4 * Complete reimplementation
5 * (C) 1997 Thomas Schoebel-Theuer,
6 * with heavy changes by Linus Torvalds
10 * Notes on the allocation strategy:
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.
17 #include <linux/ratelimit.h>
18 #include <linux/string.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/security.h>
28 #include <linux/seqlock.h>
29 #include <linux/bootmem.h>
30 #include <linux/bit_spinlock.h>
31 #include <linux/rculist_bl.h>
32 #include <linux/list_lru.h>
38 * dcache->d_inode->i_lock protects:
39 * - i_dentry, d_u.d_alias, d_inode of aliases
40 * dcache_hash_bucket lock protects:
41 * - the dcache hash table
42 * s_roots bl list spinlock protects:
43 * - the s_roots list (see __d_drop)
44 * dentry->d_sb->s_dentry_lru_lock protects:
45 * - the dcache lru lists and counters
52 * - d_parent and d_subdirs
53 * - childrens' d_child and d_parent
54 * - d_u.d_alias, d_inode
57 * dentry->d_inode->i_lock
59 * dentry->d_sb->s_dentry_lru_lock
60 * dcache_hash_bucket lock
63 * If there is an ancestor relationship:
64 * dentry->d_parent->...->d_parent->d_lock
66 * dentry->d_parent->d_lock
69 * If no ancestor relationship:
70 * arbitrary, since it's serialized on rename_lock
72 int sysctl_vfs_cache_pressure __read_mostly = 100;
73 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
75 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
77 EXPORT_SYMBOL(rename_lock);
79 static struct kmem_cache *dentry_cache __read_mostly;
81 const struct qstr empty_name = QSTR_INIT("", 0);
82 EXPORT_SYMBOL(empty_name);
83 const struct qstr slash_name = QSTR_INIT("/", 1);
84 EXPORT_SYMBOL(slash_name);
87 * This is the single most critical data structure when it comes
88 * to the dcache: the hashtable for lookups. Somebody should try
89 * to make this good - I've just made it work.
91 * This hash-function tries to avoid losing too many bits of hash
92 * information, yet avoid using a prime hash-size or similar.
95 static unsigned int d_hash_shift __read_mostly;
97 static struct hlist_bl_head *dentry_hashtable __read_mostly;
99 static inline struct hlist_bl_head *d_hash(unsigned int hash)
101 return dentry_hashtable + (hash >> d_hash_shift);
104 #define IN_LOOKUP_SHIFT 10
105 static struct hlist_bl_head in_lookup_hashtable[1 << IN_LOOKUP_SHIFT];
107 static inline struct hlist_bl_head *in_lookup_hash(const struct dentry *parent,
110 hash += (unsigned long) parent / L1_CACHE_BYTES;
111 return in_lookup_hashtable + hash_32(hash, IN_LOOKUP_SHIFT);
115 /* Statistics gathering. */
116 struct dentry_stat_t dentry_stat = {
120 static DEFINE_PER_CPU(long, nr_dentry);
121 static DEFINE_PER_CPU(long, nr_dentry_unused);
123 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
126 * Here we resort to our own counters instead of using generic per-cpu counters
127 * for consistency with what the vfs inode code does. We are expected to harvest
128 * better code and performance by having our own specialized counters.
130 * Please note that the loop is done over all possible CPUs, not over all online
131 * CPUs. The reason for this is that we don't want to play games with CPUs going
132 * on and off. If one of them goes off, we will just keep their counters.
134 * glommer: See cffbc8a for details, and if you ever intend to change this,
135 * please update all vfs counters to match.
137 static long get_nr_dentry(void)
141 for_each_possible_cpu(i)
142 sum += per_cpu(nr_dentry, i);
143 return sum < 0 ? 0 : sum;
146 static long get_nr_dentry_unused(void)
150 for_each_possible_cpu(i)
151 sum += per_cpu(nr_dentry_unused, i);
152 return sum < 0 ? 0 : sum;
155 int proc_nr_dentry(struct ctl_table *table, int write, void __user *buffer,
156 size_t *lenp, loff_t *ppos)
158 dentry_stat.nr_dentry = get_nr_dentry();
159 dentry_stat.nr_unused = get_nr_dentry_unused();
160 return proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
165 * Compare 2 name strings, return 0 if they match, otherwise non-zero.
166 * The strings are both count bytes long, and count is non-zero.
168 #ifdef CONFIG_DCACHE_WORD_ACCESS
170 #include <asm/word-at-a-time.h>
172 * NOTE! 'cs' and 'scount' come from a dentry, so it has a
173 * aligned allocation for this particular component. We don't
174 * strictly need the load_unaligned_zeropad() safety, but it
175 * doesn't hurt either.
177 * In contrast, 'ct' and 'tcount' can be from a pathname, and do
178 * need the careful unaligned handling.
180 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
182 unsigned long a,b,mask;
185 a = read_word_at_a_time(cs);
186 b = load_unaligned_zeropad(ct);
187 if (tcount < sizeof(unsigned long))
189 if (unlikely(a != b))
191 cs += sizeof(unsigned long);
192 ct += sizeof(unsigned long);
193 tcount -= sizeof(unsigned long);
197 mask = bytemask_from_count(tcount);
198 return unlikely(!!((a ^ b) & mask));
203 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
217 static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount)
220 * Be careful about RCU walk racing with rename:
221 * use 'READ_ONCE' to fetch the name pointer.
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.
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)
235 const unsigned char *cs = READ_ONCE(dentry->d_name.name);
237 return dentry_string_cmp(cs, ct, tcount);
240 struct external_name {
243 struct rcu_head head;
245 unsigned char name[];
248 static inline struct external_name *external_name(struct dentry *dentry)
250 return container_of(dentry->d_name.name, struct external_name, name[0]);
253 static void __d_free(struct rcu_head *head)
255 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
257 kmem_cache_free(dentry_cache, dentry);
260 static void __d_free_external_name(struct rcu_head *head)
262 struct external_name *name = container_of(head, struct external_name,
265 mod_node_page_state(page_pgdat(virt_to_page(name)),
266 NR_INDIRECTLY_RECLAIMABLE_BYTES,
272 static void __d_free_external(struct rcu_head *head)
274 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
276 __d_free_external_name(&external_name(dentry)->u.head);
278 kmem_cache_free(dentry_cache, dentry);
281 static inline int dname_external(const struct dentry *dentry)
283 return dentry->d_name.name != dentry->d_iname;
286 void take_dentry_name_snapshot(struct name_snapshot *name, struct dentry *dentry)
288 spin_lock(&dentry->d_lock);
289 if (unlikely(dname_external(dentry))) {
290 struct external_name *p = external_name(dentry);
291 atomic_inc(&p->u.count);
292 spin_unlock(&dentry->d_lock);
293 name->name = p->name;
295 memcpy(name->inline_name, dentry->d_iname, DNAME_INLINE_LEN);
296 spin_unlock(&dentry->d_lock);
297 name->name = name->inline_name;
300 EXPORT_SYMBOL(take_dentry_name_snapshot);
302 void release_dentry_name_snapshot(struct name_snapshot *name)
304 if (unlikely(name->name != name->inline_name)) {
305 struct external_name *p;
306 p = container_of(name->name, struct external_name, name[0]);
307 if (unlikely(atomic_dec_and_test(&p->u.count)))
308 call_rcu(&p->u.head, __d_free_external_name);
311 EXPORT_SYMBOL(release_dentry_name_snapshot);
313 static inline void __d_set_inode_and_type(struct dentry *dentry,
319 dentry->d_inode = inode;
320 flags = READ_ONCE(dentry->d_flags);
321 flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
323 WRITE_ONCE(dentry->d_flags, flags);
326 static inline void __d_clear_type_and_inode(struct dentry *dentry)
328 unsigned flags = READ_ONCE(dentry->d_flags);
330 flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
331 WRITE_ONCE(dentry->d_flags, flags);
332 dentry->d_inode = NULL;
335 static void dentry_free(struct dentry *dentry)
337 WARN_ON(!hlist_unhashed(&dentry->d_u.d_alias));
338 if (unlikely(dname_external(dentry))) {
339 struct external_name *p = external_name(dentry);
340 if (likely(atomic_dec_and_test(&p->u.count))) {
341 call_rcu(&dentry->d_u.d_rcu, __d_free_external);
345 /* if dentry was never visible to RCU, immediate free is OK */
346 if (!(dentry->d_flags & DCACHE_RCUACCESS))
347 __d_free(&dentry->d_u.d_rcu);
349 call_rcu(&dentry->d_u.d_rcu, __d_free);
353 * Release the dentry's inode, using the filesystem
354 * d_iput() operation if defined.
356 static void dentry_unlink_inode(struct dentry * dentry)
357 __releases(dentry->d_lock)
358 __releases(dentry->d_inode->i_lock)
360 struct inode *inode = dentry->d_inode;
361 bool hashed = !d_unhashed(dentry);
364 raw_write_seqcount_begin(&dentry->d_seq);
365 __d_clear_type_and_inode(dentry);
366 hlist_del_init(&dentry->d_u.d_alias);
368 raw_write_seqcount_end(&dentry->d_seq);
369 spin_unlock(&dentry->d_lock);
370 spin_unlock(&inode->i_lock);
372 fsnotify_inoderemove(inode);
373 if (dentry->d_op && dentry->d_op->d_iput)
374 dentry->d_op->d_iput(dentry, inode);
380 * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
381 * is in use - which includes both the "real" per-superblock
382 * LRU list _and_ the DCACHE_SHRINK_LIST use.
384 * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
385 * on the shrink list (ie not on the superblock LRU list).
387 * The per-cpu "nr_dentry_unused" counters are updated with
388 * the DCACHE_LRU_LIST bit.
390 * These helper functions make sure we always follow the
391 * rules. d_lock must be held by the caller.
