1/*
2 * fs/dcache.c
3 *
4 * Complete reimplementation
5 * (C) 1997 Thomas Schoebel-Theuer,
6 * with heavy changes by Linus Torvalds
7 */
8
9/*
10 * Notes on the allocation strategy:
11 *
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.
15 */
16
17#include <linux/ratelimit.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/export.h>
27#include <linux/security.h>
28#include <linux/seqlock.h>
29#include <linux/memblock.h>
30#include <linux/bit_spinlock.h>
31#include <linux/rculist_bl.h>
32#include <linux/list_lru.h>
33#include "internal.h"
34#include "mount.h"
35
36/*
37 * Usage:
38 * dcache->d_inode->i_lock protects:
39 * - i_dentry, d_u.d_alias, d_inode of aliases
40 * dcache_hash_bucket lock protects:
41 * - the dcache hash table
42 * s_roots bl list spinlock protects:
43 * - the s_roots list (see __d_drop)
44 * dentry->d_sb->s_dentry_lru_lock protects:
45 * - the dcache lru lists and counters
46 * d_lock protects:
47 * - d_flags
48 * - d_name
49 * - d_lru
50 * - d_count
51 * - d_unhashed()
52 * - d_parent and d_subdirs
53 * - childrens' d_child and d_parent
54 * - d_u.d_alias, d_inode
55 *
56 * Ordering:
57 * dentry->d_inode->i_lock
58 * dentry->d_lock
59 * dentry->d_sb->s_dentry_lru_lock
60 * dcache_hash_bucket lock
61 * s_roots lock
62 *
63 * If there is an ancestor relationship:
64 * dentry->d_parent->...->d_parent->d_lock
65 * ...
66 * dentry->d_parent->d_lock
67 * dentry->d_lock
68 *
69 * If no ancestor relationship:
70 * arbitrary, since it's serialized on rename_lock
71 */
72int sysctl_vfs_cache_pressure __read_mostly = 100;
73EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
74
75__cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
76
77EXPORT_SYMBOL(rename_lock);
78
79static struct kmem_cache *dentry_cache __read_mostly;
80
81const struct qstr empty_name = QSTR_INIT("", 0);
82EXPORT_SYMBOL(empty_name);
83const struct qstr slash_name = QSTR_INIT("/", 1);
84EXPORT_SYMBOL(slash_name);
85
86/*
87 * This is the single most critical data structure when it comes
88 * to the dcache: the hashtable for lookups. Somebody should try
89 * to make this good - I've just made it work.
90 *
91 * This hash-function tries to avoid losing too many bits of hash
92 * information, yet avoid using a prime hash-size or similar.
93 */
94
95static unsigned int d_hash_shift __read_mostly;
96
97static struct hlist_bl_head *dentry_hashtable __read_mostly;
98
99static inline struct hlist_bl_head *d_hash(unsigned int hash)
100{
101 return dentry_hashtable + (hash >> d_hash_shift);
102}
103
104#define IN_LOOKUP_SHIFT 10
105static struct hlist_bl_head in_lookup_hashtable[1 << IN_LOOKUP_SHIFT];
106
107static inline struct hlist_bl_head *in_lookup_hash(const struct dentry *parent,
108 unsigned int hash)
109{
110 hash += (unsigned long) parent / L1_CACHE_BYTES;
111 return in_lookup_hashtable + hash_32(hash, IN_LOOKUP_SHIFT);
112}
113
114
115/* Statistics gathering. */
116struct dentry_stat_t dentry_stat = {
117 .age_limit = 45,
118};
119
120static DEFINE_PER_CPU(long, nr_dentry);
121static DEFINE_PER_CPU(long, nr_dentry_unused);
122static DEFINE_PER_CPU(long, nr_dentry_negative);
123
124#if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
125
126/*
127 * Here we resort to our own counters instead of using generic per-cpu counters
128 * for consistency with what the vfs inode code does. We are expected to harvest
129 * better code and performance by having our own specialized counters.
130 *
131 * Please note that the loop is done over all possible CPUs, not over all online
132 * CPUs. The reason for this is that we don't want to play games with CPUs going
133 * on and off. If one of them goes off, we will just keep their counters.
134 *
135 * glommer: See cffbc8a for details, and if you ever intend to change this,
136 * please update all vfs counters to match.
137 */
138static long get_nr_dentry(void)
139{
140 int i;
141 long sum = 0;
142 for_each_possible_cpu(i)
143 sum += per_cpu(nr_dentry, i);
144 return sum < 0 ? 0 : sum;
145}
146
147static long get_nr_dentry_unused(void)
148{
149 int i;
150 long sum = 0;
151 for_each_possible_cpu(i)
152 sum += per_cpu(nr_dentry_unused, i);
153 return sum < 0 ? 0 : sum;
154}
155
156static long get_nr_dentry_negative(void)
157{
158 int i;
159 long sum = 0;
160
161 for_each_possible_cpu(i)
162 sum += per_cpu(nr_dentry_negative, i);
163 return sum < 0 ? 0 : sum;
164}
165
166int proc_nr_dentry(struct ctl_table *table, int write, void __user *buffer,
167 size_t *lenp, loff_t *ppos)
168{
169 dentry_stat.nr_dentry = get_nr_dentry();
170 dentry_stat.nr_unused = get_nr_dentry_unused();
171 dentry_stat.nr_negative = get_nr_dentry_negative();
172 return proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
173}
174#endif
175
176/*
177 * Compare 2 name strings, return 0 if they match, otherwise non-zero.
178 * The strings are both count bytes long, and count is non-zero.
179 */
180#ifdef CONFIG_DCACHE_WORD_ACCESS
181
182#include <asm/word-at-a-time.h>
183/*
184 * NOTE! 'cs' and 'scount' come from a dentry, so it has a
185 * aligned allocation for this particular component. We don't
186 * strictly need the load_unaligned_zeropad() safety, but it
187 * doesn't hurt either.
188 *
189 * In contrast, 'ct' and 'tcount' can be from a pathname, and do
190 * need the careful unaligned handling.
191 */
192static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
193{
194 unsigned long a,b,mask;
195
196 for (;;) {
197 a = read_word_at_a_time(cs);
198 b = load_unaligned_zeropad(ct);
199 if (tcount < sizeof(unsigned long))
200 break;
201 if (unlikely(a != b))
202 return 1;
203 cs += sizeof(unsigned long);
204 ct += sizeof(unsigned long);
205 tcount -= sizeof(unsigned long);
206 if (!tcount)
207 return 0;
208 }
209 mask = bytemask_from_count(tcount);
210 return unlikely(!!((a ^ b) & mask));
211}
212
213#else
214
215static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
216{
217 do {
218 if (*cs != *ct)
219 return 1;
220 cs++;
221 ct++;
222 tcount--;
223 } while (tcount);
224 return 0;
225}
226
227#endif
228
229static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount)
230{
231 /*
232 * Be careful about RCU walk racing with rename:
233 * use 'READ_ONCE' to fetch the name pointer.
234 *
235 * NOTE! Even if a rename will mean that the length
236 * was not loaded atomically, we don't care. The
237 * RCU walk will check the sequence count eventually,
238 * and catch it. And we won't overrun the buffer,
239 * because we're reading the name pointer atomically,
240 * and a dentry name is guaranteed to be properly
241 * terminated with a NUL byte.
