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/slab.h>
23 #include <linux/init.h>
24 #include <linux/hash.h>
25 #include <linux/cache.h>
26 #include <linux/export.h>
27 #include <linux/mount.h>
28 #include <linux/file.h>
29 #include <asm/uaccess.h>
30 #include <linux/security.h>
31 #include <linux/seqlock.h>
32 #include <linux/swap.h>
33 #include <linux/bootmem.h>
34 #include <linux/fs_struct.h>
35 #include <linux/hardirq.h>
36 #include <linux/bit_spinlock.h>
37 #include <linux/rculist_bl.h>
38 #include <linux/prefetch.h>
39 #include <linux/ratelimit.h>
40 #include <linux/list_lru.h>
46 * dcache->d_inode->i_lock protects:
47 * - i_dentry, d_alias, d_inode of aliases
48 * dcache_hash_bucket lock protects:
49 * - the dcache hash table
50 * s_anon bl list spinlock protects:
51 * - the s_anon list (see __d_drop)
52 * dentry->d_sb->s_dentry_lru_lock protects:
53 * - the dcache lru lists and counters
60 * - d_parent and d_subdirs
61 * - childrens' d_child and d_parent
65 * dentry->d_inode->i_lock
67 * dentry->d_sb->s_dentry_lru_lock
68 * dcache_hash_bucket lock
71 * If there is an ancestor relationship:
72 * dentry->d_parent->...->d_parent->d_lock
74 * dentry->d_parent->d_lock
77 * If no ancestor relationship:
78 * if (dentry1 < dentry2)
82 int sysctl_vfs_cache_pressure __read_mostly = 100;
83 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
85 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
87 EXPORT_SYMBOL(rename_lock);
89 static struct kmem_cache *dentry_cache __read_mostly;
92 * read_seqbegin_or_lock - begin a sequence number check or locking block
93 * @lock: sequence lock
94 * @seq : sequence number to be checked
96 * First try it once optimistically without taking the lock. If that fails,
97 * take the lock. The sequence number is also used as a marker for deciding
98 * whether to be a reader (even) or writer (odd).
99 * N.B. seq must be initialized to an even number to begin with.
101 static inline void read_seqbegin_or_lock(seqlock_t *lock, int *seq)
103 if (!(*seq & 1)) /* Even */
104 *seq = read_seqbegin(lock);
106 read_seqlock_excl(lock);
109 static inline int need_seqretry(seqlock_t *lock, int seq)
111 return !(seq & 1) && read_seqretry(lock, seq);
114 static inline void done_seqretry(seqlock_t *lock, int seq)
117 read_sequnlock_excl(lock);
121 * This is the single most critical data structure when it comes
122 * to the dcache: the hashtable for lookups. Somebody should try
123 * to make this good - I've just made it work.
125 * This hash-function tries to avoid losing too many bits of hash
126 * information, yet avoid using a prime hash-size or similar.
128 #define D_HASHBITS d_hash_shift
129 #define D_HASHMASK d_hash_mask
131 static unsigned int d_hash_mask __read_mostly;
132 static unsigned int d_hash_shift __read_mostly;
134 static struct hlist_bl_head *dentry_hashtable __read_mostly;
136 static inline struct hlist_bl_head *d_hash(const struct dentry *parent,
139 hash += (unsigned long) parent / L1_CACHE_BYTES;
140 hash = hash + (hash >> D_HASHBITS);
141 return dentry_hashtable + (hash & D_HASHMASK);
144 /* Statistics gathering. */
145 struct dentry_stat_t dentry_stat = {
149 static DEFINE_PER_CPU(long, nr_dentry);
150 static DEFINE_PER_CPU(long, nr_dentry_unused);
152 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
155 * Here we resort to our own counters instead of using generic per-cpu counters
156 * for consistency with what the vfs inode code does. We are expected to harvest
157 * better code and performance by having our own specialized counters.
159 * Please note that the loop is done over all possible CPUs, not over all online
160 * CPUs. The reason for this is that we don't want to play games with CPUs going
161 * on and off. If one of them goes off, we will just keep their counters.
163 * glommer: See cffbc8a for details, and if you ever intend to change this,
164 * please update all vfs counters to match.
166 static long get_nr_dentry(void)
170 for_each_possible_cpu(i)
171 sum += per_cpu(nr_dentry, i);
172 return sum < 0 ? 0 : sum;
175 static long get_nr_dentry_unused(void)
179 for_each_possible_cpu(i)
180 sum += per_cpu(nr_dentry_unused, i);
181 return sum < 0 ? 0 : sum;
184 int proc_nr_dentry(ctl_table *table, int write, void __user *buffer,
185 size_t *lenp, loff_t *ppos)
187 dentry_stat.nr_dentry = get_nr_dentry();
188 dentry_stat.nr_unused = get_nr_dentry_unused();
189 return proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
194 * Compare 2 name strings, return 0 if they match, otherwise non-zero.
195 * The strings are both count bytes long, and count is non-zero.
197 #ifdef CONFIG_DCACHE_WORD_ACCESS
199 #include <asm/word-at-a-time.h>
201 * NOTE! 'cs' and 'scount' come from a dentry, so it has a
202 * aligned allocation for this particular component. We don't
203 * strictly need the load_unaligned_zeropad() safety, but it
204 * doesn't hurt either.
206 * In contrast, 'ct' and 'tcount' can be from a pathname, and do
207 * need the careful unaligned handling.
209 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
211 unsigned long a,b,mask;
214 a = *(unsigned long *)cs;
215 b = load_unaligned_zeropad(ct);
216 if (tcount < sizeof(unsigned long))
218 if (unlikely(a != b))
220 cs += sizeof(unsigned long);
221 ct += sizeof(unsigned long);
222 tcount -= sizeof(unsigned long);
226 mask = ~(~0ul << tcount*8);
227 return unlikely(!!((a ^ b) & mask));
232 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
246 static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount)
248 const unsigned char *cs;
250 * Be careful about RCU walk racing with rename:
251 * use ACCESS_ONCE to fetch the name pointer.
253 * NOTE! Even if a rename will mean that the length
254 * was not loaded atomically, we don't care. The
255 * RCU walk will check the sequence count eventually,
256 * and catch it. And we won't overrun the buffer,
257 * because we're reading the name pointer atomically,
258 * and a dentry name is guaranteed to be properly
259 * terminated with a NUL byte.
261 * End result: even if 'len' is wrong, we'll exit
262 * early because the data cannot match (there can
263 * be no NUL in the ct/tcount data)
265 cs = ACCESS_ONCE(dentry->d_name.name);
266 smp_read_barrier_depends();
267 return dentry_string_cmp(cs, ct, tcount);
270 static void __d_free(struct rcu_head *head)
272 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
274 WARN_ON(!hlist_unhashed(&dentry->d_alias));
275 if (dname_external(dentry))
276 kfree(dentry->d_name.name);
277 kmem_cache_free(dentry_cache, dentry);
283 static void d_free(struct dentry *dentry)
285 BUG_ON((int)dentry->d_lockref.count > 0);
286 this_cpu_dec(nr_dentry);
287 if (dentry->d_op && dentry->d_op->d_release)
288 dentry->d_op->d_release(dentry);
290 /* if dentry was never visible to RCU, immediate free is OK */
291 if (!(dentry->d_flags & DCACHE_RCUACCESS))
292 __d_free(&dentry->d_u.d_rcu);
294 call_rcu(&dentry->d_u.d_rcu, __d_free);
298 * dentry_rcuwalk_barrier - invalidate in-progress rcu-walk lookups
299 * @dentry: the target dentry
300 * After this call, in-progress rcu-walk path lookup will fail. This
301 * should be called after unhashing, and after changing d_inode (if
302 * the dentry has not already been unhashed).
304 static inline void dentry_rcuwalk_barrier(struct dentry *dentry)
306 assert_spin_locked(&dentry->d_lock);
307 /* Go through a barrier */
308 write_seqcount_barrier(&dentry->d_seq);
312 * Release the dentry's inode, using the filesystem
313 * d_iput() operation if defined. Dentry has no refcount
316 static void dentry_iput(struct dentry * dentry)
317 __releases(dentry->d_lock)
318 __releases(dentry->d_inode->i_lock)
320 struct inode *inode = dentry->d_inode;
322 dentry->d_inode = NULL;
323 hlist_del_init(&dentry->d_alias);
324 spin_unlock(&dentry->d_lock);
325 spin_unlock(&inode->i_lock);
327 fsnotify_inoderemove(inode);
328 if (dentry->d_op && dentry->d_op->d_iput)
329 dentry->d_op->d_iput(dentry, inode);
333 spin_unlock(&dentry->d_lock);
338 * Release the dentry's inode, using the filesystem
339 * d_iput() operation if defined. dentry remains in-use.
341 static void dentry_unlink_inode(struct dentry * dentry)
342 __releases(dentry->d_lock)
343 __releases(dentry->d_inode->i_lock)
345 struct inode *inode = dentry->d_inode;
346 dentry->d_inode = NULL;
347 hlist_del_init(&dentry->d_alias);
348 dentry_rcuwalk_barrier(dentry);
349 spin_unlock(&dentry->d_lock);
350 spin_unlock(&inode->i_lock);
352 fsnotify_inoderemove(inode);
353 if (dentry->d_op && dentry->d_op->d_iput)
354 dentry->d_op->d_iput(dentry, inode);
360 * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
361 * is in use - which includes both the "real" per-superblock
362 * LRU list _and_ the DCACHE_SHRINK_LIST use.
364 * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
365 * on the shrink list (ie not on the superblock LRU list).
367 * The per-cpu "nr_dentry_unused" counters are updated with
368 * the DCACHE_LRU_LIST bit.
370 * These helper functions make sure we always follow the
371 * rules. d_lock must be held by the caller.
373 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
374 static void d_lru_add(struct dentry *dentry)
376 D_FLAG_VERIFY(dentry, 0);
377 dentry->d_flags |= DCACHE_LRU_LIST;
378 this_cpu_inc(nr_dentry_unused);
379 WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
382 static void d_lru_del(struct dentry *dentry)
384 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
385 dentry->d_flags &= ~DCACHE_LRU_LIST;
386 this_cpu_dec(nr_dentry_unused);
387 WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
390 static void d_shrink_del(struct dentry *dentry)
392 D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
393 list_del_init(&dentry->d_lru);
394 dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
395 this_cpu_dec(nr_dentry_unused);
398 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
400 D_FLAG_VERIFY(dentry, 0);
401 list_add(&dentry->d_lru, list);
402 dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
403 this_cpu_inc(nr_dentry_unused);
407 * These can only be called under the global LRU lock, ie during the
408 * callback for freeing the LRU list. "isolate" removes it from the
409 * LRU lists entirely, while shrink_move moves it to the indicated
412 static void d_lru_isolate(struct dentry *dentry)
414 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
415 dentry->d_flags &= ~DCACHE_LRU_LIST;
416 this_cpu_dec(nr_dentry_unused);
417 list_del_init(&dentry->d_lru);
420 static void d_lru_shrink_move(struct dentry *dentry, struct list_head *list)
422 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
423 dentry->d_flags |= DCACHE_SHRINK_LIST;
424 list_move_tail(&dentry->d_lru, list);
428 * dentry_lru_(add|del)_list) must be called with d_lock held.
