 /* 369 * These can only be called under the global LRU lock, ie during the 370 * callback for freeing the LRU list. "isolate" removes it from the 371 * LRU lists entirely, while shrink_move moves it to the indicated 372 * private list. 373 */ 374 static void d_lru_isolate(struct dentry *dentry) 375 { 376 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST); 377 dentry->d_flags &= ~DCACHE_LRU_LIST; 378 this_cpu_dec(nr_dentry_unused); 379 list_del_init(&dentry->d_lru); 380 } 381 382 static void d_lru_shrink_move(struct dentry *dentry, struct list_head *list) 383 { 384 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST); 385 dentry->d_flags |= DCACHE_SHRINK_LIST; 386 list_move_tail(&dentry->d_lru, list); 387 } 388 389 /* 390 * dentry_lru_(add|del)_list) must be called with d_lock held. 391 */ 392 static void dentry_lru_add(struct dentry *dentry) 393 { 394 if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST))) 395 d_lru_add(dentry); 396 } 397 398 /** 399 * d_drop - drop a dentry 400 * @dentry: dentry to drop 401 * 402 * d_drop() unhashes the entry from the parent dentry hashes, so that it won't 403 * be found through a VFS lookup any more. Note that this is different from 404 * deleting the dentry - d_delete will try to mark the dentry negative if 405 * possible, giving a successful _negative_ lookup, while d_drop will 406 * just make the cache lookup fail. 407 * 408 * d_drop() is used mainly for stuff that wants to invalidate a dentry for some 409 * reason (NFS timeouts or autofs deletes). 410 * 411 * __d_drop requires dentry->d_lock. 412 */ 413 void __d_drop(struct dentry *dentry) 414 { 415 if (!d_unhashed(dentry)) { 416 struct hlist_bl_head *b; 417 /* 418 * Hashed dentries are normally on the dentry hashtable, 419 * with the exception of those newly allocated by 420 * d_obtain_alias, which are always IS_ROOT: 421 */ 422 if (unlikely(IS_ROOT(dentry))) 423 b = &dentry->d_sb->s_anon; 424 else 425 b = d_hash(dentry->d_parent, dentry->d_name.hash); 426 427 hlist_bl_lock(b); 428 __hlist_bl_del(&dentry->d_hash); 429 dentry->d_hash.pprev = NULL; 430 hlist_bl_unlock(b); 431 dentry_rcuwalk_barrier(dentry); 432 } 433 } 434 EXPORT_SYMBOL(__d_drop); 435 436 void d_drop(struct dentry *dentry) 437 { 438 spin_lock(&dentry->d_lock); 439 __d_drop(dentry); 440 spin_unlock(&dentry->d_lock); 441 } 442 EXPORT_SYMBOL(d_drop); 443 444 /* 445 * Finish off a dentry we've decided to kill. 446 * dentry->d_lock must be held, returns with it unlocked. 447 * If ref is non-zero, then decrement the refcount too. 448 * Returns dentry requiring refcount drop, or NULL if we're done. 449 */ 450 static struct dentry * 451 dentry_kill(struct dentry *dentry, int unlock_on_failure) 452 __releases(dentry->d_lock) 453 { 454 struct inode *inode; 455 struct dentry *parent = NULL; 456 bool can_free = true; 457 458 if (unlikely(dentry->d_flags & DCACHE_DENTRY_KILLED)) { 459 can_free = dentry->d_flags & DCACHE_MAY_FREE; 460 spin_unlock(&dentry->d_lock); 461 goto out; 462 } 463 464 inode = dentry->d_inode; 465 if (inode && !spin_trylock(&inode->i_lock)) { 466 relock: 467 if (unlock_on_failure) { 468 spin_unlock(&dentry->d_lock); 469 cpu_relax(); 470 } 471 return dentry; /* try again with same dentry */ 472 } 473 if (!IS_ROOT(dentry)) 474 parent = dentry->d_parent; 475 if (parent && !spin_trylock(&parent->d_lock)) { 476 if (inode) 477 spin_unlock(&inode->i_lock); 478 goto relock; 479 } 480 481 /* 482 * The dentry is now unrecoverably dead to the world. 483 */ 484 lockref_mark_dead(&dentry->d_lockref); 485 486 /* 487 * inform the fs via d_prune that this dentry is about to be 488 * unhashed and destroyed. 489 */ 490 if ((dentry->d_flags & DCACHE_OP_PRUNE) && !d_unhashed(dentry)) 491 dentry->d_op->d_prune(dentry); 492 493 if (dentry->d_flags & DCACHE_LRU_LIST) { 494 if (!(dentry->d_flags & DCACHE_SHRINK_LIST)) 495 d_lru_del(dentry); 496 } 497 /* if it was on the hash then remove it */ 498 __d_drop(dentry); 499 list_del(&dentry->d_u.d_child); 500 /* 501 * Inform d_walk() that we are no longer attached to the 502 * dentry tree 503 */ 504 dentry->d_flags |= DCACHE_DENTRY_KILLED; 505 if (parent) 506 spin_unlock(&parent->d_lock); 507 dentry_iput(dentry); 508 /* 509 * dentry_iput drops the locks, at which point nobody (except 510 * transient RCU lookups) can reach this dentry. 511 */ 512 BUG_ON((int)dentry->d_lockref.count > 0); 513 this_cpu_dec(nr_dentry); 514 if (dentry->d_op && dentry->d_op->d_release) 515 dentry->d_op->d_release(dentry); 516 517 spin_lock(&dentry->d_lock); 518 if (dentry->d_flags & DCACHE_SHRINK_LIST) { 519 dentry->d_flags |= DCACHE_MAY_FREE; 520 can_free = false; 521 } 522 spin_unlock(&dentry->d_lock); 523 out: 524 if (likely(can_free)) 525 dentry_free(dentry); 526 return parent; 527 } 528 529 /* 530 * This is dput 531 * 532 * This is complicated by the fact that we do not want to put 533 * dentries that are no longer on any hash chain on the unused 534 * list: we'd much rather just get rid of them immediately. 535 * 536 * However, that implies that we have to traverse the dentry 537 * tree upwards to the parents which might _also_ now be 538 * scheduled for deletion (it may have been only waiting for 539 * its last child to go away). 540 * 541 * This tail recursion is done by hand as we don't want to depend 542 * on the compiler to always get this right (gcc generally doesn't). 543 * Real recursion would eat up our stack space. 544 */ 545 546 /* 547 * dput - release a dentry 548 * @dentry: dentry to release 549 * 550 * Release a dentry. This will drop the usage count and if appropriate 551 * call the dentry unlink method as well as removing it from the queues and 552 * releasing its resources. If the parent dentries were scheduled for release 553 * they too may now get deleted. 554 */ 555 void dput(struct dentry *dentry) 556 { 557 if (unlikely(!dentry)) 558 return; 559 560 repeat: 561 if (lockref_put_or_lock(&dentry->d_lockref)) 562 return; 563 564 /* Unreachable? Get rid of it */ 565 if (unlikely(d_unhashed(dentry))) 566 goto kill_it; 567 568 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) { 569 if (dentry->d_op->d_delete(dentry)) 570 goto kill_it; 571 } 572 573 if (!(dentry->d_flags & DCACHE_REFERENCED)) 574 dentry->d_flags |= DCACHE_REFERENCED; 575 dentry_lru_add(dentry); 576 577 dentry->d_lockref.count--; 578 spin_unlock(&dentry->d_lock); 579 return; 580 581 kill_it: 582 dentry = dentry_kill(dentry, 1); 583 if (dentry) 584 goto repeat; 585 } 586 EXPORT_SYMBOL(dput); 587 588 /** 589 * d_invalidate - invalidate a dentry 590 * @dentry: dentry to invalidate 591 * 592 * Try to invalidate the dentry if it turns out to be 593 * possible. If there are other dentries that can be 594 * reached through this one we can't delete it and we 595 * return -EBUSY. On success we return 0. 596 * 597 * no dcache lock. 598 */ 599 600 int d_invalidate(struct dentry * dentry) 601 { 602 /* 603 * If it's already been dropped, return OK. 604 */ 605 spin_lock(&dentry->d_lock); 606 if (d_unhashed(dentry)) { 607 spin_unlock(&dentry->d_lock); 608 return 0; 609 } 610 /* 611 * Check whether to do a partial shrink_dcache 612 * to get rid of unused child entries. 613 */ 614 if (!list_empty(&dentry->d_subdirs)) { 615 spin_unlock(&dentry->d_lock); 616 shrink_dcache_parent(dentry); 617 spin_lock(&dentry->d_lock); 618 } 619 620 /* 621 * Somebody else still using it? 622 * 623 * If it's a directory, we can't drop it 624 * for fear of somebody re-populating it 625 * with children (even though dropping it 626 * would make it unreachable from the root, 627 * we might still populate it if it was a 628 * working directory or similar). 629 * We also need to leave mountpoints alone, 630 * directory or not. 631 */ 632 if (dentry->d_lockref.count > 1 && dentry->d_inode) { 633 if (S_ISDIR(dentry->d_inode->i_mode) || d_mountpoint(dentry)) { 634 spin_unlock(&dentry->d_lock); 635 return -EBUSY; 636 } 637 } 638 639 __d_drop(dentry); 640 spin_unlock(&dentry->d_lock); 641 return 0; 642 } 643 EXPORT_SYMBOL(d_invalidate); 644 645 /* This must be called with d_lock held */ 646 static inline void __dget_dlock(struct dentry *dentry) 647 { 648 dentry->d_lockref.count++; 649 } 650 651 static inline void __dget(struct dentry *dentry) 652 { 653 lockref_get(&dentry->d_lockref); 654 } 655 656 struct dentry *dget_parent(struct dentry *dentry) 657 { 658 int gotref; 659 struct dentry *ret; 660 661 /* 662 * Do optimistic parent lookup without any 663 * locking. 664 */ 665 rcu_read_lock(); 666 ret = ACCESS_ONCE(dentry->d_parent); 667 gotref = lockref_get_not_zero(&ret->d_lockref); 668 rcu_read_unlock(); 669 if (likely(gotref)) { 670 if (likely(ret == ACCESS_ONCE(dentry->d_parent))) 671 return ret; 672 dput(ret); 673 } 674 675 repeat: 676 /* 677 * Don't need rcu_dereference because we re-check it was correct under 678 * the lock. 679 */ 680 rcu_read_lock(); 681 ret = dentry->d_parent; 682 spin_lock(&ret->d_lock); 683 if (unlikely(ret != dentry->d_parent)) { 684 spin_unlock(&ret->d_lock); 685 rcu_read_unlock(); 686 goto repeat; 687 } 688 rcu_read_unlock(); 689 BUG_ON(!ret->d_lockref.count); 690 ret->d_lockref.count++; 691 spin_unlock(&ret->d_lock); 692 return ret; 693 } 694 EXPORT_SYMBOL(dget_parent); 695 696 /** 697 * d_find_alias - grab a hashed alias of inode 698 * @inode: inode in question 699 * @want_discon: flag, used by d_splice_alias, to request 700 * that only a DISCONNECTED alias be returned. 701 * 702 * If inode has a hashed alias, or is a directory and has any alias, 703 * acquire the reference to alias and return it. Otherwise return NULL. 704 * Notice that if inode is a directory there can be only one alias and 705 * it can be unhashed only if it has no children, or if it is the root 706 * of a filesystem. 707 * 708 * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer 709 * any other hashed alias over that one unless @want_discon is set, 710 * in which case only return an IS_ROOT, DCACHE_DISCONNECTED alias. 711 */ 712 static struct dentry *__d_find_alias(struct inode *inode, int want_discon) 713 { 714 struct dentry *alias, *discon_alias; 715 716 again: 717 discon_alias = NULL; 718 hlist_for_each_entry(alias, &inode->i_dentry, d_alias) { 719 spin_lock(&alias->d_lock); 720 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) { 721 if (IS_ROOT(alias) && 722 (alias->d_flags & DCACHE_DISCONNECTED)) { 723 discon_alias = alias; 724 } else if (!want_discon) { 725 __dget_dlock(alias); 726 spin_unlock(&alias->d_lock); 727 return alias; 728 } 729 } 730 spin_unlock(&alias->d_lock); 731 } 732 if (discon_alias) { 733 alias = discon_alias; 734 spin_lock(&alias->d_lock); 735 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) { 736 if (IS_ROOT(alias) && 737 (alias->d_flags & DCACHE_DISCONNECTED)) { 738 __dget_dlock(alias); 739 spin_unlock(&alias->d_lock); 740 return alias; 741 } 742 } 743 spin_unlock(&alias->d_lock); 744 goto again; 745 } 746 return NULL; 747 } 748 749 struct dentry *d_find_alias(struct inode *inode) 750 { 751 struct dentry *de = NULL; 752 753 if (!hlist_empty(&inode->i_dentry)) { 754 spin_lock(&inode->i_lock); 755 de = __d_find_alias(inode, 0); 756 spin_unlock(&inode->i_lock); 757 } 758 return de; 759 } 760 EXPORT_SYMBOL(d_find_alias); 761 762 /* 763 * Try to kill dentries associated with this inode. 