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/fscrypt.h>
22 #include <linux/fsnotify.h>
23 #include <linux/slab.h>
24 #include <linux/init.h>
25 #include <linux/hash.h>
26 #include <linux/cache.h>
27 #include <linux/export.h>
28 #include <linux/security.h>
29 #include <linux/seqlock.h>
30 #include <linux/memblock.h>
31 #include <linux/bit_spinlock.h>
32 #include <linux/rculist_bl.h>
33 #include <linux/list_lru.h>
39 * dcache->d_inode->i_lock protects:
40 * - i_dentry, d_u.d_alias, d_inode of aliases
41 * dcache_hash_bucket lock protects:
42 * - the dcache hash table
43 * s_roots bl list spinlock protects:
44 * - the s_roots list (see __d_drop)
45 * dentry->d_sb->s_dentry_lru_lock protects:
46 * - the dcache lru lists and counters
53 * - d_parent and d_subdirs
54 * - childrens' d_child and d_parent
55 * - d_u.d_alias, d_inode
58 * dentry->d_inode->i_lock
60 * dentry->d_sb->s_dentry_lru_lock
61 * dcache_hash_bucket lock
64 * If there is an ancestor relationship:
65 * dentry->d_parent->...->d_parent->d_lock
67 * dentry->d_parent->d_lock
70 * If no ancestor relationship:
71 * arbitrary, since it's serialized on rename_lock
73 int sysctl_vfs_cache_pressure __read_mostly = 100;
74 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
76 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
78 EXPORT_SYMBOL(rename_lock);
80 static struct kmem_cache *dentry_cache __read_mostly;
82 const struct qstr empty_name = QSTR_INIT("", 0);
83 EXPORT_SYMBOL(empty_name);
84 const struct qstr slash_name = QSTR_INIT("/", 1);
85 EXPORT_SYMBOL(slash_name);
88 * This is the single most critical data structure when it comes
89 * to the dcache: the hashtable for lookups. Somebody should try
90 * to make this good - I've just made it work.
92 * This hash-function tries to avoid losing too many bits of hash
93 * information, yet avoid using a prime hash-size or similar.
96 static unsigned int d_hash_shift __read_mostly;
98 static struct hlist_bl_head *dentry_hashtable __read_mostly;
100 static inline struct hlist_bl_head *d_hash(unsigned int hash)
102 return dentry_hashtable + (hash >> d_hash_shift);
105 #define IN_LOOKUP_SHIFT 10
106 static struct hlist_bl_head in_lookup_hashtable[1 << IN_LOOKUP_SHIFT];
108 static inline struct hlist_bl_head *in_lookup_hash(const struct dentry *parent,
111 hash += (unsigned long) parent / L1_CACHE_BYTES;
112 return in_lookup_hashtable + hash_32(hash, IN_LOOKUP_SHIFT);
116 /* Statistics gathering. */
117 struct dentry_stat_t dentry_stat = {
121 static DEFINE_PER_CPU(long, nr_dentry);
122 static DEFINE_PER_CPU(long, nr_dentry_unused);
123 static DEFINE_PER_CPU(long, nr_dentry_negative);
125 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
128 * Here we resort to our own counters instead of using generic per-cpu counters
129 * for consistency with what the vfs inode code does. We are expected to harvest
130 * better code and performance by having our own specialized counters.
132 * Please note that the loop is done over all possible CPUs, not over all online
133 * CPUs. The reason for this is that we don't want to play games with CPUs going
134 * on and off. If one of them goes off, we will just keep their counters.
136 * glommer: See cffbc8a for details, and if you ever intend to change this,
137 * please update all vfs counters to match.
139 static long get_nr_dentry(void)
143 for_each_possible_cpu(i)
144 sum += per_cpu(nr_dentry, i);
145 return sum < 0 ? 0 : sum;
148 static long get_nr_dentry_unused(void)
152 for_each_possible_cpu(i)
153 sum += per_cpu(nr_dentry_unused, i);
154 return sum < 0 ? 0 : sum;
157 static long get_nr_dentry_negative(void)
162 for_each_possible_cpu(i)
163 sum += per_cpu(nr_dentry_negative, i);
164 return sum < 0 ? 0 : sum;
167 int proc_nr_dentry(struct ctl_table *table, int write, void __user *buffer,
168 size_t *lenp, loff_t *ppos)
170 dentry_stat.nr_dentry = get_nr_dentry();
171 dentry_stat.nr_unused = get_nr_dentry_unused();
172 dentry_stat.nr_negative = get_nr_dentry_negative();
173 return proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
178 * Compare 2 name strings, return 0 if they match, otherwise non-zero.
179 * The strings are both count bytes long, and count is non-zero.
181 #ifdef CONFIG_DCACHE_WORD_ACCESS
183 #include <asm/word-at-a-time.h>
185 * NOTE! 'cs' and 'scount' come from a dentry, so it has a
186 * aligned allocation for this particular component. We don't
187 * strictly need the load_unaligned_zeropad() safety, but it
188 * doesn't hurt either.
190 * In contrast, 'ct' and 'tcount' can be from a pathname, and do
191 * need the careful unaligned handling.
193 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
195 unsigned long a,b,mask;
198 a = read_word_at_a_time(cs);
199 b = load_unaligned_zeropad(ct);
200 if (tcount < sizeof(unsigned long))
202 if (unlikely(a != b))
204 cs += sizeof(unsigned long);
205 ct += sizeof(unsigned long);
206 tcount -= sizeof(unsigned long);
210 mask = bytemask_from_count(tcount);
211 return unlikely(!!((a ^ b) & mask));
216 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
230 static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount)
233 * Be careful about RCU walk racing with rename:
234 * use 'READ_ONCE' to fetch the name pointer.
236 * NOTE! Even if a rename will mean that the length
237 * was not loaded atomically, we don't care. The
238 * RCU walk will check the sequence count eventually,
239 * and catch it. And we won't overrun the buffer,
240 * because we're reading the name pointer atomically,
241 * and a dentry name is guaranteed to be properly
242 * terminated with a NUL byte.
244 * End result: even if 'len' is wrong, we'll exit
245 * early because the data cannot match (there can
246 * be no NUL in the ct/tcount data)
248 const unsigned char *cs = READ_ONCE(dentry->d_name.name);
250 return dentry_string_cmp(cs, ct, tcount);
253 struct external_name {
256 struct rcu_head head;
258 unsigned char name[];
261 static inline struct external_name *external_name(struct dentry *dentry)
263 return container_of(dentry->d_name.name, struct external_name, name[0]);
266 static void __d_free(struct rcu_head *head)
268 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
270 kmem_cache_free(dentry_cache, dentry);
273 static void __d_free_external(struct rcu_head *head)
275 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
276 kfree(external_name(dentry));
277 kmem_cache_free(dentry_cache, dentry);
280 static inline int dname_external(const struct dentry *dentry)
282 return dentry->d_name.name != dentry->d_iname;
285 void take_dentry_name_snapshot(struct name_snapshot *name, struct dentry *dentry)
287 spin_lock(&dentry->d_lock);
288 name->name = dentry->d_name;
289 if (unlikely(dname_external(dentry))) {
290 atomic_inc(&external_name(dentry)->u.count);
292 memcpy(name->inline_name, dentry->d_iname,
293 dentry->d_name.len + 1);
294 name->name.name = name->inline_name;
296 spin_unlock(&dentry->d_lock);
298 EXPORT_SYMBOL(take_dentry_name_snapshot);
300 void release_dentry_name_snapshot(struct name_snapshot *name)
302 if (unlikely(name->name.name != name->inline_name)) {
303 struct external_name *p;
304 p = container_of(name->name.name, struct external_name, name[0]);
305 if (unlikely(atomic_dec_and_test(&p->u.count)))
306 kfree_rcu(p, u.head);
309 EXPORT_SYMBOL(release_dentry_name_snapshot);
311 static inline void __d_set_inode_and_type(struct dentry *dentry,
317 dentry->d_inode = inode;
318 flags = READ_ONCE(dentry->d_flags);
319 flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
321 WRITE_ONCE(dentry->d_flags, flags);
324 static inline void __d_clear_type_and_inode(struct dentry *dentry)
326 unsigned flags = READ_ONCE(dentry->d_flags);
328 flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
329 WRITE_ONCE(dentry->d_flags, flags);
330 dentry->d_inode = NULL;
331 if (dentry->d_flags & DCACHE_LRU_LIST)
332 this_cpu_inc(nr_dentry_negative);
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_NORCU)
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;
362 raw_write_seqcount_begin(&dentry->d_seq);
363 __d_clear_type_and_inode(dentry);
364 hlist_del_init(&dentry->d_u.d_alias);
365 raw_write_seqcount_end(&dentry->d_seq);
366 spin_unlock(&dentry->d_lock);
367 spin_unlock(&inode->i_lock);
369 fsnotify_inoderemove(inode);
370 if (dentry->d_op && dentry->d_op->d_iput)
371 dentry->d_op->d_iput(dentry, inode);
377 * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
378 * is in use - which includes both the "real" per-superblock
379 * LRU list _and_ the DCACHE_SHRINK_LIST use.
381 * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
382 * on the shrink list (ie not on the superblock LRU list).
384 * The per-cpu "nr_dentry_unused" counters are updated with
385 * the DCACHE_LRU_LIST bit.
387 * The per-cpu "nr_dentry_negative" counters are only updated
388 * when deleted from or added to the per-superblock LRU list, not
389 * from/to the shrink list. That is to avoid an unneeded dec/inc
390 * pair when moving from LRU to shrink list in select_collect().
392 * These helper functions make sure we always follow the
393 * rules. d_lock must be held by the caller.
