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(struct 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 static void __dentry_kill(struct dentry *dentry) 445 { 446 struct dentry *parent = NULL; 447 bool can_free = true; 448 if (!IS_ROOT(dentry)) 449 parent = dentry->d_parent; 450 451 /* 452 * The dentry is now unrecoverably dead to the world. 453 */ 454 lockref_mark_dead(&dentry->d_lockref); 455 456 /* 457 * inform the fs via d_prune that this dentry is about to be 458 * unhashed and destroyed. 459 */ 460 if ((dentry->d_flags & DCACHE_OP_PRUNE) && !d_unhashed(dentry)) 461 dentry->d_op->d_prune(dentry); 462 463 if (dentry->d_flags & DCACHE_LRU_LIST) { 464 if (!(dentry->d_flags & DCACHE_SHRINK_LIST)) 465 d_lru_del(dentry); 466 } 467 /* if it was on the hash then remove it */ 468 __d_drop(dentry); 469 list_del(&dentry->d_u.d_child); 470 /* 471 * Inform d_walk() that we are no longer attached to the 472 * dentry tree 473 */ 474 dentry->d_flags |= DCACHE_DENTRY_KILLED; 475 if (parent) 476 spin_unlock(&parent->d_lock); 477 dentry_iput(dentry); 478 /* 479 * dentry_iput drops the locks, at which point nobody (except 480 * transient RCU lookups) can reach this dentry. 481 */ 482 BUG_ON((int)dentry->d_lockref.count > 0); 483 this_cpu_dec(nr_dentry); 484 if (dentry->d_op && dentry->d_op->d_release) 485 dentry->d_op->d_release(dentry); 486 487 spin_lock(&dentry->d_lock); 488 if (dentry->d_flags & DCACHE_SHRINK_LIST) { 489 dentry->d_flags |= DCACHE_MAY_FREE; 490 can_free = false; 491 } 492 spin_unlock(&dentry->d_lock); 493 if (likely(can_free)) 494 dentry_free(dentry); 495 } 496 497 /* 498 * Finish off a dentry we've decided to kill. 499 * dentry->d_lock must be held, returns with it unlocked. 500 * If ref is non-zero, then decrement the refcount too. 501 * Returns dentry requiring refcount drop, or NULL if we're done. 502 */ 503 static struct dentry *dentry_kill(struct dentry *dentry) 504 __releases(dentry->d_lock) 505 { 506 struct inode *inode = dentry->d_inode; 507 struct dentry *parent = NULL; 508 509 if (inode && unlikely(!spin_trylock(&inode->i_lock))) 510 goto failed; 511 512 if (!IS_ROOT(dentry)) { 513 parent = dentry->d_parent; 514 if (unlikely(!spin_trylock(&parent->d_lock))) { 515 if (inode) 516 spin_unlock(&inode->i_lock); 517 goto failed; 518 } 519 } 520 521 __dentry_kill(dentry); 522 return parent; 523 524 failed: 525 spin_unlock(&dentry->d_lock); 526 cpu_relax(); 527 return dentry; /* try again with same dentry */ 528 } 529 530 static inline struct dentry *lock_parent(struct dentry *dentry) 531 { 532 struct dentry *parent = dentry->d_parent; 533 if (IS_ROOT(dentry)) 534 return NULL; 535 if (unlikely((int)dentry->d_lockref.count < 0)) 536 return NULL; 537 if (likely(spin_trylock(&parent->d_lock))) 538 return parent; 539 rcu_read_lock(); 540 spin_unlock(&dentry->d_lock); 541 again: 542 parent = ACCESS_ONCE(dentry->d_parent); 543 spin_lock(&parent->d_lock); 544 /* 545 * We can't blindly lock dentry until we are sure 546 * that we won't violate the locking order. 547 * Any changes of dentry->d_parent must have 548 * been done with parent->d_lock held, so 549 * spin_lock() above is enough of a barrier 550 * for checking if it's still our child. 551 */ 552 if (unlikely(parent != dentry->d_parent)) { 553 spin_unlock(&parent->d_lock); 554 goto again; 555 } 556 rcu_read_unlock(); 557 if (parent != dentry) 558 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED); 559 else 560 parent = NULL; 561 return parent; 562 } 563 564 /* 565 * This is dput 566 * 567 * This is complicated by the fact that we do not want to put 568 * dentries that are no longer on any hash chain on the unused 569 * list: we'd much rather just get rid of them immediately. 570 * 571 * However, that implies that we have to traverse the dentry 572 * tree upwards to the parents which might _also_ now be 573 * scheduled for deletion (it may have been only waiting for 574 * its last child to go away). 575 * 576 * This tail recursion is done by hand as we don't want to depend 577 * on the compiler to always get this right (gcc generally doesn't). 578 * Real recursion would eat up our stack space. 579 */ 580 581 /* 582 * dput - release a dentry 583 * @dentry: dentry to release 584 * 585 * Release a dentry. This will drop the usage count and if appropriate 586 * call the dentry unlink method as well as removing it from the queues and 587 * releasing its resources. If the parent dentries were scheduled for release 588 * they too may now get deleted. 589 */ 590 void dput(struct dentry *dentry) 591 { 592 if (unlikely(!dentry)) 593 return; 594 595 repeat: 596 if (lockref_put_or_lock(&dentry->d_lockref)) 597 return; 598 599 /* Unreachable? Get rid of it */ 600 if (unlikely(d_unhashed(dentry))) 601 goto kill_it; 602 603 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) { 604 if (dentry->d_op->d_delete(dentry)) 605 goto kill_it; 606 } 607 608 if (!(dentry->d_flags & DCACHE_REFERENCED)) 609 dentry->d_flags |= DCACHE_REFERENCED; 610 dentry_lru_add(dentry); 611 612 dentry->d_lockref.count--; 613 spin_unlock(&dentry->d_lock); 614 return; 615 616 kill_it: 617 dentry = dentry_kill(dentry); 618 if (dentry) 619 goto repeat; 620 } 621 EXPORT_SYMBOL(dput); 622 623 /** 624 * d_invalidate - invalidate a dentry 625 * @dentry: dentry to invalidate 626 * 627 * Try to invalidate the dentry if it turns out to be 628 * possible. If there are other dentries that can be 629 * reached through this one we can't delete it and we 630 * return -EBUSY. On success we return 0. 631 * 632 * no dcache lock. 633 */ 634 635 int d_invalidate(struct dentry * dentry) 636 { 637 /* 638 * If it's already been dropped, return OK. 639 */ 640 spin_lock(&dentry->d_lock); 641 if (d_unhashed(dentry)) { 642 spin_unlock(&dentry->d_lock); 643 return 0; 644 } 645 /* 646 * Check whether to do a partial shrink_dcache 647 * to get rid of unused child entries. 648 */ 649 if (!list_empty(&dentry->d_subdirs)) { 650 spin_unlock(&dentry->d_lock); 651 shrink_dcache_parent(dentry); 652 spin_lock(&dentry->d_lock); 653 } 654 655 /* 656 * Somebody else still using it? 657 * 658 * If it's a directory, we can't drop it 659 * for fear of somebody re-populating it 660 * with children (even though dropping it 661 * would make it unreachable from the root, 662 * we might still populate it if it was a 663 * working directory or similar). 664 * We also need to leave mountpoints alone, 665 * directory or not. 666 */ 667 if (dentry->d_lockref.count > 1 && dentry->d_inode) { 668 if (S_ISDIR(dentry->d_inode->i_mode) || d_mountpoint(dentry)) { 669 spin_unlock(&dentry->d_lock); 670 return -EBUSY; 671 } 672 } 673 674 __d_drop(dentry); 675 spin_unlock(&dentry->d_lock); 676 return 0; 677 } 678 EXPORT_SYMBOL(d_invalidate); 679 680 /* This must be called with d_lock held */ 681 static inline void __dget_dlock(struct dentry *dentry) 682 { 683 dentry->d_lockref.count++; 684 } 685 686 static inline void __dget(struct dentry *dentry) 687 { 688 lockref_get(&dentry->d_lockref); 689 } 690 691 struct dentry *dget_parent(struct dentry *dentry) 692 { 693 int gotref; 694 struct dentry *ret; 695 696 /* 697 * Do optimistic parent lookup without any 698 * locking. 699 */ 700 rcu_read_lock(); 701 ret = ACCESS_ONCE(dentry->d_parent); 702 gotref = lockref_get_not_zero(&ret->d_lockref); 703 rcu_read_unlock(); 704 if (likely(gotref)) { 705 if (likely(ret == ACCESS_ONCE(dentry->d_parent))) 706 return ret; 707 dput(ret); 708 } 709 710 repeat: 711 /* 712 * Don't need rcu_dereference because we re-check it was correct under 713 * the lock. 714 */ 715 rcu_read_lock(); 716 ret = dentry->d_parent; 717 spin_lock(&ret->d_lock); 718 if (unlikely(ret != dentry->d_parent)) { 719 spin_unlock(&ret->d_lock); 720 rcu_read_unlock(); 721 goto repeat; 722 } 723 rcu_read_unlock(); 724 BUG_ON(!ret->d_lockref.count); 725 ret->d_lockref.count++; 726 spin_unlock(&ret->d_lock); 727 return ret; 728 } 729 EXPORT_SYMBOL(dget_parent); 730 731 /** 732 * d_find_alias - grab a hashed alias of inode 733 * @inode: inode in question 734 * 735 * If inode has a hashed alias, or is a directory and has any alias, 736 * acquire the reference to alias and return it. Otherwise return NULL. 737 * Notice that if inode is a directory there can be only one alias and 738 * it can be unhashed only if it has no children, or if it is the root 739 * of a filesystem. 740 * 741 * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer 742 * any other hashed alias over that one. 743 */ 744 static struct dentry *__d_find_alias(struct inode *inode) 745 { 746 struct dentry *alias, *discon_alias; 747 748 again: 749 discon_alias = NULL; 750 hlist_for_each_entry(alias, &inode->i_dentry, d_alias) { 751 spin_lock(&alias->d_lock); 752 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) { 753 if (IS_ROOT(alias) && 754 (alias->d_flags & DCACHE_DISCONNECTED)) { 755 discon_alias = alias; 756 } else { 757 __dget_dlock(alias); 758 spin_unlock(&alias->d_lock); 759 return alias; 760 } 761 } 762 spin_unlock(&alias->d_lock); 763 } 764 if (discon_alias) { 765 alias = discon_alias; 766 spin_lock(&alias->d_lock); 767 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) { 768 __dget_dlock(alias); 769 spin_unlock(&alias->d_lock); 770 return alias; 771 } 772 spin_unlock(&alias->d_lock); 773 goto again; 774 } 775 return NULL; 776 } 777 778 struct dentry *d_find_alias(struct inode *inode) 779 { 780 struct dentry *de = NULL; 781 782 if (!hlist_empty(&inode->i_dentry)) { 783 spin_lock(&inode->i_lock); 784 de = __d_find_alias(inode); 785 spin_unlock(&inode->i_lock); 786 } 787 return de; 788 } 789 EXPORT_SYMBOL(d_find_alias); 790 791 /* 792 * Try to kill dentries associated with this inode. 793 * WARNING: you must own a reference to inode. 794 */ 795 void d_prune_aliases(struct inode *inode) 796 { 797 struct dentry *dentry; 798 restart: 799 spin_lock(&inode->i_lock); 800 hlist_for_each_entry(dentry, &inode->i_dentry, d_alias) { 801 spin_lock(&dentry->d_lock); 802 if (!dentry->d_lockref.count) { 803 /* 804 * inform the fs via d_prune that this dentry 805 * is about to be unhashed and destroyed. 