764 * WARNING: you must own a reference to inode. 765 */ 766 void d_prune_aliases(struct inode *inode) 767 { 768 struct dentry *dentry; 769 restart: 770 spin_lock(&inode->i_lock); 771 hlist_for_each_entry(dentry, &inode->i_dentry, d_alias) { 772 spin_lock(&dentry->d_lock); 773 if (!dentry->d_lockref.count) { 774 /* 775 * inform the fs via d_prune that this dentry 776 * is about to be unhashed and destroyed. 777 */ 778 if ((dentry->d_flags & DCACHE_OP_PRUNE) && 779 !d_unhashed(dentry)) 780 dentry->d_op->d_prune(dentry); 781 782 __dget_dlock(dentry); 783 __d_drop(dentry); 784 spin_unlock(&dentry->d_lock); 785 spin_unlock(&inode->i_lock); 786 dput(dentry); 787 goto restart; 788 } 789 spin_unlock(&dentry->d_lock); 790 } 791 spin_unlock(&inode->i_lock); 792 } 793 EXPORT_SYMBOL(d_prune_aliases); 794 795 static void shrink_dentry_list(struct list_head *list) 796 { 797 struct dentry *dentry, *parent; 798 799 while (!list_empty(list)) { 800 dentry = list_entry(list->prev, struct dentry, d_lru); 801 spin_lock(&dentry->d_lock); 802 /* 803 * The dispose list is isolated and dentries are not accounted 804 * to the LRU here, so we can simply remove it from the list 805 * here regardless of whether it is referenced or not. 806 */ 807 d_shrink_del(dentry); 808 809 /* 810 * We found an inuse dentry which was not removed from 811 * the LRU because of laziness during lookup. Do not free it. 812 */ 813 if ((int)dentry->d_lockref.count > 0) { 814 spin_unlock(&dentry->d_lock); 815 continue; 816 } 817 818 parent = dentry_kill(dentry, 0); 819 /* 820 * If dentry_kill returns NULL, we have nothing more to do. 821 */ 822 if (!parent) 823 continue; 824 825 if (unlikely(parent == dentry)) { 826 /* 827 * trylocks have failed and d_lock has been held the 828 * whole time, so it could not have been added to any 829 * other lists. Just add it back to the shrink list. 830 */ 831 d_shrink_add(dentry, list); 832 spin_unlock(&dentry->d_lock); 833 continue; 834 } 835 /* 836 * We need to prune ancestors too. This is necessary to prevent 837 * quadratic behavior of shrink_dcache_parent(), but is also 838 * expected to be beneficial in reducing dentry cache 839 * fragmentation. 840 */ 841 dentry = parent; 842 while (dentry && !lockref_put_or_lock(&dentry->d_lockref)) 843 dentry = dentry_kill(dentry, 1); 844 } 845 } 846 847 static enum lru_status 848 dentry_lru_isolate(struct list_head *item, spinlock_t *lru_lock, void *arg) 849 { 850 struct list_head *freeable = arg; 851 struct dentry *dentry = container_of(item, struct dentry, d_lru); 852 853 854 /* 855 * we are inverting the lru lock/dentry->d_lock here, 856 * so use a trylock. If we fail to get the lock, just skip 857 * it 858 */ 859 if (!spin_trylock(&dentry->d_lock)) 860 return LRU_SKIP; 861 862 /* 863 * Referenced dentries are still in use. If they have active 864 * counts, just remove them from the LRU. Otherwise give them 865 * another pass through the LRU. 866 */ 867 if (dentry->d_lockref.count) { 868 d_lru_isolate(dentry); 869 spin_unlock(&dentry->d_lock); 870 return LRU_REMOVED; 871 } 872 873 if (dentry->d_flags & DCACHE_REFERENCED) { 874 dentry->d_flags &= ~DCACHE_REFERENCED; 875 spin_unlock(&dentry->d_lock); 876 877 /* 878 * The list move itself will be made by the common LRU code. At 879 * this point, we've dropped the dentry->d_lock but keep the 880 * lru lock. This is safe to do, since every list movement is 881 * protected by the lru lock even if both locks are held. 882 * 883 * This is guaranteed by the fact that all LRU management 884 * functions are intermediated by the LRU API calls like 885 * list_lru_add and list_lru_del. List movement in this file 886 * only ever occur through this functions or through callbacks 887 * like this one, that are called from the LRU API. 888 * 889 * The only exceptions to this are functions like 890 * shrink_dentry_list, and code that first checks for the 891 * DCACHE_SHRINK_LIST flag. Those are guaranteed to be 892 * operating only with stack provided lists after they are 893 * properly isolated from the main list. It is thus, always a 894 * local access. 895 */ 896 return LRU_ROTATE; 897 } 898 899 d_lru_shrink_move(dentry, freeable); 900 spin_unlock(&dentry->d_lock); 901 902 return LRU_REMOVED; 903 } 904 905 /** 906 * prune_dcache_sb - shrink the dcache 907 * @sb: superblock 908 * @nr_to_scan : number of entries to try to free 909 * @nid: which node to scan for freeable entities 910 * 911 * Attempt to shrink the superblock dcache LRU by @nr_to_scan entries. This is 912 * done when we need more memory an called from the superblock shrinker 913 * function. 914 * 915 * This function may fail to free any resources if all the dentries are in 916 * use. 917 */ 918 long prune_dcache_sb(struct super_block *sb, unsigned long nr_to_scan, 919 int nid) 920 { 921 LIST_HEAD(dispose); 922 long freed; 923 924 freed = list_lru_walk_node(&sb->s_dentry_lru, nid, dentry_lru_isolate, 925 &dispose, &nr_to_scan); 926 shrink_dentry_list(&dispose); 927 return freed; 928 } 929 930 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item, 931 spinlock_t *lru_lock, void *arg) 932 { 933 struct list_head *freeable = arg; 934 struct dentry *dentry = container_of(item, struct dentry, d_lru); 935 936 /* 937 * we are inverting the lru lock/dentry->d_lock here, 938 * so use a trylock. If we fail to get the lock, just skip 939 * it 940 */ 941 if (!spin_trylock(&dentry->d_lock)) 942 return LRU_SKIP; 943 944 d_lru_shrink_move(dentry, freeable); 945 spin_unlock(&dentry->d_lock); 946 947 return LRU_REMOVED; 948 } 949 950 951 /** 952 * shrink_dcache_sb - shrink dcache for a superblock 953 * @sb: superblock 954 * 955 * Shrink the dcache for the specified super block. This is used to free 956 * the dcache before unmounting a file system. 957 */ 958 void shrink_dcache_sb(struct super_block *sb) 959 { 960 long freed; 961 962 do { 963 LIST_HEAD(dispose); 964 965 freed = list_lru_walk(&sb->s_dentry_lru, 966 dentry_lru_isolate_shrink, &dispose, UINT_MAX); 967 968 this_cpu_sub(nr_dentry_unused, freed); 969 shrink_dentry_list(&dispose); 970 } while (freed > 0); 971 } 972 EXPORT_SYMBOL(shrink_dcache_sb); 973 974 /** 975 * enum d_walk_ret - action to talke during tree walk 976 * @D_WALK_CONTINUE: contrinue walk 977 * @D_WALK_QUIT: quit walk 978 * @D_WALK_NORETRY: quit when retry is needed 979 * @D_WALK_SKIP: skip this dentry and its children 980 */ 981 enum d_walk_ret { 982 D_WALK_CONTINUE, 983 D_WALK_QUIT, 984 D_WALK_NORETRY, 985 D_WALK_SKIP, 986 }; 987 988 /** 989 * d_walk - walk the dentry tree 990 * @parent: start of walk 991 * @data: data passed to @enter() and @finish() 992 * @enter: callback when first entering the dentry 993 * @finish: callback when successfully finished the walk 994 * 995 * The @enter() and @finish() callbacks are called with d_lock held. 996 */ 997 static void d_walk(struct dentry *parent, void *data, 998 enum d_walk_ret (*enter)(void *, struct dentry *), 999 void (*finish)(void *)) 1000 { 1001 struct dentry *this_parent; 1002 struct list_head *next; 1003 unsigned seq = 0; 1004 enum d_walk_ret ret; 1005 bool retry = true; 1006 1007 again: 1008 read_seqbegin_or_lock(&rename_lock, &seq); 1009 this_parent = parent; 1010 spin_lock(&this_parent->d_lock); 1011 1012 ret = enter(data, this_parent); 1013 switch (ret) { 1014 case D_WALK_CONTINUE: 1015 break; 1016 case D_WALK_QUIT: 1017 case D_WALK_SKIP: 1018 goto out_unlock; 1019 case D_WALK_NORETRY: 1020 retry = false; 1021 break; 1022 } 1023 repeat: 1024 next = this_parent->d_subdirs.next; 1025 resume: 1026 while (next != &this_parent->d_subdirs) { 1027 struct list_head *tmp = next; 1028 struct dentry *dentry = list_entry(tmp, struct dentry, d_u.d_child); 1029 next = tmp->next; 1030 1031 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED); 1032 1033 ret = enter(data, dentry); 1034 switch (ret) { 1035 case D_WALK_CONTINUE: 1036 break; 1037 case D_WALK_QUIT: 1038 spin_unlock(&dentry->d_lock); 1039 goto out_unlock; 1040 case D_WALK_NORETRY: 1041 retry = false; 1042 break; 1043 case D_WALK_SKIP: 1044 spin_unlock(&dentry->d_lock); 1045 continue; 1046 } 1047 1048 if (!list_empty(&dentry->d_subdirs)) { 1049 spin_unlock(&this_parent->d_lock); 1050 spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_); 1051 this_parent = dentry; 1052 spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_); 1053 goto repeat; 1054 } 1055 spin_unlock(&dentry->d_lock); 1056 } 1057 /* 1058 * All done at this level ... ascend and resume the search. 1059 */ 1060 if (this_parent != parent) { 1061 struct dentry *child = this_parent; 1062 this_parent = child->d_parent; 1063 1064 rcu_read_lock(); 1065 spin_unlock(&child->d_lock); 1066 spin_lock(&this_parent->d_lock); 1067 1068 /* 1069 * might go back up the wrong parent if we have had a rename 1070 * or deletion 1071 */ 1072 if (this_parent != child->d_parent || 1073 (child->d_flags & DCACHE_DENTRY_KILLED) || 1074 need_seqretry(&rename_lock, seq)) { 1075 spin_unlock(&this_parent->d_lock); 1076 rcu_read_unlock(); 1077 goto rename_retry; 1078 } 1079 rcu_read_unlock(); 1080 next = child->d_u.d_child.next; 1081 goto resume; 1082 } 1083 if (need_seqretry(&rename_lock, seq)) { 1084 spin_unlock(&this_parent->d_lock); 1085 goto rename_retry; 1086 } 1087 if (finish) 1088 finish(data); 1089 1090 out_unlock: 1091 spin_unlock(&this_parent->d_lock); 1092 done_seqretry(&rename_lock, seq); 1093 return; 1094 1095 rename_retry: 1096 if (!retry) 1097 return; 1098 seq = 1; 1099 goto again; 1100 } 1101 1102 /* 1103 * Search for at least 1 mount point in the dentry's subdirs. 1104 * We descend to the next level whenever the d_subdirs 1105 * list is non-empty and continue searching. 1106 */ 1107 1108 static enum d_walk_ret check_mount(void *data, struct dentry *dentry) 1109 { 1110 int *ret = data; 1111 if (d_mountpoint(dentry)) { 1112 *ret = 1; 1113 return D_WALK_QUIT; 1114 } 1115 return D_WALK_CONTINUE; 1116 } 1117 1118 /** 1119 * have_submounts - check for mounts over a dentry 1120 * @parent: dentry to check. 1121 * 1122 * Return true if the parent or its subdirectories contain 1123 * a mount point 1124 */ 1125 int have_submounts(struct dentry *parent) 1126 { 1127 int ret = 0; 1128 1129 d_walk(parent, &ret, check_mount, NULL); 1130 1131 return ret; 1132 } 1133 EXPORT_SYMBOL(have_submounts); 1134 1135 /* 1136 * Called by mount code to set a mountpoint and check if the mountpoint is 1137 * reachable (e.g. NFS can unhash a directory dentry and then the complete 1138 * subtree can become unreachable). 1139 * 1140 * Only one of check_submounts_and_drop() and d_set_mounted() must succeed. For 1141 * this reason take rename_lock and d_lock on dentry and ancestors. 