242 *
243 * End result: even if 'len' is wrong, we'll exit
244 * early because the data cannot match (there can
245 * be no NUL in the ct/tcount data)
246 */
247 const unsigned char *cs = READ_ONCE(dentry->d_name.name);
248
249 return dentry_string_cmp(cs, ct, tcount);
250}
251
252struct external_name {
253 union {
254 atomic_t count;
255 struct rcu_head head;
256 } u;
257 unsigned char name[];
258};
259
260static inline struct external_name *external_name(struct dentry *dentry)
261{
262 return container_of(dentry->d_name.name, struct external_name, name[0]);
263}
264
265static void __d_free(struct rcu_head *head)
266{
267 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
268
269 kmem_cache_free(dentry_cache, dentry);
270}
271
272static void __d_free_external(struct rcu_head *head)
273{
274 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
275 kfree(external_name(dentry));
276 kmem_cache_free(dentry_cache, dentry);
277}
278
279static inline int dname_external(const struct dentry *dentry)
280{
281 return dentry->d_name.name != dentry->d_iname;
282}
283
284void take_dentry_name_snapshot(struct name_snapshot *name, struct dentry *dentry)
285{
286 spin_lock(&dentry->d_lock);
287 if (unlikely(dname_external(dentry))) {
288 struct external_name *p = external_name(dentry);
289 atomic_inc(&p->u.count);
290 spin_unlock(&dentry->d_lock);
291 name->name = p->name;
292 } else {
293 memcpy(name->inline_name, dentry->d_iname,
294 dentry->d_name.len + 1);
295 spin_unlock(&dentry->d_lock);
296 name->name = name->inline_name;
297 }
298}
299EXPORT_SYMBOL(take_dentry_name_snapshot);
300
301void release_dentry_name_snapshot(struct name_snapshot *name)
302{
303 if (unlikely(name->name != name->inline_name)) {
304 struct external_name *p;
305 p = container_of(name->name, struct external_name, name[0]);
306 if (unlikely(atomic_dec_and_test(&p->u.count)))
307 kfree_rcu(p, u.head);
308 }
309}
310EXPORT_SYMBOL(release_dentry_name_snapshot);
311
312static inline void __d_set_inode_and_type(struct dentry *dentry,
313 struct inode *inode,
314 unsigned type_flags)
315{
316 unsigned flags;
317
318 dentry->d_inode = inode;
319 flags = READ_ONCE(dentry->d_flags);
320 flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
321 flags |= type_flags;
322 WRITE_ONCE(dentry->d_flags, flags);
323}
324
325static inline void __d_clear_type_and_inode(struct dentry *dentry)
326{
327 unsigned flags = READ_ONCE(dentry->d_flags);
328
329 flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
330 WRITE_ONCE(dentry->d_flags, flags);
331 dentry->d_inode = NULL;
332 if (dentry->d_flags & DCACHE_LRU_LIST)
333 this_cpu_inc(nr_dentry_negative);
334}
335
336static void dentry_free(struct dentry *dentry)
337{
338 WARN_ON(!hlist_unhashed(&dentry->d_u.d_alias));
339 if (unlikely(dname_external(dentry))) {
340 struct external_name *p = external_name(dentry);
341 if (likely(atomic_dec_and_test(&p->u.count))) {
342 call_rcu(&dentry->d_u.d_rcu, __d_free_external);
343 return;
344 }
345 }
346 /* if dentry was never visible to RCU, immediate free is OK */
347 if (!(dentry->d_flags & DCACHE_RCUACCESS))
348 __d_free(&dentry->d_u.d_rcu);
349 else
350 call_rcu(&dentry->d_u.d_rcu, __d_free);
351}
352
353/*
354 * Release the dentry's inode, using the filesystem
355 * d_iput() operation if defined.
356 */
357static void dentry_unlink_inode(struct dentry * dentry)
358 __releases(dentry->d_lock)
359 __releases(dentry->d_inode->i_lock)
360{
361 struct inode *inode = dentry->d_inode;
362
363 raw_write_seqcount_begin(&dentry->d_seq);
364 __d_clear_type_and_inode(dentry);
365 hlist_del_init(&dentry->d_u.d_alias);
366 raw_write_seqcount_end(&dentry->d_seq);
367 spin_unlock(&dentry->d_lock);
368 spin_unlock(&inode->i_lock);
369 if (!inode->i_nlink)
370 fsnotify_inoderemove(inode);
371 if (dentry->d_op && dentry->d_op->d_iput)
372 dentry->d_op->d_iput(dentry, inode);
373 else
374 iput(inode);
375}
376
377/*
378 * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
379 * is in use - which includes both the "real" per-superblock
380 * LRU list _and_ the DCACHE_SHRINK_LIST use.
381 *
382 * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
383 * on the shrink list (ie not on the superblock LRU list).
384 *
385 * The per-cpu "nr_dentry_unused" counters are updated with
386 * the DCACHE_LRU_LIST bit.
387 *
388 * The per-cpu "nr_dentry_negative" counters are only updated
389 * when deleted from or added to the per-superblock LRU list, not
390 * from/to the shrink list. That is to avoid an unneeded dec/inc
391 * pair when moving from LRU to shrink list in select_collect().
392 *
393 * These helper functions make sure we always follow the
394 * rules. d_lock must be held by the caller.
395 */
396#define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
397static void d_lru_add(struct dentry *dentry)
398{
399 D_FLAG_VERIFY(dentry, 0);
400 dentry->d_flags |= DCACHE_LRU_LIST;
401 this_cpu_inc(nr_dentry_unused);
402 if (d_is_negative(dentry))
403 this_cpu_inc(nr_dentry_negative);
404 WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
405}
406
407static void d_lru_del(struct dentry *dentry)
408{
409 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
410 dentry->d_flags &= ~DCACHE_LRU_LIST;
411 this_cpu_dec(nr_dentry_unused);
412 if (d_is_negative(dentry))
413 this_cpu_dec(nr_dentry_negative);
414 WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
415}
416
417static void d_shrink_del(struct dentry *dentry)
418{
419 D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
420 list_del_init(&dentry->d_lru);
421 dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
422 this_cpu_dec(nr_dentry_unused);
423}
424
425static void d_shrink_add(struct dentry *dentry, struct list_head *list)
426{
427 D_FLAG_VERIFY(dentry, 0);
428 list_add(&dentry->d_lru, list);
429 dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
430 this_cpu_inc(nr_dentry_unused);
431}
432
433/*
434 * These can only be called under the global LRU lock, ie during the
435 * callback for freeing the LRU list. "isolate" removes it from the
436 * LRU lists entirely, while shrink_move moves it to the indicated
437 * private list.
438 */
439static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
440{
441 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
442 dentry->d_flags &= ~DCACHE_LRU_LIST;
443 this_cpu_dec(nr_dentry_unused);
444 if (d_is_negative(dentry))
445 this_cpu_dec(nr_dentry_negative);
446 list_lru_isolate(lru, &dentry->d_lru);
447}
448
449static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
450 struct list_head *list)
451{
452 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
453 dentry->d_flags |= DCACHE_SHRINK_LIST;
454 if (d_is_negative(dentry))
455 this_cpu_dec(nr_dentry_negative);
456 list_lru_isolate_move(lru, &dentry->d_lru, list);
457}
458
459/**
460 * d_drop - drop a dentry
461 * @dentry: dentry to drop
462 *
463 * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
464 * be found through a VFS lookup any more. Note that this is different from
465 * deleting the dentry - d_delete will try to mark the dentry negative if
466 * possible, giving a successful _negative_ lookup, while d_drop will
467 * just make the cache lookup fail.
468 *
469 * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
470 * reason (NFS timeouts or autofs deletes).
471 *
472 * __d_drop requires dentry->d_lock
473 * ___d_drop doesn't mark dentry as "unhashed"
474 * (dentry->d_hash.pprev will be LIST_POISON2, not NULL).
