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/syscalls.h>
18 #include <linux/string.h>
21 #include <linux/fsnotify.h>
22 #include <linux/delay.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/mount.h>
29 #include <linux/file.h>
30 #include <asm/uaccess.h>
31 #include <linux/security.h>
32 #include <linux/seqlock.h>
33 #include <linux/swap.h>
34 #include <linux/bootmem.h>
35 #include <linux/fs_struct.h>
36 #include <linux/hardirq.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rculist_bl.h>
39 #include <linux/prefetch.h>
40 #include <linux/ratelimit.h>
41 #include <linux/list_lru.h>
42 #include <linux/kasan.h>
49 * dcache->d_inode->i_lock protects:
50 * - i_dentry, d_u.d_alias, d_inode of aliases
51 * dcache_hash_bucket lock protects:
52 * - the dcache hash table
53 * s_anon bl list spinlock protects:
54 * - the s_anon list (see __d_drop)
55 * dentry->d_sb->s_dentry_lru_lock protects:
56 * - the dcache lru lists and counters
63 * - d_parent and d_subdirs
64 * - childrens' d_child and d_parent
65 * - d_u.d_alias, d_inode
68 * dentry->d_inode->i_lock
70 * dentry->d_sb->s_dentry_lru_lock
71 * dcache_hash_bucket lock
74 * If there is an ancestor relationship:
75 * dentry->d_parent->...->d_parent->d_lock
77 * dentry->d_parent->d_lock
80 * If no ancestor relationship:
81 * if (dentry1 < dentry2)
85 int sysctl_vfs_cache_pressure __read_mostly = 100;
86 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
88 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
90 EXPORT_SYMBOL(rename_lock);
92 static struct kmem_cache *dentry_cache __read_mostly;
95 * This is the single most critical data structure when it comes
96 * to the dcache: the hashtable for lookups. Somebody should try
97 * to make this good - I've just made it work.
99 * This hash-function tries to avoid losing too many bits of hash
100 * information, yet avoid using a prime hash-size or similar.
103 static unsigned int d_hash_mask __read_mostly;
104 static unsigned int d_hash_shift __read_mostly;
106 static struct hlist_bl_head *dentry_hashtable __read_mostly;
108 static inline struct hlist_bl_head *d_hash(const struct dentry *parent,
111 hash += (unsigned long) parent / L1_CACHE_BYTES;
112 return dentry_hashtable + hash_32(hash, d_hash_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 = *(unsigned long *)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)
219 const unsigned char *cs;
221 * Be careful about RCU walk racing with rename:
222 * use ACCESS_ONCE to fetch the name pointer.
224 * NOTE! Even if a rename will mean that the length
225 * was not loaded atomically, we don't care. The
226 * RCU walk will check the sequence count eventually,
227 * and catch it. And we won't overrun the buffer,
228 * because we're reading the name pointer atomically,
229 * and a dentry name is guaranteed to be properly
230 * terminated with a NUL byte.
232 * End result: even if 'len' is wrong, we'll exit
233 * early because the data cannot match (there can
234 * be no NUL in the ct/tcount data)
236 cs = ACCESS_ONCE(dentry->d_name.name);
237 smp_read_barrier_depends();
238 return dentry_string_cmp(cs, ct, tcount);
241 struct external_name {
244 struct rcu_head head;
246 unsigned char name[];
249 static inline struct external_name *external_name(struct dentry *dentry)
251 return container_of(dentry->d_name.name, struct external_name, name[0]);
254 static void __d_free(struct rcu_head *head)
256 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
258 kmem_cache_free(dentry_cache, dentry);
261 static void __d_free_external(struct rcu_head *head)
263 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
264 kfree(external_name(dentry));
265 kmem_cache_free(dentry_cache, dentry);
268 static inline int dname_external(const struct dentry *dentry)
270 return dentry->d_name.name != dentry->d_iname;
274 * Make sure other CPUs see the inode attached before the type is set.
276 static inline void __d_set_inode_and_type(struct dentry *dentry,
282 dentry->d_inode = inode;
284 flags = READ_ONCE(dentry->d_flags);
285 flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
287 WRITE_ONCE(dentry->d_flags, flags);
291 * Ideally, we want to make sure that other CPUs see the flags cleared before
292 * the inode is detached, but this is really a violation of RCU principles
293 * since the ordering suggests we should always set inode before flags.
295 * We should instead replace or discard the entire dentry - but that sucks
296 * performancewise on mass deletion/rename.
298 static inline void __d_clear_type_and_inode(struct dentry *dentry)
300 unsigned flags = READ_ONCE(dentry->d_flags);
302 flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
303 WRITE_ONCE(dentry->d_flags, flags);
305 dentry->d_inode = NULL;
308 static void dentry_free(struct dentry *dentry)
310 WARN_ON(!hlist_unhashed(&dentry->d_u.d_alias));
311 if (unlikely(dname_external(dentry))) {
312 struct external_name *p = external_name(dentry);
313 if (likely(atomic_dec_and_test(&p->u.count))) {
314 call_rcu(&dentry->d_u.d_rcu, __d_free_external);
318 /* if dentry was never visible to RCU, immediate free is OK */
319 if (!(dentry->d_flags & DCACHE_RCUACCESS))
320 __d_free(&dentry->d_u.d_rcu);
322 call_rcu(&dentry->d_u.d_rcu, __d_free);
326 * dentry_rcuwalk_barrier - invalidate in-progress rcu-walk lookups
327 * @dentry: the target dentry
328 * After this call, in-progress rcu-walk path lookup will fail. This
329 * should be called after unhashing, and after changing d_inode (if
330 * the dentry has not already been unhashed).
332 static inline void dentry_rcuwalk_barrier(struct dentry *dentry)
334 assert_spin_locked(&dentry->d_lock);
335 /* Go through a barrier */
336 write_seqcount_barrier(&dentry->d_seq);
340 * Release the dentry's inode, using the filesystem
341 * d_iput() operation if defined. Dentry has no refcount
344 static void dentry_iput(struct dentry * dentry)
345 __releases(dentry->d_lock)
346 __releases(dentry->d_inode->i_lock)
348 struct inode *inode = dentry->d_inode;
350 __d_clear_type_and_inode(dentry);
351 hlist_del_init(&dentry->d_u.d_alias);
352 spin_unlock(&dentry->d_lock);
353 spin_unlock(&inode->i_lock);
355 fsnotify_inoderemove(inode);
356 if (dentry->d_op && dentry->d_op->d_iput)
357 dentry->d_op->d_iput(dentry, inode);
361 spin_unlock(&dentry->d_lock);
366 * Release the dentry's inode, using the filesystem
367 * d_iput() operation if defined. dentry remains in-use.
369 static void dentry_unlink_inode(struct dentry * dentry)
370 __releases(dentry->d_lock)
371 __releases(dentry->d_inode->i_lock)
373 struct inode *inode = dentry->d_inode;
374 __d_clear_type_and_inode(dentry);
375 hlist_del_init(&dentry->d_u.d_alias);
376 dentry_rcuwalk_barrier(dentry);
377 spin_unlock(&dentry->d_lock);
378 spin_unlock(&inode->i_lock);
380 fsnotify_inoderemove(inode);
381 if (dentry->d_op && dentry->d_op->d_iput)
382 dentry->d_op->d_iput(dentry, inode);
388 * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
389 * is in use - which includes both the "real" per-superblock
390 * LRU list _and_ the DCACHE_SHRINK_LIST use.
392 * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
393 * on the shrink list (ie not on the superblock LRU list).
395 * The per-cpu "nr_dentry_unused" counters are updated with
396 * the DCACHE_LRU_LIST bit.
398 * These helper functions make sure we always follow the
399 * rules. d_lock must be held by the caller.
401 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
402 static void d_lru_add(struct dentry *dentry)
404 D_FLAG_VERIFY(dentry, 0);
405 dentry->d_flags |= DCACHE_LRU_LIST;
406 this_cpu_inc(nr_dentry_unused);
407 WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
410 static void d_lru_del(struct dentry *dentry)
412 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
413 dentry->d_flags &= ~DCACHE_LRU_LIST;
414 this_cpu_dec(nr_dentry_unused);
415 WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
418 static void d_shrink_del(struct dentry *dentry)
420 D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
421 list_del_init(&dentry->d_lru);
422 dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
423 this_cpu_dec(nr_dentry_unused);
426 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
428 D_FLAG_VERIFY(dentry, 0);
429 list_add(&dentry->d_lru, list);
430 dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
431 this_cpu_inc(nr_dentry_unused);
435 * These can only be called under the global LRU lock, ie during the
436 * callback for freeing the LRU list. "isolate" removes it from the
437 * LRU lists entirely, while shrink_move moves it to the indicated
440 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
442 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
443 dentry->d_flags &= ~DCACHE_LRU_LIST;
444 this_cpu_dec(nr_dentry_unused);
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 list_lru_isolate_move(lru, &dentry->d_lru, list);
457 * dentry_lru_(add|del)_list) must be called with d_lock held.
459 static void dentry_lru_add(struct dentry *dentry)
461 if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
466 * d_drop - drop a dentry
467 * @dentry: dentry to drop
469 * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
470 * be found through a VFS lookup any more. Note that this is different from
471 * deleting the dentry - d_delete will try to mark the dentry negative if
472 * possible, giving a successful _negative_ lookup, while d_drop will
473 * just make the cache lookup fail.
475 * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
476 * reason (NFS timeouts or autofs deletes).
478 * __d_drop requires dentry->d_lock.
480 void __d_drop(struct dentry *dentry)
482 if (!d_unhashed(dentry)) {
483 struct hlist_bl_head *b;
485 * Hashed dentries are normally on the dentry hashtable,
486 * with the exception of those newly allocated by
487 * d_obtain_alias, which are always IS_ROOT:
489 if (unlikely(IS_ROOT(dentry)))
490 b = &dentry->d_sb->s_anon;
492 b = d_hash(dentry->d_parent, dentry->d_name.hash);
495 __hlist_bl_del(&dentry->d_hash);
496 dentry->d_hash.pprev = NULL;
498 dentry_rcuwalk_barrier(dentry);
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 void __dentry_kill(struct dentry *dentry)
513 struct dentry *parent = NULL;
514 bool can_free = true;
515 if (!IS_ROOT(dentry))
516 parent = dentry->d_parent;
519 * The dentry is now unrecoverably dead to the world.
521 lockref_mark_dead(&dentry->d_lockref);
524 * inform the fs via d_prune that this dentry is about to be
525 * unhashed and destroyed.
