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 <linux/kasan.h> 42 43 #include "internal.h" 44 #include "mount.h" 45 46 /* 47 * Usage: 48 * dcache->d_inode->i_lock protects: 49 * - i_dentry, d_u.d_alias, d_inode of aliases 50 * dcache_hash_bucket lock protects: 51 * - the dcache hash table 52 * s_anon bl list spinlock protects: 53 * - the s_anon list (see __d_drop) 54 * dentry->d_sb->s_dentry_lru_lock protects: 55 * - the dcache lru lists and counters 56 * d_lock protects: 57 * - d_flags 58 * - d_name 59 * - d_lru 60 * - d_count 61 * - d_unhashed() 62 * - d_parent and d_subdirs 63 * - childrens' d_child and d_parent 64 * - d_u.d_alias, d_inode 65 * 66 * Ordering: 67 * dentry->d_inode->i_lock 68 * dentry->d_lock 69 * dentry->d_sb->s_dentry_lru_lock 70 * dcache_hash_bucket lock 71 * s_anon lock 72 * 73 * If there is an ancestor relationship: 74 * dentry->d_parent->...->d_parent->d_lock 75 * ... 76 * dentry->d_parent->d_lock 77 * dentry->d_lock 78 * 79 * If no ancestor relationship: 80 * if (dentry1 < dentry2) 81 * dentry1->d_lock 82 * dentry2->d_lock 83 */ 84 int sysctl_vfs_cache_pressure __read_mostly = 100; 85 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure); 86 87 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock); 88 89 EXPORT_SYMBOL(rename_lock); 90 91 static struct kmem_cache *dentry_cache __read_mostly; 92 93 /* 94 * This is the single most critical data structure when it comes 95 * to the dcache: the hashtable for lookups. Somebody should try 96 * to make this good - I've just made it work. 97 * 98 * This hash-function tries to avoid losing too many bits of hash 99 * information, yet avoid using a prime hash-size or similar. 100 */ 101 102 static unsigned int d_hash_mask __read_mostly; 103 static unsigned int d_hash_shift __read_mostly; 104 105 static struct hlist_bl_head *dentry_hashtable __read_mostly; 106 107 static inline struct hlist_bl_head *d_hash(const struct dentry *parent, 108 unsigned int hash) 109 { 110 hash += (unsigned long) parent / L1_CACHE_BYTES; 111 return dentry_hashtable + hash_32(hash, d_hash_shift); 112 } 113 114 /* Statistics gathering. */ 115 struct dentry_stat_t dentry_stat = { 116 .age_limit = 45, 117 }; 118 119 static DEFINE_PER_CPU(long, nr_dentry); 120 static DEFINE_PER_CPU(long, nr_dentry_unused); 121 122 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS) 123 124 /* 125 * Here we resort to our own counters instead of using generic per-cpu counters 126 * for consistency with what the vfs inode code does. We are expected to harvest 127 * better code and performance by having our own specialized counters. 128 * 129 * Please note that the loop is done over all possible CPUs, not over all online 130 * CPUs. The reason for this is that we don't want to play games with CPUs going 131 * on and off. If one of them goes off, we will just keep their counters. 132 * 133 * glommer: See cffbc8a for details, and if you ever intend to change this, 134 * please update all vfs counters to match. 135 */ 136 static long get_nr_dentry(void) 137 { 138 int i; 139 long sum = 0; 140 for_each_possible_cpu(i) 141 sum += per_cpu(nr_dentry, i); 142 return sum < 0 ? 0 : sum; 143 } 144 145 static long get_nr_dentry_unused(void) 146 { 147 int i; 148 long sum = 0; 149 for_each_possible_cpu(i) 150 sum += per_cpu(nr_dentry_unused, i); 151 return sum < 0 ? 0 : sum; 152 } 153 154 int proc_nr_dentry(struct ctl_table *table, int write, void __user *buffer, 155 size_t *lenp, loff_t *ppos) 156 { 157 dentry_stat.nr_dentry = get_nr_dentry(); 158 dentry_stat.nr_unused = get_nr_dentry_unused(); 159 return proc_doulongvec_minmax(table, write, buffer, lenp, ppos); 160 } 161 #endif 162 163 /* 164 * Compare 2 name strings, return 0 if they match, otherwise non-zero. 165 * The strings are both count bytes long, and count is non-zero. 166 */ 167 #ifdef CONFIG_DCACHE_WORD_ACCESS 168 169 #include <asm/word-at-a-time.h> 170 /* 171 * NOTE! 'cs' and 'scount' come from a dentry, so it has a 172 * aligned allocation for this particular component. We don't 173 * strictly need the load_unaligned_zeropad() safety, but it 174 * doesn't hurt either. 175 * 176 * In contrast, 'ct' and 'tcount' can be from a pathname, and do 177 * need the careful unaligned handling. 178 */ 179 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount) 180 { 181 unsigned long a,b,mask; 182 183 for (;;) { 184 a = *(unsigned long *)cs; 185 b = load_unaligned_zeropad(ct); 186 if (tcount < sizeof(unsigned long)) 187 break; 188 if (unlikely(a != b)) 189 return 1; 190 cs += sizeof(unsigned long); 191 ct += sizeof(unsigned long); 192 tcount -= sizeof(unsigned long); 193 if (!tcount) 194 return 0; 195 } 196 mask = bytemask_from_count(tcount); 197 return unlikely(!!((a ^ b) & mask)); 198 } 199 200 #else 201 202 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount) 203 { 204 do { 205 if (*cs != *ct) 206 return 1; 207 cs++; 208 ct++; 209 tcount--; 210 } while (tcount); 211 return 0; 212 } 213 214 #endif 215 216 static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount) 217 { 218 const unsigned char *cs; 219 /* 220 * Be careful about RCU walk racing with rename: 221 * use ACCESS_ONCE to fetch the name pointer. 222 * 223 * NOTE! Even if a rename will mean that the length 224 * was not loaded atomically, we don't care. The 225 * RCU walk will check the sequence count eventually, 226 * and catch it. And we won't overrun the buffer, 227 * because we're reading the name pointer atomically, 228 * and a dentry name is guaranteed to be properly 229 * terminated with a NUL byte. 230 * 231 * End result: even if 'len' is wrong, we'll exit 232 * early because the data cannot match (there can 233 * be no NUL in the ct/tcount data) 234 */ 235 cs = ACCESS_ONCE(dentry->d_name.name); 236 smp_read_barrier_depends(); 237 return dentry_string_cmp(cs, ct, tcount); 238 } 239 240 struct external_name { 241 union { 242 atomic_t count; 243 struct rcu_head head; 244 } u; 245 unsigned char name[]; 246 }; 247 248 static inline struct external_name *external_name(struct dentry *dentry) 249 { 250 return container_of(dentry->d_name.name, struct external_name, name[0]); 251 } 252 253 static void __d_free(struct rcu_head *head) 254 { 255 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu); 256 257 kmem_cache_free(dentry_cache, dentry); 258 } 259 260 static void __d_free_external(struct rcu_head *head) 261 { 262 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu); 263 kfree(external_name(dentry)); 264 kmem_cache_free(dentry_cache, dentry); 265 } 266 267 static inline int dname_external(const struct dentry *dentry) 268 { 269 return dentry->d_name.name != dentry->d_iname; 270 } 271 272 /* 273 * Make sure other CPUs see the inode attached before the type is set. 274 */ 275 static inline void __d_set_inode_and_type(struct dentry *dentry, 276 struct inode *inode, 277 unsigned type_flags) 278 { 279 unsigned flags; 280 281 dentry->d_inode = inode; 282 smp_wmb(); 283 flags = READ_ONCE(dentry->d_flags); 284 flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU); 285 flags |= type_flags; 286 WRITE_ONCE(dentry->d_flags, flags); 287 } 288 289 /* 290 * Ideally, we want to make sure that other CPUs see the flags cleared before 291 * the inode is detached, but this is really a violation of RCU principles 292 * since the ordering suggests we should always set inode before flags. 293 * 294 * We should instead replace or discard the entire dentry - but that sucks 295 * performancewise on mass deletion/rename. 296 */ 297 static inline void __d_clear_type_and_inode(struct dentry *dentry) 298 { 299 unsigned flags = READ_ONCE(dentry->d_flags); 300 301 flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU); 302 WRITE_ONCE(dentry->d_flags, flags); 303 smp_wmb(); 304 dentry->d_inode = NULL; 305 } 306 307 static void dentry_free(struct dentry *dentry) 308 { 309 WARN_ON(!hlist_unhashed(&dentry->d_u.d_alias)); 310 if (unlikely(dname_external(dentry))) { 311 struct external_name *p = external_name(dentry); 312 if (likely(atomic_dec_and_test(&p->u.count))) { 313 call_rcu(&dentry->d_u.d_rcu, __d_free_external); 314 return; 315 } 316 } 317 /* if dentry was never visible to RCU, immediate free is OK */ 318 if (!(dentry->d_flags & DCACHE_RCUACCESS)) 319 __d_free(&dentry->d_u.d_rcu); 320 else 321 call_rcu(&dentry->d_u.d_rcu, __d_free); 322 } 323 324 /** 325 * dentry_rcuwalk_invalidate - invalidate in-progress rcu-walk lookups 326 * @dentry: the target dentry 327 * After this call, in-progress rcu-walk path lookup will fail. This 328 * should be called after unhashing, and after changing d_inode (if 329 * the dentry has not already been unhashed). 330 */ 331 static inline void dentry_rcuwalk_invalidate(struct dentry *dentry) 332 { 333 lockdep_assert_held(&dentry->d_lock); 334 /* Go through am invalidation barrier */ 335 write_seqcount_invalidate(&dentry->d_seq); 336 } 337 338 /* 339 * Release the dentry's inode, using the filesystem 340 * d_iput() operation if defined. Dentry has no refcount 341 * and is unhashed. 342 */ 343 static void dentry_iput(struct dentry * dentry) 344 __releases(dentry->d_lock) 345 __releases(dentry->d_inode->i_lock) 346 { 347 struct inode *inode = dentry->d_inode; 348 if (inode) { 349 __d_clear_type_and_inode(dentry); 350 hlist_del_init(&dentry->d_u.d_alias); 351 spin_unlock(&dentry->d_lock); 352 spin_unlock(&inode->i_lock); 353 if (!inode->i_nlink) 354 fsnotify_inoderemove(inode); 355 if (dentry->d_op && dentry->d_op->d_iput) 356 dentry->d_op->d_iput(dentry, inode); 357 else 358 iput(inode); 359 } else { 360 spin_unlock(&dentry->d_lock); 361 } 362 } 363 364 /* 365 * Release the dentry's inode, using the filesystem 366 * d_iput() operation if defined. dentry remains in-use. 367 */ 368 static void dentry_unlink_inode(struct dentry * dentry) 369 __releases(dentry->d_lock) 370 __releases(dentry->d_inode->i_lock) 371 { 372 struct inode *inode = dentry->d_inode; 373 __d_clear_type_and_inode(dentry); 374 hlist_del_init(&dentry->d_u.d_alias); 375 dentry_rcuwalk_invalidate(dentry); 376 spin_unlock(&dentry->d_lock); 377 spin_unlock(&inode->i_lock); 378 if (!inode->i_nlink) 379 fsnotify_inoderemove(inode); 380 if (dentry->d_op && dentry->d_op->d_iput) 381 dentry->d_op->d_iput(dentry, inode); 382 else 383 iput(inode); 384 } 385 386 /* 387 * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry 388 * is in use - which includes both the "real" per-superblock 389 * LRU list _and_ the DCACHE_SHRINK_LIST use. 390 * 391 * The DCACHE_SHRINK_LIST bit is set whenever the dentry is 392 * on the shrink list (ie not on the superblock LRU list). 393 * 394 * The per-cpu "nr_dentry_unused" counters are updated with 395 * the DCACHE_LRU_LIST bit. 396 * 397 * These helper functions make sure we always follow the 398 * rules. d_lock must be held by the caller. 399 */ 400 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x)) 401 static void d_lru_add(struct dentry *dentry) 402 { 403 D_FLAG_VERIFY(dentry, 0); 404 dentry->d_flags |= DCACHE_LRU_LIST; 405 this_cpu_inc(nr_dentry_unused); 406 WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru)); 407 } 408 409 static void d_lru_del(struct dentry *dentry) 410 { 411 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST); 412 dentry->d_flags &= ~DCACHE_LRU_LIST; 413 this_cpu_dec(nr_dentry_unused); 414 WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru)); 415 } 416 417 static void d_shrink_del(struct dentry *dentry) 418 { 419 D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST); 420 list_del_init(&dentry->d_lru); 421 dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST); 422 this_cpu_dec(nr_dentry_unused); 423 } 424 425 static void d_shrink_add(struct dentry *dentry, struct list_head *list) 426 { 427 D_FLAG_VERIFY(dentry, 0); 428 list_add(&dentry->d_lru, list); 429 dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST; 430 this_cpu_inc(nr_dentry_unused); 431 } 432 433 /* 434 * These can only be called under the global LRU lock, ie during the 435 * callback for freeing the LRU list. "isolate" removes it from the 436 * LRU lists entirely, while shrink_move moves it to the indicated 437 * private list. 438 */ 439 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry) 440 { 441 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST); 442 dentry->d_flags &= ~DCACHE_LRU_LIST; 443 this_cpu_dec(nr_dentry_unused); 444 list_lru_isolate(lru, &dentry->d_lru); 445 } 446 447 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry, 448 struct list_head *list) 449 { 450 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST); 451 dentry->d_flags |= DCACHE_SHRINK_LIST; 452 list_lru_isolate_move(lru, &dentry->d_lru, list); 453 } 454 455 /* 456 * dentry_lru_(add|del)_list) must be called with d_lock held. 457 */ 458 static void dentry_lru_add(struct dentry *dentry) 459 { 460 if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST))) 461 d_lru_add(dentry); 462 } 463 464 /** 465 * d_drop - drop a dentry 466 * @dentry: dentry to drop 467 * 468 * d_drop() unhashes the entry from the parent dentry hashes, so that it won't 469 * be found through a VFS lookup any more. Note that this is different from 470 * deleting the dentry - d_delete will try to mark the dentry negative if 471 * possible, giving a successful _negative_ lookup, while d_drop will 472 * just make the cache lookup fail. 473 * 474 * d_drop() is used mainly for stuff that wants to invalidate a dentry for some 475 * reason (NFS timeouts or autofs deletes). 