475 */
476static void ___d_drop(struct dentry *dentry)
477{
478 struct hlist_bl_head *b;
479 /*
480 * Hashed dentries are normally on the dentry hashtable,
481 * with the exception of those newly allocated by
482 * d_obtain_root, which are always IS_ROOT:
483 */
484 if (unlikely(IS_ROOT(dentry)))
485 b = &dentry->d_sb->s_roots;
486 else
487 b = d_hash(dentry->d_name.hash);
488
489 hlist_bl_lock(b);
490 __hlist_bl_del(&dentry->d_hash);
491 hlist_bl_unlock(b);
492}
493
494void __d_drop(struct dentry *dentry)
495{
496 if (!d_unhashed(dentry)) {
497 ___d_drop(dentry);
498 dentry->d_hash.pprev = NULL;
499 write_seqcount_invalidate(&dentry->d_seq);
500 }
501}
502EXPORT_SYMBOL(__d_drop);
503
504void d_drop(struct dentry *dentry)
505{
506 spin_lock(&dentry->d_lock);
507 __d_drop(dentry);
508 spin_unlock(&dentry->d_lock);
509}
510EXPORT_SYMBOL(d_drop);
511
512static inline void dentry_unlist(struct dentry *dentry, struct dentry *parent)
513{
514 struct dentry *next;
515 /*
516 * Inform d_walk() and shrink_dentry_list() that we are no longer
517 * attached to the dentry tree
518 */
519 dentry->d_flags |= DCACHE_DENTRY_KILLED;
520 if (unlikely(list_empty(&dentry->d_child)))
521 return;
522 __list_del_entry(&dentry->d_child);
523 /*
524 * Cursors can move around the list of children. While we'd been
525 * a normal list member, it didn't matter - ->d_child.next would've
526 * been updated. However, from now on it won't be and for the
527 * things like d_walk() it might end up with a nasty surprise.
528 * Normally d_walk() doesn't care about cursors moving around -
529 * ->d_lock on parent prevents that and since a cursor has no children
530 * of its own, we get through it without ever unlocking the parent.
531 * There is one exception, though - if we ascend from a child that
532 * gets killed as soon as we unlock it, the next sibling is found
533 * using the value left in its ->d_child.next. And if _that_
534 * pointed to a cursor, and cursor got moved (e.g. by lseek())
535 * before d_walk() regains parent->d_lock, we'll end up skipping
536 * everything the cursor had been moved past.
537 *
538 * Solution: make sure that the pointer left behind in ->d_child.next
539 * points to something that won't be moving around. I.e. skip the
540 * cursors.
541 */
542 while (dentry->d_child.next != &parent->d_subdirs) {
543 next = list_entry(dentry->d_child.next, struct dentry, d_child);
544 if (likely(!(next->d_flags & DCACHE_DENTRY_CURSOR)))
545 break;
546 dentry->d_child.next = next->d_child.next;
547 }
548}
549
550static void __dentry_kill(struct dentry *dentry)
551{
552 struct dentry *parent = NULL;
553 bool can_free = true;
554 if (!IS_ROOT(dentry))
555 parent = dentry->d_parent;
556
557 /*
558 * The dentry is now unrecoverably dead to the world.
559 */
560 lockref_mark_dead(&dentry->d_lockref);
561
562 /*
563 * inform the fs via d_prune that this dentry is about to be
564 * unhashed and destroyed.
565 */
566 if (dentry->d_flags & DCACHE_OP_PRUNE)
567 dentry->d_op->d_prune(dentry);
568
569 if (dentry->d_flags & DCACHE_LRU_LIST) {
570 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
571 d_lru_del(dentry);
572 }
573 /* if it was on the hash then remove it */
574 __d_drop(dentry);
575 dentry_unlist(dentry, parent);
576 if (parent)
577 spin_unlock(&parent->d_lock);
578 if (dentry->d_inode)
579 dentry_unlink_inode(dentry);
580 else
581 spin_unlock(&dentry->d_lock);
582 this_cpu_dec(nr_dentry);
583 if (dentry->d_op && dentry->d_op->d_release)
584 dentry->d_op->d_release(dentry);
585
586 spin_lock(&dentry->d_lock);
587 if (dentry->d_flags & DCACHE_SHRINK_LIST) {
588 dentry->d_flags |= DCACHE_MAY_FREE;
589 can_free = false;
590 }
591 spin_unlock(&dentry->d_lock);
592 if (likely(can_free))
593 dentry_free(dentry);
594 cond_resched();
595}
596
597static struct dentry *__lock_parent(struct dentry *dentry)
598{
599 struct dentry *parent;
600 rcu_read_lock();
601 spin_unlock(&dentry->d_lock);
602again:
603 parent = READ_ONCE(dentry->d_parent);
604 spin_lock(&parent->d_lock);
605 /*
606 * We can't blindly lock dentry until we are sure
607 * that we won't violate the locking order.
608 * Any changes of dentry->d_parent must have
609 * been done with parent->d_lock held, so
610 * spin_lock() above is enough of a barrier
611 * for checking if it's still our child.
612 */
613 if (unlikely(parent != dentry->d_parent)) {
614 spin_unlock(&parent->d_lock);
615 goto again;
616 }
617 rcu_read_unlock();
618 if (parent != dentry)
619 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
620 else
621 parent = NULL;
622 return parent;
623}
624
625static inline struct dentry *lock_parent(struct dentry *dentry)
626{
627 struct dentry *parent = dentry->d_parent;
628 if (IS_ROOT(dentry))
629 return NULL;
630 if (likely(spin_trylock(&parent->d_lock)))
631 return parent;
632 return __lock_parent(dentry);
633}
634
635static inline bool retain_dentry(struct dentry *dentry)
636{
637 WARN_ON(d_in_lookup(dentry));
638
639 /* Unreachable? Get rid of it */
640 if (unlikely(d_unhashed(dentry)))
641 return false;
642
643 if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
644 return false;
645
646 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
647 if (dentry->d_op->d_delete(dentry))
648 return false;
649 }
650 /* retain; LRU fodder */
651 dentry->d_lockref.count--;
652 if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
653 d_lru_add(dentry);
654 else if (unlikely(!(dentry->d_flags & DCACHE_REFERENCED)))
655 dentry->d_flags |= DCACHE_REFERENCED;
656 return true;
657}
658
659/*
660 * Finish off a dentry we've decided to kill.
661 * dentry->d_lock must be held, returns with it unlocked.
662 * Returns dentry requiring refcount drop, or NULL if we're done.
663 */
664static struct dentry *dentry_kill(struct dentry *dentry)
665 __releases(dentry->d_lock)
666{
667 struct inode *inode = dentry->d_inode;
668 struct dentry *parent = NULL;
669
670 if (inode && unlikely(!spin_trylock(&inode->i_lock)))
671 goto slow_positive;
672
673 if (!IS_ROOT(dentry)) {
674 parent = dentry->d_parent;
675 if (unlikely(!spin_trylock(&parent->d_lock))) {
676 parent = __lock_parent(dentry);
677 if (likely(inode || !dentry->d_inode))
678 goto got_locks;
679 /* negative that became positive */
680 if (parent)
681 spin_unlock(&parent->d_lock);
682 inode = dentry->d_inode;
683 goto slow_positive;
684 }
685 }
686 __dentry_kill(dentry);
687 return parent;
688
689slow_positive:
690 spin_unlock(&dentry->d_lock);
691 spin_lock(&inode->i_lock);
692 spin_lock(&dentry->d_lock);
693 parent = lock_parent(dentry);
694got_locks:
695 if (unlikely(dentry->d_lockref.count != 1)) {
696 dentry->d_lockref.count--;
697 } else if (likely(!retain_dentry(dentry))) {
698 __dentry_kill(dentry);
699 return parent;
700 }
701 /* we are keeping it, after all */
702 if (inode)
703 spin_unlock(&inode->i_lock);
704 if (parent)
705 spin_unlock(&parent->d_lock);
706 spin_unlock(&dentry->d_lock);
707 return NULL;
708}
709
710/*
711 * Try to do a lockless dput(), and return whether that was successful.