806 */ 807 if ((dentry->d_flags & DCACHE_OP_PRUNE) && 808 !d_unhashed(dentry)) 809 dentry->d_op->d_prune(dentry); 810 811 __dget_dlock(dentry); 812 __d_drop(dentry); 813 spin_unlock(&dentry->d_lock); 814 spin_unlock(&inode->i_lock); 815 dput(dentry); 816 goto restart; 817 } 818 spin_unlock(&dentry->d_lock); 819 } 820 spin_unlock(&inode->i_lock); 821 } 822 EXPORT_SYMBOL(d_prune_aliases); 823 824 static void shrink_dentry_list(struct list_head *list) 825 { 826 struct dentry *dentry, *parent; 827 828 while (!list_empty(list)) { 829 struct inode *inode; 830 dentry = list_entry(list->prev, struct dentry, d_lru); 831 spin_lock(&dentry->d_lock); 832 parent = lock_parent(dentry); 833 834 /* 835 * The dispose list is isolated and dentries are not accounted 836 * to the LRU here, so we can simply remove it from the list 837 * here regardless of whether it is referenced or not. 838 */ 839 d_shrink_del(dentry); 840 841 /* 842 * We found an inuse dentry which was not removed from 843 * the LRU because of laziness during lookup. Do not free it. 844 */ 845 if ((int)dentry->d_lockref.count > 0) { 846 spin_unlock(&dentry->d_lock); 847 if (parent) 848 spin_unlock(&parent->d_lock); 849 continue; 850 } 851 852 853 if (unlikely(dentry->d_flags & DCACHE_DENTRY_KILLED)) { 854 bool can_free = dentry->d_flags & DCACHE_MAY_FREE; 855 spin_unlock(&dentry->d_lock); 856 if (parent) 857 spin_unlock(&parent->d_lock); 858 if (can_free) 859 dentry_free(dentry); 860 continue; 861 } 862 863 inode = dentry->d_inode; 864 if (inode && unlikely(!spin_trylock(&inode->i_lock))) { 865 d_shrink_add(dentry, list); 866 spin_unlock(&dentry->d_lock); 867 if (parent) 868 spin_unlock(&parent->d_lock); 869 continue; 870 } 871 872 __dentry_kill(dentry); 873 874 /* 875 * We need to prune ancestors too. This is necessary to prevent 876 * quadratic behavior of shrink_dcache_parent(), but is also 877 * expected to be beneficial in reducing dentry cache 878 * fragmentation. 879 */ 880 dentry = parent; 881 while (dentry && !lockref_put_or_lock(&dentry->d_lockref)) { 882 parent = lock_parent(dentry); 883 if (dentry->d_lockref.count != 1) { 884 dentry->d_lockref.count--; 885 spin_unlock(&dentry->d_lock); 886 if (parent) 887 spin_unlock(&parent->d_lock); 888 break; 889 } 890 inode = dentry->d_inode; /* can't be NULL */ 891 if (unlikely(!spin_trylock(&inode->i_lock))) { 892 spin_unlock(&dentry->d_lock); 893 if (parent) 894 spin_unlock(&parent->d_lock); 895 cpu_relax(); 896 continue; 897 } 898 __dentry_kill(dentry); 899 dentry = parent; 900 } 901 } 902 } 903 904 static enum lru_status 905 dentry_lru_isolate(struct list_head *item, spinlock_t *lru_lock, void *arg) 906 { 907 struct list_head *freeable = arg; 908 struct dentry *dentry = container_of(item, struct dentry, d_lru); 909 910 911 /* 912 * we are inverting the lru lock/dentry->d_lock here, 913 * so use a trylock. If we fail to get the lock, just skip 914 * it 915 */ 916 if (!spin_trylock(&dentry->d_lock)) 917 return LRU_SKIP; 918 919 /* 920 * Referenced dentries are still in use. If they have active 921 * counts, just remove them from the LRU. Otherwise give them 922 * another pass through the LRU. 923 */ 924 if (dentry->d_lockref.count) { 925 d_lru_isolate(dentry); 926 spin_unlock(&dentry->d_lock); 927 return LRU_REMOVED; 928 } 929 930 if (dentry->d_flags & DCACHE_REFERENCED) { 931 dentry->d_flags &= ~DCACHE_REFERENCED; 932 spin_unlock(&dentry->d_lock); 933 934 /* 935 * The list move itself will be made by the common LRU code. At 936 * this point, we've dropped the dentry->d_lock but keep the 937 * lru lock. This is safe to do, since every list movement is 938 * protected by the lru lock even if both locks are held. 939 * 940 * This is guaranteed by the fact that all LRU management 941 * functions are intermediated by the LRU API calls like 942 * list_lru_add and list_lru_del. List movement in this file 943 * only ever occur through this functions or through callbacks 944 * like this one, that are called from the LRU API. 945 * 946 * The only exceptions to this are functions like 947 * shrink_dentry_list, and code that first checks for the 948 * DCACHE_SHRINK_LIST flag. Those are guaranteed to be 949 * operating only with stack provided lists after they are 950 * properly isolated from the main list. It is thus, always a 951 * local access. 952 */ 953 return LRU_ROTATE; 954 } 955 956 d_lru_shrink_move(dentry, freeable); 957 spin_unlock(&dentry->d_lock); 958 959 return LRU_REMOVED; 960 } 961 962 /** 963 * prune_dcache_sb - shrink the dcache 964 * @sb: superblock 965 * @nr_to_scan : number of entries to try to free 966 * @nid: which node to scan for freeable entities 967 * 968 * Attempt to shrink the superblock dcache LRU by @nr_to_scan entries. This is 969 * done when we need more memory an called from the superblock shrinker 970 * function. 971 * 972 * This function may fail to free any resources if all the dentries are in 973 * use. 974 */ 975 long prune_dcache_sb(struct super_block *sb, unsigned long nr_to_scan, 976 int nid) 977 { 978 LIST_HEAD(dispose); 979 long freed; 980 981 freed = list_lru_walk_node(&sb->s_dentry_lru, nid, dentry_lru_isolate, 982 &dispose, &nr_to_scan); 983 shrink_dentry_list(&dispose); 984 return freed; 985 } 986 987 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item, 988 spinlock_t *lru_lock, void *arg) 989 { 990 struct list_head *freeable = arg; 991 struct dentry *dentry = container_of(item, struct dentry, d_lru); 992 993 /* 994 * we are inverting the lru lock/dentry->d_lock here, 995 * so use a trylock. If we fail to get the lock, just skip 996 * it 997 */ 998 if (!spin_trylock(&dentry->d_lock)) 999 return LRU_SKIP; 1000 1001 d_lru_shrink_move(dentry, freeable); 1002 spin_unlock(&dentry->d_lock); 1003 1004 return LRU_REMOVED; 1005 } 1006 1007 1008 /** 1009 * shrink_dcache_sb - shrink dcache for a superblock 1010 * @sb: superblock 1011 * 1012 * Shrink the dcache for the specified super block. This is used to free 1013 * the dcache before unmounting a file system. 1014 */ 1015 void shrink_dcache_sb(struct super_block *sb) 1016 { 1017 long freed; 1018 1019 do { 1020 LIST_HEAD(dispose); 1021 1022 freed = list_lru_walk(&sb->s_dentry_lru, 1023 dentry_lru_isolate_shrink, &dispose, UINT_MAX); 1024 1025 this_cpu_sub(nr_dentry_unused, freed); 1026 shrink_dentry_list(&dispose); 1027 } while (freed > 0); 1028 } 1029 EXPORT_SYMBOL(shrink_dcache_sb); 1030 1031 /** 1032 * enum d_walk_ret - action to talke during tree walk 1033 * @D_WALK_CONTINUE: contrinue walk 1034 * @D_WALK_QUIT: quit walk 1035 * @D_WALK_NORETRY: quit when retry is needed 1036 * @D_WALK_SKIP: skip this dentry and its children 1037 */ 1038 enum d_walk_ret { 1039 D_WALK_CONTINUE, 1040 D_WALK_QUIT, 1041 D_WALK_NORETRY, 1042 D_WALK_SKIP, 1043 }; 1044 1045 /** 1046 * d_walk - walk the dentry tree 1047 * @parent: start of walk 1048 * @data: data passed to @enter() and @finish() 1049 * @enter: callback when first entering the dentry 1050 * @finish: callback when successfully finished the walk 1051 * 1052 * The @enter() and @finish() callbacks are called with d_lock held. 1053 */ 1054 static void d_walk(struct dentry *parent, void *data, 1055 enum d_walk_ret (*enter)(void *, struct dentry *), 1056 void (*finish)(void *)) 1057 { 1058 struct dentry *this_parent; 1059 struct list_head *next; 1060 unsigned seq = 0; 1061 enum d_walk_ret ret; 1062 bool retry = true; 1063 1064 again: 1065 read_seqbegin_or_lock(&rename_lock, &seq); 1066 this_parent = parent; 1067 spin_lock(&this_parent->d_lock); 1068 1069 ret = enter(data, this_parent); 1070 switch (ret) { 1071 case D_WALK_CONTINUE: 1072 break; 1073 case D_WALK_QUIT: 1074 case D_WALK_SKIP: 1075 goto out_unlock; 1076 case D_WALK_NORETRY: 1077 retry = false; 1078 break; 1079 } 1080 repeat: 1081 next = this_parent->d_subdirs.next; 1082 resume: 1083 while (next != &this_parent->d_subdirs) { 1084 struct list_head *tmp = next; 1085 struct dentry *dentry = list_entry(tmp, struct dentry, d_u.d_child); 1086 next = tmp->next; 1087 1088 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED); 1089 1090 ret = enter(data, dentry); 1091 switch (ret) { 1092 case D_WALK_CONTINUE: 1093 break; 1094 case D_WALK_QUIT: 1095 spin_unlock(&dentry->d_lock); 1096 goto out_unlock; 1097 case D_WALK_NORETRY: 1098 retry = false; 1099 break; 1100 case D_WALK_SKIP: 1101 spin_unlock(&dentry->d_lock); 1102 continue; 1103 } 1104 1105 if (!list_empty(&dentry->d_subdirs)) { 1106 spin_unlock(&this_parent->d_lock); 1107 spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_); 1108 this_parent = dentry; 1109 spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_); 1110 goto repeat; 1111 } 1112 spin_unlock(&dentry->d_lock); 1113 } 1114 /* 1115 * All done at this level ... ascend and resume the search. 1116 */ 1117 if (this_parent != parent) { 1118 struct dentry *child = this_parent; 1119 this_parent = child->d_parent; 1120 1121 rcu_read_lock(); 1122 spin_unlock(&child->d_lock); 1123 spin_lock(&this_parent->d_lock); 1124 1125 /* 1126 * might go back up the wrong parent if we have had a rename 1127 * or deletion 1128 */ 1129 if (this_parent != child->d_parent || 1130 (child->d_flags & DCACHE_DENTRY_KILLED) || 1131 need_seqretry(&rename_lock, seq)) { 1132 spin_unlock(&this_parent->d_lock); 1133 rcu_read_unlock(); 1134 goto rename_retry; 1135 } 1136 rcu_read_unlock(); 1137 next = child->d_u.d_child.next; 1138 goto resume; 1139 } 1140 if (need_seqretry(&rename_lock, seq)) { 1141 spin_unlock(&this_parent->d_lock); 1142 goto rename_retry; 1143 } 1144 if (finish) 1145 finish(data); 1146 1147 out_unlock: 1148 spin_unlock(&this_parent->d_lock); 1149 done_seqretry(&rename_lock, seq); 1150 return; 1151 1152 rename_retry: 1153 if (!retry) 1154 return; 1155 seq = 1; 1156 goto again; 1157 } 1158 1159 /* 1160 * Search for at least 1 mount point in the dentry's subdirs. 1161 * We descend to the next level whenever the d_subdirs 1162 * list is non-empty and continue searching. 1163 */ 1164 1165 static enum d_walk_ret check_mount(void *data, struct dentry *dentry) 1166 { 1167 int *ret = data; 1168 if (d_mountpoint(dentry)) { 1169 *ret = 1; 1170 return D_WALK_QUIT; 1171 } 1172 return D_WALK_CONTINUE; 1173 } 1174 1175 /** 1176 * have_submounts - check for mounts over a dentry 1177 * @parent: dentry to check. 1178 * 1179 * Return true if the parent or its subdirectories contain 1180 * a mount point 1181 */ 1182 int have_submounts(struct dentry *parent) 1183 { 1184 int ret = 0; 1185 1186 d_walk(parent, &ret, check_mount, NULL); 1187 1188 return ret; 1189 } 1190 EXPORT_SYMBOL(have_submounts); 1191 1192 /* 1193 * Called by mount code to set a mountpoint and check if the mountpoint is 1194 * reachable (e.g. NFS can unhash a directory dentry and then the complete 1195 * subtree can become unreachable). 1196 * 1197 * Only one of check_submounts_and_drop() and d_set_mounted() must succeed. For 1198 * this reason take rename_lock and d_lock on dentry and ancestors. 1199 */ 1200 int d_set_mounted(struct dentry *dentry) 1201 { 1202 struct dentry *p; 1203 int ret = -ENOENT; 1204 write_seqlock(&rename_lock); 1205 for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) { 1206 /* Need exclusion wrt. check_submounts_and_drop() */ 1207 spin_lock(&p->d_lock); 1208 if (unlikely(d_unhashed(p))) { 1209 spin_unlock(&p->d_lock); 1210 goto out; 1211 } 1212 spin_unlock(&p->d_lock); 1213 } 1214 spin_lock(&dentry->d_lock); 1215 if (!d_unlinked(dentry)) { 1216 dentry->d_flags |= DCACHE_MOUNTED; 1217 ret = 0; 1218 } 1219 spin_unlock(&dentry->d_lock); 1220 out: 1221 write_sequnlock(&rename_lock); 1222 return ret; 1223 } 1224 1225 /* 1226 * Search the dentry child list of the specified parent, 1227 * and move any unused dentries to the end of the unused 1228 * list for prune_dcache(). We descend to the next level 1229 * whenever the d_subdirs list is non-empty and continue 1230 * searching. 