476 * 477 * __d_drop requires dentry->d_lock. 478 */ 479 void __d_drop(struct dentry *dentry) 480 { 481 if (!d_unhashed(dentry)) { 482 struct hlist_bl_head *b; 483 /* 484 * Hashed dentries are normally on the dentry hashtable, 485 * with the exception of those newly allocated by 486 * d_obtain_alias, which are always IS_ROOT: 487 */ 488 if (unlikely(IS_ROOT(dentry))) 489 b = &dentry->d_sb->s_anon; 490 else 491 b = d_hash(dentry->d_parent, dentry->d_name.hash); 492 493 hlist_bl_lock(b); 494 __hlist_bl_del(&dentry->d_hash); 495 dentry->d_hash.pprev = NULL; 496 hlist_bl_unlock(b); 497 dentry_rcuwalk_invalidate(dentry); 498 } 499 } 500 EXPORT_SYMBOL(__d_drop); 501 502 void d_drop(struct dentry *dentry) 503 { 504 spin_lock(&dentry->d_lock); 505 __d_drop(dentry); 506 spin_unlock(&dentry->d_lock); 507 } 508 EXPORT_SYMBOL(d_drop); 509 510 static void __dentry_kill(struct dentry *dentry) 511 { 512 struct dentry *parent = NULL; 513 bool can_free = true; 514 if (!IS_ROOT(dentry)) 515 parent = dentry->d_parent; 516 517 /* 518 * The dentry is now unrecoverably dead to the world. 519 */ 520 lockref_mark_dead(&dentry->d_lockref); 521 522 /* 523 * inform the fs via d_prune that this dentry is about to be 524 * unhashed and destroyed. 525 */ 526 if (dentry->d_flags & DCACHE_OP_PRUNE) 527 dentry->d_op->d_prune(dentry); 528 529 if (dentry->d_flags & DCACHE_LRU_LIST) { 530 if (!(dentry->d_flags & DCACHE_SHRINK_LIST)) 531 d_lru_del(dentry); 532 } 533 /* if it was on the hash then remove it */ 534 __d_drop(dentry); 535 __list_del_entry(&dentry->d_child); 536 /* 537 * Inform d_walk() that we are no longer attached to the 538 * dentry tree 539 */ 540 dentry->d_flags |= DCACHE_DENTRY_KILLED; 541 if (parent) 542 spin_unlock(&parent->d_lock); 543 dentry_iput(dentry); 544 /* 545 * dentry_iput drops the locks, at which point nobody (except 546 * transient RCU lookups) can reach this dentry. 547 */ 548 BUG_ON(dentry->d_lockref.count > 0); 549 this_cpu_dec(nr_dentry); 550 if (dentry->d_op && dentry->d_op->d_release) 551 dentry->d_op->d_release(dentry); 552 553 spin_lock(&dentry->d_lock); 554 if (dentry->d_flags & DCACHE_SHRINK_LIST) { 555 dentry->d_flags |= DCACHE_MAY_FREE; 556 can_free = false; 557 } 558 spin_unlock(&dentry->d_lock); 559 if (likely(can_free)) 560 dentry_free(dentry); 561 } 562 563 /* 564 * Finish off a dentry we've decided to kill. 565 * dentry->d_lock must be held, returns with it unlocked. 566 * If ref is non-zero, then decrement the refcount too. 567 * Returns dentry requiring refcount drop, or NULL if we're done. 568 */ 569 static struct dentry *dentry_kill(struct dentry *dentry) 570 __releases(dentry->d_lock) 571 { 572 struct inode *inode = dentry->d_inode; 573 struct dentry *parent = NULL; 574 575 if (inode && unlikely(!spin_trylock(&inode->i_lock))) 576 goto failed; 577 578 if (!IS_ROOT(dentry)) { 579 parent = dentry->d_parent; 580 if (unlikely(!spin_trylock(&parent->d_lock))) { 581 if (inode) 582 spin_unlock(&inode->i_lock); 583 goto failed; 584 } 585 } 586 587 __dentry_kill(dentry); 588 return parent; 589 590 failed: 591 spin_unlock(&dentry->d_lock); 592 cpu_relax(); 593 return dentry; /* try again with same dentry */ 594 } 595 596 static inline struct dentry *lock_parent(struct dentry *dentry) 597 { 598 struct dentry *parent = dentry->d_parent; 599 if (IS_ROOT(dentry)) 600 return NULL; 601 if (unlikely(dentry->d_lockref.count < 0)) 602 return NULL; 603 if (likely(spin_trylock(&parent->d_lock))) 604 return parent; 605 rcu_read_lock(); 606 spin_unlock(&dentry->d_lock); 607 again: 608 parent = ACCESS_ONCE(dentry->d_parent); 609 spin_lock(&parent->d_lock); 610 /* 611 * We can't blindly lock dentry until we are sure 612 * that we won't violate the locking order. 613 * Any changes of dentry->d_parent must have 614 * been done with parent->d_lock held, so 615 * spin_lock() above is enough of a barrier 616 * for checking if it's still our child. 617 */ 618 if (unlikely(parent != dentry->d_parent)) { 619 spin_unlock(&parent->d_lock); 620 goto again; 621 } 622 rcu_read_unlock(); 623 if (parent != dentry) 624 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED); 625 else 626 parent = NULL; 627 return parent; 628 } 629 630 /* 631 * Try to do a lockless dput(), and return whether that was successful. 632 * 633 * If unsuccessful, we return false, having already taken the dentry lock. 634 * 635 * The caller needs to hold the RCU read lock, so that the dentry is 636 * guaranteed to stay around even if the refcount goes down to zero! 637 */ 638 static inline bool fast_dput(struct dentry *dentry) 639 { 640 int ret; 641 unsigned int d_flags; 642 643 /* 644 * If we have a d_op->d_delete() operation, we sould not 645 * let the dentry count go to zero, so use "put__or_lock". 646 */ 647 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) 648 return lockref_put_or_lock(&dentry->d_lockref); 649 650 /* 651 * .. otherwise, we can try to just decrement the 652 * lockref optimistically. 653 */ 654 ret = lockref_put_return(&dentry->d_lockref); 655 656 /* 657 * If the lockref_put_return() failed due to the lock being held 658 * by somebody else, the fast path has failed. We will need to 659 * get the lock, and then check the count again. 660 */ 661 if (unlikely(ret < 0)) { 662 spin_lock(&dentry->d_lock); 663 if (dentry->d_lockref.count > 1) { 664 dentry->d_lockref.count--; 665 spin_unlock(&dentry->d_lock); 666 return 1; 667 } 668 return 0; 669 } 670 671 /* 672 * If we weren't the last ref, we're done. 673 */ 674 if (ret) 675 return 1; 676 677 /* 678 * Careful, careful. The reference count went down 679 * to zero, but we don't hold the dentry lock, so 680 * somebody else could get it again, and do another 681 * dput(), and we need to not race with that. 682 * 683 * However, there is a very special and common case 684 * where we don't care, because there is nothing to 685 * do: the dentry is still hashed, it does not have 686 * a 'delete' op, and it's referenced and already on 687 * the LRU list. 688 * 689 * NOTE! Since we aren't locked, these values are 690 * not "stable". However, it is sufficient that at 691 * some point after we dropped the reference the 692 * dentry was hashed and the flags had the proper 693 * value. Other dentry users may have re-gotten 694 * a reference to the dentry and change that, but 695 * our work is done - we can leave the dentry 696 * around with a zero refcount. 697 */ 698 smp_rmb(); 699 d_flags = ACCESS_ONCE(dentry->d_flags); 700 d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST; 701 702 /* Nothing to do? Dropping the reference was all we needed? */ 703 if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry)) 704 return 1; 705 706 /* 707 * Not the fast normal case? Get the lock. We've already decremented 708 * the refcount, but we'll need to re-check the situation after 709 * getting the lock. 710 */ 711 spin_lock(&dentry->d_lock); 712 713 /* 714 * Did somebody else grab a reference to it in the meantime, and 715 * we're no longer the last user after all? Alternatively, somebody 716 * else could have killed it and marked it dead. Either way, we 717 * don't need to do anything else. 718 */ 719 if (dentry->d_lockref.count) { 720 spin_unlock(&dentry->d_lock); 721 return 1; 722 } 723 724 /* 725 * Re-get the reference we optimistically dropped. We hold the 726 * lock, and we just tested that it was zero, so we can just 727 * set it to 1. 728 */ 729 dentry->d_lockref.count = 1; 730 return 0; 731 } 732 733 734 /* 735 * This is dput 736 * 737 * This is complicated by the fact that we do not want to put 738 * dentries that are no longer on any hash chain on the unused 739 * list: we'd much rather just get rid of them immediately. 740 * 741 * However, that implies that we have to traverse the dentry 742 * tree upwards to the parents which might _also_ now be 743 * scheduled for deletion (it may have been only waiting for 744 * its last child to go away). 745 * 746 * This tail recursion is done by hand as we don't want to depend 747 * on the compiler to always get this right (gcc generally doesn't). 748 * Real recursion would eat up our stack space. 749 */ 750 751 /* 752 * dput - release a dentry 753 * @dentry: dentry to release 754 * 755 * Release a dentry. This will drop the usage count and if appropriate 756 * call the dentry unlink method as well as removing it from the queues and 757 * releasing its resources. If the parent dentries were scheduled for release 758 * they too may now get deleted. 759 */ 760 void dput(struct dentry *dentry) 761 { 762 if (unlikely(!dentry)) 763 return; 764 765 repeat: 766 rcu_read_lock(); 767 if (likely(fast_dput(dentry))) { 768 rcu_read_unlock(); 769 return; 770 } 771 772 /* Slow case: now with the dentry lock held */ 773 rcu_read_unlock(); 774 775 /* Unreachable? Get rid of it */ 776 if (unlikely(d_unhashed(dentry))) 777 goto kill_it; 778 779 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) { 780 if (dentry->d_op->d_delete(dentry)) 781 goto kill_it; 782 } 783 784 if (!(dentry->d_flags & DCACHE_REFERENCED)) 785 dentry->d_flags |= DCACHE_REFERENCED; 786 dentry_lru_add(dentry); 787 788 dentry->d_lockref.count--; 789 spin_unlock(&dentry->d_lock); 790 return; 791 792 kill_it: 793 dentry = dentry_kill(dentry); 794 if (dentry) 795 goto repeat; 796 } 797 EXPORT_SYMBOL(dput); 798 799 800 /* This must be called with d_lock held */ 801 static inline void __dget_dlock(struct dentry *dentry) 802 { 803 dentry->d_lockref.count++; 804 } 805 806 static inline void __dget(struct dentry *dentry) 807 { 808 lockref_get(&dentry->d_lockref); 809 } 810 811 struct dentry *dget_parent(struct dentry *dentry) 812 { 813 int gotref; 814 struct dentry *ret; 815 816 /* 817 * Do optimistic parent lookup without any 818 * locking. 819 */ 820 rcu_read_lock(); 821 ret = ACCESS_ONCE(dentry->d_parent); 822 gotref = lockref_get_not_zero(&ret->d_lockref); 823 rcu_read_unlock(); 824 if (likely(gotref)) { 825 if (likely(ret == ACCESS_ONCE(dentry->d_parent))) 826 return ret; 827 dput(ret); 828 } 829 830 repeat: 831 /* 832 * Don't need rcu_dereference because we re-check it was correct under 833 * the lock. 834 */ 835 rcu_read_lock(); 836 ret = dentry->d_parent; 837 spin_lock(&ret->d_lock); 838 if (unlikely(ret != dentry->d_parent)) { 839 spin_unlock(&ret->d_lock); 840 rcu_read_unlock(); 841 goto repeat; 842 } 843 rcu_read_unlock(); 844 BUG_ON(!ret->d_lockref.count); 845 ret->d_lockref.count++; 846 spin_unlock(&ret->d_lock); 847 return ret; 848 } 849 EXPORT_SYMBOL(dget_parent); 850 851 /** 852 * d_find_alias - grab a hashed alias of inode 853 * @inode: inode in question 854 * 855 * If inode has a hashed alias, or is a directory and has any alias, 856 * acquire the reference to alias and return it. Otherwise return NULL. 857 * Notice that if inode is a directory there can be only one alias and 858 * it can be unhashed only if it has no children, or if it is the root 859 * of a filesystem, or if the directory was renamed and d_revalidate 860 * was the first vfs operation to notice. 861 * 862 * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer 863 * any other hashed alias over that one. 864 */ 865 static struct dentry *__d_find_alias(struct inode *inode) 866 { 867 struct dentry *alias, *discon_alias; 868 869 again: 870 discon_alias = NULL; 871 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) { 872 spin_lock(&alias->d_lock); 873 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) { 874 if (IS_ROOT(alias) && 875 (alias->d_flags & DCACHE_DISCONNECTED)) { 876 discon_alias = alias; 877 } else { 878 __dget_dlock(alias); 879 spin_unlock(&alias->d_lock); 880 return alias; 881 } 882 } 883 spin_unlock(&alias->d_lock); 884 } 885 if (discon_alias) { 886 alias = discon_alias; 887 spin_lock(&alias->d_lock); 888 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) { 889 __dget_dlock(alias); 890 spin_unlock(&alias->d_lock); 891 return alias; 892 } 893 spin_unlock(&alias->d_lock); 894 goto again; 895 } 896 return NULL; 897 } 898 899 struct dentry *d_find_alias(struct inode *inode) 900 { 901 struct dentry *de = NULL; 902 903 if (!hlist_empty(&inode->i_dentry)) { 904 spin_lock(&inode->i_lock); 905 de = __d_find_alias(inode); 906 spin_unlock(&inode->i_lock); 907 } 908 return de; 909 } 910 EXPORT_SYMBOL(d_find_alias); 911 912 /* 913 * Try to kill dentries associated with this inode. 