712 *
713 * If unsuccessful, we return false, having already taken the dentry lock.
714 *
715 * The caller needs to hold the RCU read lock, so that the dentry is
716 * guaranteed to stay around even if the refcount goes down to zero!
717 */
718static inline bool fast_dput(struct dentry *dentry)
719{
720 int ret;
721 unsigned int d_flags;
722
723 /*
724 * If we have a d_op->d_delete() operation, we sould not
725 * let the dentry count go to zero, so use "put_or_lock".
726 */
727 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE))
728 return lockref_put_or_lock(&dentry->d_lockref);
729
730 /*
731 * .. otherwise, we can try to just decrement the
732 * lockref optimistically.
733 */
734 ret = lockref_put_return(&dentry->d_lockref);
735
736 /*
737 * If the lockref_put_return() failed due to the lock being held
738 * by somebody else, the fast path has failed. We will need to
739 * get the lock, and then check the count again.
740 */
741 if (unlikely(ret < 0)) {
742 spin_lock(&dentry->d_lock);
743 if (dentry->d_lockref.count > 1) {
744 dentry->d_lockref.count--;
745 spin_unlock(&dentry->d_lock);
746 return true;
747 }
748 return false;
749 }
750
751 /*
752 * If we weren't the last ref, we're done.
753 */
754 if (ret)
755 return true;
756
757 /*
758 * Careful, careful. The reference count went down
759 * to zero, but we don't hold the dentry lock, so
760 * somebody else could get it again, and do another
761 * dput(), and we need to not race with that.
762 *
763 * However, there is a very special and common case
764 * where we don't care, because there is nothing to
765 * do: the dentry is still hashed, it does not have
766 * a 'delete' op, and it's referenced and already on
767 * the LRU list.
768 *
769 * NOTE! Since we aren't locked, these values are
770 * not "stable". However, it is sufficient that at
771 * some point after we dropped the reference the
772 * dentry was hashed and the flags had the proper
773 * value. Other dentry users may have re-gotten
774 * a reference to the dentry and change that, but
775 * our work is done - we can leave the dentry
776 * around with a zero refcount.
777 */
778 smp_rmb();
779 d_flags = READ_ONCE(dentry->d_flags);
780 d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST | DCACHE_DISCONNECTED;
781
782 /* Nothing to do? Dropping the reference was all we needed? */
783 if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry))
784 return true;
785
786 /*
787 * Not the fast normal case? Get the lock. We've already decremented
788 * the refcount, but we'll need to re-check the situation after
789 * getting the lock.
790 */
791 spin_lock(&dentry->d_lock);
792
793 /*
794 * Did somebody else grab a reference to it in the meantime, and
795 * we're no longer the last user after all? Alternatively, somebody
796 * else could have killed it and marked it dead. Either way, we
797 * don't need to do anything else.
798 */
799 if (dentry->d_lockref.count) {
800 spin_unlock(&dentry->d_lock);
801 return true;
802 }
803
804 /*
805 * Re-get the reference we optimistically dropped. We hold the
806 * lock, and we just tested that it was zero, so we can just
807 * set it to 1.
808 */
809 dentry->d_lockref.count = 1;
810 return false;
811}
812
813
814/*
815 * This is dput
816 *
817 * This is complicated by the fact that we do not want to put
818 * dentries that are no longer on any hash chain on the unused
819 * list: we'd much rather just get rid of them immediately.
820 *
821 * However, that implies that we have to traverse the dentry
822 * tree upwards to the parents which might _also_ now be
823 * scheduled for deletion (it may have been only waiting for
824 * its last child to go away).
825 *
826 * This tail recursion is done by hand as we don't want to depend
827 * on the compiler to always get this right (gcc generally doesn't).
828 * Real recursion would eat up our stack space.
829 */
830
831/*
832 * dput - release a dentry
833 * @dentry: dentry to release
834 *
835 * Release a dentry. This will drop the usage count and if appropriate
836 * call the dentry unlink method as well as removing it from the queues and
837 * releasing its resources. If the parent dentries were scheduled for release
838 * they too may now get deleted.
839 */
840void dput(struct dentry *dentry)
841{
842 while (dentry) {
843 might_sleep();
844
845 rcu_read_lock();
846 if (likely(fast_dput(dentry))) {
847 rcu_read_unlock();
848 return;
849 }
850
851 /* Slow case: now with the dentry lock held */
852 rcu_read_unlock();
853
854 if (likely(retain_dentry(dentry))) {
855 spin_unlock(&dentry->d_lock);
856 return;
857 }
858
859 dentry = dentry_kill(dentry);
860 }
861}
862EXPORT_SYMBOL(dput);
863
864
865/* This must be called with d_lock held */
866static inline void __dget_dlock(struct dentry *dentry)
867{
868 dentry->d_lockref.count++;
869}
870
871static inline void __dget(struct dentry *dentry)
872{
873 lockref_get(&dentry->d_lockref);
874}
875
876struct dentry *dget_parent(struct dentry *dentry)
877{
878 int gotref;
879 struct dentry *ret;
880
881 /*
882 * Do optimistic parent lookup without any
883 * locking.
884 */
885 rcu_read_lock();
886 ret = READ_ONCE(dentry->d_parent);
887 gotref = lockref_get_not_zero(&ret->d_lockref);
888 rcu_read_unlock();
889 if (likely(gotref)) {
890 if (likely(ret == READ_ONCE(dentry->d_parent)))
891 return ret;
892 dput(ret);
893 }
894
895repeat:
896 /*
897 * Don't need rcu_dereference because we re-check it was correct under
898 * the lock.
899 */
900 rcu_read_lock();
901 ret = dentry->d_parent;
902 spin_lock(&ret->d_lock);
903 if (unlikely(ret != dentry->d_parent)) {
904 spin_unlock(&ret->d_lock);
905 rcu_read_unlock();
906 goto repeat;
907 }
908 rcu_read_unlock();
909 BUG_ON(!ret->d_lockref.count);
910 ret->d_lockref.count++;
911 spin_unlock(&ret->d_lock);
912 return ret;
913}
914EXPORT_SYMBOL(dget_parent);
915
916static struct dentry * __d_find_any_alias(struct inode *inode)
917{
918 struct dentry *alias;
919
920 if (hlist_empty(&inode->i_dentry))
921 return NULL;
922 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
923 __dget(alias);
924 return alias;
925}
926
927/**
928 * d_find_any_alias - find any alias for a given inode
929 * @inode: inode to find an alias for
930 *
931 * If any aliases exist for the given inode, take and return a
932 * reference for one of them. If no aliases exist, return %NULL.
933 */
934struct dentry *d_find_any_alias(struct inode *inode)
935{
936 struct dentry *de;
937
938 spin_lock(&inode->i_lock);
939 de = __d_find_any_alias(inode);
940 spin_unlock(&inode->i_lock);
941 return de;
942}
943EXPORT_SYMBOL(d_find_any_alias);
944
945/**
946 * d_find_alias - grab a hashed alias of inode
947 * @inode: inode in question
948 *
949 * If inode has a hashed alias, or is a directory and has any alias,
950 * acquire the reference to alias and return it. Otherwise return NULL.
951 * Notice that if inode is a directory there can be only one alias and
952 * it can be unhashed only if it has no children, or if it is the root
953 * of a filesystem, or if the directory was renamed and d_revalidate
954 * was the first vfs operation to notice.