430 static void dentry_lru_add(struct dentry *dentry)
432 if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
437 * Remove a dentry with references from the LRU.
439 * If we are on the shrink list, then we can get to try_prune_one_dentry() and
440 * lose our last reference through the parent walk. In this case, we need to
441 * remove ourselves from the shrink list, not the LRU.
443 static void dentry_lru_del(struct dentry *dentry)
445 if (dentry->d_flags & DCACHE_LRU_LIST) {
446 if (dentry->d_flags & DCACHE_SHRINK_LIST)
447 return d_shrink_del(dentry);
453 * d_kill - kill dentry and return parent
454 * @dentry: dentry to kill
455 * @parent: parent dentry
457 * The dentry must already be unhashed and removed from the LRU.
459 * If this is the root of the dentry tree, return NULL.
461 * dentry->d_lock and parent->d_lock must be held by caller, and are dropped by
464 static struct dentry *d_kill(struct dentry *dentry, struct dentry *parent)
465 __releases(dentry->d_lock)
466 __releases(parent->d_lock)
467 __releases(dentry->d_inode->i_lock)
469 list_del(&dentry->d_u.d_child);
471 * Inform try_to_ascend() that we are no longer attached to the
474 dentry->d_flags |= DCACHE_DENTRY_KILLED;
476 spin_unlock(&parent->d_lock);
479 * dentry_iput drops the locks, at which point nobody (except
480 * transient RCU lookups) can reach this dentry.
487 * Unhash a dentry without inserting an RCU walk barrier or checking that
488 * dentry->d_lock is locked. The caller must take care of that, if
491 static void __d_shrink(struct dentry *dentry)
493 if (!d_unhashed(dentry)) {
494 struct hlist_bl_head *b;
495 if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
496 b = &dentry->d_sb->s_anon;
498 b = d_hash(dentry->d_parent, dentry->d_name.hash);
501 __hlist_bl_del(&dentry->d_hash);
502 dentry->d_hash.pprev = NULL;
508 * d_drop - drop a dentry
509 * @dentry: dentry to drop
511 * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
512 * be found through a VFS lookup any more. Note that this is different from
513 * deleting the dentry - d_delete will try to mark the dentry negative if
514 * possible, giving a successful _negative_ lookup, while d_drop will
515 * just make the cache lookup fail.
517 * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
518 * reason (NFS timeouts or autofs deletes).
520 * __d_drop requires dentry->d_lock.
522 void __d_drop(struct dentry *dentry)
524 if (!d_unhashed(dentry)) {
526 dentry_rcuwalk_barrier(dentry);
529 EXPORT_SYMBOL(__d_drop);
531 void d_drop(struct dentry *dentry)
533 spin_lock(&dentry->d_lock);
535 spin_unlock(&dentry->d_lock);
537 EXPORT_SYMBOL(d_drop);
540 * Finish off a dentry we've decided to kill.
541 * dentry->d_lock must be held, returns with it unlocked.
542 * If ref is non-zero, then decrement the refcount too.
543 * Returns dentry requiring refcount drop, or NULL if we're done.
545 static inline struct dentry *
546 dentry_kill(struct dentry *dentry, int unlock_on_failure)
547 __releases(dentry->d_lock)
550 struct dentry *parent;
552 inode = dentry->d_inode;
553 if (inode && !spin_trylock(&inode->i_lock)) {
555 if (unlock_on_failure) {
556 spin_unlock(&dentry->d_lock);
559 return dentry; /* try again with same dentry */
564 parent = dentry->d_parent;
565 if (parent && !spin_trylock(&parent->d_lock)) {
567 spin_unlock(&inode->i_lock);
572 * The dentry is now unrecoverably dead to the world.
574 lockref_mark_dead(&dentry->d_lockref);
577 * inform the fs via d_prune that this dentry is about to be
578 * unhashed and destroyed.
580 if ((dentry->d_flags & DCACHE_OP_PRUNE) && !d_unhashed(dentry))
581 dentry->d_op->d_prune(dentry);
583 dentry_lru_del(dentry);
584 /* if it was on the hash then remove it */
586 return d_kill(dentry, parent);
592 * This is complicated by the fact that we do not want to put
593 * dentries that are no longer on any hash chain on the unused
594 * list: we'd much rather just get rid of them immediately.
596 * However, that implies that we have to traverse the dentry
597 * tree upwards to the parents which might _also_ now be
598 * scheduled for deletion (it may have been only waiting for
599 * its last child to go away).
601 * This tail recursion is done by hand as we don't want to depend
602 * on the compiler to always get this right (gcc generally doesn't).
603 * Real recursion would eat up our stack space.
607 * dput - release a dentry
608 * @dentry: dentry to release
610 * Release a dentry. This will drop the usage count and if appropriate
611 * call the dentry unlink method as well as removing it from the queues and
612 * releasing its resources. If the parent dentries were scheduled for release
613 * they too may now get deleted.
615 void dput(struct dentry *dentry)
617 if (unlikely(!dentry))
621 if (lockref_put_or_lock(&dentry->d_lockref))
624 /* Unreachable? Get rid of it */
625 if (unlikely(d_unhashed(dentry)))
628 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
629 if (dentry->d_op->d_delete(dentry))
633 dentry->d_flags |= DCACHE_REFERENCED;
634 dentry_lru_add(dentry);
636 dentry->d_lockref.count--;
637 spin_unlock(&dentry->d_lock);
641 dentry = dentry_kill(dentry, 1);
648 * d_invalidate - invalidate a dentry
649 * @dentry: dentry to invalidate
651 * Try to invalidate the dentry if it turns out to be
652 * possible. If there are other dentries that can be
653 * reached through this one we can't delete it and we
654 * return -EBUSY. On success we return 0.
659 int d_invalidate(struct dentry * dentry)
662 * If it's already been dropped, return OK.
664 spin_lock(&dentry->d_lock);
665 if (d_unhashed(dentry)) {
666 spin_unlock(&dentry->d_lock);
670 * Check whether to do a partial shrink_dcache
671 * to get rid of unused child entries.
673 if (!list_empty(&dentry->d_subdirs)) {
674 spin_unlock(&dentry->d_lock);
675 shrink_dcache_parent(dentry);
676 spin_lock(&dentry->d_lock);
680 * Somebody else still using it?
682 * If it's a directory, we can't drop it
683 * for fear of somebody re-populating it
684 * with children (even though dropping it
685 * would make it unreachable from the root,
686 * we might still populate it if it was a
687 * working directory or similar).
688 * We also need to leave mountpoints alone,
691 if (dentry->d_lockref.count > 1 && dentry->d_inode) {
692 if (S_ISDIR(dentry->d_inode->i_mode) || d_mountpoint(dentry)) {
693 spin_unlock(&dentry->d_lock);
699 spin_unlock(&dentry->d_lock);
702 EXPORT_SYMBOL(d_invalidate);
704 /* This must be called with d_lock held */
705 static inline void __dget_dlock(struct dentry *dentry)
707 dentry->d_lockref.count++;
710 static inline void __dget(struct dentry *dentry)
712 lockref_get(&dentry->d_lockref);
715 struct dentry *dget_parent(struct dentry *dentry)
721 * Do optimistic parent lookup without any
725 ret = ACCESS_ONCE(dentry->d_parent);
726 gotref = lockref_get_not_zero(&ret->d_lockref);
728 if (likely(gotref)) {
729 if (likely(ret == ACCESS_ONCE(dentry->d_parent)))
736 * Don't need rcu_dereference because we re-check it was correct under
740 ret = dentry->d_parent;
741 spin_lock(&ret->d_lock);
742 if (unlikely(ret != dentry->d_parent)) {
743 spin_unlock(&ret->d_lock);
748 BUG_ON(!ret->d_lockref.count);
749 ret->d_lockref.count++;
750 spin_unlock(&ret->d_lock);
753 EXPORT_SYMBOL(dget_parent);
756 * d_find_alias - grab a hashed alias of inode
757 * @inode: inode in question
758 * @want_discon: flag, used by d_splice_alias, to request
759 * that only a DISCONNECTED alias be returned.
761 * If inode has a hashed alias, or is a directory and has any alias,
762 * acquire the reference to alias and return it. Otherwise return NULL.
763 * Notice that if inode is a directory there can be only one alias and
764 * it can be unhashed only if it has no children, or if it is the root
767 * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
768 * any other hashed alias over that one unless @want_discon is set,
769 * in which case only return an IS_ROOT, DCACHE_DISCONNECTED alias.
771 static struct dentry *__d_find_alias(struct inode *inode, int want_discon)
773 struct dentry *alias, *discon_alias;
777 hlist_for_each_entry(alias, &inode->i_dentry, d_alias) {
778 spin_lock(&alias->d_lock);
779 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
780 if (IS_ROOT(alias) &&
781 (alias->d_flags & DCACHE_DISCONNECTED)) {
782 discon_alias = alias;
783 } else if (!want_discon) {
785 spin_unlock(&alias->d_lock);
789 spin_unlock(&alias->d_lock);
792 alias = discon_alias;
793 spin_lock(&alias->d_lock);
794 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
795 if (IS_ROOT(alias) &&
796 (alias->d_flags & DCACHE_DISCONNECTED)) {
798 spin_unlock(&alias->d_lock);
802 spin_unlock(&alias->d_lock);
808 struct dentry *d_find_alias(struct inode *inode)
810 struct dentry *de = NULL;
812 if (!hlist_empty(&inode->i_dentry)) {
813 spin_lock(&inode->i_lock);
814 de = __d_find_alias(inode, 0);
815 spin_unlock(&inode->i_lock);
819 EXPORT_SYMBOL(d_find_alias);
822 * Try to kill dentries associated with this inode.
823 * WARNING: you must own a reference to inode.
825 void d_prune_aliases(struct inode *inode)
827 struct dentry *dentry;
829 spin_lock(&inode->i_lock);
830 hlist_for_each_entry(dentry, &inode->i_dentry, d_alias) {
831 spin_lock(&dentry->d_lock);
832 if (!dentry->d_lockref.count) {
834 * inform the fs via d_prune that this dentry
835 * is about to be unhashed and destroyed.
837 if ((dentry->d_flags & DCACHE_OP_PRUNE) &&
839 dentry->d_op->d_prune(dentry);
841 __dget_dlock(dentry);
843 spin_unlock(&dentry->d_lock);
844 spin_unlock(&inode->i_lock);
848 spin_unlock(&dentry->d_lock);
850 spin_unlock(&inode->i_lock);
852 EXPORT_SYMBOL(d_prune_aliases);
855 * Try to throw away a dentry - free the inode, dput the parent.
856 * Requires dentry->d_lock is held, and dentry->d_count == 0.
857 * Releases dentry->d_lock.
859 * This may fail if locks cannot be acquired no problem, just try again.