527 if (dentry->d_flags & DCACHE_OP_PRUNE)
528 dentry->d_op->d_prune(dentry);
530 if (dentry->d_flags & DCACHE_LRU_LIST) {
531 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
534 /* if it was on the hash then remove it */
536 __list_del_entry(&dentry->d_child);
538 * Inform d_walk() that we are no longer attached to the
541 dentry->d_flags |= DCACHE_DENTRY_KILLED;
543 spin_unlock(&parent->d_lock);
546 * dentry_iput drops the locks, at which point nobody (except
547 * transient RCU lookups) can reach this dentry.
549 BUG_ON(dentry->d_lockref.count > 0);
550 this_cpu_dec(nr_dentry);
551 if (dentry->d_op && dentry->d_op->d_release)
552 dentry->d_op->d_release(dentry);
554 spin_lock(&dentry->d_lock);
555 if (dentry->d_flags & DCACHE_SHRINK_LIST) {
556 dentry->d_flags |= DCACHE_MAY_FREE;
559 spin_unlock(&dentry->d_lock);
560 if (likely(can_free))
565 * Finish off a dentry we've decided to kill.
566 * dentry->d_lock must be held, returns with it unlocked.
567 * If ref is non-zero, then decrement the refcount too.
568 * Returns dentry requiring refcount drop, or NULL if we're done.
570 static struct dentry *dentry_kill(struct dentry *dentry)
571 __releases(dentry->d_lock)
573 struct inode *inode = dentry->d_inode;
574 struct dentry *parent = NULL;
576 if (inode && unlikely(!spin_trylock(&inode->i_lock)))
579 if (!IS_ROOT(dentry)) {
580 parent = dentry->d_parent;
581 if (unlikely(!spin_trylock(&parent->d_lock))) {
583 spin_unlock(&inode->i_lock);
588 __dentry_kill(dentry);
592 spin_unlock(&dentry->d_lock);
594 return dentry; /* try again with same dentry */
597 static inline struct dentry *lock_parent(struct dentry *dentry)
599 struct dentry *parent = dentry->d_parent;
602 if (unlikely(dentry->d_lockref.count < 0))
604 if (likely(spin_trylock(&parent->d_lock)))
607 spin_unlock(&dentry->d_lock);
609 parent = ACCESS_ONCE(dentry->d_parent);
610 spin_lock(&parent->d_lock);
612 * We can't blindly lock dentry until we are sure
613 * that we won't violate the locking order.
614 * Any changes of dentry->d_parent must have
615 * been done with parent->d_lock held, so
616 * spin_lock() above is enough of a barrier
617 * for checking if it's still our child.
619 if (unlikely(parent != dentry->d_parent)) {
620 spin_unlock(&parent->d_lock);
624 if (parent != dentry)
625 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
632 * Try to do a lockless dput(), and return whether that was successful.
634 * If unsuccessful, we return false, having already taken the dentry lock.
636 * The caller needs to hold the RCU read lock, so that the dentry is
637 * guaranteed to stay around even if the refcount goes down to zero!
639 static inline bool fast_dput(struct dentry *dentry)
642 unsigned int d_flags;
645 * If we have a d_op->d_delete() operation, we sould not
646 * let the dentry count go to zero, so use "put__or_lock".
648 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE))
649 return lockref_put_or_lock(&dentry->d_lockref);
652 * .. otherwise, we can try to just decrement the
653 * lockref optimistically.
655 ret = lockref_put_return(&dentry->d_lockref);
658 * If the lockref_put_return() failed due to the lock being held
659 * by somebody else, the fast path has failed. We will need to
660 * get the lock, and then check the count again.
662 if (unlikely(ret < 0)) {
663 spin_lock(&dentry->d_lock);
664 if (dentry->d_lockref.count > 1) {
665 dentry->d_lockref.count--;
666 spin_unlock(&dentry->d_lock);
673 * If we weren't the last ref, we're done.
679 * Careful, careful. The reference count went down
680 * to zero, but we don't hold the dentry lock, so
681 * somebody else could get it again, and do another
682 * dput(), and we need to not race with that.
684 * However, there is a very special and common case
685 * where we don't care, because there is nothing to
686 * do: the dentry is still hashed, it does not have
687 * a 'delete' op, and it's referenced and already on
690 * NOTE! Since we aren't locked, these values are
691 * not "stable". However, it is sufficient that at
692 * some point after we dropped the reference the
693 * dentry was hashed and the flags had the proper
694 * value. Other dentry users may have re-gotten
695 * a reference to the dentry and change that, but
696 * our work is done - we can leave the dentry
697 * around with a zero refcount.
700 d_flags = ACCESS_ONCE(dentry->d_flags);
701 d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST;
703 /* Nothing to do? Dropping the reference was all we needed? */
704 if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry))
708 * Not the fast normal case? Get the lock. We've already decremented
709 * the refcount, but we'll need to re-check the situation after
712 spin_lock(&dentry->d_lock);
715 * Did somebody else grab a reference to it in the meantime, and
716 * we're no longer the last user after all? Alternatively, somebody
717 * else could have killed it and marked it dead. Either way, we
718 * don't need to do anything else.
720 if (dentry->d_lockref.count) {
721 spin_unlock(&dentry->d_lock);
726 * Re-get the reference we optimistically dropped. We hold the
727 * lock, and we just tested that it was zero, so we can just
730 dentry->d_lockref.count = 1;
738 * This is complicated by the fact that we do not want to put
739 * dentries that are no longer on any hash chain on the unused
740 * list: we'd much rather just get rid of them immediately.
742 * However, that implies that we have to traverse the dentry
743 * tree upwards to the parents which might _also_ now be
744 * scheduled for deletion (it may have been only waiting for
745 * its last child to go away).
747 * This tail recursion is done by hand as we don't want to depend
748 * on the compiler to always get this right (gcc generally doesn't).
749 * Real recursion would eat up our stack space.
753 * dput - release a dentry
754 * @dentry: dentry to release
756 * Release a dentry. This will drop the usage count and if appropriate
757 * call the dentry unlink method as well as removing it from the queues and
758 * releasing its resources. If the parent dentries were scheduled for release
759 * they too may now get deleted.
761 void dput(struct dentry *dentry)
763 if (unlikely(!dentry))
768 if (likely(fast_dput(dentry))) {
773 /* Slow case: now with the dentry lock held */
776 /* Unreachable? Get rid of it */
777 if (unlikely(d_unhashed(dentry)))
780 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
781 if (dentry->d_op->d_delete(dentry))
785 if (!(dentry->d_flags & DCACHE_REFERENCED))
786 dentry->d_flags |= DCACHE_REFERENCED;
787 dentry_lru_add(dentry);
789 dentry->d_lockref.count--;
790 spin_unlock(&dentry->d_lock);
794 dentry = dentry_kill(dentry);
801 /* This must be called with d_lock held */
802 static inline void __dget_dlock(struct dentry *dentry)
804 dentry->d_lockref.count++;
807 static inline void __dget(struct dentry *dentry)
809 lockref_get(&dentry->d_lockref);
812 struct dentry *dget_parent(struct dentry *dentry)
818 * Do optimistic parent lookup without any
822 ret = ACCESS_ONCE(dentry->d_parent);
823 gotref = lockref_get_not_zero(&ret->d_lockref);
825 if (likely(gotref)) {
826 if (likely(ret == ACCESS_ONCE(dentry->d_parent)))
833 * Don't need rcu_dereference because we re-check it was correct under
837 ret = dentry->d_parent;
838 spin_lock(&ret->d_lock);
839 if (unlikely(ret != dentry->d_parent)) {
840 spin_unlock(&ret->d_lock);
845 BUG_ON(!ret->d_lockref.count);
846 ret->d_lockref.count++;
847 spin_unlock(&ret->d_lock);
850 EXPORT_SYMBOL(dget_parent);
853 * d_find_alias - grab a hashed alias of inode
854 * @inode: inode in question
856 * If inode has a hashed alias, or is a directory and has any alias,
857 * acquire the reference to alias and return it. Otherwise return NULL.
858 * Notice that if inode is a directory there can be only one alias and
859 * it can be unhashed only if it has no children, or if it is the root
860 * of a filesystem, or if the directory was renamed and d_revalidate
861 * was the first vfs operation to notice.
863 * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
864 * any other hashed alias over that one.
866 static struct dentry *__d_find_alias(struct inode *inode)
868 struct dentry *alias, *discon_alias;
872 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
873 spin_lock(&alias->d_lock);
874 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
875 if (IS_ROOT(alias) &&
876 (alias->d_flags & DCACHE_DISCONNECTED)) {
877 discon_alias = alias;
880 spin_unlock(&alias->d_lock);
884 spin_unlock(&alias->d_lock);
887 alias = discon_alias;
888 spin_lock(&alias->d_lock);
889 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
891 spin_unlock(&alias->d_lock);
894 spin_unlock(&alias->d_lock);
900 struct dentry *d_find_alias(struct inode *inode)
902 struct dentry *de = NULL;
904 if (!hlist_empty(&inode->i_dentry)) {
905 spin_lock(&inode->i_lock);
906 de = __d_find_alias(inode);
907 spin_unlock(&inode->i_lock);
911 EXPORT_SYMBOL(d_find_alias);
914 * Try to kill dentries associated with this inode.
915 * WARNING: you must own a reference to inode.
917 void d_prune_aliases(struct inode *inode)
919 struct dentry *dentry;
921 spin_lock(&inode->i_lock);
922 hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
923 spin_lock(&dentry->d_lock);
924 if (!dentry->d_lockref.count) {
925 struct dentry *parent = lock_parent(dentry);
926 if (likely(!dentry->d_lockref.count)) {
927 __dentry_kill(dentry);
932 spin_unlock(&parent->d_lock);
934 spin_unlock(&dentry->d_lock);
936 spin_unlock(&inode->i_lock);
938 EXPORT_SYMBOL(d_prune_aliases);
940 static void shrink_dentry_list(struct list_head *list)
942 struct dentry *dentry, *parent;
944 while (!list_empty(list)) {
946 dentry = list_entry(list->prev, struct dentry, d_lru);
947 spin_lock(&dentry->d_lock);
948 parent = lock_parent(dentry);
951 * The dispose list is isolated and dentries are not accounted
952 * to the LRU here, so we can simply remove it from the list
953 * here regardless of whether it is referenced or not.
955 d_shrink_del(dentry);
958 * We found an inuse dentry which was not removed from
959 * the LRU because of laziness during lookup. Do not free it.
