mm: compaction: prevent kswapd compacting memory to reduce CPU usage
[linux/fpc-iii.git] / fs / dcache.c
blobad25c4cec7d50a7185459609bf6e5f63b9e9dcbd
1 /*
2 * fs/dcache.c
4 * Complete reimplementation
5 * (C) 1997 Thomas Schoebel-Theuer,
6 * with heavy changes by Linus Torvalds
7 */
9 /*
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>
19 #include <linux/mm.h>
20 #include <linux/fs.h>
21 #include <linux/fsnotify.h>
22 #include <linux/slab.h>
23 #include <linux/init.h>
24 #include <linux/hash.h>
25 #include <linux/cache.h>
26 #include <linux/module.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 "internal.h"
41 * Usage:
42 * dcache->d_inode->i_lock protects:
43 * - i_dentry, d_alias, d_inode of aliases
44 * dcache_hash_bucket lock protects:
45 * - the dcache hash table
46 * s_anon bl list spinlock protects:
47 * - the s_anon list (see __d_drop)
48 * dcache_lru_lock protects:
49 * - the dcache lru lists and counters
50 * d_lock protects:
51 * - d_flags
52 * - d_name
53 * - d_lru
54 * - d_count
55 * - d_unhashed()
56 * - d_parent and d_subdirs
57 * - childrens' d_child and d_parent
58 * - d_alias, d_inode
60 * Ordering:
61 * dentry->d_inode->i_lock
62 * dentry->d_lock
63 * dcache_lru_lock
64 * dcache_hash_bucket lock
65 * s_anon lock
67 * If there is an ancestor relationship:
68 * dentry->d_parent->...->d_parent->d_lock
69 * ...
70 * dentry->d_parent->d_lock
71 * dentry->d_lock
73 * If no ancestor relationship:
74 * if (dentry1 < dentry2)
75 * dentry1->d_lock
76 * dentry2->d_lock
78 int sysctl_vfs_cache_pressure __read_mostly = 100;
79 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
81 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(dcache_lru_lock);
82 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
84 EXPORT_SYMBOL(rename_lock);
86 static struct kmem_cache *dentry_cache __read_mostly;
89 * This is the single most critical data structure when it comes
90 * to the dcache: the hashtable for lookups. Somebody should try
91 * to make this good - I've just made it work.
93 * This hash-function tries to avoid losing too many bits of hash
94 * information, yet avoid using a prime hash-size or similar.
96 #define D_HASHBITS d_hash_shift
97 #define D_HASHMASK d_hash_mask
99 static unsigned int d_hash_mask __read_mostly;
100 static unsigned int d_hash_shift __read_mostly;
102 struct dcache_hash_bucket {
103 struct hlist_bl_head head;
105 static struct dcache_hash_bucket *dentry_hashtable __read_mostly;
107 static inline struct dcache_hash_bucket *d_hash(struct dentry *parent,
108 unsigned long hash)
110 hash += ((unsigned long) parent ^ GOLDEN_RATIO_PRIME) / L1_CACHE_BYTES;
111 hash = hash ^ ((hash ^ GOLDEN_RATIO_PRIME) >> D_HASHBITS);
112 return dentry_hashtable + (hash & D_HASHMASK);
115 static inline void spin_lock_bucket(struct dcache_hash_bucket *b)
117 bit_spin_lock(0, (unsigned long *)&b->head.first);
120 static inline void spin_unlock_bucket(struct dcache_hash_bucket *b)
122 __bit_spin_unlock(0, (unsigned long *)&b->head.first);
125 /* Statistics gathering. */
126 struct dentry_stat_t dentry_stat = {
127 .age_limit = 45,
130 static DEFINE_PER_CPU(unsigned int, nr_dentry);
132 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
133 static int get_nr_dentry(void)
135 int i;
136 int sum = 0;
137 for_each_possible_cpu(i)
138 sum += per_cpu(nr_dentry, i);
139 return sum < 0 ? 0 : sum;
142 int proc_nr_dentry(ctl_table *table, int write, void __user *buffer,
143 size_t *lenp, loff_t *ppos)
145 dentry_stat.nr_dentry = get_nr_dentry();
146 return proc_dointvec(table, write, buffer, lenp, ppos);
148 #endif
150 static void __d_free(struct rcu_head *head)
152 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
154 WARN_ON(!list_empty(&dentry->d_alias));
155 if (dname_external(dentry))
156 kfree(dentry->d_name.name);
157 kmem_cache_free(dentry_cache, dentry);
161 * no locks, please.
163 static void d_free(struct dentry *dentry)
165 BUG_ON(dentry->d_count);
166 this_cpu_dec(nr_dentry);
167 if (dentry->d_op && dentry->d_op->d_release)
168 dentry->d_op->d_release(dentry);
170 /* if dentry was never inserted into hash, immediate free is OK */
171 if (hlist_bl_unhashed(&dentry->d_hash))
172 __d_free(&dentry->d_u.d_rcu);
173 else
174 call_rcu(&dentry->d_u.d_rcu, __d_free);
178 * dentry_rcuwalk_barrier - invalidate in-progress rcu-walk lookups
179 * @dentry: the target dentry
180 * After this call, in-progress rcu-walk path lookup will fail. This
181 * should be called after unhashing, and after changing d_inode (if
182 * the dentry has not already been unhashed).
184 static inline void dentry_rcuwalk_barrier(struct dentry *dentry)
186 assert_spin_locked(&dentry->d_lock);
187 /* Go through a barrier */
188 write_seqcount_barrier(&dentry->d_seq);
192 * Release the dentry's inode, using the filesystem
193 * d_iput() operation if defined. Dentry has no refcount
194 * and is unhashed.
196 static void dentry_iput(struct dentry * dentry)
197 __releases(dentry->d_lock)
198 __releases(dentry->d_inode->i_lock)
200 struct inode *inode = dentry->d_inode;
201 if (inode) {
202 dentry->d_inode = NULL;
203 list_del_init(&dentry->d_alias);
204 spin_unlock(&dentry->d_lock);
205 spin_unlock(&inode->i_lock);
206 if (!inode->i_nlink)
207 fsnotify_inoderemove(inode);
208 if (dentry->d_op && dentry->d_op->d_iput)
209 dentry->d_op->d_iput(dentry, inode);
210 else
211 iput(inode);
212 } else {
213 spin_unlock(&dentry->d_lock);
218 * Release the dentry's inode, using the filesystem
219 * d_iput() operation if defined. dentry remains in-use.
221 static void dentry_unlink_inode(struct dentry * dentry)
222 __releases(dentry->d_lock)
223 __releases(dentry->d_inode->i_lock)
225 struct inode *inode = dentry->d_inode;
226 dentry->d_inode = NULL;
227 list_del_init(&dentry->d_alias);
228 dentry_rcuwalk_barrier(dentry);
229 spin_unlock(&dentry->d_lock);
230 spin_unlock(&inode->i_lock);
231 if (!inode->i_nlink)
232 fsnotify_inoderemove(inode);
233 if (dentry->d_op && dentry->d_op->d_iput)
234 dentry->d_op->d_iput(dentry, inode);
235 else
236 iput(inode);
240 * dentry_lru_(add|del|move_tail) must be called with d_lock held.
242 static void dentry_lru_add(struct dentry *dentry)
244 if (list_empty(&dentry->d_lru)) {
245 spin_lock(&dcache_lru_lock);
246 list_add(&dentry->d_lru, &dentry->d_sb->s_dentry_lru);
247 dentry->d_sb->s_nr_dentry_unused++;
248 dentry_stat.nr_unused++;
249 spin_unlock(&dcache_lru_lock);
253 static void __dentry_lru_del(struct dentry *dentry)
255 list_del_init(&dentry->d_lru);
256 dentry->d_sb->s_nr_dentry_unused--;
257 dentry_stat.nr_unused--;
260 static void dentry_lru_del(struct dentry *dentry)
262 if (!list_empty(&dentry->d_lru)) {
263 spin_lock(&dcache_lru_lock);
264 __dentry_lru_del(dentry);
265 spin_unlock(&dcache_lru_lock);
269 static void dentry_lru_move_tail(struct dentry *dentry)
271 spin_lock(&dcache_lru_lock);
272 if (list_empty(&dentry->d_lru)) {
273 list_add_tail(&dentry->d_lru, &dentry->d_sb->s_dentry_lru);
274 dentry->d_sb->s_nr_dentry_unused++;
275 dentry_stat.nr_unused++;
276 } else {
277 list_move_tail(&dentry->d_lru, &dentry->d_sb->s_dentry_lru);
279 spin_unlock(&dcache_lru_lock);
283 * d_kill - kill dentry and return parent
284 * @dentry: dentry to kill
285 * @parent: parent dentry
287 * The dentry must already be unhashed and removed from the LRU.
289 * If this is the root of the dentry tree, return NULL.
291 * dentry->d_lock and parent->d_lock must be held by caller, and are dropped by
292 * d_kill.
294 static struct dentry *d_kill(struct dentry *dentry, struct dentry *parent)
295 __releases(dentry->d_lock)
296 __releases(parent->d_lock)
297 __releases(dentry->d_inode->i_lock)
299 list_del(&dentry->d_u.d_child);
301 * Inform try_to_ascend() that we are no longer attached to the
302 * dentry tree
304 dentry->d_flags |= DCACHE_DISCONNECTED;
305 if (parent)
306 spin_unlock(&parent->d_lock);
307 dentry_iput(dentry);
309 * dentry_iput drops the locks, at which point nobody (except
310 * transient RCU lookups) can reach this dentry.
312 d_free(dentry);
313 return parent;
317 * d_drop - drop a dentry
318 * @dentry: dentry to drop
320 * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
321 * be found through a VFS lookup any more. Note that this is different from
322 * deleting the dentry - d_delete will try to mark the dentry negative if
323 * possible, giving a successful _negative_ lookup, while d_drop will
324 * just make the cache lookup fail.
326 * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
327 * reason (NFS timeouts or autofs deletes).
329 * __d_drop requires dentry->d_lock.
331 void __d_drop(struct dentry *dentry)
333 if (!(dentry->d_flags & DCACHE_UNHASHED)) {
334 if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED)) {
335 bit_spin_lock(0,
336 (unsigned long *)&dentry->d_sb->s_anon.first);
337 dentry->d_flags |= DCACHE_UNHASHED;
338 hlist_bl_del_init(&dentry->d_hash);
339 __bit_spin_unlock(0,
340 (unsigned long *)&dentry->d_sb->s_anon.first);
341 } else {
342 struct dcache_hash_bucket *b;
343 b = d_hash(dentry->d_parent, dentry->d_name.hash);
344 spin_lock_bucket(b);
346 * We may not actually need to put DCACHE_UNHASHED
347 * manipulations under the hash lock, but follow
348 * the principle of least surprise.
350 dentry->d_flags |= DCACHE_UNHASHED;
351 hlist_bl_del_rcu(&dentry->d_hash);
352 spin_unlock_bucket(b);
353 dentry_rcuwalk_barrier(dentry);
357 EXPORT_SYMBOL(__d_drop);
359 void d_drop(struct dentry *dentry)
361 spin_lock(&dentry->d_lock);
362 __d_drop(dentry);
363 spin_unlock(&dentry->d_lock);
365 EXPORT_SYMBOL(d_drop);
368 * Finish off a dentry we've decided to kill.
369 * dentry->d_lock must be held, returns with it unlocked.
370 * If ref is non-zero, then decrement the refcount too.
371 * Returns dentry requiring refcount drop, or NULL if we're done.