1231 * 1232 * It returns zero iff there are no unused children, 1233 * otherwise it returns the number of children moved to 1234 * the end of the unused list. This may not be the total 1235 * number of unused children, because select_parent can 1236 * drop the lock and return early due to latency 1237 * constraints. 1238 */ 1239 1240 struct select_data { 1241 struct dentry *start; 1242 struct list_head dispose; 1243 int found; 1244 }; 1245 1246 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry) 1247 { 1248 struct select_data *data = _data; 1249 enum d_walk_ret ret = D_WALK_CONTINUE; 1250 1251 if (data->start == dentry) 1252 goto out; 1253 1254 if (dentry->d_flags & DCACHE_SHRINK_LIST) { 1255 data->found++; 1256 } else { 1257 if (dentry->d_flags & DCACHE_LRU_LIST) 1258 d_lru_del(dentry); 1259 if (!dentry->d_lockref.count) { 1260 d_shrink_add(dentry, &data->dispose); 1261 data->found++; 1262 } 1263 } 1264 /* 1265 * We can return to the caller if we have found some (this 1266 * ensures forward progress). We'll be coming back to find 1267 * the rest. 1268 */ 1269 if (!list_empty(&data->dispose)) 1270 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY; 1271 out: 1272 return ret; 1273 } 1274 1275 /** 1276 * shrink_dcache_parent - prune dcache 1277 * @parent: parent of entries to prune 1278 * 1279 * Prune the dcache to remove unused children of the parent dentry. 1280 */ 1281 void shrink_dcache_parent(struct dentry *parent) 1282 { 1283 for (;;) { 1284 struct select_data data; 1285 1286 INIT_LIST_HEAD(&data.dispose); 1287 data.start = parent; 1288 data.found = 0; 1289 1290 d_walk(parent, &data, select_collect, NULL); 1291 if (!data.found) 1292 break; 1293 1294 shrink_dentry_list(&data.dispose); 1295 cond_resched(); 1296 } 1297 } 1298 EXPORT_SYMBOL(shrink_dcache_parent); 1299 1300 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry) 1301 { 1302 /* it has busy descendents; complain about those instead */ 1303 if (!list_empty(&dentry->d_subdirs)) 1304 return D_WALK_CONTINUE; 1305 1306 /* root with refcount 1 is fine */ 1307 if (dentry == _data && dentry->d_lockref.count == 1) 1308 return D_WALK_CONTINUE; 1309 1310 printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} " 1311 " still in use (%d) [unmount of %s %s]\n", 1312 dentry, 1313 dentry->d_inode ? 1314 dentry->d_inode->i_ino : 0UL, 1315 dentry, 1316 dentry->d_lockref.count, 1317 dentry->d_sb->s_type->name, 1318 dentry->d_sb->s_id); 1319 WARN_ON(1); 1320 return D_WALK_CONTINUE; 1321 } 1322 1323 static void do_one_tree(struct dentry *dentry) 1324 { 1325 shrink_dcache_parent(dentry); 1326 d_walk(dentry, dentry, umount_check, NULL); 1327 d_drop(dentry); 1328 dput(dentry); 1329 } 1330 1331 /* 1332 * destroy the dentries attached to a superblock on unmounting 1333 */ 1334 void shrink_dcache_for_umount(struct super_block *sb) 1335 { 1336 struct dentry *dentry; 1337 1338 WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked"); 1339 1340 dentry = sb->s_root; 1341 sb->s_root = NULL; 1342 do_one_tree(dentry); 1343 1344 while (!hlist_bl_empty(&sb->s_anon)) { 1345 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_anon), struct dentry, d_hash)); 1346 do_one_tree(dentry); 1347 } 1348 } 1349 1350 static enum d_walk_ret check_and_collect(void *_data, struct dentry *dentry) 1351 { 1352 struct select_data *data = _data; 1353 1354 if (d_mountpoint(dentry)) { 1355 data->found = -EBUSY; 1356 return D_WALK_QUIT; 1357 } 1358 1359 return select_collect(_data, dentry); 1360 } 1361 1362 static void check_and_drop(void *_data) 1363 { 1364 struct select_data *data = _data; 1365 1366 if (d_mountpoint(data->start)) 1367 data->found = -EBUSY; 1368 if (!data->found) 1369 __d_drop(data->start); 1370 } 1371 1372 /** 1373 * check_submounts_and_drop - prune dcache, check for submounts and drop 1374 * 1375 * All done as a single atomic operation relative to has_unlinked_ancestor(). 1376 * Returns 0 if successfully unhashed @parent. If there were submounts then 1377 * return -EBUSY. 1378 * 1379 * @dentry: dentry to prune and drop 1380 */ 1381 int check_submounts_and_drop(struct dentry *dentry) 1382 { 1383 int ret = 0; 1384 1385 /* Negative dentries can be dropped without further checks */ 1386 if (!dentry->d_inode) { 1387 d_drop(dentry); 1388 goto out; 1389 } 1390 1391 for (;;) { 1392 struct select_data data; 1393 1394 INIT_LIST_HEAD(&data.dispose); 1395 data.start = dentry; 1396 data.found = 0; 1397 1398 d_walk(dentry, &data, check_and_collect, check_and_drop); 1399 ret = data.found; 1400 1401 if (!list_empty(&data.dispose)) 1402 shrink_dentry_list(&data.dispose); 1403 1404 if (ret <= 0) 1405 break; 1406 1407 cond_resched(); 1408 } 1409 1410 out: 1411 return ret; 1412 } 1413 EXPORT_SYMBOL(check_submounts_and_drop); 1414 1415 /** 1416 * __d_alloc - allocate a dcache entry 1417 * @sb: filesystem it will belong to 1418 * @name: qstr of the name 1419 * 1420 * Allocates a dentry. It returns %NULL if there is insufficient memory 1421 * available. On a success the dentry is returned. The name passed in is 1422 * copied and the copy passed in may be reused after this call. 1423 */ 1424 1425 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name) 1426 { 1427 struct dentry *dentry; 1428 char *dname; 1429 1430 dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL); 1431 if (!dentry) 1432 return NULL; 1433 1434 /* 1435 * We guarantee that the inline name is always NUL-terminated. 1436 * This way the memcpy() done by the name switching in rename 1437 * will still always have a NUL at the end, even if we might 1438 * be overwriting an internal NUL character 1439 */ 1440 dentry->d_iname[DNAME_INLINE_LEN-1] = 0; 1441 if (name->len > DNAME_INLINE_LEN-1) { 1442 dname = kmalloc(name->len + 1, GFP_KERNEL); 1443 if (!dname) { 1444 kmem_cache_free(dentry_cache, dentry); 1445 return NULL; 1446 } 1447 } else { 1448 dname = dentry->d_iname; 1449 } 1450 1451 dentry->d_name.len = name->len; 1452 dentry->d_name.hash = name->hash; 1453 memcpy(dname, name->name, name->len); 1454 dname[name->len] = 0; 1455 1456 /* Make sure we always see the terminating NUL character */ 1457 smp_wmb(); 1458 dentry->d_name.name = dname; 1459 1460 dentry->d_lockref.count = 1; 1461 dentry->d_flags = 0; 1462 spin_lock_init(&dentry->d_lock); 1463 seqcount_init(&dentry->d_seq); 1464 dentry->d_inode = NULL; 1465 dentry->d_parent = dentry; 1466 dentry->d_sb = sb; 1467 dentry->d_op = NULL; 1468 dentry->d_fsdata = NULL; 1469 INIT_HLIST_BL_NODE(&dentry->d_hash); 1470 INIT_LIST_HEAD(&dentry->d_lru); 1471 INIT_LIST_HEAD(&dentry->d_subdirs); 1472 INIT_HLIST_NODE(&dentry->d_alias); 1473 INIT_LIST_HEAD(&dentry->d_u.d_child); 1474 d_set_d_op(dentry, dentry->d_sb->s_d_op); 1475 1476 this_cpu_inc(nr_dentry); 1477 1478 return dentry; 1479 } 1480 1481 /** 1482 * d_alloc - allocate a dcache entry 1483 * @parent: parent of entry to allocate 1484 * @name: qstr of the name 1485 * 1486 * Allocates a dentry. It returns %NULL if there is insufficient memory 1487 * available. On a success the dentry is returned. The name passed in is 1488 * copied and the copy passed in may be reused after this call. 1489 */ 1490 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name) 1491 { 1492 struct dentry *dentry = __d_alloc(parent->d_sb, name); 1493 if (!dentry) 1494 return NULL; 1495 1496 spin_lock(&parent->d_lock); 1497 /* 1498 * don't need child lock because it is not subject 1499 * to concurrency here 1500 */ 1501 __dget_dlock(parent); 1502 dentry->d_parent = parent; 1503 list_add(&dentry->d_u.d_child, &parent->d_subdirs); 1504 spin_unlock(&parent->d_lock); 1505 1506 return dentry; 1507 } 1508 EXPORT_SYMBOL(d_alloc); 1509 1510 /** 1511 * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems) 1512 * @sb: the superblock 1513 * @name: qstr of the name 1514 * 1515 * For a filesystem that just pins its dentries in memory and never 1516 * performs lookups at all, return an unhashed IS_ROOT dentry. 1517 */ 1518 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name) 1519 { 1520 return __d_alloc(sb, name); 1521 } 1522 EXPORT_SYMBOL(d_alloc_pseudo); 1523 1524 struct dentry *d_alloc_name(struct dentry *parent, const char *name) 1525 { 1526 struct qstr q; 1527 1528 q.name = name; 1529 q.len = strlen(name); 1530 q.hash = full_name_hash(q.name, q.len); 1531 return d_alloc(parent, &q); 1532 } 1533 EXPORT_SYMBOL(d_alloc_name); 1534 1535 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op) 1536 { 1537 WARN_ON_ONCE(dentry->d_op); 1538 WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH | 1539 DCACHE_OP_COMPARE | 1540 DCACHE_OP_REVALIDATE | 1541 DCACHE_OP_WEAK_REVALIDATE | 1542 DCACHE_OP_DELETE )); 1543 dentry->d_op = op; 1544 if (!op) 1545 return; 1546 if (op->d_hash) 1547 dentry->d_flags |= DCACHE_OP_HASH; 1548 if (op->d_compare) 1549 dentry->d_flags |= DCACHE_OP_COMPARE; 1550 if (op->d_revalidate) 1551 dentry->d_flags |= DCACHE_OP_REVALIDATE; 1552 if (op->d_weak_revalidate) 1553 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE; 1554 if (op->d_delete) 1555 dentry->d_flags |= DCACHE_OP_DELETE; 1556 if (op->d_prune) 1557 dentry->d_flags |= DCACHE_OP_PRUNE; 1558 1559 } 1560 EXPORT_SYMBOL(d_set_d_op); 1561 1562 static unsigned d_flags_for_inode(struct inode *inode) 1563 { 1564 unsigned add_flags = DCACHE_FILE_TYPE; 1565 1566 if (!inode) 1567 return DCACHE_MISS_TYPE; 1568 1569 if (S_ISDIR(inode->i_mode)) { 1570 add_flags = DCACHE_DIRECTORY_TYPE; 1571 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) { 1572 if (unlikely(!inode->i_op->lookup)) 1573 add_flags = DCACHE_AUTODIR_TYPE; 1574 else 1575 inode->i_opflags |= IOP_LOOKUP; 1576 } 1577 } else if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) { 1578 if (unlikely(inode->i_op->follow_link)) 1579 add_flags = DCACHE_SYMLINK_TYPE; 1580 else 1581 inode->i_opflags |= IOP_NOFOLLOW; 1582 } 1583 1584 if (unlikely(IS_AUTOMOUNT(inode))) 1585 add_flags |= DCACHE_NEED_AUTOMOUNT; 1586 return add_flags; 1587 } 1588 1589 static void __d_instantiate(struct dentry *dentry, struct inode *inode) 1590 { 1591 unsigned add_flags = d_flags_for_inode(inode); 1592 1593 spin_lock(&dentry->d_lock); 1594 __d_set_type(dentry, add_flags); 1595 if (inode) 1596 hlist_add_head(&dentry->d_alias, &inode->i_dentry); 1597 dentry->d_inode = inode; 1598 dentry_rcuwalk_barrier(dentry); 1599 spin_unlock(&dentry->d_lock); 1600 fsnotify_d_instantiate(dentry, inode); 1601 } 1602 1603 /** 1604 * d_instantiate - fill in inode information for a dentry 1605 * @entry: dentry to complete 1606 * @inode: inode to attach to this dentry 1607 * 1608 * Fill in inode information in the entry. 1609 * 1610 * This turns negative dentries into productive full members 1611 * of society. 1612 * 1613 * NOTE! This assumes that the inode count has been incremented 1614 * (or otherwise set) by the caller to indicate that it is now 1615 * in use by the dcache. 