395 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
396 static void d_lru_add(struct dentry *dentry)
398 D_FLAG_VERIFY(dentry, 0);
399 dentry->d_flags |= DCACHE_LRU_LIST;
400 this_cpu_inc(nr_dentry_unused);
401 if (d_is_negative(dentry))
402 this_cpu_inc(nr_dentry_negative);
403 WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
406 static void d_lru_del(struct dentry *dentry)
408 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
409 dentry->d_flags &= ~DCACHE_LRU_LIST;
410 this_cpu_dec(nr_dentry_unused);
411 if (d_is_negative(dentry))
412 this_cpu_dec(nr_dentry_negative);
413 WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
416 static void d_shrink_del(struct dentry *dentry)
418 D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
419 list_del_init(&dentry->d_lru);
420 dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
421 this_cpu_dec(nr_dentry_unused);
424 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
426 D_FLAG_VERIFY(dentry, 0);
427 list_add(&dentry->d_lru, list);
428 dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
429 this_cpu_inc(nr_dentry_unused);
433 * These can only be called under the global LRU lock, ie during the
434 * callback for freeing the LRU list. "isolate" removes it from the
435 * LRU lists entirely, while shrink_move moves it to the indicated
438 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
440 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
441 dentry->d_flags &= ~DCACHE_LRU_LIST;
442 this_cpu_dec(nr_dentry_unused);
443 if (d_is_negative(dentry))
444 this_cpu_dec(nr_dentry_negative);
445 list_lru_isolate(lru, &dentry->d_lru);
448 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
449 struct list_head *list)
451 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
452 dentry->d_flags |= DCACHE_SHRINK_LIST;
453 if (d_is_negative(dentry))
454 this_cpu_dec(nr_dentry_negative);
455 list_lru_isolate_move(lru, &dentry->d_lru, list);
459 * d_drop - drop a dentry
460 * @dentry: dentry to drop
462 * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
463 * be found through a VFS lookup any more. Note that this is different from
464 * deleting the dentry - d_delete will try to mark the dentry negative if
465 * possible, giving a successful _negative_ lookup, while d_drop will
466 * just make the cache lookup fail.
468 * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
469 * reason (NFS timeouts or autofs deletes).
471 * __d_drop requires dentry->d_lock
472 * ___d_drop doesn't mark dentry as "unhashed"
473 * (dentry->d_hash.pprev will be LIST_POISON2, not NULL).
475 static void ___d_drop(struct dentry *dentry)
477 struct hlist_bl_head *b;
479 * Hashed dentries are normally on the dentry hashtable,
480 * with the exception of those newly allocated by
481 * d_obtain_root, which are always IS_ROOT:
483 if (unlikely(IS_ROOT(dentry)))
484 b = &dentry->d_sb->s_roots;
486 b = d_hash(dentry->d_name.hash);
489 __hlist_bl_del(&dentry->d_hash);
493 void __d_drop(struct dentry *dentry)
495 if (!d_unhashed(dentry)) {
497 dentry->d_hash.pprev = NULL;
498 write_seqcount_invalidate(&dentry->d_seq);
501 EXPORT_SYMBOL(__d_drop);
503 void d_drop(struct dentry *dentry)
505 spin_lock(&dentry->d_lock);
507 spin_unlock(&dentry->d_lock);
509 EXPORT_SYMBOL(d_drop);
511 static inline void dentry_unlist(struct dentry *dentry, struct dentry *parent)
515 * Inform d_walk() and shrink_dentry_list() that we are no longer
516 * attached to the dentry tree
518 dentry->d_flags |= DCACHE_DENTRY_KILLED;
519 if (unlikely(list_empty(&dentry->d_child)))
521 __list_del_entry(&dentry->d_child);
523 * Cursors can move around the list of children. While we'd been
524 * a normal list member, it didn't matter - ->d_child.next would've
525 * been updated. However, from now on it won't be and for the
526 * things like d_walk() it might end up with a nasty surprise.
527 * Normally d_walk() doesn't care about cursors moving around -
528 * ->d_lock on parent prevents that and since a cursor has no children
529 * of its own, we get through it without ever unlocking the parent.
530 * There is one exception, though - if we ascend from a child that
531 * gets killed as soon as we unlock it, the next sibling is found
532 * using the value left in its ->d_child.next. And if _that_
533 * pointed to a cursor, and cursor got moved (e.g. by lseek())
534 * before d_walk() regains parent->d_lock, we'll end up skipping
535 * everything the cursor had been moved past.
537 * Solution: make sure that the pointer left behind in ->d_child.next
538 * points to something that won't be moving around. I.e. skip the
541 while (dentry->d_child.next != &parent->d_subdirs) {
542 next = list_entry(dentry->d_child.next, struct dentry, d_child);
543 if (likely(!(next->d_flags & DCACHE_DENTRY_CURSOR)))
545 dentry->d_child.next = next->d_child.next;
549 static void __dentry_kill(struct dentry *dentry)
551 struct dentry *parent = NULL;
552 bool can_free = true;
553 if (!IS_ROOT(dentry))
554 parent = dentry->d_parent;
557 * The dentry is now unrecoverably dead to the world.
559 lockref_mark_dead(&dentry->d_lockref);
562 * inform the fs via d_prune that this dentry is about to be
563 * unhashed and destroyed.
565 if (dentry->d_flags & DCACHE_OP_PRUNE)
566 dentry->d_op->d_prune(dentry);
568 if (dentry->d_flags & DCACHE_LRU_LIST) {
569 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
572 /* if it was on the hash then remove it */
574 dentry_unlist(dentry, parent);
576 spin_unlock(&parent->d_lock);
578 dentry_unlink_inode(dentry);
580 spin_unlock(&dentry->d_lock);
581 this_cpu_dec(nr_dentry);
582 if (dentry->d_op && dentry->d_op->d_release)
583 dentry->d_op->d_release(dentry);
585 spin_lock(&dentry->d_lock);
586 if (dentry->d_flags & DCACHE_SHRINK_LIST) {
587 dentry->d_flags |= DCACHE_MAY_FREE;
590 spin_unlock(&dentry->d_lock);
591 if (likely(can_free))
596 static struct dentry *__lock_parent(struct dentry *dentry)
598 struct dentry *parent;
600 spin_unlock(&dentry->d_lock);
602 parent = READ_ONCE(dentry->d_parent);
603 spin_lock(&parent->d_lock);
605 * We can't blindly lock dentry until we are sure
606 * that we won't violate the locking order.
607 * Any changes of dentry->d_parent must have
608 * been done with parent->d_lock held, so
609 * spin_lock() above is enough of a barrier
610 * for checking if it's still our child.
612 if (unlikely(parent != dentry->d_parent)) {
613 spin_unlock(&parent->d_lock);
617 if (parent != dentry)
618 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
624 static inline struct dentry *lock_parent(struct dentry *dentry)
626 struct dentry *parent = dentry->d_parent;
629 if (likely(spin_trylock(&parent->d_lock)))
631 return __lock_parent(dentry);
634 static inline bool retain_dentry(struct dentry *dentry)
636 WARN_ON(d_in_lookup(dentry));
638 /* Unreachable? Get rid of it */
639 if (unlikely(d_unhashed(dentry)))
642 if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
645 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
646 if (dentry->d_op->d_delete(dentry))
649 /* retain; LRU fodder */
650 dentry->d_lockref.count--;
651 if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
653 else if (unlikely(!(dentry->d_flags & DCACHE_REFERENCED)))
654 dentry->d_flags |= DCACHE_REFERENCED;
659 * Finish off a dentry we've decided to kill.
660 * dentry->d_lock must be held, returns with it unlocked.
661 * Returns dentry requiring refcount drop, or NULL if we're done.
663 static struct dentry *dentry_kill(struct dentry *dentry)
664 __releases(dentry->d_lock)
666 struct inode *inode = dentry->d_inode;
667 struct dentry *parent = NULL;
669 if (inode && unlikely(!spin_trylock(&inode->i_lock)))
672 if (!IS_ROOT(dentry)) {
673 parent = dentry->d_parent;
674 if (unlikely(!spin_trylock(&parent->d_lock))) {
675 parent = __lock_parent(dentry);
676 if (likely(inode || !dentry->d_inode))
678 /* negative that became positive */
680 spin_unlock(&parent->d_lock);
681 inode = dentry->d_inode;
685 __dentry_kill(dentry);
689 spin_unlock(&dentry->d_lock);
690 spin_lock(&inode->i_lock);
691 spin_lock(&dentry->d_lock);
692 parent = lock_parent(dentry);
694 if (unlikely(dentry->d_lockref.count != 1)) {
695 dentry->d_lockref.count--;
696 } else if (likely(!retain_dentry(dentry))) {
697 __dentry_kill(dentry);
700 /* we are keeping it, after all */
702 spin_unlock(&inode->i_lock);
704 spin_unlock(&parent->d_lock);
705 spin_unlock(&dentry->d_lock);
710 * Try to do a lockless dput(), and return whether that was successful.
712 * If unsuccessful, we return false, having already taken the dentry lock.
714 * The caller needs to hold the RCU read lock, so that the dentry is
715 * guaranteed to stay around even if the refcount goes down to zero!
717 static inline bool fast_dput(struct dentry *dentry)
720 unsigned int d_flags;
723 * If we have a d_op->d_delete() operation, we sould not
724 * let the dentry count go to zero, so use "put_or_lock".
726 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE))
727 return lockref_put_or_lock(&dentry->d_lockref);
730 * .. otherwise, we can try to just decrement the
731 * lockref optimistically.
733 ret = lockref_put_return(&dentry->d_lockref);
736 * If the lockref_put_return() failed due to the lock being held
737 * by somebody else, the fast path has failed. We will need to
738 * get the lock, and then check the count again.
740 if (unlikely(ret < 0)) {
741 spin_lock(&dentry->d_lock);
742 if (dentry->d_lockref.count > 1) {
743 dentry->d_lockref.count--;
744 spin_unlock(&dentry->d_lock);
751 * If we weren't the last ref, we're done.
757 * Careful, careful. The reference count went down
758 * to zero, but we don't hold the dentry lock, so
759 * somebody else could get it again, and do another
760 * dput(), and we need to not race with that.
762 * However, there is a very special and common case
763 * where we don't care, because there is nothing to
764 * do: the dentry is still hashed, it does not have
765 * a 'delete' op, and it's referenced and already on
768 * NOTE! Since we aren't locked, these values are
769 * not "stable". However, it is sufficient that at
770 * some point after we dropped the reference the
771 * dentry was hashed and the flags had the proper
772 * value. Other dentry users may have re-gotten
773 * a reference to the dentry and change that, but
774 * our work is done - we can leave the dentry
775 * around with a zero refcount.