373 static inline struct dentry *dentry_kill(struct dentry *dentry, int ref)
374 __releases(dentry->d_lock)
376 struct inode *inode;
377 struct dentry *parent;
379 inode = dentry->d_inode;
380 if (inode && !spin_trylock(&inode->i_lock)) {
381 relock:
382 spin_unlock(&dentry->d_lock);
383 cpu_relax();
384 return dentry; /* try again with same dentry */
386 if (IS_ROOT(dentry))
387 parent = NULL;
388 else
389 parent = dentry->d_parent;
390 if (parent && !spin_trylock(&parent->d_lock)) {
391 if (inode)
392 spin_unlock(&inode->i_lock);
393 goto relock;
396 if (ref)
397 dentry->d_count--;
398 /* if dentry was on the d_lru list delete it from there */
399 dentry_lru_del(dentry);
400 /* if it was on the hash then remove it */
401 __d_drop(dentry);
402 return d_kill(dentry, parent);
406 * This is dput
408 * This is complicated by the fact that we do not want to put
409 * dentries that are no longer on any hash chain on the unused
410 * list: we'd much rather just get rid of them immediately.
412 * However, that implies that we have to traverse the dentry
413 * tree upwards to the parents which might _also_ now be
414 * scheduled for deletion (it may have been only waiting for
415 * its last child to go away).
417 * This tail recursion is done by hand as we don't want to depend
418 * on the compiler to always get this right (gcc generally doesn't).
419 * Real recursion would eat up our stack space.
423 * dput - release a dentry
424 * @dentry: dentry to release
426 * Release a dentry. This will drop the usage count and if appropriate
427 * call the dentry unlink method as well as removing it from the queues and
428 * releasing its resources. If the parent dentries were scheduled for release
429 * they too may now get deleted.
431 void dput(struct dentry *dentry)
433 if (!dentry)
434 return;
436 repeat:
437 if (dentry->d_count == 1)
438 might_sleep();
439 spin_lock(&dentry->d_lock);
440 BUG_ON(!dentry->d_count);
441 if (dentry->d_count > 1) {
442 dentry->d_count--;
443 spin_unlock(&dentry->d_lock);
444 return;
447 if (dentry->d_flags & DCACHE_OP_DELETE) {
448 if (dentry->d_op->d_delete(dentry))
449 goto kill_it;
452 /* Unreachable? Get rid of it */
453 if (d_unhashed(dentry))
454 goto kill_it;
456 /* Otherwise leave it cached and ensure it's on the LRU */
457 dentry->d_flags |= DCACHE_REFERENCED;
458 dentry_lru_add(dentry);
460 dentry->d_count--;
461 spin_unlock(&dentry->d_lock);
462 return;
464 kill_it:
465 dentry = dentry_kill(dentry, 1);
466 if (dentry)
467 goto repeat;
469 EXPORT_SYMBOL(dput);
472 * d_invalidate - invalidate a dentry
473 * @dentry: dentry to invalidate
475 * Try to invalidate the dentry if it turns out to be
476 * possible. If there are other dentries that can be
477 * reached through this one we can't delete it and we
478 * return -EBUSY. On success we return 0.
480 * no dcache lock.
483 int d_invalidate(struct dentry * dentry)
486 * If it's already been dropped, return OK.
488 spin_lock(&dentry->d_lock);
489 if (d_unhashed(dentry)) {
490 spin_unlock(&dentry->d_lock);
491 return 0;
494 * Check whether to do a partial shrink_dcache
495 * to get rid of unused child entries.
497 if (!list_empty(&dentry->d_subdirs)) {
498 spin_unlock(&dentry->d_lock);
499 shrink_dcache_parent(dentry);
500 spin_lock(&dentry->d_lock);
504 * Somebody else still using it?
506 * If it's a directory, we can't drop it
507 * for fear of somebody re-populating it
508 * with children (even though dropping it
509 * would make it unreachable from the root,
510 * we might still populate it if it was a
511 * working directory or similar).
513 if (dentry->d_count > 1) {
514 if (dentry->d_inode && S_ISDIR(dentry->d_inode->i_mode)) {
515 spin_unlock(&dentry->d_lock);
516 return -EBUSY;
520 __d_drop(dentry);
521 spin_unlock(&dentry->d_lock);
522 return 0;
524 EXPORT_SYMBOL(d_invalidate);
526 /* This must be called with d_lock held */
527 static inline void __dget_dlock(struct dentry *dentry)
529 dentry->d_count++;
532 static inline void __dget(struct dentry *dentry)
534 spin_lock(&dentry->d_lock);
535 __dget_dlock(dentry);
536 spin_unlock(&dentry->d_lock);
539 struct dentry *dget_parent(struct dentry *dentry)
541 struct dentry *ret;
543 repeat:
545 * Don't need rcu_dereference because we re-check it was correct under
546 * the lock.
548 rcu_read_lock();
549 ret = dentry->d_parent;
550 if (!ret) {
551 rcu_read_unlock();
552 goto out;
554 spin_lock(&ret->d_lock);
555 if (unlikely(ret != dentry->d_parent)) {
556 spin_unlock(&ret->d_lock);
557 rcu_read_unlock();
558 goto repeat;
560 rcu_read_unlock();
561 BUG_ON(!ret->d_count);
562 ret->d_count++;
563 spin_unlock(&ret->d_lock);
564 out:
565 return ret;
567 EXPORT_SYMBOL(dget_parent);
570 * d_find_alias - grab a hashed alias of inode
571 * @inode: inode in question
572 * @want_discon: flag, used by d_splice_alias, to request
573 * that only a DISCONNECTED alias be returned.
575 * If inode has a hashed alias, or is a directory and has any alias,
576 * acquire the reference to alias and return it. Otherwise return NULL.
577 * Notice that if inode is a directory there can be only one alias and
578 * it can be unhashed only if it has no children, or if it is the root
579 * of a filesystem.
581 * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
582 * any other hashed alias over that one unless @want_discon is set,
583 * in which case only return an IS_ROOT, DCACHE_DISCONNECTED alias.
585 static struct dentry *__d_find_alias(struct inode *inode, int want_discon)
587 struct dentry *alias, *discon_alias;
589 again:
590 discon_alias = NULL;
591 list_for_each_entry(alias, &inode->i_dentry, d_alias) {
592 spin_lock(&alias->d_lock);
593 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
594 if (IS_ROOT(alias) &&
595 (alias->d_flags & DCACHE_DISCONNECTED)) {
596 discon_alias = alias;
597 } else if (!want_discon) {
598 __dget_dlock(alias);
599 spin_unlock(&alias->d_lock);
600 return alias;
603 spin_unlock(&alias->d_lock);
605 if (discon_alias) {
606 alias = discon_alias;
607 spin_lock(&alias->d_lock);
608 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
609 if (IS_ROOT(alias) &&
610 (alias->d_flags & DCACHE_DISCONNECTED)) {
611 __dget_dlock(alias);
612 spin_unlock(&alias->d_lock);
613 return alias;
616 spin_unlock(&alias->d_lock);
617 goto again;
619 return NULL;
622 struct dentry *d_find_alias(struct inode *inode)
624 struct dentry *de = NULL;
626 if (!list_empty(&inode->i_dentry)) {
627 spin_lock(&inode->i_lock);
628 de = __d_find_alias(inode, 0);
629 spin_unlock(&inode->i_lock);
631 return de;
633 EXPORT_SYMBOL(d_find_alias);
636 * Try to kill dentries associated with this inode.
637 * WARNING: you must own a reference to inode.
639 void d_prune_aliases(struct inode *inode)
641 struct dentry *dentry;
642 restart:
643 spin_lock(&inode->i_lock);
644 list_for_each_entry(dentry, &inode->i_dentry, d_alias) {
645 spin_lock(&dentry->d_lock);
646 if (!dentry->d_count) {
647 __dget_dlock(dentry);
648 __d_drop(dentry);
649 spin_unlock(&dentry->d_lock);
650 spin_unlock(&inode->i_lock);
651 dput(dentry);
652 goto restart;
654 spin_unlock(&dentry->d_lock);
656 spin_unlock(&inode->i_lock);
658 EXPORT_SYMBOL(d_prune_aliases);
661 * Try to throw away a dentry - free the inode, dput the parent.
662 * Requires dentry->d_lock is held, and dentry->d_count == 0.
663 * Releases dentry->d_lock.
665 * This may fail if locks cannot be acquired no problem, just try again.
667 static void try_prune_one_dentry(struct dentry *dentry)
668 __releases(dentry->d_lock)
670 struct dentry *parent;
672 parent = dentry_kill(dentry, 0);
674 * If dentry_kill returns NULL, we have nothing more to do.
675 * if it returns the same dentry, trylocks failed. In either
676 * case, just loop again.
678 * Otherwise, we need to prune ancestors too. This is necessary
679 * to prevent quadratic behavior of shrink_dcache_parent(), but
680 * is also expected to be beneficial in reducing dentry cache
681 * fragmentation.
683 if (!parent)
684 return;
685 if (parent == dentry)
686 return;
688 /* Prune ancestors. */
689 dentry = parent;
690 while (dentry) {
691 spin_lock(&dentry->d_lock);
692 if (dentry->d_count > 1) {
693 dentry->d_count--;
694 spin_unlock(&dentry->d_lock);
695 return;
697 dentry = dentry_kill(dentry, 1);
701 static void shrink_dentry_list(struct list_head *list)
703 struct dentry *dentry;
705 rcu_read_lock();
706 for (;;) {
707 dentry = list_entry_rcu(list->prev, struct dentry, d_lru);
708 if (&dentry->d_lru == list)
709 break; /* empty */
710 spin_lock(&dentry->d_lock);
711 if (dentry != list_entry(list->prev, struct dentry, d_lru)) {
712 spin_unlock(&dentry->d_lock);
713 continue;
717 * We found an inuse dentry which was not removed from
718 * the LRU because of laziness during lookup. Do not free
719 * it - just keep it off the LRU list.
721 if (dentry->d_count) {
722 dentry_lru_del(dentry);
723 spin_unlock(&dentry->d_lock);
724 continue;
727 rcu_read_unlock();
729 try_prune_one_dentry(dentry);
731 rcu_read_lock();
733 rcu_read_unlock();
737 * __shrink_dcache_sb - shrink the dentry LRU on a given superblock
738 * @sb: superblock to shrink dentry LRU.
739 * @count: number of entries to prune
740 * @flags: flags to control the dentry processing
742 * If flags contains DCACHE_REFERENCED reference dentries will not be pruned.
744 static void __shrink_dcache_sb(struct super_block *sb, int *count, int flags)
746 /* called from prune_dcache() and shrink_dcache_parent() */
747 struct dentry *dentry;
748 LIST_HEAD(referenced);
749 LIST_HEAD(tmp);
750 int cnt = *count;
752 relock:
753 spin_lock(&dcache_lru_lock);
754 while (!list_empty(&sb->s_dentry_lru)) {
755 dentry = list_entry(sb->s_dentry_lru.prev,
756 struct dentry, d_lru);
757 BUG_ON(dentry->d_sb != sb);
759 if (!spin_trylock(&dentry->d_lock)) {
760 spin_unlock(&dcache_lru_lock);
761 cpu_relax();
762 goto relock;
766 * If we are honouring the DCACHE_REFERENCED flag and the
767 * dentry has this flag set, don't free it. Clear the flag
768 * and put it back on the LRU.
770 if (flags & DCACHE_REFERENCED &&
771 dentry->d_flags & DCACHE_REFERENCED) {
772 dentry->d_flags &= ~DCACHE_REFERENCED;
773 list_move(&dentry->d_lru, &referenced);
774 spin_unlock(&dentry->d_lock);
775 } else {
776 list_move_tail(&dentry->d_lru, &tmp);
777 spin_unlock(&dentry->d_lock);
778 if (!--cnt)
779 break;
781 cond_resched_lock(&dcache_lru_lock);
783 if (!list_empty(&referenced))
784 list_splice(&referenced, &sb->s_dentry_lru);
785 spin_unlock(&dcache_lru_lock);
787 shrink_dentry_list(&tmp);
789 *count = cnt;
793 * prune_dcache - shrink the dcache
794 * @count: number of entries to try to free
796 * Shrink the dcache. This is done when we need more memory, or simply when we
797 * need to unmount something (at which point we need to unuse all dentries).
799 * This function may fail to free any resources if all the dentries are in use.