861 static struct dentry * try_prune_one_dentry(struct dentry *dentry)
862 __releases(dentry->d_lock)
864 struct dentry *parent;
866 parent = dentry_kill(dentry, 0);
868 * If dentry_kill returns NULL, we have nothing more to do.
869 * if it returns the same dentry, trylocks failed. In either
870 * case, just loop again.
872 * Otherwise, we need to prune ancestors too. This is necessary
873 * to prevent quadratic behavior of shrink_dcache_parent(), but
874 * is also expected to be beneficial in reducing dentry cache
879 if (parent == dentry)
882 /* Prune ancestors. */
885 if (lockref_put_or_lock(&dentry->d_lockref))
887 dentry = dentry_kill(dentry, 1);
892 static void shrink_dentry_list(struct list_head *list)
894 struct dentry *dentry;
898 dentry = list_entry_rcu(list->prev, struct dentry, d_lru);
899 if (&dentry->d_lru == list)
903 * Get the dentry lock, and re-verify that the dentry is
904 * this on the shrinking list. If it is, we know that
905 * DCACHE_SHRINK_LIST and DCACHE_LRU_LIST are set.
907 spin_lock(&dentry->d_lock);
908 if (dentry != list_entry(list->prev, struct dentry, d_lru)) {
909 spin_unlock(&dentry->d_lock);
914 * The dispose list is isolated and dentries are not accounted
915 * to the LRU here, so we can simply remove it from the list
916 * here regardless of whether it is referenced or not.
918 d_shrink_del(dentry);
921 * We found an inuse dentry which was not removed from
922 * the LRU because of laziness during lookup. Do not free it.
924 if (dentry->d_lockref.count) {
925 spin_unlock(&dentry->d_lock);
931 * If 'try_to_prune()' returns a dentry, it will
932 * be the same one we passed in, and d_lock will
933 * have been held the whole time, so it will not
934 * have been added to any other lists. We failed
935 * to get the inode lock.
937 * We just add it back to the shrink list.
939 dentry = try_prune_one_dentry(dentry);
943 d_shrink_add(dentry, list);
944 spin_unlock(&dentry->d_lock);
950 static enum lru_status
951 dentry_lru_isolate(struct list_head *item, spinlock_t *lru_lock, void *arg)
953 struct list_head *freeable = arg;
954 struct dentry *dentry = container_of(item, struct dentry, d_lru);
958 * we are inverting the lru lock/dentry->d_lock here,
959 * so use a trylock. If we fail to get the lock, just skip
962 if (!spin_trylock(&dentry->d_lock))
966 * Referenced dentries are still in use. If they have active
967 * counts, just remove them from the LRU. Otherwise give them
968 * another pass through the LRU.
970 if (dentry->d_lockref.count) {
971 d_lru_isolate(dentry);
972 spin_unlock(&dentry->d_lock);
976 if (dentry->d_flags & DCACHE_REFERENCED) {
977 dentry->d_flags &= ~DCACHE_REFERENCED;
978 spin_unlock(&dentry->d_lock);
981 * The list move itself will be made by the common LRU code. At
982 * this point, we've dropped the dentry->d_lock but keep the
983 * lru lock. This is safe to do, since every list movement is
984 * protected by the lru lock even if both locks are held.
986 * This is guaranteed by the fact that all LRU management
987 * functions are intermediated by the LRU API calls like
988 * list_lru_add and list_lru_del. List movement in this file
989 * only ever occur through this functions or through callbacks
990 * like this one, that are called from the LRU API.
992 * The only exceptions to this are functions like
993 * shrink_dentry_list, and code that first checks for the
994 * DCACHE_SHRINK_LIST flag. Those are guaranteed to be
995 * operating only with stack provided lists after they are
996 * properly isolated from the main list. It is thus, always a
1002 d_lru_shrink_move(dentry, freeable);
1003 spin_unlock(&dentry->d_lock);
1009 * prune_dcache_sb - shrink the dcache
1011 * @nr_to_scan : number of entries to try to free
1012 * @nid: which node to scan for freeable entities
1014 * Attempt to shrink the superblock dcache LRU by @nr_to_scan entries. This is
1015 * done when we need more memory an called from the superblock shrinker
1018 * This function may fail to free any resources if all the dentries are in
1021 long prune_dcache_sb(struct super_block *sb, unsigned long nr_to_scan,
1027 freed = list_lru_walk_node(&sb->s_dentry_lru, nid, dentry_lru_isolate,
1028 &dispose, &nr_to_scan);
1029 shrink_dentry_list(&dispose);
1033 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1034 spinlock_t *lru_lock, void *arg)
1036 struct list_head *freeable = arg;
1037 struct dentry *dentry = container_of(item, struct dentry, d_lru);
1040 * we are inverting the lru lock/dentry->d_lock here,
1041 * so use a trylock. If we fail to get the lock, just skip
1044 if (!spin_trylock(&dentry->d_lock))
1047 d_lru_shrink_move(dentry, freeable);
1048 spin_unlock(&dentry->d_lock);
1055 * shrink_dcache_sb - shrink dcache for a superblock
1058 * Shrink the dcache for the specified super block. This is used to free
1059 * the dcache before unmounting a file system.
1061 void shrink_dcache_sb(struct super_block *sb)
1068 freed = list_lru_walk(&sb->s_dentry_lru,
1069 dentry_lru_isolate_shrink, &dispose, UINT_MAX);
1071 this_cpu_sub(nr_dentry_unused, freed);
1072 shrink_dentry_list(&dispose);
1073 } while (freed > 0);
1075 EXPORT_SYMBOL(shrink_dcache_sb);
1078 * destroy a single subtree of dentries for unmount
1079 * - see the comments on shrink_dcache_for_umount() for a description of the
1082 static void shrink_dcache_for_umount_subtree(struct dentry *dentry)
1084 struct dentry *parent;
1086 BUG_ON(!IS_ROOT(dentry));
1089 /* descend to the first leaf in the current subtree */
1090 while (!list_empty(&dentry->d_subdirs))
1091 dentry = list_entry(dentry->d_subdirs.next,
1092 struct dentry, d_u.d_child);
1094 /* consume the dentries from this leaf up through its parents
1095 * until we find one with children or run out altogether */
1097 struct inode *inode;
1100 * inform the fs that this dentry is about to be
1101 * unhashed and destroyed.
1103 if ((dentry->d_flags & DCACHE_OP_PRUNE) &&
1104 !d_unhashed(dentry))
1105 dentry->d_op->d_prune(dentry);
1107 dentry_lru_del(dentry);
1110 if (dentry->d_lockref.count != 0) {
1112 "BUG: Dentry %p{i=%lx,n=%s}"
1113 " still in use (%d)"
1114 " [unmount of %s %s]\n",
1117 dentry->d_inode->i_ino : 0UL,
1118 dentry->d_name.name,
1119 dentry->d_lockref.count,
1120 dentry->d_sb->s_type->name,
1121 dentry->d_sb->s_id);
1125 if (IS_ROOT(dentry)) {
1127 list_del(&dentry->d_u.d_child);
1129 parent = dentry->d_parent;
1130 parent->d_lockref.count--;
1131 list_del(&dentry->d_u.d_child);
1134 inode = dentry->d_inode;
1136 dentry->d_inode = NULL;
1137 hlist_del_init(&dentry->d_alias);
1138 if (dentry->d_op && dentry->d_op->d_iput)
1139 dentry->d_op->d_iput(dentry, inode);
1146 /* finished when we fall off the top of the tree,
1147 * otherwise we ascend to the parent and move to the
1148 * next sibling if there is one */
1152 } while (list_empty(&dentry->d_subdirs));
1154 dentry = list_entry(dentry->d_subdirs.next,
1155 struct dentry, d_u.d_child);
1160 * destroy the dentries attached to a superblock on unmounting
1161 * - we don't need to use dentry->d_lock because:
1162 * - the superblock is detached from all mountings and open files, so the
1163 * dentry trees will not be rearranged by the VFS
1164 * - s_umount is write-locked, so the memory pressure shrinker will ignore
1165 * any dentries belonging to this superblock that it comes across
1166 * - the filesystem itself is no longer permitted to rearrange the dentries
1167 * in this superblock
1169 void shrink_dcache_for_umount(struct super_block *sb)
1171 struct dentry *dentry;
1173 if (down_read_trylock(&sb->s_umount))
1176 dentry = sb->s_root;
1178 dentry->d_lockref.count--;
1179 shrink_dcache_for_umount_subtree(dentry);
1181 while (!hlist_bl_empty(&sb->s_anon)) {
1182 dentry = hlist_bl_entry(hlist_bl_first(&sb->s_anon), struct dentry, d_hash);
1183 shrink_dcache_for_umount_subtree(dentry);
1188 * This tries to ascend one level of parenthood, but
1189 * we can race with renaming, so we need to re-check
1190 * the parenthood after dropping the lock and check
1191 * that the sequence number still matches.
1193 static struct dentry *try_to_ascend(struct dentry *old, unsigned seq)
1195 struct dentry *new = old->d_parent;
1198 spin_unlock(&old->d_lock);
1199 spin_lock(&new->d_lock);
1202 * might go back up the wrong parent if we have had a rename
1205 if (new != old->d_parent ||
1206 (old->d_flags & DCACHE_DENTRY_KILLED) ||
1207 need_seqretry(&rename_lock, seq)) {
1208 spin_unlock(&new->d_lock);
1216 * enum d_walk_ret - action to talke during tree walk
1217 * @D_WALK_CONTINUE: contrinue walk
1218 * @D_WALK_QUIT: quit walk
1219 * @D_WALK_NORETRY: quit when retry is needed
1220 * @D_WALK_SKIP: skip this dentry and its children
1230 * d_walk - walk the dentry tree
1231 * @parent: start of walk
1232 * @data: data passed to @enter() and @finish()
1233 * @enter: callback when first entering the dentry
1234 * @finish: callback when successfully finished the walk
1236 * The @enter() and @finish() callbacks are called with d_lock held.
1238 static void d_walk(struct dentry *parent, void *data,
1239 enum d_walk_ret (*enter)(void *, struct dentry *),
1240 void (*finish)(void *))
1242 struct dentry *this_parent;
1243 struct list_head *next;
1245 enum d_walk_ret ret;
1249 read_seqbegin_or_lock(&rename_lock, &seq);
1250 this_parent = parent;
1251 spin_lock(&this_parent->d_lock);
1253 ret = enter(data, this_parent);
1255 case D_WALK_CONTINUE:
1260 case D_WALK_NORETRY:
1265 next = this_parent->d_subdirs.next;
1267 while (next != &this_parent->d_subdirs) {
1268 struct list_head *tmp = next;
1269 struct dentry *dentry = list_entry(tmp, struct dentry, d_u.d_child);
1272 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1274 ret = enter(data, dentry);
1276 case D_WALK_CONTINUE:
1279 spin_unlock(&dentry->d_lock);
1281 case D_WALK_NORETRY:
1285 spin_unlock(&dentry->d_lock);
1289 if (!list_empty(&dentry->d_subdirs)) {
1290 spin_unlock(&this_parent->d_lock);
1291 spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1292 this_parent = dentry;
1293 spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1296 spin_unlock(&dentry->d_lock);
1299 * All done at this level ... ascend and resume the search.