778 d_flags = READ_ONCE(dentry->d_flags);
779 d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST | DCACHE_DISCONNECTED;
781 /* Nothing to do? Dropping the reference was all we needed? */
782 if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry))
786 * Not the fast normal case? Get the lock. We've already decremented
787 * the refcount, but we'll need to re-check the situation after
790 spin_lock(&dentry->d_lock);
793 * Did somebody else grab a reference to it in the meantime, and
794 * we're no longer the last user after all? Alternatively, somebody
795 * else could have killed it and marked it dead. Either way, we
796 * don't need to do anything else.
798 if (dentry->d_lockref.count) {
799 spin_unlock(&dentry->d_lock);
804 * Re-get the reference we optimistically dropped. We hold the
805 * lock, and we just tested that it was zero, so we can just
808 dentry->d_lockref.count = 1;
816 * This is complicated by the fact that we do not want to put
817 * dentries that are no longer on any hash chain on the unused
818 * list: we'd much rather just get rid of them immediately.
820 * However, that implies that we have to traverse the dentry
821 * tree upwards to the parents which might _also_ now be
822 * scheduled for deletion (it may have been only waiting for
823 * its last child to go away).
825 * This tail recursion is done by hand as we don't want to depend
826 * on the compiler to always get this right (gcc generally doesn't).
827 * Real recursion would eat up our stack space.
831 * dput - release a dentry
832 * @dentry: dentry to release
834 * Release a dentry. This will drop the usage count and if appropriate
835 * call the dentry unlink method as well as removing it from the queues and
836 * releasing its resources. If the parent dentries were scheduled for release
837 * they too may now get deleted.
839 void dput(struct dentry *dentry)
845 if (likely(fast_dput(dentry))) {
850 /* Slow case: now with the dentry lock held */
853 if (likely(retain_dentry(dentry))) {
854 spin_unlock(&dentry->d_lock);
858 dentry = dentry_kill(dentry);
864 /* This must be called with d_lock held */
865 static inline void __dget_dlock(struct dentry *dentry)
867 dentry->d_lockref.count++;
870 static inline void __dget(struct dentry *dentry)
872 lockref_get(&dentry->d_lockref);
875 struct dentry *dget_parent(struct dentry *dentry)
881 * Do optimistic parent lookup without any
885 ret = READ_ONCE(dentry->d_parent);
886 gotref = lockref_get_not_zero(&ret->d_lockref);
888 if (likely(gotref)) {
889 if (likely(ret == READ_ONCE(dentry->d_parent)))
896 * Don't need rcu_dereference because we re-check it was correct under
900 ret = dentry->d_parent;
901 spin_lock(&ret->d_lock);
902 if (unlikely(ret != dentry->d_parent)) {
903 spin_unlock(&ret->d_lock);
908 BUG_ON(!ret->d_lockref.count);
909 ret->d_lockref.count++;
910 spin_unlock(&ret->d_lock);
913 EXPORT_SYMBOL(dget_parent);
915 static struct dentry * __d_find_any_alias(struct inode *inode)
917 struct dentry *alias;
919 if (hlist_empty(&inode->i_dentry))
921 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
927 * d_find_any_alias - find any alias for a given inode
928 * @inode: inode to find an alias for
930 * If any aliases exist for the given inode, take and return a
931 * reference for one of them. If no aliases exist, return %NULL.
933 struct dentry *d_find_any_alias(struct inode *inode)
937 spin_lock(&inode->i_lock);
938 de = __d_find_any_alias(inode);
939 spin_unlock(&inode->i_lock);
942 EXPORT_SYMBOL(d_find_any_alias);
945 * d_find_alias - grab a hashed alias of inode
946 * @inode: inode in question
948 * If inode has a hashed alias, or is a directory and has any alias,
949 * acquire the reference to alias and return it. Otherwise return NULL.
950 * Notice that if inode is a directory there can be only one alias and
951 * it can be unhashed only if it has no children, or if it is the root
952 * of a filesystem, or if the directory was renamed and d_revalidate
953 * was the first vfs operation to notice.
955 * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
956 * any other hashed alias over that one.
958 static struct dentry *__d_find_alias(struct inode *inode)
960 struct dentry *alias;
962 if (S_ISDIR(inode->i_mode))
963 return __d_find_any_alias(inode);
965 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
966 spin_lock(&alias->d_lock);
967 if (!d_unhashed(alias)) {
969 spin_unlock(&alias->d_lock);
972 spin_unlock(&alias->d_lock);
977 struct dentry *d_find_alias(struct inode *inode)
979 struct dentry *de = NULL;
981 if (!hlist_empty(&inode->i_dentry)) {
982 spin_lock(&inode->i_lock);
983 de = __d_find_alias(inode);
984 spin_unlock(&inode->i_lock);
988 EXPORT_SYMBOL(d_find_alias);
991 * Try to kill dentries associated with this inode.
992 * WARNING: you must own a reference to inode.
994 void d_prune_aliases(struct inode *inode)
996 struct dentry *dentry;
998 spin_lock(&inode->i_lock);
999 hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
1000 spin_lock(&dentry->d_lock);
1001 if (!dentry->d_lockref.count) {
1002 struct dentry *parent = lock_parent(dentry);
1003 if (likely(!dentry->d_lockref.count)) {
1004 __dentry_kill(dentry);
1009 spin_unlock(&parent->d_lock);
1011 spin_unlock(&dentry->d_lock);
1013 spin_unlock(&inode->i_lock);
1015 EXPORT_SYMBOL(d_prune_aliases);
1018 * Lock a dentry from shrink list.
1019 * Called under rcu_read_lock() and dentry->d_lock; the former
1020 * guarantees that nothing we access will be freed under us.
1021 * Note that dentry is *not* protected from concurrent dentry_kill(),
1024 * Return false if dentry has been disrupted or grabbed, leaving
1025 * the caller to kick it off-list. Otherwise, return true and have
1026 * that dentry's inode and parent both locked.
1028 static bool shrink_lock_dentry(struct dentry *dentry)
1030 struct inode *inode;
1031 struct dentry *parent;
1033 if (dentry->d_lockref.count)
1036 inode = dentry->d_inode;
1037 if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
1038 spin_unlock(&dentry->d_lock);
1039 spin_lock(&inode->i_lock);
1040 spin_lock(&dentry->d_lock);
1041 if (unlikely(dentry->d_lockref.count))
1043 /* changed inode means that somebody had grabbed it */
1044 if (unlikely(inode != dentry->d_inode))
1048 parent = dentry->d_parent;
1049 if (IS_ROOT(dentry) || likely(spin_trylock(&parent->d_lock)))
1052 spin_unlock(&dentry->d_lock);
1053 spin_lock(&parent->d_lock);
1054 if (unlikely(parent != dentry->d_parent)) {
1055 spin_unlock(&parent->d_lock);
1056 spin_lock(&dentry->d_lock);
1059 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1060 if (likely(!dentry->d_lockref.count))
1062 spin_unlock(&parent->d_lock);
1065 spin_unlock(&inode->i_lock);
1069 static void shrink_dentry_list(struct list_head *list)
1071 while (!list_empty(list)) {
1072 struct dentry *dentry, *parent;
1074 dentry = list_entry(list->prev, struct dentry, d_lru);
1075 spin_lock(&dentry->d_lock);
1077 if (!shrink_lock_dentry(dentry)) {
1078 bool can_free = false;
1080 d_shrink_del(dentry);
1081 if (dentry->d_lockref.count < 0)
1082 can_free = dentry->d_flags & DCACHE_MAY_FREE;
1083 spin_unlock(&dentry->d_lock);
1085 dentry_free(dentry);
1089 d_shrink_del(dentry);
1090 parent = dentry->d_parent;
1091 __dentry_kill(dentry);
1092 if (parent == dentry)
1095 * We need to prune ancestors too. This is necessary to prevent
1096 * quadratic behavior of shrink_dcache_parent(), but is also
1097 * expected to be beneficial in reducing dentry cache
1101 while (dentry && !lockref_put_or_lock(&dentry->d_lockref))
1102 dentry = dentry_kill(dentry);
1106 static enum lru_status dentry_lru_isolate(struct list_head *item,
1107 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1109 struct list_head *freeable = arg;
1110 struct dentry *dentry = container_of(item, struct dentry, d_lru);
1114 * we are inverting the lru lock/dentry->d_lock here,
1115 * so use a trylock. If we fail to get the lock, just skip
1118 if (!spin_trylock(&dentry->d_lock))
1122 * Referenced dentries are still in use. If they have active
1123 * counts, just remove them from the LRU. Otherwise give them
1124 * another pass through the LRU.
1126 if (dentry->d_lockref.count) {
1127 d_lru_isolate(lru, dentry);
1128 spin_unlock(&dentry->d_lock);
1132 if (dentry->d_flags & DCACHE_REFERENCED) {
1133 dentry->d_flags &= ~DCACHE_REFERENCED;
1134 spin_unlock(&dentry->d_lock);
1137 * The list move itself will be made by the common LRU code. At
1138 * this point, we've dropped the dentry->d_lock but keep the
1139 * lru lock. This is safe to do, since every list movement is
1140 * protected by the lru lock even if both locks are held.
1142 * This is guaranteed by the fact that all LRU management
1143 * functions are intermediated by the LRU API calls like
1144 * list_lru_add and list_lru_del. List movement in this file
1145 * only ever occur through this functions or through callbacks
1146 * like this one, that are called from the LRU API.
1148 * The only exceptions to this are functions like
1149 * shrink_dentry_list, and code that first checks for the
1150 * DCACHE_SHRINK_LIST flag. Those are guaranteed to be
1151 * operating only with stack provided lists after they are
1152 * properly isolated from the main list. It is thus, always a
1158 d_lru_shrink_move(lru, dentry, freeable);
1159 spin_unlock(&dentry->d_lock);
1165 * prune_dcache_sb - shrink the dcache
1167 * @sc: shrink control, passed to list_lru_shrink_walk()
1169 * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1170 * is done when we need more memory and called from the superblock shrinker
1173 * This function may fail to free any resources if all the dentries are in
1176 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1181 freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1182 dentry_lru_isolate, &dispose);
1183 shrink_dentry_list(&dispose);
1187 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1188 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1190 struct list_head *freeable = arg;
1191 struct dentry *dentry = container_of(item, struct dentry, d_lru);
1194 * we are inverting the lru lock/dentry->d_lock here,
1195 * so use a trylock. If we fail to get the lock, just skip
1198 if (!spin_trylock(&dentry->d_lock))
1201 d_lru_shrink_move(lru, dentry, freeable);
1202 spin_unlock(&dentry->d_lock);
1209 * shrink_dcache_sb - shrink dcache for a superblock
1212 * Shrink the dcache for the specified super block. This is used to free
1213 * the dcache before unmounting a file system.