801 static void prune_dcache(int count)
803 struct super_block *sb, *p = NULL;
804 int w_count;
805 int unused = dentry_stat.nr_unused;
806 int prune_ratio;
807 int pruned;
809 if (unused == 0 || count == 0)
810 return;
811 if (count >= unused)
812 prune_ratio = 1;
813 else
814 prune_ratio = unused / count;
815 spin_lock(&sb_lock);
816 list_for_each_entry(sb, &super_blocks, s_list) {
817 if (list_empty(&sb->s_instances))
818 continue;
819 if (sb->s_nr_dentry_unused == 0)
820 continue;
821 sb->s_count++;
822 /* Now, we reclaim unused dentrins with fairness.
823 * We reclaim them same percentage from each superblock.
824 * We calculate number of dentries to scan on this sb
825 * as follows, but the implementation is arranged to avoid
826 * overflows:
827 * number of dentries to scan on this sb =
828 * count * (number of dentries on this sb /
829 * number of dentries in the machine)
831 spin_unlock(&sb_lock);
832 if (prune_ratio != 1)
833 w_count = (sb->s_nr_dentry_unused / prune_ratio) + 1;
834 else
835 w_count = sb->s_nr_dentry_unused;
836 pruned = w_count;
838 * We need to be sure this filesystem isn't being unmounted,
839 * otherwise we could race with generic_shutdown_super(), and
840 * end up holding a reference to an inode while the filesystem
841 * is unmounted. So we try to get s_umount, and make sure
842 * s_root isn't NULL.
844 if (down_read_trylock(&sb->s_umount)) {
845 if ((sb->s_root != NULL) &&
846 (!list_empty(&sb->s_dentry_lru))) {
847 __shrink_dcache_sb(sb, &w_count,
848 DCACHE_REFERENCED);
849 pruned -= w_count;
851 up_read(&sb->s_umount);
853 spin_lock(&sb_lock);
854 if (p)
855 __put_super(p);
856 count -= pruned;
857 p = sb;
858 /* more work left to do? */
859 if (count <= 0)
860 break;
862 if (p)
863 __put_super(p);
864 spin_unlock(&sb_lock);
868 * shrink_dcache_sb - shrink dcache for a superblock
869 * @sb: superblock
871 * Shrink the dcache for the specified super block. This is used to free
872 * the dcache before unmounting a file system.
874 void shrink_dcache_sb(struct super_block *sb)
876 LIST_HEAD(tmp);
878 spin_lock(&dcache_lru_lock);
879 while (!list_empty(&sb->s_dentry_lru)) {
880 list_splice_init(&sb->s_dentry_lru, &tmp);
881 spin_unlock(&dcache_lru_lock);
882 shrink_dentry_list(&tmp);
883 spin_lock(&dcache_lru_lock);
885 spin_unlock(&dcache_lru_lock);
887 EXPORT_SYMBOL(shrink_dcache_sb);
890 * destroy a single subtree of dentries for unmount
891 * - see the comments on shrink_dcache_for_umount() for a description of the
892 * locking
894 static void shrink_dcache_for_umount_subtree(struct dentry *dentry)
896 struct dentry *parent;
897 unsigned detached = 0;
899 BUG_ON(!IS_ROOT(dentry));
901 /* detach this root from the system */
902 spin_lock(&dentry->d_lock);
903 dentry_lru_del(dentry);
904 __d_drop(dentry);
905 spin_unlock(&dentry->d_lock);
907 for (;;) {
908 /* descend to the first leaf in the current subtree */
909 while (!list_empty(&dentry->d_subdirs)) {
910 struct dentry *loop;
912 /* this is a branch with children - detach all of them
913 * from the system in one go */
914 spin_lock(&dentry->d_lock);
915 list_for_each_entry(loop, &dentry->d_subdirs,
916 d_u.d_child) {
917 spin_lock_nested(&loop->d_lock,
918 DENTRY_D_LOCK_NESTED);
919 dentry_lru_del(loop);
920 __d_drop(loop);
921 spin_unlock(&loop->d_lock);
923 spin_unlock(&dentry->d_lock);
925 /* move to the first child */
926 dentry = list_entry(dentry->d_subdirs.next,
927 struct dentry, d_u.d_child);
930 /* consume the dentries from this leaf up through its parents
931 * until we find one with children or run out altogether */
932 do {
933 struct inode *inode;
935 if (dentry->d_count != 0) {
936 printk(KERN_ERR
937 "BUG: Dentry %p{i=%lx,n=%s}"
938 " still in use (%d)"
939 " [unmount of %s %s]\n",
940 dentry,
941 dentry->d_inode ?
942 dentry->d_inode->i_ino : 0UL,
943 dentry->d_name.name,
944 dentry->d_count,
945 dentry->d_sb->s_type->name,
946 dentry->d_sb->s_id);
947 BUG();
950 if (IS_ROOT(dentry)) {
951 parent = NULL;
952 list_del(&dentry->d_u.d_child);
953 } else {
954 parent = dentry->d_parent;
955 spin_lock(&parent->d_lock);
956 parent->d_count--;
957 list_del(&dentry->d_u.d_child);
958 spin_unlock(&parent->d_lock);
961 detached++;
963 inode = dentry->d_inode;
964 if (inode) {
965 dentry->d_inode = NULL;
966 list_del_init(&dentry->d_alias);
967 if (dentry->d_op && dentry->d_op->d_iput)
968 dentry->d_op->d_iput(dentry, inode);
969 else
970 iput(inode);
973 d_free(dentry);
975 /* finished when we fall off the top of the tree,
976 * otherwise we ascend to the parent and move to the
977 * next sibling if there is one */
978 if (!parent)
979 return;
980 dentry = parent;
981 } while (list_empty(&dentry->d_subdirs));
983 dentry = list_entry(dentry->d_subdirs.next,
984 struct dentry, d_u.d_child);
989 * destroy the dentries attached to a superblock on unmounting
990 * - we don't need to use dentry->d_lock because:
991 * - the superblock is detached from all mountings and open files, so the
992 * dentry trees will not be rearranged by the VFS
993 * - s_umount is write-locked, so the memory pressure shrinker will ignore
994 * any dentries belonging to this superblock that it comes across
995 * - the filesystem itself is no longer permitted to rearrange the dentries
996 * in this superblock
998 void shrink_dcache_for_umount(struct super_block *sb)
1000 struct dentry *dentry;
1002 if (down_read_trylock(&sb->s_umount))
1003 BUG();
1005 dentry = sb->s_root;
1006 sb->s_root = NULL;
1007 spin_lock(&dentry->d_lock);
1008 dentry->d_count--;
1009 spin_unlock(&dentry->d_lock);
1010 shrink_dcache_for_umount_subtree(dentry);
1012 while (!hlist_bl_empty(&sb->s_anon)) {
1013 dentry = hlist_bl_entry(hlist_bl_first(&sb->s_anon), struct dentry, d_hash);
1014 shrink_dcache_for_umount_subtree(dentry);
1019 * This tries to ascend one level of parenthood, but
1020 * we can race with renaming, so we need to re-check
1021 * the parenthood after dropping the lock and check
1022 * that the sequence number still matches.
1024 static struct dentry *try_to_ascend(struct dentry *old, int locked, unsigned seq)
1026 struct dentry *new = old->d_parent;
1028 rcu_read_lock();
1029 spin_unlock(&old->d_lock);
1030 spin_lock(&new->d_lock);
1033 * might go back up the wrong parent if we have had a rename
1034 * or deletion
1036 if (new != old->d_parent ||
1037 (old->d_flags & DCACHE_DISCONNECTED) ||
1038 (!locked && read_seqretry(&rename_lock, seq))) {
1039 spin_unlock(&new->d_lock);
1040 new = NULL;
1042 rcu_read_unlock();
1043 return new;
1048 * Search for at least 1 mount point in the dentry's subdirs.
1049 * We descend to the next level whenever the d_subdirs
1050 * list is non-empty and continue searching.
1054 * have_submounts - check for mounts over a dentry
1055 * @parent: dentry to check.
1057 * Return true if the parent or its subdirectories contain
1058 * a mount point
1060 int have_submounts(struct dentry *parent)
1062 struct dentry *this_parent;
1063 struct list_head *next;
1064 unsigned seq;
1065 int locked = 0;
1067 seq = read_seqbegin(&rename_lock);
1068 again:
1069 this_parent = parent;
1071 if (d_mountpoint(parent))
1072 goto positive;
1073 spin_lock(&this_parent->d_lock);
1074 repeat:
1075 next = this_parent->d_subdirs.next;
1076 resume:
1077 while (next != &this_parent->d_subdirs) {
1078 struct list_head *tmp = next;
1079 struct dentry *dentry = list_entry(tmp, struct dentry, d_u.d_child);
1080 next = tmp->next;
1082 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1083 /* Have we found a mount point ? */
1084 if (d_mountpoint(dentry)) {
1085 spin_unlock(&dentry->d_lock);
1086 spin_unlock(&this_parent->d_lock);
1087 goto positive;
1089 if (!list_empty(&dentry->d_subdirs)) {
1090 spin_unlock(&this_parent->d_lock);
1091 spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1092 this_parent = dentry;
1093 spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1094 goto repeat;
1096 spin_unlock(&dentry->d_lock);
1099 * All done at this level ... ascend and resume the search.
1101 if (this_parent != parent) {
1102 struct dentry *child = this_parent;
1103 this_parent = try_to_ascend(this_parent, locked, seq);
1104 if (!this_parent)
1105 goto rename_retry;
1106 next = child->d_u.d_child.next;
1107 goto resume;
1109 spin_unlock(&this_parent->d_lock);
1110 if (!locked && read_seqretry(&rename_lock, seq))
1111 goto rename_retry;
1112 if (locked)
1113 write_sequnlock(&rename_lock);
1114 return 0; /* No mount points found in tree */
1115 positive:
1116 if (!locked && read_seqretry(&rename_lock, seq))
1117 goto rename_retry;
1118 if (locked)
1119 write_sequnlock(&rename_lock);
1120 return 1;
1122 rename_retry:
1123 locked = 1;
1124 write_seqlock(&rename_lock);
1125 goto again;
1127 EXPORT_SYMBOL(have_submounts);
1130 * Search the dentry child list for the specified parent,
1131 * and move any unused dentries to the end of the unused
1132 * list for prune_dcache(). We descend to the next level
1133 * whenever the d_subdirs list is non-empty and continue
1134 * searching.
1136 * It returns zero iff there are no unused children,
1137 * otherwise it returns the number of children moved to
1138 * the end of the unused list. This may not be the total
1139 * number of unused children, because select_parent can
1140 * drop the lock and return early due to latency
1141 * constraints.
1143 static int select_parent(struct dentry * parent)
1145 struct dentry *this_parent;
1146 struct list_head *next;
1147 unsigned seq;
1148 int found = 0;
1149 int locked = 0;
1151 seq = read_seqbegin(&rename_lock);
1152 again:
1153 this_parent = parent;
1154 spin_lock(&this_parent->d_lock);
1155 repeat:
1156 next = this_parent->d_subdirs.next;
1157 resume:
1158 while (next != &this_parent->d_subdirs) {
1159 struct list_head *tmp = next;
1160 struct dentry *dentry = list_entry(tmp, struct dentry, d_u.d_child);
1161 next = tmp->next;
1163 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1166 * move only zero ref count dentries to the end
1167 * of the unused list for prune_dcache
1169 if (!dentry->d_count) {
1170 dentry_lru_move_tail(dentry);
1171 found++;
1172 } else {
1173 dentry_lru_del(dentry);
1177 * We can return to the caller if we have found some (this
1178 * ensures forward progress). We'll be coming back to find
1179 * the rest.
1181 if (found && need_resched()) {
1182 spin_unlock(&dentry->d_lock);
1183 goto out;
1187 * Descend a level if the d_subdirs list is non-empty.
1189 if (!list_empty(&dentry->d_subdirs)) {
1190 spin_unlock(&this_parent->d_lock);
1191 spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1192 this_parent = dentry;
1193 spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1194 goto repeat;
1197 spin_unlock(&dentry->d_lock);
1200 * All done at this level ... ascend and resume the search.