1142 */ 1143 int d_set_mounted(struct dentry *dentry) 1144 { 1145 struct dentry *p; 1146 int ret = -ENOENT; 1147 write_seqlock(&rename_lock); 1148 for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) { 1149 /* Need exclusion wrt. check_submounts_and_drop() */ 1150 spin_lock(&p->d_lock); 1151 if (unlikely(d_unhashed(p))) { 1152 spin_unlock(&p->d_lock); 1153 goto out; 1154 } 1155 spin_unlock(&p->d_lock); 1156 } 1157 spin_lock(&dentry->d_lock); 1158 if (!d_unlinked(dentry)) { 1159 dentry->d_flags |= DCACHE_MOUNTED; 1160 ret = 0; 1161 } 1162 spin_unlock(&dentry->d_lock); 1163 out: 1164 write_sequnlock(&rename_lock); 1165 return ret; 1166 } 1167 1168 /* 1169 * Search the dentry child list of the specified parent, 1170 * and move any unused dentries to the end of the unused 1171 * list for prune_dcache(). We descend to the next level 1172 * whenever the d_subdirs list is non-empty and continue 1173 * searching. 1174 * 1175 * It returns zero iff there are no unused children, 1176 * otherwise it returns the number of children moved to 1177 * the end of the unused list. This may not be the total 1178 * number of unused children, because select_parent can 1179 * drop the lock and return early due to latency 1180 * constraints. 1181 */ 1182 1183 struct select_data { 1184 struct dentry *start; 1185 struct list_head dispose; 1186 int found; 1187 }; 1188 1189 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry) 1190 { 1191 struct select_data *data = _data; 1192 enum d_walk_ret ret = D_WALK_CONTINUE; 1193 1194 if (data->start == dentry) 1195 goto out; 1196 1197 if (dentry->d_flags & DCACHE_SHRINK_LIST) { 1198 data->found++; 1199 } else { 1200 if (dentry->d_flags & DCACHE_LRU_LIST) 1201 d_lru_del(dentry); 1202 if (!dentry->d_lockref.count) { 1203 d_shrink_add(dentry, &data->dispose); 1204 data->found++; 1205 } 1206 } 1207 /* 1208 * We can return to the caller if we have found some (this 1209 * ensures forward progress). We'll be coming back to find 1210 * the rest. 1211 */ 1212 if (!list_empty(&data->dispose)) 1213 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY; 1214 out: 1215 return ret; 1216 } 1217 1218 /** 1219 * shrink_dcache_parent - prune dcache 1220 * @parent: parent of entries to prune 1221 * 1222 * Prune the dcache to remove unused children of the parent dentry. 1223 */ 1224 void shrink_dcache_parent(struct dentry *parent) 1225 { 1226 for (;;) { 1227 struct select_data data; 1228 1229 INIT_LIST_HEAD(&data.dispose); 1230 data.start = parent; 1231 data.found = 0; 1232 1233 d_walk(parent, &data, select_collect, NULL); 1234 if (!data.found) 1235 break; 1236 1237 shrink_dentry_list(&data.dispose); 1238 cond_resched(); 1239 } 1240 } 1241 EXPORT_SYMBOL(shrink_dcache_parent); 1242 1243 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry) 1244 { 1245 /* it has busy descendents; complain about those instead */ 1246 if (!list_empty(&dentry->d_subdirs)) 1247 return D_WALK_CONTINUE; 1248 1249 /* root with refcount 1 is fine */ 1250 if (dentry == _data && dentry->d_lockref.count == 1) 1251 return D_WALK_CONTINUE; 1252 1253 printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} " 1254 " still in use (%d) [unmount of %s %s]\n", 1255 dentry, 1256 dentry->d_inode ? 1257 dentry->d_inode->i_ino : 0UL, 1258 dentry, 1259 dentry->d_lockref.count, 1260 dentry->d_sb->s_type->name, 1261 dentry->d_sb->s_id); 1262 WARN_ON(1); 1263 return D_WALK_CONTINUE; 1264 } 1265 1266 static void do_one_tree(struct dentry *dentry) 1267 { 1268 shrink_dcache_parent(dentry); 1269 d_walk(dentry, dentry, umount_check, NULL); 1270 d_drop(dentry); 1271 dput(dentry); 1272 } 1273 1274 /* 1275 * destroy the dentries attached to a superblock on unmounting 1276 */ 1277 void shrink_dcache_for_umount(struct super_block *sb) 1278 { 1279 struct dentry *dentry; 1280 1281 WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked"); 1282 1283 dentry = sb->s_root; 1284 sb->s_root = NULL; 1285 do_one_tree(dentry); 1286 1287 while (!hlist_bl_empty(&sb->s_anon)) { 1288 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_anon), struct dentry, d_hash)); 1289 do_one_tree(dentry); 1290 } 1291 } 1292 1293 static enum d_walk_ret check_and_collect(void *_data, struct dentry *dentry) 1294 { 1295 struct select_data *data = _data; 1296 1297 if (d_mountpoint(dentry)) { 1298 data->found = -EBUSY; 1299 return D_WALK_QUIT; 1300 } 1301 1302 return select_collect(_data, dentry); 1303 } 1304 1305 static void check_and_drop(void *_data) 1306 { 1307 struct select_data *data = _data; 1308 1309 if (d_mountpoint(data->start)) 1310 data->found = -EBUSY; 1311 if (!data->found) 1312 __d_drop(data->start); 1313 } 1314 1315 /** 1316 * check_submounts_and_drop - prune dcache, check for submounts and drop 1317 * 1318 * All done as a single atomic operation relative to has_unlinked_ancestor(). 1319 * Returns 0 if successfully unhashed @parent. If there were submounts then 1320 * return -EBUSY. 1321 * 1322 * @dentry: dentry to prune and drop 1323 */ 1324 int check_submounts_and_drop(struct dentry *dentry) 1325 { 1326 int ret = 0; 1327 1328 /* Negative dentries can be dropped without further checks */ 1329 if (!dentry->d_inode) { 1330 d_drop(dentry); 1331 goto out; 1332 } 1333 1334 for (;;) { 1335 struct select_data data; 1336 1337 INIT_LIST_HEAD(&data.dispose); 1338 data.start = dentry; 1339 data.found = 0; 1340 1341 d_walk(dentry, &data, check_and_collect, check_and_drop); 1342 ret = data.found; 1343 1344 if (!list_empty(&data.dispose)) 1345 shrink_dentry_list(&data.dispose); 1346 1347 if (ret <= 0) 1348 break; 1349 1350 cond_resched(); 1351 } 1352 1353 out: 1354 return ret; 1355 } 1356 EXPORT_SYMBOL(check_submounts_and_drop); 1357 1358 /** 1359 * __d_alloc - allocate a dcache entry 1360 * @sb: filesystem it will belong to 1361 * @name: qstr of the name 1362 * 1363 * Allocates a dentry. It returns %NULL if there is insufficient memory 1364 * available. On a success the dentry is returned. The name passed in is 1365 * copied and the copy passed in may be reused after this call. 1366 */ 1367 1368 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name) 1369 { 1370 struct dentry *dentry; 1371 char *dname; 1372 1373 dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL); 1374 if (!dentry) 1375 return NULL; 1376 1377 /* 1378 * We guarantee that the inline name is always NUL-terminated. 1379 * This way the memcpy() done by the name switching in rename 1380 * will still always have a NUL at the end, even if we might 1381 * be overwriting an internal NUL character 1382 */ 1383 dentry->d_iname[DNAME_INLINE_LEN-1] = 0; 1384 if (name->len > DNAME_INLINE_LEN-1) { 1385 dname = kmalloc(name->len + 1, GFP_KERNEL); 1386 if (!dname) { 1387 kmem_cache_free(dentry_cache, dentry); 1388 return NULL; 1389 } 1390 } else { 1391 dname = dentry->d_iname; 1392 } 1393 1394 dentry->d_name.len = name->len; 1395 dentry->d_name.hash = name->hash; 1396 memcpy(dname, name->name, name->len); 1397 dname[name->len] = 0; 1398 1399 /* Make sure we always see the terminating NUL character */ 1400 smp_wmb(); 1401 dentry->d_name.name = dname; 1402 1403 dentry->d_lockref.count = 1; 1404 dentry->d_flags = 0; 1405 spin_lock_init(&dentry->d_lock); 1406 seqcount_init(&dentry->d_seq); 1407 dentry->d_inode = NULL; 1408 dentry->d_parent = dentry; 1409 dentry->d_sb = sb; 1410 dentry->d_op = NULL; 1411 dentry->d_fsdata = NULL; 1412 INIT_HLIST_BL_NODE(&dentry->d_hash); 1413 INIT_LIST_HEAD(&dentry->d_lru); 1414 INIT_LIST_HEAD(&dentry->d_subdirs); 1415 INIT_HLIST_NODE(&dentry->d_alias); 1416 INIT_LIST_HEAD(&dentry->d_u.d_child); 1417 d_set_d_op(dentry, dentry->d_sb->s_d_op); 1418 1419 this_cpu_inc(nr_dentry); 1420 1421 return dentry; 1422 } 1423 1424 /** 1425 * d_alloc - allocate a dcache entry 1426 * @parent: parent of entry to allocate 1427 * @name: qstr of the name 1428 * 1429 * Allocates a dentry. It returns %NULL if there is insufficient memory 1430 * available. On a success the dentry is returned. The name passed in is 1431 * copied and the copy passed in may be reused after this call. 1432 */ 1433 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name) 1434 { 1435 struct dentry *dentry = __d_alloc(parent->d_sb, name); 1436 if (!dentry) 1437 return NULL; 1438 1439 spin_lock(&parent->d_lock); 1440 /* 1441 * don't need child lock because it is not subject 1442 * to concurrency here 1443 */ 1444 __dget_dlock(parent); 1445 dentry->d_parent = parent; 1446 list_add(&dentry->d_u.d_child, &parent->d_subdirs); 1447 spin_unlock(&parent->d_lock); 1448 1449 return dentry; 1450 } 1451 EXPORT_SYMBOL(d_alloc); 1452 1453 /** 1454 * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems) 1455 * @sb: the superblock 1456 * @name: qstr of the name 1457 * 1458 * For a filesystem that just pins its dentries in memory and never 1459 * performs lookups at all, return an unhashed IS_ROOT dentry. 1460 */ 1461 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name) 1462 { 1463 return __d_alloc(sb, name); 1464 } 1465 EXPORT_SYMBOL(d_alloc_pseudo); 1466 1467 struct dentry *d_alloc_name(struct dentry *parent, const char *name) 1468 { 1469 struct qstr q; 1470 1471 q.name = name; 1472 q.len = strlen(name); 1473 q.hash = full_name_hash(q.name, q.len); 1474 return d_alloc(parent, &q); 1475 } 1476 EXPORT_SYMBOL(d_alloc_name); 1477 1478 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op) 1479 { 1480 WARN_ON_ONCE(dentry->d_op); 1481 WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH | 1482 DCACHE_OP_COMPARE | 1483 DCACHE_OP_REVALIDATE | 1484 DCACHE_OP_WEAK_REVALIDATE | 1485 DCACHE_OP_DELETE )); 1486 dentry->d_op = op; 1487 if (!op) 1488 return; 1489 if (op->d_hash) 1490 dentry->d_flags |= DCACHE_OP_HASH; 1491 if (op->d_compare) 1492 dentry->d_flags |= DCACHE_OP_COMPARE; 1493 if (op->d_revalidate) 1494 dentry->d_flags |= DCACHE_OP_REVALIDATE; 1495 if (op->d_weak_revalidate) 1496 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE; 1497 if (op->d_delete) 1498 dentry->d_flags |= DCACHE_OP_DELETE; 1499 if (op->d_prune) 1500 dentry->d_flags |= DCACHE_OP_PRUNE; 1501 1502 } 1503 EXPORT_SYMBOL(d_set_d_op); 1504 1505 static unsigned d_flags_for_inode(struct inode *inode) 1506 { 1507 unsigned add_flags = DCACHE_FILE_TYPE; 1508 1509 if (!inode) 1510 return DCACHE_MISS_TYPE; 1511 1512 if (S_ISDIR(inode->i_mode)) { 1513 add_flags = DCACHE_DIRECTORY_TYPE; 1514 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) { 1515 if (unlikely(!inode->i_op->lookup)) 1516 add_flags = DCACHE_AUTODIR_TYPE; 1517 else 1518 inode->i_opflags |= IOP_LOOKUP; 1519 } 1520 } else if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) { 1521 if (unlikely(inode->i_op->follow_link)) 1522 add_flags = DCACHE_SYMLINK_TYPE; 1523 else 1524 inode->i_opflags |= IOP_NOFOLLOW; 1525 } 1526 1527 if (unlikely(IS_AUTOMOUNT(inode))) 1528 add_flags |= DCACHE_NEED_AUTOMOUNT; 1529 return add_flags; 1530 } 1531 1532 static void __d_instantiate(struct dentry *dentry, struct inode *inode) 1533 { 1534 unsigned add_flags = d_flags_for_inode(inode); 1535 1536 spin_lock(&dentry->d_lock); 1537 __d_set_type(dentry, add_flags); 1538 if (inode) 1539 hlist_add_head(&dentry->d_alias, &inode->i_dentry); 1540 dentry->d_inode = inode; 1541 dentry_rcuwalk_barrier(dentry); 1542 spin_unlock(&dentry->d_lock); 1543 fsnotify_d_instantiate(dentry, inode); 1544 } 1545 1546 /** 1547 * d_instantiate - fill in inode information for a dentry 1548 * @entry: dentry to complete 1549 * @inode: inode to attach to this dentry 1550 * 1551 * Fill in inode information in the entry. 