1215 void shrink_dcache_sb(struct super_block *sb)
1220 list_lru_walk(&sb->s_dentry_lru,
1221 dentry_lru_isolate_shrink, &dispose, 1024);
1222 shrink_dentry_list(&dispose);
1223 } while (list_lru_count(&sb->s_dentry_lru) > 0);
1225 EXPORT_SYMBOL(shrink_dcache_sb);
1228 * enum d_walk_ret - action to talke during tree walk
1229 * @D_WALK_CONTINUE: contrinue walk
1230 * @D_WALK_QUIT: quit walk
1231 * @D_WALK_NORETRY: quit when retry is needed
1232 * @D_WALK_SKIP: skip this dentry and its children
1242 * d_walk - walk the dentry tree
1243 * @parent: start of walk
1244 * @data: data passed to @enter() and @finish()
1245 * @enter: callback when first entering the dentry
1247 * The @enter() callbacks are called with d_lock held.
1249 static void d_walk(struct dentry *parent, void *data,
1250 enum d_walk_ret (*enter)(void *, struct dentry *))
1252 struct dentry *this_parent;
1253 struct list_head *next;
1255 enum d_walk_ret ret;
1259 read_seqbegin_or_lock(&rename_lock, &seq);
1260 this_parent = parent;
1261 spin_lock(&this_parent->d_lock);
1263 ret = enter(data, this_parent);
1265 case D_WALK_CONTINUE:
1270 case D_WALK_NORETRY:
1275 next = this_parent->d_subdirs.next;
1277 while (next != &this_parent->d_subdirs) {
1278 struct list_head *tmp = next;
1279 struct dentry *dentry = list_entry(tmp, struct dentry, d_child);
1282 if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
1285 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1287 ret = enter(data, dentry);
1289 case D_WALK_CONTINUE:
1292 spin_unlock(&dentry->d_lock);
1294 case D_WALK_NORETRY:
1298 spin_unlock(&dentry->d_lock);
1302 if (!list_empty(&dentry->d_subdirs)) {
1303 spin_unlock(&this_parent->d_lock);
1304 spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1305 this_parent = dentry;
1306 spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1309 spin_unlock(&dentry->d_lock);
1312 * All done at this level ... ascend and resume the search.
1316 if (this_parent != parent) {
1317 struct dentry *child = this_parent;
1318 this_parent = child->d_parent;
1320 spin_unlock(&child->d_lock);
1321 spin_lock(&this_parent->d_lock);
1323 /* might go back up the wrong parent if we have had a rename. */
1324 if (need_seqretry(&rename_lock, seq))
1326 /* go into the first sibling still alive */
1328 next = child->d_child.next;
1329 if (next == &this_parent->d_subdirs)
1331 child = list_entry(next, struct dentry, d_child);
1332 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED));
1336 if (need_seqretry(&rename_lock, seq))
1341 spin_unlock(&this_parent->d_lock);
1342 done_seqretry(&rename_lock, seq);
1346 spin_unlock(&this_parent->d_lock);
1355 struct check_mount {
1356 struct vfsmount *mnt;
1357 unsigned int mounted;
1360 static enum d_walk_ret path_check_mount(void *data, struct dentry *dentry)
1362 struct check_mount *info = data;
1363 struct path path = { .mnt = info->mnt, .dentry = dentry };
1365 if (likely(!d_mountpoint(dentry)))
1366 return D_WALK_CONTINUE;
1367 if (__path_is_mountpoint(&path)) {
1371 return D_WALK_CONTINUE;
1375 * path_has_submounts - check for mounts over a dentry in the
1376 * current namespace.
1377 * @parent: path to check.
1379 * Return true if the parent or its subdirectories contain
1380 * a mount point in the current namespace.
1382 int path_has_submounts(const struct path *parent)
1384 struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };
1386 read_seqlock_excl(&mount_lock);
1387 d_walk(parent->dentry, &data, path_check_mount);
1388 read_sequnlock_excl(&mount_lock);
1390 return data.mounted;
1392 EXPORT_SYMBOL(path_has_submounts);
1395 * Called by mount code to set a mountpoint and check if the mountpoint is
1396 * reachable (e.g. NFS can unhash a directory dentry and then the complete
1397 * subtree can become unreachable).
1399 * Only one of d_invalidate() and d_set_mounted() must succeed. For
1400 * this reason take rename_lock and d_lock on dentry and ancestors.
1402 int d_set_mounted(struct dentry *dentry)
1406 write_seqlock(&rename_lock);
1407 for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1408 /* Need exclusion wrt. d_invalidate() */
1409 spin_lock(&p->d_lock);
1410 if (unlikely(d_unhashed(p))) {
1411 spin_unlock(&p->d_lock);
1414 spin_unlock(&p->d_lock);
1416 spin_lock(&dentry->d_lock);
1417 if (!d_unlinked(dentry)) {
1419 if (!d_mountpoint(dentry)) {
1420 dentry->d_flags |= DCACHE_MOUNTED;
1424 spin_unlock(&dentry->d_lock);
1426 write_sequnlock(&rename_lock);
1431 * Search the dentry child list of the specified parent,
1432 * and move any unused dentries to the end of the unused
1433 * list for prune_dcache(). We descend to the next level
1434 * whenever the d_subdirs list is non-empty and continue
1437 * It returns zero iff there are no unused children,
1438 * otherwise it returns the number of children moved to
1439 * the end of the unused list. This may not be the total
1440 * number of unused children, because select_parent can
1441 * drop the lock and return early due to latency
1445 struct select_data {
1446 struct dentry *start;
1447 struct list_head dispose;
1451 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1453 struct select_data *data = _data;
1454 enum d_walk_ret ret = D_WALK_CONTINUE;
1456 if (data->start == dentry)
1459 if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1462 if (dentry->d_flags & DCACHE_LRU_LIST)
1464 if (!dentry->d_lockref.count) {
1465 d_shrink_add(dentry, &data->dispose);
1470 * We can return to the caller if we have found some (this
1471 * ensures forward progress). We'll be coming back to find
1474 if (!list_empty(&data->dispose))
1475 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1481 * shrink_dcache_parent - prune dcache
1482 * @parent: parent of entries to prune
1484 * Prune the dcache to remove unused children of the parent dentry.
1486 void shrink_dcache_parent(struct dentry *parent)
1489 struct select_data data;
1491 INIT_LIST_HEAD(&data.dispose);
1492 data.start = parent;
1495 d_walk(parent, &data, select_collect);
1497 if (!list_empty(&data.dispose)) {
1498 shrink_dentry_list(&data.dispose);
1507 EXPORT_SYMBOL(shrink_dcache_parent);
1509 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1511 /* it has busy descendents; complain about those instead */
1512 if (!list_empty(&dentry->d_subdirs))
1513 return D_WALK_CONTINUE;
1515 /* root with refcount 1 is fine */
1516 if (dentry == _data && dentry->d_lockref.count == 1)
1517 return D_WALK_CONTINUE;
1519 printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} "
1520 " still in use (%d) [unmount of %s %s]\n",
1523 dentry->d_inode->i_ino : 0UL,
1525 dentry->d_lockref.count,
1526 dentry->d_sb->s_type->name,
1527 dentry->d_sb->s_id);
1529 return D_WALK_CONTINUE;
1532 static void do_one_tree(struct dentry *dentry)
1534 shrink_dcache_parent(dentry);
1535 d_walk(dentry, dentry, umount_check);
1541 * destroy the dentries attached to a superblock on unmounting
1543 void shrink_dcache_for_umount(struct super_block *sb)
1545 struct dentry *dentry;
1547 WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
1549 dentry = sb->s_root;
1551 do_one_tree(dentry);
1553 while (!hlist_bl_empty(&sb->s_roots)) {
1554 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_roots), struct dentry, d_hash));
1555 do_one_tree(dentry);
1559 static enum d_walk_ret find_submount(void *_data, struct dentry *dentry)
1561 struct dentry **victim = _data;
1562 if (d_mountpoint(dentry)) {
1563 __dget_dlock(dentry);
1567 return D_WALK_CONTINUE;
1571 * d_invalidate - detach submounts, prune dcache, and drop
1572 * @dentry: dentry to invalidate (aka detach, prune and drop)
1574 void d_invalidate(struct dentry *dentry)
1576 bool had_submounts = false;
1577 spin_lock(&dentry->d_lock);
1578 if (d_unhashed(dentry)) {
1579 spin_unlock(&dentry->d_lock);
1583 spin_unlock(&dentry->d_lock);
1585 /* Negative dentries can be dropped without further checks */
1586 if (!dentry->d_inode)
1589 shrink_dcache_parent(dentry);
1591 struct dentry *victim = NULL;
1592 d_walk(dentry, &victim, find_submount);
1595 shrink_dcache_parent(dentry);
1598 had_submounts = true;
1599 detach_mounts(victim);
1603 EXPORT_SYMBOL(d_invalidate);
1606 * __d_alloc - allocate a dcache entry
1607 * @sb: filesystem it will belong to
1608 * @name: qstr of the name
1610 * Allocates a dentry. It returns %NULL if there is insufficient memory
1611 * available. On a success the dentry is returned. The name passed in is
1612 * copied and the copy passed in may be reused after this call.
1615 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1617 struct dentry *dentry;
1621 dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1626 * We guarantee that the inline name is always NUL-terminated.