955 *
956 * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
957 * any other hashed alias over that one.
958 */
959static struct dentry *__d_find_alias(struct inode *inode)
960{
961 struct dentry *alias;
962
963 if (S_ISDIR(inode->i_mode))
964 return __d_find_any_alias(inode);
965
966 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
967 spin_lock(&alias->d_lock);
968 if (!d_unhashed(alias)) {
969 __dget_dlock(alias);
970 spin_unlock(&alias->d_lock);
971 return alias;
972 }
973 spin_unlock(&alias->d_lock);
974 }
975 return NULL;
976}
977
978struct dentry *d_find_alias(struct inode *inode)
979{
980 struct dentry *de = NULL;
981
982 if (!hlist_empty(&inode->i_dentry)) {
983 spin_lock(&inode->i_lock);
984 de = __d_find_alias(inode);
985 spin_unlock(&inode->i_lock);
986 }
987 return de;
988}
989EXPORT_SYMBOL(d_find_alias);
990
991/*
992 * Try to kill dentries associated with this inode.
993 * WARNING: you must own a reference to inode.
994 */
995void d_prune_aliases(struct inode *inode)
996{
997 struct dentry *dentry;
998restart:
999 spin_lock(&inode->i_lock);
1000 hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
1001 spin_lock(&dentry->d_lock);
1002 if (!dentry->d_lockref.count) {
1003 struct dentry *parent = lock_parent(dentry);
1004 if (likely(!dentry->d_lockref.count)) {
1005 __dentry_kill(dentry);
1006 dput(parent);
1007 goto restart;
1008 }
1009 if (parent)
1010 spin_unlock(&parent->d_lock);
1011 }
1012 spin_unlock(&dentry->d_lock);
1013 }
1014 spin_unlock(&inode->i_lock);
1015}
1016EXPORT_SYMBOL(d_prune_aliases);
1017
1018/*
1019 * Lock a dentry from shrink list.
1020 * Called under rcu_read_lock() and dentry->d_lock; the former
1021 * guarantees that nothing we access will be freed under us.
1022 * Note that dentry is *not* protected from concurrent dentry_kill(),
1023 * d_delete(), etc.
1024 *
1025 * Return false if dentry has been disrupted or grabbed, leaving
1026 * the caller to kick it off-list. Otherwise, return true and have
1027 * that dentry's inode and parent both locked.
1028 */
1029static bool shrink_lock_dentry(struct dentry *dentry)
1030{
1031 struct inode *inode;
1032 struct dentry *parent;
1033
1034 if (dentry->d_lockref.count)
1035 return false;
1036
1037 inode = dentry->d_inode;
1038 if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
1039 spin_unlock(&dentry->d_lock);
1040 spin_lock(&inode->i_lock);
1041 spin_lock(&dentry->d_lock);
1042 if (unlikely(dentry->d_lockref.count))
1043 goto out;
1044 /* changed inode means that somebody had grabbed it */
1045 if (unlikely(inode != dentry->d_inode))
1046 goto out;
1047 }
1048
1049 parent = dentry->d_parent;
1050 if (IS_ROOT(dentry) || likely(spin_trylock(&parent->d_lock)))
1051 return true;
1052
1053 spin_unlock(&dentry->d_lock);
1054 spin_lock(&parent->d_lock);
1055 if (unlikely(parent != dentry->d_parent)) {
1056 spin_unlock(&parent->d_lock);
1057 spin_lock(&dentry->d_lock);
1058 goto out;
1059 }
1060 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1061 if (likely(!dentry->d_lockref.count))
1062 return true;
1063 spin_unlock(&parent->d_lock);
1064out:
1065 if (inode)
1066 spin_unlock(&inode->i_lock);
1067 return false;
1068}
1069
1070static void shrink_dentry_list(struct list_head *list)
1071{
1072 while (!list_empty(list)) {
1073 struct dentry *dentry, *parent;
1074
1075 dentry = list_entry(list->prev, struct dentry, d_lru);
1076 spin_lock(&dentry->d_lock);
1077 rcu_read_lock();
1078 if (!shrink_lock_dentry(dentry)) {
1079 bool can_free = false;
1080 rcu_read_unlock();
1081 d_shrink_del(dentry);
1082 if (dentry->d_lockref.count < 0)
1083 can_free = dentry->d_flags & DCACHE_MAY_FREE;
1084 spin_unlock(&dentry->d_lock);
1085 if (can_free)
1086 dentry_free(dentry);
1087 continue;
1088 }
1089 rcu_read_unlock();
1090 d_shrink_del(dentry);
1091 parent = dentry->d_parent;
1092 __dentry_kill(dentry);
1093 if (parent == dentry)
1094 continue;
1095 /*
1096 * We need to prune ancestors too. This is necessary to prevent
1097 * quadratic behavior of shrink_dcache_parent(), but is also
1098 * expected to be beneficial in reducing dentry cache
1099 * fragmentation.
1100 */
1101 dentry = parent;
1102 while (dentry && !lockref_put_or_lock(&dentry->d_lockref))
1103 dentry = dentry_kill(dentry);
1104 }
1105}
1106
1107static enum lru_status dentry_lru_isolate(struct list_head *item,
1108 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1109{
1110 struct list_head *freeable = arg;
1111 struct dentry *dentry = container_of(item, struct dentry, d_lru);
1112
1113
1114 /*
1115 * we are inverting the lru lock/dentry->d_lock here,
1116 * so use a trylock. If we fail to get the lock, just skip
1117 * it
1118 */
1119 if (!spin_trylock(&dentry->d_lock))
1120 return LRU_SKIP;
1121
1122 /*
1123 * Referenced dentries are still in use. If they have active
1124 * counts, just remove them from the LRU. Otherwise give them
1125 * another pass through the LRU.
1126 */
1127 if (dentry->d_lockref.count) {
1128 d_lru_isolate(lru, dentry);
1129 spin_unlock(&dentry->d_lock);
1130 return LRU_REMOVED;
1131 }
1132
1133 if (dentry->d_flags & DCACHE_REFERENCED) {
1134 dentry->d_flags &= ~DCACHE_REFERENCED;
1135 spin_unlock(&dentry->d_lock);
1136
1137 /*
1138 * The list move itself will be made by the common LRU code. At
1139 * this point, we've dropped the dentry->d_lock but keep the
1140 * lru lock. This is safe to do, since every list movement is
1141 * protected by the lru lock even if both locks are held.
1142 *
1143 * This is guaranteed by the fact that all LRU management
1144 * functions are intermediated by the LRU API calls like
1145 * list_lru_add and list_lru_del. List movement in this file
1146 * only ever occur through this functions or through callbacks
1147 * like this one, that are called from the LRU API.
1148 *
1149 * The only exceptions to this are functions like
1150 * shrink_dentry_list, and code that first checks for the
1151 * DCACHE_SHRINK_LIST flag. Those are guaranteed to be
1152 * operating only with stack provided lists after they are
1153 * properly isolated from the main list. It is thus, always a
1154 * local access.
1155 */
1156 return LRU_ROTATE;
1157 }
1158
1159 d_lru_shrink_move(lru, dentry, freeable);
1160 spin_unlock(&dentry->d_lock);
1161
1162 return LRU_REMOVED;
1163}
1164
1165/**
1166 * prune_dcache_sb - shrink the dcache
1167 * @sb: superblock
1168 * @sc: shrink control, passed to list_lru_shrink_walk()
1169 *
1170 * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1171 * is done when we need more memory and called from the superblock shrinker
1172 * function.
1173 *
1174 * This function may fail to free any resources if all the dentries are in
1175 * use.