1552 * 1553 * This turns negative dentries into productive full members 1554 * of society. 1555 * 1556 * NOTE! This assumes that the inode count has been incremented 1557 * (or otherwise set) by the caller to indicate that it is now 1558 * in use by the dcache. 1559 */ 1560 1561 void d_instantiate(struct dentry *entry, struct inode * inode) 1562 { 1563 BUG_ON(!hlist_unhashed(&entry->d_alias)); 1564 if (inode) 1565 spin_lock(&inode->i_lock); 1566 __d_instantiate(entry, inode); 1567 if (inode) 1568 spin_unlock(&inode->i_lock); 1569 security_d_instantiate(entry, inode); 1570 } 1571 EXPORT_SYMBOL(d_instantiate); 1572 1573 /** 1574 * d_instantiate_unique - instantiate a non-aliased dentry 1575 * @entry: dentry to instantiate 1576 * @inode: inode to attach to this dentry 1577 * 1578 * Fill in inode information in the entry. On success, it returns NULL. 1579 * If an unhashed alias of "entry" already exists, then we return the 1580 * aliased dentry instead and drop one reference to inode. 1581 * 1582 * Note that in order to avoid conflicts with rename() etc, the caller 1583 * had better be holding the parent directory semaphore. 1584 * 1585 * This also assumes that the inode count has been incremented 1586 * (or otherwise set) by the caller to indicate that it is now 1587 * in use by the dcache. 1588 */ 1589 static struct dentry *__d_instantiate_unique(struct dentry *entry, 1590 struct inode *inode) 1591 { 1592 struct dentry *alias; 1593 int len = entry->d_name.len; 1594 const char *name = entry->d_name.name; 1595 unsigned int hash = entry->d_name.hash; 1596 1597 if (!inode) { 1598 __d_instantiate(entry, NULL); 1599 return NULL; 1600 } 1601 1602 hlist_for_each_entry(alias, &inode->i_dentry, d_alias) { 1603 /* 1604 * Don't need alias->d_lock here, because aliases with 1605 * d_parent == entry->d_parent are not subject to name or 1606 * parent changes, because the parent inode i_mutex is held. 1607 */ 1608 if (alias->d_name.hash != hash) 1609 continue; 1610 if (alias->d_parent != entry->d_parent) 1611 continue; 1612 if (alias->d_name.len != len) 1613 continue; 1614 if (dentry_cmp(alias, name, len)) 1615 continue; 1616 __dget(alias); 1617 return alias; 1618 } 1619 1620 __d_instantiate(entry, inode); 1621 return NULL; 1622 } 1623 1624 struct dentry *d_instantiate_unique(struct dentry *entry, struct inode *inode) 1625 { 1626 struct dentry *result; 1627 1628 BUG_ON(!hlist_unhashed(&entry->d_alias)); 1629 1630 if (inode) 1631 spin_lock(&inode->i_lock); 1632 result = __d_instantiate_unique(entry, inode); 1633 if (inode) 1634 spin_unlock(&inode->i_lock); 1635 1636 if (!result) { 1637 security_d_instantiate(entry, inode); 1638 return NULL; 1639 } 1640 1641 BUG_ON(!d_unhashed(result)); 1642 iput(inode); 1643 return result; 1644 } 1645 1646 EXPORT_SYMBOL(d_instantiate_unique); 1647 1648 /** 1649 * d_instantiate_no_diralias - instantiate a non-aliased dentry 1650 * @entry: dentry to complete 1651 * @inode: inode to attach to this dentry 1652 * 1653 * Fill in inode information in the entry. If a directory alias is found, then 1654 * return an error (and drop inode). Together with d_materialise_unique() this 1655 * guarantees that a directory inode may never have more than one alias. 1656 */ 1657 int d_instantiate_no_diralias(struct dentry *entry, struct inode *inode) 1658 { 1659 BUG_ON(!hlist_unhashed(&entry->d_alias)); 1660 1661 spin_lock(&inode->i_lock); 1662 if (S_ISDIR(inode->i_mode) && !hlist_empty(&inode->i_dentry)) { 1663 spin_unlock(&inode->i_lock); 1664 iput(inode); 1665 return -EBUSY; 1666 } 1667 __d_instantiate(entry, inode); 1668 spin_unlock(&inode->i_lock); 1669 security_d_instantiate(entry, inode); 1670 1671 return 0; 1672 } 1673 EXPORT_SYMBOL(d_instantiate_no_diralias); 1674 1675 struct dentry *d_make_root(struct inode *root_inode) 1676 { 1677 struct dentry *res = NULL; 1678 1679 if (root_inode) { 1680 static const struct qstr name = QSTR_INIT("/", 1); 1681 1682 res = __d_alloc(root_inode->i_sb, &name); 1683 if (res) 1684 d_instantiate(res, root_inode); 1685 else 1686 iput(root_inode); 1687 } 1688 return res; 1689 } 1690 EXPORT_SYMBOL(d_make_root); 1691 1692 static struct dentry * __d_find_any_alias(struct inode *inode) 1693 { 1694 struct dentry *alias; 1695 1696 if (hlist_empty(&inode->i_dentry)) 1697 return NULL; 1698 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_alias); 1699 __dget(alias); 1700 return alias; 1701 } 1702 1703 /** 1704 * d_find_any_alias - find any alias for a given inode 1705 * @inode: inode to find an alias for 1706 * 1707 * If any aliases exist for the given inode, take and return a 1708 * reference for one of them. If no aliases exist, return %NULL. 1709 */ 1710 struct dentry *d_find_any_alias(struct inode *inode) 1711 { 1712 struct dentry *de; 1713 1714 spin_lock(&inode->i_lock); 1715 de = __d_find_any_alias(inode); 1716 spin_unlock(&inode->i_lock); 1717 return de; 1718 } 1719 EXPORT_SYMBOL(d_find_any_alias); 1720 1721 /** 1722 * d_obtain_alias - find or allocate a dentry for a given inode 1723 * @inode: inode to allocate the dentry for 1724 * 1725 * Obtain a dentry for an inode resulting from NFS filehandle conversion or 1726 * similar open by handle operations. The returned dentry may be anonymous, 1727 * or may have a full name (if the inode was already in the cache). 1728 * 1729 * When called on a directory inode, we must ensure that the inode only ever 1730 * has one dentry. If a dentry is found, that is returned instead of 1731 * allocating a new one. 1732 * 1733 * On successful return, the reference to the inode has been transferred 1734 * to the dentry. In case of an error the reference on the inode is released. 1735 * To make it easier to use in export operations a %NULL or IS_ERR inode may 1736 * be passed in and will be the error will be propagate to the return value, 1737 * with a %NULL @inode replaced by ERR_PTR(-ESTALE). 1738 */ 1739 struct dentry *d_obtain_alias(struct inode *inode) 1740 { 1741 static const struct qstr anonstring = QSTR_INIT("/", 1); 1742 struct dentry *tmp; 1743 struct dentry *res; 1744 unsigned add_flags; 1745 1746 if (!inode) 1747 return ERR_PTR(-ESTALE); 1748 if (IS_ERR(inode)) 1749 return ERR_CAST(inode); 1750 1751 res = d_find_any_alias(inode); 1752 if (res) 1753 goto out_iput; 1754 1755 tmp = __d_alloc(inode->i_sb, &anonstring); 1756 if (!tmp) { 1757 res = ERR_PTR(-ENOMEM); 1758 goto out_iput; 1759 } 1760 1761 spin_lock(&inode->i_lock); 1762 res = __d_find_any_alias(inode); 1763 if (res) { 1764 spin_unlock(&inode->i_lock); 1765 dput(tmp); 1766 goto out_iput; 1767 } 1768 1769 /* attach a disconnected dentry */ 1770 add_flags = d_flags_for_inode(inode) | DCACHE_DISCONNECTED; 1771 1772 spin_lock(&tmp->d_lock); 1773 tmp->d_inode = inode; 1774 tmp->d_flags |= add_flags; 1775 hlist_add_head(&tmp->d_alias, &inode->i_dentry); 1776 hlist_bl_lock(&tmp->d_sb->s_anon); 1777 hlist_bl_add_head(&tmp->d_hash, &tmp->d_sb->s_anon); 1778 hlist_bl_unlock(&tmp->d_sb->s_anon); 1779 spin_unlock(&tmp->d_lock); 1780 spin_unlock(&inode->i_lock); 1781 security_d_instantiate(tmp, inode); 1782 1783 return tmp; 1784 1785 out_iput: 1786 if (res && !IS_ERR(res)) 1787 security_d_instantiate(res, inode); 1788 iput(inode); 1789 return res; 1790 } 1791 EXPORT_SYMBOL(d_obtain_alias); 1792 1793 /** 1794 * d_splice_alias - splice a disconnected dentry into the tree if one exists 1795 * @inode: the inode which may have a disconnected dentry 1796 * @dentry: a negative dentry which we want to point to the inode. 1797 * 1798 * If inode is a directory and has a 'disconnected' dentry (i.e. IS_ROOT and 1799 * DCACHE_DISCONNECTED), then d_move that in place of the given dentry 1800 * and return it, else simply d_add the inode to the dentry and return NULL. 1801 * 1802 * This is needed in the lookup routine of any filesystem that is exportable 1803 * (via knfsd) so that we can build dcache paths to directories effectively. 1804 * 1805 * If a dentry was found and moved, then it is returned. Otherwise NULL 1806 * is returned. This matches the expected return value of ->lookup. 1807 * 1808 * Cluster filesystems may call this function with a negative, hashed dentry. 1809 * In that case, we know that the inode will be a regular file, and also this 1810 * will only occur during atomic_open. So we need to check for the dentry 1811 * being already hashed only in the final case. 1812 */ 1813 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry) 1814 { 1815 struct dentry *new = NULL; 1816 1817 if (IS_ERR(inode)) 1818 return ERR_CAST(inode); 1819 1820 if (inode && S_ISDIR(inode->i_mode)) { 1821 spin_lock(&inode->i_lock); 1822 new = __d_find_alias(inode, 1); 1823 if (new) { 1824 BUG_ON(!(new->d_flags & DCACHE_DISCONNECTED)); 1825 spin_unlock(&inode->i_lock); 1826 security_d_instantiate(new, inode); 1827 d_move(new, dentry); 1828 iput(inode); 1829 } else { 1830 /* already taking inode->i_lock, so d_add() by hand */ 1831 __d_instantiate(dentry, inode); 1832 spin_unlock(&inode->i_lock); 1833 security_d_instantiate(dentry, inode); 1834 d_rehash(dentry); 1835 } 1836 } else { 1837 d_instantiate(dentry, inode); 1838 if (d_unhashed(dentry)) 1839 d_rehash(dentry); 1840 } 1841 return new; 1842 } 1843 EXPORT_SYMBOL(d_splice_alias); 1844 1845 /** 1846 * d_add_ci - lookup or allocate new dentry with case-exact name 1847 * @inode: the inode case-insensitive lookup has found 1848 * @dentry: the negative dentry that was passed to the parent's lookup func 1849 * @name: the case-exact name to be associated with the returned dentry 1850 * 1851 * This is to avoid filling the dcache with case-insensitive names to the 1852 * same inode, only the actual correct case is stored in the dcache for 1853 * case-insensitive filesystems. 1854 * 1855 * For a case-insensitive lookup match and if the the case-exact dentry 1856 * already exists in in the dcache, use it and return it. 1857 * 1858 * If no entry exists with the exact case name, allocate new dentry with 1859 * the exact case, and return the spliced entry. 