914 * WARNING: you must own a reference to inode. 915 */ 916 void d_prune_aliases(struct inode *inode) 917 { 918 struct dentry *dentry; 919 restart: 920 spin_lock(&inode->i_lock); 921 hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) { 922 spin_lock(&dentry->d_lock); 923 if (!dentry->d_lockref.count) { 924 struct dentry *parent = lock_parent(dentry); 925 if (likely(!dentry->d_lockref.count)) { 926 __dentry_kill(dentry); 927 dput(parent); 928 goto restart; 929 } 930 if (parent) 931 spin_unlock(&parent->d_lock); 932 } 933 spin_unlock(&dentry->d_lock); 934 } 935 spin_unlock(&inode->i_lock); 936 } 937 EXPORT_SYMBOL(d_prune_aliases); 938 939 static void shrink_dentry_list(struct list_head *list) 940 { 941 struct dentry *dentry, *parent; 942 943 while (!list_empty(list)) { 944 struct inode *inode; 945 dentry = list_entry(list->prev, struct dentry, d_lru); 946 spin_lock(&dentry->d_lock); 947 parent = lock_parent(dentry); 948 949 /* 950 * The dispose list is isolated and dentries are not accounted 951 * to the LRU here, so we can simply remove it from the list 952 * here regardless of whether it is referenced or not. 953 */ 954 d_shrink_del(dentry); 955 956 /* 957 * We found an inuse dentry which was not removed from 958 * the LRU because of laziness during lookup. Do not free it. 959 */ 960 if (dentry->d_lockref.count > 0) { 961 spin_unlock(&dentry->d_lock); 962 if (parent) 963 spin_unlock(&parent->d_lock); 964 continue; 965 } 966 967 968 if (unlikely(dentry->d_flags & DCACHE_DENTRY_KILLED)) { 969 bool can_free = dentry->d_flags & DCACHE_MAY_FREE; 970 spin_unlock(&dentry->d_lock); 971 if (parent) 972 spin_unlock(&parent->d_lock); 973 if (can_free) 974 dentry_free(dentry); 975 continue; 976 } 977 978 inode = dentry->d_inode; 979 if (inode && unlikely(!spin_trylock(&inode->i_lock))) { 980 d_shrink_add(dentry, list); 981 spin_unlock(&dentry->d_lock); 982 if (parent) 983 spin_unlock(&parent->d_lock); 984 continue; 985 } 986 987 __dentry_kill(dentry); 988 989 /* 990 * We need to prune ancestors too. This is necessary to prevent 991 * quadratic behavior of shrink_dcache_parent(), but is also 992 * expected to be beneficial in reducing dentry cache 993 * fragmentation. 994 */ 995 dentry = parent; 996 while (dentry && !lockref_put_or_lock(&dentry->d_lockref)) { 997 parent = lock_parent(dentry); 998 if (dentry->d_lockref.count != 1) { 999 dentry->d_lockref.count--; 1000 spin_unlock(&dentry->d_lock); 1001 if (parent) 1002 spin_unlock(&parent->d_lock); 1003 break; 1004 } 1005 inode = dentry->d_inode; /* can't be NULL */ 1006 if (unlikely(!spin_trylock(&inode->i_lock))) { 1007 spin_unlock(&dentry->d_lock); 1008 if (parent) 1009 spin_unlock(&parent->d_lock); 1010 cpu_relax(); 1011 continue; 1012 } 1013 __dentry_kill(dentry); 1014 dentry = parent; 1015 } 1016 } 1017 } 1018 1019 static enum lru_status dentry_lru_isolate(struct list_head *item, 1020 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg) 1021 { 1022 struct list_head *freeable = arg; 1023 struct dentry *dentry = container_of(item, struct dentry, d_lru); 1024 1025 1026 /* 1027 * we are inverting the lru lock/dentry->d_lock here, 1028 * so use a trylock. If we fail to get the lock, just skip 1029 * it 1030 */ 1031 if (!spin_trylock(&dentry->d_lock)) 1032 return LRU_SKIP; 1033 1034 /* 1035 * Referenced dentries are still in use. If they have active 1036 * counts, just remove them from the LRU. Otherwise give them 1037 * another pass through the LRU. 1038 */ 1039 if (dentry->d_lockref.count) { 1040 d_lru_isolate(lru, dentry); 1041 spin_unlock(&dentry->d_lock); 1042 return LRU_REMOVED; 1043 } 1044 1045 if (dentry->d_flags & DCACHE_REFERENCED) { 1046 dentry->d_flags &= ~DCACHE_REFERENCED; 1047 spin_unlock(&dentry->d_lock); 1048 1049 /* 1050 * The list move itself will be made by the common LRU code. At 1051 * this point, we've dropped the dentry->d_lock but keep the 1052 * lru lock. This is safe to do, since every list movement is 1053 * protected by the lru lock even if both locks are held. 1054 * 1055 * This is guaranteed by the fact that all LRU management 1056 * functions are intermediated by the LRU API calls like 1057 * list_lru_add and list_lru_del. List movement in this file 1058 * only ever occur through this functions or through callbacks 1059 * like this one, that are called from the LRU API. 1060 * 1061 * The only exceptions to this are functions like 1062 * shrink_dentry_list, and code that first checks for the 1063 * DCACHE_SHRINK_LIST flag. Those are guaranteed to be 1064 * operating only with stack provided lists after they are 1065 * properly isolated from the main list. It is thus, always a 1066 * local access. 1067 */ 1068 return LRU_ROTATE; 1069 } 1070 1071 d_lru_shrink_move(lru, dentry, freeable); 1072 spin_unlock(&dentry->d_lock); 1073 1074 return LRU_REMOVED; 1075 } 1076 1077 /** 1078 * prune_dcache_sb - shrink the dcache 1079 * @sb: superblock 1080 * @sc: shrink control, passed to list_lru_shrink_walk() 1081 * 1082 * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This 1083 * is done when we need more memory and called from the superblock shrinker 1084 * function. 1085 * 1086 * This function may fail to free any resources if all the dentries are in 1087 * use. 1088 */ 1089 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc) 1090 { 1091 LIST_HEAD(dispose); 1092 long freed; 1093 1094 freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc, 1095 dentry_lru_isolate, &dispose); 1096 shrink_dentry_list(&dispose); 1097 return freed; 1098 } 1099 1100 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item, 1101 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg) 1102 { 1103 struct list_head *freeable = arg; 1104 struct dentry *dentry = container_of(item, struct dentry, d_lru); 1105 1106 /* 1107 * we are inverting the lru lock/dentry->d_lock here, 1108 * so use a trylock. If we fail to get the lock, just skip 1109 * it 1110 */ 1111 if (!spin_trylock(&dentry->d_lock)) 1112 return LRU_SKIP; 1113 1114 d_lru_shrink_move(lru, dentry, freeable); 1115 spin_unlock(&dentry->d_lock); 1116 1117 return LRU_REMOVED; 1118 } 1119 1120 1121 /** 1122 * shrink_dcache_sb - shrink dcache for a superblock 1123 * @sb: superblock 1124 * 1125 * Shrink the dcache for the specified super block. This is used to free 1126 * the dcache before unmounting a file system. 1127 */ 1128 void shrink_dcache_sb(struct super_block *sb) 1129 { 1130 long freed; 1131 1132 do { 1133 LIST_HEAD(dispose); 1134 1135 freed = list_lru_walk(&sb->s_dentry_lru, 1136 dentry_lru_isolate_shrink, &dispose, UINT_MAX); 1137 1138 this_cpu_sub(nr_dentry_unused, freed); 1139 shrink_dentry_list(&dispose); 1140 } while (freed > 0); 1141 } 1142 EXPORT_SYMBOL(shrink_dcache_sb); 1143 1144 /** 1145 * enum d_walk_ret - action to talke during tree walk 1146 * @D_WALK_CONTINUE: contrinue walk 1147 * @D_WALK_QUIT: quit walk 1148 * @D_WALK_NORETRY: quit when retry is needed 1149 * @D_WALK_SKIP: skip this dentry and its children 1150 */ 1151 enum d_walk_ret { 1152 D_WALK_CONTINUE, 1153 D_WALK_QUIT, 1154 D_WALK_NORETRY, 1155 D_WALK_SKIP, 1156 }; 1157 1158 /** 1159 * d_walk - walk the dentry tree 1160 * @parent: start of walk 1161 * @data: data passed to @enter() and @finish() 1162 * @enter: callback when first entering the dentry 1163 * @finish: callback when successfully finished the walk 1164 * 1165 * The @enter() and @finish() callbacks are called with d_lock held. 1166 */ 1167 static void d_walk(struct dentry *parent, void *data, 1168 enum d_walk_ret (*enter)(void *, struct dentry *), 1169 void (*finish)(void *)) 1170 { 1171 struct dentry *this_parent; 1172 struct list_head *next; 1173 unsigned seq = 0; 1174 enum d_walk_ret ret; 1175 bool retry = true; 1176 1177 again: 1178 read_seqbegin_or_lock(&rename_lock, &seq); 1179 this_parent = parent; 1180 spin_lock(&this_parent->d_lock); 1181 1182 ret = enter(data, this_parent); 1183 switch (ret) { 1184 case D_WALK_CONTINUE: 1185 break; 1186 case D_WALK_QUIT: 1187 case D_WALK_SKIP: 1188 goto out_unlock; 1189 case D_WALK_NORETRY: 1190 retry = false; 1191 break; 1192 } 1193 repeat: 1194 next = this_parent->d_subdirs.next; 1195 resume: 1196 while (next != &this_parent->d_subdirs) { 1197 struct list_head *tmp = next; 1198 struct dentry *dentry = list_entry(tmp, struct dentry, d_child); 1199 next = tmp->next; 1200 1201 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED); 1202 1203 ret = enter(data, dentry); 1204 switch (ret) { 1205 case D_WALK_CONTINUE: 1206 break; 1207 case D_WALK_QUIT: 1208 spin_unlock(&dentry->d_lock); 1209 goto out_unlock; 1210 case D_WALK_NORETRY: 1211 retry = false; 1212 break; 1213 case D_WALK_SKIP: 1214 spin_unlock(&dentry->d_lock); 1215 continue; 1216 } 1217 1218 if (!list_empty(&dentry->d_subdirs)) { 1219 spin_unlock(&this_parent->d_lock); 1220 spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_); 1221 this_parent = dentry; 1222 spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_); 1223 goto repeat; 1224 } 1225 spin_unlock(&dentry->d_lock); 1226 } 1227 /* 1228 * All done at this level ... ascend and resume the search. 1229 */ 1230 rcu_read_lock(); 1231 ascend: 1232 if (this_parent != parent) { 1233 struct dentry *child = this_parent; 1234 this_parent = child->d_parent; 1235 1236 spin_unlock(&child->d_lock); 1237 spin_lock(&this_parent->d_lock); 1238 1239 /* might go back up the wrong parent if we have had a rename. */ 1240 if (need_seqretry(&rename_lock, seq)) 1241 goto rename_retry; 1242 /* go into the first sibling still alive */ 1243 do { 1244 next = child->d_child.next; 1245 if (next == &this_parent->d_subdirs) 1246 goto ascend; 1247 child = list_entry(next, struct dentry, d_child); 1248 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED)); 1249 rcu_read_unlock(); 1250 goto resume; 1251 } 1252 if (need_seqretry(&rename_lock, seq)) 1253 goto rename_retry; 1254 rcu_read_unlock(); 1255 if (finish) 1256 finish(data); 1257 1258 out_unlock: 1259 spin_unlock(&this_parent->d_lock); 1260 done_seqretry(&rename_lock, seq); 1261 return; 1262 1263 rename_retry: 1264 spin_unlock(&this_parent->d_lock); 1265 rcu_read_unlock(); 1266 BUG_ON(seq & 1); 1267 if (!retry) 1268 return; 1269 seq = 1; 1270 goto again; 1271 } 1272 1273 /* 1274 * Search for at least 1 mount point in the dentry's subdirs. 1275 * We descend to the next level whenever the d_subdirs 1276 * list is non-empty and continue searching. 1277 */ 1278 1279 static enum d_walk_ret check_mount(void *data, struct dentry *dentry) 1280 { 1281 int *ret = data; 1282 if (d_mountpoint(dentry)) { 1283 *ret = 1; 1284 return D_WALK_QUIT; 1285 } 1286 return D_WALK_CONTINUE; 1287 } 1288 1289 /** 1290 * have_submounts - check for mounts over a dentry 1291 * @parent: dentry to check. 1292 * 1293 * Return true if the parent or its subdirectories contain 1294 * a mount point 1295 */ 1296 int have_submounts(struct dentry *parent) 1297 { 1298 int ret = 0; 1299 1300 d_walk(parent, &ret, check_mount, NULL); 1301 1302 return ret; 1303 } 1304 EXPORT_SYMBOL(have_submounts); 1305 1306 /* 1307 * Called by mount code to set a mountpoint and check if the mountpoint is 1308 * reachable (e.g. NFS can unhash a directory dentry and then the complete 1309 * subtree can become unreachable). 1310 * 1311 * Only one of d_invalidate() and d_set_mounted() must succeed. For 1312 * this reason take rename_lock and d_lock on dentry and ancestors. 1313 */ 1314 int d_set_mounted(struct dentry *dentry) 1315 { 1316 struct dentry *p; 1317 int ret = -ENOENT; 1318 write_seqlock(&rename_lock); 1319 for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) { 1320 /* Need exclusion wrt. d_invalidate() */ 1321 spin_lock(&p->d_lock); 1322 if (unlikely(d_unhashed(p))) { 1323 spin_unlock(&p->d_lock); 1324 goto out; 1325 } 1326 spin_unlock(&p->d_lock); 1327 } 1328 spin_lock(&dentry->d_lock); 1329 if (!d_unlinked(dentry)) { 1330 dentry->d_flags |= DCACHE_MOUNTED; 1331 ret = 0; 1332 } 1333 spin_unlock(&dentry->d_lock); 1334 out: 1335 write_sequnlock(&rename_lock); 1336 return ret; 1337 } 1338 1339 /* 1340 * Search the dentry child list of the specified parent, 1341 * and move any unused dentries to the end of the unused 1342 * list for prune_dcache(). We descend to the next level 1343 * whenever the d_subdirs list is non-empty and continue 1344 * searching. 