393 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
394 static void d_lru_add(struct dentry *dentry)
396 D_FLAG_VERIFY(dentry, 0);
397 dentry->d_flags |= DCACHE_LRU_LIST;
398 this_cpu_inc(nr_dentry_unused);
399 WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
402 static void d_lru_del(struct dentry *dentry)
404 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
405 dentry->d_flags &= ~DCACHE_LRU_LIST;
406 this_cpu_dec(nr_dentry_unused);
407 WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
410 static void d_shrink_del(struct dentry *dentry)
412 D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
413 list_del_init(&dentry->d_lru);
414 dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
415 this_cpu_dec(nr_dentry_unused);
418 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
420 D_FLAG_VERIFY(dentry, 0);
421 list_add(&dentry->d_lru, list);
422 dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
423 this_cpu_inc(nr_dentry_unused);
427 * These can only be called under the global LRU lock, ie during the
428 * callback for freeing the LRU list. "isolate" removes it from the
429 * LRU lists entirely, while shrink_move moves it to the indicated
432 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
434 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
435 dentry->d_flags &= ~DCACHE_LRU_LIST;
436 this_cpu_dec(nr_dentry_unused);
437 list_lru_isolate(lru, &dentry->d_lru);
440 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
441 struct list_head *list)
443 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
444 dentry->d_flags |= DCACHE_SHRINK_LIST;
445 list_lru_isolate_move(lru, &dentry->d_lru, list);
449 * d_drop - drop a dentry
450 * @dentry: dentry to drop
452 * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
453 * be found through a VFS lookup any more. Note that this is different from
454 * deleting the dentry - d_delete will try to mark the dentry negative if
455 * possible, giving a successful _negative_ lookup, while d_drop will
456 * just make the cache lookup fail.
458 * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
459 * reason (NFS timeouts or autofs deletes).
461 * __d_drop requires dentry->d_lock
462 * ___d_drop doesn't mark dentry as "unhashed"
463 * (dentry->d_hash.pprev will be LIST_POISON2, not NULL).
465 static void ___d_drop(struct dentry *dentry)
467 struct hlist_bl_head *b;
469 * Hashed dentries are normally on the dentry hashtable,
470 * with the exception of those newly allocated by
471 * d_obtain_root, which are always IS_ROOT:
473 if (unlikely(IS_ROOT(dentry)))
474 b = &dentry->d_sb->s_roots;
476 b = d_hash(dentry->d_name.hash);
479 __hlist_bl_del(&dentry->d_hash);
483 void __d_drop(struct dentry *dentry)
485 if (!d_unhashed(dentry)) {
487 dentry->d_hash.pprev = NULL;
488 write_seqcount_invalidate(&dentry->d_seq);
491 EXPORT_SYMBOL(__d_drop);
493 void d_drop(struct dentry *dentry)
495 spin_lock(&dentry->d_lock);
497 spin_unlock(&dentry->d_lock);
499 EXPORT_SYMBOL(d_drop);
501 static inline void dentry_unlist(struct dentry *dentry, struct dentry *parent)
505 * Inform d_walk() and shrink_dentry_list() that we are no longer
506 * attached to the dentry tree
508 dentry->d_flags |= DCACHE_DENTRY_KILLED;
509 if (unlikely(list_empty(&dentry->d_child)))
511 __list_del_entry(&dentry->d_child);
513 * Cursors can move around the list of children. While we'd been
514 * a normal list member, it didn't matter - ->d_child.next would've
515 * been updated. However, from now on it won't be and for the
516 * things like d_walk() it might end up with a nasty surprise.
517 * Normally d_walk() doesn't care about cursors moving around -
518 * ->d_lock on parent prevents that and since a cursor has no children
519 * of its own, we get through it without ever unlocking the parent.
520 * There is one exception, though - if we ascend from a child that
521 * gets killed as soon as we unlock it, the next sibling is found
522 * using the value left in its ->d_child.next. And if _that_
523 * pointed to a cursor, and cursor got moved (e.g. by lseek())
524 * before d_walk() regains parent->d_lock, we'll end up skipping
525 * everything the cursor had been moved past.
527 * Solution: make sure that the pointer left behind in ->d_child.next
528 * points to something that won't be moving around. I.e. skip the
531 while (dentry->d_child.next != &parent->d_subdirs) {
532 next = list_entry(dentry->d_child.next, struct dentry, d_child);
533 if (likely(!(next->d_flags & DCACHE_DENTRY_CURSOR)))
535 dentry->d_child.next = next->d_child.next;
539 static void __dentry_kill(struct dentry *dentry)
541 struct dentry *parent = NULL;
542 bool can_free = true;
543 if (!IS_ROOT(dentry))
544 parent = dentry->d_parent;
547 * The dentry is now unrecoverably dead to the world.
549 lockref_mark_dead(&dentry->d_lockref);
552 * inform the fs via d_prune that this dentry is about to be
553 * unhashed and destroyed.
555 if (dentry->d_flags & DCACHE_OP_PRUNE)
556 dentry->d_op->d_prune(dentry);
558 if (dentry->d_flags & DCACHE_LRU_LIST) {
559 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
562 /* if it was on the hash then remove it */
564 dentry_unlist(dentry, parent);
566 spin_unlock(&parent->d_lock);
568 dentry_unlink_inode(dentry);
570 spin_unlock(&dentry->d_lock);
571 this_cpu_dec(nr_dentry);
572 if (dentry->d_op && dentry->d_op->d_release)
573 dentry->d_op->d_release(dentry);
575 spin_lock(&dentry->d_lock);
576 if (dentry->d_flags & DCACHE_SHRINK_LIST) {
577 dentry->d_flags |= DCACHE_MAY_FREE;
580 spin_unlock(&dentry->d_lock);
581 if (likely(can_free))
585 static struct dentry *__lock_parent(struct dentry *dentry)
587 struct dentry *parent;
589 spin_unlock(&dentry->d_lock);
591 parent = READ_ONCE(dentry->d_parent);
592 spin_lock(&parent->d_lock);
594 * We can't blindly lock dentry until we are sure
595 * that we won't violate the locking order.
596 * Any changes of dentry->d_parent must have
597 * been done with parent->d_lock held, so
598 * spin_lock() above is enough of a barrier
599 * for checking if it's still our child.
601 if (unlikely(parent != dentry->d_parent)) {
602 spin_unlock(&parent->d_lock);
606 if (parent != dentry)
607 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
613 static inline struct dentry *lock_parent(struct dentry *dentry)
615 struct dentry *parent = dentry->d_parent;
618 if (likely(spin_trylock(&parent->d_lock)))
620 return __lock_parent(dentry);
623 static inline bool retain_dentry(struct dentry *dentry)
625 WARN_ON(d_in_lookup(dentry));
627 /* Unreachable? Get rid of it */
628 if (unlikely(d_unhashed(dentry)))
631 if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
634 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
635 if (dentry->d_op->d_delete(dentry))
638 /* retain; LRU fodder */
639 dentry->d_lockref.count--;
640 if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
642 else if (unlikely(!(dentry->d_flags & DCACHE_REFERENCED)))
643 dentry->d_flags |= DCACHE_REFERENCED;
648 * Finish off a dentry we've decided to kill.
649 * dentry->d_lock must be held, returns with it unlocked.
650 * Returns dentry requiring refcount drop, or NULL if we're done.
652 static struct dentry *dentry_kill(struct dentry *dentry)
653 __releases(dentry->d_lock)
655 struct inode *inode = dentry->d_inode;
656 struct dentry *parent = NULL;
658 if (inode && unlikely(!spin_trylock(&inode->i_lock)))
661 if (!IS_ROOT(dentry)) {
662 parent = dentry->d_parent;
663 if (unlikely(!spin_trylock(&parent->d_lock))) {
664 parent = __lock_parent(dentry);
665 if (likely(inode || !dentry->d_inode))
667 /* negative that became positive */
669 spin_unlock(&parent->d_lock);
670 inode = dentry->d_inode;
674 __dentry_kill(dentry);
678 spin_unlock(&dentry->d_lock);
679 spin_lock(&inode->i_lock);
680 spin_lock(&dentry->d_lock);
681 parent = lock_parent(dentry);
683 if (unlikely(dentry->d_lockref.count != 1)) {
684 dentry->d_lockref.count--;
685 } else if (likely(!retain_dentry(dentry))) {
686 __dentry_kill(dentry);
689 /* we are keeping it, after all */
691 spin_unlock(&inode->i_lock);
693 spin_unlock(&parent->d_lock);
694 spin_unlock(&dentry->d_lock);
699 * Try to do a lockless dput(), and return whether that was successful.
701 * If unsuccessful, we return false, having already taken the dentry lock.
703 * The caller needs to hold the RCU read lock, so that the dentry is
704 * guaranteed to stay around even if the refcount goes down to zero!
706 static inline bool fast_dput(struct dentry *dentry)
709 unsigned int d_flags;
712 * If we have a d_op->d_delete() operation, we sould not
713 * let the dentry count go to zero, so use "put_or_lock".
715 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE))
716 return lockref_put_or_lock(&dentry->d_lockref);
719 * .. otherwise, we can try to just decrement the
720 * lockref optimistically.
722 ret = lockref_put_return(&dentry->d_lockref);
725 * If the lockref_put_return() failed due to the lock being held
726 * by somebody else, the fast path has failed. We will need to
727 * get the lock, and then check the count again.
729 if (unlikely(ret < 0)) {
730 spin_lock(&dentry->d_lock);
731 if (dentry->d_lockref.count > 1) {
732 dentry->d_lockref.count--;
733 spin_unlock(&dentry->d_lock);
740 * If we weren't the last ref, we're done.
746 * Careful, careful. The reference count went down
747 * to zero, but we don't hold the dentry lock, so
748 * somebody else could get it again, and do another
749 * dput(), and we need to not race with that.
751 * However, there is a very special and common case
752 * where we don't care, because there is nothing to
753 * do: the dentry is still hashed, it does not have
754 * a 'delete' op, and it's referenced and already on
757 * NOTE! Since we aren't locked, these values are
758 * not "stable". However, it is sufficient that at
759 * some point after we dropped the reference the
760 * dentry was hashed and the flags had the proper
761 * value. Other dentry users may have re-gotten
762 * a reference to the dentry and change that, but
763 * our work is done - we can leave the dentry
764 * around with a zero refcount.
767 d_flags = READ_ONCE(dentry->d_flags);
768 d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST | DCACHE_DISCONNECTED;
770 /* Nothing to do? Dropping the reference was all we needed? */
771 if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry))
775 * Not the fast normal case? Get the lock. We've already decremented
776 * the refcount, but we'll need to re-check the situation after
779 spin_lock(&dentry->d_lock);
782 * Did somebody else grab a reference to it in the meantime, and
783 * we're no longer the last user after all? Alternatively, somebody
784 * else could have killed it and marked it dead. Either way, we
785 * don't need to do anything else.
787 if (dentry->d_lockref.count) {
788 spin_unlock(&dentry->d_lock);
793 * Re-get the reference we optimistically dropped. We hold the
794 * lock, and we just tested that it was zero, so we can just
797 dentry->d_lockref.count = 1;
805 * This is complicated by the fact that we do not want to put
806 * dentries that are no longer on any hash chain on the unused
807 * list: we'd much rather just get rid of them immediately.