1616 */ 1617 1618 void d_instantiate(struct dentry *entry, struct inode * inode) 1619 { 1620 BUG_ON(!hlist_unhashed(&entry->d_alias)); 1621 if (inode) 1622 spin_lock(&inode->i_lock); 1623 __d_instantiate(entry, inode); 1624 if (inode) 1625 spin_unlock(&inode->i_lock); 1626 security_d_instantiate(entry, inode); 1627 } 1628 EXPORT_SYMBOL(d_instantiate); 1629 1630 /** 1631 * d_instantiate_unique - instantiate a non-aliased dentry 1632 * @entry: dentry to instantiate 1633 * @inode: inode to attach to this dentry 1634 * 1635 * Fill in inode information in the entry. On success, it returns NULL. 1636 * If an unhashed alias of "entry" already exists, then we return the 1637 * aliased dentry instead and drop one reference to inode. 1638 * 1639 * Note that in order to avoid conflicts with rename() etc, the caller 1640 * had better be holding the parent directory semaphore. 1641 * 1642 * This also assumes that the inode count has been incremented 1643 * (or otherwise set) by the caller to indicate that it is now 1644 * in use by the dcache. 1645 */ 1646 static struct dentry *__d_instantiate_unique(struct dentry *entry, 1647 struct inode *inode) 1648 { 1649 struct dentry *alias; 1650 int len = entry->d_name.len; 1651 const char *name = entry->d_name.name; 1652 unsigned int hash = entry->d_name.hash; 1653 1654 if (!inode) { 1655 __d_instantiate(entry, NULL); 1656 return NULL; 1657 } 1658 1659 hlist_for_each_entry(alias, &inode->i_dentry, d_alias) { 1660 /* 1661 * Don't need alias->d_lock here, because aliases with 1662 * d_parent == entry->d_parent are not subject to name or 1663 * parent changes, because the parent inode i_mutex is held. 1664 */ 1665 if (alias->d_name.hash != hash) 1666 continue; 1667 if (alias->d_parent != entry->d_parent) 1668 continue; 1669 if (alias->d_name.len != len) 1670 continue; 1671 if (dentry_cmp(alias, name, len)) 1672 continue; 1673 __dget(alias); 1674 return alias; 1675 } 1676 1677 __d_instantiate(entry, inode); 1678 return NULL; 1679 } 1680 1681 struct dentry *d_instantiate_unique(struct dentry *entry, struct inode *inode) 1682 { 1683 struct dentry *result; 1684 1685 BUG_ON(!hlist_unhashed(&entry->d_alias)); 1686 1687 if (inode) 1688 spin_lock(&inode->i_lock); 1689 result = __d_instantiate_unique(entry, inode); 1690 if (inode) 1691 spin_unlock(&inode->i_lock); 1692 1693 if (!result) { 1694 security_d_instantiate(entry, inode); 1695 return NULL; 1696 } 1697 1698 BUG_ON(!d_unhashed(result)); 1699 iput(inode); 1700 return result; 1701 } 1702 1703 EXPORT_SYMBOL(d_instantiate_unique); 1704 1705 /** 1706 * d_instantiate_no_diralias - instantiate a non-aliased dentry 1707 * @entry: dentry to complete 1708 * @inode: inode to attach to this dentry 1709 * 1710 * Fill in inode information in the entry. If a directory alias is found, then 1711 * return an error (and drop inode). Together with d_materialise_unique() this 1712 * guarantees that a directory inode may never have more than one alias. 1713 */ 1714 int d_instantiate_no_diralias(struct dentry *entry, struct inode *inode) 1715 { 1716 BUG_ON(!hlist_unhashed(&entry->d_alias)); 1717 1718 spin_lock(&inode->i_lock); 1719 if (S_ISDIR(inode->i_mode) && !hlist_empty(&inode->i_dentry)) { 1720 spin_unlock(&inode->i_lock); 1721 iput(inode); 1722 return -EBUSY; 1723 } 1724 __d_instantiate(entry, inode); 1725 spin_unlock(&inode->i_lock); 1726 security_d_instantiate(entry, inode); 1727 1728 return 0; 1729 } 1730 EXPORT_SYMBOL(d_instantiate_no_diralias); 1731 1732 struct dentry *d_make_root(struct inode *root_inode) 1733 { 1734 struct dentry *res = NULL; 1735 1736 if (root_inode) { 1737 static const struct qstr name = QSTR_INIT("/", 1); 1738 1739 res = __d_alloc(root_inode->i_sb, &name); 1740 if (res) 1741 d_instantiate(res, root_inode); 1742 else 1743 iput(root_inode); 1744 } 1745 return res; 1746 } 1747 EXPORT_SYMBOL(d_make_root); 1748 1749 static struct dentry * __d_find_any_alias(struct inode *inode) 1750 { 1751 struct dentry *alias; 1752 1753 if (hlist_empty(&inode->i_dentry)) 1754 return NULL; 1755 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_alias); 1756 __dget(alias); 1757 return alias; 1758 } 1759 1760 /** 1761 * d_find_any_alias - find any alias for a given inode 1762 * @inode: inode to find an alias for 1763 * 1764 * If any aliases exist for the given inode, take and return a 1765 * reference for one of them. If no aliases exist, return %NULL. 1766 */ 1767 struct dentry *d_find_any_alias(struct inode *inode) 1768 { 1769 struct dentry *de; 1770 1771 spin_lock(&inode->i_lock); 1772 de = __d_find_any_alias(inode); 1773 spin_unlock(&inode->i_lock); 1774 return de; 1775 } 1776 EXPORT_SYMBOL(d_find_any_alias); 1777 1778 static struct dentry *__d_obtain_alias(struct inode *inode, int disconnected) 1779 { 1780 static const struct qstr anonstring = QSTR_INIT("/", 1); 1781 struct dentry *tmp; 1782 struct dentry *res; 1783 unsigned add_flags; 1784 1785 if (!inode) 1786 return ERR_PTR(-ESTALE); 1787 if (IS_ERR(inode)) 1788 return ERR_CAST(inode); 1789 1790 res = d_find_any_alias(inode); 1791 if (res) 1792 goto out_iput; 1793 1794 tmp = __d_alloc(inode->i_sb, &anonstring); 1795 if (!tmp) { 1796 res = ERR_PTR(-ENOMEM); 1797 goto out_iput; 1798 } 1799 1800 spin_lock(&inode->i_lock); 1801 res = __d_find_any_alias(inode); 1802 if (res) { 1803 spin_unlock(&inode->i_lock); 1804 dput(tmp); 1805 goto out_iput; 1806 } 1807 1808 /* attach a disconnected dentry */ 1809 add_flags = d_flags_for_inode(inode); 1810 1811 if (disconnected) 1812 add_flags |= DCACHE_DISCONNECTED; 1813 1814 spin_lock(&tmp->d_lock); 1815 tmp->d_inode = inode; 1816 tmp->d_flags |= add_flags; 1817 hlist_add_head(&tmp->d_alias, &inode->i_dentry); 1818 hlist_bl_lock(&tmp->d_sb->s_anon); 1819 hlist_bl_add_head(&tmp->d_hash, &tmp->d_sb->s_anon); 1820 hlist_bl_unlock(&tmp->d_sb->s_anon); 1821 spin_unlock(&tmp->d_lock); 1822 spin_unlock(&inode->i_lock); 1823 security_d_instantiate(tmp, inode); 1824 1825 return tmp; 1826 1827 out_iput: 1828 if (res && !IS_ERR(res)) 1829 security_d_instantiate(res, inode); 1830 iput(inode); 1831 return res; 1832 } 1833 1834 /** 1835 * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode 1836 * @inode: inode to allocate the dentry for 1837 * 1838 * Obtain a dentry for an inode resulting from NFS filehandle conversion or 1839 * similar open by handle operations. The returned dentry may be anonymous, 1840 * or may have a full name (if the inode was already in the cache). 1841 * 1842 * When called on a directory inode, we must ensure that the inode only ever 1843 * has one dentry. If a dentry is found, that is returned instead of 1844 * allocating a new one. 1845 * 1846 * On successful return, the reference to the inode has been transferred 1847 * to the dentry. In case of an error the reference on the inode is released. 1848 * To make it easier to use in export operations a %NULL or IS_ERR inode may 1849 * be passed in and the error will be propagated to the return value, 1850 * with a %NULL @inode replaced by ERR_PTR(-ESTALE). 1851 */ 1852 struct dentry *d_obtain_alias(struct inode *inode) 1853 { 1854 return __d_obtain_alias(inode, 1); 1855 } 1856 EXPORT_SYMBOL(d_obtain_alias); 1857 1858 /** 1859 * d_obtain_root - find or allocate a dentry for a given inode 1860 * @inode: inode to allocate the dentry for 1861 * 1862 * Obtain an IS_ROOT dentry for the root of a filesystem. 1863 * 1864 * We must ensure that directory inodes only ever have one dentry. If a 1865 * dentry is found, that is returned instead of allocating a new one. 1866 * 1867 * On successful return, the reference to the inode has been transferred 1868 * to the dentry. In case of an error the reference on the inode is 1869 * released. A %NULL or IS_ERR inode may be passed in and will be the 1870 * error will be propagate to the return value, with a %NULL @inode 1871 * replaced by ERR_PTR(-ESTALE). 1872 */ 1873 struct dentry *d_obtain_root(struct inode *inode) 1874 { 1875 return __d_obtain_alias(inode, 0); 1876 } 1877 EXPORT_SYMBOL(d_obtain_root); 1878 1879 /** 1880 * d_add_ci - lookup or allocate new dentry with case-exact name 1881 * @inode: the inode case-insensitive lookup has found 1882 * @dentry: the negative dentry that was passed to the parent's lookup func 1883 * @name: the case-exact name to be associated with the returned dentry 1884 * 1885 * This is to avoid filling the dcache with case-insensitive names to the 1886 * same inode, only the actual correct case is stored in the dcache for 1887 * case-insensitive filesystems. 1888 * 1889 * For a case-insensitive lookup match and if the the case-exact dentry 1890 * already exists in in the dcache, use it and return it. 1891 * 1892 * If no entry exists with the exact case name, allocate new dentry with 1893 * the exact case, and return the spliced entry. 1894 */ 1895 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode, 1896 struct qstr *name) 1897 { 1898 struct dentry *found; 1899 struct dentry *new; 1900 1901 /* 1902 * First check if a dentry matching the name already exists, 1903 * if not go ahead and create it now. 1904 */ 1905 found = d_hash_and_lookup(dentry->d_parent, name); 1906 if (unlikely(IS_ERR(found))) 1907 goto err_out; 1908 if (!found) { 1909 new = d_alloc(dentry->d_parent, name); 1910 if (!new) { 1911 found = ERR_PTR(-ENOMEM); 1912 goto err_out; 1913 } 1914 1915 found = d_splice_alias(inode, new); 1916 if (found) { 1917 dput(new); 1918 return found; 1919 } 1920 return new; 1921 } 1922 1923 /* 1924 * If a matching dentry exists, and it's not negative use it. 1925 * 1926 * Decrement the reference count to balance the iget() done 1927 * earlier on. 1928 */ 1929 if (found->d_inode) { 1930 if (unlikely(found->d_inode != inode)) { 1931 /* This can't happen because bad inodes are unhashed. */ 1932 BUG_ON(!is_bad_inode(inode)); 1933 BUG_ON(!is_bad_inode(found->d_inode)); 1934 } 1935 iput(inode); 1936 return found; 1937 } 1938 1939 /* 1940 * Negative dentry: instantiate it unless the inode is a directory and 1941 * already has a dentry. 1942 */ 1943 new = d_splice_alias(inode, found); 1944 if (new) { 1945 dput(found); 1946 found = new; 1947 } 1948 return found; 1949 1950 err_out: 1951 iput(inode); 1952 return found; 1953 } 1954 EXPORT_SYMBOL(d_add_ci); 1955 1956 /* 1957 * Do the slow-case of the dentry name compare. 1958 * 1959 * Unlike the dentry_cmp() function, we need to atomically 1960 * load the name and length information, so that the 1961 * filesystem can rely on them, and can use the 'name' and 1962 * 'len' information without worrying about walking off the 1963 * end of memory etc. 1964 * 1965 * Thus the read_seqcount_retry() and the "duplicate" info 1966 * in arguments (the low-level filesystem should not look 1967 * at the dentry inode or name contents directly, since 1968 * rename can change them while we're in RCU mode). 