1301 if (this_parent != parent) {
1302 struct dentry *child = this_parent;
1303 this_parent = try_to_ascend(this_parent, seq);
1306 next = child->d_u.d_child.next;
1309 if (need_seqretry(&rename_lock, seq)) {
1310 spin_unlock(&this_parent->d_lock);
1317 spin_unlock(&this_parent->d_lock);
1318 done_seqretry(&rename_lock, seq);
1329 * Search for at least 1 mount point in the dentry's subdirs.
1330 * We descend to the next level whenever the d_subdirs
1331 * list is non-empty and continue searching.
1335 * have_submounts - check for mounts over a dentry
1336 * @parent: dentry to check.
1338 * Return true if the parent or its subdirectories contain
1342 static enum d_walk_ret check_mount(void *data, struct dentry *dentry)
1345 if (d_mountpoint(dentry)) {
1349 return D_WALK_CONTINUE;
1352 int have_submounts(struct dentry *parent)
1356 d_walk(parent, &ret, check_mount, NULL);
1360 EXPORT_SYMBOL(have_submounts);
1363 * Called by mount code to set a mountpoint and check if the mountpoint is
1364 * reachable (e.g. NFS can unhash a directory dentry and then the complete
1365 * subtree can become unreachable).
1367 * Only one of check_submounts_and_drop() and d_set_mounted() must succeed. For
1368 * this reason take rename_lock and d_lock on dentry and ancestors.
1370 int d_set_mounted(struct dentry *dentry)
1374 write_seqlock(&rename_lock);
1375 for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1376 /* Need exclusion wrt. shrink_submounts_and_drop() */
1377 spin_lock(&p->d_lock);
1378 if (unlikely(d_unhashed(p))) {
1379 spin_unlock(&p->d_lock);
1382 spin_unlock(&p->d_lock);
1384 spin_lock(&dentry->d_lock);
1385 if (!d_unlinked(dentry)) {
1386 dentry->d_flags |= DCACHE_MOUNTED;
1389 spin_unlock(&dentry->d_lock);
1391 write_sequnlock(&rename_lock);
1396 * Search the dentry child list of the specified parent,
1397 * and move any unused dentries to the end of the unused
1398 * list for prune_dcache(). We descend to the next level
1399 * whenever the d_subdirs list is non-empty and continue
1402 * It returns zero iff there are no unused children,
1403 * otherwise it returns the number of children moved to
1404 * the end of the unused list. This may not be the total
1405 * number of unused children, because select_parent can
1406 * drop the lock and return early due to latency
1410 struct select_data {
1411 struct dentry *start;
1412 struct list_head dispose;
1416 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1418 struct select_data *data = _data;
1419 enum d_walk_ret ret = D_WALK_CONTINUE;
1421 if (data->start == dentry)
1425 * move only zero ref count dentries to the dispose list.
1427 * Those which are presently on the shrink list, being processed
1428 * by shrink_dentry_list(), shouldn't be moved. Otherwise the
1429 * loop in shrink_dcache_parent() might not make any progress
1432 if (dentry->d_lockref.count) {
1433 dentry_lru_del(dentry);
1434 } else if (!(dentry->d_flags & DCACHE_SHRINK_LIST)) {
1436 * We can't use d_lru_shrink_move() because we
1437 * need to get the global LRU lock and do the
1441 d_shrink_add(dentry, &data->dispose);
1443 ret = D_WALK_NORETRY;
1446 * We can return to the caller if we have found some (this
1447 * ensures forward progress). We'll be coming back to find
1450 if (data->found && need_resched())
1457 * shrink_dcache_parent - prune dcache
1458 * @parent: parent of entries to prune
1460 * Prune the dcache to remove unused children of the parent dentry.
1462 void shrink_dcache_parent(struct dentry *parent)
1465 struct select_data data;
1467 INIT_LIST_HEAD(&data.dispose);
1468 data.start = parent;
1471 d_walk(parent, &data, select_collect, NULL);
1475 shrink_dentry_list(&data.dispose);
1479 EXPORT_SYMBOL(shrink_dcache_parent);
1481 struct detach_data {
1482 struct dentry *found;
1484 static enum d_walk_ret do_detach_submounts(void *ptr, struct dentry *dentry)
1486 struct detach_data *data = ptr;
1488 if (d_mountpoint(dentry))
1489 data->found = dentry;
1491 return data->found ? D_WALK_QUIT : D_WALK_CONTINUE;
1495 * detach_submounts - check for submounts and detach them.
1497 * @dentry: dentry to find mount points under.
1499 * If dentry or any of it's children is a mount point detach those mounts.
1501 void detach_submounts(struct dentry *dentry)
1503 struct detach_data data;
1506 d_walk(dentry, &data, do_detach_submounts, NULL);
1511 detach_mounts(data.found);
1514 detach_mounts(dentry);
1518 * shrink_submounts_and_drop - detach submounts, prune dcache, and drop
1520 * All done as a single atomic operation reletaive to d_set_mounted().
1522 * @dentry: dentry to detach, prune and drop
1524 void shrink_submounts_and_drop(struct dentry *dentry)
1527 detach_submounts(dentry);
1528 shrink_dcache_parent(dentry);
1530 EXPORT_SYMBOL(shrink_submounts_and_drop);
1533 * __d_alloc - allocate a dcache entry
1534 * @sb: filesystem it will belong to
1535 * @name: qstr of the name
1537 * Allocates a dentry. It returns %NULL if there is insufficient memory
1538 * available. On a success the dentry is returned. The name passed in is
1539 * copied and the copy passed in may be reused after this call.
1542 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1544 struct dentry *dentry;
1547 dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1552 * We guarantee that the inline name is always NUL-terminated.
1553 * This way the memcpy() done by the name switching in rename
1554 * will still always have a NUL at the end, even if we might
1555 * be overwriting an internal NUL character
1557 dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1558 if (name->len > DNAME_INLINE_LEN-1) {
1559 dname = kmalloc(name->len + 1, GFP_KERNEL);
1561 kmem_cache_free(dentry_cache, dentry);
1565 dname = dentry->d_iname;
1568 dentry->d_name.len = name->len;
1569 dentry->d_name.hash = name->hash;
1570 memcpy(dname, name->name, name->len);
1571 dname[name->len] = 0;
1573 /* Make sure we always see the terminating NUL character */
1575 dentry->d_name.name = dname;
1577 dentry->d_lockref.count = 1;
1578 dentry->d_flags = 0;
1579 spin_lock_init(&dentry->d_lock);
1580 seqcount_init(&dentry->d_seq);
1581 dentry->d_inode = NULL;
1582 dentry->d_parent = dentry;
1584 dentry->d_op = NULL;
1585 dentry->d_fsdata = NULL;
1586 INIT_HLIST_BL_NODE(&dentry->d_hash);
1587 INIT_LIST_HEAD(&dentry->d_lru);
1588 INIT_LIST_HEAD(&dentry->d_subdirs);
1589 INIT_HLIST_NODE(&dentry->d_alias);
1590 INIT_LIST_HEAD(&dentry->d_u.d_child);
1591 d_set_d_op(dentry, dentry->d_sb->s_d_op);
1593 this_cpu_inc(nr_dentry);
1599 * d_alloc - allocate a dcache entry
1600 * @parent: parent of entry to allocate
1601 * @name: qstr of the name
1603 * Allocates a dentry. It returns %NULL if there is insufficient memory
1604 * available. On a success the dentry is returned. The name passed in is
1605 * copied and the copy passed in may be reused after this call.
1607 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1609 struct dentry *dentry = __d_alloc(parent->d_sb, name);
1613 spin_lock(&parent->d_lock);
1615 * don't need child lock because it is not subject
1616 * to concurrency here
1618 __dget_dlock(parent);
1619 dentry->d_parent = parent;
1620 list_add(&dentry->d_u.d_child, &parent->d_subdirs);
1621 spin_unlock(&parent->d_lock);
1625 EXPORT_SYMBOL(d_alloc);
1627 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1629 struct dentry *dentry = __d_alloc(sb, name);
1631 dentry->d_flags |= DCACHE_DISCONNECTED;
1634 EXPORT_SYMBOL(d_alloc_pseudo);
1636 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1641 q.len = strlen(name);
1642 q.hash = full_name_hash(q.name, q.len);
1643 return d_alloc(parent, &q);
1645 EXPORT_SYMBOL(d_alloc_name);
1647 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1649 WARN_ON_ONCE(dentry->d_op);
1650 WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH |
1652 DCACHE_OP_REVALIDATE |
1653 DCACHE_OP_WEAK_REVALIDATE |
1654 DCACHE_OP_DELETE ));
1659 dentry->d_flags |= DCACHE_OP_HASH;
1661 dentry->d_flags |= DCACHE_OP_COMPARE;
1662 if (op->d_revalidate)
1663 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1664 if (op->d_weak_revalidate)
1665 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1667 dentry->d_flags |= DCACHE_OP_DELETE;
1669 dentry->d_flags |= DCACHE_OP_PRUNE;
1672 EXPORT_SYMBOL(d_set_d_op);
1674 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1676 spin_lock(&dentry->d_lock);
1678 if (unlikely(IS_AUTOMOUNT(inode)))
1679 dentry->d_flags |= DCACHE_NEED_AUTOMOUNT;
1680 hlist_add_head(&dentry->d_alias, &inode->i_dentry);
1682 dentry->d_inode = inode;
1683 dentry_rcuwalk_barrier(dentry);
1684 spin_unlock(&dentry->d_lock);
1685 fsnotify_d_instantiate(dentry, inode);
1689 * d_instantiate - fill in inode information for a dentry
1690 * @entry: dentry to complete
1691 * @inode: inode to attach to this dentry
1693 * Fill in inode information in the entry.
1695 * This turns negative dentries into productive full members
1698 * NOTE! This assumes that the inode count has been incremented
1699 * (or otherwise set) by the caller to indicate that it is now
1700 * in use by the dcache.
1703 void d_instantiate(struct dentry *entry, struct inode * inode)
1705 BUG_ON(!hlist_unhashed(&entry->d_alias));
1707 spin_lock(&inode->i_lock);
1708 __d_instantiate(entry, inode);
1710 spin_unlock(&inode->i_lock);
1711 security_d_instantiate(entry, inode);
1713 EXPORT_SYMBOL(d_instantiate);
1716 * d_instantiate_unique - instantiate a non-aliased dentry
1717 * @entry: dentry to instantiate
1718 * @inode: inode to attach to this dentry
1720 * Fill in inode information in the entry. On success, it returns NULL.
1721 * If an unhashed alias of "entry" already exists, then we return the
1722 * aliased dentry instead and drop one reference to inode.
1724 * Note that in order to avoid conflicts with rename() etc, the caller
1725 * had better be holding the parent directory semaphore.
1727 * This also assumes that the inode count has been incremented
1728 * (or otherwise set) by the caller to indicate that it is now
1729 * in use by the dcache.