961 if (dentry->d_lockref.count > 0) {
962 spin_unlock(&dentry->d_lock);
964 spin_unlock(&parent->d_lock);
969 if (unlikely(dentry->d_flags & DCACHE_DENTRY_KILLED)) {
970 bool can_free = dentry->d_flags & DCACHE_MAY_FREE;
971 spin_unlock(&dentry->d_lock);
973 spin_unlock(&parent->d_lock);
979 inode = dentry->d_inode;
980 if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
981 d_shrink_add(dentry, list);
982 spin_unlock(&dentry->d_lock);
984 spin_unlock(&parent->d_lock);
988 __dentry_kill(dentry);
991 * We need to prune ancestors too. This is necessary to prevent
992 * quadratic behavior of shrink_dcache_parent(), but is also
993 * expected to be beneficial in reducing dentry cache
997 while (dentry && !lockref_put_or_lock(&dentry->d_lockref)) {
998 parent = lock_parent(dentry);
999 if (dentry->d_lockref.count != 1) {
1000 dentry->d_lockref.count--;
1001 spin_unlock(&dentry->d_lock);
1003 spin_unlock(&parent->d_lock);
1006 inode = dentry->d_inode; /* can't be NULL */
1007 if (unlikely(!spin_trylock(&inode->i_lock))) {
1008 spin_unlock(&dentry->d_lock);
1010 spin_unlock(&parent->d_lock);
1014 __dentry_kill(dentry);
1020 static enum lru_status dentry_lru_isolate(struct list_head *item,
1021 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1023 struct list_head *freeable = arg;
1024 struct dentry *dentry = container_of(item, struct dentry, d_lru);
1028 * we are inverting the lru lock/dentry->d_lock here,
1029 * so use a trylock. If we fail to get the lock, just skip
1032 if (!spin_trylock(&dentry->d_lock))
1036 * Referenced dentries are still in use. If they have active
1037 * counts, just remove them from the LRU. Otherwise give them
1038 * another pass through the LRU.
1040 if (dentry->d_lockref.count) {
1041 d_lru_isolate(lru, dentry);
1042 spin_unlock(&dentry->d_lock);
1046 if (dentry->d_flags & DCACHE_REFERENCED) {
1047 dentry->d_flags &= ~DCACHE_REFERENCED;
1048 spin_unlock(&dentry->d_lock);
1051 * The list move itself will be made by the common LRU code. At
1052 * this point, we've dropped the dentry->d_lock but keep the
1053 * lru lock. This is safe to do, since every list movement is
1054 * protected by the lru lock even if both locks are held.
1056 * This is guaranteed by the fact that all LRU management
1057 * functions are intermediated by the LRU API calls like
1058 * list_lru_add and list_lru_del. List movement in this file
1059 * only ever occur through this functions or through callbacks
1060 * like this one, that are called from the LRU API.
1062 * The only exceptions to this are functions like
1063 * shrink_dentry_list, and code that first checks for the
1064 * DCACHE_SHRINK_LIST flag. Those are guaranteed to be
1065 * operating only with stack provided lists after they are
1066 * properly isolated from the main list. It is thus, always a
1072 d_lru_shrink_move(lru, dentry, freeable);
1073 spin_unlock(&dentry->d_lock);
1079 * prune_dcache_sb - shrink the dcache
1081 * @sc: shrink control, passed to list_lru_shrink_walk()
1083 * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1084 * is done when we need more memory and called from the superblock shrinker
1087 * This function may fail to free any resources if all the dentries are in
1090 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1095 freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1096 dentry_lru_isolate, &dispose);
1097 shrink_dentry_list(&dispose);
1101 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1102 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1104 struct list_head *freeable = arg;
1105 struct dentry *dentry = container_of(item, struct dentry, d_lru);
1108 * we are inverting the lru lock/dentry->d_lock here,
1109 * so use a trylock. If we fail to get the lock, just skip
1112 if (!spin_trylock(&dentry->d_lock))
1115 d_lru_shrink_move(lru, dentry, freeable);
1116 spin_unlock(&dentry->d_lock);
1123 * shrink_dcache_sb - shrink dcache for a superblock
1126 * Shrink the dcache for the specified super block. This is used to free
1127 * the dcache before unmounting a file system.
1129 void shrink_dcache_sb(struct super_block *sb)
1136 freed = list_lru_walk(&sb->s_dentry_lru,
1137 dentry_lru_isolate_shrink, &dispose, UINT_MAX);
1139 this_cpu_sub(nr_dentry_unused, freed);
1140 shrink_dentry_list(&dispose);
1141 } while (freed > 0);
1143 EXPORT_SYMBOL(shrink_dcache_sb);
1146 * enum d_walk_ret - action to talke during tree walk
1147 * @D_WALK_CONTINUE: contrinue walk
1148 * @D_WALK_QUIT: quit walk
1149 * @D_WALK_NORETRY: quit when retry is needed
1150 * @D_WALK_SKIP: skip this dentry and its children
1160 * d_walk - walk the dentry tree
1161 * @parent: start of walk
1162 * @data: data passed to @enter() and @finish()
1163 * @enter: callback when first entering the dentry
1164 * @finish: callback when successfully finished the walk
1166 * The @enter() and @finish() callbacks are called with d_lock held.
1168 static void d_walk(struct dentry *parent, void *data,
1169 enum d_walk_ret (*enter)(void *, struct dentry *),
1170 void (*finish)(void *))
1172 struct dentry *this_parent;
1173 struct list_head *next;
1175 enum d_walk_ret ret;
1179 read_seqbegin_or_lock(&rename_lock, &seq);
1180 this_parent = parent;
1181 spin_lock(&this_parent->d_lock);
1183 ret = enter(data, this_parent);
1185 case D_WALK_CONTINUE:
1190 case D_WALK_NORETRY:
1195 next = this_parent->d_subdirs.next;
1197 while (next != &this_parent->d_subdirs) {
1198 struct list_head *tmp = next;
1199 struct dentry *dentry = list_entry(tmp, struct dentry, d_child);
1202 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1204 ret = enter(data, dentry);
1206 case D_WALK_CONTINUE:
1209 spin_unlock(&dentry->d_lock);
1211 case D_WALK_NORETRY:
1215 spin_unlock(&dentry->d_lock);
1219 if (!list_empty(&dentry->d_subdirs)) {
1220 spin_unlock(&this_parent->d_lock);
1221 spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1222 this_parent = dentry;
1223 spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1226 spin_unlock(&dentry->d_lock);
1229 * All done at this level ... ascend and resume the search.
1233 if (this_parent != parent) {
1234 struct dentry *child = this_parent;
1235 this_parent = child->d_parent;
1237 spin_unlock(&child->d_lock);
1238 spin_lock(&this_parent->d_lock);
1240 /* might go back up the wrong parent if we have had a rename. */
1241 if (need_seqretry(&rename_lock, seq))
1243 /* go into the first sibling still alive */
1245 next = child->d_child.next;
1246 if (next == &this_parent->d_subdirs)
1248 child = list_entry(next, struct dentry, d_child);
1249 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED));
1253 if (need_seqretry(&rename_lock, seq))
1260 spin_unlock(&this_parent->d_lock);
1261 done_seqretry(&rename_lock, seq);
1265 spin_unlock(&this_parent->d_lock);
1275 * Search for at least 1 mount point in the dentry's subdirs.
1276 * We descend to the next level whenever the d_subdirs
1277 * list is non-empty and continue searching.
1280 static enum d_walk_ret check_mount(void *data, struct dentry *dentry)
1283 if (d_mountpoint(dentry)) {
1287 return D_WALK_CONTINUE;
1291 * have_submounts - check for mounts over a dentry
1292 * @parent: dentry to check.
1294 * Return true if the parent or its subdirectories contain
1297 int have_submounts(struct dentry *parent)
1301 d_walk(parent, &ret, check_mount, NULL);
1305 EXPORT_SYMBOL(have_submounts);
1308 * Called by mount code to set a mountpoint and check if the mountpoint is
1309 * reachable (e.g. NFS can unhash a directory dentry and then the complete
1310 * subtree can become unreachable).
1312 * Only one of d_invalidate() and d_set_mounted() must succeed. For
1313 * this reason take rename_lock and d_lock on dentry and ancestors.
1315 int d_set_mounted(struct dentry *dentry)
1319 write_seqlock(&rename_lock);
1320 for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1321 /* Need exclusion wrt. d_invalidate() */
1322 spin_lock(&p->d_lock);
1323 if (unlikely(d_unhashed(p))) {
1324 spin_unlock(&p->d_lock);
1327 spin_unlock(&p->d_lock);
1329 spin_lock(&dentry->d_lock);
1330 if (!d_unlinked(dentry)) {
1331 dentry->d_flags |= DCACHE_MOUNTED;
1334 spin_unlock(&dentry->d_lock);
1336 write_sequnlock(&rename_lock);
1341 * Search the dentry child list of the specified parent,
1342 * and move any unused dentries to the end of the unused
1343 * list for prune_dcache(). We descend to the next level
1344 * whenever the d_subdirs list is non-empty and continue
1347 * It returns zero iff there are no unused children,
1348 * otherwise it returns the number of children moved to
1349 * the end of the unused list. This may not be the total
1350 * number of unused children, because select_parent can
1351 * drop the lock and return early due to latency
1355 struct select_data {
1356 struct dentry *start;
1357 struct list_head dispose;
1361 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1363 struct select_data *data = _data;
1364 enum d_walk_ret ret = D_WALK_CONTINUE;
1366 if (data->start == dentry)
1369 if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1372 if (dentry->d_flags & DCACHE_LRU_LIST)
1374 if (!dentry->d_lockref.count) {
1375 d_shrink_add(dentry, &data->dispose);
1380 * We can return to the caller if we have found some (this
1381 * ensures forward progress). We'll be coming back to find
1384 if (!list_empty(&data->dispose))
1385 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1391 * shrink_dcache_parent - prune dcache
1392 * @parent: parent of entries to prune
1394 * Prune the dcache to remove unused children of the parent dentry.