1969 */ 1970 enum slow_d_compare { 1971 D_COMP_OK, 1972 D_COMP_NOMATCH, 1973 D_COMP_SEQRETRY, 1974 }; 1975 1976 static noinline enum slow_d_compare slow_dentry_cmp( 1977 const struct dentry *parent, 1978 struct dentry *dentry, 1979 unsigned int seq, 1980 const struct qstr *name) 1981 { 1982 int tlen = dentry->d_name.len; 1983 const char *tname = dentry->d_name.name; 1984 1985 if (read_seqcount_retry(&dentry->d_seq, seq)) { 1986 cpu_relax(); 1987 return D_COMP_SEQRETRY; 1988 } 1989 if (parent->d_op->d_compare(parent, dentry, tlen, tname, name)) 1990 return D_COMP_NOMATCH; 1991 return D_COMP_OK; 1992 } 1993 1994 /** 1995 * __d_lookup_rcu - search for a dentry (racy, store-free) 1996 * @parent: parent dentry 1997 * @name: qstr of name we wish to find 1998 * @seqp: returns d_seq value at the point where the dentry was found 1999 * Returns: dentry, or NULL 2000 * 2001 * __d_lookup_rcu is the dcache lookup function for rcu-walk name 2002 * resolution (store-free path walking) design described in 2003 * Documentation/filesystems/path-lookup.txt. 2004 * 2005 * This is not to be used outside core vfs. 2006 * 2007 * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock 2008 * held, and rcu_read_lock held. The returned dentry must not be stored into 2009 * without taking d_lock and checking d_seq sequence count against @seq 2010 * returned here. 2011 * 2012 * A refcount may be taken on the found dentry with the d_rcu_to_refcount 2013 * function. 2014 * 2015 * Alternatively, __d_lookup_rcu may be called again to look up the child of 2016 * the returned dentry, so long as its parent's seqlock is checked after the 2017 * child is looked up. Thus, an interlocking stepping of sequence lock checks 2018 * is formed, giving integrity down the path walk. 2019 * 2020 * NOTE! The caller *has* to check the resulting dentry against the sequence 2021 * number we've returned before using any of the resulting dentry state! 2022 */ 2023 struct dentry *__d_lookup_rcu(const struct dentry *parent, 2024 const struct qstr *name, 2025 unsigned *seqp) 2026 { 2027 u64 hashlen = name->hash_len; 2028 const unsigned char *str = name->name; 2029 struct hlist_bl_head *b = d_hash(parent, hashlen_hash(hashlen)); 2030 struct hlist_bl_node *node; 2031 struct dentry *dentry; 2032 2033 /* 2034 * Note: There is significant duplication with __d_lookup_rcu which is 2035 * required to prevent single threaded performance regressions 2036 * especially on architectures where smp_rmb (in seqcounts) are costly. 2037 * Keep the two functions in sync. 2038 */ 2039 2040 /* 2041 * The hash list is protected using RCU. 2042 * 2043 * Carefully use d_seq when comparing a candidate dentry, to avoid 2044 * races with d_move(). 2045 * 2046 * It is possible that concurrent renames can mess up our list 2047 * walk here and result in missing our dentry, resulting in the 2048 * false-negative result. d_lookup() protects against concurrent 2049 * renames using rename_lock seqlock. 2050 * 2051 * See Documentation/filesystems/path-lookup.txt for more details. 2052 */ 2053 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) { 2054 unsigned seq; 2055 2056 seqretry: 2057 /* 2058 * The dentry sequence count protects us from concurrent 2059 * renames, and thus protects parent and name fields. 2060 * 2061 * The caller must perform a seqcount check in order 2062 * to do anything useful with the returned dentry. 2063 * 2064 * NOTE! We do a "raw" seqcount_begin here. That means that 2065 * we don't wait for the sequence count to stabilize if it 2066 * is in the middle of a sequence change. If we do the slow 2067 * dentry compare, we will do seqretries until it is stable, 2068 * and if we end up with a successful lookup, we actually 2069 * want to exit RCU lookup anyway. 2070 */ 2071 seq = raw_seqcount_begin(&dentry->d_seq); 2072 if (dentry->d_parent != parent) 2073 continue; 2074 if (d_unhashed(dentry)) 2075 continue; 2076 2077 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) { 2078 if (dentry->d_name.hash != hashlen_hash(hashlen)) 2079 continue; 2080 *seqp = seq; 2081 switch (slow_dentry_cmp(parent, dentry, seq, name)) { 2082 case D_COMP_OK: 2083 return dentry; 2084 case D_COMP_NOMATCH: 2085 continue; 2086 default: 2087 goto seqretry; 2088 } 2089 } 2090 2091 if (dentry->d_name.hash_len != hashlen) 2092 continue; 2093 *seqp = seq; 2094 if (!dentry_cmp(dentry, str, hashlen_len(hashlen))) 2095 return dentry; 2096 } 2097 return NULL; 2098 } 2099 2100 /** 2101 * d_lookup - search for a dentry 2102 * @parent: parent dentry 2103 * @name: qstr of name we wish to find 2104 * Returns: dentry, or NULL 2105 * 2106 * d_lookup searches the children of the parent dentry for the name in 2107 * question. If the dentry is found its reference count is incremented and the 2108 * dentry is returned. The caller must use dput to free the entry when it has 2109 * finished using it. %NULL is returned if the dentry does not exist. 2110 */ 2111 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name) 2112 { 2113 struct dentry *dentry; 2114 unsigned seq; 2115 2116 do { 2117 seq = read_seqbegin(&rename_lock); 2118 dentry = __d_lookup(parent, name); 2119 if (dentry) 2120 break; 2121 } while (read_seqretry(&rename_lock, seq)); 2122 return dentry; 2123 } 2124 EXPORT_SYMBOL(d_lookup); 2125 2126 /** 2127 * __d_lookup - search for a dentry (racy) 2128 * @parent: parent dentry 2129 * @name: qstr of name we wish to find 2130 * Returns: dentry, or NULL 2131 * 2132 * __d_lookup is like d_lookup, however it may (rarely) return a 2133 * false-negative result due to unrelated rename activity. 2134 * 2135 * __d_lookup is slightly faster by avoiding rename_lock read seqlock, 2136 * however it must be used carefully, eg. with a following d_lookup in 2137 * the case of failure. 2138 * 2139 * __d_lookup callers must be commented. 2140 */ 2141 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name) 2142 { 2143 unsigned int len = name->len; 2144 unsigned int hash = name->hash; 2145 const unsigned char *str = name->name; 2146 struct hlist_bl_head *b = d_hash(parent, hash); 2147 struct hlist_bl_node *node; 2148 struct dentry *found = NULL; 2149 struct dentry *dentry; 2150 2151 /* 2152 * Note: There is significant duplication with __d_lookup_rcu which is 2153 * required to prevent single threaded performance regressions 2154 * especially on architectures where smp_rmb (in seqcounts) are costly. 2155 * Keep the two functions in sync. 2156 */ 2157 2158 /* 2159 * The hash list is protected using RCU. 2160 * 2161 * Take d_lock when comparing a candidate dentry, to avoid races 2162 * with d_move(). 2163 * 2164 * It is possible that concurrent renames can mess up our list 2165 * walk here and result in missing our dentry, resulting in the 2166 * false-negative result. d_lookup() protects against concurrent 2167 * renames using rename_lock seqlock. 2168 * 2169 * See Documentation/filesystems/path-lookup.txt for more details. 2170 */ 2171 rcu_read_lock(); 2172 2173 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) { 2174 2175 if (dentry->d_name.hash != hash) 2176 continue; 2177 2178 spin_lock(&dentry->d_lock); 2179 if (dentry->d_parent != parent) 2180 goto next; 2181 if (d_unhashed(dentry)) 2182 goto next; 2183 2184 /* 2185 * It is safe to compare names since d_move() cannot 2186 * change the qstr (protected by d_lock). 2187 */ 2188 if (parent->d_flags & DCACHE_OP_COMPARE) { 2189 int tlen = dentry->d_name.len; 2190 const char *tname = dentry->d_name.name; 2191 if (parent->d_op->d_compare(parent, dentry, tlen, tname, name)) 2192 goto next; 2193 } else { 2194 if (dentry->d_name.len != len) 2195 goto next; 2196 if (dentry_cmp(dentry, str, len)) 2197 goto next; 2198 } 2199 2200 dentry->d_lockref.count++; 2201 found = dentry; 2202 spin_unlock(&dentry->d_lock); 2203 break; 2204 next: 2205 spin_unlock(&dentry->d_lock); 2206 } 2207 rcu_read_unlock(); 2208 2209 return found; 2210 } 2211 2212 /** 2213 * d_hash_and_lookup - hash the qstr then search for a dentry 2214 * @dir: Directory to search in 2215 * @name: qstr of name we wish to find 2216 * 2217 * On lookup failure NULL is returned; on bad name - ERR_PTR(-error) 2218 */ 2219 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name) 2220 { 2221 /* 2222 * Check for a fs-specific hash function. Note that we must 2223 * calculate the standard hash first, as the d_op->d_hash() 2224 * routine may choose to leave the hash value unchanged. 2225 */ 2226 name->hash = full_name_hash(name->name, name->len); 2227 if (dir->d_flags & DCACHE_OP_HASH) { 2228 int err = dir->d_op->d_hash(dir, name); 2229 if (unlikely(err < 0)) 2230 return ERR_PTR(err); 2231 } 2232 return d_lookup(dir, name); 2233 } 2234 EXPORT_SYMBOL(d_hash_and_lookup); 2235 2236 /** 2237 * d_validate - verify dentry provided from insecure source (deprecated) 2238 * @dentry: The dentry alleged to be valid child of @dparent 2239 * @dparent: The parent dentry (known to be valid) 2240 * 2241 * An insecure source has sent us a dentry, here we verify it and dget() it. 2242 * This is used by ncpfs in its readdir implementation. 2243 * Zero is returned in the dentry is invalid. 2244 * 2245 * This function is slow for big directories, and deprecated, do not use it. 2246 */ 2247 int d_validate(struct dentry *dentry, struct dentry *dparent) 2248 { 2249 struct dentry *child; 2250 2251 spin_lock(&dparent->d_lock); 2252 list_for_each_entry(child, &dparent->d_subdirs, d_u.d_child) { 2253 if (dentry == child) { 2254 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED); 2255 __dget_dlock(dentry); 2256 spin_unlock(&dentry->d_lock); 2257 spin_unlock(&dparent->d_lock); 2258 return 1; 2259 } 2260 } 2261 spin_unlock(&dparent->d_lock); 2262 2263 return 0; 2264 } 2265 EXPORT_SYMBOL(d_validate); 2266 2267 /* 2268 * When a file is deleted, we have two options: 2269 * - turn this dentry into a negative dentry 2270 * - unhash this dentry and free it. 2271 * 2272 * Usually, we want to just turn this into 2273 * a negative dentry, but if anybody else is 2274 * currently using the dentry or the inode 2275 * we can't do that and we fall back on removing 2276 * it from the hash queues and waiting for 2277 * it to be deleted later when it has no users 2278 */ 2279 2280 /** 2281 * d_delete - delete a dentry 2282 * @dentry: The dentry to delete 2283 * 2284 * Turn the dentry into a negative dentry if possible, otherwise 2285 * remove it from the hash queues so it can be deleted later 2286 */ 2287 2288 void d_delete(struct dentry * dentry) 2289 { 2290 struct inode *inode; 2291 int isdir = 0; 2292 /* 2293 * Are we the only user? 2294 */ 2295 again: 2296 spin_lock(&dentry->d_lock); 2297 inode = dentry->d_inode; 2298 isdir = S_ISDIR(inode->i_mode); 2299 if (dentry->d_lockref.count == 1) { 2300 if (!spin_trylock(&inode->i_lock)) { 2301 spin_unlock(&dentry->d_lock); 2302 cpu_relax(); 2303 goto again; 2304 } 2305 dentry->d_flags &= ~DCACHE_CANT_MOUNT; 2306 dentry_unlink_inode(dentry); 2307 fsnotify_nameremove(dentry, isdir); 2308 return; 2309 } 2310 2311 if (!d_unhashed(dentry)) 2312 __d_drop(dentry); 2313 2314 spin_unlock(&dentry->d_lock); 2315 2316 fsnotify_nameremove(dentry, isdir); 2317 } 2318 EXPORT_SYMBOL(d_delete); 2319 2320 static void __d_rehash(struct dentry * entry, struct hlist_bl_head *b) 2321 { 2322 BUG_ON(!d_unhashed(entry)); 2323 hlist_bl_lock(b); 2324 entry->d_flags |= DCACHE_RCUACCESS; 2325 hlist_bl_add_head_rcu(&entry->d_hash, b); 2326 hlist_bl_unlock(b); 2327 } 2328 2329 static void _d_rehash(struct dentry * entry) 2330 { 2331 __d_rehash(entry, d_hash(entry->d_parent, entry->d_name.hash)); 2332 } 2333 2334 /** 2335 * d_rehash - add an entry back to the hash 2336 * @entry: dentry to add to the hash 2337 * 2338 * Adds a dentry to the hash according to its name. 2339 */ 2340 2341 void d_rehash(struct dentry * entry) 2342 { 2343 spin_lock(&entry->d_lock); 2344 _d_rehash(entry); 2345 spin_unlock(&entry->d_lock); 2346 } 2347 EXPORT_SYMBOL(d_rehash); 2348 2349 /** 2350 * dentry_update_name_case - update case insensitive dentry with a new name 2351 * @dentry: dentry to be updated 2352 * @name: new name 2353 * 2354 * Update a case insensitive dentry with new case of name. 2355 * 2356 * dentry must have been returned by d_lookup with name @name. Old and new 2357 * name lengths must match (ie. no d_compare which allows mismatched name 2358 * lengths). 