1202 if (this_parent != parent) {
1203 struct dentry *child = this_parent;
1204 this_parent = try_to_ascend(this_parent, locked, seq);
1205 if (!this_parent)
1206 goto rename_retry;
1207 next = child->d_u.d_child.next;
1208 goto resume;
1210 out:
1211 spin_unlock(&this_parent->d_lock);
1212 if (!locked && read_seqretry(&rename_lock, seq))
1213 goto rename_retry;
1214 if (locked)
1215 write_sequnlock(&rename_lock);
1216 return found;
1218 rename_retry:
1219 if (found)
1220 return found;
1221 locked = 1;
1222 write_seqlock(&rename_lock);
1223 goto again;
1227 * shrink_dcache_parent - prune dcache
1228 * @parent: parent of entries to prune
1230 * Prune the dcache to remove unused children of the parent dentry.
1233 void shrink_dcache_parent(struct dentry * parent)
1235 struct super_block *sb = parent->d_sb;
1236 int found;
1238 while ((found = select_parent(parent)) != 0)
1239 __shrink_dcache_sb(sb, &found, 0);
1241 EXPORT_SYMBOL(shrink_dcache_parent);
1244 * Scan `nr' dentries and return the number which remain.
1246 * We need to avoid reentering the filesystem if the caller is performing a
1247 * GFP_NOFS allocation attempt. One example deadlock is:
1249 * ext2_new_block->getblk->GFP->shrink_dcache_memory->prune_dcache->
1250 * prune_one_dentry->dput->dentry_iput->iput->inode->i_sb->s_op->put_inode->
1251 * ext2_discard_prealloc->ext2_free_blocks->lock_super->DEADLOCK.
1253 * In this case we return -1 to tell the caller that we baled.
1255 static int shrink_dcache_memory(struct shrinker *shrink, int nr, gfp_t gfp_mask)
1257 if (nr) {
1258 if (!(gfp_mask & __GFP_FS))
1259 return -1;
1260 prune_dcache(nr);
1263 return (dentry_stat.nr_unused / 100) * sysctl_vfs_cache_pressure;
1266 static struct shrinker dcache_shrinker = {
1267 .shrink = shrink_dcache_memory,
1268 .seeks = DEFAULT_SEEKS,
1272 * d_alloc - allocate a dcache entry
1273 * @parent: parent of entry to allocate
1274 * @name: qstr of the name
1276 * Allocates a dentry. It returns %NULL if there is insufficient memory
1277 * available. On a success the dentry is returned. The name passed in is
1278 * copied and the copy passed in may be reused after this call.
1281 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1283 struct dentry *dentry;
1284 char *dname;
1286 dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1287 if (!dentry)
1288 return NULL;
1290 if (name->len > DNAME_INLINE_LEN-1) {
1291 dname = kmalloc(name->len + 1, GFP_KERNEL);
1292 if (!dname) {
1293 kmem_cache_free(dentry_cache, dentry);
1294 return NULL;
1296 } else {
1297 dname = dentry->d_iname;
1299 dentry->d_name.name = dname;
1301 dentry->d_name.len = name->len;
1302 dentry->d_name.hash = name->hash;
1303 memcpy(dname, name->name, name->len);
1304 dname[name->len] = 0;
1306 dentry->d_count = 1;
1307 dentry->d_flags = DCACHE_UNHASHED;
1308 spin_lock_init(&dentry->d_lock);
1309 seqcount_init(&dentry->d_seq);
1310 dentry->d_inode = NULL;
1311 dentry->d_parent = NULL;
1312 dentry->d_sb = NULL;
1313 dentry->d_op = NULL;
1314 dentry->d_fsdata = NULL;
1315 INIT_HLIST_BL_NODE(&dentry->d_hash);
1316 INIT_LIST_HEAD(&dentry->d_lru);
1317 INIT_LIST_HEAD(&dentry->d_subdirs);
1318 INIT_LIST_HEAD(&dentry->d_alias);
1319 INIT_LIST_HEAD(&dentry->d_u.d_child);
1321 if (parent) {
1322 spin_lock(&parent->d_lock);
1324 * don't need child lock because it is not subject
1325 * to concurrency here
1327 __dget_dlock(parent);
1328 dentry->d_parent = parent;
1329 dentry->d_sb = parent->d_sb;
1330 d_set_d_op(dentry, dentry->d_sb->s_d_op);
1331 list_add(&dentry->d_u.d_child, &parent->d_subdirs);
1332 spin_unlock(&parent->d_lock);
1335 this_cpu_inc(nr_dentry);
1337 return dentry;
1339 EXPORT_SYMBOL(d_alloc);
1341 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1343 struct dentry *dentry = d_alloc(NULL, name);
1344 if (dentry) {
1345 dentry->d_sb = sb;
1346 d_set_d_op(dentry, dentry->d_sb->s_d_op);
1347 dentry->d_parent = dentry;
1348 dentry->d_flags |= DCACHE_DISCONNECTED;
1350 return dentry;
1352 EXPORT_SYMBOL(d_alloc_pseudo);
1354 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1356 struct qstr q;
1358 q.name = name;
1359 q.len = strlen(name);
1360 q.hash = full_name_hash(q.name, q.len);
1361 return d_alloc(parent, &q);
1363 EXPORT_SYMBOL(d_alloc_name);
1365 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1367 WARN_ON_ONCE(dentry->d_op);
1368 WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH |
1369 DCACHE_OP_COMPARE |
1370 DCACHE_OP_REVALIDATE |
1371 DCACHE_OP_DELETE ));
1372 dentry->d_op = op;
1373 if (!op)
1374 return;
1375 if (op->d_hash)
1376 dentry->d_flags |= DCACHE_OP_HASH;
1377 if (op->d_compare)
1378 dentry->d_flags |= DCACHE_OP_COMPARE;
1379 if (op->d_revalidate)
1380 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1381 if (op->d_delete)
1382 dentry->d_flags |= DCACHE_OP_DELETE;
1385 EXPORT_SYMBOL(d_set_d_op);
1387 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1389 spin_lock(&dentry->d_lock);
1390 if (inode) {
1391 if (unlikely(IS_AUTOMOUNT(inode)))
1392 dentry->d_flags |= DCACHE_NEED_AUTOMOUNT;
1393 list_add(&dentry->d_alias, &inode->i_dentry);
1395 dentry->d_inode = inode;
1396 dentry_rcuwalk_barrier(dentry);
1397 spin_unlock(&dentry->d_lock);
1398 fsnotify_d_instantiate(dentry, inode);
1402 * d_instantiate - fill in inode information for a dentry
1403 * @entry: dentry to complete
1404 * @inode: inode to attach to this dentry
1406 * Fill in inode information in the entry.
1408 * This turns negative dentries into productive full members
1409 * of society.
1411 * NOTE! This assumes that the inode count has been incremented
1412 * (or otherwise set) by the caller to indicate that it is now
1413 * in use by the dcache.
1416 void d_instantiate(struct dentry *entry, struct inode * inode)
1418 BUG_ON(!list_empty(&entry->d_alias));
1419 if (inode)
1420 spin_lock(&inode->i_lock);
1421 __d_instantiate(entry, inode);
1422 if (inode)
1423 spin_unlock(&inode->i_lock);
1424 security_d_instantiate(entry, inode);
1426 EXPORT_SYMBOL(d_instantiate);
1429 * d_instantiate_unique - instantiate a non-aliased dentry
1430 * @entry: dentry to instantiate
1431 * @inode: inode to attach to this dentry
1433 * Fill in inode information in the entry. On success, it returns NULL.
1434 * If an unhashed alias of "entry" already exists, then we return the
1435 * aliased dentry instead and drop one reference to inode.
1437 * Note that in order to avoid conflicts with rename() etc, the caller
1438 * had better be holding the parent directory semaphore.
1440 * This also assumes that the inode count has been incremented
1441 * (or otherwise set) by the caller to indicate that it is now
1442 * in use by the dcache.
1444 static struct dentry *__d_instantiate_unique(struct dentry *entry,
1445 struct inode *inode)
1447 struct dentry *alias;
1448 int len = entry->d_name.len;
1449 const char *name = entry->d_name.name;
1450 unsigned int hash = entry->d_name.hash;
1452 if (!inode) {
1453 __d_instantiate(entry, NULL);
1454 return NULL;
1457 list_for_each_entry(alias, &inode->i_dentry, d_alias) {
1458 struct qstr *qstr = &alias->d_name;
1461 * Don't need alias->d_lock here, because aliases with
1462 * d_parent == entry->d_parent are not subject to name or
1463 * parent changes, because the parent inode i_mutex is held.
1465 if (qstr->hash != hash)
1466 continue;
1467 if (alias->d_parent != entry->d_parent)
1468 continue;
1469 if (dentry_cmp(qstr->name, qstr->len, name, len))
1470 continue;
1471 __dget(alias);
1472 return alias;
1475 __d_instantiate(entry, inode);
1476 return NULL;
1479 struct dentry *d_instantiate_unique(struct dentry *entry, struct inode *inode)
1481 struct dentry *result;
1483 BUG_ON(!list_empty(&entry->d_alias));
1485 if (inode)
1486 spin_lock(&inode->i_lock);
1487 result = __d_instantiate_unique(entry, inode);
1488 if (inode)
1489 spin_unlock(&inode->i_lock);
1491 if (!result) {
1492 security_d_instantiate(entry, inode);
1493 return NULL;
1496 BUG_ON(!d_unhashed(result));
1497 iput(inode);
1498 return result;
1501 EXPORT_SYMBOL(d_instantiate_unique);
1504 * d_alloc_root - allocate root dentry
1505 * @root_inode: inode to allocate the root for
1507 * Allocate a root ("/") dentry for the inode given. The inode is
1508 * instantiated and returned. %NULL is returned if there is insufficient
1509 * memory or the inode passed is %NULL.
1512 struct dentry * d_alloc_root(struct inode * root_inode)
1514 struct dentry *res = NULL;
1516 if (root_inode) {
1517 static const struct qstr name = { .name = "/", .len = 1 };
1519 res = d_alloc(NULL, &name);
1520 if (res) {
1521 res->d_sb = root_inode->i_sb;
1522 d_set_d_op(res, res->d_sb->s_d_op);
1523 res->d_parent = res;
1524 d_instantiate(res, root_inode);
1527 return res;
1529 EXPORT_SYMBOL(d_alloc_root);
1531 static struct dentry * __d_find_any_alias(struct inode *inode)
1533 struct dentry *alias;
1535 if (list_empty(&inode->i_dentry))
1536 return NULL;
1537 alias = list_first_entry(&inode->i_dentry, struct dentry, d_alias);
1538 __dget(alias);
1539 return alias;
1542 static struct dentry * d_find_any_alias(struct inode *inode)
1544 struct dentry *de;
1546 spin_lock(&inode->i_lock);
1547 de = __d_find_any_alias(inode);
1548 spin_unlock(&inode->i_lock);
1549 return de;
1554 * d_obtain_alias - find or allocate a dentry for a given inode
1555 * @inode: inode to allocate the dentry for
1557 * Obtain a dentry for an inode resulting from NFS filehandle conversion or
1558 * similar open by handle operations. The returned dentry may be anonymous,
1559 * or may have a full name (if the inode was already in the cache).
1561 * When called on a directory inode, we must ensure that the inode only ever
1562 * has one dentry. If a dentry is found, that is returned instead of
1563 * allocating a new one.
1565 * On successful return, the reference to the inode has been transferred
1566 * to the dentry. In case of an error the reference on the inode is released.