1627 * This way the memcpy() done by the name switching in rename
1628 * will still always have a NUL at the end, even if we might
1629 * be overwriting an internal NUL character
1631 dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1632 if (unlikely(!name)) {
1634 dname = dentry->d_iname;
1635 } else if (name->len > DNAME_INLINE_LEN-1) {
1636 size_t size = offsetof(struct external_name, name[1]);
1637 struct external_name *p = kmalloc(size + name->len,
1638 GFP_KERNEL_ACCOUNT |
1641 kmem_cache_free(dentry_cache, dentry);
1644 atomic_set(&p->u.count, 1);
1647 dname = dentry->d_iname;
1650 dentry->d_name.len = name->len;
1651 dentry->d_name.hash = name->hash;
1652 memcpy(dname, name->name, name->len);
1653 dname[name->len] = 0;
1655 /* Make sure we always see the terminating NUL character */
1656 smp_store_release(&dentry->d_name.name, dname); /* ^^^ */
1658 dentry->d_lockref.count = 1;
1659 dentry->d_flags = 0;
1660 spin_lock_init(&dentry->d_lock);
1661 seqcount_init(&dentry->d_seq);
1662 dentry->d_inode = NULL;
1663 dentry->d_parent = dentry;
1665 dentry->d_op = NULL;
1666 dentry->d_fsdata = NULL;
1667 INIT_HLIST_BL_NODE(&dentry->d_hash);
1668 INIT_LIST_HEAD(&dentry->d_lru);
1669 INIT_LIST_HEAD(&dentry->d_subdirs);
1670 INIT_HLIST_NODE(&dentry->d_u.d_alias);
1671 INIT_LIST_HEAD(&dentry->d_child);
1672 d_set_d_op(dentry, dentry->d_sb->s_d_op);
1674 if (dentry->d_op && dentry->d_op->d_init) {
1675 err = dentry->d_op->d_init(dentry);
1677 if (dname_external(dentry))
1678 kfree(external_name(dentry));
1679 kmem_cache_free(dentry_cache, dentry);
1684 this_cpu_inc(nr_dentry);
1690 * d_alloc - allocate a dcache entry
1691 * @parent: parent of entry to allocate
1692 * @name: qstr of the name
1694 * Allocates a dentry. It returns %NULL if there is insufficient memory
1695 * available. On a success the dentry is returned. The name passed in is
1696 * copied and the copy passed in may be reused after this call.
1698 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1700 struct dentry *dentry = __d_alloc(parent->d_sb, name);
1703 spin_lock(&parent->d_lock);
1705 * don't need child lock because it is not subject
1706 * to concurrency here
1708 __dget_dlock(parent);
1709 dentry->d_parent = parent;
1710 list_add(&dentry->d_child, &parent->d_subdirs);
1711 spin_unlock(&parent->d_lock);
1715 EXPORT_SYMBOL(d_alloc);
1717 struct dentry *d_alloc_anon(struct super_block *sb)
1719 return __d_alloc(sb, NULL);
1721 EXPORT_SYMBOL(d_alloc_anon);
1723 struct dentry *d_alloc_cursor(struct dentry * parent)
1725 struct dentry *dentry = d_alloc_anon(parent->d_sb);
1727 dentry->d_flags |= DCACHE_DENTRY_CURSOR;
1728 dentry->d_parent = dget(parent);
1734 * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1735 * @sb: the superblock
1736 * @name: qstr of the name
1738 * For a filesystem that just pins its dentries in memory and never
1739 * performs lookups at all, return an unhashed IS_ROOT dentry.
1740 * This is used for pipes, sockets et.al. - the stuff that should
1741 * never be anyone's children or parents. Unlike all other
1742 * dentries, these will not have RCU delay between dropping the
1743 * last reference and freeing them.
1745 * The only user is alloc_file_pseudo() and that's what should
1746 * be considered a public interface. Don't use directly.
1748 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1750 struct dentry *dentry = __d_alloc(sb, name);
1752 dentry->d_flags |= DCACHE_NORCU;
1756 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1761 q.hash_len = hashlen_string(parent, name);
1762 return d_alloc(parent, &q);
1764 EXPORT_SYMBOL(d_alloc_name);
1766 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1768 WARN_ON_ONCE(dentry->d_op);
1769 WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH |
1771 DCACHE_OP_REVALIDATE |
1772 DCACHE_OP_WEAK_REVALIDATE |
1779 dentry->d_flags |= DCACHE_OP_HASH;
1781 dentry->d_flags |= DCACHE_OP_COMPARE;
1782 if (op->d_revalidate)
1783 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1784 if (op->d_weak_revalidate)
1785 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1787 dentry->d_flags |= DCACHE_OP_DELETE;
1789 dentry->d_flags |= DCACHE_OP_PRUNE;
1791 dentry->d_flags |= DCACHE_OP_REAL;
1794 EXPORT_SYMBOL(d_set_d_op);
1798 * d_set_fallthru - Mark a dentry as falling through to a lower layer
1799 * @dentry - The dentry to mark
1801 * Mark a dentry as falling through to the lower layer (as set with
1802 * d_pin_lower()). This flag may be recorded on the medium.
1804 void d_set_fallthru(struct dentry *dentry)
1806 spin_lock(&dentry->d_lock);
1807 dentry->d_flags |= DCACHE_FALLTHRU;
1808 spin_unlock(&dentry->d_lock);
1810 EXPORT_SYMBOL(d_set_fallthru);
1812 static unsigned d_flags_for_inode(struct inode *inode)
1814 unsigned add_flags = DCACHE_REGULAR_TYPE;
1817 return DCACHE_MISS_TYPE;
1819 if (S_ISDIR(inode->i_mode)) {
1820 add_flags = DCACHE_DIRECTORY_TYPE;
1821 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1822 if (unlikely(!inode->i_op->lookup))
1823 add_flags = DCACHE_AUTODIR_TYPE;
1825 inode->i_opflags |= IOP_LOOKUP;
1827 goto type_determined;
1830 if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1831 if (unlikely(inode->i_op->get_link)) {
1832 add_flags = DCACHE_SYMLINK_TYPE;
1833 goto type_determined;
1835 inode->i_opflags |= IOP_NOFOLLOW;
1838 if (unlikely(!S_ISREG(inode->i_mode)))
1839 add_flags = DCACHE_SPECIAL_TYPE;
1842 if (unlikely(IS_AUTOMOUNT(inode)))
1843 add_flags |= DCACHE_NEED_AUTOMOUNT;
1847 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1849 unsigned add_flags = d_flags_for_inode(inode);
1850 WARN_ON(d_in_lookup(dentry));
1852 spin_lock(&dentry->d_lock);
1854 * Decrement negative dentry count if it was in the LRU list.
1856 if (dentry->d_flags & DCACHE_LRU_LIST)
1857 this_cpu_dec(nr_dentry_negative);
1858 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1859 raw_write_seqcount_begin(&dentry->d_seq);
1860 __d_set_inode_and_type(dentry, inode, add_flags);
1861 raw_write_seqcount_end(&dentry->d_seq);
1862 fsnotify_update_flags(dentry);
1863 spin_unlock(&dentry->d_lock);
1867 * d_instantiate - fill in inode information for a dentry
1868 * @entry: dentry to complete
1869 * @inode: inode to attach to this dentry
1871 * Fill in inode information in the entry.
1873 * This turns negative dentries into productive full members
1876 * NOTE! This assumes that the inode count has been incremented
1877 * (or otherwise set) by the caller to indicate that it is now
1878 * in use by the dcache.
1881 void d_instantiate(struct dentry *entry, struct inode * inode)
1883 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1885 security_d_instantiate(entry, inode);
1886 spin_lock(&inode->i_lock);
1887 __d_instantiate(entry, inode);
1888 spin_unlock(&inode->i_lock);
1891 EXPORT_SYMBOL(d_instantiate);
1894 * This should be equivalent to d_instantiate() + unlock_new_inode(),
1895 * with lockdep-related part of unlock_new_inode() done before
1896 * anything else. Use that instead of open-coding d_instantiate()/
1897 * unlock_new_inode() combinations.
1899 void d_instantiate_new(struct dentry *entry, struct inode *inode)
1901 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1903 lockdep_annotate_inode_mutex_key(inode);
1904 security_d_instantiate(entry, inode);
1905 spin_lock(&inode->i_lock);
1906 __d_instantiate(entry, inode);
1907 WARN_ON(!(inode->i_state & I_NEW));
1908 inode->i_state &= ~I_NEW & ~I_CREATING;
1910 wake_up_bit(&inode->i_state, __I_NEW);
1911 spin_unlock(&inode->i_lock);
1913 EXPORT_SYMBOL(d_instantiate_new);
1915 struct dentry *d_make_root(struct inode *root_inode)
1917 struct dentry *res = NULL;
1920 res = d_alloc_anon(root_inode->i_sb);
1922 d_instantiate(res, root_inode);
1928 EXPORT_SYMBOL(d_make_root);
1930 static struct dentry *__d_instantiate_anon(struct dentry *dentry,
1931 struct inode *inode,
1937 security_d_instantiate(dentry, inode);
1938 spin_lock(&inode->i_lock);
1939 res = __d_find_any_alias(inode);
1941 spin_unlock(&inode->i_lock);
1946 /* attach a disconnected dentry */
1947 add_flags = d_flags_for_inode(inode);
1950 add_flags |= DCACHE_DISCONNECTED;
1952 spin_lock(&dentry->d_lock);
1953 __d_set_inode_and_type(dentry, inode, add_flags);
1954 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1955 if (!disconnected) {
1956 hlist_bl_lock(&dentry->d_sb->s_roots);
1957 hlist_bl_add_head(&dentry->d_hash, &dentry->d_sb->s_roots);
1958 hlist_bl_unlock(&dentry->d_sb->s_roots);
1960 spin_unlock(&dentry->d_lock);
1961 spin_unlock(&inode->i_lock);
1970 struct dentry *d_instantiate_anon(struct dentry *dentry, struct inode *inode)
1972 return __d_instantiate_anon(dentry, inode, true);
1974 EXPORT_SYMBOL(d_instantiate_anon);
1976 static struct dentry *__d_obtain_alias(struct inode *inode, bool disconnected)
1982 return ERR_PTR(-ESTALE);
1984 return ERR_CAST(inode);
1986 res = d_find_any_alias(inode);
1990 tmp = d_alloc_anon(inode->i_sb);
1992 res = ERR_PTR(-ENOMEM);
1996 return __d_instantiate_anon(tmp, inode, disconnected);
2004 * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
2005 * @inode: inode to allocate the dentry for
2007 * Obtain a dentry for an inode resulting from NFS filehandle conversion or
2008 * similar open by handle operations. The returned dentry may be anonymous,
2009 * or may have a full name (if the inode was already in the cache).