1860 */ 1861 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode, 1862 struct qstr *name) 1863 { 1864 struct dentry *found; 1865 struct dentry *new; 1866 1867 /* 1868 * First check if a dentry matching the name already exists, 1869 * if not go ahead and create it now. 1870 */ 1871 found = d_hash_and_lookup(dentry->d_parent, name); 1872 if (unlikely(IS_ERR(found))) 1873 goto err_out; 1874 if (!found) { 1875 new = d_alloc(dentry->d_parent, name); 1876 if (!new) { 1877 found = ERR_PTR(-ENOMEM); 1878 goto err_out; 1879 } 1880 1881 found = d_splice_alias(inode, new); 1882 if (found) { 1883 dput(new); 1884 return found; 1885 } 1886 return new; 1887 } 1888 1889 /* 1890 * If a matching dentry exists, and it's not negative use it. 1891 * 1892 * Decrement the reference count to balance the iget() done 1893 * earlier on. 1894 */ 1895 if (found->d_inode) { 1896 if (unlikely(found->d_inode != inode)) { 1897 /* This can't happen because bad inodes are unhashed. */ 1898 BUG_ON(!is_bad_inode(inode)); 1899 BUG_ON(!is_bad_inode(found->d_inode)); 1900 } 1901 iput(inode); 1902 return found; 1903 } 1904 1905 /* 1906 * Negative dentry: instantiate it unless the inode is a directory and 1907 * already has a dentry. 1908 */ 1909 new = d_splice_alias(inode, found); 1910 if (new) { 1911 dput(found); 1912 found = new; 1913 } 1914 return found; 1915 1916 err_out: 1917 iput(inode); 1918 return found; 1919 } 1920 EXPORT_SYMBOL(d_add_ci); 1921 1922 /* 1923 * Do the slow-case of the dentry name compare. 1924 * 1925 * Unlike the dentry_cmp() function, we need to atomically 1926 * load the name and length information, so that the 1927 * filesystem can rely on them, and can use the 'name' and 1928 * 'len' information without worrying about walking off the 1929 * end of memory etc. 1930 * 1931 * Thus the read_seqcount_retry() and the "duplicate" info 1932 * in arguments (the low-level filesystem should not look 1933 * at the dentry inode or name contents directly, since 1934 * rename can change them while we're in RCU mode). 1935 */ 1936 enum slow_d_compare { 1937 D_COMP_OK, 1938 D_COMP_NOMATCH, 1939 D_COMP_SEQRETRY, 1940 }; 1941 1942 static noinline enum slow_d_compare slow_dentry_cmp( 1943 const struct dentry *parent, 1944 struct dentry *dentry, 1945 unsigned int seq, 1946 const struct qstr *name) 1947 { 1948 int tlen = dentry->d_name.len; 1949 const char *tname = dentry->d_name.name; 1950 1951 if (read_seqcount_retry(&dentry->d_seq, seq)) { 1952 cpu_relax(); 1953 return D_COMP_SEQRETRY; 1954 } 1955 if (parent->d_op->d_compare(parent, dentry, tlen, tname, name)) 1956 return D_COMP_NOMATCH; 1957 return D_COMP_OK; 1958 } 1959 1960 /** 1961 * __d_lookup_rcu - search for a dentry (racy, store-free) 1962 * @parent: parent dentry 1963 * @name: qstr of name we wish to find 1964 * @seqp: returns d_seq value at the point where the dentry was found 1965 * Returns: dentry, or NULL 1966 * 1967 * __d_lookup_rcu is the dcache lookup function for rcu-walk name 1968 * resolution (store-free path walking) design described in 1969 * Documentation/filesystems/path-lookup.txt. 1970 * 1971 * This is not to be used outside core vfs. 1972 * 1973 * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock 1974 * held, and rcu_read_lock held. The returned dentry must not be stored into 1975 * without taking d_lock and checking d_seq sequence count against @seq 1976 * returned here. 1977 * 1978 * A refcount may be taken on the found dentry with the d_rcu_to_refcount 1979 * function. 1980 * 1981 * Alternatively, __d_lookup_rcu may be called again to look up the child of 1982 * the returned dentry, so long as its parent's seqlock is checked after the 1983 * child is looked up. Thus, an interlocking stepping of sequence lock checks 1984 * is formed, giving integrity down the path walk. 1985 * 1986 * NOTE! The caller *has* to check the resulting dentry against the sequence 1987 * number we've returned before using any of the resulting dentry state! 1988 */ 1989 struct dentry *__d_lookup_rcu(const struct dentry *parent, 1990 const struct qstr *name, 1991 unsigned *seqp) 1992 { 1993 u64 hashlen = name->hash_len; 1994 const unsigned char *str = name->name; 1995 struct hlist_bl_head *b = d_hash(parent, hashlen_hash(hashlen)); 1996 struct hlist_bl_node *node; 1997 struct dentry *dentry; 1998 1999 /* 2000 * Note: There is significant duplication with __d_lookup_rcu which is 2001 * required to prevent single threaded performance regressions 2002 * especially on architectures where smp_rmb (in seqcounts) are costly. 2003 * Keep the two functions in sync. 2004 */ 2005 2006 /* 2007 * The hash list is protected using RCU. 2008 * 2009 * Carefully use d_seq when comparing a candidate dentry, to avoid 2010 * races with d_move(). 2011 * 2012 * It is possible that concurrent renames can mess up our list 2013 * walk here and result in missing our dentry, resulting in the 2014 * false-negative result. d_lookup() protects against concurrent 2015 * renames using rename_lock seqlock. 2016 * 2017 * See Documentation/filesystems/path-lookup.txt for more details. 2018 */ 2019 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) { 2020 unsigned seq; 2021 2022 seqretry: 2023 /* 2024 * The dentry sequence count protects us from concurrent 2025 * renames, and thus protects parent and name fields. 2026 * 2027 * The caller must perform a seqcount check in order 2028 * to do anything useful with the returned dentry. 2029 * 2030 * NOTE! We do a "raw" seqcount_begin here. That means that 2031 * we don't wait for the sequence count to stabilize if it 2032 * is in the middle of a sequence change. If we do the slow 2033 * dentry compare, we will do seqretries until it is stable, 2034 * and if we end up with a successful lookup, we actually 2035 * want to exit RCU lookup anyway. 2036 */ 2037 seq = raw_seqcount_begin(&dentry->d_seq); 2038 if (dentry->d_parent != parent) 2039 continue; 2040 if (d_unhashed(dentry)) 2041 continue; 2042 2043 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) { 2044 if (dentry->d_name.hash != hashlen_hash(hashlen)) 2045 continue; 2046 *seqp = seq; 2047 switch (slow_dentry_cmp(parent, dentry, seq, name)) { 2048 case D_COMP_OK: 2049 return dentry; 2050 case D_COMP_NOMATCH: 2051 continue; 2052 default: 2053 goto seqretry; 2054 } 2055 } 2056 2057 if (dentry->d_name.hash_len != hashlen) 2058 continue; 2059 *seqp = seq; 2060 if (!dentry_cmp(dentry, str, hashlen_len(hashlen))) 2061 return dentry; 2062 } 2063 return NULL; 2064 } 2065 2066 /** 2067 * d_lookup - search for a dentry 2068 * @parent: parent dentry 2069 * @name: qstr of name we wish to find 2070 * Returns: dentry, or NULL 2071 * 2072 * d_lookup searches the children of the parent dentry for the name in 2073 * question. If the dentry is found its reference count is incremented and the 2074 * dentry is returned. The caller must use dput to free the entry when it has 2075 * finished using it. %NULL is returned if the dentry does not exist. 2076 */ 2077 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name) 2078 { 2079 struct dentry *dentry; 2080 unsigned seq; 2081 2082 do { 2083 seq = read_seqbegin(&rename_lock); 2084 dentry = __d_lookup(parent, name); 2085 if (dentry) 2086 break; 2087 } while (read_seqretry(&rename_lock, seq)); 2088 return dentry; 2089 } 2090 EXPORT_SYMBOL(d_lookup); 2091 2092 /** 2093 * __d_lookup - search for a dentry (racy) 2094 * @parent: parent dentry 2095 * @name: qstr of name we wish to find 2096 * Returns: dentry, or NULL 2097 * 2098 * __d_lookup is like d_lookup, however it may (rarely) return a 2099 * false-negative result due to unrelated rename activity. 2100 * 2101 * __d_lookup is slightly faster by avoiding rename_lock read seqlock, 2102 * however it must be used carefully, eg. with a following d_lookup in 2103 * the case of failure. 2104 * 2105 * __d_lookup callers must be commented. 2106 */ 2107 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name) 2108 { 2109 unsigned int len = name->len; 2110 unsigned int hash = name->hash; 2111 const unsigned char *str = name->name; 2112 struct hlist_bl_head *b = d_hash(parent, hash); 2113 struct hlist_bl_node *node; 2114 struct dentry *found = NULL; 2115 struct dentry *dentry; 2116 2117 /* 2118 * Note: There is significant duplication with __d_lookup_rcu which is 2119 * required to prevent single threaded performance regressions 2120 * especially on architectures where smp_rmb (in seqcounts) are costly. 2121 * Keep the two functions in sync. 2122 */ 2123 2124 /* 2125 * The hash list is protected using RCU. 2126 * 2127 * Take d_lock when comparing a candidate dentry, to avoid races 2128 * with d_move(). 2129 * 2130 * It is possible that concurrent renames can mess up our list 2131 * walk here and result in missing our dentry, resulting in the 2132 * false-negative result. d_lookup() protects against concurrent 2133 * renames using rename_lock seqlock. 2134 * 2135 * See Documentation/filesystems/path-lookup.txt for more details. 2136 */ 2137 rcu_read_lock(); 2138 2139 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) { 2140 2141 if (dentry->d_name.hash != hash) 2142 continue; 2143 2144 spin_lock(&dentry->d_lock); 2145 if (dentry->d_parent != parent) 2146 goto next; 2147 if (d_unhashed(dentry)) 2148 goto next; 2149 2150 /* 2151 * It is safe to compare names since d_move() cannot 2152 * change the qstr (protected by d_lock). 2153 */ 2154 if (parent->d_flags & DCACHE_OP_COMPARE) { 2155 int tlen = dentry->d_name.len; 2156 const char *tname = dentry->d_name.name; 2157 if (parent->d_op->d_compare(parent, dentry, tlen, tname, name)) 2158 goto next; 2159 } else { 2160 if (dentry->d_name.len != len) 2161 goto next; 2162 if (dentry_cmp(dentry, str, len)) 2163 goto next; 2164 } 2165 2166 dentry->d_lockref.count++; 2167 found = dentry; 2168 spin_unlock(&dentry->d_lock); 2169 break; 2170 next: 2171 spin_unlock(&dentry->d_lock); 2172 } 2173 rcu_read_unlock(); 2174 2175 return found; 2176 } 2177 2178 /** 2179 * d_hash_and_lookup - hash the qstr then search for a dentry 2180 * @dir: Directory to search in 2181 * @name: qstr of name we wish to find 2182 * 2183 * On lookup failure NULL is returned; on bad name - ERR_PTR(-error) 2184 */ 2185 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name) 2186 { 2187 /* 2188 * Check for a fs-specific hash function. Note that we must 2189 * calculate the standard hash first, as the d_op->d_hash() 2190 * routine may choose to leave the hash value unchanged. 2191 */ 2192 name->hash = full_name_hash(name->name, name->len); 2193 if (dir->d_flags & DCACHE_OP_HASH) { 2194 int err = dir->d_op->d_hash(dir, name); 2195 if (unlikely(err < 0)) 2196 return ERR_PTR(err); 2197 } 2198 return d_lookup(dir, name); 2199 } 2200 EXPORT_SYMBOL(d_hash_and_lookup); 2201 2202 /** 2203 * d_validate - verify dentry provided from insecure source (deprecated) 2204 * @dentry: The dentry alleged to be valid child of @dparent 2205 * @dparent: The parent dentry (known to be valid) 2206 * 2207 * An insecure source has sent us a dentry, here we verify it and dget() it. 2208 * This is used by ncpfs in its readdir implementation. 2209 * Zero is returned in the dentry is invalid. 