809 * However, that implies that we have to traverse the dentry
810 * tree upwards to the parents which might _also_ now be
811 * scheduled for deletion (it may have been only waiting for
812 * its last child to go away).
814 * This tail recursion is done by hand as we don't want to depend
815 * on the compiler to always get this right (gcc generally doesn't).
816 * Real recursion would eat up our stack space.
820 * dput - release a dentry
821 * @dentry: dentry to release
823 * Release a dentry. This will drop the usage count and if appropriate
824 * call the dentry unlink method as well as removing it from the queues and
825 * releasing its resources. If the parent dentries were scheduled for release
826 * they too may now get deleted.
828 void dput(struct dentry *dentry)
830 if (unlikely(!dentry))
837 if (likely(fast_dput(dentry))) {
842 /* Slow case: now with the dentry lock held */
845 if (likely(retain_dentry(dentry))) {
846 spin_unlock(&dentry->d_lock);
850 dentry = dentry_kill(dentry);
859 /* This must be called with d_lock held */
860 static inline void __dget_dlock(struct dentry *dentry)
862 dentry->d_lockref.count++;
865 static inline void __dget(struct dentry *dentry)
867 lockref_get(&dentry->d_lockref);
870 struct dentry *dget_parent(struct dentry *dentry)
876 * Do optimistic parent lookup without any
880 ret = READ_ONCE(dentry->d_parent);
881 gotref = lockref_get_not_zero(&ret->d_lockref);
883 if (likely(gotref)) {
884 if (likely(ret == READ_ONCE(dentry->d_parent)))
891 * Don't need rcu_dereference because we re-check it was correct under
895 ret = dentry->d_parent;
896 spin_lock(&ret->d_lock);
897 if (unlikely(ret != dentry->d_parent)) {
898 spin_unlock(&ret->d_lock);
903 BUG_ON(!ret->d_lockref.count);
904 ret->d_lockref.count++;
905 spin_unlock(&ret->d_lock);
908 EXPORT_SYMBOL(dget_parent);
911 * d_find_alias - grab a hashed alias of inode
912 * @inode: inode in question
914 * If inode has a hashed alias, or is a directory and has any alias,
915 * acquire the reference to alias and return it. Otherwise return NULL.
916 * Notice that if inode is a directory there can be only one alias and
917 * it can be unhashed only if it has no children, or if it is the root
918 * of a filesystem, or if the directory was renamed and d_revalidate
919 * was the first vfs operation to notice.
921 * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
922 * any other hashed alias over that one.
924 static struct dentry *__d_find_alias(struct inode *inode)
926 struct dentry *alias, *discon_alias;
930 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
931 spin_lock(&alias->d_lock);
932 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
933 if (IS_ROOT(alias) &&
934 (alias->d_flags & DCACHE_DISCONNECTED)) {
935 discon_alias = alias;
938 spin_unlock(&alias->d_lock);
942 spin_unlock(&alias->d_lock);
945 alias = discon_alias;
946 spin_lock(&alias->d_lock);
947 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
949 spin_unlock(&alias->d_lock);
952 spin_unlock(&alias->d_lock);
958 struct dentry *d_find_alias(struct inode *inode)
960 struct dentry *de = NULL;
962 if (!hlist_empty(&inode->i_dentry)) {
963 spin_lock(&inode->i_lock);
964 de = __d_find_alias(inode);
965 spin_unlock(&inode->i_lock);
969 EXPORT_SYMBOL(d_find_alias);
972 * Try to kill dentries associated with this inode.
973 * WARNING: you must own a reference to inode.
975 void d_prune_aliases(struct inode *inode)
977 struct dentry *dentry;
979 spin_lock(&inode->i_lock);
980 hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
981 spin_lock(&dentry->d_lock);
982 if (!dentry->d_lockref.count) {
983 struct dentry *parent = lock_parent(dentry);
984 if (likely(!dentry->d_lockref.count)) {
985 __dentry_kill(dentry);
990 spin_unlock(&parent->d_lock);
992 spin_unlock(&dentry->d_lock);
994 spin_unlock(&inode->i_lock);
996 EXPORT_SYMBOL(d_prune_aliases);
999 * Lock a dentry from shrink list.
1000 * Called under rcu_read_lock() and dentry->d_lock; the former
1001 * guarantees that nothing we access will be freed under us.
1002 * Note that dentry is *not* protected from concurrent dentry_kill(),
1005 * Return false if dentry has been disrupted or grabbed, leaving
1006 * the caller to kick it off-list. Otherwise, return true and have
1007 * that dentry's inode and parent both locked.
1009 static bool shrink_lock_dentry(struct dentry *dentry)
1011 struct inode *inode;
1012 struct dentry *parent;
1014 if (dentry->d_lockref.count)
1017 inode = dentry->d_inode;
1018 if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
1019 spin_unlock(&dentry->d_lock);
1020 spin_lock(&inode->i_lock);
1021 spin_lock(&dentry->d_lock);
1022 if (unlikely(dentry->d_lockref.count))
1024 /* changed inode means that somebody had grabbed it */
1025 if (unlikely(inode != dentry->d_inode))
1029 parent = dentry->d_parent;
1030 if (IS_ROOT(dentry) || likely(spin_trylock(&parent->d_lock)))
1033 spin_unlock(&dentry->d_lock);
1034 spin_lock(&parent->d_lock);
1035 if (unlikely(parent != dentry->d_parent)) {
1036 spin_unlock(&parent->d_lock);
1037 spin_lock(&dentry->d_lock);
1040 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1041 if (likely(!dentry->d_lockref.count))
1043 spin_unlock(&parent->d_lock);
1046 spin_unlock(&inode->i_lock);
1050 static void shrink_dentry_list(struct list_head *list)
1052 while (!list_empty(list)) {
1053 struct dentry *dentry, *parent;
1057 dentry = list_entry(list->prev, struct dentry, d_lru);
1058 spin_lock(&dentry->d_lock);
1060 if (!shrink_lock_dentry(dentry)) {
1061 bool can_free = false;
1063 d_shrink_del(dentry);
1064 if (dentry->d_lockref.count < 0)
1065 can_free = dentry->d_flags & DCACHE_MAY_FREE;
1066 spin_unlock(&dentry->d_lock);
1068 dentry_free(dentry);
1072 d_shrink_del(dentry);
1073 parent = dentry->d_parent;
1074 __dentry_kill(dentry);
1075 if (parent == dentry)
1078 * We need to prune ancestors too. This is necessary to prevent
1079 * quadratic behavior of shrink_dcache_parent(), but is also
1080 * expected to be beneficial in reducing dentry cache
1084 while (dentry && !lockref_put_or_lock(&dentry->d_lockref))
1085 dentry = dentry_kill(dentry);
1089 static enum lru_status dentry_lru_isolate(struct list_head *item,
1090 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1092 struct list_head *freeable = arg;
1093 struct dentry *dentry = container_of(item, struct dentry, d_lru);
1097 * we are inverting the lru lock/dentry->d_lock here,
1098 * so use a trylock. If we fail to get the lock, just skip
1101 if (!spin_trylock(&dentry->d_lock))
1105 * Referenced dentries are still in use. If they have active
1106 * counts, just remove them from the LRU. Otherwise give them
1107 * another pass through the LRU.
1109 if (dentry->d_lockref.count) {
1110 d_lru_isolate(lru, dentry);
1111 spin_unlock(&dentry->d_lock);
1115 if (dentry->d_flags & DCACHE_REFERENCED) {
1116 dentry->d_flags &= ~DCACHE_REFERENCED;
1117 spin_unlock(&dentry->d_lock);
1120 * The list move itself will be made by the common LRU code. At
1121 * this point, we've dropped the dentry->d_lock but keep the
1122 * lru lock. This is safe to do, since every list movement is
1123 * protected by the lru lock even if both locks are held.
1125 * This is guaranteed by the fact that all LRU management
1126 * functions are intermediated by the LRU API calls like
1127 * list_lru_add and list_lru_del. List movement in this file
1128 * only ever occur through this functions or through callbacks
1129 * like this one, that are called from the LRU API.
1131 * The only exceptions to this are functions like
1132 * shrink_dentry_list, and code that first checks for the
1133 * DCACHE_SHRINK_LIST flag. Those are guaranteed to be
1134 * operating only with stack provided lists after they are
1135 * properly isolated from the main list. It is thus, always a
1141 d_lru_shrink_move(lru, dentry, freeable);
1142 spin_unlock(&dentry->d_lock);
1148 * prune_dcache_sb - shrink the dcache
1150 * @sc: shrink control, passed to list_lru_shrink_walk()
1152 * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1153 * is done when we need more memory and called from the superblock shrinker
1156 * This function may fail to free any resources if all the dentries are in
1159 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1164 freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1165 dentry_lru_isolate, &dispose);
1166 shrink_dentry_list(&dispose);
1170 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1171 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1173 struct list_head *freeable = arg;
1174 struct dentry *dentry = container_of(item, struct dentry, d_lru);
1177 * we are inverting the lru lock/dentry->d_lock here,
1178 * so use a trylock. If we fail to get the lock, just skip
1181 if (!spin_trylock(&dentry->d_lock))
1184 d_lru_shrink_move(lru, dentry, freeable);
1185 spin_unlock(&dentry->d_lock);
1192 * shrink_dcache_sb - shrink dcache for a superblock
1195 * Shrink the dcache for the specified super block. This is used to free
1196 * the dcache before unmounting a file system.
1198 void shrink_dcache_sb(struct super_block *sb)
1205 freed = list_lru_walk(&sb->s_dentry_lru,
1206 dentry_lru_isolate_shrink, &dispose, 1024);
1208 this_cpu_sub(nr_dentry_unused, freed);
1209 shrink_dentry_list(&dispose);
1210 } while (list_lru_count(&sb->s_dentry_lru) > 0);
1212 EXPORT_SYMBOL(shrink_dcache_sb);
1215 * enum d_walk_ret - action to talke during tree walk
1216 * @D_WALK_CONTINUE: contrinue walk
1217 * @D_WALK_QUIT: quit walk
1218 * @D_WALK_NORETRY: quit when retry is needed
1219 * @D_WALK_SKIP: skip this dentry and its children
1229 * d_walk - walk the dentry tree
1230 * @parent: start of walk
1231 * @data: data passed to @enter() and @finish()
1232 * @enter: callback when first entering the dentry
1233 * @finish: callback when successfully finished the walk
1235 * The @enter() and @finish() callbacks are called with d_lock held.