1345 * 1346 * It returns zero iff there are no unused children, 1347 * otherwise it returns the number of children moved to 1348 * the end of the unused list. This may not be the total 1349 * number of unused children, because select_parent can 1350 * drop the lock and return early due to latency 1351 * constraints. 1352 */ 1353 1354 struct select_data { 1355 struct dentry *start; 1356 struct list_head dispose; 1357 int found; 1358 }; 1359 1360 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry) 1361 { 1362 struct select_data *data = _data; 1363 enum d_walk_ret ret = D_WALK_CONTINUE; 1364 1365 if (data->start == dentry) 1366 goto out; 1367 1368 if (dentry->d_flags & DCACHE_SHRINK_LIST) { 1369 data->found++; 1370 } else { 1371 if (dentry->d_flags & DCACHE_LRU_LIST) 1372 d_lru_del(dentry); 1373 if (!dentry->d_lockref.count) { 1374 d_shrink_add(dentry, &data->dispose); 1375 data->found++; 1376 } 1377 } 1378 /* 1379 * We can return to the caller if we have found some (this 1380 * ensures forward progress). We'll be coming back to find 1381 * the rest. 1382 */ 1383 if (!list_empty(&data->dispose)) 1384 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY; 1385 out: 1386 return ret; 1387 } 1388 1389 /** 1390 * shrink_dcache_parent - prune dcache 1391 * @parent: parent of entries to prune 1392 * 1393 * Prune the dcache to remove unused children of the parent dentry. 1394 */ 1395 void shrink_dcache_parent(struct dentry *parent) 1396 { 1397 for (;;) { 1398 struct select_data data; 1399 1400 INIT_LIST_HEAD(&data.dispose); 1401 data.start = parent; 1402 data.found = 0; 1403 1404 d_walk(parent, &data, select_collect, NULL); 1405 if (!data.found) 1406 break; 1407 1408 shrink_dentry_list(&data.dispose); 1409 cond_resched(); 1410 } 1411 } 1412 EXPORT_SYMBOL(shrink_dcache_parent); 1413 1414 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry) 1415 { 1416 /* it has busy descendents; complain about those instead */ 1417 if (!list_empty(&dentry->d_subdirs)) 1418 return D_WALK_CONTINUE; 1419 1420 /* root with refcount 1 is fine */ 1421 if (dentry == _data && dentry->d_lockref.count == 1) 1422 return D_WALK_CONTINUE; 1423 1424 printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} " 1425 " still in use (%d) [unmount of %s %s]\n", 1426 dentry, 1427 dentry->d_inode ? 1428 dentry->d_inode->i_ino : 0UL, 1429 dentry, 1430 dentry->d_lockref.count, 1431 dentry->d_sb->s_type->name, 1432 dentry->d_sb->s_id); 1433 WARN_ON(1); 1434 return D_WALK_CONTINUE; 1435 } 1436 1437 static void do_one_tree(struct dentry *dentry) 1438 { 1439 shrink_dcache_parent(dentry); 1440 d_walk(dentry, dentry, umount_check, NULL); 1441 d_drop(dentry); 1442 dput(dentry); 1443 } 1444 1445 /* 1446 * destroy the dentries attached to a superblock on unmounting 1447 */ 1448 void shrink_dcache_for_umount(struct super_block *sb) 1449 { 1450 struct dentry *dentry; 1451 1452 WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked"); 1453 1454 dentry = sb->s_root; 1455 sb->s_root = NULL; 1456 do_one_tree(dentry); 1457 1458 while (!hlist_bl_empty(&sb->s_anon)) { 1459 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_anon), struct dentry, d_hash)); 1460 do_one_tree(dentry); 1461 } 1462 } 1463 1464 struct detach_data { 1465 struct select_data select; 1466 struct dentry *mountpoint; 1467 }; 1468 static enum d_walk_ret detach_and_collect(void *_data, struct dentry *dentry) 1469 { 1470 struct detach_data *data = _data; 1471 1472 if (d_mountpoint(dentry)) { 1473 __dget_dlock(dentry); 1474 data->mountpoint = dentry; 1475 return D_WALK_QUIT; 1476 } 1477 1478 return select_collect(&data->select, dentry); 1479 } 1480 1481 static void check_and_drop(void *_data) 1482 { 1483 struct detach_data *data = _data; 1484 1485 if (!data->mountpoint && !data->select.found) 1486 __d_drop(data->select.start); 1487 } 1488 1489 /** 1490 * d_invalidate - detach submounts, prune dcache, and drop 1491 * @dentry: dentry to invalidate (aka detach, prune and drop) 1492 * 1493 * no dcache lock. 1494 * 1495 * The final d_drop is done as an atomic operation relative to 1496 * rename_lock ensuring there are no races with d_set_mounted. This 1497 * ensures there are no unhashed dentries on the path to a mountpoint. 1498 */ 1499 void d_invalidate(struct dentry *dentry) 1500 { 1501 /* 1502 * If it's already been dropped, return OK. 1503 */ 1504 spin_lock(&dentry->d_lock); 1505 if (d_unhashed(dentry)) { 1506 spin_unlock(&dentry->d_lock); 1507 return; 1508 } 1509 spin_unlock(&dentry->d_lock); 1510 1511 /* Negative dentries can be dropped without further checks */ 1512 if (!dentry->d_inode) { 1513 d_drop(dentry); 1514 return; 1515 } 1516 1517 for (;;) { 1518 struct detach_data data; 1519 1520 data.mountpoint = NULL; 1521 INIT_LIST_HEAD(&data.select.dispose); 1522 data.select.start = dentry; 1523 data.select.found = 0; 1524 1525 d_walk(dentry, &data, detach_and_collect, check_and_drop); 1526 1527 if (data.select.found) 1528 shrink_dentry_list(&data.select.dispose); 1529 1530 if (data.mountpoint) { 1531 detach_mounts(data.mountpoint); 1532 dput(data.mountpoint); 1533 } 1534 1535 if (!data.mountpoint && !data.select.found) 1536 break; 1537 1538 cond_resched(); 1539 } 1540 } 1541 EXPORT_SYMBOL(d_invalidate); 1542 1543 /** 1544 * __d_alloc - allocate a dcache entry 1545 * @sb: filesystem it will belong to 1546 * @name: qstr of the name 1547 * 1548 * Allocates a dentry. It returns %NULL if there is insufficient memory 1549 * available. On a success the dentry is returned. The name passed in is 1550 * copied and the copy passed in may be reused after this call. 1551 */ 1552 1553 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name) 1554 { 1555 struct dentry *dentry; 1556 char *dname; 1557 1558 dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL); 1559 if (!dentry) 1560 return NULL; 1561 1562 /* 1563 * We guarantee that the inline name is always NUL-terminated. 1564 * This way the memcpy() done by the name switching in rename 1565 * will still always have a NUL at the end, even if we might 1566 * be overwriting an internal NUL character 1567 */ 1568 dentry->d_iname[DNAME_INLINE_LEN-1] = 0; 1569 if (name->len > DNAME_INLINE_LEN-1) { 1570 size_t size = offsetof(struct external_name, name[1]); 1571 struct external_name *p = kmalloc(size + name->len, GFP_KERNEL); 1572 if (!p) { 1573 kmem_cache_free(dentry_cache, dentry); 1574 return NULL; 1575 } 1576 atomic_set(&p->u.count, 1); 1577 dname = p->name; 1578 if (IS_ENABLED(CONFIG_DCACHE_WORD_ACCESS)) 1579 kasan_unpoison_shadow(dname, 1580 round_up(name->len + 1, sizeof(unsigned long))); 1581 } else { 1582 dname = dentry->d_iname; 1583 } 1584 1585 dentry->d_name.len = name->len; 1586 dentry->d_name.hash = name->hash; 1587 memcpy(dname, name->name, name->len); 1588 dname[name->len] = 0; 1589 1590 /* Make sure we always see the terminating NUL character */ 1591 smp_wmb(); 1592 dentry->d_name.name = dname; 1593 1594 dentry->d_lockref.count = 1; 1595 dentry->d_flags = 0; 1596 spin_lock_init(&dentry->d_lock); 1597 seqcount_init(&dentry->d_seq); 1598 dentry->d_inode = NULL; 1599 dentry->d_parent = dentry; 1600 dentry->d_sb = sb; 1601 dentry->d_op = NULL; 1602 dentry->d_fsdata = NULL; 1603 INIT_HLIST_BL_NODE(&dentry->d_hash); 1604 INIT_LIST_HEAD(&dentry->d_lru); 1605 INIT_LIST_HEAD(&dentry->d_subdirs); 1606 INIT_HLIST_NODE(&dentry->d_u.d_alias); 1607 INIT_LIST_HEAD(&dentry->d_child); 1608 d_set_d_op(dentry, dentry->d_sb->s_d_op); 1609 1610 this_cpu_inc(nr_dentry); 1611 1612 return dentry; 1613 } 1614 1615 /** 1616 * d_alloc - allocate a dcache entry 1617 * @parent: parent of entry to allocate 1618 * @name: qstr of the name 1619 * 1620 * Allocates a dentry. It returns %NULL if there is insufficient memory 1621 * available. On a success the dentry is returned. The name passed in is 1622 * copied and the copy passed in may be reused after this call. 1623 */ 1624 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name) 1625 { 1626 struct dentry *dentry = __d_alloc(parent->d_sb, name); 1627 if (!dentry) 1628 return NULL; 1629 1630 spin_lock(&parent->d_lock); 1631 /* 1632 * don't need child lock because it is not subject 1633 * to concurrency here 1634 */ 1635 __dget_dlock(parent); 1636 dentry->d_parent = parent; 1637 list_add(&dentry->d_child, &parent->d_subdirs); 1638 spin_unlock(&parent->d_lock); 1639 1640 return dentry; 1641 } 1642 EXPORT_SYMBOL(d_alloc); 1643 1644 /** 1645 * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems) 1646 * @sb: the superblock 1647 * @name: qstr of the name 1648 * 1649 * For a filesystem that just pins its dentries in memory and never 1650 * performs lookups at all, return an unhashed IS_ROOT dentry. 1651 */ 1652 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name) 1653 { 1654 return __d_alloc(sb, name); 1655 } 1656 EXPORT_SYMBOL(d_alloc_pseudo); 1657 1658 struct dentry *d_alloc_name(struct dentry *parent, const char *name) 1659 { 1660 struct qstr q; 1661 1662 q.name = name; 1663 q.len = strlen(name); 1664 q.hash = full_name_hash(q.name, q.len); 1665 return d_alloc(parent, &q); 1666 } 1667 EXPORT_SYMBOL(d_alloc_name); 1668 1669 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op) 1670 { 1671 WARN_ON_ONCE(dentry->d_op); 1672 WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH | 1673 DCACHE_OP_COMPARE | 1674 DCACHE_OP_REVALIDATE | 1675 DCACHE_OP_WEAK_REVALIDATE | 1676 DCACHE_OP_DELETE | 1677 DCACHE_OP_SELECT_INODE)); 1678 dentry->d_op = op; 1679 if (!op) 1680 return; 1681 if (op->d_hash) 1682 dentry->d_flags |= DCACHE_OP_HASH; 1683 if (op->d_compare) 1684 dentry->d_flags |= DCACHE_OP_COMPARE; 1685 if (op->d_revalidate) 1686 dentry->d_flags |= DCACHE_OP_REVALIDATE; 1687 if (op->d_weak_revalidate) 1688 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE; 1689 if (op->d_delete) 1690 dentry->d_flags |= DCACHE_OP_DELETE; 1691 if (op->d_prune) 1692 dentry->d_flags |= DCACHE_OP_PRUNE; 1693 if (op->d_select_inode) 1694 dentry->d_flags |= DCACHE_OP_SELECT_INODE; 1695 1696 } 1697 EXPORT_SYMBOL(d_set_d_op); 1698 1699 1700 /* 1701 * d_set_fallthru - Mark a dentry as falling through to a lower layer 1702 * @dentry - The dentry to mark 1703 * 1704 * Mark a dentry as falling through to the lower layer (as set with 1705 * d_pin_lower()). This flag may be recorded on the medium. 1706 */ 1707 void d_set_fallthru(struct dentry *dentry) 1708 { 1709 spin_lock(&dentry->d_lock); 1710 dentry->d_flags |= DCACHE_FALLTHRU; 1711 spin_unlock(&dentry->d_lock); 1712 } 1713 EXPORT_SYMBOL(d_set_fallthru); 1714 1715 static unsigned d_flags_for_inode(struct inode *inode) 1716 { 1717 unsigned add_flags = DCACHE_REGULAR_TYPE; 1718 1719 if (!inode) 1720 return DCACHE_MISS_TYPE; 1721 1722 if (S_ISDIR(inode->i_mode)) { 1723 add_flags = DCACHE_DIRECTORY_TYPE; 1724 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) { 1725 if (unlikely(!inode->i_op->lookup)) 1726 add_flags = DCACHE_AUTODIR_TYPE; 1727 else 1728 inode->i_opflags |= IOP_LOOKUP; 1729 } 1730 goto type_determined; 1731 } 1732 1733 if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) { 1734 if (unlikely(inode->i_op->follow_link)) { 1735 add_flags = DCACHE_SYMLINK_TYPE; 1736 goto type_determined; 1737 } 1738 inode->i_opflags |= IOP_NOFOLLOW; 1739 } 1740 1741 if (unlikely(!S_ISREG(inode->i_mode))) 1742 add_flags = DCACHE_SPECIAL_TYPE; 1743 1744 type_determined: 1745 if (unlikely(IS_AUTOMOUNT(inode))) 1746 add_flags |= DCACHE_NEED_AUTOMOUNT; 1747 return add_flags; 1748 } 1749 1750 static void __d_instantiate(struct dentry *dentry, struct inode *inode) 1751 { 1752 unsigned add_flags = d_flags_for_inode(inode); 1753 1754 spin_lock(&dentry->d_lock); 1755 if (inode) 1756 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry); 1757 __d_set_inode_and_type(dentry, inode, add_flags); 1758 dentry_rcuwalk_invalidate(dentry); 1759 spin_unlock(&dentry->d_lock); 1760 fsnotify_d_instantiate(dentry, inode); 1761 } 1762 1763 /** 1764 * d_instantiate - fill in inode information for a dentry 1765 * @entry: dentry to complete 1766 * @inode: inode to attach to this dentry 1767 * 1768 * Fill in inode information in the entry. 1769 * 1770 * This turns negative dentries into productive full members 1771 * of society. 1772 * 1773 * NOTE! This assumes that the inode count has been incremented 1774 * (or otherwise set) by the caller to indicate that it is now 1775 * in use by the dcache. 1776 */ 1777 1778 void d_instantiate(struct dentry *entry, struct inode * inode) 1779 { 1780 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias)); 1781 if (inode) 1782 spin_lock(&inode->i_lock); 1783 __d_instantiate(entry, inode); 1784 if (inode) 1785 spin_unlock(&inode->i_lock); 1786 security_d_instantiate(entry, inode); 1787 } 1788 EXPORT_SYMBOL(d_instantiate); 1789 1790 /** 1791 * d_instantiate_unique - instantiate a non-aliased dentry 1792 * @entry: dentry to instantiate 1793 * @inode: inode to attach to this dentry 1794 * 1795 * Fill in inode information in the entry. On success, it returns NULL. 1796 * If an unhashed alias of "entry" already exists, then we return the 1797 * aliased dentry instead and drop one reference to inode. 