1567 * To make it easier to use in export operations a %NULL or IS_ERR inode may
1568 * be passed in and will be the error will be propagate to the return value,
1569 * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
1571 struct dentry *d_obtain_alias(struct inode *inode)
1573 static const struct qstr anonstring = { .name = "" };
1574 struct dentry *tmp;
1575 struct dentry *res;
1577 if (!inode)
1578 return ERR_PTR(-ESTALE);
1579 if (IS_ERR(inode))
1580 return ERR_CAST(inode);
1582 res = d_find_any_alias(inode);
1583 if (res)
1584 goto out_iput;
1586 tmp = d_alloc(NULL, &anonstring);
1587 if (!tmp) {
1588 res = ERR_PTR(-ENOMEM);
1589 goto out_iput;
1591 tmp->d_parent = tmp; /* make sure dput doesn't croak */
1594 spin_lock(&inode->i_lock);
1595 res = __d_find_any_alias(inode);
1596 if (res) {
1597 spin_unlock(&inode->i_lock);
1598 dput(tmp);
1599 goto out_iput;
1602 /* attach a disconnected dentry */
1603 spin_lock(&tmp->d_lock);
1604 tmp->d_sb = inode->i_sb;
1605 d_set_d_op(tmp, tmp->d_sb->s_d_op);
1606 tmp->d_inode = inode;
1607 tmp->d_flags |= DCACHE_DISCONNECTED;
1608 list_add(&tmp->d_alias, &inode->i_dentry);
1609 bit_spin_lock(0, (unsigned long *)&tmp->d_sb->s_anon.first);
1610 tmp->d_flags &= ~DCACHE_UNHASHED;
1611 hlist_bl_add_head(&tmp->d_hash, &tmp->d_sb->s_anon);
1612 __bit_spin_unlock(0, (unsigned long *)&tmp->d_sb->s_anon.first);
1613 spin_unlock(&tmp->d_lock);
1614 spin_unlock(&inode->i_lock);
1615 security_d_instantiate(tmp, inode);
1617 return tmp;
1619 out_iput:
1620 if (res && !IS_ERR(res))
1621 security_d_instantiate(res, inode);
1622 iput(inode);
1623 return res;
1625 EXPORT_SYMBOL(d_obtain_alias);
1628 * d_splice_alias - splice a disconnected dentry into the tree if one exists
1629 * @inode: the inode which may have a disconnected dentry
1630 * @dentry: a negative dentry which we want to point to the inode.
1632 * If inode is a directory and has a 'disconnected' dentry (i.e. IS_ROOT and
1633 * DCACHE_DISCONNECTED), then d_move that in place of the given dentry
1634 * and return it, else simply d_add the inode to the dentry and return NULL.
1636 * This is needed in the lookup routine of any filesystem that is exportable
1637 * (via knfsd) so that we can build dcache paths to directories effectively.
1639 * If a dentry was found and moved, then it is returned. Otherwise NULL
1640 * is returned. This matches the expected return value of ->lookup.
1643 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
1645 struct dentry *new = NULL;
1647 if (inode && S_ISDIR(inode->i_mode)) {
1648 spin_lock(&inode->i_lock);
1649 new = __d_find_alias(inode, 1);
1650 if (new) {
1651 BUG_ON(!(new->d_flags & DCACHE_DISCONNECTED));
1652 spin_unlock(&inode->i_lock);
1653 security_d_instantiate(new, inode);
1654 d_move(new, dentry);
1655 iput(inode);
1656 } else {
1657 /* already taking inode->i_lock, so d_add() by hand */
1658 __d_instantiate(dentry, inode);
1659 spin_unlock(&inode->i_lock);
1660 security_d_instantiate(dentry, inode);
1661 d_rehash(dentry);
1663 } else
1664 d_add(dentry, inode);
1665 return new;
1667 EXPORT_SYMBOL(d_splice_alias);
1670 * d_add_ci - lookup or allocate new dentry with case-exact name
1671 * @inode: the inode case-insensitive lookup has found
1672 * @dentry: the negative dentry that was passed to the parent's lookup func
1673 * @name: the case-exact name to be associated with the returned dentry
1675 * This is to avoid filling the dcache with case-insensitive names to the
1676 * same inode, only the actual correct case is stored in the dcache for
1677 * case-insensitive filesystems.
1679 * For a case-insensitive lookup match and if the the case-exact dentry
1680 * already exists in in the dcache, use it and return it.
1682 * If no entry exists with the exact case name, allocate new dentry with
1683 * the exact case, and return the spliced entry.
1685 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
1686 struct qstr *name)
1688 int error;
1689 struct dentry *found;
1690 struct dentry *new;
1693 * First check if a dentry matching the name already exists,
1694 * if not go ahead and create it now.
1696 found = d_hash_and_lookup(dentry->d_parent, name);
1697 if (!found) {
1698 new = d_alloc(dentry->d_parent, name);
1699 if (!new) {
1700 error = -ENOMEM;
1701 goto err_out;
1704 found = d_splice_alias(inode, new);
1705 if (found) {
1706 dput(new);
1707 return found;
1709 return new;
1713 * If a matching dentry exists, and it's not negative use it.
1715 * Decrement the reference count to balance the iget() done
1716 * earlier on.
1718 if (found->d_inode) {
1719 if (unlikely(found->d_inode != inode)) {
1720 /* This can't happen because bad inodes are unhashed. */
1721 BUG_ON(!is_bad_inode(inode));
1722 BUG_ON(!is_bad_inode(found->d_inode));
1724 iput(inode);
1725 return found;
1729 * Negative dentry: instantiate it unless the inode is a directory and
1730 * already has a dentry.
1732 spin_lock(&inode->i_lock);
1733 if (!S_ISDIR(inode->i_mode) || list_empty(&inode->i_dentry)) {
1734 __d_instantiate(found, inode);
1735 spin_unlock(&inode->i_lock);
1736 security_d_instantiate(found, inode);
1737 return found;
1741 * In case a directory already has a (disconnected) entry grab a
1742 * reference to it, move it in place and use it.
1744 new = list_entry(inode->i_dentry.next, struct dentry, d_alias);
1745 __dget(new);
1746 spin_unlock(&inode->i_lock);
1747 security_d_instantiate(found, inode);
1748 d_move(new, found);
1749 iput(inode);
1750 dput(found);
1751 return new;
1753 err_out:
1754 iput(inode);
1755 return ERR_PTR(error);
1757 EXPORT_SYMBOL(d_add_ci);
1760 * __d_lookup_rcu - search for a dentry (racy, store-free)
1761 * @parent: parent dentry
1762 * @name: qstr of name we wish to find
1763 * @seq: returns d_seq value at the point where the dentry was found
1764 * @inode: returns dentry->d_inode when the inode was found valid.
1765 * Returns: dentry, or NULL
1767 * __d_lookup_rcu is the dcache lookup function for rcu-walk name
1768 * resolution (store-free path walking) design described in
1769 * Documentation/filesystems/path-lookup.txt.
1771 * This is not to be used outside core vfs.
1773 * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
1774 * held, and rcu_read_lock held. The returned dentry must not be stored into
1775 * without taking d_lock and checking d_seq sequence count against @seq
1776 * returned here.
1778 * A refcount may be taken on the found dentry with the __d_rcu_to_refcount
1779 * function.
1781 * Alternatively, __d_lookup_rcu may be called again to look up the child of
1782 * the returned dentry, so long as its parent's seqlock is checked after the
1783 * child is looked up. Thus, an interlocking stepping of sequence lock checks
1784 * is formed, giving integrity down the path walk.
1786 struct dentry *__d_lookup_rcu(struct dentry *parent, struct qstr *name,
1787 unsigned *seq, struct inode **inode)
1789 unsigned int len = name->len;
1790 unsigned int hash = name->hash;
1791 const unsigned char *str = name->name;
1792 struct dcache_hash_bucket *b = d_hash(parent, hash);
1793 struct hlist_bl_node *node;
1794 struct dentry *dentry;
1797 * Note: There is significant duplication with __d_lookup_rcu which is
1798 * required to prevent single threaded performance regressions
1799 * especially on architectures where smp_rmb (in seqcounts) are costly.
1800 * Keep the two functions in sync.
1804 * The hash list is protected using RCU.
1806 * Carefully use d_seq when comparing a candidate dentry, to avoid
1807 * races with d_move().
1809 * It is possible that concurrent renames can mess up our list
1810 * walk here and result in missing our dentry, resulting in the
1811 * false-negative result. d_lookup() protects against concurrent
1812 * renames using rename_lock seqlock.
1814 * See Documentation/filesystems/path-lookup.txt for more details.
1816 hlist_bl_for_each_entry_rcu(dentry, node, &b->head, d_hash) {
1817 struct inode *i;
1818 const char *tname;
1819 int tlen;
1821 if (dentry->d_name.hash != hash)
1822 continue;
1824 seqretry:
1825 *seq = read_seqcount_begin(&dentry->d_seq);
1826 if (dentry->d_parent != parent)
1827 continue;
1828 if (d_unhashed(dentry))
1829 continue;
1830 tlen = dentry->d_name.len;
1831 tname = dentry->d_name.name;
1832 i = dentry->d_inode;
1833 prefetch(tname);
1834 if (i)
1835 prefetch(i);
1837 * This seqcount check is required to ensure name and
1838 * len are loaded atomically, so as not to walk off the
1839 * edge of memory when walking. If we could load this
1840 * atomically some other way, we could drop this check.
1842 if (read_seqcount_retry(&dentry->d_seq, *seq))
1843 goto seqretry;
1844 if (parent->d_flags & DCACHE_OP_COMPARE) {
1845 if (parent->d_op->d_compare(parent, *inode,
1846 dentry, i,
1847 tlen, tname, name))
1848 continue;
1849 } else {
1850 if (dentry_cmp(tname, tlen, str, len))
1851 continue;
1854 * No extra seqcount check is required after the name
1855 * compare. The caller must perform a seqcount check in
1856 * order to do anything useful with the returned dentry
1857 * anyway.
1859 *inode = i;
1860 return dentry;
1862 return NULL;
1866 * d_lookup - search for a dentry
1867 * @parent: parent dentry
1868 * @name: qstr of name we wish to find
1869 * Returns: dentry, or NULL
1871 * d_lookup searches the children of the parent dentry for the name in
1872 * question. If the dentry is found its reference count is incremented and the
1873 * dentry is returned. The caller must use dput to free the entry when it has
1874 * finished using it. %NULL is returned if the dentry does not exist.
1876 struct dentry *d_lookup(struct dentry *parent, struct qstr *name)
1878 struct dentry *dentry;
1879 unsigned seq;
1881 do {
1882 seq = read_seqbegin(&rename_lock);
1883 dentry = __d_lookup(parent, name);
1884 if (dentry)
1885 break;
1886 } while (read_seqretry(&rename_lock, seq));
1887 return dentry;
1889 EXPORT_SYMBOL(d_lookup);
1892 * __d_lookup - search for a dentry (racy)
1893 * @parent: parent dentry
1894 * @name: qstr of name we wish to find
1895 * Returns: dentry, or NULL
1897 * __d_lookup is like d_lookup, however it may (rarely) return a
1898 * false-negative result due to unrelated rename activity.
1900 * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
1901 * however it must be used carefully, eg. with a following d_lookup in
1902 * the case of failure.
1904 * __d_lookup callers must be commented.
1906 struct dentry *__d_lookup(struct dentry *parent, struct qstr *name)
1908 unsigned int len = name->len;
1909 unsigned int hash = name->hash;
1910 const unsigned char *str = name->name;
1911 struct dcache_hash_bucket *b = d_hash(parent, hash);
1912 struct hlist_bl_node *node;
1913 struct dentry *found = NULL;
1914 struct dentry *dentry;
1917 * Note: There is significant duplication with __d_lookup_rcu which is
1918 * required to prevent single threaded performance regressions
1919 * especially on architectures where smp_rmb (in seqcounts) are costly.
1920 * Keep the two functions in sync.
1924 * The hash list is protected using RCU.
1926 * Take d_lock when comparing a candidate dentry, to avoid races
1927 * with d_move().
1929 * It is possible that concurrent renames can mess up our list
1930 * walk here and result in missing our dentry, resulting in the
1931 * false-negative result. d_lookup() protects against concurrent
1932 * renames using rename_lock seqlock.