1176 */
1177long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1178{
1179 LIST_HEAD(dispose);
1180 long freed;
1181
1182 freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1183 dentry_lru_isolate, &dispose);
1184 shrink_dentry_list(&dispose);
1185 return freed;
1186}
1187
1188static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1189 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1190{
1191 struct list_head *freeable = arg;
1192 struct dentry *dentry = container_of(item, struct dentry, d_lru);
1193
1194 /*
1195 * we are inverting the lru lock/dentry->d_lock here,
1196 * so use a trylock. If we fail to get the lock, just skip
1197 * it
1198 */
1199 if (!spin_trylock(&dentry->d_lock))
1200 return LRU_SKIP;
1201
1202 d_lru_shrink_move(lru, dentry, freeable);
1203 spin_unlock(&dentry->d_lock);
1204
1205 return LRU_REMOVED;
1206}
1207
1208
1209/**
1210 * shrink_dcache_sb - shrink dcache for a superblock
1211 * @sb: superblock
1212 *
1213 * Shrink the dcache for the specified super block. This is used to free
1214 * the dcache before unmounting a file system.
1215 */
1216void shrink_dcache_sb(struct super_block *sb)
1217{
1218 do {
1219 LIST_HEAD(dispose);
1220
1221 list_lru_walk(&sb->s_dentry_lru,
1222 dentry_lru_isolate_shrink, &dispose, 1024);
1223 shrink_dentry_list(&dispose);
1224 } while (list_lru_count(&sb->s_dentry_lru) > 0);
1225}
1226EXPORT_SYMBOL(shrink_dcache_sb);
1227
1228/**
1229 * enum d_walk_ret - action to talke during tree walk
1230 * @D_WALK_CONTINUE: contrinue walk
1231 * @D_WALK_QUIT: quit walk
1232 * @D_WALK_NORETRY: quit when retry is needed
1233 * @D_WALK_SKIP: skip this dentry and its children
1234 */
1235enum d_walk_ret {
1236 D_WALK_CONTINUE,
1237 D_WALK_QUIT,
1238 D_WALK_NORETRY,
1239 D_WALK_SKIP,
1240};
1241
1242/**
1243 * d_walk - walk the dentry tree
1244 * @parent: start of walk
1245 * @data: data passed to @enter() and @finish()
1246 * @enter: callback when first entering the dentry
1247 *
1248 * The @enter() callbacks are called with d_lock held.
1249 */
1250static void d_walk(struct dentry *parent, void *data,
1251 enum d_walk_ret (*enter)(void *, struct dentry *))
1252{
1253 struct dentry *this_parent;
1254 struct list_head *next;
1255 unsigned seq = 0;
1256 enum d_walk_ret ret;
1257 bool retry = true;
1258
1259again:
1260 read_seqbegin_or_lock(&rename_lock, &seq);
1261 this_parent = parent;
1262 spin_lock(&this_parent->d_lock);
1263
1264 ret = enter(data, this_parent);
1265 switch (ret) {
1266 case D_WALK_CONTINUE:
1267 break;
1268 case D_WALK_QUIT:
1269 case D_WALK_SKIP:
1270 goto out_unlock;
1271 case D_WALK_NORETRY:
1272 retry = false;
1273 break;
1274 }
1275repeat:
1276 next = this_parent->d_subdirs.next;
1277resume:
1278 while (next != &this_parent->d_subdirs) {
1279 struct list_head *tmp = next;
1280 struct dentry *dentry = list_entry(tmp, struct dentry, d_child);
1281 next = tmp->next;
1282
1283 if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
1284 continue;
1285
1286 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1287
1288 ret = enter(data, dentry);
1289 switch (ret) {
1290 case D_WALK_CONTINUE:
1291 break;
1292 case D_WALK_QUIT:
1293 spin_unlock(&dentry->d_lock);
1294 goto out_unlock;
1295 case D_WALK_NORETRY:
1296 retry = false;
1297 break;
1298 case D_WALK_SKIP:
1299 spin_unlock(&dentry->d_lock);
1300 continue;
1301 }
1302
1303 if (!list_empty(&dentry->d_subdirs)) {
1304 spin_unlock(&this_parent->d_lock);
1305 spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1306 this_parent = dentry;
1307 spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1308 goto repeat;
1309 }
1310 spin_unlock(&dentry->d_lock);
1311 }
1312 /*
1313 * All done at this level ... ascend and resume the search.
1314 */
1315 rcu_read_lock();
1316ascend:
1317 if (this_parent != parent) {
1318 struct dentry *child = this_parent;
1319 this_parent = child->d_parent;
1320
1321 spin_unlock(&child->d_lock);
1322 spin_lock(&this_parent->d_lock);
1323
1324 /* might go back up the wrong parent if we have had a rename. */
1325 if (need_seqretry(&rename_lock, seq))
1326 goto rename_retry;
1327 /* go into the first sibling still alive */
1328 do {
1329 next = child->d_child.next;
1330 if (next == &this_parent->d_subdirs)
1331 goto ascend;
1332 child = list_entry(next, struct dentry, d_child);
1333 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED));
1334 rcu_read_unlock();
1335 goto resume;
1336 }
1337 if (need_seqretry(&rename_lock, seq))
1338 goto rename_retry;
1339 rcu_read_unlock();
1340
1341out_unlock:
1342 spin_unlock(&this_parent->d_lock);
1343 done_seqretry(&rename_lock, seq);
1344 return;
1345
1346rename_retry:
1347 spin_unlock(&this_parent->d_lock);
1348 rcu_read_unlock();
1349 BUG_ON(seq & 1);
1350 if (!retry)
1351 return;
1352 seq = 1;
1353 goto again;
1354}
1355
1356struct check_mount {
1357 struct vfsmount *mnt;
1358 unsigned int mounted;
1359};
1360
1361static enum d_walk_ret path_check_mount(void *data, struct dentry *dentry)
1362{
1363 struct check_mount *info = data;
1364 struct path path = { .mnt = info->mnt, .dentry = dentry };
1365
1366 if (likely(!d_mountpoint(dentry)))
1367 return D_WALK_CONTINUE;
1368 if (__path_is_mountpoint(&path)) {
1369 info->mounted = 1;
1370 return D_WALK_QUIT;
1371 }
1372 return D_WALK_CONTINUE;
1373}
1374
1375/**
1376 * path_has_submounts - check for mounts over a dentry in the
1377 * current namespace.
1378 * @parent: path to check.
1379 *
1380 * Return true if the parent or its subdirectories contain
1381 * a mount point in the current namespace.
1382 */
1383int path_has_submounts(const struct path *parent)
1384{
1385 struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };
1386
1387 read_seqlock_excl(&mount_lock);
1388 d_walk(parent->dentry, &data, path_check_mount);
1389 read_sequnlock_excl(&mount_lock);
1390
1391 return data.mounted;
1392}
1393EXPORT_SYMBOL(path_has_submounts);
1394
1395/*
1396 * Called by mount code to set a mountpoint and check if the mountpoint is
1397 * reachable (e.g. NFS can unhash a directory dentry and then the complete
1398 * subtree can become unreachable).
1399 *
1400 * Only one of d_invalidate() and d_set_mounted() must succeed. For
1401 * this reason take rename_lock and d_lock on dentry and ancestors.