1798 * 1799 * Note that in order to avoid conflicts with rename() etc, the caller 1800 * had better be holding the parent directory semaphore. 1801 * 1802 * This also assumes that the inode count has been incremented 1803 * (or otherwise set) by the caller to indicate that it is now 1804 * in use by the dcache. 1805 */ 1806 static struct dentry *__d_instantiate_unique(struct dentry *entry, 1807 struct inode *inode) 1808 { 1809 struct dentry *alias; 1810 int len = entry->d_name.len; 1811 const char *name = entry->d_name.name; 1812 unsigned int hash = entry->d_name.hash; 1813 1814 if (!inode) { 1815 __d_instantiate(entry, NULL); 1816 return NULL; 1817 } 1818 1819 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) { 1820 /* 1821 * Don't need alias->d_lock here, because aliases with 1822 * d_parent == entry->d_parent are not subject to name or 1823 * parent changes, because the parent inode i_mutex is held. 1824 */ 1825 if (alias->d_name.hash != hash) 1826 continue; 1827 if (alias->d_parent != entry->d_parent) 1828 continue; 1829 if (alias->d_name.len != len) 1830 continue; 1831 if (dentry_cmp(alias, name, len)) 1832 continue; 1833 __dget(alias); 1834 return alias; 1835 } 1836 1837 __d_instantiate(entry, inode); 1838 return NULL; 1839 } 1840 1841 struct dentry *d_instantiate_unique(struct dentry *entry, struct inode *inode) 1842 { 1843 struct dentry *result; 1844 1845 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias)); 1846 1847 if (inode) 1848 spin_lock(&inode->i_lock); 1849 result = __d_instantiate_unique(entry, inode); 1850 if (inode) 1851 spin_unlock(&inode->i_lock); 1852 1853 if (!result) { 1854 security_d_instantiate(entry, inode); 1855 return NULL; 1856 } 1857 1858 BUG_ON(!d_unhashed(result)); 1859 iput(inode); 1860 return result; 1861 } 1862 1863 EXPORT_SYMBOL(d_instantiate_unique); 1864 1865 /** 1866 * d_instantiate_no_diralias - instantiate a non-aliased dentry 1867 * @entry: dentry to complete 1868 * @inode: inode to attach to this dentry 1869 * 1870 * Fill in inode information in the entry. If a directory alias is found, then 1871 * return an error (and drop inode). Together with d_materialise_unique() this 1872 * guarantees that a directory inode may never have more than one alias. 1873 */ 1874 int d_instantiate_no_diralias(struct dentry *entry, struct inode *inode) 1875 { 1876 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias)); 1877 1878 spin_lock(&inode->i_lock); 1879 if (S_ISDIR(inode->i_mode) && !hlist_empty(&inode->i_dentry)) { 1880 spin_unlock(&inode->i_lock); 1881 iput(inode); 1882 return -EBUSY; 1883 } 1884 __d_instantiate(entry, inode); 1885 spin_unlock(&inode->i_lock); 1886 security_d_instantiate(entry, inode); 1887 1888 return 0; 1889 } 1890 EXPORT_SYMBOL(d_instantiate_no_diralias); 1891 1892 struct dentry *d_make_root(struct inode *root_inode) 1893 { 1894 struct dentry *res = NULL; 1895 1896 if (root_inode) { 1897 static const struct qstr name = QSTR_INIT("/", 1); 1898 1899 res = __d_alloc(root_inode->i_sb, &name); 1900 if (res) 1901 d_instantiate(res, root_inode); 1902 else 1903 iput(root_inode); 1904 } 1905 return res; 1906 } 1907 EXPORT_SYMBOL(d_make_root); 1908 1909 static struct dentry * __d_find_any_alias(struct inode *inode) 1910 { 1911 struct dentry *alias; 1912 1913 if (hlist_empty(&inode->i_dentry)) 1914 return NULL; 1915 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias); 1916 __dget(alias); 1917 return alias; 1918 } 1919 1920 /** 1921 * d_find_any_alias - find any alias for a given inode 1922 * @inode: inode to find an alias for 1923 * 1924 * If any aliases exist for the given inode, take and return a 1925 * reference for one of them. If no aliases exist, return %NULL. 1926 */ 1927 struct dentry *d_find_any_alias(struct inode *inode) 1928 { 1929 struct dentry *de; 1930 1931 spin_lock(&inode->i_lock); 1932 de = __d_find_any_alias(inode); 1933 spin_unlock(&inode->i_lock); 1934 return de; 1935 } 1936 EXPORT_SYMBOL(d_find_any_alias); 1937 1938 static struct dentry *__d_obtain_alias(struct inode *inode, int disconnected) 1939 { 1940 static const struct qstr anonstring = QSTR_INIT("/", 1); 1941 struct dentry *tmp; 1942 struct dentry *res; 1943 unsigned add_flags; 1944 1945 if (!inode) 1946 return ERR_PTR(-ESTALE); 1947 if (IS_ERR(inode)) 1948 return ERR_CAST(inode); 1949 1950 res = d_find_any_alias(inode); 1951 if (res) 1952 goto out_iput; 1953 1954 tmp = __d_alloc(inode->i_sb, &anonstring); 1955 if (!tmp) { 1956 res = ERR_PTR(-ENOMEM); 1957 goto out_iput; 1958 } 1959 1960 spin_lock(&inode->i_lock); 1961 res = __d_find_any_alias(inode); 1962 if (res) { 1963 spin_unlock(&inode->i_lock); 1964 dput(tmp); 1965 goto out_iput; 1966 } 1967 1968 /* attach a disconnected dentry */ 1969 add_flags = d_flags_for_inode(inode); 1970 1971 if (disconnected) 1972 add_flags |= DCACHE_DISCONNECTED; 1973 1974 spin_lock(&tmp->d_lock); 1975 __d_set_inode_and_type(tmp, inode, add_flags); 1976 hlist_add_head(&tmp->d_u.d_alias, &inode->i_dentry); 1977 hlist_bl_lock(&tmp->d_sb->s_anon); 1978 hlist_bl_add_head(&tmp->d_hash, &tmp->d_sb->s_anon); 1979 hlist_bl_unlock(&tmp->d_sb->s_anon); 1980 spin_unlock(&tmp->d_lock); 1981 spin_unlock(&inode->i_lock); 1982 security_d_instantiate(tmp, inode); 1983 1984 return tmp; 1985 1986 out_iput: 1987 if (res && !IS_ERR(res)) 1988 security_d_instantiate(res, inode); 1989 iput(inode); 1990 return res; 1991 } 1992 1993 /** 1994 * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode 1995 * @inode: inode to allocate the dentry for 1996 * 1997 * Obtain a dentry for an inode resulting from NFS filehandle conversion or 1998 * similar open by handle operations. The returned dentry may be anonymous, 1999 * or may have a full name (if the inode was already in the cache). 2000 * 2001 * When called on a directory inode, we must ensure that the inode only ever 2002 * has one dentry. If a dentry is found, that is returned instead of 2003 * allocating a new one. 2004 * 2005 * On successful return, the reference to the inode has been transferred 2006 * to the dentry. In case of an error the reference on the inode is released. 2007 * To make it easier to use in export operations a %NULL or IS_ERR inode may 2008 * be passed in and the error will be propagated to the return value, 2009 * with a %NULL @inode replaced by ERR_PTR(-ESTALE). 2010 */ 2011 struct dentry *d_obtain_alias(struct inode *inode) 2012 { 2013 return __d_obtain_alias(inode, 1); 2014 } 2015 EXPORT_SYMBOL(d_obtain_alias); 2016 2017 /** 2018 * d_obtain_root - find or allocate a dentry for a given inode 2019 * @inode: inode to allocate the dentry for 2020 * 2021 * Obtain an IS_ROOT dentry for the root of a filesystem. 2022 * 2023 * We must ensure that directory inodes only ever have one dentry. If a 2024 * dentry is found, that is returned instead of allocating a new one. 2025 * 2026 * On successful return, the reference to the inode has been transferred 2027 * to the dentry. In case of an error the reference on the inode is 2028 * released. A %NULL or IS_ERR inode may be passed in and will be the 2029 * error will be propagate to the return value, with a %NULL @inode 2030 * replaced by ERR_PTR(-ESTALE). 2031 */ 2032 struct dentry *d_obtain_root(struct inode *inode) 2033 { 2034 return __d_obtain_alias(inode, 0); 2035 } 2036 EXPORT_SYMBOL(d_obtain_root); 2037 2038 /** 2039 * d_add_ci - lookup or allocate new dentry with case-exact name 2040 * @inode: the inode case-insensitive lookup has found 2041 * @dentry: the negative dentry that was passed to the parent's lookup func 2042 * @name: the case-exact name to be associated with the returned dentry 2043 * 2044 * This is to avoid filling the dcache with case-insensitive names to the 2045 * same inode, only the actual correct case is stored in the dcache for 2046 * case-insensitive filesystems. 2047 * 2048 * For a case-insensitive lookup match and if the the case-exact dentry 2049 * already exists in in the dcache, use it and return it. 2050 * 2051 * If no entry exists with the exact case name, allocate new dentry with 2052 * the exact case, and return the spliced entry. 2053 */ 2054 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode, 2055 struct qstr *name) 2056 { 2057 struct dentry *found; 2058 struct dentry *new; 2059 2060 /* 2061 * First check if a dentry matching the name already exists, 2062 * if not go ahead and create it now. 2063 */ 2064 found = d_hash_and_lookup(dentry->d_parent, name); 2065 if (!found) { 2066 new = d_alloc(dentry->d_parent, name); 2067 if (!new) { 2068 found = ERR_PTR(-ENOMEM); 2069 } else { 2070 found = d_splice_alias(inode, new); 2071 if (found) { 2072 dput(new); 2073 return found; 2074 } 2075 return new; 2076 } 2077 } 2078 iput(inode); 2079 return found; 2080 } 2081 EXPORT_SYMBOL(d_add_ci); 2082 2083 /* 2084 * Do the slow-case of the dentry name compare. 2085 * 2086 * Unlike the dentry_cmp() function, we need to atomically 2087 * load the name and length information, so that the 2088 * filesystem can rely on them, and can use the 'name' and 2089 * 'len' information without worrying about walking off the 2090 * end of memory etc. 2091 * 2092 * Thus the read_seqcount_retry() and the "duplicate" info 2093 * in arguments (the low-level filesystem should not look 2094 * at the dentry inode or name contents directly, since 2095 * rename can change them while we're in RCU mode). 2096 */ 2097 enum slow_d_compare { 2098 D_COMP_OK, 2099 D_COMP_NOMATCH, 2100 D_COMP_SEQRETRY, 2101 }; 2102 2103 static noinline enum slow_d_compare slow_dentry_cmp( 2104 const struct dentry *parent, 2105 struct dentry *dentry, 2106 unsigned int seq, 2107 const struct qstr *name) 2108 { 2109 int tlen = dentry->d_name.len; 2110 const char *tname = dentry->d_name.name; 2111 2112 if (read_seqcount_retry(&dentry->d_seq, seq)) { 2113 cpu_relax(); 2114 return D_COMP_SEQRETRY; 2115 } 2116 if (parent->d_op->d_compare(parent, dentry, tlen, tname, name)) 2117 return D_COMP_NOMATCH; 2118 return D_COMP_OK; 2119 } 2120 2121 /** 2122 * __d_lookup_rcu - search for a dentry (racy, store-free) 2123 * @parent: parent dentry 2124 * @name: qstr of name we wish to find 2125 * @seqp: returns d_seq value at the point where the dentry was found 2126 * Returns: dentry, or NULL 2127 * 2128 * __d_lookup_rcu is the dcache lookup function for rcu-walk name 2129 * resolution (store-free path walking) design described in 2130 * Documentation/filesystems/path-lookup.txt. 2131 * 2132 * This is not to be used outside core vfs. 2133 * 2134 * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock 2135 * held, and rcu_read_lock held. The returned dentry must not be stored into 2136 * without taking d_lock and checking d_seq sequence count against @seq 2137 * returned here. 2138 * 2139 * A refcount may be taken on the found dentry with the d_rcu_to_refcount 2140 * function. 2141 * 2142 * Alternatively, __d_lookup_rcu may be called again to look up the child of 2143 * the returned dentry, so long as its parent's seqlock is checked after the 2144 * child is looked up. Thus, an interlocking stepping of sequence lock checks 2145 * is formed, giving integrity down the path walk. 2146 * 2147 * NOTE! The caller *has* to check the resulting dentry against the sequence 2148 * number we've returned before using any of the resulting dentry state! 2149 */ 2150 struct dentry *__d_lookup_rcu(const struct dentry *parent, 2151 const struct qstr *name, 2152 unsigned *seqp) 2153 { 2154 u64 hashlen = name->hash_len; 2155 const unsigned char *str = name->name; 2156 struct hlist_bl_head *b = d_hash(parent, hashlen_hash(hashlen)); 2157 struct hlist_bl_node *node; 2158 struct dentry *dentry; 2159 2160 /* 2161 * Note: There is significant duplication with __d_lookup_rcu which is 2162 * required to prevent single threaded performance regressions 2163 * especially on architectures where smp_rmb (in seqcounts) are costly. 2164 * Keep the two functions in sync. 2165 */ 2166 2167 /* 2168 * The hash list is protected using RCU. 2169 * 2170 * Carefully use d_seq when comparing a candidate dentry, to avoid 2171 * races with d_move(). 2172 * 2173 * It is possible that concurrent renames can mess up our list 2174 * walk here and result in missing our dentry, resulting in the 2175 * false-negative result. d_lookup() protects against concurrent 2176 * renames using rename_lock seqlock. 2177 * 2178 * See Documentation/filesystems/path-lookup.txt for more details. 