1396 void shrink_dcache_parent(struct dentry *parent)
1399 struct select_data data;
1401 INIT_LIST_HEAD(&data.dispose);
1402 data.start = parent;
1405 d_walk(parent, &data, select_collect, NULL);
1409 shrink_dentry_list(&data.dispose);
1413 EXPORT_SYMBOL(shrink_dcache_parent);
1415 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1417 /* it has busy descendents; complain about those instead */
1418 if (!list_empty(&dentry->d_subdirs))
1419 return D_WALK_CONTINUE;
1421 /* root with refcount 1 is fine */
1422 if (dentry == _data && dentry->d_lockref.count == 1)
1423 return D_WALK_CONTINUE;
1425 printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} "
1426 " still in use (%d) [unmount of %s %s]\n",
1429 dentry->d_inode->i_ino : 0UL,
1431 dentry->d_lockref.count,
1432 dentry->d_sb->s_type->name,
1433 dentry->d_sb->s_id);
1435 return D_WALK_CONTINUE;
1438 static void do_one_tree(struct dentry *dentry)
1440 shrink_dcache_parent(dentry);
1441 d_walk(dentry, dentry, umount_check, NULL);
1447 * destroy the dentries attached to a superblock on unmounting
1449 void shrink_dcache_for_umount(struct super_block *sb)
1451 struct dentry *dentry;
1453 WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
1455 dentry = sb->s_root;
1457 do_one_tree(dentry);
1459 while (!hlist_bl_empty(&sb->s_anon)) {
1460 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_anon), struct dentry, d_hash));
1461 do_one_tree(dentry);
1465 struct detach_data {
1466 struct select_data select;
1467 struct dentry *mountpoint;
1469 static enum d_walk_ret detach_and_collect(void *_data, struct dentry *dentry)
1471 struct detach_data *data = _data;
1473 if (d_mountpoint(dentry)) {
1474 __dget_dlock(dentry);
1475 data->mountpoint = dentry;
1479 return select_collect(&data->select, dentry);
1482 static void check_and_drop(void *_data)
1484 struct detach_data *data = _data;
1486 if (!data->mountpoint && !data->select.found)
1487 __d_drop(data->select.start);
1491 * d_invalidate - detach submounts, prune dcache, and drop
1492 * @dentry: dentry to invalidate (aka detach, prune and drop)
1496 * The final d_drop is done as an atomic operation relative to
1497 * rename_lock ensuring there are no races with d_set_mounted. This
1498 * ensures there are no unhashed dentries on the path to a mountpoint.
1500 void d_invalidate(struct dentry *dentry)
1503 * If it's already been dropped, return OK.
1505 spin_lock(&dentry->d_lock);
1506 if (d_unhashed(dentry)) {
1507 spin_unlock(&dentry->d_lock);
1510 spin_unlock(&dentry->d_lock);
1512 /* Negative dentries can be dropped without further checks */
1513 if (!dentry->d_inode) {
1519 struct detach_data data;
1521 data.mountpoint = NULL;
1522 INIT_LIST_HEAD(&data.select.dispose);
1523 data.select.start = dentry;
1524 data.select.found = 0;
1526 d_walk(dentry, &data, detach_and_collect, check_and_drop);
1528 if (data.select.found)
1529 shrink_dentry_list(&data.select.dispose);
1531 if (data.mountpoint) {
1532 detach_mounts(data.mountpoint);
1533 dput(data.mountpoint);
1536 if (!data.mountpoint && !data.select.found)
1542 EXPORT_SYMBOL(d_invalidate);
1545 * __d_alloc - allocate a dcache entry
1546 * @sb: filesystem it will belong to
1547 * @name: qstr of the name
1549 * Allocates a dentry. It returns %NULL if there is insufficient memory
1550 * available. On a success the dentry is returned. The name passed in is
1551 * copied and the copy passed in may be reused after this call.
1554 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1556 struct dentry *dentry;
1559 dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1564 * We guarantee that the inline name is always NUL-terminated.
1565 * This way the memcpy() done by the name switching in rename
1566 * will still always have a NUL at the end, even if we might
1567 * be overwriting an internal NUL character
1569 dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1570 if (name->len > DNAME_INLINE_LEN-1) {
1571 size_t size = offsetof(struct external_name, name[1]);
1572 struct external_name *p = kmalloc(size + name->len, GFP_KERNEL);
1574 kmem_cache_free(dentry_cache, dentry);
1577 atomic_set(&p->u.count, 1);
1579 if (IS_ENABLED(CONFIG_DCACHE_WORD_ACCESS))
1580 kasan_unpoison_shadow(dname,
1581 round_up(name->len + 1, sizeof(unsigned long)));
1583 dname = dentry->d_iname;
1586 dentry->d_name.len = name->len;
1587 dentry->d_name.hash = name->hash;
1588 memcpy(dname, name->name, name->len);
1589 dname[name->len] = 0;
1591 /* Make sure we always see the terminating NUL character */
1593 dentry->d_name.name = dname;
1595 dentry->d_lockref.count = 1;
1596 dentry->d_flags = 0;
1597 spin_lock_init(&dentry->d_lock);
1598 seqcount_init(&dentry->d_seq);
1599 dentry->d_inode = NULL;
1600 dentry->d_parent = dentry;
1602 dentry->d_op = NULL;
1603 dentry->d_fsdata = NULL;
1604 INIT_HLIST_BL_NODE(&dentry->d_hash);
1605 INIT_LIST_HEAD(&dentry->d_lru);
1606 INIT_LIST_HEAD(&dentry->d_subdirs);
1607 INIT_HLIST_NODE(&dentry->d_u.d_alias);
1608 INIT_LIST_HEAD(&dentry->d_child);
1609 d_set_d_op(dentry, dentry->d_sb->s_d_op);
1611 this_cpu_inc(nr_dentry);
1617 * d_alloc - allocate a dcache entry
1618 * @parent: parent of entry to allocate
1619 * @name: qstr of the name
1621 * Allocates a dentry. It returns %NULL if there is insufficient memory
1622 * available. On a success the dentry is returned. The name passed in is
1623 * copied and the copy passed in may be reused after this call.
1625 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1627 struct dentry *dentry = __d_alloc(parent->d_sb, name);
1631 spin_lock(&parent->d_lock);
1633 * don't need child lock because it is not subject
1634 * to concurrency here
1636 __dget_dlock(parent);
1637 dentry->d_parent = parent;
1638 list_add(&dentry->d_child, &parent->d_subdirs);
1639 spin_unlock(&parent->d_lock);
1643 EXPORT_SYMBOL(d_alloc);
1646 * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1647 * @sb: the superblock
1648 * @name: qstr of the name
1650 * For a filesystem that just pins its dentries in memory and never
1651 * performs lookups at all, return an unhashed IS_ROOT dentry.
1653 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1655 return __d_alloc(sb, name);
1657 EXPORT_SYMBOL(d_alloc_pseudo);
1659 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1664 q.len = strlen(name);
1665 q.hash = full_name_hash(q.name, q.len);
1666 return d_alloc(parent, &q);
1668 EXPORT_SYMBOL(d_alloc_name);
1670 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1672 WARN_ON_ONCE(dentry->d_op);
1673 WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH |
1675 DCACHE_OP_REVALIDATE |
1676 DCACHE_OP_WEAK_REVALIDATE |
1677 DCACHE_OP_DELETE ));
1682 dentry->d_flags |= DCACHE_OP_HASH;
1684 dentry->d_flags |= DCACHE_OP_COMPARE;
1685 if (op->d_revalidate)
1686 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1687 if (op->d_weak_revalidate)
1688 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1690 dentry->d_flags |= DCACHE_OP_DELETE;
1692 dentry->d_flags |= DCACHE_OP_PRUNE;
1695 EXPORT_SYMBOL(d_set_d_op);
1699 * d_set_fallthru - Mark a dentry as falling through to a lower layer
1700 * @dentry - The dentry to mark
1702 * Mark a dentry as falling through to the lower layer (as set with
1703 * d_pin_lower()). This flag may be recorded on the medium.
1705 void d_set_fallthru(struct dentry *dentry)
1707 spin_lock(&dentry->d_lock);
1708 dentry->d_flags |= DCACHE_FALLTHRU;
1709 spin_unlock(&dentry->d_lock);
1711 EXPORT_SYMBOL(d_set_fallthru);
1713 static unsigned d_flags_for_inode(struct inode *inode)
1715 unsigned add_flags = DCACHE_REGULAR_TYPE;
1718 return DCACHE_MISS_TYPE;
1720 if (S_ISDIR(inode->i_mode)) {
1721 add_flags = DCACHE_DIRECTORY_TYPE;
1722 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1723 if (unlikely(!inode->i_op->lookup))
1724 add_flags = DCACHE_AUTODIR_TYPE;
1726 inode->i_opflags |= IOP_LOOKUP;
1728 goto type_determined;
1731 if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1732 if (unlikely(inode->i_op->follow_link)) {
1733 add_flags = DCACHE_SYMLINK_TYPE;
1734 goto type_determined;
1736 inode->i_opflags |= IOP_NOFOLLOW;
1739 if (unlikely(!S_ISREG(inode->i_mode)))
1740 add_flags = DCACHE_SPECIAL_TYPE;
1743 if (unlikely(IS_AUTOMOUNT(inode)))
1744 add_flags |= DCACHE_NEED_AUTOMOUNT;
1748 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1750 unsigned add_flags = d_flags_for_inode(inode);
1752 spin_lock(&dentry->d_lock);
1754 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1755 __d_set_inode_and_type(dentry, inode, add_flags);
1756 dentry_rcuwalk_barrier(dentry);
1757 spin_unlock(&dentry->d_lock);
1758 fsnotify_d_instantiate(dentry, inode);
1762 * d_instantiate - fill in inode information for a dentry
1763 * @entry: dentry to complete
1764 * @inode: inode to attach to this dentry
1766 * Fill in inode information in the entry.
1768 * This turns negative dentries into productive full members
1771 * NOTE! This assumes that the inode count has been incremented
1772 * (or otherwise set) by the caller to indicate that it is now
1773 * in use by the dcache.
1776 void d_instantiate(struct dentry *entry, struct inode * inode)
1778 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1780 spin_lock(&inode->i_lock);
1781 __d_instantiate(entry, inode);
1783 spin_unlock(&inode->i_lock);
1784 security_d_instantiate(entry, inode);
1786 EXPORT_SYMBOL(d_instantiate);
1789 * d_instantiate_unique - instantiate a non-aliased dentry
1790 * @entry: dentry to instantiate
1791 * @inode: inode to attach to this dentry
1793 * Fill in inode information in the entry. On success, it returns NULL.
1794 * If an unhashed alias of "entry" already exists, then we return the
1795 * aliased dentry instead and drop one reference to inode.
1797 * Note that in order to avoid conflicts with rename() etc, the caller
1798 * had better be holding the parent directory semaphore.
1800 * This also assumes that the inode count has been incremented
1801 * (or otherwise set) by the caller to indicate that it is now
1802 * in use by the dcache.
1804 static struct dentry *__d_instantiate_unique(struct dentry *entry,
1805 struct inode *inode)
1807 struct dentry *alias;
1808 int len = entry->d_name.len;
1809 const char *name = entry->d_name.name;
1810 unsigned int hash = entry->d_name.hash;
1813 __d_instantiate(entry, NULL);
1817 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
1819 * Don't need alias->d_lock here, because aliases with
1820 * d_parent == entry->d_parent are not subject to name or
1821 * parent changes, because the parent inode i_mutex is held.