1237 static void d_walk(struct dentry *parent, void *data,
1238 enum d_walk_ret (*enter)(void *, struct dentry *),
1239 void (*finish)(void *))
1241 struct dentry *this_parent;
1242 struct list_head *next;
1244 enum d_walk_ret ret;
1248 read_seqbegin_or_lock(&rename_lock, &seq);
1249 this_parent = parent;
1250 spin_lock(&this_parent->d_lock);
1252 ret = enter(data, this_parent);
1254 case D_WALK_CONTINUE:
1259 case D_WALK_NORETRY:
1264 next = this_parent->d_subdirs.next;
1266 while (next != &this_parent->d_subdirs) {
1267 struct list_head *tmp = next;
1268 struct dentry *dentry = list_entry(tmp, struct dentry, d_child);
1271 if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
1274 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1276 ret = enter(data, dentry);
1278 case D_WALK_CONTINUE:
1281 spin_unlock(&dentry->d_lock);
1283 case D_WALK_NORETRY:
1287 spin_unlock(&dentry->d_lock);
1291 if (!list_empty(&dentry->d_subdirs)) {
1292 spin_unlock(&this_parent->d_lock);
1293 spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1294 this_parent = dentry;
1295 spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1298 spin_unlock(&dentry->d_lock);
1301 * All done at this level ... ascend and resume the search.
1305 if (this_parent != parent) {
1306 struct dentry *child = this_parent;
1307 this_parent = child->d_parent;
1309 spin_unlock(&child->d_lock);
1310 spin_lock(&this_parent->d_lock);
1312 /* might go back up the wrong parent if we have had a rename. */
1313 if (need_seqretry(&rename_lock, seq))
1315 /* go into the first sibling still alive */
1317 next = child->d_child.next;
1318 if (next == &this_parent->d_subdirs)
1320 child = list_entry(next, struct dentry, d_child);
1321 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED));
1325 if (need_seqretry(&rename_lock, seq))
1332 spin_unlock(&this_parent->d_lock);
1333 done_seqretry(&rename_lock, seq);
1337 spin_unlock(&this_parent->d_lock);
1346 struct check_mount {
1347 struct vfsmount *mnt;
1348 unsigned int mounted;
1351 static enum d_walk_ret path_check_mount(void *data, struct dentry *dentry)
1353 struct check_mount *info = data;
1354 struct path path = { .mnt = info->mnt, .dentry = dentry };
1356 if (likely(!d_mountpoint(dentry)))
1357 return D_WALK_CONTINUE;
1358 if (__path_is_mountpoint(&path)) {
1362 return D_WALK_CONTINUE;
1366 * path_has_submounts - check for mounts over a dentry in the
1367 * current namespace.
1368 * @parent: path to check.
1370 * Return true if the parent or its subdirectories contain
1371 * a mount point in the current namespace.
1373 int path_has_submounts(const struct path *parent)
1375 struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };
1377 read_seqlock_excl(&mount_lock);
1378 d_walk(parent->dentry, &data, path_check_mount, NULL);
1379 read_sequnlock_excl(&mount_lock);
1381 return data.mounted;
1383 EXPORT_SYMBOL(path_has_submounts);
1386 * Called by mount code to set a mountpoint and check if the mountpoint is
1387 * reachable (e.g. NFS can unhash a directory dentry and then the complete
1388 * subtree can become unreachable).
1390 * Only one of d_invalidate() and d_set_mounted() must succeed. For
1391 * this reason take rename_lock and d_lock on dentry and ancestors.
1393 int d_set_mounted(struct dentry *dentry)
1397 write_seqlock(&rename_lock);
1398 for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1399 /* Need exclusion wrt. d_invalidate() */
1400 spin_lock(&p->d_lock);
1401 if (unlikely(d_unhashed(p))) {
1402 spin_unlock(&p->d_lock);
1405 spin_unlock(&p->d_lock);
1407 spin_lock(&dentry->d_lock);
1408 if (!d_unlinked(dentry)) {
1410 if (!d_mountpoint(dentry)) {
1411 dentry->d_flags |= DCACHE_MOUNTED;
1415 spin_unlock(&dentry->d_lock);
1417 write_sequnlock(&rename_lock);
1422 * Search the dentry child list of the specified parent,
1423 * and move any unused dentries to the end of the unused
1424 * list for prune_dcache(). We descend to the next level
1425 * whenever the d_subdirs list is non-empty and continue
1428 * It returns zero iff there are no unused children,
1429 * otherwise it returns the number of children moved to
1430 * the end of the unused list. This may not be the total
1431 * number of unused children, because select_parent can
1432 * drop the lock and return early due to latency
1436 struct select_data {
1437 struct dentry *start;
1438 struct list_head dispose;
1442 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1444 struct select_data *data = _data;
1445 enum d_walk_ret ret = D_WALK_CONTINUE;
1447 if (data->start == dentry)
1450 if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1453 if (dentry->d_flags & DCACHE_LRU_LIST)
1455 if (!dentry->d_lockref.count) {
1456 d_shrink_add(dentry, &data->dispose);
1461 * We can return to the caller if we have found some (this
1462 * ensures forward progress). We'll be coming back to find
1465 if (!list_empty(&data->dispose))
1466 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1472 * shrink_dcache_parent - prune dcache
1473 * @parent: parent of entries to prune
1475 * Prune the dcache to remove unused children of the parent dentry.
1477 void shrink_dcache_parent(struct dentry *parent)
1480 struct select_data data;
1482 INIT_LIST_HEAD(&data.dispose);
1483 data.start = parent;
1486 d_walk(parent, &data, select_collect, NULL);
1490 shrink_dentry_list(&data.dispose);
1493 EXPORT_SYMBOL(shrink_dcache_parent);
1495 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1497 /* it has busy descendents; complain about those instead */
1498 if (!list_empty(&dentry->d_subdirs))
1499 return D_WALK_CONTINUE;
1501 /* root with refcount 1 is fine */
1502 if (dentry == _data && dentry->d_lockref.count == 1)
1503 return D_WALK_CONTINUE;
1505 printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} "
1506 " still in use (%d) [unmount of %s %s]\n",
1509 dentry->d_inode->i_ino : 0UL,
1511 dentry->d_lockref.count,
1512 dentry->d_sb->s_type->name,
1513 dentry->d_sb->s_id);
1515 return D_WALK_CONTINUE;
1518 static void do_one_tree(struct dentry *dentry)
1520 shrink_dcache_parent(dentry);
1521 d_walk(dentry, dentry, umount_check, NULL);
1527 * destroy the dentries attached to a superblock on unmounting
1529 void shrink_dcache_for_umount(struct super_block *sb)
1531 struct dentry *dentry;
1533 WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
1535 dentry = sb->s_root;
1537 do_one_tree(dentry);
1539 while (!hlist_bl_empty(&sb->s_roots)) {
1540 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_roots), struct dentry, d_hash));
1541 do_one_tree(dentry);
1545 struct detach_data {
1546 struct select_data select;
1547 struct dentry *mountpoint;
1549 static enum d_walk_ret detach_and_collect(void *_data, struct dentry *dentry)
1551 struct detach_data *data = _data;
1553 if (d_mountpoint(dentry)) {
1554 __dget_dlock(dentry);
1555 data->mountpoint = dentry;
1559 return select_collect(&data->select, dentry);
1562 static void check_and_drop(void *_data)
1564 struct detach_data *data = _data;
1566 if (!data->mountpoint && list_empty(&data->select.dispose))
1567 __d_drop(data->select.start);
1571 * d_invalidate - detach submounts, prune dcache, and drop
1572 * @dentry: dentry to invalidate (aka detach, prune and drop)
1576 * The final d_drop is done as an atomic operation relative to
1577 * rename_lock ensuring there are no races with d_set_mounted. This
1578 * ensures there are no unhashed dentries on the path to a mountpoint.
1580 void d_invalidate(struct dentry *dentry)
1583 * If it's already been dropped, return OK.
1585 spin_lock(&dentry->d_lock);
1586 if (d_unhashed(dentry)) {
1587 spin_unlock(&dentry->d_lock);
1590 spin_unlock(&dentry->d_lock);
1592 /* Negative dentries can be dropped without further checks */
1593 if (!dentry->d_inode) {
1599 struct detach_data data;
1601 data.mountpoint = NULL;
1602 INIT_LIST_HEAD(&data.select.dispose);
1603 data.select.start = dentry;
1604 data.select.found = 0;
1606 d_walk(dentry, &data, detach_and_collect, check_and_drop);
1608 if (!list_empty(&data.select.dispose))
1609 shrink_dentry_list(&data.select.dispose);
1610 else if (!data.mountpoint)
1613 if (data.mountpoint) {
1614 detach_mounts(data.mountpoint);
1615 dput(data.mountpoint);
1619 EXPORT_SYMBOL(d_invalidate);
1622 * __d_alloc - allocate a dcache entry
1623 * @sb: filesystem it will belong to
1624 * @name: qstr of the name
1626 * Allocates a dentry. It returns %NULL if there is insufficient memory
1627 * available. On a success the dentry is returned. The name passed in is
1628 * copied and the copy passed in may be reused after this call.
1631 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1633 struct external_name *ext = NULL;
1634 struct dentry *dentry;
1638 dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1643 * We guarantee that the inline name is always NUL-terminated.
1644 * This way the memcpy() done by the name switching in rename
1645 * will still always have a NUL at the end, even if we might
1646 * be overwriting an internal NUL character
1648 dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1649 if (unlikely(!name)) {
1651 dname = dentry->d_iname;
1652 } else if (name->len > DNAME_INLINE_LEN-1) {
1653 size_t size = offsetof(struct external_name, name[1]);
1655 ext = kmalloc(size + name->len, GFP_KERNEL_ACCOUNT);
1657 kmem_cache_free(dentry_cache, dentry);
1660 atomic_set(&ext->u.count, 1);
1663 dname = dentry->d_iname;
1666 dentry->d_name.len = name->len;
1667 dentry->d_name.hash = name->hash;
1668 memcpy(dname, name->name, name->len);
1669 dname[name->len] = 0;
1671 /* Make sure we always see the terminating NUL character */
1672 smp_store_release(&dentry->d_name.name, dname); /* ^^^ */
1674 dentry->d_lockref.count = 1;
1675 dentry->d_flags = 0;
1676 spin_lock_init(&dentry->d_lock);
1677 seqcount_init(&dentry->d_seq);
1678 dentry->d_inode = NULL;
1679 dentry->d_parent = dentry;
1681 dentry->d_op = NULL;
1682 dentry->d_fsdata = NULL;
1683 INIT_HLIST_BL_NODE(&dentry->d_hash);
1684 INIT_LIST_HEAD(&dentry->d_lru);
1685 INIT_LIST_HEAD(&dentry->d_subdirs);
1686 INIT_HLIST_NODE(&dentry->d_u.d_alias);
1687 INIT_LIST_HEAD(&dentry->d_child);
1688 d_set_d_op(dentry, dentry->d_sb->s_d_op);
1690 if (dentry->d_op && dentry->d_op->d_init) {
1691 err = dentry->d_op->d_init(dentry);
1693 if (dname_external(dentry))
1694 kfree(external_name(dentry));
1695 kmem_cache_free(dentry_cache, dentry);
1700 if (unlikely(ext)) {
1701 pg_data_t *pgdat = page_pgdat(virt_to_page(ext));
1702 mod_node_page_state(pgdat, NR_INDIRECTLY_RECLAIMABLE_BYTES,
1706 this_cpu_inc(nr_dentry);
1712 * d_alloc - allocate a dcache entry
1713 * @parent: parent of entry to allocate
1714 * @name: qstr of the name
1716 * Allocates a dentry. It returns %NULL if there is insufficient memory
1717 * available. On a success the dentry is returned. The name passed in is
1718 * copied and the copy passed in may be reused after this call.