1731 static struct dentry *__d_instantiate_unique(struct dentry *entry,
1732 struct inode *inode)
1734 struct dentry *alias;
1735 int len = entry->d_name.len;
1736 const char *name = entry->d_name.name;
1737 unsigned int hash = entry->d_name.hash;
1740 __d_instantiate(entry, NULL);
1744 hlist_for_each_entry(alias, &inode->i_dentry, d_alias) {
1746 * Don't need alias->d_lock here, because aliases with
1747 * d_parent == entry->d_parent are not subject to name or
1748 * parent changes, because the parent inode i_mutex is held.
1750 if (alias->d_name.hash != hash)
1752 if (alias->d_parent != entry->d_parent)
1754 if (alias->d_name.len != len)
1756 if (dentry_cmp(alias, name, len))
1762 __d_instantiate(entry, inode);
1766 struct dentry *d_instantiate_unique(struct dentry *entry, struct inode *inode)
1768 struct dentry *result;
1770 BUG_ON(!hlist_unhashed(&entry->d_alias));
1773 spin_lock(&inode->i_lock);
1774 result = __d_instantiate_unique(entry, inode);
1776 spin_unlock(&inode->i_lock);
1779 security_d_instantiate(entry, inode);
1783 BUG_ON(!d_unhashed(result));
1788 EXPORT_SYMBOL(d_instantiate_unique);
1790 struct dentry *d_make_root(struct inode *root_inode)
1792 struct dentry *res = NULL;
1795 static const struct qstr name = QSTR_INIT("/", 1);
1797 res = __d_alloc(root_inode->i_sb, &name);
1799 d_instantiate(res, root_inode);
1805 EXPORT_SYMBOL(d_make_root);
1807 static struct dentry * __d_find_any_alias(struct inode *inode)
1809 struct dentry *alias;
1811 if (hlist_empty(&inode->i_dentry))
1813 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_alias);
1819 * d_find_any_alias - find any alias for a given inode
1820 * @inode: inode to find an alias for
1822 * If any aliases exist for the given inode, take and return a
1823 * reference for one of them. If no aliases exist, return %NULL.
1825 struct dentry *d_find_any_alias(struct inode *inode)
1829 spin_lock(&inode->i_lock);
1830 de = __d_find_any_alias(inode);
1831 spin_unlock(&inode->i_lock);
1834 EXPORT_SYMBOL(d_find_any_alias);
1837 * d_obtain_alias - find or allocate a dentry for a given inode
1838 * @inode: inode to allocate the dentry for
1840 * Obtain a dentry for an inode resulting from NFS filehandle conversion or
1841 * similar open by handle operations. The returned dentry may be anonymous,
1842 * or may have a full name (if the inode was already in the cache).
1844 * When called on a directory inode, we must ensure that the inode only ever
1845 * has one dentry. If a dentry is found, that is returned instead of
1846 * allocating a new one.
1848 * On successful return, the reference to the inode has been transferred
1849 * to the dentry. In case of an error the reference on the inode is released.
1850 * To make it easier to use in export operations a %NULL or IS_ERR inode may
1851 * be passed in and will be the error will be propagate to the return value,
1852 * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
1854 struct dentry *d_obtain_alias(struct inode *inode)
1856 static const struct qstr anonstring = QSTR_INIT("/", 1);
1861 return ERR_PTR(-ESTALE);
1863 return ERR_CAST(inode);
1865 res = d_find_any_alias(inode);
1869 tmp = __d_alloc(inode->i_sb, &anonstring);
1871 res = ERR_PTR(-ENOMEM);
1875 spin_lock(&inode->i_lock);
1876 res = __d_find_any_alias(inode);
1878 spin_unlock(&inode->i_lock);
1883 /* attach a disconnected dentry */
1884 spin_lock(&tmp->d_lock);
1885 tmp->d_inode = inode;
1886 tmp->d_flags |= DCACHE_DISCONNECTED;
1887 hlist_add_head(&tmp->d_alias, &inode->i_dentry);
1888 hlist_bl_lock(&tmp->d_sb->s_anon);
1889 hlist_bl_add_head(&tmp->d_hash, &tmp->d_sb->s_anon);
1890 hlist_bl_unlock(&tmp->d_sb->s_anon);
1891 spin_unlock(&tmp->d_lock);
1892 spin_unlock(&inode->i_lock);
1893 security_d_instantiate(tmp, inode);
1898 if (res && !IS_ERR(res))
1899 security_d_instantiate(res, inode);
1903 EXPORT_SYMBOL(d_obtain_alias);
1906 * d_splice_alias - splice a disconnected dentry into the tree if one exists
1907 * @inode: the inode which may have a disconnected dentry
1908 * @dentry: a negative dentry which we want to point to the inode.
1910 * If inode is a directory and has a 'disconnected' dentry (i.e. IS_ROOT and
1911 * DCACHE_DISCONNECTED), then d_move that in place of the given dentry
1912 * and return it, else simply d_add the inode to the dentry and return NULL.
1914 * This is needed in the lookup routine of any filesystem that is exportable
1915 * (via knfsd) so that we can build dcache paths to directories effectively.
1917 * If a dentry was found and moved, then it is returned. Otherwise NULL
1918 * is returned. This matches the expected return value of ->lookup.
1920 * Cluster filesystems may call this function with a negative, hashed dentry.
1921 * In that case, we know that the inode will be a regular file, and also this
1922 * will only occur during atomic_open. So we need to check for the dentry
1923 * being already hashed only in the final case.
1925 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
1927 struct dentry *new = NULL;
1930 return ERR_CAST(inode);
1932 if (inode && S_ISDIR(inode->i_mode)) {
1933 spin_lock(&inode->i_lock);
1934 new = __d_find_alias(inode, 1);
1936 BUG_ON(!(new->d_flags & DCACHE_DISCONNECTED));
1937 spin_unlock(&inode->i_lock);
1938 security_d_instantiate(new, inode);
1939 d_move(new, dentry);
1942 /* already taking inode->i_lock, so d_add() by hand */
1943 __d_instantiate(dentry, inode);
1944 spin_unlock(&inode->i_lock);
1945 security_d_instantiate(dentry, inode);
1949 d_instantiate(dentry, inode);
1950 if (d_unhashed(dentry))
1955 EXPORT_SYMBOL(d_splice_alias);
1958 * d_add_ci - lookup or allocate new dentry with case-exact name
1959 * @inode: the inode case-insensitive lookup has found
1960 * @dentry: the negative dentry that was passed to the parent's lookup func
1961 * @name: the case-exact name to be associated with the returned dentry
1963 * This is to avoid filling the dcache with case-insensitive names to the
1964 * same inode, only the actual correct case is stored in the dcache for
1965 * case-insensitive filesystems.
1967 * For a case-insensitive lookup match and if the the case-exact dentry
1968 * already exists in in the dcache, use it and return it.
1970 * If no entry exists with the exact case name, allocate new dentry with
1971 * the exact case, and return the spliced entry.
1973 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
1976 struct dentry *found;
1980 * First check if a dentry matching the name already exists,
1981 * if not go ahead and create it now.
1983 found = d_hash_and_lookup(dentry->d_parent, name);
1984 if (unlikely(IS_ERR(found)))
1987 new = d_alloc(dentry->d_parent, name);
1989 found = ERR_PTR(-ENOMEM);
1993 found = d_splice_alias(inode, new);
2002 * If a matching dentry exists, and it's not negative use it.
2004 * Decrement the reference count to balance the iget() done
2007 if (found->d_inode) {
2008 if (unlikely(found->d_inode != inode)) {
2009 /* This can't happen because bad inodes are unhashed. */
2010 BUG_ON(!is_bad_inode(inode));
2011 BUG_ON(!is_bad_inode(found->d_inode));
2018 * Negative dentry: instantiate it unless the inode is a directory and
2019 * already has a dentry.
2021 new = d_splice_alias(inode, found);
2032 EXPORT_SYMBOL(d_add_ci);
2035 * Do the slow-case of the dentry name compare.
2037 * Unlike the dentry_cmp() function, we need to atomically
2038 * load the name and length information, so that the
2039 * filesystem can rely on them, and can use the 'name' and
2040 * 'len' information without worrying about walking off the
2041 * end of memory etc.
2043 * Thus the read_seqcount_retry() and the "duplicate" info
2044 * in arguments (the low-level filesystem should not look
2045 * at the dentry inode or name contents directly, since
2046 * rename can change them while we're in RCU mode).
2048 enum slow_d_compare {
2054 static noinline enum slow_d_compare slow_dentry_cmp(
2055 const struct dentry *parent,
2056 struct dentry *dentry,
2058 const struct qstr *name)
2060 int tlen = dentry->d_name.len;
2061 const char *tname = dentry->d_name.name;
2063 if (read_seqcount_retry(&dentry->d_seq, seq)) {
2065 return D_COMP_SEQRETRY;
2067 if (parent->d_op->d_compare(parent, dentry, tlen, tname, name))
2068 return D_COMP_NOMATCH;
2073 * __d_lookup_rcu - search for a dentry (racy, store-free)
2074 * @parent: parent dentry
2075 * @name: qstr of name we wish to find
2076 * @seqp: returns d_seq value at the point where the dentry was found
2077 * Returns: dentry, or NULL
2079 * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2080 * resolution (store-free path walking) design described in
2081 * Documentation/filesystems/path-lookup.txt.
2083 * This is not to be used outside core vfs.
2085 * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2086 * held, and rcu_read_lock held. The returned dentry must not be stored into
2087 * without taking d_lock and checking d_seq sequence count against @seq
2090 * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2093 * Alternatively, __d_lookup_rcu may be called again to look up the child of
2094 * the returned dentry, so long as its parent's seqlock is checked after the
2095 * child is looked up. Thus, an interlocking stepping of sequence lock checks
2096 * is formed, giving integrity down the path walk.
2098 * NOTE! The caller *has* to check the resulting dentry against the sequence
2099 * number we've returned before using any of the resulting dentry state!
2101 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2102 const struct qstr *name,
2105 u64 hashlen = name->hash_len;
2106 const unsigned char *str = name->name;
2107 struct hlist_bl_head *b = d_hash(parent, hashlen_hash(hashlen));
2108 struct hlist_bl_node *node;
2109 struct dentry *dentry;
2112 * Note: There is significant duplication with __d_lookup_rcu which is
2113 * required to prevent single threaded performance regressions
2114 * especially on architectures where smp_rmb (in seqcounts) are costly.
2115 * Keep the two functions in sync.
2119 * The hash list is protected using RCU.
2121 * Carefully use d_seq when comparing a candidate dentry, to avoid
2122 * races with d_move().
2124 * It is possible that concurrent renames can mess up our list
2125 * walk here and result in missing our dentry, resulting in the
2126 * false-negative result. d_lookup() protects against concurrent
2127 * renames using rename_lock seqlock.
2129 * See Documentation/filesystems/path-lookup.txt for more details.
2131 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2136 * The dentry sequence count protects us from concurrent
2137 * renames, and thus protects parent and name fields.
2139 * The caller must perform a seqcount check in order
2140 * to do anything useful with the returned dentry.