1402 */
1403int d_set_mounted(struct dentry *dentry)
1404{
1405 struct dentry *p;
1406 int ret = -ENOENT;
1407 write_seqlock(&rename_lock);
1408 for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1409 /* Need exclusion wrt. d_invalidate() */
1410 spin_lock(&p->d_lock);
1411 if (unlikely(d_unhashed(p))) {
1412 spin_unlock(&p->d_lock);
1413 goto out;
1414 }
1415 spin_unlock(&p->d_lock);
1416 }
1417 spin_lock(&dentry->d_lock);
1418 if (!d_unlinked(dentry)) {
1419 ret = -EBUSY;
1420 if (!d_mountpoint(dentry)) {
1421 dentry->d_flags |= DCACHE_MOUNTED;
1422 ret = 0;
1423 }
1424 }
1425 spin_unlock(&dentry->d_lock);
1426out:
1427 write_sequnlock(&rename_lock);
1428 return ret;
1429}
1430
1431/*
1432 * Search the dentry child list of the specified parent,
1433 * and move any unused dentries to the end of the unused
1434 * list for prune_dcache(). We descend to the next level
1435 * whenever the d_subdirs list is non-empty and continue
1436 * searching.
1437 *
1438 * It returns zero iff there are no unused children,
1439 * otherwise it returns the number of children moved to
1440 * the end of the unused list. This may not be the total
1441 * number of unused children, because select_parent can
1442 * drop the lock and return early due to latency
1443 * constraints.
1444 */
1445
1446struct select_data {
1447 struct dentry *start;
1448 struct list_head dispose;
1449 int found;
1450};
1451
1452static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1453{
1454 struct select_data *data = _data;
1455 enum d_walk_ret ret = D_WALK_CONTINUE;
1456
1457 if (data->start == dentry)
1458 goto out;
1459
1460 if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1461 data->found++;
1462 } else {
1463 if (dentry->d_flags & DCACHE_LRU_LIST)
1464 d_lru_del(dentry);
1465 if (!dentry->d_lockref.count) {
1466 d_shrink_add(dentry, &data->dispose);
1467 data->found++;
1468 }
1469 }
1470 /*
1471 * We can return to the caller if we have found some (this
1472 * ensures forward progress). We'll be coming back to find
1473 * the rest.
1474 */
1475 if (!list_empty(&data->dispose))
1476 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1477out:
1478 return ret;
1479}
1480
1481/**
1482 * shrink_dcache_parent - prune dcache
1483 * @parent: parent of entries to prune
1484 *
1485 * Prune the dcache to remove unused children of the parent dentry.
1486 */
1487void shrink_dcache_parent(struct dentry *parent)
1488{
1489 for (;;) {
1490 struct select_data data;
1491
1492 INIT_LIST_HEAD(&data.dispose);
1493 data.start = parent;
1494 data.found = 0;
1495
1496 d_walk(parent, &data, select_collect);
1497
1498 if (!list_empty(&data.dispose)) {
1499 shrink_dentry_list(&data.dispose);
1500 continue;
1501 }
1502
1503 cond_resched();
1504 if (!data.found)
1505 break;
1506 }
1507}
1508EXPORT_SYMBOL(shrink_dcache_parent);
1509
1510static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1511{
1512 /* it has busy descendents; complain about those instead */
1513 if (!list_empty(&dentry->d_subdirs))
1514 return D_WALK_CONTINUE;
1515
1516 /* root with refcount 1 is fine */
1517 if (dentry == _data && dentry->d_lockref.count == 1)
1518 return D_WALK_CONTINUE;
1519
1520 printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} "
1521 " still in use (%d) [unmount of %s %s]\n",
1522 dentry,
1523 dentry->d_inode ?
1524 dentry->d_inode->i_ino : 0UL,
1525 dentry,
1526 dentry->d_lockref.count,
1527 dentry->d_sb->s_type->name,
1528 dentry->d_sb->s_id);
1529 WARN_ON(1);
1530 return D_WALK_CONTINUE;
1531}
1532
1533static void do_one_tree(struct dentry *dentry)
1534{
1535 shrink_dcache_parent(dentry);
1536 d_walk(dentry, dentry, umount_check);
1537 d_drop(dentry);
1538 dput(dentry);
1539}
1540
1541/*
1542 * destroy the dentries attached to a superblock on unmounting
1543 */
1544void shrink_dcache_for_umount(struct super_block *sb)
1545{
1546 struct dentry *dentry;
1547
1548 WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
1549
1550 dentry = sb->s_root;
1551 sb->s_root = NULL;
1552 do_one_tree(dentry);
1553
1554 while (!hlist_bl_empty(&sb->s_roots)) {
1555 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_roots), struct dentry, d_hash));
1556 do_one_tree(dentry);
1557 }
1558}
1559
1560static enum d_walk_ret find_submount(void *_data, struct dentry *dentry)
1561{
1562 struct dentry **victim = _data;
1563 if (d_mountpoint(dentry)) {
1564 __dget_dlock(dentry);
1565 *victim = dentry;
1566 return D_WALK_QUIT;
1567 }
1568 return D_WALK_CONTINUE;
1569}
1570
1571/**
1572 * d_invalidate - detach submounts, prune dcache, and drop
1573 * @dentry: dentry to invalidate (aka detach, prune and drop)
1574 */
1575void d_invalidate(struct dentry *dentry)
1576{
1577 bool had_submounts = false;
1578 spin_lock(&dentry->d_lock);
1579 if (d_unhashed(dentry)) {
1580 spin_unlock(&dentry->d_lock);
1581 return;
1582 }
1583 __d_drop(dentry);
1584 spin_unlock(&dentry->d_lock);
1585
1586 /* Negative dentries can be dropped without further checks */
1587 if (!dentry->d_inode)
1588 return;
1589
1590 shrink_dcache_parent(dentry);
1591 for (;;) {
1592 struct dentry *victim = NULL;
1593 d_walk(dentry, &victim, find_submount);
1594 if (!victim) {
1595 if (had_submounts)
1596 shrink_dcache_parent(dentry);
1597 return;
1598 }
1599 had_submounts = true;
1600 detach_mounts(victim);
1601 dput(victim);
1602 }
1603}
1604EXPORT_SYMBOL(d_invalidate);
1605
1606/**
1607 * __d_alloc - allocate a dcache entry
1608 * @sb: filesystem it will belong to
1609 * @name: qstr of the name
1610 *
1611 * Allocates a dentry. It returns %NULL if there is insufficient memory
1612 * available. On a success the dentry is returned. The name passed in is
1613 * copied and the copy passed in may be reused after this call.
1614 */
1615
1616struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1617{
1618 struct dentry *dentry;
1619 char *dname;
1620 int err;
1621
1622 dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1623 if (!dentry)
1624 return NULL;
1625
1626 /*
1627 * We guarantee that the inline name is always NUL-terminated.