1934 * See Documentation/filesystems/path-lookup.txt for more details.
1936 rcu_read_lock();
1938 hlist_bl_for_each_entry_rcu(dentry, node, &b->head, d_hash) {
1939 const char *tname;
1940 int tlen;
1942 if (dentry->d_name.hash != hash)
1943 continue;
1945 spin_lock(&dentry->d_lock);
1946 if (dentry->d_parent != parent)
1947 goto next;
1948 if (d_unhashed(dentry))
1949 goto next;
1952 * It is safe to compare names since d_move() cannot
1953 * change the qstr (protected by d_lock).
1955 tlen = dentry->d_name.len;
1956 tname = dentry->d_name.name;
1957 if (parent->d_flags & DCACHE_OP_COMPARE) {
1958 if (parent->d_op->d_compare(parent, parent->d_inode,
1959 dentry, dentry->d_inode,
1960 tlen, tname, name))
1961 goto next;
1962 } else {
1963 if (dentry_cmp(tname, tlen, str, len))
1964 goto next;
1967 dentry->d_count++;
1968 found = dentry;
1969 spin_unlock(&dentry->d_lock);
1970 break;
1971 next:
1972 spin_unlock(&dentry->d_lock);
1974 rcu_read_unlock();
1976 return found;
1980 * d_hash_and_lookup - hash the qstr then search for a dentry
1981 * @dir: Directory to search in
1982 * @name: qstr of name we wish to find
1984 * On hash failure or on lookup failure NULL is returned.
1986 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
1988 struct dentry *dentry = NULL;
1991 * Check for a fs-specific hash function. Note that we must
1992 * calculate the standard hash first, as the d_op->d_hash()
1993 * routine may choose to leave the hash value unchanged.
1995 name->hash = full_name_hash(name->name, name->len);
1996 if (dir->d_flags & DCACHE_OP_HASH) {
1997 if (dir->d_op->d_hash(dir, dir->d_inode, name) < 0)
1998 goto out;
2000 dentry = d_lookup(dir, name);
2001 out:
2002 return dentry;
2006 * d_validate - verify dentry provided from insecure source (deprecated)
2007 * @dentry: The dentry alleged to be valid child of @dparent
2008 * @dparent: The parent dentry (known to be valid)
2010 * An insecure source has sent us a dentry, here we verify it and dget() it.
2011 * This is used by ncpfs in its readdir implementation.
2012 * Zero is returned in the dentry is invalid.
2014 * This function is slow for big directories, and deprecated, do not use it.
2016 int d_validate(struct dentry *dentry, struct dentry *dparent)
2018 struct dentry *child;
2020 spin_lock(&dparent->d_lock);
2021 list_for_each_entry(child, &dparent->d_subdirs, d_u.d_child) {
2022 if (dentry == child) {
2023 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
2024 __dget_dlock(dentry);
2025 spin_unlock(&dentry->d_lock);
2026 spin_unlock(&dparent->d_lock);
2027 return 1;
2030 spin_unlock(&dparent->d_lock);
2032 return 0;
2034 EXPORT_SYMBOL(d_validate);
2037 * When a file is deleted, we have two options:
2038 * - turn this dentry into a negative dentry
2039 * - unhash this dentry and free it.
2041 * Usually, we want to just turn this into
2042 * a negative dentry, but if anybody else is
2043 * currently using the dentry or the inode
2044 * we can't do that and we fall back on removing
2045 * it from the hash queues and waiting for
2046 * it to be deleted later when it has no users
2050 * d_delete - delete a dentry
2051 * @dentry: The dentry to delete
2053 * Turn the dentry into a negative dentry if possible, otherwise
2054 * remove it from the hash queues so it can be deleted later
2057 void d_delete(struct dentry * dentry)
2059 struct inode *inode;
2060 int isdir = 0;
2062 * Are we the only user?
2064 again:
2065 spin_lock(&dentry->d_lock);
2066 inode = dentry->d_inode;
2067 isdir = S_ISDIR(inode->i_mode);
2068 if (dentry->d_count == 1) {
2069 if (inode && !spin_trylock(&inode->i_lock)) {
2070 spin_unlock(&dentry->d_lock);
2071 cpu_relax();
2072 goto again;
2074 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2075 dentry_unlink_inode(dentry);
2076 fsnotify_nameremove(dentry, isdir);
2077 return;
2080 if (!d_unhashed(dentry))
2081 __d_drop(dentry);
2083 spin_unlock(&dentry->d_lock);
2085 fsnotify_nameremove(dentry, isdir);
2087 EXPORT_SYMBOL(d_delete);
2089 static void __d_rehash(struct dentry * entry, struct dcache_hash_bucket *b)
2091 BUG_ON(!d_unhashed(entry));
2092 spin_lock_bucket(b);
2093 entry->d_flags &= ~DCACHE_UNHASHED;
2094 hlist_bl_add_head_rcu(&entry->d_hash, &b->head);
2095 spin_unlock_bucket(b);
2098 static void _d_rehash(struct dentry * entry)
2100 __d_rehash(entry, d_hash(entry->d_parent, entry->d_name.hash));
2104 * d_rehash - add an entry back to the hash
2105 * @entry: dentry to add to the hash
2107 * Adds a dentry to the hash according to its name.
2110 void d_rehash(struct dentry * entry)
2112 spin_lock(&entry->d_lock);
2113 _d_rehash(entry);
2114 spin_unlock(&entry->d_lock);
2116 EXPORT_SYMBOL(d_rehash);
2119 * dentry_update_name_case - update case insensitive dentry with a new name
2120 * @dentry: dentry to be updated
2121 * @name: new name
2123 * Update a case insensitive dentry with new case of name.
2125 * dentry must have been returned by d_lookup with name @name. Old and new
2126 * name lengths must match (ie. no d_compare which allows mismatched name
2127 * lengths).
2129 * Parent inode i_mutex must be held over d_lookup and into this call (to
2130 * keep renames and concurrent inserts, and readdir(2) away).
2132 void dentry_update_name_case(struct dentry *dentry, struct qstr *name)
2134 BUG_ON(!mutex_is_locked(&dentry->d_inode->i_mutex));
2135 BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */
2137 spin_lock(&dentry->d_lock);
2138 write_seqcount_begin(&dentry->d_seq);
2139 memcpy((unsigned char *)dentry->d_name.name, name->name, name->len);
2140 write_seqcount_end(&dentry->d_seq);
2141 spin_unlock(&dentry->d_lock);
2143 EXPORT_SYMBOL(dentry_update_name_case);
2145 static void switch_names(struct dentry *dentry, struct dentry *target)
2147 if (dname_external(target)) {
2148 if (dname_external(dentry)) {
2150 * Both external: swap the pointers
2152 swap(target->d_name.name, dentry->d_name.name);
2153 } else {
2155 * dentry:internal, target:external. Steal target's
2156 * storage and make target internal.
2158 memcpy(target->d_iname, dentry->d_name.name,
2159 dentry->d_name.len + 1);
2160 dentry->d_name.name = target->d_name.name;
2161 target->d_name.name = target->d_iname;
2163 } else {
2164 if (dname_external(dentry)) {
2166 * dentry:external, target:internal. Give dentry's
2167 * storage to target and make dentry internal
2169 memcpy(dentry->d_iname, target->d_name.name,
2170 target->d_name.len + 1);
2171 target->d_name.name = dentry->d_name.name;
2172 dentry->d_name.name = dentry->d_iname;
2173 } else {
2175 * Both are internal. Just copy target to dentry
2177 memcpy(dentry->d_iname, target->d_name.name,
2178 target->d_name.len + 1);
2179 dentry->d_name.len = target->d_name.len;
2180 return;
2183 swap(dentry->d_name.len, target->d_name.len);
2186 static void dentry_lock_for_move(struct dentry *dentry, struct dentry *target)
2189 * XXXX: do we really need to take target->d_lock?
2191 if (IS_ROOT(dentry) || dentry->d_parent == target->d_parent)
2192 spin_lock(&target->d_parent->d_lock);
2193 else {
2194 if (d_ancestor(dentry->d_parent, target->d_parent)) {
2195 spin_lock(&dentry->d_parent->d_lock);
2196 spin_lock_nested(&target->d_parent->d_lock,
2197 DENTRY_D_LOCK_NESTED);
2198 } else {
2199 spin_lock(&target->d_parent->d_lock);
2200 spin_lock_nested(&dentry->d_parent->d_lock,
2201 DENTRY_D_LOCK_NESTED);
2204 if (target < dentry) {
2205 spin_lock_nested(&target->d_lock, 2);
2206 spin_lock_nested(&dentry->d_lock, 3);
2207 } else {
2208 spin_lock_nested(&dentry->d_lock, 2);
2209 spin_lock_nested(&target->d_lock, 3);
2213 static void dentry_unlock_parents_for_move(struct dentry *dentry,
2214 struct dentry *target)
2216 if (target->d_parent != dentry->d_parent)
2217 spin_unlock(&dentry->d_parent->d_lock);
2218 if (target->d_parent != target)
2219 spin_unlock(&target->d_parent->d_lock);
2223 * When switching names, the actual string doesn't strictly have to
2224 * be preserved in the target - because we're dropping the target
2225 * anyway. As such, we can just do a simple memcpy() to copy over
2226 * the new name before we switch.
2228 * Note that we have to be a lot more careful about getting the hash
2229 * switched - we have to switch the hash value properly even if it
2230 * then no longer matches the actual (corrupted) string of the target.
2231 * The hash value has to match the hash queue that the dentry is on..
2234 * d_move - move a dentry
2235 * @dentry: entry to move
2236 * @target: new dentry
2238 * Update the dcache to reflect the move of a file name. Negative
2239 * dcache entries should not be moved in this way.
2241 void d_move(struct dentry * dentry, struct dentry * target)
2243 if (!dentry->d_inode)
2244 printk(KERN_WARNING "VFS: moving negative dcache entry\n");
2246 BUG_ON(d_ancestor(dentry, target));
2247 BUG_ON(d_ancestor(target, dentry));
2249 write_seqlock(&rename_lock);
2251 dentry_lock_for_move(dentry, target);
2253 write_seqcount_begin(&dentry->d_seq);
2254 write_seqcount_begin(&target->d_seq);
2256 /* __d_drop does write_seqcount_barrier, but they're OK to nest. */
2259 * Move the dentry to the target hash queue. Don't bother checking
2260 * for the same hash queue because of how unlikely it is.
2262 __d_drop(dentry);
2263 __d_rehash(dentry, d_hash(target->d_parent, target->d_name.hash));
2265 /* Unhash the target: dput() will then get rid of it */
2266 __d_drop(target);
2268 list_del(&dentry->d_u.d_child);
2269 list_del(&target->d_u.d_child);
2271 /* Switch the names.. */
2272 switch_names(dentry, target);
2273 swap(dentry->d_name.hash, target->d_name.hash);
2275 /* ... and switch the parents */
2276 if (IS_ROOT(dentry)) {
2277 dentry->d_parent = target->d_parent;
2278 target->d_parent = target;
2279 INIT_LIST_HEAD(&target->d_u.d_child);
2280 } else {
2281 swap(dentry->d_parent, target->d_parent);
2283 /* And add them back to the (new) parent lists */
2284 list_add(&target->d_u.d_child, &target->d_parent->d_subdirs);
2287 list_add(&dentry->d_u.d_child, &dentry->d_parent->d_subdirs);
2289 write_seqcount_end(&target->d_seq);
2290 write_seqcount_end(&dentry->d_seq);
2292 dentry_unlock_parents_for_move(dentry, target);
2293 spin_unlock(&target->d_lock);
2294 fsnotify_d_move(dentry);
2295 spin_unlock(&dentry->d_lock);
2296 write_sequnlock(&rename_lock);
2298 EXPORT_SYMBOL(d_move);
2301 * d_ancestor - search for an ancestor
2302 * @p1: ancestor dentry
2303 * @p2: child dentry
2305 * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2306 * an ancestor of p2, else NULL.