2210 * 2211 * This function is slow for big directories, and deprecated, do not use it. 2212 */ 2213 int d_validate(struct dentry *dentry, struct dentry *dparent) 2214 { 2215 struct dentry *child; 2216 2217 spin_lock(&dparent->d_lock); 2218 list_for_each_entry(child, &dparent->d_subdirs, d_u.d_child) { 2219 if (dentry == child) { 2220 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED); 2221 __dget_dlock(dentry); 2222 spin_unlock(&dentry->d_lock); 2223 spin_unlock(&dparent->d_lock); 2224 return 1; 2225 } 2226 } 2227 spin_unlock(&dparent->d_lock); 2228 2229 return 0; 2230 } 2231 EXPORT_SYMBOL(d_validate); 2232 2233 /* 2234 * When a file is deleted, we have two options: 2235 * - turn this dentry into a negative dentry 2236 * - unhash this dentry and free it. 2237 * 2238 * Usually, we want to just turn this into 2239 * a negative dentry, but if anybody else is 2240 * currently using the dentry or the inode 2241 * we can't do that and we fall back on removing 2242 * it from the hash queues and waiting for 2243 * it to be deleted later when it has no users 2244 */ 2245 2246 /** 2247 * d_delete - delete a dentry 2248 * @dentry: The dentry to delete 2249 * 2250 * Turn the dentry into a negative dentry if possible, otherwise 2251 * remove it from the hash queues so it can be deleted later 2252 */ 2253 2254 void d_delete(struct dentry * dentry) 2255 { 2256 struct inode *inode; 2257 int isdir = 0; 2258 /* 2259 * Are we the only user? 2260 */ 2261 again: 2262 spin_lock(&dentry->d_lock); 2263 inode = dentry->d_inode; 2264 isdir = S_ISDIR(inode->i_mode); 2265 if (dentry->d_lockref.count == 1) { 2266 if (!spin_trylock(&inode->i_lock)) { 2267 spin_unlock(&dentry->d_lock); 2268 cpu_relax(); 2269 goto again; 2270 } 2271 dentry->d_flags &= ~DCACHE_CANT_MOUNT; 2272 dentry_unlink_inode(dentry); 2273 fsnotify_nameremove(dentry, isdir); 2274 return; 2275 } 2276 2277 if (!d_unhashed(dentry)) 2278 __d_drop(dentry); 2279 2280 spin_unlock(&dentry->d_lock); 2281 2282 fsnotify_nameremove(dentry, isdir); 2283 } 2284 EXPORT_SYMBOL(d_delete); 2285 2286 static void __d_rehash(struct dentry * entry, struct hlist_bl_head *b) 2287 { 2288 BUG_ON(!d_unhashed(entry)); 2289 hlist_bl_lock(b); 2290 entry->d_flags |= DCACHE_RCUACCESS; 2291 hlist_bl_add_head_rcu(&entry->d_hash, b); 2292 hlist_bl_unlock(b); 2293 } 2294 2295 static void _d_rehash(struct dentry * entry) 2296 { 2297 __d_rehash(entry, d_hash(entry->d_parent, entry->d_name.hash)); 2298 } 2299 2300 /** 2301 * d_rehash - add an entry back to the hash 2302 * @entry: dentry to add to the hash 2303 * 2304 * Adds a dentry to the hash according to its name. 2305 */ 2306 2307 void d_rehash(struct dentry * entry) 2308 { 2309 spin_lock(&entry->d_lock); 2310 _d_rehash(entry); 2311 spin_unlock(&entry->d_lock); 2312 } 2313 EXPORT_SYMBOL(d_rehash); 2314 2315 /** 2316 * dentry_update_name_case - update case insensitive dentry with a new name 2317 * @dentry: dentry to be updated 2318 * @name: new name 2319 * 2320 * Update a case insensitive dentry with new case of name. 2321 * 2322 * dentry must have been returned by d_lookup with name @name. Old and new 2323 * name lengths must match (ie. no d_compare which allows mismatched name 2324 * lengths). 2325 * 2326 * Parent inode i_mutex must be held over d_lookup and into this call (to 2327 * keep renames and concurrent inserts, and readdir(2) away). 2328 */ 2329 void dentry_update_name_case(struct dentry *dentry, struct qstr *name) 2330 { 2331 BUG_ON(!mutex_is_locked(&dentry->d_parent->d_inode->i_mutex)); 2332 BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */ 2333 2334 spin_lock(&dentry->d_lock); 2335 write_seqcount_begin(&dentry->d_seq); 2336 memcpy((unsigned char *)dentry->d_name.name, name->name, name->len); 2337 write_seqcount_end(&dentry->d_seq); 2338 spin_unlock(&dentry->d_lock); 2339 } 2340 EXPORT_SYMBOL(dentry_update_name_case); 2341 2342 static void switch_names(struct dentry *dentry, struct dentry *target) 2343 { 2344 if (dname_external(target)) { 2345 if (dname_external(dentry)) { 2346 /* 2347 * Both external: swap the pointers 2348 */ 2349 swap(target->d_name.name, dentry->d_name.name); 2350 } else { 2351 /* 2352 * dentry:internal, target:external. Steal target's 2353 * storage and make target internal. 2354 */ 2355 memcpy(target->d_iname, dentry->d_name.name, 2356 dentry->d_name.len + 1); 2357 dentry->d_name.name = target->d_name.name; 2358 target->d_name.name = target->d_iname; 2359 } 2360 } else { 2361 if (dname_external(dentry)) { 2362 /* 2363 * dentry:external, target:internal. Give dentry's 2364 * storage to target and make dentry internal 2365 */ 2366 memcpy(dentry->d_iname, target->d_name.name, 2367 target->d_name.len + 1); 2368 target->d_name.name = dentry->d_name.name; 2369 dentry->d_name.name = dentry->d_iname; 2370 } else { 2371 /* 2372 * Both are internal. 2373 */ 2374 unsigned int i; 2375 BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long))); 2376 for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) { 2377 swap(((long *) &dentry->d_iname)[i], 2378 ((long *) &target->d_iname)[i]); 2379 } 2380 } 2381 } 2382 swap(dentry->d_name.len, target->d_name.len); 2383 } 2384 2385 static void dentry_lock_for_move(struct dentry *dentry, struct dentry *target) 2386 { 2387 /* 2388 * XXXX: do we really need to take target->d_lock? 2389 */ 2390 if (IS_ROOT(dentry) || dentry->d_parent == target->d_parent) 2391 spin_lock(&target->d_parent->d_lock); 2392 else { 2393 if (d_ancestor(dentry->d_parent, target->d_parent)) { 2394 spin_lock(&dentry->d_parent->d_lock); 2395 spin_lock_nested(&target->d_parent->d_lock, 2396 DENTRY_D_LOCK_NESTED); 2397 } else { 2398 spin_lock(&target->d_parent->d_lock); 2399 spin_lock_nested(&dentry->d_parent->d_lock, 2400 DENTRY_D_LOCK_NESTED); 2401 } 2402 } 2403 if (target < dentry) { 2404 spin_lock_nested(&target->d_lock, 2); 2405 spin_lock_nested(&dentry->d_lock, 3); 2406 } else { 2407 spin_lock_nested(&dentry->d_lock, 2); 2408 spin_lock_nested(&target->d_lock, 3); 2409 } 2410 } 2411 2412 static void dentry_unlock_parents_for_move(struct dentry *dentry, 2413 struct dentry *target) 2414 { 2415 if (target->d_parent != dentry->d_parent) 2416 spin_unlock(&dentry->d_parent->d_lock); 2417 if (target->d_parent != target) 2418 spin_unlock(&target->d_parent->d_lock); 2419 } 2420 2421 /* 2422 * When switching names, the actual string doesn't strictly have to 2423 * be preserved in the target - because we're dropping the target 2424 * anyway. As such, we can just do a simple memcpy() to copy over 2425 * the new name before we switch. 2426 * 2427 * Note that we have to be a lot more careful about getting the hash 2428 * switched - we have to switch the hash value properly even if it 2429 * then no longer matches the actual (corrupted) string of the target. 2430 * The hash value has to match the hash queue that the dentry is on.. 2431 */ 2432 /* 2433 * __d_move - move a dentry 2434 * @dentry: entry to move 2435 * @target: new dentry 2436 * @exchange: exchange the two dentries 2437 * 2438 * Update the dcache to reflect the move of a file name. Negative 2439 * dcache entries should not be moved in this way. Caller must hold 2440 * rename_lock, the i_mutex of the source and target directories, 2441 * and the sb->s_vfs_rename_mutex if they differ. See lock_rename(). 2442 */ 2443 static void __d_move(struct dentry *dentry, struct dentry *target, 2444 bool exchange) 2445 { 2446 if (!dentry->d_inode) 2447 printk(KERN_WARNING "VFS: moving negative dcache entry\n"); 2448 2449 BUG_ON(d_ancestor(dentry, target)); 2450 BUG_ON(d_ancestor(target, dentry)); 2451 2452 dentry_lock_for_move(dentry, target); 2453 2454 write_seqcount_begin(&dentry->d_seq); 2455 write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED); 2456 2457 /* __d_drop does write_seqcount_barrier, but they're OK to nest. */ 2458 2459 /* 2460 * Move the dentry to the target hash queue. Don't bother checking 2461 * for the same hash queue because of how unlikely it is. 2462 */ 2463 __d_drop(dentry); 2464 __d_rehash(dentry, d_hash(target->d_parent, target->d_name.hash)); 2465 2466 /* 2467 * Unhash the target (d_delete() is not usable here). If exchanging 2468 * the two dentries, then rehash onto the other's hash queue. 2469 */ 2470 __d_drop(target); 2471 if (exchange) { 2472 __d_rehash(target, 2473 d_hash(dentry->d_parent, dentry->d_name.hash)); 2474 } 2475 2476 list_del(&dentry->d_u.d_child); 2477 list_del(&target->d_u.d_child); 2478 2479 /* Switch the names.. */ 2480 switch_names(dentry, target); 2481 swap(dentry->d_name.hash, target->d_name.hash); 2482 2483 /* ... and switch the parents */ 2484 if (IS_ROOT(dentry)) { 2485 dentry->d_parent = target->d_parent; 2486 target->d_parent = target; 2487 INIT_LIST_HEAD(&target->d_u.d_child); 2488 } else { 2489 swap(dentry->d_parent, target->d_parent); 2490 2491 /* And add them back to the (new) parent lists */ 2492 list_add(&target->d_u.d_child, &target->d_parent->d_subdirs); 2493 } 2494 2495 list_add(&dentry->d_u.d_child, &dentry->d_parent->d_subdirs); 2496 2497 write_seqcount_end(&target->d_seq); 2498 write_seqcount_end(&dentry->d_seq); 2499 2500 dentry_unlock_parents_for_move(dentry, target); 2501 if (exchange) 2502 fsnotify_d_move(target); 2503 spin_unlock(&target->d_lock); 2504 fsnotify_d_move(dentry); 2505 spin_unlock(&dentry->d_lock); 2506 } 2507 2508 /* 2509 * d_move - move a dentry 2510 * @dentry: entry to move 2511 * @target: new dentry 2512 * 2513 * Update the dcache to reflect the move of a file name. Negative 2514 * dcache entries should not be moved in this way. See the locking 2515 * requirements for __d_move. 2516 */ 2517 void d_move(struct dentry *dentry, struct dentry *target) 2518 { 2519 write_seqlock(&rename_lock); 2520 __d_move(dentry, target, false); 2521 write_sequnlock(&rename_lock); 2522 } 2523 EXPORT_SYMBOL(d_move); 2524 2525 /* 2526 * d_exchange - exchange two dentries 2527 * @dentry1: first dentry 2528 * @dentry2: second dentry 2529 */ 2530 void d_exchange(struct dentry *dentry1, struct dentry *dentry2) 2531 { 2532 write_seqlock(&rename_lock); 2533 2534 WARN_ON(!dentry1->d_inode); 2535 WARN_ON(!dentry2->d_inode); 2536 WARN_ON(IS_ROOT(dentry1)); 2537 WARN_ON(IS_ROOT(dentry2)); 2538 2539 __d_move(dentry1, dentry2, true); 2540 2541 write_sequnlock(&rename_lock); 2542 } 2543 2544 /** 2545 * d_ancestor - search for an ancestor 2546 * @p1: ancestor dentry 2547 * @p2: child dentry 2548 * 2549 * Returns the ancestor dentry of p2 which is a child of p1, if p1 is 2550 * an ancestor of p2, else NULL. 2551 */ 2552 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2) 2553 { 2554 struct dentry *p; 2555 2556 for (p = p2; !IS_ROOT(p); p = p->d_parent) { 2557 if (p->d_parent == p1) 2558 return p; 2559 } 2560 return NULL; 2561 } 2562 2563 /* 2564 * This helper attempts to cope with remotely renamed directories 2565 * 2566 * It assumes that the caller is already holding 2567 * dentry->d_parent->d_inode->i_mutex, inode->i_lock and rename_lock 2568 * 2569 * Note: If ever the locking in lock_rename() changes, then please 2570 * remember to update this too... 2571 */ 2572 static struct dentry *__d_unalias(struct inode *inode, 2573 struct dentry *dentry, struct dentry *alias) 2574 { 2575 struct mutex *m1 = NULL, *m2 = NULL; 2576 struct dentry *ret = ERR_PTR(-EBUSY); 2577 2578 /* If alias and dentry share a parent, then no extra locks required */ 2579 if (alias->d_parent == dentry->d_parent) 2580 goto out_unalias; 2581 2582 /* See lock_rename() */ 2583 if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex)) 2584 goto out_err; 2585 m1 = &dentry->d_sb->s_vfs_rename_mutex; 2586 if (!mutex_trylock(&alias->d_parent->d_inode->i_mutex)) 2587 goto out_err; 2588 m2 = &alias->d_parent->d_inode->i_mutex; 2589 out_unalias: 2590 if (likely(!d_mountpoint(alias))) { 2591 __d_move(alias, dentry, false); 2592 ret = alias; 2593 } 2594 out_err: 2595 spin_unlock(&inode->i_lock); 2596 if (m2) 2597 mutex_unlock(m2); 2598 if (m1) 2599 mutex_unlock(m1); 2600 return ret; 2601 } 2602 2603 /* 2604 * Prepare an anonymous dentry for life in the superblock's dentry tree as a 2605 * named dentry in place of the dentry to be replaced. 