2179 */ 2180 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) { 2181 unsigned seq; 2182 2183 seqretry: 2184 /* 2185 * The dentry sequence count protects us from concurrent 2186 * renames, and thus protects parent and name fields. 2187 * 2188 * The caller must perform a seqcount check in order 2189 * to do anything useful with the returned dentry. 2190 * 2191 * NOTE! We do a "raw" seqcount_begin here. That means that 2192 * we don't wait for the sequence count to stabilize if it 2193 * is in the middle of a sequence change. If we do the slow 2194 * dentry compare, we will do seqretries until it is stable, 2195 * and if we end up with a successful lookup, we actually 2196 * want to exit RCU lookup anyway. 2197 */ 2198 seq = raw_seqcount_begin(&dentry->d_seq); 2199 if (dentry->d_parent != parent) 2200 continue; 2201 if (d_unhashed(dentry)) 2202 continue; 2203 2204 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) { 2205 if (dentry->d_name.hash != hashlen_hash(hashlen)) 2206 continue; 2207 *seqp = seq; 2208 switch (slow_dentry_cmp(parent, dentry, seq, name)) { 2209 case D_COMP_OK: 2210 return dentry; 2211 case D_COMP_NOMATCH: 2212 continue; 2213 default: 2214 goto seqretry; 2215 } 2216 } 2217 2218 if (dentry->d_name.hash_len != hashlen) 2219 continue; 2220 *seqp = seq; 2221 if (!dentry_cmp(dentry, str, hashlen_len(hashlen))) 2222 return dentry; 2223 } 2224 return NULL; 2225 } 2226 2227 /** 2228 * d_lookup - search for a dentry 2229 * @parent: parent dentry 2230 * @name: qstr of name we wish to find 2231 * Returns: dentry, or NULL 2232 * 2233 * d_lookup searches the children of the parent dentry for the name in 2234 * question. If the dentry is found its reference count is incremented and the 2235 * dentry is returned. The caller must use dput to free the entry when it has 2236 * finished using it. %NULL is returned if the dentry does not exist. 2237 */ 2238 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name) 2239 { 2240 struct dentry *dentry; 2241 unsigned seq; 2242 2243 do { 2244 seq = read_seqbegin(&rename_lock); 2245 dentry = __d_lookup(parent, name); 2246 if (dentry) 2247 break; 2248 } while (read_seqretry(&rename_lock, seq)); 2249 return dentry; 2250 } 2251 EXPORT_SYMBOL(d_lookup); 2252 2253 /** 2254 * __d_lookup - search for a dentry (racy) 2255 * @parent: parent dentry 2256 * @name: qstr of name we wish to find 2257 * Returns: dentry, or NULL 2258 * 2259 * __d_lookup is like d_lookup, however it may (rarely) return a 2260 * false-negative result due to unrelated rename activity. 2261 * 2262 * __d_lookup is slightly faster by avoiding rename_lock read seqlock, 2263 * however it must be used carefully, eg. with a following d_lookup in 2264 * the case of failure. 2265 * 2266 * __d_lookup callers must be commented. 2267 */ 2268 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name) 2269 { 2270 unsigned int len = name->len; 2271 unsigned int hash = name->hash; 2272 const unsigned char *str = name->name; 2273 struct hlist_bl_head *b = d_hash(parent, hash); 2274 struct hlist_bl_node *node; 2275 struct dentry *found = NULL; 2276 struct dentry *dentry; 2277 2278 /* 2279 * Note: There is significant duplication with __d_lookup_rcu which is 2280 * required to prevent single threaded performance regressions 2281 * especially on architectures where smp_rmb (in seqcounts) are costly. 2282 * Keep the two functions in sync. 2283 */ 2284 2285 /* 2286 * The hash list is protected using RCU. 2287 * 2288 * Take d_lock when comparing a candidate dentry, to avoid races 2289 * with d_move(). 2290 * 2291 * It is possible that concurrent renames can mess up our list 2292 * walk here and result in missing our dentry, resulting in the 2293 * false-negative result. d_lookup() protects against concurrent 2294 * renames using rename_lock seqlock. 2295 * 2296 * See Documentation/filesystems/path-lookup.txt for more details. 2297 */ 2298 rcu_read_lock(); 2299 2300 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) { 2301 2302 if (dentry->d_name.hash != hash) 2303 continue; 2304 2305 spin_lock(&dentry->d_lock); 2306 if (dentry->d_parent != parent) 2307 goto next; 2308 if (d_unhashed(dentry)) 2309 goto next; 2310 2311 /* 2312 * It is safe to compare names since d_move() cannot 2313 * change the qstr (protected by d_lock). 2314 */ 2315 if (parent->d_flags & DCACHE_OP_COMPARE) { 2316 int tlen = dentry->d_name.len; 2317 const char *tname = dentry->d_name.name; 2318 if (parent->d_op->d_compare(parent, dentry, tlen, tname, name)) 2319 goto next; 2320 } else { 2321 if (dentry->d_name.len != len) 2322 goto next; 2323 if (dentry_cmp(dentry, str, len)) 2324 goto next; 2325 } 2326 2327 dentry->d_lockref.count++; 2328 found = dentry; 2329 spin_unlock(&dentry->d_lock); 2330 break; 2331 next: 2332 spin_unlock(&dentry->d_lock); 2333 } 2334 rcu_read_unlock(); 2335 2336 return found; 2337 } 2338 2339 /** 2340 * d_hash_and_lookup - hash the qstr then search for a dentry 2341 * @dir: Directory to search in 2342 * @name: qstr of name we wish to find 2343 * 2344 * On lookup failure NULL is returned; on bad name - ERR_PTR(-error) 2345 */ 2346 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name) 2347 { 2348 /* 2349 * Check for a fs-specific hash function. Note that we must 2350 * calculate the standard hash first, as the d_op->d_hash() 2351 * routine may choose to leave the hash value unchanged. 2352 */ 2353 name->hash = full_name_hash(name->name, name->len); 2354 if (dir->d_flags & DCACHE_OP_HASH) { 2355 int err = dir->d_op->d_hash(dir, name); 2356 if (unlikely(err < 0)) 2357 return ERR_PTR(err); 2358 } 2359 return d_lookup(dir, name); 2360 } 2361 EXPORT_SYMBOL(d_hash_and_lookup); 2362 2363 /* 2364 * When a file is deleted, we have two options: 2365 * - turn this dentry into a negative dentry 2366 * - unhash this dentry and free it. 2367 * 2368 * Usually, we want to just turn this into 2369 * a negative dentry, but if anybody else is 2370 * currently using the dentry or the inode 2371 * we can't do that and we fall back on removing 2372 * it from the hash queues and waiting for 2373 * it to be deleted later when it has no users 2374 */ 2375 2376 /** 2377 * d_delete - delete a dentry 2378 * @dentry: The dentry to delete 2379 * 2380 * Turn the dentry into a negative dentry if possible, otherwise 2381 * remove it from the hash queues so it can be deleted later 2382 */ 2383 2384 void d_delete(struct dentry * dentry) 2385 { 2386 struct inode *inode; 2387 int isdir = 0; 2388 /* 2389 * Are we the only user? 2390 */ 2391 again: 2392 spin_lock(&dentry->d_lock); 2393 inode = dentry->d_inode; 2394 isdir = S_ISDIR(inode->i_mode); 2395 if (dentry->d_lockref.count == 1) { 2396 if (!spin_trylock(&inode->i_lock)) { 2397 spin_unlock(&dentry->d_lock); 2398 cpu_relax(); 2399 goto again; 2400 } 2401 dentry->d_flags &= ~DCACHE_CANT_MOUNT; 2402 dentry_unlink_inode(dentry); 2403 fsnotify_nameremove(dentry, isdir); 2404 return; 2405 } 2406 2407 if (!d_unhashed(dentry)) 2408 __d_drop(dentry); 2409 2410 spin_unlock(&dentry->d_lock); 2411 2412 fsnotify_nameremove(dentry, isdir); 2413 } 2414 EXPORT_SYMBOL(d_delete); 2415 2416 static void __d_rehash(struct dentry * entry, struct hlist_bl_head *b) 2417 { 2418 BUG_ON(!d_unhashed(entry)); 2419 hlist_bl_lock(b); 2420 entry->d_flags |= DCACHE_RCUACCESS; 2421 hlist_bl_add_head_rcu(&entry->d_hash, b); 2422 hlist_bl_unlock(b); 2423 } 2424 2425 static void _d_rehash(struct dentry * entry) 2426 { 2427 __d_rehash(entry, d_hash(entry->d_parent, entry->d_name.hash)); 2428 } 2429 2430 /** 2431 * d_rehash - add an entry back to the hash 2432 * @entry: dentry to add to the hash 2433 * 2434 * Adds a dentry to the hash according to its name. 2435 */ 2436 2437 void d_rehash(struct dentry * entry) 2438 { 2439 spin_lock(&entry->d_lock); 2440 _d_rehash(entry); 2441 spin_unlock(&entry->d_lock); 2442 } 2443 EXPORT_SYMBOL(d_rehash); 2444 2445 /** 2446 * dentry_update_name_case - update case insensitive dentry with a new name 2447 * @dentry: dentry to be updated 2448 * @name: new name 2449 * 2450 * Update a case insensitive dentry with new case of name. 2451 * 2452 * dentry must have been returned by d_lookup with name @name. Old and new 2453 * name lengths must match (ie. no d_compare which allows mismatched name 2454 * lengths). 2455 * 2456 * Parent inode i_mutex must be held over d_lookup and into this call (to 2457 * keep renames and concurrent inserts, and readdir(2) away). 2458 */ 2459 void dentry_update_name_case(struct dentry *dentry, struct qstr *name) 2460 { 2461 BUG_ON(!mutex_is_locked(&dentry->d_parent->d_inode->i_mutex)); 2462 BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */ 2463 2464 spin_lock(&dentry->d_lock); 2465 write_seqcount_begin(&dentry->d_seq); 2466 memcpy((unsigned char *)dentry->d_name.name, name->name, name->len); 2467 write_seqcount_end(&dentry->d_seq); 2468 spin_unlock(&dentry->d_lock); 2469 } 2470 EXPORT_SYMBOL(dentry_update_name_case); 2471 2472 static void swap_names(struct dentry *dentry, struct dentry *target) 2473 { 2474 if (unlikely(dname_external(target))) { 2475 if (unlikely(dname_external(dentry))) { 2476 /* 2477 * Both external: swap the pointers 2478 */ 2479 swap(target->d_name.name, dentry->d_name.name); 2480 } else { 2481 /* 2482 * dentry:internal, target:external. Steal target's 2483 * storage and make target internal. 2484 */ 2485 memcpy(target->d_iname, dentry->d_name.name, 2486 dentry->d_name.len + 1); 2487 dentry->d_name.name = target->d_name.name; 2488 target->d_name.name = target->d_iname; 2489 } 2490 } else { 2491 if (unlikely(dname_external(dentry))) { 2492 /* 2493 * dentry:external, target:internal. Give dentry's 2494 * storage to target and make dentry internal 2495 */ 2496 memcpy(dentry->d_iname, target->d_name.name, 2497 target->d_name.len + 1); 2498 target->d_name.name = dentry->d_name.name; 2499 dentry->d_name.name = dentry->d_iname; 2500 } else { 2501 /* 2502 * Both are internal. 2503 */ 2504 unsigned int i; 2505 BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long))); 2506 kmemcheck_mark_initialized(dentry->d_iname, DNAME_INLINE_LEN); 2507 kmemcheck_mark_initialized(target->d_iname, DNAME_INLINE_LEN); 2508 for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) { 2509 swap(((long *) &dentry->d_iname)[i], 2510 ((long *) &target->d_iname)[i]); 2511 } 2512 } 2513 } 2514 swap(dentry->d_name.hash_len, target->d_name.hash_len); 2515 } 2516 2517 static void copy_name(struct dentry *dentry, struct dentry *target) 2518 { 2519 struct external_name *old_name = NULL; 2520 if (unlikely(dname_external(dentry))) 2521 old_name = external_name(dentry); 2522 if (unlikely(dname_external(target))) { 2523 atomic_inc(&external_name(target)->u.count); 2524 dentry->d_name = target->d_name; 2525 } else { 2526 memcpy(dentry->d_iname, target->d_name.name, 2527 target->d_name.len + 1); 2528 dentry->d_name.name = dentry->d_iname; 2529 dentry->d_name.hash_len = target->d_name.hash_len; 2530 } 2531 if (old_name && likely(atomic_dec_and_test(&old_name->u.count))) 2532 kfree_rcu(old_name, u.head); 2533 } 2534 2535 static void dentry_lock_for_move(struct dentry *dentry, struct dentry *target) 2536 { 2537 /* 2538 * XXXX: do we really need to take target->d_lock? 2539 */ 2540 if (IS_ROOT(dentry) || dentry->d_parent == target->d_parent) 2541 spin_lock(&target->d_parent->d_lock); 2542 else { 2543 if (d_ancestor(dentry->d_parent, target->d_parent)) { 2544 spin_lock(&dentry->d_parent->d_lock); 2545 spin_lock_nested(&target->d_parent->d_lock, 2546 DENTRY_D_LOCK_NESTED); 2547 } else { 2548 spin_lock(&target->d_parent->d_lock); 2549 spin_lock_nested(&dentry->d_parent->d_lock, 2550 DENTRY_D_LOCK_NESTED); 2551 } 2552 } 2553 if (target < dentry) { 2554 spin_lock_nested(&target->d_lock, 2); 2555 spin_lock_nested(&dentry->d_lock, 3); 2556 } else { 2557 spin_lock_nested(&dentry->d_lock, 2); 2558 spin_lock_nested(&target->d_lock, 3); 2559 } 2560 } 2561 2562 static void dentry_unlock_for_move(struct dentry *dentry, struct dentry *target) 2563 { 2564 if (target->d_parent != dentry->d_parent) 2565 spin_unlock(&dentry->d_parent->d_lock); 2566 if (target->d_parent != target) 2567 spin_unlock(&target->d_parent->d_lock); 2568 spin_unlock(&target->d_lock); 2569 spin_unlock(&dentry->d_lock); 2570 } 2571 2572 /* 2573 * When switching names, the actual string doesn't strictly have to 2574 * be preserved in the target - because we're dropping the target 2575 * anyway. As such, we can just do a simple memcpy() to copy over 2576 * the new name before we switch, unless we are going to rehash 2577 * it. Note that if we *do* unhash the target, we are not allowed 2578 * to rehash it without giving it a new name/hash key - whether 2579 * we swap or overwrite the names here, resulting name won't match 2580 * the reality in filesystem; it's only there for d_path() purposes. 