2359 * 2360 * Parent inode i_mutex must be held over d_lookup and into this call (to 2361 * keep renames and concurrent inserts, and readdir(2) away). 2362 */ 2363 void dentry_update_name_case(struct dentry *dentry, struct qstr *name) 2364 { 2365 BUG_ON(!mutex_is_locked(&dentry->d_parent->d_inode->i_mutex)); 2366 BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */ 2367 2368 spin_lock(&dentry->d_lock); 2369 write_seqcount_begin(&dentry->d_seq); 2370 memcpy((unsigned char *)dentry->d_name.name, name->name, name->len); 2371 write_seqcount_end(&dentry->d_seq); 2372 spin_unlock(&dentry->d_lock); 2373 } 2374 EXPORT_SYMBOL(dentry_update_name_case); 2375 2376 static void switch_names(struct dentry *dentry, struct dentry *target) 2377 { 2378 if (dname_external(target)) { 2379 if (dname_external(dentry)) { 2380 /* 2381 * Both external: swap the pointers 2382 */ 2383 swap(target->d_name.name, dentry->d_name.name); 2384 } else { 2385 /* 2386 * dentry:internal, target:external. Steal target's 2387 * storage and make target internal. 2388 */ 2389 memcpy(target->d_iname, dentry->d_name.name, 2390 dentry->d_name.len + 1); 2391 dentry->d_name.name = target->d_name.name; 2392 target->d_name.name = target->d_iname; 2393 } 2394 } else { 2395 if (dname_external(dentry)) { 2396 /* 2397 * dentry:external, target:internal. Give dentry's 2398 * storage to target and make dentry internal 2399 */ 2400 memcpy(dentry->d_iname, target->d_name.name, 2401 target->d_name.len + 1); 2402 target->d_name.name = dentry->d_name.name; 2403 dentry->d_name.name = dentry->d_iname; 2404 } else { 2405 /* 2406 * Both are internal. 2407 */ 2408 unsigned int i; 2409 BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long))); 2410 for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) { 2411 swap(((long *) &dentry->d_iname)[i], 2412 ((long *) &target->d_iname)[i]); 2413 } 2414 } 2415 } 2416 swap(dentry->d_name.len, target->d_name.len); 2417 } 2418 2419 static void dentry_lock_for_move(struct dentry *dentry, struct dentry *target) 2420 { 2421 /* 2422 * XXXX: do we really need to take target->d_lock? 2423 */ 2424 if (IS_ROOT(dentry) || dentry->d_parent == target->d_parent) 2425 spin_lock(&target->d_parent->d_lock); 2426 else { 2427 if (d_ancestor(dentry->d_parent, target->d_parent)) { 2428 spin_lock(&dentry->d_parent->d_lock); 2429 spin_lock_nested(&target->d_parent->d_lock, 2430 DENTRY_D_LOCK_NESTED); 2431 } else { 2432 spin_lock(&target->d_parent->d_lock); 2433 spin_lock_nested(&dentry->d_parent->d_lock, 2434 DENTRY_D_LOCK_NESTED); 2435 } 2436 } 2437 if (target < dentry) { 2438 spin_lock_nested(&target->d_lock, 2); 2439 spin_lock_nested(&dentry->d_lock, 3); 2440 } else { 2441 spin_lock_nested(&dentry->d_lock, 2); 2442 spin_lock_nested(&target->d_lock, 3); 2443 } 2444 } 2445 2446 static void dentry_unlock_parents_for_move(struct dentry *dentry, 2447 struct dentry *target) 2448 { 2449 if (target->d_parent != dentry->d_parent) 2450 spin_unlock(&dentry->d_parent->d_lock); 2451 if (target->d_parent != target) 2452 spin_unlock(&target->d_parent->d_lock); 2453 } 2454 2455 /* 2456 * When switching names, the actual string doesn't strictly have to 2457 * be preserved in the target - because we're dropping the target 2458 * anyway. As such, we can just do a simple memcpy() to copy over 2459 * the new name before we switch. 2460 * 2461 * Note that we have to be a lot more careful about getting the hash 2462 * switched - we have to switch the hash value properly even if it 2463 * then no longer matches the actual (corrupted) string of the target. 2464 * The hash value has to match the hash queue that the dentry is on.. 2465 */ 2466 /* 2467 * __d_move - move a dentry 2468 * @dentry: entry to move 2469 * @target: new dentry 2470 * @exchange: exchange the two dentries 2471 * 2472 * Update the dcache to reflect the move of a file name. Negative 2473 * dcache entries should not be moved in this way. Caller must hold 2474 * rename_lock, the i_mutex of the source and target directories, 2475 * and the sb->s_vfs_rename_mutex if they differ. See lock_rename(). 2476 */ 2477 static void __d_move(struct dentry *dentry, struct dentry *target, 2478 bool exchange) 2479 { 2480 if (!dentry->d_inode) 2481 printk(KERN_WARNING "VFS: moving negative dcache entry\n"); 2482 2483 BUG_ON(d_ancestor(dentry, target)); 2484 BUG_ON(d_ancestor(target, dentry)); 2485 2486 dentry_lock_for_move(dentry, target); 2487 2488 write_seqcount_begin(&dentry->d_seq); 2489 write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED); 2490 2491 /* __d_drop does write_seqcount_barrier, but they're OK to nest. */ 2492 2493 /* 2494 * Move the dentry to the target hash queue. Don't bother checking 2495 * for the same hash queue because of how unlikely it is. 2496 */ 2497 __d_drop(dentry); 2498 __d_rehash(dentry, d_hash(target->d_parent, target->d_name.hash)); 2499 2500 /* 2501 * Unhash the target (d_delete() is not usable here). If exchanging 2502 * the two dentries, then rehash onto the other's hash queue. 2503 */ 2504 __d_drop(target); 2505 if (exchange) { 2506 __d_rehash(target, 2507 d_hash(dentry->d_parent, dentry->d_name.hash)); 2508 } 2509 2510 list_del(&dentry->d_u.d_child); 2511 list_del(&target->d_u.d_child); 2512 2513 /* Switch the names.. */ 2514 switch_names(dentry, target); 2515 swap(dentry->d_name.hash, target->d_name.hash); 2516 2517 /* ... and switch the parents */ 2518 if (IS_ROOT(dentry)) { 2519 dentry->d_parent = target->d_parent; 2520 target->d_parent = target; 2521 INIT_LIST_HEAD(&target->d_u.d_child); 2522 } else { 2523 swap(dentry->d_parent, target->d_parent); 2524 2525 /* And add them back to the (new) parent lists */ 2526 list_add(&target->d_u.d_child, &target->d_parent->d_subdirs); 2527 } 2528 2529 list_add(&dentry->d_u.d_child, &dentry->d_parent->d_subdirs); 2530 2531 write_seqcount_end(&target->d_seq); 2532 write_seqcount_end(&dentry->d_seq); 2533 2534 dentry_unlock_parents_for_move(dentry, target); 2535 if (exchange) 2536 fsnotify_d_move(target); 2537 spin_unlock(&target->d_lock); 2538 fsnotify_d_move(dentry); 2539 spin_unlock(&dentry->d_lock); 2540 } 2541 2542 /* 2543 * d_move - move a dentry 2544 * @dentry: entry to move 2545 * @target: new dentry 2546 * 2547 * Update the dcache to reflect the move of a file name. Negative 2548 * dcache entries should not be moved in this way. See the locking 2549 * requirements for __d_move. 2550 */ 2551 void d_move(struct dentry *dentry, struct dentry *target) 2552 { 2553 write_seqlock(&rename_lock); 2554 __d_move(dentry, target, false); 2555 write_sequnlock(&rename_lock); 2556 } 2557 EXPORT_SYMBOL(d_move); 2558 2559 /* 2560 * d_exchange - exchange two dentries 2561 * @dentry1: first dentry 2562 * @dentry2: second dentry 2563 */ 2564 void d_exchange(struct dentry *dentry1, struct dentry *dentry2) 2565 { 2566 write_seqlock(&rename_lock); 2567 2568 WARN_ON(!dentry1->d_inode); 2569 WARN_ON(!dentry2->d_inode); 2570 WARN_ON(IS_ROOT(dentry1)); 2571 WARN_ON(IS_ROOT(dentry2)); 2572 2573 __d_move(dentry1, dentry2, true); 2574 2575 write_sequnlock(&rename_lock); 2576 } 2577 2578 /** 2579 * d_ancestor - search for an ancestor 2580 * @p1: ancestor dentry 2581 * @p2: child dentry 2582 * 2583 * Returns the ancestor dentry of p2 which is a child of p1, if p1 is 2584 * an ancestor of p2, else NULL. 2585 */ 2586 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2) 2587 { 2588 struct dentry *p; 2589 2590 for (p = p2; !IS_ROOT(p); p = p->d_parent) { 2591 if (p->d_parent == p1) 2592 return p; 2593 } 2594 return NULL; 2595 } 2596 2597 /* 2598 * This helper attempts to cope with remotely renamed directories 2599 * 2600 * It assumes that the caller is already holding 2601 * dentry->d_parent->d_inode->i_mutex, inode->i_lock and rename_lock 2602 * 2603 * Note: If ever the locking in lock_rename() changes, then please 2604 * remember to update this too... 2605 */ 2606 static struct dentry *__d_unalias(struct inode *inode, 2607 struct dentry *dentry, struct dentry *alias) 2608 { 2609 struct mutex *m1 = NULL, *m2 = NULL; 2610 struct dentry *ret = ERR_PTR(-EBUSY); 2611 2612 /* If alias and dentry share a parent, then no extra locks required */ 2613 if (alias->d_parent == dentry->d_parent) 2614 goto out_unalias; 2615 2616 /* See lock_rename() */ 2617 if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex)) 2618 goto out_err; 2619 m1 = &dentry->d_sb->s_vfs_rename_mutex; 2620 if (!mutex_trylock(&alias->d_parent->d_inode->i_mutex)) 2621 goto out_err; 2622 m2 = &alias->d_parent->d_inode->i_mutex; 2623 out_unalias: 2624 if (likely(!d_mountpoint(alias))) { 2625 __d_move(alias, dentry, false); 2626 ret = alias; 2627 } 2628 out_err: 2629 spin_unlock(&inode->i_lock); 2630 if (m2) 2631 mutex_unlock(m2); 2632 if (m1) 2633 mutex_unlock(m1); 2634 return ret; 2635 } 2636 2637 /* 2638 * Prepare an anonymous dentry for life in the superblock's dentry tree as a 2639 * named dentry in place of the dentry to be replaced. 2640 * returns with anon->d_lock held! 2641 */ 2642 static void __d_materialise_dentry(struct dentry *dentry, struct dentry *anon) 2643 { 2644 struct dentry *dparent; 2645 2646 dentry_lock_for_move(anon, dentry); 2647 2648 write_seqcount_begin(&dentry->d_seq); 2649 write_seqcount_begin_nested(&anon->d_seq, DENTRY_D_LOCK_NESTED); 2650 2651 dparent = dentry->d_parent; 2652 2653 switch_names(dentry, anon); 2654 swap(dentry->d_name.hash, anon->d_name.hash); 2655 2656 dentry->d_parent = dentry; 2657 list_del_init(&dentry->d_u.d_child); 2658 anon->d_parent = dparent; 2659 list_move(&anon->d_u.d_child, &dparent->d_subdirs); 2660 2661 write_seqcount_end(&dentry->d_seq); 2662 write_seqcount_end(&anon->d_seq); 2663 2664 dentry_unlock_parents_for_move(anon, dentry); 2665 spin_unlock(&dentry->d_lock); 2666 2667 /* anon->d_lock still locked, returns locked */ 2668 } 2669 2670 /** 2671 * d_splice_alias - splice a disconnected dentry into the tree if one exists 2672 * @inode: the inode which may have a disconnected dentry 2673 * @dentry: a negative dentry which we want to point to the inode. 2674 * 2675 * If inode is a directory and has an IS_ROOT alias, then d_move that in 2676 * place of the given dentry and return it, else simply d_add the inode 2677 * to the dentry and return NULL. 2678 * 2679 * If a non-IS_ROOT directory is found, the filesystem is corrupt, and 2680 * we should error out: directories can't have multiple aliases. 