2142 * NOTE! We do a "raw" seqcount_begin here. That means that
2143 * we don't wait for the sequence count to stabilize if it
2144 * is in the middle of a sequence change. If we do the slow
2145 * dentry compare, we will do seqretries until it is stable,
2146 * and if we end up with a successful lookup, we actually
2147 * want to exit RCU lookup anyway.
2149 seq = raw_seqcount_begin(&dentry->d_seq);
2150 if (dentry->d_parent != parent)
2152 if (d_unhashed(dentry))
2155 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) {
2156 if (dentry->d_name.hash != hashlen_hash(hashlen))
2159 switch (slow_dentry_cmp(parent, dentry, seq, name)) {
2162 case D_COMP_NOMATCH:
2169 if (dentry->d_name.hash_len != hashlen)
2172 if (!dentry_cmp(dentry, str, hashlen_len(hashlen)))
2179 * d_lookup - search for a dentry
2180 * @parent: parent dentry
2181 * @name: qstr of name we wish to find
2182 * Returns: dentry, or NULL
2184 * d_lookup searches the children of the parent dentry for the name in
2185 * question. If the dentry is found its reference count is incremented and the
2186 * dentry is returned. The caller must use dput to free the entry when it has
2187 * finished using it. %NULL is returned if the dentry does not exist.
2189 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2191 struct dentry *dentry;
2195 seq = read_seqbegin(&rename_lock);
2196 dentry = __d_lookup(parent, name);
2199 } while (read_seqretry(&rename_lock, seq));
2202 EXPORT_SYMBOL(d_lookup);
2205 * __d_lookup - search for a dentry (racy)
2206 * @parent: parent dentry
2207 * @name: qstr of name we wish to find
2208 * Returns: dentry, or NULL
2210 * __d_lookup is like d_lookup, however it may (rarely) return a
2211 * false-negative result due to unrelated rename activity.
2213 * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2214 * however it must be used carefully, eg. with a following d_lookup in
2215 * the case of failure.
2217 * __d_lookup callers must be commented.
2219 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2221 unsigned int len = name->len;
2222 unsigned int hash = name->hash;
2223 const unsigned char *str = name->name;
2224 struct hlist_bl_head *b = d_hash(parent, hash);
2225 struct hlist_bl_node *node;
2226 struct dentry *found = NULL;
2227 struct dentry *dentry;
2230 * Note: There is significant duplication with __d_lookup_rcu which is
2231 * required to prevent single threaded performance regressions
2232 * especially on architectures where smp_rmb (in seqcounts) are costly.
2233 * Keep the two functions in sync.
2237 * The hash list is protected using RCU.
2239 * Take d_lock when comparing a candidate dentry, to avoid races
2242 * It is possible that concurrent renames can mess up our list
2243 * walk here and result in missing our dentry, resulting in the
2244 * false-negative result. d_lookup() protects against concurrent
2245 * renames using rename_lock seqlock.
2247 * See Documentation/filesystems/path-lookup.txt for more details.
2251 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2253 if (dentry->d_name.hash != hash)
2256 spin_lock(&dentry->d_lock);
2257 if (dentry->d_parent != parent)
2259 if (d_unhashed(dentry))
2263 * It is safe to compare names since d_move() cannot
2264 * change the qstr (protected by d_lock).
2266 if (parent->d_flags & DCACHE_OP_COMPARE) {
2267 int tlen = dentry->d_name.len;
2268 const char *tname = dentry->d_name.name;
2269 if (parent->d_op->d_compare(parent, dentry, tlen, tname, name))
2272 if (dentry->d_name.len != len)
2274 if (dentry_cmp(dentry, str, len))
2278 dentry->d_lockref.count++;
2280 spin_unlock(&dentry->d_lock);
2283 spin_unlock(&dentry->d_lock);
2291 * d_hash_and_lookup - hash the qstr then search for a dentry
2292 * @dir: Directory to search in
2293 * @name: qstr of name we wish to find
2295 * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2297 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2300 * Check for a fs-specific hash function. Note that we must
2301 * calculate the standard hash first, as the d_op->d_hash()
2302 * routine may choose to leave the hash value unchanged.
2304 name->hash = full_name_hash(name->name, name->len);
2305 if (dir->d_flags & DCACHE_OP_HASH) {
2306 int err = dir->d_op->d_hash(dir, name);
2307 if (unlikely(err < 0))
2308 return ERR_PTR(err);
2310 return d_lookup(dir, name);
2312 EXPORT_SYMBOL(d_hash_and_lookup);
2315 * d_validate - verify dentry provided from insecure source (deprecated)
2316 * @dentry: The dentry alleged to be valid child of @dparent
2317 * @dparent: The parent dentry (known to be valid)
2319 * An insecure source has sent us a dentry, here we verify it and dget() it.
2320 * This is used by ncpfs in its readdir implementation.
2321 * Zero is returned in the dentry is invalid.
2323 * This function is slow for big directories, and deprecated, do not use it.
2325 int d_validate(struct dentry *dentry, struct dentry *dparent)
2327 struct dentry *child;
2329 spin_lock(&dparent->d_lock);
2330 list_for_each_entry(child, &dparent->d_subdirs, d_u.d_child) {
2331 if (dentry == child) {
2332 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
2333 __dget_dlock(dentry);
2334 spin_unlock(&dentry->d_lock);
2335 spin_unlock(&dparent->d_lock);
2339 spin_unlock(&dparent->d_lock);
2343 EXPORT_SYMBOL(d_validate);
2346 * When a file is deleted, we have two options:
2347 * - turn this dentry into a negative dentry
2348 * - unhash this dentry and free it.
2350 * Usually, we want to just turn this into
2351 * a negative dentry, but if anybody else is
2352 * currently using the dentry or the inode
2353 * we can't do that and we fall back on removing
2354 * it from the hash queues and waiting for
2355 * it to be deleted later when it has no users
2359 * d_delete - delete a dentry
2360 * @dentry: The dentry to delete
2362 * Turn the dentry into a negative dentry if possible, otherwise
2363 * remove it from the hash queues so it can be deleted later
2366 void d_delete(struct dentry * dentry)
2368 struct inode *inode;
2371 * Are we the only user?
2374 spin_lock(&dentry->d_lock);
2375 inode = dentry->d_inode;
2376 isdir = S_ISDIR(inode->i_mode);
2377 if (dentry->d_lockref.count == 1) {
2378 if (!spin_trylock(&inode->i_lock)) {
2379 spin_unlock(&dentry->d_lock);
2383 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2384 dentry_unlink_inode(dentry);
2385 fsnotify_nameremove(dentry, isdir);
2389 if (!d_unhashed(dentry))
2392 spin_unlock(&dentry->d_lock);
2394 fsnotify_nameremove(dentry, isdir);
2396 EXPORT_SYMBOL(d_delete);
2398 static void __d_rehash(struct dentry * entry, struct hlist_bl_head *b)
2400 BUG_ON(!d_unhashed(entry));
2402 entry->d_flags |= DCACHE_RCUACCESS;
2403 hlist_bl_add_head_rcu(&entry->d_hash, b);
2407 static void _d_rehash(struct dentry * entry)
2409 __d_rehash(entry, d_hash(entry->d_parent, entry->d_name.hash));
2413 * d_rehash - add an entry back to the hash
2414 * @entry: dentry to add to the hash
2416 * Adds a dentry to the hash according to its name.
2419 void d_rehash(struct dentry * entry)
2421 spin_lock(&entry->d_lock);
2423 spin_unlock(&entry->d_lock);
2425 EXPORT_SYMBOL(d_rehash);
2428 * dentry_update_name_case - update case insensitive dentry with a new name
2429 * @dentry: dentry to be updated
2432 * Update a case insensitive dentry with new case of name.
2434 * dentry must have been returned by d_lookup with name @name. Old and new
2435 * name lengths must match (ie. no d_compare which allows mismatched name
2438 * Parent inode i_mutex must be held over d_lookup and into this call (to
2439 * keep renames and concurrent inserts, and readdir(2) away).
2441 void dentry_update_name_case(struct dentry *dentry, struct qstr *name)
2443 BUG_ON(!mutex_is_locked(&dentry->d_parent->d_inode->i_mutex));
2444 BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */
2446 spin_lock(&dentry->d_lock);
2447 write_seqcount_begin(&dentry->d_seq);
2448 memcpy((unsigned char *)dentry->d_name.name, name->name, name->len);
2449 write_seqcount_end(&dentry->d_seq);
2450 spin_unlock(&dentry->d_lock);
2452 EXPORT_SYMBOL(dentry_update_name_case);
2454 static void switch_names(struct dentry *dentry, struct dentry *target)
2456 if (dname_external(target)) {
2457 if (dname_external(dentry)) {
2459 * Both external: swap the pointers
2461 swap(target->d_name.name, dentry->d_name.name);
2464 * dentry:internal, target:external. Steal target's
2465 * storage and make target internal.
2467 memcpy(target->d_iname, dentry->d_name.name,
2468 dentry->d_name.len + 1);
2469 dentry->d_name.name = target->d_name.name;
2470 target->d_name.name = target->d_iname;
2473 if (dname_external(dentry)) {
2475 * dentry:external, target:internal. Give dentry's
2476 * storage to target and make dentry internal
2478 memcpy(dentry->d_iname, target->d_name.name,
2479 target->d_name.len + 1);
2480 target->d_name.name = dentry->d_name.name;
2481 dentry->d_name.name = dentry->d_iname;
2484 * Both are internal. Just copy target to dentry
2486 memcpy(dentry->d_iname, target->d_name.name,
2487 target->d_name.len + 1);
2488 dentry->d_name.len = target->d_name.len;
2492 swap(dentry->d_name.len, target->d_name.len);
2495 static void dentry_lock_for_move(struct dentry *dentry, struct dentry *target)
2498 * XXXX: do we really need to take target->d_lock?
2500 if (IS_ROOT(dentry) || dentry->d_parent == target->d_parent)
2501 spin_lock(&target->d_parent->d_lock);
2503 if (d_ancestor(dentry->d_parent, target->d_parent)) {
2504 spin_lock(&dentry->d_parent->d_lock);
2505 spin_lock_nested(&target->d_parent->d_lock,
2506 DENTRY_D_LOCK_NESTED);
2508 spin_lock(&target->d_parent->d_lock);
2509 spin_lock_nested(&dentry->d_parent->d_lock,
2510 DENTRY_D_LOCK_NESTED);
2513 if (target < dentry) {
2514 spin_lock_nested(&target->d_lock, 2);
2515 spin_lock_nested(&dentry->d_lock, 3);
2517 spin_lock_nested(&dentry->d_lock, 2);
2518 spin_lock_nested(&target->d_lock, 3);
2522 static void dentry_unlock_parents_for_move(struct dentry *dentry,
2523 struct dentry *target)
2525 if (target->d_parent != dentry->d_parent)
2526 spin_unlock(&dentry->d_parent->d_lock);
2527 if (target->d_parent != target)
2528 spin_unlock(&target->d_parent->d_lock);
2532 * When switching names, the actual string doesn't strictly have to
2533 * be preserved in the target - because we're dropping the target
2534 * anyway. As such, we can just do a simple memcpy() to copy over
2535 * the new name before we switch.