2308 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2310 struct dentry *p;
2312 for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2313 if (p->d_parent == p1)
2314 return p;
2316 return NULL;
2320 * This helper attempts to cope with remotely renamed directories
2322 * It assumes that the caller is already holding
2323 * dentry->d_parent->d_inode->i_mutex and the inode->i_lock
2325 * Note: If ever the locking in lock_rename() changes, then please
2326 * remember to update this too...
2328 static struct dentry *__d_unalias(struct inode *inode,
2329 struct dentry *dentry, struct dentry *alias)
2331 struct mutex *m1 = NULL, *m2 = NULL;
2332 struct dentry *ret;
2334 /* If alias and dentry share a parent, then no extra locks required */
2335 if (alias->d_parent == dentry->d_parent)
2336 goto out_unalias;
2338 /* Check for loops */
2339 ret = ERR_PTR(-ELOOP);
2340 if (d_ancestor(alias, dentry))
2341 goto out_err;
2343 /* See lock_rename() */
2344 ret = ERR_PTR(-EBUSY);
2345 if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2346 goto out_err;
2347 m1 = &dentry->d_sb->s_vfs_rename_mutex;
2348 if (!mutex_trylock(&alias->d_parent->d_inode->i_mutex))
2349 goto out_err;
2350 m2 = &alias->d_parent->d_inode->i_mutex;
2351 out_unalias:
2352 d_move(alias, dentry);
2353 ret = alias;
2354 out_err:
2355 spin_unlock(&inode->i_lock);
2356 if (m2)
2357 mutex_unlock(m2);
2358 if (m1)
2359 mutex_unlock(m1);
2360 return ret;
2364 * Prepare an anonymous dentry for life in the superblock's dentry tree as a
2365 * named dentry in place of the dentry to be replaced.
2366 * returns with anon->d_lock held!
2368 static void __d_materialise_dentry(struct dentry *dentry, struct dentry *anon)
2370 struct dentry *dparent, *aparent;
2372 dentry_lock_for_move(anon, dentry);
2374 write_seqcount_begin(&dentry->d_seq);
2375 write_seqcount_begin(&anon->d_seq);
2377 dparent = dentry->d_parent;
2378 aparent = anon->d_parent;
2380 switch_names(dentry, anon);
2381 swap(dentry->d_name.hash, anon->d_name.hash);
2383 dentry->d_parent = (aparent == anon) ? dentry : aparent;
2384 list_del(&dentry->d_u.d_child);
2385 if (!IS_ROOT(dentry))
2386 list_add(&dentry->d_u.d_child, &dentry->d_parent->d_subdirs);
2387 else
2388 INIT_LIST_HEAD(&dentry->d_u.d_child);
2390 anon->d_parent = (dparent == dentry) ? anon : dparent;
2391 list_del(&anon->d_u.d_child);
2392 if (!IS_ROOT(anon))
2393 list_add(&anon->d_u.d_child, &anon->d_parent->d_subdirs);
2394 else
2395 INIT_LIST_HEAD(&anon->d_u.d_child);
2397 write_seqcount_end(&dentry->d_seq);
2398 write_seqcount_end(&anon->d_seq);
2400 dentry_unlock_parents_for_move(anon, dentry);
2401 spin_unlock(&dentry->d_lock);
2403 /* anon->d_lock still locked, returns locked */
2404 anon->d_flags &= ~DCACHE_DISCONNECTED;
2408 * d_materialise_unique - introduce an inode into the tree
2409 * @dentry: candidate dentry
2410 * @inode: inode to bind to the dentry, to which aliases may be attached
2412 * Introduces an dentry into the tree, substituting an extant disconnected
2413 * root directory alias in its place if there is one
2415 struct dentry *d_materialise_unique(struct dentry *dentry, struct inode *inode)
2417 struct dentry *actual;
2419 BUG_ON(!d_unhashed(dentry));
2421 if (!inode) {
2422 actual = dentry;
2423 __d_instantiate(dentry, NULL);
2424 d_rehash(actual);
2425 goto out_nolock;
2428 spin_lock(&inode->i_lock);
2430 if (S_ISDIR(inode->i_mode)) {
2431 struct dentry *alias;
2433 /* Does an aliased dentry already exist? */
2434 alias = __d_find_alias(inode, 0);
2435 if (alias) {
2436 actual = alias;
2437 /* Is this an anonymous mountpoint that we could splice
2438 * into our tree? */
2439 if (IS_ROOT(alias)) {
2440 __d_materialise_dentry(dentry, alias);
2441 __d_drop(alias);
2442 goto found;
2444 /* Nope, but we must(!) avoid directory aliasing */
2445 actual = __d_unalias(inode, dentry, alias);
2446 if (IS_ERR(actual))
2447 dput(alias);
2448 goto out_nolock;
2452 /* Add a unique reference */
2453 actual = __d_instantiate_unique(dentry, inode);
2454 if (!actual)
2455 actual = dentry;
2456 else
2457 BUG_ON(!d_unhashed(actual));
2459 spin_lock(&actual->d_lock);
2460 found:
2461 _d_rehash(actual);
2462 spin_unlock(&actual->d_lock);
2463 spin_unlock(&inode->i_lock);
2464 out_nolock:
2465 if (actual == dentry) {
2466 security_d_instantiate(dentry, inode);
2467 return NULL;
2470 iput(inode);
2471 return actual;
2473 EXPORT_SYMBOL_GPL(d_materialise_unique);
2475 static int prepend(char **buffer, int *buflen, const char *str, int namelen)
2477 *buflen -= namelen;
2478 if (*buflen < 0)
2479 return -ENAMETOOLONG;
2480 *buffer -= namelen;
2481 memcpy(*buffer, str, namelen);
2482 return 0;
2485 static int prepend_name(char **buffer, int *buflen, struct qstr *name)
2487 return prepend(buffer, buflen, name->name, name->len);
2491 * prepend_path - Prepend path string to a buffer
2492 * @path: the dentry/vfsmount to report
2493 * @root: root vfsmnt/dentry (may be modified by this function)
2494 * @buffer: pointer to the end of the buffer
2495 * @buflen: pointer to buffer length
2497 * Caller holds the rename_lock.
2499 * If path is not reachable from the supplied root, then the value of
2500 * root is changed (without modifying refcounts).
2502 static int prepend_path(const struct path *path, struct path *root,
2503 char **buffer, int *buflen)
2505 struct dentry *dentry = path->dentry;
2506 struct vfsmount *vfsmnt = path->mnt;
2507 bool slash = false;
2508 int error = 0;
2510 br_read_lock(vfsmount_lock);
2511 while (dentry != root->dentry || vfsmnt != root->mnt) {
2512 struct dentry * parent;
2514 if (dentry == vfsmnt->mnt_root || IS_ROOT(dentry)) {
2515 /* Global root? */
2516 if (vfsmnt->mnt_parent == vfsmnt) {
2517 goto global_root;
2519 dentry = vfsmnt->mnt_mountpoint;
2520 vfsmnt = vfsmnt->mnt_parent;
2521 continue;
2523 parent = dentry->d_parent;
2524 prefetch(parent);
2525 spin_lock(&dentry->d_lock);
2526 error = prepend_name(buffer, buflen, &dentry->d_name);
2527 spin_unlock(&dentry->d_lock);
2528 if (!error)
2529 error = prepend(buffer, buflen, "/", 1);
2530 if (error)
2531 break;
2533 slash = true;
2534 dentry = parent;
2537 out:
2538 if (!error && !slash)
2539 error = prepend(buffer, buflen, "/", 1);
2541 br_read_unlock(vfsmount_lock);
2542 return error;
2544 global_root:
2546 * Filesystems needing to implement special "root names"
2547 * should do so with ->d_dname()
2549 if (IS_ROOT(dentry) &&
2550 (dentry->d_name.len != 1 || dentry->d_name.name[0] != '/')) {
2551 WARN(1, "Root dentry has weird name <%.*s>\n",
2552 (int) dentry->d_name.len, dentry->d_name.name);
2554 root->mnt = vfsmnt;
2555 root->dentry = dentry;
2556 goto out;
2560 * __d_path - return the path of a dentry
2561 * @path: the dentry/vfsmount to report
2562 * @root: root vfsmnt/dentry (may be modified by this function)
2563 * @buf: buffer to return value in
2564 * @buflen: buffer length
2566 * Convert a dentry into an ASCII path name.
2568 * Returns a pointer into the buffer or an error code if the
2569 * path was too long.
2571 * "buflen" should be positive.
2573 * If path is not reachable from the supplied root, then the value of
2574 * root is changed (without modifying refcounts).
2576 char *__d_path(const struct path *path, struct path *root,
2577 char *buf, int buflen)
2579 char *res = buf + buflen;
2580 int error;
2582 prepend(&res, &buflen, "\0", 1);
2583 write_seqlock(&rename_lock);
2584 error = prepend_path(path, root, &res, &buflen);
2585 write_sequnlock(&rename_lock);
2587 if (error)
2588 return ERR_PTR(error);
2589 return res;
2593 * same as __d_path but appends "(deleted)" for unlinked files.
2595 static int path_with_deleted(const struct path *path, struct path *root,
2596 char **buf, int *buflen)
2598 prepend(buf, buflen, "\0", 1);
2599 if (d_unlinked(path->dentry)) {
2600 int error = prepend(buf, buflen, " (deleted)", 10);
2601 if (error)
2602 return error;
2605 return prepend_path(path, root, buf, buflen);
2608 static int prepend_unreachable(char **buffer, int *buflen)
2610 return prepend(buffer, buflen, "(unreachable)", 13);
2614 * d_path - return the path of a dentry
2615 * @path: path to report
2616 * @buf: buffer to return value in
2617 * @buflen: buffer length
2619 * Convert a dentry into an ASCII path name. If the entry has been deleted
2620 * the string " (deleted)" is appended. Note that this is ambiguous.
2622 * Returns a pointer into the buffer or an error code if the path was
2623 * too long. Note: Callers should use the returned pointer, not the passed
2624 * in buffer, to use the name! The implementation often starts at an offset
2625 * into the buffer, and may leave 0 bytes at the start.
2627 * "buflen" should be positive.
2629 char *d_path(const struct path *path, char *buf, int buflen)
2631 char *res = buf + buflen;
2632 struct path root;
2633 struct path tmp;
2634 int error;
2637 * We have various synthetic filesystems that never get mounted. On
2638 * these filesystems dentries are never used for lookup purposes, and
2639 * thus don't need to be hashed. They also don't need a name until a
2640 * user wants to identify the object in /proc/pid/fd/. The little hack
2641 * below allows us to generate a name for these objects on demand:
2643 if (path->dentry->d_op && path->dentry->d_op->d_dname)
2644 return path->dentry->d_op->d_dname(path->dentry, buf, buflen);
2646 get_fs_root(current->fs, &root);
2647 write_seqlock(&rename_lock);
2648 tmp = root;
2649 error = path_with_deleted(path, &tmp, &res, &buflen);
2650 if (error)
2651 res = ERR_PTR(error);
2652 write_sequnlock(&rename_lock);
2653 path_put(&root);
2654 return res;
2656 EXPORT_SYMBOL(d_path);
2659 * d_path_with_unreachable - return the path of a dentry
2660 * @path: path to report
2661 * @buf: buffer to return value in
2662 * @buflen: buffer length
2664 * The difference from d_path() is that this prepends "(unreachable)"
2665 * to paths which are unreachable from the current process' root.
2667 char *d_path_with_unreachable(const struct path *path, char *buf, int buflen)
2669 char *res = buf + buflen;
2670 struct path root;
2671 struct path tmp;
2672 int error;
2674 if (path->dentry->d_op && path->dentry->d_op->d_dname)
2675 return path->dentry->d_op->d_dname(path->dentry, buf, buflen);
2677 get_fs_root(current->fs, &root);
2678 write_seqlock(&rename_lock);
2679 tmp = root;
2680 error = path_with_deleted(path, &tmp, &res, &buflen);
2681 if (!error && !path_equal(&tmp, &root))
2682 error = prepend_unreachable(&res, &buflen);
2683 write_sequnlock(&rename_lock);
2684 path_put(&root);
2685 if (error)
2686 res = ERR_PTR(error);
2688 return res;
2692 * Helper function for dentry_operations.d_dname() members
2694 char *dynamic_dname(struct dentry *dentry, char *buffer, int buflen,
2695 const char *fmt, ...)