2681 * 2682 * This is needed in the lookup routine of any filesystem that is exportable 2683 * (via knfsd) so that we can build dcache paths to directories effectively. 2684 * 2685 * If a dentry was found and moved, then it is returned. Otherwise NULL 2686 * is returned. This matches the expected return value of ->lookup. 2687 * 2688 * Cluster filesystems may call this function with a negative, hashed dentry. 2689 * In that case, we know that the inode will be a regular file, and also this 2690 * will only occur during atomic_open. So we need to check for the dentry 2691 * being already hashed only in the final case. 2692 */ 2693 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry) 2694 { 2695 struct dentry *new = NULL; 2696 2697 if (IS_ERR(inode)) 2698 return ERR_CAST(inode); 2699 2700 if (inode && S_ISDIR(inode->i_mode)) { 2701 spin_lock(&inode->i_lock); 2702 new = __d_find_any_alias(inode); 2703 if (new) { 2704 if (!IS_ROOT(new)) { 2705 spin_unlock(&inode->i_lock); 2706 dput(new); 2707 return ERR_PTR(-EIO); 2708 } 2709 if (d_ancestor(new, dentry)) { 2710 spin_unlock(&inode->i_lock); 2711 dput(new); 2712 return ERR_PTR(-EIO); 2713 } 2714 write_seqlock(&rename_lock); 2715 __d_materialise_dentry(dentry, new); 2716 write_sequnlock(&rename_lock); 2717 __d_drop(new); 2718 _d_rehash(new); 2719 spin_unlock(&new->d_lock); 2720 spin_unlock(&inode->i_lock); 2721 security_d_instantiate(new, inode); 2722 iput(inode); 2723 } else { 2724 /* already taking inode->i_lock, so d_add() by hand */ 2725 __d_instantiate(dentry, inode); 2726 spin_unlock(&inode->i_lock); 2727 security_d_instantiate(dentry, inode); 2728 d_rehash(dentry); 2729 } 2730 } else { 2731 d_instantiate(dentry, inode); 2732 if (d_unhashed(dentry)) 2733 d_rehash(dentry); 2734 } 2735 return new; 2736 } 2737 EXPORT_SYMBOL(d_splice_alias); 2738 2739 /** 2740 * d_materialise_unique - introduce an inode into the tree 2741 * @dentry: candidate dentry 2742 * @inode: inode to bind to the dentry, to which aliases may be attached 2743 * 2744 * Introduces an dentry into the tree, substituting an extant disconnected 2745 * root directory alias in its place if there is one. Caller must hold the 2746 * i_mutex of the parent directory. 2747 */ 2748 struct dentry *d_materialise_unique(struct dentry *dentry, struct inode *inode) 2749 { 2750 struct dentry *actual; 2751 2752 BUG_ON(!d_unhashed(dentry)); 2753 2754 if (!inode) { 2755 actual = dentry; 2756 __d_instantiate(dentry, NULL); 2757 d_rehash(actual); 2758 goto out_nolock; 2759 } 2760 2761 spin_lock(&inode->i_lock); 2762 2763 if (S_ISDIR(inode->i_mode)) { 2764 struct dentry *alias; 2765 2766 /* Does an aliased dentry already exist? */ 2767 alias = __d_find_alias(inode); 2768 if (alias) { 2769 actual = alias; 2770 write_seqlock(&rename_lock); 2771 2772 if (d_ancestor(alias, dentry)) { 2773 /* Check for loops */ 2774 actual = ERR_PTR(-ELOOP); 2775 spin_unlock(&inode->i_lock); 2776 } else if (IS_ROOT(alias)) { 2777 /* Is this an anonymous mountpoint that we 2778 * could splice into our tree? */ 2779 __d_materialise_dentry(dentry, alias); 2780 write_sequnlock(&rename_lock); 2781 __d_drop(alias); 2782 goto found; 2783 } else { 2784 /* Nope, but we must(!) avoid directory 2785 * aliasing. This drops inode->i_lock */ 2786 actual = __d_unalias(inode, dentry, alias); 2787 } 2788 write_sequnlock(&rename_lock); 2789 if (IS_ERR(actual)) { 2790 if (PTR_ERR(actual) == -ELOOP) 2791 pr_warn_ratelimited( 2792 "VFS: Lookup of '%s' in %s %s" 2793 " would have caused loop\n", 2794 dentry->d_name.name, 2795 inode->i_sb->s_type->name, 2796 inode->i_sb->s_id); 2797 dput(alias); 2798 } 2799 goto out_nolock; 2800 } 2801 } 2802 2803 /* Add a unique reference */ 2804 actual = __d_instantiate_unique(dentry, inode); 2805 if (!actual) 2806 actual = dentry; 2807 else 2808 BUG_ON(!d_unhashed(actual)); 2809 2810 spin_lock(&actual->d_lock); 2811 found: 2812 _d_rehash(actual); 2813 spin_unlock(&actual->d_lock); 2814 spin_unlock(&inode->i_lock); 2815 out_nolock: 2816 if (actual == dentry) { 2817 security_d_instantiate(dentry, inode); 2818 return NULL; 2819 } 2820 2821 iput(inode); 2822 return actual; 2823 } 2824 EXPORT_SYMBOL_GPL(d_materialise_unique); 2825 2826 static int prepend(char **buffer, int *buflen, const char *str, int namelen) 2827 { 2828 *buflen -= namelen; 2829 if (*buflen < 0) 2830 return -ENAMETOOLONG; 2831 *buffer -= namelen; 2832 memcpy(*buffer, str, namelen); 2833 return 0; 2834 } 2835 2836 /** 2837 * prepend_name - prepend a pathname in front of current buffer pointer 2838 * @buffer: buffer pointer 2839 * @buflen: allocated length of the buffer 2840 * @name: name string and length qstr structure 2841 * 2842 * With RCU path tracing, it may race with d_move(). Use ACCESS_ONCE() to 2843 * make sure that either the old or the new name pointer and length are 2844 * fetched. However, there may be mismatch between length and pointer. 2845 * The length cannot be trusted, we need to copy it byte-by-byte until 2846 * the length is reached or a null byte is found. It also prepends "/" at 2847 * the beginning of the name. The sequence number check at the caller will 2848 * retry it again when a d_move() does happen. So any garbage in the buffer 2849 * due to mismatched pointer and length will be discarded. 2850 */ 2851 static int prepend_name(char **buffer, int *buflen, struct qstr *name) 2852 { 2853 const char *dname = ACCESS_ONCE(name->name); 2854 u32 dlen = ACCESS_ONCE(name->len); 2855 char *p; 2856 2857 *buflen -= dlen + 1; 2858 if (*buflen < 0) 2859 return -ENAMETOOLONG; 2860 p = *buffer -= dlen + 1; 2861 *p++ = '/'; 2862 while (dlen--) { 2863 char c = *dname++; 2864 if (!c) 2865 break; 2866 *p++ = c; 2867 } 2868 return 0; 2869 } 2870 2871 /** 2872 * prepend_path - Prepend path string to a buffer 2873 * @path: the dentry/vfsmount to report 2874 * @root: root vfsmnt/dentry 2875 * @buffer: pointer to the end of the buffer 2876 * @buflen: pointer to buffer length 2877 * 2878 * The function will first try to write out the pathname without taking any 2879 * lock other than the RCU read lock to make sure that dentries won't go away. 2880 * It only checks the sequence number of the global rename_lock as any change 2881 * in the dentry's d_seq will be preceded by changes in the rename_lock 2882 * sequence number. If the sequence number had been changed, it will restart 2883 * the whole pathname back-tracing sequence again by taking the rename_lock. 2884 * In this case, there is no need to take the RCU read lock as the recursive 2885 * parent pointer references will keep the dentry chain alive as long as no 2886 * rename operation is performed. 2887 */ 2888 static int prepend_path(const struct path *path, 2889 const struct path *root, 2890 char **buffer, int *buflen) 2891 { 2892 struct dentry *dentry; 2893 struct vfsmount *vfsmnt; 2894 struct mount *mnt; 2895 int error = 0; 2896 unsigned seq, m_seq = 0; 2897 char *bptr; 2898 int blen; 2899 2900 rcu_read_lock(); 2901 restart_mnt: 2902 read_seqbegin_or_lock(&mount_lock, &m_seq); 2903 seq = 0; 2904 rcu_read_lock(); 2905 restart: 2906 bptr = *buffer; 2907 blen = *buflen; 2908 error = 0; 2909 dentry = path->dentry; 2910 vfsmnt = path->mnt; 2911 mnt = real_mount(vfsmnt); 2912 read_seqbegin_or_lock(&rename_lock, &seq); 2913 while (dentry != root->dentry || vfsmnt != root->mnt) { 2914 struct dentry * parent; 2915 2916 if (dentry == vfsmnt->mnt_root || IS_ROOT(dentry)) { 2917 struct mount *parent = ACCESS_ONCE(mnt->mnt_parent); 2918 /* Global root? */ 2919 if (mnt != parent) { 2920 dentry = ACCESS_ONCE(mnt->mnt_mountpoint); 2921 mnt = parent; 2922 vfsmnt = &mnt->mnt; 2923 continue; 2924 } 2925 /* 2926 * Filesystems needing to implement special "root names" 2927 * should do so with ->d_dname() 2928 */ 2929 if (IS_ROOT(dentry) && 2930 (dentry->d_name.len != 1 || 2931 dentry->d_name.name[0] != '/')) { 2932 WARN(1, "Root dentry has weird name <%.*s>\n", 2933 (int) dentry->d_name.len, 2934 dentry->d_name.name); 2935 } 2936 if (!error) 2937 error = is_mounted(vfsmnt) ? 1 : 2; 2938 break; 2939 } 2940 parent = dentry->d_parent; 2941 prefetch(parent); 2942 error = prepend_name(&bptr, &blen, &dentry->d_name); 2943 if (error) 2944 break; 2945 2946 dentry = parent; 2947 } 2948 if (!(seq & 1)) 2949 rcu_read_unlock(); 2950 if (need_seqretry(&rename_lock, seq)) { 2951 seq = 1; 2952 goto restart; 2953 } 2954 done_seqretry(&rename_lock, seq); 2955 2956 if (!(m_seq & 1)) 2957 rcu_read_unlock(); 2958 if (need_seqretry(&mount_lock, m_seq)) { 2959 m_seq = 1; 2960 goto restart_mnt; 2961 } 2962 done_seqretry(&mount_lock, m_seq); 2963 2964 if (error >= 0 && bptr == *buffer) { 2965 if (--blen < 0) 2966 error = -ENAMETOOLONG; 2967 else 2968 *--bptr = '/'; 2969 } 2970 *buffer = bptr; 2971 *buflen = blen; 2972 return error; 2973 } 2974 2975 /** 2976 * __d_path - return the path of a dentry 2977 * @path: the dentry/vfsmount to report 2978 * @root: root vfsmnt/dentry 2979 * @buf: buffer to return value in 2980 * @buflen: buffer length 2981 * 2982 * Convert a dentry into an ASCII path name. 2983 * 2984 * Returns a pointer into the buffer or an error code if the 2985 * path was too long. 2986 * 2987 * "buflen" should be positive. 2988 * 2989 * If the path is not reachable from the supplied root, return %NULL. 2990 */ 2991 char *__d_path(const struct path *path, 2992 const struct path *root, 2993 char *buf, int buflen) 2994 { 2995 char *res = buf + buflen; 2996 int error; 2997 2998 prepend(&res, &buflen, "\0", 1); 2999 error = prepend_path(path, root, &res, &buflen); 3000 3001 if (error < 0) 3002 return ERR_PTR(error); 3003 if (error > 0) 3004 return NULL; 3005 return res; 3006 } 3007 3008 char *d_absolute_path(const struct path *path, 3009 char *buf, int buflen) 3010 { 3011 struct path root = {}; 3012 char *res = buf + buflen; 3013 int error; 3014 3015 prepend(&res, &buflen, "\0", 1); 3016 error = prepend_path(path, &root, &res, &buflen); 3017 3018 if (error > 1) 3019 error = -EINVAL; 3020 if (error < 0) 3021 return ERR_PTR(error); 3022 return res; 3023 } 3024 3025 /* 3026 * same as __d_path but appends "(deleted)" for unlinked files. 3027 */ 3028 static int path_with_deleted(const struct path *path, 3029 const struct path *root, 3030 char **buf, int *buflen) 3031 { 3032 prepend(buf, buflen, "\0", 1); 3033 if (d_unlinked(path->dentry)) { 3034 int error = prepend(buf, buflen, " (deleted)", 10); 3035 if (error) 3036 return error; 3037 } 3038 3039 return prepend_path(path, root, buf, buflen); 3040 } 3041 3042 static int prepend_unreachable(char **buffer, int *buflen) 3043 { 3044 return prepend(buffer, buflen, "(unreachable)", 13); 3045 } 3046 3047 static void get_fs_root_rcu(struct fs_struct *fs, struct path *root) 3048 { 3049 unsigned seq; 3050 3051 do { 3052 seq = read_seqcount_begin(&fs->seq); 3053 *root = fs->root; 3054 } while (read_seqcount_retry(&fs->seq, seq)); 3055 } 3056 3057 /** 3058 * d_path - return the path of a dentry 3059 * @path: path to report 3060 * @buf: buffer to return value in 3061 * @buflen: buffer length 3062 * 3063 * Convert a dentry into an ASCII path name. If the entry has been deleted 3064 * the string " (deleted)" is appended. Note that this is ambiguous. 