1823 if (alias->d_name.hash != hash)
1825 if (alias->d_parent != entry->d_parent)
1827 if (alias->d_name.len != len)
1829 if (dentry_cmp(alias, name, len))
1835 __d_instantiate(entry, inode);
1839 struct dentry *d_instantiate_unique(struct dentry *entry, struct inode *inode)
1841 struct dentry *result;
1843 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1846 spin_lock(&inode->i_lock);
1847 result = __d_instantiate_unique(entry, inode);
1849 spin_unlock(&inode->i_lock);
1852 security_d_instantiate(entry, inode);
1856 BUG_ON(!d_unhashed(result));
1861 EXPORT_SYMBOL(d_instantiate_unique);
1864 * d_instantiate_no_diralias - instantiate a non-aliased dentry
1865 * @entry: dentry to complete
1866 * @inode: inode to attach to this dentry
1868 * Fill in inode information in the entry. If a directory alias is found, then
1869 * return an error (and drop inode). Together with d_materialise_unique() this
1870 * guarantees that a directory inode may never have more than one alias.
1872 int d_instantiate_no_diralias(struct dentry *entry, struct inode *inode)
1874 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1876 spin_lock(&inode->i_lock);
1877 if (S_ISDIR(inode->i_mode) && !hlist_empty(&inode->i_dentry)) {
1878 spin_unlock(&inode->i_lock);
1882 __d_instantiate(entry, inode);
1883 spin_unlock(&inode->i_lock);
1884 security_d_instantiate(entry, inode);
1888 EXPORT_SYMBOL(d_instantiate_no_diralias);
1890 struct dentry *d_make_root(struct inode *root_inode)
1892 struct dentry *res = NULL;
1895 static const struct qstr name = QSTR_INIT("/", 1);
1897 res = __d_alloc(root_inode->i_sb, &name);
1899 d_instantiate(res, root_inode);
1905 EXPORT_SYMBOL(d_make_root);
1907 static struct dentry * __d_find_any_alias(struct inode *inode)
1909 struct dentry *alias;
1911 if (hlist_empty(&inode->i_dentry))
1913 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
1919 * d_find_any_alias - find any alias for a given inode
1920 * @inode: inode to find an alias for
1922 * If any aliases exist for the given inode, take and return a
1923 * reference for one of them. If no aliases exist, return %NULL.
1925 struct dentry *d_find_any_alias(struct inode *inode)
1929 spin_lock(&inode->i_lock);
1930 de = __d_find_any_alias(inode);
1931 spin_unlock(&inode->i_lock);
1934 EXPORT_SYMBOL(d_find_any_alias);
1936 static struct dentry *__d_obtain_alias(struct inode *inode, int disconnected)
1938 static const struct qstr anonstring = QSTR_INIT("/", 1);
1944 return ERR_PTR(-ESTALE);
1946 return ERR_CAST(inode);
1948 res = d_find_any_alias(inode);
1952 tmp = __d_alloc(inode->i_sb, &anonstring);
1954 res = ERR_PTR(-ENOMEM);
1958 spin_lock(&inode->i_lock);
1959 res = __d_find_any_alias(inode);
1961 spin_unlock(&inode->i_lock);
1966 /* attach a disconnected dentry */
1967 add_flags = d_flags_for_inode(inode);
1970 add_flags |= DCACHE_DISCONNECTED;
1972 spin_lock(&tmp->d_lock);
1973 __d_set_inode_and_type(tmp, inode, add_flags);
1974 hlist_add_head(&tmp->d_u.d_alias, &inode->i_dentry);
1975 hlist_bl_lock(&tmp->d_sb->s_anon);
1976 hlist_bl_add_head(&tmp->d_hash, &tmp->d_sb->s_anon);
1977 hlist_bl_unlock(&tmp->d_sb->s_anon);
1978 spin_unlock(&tmp->d_lock);
1979 spin_unlock(&inode->i_lock);
1980 security_d_instantiate(tmp, inode);
1985 if (res && !IS_ERR(res))
1986 security_d_instantiate(res, inode);
1992 * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
1993 * @inode: inode to allocate the dentry for
1995 * Obtain a dentry for an inode resulting from NFS filehandle conversion or
1996 * similar open by handle operations. The returned dentry may be anonymous,
1997 * or may have a full name (if the inode was already in the cache).
1999 * When called on a directory inode, we must ensure that the inode only ever
2000 * has one dentry. If a dentry is found, that is returned instead of
2001 * allocating a new one.
2003 * On successful return, the reference to the inode has been transferred
2004 * to the dentry. In case of an error the reference on the inode is released.
2005 * To make it easier to use in export operations a %NULL or IS_ERR inode may
2006 * be passed in and the error will be propagated to the return value,
2007 * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
2009 struct dentry *d_obtain_alias(struct inode *inode)
2011 return __d_obtain_alias(inode, 1);
2013 EXPORT_SYMBOL(d_obtain_alias);
2016 * d_obtain_root - find or allocate a dentry for a given inode
2017 * @inode: inode to allocate the dentry for
2019 * Obtain an IS_ROOT dentry for the root of a filesystem.
2021 * We must ensure that directory inodes only ever have one dentry. If a
2022 * dentry is found, that is returned instead of allocating a new one.
2024 * On successful return, the reference to the inode has been transferred
2025 * to the dentry. In case of an error the reference on the inode is
2026 * released. A %NULL or IS_ERR inode may be passed in and will be the
2027 * error will be propagate to the return value, with a %NULL @inode
2028 * replaced by ERR_PTR(-ESTALE).
2030 struct dentry *d_obtain_root(struct inode *inode)
2032 return __d_obtain_alias(inode, 0);
2034 EXPORT_SYMBOL(d_obtain_root);
2037 * d_add_ci - lookup or allocate new dentry with case-exact name
2038 * @inode: the inode case-insensitive lookup has found
2039 * @dentry: the negative dentry that was passed to the parent's lookup func
2040 * @name: the case-exact name to be associated with the returned dentry
2042 * This is to avoid filling the dcache with case-insensitive names to the
2043 * same inode, only the actual correct case is stored in the dcache for
2044 * case-insensitive filesystems.
2046 * For a case-insensitive lookup match and if the the case-exact dentry
2047 * already exists in in the dcache, use it and return it.
2049 * If no entry exists with the exact case name, allocate new dentry with
2050 * the exact case, and return the spliced entry.
2052 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2055 struct dentry *found;
2059 * First check if a dentry matching the name already exists,
2060 * if not go ahead and create it now.
2062 found = d_hash_and_lookup(dentry->d_parent, name);
2064 new = d_alloc(dentry->d_parent, name);
2066 found = ERR_PTR(-ENOMEM);
2068 found = d_splice_alias(inode, new);
2079 EXPORT_SYMBOL(d_add_ci);
2082 * Do the slow-case of the dentry name compare.
2084 * Unlike the dentry_cmp() function, we need to atomically
2085 * load the name and length information, so that the
2086 * filesystem can rely on them, and can use the 'name' and
2087 * 'len' information without worrying about walking off the
2088 * end of memory etc.
2090 * Thus the read_seqcount_retry() and the "duplicate" info
2091 * in arguments (the low-level filesystem should not look
2092 * at the dentry inode or name contents directly, since
2093 * rename can change them while we're in RCU mode).
2095 enum slow_d_compare {
2101 static noinline enum slow_d_compare slow_dentry_cmp(
2102 const struct dentry *parent,
2103 struct dentry *dentry,
2105 const struct qstr *name)
2107 int tlen = dentry->d_name.len;
2108 const char *tname = dentry->d_name.name;
2110 if (read_seqcount_retry(&dentry->d_seq, seq)) {
2112 return D_COMP_SEQRETRY;
2114 if (parent->d_op->d_compare(parent, dentry, tlen, tname, name))
2115 return D_COMP_NOMATCH;
2120 * __d_lookup_rcu - search for a dentry (racy, store-free)
2121 * @parent: parent dentry
2122 * @name: qstr of name we wish to find
2123 * @seqp: returns d_seq value at the point where the dentry was found
2124 * Returns: dentry, or NULL
2126 * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2127 * resolution (store-free path walking) design described in
2128 * Documentation/filesystems/path-lookup.txt.
2130 * This is not to be used outside core vfs.
2132 * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2133 * held, and rcu_read_lock held. The returned dentry must not be stored into
2134 * without taking d_lock and checking d_seq sequence count against @seq
2137 * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2140 * Alternatively, __d_lookup_rcu may be called again to look up the child of
2141 * the returned dentry, so long as its parent's seqlock is checked after the
2142 * child is looked up. Thus, an interlocking stepping of sequence lock checks
2143 * is formed, giving integrity down the path walk.
2145 * NOTE! The caller *has* to check the resulting dentry against the sequence
2146 * number we've returned before using any of the resulting dentry state!
2148 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2149 const struct qstr *name,
2152 u64 hashlen = name->hash_len;
2153 const unsigned char *str = name->name;
2154 struct hlist_bl_head *b = d_hash(parent, hashlen_hash(hashlen));
2155 struct hlist_bl_node *node;
2156 struct dentry *dentry;
2159 * Note: There is significant duplication with __d_lookup_rcu which is
2160 * required to prevent single threaded performance regressions
2161 * especially on architectures where smp_rmb (in seqcounts) are costly.
2162 * Keep the two functions in sync.
2166 * The hash list is protected using RCU.
2168 * Carefully use d_seq when comparing a candidate dentry, to avoid
2169 * races with d_move().
2171 * It is possible that concurrent renames can mess up our list
2172 * walk here and result in missing our dentry, resulting in the
2173 * false-negative result. d_lookup() protects against concurrent
2174 * renames using rename_lock seqlock.
2176 * See Documentation/filesystems/path-lookup.txt for more details.
2178 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2183 * The dentry sequence count protects us from concurrent
2184 * renames, and thus protects parent and name fields.
2186 * The caller must perform a seqcount check in order
2187 * to do anything useful with the returned dentry.
2189 * NOTE! We do a "raw" seqcount_begin here. That means that
2190 * we don't wait for the sequence count to stabilize if it
2191 * is in the middle of a sequence change. If we do the slow
2192 * dentry compare, we will do seqretries until it is stable,
2193 * and if we end up with a successful lookup, we actually
2194 * want to exit RCU lookup anyway.