2606 * returns with anon->d_lock held! 2607 */ 2608 static void __d_materialise_dentry(struct dentry *dentry, struct dentry *anon) 2609 { 2610 struct dentry *dparent; 2611 2612 dentry_lock_for_move(anon, dentry); 2613 2614 write_seqcount_begin(&dentry->d_seq); 2615 write_seqcount_begin_nested(&anon->d_seq, DENTRY_D_LOCK_NESTED); 2616 2617 dparent = dentry->d_parent; 2618 2619 switch_names(dentry, anon); 2620 swap(dentry->d_name.hash, anon->d_name.hash); 2621 2622 dentry->d_parent = dentry; 2623 list_del_init(&dentry->d_u.d_child); 2624 anon->d_parent = dparent; 2625 list_move(&anon->d_u.d_child, &dparent->d_subdirs); 2626 2627 write_seqcount_end(&dentry->d_seq); 2628 write_seqcount_end(&anon->d_seq); 2629 2630 dentry_unlock_parents_for_move(anon, dentry); 2631 spin_unlock(&dentry->d_lock); 2632 2633 /* anon->d_lock still locked, returns locked */ 2634 } 2635 2636 /** 2637 * d_materialise_unique - introduce an inode into the tree 2638 * @dentry: candidate dentry 2639 * @inode: inode to bind to the dentry, to which aliases may be attached 2640 * 2641 * Introduces an dentry into the tree, substituting an extant disconnected 2642 * root directory alias in its place if there is one. Caller must hold the 2643 * i_mutex of the parent directory. 2644 */ 2645 struct dentry *d_materialise_unique(struct dentry *dentry, struct inode *inode) 2646 { 2647 struct dentry *actual; 2648 2649 BUG_ON(!d_unhashed(dentry)); 2650 2651 if (!inode) { 2652 actual = dentry; 2653 __d_instantiate(dentry, NULL); 2654 d_rehash(actual); 2655 goto out_nolock; 2656 } 2657 2658 spin_lock(&inode->i_lock); 2659 2660 if (S_ISDIR(inode->i_mode)) { 2661 struct dentry *alias; 2662 2663 /* Does an aliased dentry already exist? */ 2664 alias = __d_find_alias(inode, 0); 2665 if (alias) { 2666 actual = alias; 2667 write_seqlock(&rename_lock); 2668 2669 if (d_ancestor(alias, dentry)) { 2670 /* Check for loops */ 2671 actual = ERR_PTR(-ELOOP); 2672 spin_unlock(&inode->i_lock); 2673 } else if (IS_ROOT(alias)) { 2674 /* Is this an anonymous mountpoint that we 2675 * could splice into our tree? */ 2676 __d_materialise_dentry(dentry, alias); 2677 write_sequnlock(&rename_lock); 2678 __d_drop(alias); 2679 goto found; 2680 } else { 2681 /* Nope, but we must(!) avoid directory 2682 * aliasing. This drops inode->i_lock */ 2683 actual = __d_unalias(inode, dentry, alias); 2684 } 2685 write_sequnlock(&rename_lock); 2686 if (IS_ERR(actual)) { 2687 if (PTR_ERR(actual) == -ELOOP) 2688 pr_warn_ratelimited( 2689 "VFS: Lookup of '%s' in %s %s" 2690 " would have caused loop\n", 2691 dentry->d_name.name, 2692 inode->i_sb->s_type->name, 2693 inode->i_sb->s_id); 2694 dput(alias); 2695 } 2696 goto out_nolock; 2697 } 2698 } 2699 2700 /* Add a unique reference */ 2701 actual = __d_instantiate_unique(dentry, inode); 2702 if (!actual) 2703 actual = dentry; 2704 else 2705 BUG_ON(!d_unhashed(actual)); 2706 2707 spin_lock(&actual->d_lock); 2708 found: 2709 _d_rehash(actual); 2710 spin_unlock(&actual->d_lock); 2711 spin_unlock(&inode->i_lock); 2712 out_nolock: 2713 if (actual == dentry) { 2714 security_d_instantiate(dentry, inode); 2715 return NULL; 2716 } 2717 2718 iput(inode); 2719 return actual; 2720 } 2721 EXPORT_SYMBOL_GPL(d_materialise_unique); 2722 2723 static int prepend(char **buffer, int *buflen, const char *str, int namelen) 2724 { 2725 *buflen -= namelen; 2726 if (*buflen < 0) 2727 return -ENAMETOOLONG; 2728 *buffer -= namelen; 2729 memcpy(*buffer, str, namelen); 2730 return 0; 2731 } 2732 2733 /** 2734 * prepend_name - prepend a pathname in front of current buffer pointer 2735 * @buffer: buffer pointer 2736 * @buflen: allocated length of the buffer 2737 * @name: name string and length qstr structure 2738 * 2739 * With RCU path tracing, it may race with d_move(). Use ACCESS_ONCE() to 2740 * make sure that either the old or the new name pointer and length are 2741 * fetched. However, there may be mismatch between length and pointer. 2742 * The length cannot be trusted, we need to copy it byte-by-byte until 2743 * the length is reached or a null byte is found. It also prepends "/" at 2744 * the beginning of the name. The sequence number check at the caller will 2745 * retry it again when a d_move() does happen. So any garbage in the buffer 2746 * due to mismatched pointer and length will be discarded. 2747 */ 2748 static int prepend_name(char **buffer, int *buflen, struct qstr *name) 2749 { 2750 const char *dname = ACCESS_ONCE(name->name); 2751 u32 dlen = ACCESS_ONCE(name->len); 2752 char *p; 2753 2754 *buflen -= dlen + 1; 2755 if (*buflen < 0) 2756 return -ENAMETOOLONG; 2757 p = *buffer -= dlen + 1; 2758 *p++ = '/'; 2759 while (dlen--) { 2760 char c = *dname++; 2761 if (!c) 2762 break; 2763 *p++ = c; 2764 } 2765 return 0; 2766 } 2767 2768 /** 2769 * prepend_path - Prepend path string to a buffer 2770 * @path: the dentry/vfsmount to report 2771 * @root: root vfsmnt/dentry 2772 * @buffer: pointer to the end of the buffer 2773 * @buflen: pointer to buffer length 2774 * 2775 * The function will first try to write out the pathname without taking any 2776 * lock other than the RCU read lock to make sure that dentries won't go away. 2777 * It only checks the sequence number of the global rename_lock as any change 2778 * in the dentry's d_seq will be preceded by changes in the rename_lock 2779 * sequence number. If the sequence number had been changed, it will restart 2780 * the whole pathname back-tracing sequence again by taking the rename_lock. 2781 * In this case, there is no need to take the RCU read lock as the recursive 2782 * parent pointer references will keep the dentry chain alive as long as no 2783 * rename operation is performed. 2784 */ 2785 static int prepend_path(const struct path *path, 2786 const struct path *root, 2787 char **buffer, int *buflen) 2788 { 2789 struct dentry *dentry; 2790 struct vfsmount *vfsmnt; 2791 struct mount *mnt; 2792 int error = 0; 2793 unsigned seq, m_seq = 0; 2794 char *bptr; 2795 int blen; 2796 2797 rcu_read_lock(); 2798 restart_mnt: 2799 read_seqbegin_or_lock(&mount_lock, &m_seq); 2800 seq = 0; 2801 rcu_read_lock(); 2802 restart: 2803 bptr = *buffer; 2804 blen = *buflen; 2805 error = 0; 2806 dentry = path->dentry; 2807 vfsmnt = path->mnt; 2808 mnt = real_mount(vfsmnt); 2809 read_seqbegin_or_lock(&rename_lock, &seq); 2810 while (dentry != root->dentry || vfsmnt != root->mnt) { 2811 struct dentry * parent; 2812 2813 if (dentry == vfsmnt->mnt_root || IS_ROOT(dentry)) { 2814 struct mount *parent = ACCESS_ONCE(mnt->mnt_parent); 2815 /* Global root? */ 2816 if (mnt != parent) { 2817 dentry = ACCESS_ONCE(mnt->mnt_mountpoint); 2818 mnt = parent; 2819 vfsmnt = &mnt->mnt; 2820 continue; 2821 } 2822 /* 2823 * Filesystems needing to implement special "root names" 2824 * should do so with ->d_dname() 2825 */ 2826 if (IS_ROOT(dentry) && 2827 (dentry->d_name.len != 1 || 2828 dentry->d_name.name[0] != '/')) { 2829 WARN(1, "Root dentry has weird name <%.*s>\n", 2830 (int) dentry->d_name.len, 2831 dentry->d_name.name); 2832 } 2833 if (!error) 2834 error = is_mounted(vfsmnt) ? 1 : 2; 2835 break; 2836 } 2837 parent = dentry->d_parent; 2838 prefetch(parent); 2839 error = prepend_name(&bptr, &blen, &dentry->d_name); 2840 if (error) 2841 break; 2842 2843 dentry = parent; 2844 } 2845 if (!(seq & 1)) 2846 rcu_read_unlock(); 2847 if (need_seqretry(&rename_lock, seq)) { 2848 seq = 1; 2849 goto restart; 2850 } 2851 done_seqretry(&rename_lock, seq); 2852 2853 if (!(m_seq & 1)) 2854 rcu_read_unlock(); 2855 if (need_seqretry(&mount_lock, m_seq)) { 2856 m_seq = 1; 2857 goto restart_mnt; 2858 } 2859 done_seqretry(&mount_lock, m_seq); 2860 2861 if (error >= 0 && bptr == *buffer) { 2862 if (--blen < 0) 2863 error = -ENAMETOOLONG; 2864 else 2865 *--bptr = '/'; 2866 } 2867 *buffer = bptr; 2868 *buflen = blen; 2869 return error; 2870 } 2871 2872 /** 2873 * __d_path - return the path of a dentry 2874 * @path: the dentry/vfsmount to report 2875 * @root: root vfsmnt/dentry 2876 * @buf: buffer to return value in 2877 * @buflen: buffer length 2878 * 2879 * Convert a dentry into an ASCII path name. 2880 * 2881 * Returns a pointer into the buffer or an error code if the 2882 * path was too long. 2883 * 2884 * "buflen" should be positive. 2885 * 2886 * If the path is not reachable from the supplied root, return %NULL. 2887 */ 2888 char *__d_path(const struct path *path, 2889 const struct path *root, 2890 char *buf, int buflen) 2891 { 2892 char *res = buf + buflen; 2893 int error; 2894 2895 prepend(&res, &buflen, "\0", 1); 2896 error = prepend_path(path, root, &res, &buflen); 2897 2898 if (error < 0) 2899 return ERR_PTR(error); 2900 if (error > 0) 2901 return NULL; 2902 return res; 2903 } 2904 2905 char *d_absolute_path(const struct path *path, 2906 char *buf, int buflen) 2907 { 2908 struct path root = {}; 2909 char *res = buf + buflen; 2910 int error; 2911 2912 prepend(&res, &buflen, "\0", 1); 2913 error = prepend_path(path, &root, &res, &buflen); 2914 2915 if (error > 1) 2916 error = -EINVAL; 2917 if (error < 0) 2918 return ERR_PTR(error); 2919 return res; 2920 } 2921 2922 /* 2923 * same as __d_path but appends "(deleted)" for unlinked files. 2924 */ 2925 static int path_with_deleted(const struct path *path, 2926 const struct path *root, 2927 char **buf, int *buflen) 2928 { 2929 prepend(buf, buflen, "\0", 1); 2930 if (d_unlinked(path->dentry)) { 2931 int error = prepend(buf, buflen, " (deleted)", 10); 2932 if (error) 2933 return error; 2934 } 2935 2936 return prepend_path(path, root, buf, buflen); 2937 } 2938 2939 static int prepend_unreachable(char **buffer, int *buflen) 2940 { 2941 return prepend(buffer, buflen, "(unreachable)", 13); 2942 } 2943 2944 static void get_fs_root_rcu(struct fs_struct *fs, struct path *root) 2945 { 2946 unsigned seq; 2947 2948 do { 2949 seq = read_seqcount_begin(&fs->seq); 2950 *root = fs->root; 2951 } while (read_seqcount_retry(&fs->seq, seq)); 2952 } 2953 2954 /** 2955 * d_path - return the path of a dentry 2956 * @path: path to report 2957 * @buf: buffer to return value in 2958 * @buflen: buffer length 2959 * 2960 * Convert a dentry into an ASCII path name. If the entry has been deleted 2961 * the string " (deleted)" is appended. Note that this is ambiguous. 