1720 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1722 struct dentry *dentry = __d_alloc(parent->d_sb, name);
1725 dentry->d_flags |= DCACHE_RCUACCESS;
1726 spin_lock(&parent->d_lock);
1728 * don't need child lock because it is not subject
1729 * to concurrency here
1731 __dget_dlock(parent);
1732 dentry->d_parent = parent;
1733 list_add(&dentry->d_child, &parent->d_subdirs);
1734 spin_unlock(&parent->d_lock);
1738 EXPORT_SYMBOL(d_alloc);
1740 struct dentry *d_alloc_anon(struct super_block *sb)
1742 return __d_alloc(sb, NULL);
1744 EXPORT_SYMBOL(d_alloc_anon);
1746 struct dentry *d_alloc_cursor(struct dentry * parent)
1748 struct dentry *dentry = d_alloc_anon(parent->d_sb);
1750 dentry->d_flags |= DCACHE_RCUACCESS | DCACHE_DENTRY_CURSOR;
1751 dentry->d_parent = dget(parent);
1757 * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1758 * @sb: the superblock
1759 * @name: qstr of the name
1761 * For a filesystem that just pins its dentries in memory and never
1762 * performs lookups at all, return an unhashed IS_ROOT dentry.
1764 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1766 return __d_alloc(sb, name);
1768 EXPORT_SYMBOL(d_alloc_pseudo);
1770 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1775 q.hash_len = hashlen_string(parent, name);
1776 return d_alloc(parent, &q);
1778 EXPORT_SYMBOL(d_alloc_name);
1780 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1782 WARN_ON_ONCE(dentry->d_op);
1783 WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH |
1785 DCACHE_OP_REVALIDATE |
1786 DCACHE_OP_WEAK_REVALIDATE |
1793 dentry->d_flags |= DCACHE_OP_HASH;
1795 dentry->d_flags |= DCACHE_OP_COMPARE;
1796 if (op->d_revalidate)
1797 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1798 if (op->d_weak_revalidate)
1799 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1801 dentry->d_flags |= DCACHE_OP_DELETE;
1803 dentry->d_flags |= DCACHE_OP_PRUNE;
1805 dentry->d_flags |= DCACHE_OP_REAL;
1808 EXPORT_SYMBOL(d_set_d_op);
1812 * d_set_fallthru - Mark a dentry as falling through to a lower layer
1813 * @dentry - The dentry to mark
1815 * Mark a dentry as falling through to the lower layer (as set with
1816 * d_pin_lower()). This flag may be recorded on the medium.
1818 void d_set_fallthru(struct dentry *dentry)
1820 spin_lock(&dentry->d_lock);
1821 dentry->d_flags |= DCACHE_FALLTHRU;
1822 spin_unlock(&dentry->d_lock);
1824 EXPORT_SYMBOL(d_set_fallthru);
1826 static unsigned d_flags_for_inode(struct inode *inode)
1828 unsigned add_flags = DCACHE_REGULAR_TYPE;
1831 return DCACHE_MISS_TYPE;
1833 if (S_ISDIR(inode->i_mode)) {
1834 add_flags = DCACHE_DIRECTORY_TYPE;
1835 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1836 if (unlikely(!inode->i_op->lookup))
1837 add_flags = DCACHE_AUTODIR_TYPE;
1839 inode->i_opflags |= IOP_LOOKUP;
1841 goto type_determined;
1844 if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1845 if (unlikely(inode->i_op->get_link)) {
1846 add_flags = DCACHE_SYMLINK_TYPE;
1847 goto type_determined;
1849 inode->i_opflags |= IOP_NOFOLLOW;
1852 if (unlikely(!S_ISREG(inode->i_mode)))
1853 add_flags = DCACHE_SPECIAL_TYPE;
1856 if (unlikely(IS_AUTOMOUNT(inode)))
1857 add_flags |= DCACHE_NEED_AUTOMOUNT;
1861 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1863 unsigned add_flags = d_flags_for_inode(inode);
1864 WARN_ON(d_in_lookup(dentry));
1866 spin_lock(&dentry->d_lock);
1867 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1868 raw_write_seqcount_begin(&dentry->d_seq);
1869 __d_set_inode_and_type(dentry, inode, add_flags);
1870 raw_write_seqcount_end(&dentry->d_seq);
1871 fsnotify_update_flags(dentry);
1872 spin_unlock(&dentry->d_lock);
1876 * d_instantiate - fill in inode information for a dentry
1877 * @entry: dentry to complete
1878 * @inode: inode to attach to this dentry
1880 * Fill in inode information in the entry.
1882 * This turns negative dentries into productive full members
1885 * NOTE! This assumes that the inode count has been incremented
1886 * (or otherwise set) by the caller to indicate that it is now
1887 * in use by the dcache.
1890 void d_instantiate(struct dentry *entry, struct inode * inode)
1892 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1894 security_d_instantiate(entry, inode);
1895 spin_lock(&inode->i_lock);
1896 __d_instantiate(entry, inode);
1897 spin_unlock(&inode->i_lock);
1900 EXPORT_SYMBOL(d_instantiate);
1903 * This should be equivalent to d_instantiate() + unlock_new_inode(),
1904 * with lockdep-related part of unlock_new_inode() done before
1905 * anything else. Use that instead of open-coding d_instantiate()/
1906 * unlock_new_inode() combinations.
1908 void d_instantiate_new(struct dentry *entry, struct inode *inode)
1910 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1912 lockdep_annotate_inode_mutex_key(inode);
1913 security_d_instantiate(entry, inode);
1914 spin_lock(&inode->i_lock);
1915 __d_instantiate(entry, inode);
1916 WARN_ON(!(inode->i_state & I_NEW));
1917 inode->i_state &= ~I_NEW;
1919 wake_up_bit(&inode->i_state, __I_NEW);
1920 spin_unlock(&inode->i_lock);
1922 EXPORT_SYMBOL(d_instantiate_new);
1925 * d_instantiate_no_diralias - instantiate a non-aliased dentry
1926 * @entry: dentry to complete
1927 * @inode: inode to attach to this dentry
1929 * Fill in inode information in the entry. If a directory alias is found, then
1930 * return an error (and drop inode). Together with d_materialise_unique() this
1931 * guarantees that a directory inode may never have more than one alias.
1933 int d_instantiate_no_diralias(struct dentry *entry, struct inode *inode)
1935 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1937 security_d_instantiate(entry, inode);
1938 spin_lock(&inode->i_lock);
1939 if (S_ISDIR(inode->i_mode) && !hlist_empty(&inode->i_dentry)) {
1940 spin_unlock(&inode->i_lock);
1944 __d_instantiate(entry, inode);
1945 spin_unlock(&inode->i_lock);
1949 EXPORT_SYMBOL(d_instantiate_no_diralias);
1951 struct dentry *d_make_root(struct inode *root_inode)
1953 struct dentry *res = NULL;
1956 res = d_alloc_anon(root_inode->i_sb);
1958 d_instantiate(res, root_inode);
1964 EXPORT_SYMBOL(d_make_root);
1966 static struct dentry * __d_find_any_alias(struct inode *inode)
1968 struct dentry *alias;
1970 if (hlist_empty(&inode->i_dentry))
1972 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
1978 * d_find_any_alias - find any alias for a given inode
1979 * @inode: inode to find an alias for
1981 * If any aliases exist for the given inode, take and return a
1982 * reference for one of them. If no aliases exist, return %NULL.
1984 struct dentry *d_find_any_alias(struct inode *inode)
1988 spin_lock(&inode->i_lock);
1989 de = __d_find_any_alias(inode);
1990 spin_unlock(&inode->i_lock);
1993 EXPORT_SYMBOL(d_find_any_alias);
1995 static struct dentry *__d_instantiate_anon(struct dentry *dentry,
1996 struct inode *inode,
2002 security_d_instantiate(dentry, inode);
2003 spin_lock(&inode->i_lock);
2004 res = __d_find_any_alias(inode);
2006 spin_unlock(&inode->i_lock);
2011 /* attach a disconnected dentry */
2012 add_flags = d_flags_for_inode(inode);
2015 add_flags |= DCACHE_DISCONNECTED;
2017 spin_lock(&dentry->d_lock);
2018 __d_set_inode_and_type(dentry, inode, add_flags);
2019 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2020 if (!disconnected) {
2021 hlist_bl_lock(&dentry->d_sb->s_roots);
2022 hlist_bl_add_head(&dentry->d_hash, &dentry->d_sb->s_roots);
2023 hlist_bl_unlock(&dentry->d_sb->s_roots);
2025 spin_unlock(&dentry->d_lock);
2026 spin_unlock(&inode->i_lock);
2035 struct dentry *d_instantiate_anon(struct dentry *dentry, struct inode *inode)
2037 return __d_instantiate_anon(dentry, inode, true);
2039 EXPORT_SYMBOL(d_instantiate_anon);
2041 static struct dentry *__d_obtain_alias(struct inode *inode, bool disconnected)
2047 return ERR_PTR(-ESTALE);
2049 return ERR_CAST(inode);
2051 res = d_find_any_alias(inode);
2055 tmp = d_alloc_anon(inode->i_sb);
2057 res = ERR_PTR(-ENOMEM);
2061 return __d_instantiate_anon(tmp, inode, disconnected);
2069 * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
2070 * @inode: inode to allocate the dentry for
2072 * Obtain a dentry for an inode resulting from NFS filehandle conversion or
2073 * similar open by handle operations. The returned dentry may be anonymous,
2074 * or may have a full name (if the inode was already in the cache).