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)) {
2203 if (dentry->d_name.hash != hashlen_hash(hashlen))
2206 switch (slow_dentry_cmp(parent, dentry, seq, name)) {
2209 case D_COMP_NOMATCH:
2216 if (dentry->d_name.hash_len != hashlen)
2219 if (!dentry_cmp(dentry, str, hashlen_len(hashlen)))
2226 * d_lookup - search for a dentry
2227 * @parent: parent dentry
2228 * @name: qstr of name we wish to find
2229 * Returns: dentry, or NULL
2231 * d_lookup searches the children of the parent dentry for the name in
2232 * question. If the dentry is found its reference count is incremented and the
2233 * dentry is returned. The caller must use dput to free the entry when it has
2234 * finished using it. %NULL is returned if the dentry does not exist.
2236 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2238 struct dentry *dentry;
2242 seq = read_seqbegin(&rename_lock);
2243 dentry = __d_lookup(parent, name);
2246 } while (read_seqretry(&rename_lock, seq));
2249 EXPORT_SYMBOL(d_lookup);
2252 * __d_lookup - search for a dentry (racy)
2253 * @parent: parent dentry
2254 * @name: qstr of name we wish to find
2255 * Returns: dentry, or NULL
2257 * __d_lookup is like d_lookup, however it may (rarely) return a
2258 * false-negative result due to unrelated rename activity.
2260 * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2261 * however it must be used carefully, eg. with a following d_lookup in
2262 * the case of failure.
2264 * __d_lookup callers must be commented.
2266 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2268 unsigned int len = name->len;
2269 unsigned int hash = name->hash;
2270 const unsigned char *str = name->name;
2271 struct hlist_bl_head *b = d_hash(parent, hash);
2272 struct hlist_bl_node *node;
2273 struct dentry *found = NULL;
2274 struct dentry *dentry;
2277 * Note: There is significant duplication with __d_lookup_rcu which is
2278 * required to prevent single threaded performance regressions
2279 * especially on architectures where smp_rmb (in seqcounts) are costly.
2280 * Keep the two functions in sync.
2284 * The hash list is protected using RCU.
2286 * Take d_lock when comparing a candidate dentry, to avoid races
2289 * It is possible that concurrent renames can mess up our list
2290 * walk here and result in missing our dentry, resulting in the
2291 * false-negative result. d_lookup() protects against concurrent
2292 * renames using rename_lock seqlock.
2294 * See Documentation/filesystems/path-lookup.txt for more details.
2298 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2300 if (dentry->d_name.hash != hash)
2303 spin_lock(&dentry->d_lock);
2304 if (dentry->d_parent != parent)
2306 if (d_unhashed(dentry))
2310 * It is safe to compare names since d_move() cannot
2311 * change the qstr (protected by d_lock).
2313 if (parent->d_flags & DCACHE_OP_COMPARE) {
2314 int tlen = dentry->d_name.len;
2315 const char *tname = dentry->d_name.name;
2316 if (parent->d_op->d_compare(parent, dentry, tlen, tname, name))
2319 if (dentry->d_name.len != len)
2321 if (dentry_cmp(dentry, str, len))
2325 dentry->d_lockref.count++;
2327 spin_unlock(&dentry->d_lock);
2330 spin_unlock(&dentry->d_lock);
2338 * d_hash_and_lookup - hash the qstr then search for a dentry
2339 * @dir: Directory to search in
2340 * @name: qstr of name we wish to find
2342 * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2344 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2347 * Check for a fs-specific hash function. Note that we must
2348 * calculate the standard hash first, as the d_op->d_hash()
2349 * routine may choose to leave the hash value unchanged.
2351 name->hash = full_name_hash(name->name, name->len);
2352 if (dir->d_flags & DCACHE_OP_HASH) {
2353 int err = dir->d_op->d_hash(dir, name);
2354 if (unlikely(err < 0))
2355 return ERR_PTR(err);
2357 return d_lookup(dir, name);
2359 EXPORT_SYMBOL(d_hash_and_lookup);
2362 * When a file is deleted, we have two options:
2363 * - turn this dentry into a negative dentry
2364 * - unhash this dentry and free it.
2366 * Usually, we want to just turn this into
2367 * a negative dentry, but if anybody else is
2368 * currently using the dentry or the inode
2369 * we can't do that and we fall back on removing
2370 * it from the hash queues and waiting for
2371 * it to be deleted later when it has no users
2375 * d_delete - delete a dentry
2376 * @dentry: The dentry to delete
2378 * Turn the dentry into a negative dentry if possible, otherwise
2379 * remove it from the hash queues so it can be deleted later
2382 void d_delete(struct dentry * dentry)
2384 struct inode *inode;
2387 * Are we the only user?
2390 spin_lock(&dentry->d_lock);
2391 inode = dentry->d_inode;
2392 isdir = S_ISDIR(inode->i_mode);
2393 if (dentry->d_lockref.count == 1) {
2394 if (!spin_trylock(&inode->i_lock)) {
2395 spin_unlock(&dentry->d_lock);
2399 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2400 dentry_unlink_inode(dentry);
2401 fsnotify_nameremove(dentry, isdir);
2405 if (!d_unhashed(dentry))
2408 spin_unlock(&dentry->d_lock);
2410 fsnotify_nameremove(dentry, isdir);
2412 EXPORT_SYMBOL(d_delete);
2414 static void __d_rehash(struct dentry * entry, struct hlist_bl_head *b)
2416 BUG_ON(!d_unhashed(entry));
2418 entry->d_flags |= DCACHE_RCUACCESS;
2419 hlist_bl_add_head_rcu(&entry->d_hash, b);
2423 static void _d_rehash(struct dentry * entry)
2425 __d_rehash(entry, d_hash(entry->d_parent, entry->d_name.hash));
2429 * d_rehash - add an entry back to the hash
2430 * @entry: dentry to add to the hash
2432 * Adds a dentry to the hash according to its name.
2435 void d_rehash(struct dentry * entry)
2437 spin_lock(&entry->d_lock);
2439 spin_unlock(&entry->d_lock);
2441 EXPORT_SYMBOL(d_rehash);
2444 * dentry_update_name_case - update case insensitive dentry with a new name
2445 * @dentry: dentry to be updated
2448 * Update a case insensitive dentry with new case of name.
2450 * dentry must have been returned by d_lookup with name @name. Old and new
2451 * name lengths must match (ie. no d_compare which allows mismatched name
2454 * Parent inode i_mutex must be held over d_lookup and into this call (to
2455 * keep renames and concurrent inserts, and readdir(2) away).
2457 void dentry_update_name_case(struct dentry *dentry, struct qstr *name)
2459 BUG_ON(!mutex_is_locked(&dentry->d_parent->d_inode->i_mutex));
2460 BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */
2462 spin_lock(&dentry->d_lock);
2463 write_seqcount_begin(&dentry->d_seq);
2464 memcpy((unsigned char *)dentry->d_name.name, name->name, name->len);
2465 write_seqcount_end(&dentry->d_seq);
2466 spin_unlock(&dentry->d_lock);
2468 EXPORT_SYMBOL(dentry_update_name_case);
2470 static void swap_names(struct dentry *dentry, struct dentry *target)
2472 if (unlikely(dname_external(target))) {
2473 if (unlikely(dname_external(dentry))) {
2475 * Both external: swap the pointers
2477 swap(target->d_name.name, dentry->d_name.name);
2480 * dentry:internal, target:external. Steal target's
2481 * storage and make target internal.
2483 memcpy(target->d_iname, dentry->d_name.name,
2484 dentry->d_name.len + 1);
2485 dentry->d_name.name = target->d_name.name;
2486 target->d_name.name = target->d_iname;
2489 if (unlikely(dname_external(dentry))) {
2491 * dentry:external, target:internal. Give dentry's
2492 * storage to target and make dentry internal
2494 memcpy(dentry->d_iname, target->d_name.name,
2495 target->d_name.len + 1);
2496 target->d_name.name = dentry->d_name.name;
2497 dentry->d_name.name = dentry->d_iname;
2500 * Both are internal.
2503 BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2504 kmemcheck_mark_initialized(dentry->d_iname, DNAME_INLINE_LEN);
2505 kmemcheck_mark_initialized(target->d_iname, DNAME_INLINE_LEN);
2506 for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2507 swap(((long *) &dentry->d_iname)[i],
2508 ((long *) &target->d_iname)[i]);
2512 swap(dentry->d_name.hash_len, target->d_name.hash_len);
2515 static void copy_name(struct dentry *dentry, struct dentry *target)
2517 struct external_name *old_name = NULL;
2518 if (unlikely(dname_external(dentry)))
2519 old_name = external_name(dentry);
2520 if (unlikely(dname_external(target))) {
2521 atomic_inc(&external_name(target)->u.count);
2522 dentry->d_name = target->d_name;
2524 memcpy(dentry->d_iname, target->d_name.name,
2525 target->d_name.len + 1);
2526 dentry->d_name.name = dentry->d_iname;
2527 dentry->d_name.hash_len = target->d_name.hash_len;
2529 if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2530 kfree_rcu(old_name, u.head);
2533 static void dentry_lock_for_move(struct dentry *dentry, struct dentry *target)
2536 * XXXX: do we really need to take target->d_lock?
2538 if (IS_ROOT(dentry) || dentry->d_parent == target->d_parent)
2539 spin_lock(&target->d_parent->d_lock);
2541 if (d_ancestor(dentry->d_parent, target->d_parent)) {
2542 spin_lock(&dentry->d_parent->d_lock);
2543 spin_lock_nested(&target->d_parent->d_lock,
2544 DENTRY_D_LOCK_NESTED);
2546 spin_lock(&target->d_parent->d_lock);
2547 spin_lock_nested(&dentry->d_parent->d_lock,
2548 DENTRY_D_LOCK_NESTED);
2551 if (target < dentry) {
2552 spin_lock_nested(&target->d_lock, 2);
2553 spin_lock_nested(&dentry->d_lock, 3);
2555 spin_lock_nested(&dentry->d_lock, 2);
2556 spin_lock_nested(&target->d_lock, 3);
2560 static void dentry_unlock_for_move(struct dentry *dentry, struct dentry *target)
2562 if (target->d_parent != dentry->d_parent)
2563 spin_unlock(&dentry->d_parent->d_lock);
2564 if (target->d_parent != target)
2565 spin_unlock(&target->d_parent->d_lock);
2566 spin_unlock(&target->d_lock);
2567 spin_unlock(&dentry->d_lock);
2571 * When switching names, the actual string doesn't strictly have to
2572 * be preserved in the target - because we're dropping the target
2573 * anyway. As such, we can just do a simple memcpy() to copy over
2574 * the new name before we switch, unless we are going to rehash
2575 * it. Note that if we *do* unhash the target, we are not allowed
2576 * to rehash it without giving it a new name/hash key - whether
2577 * we swap or overwrite the names here, resulting name won't match
2578 * the reality in filesystem; it's only there for d_path() purposes.