2962 * 2963 * Returns a pointer into the buffer or an error code if the path was 2964 * too long. Note: Callers should use the returned pointer, not the passed 2965 * in buffer, to use the name! The implementation often starts at an offset 2966 * into the buffer, and may leave 0 bytes at the start. 2967 * 2968 * "buflen" should be positive. 2969 */ 2970 char *d_path(const struct path *path, char *buf, int buflen) 2971 { 2972 char *res = buf + buflen; 2973 struct path root; 2974 int error; 2975 2976 /* 2977 * We have various synthetic filesystems that never get mounted. On 2978 * these filesystems dentries are never used for lookup purposes, and 2979 * thus don't need to be hashed. They also don't need a name until a 2980 * user wants to identify the object in /proc/pid/fd/. The little hack 2981 * below allows us to generate a name for these objects on demand: 2982 * 2983 * Some pseudo inodes are mountable. When they are mounted 2984 * path->dentry == path->mnt->mnt_root. In that case don't call d_dname 2985 * and instead have d_path return the mounted path. 2986 */ 2987 if (path->dentry->d_op && path->dentry->d_op->d_dname && 2988 (!IS_ROOT(path->dentry) || path->dentry != path->mnt->mnt_root)) 2989 return path->dentry->d_op->d_dname(path->dentry, buf, buflen); 2990 2991 rcu_read_lock(); 2992 get_fs_root_rcu(current->fs, &root); 2993 error = path_with_deleted(path, &root, &res, &buflen); 2994 rcu_read_unlock(); 2995 2996 if (error < 0) 2997 res = ERR_PTR(error); 2998 return res; 2999 } 3000 EXPORT_SYMBOL(d_path); 3001 3002 /* 3003 * Helper function for dentry_operations.d_dname() members 3004 */ 3005 char *dynamic_dname(struct dentry *dentry, char *buffer, int buflen, 3006 const char *fmt, ...) 3007 { 3008 va_list args; 3009 char temp[64]; 3010 int sz; 3011 3012 va_start(args, fmt); 3013 sz = vsnprintf(temp, sizeof(temp), fmt, args) + 1; 3014 va_end(args); 3015 3016 if (sz > sizeof(temp) || sz > buflen) 3017 return ERR_PTR(-ENAMETOOLONG); 3018 3019 buffer += buflen - sz; 3020 return memcpy(buffer, temp, sz); 3021 } 3022 3023 char *simple_dname(struct dentry *dentry, char *buffer, int buflen) 3024 { 3025 char *end = buffer + buflen; 3026 /* these dentries are never renamed, so d_lock is not needed */ 3027 if (prepend(&end, &buflen, " (deleted)", 11) || 3028 prepend(&end, &buflen, dentry->d_name.name, dentry->d_name.len) || 3029 prepend(&end, &buflen, "/", 1)) 3030 end = ERR_PTR(-ENAMETOOLONG); 3031 return end; 3032 } 3033 EXPORT_SYMBOL(simple_dname); 3034 3035 /* 3036 * Write full pathname from the root of the filesystem into the buffer. 3037 */ 3038 static char *__dentry_path(struct dentry *d, char *buf, int buflen) 3039 { 3040 struct dentry *dentry; 3041 char *end, *retval; 3042 int len, seq = 0; 3043 int error = 0; 3044 3045 if (buflen < 2) 3046 goto Elong; 3047 3048 rcu_read_lock(); 3049 restart: 3050 dentry = d; 3051 end = buf + buflen; 3052 len = buflen; 3053 prepend(&end, &len, "\0", 1); 3054 /* Get '/' right */ 3055 retval = end-1; 3056 *retval = '/'; 3057 read_seqbegin_or_lock(&rename_lock, &seq); 3058 while (!IS_ROOT(dentry)) { 3059 struct dentry *parent = dentry->d_parent; 3060 3061 prefetch(parent); 3062 error = prepend_name(&end, &len, &dentry->d_name); 3063 if (error) 3064 break; 3065 3066 retval = end; 3067 dentry = parent; 3068 } 3069 if (!(seq & 1)) 3070 rcu_read_unlock(); 3071 if (need_seqretry(&rename_lock, seq)) { 3072 seq = 1; 3073 goto restart; 3074 } 3075 done_seqretry(&rename_lock, seq); 3076 if (error) 3077 goto Elong; 3078 return retval; 3079 Elong: 3080 return ERR_PTR(-ENAMETOOLONG); 3081 } 3082 3083 char *dentry_path_raw(struct dentry *dentry, char *buf, int buflen) 3084 { 3085 return __dentry_path(dentry, buf, buflen); 3086 } 3087 EXPORT_SYMBOL(dentry_path_raw); 3088 3089 char *dentry_path(struct dentry *dentry, char *buf, int buflen) 3090 { 3091 char *p = NULL; 3092 char *retval; 3093 3094 if (d_unlinked(dentry)) { 3095 p = buf + buflen; 3096 if (prepend(&p, &buflen, "//deleted", 10) != 0) 3097 goto Elong; 3098 buflen++; 3099 } 3100 retval = __dentry_path(dentry, buf, buflen); 3101 if (!IS_ERR(retval) && p) 3102 *p = '/'; /* restore '/' overriden with '\0' */ 3103 return retval; 3104 Elong: 3105 return ERR_PTR(-ENAMETOOLONG); 3106 } 3107 3108 static void get_fs_root_and_pwd_rcu(struct fs_struct *fs, struct path *root, 3109 struct path *pwd) 3110 { 3111 unsigned seq; 3112 3113 do { 3114 seq = read_seqcount_begin(&fs->seq); 3115 *root = fs->root; 3116 *pwd = fs->pwd; 3117 } while (read_seqcount_retry(&fs->seq, seq)); 3118 } 3119 3120 /* 3121 * NOTE! The user-level library version returns a 3122 * character pointer. The kernel system call just 3123 * returns the length of the buffer filled (which 3124 * includes the ending '\0' character), or a negative 3125 * error value. So libc would do something like 3126 * 3127 * char *getcwd(char * buf, size_t size) 3128 * { 3129 * int retval; 3130 * 3131 * retval = sys_getcwd(buf, size); 3132 * if (retval >= 0) 3133 * return buf; 3134 * errno = -retval; 3135 * return NULL; 3136 * } 3137 */ 3138 SYSCALL_DEFINE2(getcwd, char __user *, buf, unsigned long, size) 3139 { 3140 int error; 3141 struct path pwd, root; 3142 char *page = __getname(); 3143 3144 if (!page) 3145 return -ENOMEM; 3146 3147 rcu_read_lock(); 3148 get_fs_root_and_pwd_rcu(current->fs, &root, &pwd); 3149 3150 error = -ENOENT; 3151 if (!d_unlinked(pwd.dentry)) { 3152 unsigned long len; 3153 char *cwd = page + PATH_MAX; 3154 int buflen = PATH_MAX; 3155 3156 prepend(&cwd, &buflen, "\0", 1); 3157 error = prepend_path(&pwd, &root, &cwd, &buflen); 3158 rcu_read_unlock(); 3159 3160 if (error < 0) 3161 goto out; 3162 3163 /* Unreachable from current root */ 3164 if (error > 0) { 3165 error = prepend_unreachable(&cwd, &buflen); 3166 if (error) 3167 goto out; 3168 } 3169 3170 error = -ERANGE; 3171 len = PATH_MAX + page - cwd; 3172 if (len <= size) { 3173 error = len; 3174 if (copy_to_user(buf, cwd, len)) 3175 error = -EFAULT; 3176 } 3177 } else { 3178 rcu_read_unlock(); 3179 } 3180 3181 out: 3182 __putname(page); 3183 return error; 3184 } 3185 3186 /* 3187 * Test whether new_dentry is a subdirectory of old_dentry. 3188 * 3189 * Trivially implemented using the dcache structure 3190 */ 3191 3192 /** 3193 * is_subdir - is new dentry a subdirectory of old_dentry 3194 * @new_dentry: new dentry 3195 * @old_dentry: old dentry 3196 * 3197 * Returns 1 if new_dentry is a subdirectory of the parent (at any depth). 3198 * Returns 0 otherwise. 3199 * Caller must ensure that "new_dentry" is pinned before calling is_subdir() 3200 */ 3201 3202 int is_subdir(struct dentry *new_dentry, struct dentry *old_dentry) 3203 { 3204 int result; 3205 unsigned seq; 3206 3207 if (new_dentry == old_dentry) 3208 return 1; 3209 3210 do { 3211 /* for restarting inner loop in case of seq retry */ 3212 seq = read_seqbegin(&rename_lock); 3213 /* 3214 * Need rcu_readlock to protect against the d_parent trashing 3215 * due to d_move 3216 */ 3217 rcu_read_lock(); 3218 if (d_ancestor(old_dentry, new_dentry)) 3219 result = 1; 3220 else 3221 result = 0; 3222 rcu_read_unlock(); 3223 } while (read_seqretry(&rename_lock, seq)); 3224 3225 return result; 3226 } 3227 3228 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry) 3229 { 3230 struct dentry *root = data; 3231 if (dentry != root) { 3232 if (d_unhashed(dentry) || !dentry->d_inode) 3233 return D_WALK_SKIP; 3234 3235 if (!(dentry->d_flags & DCACHE_GENOCIDE)) { 3236 dentry->d_flags |= DCACHE_GENOCIDE; 3237 dentry->d_lockref.count--; 3238 } 3239 } 3240 return D_WALK_CONTINUE; 3241 } 3242 3243 void d_genocide(struct dentry *parent) 3244 { 3245 d_walk(parent, parent, d_genocide_kill, NULL); 3246 } 3247 3248 void d_tmpfile(struct dentry *dentry, struct inode *inode) 3249 { 3250 inode_dec_link_count(inode); 3251 BUG_ON(dentry->d_name.name != dentry->d_iname || 3252 !hlist_unhashed(&dentry->d_alias) || 3253 !d_unlinked(dentry)); 3254 spin_lock(&dentry->d_parent->d_lock); 3255 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED); 3256 dentry->d_name.len = sprintf(dentry->d_iname, "#%llu", 3257 (unsigned long long)inode->i_ino); 3258 spin_unlock(&dentry->d_lock); 3259 spin_unlock(&dentry->d_parent->d_lock); 3260 d_instantiate(dentry, inode); 3261 } 3262 EXPORT_SYMBOL(d_tmpfile); 3263 3264 static __initdata unsigned long dhash_entries; 3265 static int __init set_dhash_entries(char *str) 3266 { 3267 if (!str) 3268 return 0; 3269 dhash_entries = simple_strtoul(str, &str, 0); 3270 return 1; 3271 } 3272 __setup("dhash_entries=", set_dhash_entries); 3273 3274 static void __init dcache_init_early(void) 3275 { 3276 unsigned int loop; 3277 3278 /* If hashes are distributed across NUMA nodes, defer 3279 * hash allocation until vmalloc space is available. 3280 */ 3281 if (hashdist) 3282 return; 3283 3284 dentry_hashtable = 3285 alloc_large_system_hash("Dentry cache", 3286 sizeof(struct hlist_bl_head), 3287 dhash_entries, 3288 13, 3289 HASH_EARLY, 3290 &d_hash_shift, 3291 &d_hash_mask, 3292 0, 3293 0); 3294 3295 for (loop = 0; loop < (1U << d_hash_shift); loop++) 3296 INIT_HLIST_BL_HEAD(dentry_hashtable + loop); 3297 } 3298 3299 static void __init dcache_init(void) 3300 { 3301 unsigned int loop; 3302 3303 /* 3304 * A constructor could be added for stable state like the lists, 3305 * but it is probably not worth it because of the cache nature 3306 * of the dcache. 3307 */ 3308 dentry_cache = KMEM_CACHE(dentry, 3309 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD); 3310 3311 /* Hash may have been set up in dcache_init_early */ 3312 if (!hashdist) 3313 return; 3314 3315 dentry_hashtable = 3316 alloc_large_system_hash("Dentry cache", 3317 sizeof(struct hlist_bl_head), 3318 dhash_entries, 3319 13, 3320 0, 3321 &d_hash_shift, 3322 &d_hash_mask, 3323 0, 3324 0); 3325 3326 for (loop = 0; loop < (1U << d_hash_shift); loop++) 3327 INIT_HLIST_BL_HEAD(dentry_hashtable + loop); 3328 } 3329 3330 /* SLAB cache for __getname() consumers */ 3331 struct kmem_cache *names_cachep __read_mostly; 3332 EXPORT_SYMBOL(names_cachep); 3333 3334 EXPORT_SYMBOL(d_genocide); 3335 3336 void __init vfs_caches_init_early(void) 3337 { 3338 dcache_init_early(); 3339 inode_init_early(); 3340 } 3341 3342 void __init vfs_caches_init(unsigned long mempages) 3343 { 3344 unsigned long reserve; 3345 3346 /* Base hash sizes on available memory, with a reserve equal to 3347 150% of current kernel size */ 3348 3349 reserve = min((mempages - nr_free_pages()) * 3/2, mempages - 1); 3350 mempages -= reserve; 3351 3352 names_cachep = kmem_cache_create("names_cache", PATH_MAX, 0, 3353 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3354 3355 dcache_init(); 3356 inode_init(); 3357 files_init(mempages); 3358 mnt_init(); 3359 bdev_cache_init(); 3360 chrdev_init(); 3361 } 3362