2537 * Note that we have to be a lot more careful about getting the hash
2538 * switched - we have to switch the hash value properly even if it
2539 * then no longer matches the actual (corrupted) string of the target.
2540 * The hash value has to match the hash queue that the dentry is on..
2543 * __d_move - move a dentry
2544 * @dentry: entry to move
2545 * @target: new dentry
2547 * Update the dcache to reflect the move of a file name. Negative
2548 * dcache entries should not be moved in this way. Caller must hold
2549 * rename_lock, the i_mutex of the source and target directories,
2550 * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2552 static void __d_move(struct dentry * dentry, struct dentry * target)
2554 if (!dentry->d_inode)
2555 printk(KERN_WARNING "VFS: moving negative dcache entry\n");
2557 BUG_ON(d_ancestor(dentry, target));
2558 BUG_ON(d_ancestor(target, dentry));
2560 dentry_lock_for_move(dentry, target);
2562 write_seqcount_begin(&dentry->d_seq);
2563 write_seqcount_begin(&target->d_seq);
2565 /* __d_drop does write_seqcount_barrier, but they're OK to nest. */
2568 * Move the dentry to the target hash queue. Don't bother checking
2569 * for the same hash queue because of how unlikely it is.
2572 __d_rehash(dentry, d_hash(target->d_parent, target->d_name.hash));
2574 /* Unhash the target: dput() will then get rid of it */
2577 list_del(&dentry->d_u.d_child);
2578 list_del(&target->d_u.d_child);
2580 /* Switch the names.. */
2581 switch_names(dentry, target);
2582 swap(dentry->d_name.hash, target->d_name.hash);
2584 /* ... and switch the parents */
2585 if (IS_ROOT(dentry)) {
2586 dentry->d_parent = target->d_parent;
2587 target->d_parent = target;
2588 INIT_LIST_HEAD(&target->d_u.d_child);
2590 swap(dentry->d_parent, target->d_parent);
2592 /* And add them back to the (new) parent lists */
2593 list_add(&target->d_u.d_child, &target->d_parent->d_subdirs);
2596 list_add(&dentry->d_u.d_child, &dentry->d_parent->d_subdirs);
2598 write_seqcount_end(&target->d_seq);
2599 write_seqcount_end(&dentry->d_seq);
2601 dentry_unlock_parents_for_move(dentry, target);
2602 spin_unlock(&target->d_lock);
2603 fsnotify_d_move(dentry);
2604 spin_unlock(&dentry->d_lock);
2608 * d_move - move a dentry
2609 * @dentry: entry to move
2610 * @target: new dentry
2612 * Update the dcache to reflect the move of a file name. Negative
2613 * dcache entries should not be moved in this way. See the locking
2614 * requirements for __d_move.
2616 void d_move(struct dentry *dentry, struct dentry *target)
2618 write_seqlock(&rename_lock);
2619 __d_move(dentry, target);
2620 write_sequnlock(&rename_lock);
2622 EXPORT_SYMBOL(d_move);
2625 * d_ancestor - search for an ancestor
2626 * @p1: ancestor dentry
2629 * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2630 * an ancestor of p2, else NULL.
2632 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2636 for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2637 if (p->d_parent == p1)
2644 * This helper attempts to cope with remotely renamed directories
2646 * It assumes that the caller is already holding
2647 * dentry->d_parent->d_inode->i_mutex, inode->i_lock and rename_lock
2649 * Note: If ever the locking in lock_rename() changes, then please
2650 * remember to update this too...
2652 static struct dentry *__d_unalias(struct inode *inode,
2653 struct dentry *dentry, struct dentry *alias)
2655 struct mutex *m1 = NULL, *m2 = NULL;
2656 struct dentry *ret = ERR_PTR(-EBUSY);
2658 /* If alias and dentry share a parent, then no extra locks required */
2659 if (alias->d_parent == dentry->d_parent)
2662 /* See lock_rename() */
2663 if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2665 m1 = &dentry->d_sb->s_vfs_rename_mutex;
2666 if (!mutex_trylock(&alias->d_parent->d_inode->i_mutex))
2668 m2 = &alias->d_parent->d_inode->i_mutex;
2670 if (likely(!d_mountpoint(alias))) {
2671 __d_move(alias, dentry);
2675 spin_unlock(&inode->i_lock);
2684 * Prepare an anonymous dentry for life in the superblock's dentry tree as a
2685 * named dentry in place of the dentry to be replaced.
2686 * returns with anon->d_lock held!
2688 static void __d_materialise_dentry(struct dentry *dentry, struct dentry *anon)
2690 struct dentry *dparent;
2692 dentry_lock_for_move(anon, dentry);
2694 write_seqcount_begin(&dentry->d_seq);
2695 write_seqcount_begin(&anon->d_seq);
2697 dparent = dentry->d_parent;
2699 switch_names(dentry, anon);
2700 swap(dentry->d_name.hash, anon->d_name.hash);
2702 dentry->d_parent = dentry;
2703 list_del_init(&dentry->d_u.d_child);
2704 anon->d_parent = dparent;
2705 list_move(&anon->d_u.d_child, &dparent->d_subdirs);
2707 write_seqcount_end(&dentry->d_seq);
2708 write_seqcount_end(&anon->d_seq);
2710 dentry_unlock_parents_for_move(anon, dentry);
2711 spin_unlock(&dentry->d_lock);
2713 /* anon->d_lock still locked, returns locked */
2714 anon->d_flags &= ~DCACHE_DISCONNECTED;
2718 * d_materialise_unique - introduce an inode into the tree
2719 * @dentry: candidate dentry
2720 * @inode: inode to bind to the dentry, to which aliases may be attached
2722 * Introduces an dentry into the tree, substituting an extant disconnected
2723 * root directory alias in its place if there is one. Caller must hold the
2724 * i_mutex of the parent directory.
2726 struct dentry *d_materialise_unique(struct dentry *dentry, struct inode *inode)
2728 struct dentry *actual;
2730 BUG_ON(!d_unhashed(dentry));
2734 __d_instantiate(dentry, NULL);
2739 spin_lock(&inode->i_lock);
2741 if (S_ISDIR(inode->i_mode)) {
2742 struct dentry *alias;
2744 /* Does an aliased dentry already exist? */
2745 alias = __d_find_alias(inode, 0);
2748 write_seqlock(&rename_lock);
2750 if (d_ancestor(alias, dentry)) {
2751 /* Check for loops */
2752 actual = ERR_PTR(-ELOOP);
2753 spin_unlock(&inode->i_lock);
2754 } else if (IS_ROOT(alias)) {
2755 /* Is this an anonymous mountpoint that we
2756 * could splice into our tree? */
2757 __d_materialise_dentry(dentry, alias);
2758 write_sequnlock(&rename_lock);
2762 /* Nope, but we must(!) avoid directory
2763 * aliasing. This drops inode->i_lock */
2764 actual = __d_unalias(inode, dentry, alias);
2766 write_sequnlock(&rename_lock);
2767 if (IS_ERR(actual)) {
2768 if (PTR_ERR(actual) == -ELOOP)
2769 pr_warn_ratelimited(
2770 "VFS: Lookup of '%s' in %s %s"
2771 " would have caused loop\n",
2772 dentry->d_name.name,
2773 inode->i_sb->s_type->name,
2781 /* Add a unique reference */
2782 actual = __d_instantiate_unique(dentry, inode);
2786 BUG_ON(!d_unhashed(actual));
2788 spin_lock(&actual->d_lock);
2791 spin_unlock(&actual->d_lock);
2792 spin_unlock(&inode->i_lock);
2794 if (actual == dentry) {
2795 security_d_instantiate(dentry, inode);
2802 EXPORT_SYMBOL_GPL(d_materialise_unique);
2804 static int prepend(char **buffer, int *buflen, const char *str, int namelen)
2808 return -ENAMETOOLONG;
2810 memcpy(*buffer, str, namelen);
2815 * prepend_name - prepend a pathname in front of current buffer pointer
2816 * @buffer: buffer pointer
2817 * @buflen: allocated length of the buffer
2818 * @name: name string and length qstr structure
2820 * With RCU path tracing, it may race with d_move(). Use ACCESS_ONCE() to
2821 * make sure that either the old or the new name pointer and length are
2822 * fetched. However, there may be mismatch between length and pointer.
2823 * The length cannot be trusted, we need to copy it byte-by-byte until
2824 * the length is reached or a null byte is found. It also prepends "/" at
2825 * the beginning of the name. The sequence number check at the caller will
2826 * retry it again when a d_move() does happen. So any garbage in the buffer
2827 * due to mismatched pointer and length will be discarded.
2829 static int prepend_name(char **buffer, int *buflen, struct qstr *name)
2831 const char *dname = ACCESS_ONCE(name->name);
2832 u32 dlen = ACCESS_ONCE(name->len);
2835 if (*buflen < dlen + 1)
2836 return -ENAMETOOLONG;
2837 *buflen -= dlen + 1;
2838 p = *buffer -= dlen + 1;
2850 * prepend_path - Prepend path string to a buffer
2851 * @path: the dentry/vfsmount to report
2852 * @root: root vfsmnt/dentry
2853 * @buffer: pointer to the end of the buffer
2854 * @buflen: pointer to buffer length
2856 * The function will first try to write out the pathname without taking any
2857 * lock other than the RCU read lock to make sure that dentries won't go away.
2858 * It only checks the sequence number of the global rename_lock as any change
2859 * in the dentry's d_seq will be preceded by changes in the rename_lock
2860 * sequence number. If the sequence number had been changed, it will restart
2861 * the whole pathname back-tracing sequence again by taking the rename_lock.
2862 * In this case, there is no need to take the RCU read lock as the recursive
2863 * parent pointer references will keep the dentry chain alive as long as no
2864 * rename operation is performed.
2866 static int prepend_path(const struct path *path,
2867 const struct path *root,
2868 char **buffer, int *buflen)
2870 struct dentry *dentry = path->dentry;
2871 struct vfsmount *vfsmnt = path->mnt;
2872 struct mount *mnt = real_mount(vfsmnt);
2882 read_seqbegin_or_lock(&rename_lock, &seq);
2883 while (dentry != root->dentry || vfsmnt != root->mnt) {
2884 struct dentry * parent;
2886 if (dentry == vfsmnt->mnt_root || IS_ROOT(dentry)) {
2888 if (mnt_has_parent(mnt)) {
2889 dentry = mnt->mnt_mountpoint;
2890 mnt = mnt->mnt_parent;
2895 * Filesystems needing to implement special "root names"
2896 * should do so with ->d_dname()
2898 if (IS_ROOT(dentry) &&
2899 (dentry->d_name.len != 1 ||
2900 dentry->d_name.name[0] != '/')) {
2901 WARN(1, "Root dentry has weird name <%.*s>\n",
2902 (int) dentry->d_name.len,
2903 dentry->d_name.name);
2906 error = is_mounted(vfsmnt) ? 1 : 2;
2909 parent = dentry->d_parent;
2911 error = prepend_name(&bptr, &blen, &dentry->d_name);
2919 if (need_seqretry(&rename_lock, seq)) {
2923 done_seqretry(&rename_lock, seq);
2925 if (error >= 0 && bptr == *buffer) {
2927 error = -ENAMETOOLONG;
2937 * __d_path - return the path of a dentry
2938 * @path: the dentry/vfsmount to report
2939 * @root: root vfsmnt/dentry
2940 * @buf: buffer to return value in
2941 * @buflen: buffer length
2943 * Convert a dentry into an ASCII path name.