2076 * When called on a directory inode, we must ensure that the inode only ever
2077 * has one dentry. If a dentry is found, that is returned instead of
2078 * allocating a new one.
2080 * On successful return, the reference to the inode has been transferred
2081 * to the dentry. In case of an error the reference on the inode is released.
2082 * To make it easier to use in export operations a %NULL or IS_ERR inode may
2083 * be passed in and the error will be propagated to the return value,
2084 * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
2086 struct dentry *d_obtain_alias(struct inode *inode)
2088 return __d_obtain_alias(inode, true);
2090 EXPORT_SYMBOL(d_obtain_alias);
2093 * d_obtain_root - find or allocate a dentry for a given inode
2094 * @inode: inode to allocate the dentry for
2096 * Obtain an IS_ROOT dentry for the root of a filesystem.
2098 * We must ensure that directory inodes only ever have one dentry. If a
2099 * dentry is found, that is returned instead of allocating a new one.
2101 * On successful return, the reference to the inode has been transferred
2102 * to the dentry. In case of an error the reference on the inode is
2103 * released. A %NULL or IS_ERR inode may be passed in and will be the
2104 * error will be propagate to the return value, with a %NULL @inode
2105 * replaced by ERR_PTR(-ESTALE).
2107 struct dentry *d_obtain_root(struct inode *inode)
2109 return __d_obtain_alias(inode, false);
2111 EXPORT_SYMBOL(d_obtain_root);
2114 * d_add_ci - lookup or allocate new dentry with case-exact name
2115 * @inode: the inode case-insensitive lookup has found
2116 * @dentry: the negative dentry that was passed to the parent's lookup func
2117 * @name: the case-exact name to be associated with the returned dentry
2119 * This is to avoid filling the dcache with case-insensitive names to the
2120 * same inode, only the actual correct case is stored in the dcache for
2121 * case-insensitive filesystems.
2123 * For a case-insensitive lookup match and if the the case-exact dentry
2124 * already exists in in the dcache, use it and return it.
2126 * If no entry exists with the exact case name, allocate new dentry with
2127 * the exact case, and return the spliced entry.
2129 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2132 struct dentry *found, *res;
2135 * First check if a dentry matching the name already exists,
2136 * if not go ahead and create it now.
2138 found = d_hash_and_lookup(dentry->d_parent, name);
2143 if (d_in_lookup(dentry)) {
2144 found = d_alloc_parallel(dentry->d_parent, name,
2146 if (IS_ERR(found) || !d_in_lookup(found)) {
2151 found = d_alloc(dentry->d_parent, name);
2154 return ERR_PTR(-ENOMEM);
2157 res = d_splice_alias(inode, found);
2164 EXPORT_SYMBOL(d_add_ci);
2167 static inline bool d_same_name(const struct dentry *dentry,
2168 const struct dentry *parent,
2169 const struct qstr *name)
2171 if (likely(!(parent->d_flags & DCACHE_OP_COMPARE))) {
2172 if (dentry->d_name.len != name->len)
2174 return dentry_cmp(dentry, name->name, name->len) == 0;
2176 return parent->d_op->d_compare(dentry,
2177 dentry->d_name.len, dentry->d_name.name,
2182 * __d_lookup_rcu - search for a dentry (racy, store-free)
2183 * @parent: parent dentry
2184 * @name: qstr of name we wish to find
2185 * @seqp: returns d_seq value at the point where the dentry was found
2186 * Returns: dentry, or NULL
2188 * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2189 * resolution (store-free path walking) design described in
2190 * Documentation/filesystems/path-lookup.txt.
2192 * This is not to be used outside core vfs.
2194 * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2195 * held, and rcu_read_lock held. The returned dentry must not be stored into
2196 * without taking d_lock and checking d_seq sequence count against @seq
2199 * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2202 * Alternatively, __d_lookup_rcu may be called again to look up the child of
2203 * the returned dentry, so long as its parent's seqlock is checked after the
2204 * child is looked up. Thus, an interlocking stepping of sequence lock checks
2205 * is formed, giving integrity down the path walk.
2207 * NOTE! The caller *has* to check the resulting dentry against the sequence
2208 * number we've returned before using any of the resulting dentry state!
2210 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2211 const struct qstr *name,
2214 u64 hashlen = name->hash_len;
2215 const unsigned char *str = name->name;
2216 struct hlist_bl_head *b = d_hash(hashlen_hash(hashlen));
2217 struct hlist_bl_node *node;
2218 struct dentry *dentry;
2221 * Note: There is significant duplication with __d_lookup_rcu which is
2222 * required to prevent single threaded performance regressions
2223 * especially on architectures where smp_rmb (in seqcounts) are costly.
2224 * Keep the two functions in sync.
2228 * The hash list is protected using RCU.
2230 * Carefully use d_seq when comparing a candidate dentry, to avoid
2231 * races with d_move().
2233 * It is possible that concurrent renames can mess up our list
2234 * walk here and result in missing our dentry, resulting in the
2235 * false-negative result. d_lookup() protects against concurrent
2236 * renames using rename_lock seqlock.
2238 * See Documentation/filesystems/path-lookup.txt for more details.
2240 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2245 * The dentry sequence count protects us from concurrent
2246 * renames, and thus protects parent and name fields.
2248 * The caller must perform a seqcount check in order
2249 * to do anything useful with the returned dentry.
2251 * NOTE! We do a "raw" seqcount_begin here. That means that
2252 * we don't wait for the sequence count to stabilize if it
2253 * is in the middle of a sequence change. If we do the slow
2254 * dentry compare, we will do seqretries until it is stable,
2255 * and if we end up with a successful lookup, we actually
2256 * want to exit RCU lookup anyway.
2258 * Note that raw_seqcount_begin still *does* smp_rmb(), so
2259 * we are still guaranteed NUL-termination of ->d_name.name.
2261 seq = raw_seqcount_begin(&dentry->d_seq);
2262 if (dentry->d_parent != parent)
2264 if (d_unhashed(dentry))
2267 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) {
2270 if (dentry->d_name.hash != hashlen_hash(hashlen))
2272 tlen = dentry->d_name.len;
2273 tname = dentry->d_name.name;
2274 /* we want a consistent (name,len) pair */
2275 if (read_seqcount_retry(&dentry->d_seq, seq)) {
2279 if (parent->d_op->d_compare(dentry,
2280 tlen, tname, name) != 0)
2283 if (dentry->d_name.hash_len != hashlen)
2285 if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
2295 * d_lookup - search for a dentry
2296 * @parent: parent dentry
2297 * @name: qstr of name we wish to find
2298 * Returns: dentry, or NULL
2300 * d_lookup searches the children of the parent dentry for the name in
2301 * question. If the dentry is found its reference count is incremented and the
2302 * dentry is returned. The caller must use dput to free the entry when it has
2303 * finished using it. %NULL is returned if the dentry does not exist.
2305 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2307 struct dentry *dentry;
2311 seq = read_seqbegin(&rename_lock);
2312 dentry = __d_lookup(parent, name);
2315 } while (read_seqretry(&rename_lock, seq));
2318 EXPORT_SYMBOL(d_lookup);
2321 * __d_lookup - search for a dentry (racy)
2322 * @parent: parent dentry
2323 * @name: qstr of name we wish to find
2324 * Returns: dentry, or NULL
2326 * __d_lookup is like d_lookup, however it may (rarely) return a
2327 * false-negative result due to unrelated rename activity.
2329 * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2330 * however it must be used carefully, eg. with a following d_lookup in
2331 * the case of failure.
2333 * __d_lookup callers must be commented.
2335 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2337 unsigned int hash = name->hash;
2338 struct hlist_bl_head *b = d_hash(hash);
2339 struct hlist_bl_node *node;
2340 struct dentry *found = NULL;
2341 struct dentry *dentry;
2344 * Note: There is significant duplication with __d_lookup_rcu which is
2345 * required to prevent single threaded performance regressions
2346 * especially on architectures where smp_rmb (in seqcounts) are costly.
2347 * Keep the two functions in sync.
2351 * The hash list is protected using RCU.
2353 * Take d_lock when comparing a candidate dentry, to avoid races
2356 * It is possible that concurrent renames can mess up our list
2357 * walk here and result in missing our dentry, resulting in the
2358 * false-negative result. d_lookup() protects against concurrent
2359 * renames using rename_lock seqlock.
2361 * See Documentation/filesystems/path-lookup.txt for more details.
2365 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2367 if (dentry->d_name.hash != hash)
2370 spin_lock(&dentry->d_lock);
2371 if (dentry->d_parent != parent)
2373 if (d_unhashed(dentry))
2376 if (!d_same_name(dentry, parent, name))
2379 dentry->d_lockref.count++;
2381 spin_unlock(&dentry->d_lock);
2384 spin_unlock(&dentry->d_lock);
2392 * d_hash_and_lookup - hash the qstr then search for a dentry
2393 * @dir: Directory to search in
2394 * @name: qstr of name we wish to find
2396 * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2398 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2401 * Check for a fs-specific hash function. Note that we must
2402 * calculate the standard hash first, as the d_op->d_hash()
2403 * routine may choose to leave the hash value unchanged.
2405 name->hash = full_name_hash(dir, name->name, name->len);
2406 if (dir->d_flags & DCACHE_OP_HASH) {
2407 int err = dir->d_op->d_hash(dir, name);
2408 if (unlikely(err < 0))
2409 return ERR_PTR(err);
2411 return d_lookup(dir, name);
2413 EXPORT_SYMBOL(d_hash_and_lookup);
2416 * When a file is deleted, we have two options:
2417 * - turn this dentry into a negative dentry
2418 * - unhash this dentry and free it.
2420 * Usually, we want to just turn this into
2421 * a negative dentry, but if anybody else is
2422 * currently using the dentry or the inode
2423 * we can't do that and we fall back on removing
2424 * it from the hash queues and waiting for
2425 * it to be deleted later when it has no users
2429 * d_delete - delete a dentry
2430 * @dentry: The dentry to delete
2432 * Turn the dentry into a negative dentry if possible, otherwise
2433 * remove it from the hash queues so it can be deleted later
2436 void d_delete(struct dentry * dentry)
2438 struct inode *inode = dentry->d_inode;
2439 int isdir = d_is_dir(dentry);
2441 spin_lock(&inode->i_lock);
2442 spin_lock(&dentry->d_lock);
2444 * Are we the only user?