2011 * When called on a directory inode, we must ensure that the inode only ever
2012 * has one dentry. If a dentry is found, that is returned instead of
2013 * allocating a new one.
2015 * On successful return, the reference to the inode has been transferred
2016 * to the dentry. In case of an error the reference on the inode is released.
2017 * To make it easier to use in export operations a %NULL or IS_ERR inode may
2018 * be passed in and the error will be propagated to the return value,
2019 * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
2021 struct dentry *d_obtain_alias(struct inode *inode)
2023 return __d_obtain_alias(inode, true);
2025 EXPORT_SYMBOL(d_obtain_alias);
2028 * d_obtain_root - find or allocate a dentry for a given inode
2029 * @inode: inode to allocate the dentry for
2031 * Obtain an IS_ROOT dentry for the root of a filesystem.
2033 * We must ensure that directory inodes only ever have one dentry. If a
2034 * dentry is found, that is returned instead of allocating a new one.
2036 * On successful return, the reference to the inode has been transferred
2037 * to the dentry. In case of an error the reference on the inode is
2038 * released. A %NULL or IS_ERR inode may be passed in and will be the
2039 * error will be propagate to the return value, with a %NULL @inode
2040 * replaced by ERR_PTR(-ESTALE).
2042 struct dentry *d_obtain_root(struct inode *inode)
2044 return __d_obtain_alias(inode, false);
2046 EXPORT_SYMBOL(d_obtain_root);
2049 * d_add_ci - lookup or allocate new dentry with case-exact name
2050 * @inode: the inode case-insensitive lookup has found
2051 * @dentry: the negative dentry that was passed to the parent's lookup func
2052 * @name: the case-exact name to be associated with the returned dentry
2054 * This is to avoid filling the dcache with case-insensitive names to the
2055 * same inode, only the actual correct case is stored in the dcache for
2056 * case-insensitive filesystems.
2058 * For a case-insensitive lookup match and if the the case-exact dentry
2059 * already exists in in the dcache, use it and return it.
2061 * If no entry exists with the exact case name, allocate new dentry with
2062 * the exact case, and return the spliced entry.
2064 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2067 struct dentry *found, *res;
2070 * First check if a dentry matching the name already exists,
2071 * if not go ahead and create it now.
2073 found = d_hash_and_lookup(dentry->d_parent, name);
2078 if (d_in_lookup(dentry)) {
2079 found = d_alloc_parallel(dentry->d_parent, name,
2081 if (IS_ERR(found) || !d_in_lookup(found)) {
2086 found = d_alloc(dentry->d_parent, name);
2089 return ERR_PTR(-ENOMEM);
2092 res = d_splice_alias(inode, found);
2099 EXPORT_SYMBOL(d_add_ci);
2102 static inline bool d_same_name(const struct dentry *dentry,
2103 const struct dentry *parent,
2104 const struct qstr *name)
2106 if (likely(!(parent->d_flags & DCACHE_OP_COMPARE))) {
2107 if (dentry->d_name.len != name->len)
2109 return dentry_cmp(dentry, name->name, name->len) == 0;
2111 return parent->d_op->d_compare(dentry,
2112 dentry->d_name.len, dentry->d_name.name,
2117 * __d_lookup_rcu - search for a dentry (racy, store-free)
2118 * @parent: parent dentry
2119 * @name: qstr of name we wish to find
2120 * @seqp: returns d_seq value at the point where the dentry was found
2121 * Returns: dentry, or NULL
2123 * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2124 * resolution (store-free path walking) design described in
2125 * Documentation/filesystems/path-lookup.txt.
2127 * This is not to be used outside core vfs.
2129 * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2130 * held, and rcu_read_lock held. The returned dentry must not be stored into
2131 * without taking d_lock and checking d_seq sequence count against @seq
2134 * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2137 * Alternatively, __d_lookup_rcu may be called again to look up the child of
2138 * the returned dentry, so long as its parent's seqlock is checked after the
2139 * child is looked up. Thus, an interlocking stepping of sequence lock checks
2140 * is formed, giving integrity down the path walk.
2142 * NOTE! The caller *has* to check the resulting dentry against the sequence
2143 * number we've returned before using any of the resulting dentry state!
2145 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2146 const struct qstr *name,
2149 u64 hashlen = name->hash_len;
2150 const unsigned char *str = name->name;
2151 struct hlist_bl_head *b = d_hash(hashlen_hash(hashlen));
2152 struct hlist_bl_node *node;
2153 struct dentry *dentry;
2156 * Note: There is significant duplication with __d_lookup_rcu which is
2157 * required to prevent single threaded performance regressions
2158 * especially on architectures where smp_rmb (in seqcounts) are costly.
2159 * Keep the two functions in sync.
2163 * The hash list is protected using RCU.
2165 * Carefully use d_seq when comparing a candidate dentry, to avoid
2166 * races with d_move().
2168 * It is possible that concurrent renames can mess up our list
2169 * walk here and result in missing our dentry, resulting in the
2170 * false-negative result. d_lookup() protects against concurrent
2171 * renames using rename_lock seqlock.
2173 * See Documentation/filesystems/path-lookup.txt for more details.
2175 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2180 * The dentry sequence count protects us from concurrent
2181 * renames, and thus protects parent and name fields.
2183 * The caller must perform a seqcount check in order
2184 * to do anything useful with the returned dentry.
2186 * NOTE! We do a "raw" seqcount_begin here. That means that
2187 * we don't wait for the sequence count to stabilize if it
2188 * is in the middle of a sequence change. If we do the slow
2189 * dentry compare, we will do seqretries until it is stable,
2190 * and if we end up with a successful lookup, we actually
2191 * want to exit RCU lookup anyway.
2193 * Note that raw_seqcount_begin still *does* smp_rmb(), so
2194 * we are still guaranteed NUL-termination of ->d_name.name.
2196 seq = raw_seqcount_begin(&dentry->d_seq);
2197 if (dentry->d_parent != parent)
2199 if (d_unhashed(dentry))
2202 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) {
2205 if (dentry->d_name.hash != hashlen_hash(hashlen))
2207 tlen = dentry->d_name.len;
2208 tname = dentry->d_name.name;
2209 /* we want a consistent (name,len) pair */
2210 if (read_seqcount_retry(&dentry->d_seq, seq)) {
2214 if (parent->d_op->d_compare(dentry,
2215 tlen, tname, name) != 0)
2218 if (dentry->d_name.hash_len != hashlen)
2220 if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
2230 * d_lookup - search for a dentry
2231 * @parent: parent dentry
2232 * @name: qstr of name we wish to find
2233 * Returns: dentry, or NULL
2235 * d_lookup searches the children of the parent dentry for the name in
2236 * question. If the dentry is found its reference count is incremented and the
2237 * dentry is returned. The caller must use dput to free the entry when it has
2238 * finished using it. %NULL is returned if the dentry does not exist.
2240 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2242 struct dentry *dentry;
2246 seq = read_seqbegin(&rename_lock);
2247 dentry = __d_lookup(parent, name);
2250 } while (read_seqretry(&rename_lock, seq));
2253 EXPORT_SYMBOL(d_lookup);
2256 * __d_lookup - search for a dentry (racy)
2257 * @parent: parent dentry
2258 * @name: qstr of name we wish to find
2259 * Returns: dentry, or NULL
2261 * __d_lookup is like d_lookup, however it may (rarely) return a
2262 * false-negative result due to unrelated rename activity.
2264 * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2265 * however it must be used carefully, eg. with a following d_lookup in
2266 * the case of failure.
2268 * __d_lookup callers must be commented.
2270 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2272 unsigned int hash = name->hash;
2273 struct hlist_bl_head *b = d_hash(hash);
2274 struct hlist_bl_node *node;
2275 struct dentry *found = NULL;
2276 struct dentry *dentry;
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.
2286 * The hash list is protected using RCU.
2288 * Take d_lock when comparing a candidate dentry, to avoid races
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.
2296 * See Documentation/filesystems/path-lookup.txt for more details.
2300 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2302 if (dentry->d_name.hash != hash)
2305 spin_lock(&dentry->d_lock);
2306 if (dentry->d_parent != parent)
2308 if (d_unhashed(dentry))
2311 if (!d_same_name(dentry, parent, name))
2314 dentry->d_lockref.count++;
2316 spin_unlock(&dentry->d_lock);
2319 spin_unlock(&dentry->d_lock);
2327 * d_hash_and_lookup - hash the qstr then search for a dentry
2328 * @dir: Directory to search in
2329 * @name: qstr of name we wish to find
2331 * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2333 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2336 * Check for a fs-specific hash function. Note that we must
2337 * calculate the standard hash first, as the d_op->d_hash()
2338 * routine may choose to leave the hash value unchanged.
2340 name->hash = full_name_hash(dir, name->name, name->len);
2341 if (dir->d_flags & DCACHE_OP_HASH) {
2342 int err = dir->d_op->d_hash(dir, name);
2343 if (unlikely(err < 0))
2344 return ERR_PTR(err);
2346 return d_lookup(dir, name);
2348 EXPORT_SYMBOL(d_hash_and_lookup);
2351 * When a file is deleted, we have two options:
2352 * - turn this dentry into a negative dentry
2353 * - unhash this dentry and free it.
2355 * Usually, we want to just turn this into
2356 * a negative dentry, but if anybody else is
2357 * currently using the dentry or the inode
2358 * we can't do that and we fall back on removing
2359 * it from the hash queues and waiting for
2360 * it to be deleted later when it has no users
2364 * d_delete - delete a dentry
2365 * @dentry: The dentry to delete
2367 * Turn the dentry into a negative dentry if possible, otherwise
2368 * remove it from the hash queues so it can be deleted later
2371 void d_delete(struct dentry * dentry)
2373 struct inode *inode = dentry->d_inode;
2374 int isdir = d_is_dir(dentry);
2376 spin_lock(&inode->i_lock);
2377 spin_lock(&dentry->d_lock);
2379 * Are we the only user?