2697 va_list args;
2698 char temp[64];
2699 int sz;
2701 va_start(args, fmt);
2702 sz = vsnprintf(temp, sizeof(temp), fmt, args) + 1;
2703 va_end(args);
2705 if (sz > sizeof(temp) || sz > buflen)
2706 return ERR_PTR(-ENAMETOOLONG);
2708 buffer += buflen - sz;
2709 return memcpy(buffer, temp, sz);
2713 * Write full pathname from the root of the filesystem into the buffer.
2715 static char *__dentry_path(struct dentry *dentry, char *buf, int buflen)
2717 char *end = buf + buflen;
2718 char *retval;
2720 prepend(&end, &buflen, "\0", 1);
2721 if (buflen < 1)
2722 goto Elong;
2723 /* Get '/' right */
2724 retval = end-1;
2725 *retval = '/';
2727 while (!IS_ROOT(dentry)) {
2728 struct dentry *parent = dentry->d_parent;
2729 int error;
2731 prefetch(parent);
2732 spin_lock(&dentry->d_lock);
2733 error = prepend_name(&end, &buflen, &dentry->d_name);
2734 spin_unlock(&dentry->d_lock);
2735 if (error != 0 || prepend(&end, &buflen, "/", 1) != 0)
2736 goto Elong;
2738 retval = end;
2739 dentry = parent;
2741 return retval;
2742 Elong:
2743 return ERR_PTR(-ENAMETOOLONG);
2746 char *dentry_path_raw(struct dentry *dentry, char *buf, int buflen)
2748 char *retval;
2750 write_seqlock(&rename_lock);
2751 retval = __dentry_path(dentry, buf, buflen);
2752 write_sequnlock(&rename_lock);
2754 return retval;
2756 EXPORT_SYMBOL(dentry_path_raw);
2758 char *dentry_path(struct dentry *dentry, char *buf, int buflen)
2760 char *p = NULL;
2761 char *retval;
2763 write_seqlock(&rename_lock);
2764 if (d_unlinked(dentry)) {
2765 p = buf + buflen;
2766 if (prepend(&p, &buflen, "//deleted", 10) != 0)
2767 goto Elong;
2768 buflen++;
2770 retval = __dentry_path(dentry, buf, buflen);
2771 write_sequnlock(&rename_lock);
2772 if (!IS_ERR(retval) && p)
2773 *p = '/'; /* restore '/' overriden with '\0' */
2774 return retval;
2775 Elong:
2776 return ERR_PTR(-ENAMETOOLONG);
2780 * NOTE! The user-level library version returns a
2781 * character pointer. The kernel system call just
2782 * returns the length of the buffer filled (which
2783 * includes the ending '\0' character), or a negative
2784 * error value. So libc would do something like
2786 * char *getcwd(char * buf, size_t size)
2788 * int retval;
2790 * retval = sys_getcwd(buf, size);
2791 * if (retval >= 0)
2792 * return buf;
2793 * errno = -retval;
2794 * return NULL;
2797 SYSCALL_DEFINE2(getcwd, char __user *, buf, unsigned long, size)
2799 int error;
2800 struct path pwd, root;
2801 char *page = (char *) __get_free_page(GFP_USER);
2803 if (!page)
2804 return -ENOMEM;
2806 get_fs_root_and_pwd(current->fs, &root, &pwd);
2808 error = -ENOENT;
2809 write_seqlock(&rename_lock);
2810 if (!d_unlinked(pwd.dentry)) {
2811 unsigned long len;
2812 struct path tmp = root;
2813 char *cwd = page + PAGE_SIZE;
2814 int buflen = PAGE_SIZE;
2816 prepend(&cwd, &buflen, "\0", 1);
2817 error = prepend_path(&pwd, &tmp, &cwd, &buflen);
2818 write_sequnlock(&rename_lock);
2820 if (error)
2821 goto out;
2823 /* Unreachable from current root */
2824 if (!path_equal(&tmp, &root)) {
2825 error = prepend_unreachable(&cwd, &buflen);
2826 if (error)
2827 goto out;
2830 error = -ERANGE;
2831 len = PAGE_SIZE + page - cwd;
2832 if (len <= size) {
2833 error = len;
2834 if (copy_to_user(buf, cwd, len))
2835 error = -EFAULT;
2837 } else {
2838 write_sequnlock(&rename_lock);
2841 out:
2842 path_put(&pwd);
2843 path_put(&root);
2844 free_page((unsigned long) page);
2845 return error;
2849 * Test whether new_dentry is a subdirectory of old_dentry.
2851 * Trivially implemented using the dcache structure
2855 * is_subdir - is new dentry a subdirectory of old_dentry
2856 * @new_dentry: new dentry
2857 * @old_dentry: old dentry
2859 * Returns 1 if new_dentry is a subdirectory of the parent (at any depth).
2860 * Returns 0 otherwise.
2861 * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
2864 int is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
2866 int result;
2867 unsigned seq;
2869 if (new_dentry == old_dentry)
2870 return 1;
2872 do {
2873 /* for restarting inner loop in case of seq retry */
2874 seq = read_seqbegin(&rename_lock);
2876 * Need rcu_readlock to protect against the d_parent trashing
2877 * due to d_move
2879 rcu_read_lock();
2880 if (d_ancestor(old_dentry, new_dentry))
2881 result = 1;
2882 else
2883 result = 0;
2884 rcu_read_unlock();
2885 } while (read_seqretry(&rename_lock, seq));
2887 return result;
2890 int path_is_under(struct path *path1, struct path *path2)
2892 struct vfsmount *mnt = path1->mnt;
2893 struct dentry *dentry = path1->dentry;
2894 int res;
2896 br_read_lock(vfsmount_lock);
2897 if (mnt != path2->mnt) {
2898 for (;;) {
2899 if (mnt->mnt_parent == mnt) {
2900 br_read_unlock(vfsmount_lock);
2901 return 0;
2903 if (mnt->mnt_parent == path2->mnt)
2904 break;
2905 mnt = mnt->mnt_parent;
2907 dentry = mnt->mnt_mountpoint;
2909 res = is_subdir(dentry, path2->dentry);
2910 br_read_unlock(vfsmount_lock);
2911 return res;
2913 EXPORT_SYMBOL(path_is_under);
2915 void d_genocide(struct dentry *root)
2917 struct dentry *this_parent;
2918 struct list_head *next;
2919 unsigned seq;
2920 int locked = 0;
2922 seq = read_seqbegin(&rename_lock);
2923 again:
2924 this_parent = root;
2925 spin_lock(&this_parent->d_lock);
2926 repeat:
2927 next = this_parent->d_subdirs.next;
2928 resume:
2929 while (next != &this_parent->d_subdirs) {
2930 struct list_head *tmp = next;
2931 struct dentry *dentry = list_entry(tmp, struct dentry, d_u.d_child);
2932 next = tmp->next;
2934 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
2935 if (d_unhashed(dentry) || !dentry->d_inode) {
2936 spin_unlock(&dentry->d_lock);
2937 continue;
2939 if (!list_empty(&dentry->d_subdirs)) {
2940 spin_unlock(&this_parent->d_lock);
2941 spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
2942 this_parent = dentry;
2943 spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
2944 goto repeat;
2946 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
2947 dentry->d_flags |= DCACHE_GENOCIDE;
2948 dentry->d_count--;
2950 spin_unlock(&dentry->d_lock);
2952 if (this_parent != root) {
2953 struct dentry *child = this_parent;
2954 if (!(this_parent->d_flags & DCACHE_GENOCIDE)) {
2955 this_parent->d_flags |= DCACHE_GENOCIDE;
2956 this_parent->d_count--;
2958 this_parent = try_to_ascend(this_parent, locked, seq);
2959 if (!this_parent)
2960 goto rename_retry;
2961 next = child->d_u.d_child.next;
2962 goto resume;
2964 spin_unlock(&this_parent->d_lock);
2965 if (!locked && read_seqretry(&rename_lock, seq))
2966 goto rename_retry;
2967 if (locked)
2968 write_sequnlock(&rename_lock);
2969 return;
2971 rename_retry:
2972 locked = 1;
2973 write_seqlock(&rename_lock);
2974 goto again;
2978 * find_inode_number - check for dentry with name
2979 * @dir: directory to check
2980 * @name: Name to find.
2982 * Check whether a dentry already exists for the given name,
2983 * and return the inode number if it has an inode. Otherwise
2984 * 0 is returned.
2986 * This routine is used to post-process directory listings for
2987 * filesystems using synthetic inode numbers, and is necessary
2988 * to keep getcwd() working.
2991 ino_t find_inode_number(struct dentry *dir, struct qstr *name)
2993 struct dentry * dentry;
2994 ino_t ino = 0;
2996 dentry = d_hash_and_lookup(dir, name);
2997 if (dentry) {
2998 if (dentry->d_inode)
2999 ino = dentry->d_inode->i_ino;
3000 dput(dentry);
3002 return ino;
3004 EXPORT_SYMBOL(find_inode_number);
3006 static __initdata unsigned long dhash_entries;
3007 static int __init set_dhash_entries(char *str)
3009 if (!str)
3010 return 0;
3011 dhash_entries = simple_strtoul(str, &str, 0);
3012 return 1;
3014 __setup("dhash_entries=", set_dhash_entries);
3016 static void __init dcache_init_early(void)
3018 int loop;
3020 /* If hashes are distributed across NUMA nodes, defer
3021 * hash allocation until vmalloc space is available.
3023 if (hashdist)
3024 return;
3026 dentry_hashtable =
3027 alloc_large_system_hash("Dentry cache",
3028 sizeof(struct dcache_hash_bucket),
3029 dhash_entries,
3031 HASH_EARLY,
3032 &d_hash_shift,
3033 &d_hash_mask,
3036 for (loop = 0; loop < (1 << d_hash_shift); loop++)
3037 INIT_HLIST_BL_HEAD(&dentry_hashtable[loop].head);
3040 static void __init dcache_init(void)
3042 int loop;
3045 * A constructor could be added for stable state like the lists,
3046 * but it is probably not worth it because of the cache nature
3047 * of the dcache.
3049 dentry_cache = KMEM_CACHE(dentry,
3050 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD);
3052 register_shrinker(&dcache_shrinker);
3054 /* Hash may have been set up in dcache_init_early */
3055 if (!hashdist)
3056 return;
3058 dentry_hashtable =
3059 alloc_large_system_hash("Dentry cache",
3060 sizeof(struct dcache_hash_bucket),
3061 dhash_entries,
3064 &d_hash_shift,
3065 &d_hash_mask,
3068 for (loop = 0; loop < (1 << d_hash_shift); loop++)
3069 INIT_HLIST_BL_HEAD(&dentry_hashtable[loop].head);
3072 /* SLAB cache for __getname() consumers */
3073 struct kmem_cache *names_cachep __read_mostly;
3074 EXPORT_SYMBOL(names_cachep);
3076 EXPORT_SYMBOL(d_genocide);
3078 void __init vfs_caches_init_early(void)
3080 dcache_init_early();
3081 inode_init_early();
3084 void __init vfs_caches_init(unsigned long mempages)
3086 unsigned long reserve;
3088 /* Base hash sizes on available memory, with a reserve equal to
3089 150% of current kernel size */
3091 reserve = min((mempages - nr_free_pages()) * 3/2, mempages - 1);
3092 mempages -= reserve;
3094 names_cachep = kmem_cache_create("names_cache", PATH_MAX, 0,
3095 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3097 dcache_init();
3098 inode_init();
3099 files_init(mempages);
3100 mnt_init();
3101 bdev_cache_init();
3102 chrdev_init();