3065 * 3066 * Returns a pointer into the buffer or an error code if the path was 3067 * too long. Note: Callers should use the returned pointer, not the passed 3068 * in buffer, to use the name! The implementation often starts at an offset 3069 * into the buffer, and may leave 0 bytes at the start. 3070 * 3071 * "buflen" should be positive. 3072 */ 3073 char *d_path(const struct path *path, char *buf, int buflen) 3074 { 3075 char *res = buf + buflen; 3076 struct path root; 3077 int error; 3078 3079 /* 3080 * We have various synthetic filesystems that never get mounted. On 3081 * these filesystems dentries are never used for lookup purposes, and 3082 * thus don't need to be hashed. They also don't need a name until a 3083 * user wants to identify the object in /proc/pid/fd/. The little hack 3084 * below allows us to generate a name for these objects on demand: 3085 * 3086 * Some pseudo inodes are mountable. When they are mounted 3087 * path->dentry == path->mnt->mnt_root. In that case don't call d_dname 3088 * and instead have d_path return the mounted path. 3089 */ 3090 if (path->dentry->d_op && path->dentry->d_op->d_dname && 3091 (!IS_ROOT(path->dentry) || path->dentry != path->mnt->mnt_root)) 3092 return path->dentry->d_op->d_dname(path->dentry, buf, buflen); 3093 3094 rcu_read_lock(); 3095 get_fs_root_rcu(current->fs, &root); 3096 error = path_with_deleted(path, &root, &res, &buflen); 3097 rcu_read_unlock(); 3098 3099 if (error < 0) 3100 res = ERR_PTR(error); 3101 return res; 3102 } 3103 EXPORT_SYMBOL(d_path); 3104 3105 /* 3106 * Helper function for dentry_operations.d_dname() members 3107 */ 3108 char *dynamic_dname(struct dentry *dentry, char *buffer, int buflen, 3109 const char *fmt, ...) 3110 { 3111 va_list args; 3112 char temp[64]; 3113 int sz; 3114 3115 va_start(args, fmt); 3116 sz = vsnprintf(temp, sizeof(temp), fmt, args) + 1; 3117 va_end(args); 3118 3119 if (sz > sizeof(temp) || sz > buflen) 3120 return ERR_PTR(-ENAMETOOLONG); 3121 3122 buffer += buflen - sz; 3123 return memcpy(buffer, temp, sz); 3124 } 3125 3126 char *simple_dname(struct dentry *dentry, char *buffer, int buflen) 3127 { 3128 char *end = buffer + buflen; 3129 /* these dentries are never renamed, so d_lock is not needed */ 3130 if (prepend(&end, &buflen, " (deleted)", 11) || 3131 prepend(&end, &buflen, dentry->d_name.name, dentry->d_name.len) || 3132 prepend(&end, &buflen, "/", 1)) 3133 end = ERR_PTR(-ENAMETOOLONG); 3134 return end; 3135 } 3136 EXPORT_SYMBOL(simple_dname); 3137 3138 /* 3139 * Write full pathname from the root of the filesystem into the buffer. 3140 */ 3141 static char *__dentry_path(struct dentry *d, char *buf, int buflen) 3142 { 3143 struct dentry *dentry; 3144 char *end, *retval; 3145 int len, seq = 0; 3146 int error = 0; 3147 3148 if (buflen < 2) 3149 goto Elong; 3150 3151 rcu_read_lock(); 3152 restart: 3153 dentry = d; 3154 end = buf + buflen; 3155 len = buflen; 3156 prepend(&end, &len, "\0", 1); 3157 /* Get '/' right */ 3158 retval = end-1; 3159 *retval = '/'; 3160 read_seqbegin_or_lock(&rename_lock, &seq); 3161 while (!IS_ROOT(dentry)) { 3162 struct dentry *parent = dentry->d_parent; 3163 3164 prefetch(parent); 3165 error = prepend_name(&end, &len, &dentry->d_name); 3166 if (error) 3167 break; 3168 3169 retval = end; 3170 dentry = parent; 3171 } 3172 if (!(seq & 1)) 3173 rcu_read_unlock(); 3174 if (need_seqretry(&rename_lock, seq)) { 3175 seq = 1; 3176 goto restart; 3177 } 3178 done_seqretry(&rename_lock, seq); 3179 if (error) 3180 goto Elong; 3181 return retval; 3182 Elong: 3183 return ERR_PTR(-ENAMETOOLONG); 3184 } 3185 3186 char *dentry_path_raw(struct dentry *dentry, char *buf, int buflen) 3187 { 3188 return __dentry_path(dentry, buf, buflen); 3189 } 3190 EXPORT_SYMBOL(dentry_path_raw); 3191 3192 char *dentry_path(struct dentry *dentry, char *buf, int buflen) 3193 { 3194 char *p = NULL; 3195 char *retval; 3196 3197 if (d_unlinked(dentry)) { 3198 p = buf + buflen; 3199 if (prepend(&p, &buflen, "//deleted", 10) != 0) 3200 goto Elong; 3201 buflen++; 3202 } 3203 retval = __dentry_path(dentry, buf, buflen); 3204 if (!IS_ERR(retval) && p) 3205 *p = '/'; /* restore '/' overriden with '\0' */ 3206 return retval; 3207 Elong: 3208 return ERR_PTR(-ENAMETOOLONG); 3209 } 3210 3211 static void get_fs_root_and_pwd_rcu(struct fs_struct *fs, struct path *root, 3212 struct path *pwd) 3213 { 3214 unsigned seq; 3215 3216 do { 3217 seq = read_seqcount_begin(&fs->seq); 3218 *root = fs->root; 3219 *pwd = fs->pwd; 3220 } while (read_seqcount_retry(&fs->seq, seq)); 3221 } 3222 3223 /* 3224 * NOTE! The user-level library version returns a 3225 * character pointer. The kernel system call just 3226 * returns the length of the buffer filled (which 3227 * includes the ending '\0' character), or a negative 3228 * error value. So libc would do something like 3229 * 3230 * char *getcwd(char * buf, size_t size) 3231 * { 3232 * int retval; 3233 * 3234 * retval = sys_getcwd(buf, size); 3235 * if (retval >= 0) 3236 * return buf; 3237 * errno = -retval; 3238 * return NULL; 3239 * } 3240 */ 3241 SYSCALL_DEFINE2(getcwd, char __user *, buf, unsigned long, size) 3242 { 3243 int error; 3244 struct path pwd, root; 3245 char *page = __getname(); 3246 3247 if (!page) 3248 return -ENOMEM; 3249 3250 rcu_read_lock(); 3251 get_fs_root_and_pwd_rcu(current->fs, &root, &pwd); 3252 3253 error = -ENOENT; 3254 if (!d_unlinked(pwd.dentry)) { 3255 unsigned long len; 3256 char *cwd = page + PATH_MAX; 3257 int buflen = PATH_MAX; 3258 3259 prepend(&cwd, &buflen, "\0", 1); 3260 error = prepend_path(&pwd, &root, &cwd, &buflen); 3261 rcu_read_unlock(); 3262 3263 if (error < 0) 3264 goto out; 3265 3266 /* Unreachable from current root */ 3267 if (error > 0) { 3268 error = prepend_unreachable(&cwd, &buflen); 3269 if (error) 3270 goto out; 3271 } 3272 3273 error = -ERANGE; 3274 len = PATH_MAX + page - cwd; 3275 if (len <= size) { 3276 error = len; 3277 if (copy_to_user(buf, cwd, len)) 3278 error = -EFAULT; 3279 } 3280 } else { 3281 rcu_read_unlock(); 3282 } 3283 3284 out: 3285 __putname(page); 3286 return error; 3287 } 3288 3289 /* 3290 * Test whether new_dentry is a subdirectory of old_dentry. 3291 * 3292 * Trivially implemented using the dcache structure 3293 */ 3294 3295 /** 3296 * is_subdir - is new dentry a subdirectory of old_dentry 3297 * @new_dentry: new dentry 3298 * @old_dentry: old dentry 3299 * 3300 * Returns 1 if new_dentry is a subdirectory of the parent (at any depth). 3301 * Returns 0 otherwise. 3302 * Caller must ensure that "new_dentry" is pinned before calling is_subdir() 3303 */ 3304 3305 int is_subdir(struct dentry *new_dentry, struct dentry *old_dentry) 3306 { 3307 int result; 3308 unsigned seq; 3309 3310 if (new_dentry == old_dentry) 3311 return 1; 3312 3313 do { 3314 /* for restarting inner loop in case of seq retry */ 3315 seq = read_seqbegin(&rename_lock); 3316 /* 3317 * Need rcu_readlock to protect against the d_parent trashing 3318 * due to d_move 3319 */ 3320 rcu_read_lock(); 3321 if (d_ancestor(old_dentry, new_dentry)) 3322 result = 1; 3323 else 3324 result = 0; 3325 rcu_read_unlock(); 3326 } while (read_seqretry(&rename_lock, seq)); 3327 3328 return result; 3329 } 3330 3331 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry) 3332 { 3333 struct dentry *root = data; 3334 if (dentry != root) { 3335 if (d_unhashed(dentry) || !dentry->d_inode) 3336 return D_WALK_SKIP; 3337 3338 if (!(dentry->d_flags & DCACHE_GENOCIDE)) { 3339 dentry->d_flags |= DCACHE_GENOCIDE; 3340 dentry->d_lockref.count--; 3341 } 3342 } 3343 return D_WALK_CONTINUE; 3344 } 3345 3346 void d_genocide(struct dentry *parent) 3347 { 3348 d_walk(parent, parent, d_genocide_kill, NULL); 3349 } 3350 3351 void d_tmpfile(struct dentry *dentry, struct inode *inode) 3352 { 3353 inode_dec_link_count(inode); 3354 BUG_ON(dentry->d_name.name != dentry->d_iname || 3355 !hlist_unhashed(&dentry->d_alias) || 3356 !d_unlinked(dentry)); 3357 spin_lock(&dentry->d_parent->d_lock); 3358 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED); 3359 dentry->d_name.len = sprintf(dentry->d_iname, "#%llu", 3360 (unsigned long long)inode->i_ino); 3361 spin_unlock(&dentry->d_lock); 3362 spin_unlock(&dentry->d_parent->d_lock); 3363 d_instantiate(dentry, inode); 3364 } 3365 EXPORT_SYMBOL(d_tmpfile); 3366 3367 static __initdata unsigned long dhash_entries; 3368 static int __init set_dhash_entries(char *str) 3369 { 3370 if (!str) 3371 return 0; 3372 dhash_entries = simple_strtoul(str, &str, 0); 3373 return 1; 3374 } 3375 __setup("dhash_entries=", set_dhash_entries); 3376 3377 static void __init dcache_init_early(void) 3378 { 3379 unsigned int loop; 3380 3381 /* If hashes are distributed across NUMA nodes, defer 3382 * hash allocation until vmalloc space is available. 3383 */ 3384 if (hashdist) 3385 return; 3386 3387 dentry_hashtable = 3388 alloc_large_system_hash("Dentry cache", 3389 sizeof(struct hlist_bl_head), 3390 dhash_entries, 3391 13, 3392 HASH_EARLY, 3393 &d_hash_shift, 3394 &d_hash_mask, 3395 0, 3396 0); 3397 3398 for (loop = 0; loop < (1U << d_hash_shift); loop++) 3399 INIT_HLIST_BL_HEAD(dentry_hashtable + loop); 3400 } 3401 3402 static void __init dcache_init(void) 3403 { 3404 unsigned int loop; 3405 3406 /* 3407 * A constructor could be added for stable state like the lists, 3408 * but it is probably not worth it because of the cache nature 3409 * of the dcache. 3410 */ 3411 dentry_cache = KMEM_CACHE(dentry, 3412 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD); 3413 3414 /* Hash may have been set up in dcache_init_early */ 3415 if (!hashdist) 3416 return; 3417 3418 dentry_hashtable = 3419 alloc_large_system_hash("Dentry cache", 3420 sizeof(struct hlist_bl_head), 3421 dhash_entries, 3422 13, 3423 0, 3424 &d_hash_shift, 3425 &d_hash_mask, 3426 0, 3427 0); 3428 3429 for (loop = 0; loop < (1U << d_hash_shift); loop++) 3430 INIT_HLIST_BL_HEAD(dentry_hashtable + loop); 3431 } 3432 3433 /* SLAB cache for __getname() consumers */ 3434 struct kmem_cache *names_cachep __read_mostly; 3435 EXPORT_SYMBOL(names_cachep); 3436 3437 EXPORT_SYMBOL(d_genocide); 3438 3439 void __init vfs_caches_init_early(void) 3440 { 3441 dcache_init_early(); 3442 inode_init_early(); 3443 } 3444 3445 void __init vfs_caches_init(unsigned long mempages) 3446 { 3447 unsigned long reserve; 3448 3449 /* Base hash sizes on available memory, with a reserve equal to 3450 150% of current kernel size */ 3451 3452 reserve = min((mempages - nr_free_pages()) * 3/2, mempages - 1); 3453 mempages -= reserve; 3454 3455 names_cachep = kmem_cache_create("names_cache", PATH_MAX, 0, 3456 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3457 3458 dcache_init(); 3459 inode_init(); 3460 files_init(mempages); 3461 mnt_init(); 3462 bdev_cache_init(); 3463 chrdev_init(); 3464 } 3465