2579 * Note that all of this is happening under rename_lock, so the
2580 * any hash lookup seeing it in the middle of manipulations will
2581 * be discarded anyway. So we do not care what happens to the hash
2585 * __d_move - move a dentry
2586 * @dentry: entry to move
2587 * @target: new dentry
2588 * @exchange: exchange the two dentries
2590 * Update the dcache to reflect the move of a file name. Negative
2591 * dcache entries should not be moved in this way. Caller must hold
2592 * rename_lock, the i_mutex of the source and target directories,
2593 * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2595 static void __d_move(struct dentry *dentry, struct dentry *target,
2598 if (!dentry->d_inode)
2599 printk(KERN_WARNING "VFS: moving negative dcache entry\n");
2601 BUG_ON(d_ancestor(dentry, target));
2602 BUG_ON(d_ancestor(target, dentry));
2604 dentry_lock_for_move(dentry, target);
2606 write_seqcount_begin(&dentry->d_seq);
2607 write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2609 /* __d_drop does write_seqcount_barrier, but they're OK to nest. */
2612 * Move the dentry to the target hash queue. Don't bother checking
2613 * for the same hash queue because of how unlikely it is.
2616 __d_rehash(dentry, d_hash(target->d_parent, target->d_name.hash));
2619 * Unhash the target (d_delete() is not usable here). If exchanging
2620 * the two dentries, then rehash onto the other's hash queue.
2625 d_hash(dentry->d_parent, dentry->d_name.hash));
2628 /* Switch the names.. */
2630 swap_names(dentry, target);
2632 copy_name(dentry, target);
2634 /* ... and switch them in the tree */
2635 if (IS_ROOT(dentry)) {
2636 /* splicing a tree */
2637 dentry->d_parent = target->d_parent;
2638 target->d_parent = target;
2639 list_del_init(&target->d_child);
2640 list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2642 /* swapping two dentries */
2643 swap(dentry->d_parent, target->d_parent);
2644 list_move(&target->d_child, &target->d_parent->d_subdirs);
2645 list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2647 fsnotify_d_move(target);
2648 fsnotify_d_move(dentry);
2651 write_seqcount_end(&target->d_seq);
2652 write_seqcount_end(&dentry->d_seq);
2654 dentry_unlock_for_move(dentry, target);
2658 * d_move - move a dentry
2659 * @dentry: entry to move
2660 * @target: new dentry
2662 * Update the dcache to reflect the move of a file name. Negative
2663 * dcache entries should not be moved in this way. See the locking
2664 * requirements for __d_move.
2666 void d_move(struct dentry *dentry, struct dentry *target)
2668 write_seqlock(&rename_lock);
2669 __d_move(dentry, target, false);
2670 write_sequnlock(&rename_lock);
2672 EXPORT_SYMBOL(d_move);
2675 * d_exchange - exchange two dentries
2676 * @dentry1: first dentry
2677 * @dentry2: second dentry
2679 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2681 write_seqlock(&rename_lock);
2683 WARN_ON(!dentry1->d_inode);
2684 WARN_ON(!dentry2->d_inode);
2685 WARN_ON(IS_ROOT(dentry1));
2686 WARN_ON(IS_ROOT(dentry2));
2688 __d_move(dentry1, dentry2, true);
2690 write_sequnlock(&rename_lock);
2694 * d_ancestor - search for an ancestor
2695 * @p1: ancestor dentry
2698 * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2699 * an ancestor of p2, else NULL.
2701 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2705 for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2706 if (p->d_parent == p1)
2713 * This helper attempts to cope with remotely renamed directories
2715 * It assumes that the caller is already holding
2716 * dentry->d_parent->d_inode->i_mutex, inode->i_lock and rename_lock
2718 * Note: If ever the locking in lock_rename() changes, then please
2719 * remember to update this too...
2721 static int __d_unalias(struct inode *inode,
2722 struct dentry *dentry, struct dentry *alias)
2724 struct mutex *m1 = NULL, *m2 = NULL;
2727 /* If alias and dentry share a parent, then no extra locks required */
2728 if (alias->d_parent == dentry->d_parent)
2731 /* See lock_rename() */
2732 if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2734 m1 = &dentry->d_sb->s_vfs_rename_mutex;
2735 if (!mutex_trylock(&alias->d_parent->d_inode->i_mutex))
2737 m2 = &alias->d_parent->d_inode->i_mutex;
2739 __d_move(alias, dentry, false);
2742 spin_unlock(&inode->i_lock);
2751 * d_splice_alias - splice a disconnected dentry into the tree if one exists
2752 * @inode: the inode which may have a disconnected dentry
2753 * @dentry: a negative dentry which we want to point to the inode.
2755 * If inode is a directory and has an IS_ROOT alias, then d_move that in
2756 * place of the given dentry and return it, else simply d_add the inode
2757 * to the dentry and return NULL.
2759 * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2760 * we should error out: directories can't have multiple aliases.
2762 * This is needed in the lookup routine of any filesystem that is exportable
2763 * (via knfsd) so that we can build dcache paths to directories effectively.
2765 * If a dentry was found and moved, then it is returned. Otherwise NULL
2766 * is returned. This matches the expected return value of ->lookup.
2768 * Cluster filesystems may call this function with a negative, hashed dentry.
2769 * In that case, we know that the inode will be a regular file, and also this
2770 * will only occur during atomic_open. So we need to check for the dentry
2771 * being already hashed only in the final case.
2773 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
2776 return ERR_CAST(inode);
2778 BUG_ON(!d_unhashed(dentry));
2781 __d_instantiate(dentry, NULL);
2784 spin_lock(&inode->i_lock);
2785 if (S_ISDIR(inode->i_mode)) {
2786 struct dentry *new = __d_find_any_alias(inode);
2787 if (unlikely(new)) {
2788 write_seqlock(&rename_lock);
2789 if (unlikely(d_ancestor(new, dentry))) {
2790 write_sequnlock(&rename_lock);
2791 spin_unlock(&inode->i_lock);
2793 new = ERR_PTR(-ELOOP);
2794 pr_warn_ratelimited(
2795 "VFS: Lookup of '%s' in %s %s"
2796 " would have caused loop\n",
2797 dentry->d_name.name,
2798 inode->i_sb->s_type->name,
2800 } else if (!IS_ROOT(new)) {
2801 int err = __d_unalias(inode, dentry, new);
2802 write_sequnlock(&rename_lock);
2808 __d_move(new, dentry, false);
2809 write_sequnlock(&rename_lock);
2810 spin_unlock(&inode->i_lock);
2811 security_d_instantiate(new, inode);
2817 /* already taking inode->i_lock, so d_add() by hand */
2818 __d_instantiate(dentry, inode);
2819 spin_unlock(&inode->i_lock);
2821 security_d_instantiate(dentry, inode);
2825 EXPORT_SYMBOL(d_splice_alias);
2827 static int prepend(char **buffer, int *buflen, const char *str, int namelen)
2831 return -ENAMETOOLONG;
2833 memcpy(*buffer, str, namelen);
2838 * prepend_name - prepend a pathname in front of current buffer pointer
2839 * @buffer: buffer pointer
2840 * @buflen: allocated length of the buffer
2841 * @name: name string and length qstr structure
2843 * With RCU path tracing, it may race with d_move(). Use ACCESS_ONCE() to
2844 * make sure that either the old or the new name pointer and length are
2845 * fetched. However, there may be mismatch between length and pointer.
2846 * The length cannot be trusted, we need to copy it byte-by-byte until
2847 * the length is reached or a null byte is found. It also prepends "/" at
2848 * the beginning of the name. The sequence number check at the caller will
2849 * retry it again when a d_move() does happen. So any garbage in the buffer
2850 * due to mismatched pointer and length will be discarded.
2852 * Data dependency barrier is needed to make sure that we see that terminating
2853 * NUL. Alpha strikes again, film at 11...
2855 static int prepend_name(char **buffer, int *buflen, struct qstr *name)
2857 const char *dname = ACCESS_ONCE(name->name);
2858 u32 dlen = ACCESS_ONCE(name->len);
2861 smp_read_barrier_depends();
2863 *buflen -= dlen + 1;
2865 return -ENAMETOOLONG;
2866 p = *buffer -= dlen + 1;
2878 * prepend_path - Prepend path string to a buffer
2879 * @path: the dentry/vfsmount to report
2880 * @root: root vfsmnt/dentry
2881 * @buffer: pointer to the end of the buffer
2882 * @buflen: pointer to buffer length
2884 * The function will first try to write out the pathname without taking any
2885 * lock other than the RCU read lock to make sure that dentries won't go away.
2886 * It only checks the sequence number of the global rename_lock as any change
2887 * in the dentry's d_seq will be preceded by changes in the rename_lock
2888 * sequence number. If the sequence number had been changed, it will restart
2889 * the whole pathname back-tracing sequence again by taking the rename_lock.
2890 * In this case, there is no need to take the RCU read lock as the recursive
2891 * parent pointer references will keep the dentry chain alive as long as no
2892 * rename operation is performed.
2894 static int prepend_path(const struct path *path,
2895 const struct path *root,
2896 char **buffer, int *buflen)
2898 struct dentry *dentry;
2899 struct vfsmount *vfsmnt;
2902 unsigned seq, m_seq = 0;
2908 read_seqbegin_or_lock(&mount_lock, &m_seq);
2915 dentry = path->dentry;
2917 mnt = real_mount(vfsmnt);
2918 read_seqbegin_or_lock(&rename_lock, &seq);
2919 while (dentry != root->dentry || vfsmnt != root->mnt) {
2920 struct dentry * parent;
2922 if (dentry == vfsmnt->mnt_root || IS_ROOT(dentry)) {
2923 struct mount *parent = ACCESS_ONCE(mnt->mnt_parent);
2925 if (mnt != parent) {
2926 dentry = ACCESS_ONCE(mnt->mnt_mountpoint);
2932 error = is_mounted(vfsmnt) ? 1 : 2;
2935 parent = dentry->d_parent;
2937 error = prepend_name(&bptr, &blen, &dentry->d_name);
2945 if (need_seqretry(&rename_lock, seq)) {
2949 done_seqretry(&rename_lock, seq);
2953 if (need_seqretry(&mount_lock, m_seq)) {
2957 done_seqretry(&mount_lock, m_seq);
2959 if (error >= 0 && bptr == *buffer) {
2961 error = -ENAMETOOLONG;
2971 * __d_path - return the path of a dentry
2972 * @path: the dentry/vfsmount to report
2973 * @root: root vfsmnt/dentry
2974 * @buf: buffer to return value in
2975 * @buflen: buffer length
2977 * Convert a dentry into an ASCII path name.