2446 if (dentry->d_lockref.count == 1) {
2447 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2448 dentry_unlink_inode(dentry);
2451 spin_unlock(&dentry->d_lock);
2452 spin_unlock(&inode->i_lock);
2454 fsnotify_nameremove(dentry, isdir);
2456 EXPORT_SYMBOL(d_delete);
2458 static void __d_rehash(struct dentry *entry)
2460 struct hlist_bl_head *b = d_hash(entry->d_name.hash);
2463 hlist_bl_add_head_rcu(&entry->d_hash, b);
2468 * d_rehash - add an entry back to the hash
2469 * @entry: dentry to add to the hash
2471 * Adds a dentry to the hash according to its name.
2474 void d_rehash(struct dentry * entry)
2476 spin_lock(&entry->d_lock);
2478 spin_unlock(&entry->d_lock);
2480 EXPORT_SYMBOL(d_rehash);
2482 static inline unsigned start_dir_add(struct inode *dir)
2486 unsigned n = dir->i_dir_seq;
2487 if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
2493 static inline void end_dir_add(struct inode *dir, unsigned n)
2495 smp_store_release(&dir->i_dir_seq, n + 2);
2498 static void d_wait_lookup(struct dentry *dentry)
2500 if (d_in_lookup(dentry)) {
2501 DECLARE_WAITQUEUE(wait, current);
2502 add_wait_queue(dentry->d_wait, &wait);
2504 set_current_state(TASK_UNINTERRUPTIBLE);
2505 spin_unlock(&dentry->d_lock);
2507 spin_lock(&dentry->d_lock);
2508 } while (d_in_lookup(dentry));
2512 struct dentry *d_alloc_parallel(struct dentry *parent,
2513 const struct qstr *name,
2514 wait_queue_head_t *wq)
2516 unsigned int hash = name->hash;
2517 struct hlist_bl_head *b = in_lookup_hash(parent, hash);
2518 struct hlist_bl_node *node;
2519 struct dentry *new = d_alloc(parent, name);
2520 struct dentry *dentry;
2521 unsigned seq, r_seq, d_seq;
2524 return ERR_PTR(-ENOMEM);
2528 seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
2529 r_seq = read_seqbegin(&rename_lock);
2530 dentry = __d_lookup_rcu(parent, name, &d_seq);
2531 if (unlikely(dentry)) {
2532 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2536 if (read_seqcount_retry(&dentry->d_seq, d_seq)) {
2545 if (unlikely(read_seqretry(&rename_lock, r_seq))) {
2550 if (unlikely(seq & 1)) {
2556 if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {
2562 * No changes for the parent since the beginning of d_lookup().
2563 * Since all removals from the chain happen with hlist_bl_lock(),
2564 * any potential in-lookup matches are going to stay here until
2565 * we unlock the chain. All fields are stable in everything
2568 hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
2569 if (dentry->d_name.hash != hash)
2571 if (dentry->d_parent != parent)
2573 if (!d_same_name(dentry, parent, name))
2576 /* now we can try to grab a reference */
2577 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2584 * somebody is likely to be still doing lookup for it;
2585 * wait for them to finish
2587 spin_lock(&dentry->d_lock);
2588 d_wait_lookup(dentry);
2590 * it's not in-lookup anymore; in principle we should repeat
2591 * everything from dcache lookup, but it's likely to be what
2592 * d_lookup() would've found anyway. If it is, just return it;
2593 * otherwise we really have to repeat the whole thing.
2595 if (unlikely(dentry->d_name.hash != hash))
2597 if (unlikely(dentry->d_parent != parent))
2599 if (unlikely(d_unhashed(dentry)))
2601 if (unlikely(!d_same_name(dentry, parent, name)))
2603 /* OK, it *is* a hashed match; return it */
2604 spin_unlock(&dentry->d_lock);
2609 /* we can't take ->d_lock here; it's OK, though. */
2610 new->d_flags |= DCACHE_PAR_LOOKUP;
2612 hlist_bl_add_head_rcu(&new->d_u.d_in_lookup_hash, b);
2616 spin_unlock(&dentry->d_lock);
2620 EXPORT_SYMBOL(d_alloc_parallel);
2622 void __d_lookup_done(struct dentry *dentry)
2624 struct hlist_bl_head *b = in_lookup_hash(dentry->d_parent,
2625 dentry->d_name.hash);
2627 dentry->d_flags &= ~DCACHE_PAR_LOOKUP;
2628 __hlist_bl_del(&dentry->d_u.d_in_lookup_hash);
2629 wake_up_all(dentry->d_wait);
2630 dentry->d_wait = NULL;
2632 INIT_HLIST_NODE(&dentry->d_u.d_alias);
2633 INIT_LIST_HEAD(&dentry->d_lru);
2635 EXPORT_SYMBOL(__d_lookup_done);
2637 /* inode->i_lock held if inode is non-NULL */
2639 static inline void __d_add(struct dentry *dentry, struct inode *inode)
2641 struct inode *dir = NULL;
2643 spin_lock(&dentry->d_lock);
2644 if (unlikely(d_in_lookup(dentry))) {
2645 dir = dentry->d_parent->d_inode;
2646 n = start_dir_add(dir);
2647 __d_lookup_done(dentry);
2650 unsigned add_flags = d_flags_for_inode(inode);
2651 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2652 raw_write_seqcount_begin(&dentry->d_seq);
2653 __d_set_inode_and_type(dentry, inode, add_flags);
2654 raw_write_seqcount_end(&dentry->d_seq);
2655 fsnotify_update_flags(dentry);
2659 end_dir_add(dir, n);
2660 spin_unlock(&dentry->d_lock);
2662 spin_unlock(&inode->i_lock);
2666 * d_add - add dentry to hash queues
2667 * @entry: dentry to add
2668 * @inode: The inode to attach to this dentry
2670 * This adds the entry to the hash queues and initializes @inode.
2671 * The entry was actually filled in earlier during d_alloc().
2674 void d_add(struct dentry *entry, struct inode *inode)
2677 security_d_instantiate(entry, inode);
2678 spin_lock(&inode->i_lock);
2680 __d_add(entry, inode);
2682 EXPORT_SYMBOL(d_add);
2685 * d_exact_alias - find and hash an exact unhashed alias
2686 * @entry: dentry to add
2687 * @inode: The inode to go with this dentry
2689 * If an unhashed dentry with the same name/parent and desired
2690 * inode already exists, hash and return it. Otherwise, return
2693 * Parent directory should be locked.
2695 struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2697 struct dentry *alias;
2698 unsigned int hash = entry->d_name.hash;
2700 spin_lock(&inode->i_lock);
2701 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2703 * Don't need alias->d_lock here, because aliases with
2704 * d_parent == entry->d_parent are not subject to name or
2705 * parent changes, because the parent inode i_mutex is held.
2707 if (alias->d_name.hash != hash)
2709 if (alias->d_parent != entry->d_parent)
2711 if (!d_same_name(alias, entry->d_parent, &entry->d_name))
2713 spin_lock(&alias->d_lock);
2714 if (!d_unhashed(alias)) {
2715 spin_unlock(&alias->d_lock);
2718 __dget_dlock(alias);
2720 spin_unlock(&alias->d_lock);
2722 spin_unlock(&inode->i_lock);
2725 spin_unlock(&inode->i_lock);
2728 EXPORT_SYMBOL(d_exact_alias);
2731 * dentry_update_name_case - update case insensitive dentry with a new name
2732 * @dentry: dentry to be updated
2735 * Update a case insensitive dentry with new case of name.
2737 * dentry must have been returned by d_lookup with name @name. Old and new
2738 * name lengths must match (ie. no d_compare which allows mismatched name
2741 * Parent inode i_mutex must be held over d_lookup and into this call (to
2742 * keep renames and concurrent inserts, and readdir(2) away).
2744 void dentry_update_name_case(struct dentry *dentry, const struct qstr *name)
2746 BUG_ON(!inode_is_locked(dentry->d_parent->d_inode));
2747 BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */
2749 spin_lock(&dentry->d_lock);
2750 write_seqcount_begin(&dentry->d_seq);
2751 memcpy((unsigned char *)dentry->d_name.name, name->name, name->len);
2752 write_seqcount_end(&dentry->d_seq);
2753 spin_unlock(&dentry->d_lock);
2755 EXPORT_SYMBOL(dentry_update_name_case);
2757 static void swap_names(struct dentry *dentry, struct dentry *target)
2759 if (unlikely(dname_external(target))) {
2760 if (unlikely(dname_external(dentry))) {
2762 * Both external: swap the pointers
2764 swap(target->d_name.name, dentry->d_name.name);
2767 * dentry:internal, target:external. Steal target's
2768 * storage and make target internal.
2770 memcpy(target->d_iname, dentry->d_name.name,
2771 dentry->d_name.len + 1);
2772 dentry->d_name.name = target->d_name.name;
2773 target->d_name.name = target->d_iname;
2776 if (unlikely(dname_external(dentry))) {
2778 * dentry:external, target:internal. Give dentry's
2779 * storage to target and make dentry internal
2781 memcpy(dentry->d_iname, target->d_name.name,
2782 target->d_name.len + 1);
2783 target->d_name.name = dentry->d_name.name;
2784 dentry->d_name.name = dentry->d_iname;
2787 * Both are internal.