2581 * Note that all of this is happening under rename_lock, so the 2582 * any hash lookup seeing it in the middle of manipulations will 2583 * be discarded anyway. So we do not care what happens to the hash 2584 * key in that case. 2585 */ 2586 /* 2587 * __d_move - move a dentry 2588 * @dentry: entry to move 2589 * @target: new dentry 2590 * @exchange: exchange the two dentries 2591 * 2592 * Update the dcache to reflect the move of a file name. Negative 2593 * dcache entries should not be moved in this way. Caller must hold 2594 * rename_lock, the i_mutex of the source and target directories, 2595 * and the sb->s_vfs_rename_mutex if they differ. See lock_rename(). 2596 */ 2597 static void __d_move(struct dentry *dentry, struct dentry *target, 2598 bool exchange) 2599 { 2600 if (!dentry->d_inode) 2601 printk(KERN_WARNING "VFS: moving negative dcache entry\n"); 2602 2603 BUG_ON(d_ancestor(dentry, target)); 2604 BUG_ON(d_ancestor(target, dentry)); 2605 2606 dentry_lock_for_move(dentry, target); 2607 2608 write_seqcount_begin(&dentry->d_seq); 2609 write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED); 2610 2611 /* __d_drop does write_seqcount_barrier, but they're OK to nest. */ 2612 2613 /* 2614 * Move the dentry to the target hash queue. Don't bother checking 2615 * for the same hash queue because of how unlikely it is. 2616 */ 2617 __d_drop(dentry); 2618 __d_rehash(dentry, d_hash(target->d_parent, target->d_name.hash)); 2619 2620 /* 2621 * Unhash the target (d_delete() is not usable here). If exchanging 2622 * the two dentries, then rehash onto the other's hash queue. 2623 */ 2624 __d_drop(target); 2625 if (exchange) { 2626 __d_rehash(target, 2627 d_hash(dentry->d_parent, dentry->d_name.hash)); 2628 } 2629 2630 /* Switch the names.. */ 2631 if (exchange) 2632 swap_names(dentry, target); 2633 else 2634 copy_name(dentry, target); 2635 2636 /* ... and switch them in the tree */ 2637 if (IS_ROOT(dentry)) { 2638 /* splicing a tree */ 2639 dentry->d_parent = target->d_parent; 2640 target->d_parent = target; 2641 list_del_init(&target->d_child); 2642 list_move(&dentry->d_child, &dentry->d_parent->d_subdirs); 2643 } else { 2644 /* swapping two dentries */ 2645 swap(dentry->d_parent, target->d_parent); 2646 list_move(&target->d_child, &target->d_parent->d_subdirs); 2647 list_move(&dentry->d_child, &dentry->d_parent->d_subdirs); 2648 if (exchange) 2649 fsnotify_d_move(target); 2650 fsnotify_d_move(dentry); 2651 } 2652 2653 write_seqcount_end(&target->d_seq); 2654 write_seqcount_end(&dentry->d_seq); 2655 2656 dentry_unlock_for_move(dentry, target); 2657 } 2658 2659 /* 2660 * d_move - move a dentry 2661 * @dentry: entry to move 2662 * @target: new dentry 2663 * 2664 * Update the dcache to reflect the move of a file name. Negative 2665 * dcache entries should not be moved in this way. See the locking 2666 * requirements for __d_move. 2667 */ 2668 void d_move(struct dentry *dentry, struct dentry *target) 2669 { 2670 write_seqlock(&rename_lock); 2671 __d_move(dentry, target, false); 2672 write_sequnlock(&rename_lock); 2673 } 2674 EXPORT_SYMBOL(d_move); 2675 2676 /* 2677 * d_exchange - exchange two dentries 2678 * @dentry1: first dentry 2679 * @dentry2: second dentry 2680 */ 2681 void d_exchange(struct dentry *dentry1, struct dentry *dentry2) 2682 { 2683 write_seqlock(&rename_lock); 2684 2685 WARN_ON(!dentry1->d_inode); 2686 WARN_ON(!dentry2->d_inode); 2687 WARN_ON(IS_ROOT(dentry1)); 2688 WARN_ON(IS_ROOT(dentry2)); 2689 2690 __d_move(dentry1, dentry2, true); 2691 2692 write_sequnlock(&rename_lock); 2693 } 2694 2695 /** 2696 * d_ancestor - search for an ancestor 2697 * @p1: ancestor dentry 2698 * @p2: child dentry 2699 * 2700 * Returns the ancestor dentry of p2 which is a child of p1, if p1 is 2701 * an ancestor of p2, else NULL. 2702 */ 2703 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2) 2704 { 2705 struct dentry *p; 2706 2707 for (p = p2; !IS_ROOT(p); p = p->d_parent) { 2708 if (p->d_parent == p1) 2709 return p; 2710 } 2711 return NULL; 2712 } 2713 2714 /* 2715 * This helper attempts to cope with remotely renamed directories 2716 * 2717 * It assumes that the caller is already holding 2718 * dentry->d_parent->d_inode->i_mutex, inode->i_lock and rename_lock 2719 * 2720 * Note: If ever the locking in lock_rename() changes, then please 2721 * remember to update this too... 2722 */ 2723 static int __d_unalias(struct inode *inode, 2724 struct dentry *dentry, struct dentry *alias) 2725 { 2726 struct mutex *m1 = NULL, *m2 = NULL; 2727 int ret = -ESTALE; 2728 2729 /* If alias and dentry share a parent, then no extra locks required */ 2730 if (alias->d_parent == dentry->d_parent) 2731 goto out_unalias; 2732 2733 /* See lock_rename() */ 2734 if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex)) 2735 goto out_err; 2736 m1 = &dentry->d_sb->s_vfs_rename_mutex; 2737 if (!mutex_trylock(&alias->d_parent->d_inode->i_mutex)) 2738 goto out_err; 2739 m2 = &alias->d_parent->d_inode->i_mutex; 2740 out_unalias: 2741 __d_move(alias, dentry, false); 2742 ret = 0; 2743 out_err: 2744 spin_unlock(&inode->i_lock); 2745 if (m2) 2746 mutex_unlock(m2); 2747 if (m1) 2748 mutex_unlock(m1); 2749 return ret; 2750 } 2751 2752 /** 2753 * d_splice_alias - splice a disconnected dentry into the tree if one exists 2754 * @inode: the inode which may have a disconnected dentry 2755 * @dentry: a negative dentry which we want to point to the inode. 2756 * 2757 * If inode is a directory and has an IS_ROOT alias, then d_move that in 2758 * place of the given dentry and return it, else simply d_add the inode 2759 * to the dentry and return NULL. 2760 * 2761 * If a non-IS_ROOT directory is found, the filesystem is corrupt, and 2762 * we should error out: directories can't have multiple aliases. 2763 * 2764 * This is needed in the lookup routine of any filesystem that is exportable 2765 * (via knfsd) so that we can build dcache paths to directories effectively. 2766 * 2767 * If a dentry was found and moved, then it is returned. Otherwise NULL 2768 * is returned. This matches the expected return value of ->lookup. 2769 * 2770 * Cluster filesystems may call this function with a negative, hashed dentry. 2771 * In that case, we know that the inode will be a regular file, and also this 2772 * will only occur during atomic_open. So we need to check for the dentry 2773 * being already hashed only in the final case. 2774 */ 2775 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry) 2776 { 2777 if (IS_ERR(inode)) 2778 return ERR_CAST(inode); 2779 2780 BUG_ON(!d_unhashed(dentry)); 2781 2782 if (!inode) { 2783 __d_instantiate(dentry, NULL); 2784 goto out; 2785 } 2786 spin_lock(&inode->i_lock); 2787 if (S_ISDIR(inode->i_mode)) { 2788 struct dentry *new = __d_find_any_alias(inode); 2789 if (unlikely(new)) { 2790 write_seqlock(&rename_lock); 2791 if (unlikely(d_ancestor(new, dentry))) { 2792 write_sequnlock(&rename_lock); 2793 spin_unlock(&inode->i_lock); 2794 dput(new); 2795 new = ERR_PTR(-ELOOP); 2796 pr_warn_ratelimited( 2797 "VFS: Lookup of '%s' in %s %s" 2798 " would have caused loop\n", 2799 dentry->d_name.name, 2800 inode->i_sb->s_type->name, 2801 inode->i_sb->s_id); 2802 } else if (!IS_ROOT(new)) { 2803 int err = __d_unalias(inode, dentry, new); 2804 write_sequnlock(&rename_lock); 2805 if (err) { 2806 dput(new); 2807 new = ERR_PTR(err); 2808 } 2809 } else { 2810 __d_move(new, dentry, false); 2811 write_sequnlock(&rename_lock); 2812 spin_unlock(&inode->i_lock); 2813 security_d_instantiate(new, inode); 2814 } 2815 iput(inode); 2816 return new; 2817 } 2818 } 2819 /* already taking inode->i_lock, so d_add() by hand */ 2820 __d_instantiate(dentry, inode); 2821 spin_unlock(&inode->i_lock); 2822 out: 2823 security_d_instantiate(dentry, inode); 2824 d_rehash(dentry); 2825 return NULL; 2826 } 2827 EXPORT_SYMBOL(d_splice_alias); 2828 2829 static int prepend(char **buffer, int *buflen, const char *str, int namelen) 2830 { 2831 *buflen -= namelen; 2832 if (*buflen < 0) 2833 return -ENAMETOOLONG; 2834 *buffer -= namelen; 2835 memcpy(*buffer, str, namelen); 2836 return 0; 2837 } 2838 2839 /** 2840 * prepend_name - prepend a pathname in front of current buffer pointer 2841 * @buffer: buffer pointer 2842 * @buflen: allocated length of the buffer 2843 * @name: name string and length qstr structure 2844 * 2845 * With RCU path tracing, it may race with d_move(). Use ACCESS_ONCE() to 2846 * make sure that either the old or the new name pointer and length are 2847 * fetched. However, there may be mismatch between length and pointer. 2848 * The length cannot be trusted, we need to copy it byte-by-byte until 2849 * the length is reached or a null byte is found. It also prepends "/" at 2850 * the beginning of the name. The sequence number check at the caller will 2851 * retry it again when a d_move() does happen. So any garbage in the buffer 2852 * due to mismatched pointer and length will be discarded. 2853 * 2854 * Data dependency barrier is needed to make sure that we see that terminating 2855 * NUL. Alpha strikes again, film at 11... 2856 */ 2857 static int prepend_name(char **buffer, int *buflen, struct qstr *name) 2858 { 2859 const char *dname = ACCESS_ONCE(name->name); 2860 u32 dlen = ACCESS_ONCE(name->len); 2861 char *p; 2862 2863 smp_read_barrier_depends(); 2864 2865 *buflen -= dlen + 1; 2866 if (*buflen < 0) 2867 return -ENAMETOOLONG; 2868 p = *buffer -= dlen + 1; 2869 *p++ = '/'; 2870 while (dlen--) { 2871 char c = *dname++; 2872 if (!c) 2873 break; 2874 *p++ = c; 2875 } 2876 return 0; 2877 } 2878 2879 /** 2880 * prepend_path - Prepend path string to a buffer 2881 * @path: the dentry/vfsmount to report 2882 * @root: root vfsmnt/dentry 2883 * @buffer: pointer to the end of the buffer 2884 * @buflen: pointer to buffer length 2885 * 2886 * The function will first try to write out the pathname without taking any 2887 * lock other than the RCU read lock to make sure that dentries won't go away. 2888 * It only checks the sequence number of the global rename_lock as any change 2889 * in the dentry's d_seq will be preceded by changes in the rename_lock 2890 * sequence number. If the sequence number had been changed, it will restart 2891 * the whole pathname back-tracing sequence again by taking the rename_lock. 2892 * In this case, there is no need to take the RCU read lock as the recursive 2893 * parent pointer references will keep the dentry chain alive as long as no 2894 * rename operation is performed. 2895 */ 2896 static int prepend_path(const struct path *path, 2897 const struct path *root, 2898 char **buffer, int *buflen) 2899 { 2900 struct dentry *dentry; 2901 struct vfsmount *vfsmnt; 2902 struct mount *mnt; 2903 int error = 0; 2904 unsigned seq, m_seq = 0; 2905 char *bptr; 2906 int blen; 2907 2908 rcu_read_lock(); 2909 restart_mnt: 2910 read_seqbegin_or_lock(&mount_lock, &m_seq); 2911 seq = 0; 2912 rcu_read_lock(); 2913 restart: 2914 bptr = *buffer; 2915 blen = *buflen; 2916 error = 0; 2917 dentry = path->dentry; 2918 vfsmnt = path->mnt; 2919 mnt = real_mount(vfsmnt); 2920 read_seqbegin_or_lock(&rename_lock, &seq); 2921 while (dentry != root->dentry || vfsmnt != root->mnt) { 2922 struct dentry * parent; 2923 2924 if (dentry == vfsmnt->mnt_root || IS_ROOT(dentry)) { 2925 struct mount *parent = ACCESS_ONCE(mnt->mnt_parent); 2926 /* Global root? */ 2927 if (mnt != parent) { 2928 dentry = ACCESS_ONCE(mnt->mnt_mountpoint); 2929 mnt = parent; 2930 vfsmnt = &mnt->mnt; 2931 continue; 2932 } 2933 if (!error) 2934 error = is_mounted(vfsmnt) ? 1 : 2; 2935 break; 2936 } 2937 parent = dentry->d_parent; 2938 prefetch(parent); 2939 error = prepend_name(&bptr, &blen, &dentry->d_name); 2940 if (error) 2941 break; 2942 2943 dentry = parent; 2944 } 2945 if (!(seq & 1)) 2946 rcu_read_unlock(); 2947 if (need_seqretry(&rename_lock, seq)) { 2948 seq = 1; 2949 goto restart; 2950 } 2951 done_seqretry(&rename_lock, seq); 2952 2953 if (!(m_seq & 1)) 2954 rcu_read_unlock(); 2955 if (need_seqretry(&mount_lock, m_seq)) { 2956 m_seq = 1; 2957 goto restart_mnt; 2958 } 2959 done_seqretry(&mount_lock, m_seq); 2960 2961 if (error >= 0 && bptr == *buffer) { 2962 if (--blen < 0) 2963 error = -ENAMETOOLONG; 2964 else 2965 *--bptr = '/'; 2966 } 2967 *buffer = bptr; 2968 *buflen = blen; 2969 return error; 2970 } 2971 2972 /** 2973 * __d_path - return the path of a dentry 2974 * @path: the dentry/vfsmount to report 2975 * @root: root vfsmnt/dentry 2976 * @buf: buffer to return value in 2977 * @buflen: buffer length 2978 * 2979 * Convert a dentry into an ASCII path name. 2980 * 2981 * Returns a pointer into the buffer or an error code if the 2982 * path was too long. 2983 * 2984 * "buflen" should be positive. 