2381 if (dentry->d_lockref.count == 1) {
2382 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2383 dentry_unlink_inode(dentry);
2386 spin_unlock(&dentry->d_lock);
2387 spin_unlock(&inode->i_lock);
2389 fsnotify_nameremove(dentry, isdir);
2391 EXPORT_SYMBOL(d_delete);
2393 static void __d_rehash(struct dentry *entry)
2395 struct hlist_bl_head *b = d_hash(entry->d_name.hash);
2398 hlist_bl_add_head_rcu(&entry->d_hash, b);
2403 * d_rehash - add an entry back to the hash
2404 * @entry: dentry to add to the hash
2406 * Adds a dentry to the hash according to its name.
2409 void d_rehash(struct dentry * entry)
2411 spin_lock(&entry->d_lock);
2413 spin_unlock(&entry->d_lock);
2415 EXPORT_SYMBOL(d_rehash);
2417 static inline unsigned start_dir_add(struct inode *dir)
2421 unsigned n = dir->i_dir_seq;
2422 if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
2428 static inline void end_dir_add(struct inode *dir, unsigned n)
2430 smp_store_release(&dir->i_dir_seq, n + 2);
2433 static void d_wait_lookup(struct dentry *dentry)
2435 if (d_in_lookup(dentry)) {
2436 DECLARE_WAITQUEUE(wait, current);
2437 add_wait_queue(dentry->d_wait, &wait);
2439 set_current_state(TASK_UNINTERRUPTIBLE);
2440 spin_unlock(&dentry->d_lock);
2442 spin_lock(&dentry->d_lock);
2443 } while (d_in_lookup(dentry));
2447 struct dentry *d_alloc_parallel(struct dentry *parent,
2448 const struct qstr *name,
2449 wait_queue_head_t *wq)
2451 unsigned int hash = name->hash;
2452 struct hlist_bl_head *b = in_lookup_hash(parent, hash);
2453 struct hlist_bl_node *node;
2454 struct dentry *new = d_alloc(parent, name);
2455 struct dentry *dentry;
2456 unsigned seq, r_seq, d_seq;
2459 return ERR_PTR(-ENOMEM);
2463 seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
2464 r_seq = read_seqbegin(&rename_lock);
2465 dentry = __d_lookup_rcu(parent, name, &d_seq);
2466 if (unlikely(dentry)) {
2467 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2471 if (read_seqcount_retry(&dentry->d_seq, d_seq)) {
2480 if (unlikely(read_seqretry(&rename_lock, r_seq))) {
2485 if (unlikely(seq & 1)) {
2491 if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {
2497 * No changes for the parent since the beginning of d_lookup().
2498 * Since all removals from the chain happen with hlist_bl_lock(),
2499 * any potential in-lookup matches are going to stay here until
2500 * we unlock the chain. All fields are stable in everything
2503 hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
2504 if (dentry->d_name.hash != hash)
2506 if (dentry->d_parent != parent)
2508 if (!d_same_name(dentry, parent, name))
2511 /* now we can try to grab a reference */
2512 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2519 * somebody is likely to be still doing lookup for it;
2520 * wait for them to finish
2522 spin_lock(&dentry->d_lock);
2523 d_wait_lookup(dentry);
2525 * it's not in-lookup anymore; in principle we should repeat
2526 * everything from dcache lookup, but it's likely to be what
2527 * d_lookup() would've found anyway. If it is, just return it;
2528 * otherwise we really have to repeat the whole thing.
2530 if (unlikely(dentry->d_name.hash != hash))
2532 if (unlikely(dentry->d_parent != parent))
2534 if (unlikely(d_unhashed(dentry)))
2536 if (unlikely(!d_same_name(dentry, parent, name)))
2538 /* OK, it *is* a hashed match; return it */
2539 spin_unlock(&dentry->d_lock);
2544 /* we can't take ->d_lock here; it's OK, though. */
2545 new->d_flags |= DCACHE_PAR_LOOKUP;
2547 hlist_bl_add_head_rcu(&new->d_u.d_in_lookup_hash, b);
2551 spin_unlock(&dentry->d_lock);
2555 EXPORT_SYMBOL(d_alloc_parallel);
2557 void __d_lookup_done(struct dentry *dentry)
2559 struct hlist_bl_head *b = in_lookup_hash(dentry->d_parent,
2560 dentry->d_name.hash);
2562 dentry->d_flags &= ~DCACHE_PAR_LOOKUP;
2563 __hlist_bl_del(&dentry->d_u.d_in_lookup_hash);
2564 wake_up_all(dentry->d_wait);
2565 dentry->d_wait = NULL;
2567 INIT_HLIST_NODE(&dentry->d_u.d_alias);
2568 INIT_LIST_HEAD(&dentry->d_lru);
2570 EXPORT_SYMBOL(__d_lookup_done);
2572 /* inode->i_lock held if inode is non-NULL */
2574 static inline void __d_add(struct dentry *dentry, struct inode *inode)
2576 struct inode *dir = NULL;
2578 spin_lock(&dentry->d_lock);
2579 if (unlikely(d_in_lookup(dentry))) {
2580 dir = dentry->d_parent->d_inode;
2581 n = start_dir_add(dir);
2582 __d_lookup_done(dentry);
2585 unsigned add_flags = d_flags_for_inode(inode);
2586 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2587 raw_write_seqcount_begin(&dentry->d_seq);
2588 __d_set_inode_and_type(dentry, inode, add_flags);
2589 raw_write_seqcount_end(&dentry->d_seq);
2590 fsnotify_update_flags(dentry);
2594 end_dir_add(dir, n);
2595 spin_unlock(&dentry->d_lock);
2597 spin_unlock(&inode->i_lock);
2601 * d_add - add dentry to hash queues
2602 * @entry: dentry to add
2603 * @inode: The inode to attach to this dentry
2605 * This adds the entry to the hash queues and initializes @inode.
2606 * The entry was actually filled in earlier during d_alloc().
2609 void d_add(struct dentry *entry, struct inode *inode)
2612 security_d_instantiate(entry, inode);
2613 spin_lock(&inode->i_lock);
2615 __d_add(entry, inode);
2617 EXPORT_SYMBOL(d_add);
2620 * d_exact_alias - find and hash an exact unhashed alias
2621 * @entry: dentry to add
2622 * @inode: The inode to go with this dentry
2624 * If an unhashed dentry with the same name/parent and desired
2625 * inode already exists, hash and return it. Otherwise, return
2628 * Parent directory should be locked.
2630 struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2632 struct dentry *alias;
2633 unsigned int hash = entry->d_name.hash;
2635 spin_lock(&inode->i_lock);
2636 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2638 * Don't need alias->d_lock here, because aliases with
2639 * d_parent == entry->d_parent are not subject to name or
2640 * parent changes, because the parent inode i_mutex is held.
2642 if (alias->d_name.hash != hash)
2644 if (alias->d_parent != entry->d_parent)
2646 if (!d_same_name(alias, entry->d_parent, &entry->d_name))
2648 spin_lock(&alias->d_lock);
2649 if (!d_unhashed(alias)) {
2650 spin_unlock(&alias->d_lock);
2653 __dget_dlock(alias);
2655 spin_unlock(&alias->d_lock);
2657 spin_unlock(&inode->i_lock);
2660 spin_unlock(&inode->i_lock);
2663 EXPORT_SYMBOL(d_exact_alias);
2665 static void swap_names(struct dentry *dentry, struct dentry *target)
2667 if (unlikely(dname_external(target))) {
2668 if (unlikely(dname_external(dentry))) {
2670 * Both external: swap the pointers
2672 swap(target->d_name.name, dentry->d_name.name);
2675 * dentry:internal, target:external. Steal target's
2676 * storage and make target internal.
2678 memcpy(target->d_iname, dentry->d_name.name,
2679 dentry->d_name.len + 1);
2680 dentry->d_name.name = target->d_name.name;
2681 target->d_name.name = target->d_iname;
2684 if (unlikely(dname_external(dentry))) {
2686 * dentry:external, target:internal. Give dentry's
2687 * storage to target and make dentry internal
2689 memcpy(dentry->d_iname, target->d_name.name,
2690 target->d_name.len + 1);
2691 target->d_name.name = dentry->d_name.name;
2692 dentry->d_name.name = dentry->d_iname;
2695 * Both are internal.