2945 * Returns a pointer into the buffer or an error code if the
2946 * path was too long.
2948 * "buflen" should be positive.
2950 * If the path is not reachable from the supplied root, return %NULL.
2952 char *__d_path(const struct path *path,
2953 const struct path *root,
2954 char *buf, int buflen)
2956 char *res = buf + buflen;
2959 prepend(&res, &buflen, "\0", 1);
2960 br_read_lock(&vfsmount_lock);
2961 error = prepend_path(path, root, &res, &buflen);
2962 br_read_unlock(&vfsmount_lock);
2965 return ERR_PTR(error);
2971 char *d_absolute_path(const struct path *path,
2972 char *buf, int buflen)
2974 struct path root = {};
2975 char *res = buf + buflen;
2978 prepend(&res, &buflen, "\0", 1);
2979 br_read_lock(&vfsmount_lock);
2980 error = prepend_path(path, &root, &res, &buflen);
2981 br_read_unlock(&vfsmount_lock);
2986 return ERR_PTR(error);
2991 * same as __d_path but appends "(deleted)" for unlinked files.
2993 static int path_with_deleted(const struct path *path,
2994 const struct path *root,
2995 char **buf, int *buflen)
2997 prepend(buf, buflen, "\0", 1);
2998 if (d_unlinked(path->dentry)) {
2999 int error = prepend(buf, buflen, " (deleted)", 10);
3004 return prepend_path(path, root, buf, buflen);
3007 static int prepend_unreachable(char **buffer, int *buflen)
3009 return prepend(buffer, buflen, "(unreachable)", 13);
3012 static void get_fs_root_rcu(struct fs_struct *fs, struct path *root)
3017 seq = read_seqcount_begin(&fs->seq);
3019 } while (read_seqcount_retry(&fs->seq, seq));
3023 * d_path - return the path of a dentry
3024 * @path: path to report
3025 * @buf: buffer to return value in
3026 * @buflen: buffer length
3028 * Convert a dentry into an ASCII path name. If the entry has been deleted
3029 * the string " (deleted)" is appended. Note that this is ambiguous.
3031 * Returns a pointer into the buffer or an error code if the path was
3032 * too long. Note: Callers should use the returned pointer, not the passed
3033 * in buffer, to use the name! The implementation often starts at an offset
3034 * into the buffer, and may leave 0 bytes at the start.
3036 * "buflen" should be positive.
3038 char *d_path(const struct path *path, char *buf, int buflen)
3040 char *res = buf + buflen;
3045 * We have various synthetic filesystems that never get mounted. On
3046 * these filesystems dentries are never used for lookup purposes, and
3047 * thus don't need to be hashed. They also don't need a name until a
3048 * user wants to identify the object in /proc/pid/fd/. The little hack
3049 * below allows us to generate a name for these objects on demand:
3051 if (path->dentry->d_op && path->dentry->d_op->d_dname)
3052 return path->dentry->d_op->d_dname(path->dentry, buf, buflen);
3055 get_fs_root_rcu(current->fs, &root);
3056 br_read_lock(&vfsmount_lock);
3057 error = path_with_deleted(path, &root, &res, &buflen);
3058 br_read_unlock(&vfsmount_lock);
3062 res = ERR_PTR(error);
3065 EXPORT_SYMBOL(d_path);
3068 * Helper function for dentry_operations.d_dname() members
3070 char *dynamic_dname(struct dentry *dentry, char *buffer, int buflen,
3071 const char *fmt, ...)
3077 va_start(args, fmt);
3078 sz = vsnprintf(temp, sizeof(temp), fmt, args) + 1;
3081 if (sz > sizeof(temp) || sz > buflen)
3082 return ERR_PTR(-ENAMETOOLONG);
3084 buffer += buflen - sz;
3085 return memcpy(buffer, temp, sz);
3088 char *simple_dname(struct dentry *dentry, char *buffer, int buflen)
3090 char *end = buffer + buflen;
3091 /* these dentries are never renamed, so d_lock is not needed */
3092 if (prepend(&end, &buflen, " (deleted)", 11) ||
3093 prepend(&end, &buflen, dentry->d_name.name, dentry->d_name.len) ||
3094 prepend(&end, &buflen, "/", 1))
3095 end = ERR_PTR(-ENAMETOOLONG);
3100 * Write full pathname from the root of the filesystem into the buffer.
3102 static char *__dentry_path(struct dentry *dentry, char *buf, int buflen)
3112 prepend(&end, &len, "\0", 1);
3118 read_seqbegin_or_lock(&rename_lock, &seq);
3119 while (!IS_ROOT(dentry)) {
3120 struct dentry *parent = dentry->d_parent;
3124 error = prepend_name(&end, &len, &dentry->d_name);
3133 if (need_seqretry(&rename_lock, seq)) {
3137 done_seqretry(&rename_lock, seq);
3142 return ERR_PTR(-ENAMETOOLONG);
3145 char *dentry_path_raw(struct dentry *dentry, char *buf, int buflen)
3147 return __dentry_path(dentry, buf, buflen);
3149 EXPORT_SYMBOL(dentry_path_raw);
3151 char *dentry_path(struct dentry *dentry, char *buf, int buflen)
3156 if (d_unlinked(dentry)) {
3158 if (prepend(&p, &buflen, "//deleted", 10) != 0)
3162 retval = __dentry_path(dentry, buf, buflen);
3163 if (!IS_ERR(retval) && p)
3164 *p = '/'; /* restore '/' overriden with '\0' */
3167 return ERR_PTR(-ENAMETOOLONG);
3170 static void get_fs_root_and_pwd_rcu(struct fs_struct *fs, struct path *root,
3176 seq = read_seqcount_begin(&fs->seq);
3179 } while (read_seqcount_retry(&fs->seq, seq));
3183 * NOTE! The user-level library version returns a
3184 * character pointer. The kernel system call just
3185 * returns the length of the buffer filled (which
3186 * includes the ending '\0' character), or a negative
3187 * error value. So libc would do something like
3189 * char *getcwd(char * buf, size_t size)
3193 * retval = sys_getcwd(buf, size);
3200 SYSCALL_DEFINE2(getcwd, char __user *, buf, unsigned long, size)
3203 struct path pwd, root;
3204 char *page = __getname();
3210 get_fs_root_and_pwd_rcu(current->fs, &root, &pwd);
3213 br_read_lock(&vfsmount_lock);
3214 if (!d_unlinked(pwd.dentry)) {
3216 char *cwd = page + PATH_MAX;
3217 int buflen = PATH_MAX;
3219 prepend(&cwd, &buflen, "\0", 1);
3220 error = prepend_path(&pwd, &root, &cwd, &buflen);
3221 br_read_unlock(&vfsmount_lock);
3227 /* Unreachable from current root */
3229 error = prepend_unreachable(&cwd, &buflen);
3235 len = PATH_MAX + page - cwd;
3238 if (copy_to_user(buf, cwd, len))
3242 br_read_unlock(&vfsmount_lock);
3252 * Test whether new_dentry is a subdirectory of old_dentry.
3254 * Trivially implemented using the dcache structure
3258 * is_subdir - is new dentry a subdirectory of old_dentry
3259 * @new_dentry: new dentry
3260 * @old_dentry: old dentry
3262 * Returns 1 if new_dentry is a subdirectory of the parent (at any depth).
3263 * Returns 0 otherwise.
3264 * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3267 int is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3272 if (new_dentry == old_dentry)
3276 /* for restarting inner loop in case of seq retry */
3277 seq = read_seqbegin(&rename_lock);
3279 * Need rcu_readlock to protect against the d_parent trashing
3283 if (d_ancestor(old_dentry, new_dentry))
3288 } while (read_seqretry(&rename_lock, seq));
3293 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3295 struct dentry *root = data;
3296 if (dentry != root) {
3297 if (d_unhashed(dentry) || !dentry->d_inode)
3300 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3301 dentry->d_flags |= DCACHE_GENOCIDE;
3302 dentry->d_lockref.count--;
3305 return D_WALK_CONTINUE;
3308 void d_genocide(struct dentry *parent)
3310 d_walk(parent, parent, d_genocide_kill, NULL);
3313 void d_tmpfile(struct dentry *dentry, struct inode *inode)
3315 inode_dec_link_count(inode);
3316 BUG_ON(dentry->d_name.name != dentry->d_iname ||
3317 !hlist_unhashed(&dentry->d_alias) ||
3318 !d_unlinked(dentry));
3319 spin_lock(&dentry->d_parent->d_lock);
3320 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3321 dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3322 (unsigned long long)inode->i_ino);
3323 spin_unlock(&dentry->d_lock);
3324 spin_unlock(&dentry->d_parent->d_lock);
3325 d_instantiate(dentry, inode);
3327 EXPORT_SYMBOL(d_tmpfile);
3329 static __initdata unsigned long dhash_entries;
3330 static int __init set_dhash_entries(char *str)
3334 dhash_entries = simple_strtoul(str, &str, 0);
3337 __setup("dhash_entries=", set_dhash_entries);
3339 static void __init dcache_init_early(void)
3343 /* If hashes are distributed across NUMA nodes, defer
3344 * hash allocation until vmalloc space is available.
3350 alloc_large_system_hash("Dentry cache",
3351 sizeof(struct hlist_bl_head),
3360 for (loop = 0; loop < (1U << d_hash_shift); loop++)
3361 INIT_HLIST_BL_HEAD(dentry_hashtable + loop);
3364 static void __init dcache_init(void)
3369 * A constructor could be added for stable state like the lists,
3370 * but it is probably not worth it because of the cache nature
3373 dentry_cache = KMEM_CACHE(dentry,
3374 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD);
3376 /* Hash may have been set up in dcache_init_early */
3381 alloc_large_system_hash("Dentry cache",
3382 sizeof(struct hlist_bl_head),
3391 for (loop = 0; loop < (1U << d_hash_shift); loop++)
3392 INIT_HLIST_BL_HEAD(dentry_hashtable + loop);
3395 /* SLAB cache for __getname() consumers */
3396 struct kmem_cache *names_cachep __read_mostly;
3397 EXPORT_SYMBOL(names_cachep);
3399 EXPORT_SYMBOL(d_genocide);
3401 void __init vfs_caches_init_early(void)
3403 dcache_init_early();
3407 void __init vfs_caches_init(unsigned long mempages)
3409 unsigned long reserve;
3411 /* Base hash sizes on available memory, with a reserve equal to
3412 150% of current kernel size */
3414 reserve = min((mempages - nr_free_pages()) * 3/2, mempages - 1);
3415 mempages -= reserve;
3417 names_cachep = kmem_cache_create("names_cache", PATH_MAX, 0,
3418 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3422 files_init(mempages);