2790 BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2791 for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2792 swap(((long *) &dentry->d_iname)[i],
2793 ((long *) &target->d_iname)[i]);
2797 swap(dentry->d_name.hash_len, target->d_name.hash_len);
2800 static void copy_name(struct dentry *dentry, struct dentry *target)
2802 struct external_name *old_name = NULL;
2803 if (unlikely(dname_external(dentry)))
2804 old_name = external_name(dentry);
2805 if (unlikely(dname_external(target))) {
2806 atomic_inc(&external_name(target)->u.count);
2807 dentry->d_name = target->d_name;
2809 memcpy(dentry->d_iname, target->d_name.name,
2810 target->d_name.len + 1);
2811 dentry->d_name.name = dentry->d_iname;
2812 dentry->d_name.hash_len = target->d_name.hash_len;
2814 if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2815 call_rcu(&old_name->u.head, __d_free_external_name);
2819 * __d_move - move a dentry
2820 * @dentry: entry to move
2821 * @target: new dentry
2822 * @exchange: exchange the two dentries
2824 * Update the dcache to reflect the move of a file name. Negative
2825 * dcache entries should not be moved in this way. Caller must hold
2826 * rename_lock, the i_mutex of the source and target directories,
2827 * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2829 static void __d_move(struct dentry *dentry, struct dentry *target,
2832 struct dentry *old_parent, *p;
2833 struct inode *dir = NULL;
2836 WARN_ON(!dentry->d_inode);
2837 if (WARN_ON(dentry == target))
2840 BUG_ON(d_ancestor(target, dentry));
2841 old_parent = dentry->d_parent;
2842 p = d_ancestor(old_parent, target);
2843 if (IS_ROOT(dentry)) {
2845 spin_lock(&target->d_parent->d_lock);
2847 /* target is not a descendent of dentry->d_parent */
2848 spin_lock(&target->d_parent->d_lock);
2849 spin_lock_nested(&old_parent->d_lock, DENTRY_D_LOCK_NESTED);
2851 BUG_ON(p == dentry);
2852 spin_lock(&old_parent->d_lock);
2854 spin_lock_nested(&target->d_parent->d_lock,
2855 DENTRY_D_LOCK_NESTED);
2857 spin_lock_nested(&dentry->d_lock, 2);
2858 spin_lock_nested(&target->d_lock, 3);
2860 if (unlikely(d_in_lookup(target))) {
2861 dir = target->d_parent->d_inode;
2862 n = start_dir_add(dir);
2863 __d_lookup_done(target);
2866 write_seqcount_begin(&dentry->d_seq);
2867 write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2870 if (!d_unhashed(dentry))
2872 if (!d_unhashed(target))
2875 /* ... and switch them in the tree */
2876 dentry->d_parent = target->d_parent;
2878 copy_name(dentry, target);
2879 target->d_hash.pprev = NULL;
2880 dentry->d_parent->d_lockref.count++;
2881 if (dentry == old_parent)
2882 dentry->d_flags |= DCACHE_RCUACCESS;
2884 WARN_ON(!--old_parent->d_lockref.count);
2886 target->d_parent = old_parent;
2887 swap_names(dentry, target);
2888 list_move(&target->d_child, &target->d_parent->d_subdirs);
2890 fsnotify_update_flags(target);
2892 list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2894 fsnotify_update_flags(dentry);
2896 write_seqcount_end(&target->d_seq);
2897 write_seqcount_end(&dentry->d_seq);
2900 end_dir_add(dir, n);
2902 if (dentry->d_parent != old_parent)
2903 spin_unlock(&dentry->d_parent->d_lock);
2904 if (dentry != old_parent)
2905 spin_unlock(&old_parent->d_lock);
2906 spin_unlock(&target->d_lock);
2907 spin_unlock(&dentry->d_lock);
2911 * d_move - move a dentry
2912 * @dentry: entry to move
2913 * @target: new dentry
2915 * Update the dcache to reflect the move of a file name. Negative
2916 * dcache entries should not be moved in this way. See the locking
2917 * requirements for __d_move.
2919 void d_move(struct dentry *dentry, struct dentry *target)
2921 write_seqlock(&rename_lock);
2922 __d_move(dentry, target, false);
2923 write_sequnlock(&rename_lock);
2925 EXPORT_SYMBOL(d_move);
2928 * d_exchange - exchange two dentries
2929 * @dentry1: first dentry
2930 * @dentry2: second dentry
2932 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2934 write_seqlock(&rename_lock);
2936 WARN_ON(!dentry1->d_inode);
2937 WARN_ON(!dentry2->d_inode);
2938 WARN_ON(IS_ROOT(dentry1));
2939 WARN_ON(IS_ROOT(dentry2));
2941 __d_move(dentry1, dentry2, true);
2943 write_sequnlock(&rename_lock);
2947 * d_ancestor - search for an ancestor
2948 * @p1: ancestor dentry
2951 * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2952 * an ancestor of p2, else NULL.
2954 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2958 for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2959 if (p->d_parent == p1)
2966 * This helper attempts to cope with remotely renamed directories
2968 * It assumes that the caller is already holding
2969 * dentry->d_parent->d_inode->i_mutex, and rename_lock
2971 * Note: If ever the locking in lock_rename() changes, then please
2972 * remember to update this too...
2974 static int __d_unalias(struct inode *inode,
2975 struct dentry *dentry, struct dentry *alias)
2977 struct mutex *m1 = NULL;
2978 struct rw_semaphore *m2 = NULL;
2981 /* If alias and dentry share a parent, then no extra locks required */
2982 if (alias->d_parent == dentry->d_parent)
2985 /* See lock_rename() */
2986 if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2988 m1 = &dentry->d_sb->s_vfs_rename_mutex;
2989 if (!inode_trylock_shared(alias->d_parent->d_inode))
2991 m2 = &alias->d_parent->d_inode->i_rwsem;
2993 __d_move(alias, dentry, false);
3004 * d_splice_alias - splice a disconnected dentry into the tree if one exists
3005 * @inode: the inode which may have a disconnected dentry
3006 * @dentry: a negative dentry which we want to point to the inode.
3008 * If inode is a directory and has an IS_ROOT alias, then d_move that in
3009 * place of the given dentry and return it, else simply d_add the inode
3010 * to the dentry and return NULL.
3012 * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
3013 * we should error out: directories can't have multiple aliases.
3015 * This is needed in the lookup routine of any filesystem that is exportable
3016 * (via knfsd) so that we can build dcache paths to directories effectively.
3018 * If a dentry was found and moved, then it is returned. Otherwise NULL
3019 * is returned. This matches the expected return value of ->lookup.
3021 * Cluster filesystems may call this function with a negative, hashed dentry.
3022 * In that case, we know that the inode will be a regular file, and also this
3023 * will only occur during atomic_open. So we need to check for the dentry
3024 * being already hashed only in the final case.
3026 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
3029 return ERR_CAST(inode);
3031 BUG_ON(!d_unhashed(dentry));
3036 security_d_instantiate(dentry, inode);
3037 spin_lock(&inode->i_lock);
3038 if (S_ISDIR(inode->i_mode)) {
3039 struct dentry *new = __d_find_any_alias(inode);
3040 if (unlikely(new)) {
3041 /* The reference to new ensures it remains an alias */
3042 spin_unlock(&inode->i_lock);
3043 write_seqlock(&rename_lock);
3044 if (unlikely(d_ancestor(new, dentry))) {
3045 write_sequnlock(&rename_lock);
3047 new = ERR_PTR(-ELOOP);
3048 pr_warn_ratelimited(
3049 "VFS: Lookup of '%s' in %s %s"
3050 " would have caused loop\n",
3051 dentry->d_name.name,
3052 inode->i_sb->s_type->name,
3054 } else if (!IS_ROOT(new)) {
3055 struct dentry *old_parent = dget(new->d_parent);
3056 int err = __d_unalias(inode, dentry, new);
3057 write_sequnlock(&rename_lock);
3064 __d_move(new, dentry, false);
3065 write_sequnlock(&rename_lock);
3072 __d_add(dentry, inode);
3075 EXPORT_SYMBOL(d_splice_alias);
3078 * Test whether new_dentry is a subdirectory of old_dentry.
3080 * Trivially implemented using the dcache structure
3084 * is_subdir - is new dentry a subdirectory of old_dentry
3085 * @new_dentry: new dentry
3086 * @old_dentry: old dentry
3088 * Returns true if new_dentry is a subdirectory of the parent (at any depth).
3089 * Returns false otherwise.
3090 * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3093 bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3098 if (new_dentry == old_dentry)
3102 /* for restarting inner loop in case of seq retry */
3103 seq = read_seqbegin(&rename_lock);
3105 * Need rcu_readlock to protect against the d_parent trashing
3109 if (d_ancestor(old_dentry, new_dentry))
3114 } while (read_seqretry(&rename_lock, seq));
3118 EXPORT_SYMBOL(is_subdir);
3120 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3122 struct dentry *root = data;
3123 if (dentry != root) {
3124 if (d_unhashed(dentry) || !dentry->d_inode)
3127 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3128 dentry->d_flags |= DCACHE_GENOCIDE;
3129 dentry->d_lockref.count--;
3132 return D_WALK_CONTINUE;
3135 void d_genocide(struct dentry *parent)
3137 d_walk(parent, parent, d_genocide_kill, NULL);
3140 EXPORT_SYMBOL(d_genocide);
3142 void d_tmpfile(struct dentry *dentry, struct inode *inode)
3144 inode_dec_link_count(inode);
3145 BUG_ON(dentry->d_name.name != dentry->d_iname ||
3146 !hlist_unhashed(&dentry->d_u.d_alias) ||
3147 !d_unlinked(dentry));
3148 spin_lock(&dentry->d_parent->d_lock);
3149 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3150 dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3151 (unsigned long long)inode->i_ino);
3152 spin_unlock(&dentry->d_lock);
3153 spin_unlock(&dentry->d_parent->d_lock);
3154 d_instantiate(dentry, inode);
3156 EXPORT_SYMBOL(d_tmpfile);
3158 static __initdata unsigned long dhash_entries;
3159 static int __init set_dhash_entries(char *str)
3163 dhash_entries = simple_strtoul(str, &str, 0);
3166 __setup("dhash_entries=", set_dhash_entries);
3168 static void __init dcache_init_early(void)
3170 /* If hashes are distributed across NUMA nodes, defer
3171 * hash allocation until vmalloc space is available.
3177 alloc_large_system_hash("Dentry cache",
3178 sizeof(struct hlist_bl_head),
3181 HASH_EARLY | HASH_ZERO,
3186 d_hash_shift = 32 - d_hash_shift;
3189 static void __init dcache_init(void)
3192 * A constructor could be added for stable state like the lists,
3193 * but it is probably not worth it because of the cache nature
3196 dentry_cache = KMEM_CACHE_USERCOPY(dentry,
3197 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD|SLAB_ACCOUNT,
3200 /* Hash may have been set up in dcache_init_early */
3205 alloc_large_system_hash("Dentry cache",
3206 sizeof(struct hlist_bl_head),
3214 d_hash_shift = 32 - d_hash_shift;
3217 /* SLAB cache for __getname() consumers */
3218 struct kmem_cache *names_cachep __read_mostly;
3219 EXPORT_SYMBOL(names_cachep);
3221 void __init vfs_caches_init_early(void)
3225 for (i = 0; i < ARRAY_SIZE(in_lookup_hashtable); i++)
3226 INIT_HLIST_BL_HEAD(&in_lookup_hashtable[i]);
3228 dcache_init_early();
3232 void __init vfs_caches_init(void)
3234 names_cachep = kmem_cache_create_usercopy("names_cache", PATH_MAX, 0,
3235 SLAB_HWCACHE_ALIGN|SLAB_PANIC, 0, PATH_MAX, NULL);
3240 files_maxfiles_init();