2985 * 2986 * If the path is not reachable from the supplied root, return %NULL. 2987 */ 2988 char *__d_path(const struct path *path, 2989 const struct path *root, 2990 char *buf, int buflen) 2991 { 2992 char *res = buf + buflen; 2993 int error; 2994 2995 prepend(&res, &buflen, "\0", 1); 2996 error = prepend_path(path, root, &res, &buflen); 2997 2998 if (error < 0) 2999 return ERR_PTR(error); 3000 if (error > 0) 3001 return NULL; 3002 return res; 3003 } 3004 3005 char *d_absolute_path(const struct path *path, 3006 char *buf, int buflen) 3007 { 3008 struct path root = {}; 3009 char *res = buf + buflen; 3010 int error; 3011 3012 prepend(&res, &buflen, "\0", 1); 3013 error = prepend_path(path, &root, &res, &buflen); 3014 3015 if (error > 1) 3016 error = -EINVAL; 3017 if (error < 0) 3018 return ERR_PTR(error); 3019 return res; 3020 } 3021 3022 /* 3023 * same as __d_path but appends "(deleted)" for unlinked files. 3024 */ 3025 static int path_with_deleted(const struct path *path, 3026 const struct path *root, 3027 char **buf, int *buflen) 3028 { 3029 prepend(buf, buflen, "\0", 1); 3030 if (d_unlinked(path->dentry)) { 3031 int error = prepend(buf, buflen, " (deleted)", 10); 3032 if (error) 3033 return error; 3034 } 3035 3036 return prepend_path(path, root, buf, buflen); 3037 } 3038 3039 static int prepend_unreachable(char **buffer, int *buflen) 3040 { 3041 return prepend(buffer, buflen, "(unreachable)", 13); 3042 } 3043 3044 static void get_fs_root_rcu(struct fs_struct *fs, struct path *root) 3045 { 3046 unsigned seq; 3047 3048 do { 3049 seq = read_seqcount_begin(&fs->seq); 3050 *root = fs->root; 3051 } while (read_seqcount_retry(&fs->seq, seq)); 3052 } 3053 3054 /** 3055 * d_path - return the path of a dentry 3056 * @path: path to report 3057 * @buf: buffer to return value in 3058 * @buflen: buffer length 3059 * 3060 * Convert a dentry into an ASCII path name. If the entry has been deleted 3061 * the string " (deleted)" is appended. Note that this is ambiguous. 3062 * 3063 * Returns a pointer into the buffer or an error code if the path was 3064 * too long. Note: Callers should use the returned pointer, not the passed 3065 * in buffer, to use the name! The implementation often starts at an offset 3066 * into the buffer, and may leave 0 bytes at the start. 3067 * 3068 * "buflen" should be positive. 3069 */ 3070 char *d_path(const struct path *path, char *buf, int buflen) 3071 { 3072 char *res = buf + buflen; 3073 struct path root; 3074 int error; 3075 3076 /* 3077 * We have various synthetic filesystems that never get mounted. On 3078 * these filesystems dentries are never used for lookup purposes, and 3079 * thus don't need to be hashed. They also don't need a name until a 3080 * user wants to identify the object in /proc/pid/fd/. The little hack 3081 * below allows us to generate a name for these objects on demand: 3082 * 3083 * Some pseudo inodes are mountable. When they are mounted 3084 * path->dentry == path->mnt->mnt_root. In that case don't call d_dname 3085 * and instead have d_path return the mounted path. 3086 */ 3087 if (path->dentry->d_op && path->dentry->d_op->d_dname && 3088 (!IS_ROOT(path->dentry) || path->dentry != path->mnt->mnt_root)) 3089 return path->dentry->d_op->d_dname(path->dentry, buf, buflen); 3090 3091 rcu_read_lock(); 3092 get_fs_root_rcu(current->fs, &root); 3093 error = path_with_deleted(path, &root, &res, &buflen); 3094 rcu_read_unlock(); 3095 3096 if (error < 0) 3097 res = ERR_PTR(error); 3098 return res; 3099 } 3100 EXPORT_SYMBOL(d_path); 3101 3102 /* 3103 * Helper function for dentry_operations.d_dname() members 3104 */ 3105 char *dynamic_dname(struct dentry *dentry, char *buffer, int buflen, 3106 const char *fmt, ...) 3107 { 3108 va_list args; 3109 char temp[64]; 3110 int sz; 3111 3112 va_start(args, fmt); 3113 sz = vsnprintf(temp, sizeof(temp), fmt, args) + 1; 3114 va_end(args); 3115 3116 if (sz > sizeof(temp) || sz > buflen) 3117 return ERR_PTR(-ENAMETOOLONG); 3118 3119 buffer += buflen - sz; 3120 return memcpy(buffer, temp, sz); 3121 } 3122 3123 char *simple_dname(struct dentry *dentry, char *buffer, int buflen) 3124 { 3125 char *end = buffer + buflen; 3126 /* these dentries are never renamed, so d_lock is not needed */ 3127 if (prepend(&end, &buflen, " (deleted)", 11) || 3128 prepend(&end, &buflen, dentry->d_name.name, dentry->d_name.len) || 3129 prepend(&end, &buflen, "/", 1)) 3130 end = ERR_PTR(-ENAMETOOLONG); 3131 return end; 3132 } 3133 EXPORT_SYMBOL(simple_dname); 3134 3135 /* 3136 * Write full pathname from the root of the filesystem into the buffer. 3137 */ 3138 static char *__dentry_path(struct dentry *d, char *buf, int buflen) 3139 { 3140 struct dentry *dentry; 3141 char *end, *retval; 3142 int len, seq = 0; 3143 int error = 0; 3144 3145 if (buflen < 2) 3146 goto Elong; 3147 3148 rcu_read_lock(); 3149 restart: 3150 dentry = d; 3151 end = buf + buflen; 3152 len = buflen; 3153 prepend(&end, &len, "\0", 1); 3154 /* Get '/' right */ 3155 retval = end-1; 3156 *retval = '/'; 3157 read_seqbegin_or_lock(&rename_lock, &seq); 3158 while (!IS_ROOT(dentry)) { 3159 struct dentry *parent = dentry->d_parent; 3160 3161 prefetch(parent); 3162 error = prepend_name(&end, &len, &dentry->d_name); 3163 if (error) 3164 break; 3165 3166 retval = end; 3167 dentry = parent; 3168 } 3169 if (!(seq & 1)) 3170 rcu_read_unlock(); 3171 if (need_seqretry(&rename_lock, seq)) { 3172 seq = 1; 3173 goto restart; 3174 } 3175 done_seqretry(&rename_lock, seq); 3176 if (error) 3177 goto Elong; 3178 return retval; 3179 Elong: 3180 return ERR_PTR(-ENAMETOOLONG); 3181 } 3182 3183 char *dentry_path_raw(struct dentry *dentry, char *buf, int buflen) 3184 { 3185 return __dentry_path(dentry, buf, buflen); 3186 } 3187 EXPORT_SYMBOL(dentry_path_raw); 3188 3189 char *dentry_path(struct dentry *dentry, char *buf, int buflen) 3190 { 3191 char *p = NULL; 3192 char *retval; 3193 3194 if (d_unlinked(dentry)) { 3195 p = buf + buflen; 3196 if (prepend(&p, &buflen, "//deleted", 10) != 0) 3197 goto Elong; 3198 buflen++; 3199 } 3200 retval = __dentry_path(dentry, buf, buflen); 3201 if (!IS_ERR(retval) && p) 3202 *p = '/'; /* restore '/' overriden with '\0' */ 3203 return retval; 3204 Elong: 3205 return ERR_PTR(-ENAMETOOLONG); 3206 } 3207 3208 static void get_fs_root_and_pwd_rcu(struct fs_struct *fs, struct path *root, 3209 struct path *pwd) 3210 { 3211 unsigned seq; 3212 3213 do { 3214 seq = read_seqcount_begin(&fs->seq); 3215 *root = fs->root; 3216 *pwd = fs->pwd; 3217 } while (read_seqcount_retry(&fs->seq, seq)); 3218 } 3219 3220 /* 3221 * NOTE! The user-level library version returns a 3222 * character pointer. The kernel system call just 3223 * returns the length of the buffer filled (which 3224 * includes the ending '\0' character), or a negative 3225 * error value. So libc would do something like 3226 * 3227 * char *getcwd(char * buf, size_t size) 3228 * { 3229 * int retval; 3230 * 3231 * retval = sys_getcwd(buf, size); 3232 * if (retval >= 0) 3233 * return buf; 3234 * errno = -retval; 3235 * return NULL; 3236 * } 3237 */ 3238 SYSCALL_DEFINE2(getcwd, char __user *, buf, unsigned long, size) 3239 { 3240 int error; 3241 struct path pwd, root; 3242 char *page = __getname(); 3243 3244 if (!page) 3245 return -ENOMEM; 3246 3247 rcu_read_lock(); 3248 get_fs_root_and_pwd_rcu(current->fs, &root, &pwd); 3249 3250 error = -ENOENT; 3251 if (!d_unlinked(pwd.dentry)) { 3252 unsigned long len; 3253 char *cwd = page + PATH_MAX; 3254 int buflen = PATH_MAX; 3255 3256 prepend(&cwd, &buflen, "\0", 1); 3257 error = prepend_path(&pwd, &root, &cwd, &buflen); 3258 rcu_read_unlock(); 3259 3260 if (error < 0) 3261 goto out; 3262 3263 /* Unreachable from current root */ 3264 if (error > 0) { 3265 error = prepend_unreachable(&cwd, &buflen); 3266 if (error) 3267 goto out; 3268 } 3269 3270 error = -ERANGE; 3271 len = PATH_MAX + page - cwd; 3272 if (len <= size) { 3273 error = len; 3274 if (copy_to_user(buf, cwd, len)) 3275 error = -EFAULT; 3276 } 3277 } else { 3278 rcu_read_unlock(); 3279 } 3280 3281 out: 3282 __putname(page); 3283 return error; 3284 } 3285 3286 /* 3287 * Test whether new_dentry is a subdirectory of old_dentry. 3288 * 3289 * Trivially implemented using the dcache structure 3290 */ 3291 3292 /** 3293 * is_subdir - is new dentry a subdirectory of old_dentry 3294 * @new_dentry: new dentry 3295 * @old_dentry: old dentry 3296 * 3297 * Returns 1 if new_dentry is a subdirectory of the parent (at any depth). 3298 * Returns 0 otherwise. 3299 * Caller must ensure that "new_dentry" is pinned before calling is_subdir() 3300 */ 3301 3302 int is_subdir(struct dentry *new_dentry, struct dentry *old_dentry) 3303 { 3304 int result; 3305 unsigned seq; 3306 3307 if (new_dentry == old_dentry) 3308 return 1; 3309 3310 do { 3311 /* for restarting inner loop in case of seq retry */ 3312 seq = read_seqbegin(&rename_lock); 3313 /* 3314 * Need rcu_readlock to protect against the d_parent trashing 3315 * due to d_move 3316 */ 3317 rcu_read_lock(); 3318 if (d_ancestor(old_dentry, new_dentry)) 3319 result = 1; 3320 else 3321 result = 0; 3322 rcu_read_unlock(); 3323 } while (read_seqretry(&rename_lock, seq)); 3324 3325 return result; 3326 } 3327 3328 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry) 3329 { 3330 struct dentry *root = data; 3331 if (dentry != root) { 3332 if (d_unhashed(dentry) || !dentry->d_inode) 3333 return D_WALK_SKIP; 3334 3335 if (!(dentry->d_flags & DCACHE_GENOCIDE)) { 3336 dentry->d_flags |= DCACHE_GENOCIDE; 3337 dentry->d_lockref.count--; 3338 } 3339 } 3340 return D_WALK_CONTINUE; 3341 } 3342 3343 void d_genocide(struct dentry *parent) 3344 { 3345 d_walk(parent, parent, d_genocide_kill, NULL); 3346 } 3347 3348 void d_tmpfile(struct dentry *dentry, struct inode *inode) 3349 { 3350 inode_dec_link_count(inode); 3351 BUG_ON(dentry->d_name.name != dentry->d_iname || 3352 !hlist_unhashed(&dentry->d_u.d_alias) || 3353 !d_unlinked(dentry)); 3354 spin_lock(&dentry->d_parent->d_lock); 3355 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED); 3356 dentry->d_name.len = sprintf(dentry->d_iname, "#%llu", 3357 (unsigned long long)inode->i_ino); 3358 spin_unlock(&dentry->d_lock); 3359 spin_unlock(&dentry->d_parent->d_lock); 3360 d_instantiate(dentry, inode); 3361 } 3362 EXPORT_SYMBOL(d_tmpfile); 3363 3364 static __initdata unsigned long dhash_entries; 3365 static int __init set_dhash_entries(char *str) 3366 { 3367 if (!str) 3368 return 0; 3369 dhash_entries = simple_strtoul(str, &str, 0); 3370 return 1; 3371 } 3372 __setup("dhash_entries=", set_dhash_entries); 3373 3374 static void __init dcache_init_early(void) 3375 { 3376 unsigned int loop; 3377 3378 /* If hashes are distributed across NUMA nodes, defer 3379 * hash allocation until vmalloc space is available. 3380 */ 3381 if (hashdist) 3382 return; 3383 3384 dentry_hashtable = 3385 alloc_large_system_hash("Dentry cache", 3386 sizeof(struct hlist_bl_head), 3387 dhash_entries, 3388 13, 3389 HASH_EARLY, 3390 &d_hash_shift, 3391 &d_hash_mask, 3392 0, 3393 0); 3394 3395 for (loop = 0; loop < (1U << d_hash_shift); loop++) 3396 INIT_HLIST_BL_HEAD(dentry_hashtable + loop); 3397 } 3398 3399 static void __init dcache_init(void) 3400 { 3401 unsigned int loop; 3402 3403 /* 3404 * A constructor could be added for stable state like the lists, 3405 * but it is probably not worth it because of the cache nature 3406 * of the dcache. 3407 */ 3408 dentry_cache = KMEM_CACHE(dentry, 3409 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD); 3410 3411 /* Hash may have been set up in dcache_init_early */ 3412 if (!hashdist) 3413 return; 3414 3415 dentry_hashtable = 3416 alloc_large_system_hash("Dentry cache", 3417 sizeof(struct hlist_bl_head), 3418 dhash_entries, 3419 13, 3420 0, 3421 &d_hash_shift, 3422 &d_hash_mask, 3423 0, 3424 0); 3425 3426 for (loop = 0; loop < (1U << d_hash_shift); loop++) 3427 INIT_HLIST_BL_HEAD(dentry_hashtable + loop); 3428 } 3429 3430 /* SLAB cache for __getname() consumers */ 3431 struct kmem_cache *names_cachep __read_mostly; 3432 EXPORT_SYMBOL(names_cachep); 3433 3434 EXPORT_SYMBOL(d_genocide); 3435 3436 void __init vfs_caches_init_early(void) 3437 { 3438 dcache_init_early(); 3439 inode_init_early(); 3440 } 3441 3442 void __init vfs_caches_init(unsigned long mempages) 3443 { 3444 unsigned long reserve; 3445 3446 /* Base hash sizes on available memory, with a reserve equal to 3447 150% of current kernel size */ 3448 3449 reserve = min((mempages - nr_free_pages()) * 3/2, mempages - 1); 3450 mempages -= reserve; 3451 3452 names_cachep = kmem_cache_create("names_cache", PATH_MAX, 0, 3453 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL); 3454 3455 dcache_init(); 3456 inode_init(); 3457 files_init(mempages); 3458 mnt_init(); 3459 bdev_cache_init(); 3460 chrdev_init(); 3461 } 3462