2979 * Returns a pointer into the buffer or an error code if the
2980 * path was too long.
2982 * "buflen" should be positive.
2984 * If the path is not reachable from the supplied root, return %NULL.
2986 char *__d_path(const struct path *path,
2987 const struct path *root,
2988 char *buf, int buflen)
2990 char *res = buf + buflen;
2993 prepend(&res, &buflen, "\0", 1);
2994 error = prepend_path(path, root, &res, &buflen);
2997 return ERR_PTR(error);
3003 char *d_absolute_path(const struct path *path,
3004 char *buf, int buflen)
3006 struct path root = {};
3007 char *res = buf + buflen;
3010 prepend(&res, &buflen, "\0", 1);
3011 error = prepend_path(path, &root, &res, &buflen);
3016 return ERR_PTR(error);
3021 * same as __d_path but appends "(deleted)" for unlinked files.
3023 static int path_with_deleted(const struct path *path,
3024 const struct path *root,
3025 char **buf, int *buflen)
3027 prepend(buf, buflen, "\0", 1);
3028 if (d_unlinked(path->dentry)) {
3029 int error = prepend(buf, buflen, " (deleted)", 10);
3034 return prepend_path(path, root, buf, buflen);
3037 static int prepend_unreachable(char **buffer, int *buflen)
3039 return prepend(buffer, buflen, "(unreachable)", 13);
3042 static void get_fs_root_rcu(struct fs_struct *fs, struct path *root)
3047 seq = read_seqcount_begin(&fs->seq);
3049 } while (read_seqcount_retry(&fs->seq, seq));
3053 * d_path - return the path of a dentry
3054 * @path: path to report
3055 * @buf: buffer to return value in
3056 * @buflen: buffer length
3058 * Convert a dentry into an ASCII path name. If the entry has been deleted
3059 * the string " (deleted)" is appended. Note that this is ambiguous.
3061 * Returns a pointer into the buffer or an error code if the path was
3062 * too long. Note: Callers should use the returned pointer, not the passed
3063 * in buffer, to use the name! The implementation often starts at an offset
3064 * into the buffer, and may leave 0 bytes at the start.
3066 * "buflen" should be positive.
3068 char *d_path(const struct path *path, char *buf, int buflen)
3070 char *res = buf + buflen;
3075 * We have various synthetic filesystems that never get mounted. On
3076 * these filesystems dentries are never used for lookup purposes, and
3077 * thus don't need to be hashed. They also don't need a name until a
3078 * user wants to identify the object in /proc/pid/fd/. The little hack
3079 * below allows us to generate a name for these objects on demand:
3081 * Some pseudo inodes are mountable. When they are mounted
3082 * path->dentry == path->mnt->mnt_root. In that case don't call d_dname
3083 * and instead have d_path return the mounted path.
3085 if (path->dentry->d_op && path->dentry->d_op->d_dname &&
3086 (!IS_ROOT(path->dentry) || path->dentry != path->mnt->mnt_root))
3087 return path->dentry->d_op->d_dname(path->dentry, buf, buflen);
3090 get_fs_root_rcu(current->fs, &root);
3091 error = path_with_deleted(path, &root, &res, &buflen);
3095 res = ERR_PTR(error);
3098 EXPORT_SYMBOL(d_path);
3101 * Helper function for dentry_operations.d_dname() members
3103 char *dynamic_dname(struct dentry *dentry, char *buffer, int buflen,
3104 const char *fmt, ...)
3110 va_start(args, fmt);
3111 sz = vsnprintf(temp, sizeof(temp), fmt, args) + 1;
3114 if (sz > sizeof(temp) || sz > buflen)
3115 return ERR_PTR(-ENAMETOOLONG);
3117 buffer += buflen - sz;
3118 return memcpy(buffer, temp, sz);
3121 char *simple_dname(struct dentry *dentry, char *buffer, int buflen)
3123 char *end = buffer + buflen;
3124 /* these dentries are never renamed, so d_lock is not needed */
3125 if (prepend(&end, &buflen, " (deleted)", 11) ||
3126 prepend(&end, &buflen, dentry->d_name.name, dentry->d_name.len) ||
3127 prepend(&end, &buflen, "/", 1))
3128 end = ERR_PTR(-ENAMETOOLONG);
3131 EXPORT_SYMBOL(simple_dname);
3134 * Write full pathname from the root of the filesystem into the buffer.
3136 static char *__dentry_path(struct dentry *d, char *buf, int buflen)
3138 struct dentry *dentry;
3151 prepend(&end, &len, "\0", 1);
3155 read_seqbegin_or_lock(&rename_lock, &seq);
3156 while (!IS_ROOT(dentry)) {
3157 struct dentry *parent = dentry->d_parent;
3160 error = prepend_name(&end, &len, &dentry->d_name);
3169 if (need_seqretry(&rename_lock, seq)) {
3173 done_seqretry(&rename_lock, seq);
3178 return ERR_PTR(-ENAMETOOLONG);
3181 char *dentry_path_raw(struct dentry *dentry, char *buf, int buflen)
3183 return __dentry_path(dentry, buf, buflen);
3185 EXPORT_SYMBOL(dentry_path_raw);
3187 char *dentry_path(struct dentry *dentry, char *buf, int buflen)
3192 if (d_unlinked(dentry)) {
3194 if (prepend(&p, &buflen, "//deleted", 10) != 0)
3198 retval = __dentry_path(dentry, buf, buflen);
3199 if (!IS_ERR(retval) && p)
3200 *p = '/'; /* restore '/' overriden with '\0' */
3203 return ERR_PTR(-ENAMETOOLONG);
3206 static void get_fs_root_and_pwd_rcu(struct fs_struct *fs, struct path *root,
3212 seq = read_seqcount_begin(&fs->seq);
3215 } while (read_seqcount_retry(&fs->seq, seq));
3219 * NOTE! The user-level library version returns a
3220 * character pointer. The kernel system call just
3221 * returns the length of the buffer filled (which
3222 * includes the ending '\0' character), or a negative
3223 * error value. So libc would do something like
3225 * char *getcwd(char * buf, size_t size)
3229 * retval = sys_getcwd(buf, size);
3236 SYSCALL_DEFINE2(getcwd, char __user *, buf, unsigned long, size)
3239 struct path pwd, root;
3240 char *page = __getname();
3246 get_fs_root_and_pwd_rcu(current->fs, &root, &pwd);
3249 if (!d_unlinked(pwd.dentry)) {
3251 char *cwd = page + PATH_MAX;
3252 int buflen = PATH_MAX;
3254 prepend(&cwd, &buflen, "\0", 1);
3255 error = prepend_path(&pwd, &root, &cwd, &buflen);
3261 /* Unreachable from current root */
3263 error = prepend_unreachable(&cwd, &buflen);
3269 len = PATH_MAX + page - cwd;
3272 if (copy_to_user(buf, cwd, len))
3285 * Test whether new_dentry is a subdirectory of old_dentry.
3287 * Trivially implemented using the dcache structure
3291 * is_subdir - is new dentry a subdirectory of old_dentry
3292 * @new_dentry: new dentry
3293 * @old_dentry: old dentry
3295 * Returns 1 if new_dentry is a subdirectory of the parent (at any depth).
3296 * Returns 0 otherwise.
3297 * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3300 int is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3305 if (new_dentry == old_dentry)
3309 /* for restarting inner loop in case of seq retry */
3310 seq = read_seqbegin(&rename_lock);
3312 * Need rcu_readlock to protect against the d_parent trashing
3316 if (d_ancestor(old_dentry, new_dentry))
3321 } while (read_seqretry(&rename_lock, seq));
3326 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3328 struct dentry *root = data;
3329 if (dentry != root) {
3330 if (d_unhashed(dentry) || !dentry->d_inode)
3333 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3334 dentry->d_flags |= DCACHE_GENOCIDE;
3335 dentry->d_lockref.count--;
3338 return D_WALK_CONTINUE;
3341 void d_genocide(struct dentry *parent)
3343 d_walk(parent, parent, d_genocide_kill, NULL);
3346 void d_tmpfile(struct dentry *dentry, struct inode *inode)
3348 inode_dec_link_count(inode);
3349 BUG_ON(dentry->d_name.name != dentry->d_iname ||
3350 !hlist_unhashed(&dentry->d_u.d_alias) ||
3351 !d_unlinked(dentry));
3352 spin_lock(&dentry->d_parent->d_lock);
3353 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3354 dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3355 (unsigned long long)inode->i_ino);
3356 spin_unlock(&dentry->d_lock);
3357 spin_unlock(&dentry->d_parent->d_lock);
3358 d_instantiate(dentry, inode);
3360 EXPORT_SYMBOL(d_tmpfile);
3362 static __initdata unsigned long dhash_entries;
3363 static int __init set_dhash_entries(char *str)
3367 dhash_entries = simple_strtoul(str, &str, 0);
3370 __setup("dhash_entries=", set_dhash_entries);
3372 static void __init dcache_init_early(void)
3376 /* If hashes are distributed across NUMA nodes, defer
3377 * hash allocation until vmalloc space is available.
3383 alloc_large_system_hash("Dentry cache",
3384 sizeof(struct hlist_bl_head),
3393 for (loop = 0; loop < (1U << d_hash_shift); loop++)
3394 INIT_HLIST_BL_HEAD(dentry_hashtable + loop);
3397 static void __init dcache_init(void)
3402 * A constructor could be added for stable state like the lists,
3403 * but it is probably not worth it because of the cache nature
3406 dentry_cache = KMEM_CACHE(dentry,
3407 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD);
3409 /* Hash may have been set up in dcache_init_early */
3414 alloc_large_system_hash("Dentry cache",
3415 sizeof(struct hlist_bl_head),
3424 for (loop = 0; loop < (1U << d_hash_shift); loop++)
3425 INIT_HLIST_BL_HEAD(dentry_hashtable + loop);
3428 /* SLAB cache for __getname() consumers */
3429 struct kmem_cache *names_cachep __read_mostly;
3430 EXPORT_SYMBOL(names_cachep);
3432 EXPORT_SYMBOL(d_genocide);
3434 void __init vfs_caches_init_early(void)
3436 dcache_init_early();
3440 void __init vfs_caches_init(unsigned long mempages)
3442 unsigned long reserve;
3444 /* Base hash sizes on available memory, with a reserve equal to
3445 150% of current kernel size */
3447 reserve = min((mempages - nr_free_pages()) * 3/2, mempages - 1);
3448 mempages -= reserve;
3450 names_cachep = kmem_cache_create("names_cache", PATH_MAX, 0,
3451 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3455 files_init(mempages);