2698 BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2699 for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2700 swap(((long *) &dentry->d_iname)[i],
2701 ((long *) &target->d_iname)[i]);
2705 swap(dentry->d_name.hash_len, target->d_name.hash_len);
2708 static void copy_name(struct dentry *dentry, struct dentry *target)
2710 struct external_name *old_name = NULL;
2711 if (unlikely(dname_external(dentry)))
2712 old_name = external_name(dentry);
2713 if (unlikely(dname_external(target))) {
2714 atomic_inc(&external_name(target)->u.count);
2715 dentry->d_name = target->d_name;
2717 memcpy(dentry->d_iname, target->d_name.name,
2718 target->d_name.len + 1);
2719 dentry->d_name.name = dentry->d_iname;
2720 dentry->d_name.hash_len = target->d_name.hash_len;
2722 if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2723 kfree_rcu(old_name, u.head);
2727 * __d_move - move a dentry
2728 * @dentry: entry to move
2729 * @target: new dentry
2730 * @exchange: exchange the two dentries
2732 * Update the dcache to reflect the move of a file name. Negative
2733 * dcache entries should not be moved in this way. Caller must hold
2734 * rename_lock, the i_mutex of the source and target directories,
2735 * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2737 static void __d_move(struct dentry *dentry, struct dentry *target,
2740 struct dentry *old_parent, *p;
2741 struct inode *dir = NULL;
2744 WARN_ON(!dentry->d_inode);
2745 if (WARN_ON(dentry == target))
2748 BUG_ON(d_ancestor(target, dentry));
2749 old_parent = dentry->d_parent;
2750 p = d_ancestor(old_parent, target);
2751 if (IS_ROOT(dentry)) {
2753 spin_lock(&target->d_parent->d_lock);
2755 /* target is not a descendent of dentry->d_parent */
2756 spin_lock(&target->d_parent->d_lock);
2757 spin_lock_nested(&old_parent->d_lock, DENTRY_D_LOCK_NESTED);
2759 BUG_ON(p == dentry);
2760 spin_lock(&old_parent->d_lock);
2762 spin_lock_nested(&target->d_parent->d_lock,
2763 DENTRY_D_LOCK_NESTED);
2765 spin_lock_nested(&dentry->d_lock, 2);
2766 spin_lock_nested(&target->d_lock, 3);
2768 if (unlikely(d_in_lookup(target))) {
2769 dir = target->d_parent->d_inode;
2770 n = start_dir_add(dir);
2771 __d_lookup_done(target);
2774 write_seqcount_begin(&dentry->d_seq);
2775 write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2778 if (!d_unhashed(dentry))
2780 if (!d_unhashed(target))
2783 /* ... and switch them in the tree */
2784 dentry->d_parent = target->d_parent;
2786 copy_name(dentry, target);
2787 target->d_hash.pprev = NULL;
2788 dentry->d_parent->d_lockref.count++;
2789 if (dentry != old_parent) /* wasn't IS_ROOT */
2790 WARN_ON(!--old_parent->d_lockref.count);
2792 target->d_parent = old_parent;
2793 swap_names(dentry, target);
2794 list_move(&target->d_child, &target->d_parent->d_subdirs);
2796 fsnotify_update_flags(target);
2798 list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2800 fsnotify_update_flags(dentry);
2801 fscrypt_handle_d_move(dentry);
2803 write_seqcount_end(&target->d_seq);
2804 write_seqcount_end(&dentry->d_seq);
2807 end_dir_add(dir, n);
2809 if (dentry->d_parent != old_parent)
2810 spin_unlock(&dentry->d_parent->d_lock);
2811 if (dentry != old_parent)
2812 spin_unlock(&old_parent->d_lock);
2813 spin_unlock(&target->d_lock);
2814 spin_unlock(&dentry->d_lock);
2818 * d_move - move a dentry
2819 * @dentry: entry to move
2820 * @target: new dentry
2822 * Update the dcache to reflect the move of a file name. Negative
2823 * dcache entries should not be moved in this way. See the locking
2824 * requirements for __d_move.
2826 void d_move(struct dentry *dentry, struct dentry *target)
2828 write_seqlock(&rename_lock);
2829 __d_move(dentry, target, false);
2830 write_sequnlock(&rename_lock);
2832 EXPORT_SYMBOL(d_move);
2835 * d_exchange - exchange two dentries
2836 * @dentry1: first dentry
2837 * @dentry2: second dentry
2839 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2841 write_seqlock(&rename_lock);
2843 WARN_ON(!dentry1->d_inode);
2844 WARN_ON(!dentry2->d_inode);
2845 WARN_ON(IS_ROOT(dentry1));
2846 WARN_ON(IS_ROOT(dentry2));
2848 __d_move(dentry1, dentry2, true);
2850 write_sequnlock(&rename_lock);
2854 * d_ancestor - search for an ancestor
2855 * @p1: ancestor dentry
2858 * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2859 * an ancestor of p2, else NULL.
2861 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2865 for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2866 if (p->d_parent == p1)
2873 * This helper attempts to cope with remotely renamed directories
2875 * It assumes that the caller is already holding
2876 * dentry->d_parent->d_inode->i_mutex, and rename_lock
2878 * Note: If ever the locking in lock_rename() changes, then please
2879 * remember to update this too...
2881 static int __d_unalias(struct inode *inode,
2882 struct dentry *dentry, struct dentry *alias)
2884 struct mutex *m1 = NULL;
2885 struct rw_semaphore *m2 = NULL;
2888 /* If alias and dentry share a parent, then no extra locks required */
2889 if (alias->d_parent == dentry->d_parent)
2892 /* See lock_rename() */
2893 if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2895 m1 = &dentry->d_sb->s_vfs_rename_mutex;
2896 if (!inode_trylock_shared(alias->d_parent->d_inode))
2898 m2 = &alias->d_parent->d_inode->i_rwsem;
2900 __d_move(alias, dentry, false);
2911 * d_splice_alias - splice a disconnected dentry into the tree if one exists
2912 * @inode: the inode which may have a disconnected dentry
2913 * @dentry: a negative dentry which we want to point to the inode.
2915 * If inode is a directory and has an IS_ROOT alias, then d_move that in
2916 * place of the given dentry and return it, else simply d_add the inode
2917 * to the dentry and return NULL.
2919 * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2920 * we should error out: directories can't have multiple aliases.
2922 * This is needed in the lookup routine of any filesystem that is exportable
2923 * (via knfsd) so that we can build dcache paths to directories effectively.
2925 * If a dentry was found and moved, then it is returned. Otherwise NULL
2926 * is returned. This matches the expected return value of ->lookup.
2928 * Cluster filesystems may call this function with a negative, hashed dentry.
2929 * In that case, we know that the inode will be a regular file, and also this
2930 * will only occur during atomic_open. So we need to check for the dentry
2931 * being already hashed only in the final case.
2933 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
2936 return ERR_CAST(inode);
2938 BUG_ON(!d_unhashed(dentry));
2943 security_d_instantiate(dentry, inode);
2944 spin_lock(&inode->i_lock);
2945 if (S_ISDIR(inode->i_mode)) {
2946 struct dentry *new = __d_find_any_alias(inode);
2947 if (unlikely(new)) {
2948 /* The reference to new ensures it remains an alias */
2949 spin_unlock(&inode->i_lock);
2950 write_seqlock(&rename_lock);
2951 if (unlikely(d_ancestor(new, dentry))) {
2952 write_sequnlock(&rename_lock);
2954 new = ERR_PTR(-ELOOP);
2955 pr_warn_ratelimited(
2956 "VFS: Lookup of '%s' in %s %s"
2957 " would have caused loop\n",
2958 dentry->d_name.name,
2959 inode->i_sb->s_type->name,
2961 } else if (!IS_ROOT(new)) {
2962 struct dentry *old_parent = dget(new->d_parent);
2963 int err = __d_unalias(inode, dentry, new);
2964 write_sequnlock(&rename_lock);
2971 __d_move(new, dentry, false);
2972 write_sequnlock(&rename_lock);
2979 __d_add(dentry, inode);
2982 EXPORT_SYMBOL(d_splice_alias);
2985 * Test whether new_dentry is a subdirectory of old_dentry.
2987 * Trivially implemented using the dcache structure
2991 * is_subdir - is new dentry a subdirectory of old_dentry
2992 * @new_dentry: new dentry
2993 * @old_dentry: old dentry
2995 * Returns true if new_dentry is a subdirectory of the parent (at any depth).
2996 * Returns false otherwise.
2997 * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3000 bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3005 if (new_dentry == old_dentry)
3009 /* for restarting inner loop in case of seq retry */
3010 seq = read_seqbegin(&rename_lock);
3012 * Need rcu_readlock to protect against the d_parent trashing
3016 if (d_ancestor(old_dentry, new_dentry))
3021 } while (read_seqretry(&rename_lock, seq));
3025 EXPORT_SYMBOL(is_subdir);
3027 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3029 struct dentry *root = data;
3030 if (dentry != root) {
3031 if (d_unhashed(dentry) || !dentry->d_inode)
3034 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3035 dentry->d_flags |= DCACHE_GENOCIDE;
3036 dentry->d_lockref.count--;
3039 return D_WALK_CONTINUE;
3042 void d_genocide(struct dentry *parent)
3044 d_walk(parent, parent, d_genocide_kill);
3047 EXPORT_SYMBOL(d_genocide);
3049 void d_tmpfile(struct dentry *dentry, struct inode *inode)
3051 inode_dec_link_count(inode);
3052 BUG_ON(dentry->d_name.name != dentry->d_iname ||
3053 !hlist_unhashed(&dentry->d_u.d_alias) ||
3054 !d_unlinked(dentry));
3055 spin_lock(&dentry->d_parent->d_lock);
3056 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3057 dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3058 (unsigned long long)inode->i_ino);
3059 spin_unlock(&dentry->d_lock);
3060 spin_unlock(&dentry->d_parent->d_lock);
3061 d_instantiate(dentry, inode);
3063 EXPORT_SYMBOL(d_tmpfile);
3065 static __initdata unsigned long dhash_entries;
3066 static int __init set_dhash_entries(char *str)
3070 dhash_entries = simple_strtoul(str, &str, 0);
3073 __setup("dhash_entries=", set_dhash_entries);
3075 static void __init dcache_init_early(void)
3077 /* If hashes are distributed across NUMA nodes, defer
3078 * hash allocation until vmalloc space is available.
3084 alloc_large_system_hash("Dentry cache",
3085 sizeof(struct hlist_bl_head),
3088 HASH_EARLY | HASH_ZERO,
3093 d_hash_shift = 32 - d_hash_shift;
3096 static void __init dcache_init(void)
3099 * A constructor could be added for stable state like the lists,
3100 * but it is probably not worth it because of the cache nature
3103 dentry_cache = KMEM_CACHE_USERCOPY(dentry,
3104 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD|SLAB_ACCOUNT,
3107 /* Hash may have been set up in dcache_init_early */
3112 alloc_large_system_hash("Dentry cache",
3113 sizeof(struct hlist_bl_head),
3121 d_hash_shift = 32 - d_hash_shift;
3124 /* SLAB cache for __getname() consumers */
3125 struct kmem_cache *names_cachep __read_mostly;
3126 EXPORT_SYMBOL(names_cachep);
3128 void __init vfs_caches_init_early(void)
3132 for (i = 0; i < ARRAY_SIZE(in_lookup_hashtable); i++)
3133 INIT_HLIST_BL_HEAD(&in_lookup_hashtable[i]);
3135 dcache_init_early();
3139 void __init vfs_caches_init(void)
3141 names_cachep = kmem_cache_create_usercopy("names_cache", PATH_MAX, 0,
3142 SLAB_HWCACHE_ALIGN|SLAB_PANIC, 0, PATH_MAX, NULL);
3147 files_maxfiles_init();