1628 * This way the memcpy() done by the name switching in rename
1629 * will still always have a NUL at the end, even if we might
1630 * be overwriting an internal NUL character
1631 */
1632 dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1633 if (unlikely(!name)) {
1634 name = &slash_name;
1635 dname = dentry->d_iname;
1636 } else if (name->len > DNAME_INLINE_LEN-1) {
1637 size_t size = offsetof(struct external_name, name[1]);
1638 struct external_name *p = kmalloc(size + name->len,
1639 GFP_KERNEL_ACCOUNT |
1640 __GFP_RECLAIMABLE);
1641 if (!p) {
1642 kmem_cache_free(dentry_cache, dentry);
1643 return NULL;
1644 }
1645 atomic_set(&p->u.count, 1);
1646 dname = p->name;
1647 } else {
1648 dname = dentry->d_iname;
1649 }
1650
1651 dentry->d_name.len = name->len;
1652 dentry->d_name.hash = name->hash;
1653 memcpy(dname, name->name, name->len);
1654 dname[name->len] = 0;
1655
1656 /* Make sure we always see the terminating NUL character */
1657 smp_store_release(&dentry->d_name.name, dname); /* ^^^ */
1658
1659 dentry->d_lockref.count = 1;
1660 dentry->d_flags = 0;
1661 spin_lock_init(&dentry->d_lock);
1662 seqcount_init(&dentry->d_seq);
1663 dentry->d_inode = NULL;
1664 dentry->d_parent = dentry;
1665 dentry->d_sb = sb;
1666 dentry->d_op = NULL;
1667 dentry->d_fsdata = NULL;
1668 INIT_HLIST_BL_NODE(&dentry->d_hash);
1669 INIT_LIST_HEAD(&dentry->d_lru);
1670 INIT_LIST_HEAD(&dentry->d_subdirs);
1671 INIT_HLIST_NODE(&dentry->d_u.d_alias);
1672 INIT_LIST_HEAD(&dentry->d_child);
1673 d_set_d_op(dentry, dentry->d_sb->s_d_op);
1674
1675 if (dentry->d_op && dentry->d_op->d_init) {
1676 err = dentry->d_op->d_init(dentry);
1677 if (err) {
1678 if (dname_external(dentry))
1679 kfree(external_name(dentry));
1680 kmem_cache_free(dentry_cache, dentry);
1681 return NULL;
1682 }
1683 }
1684
1685 this_cpu_inc(nr_dentry);
1686
1687 return dentry;
1688}
1689
1690/**
1691 * d_alloc - allocate a dcache entry
1692 * @parent: parent of entry to allocate
1693 * @name: qstr of the name
1694 *
1695 * Allocates a dentry. It returns %NULL if there is insufficient memory
1696 * available. On a success the dentry is returned. The name passed in is
1697 * copied and the copy passed in may be reused after this call.
1698 */
1699struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1700{
1701 struct dentry *dentry = __d_alloc(parent->d_sb, name);
1702 if (!dentry)
1703 return NULL;
1704 dentry->d_flags |= DCACHE_RCUACCESS;
1705 spin_lock(&parent->d_lock);
1706 /*
1707 * don't need child lock because it is not subject
1708 * to concurrency here
1709 */
1710 __dget_dlock(parent);
1711 dentry->d_parent = parent;
1712 list_add(&dentry->d_child, &parent->d_subdirs);
1713 spin_unlock(&parent->d_lock);
1714
1715 return dentry;
1716}
1717EXPORT_SYMBOL(d_alloc);
1718
1719struct dentry *d_alloc_anon(struct super_block *sb)
1720{
1721 return __d_alloc(sb, NULL);
1722}
1723EXPORT_SYMBOL(d_alloc_anon);
1724
1725struct dentry *d_alloc_cursor(struct dentry * parent)
1726{
1727 struct dentry *dentry = d_alloc_anon(parent->d_sb);
1728 if (dentry) {
1729 dentry->d_flags |= DCACHE_RCUACCESS | DCACHE_DENTRY_CURSOR;
1730 dentry->d_parent = dget(parent);
1731 }
1732 return dentry;
1733}
1734
1735/**
1736 * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1737 * @sb: the superblock
1738 * @name: qstr of the name
1739 *
1740 * For a filesystem that just pins its dentries in memory and never
1741 * performs lookups at all, return an unhashed IS_ROOT dentry.
1742 */
1743struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1744{
1745 return __d_alloc(sb, name);
1746}
1747EXPORT_SYMBOL(d_alloc_pseudo);
1748
1749struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1750{
1751 struct qstr q;
1752
1753 q.name = name;
1754 q.hash_len = hashlen_string(parent, name);
1755 return d_alloc(parent, &q);
1756}
1757EXPORT_SYMBOL(d_alloc_name);
1758
1759void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1760{
1761 WARN_ON_ONCE(dentry->d_op);
1762 WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH |
1763 DCACHE_OP_COMPARE |
1764 DCACHE_OP_REVALIDATE |
1765 DCACHE_OP_WEAK_REVALIDATE |
1766 DCACHE_OP_DELETE |
1767 DCACHE_OP_REAL));
1768 dentry->d_op = op;
1769 if (!op)
1770 return;
1771 if (op->d_hash)
1772 dentry->d_flags |= DCACHE_OP_HASH;
1773 if (op->d_compare)
1774 dentry->d_flags |= DCACHE_OP_COMPARE;
1775 if (op->d_revalidate)
1776 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1777 if (op->d_weak_revalidate)
1778 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1779 if (op->d_delete)
1780 dentry->d_flags |= DCACHE_OP_DELETE;
1781 if (op->d_prune)
1782 dentry->d_flags |= DCACHE_OP_PRUNE;
1783 if (op->d_real)
1784 dentry->d_flags |= DCACHE_OP_REAL;
1785
1786}
1787EXPORT_SYMBOL(d_set_d_op);
1788
1789
1790/*
1791 * d_set_fallthru - Mark a dentry as falling through to a lower layer
1792 * @dentry - The dentry to mark
1793 *
1794 * Mark a dentry as falling through to the lower layer (as set with
1795 * d_pin_lower()). This flag may be recorded on the medium.
1796 */
1797void d_set_fallthru(struct dentry *dentry)
1798{
1799 spin_lock(&dentry->d_lock);
1800 dentry->d_flags |= DCACHE_FALLTHRU;
1801 spin_unlock(&dentry->d_lock);
1802}
1803EXPORT_SYMBOL(d_set_fallthru);
1804
1805static unsigned d_flags_for_inode(struct inode *inode)
1806{
1807 unsigned add_flags = DCACHE_REGULAR_TYPE;
1808
1809 if (!inode)
1810 return DCACHE_MISS_TYPE;
1811
1812 if (S_ISDIR(inode->i_mode)) {
1813 add_flags = DCACHE_DIRECTORY_TYPE;
1814 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1815 if (unlikely(!inode->i_op->lookup))
1816 add_flags = DCACHE_AUTODIR_TYPE;
1817 else
1818 inode->i_opflags |= IOP_LOOKUP;
1819 }
1820 goto type_determined;
1821 }
1822
1823 if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1824 if (unlikely(inode->i_op->get_link)) {
1825 add_flags = DCACHE_SYMLINK_TYPE;
1826 goto type_determined;
1827 }
1828 inode->i_opflags |= IOP_NOFOLLOW;
1829 }
1830
1831 if (unlikely(!S_ISREG(inode->i_mode)))
1832 add_flags = DCACHE_SPECIAL_TYPE;
1833
1834type_determined:
1835 if (unlikely(IS_AUTOMOUNT(inode)))
1836 add_flags |= DCACHE_NEED_AUTOMOUNT;
1837 return add_flags;
1838}
1839
1840static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1841{
1842 unsigned add_flags = d_flags_for_inode(inode);
1843 WARN_ON(d_in_lookup(dentry));
1844
1845 spin_lock(&dentry->d_lock);
1846 /*
1847 * Decrement negative dentry count if it was in the LRU list.
1848 */
1849 if (dentry->d_flags & DCACHE_LRU_LIST)
1850 this_cpu_dec(nr_dentry_negative);
1851 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1852 raw_write_seqcount_begin(&dentry->d_seq);
1853 __d_set_inode_and_type(dentry, inode, add_flags);
1854 raw_write_seqcount_end(&dentry->d_seq);
1855 fsnotify_update_flags(dentry);
1856 spin_unlock(&dentry->d_lock);
1857}
1858
1859/**
1860 * d_instantiate - fill in inode information for a dentry
1861 * @entry: dentry to complete
1862 * @inode: inode to attach to this dentry
1863 *
1864 * Fill in inode information in the entry.
1865 *
1866 * This turns negative dentries into productive full members
1867 * of society.
1868 *
1869 * NOTE! This assumes that the inode count has been incremented
1870 * (or otherwise set) by the caller to indicate that it is now
1871 * in use by the dcache.
1872 */
1873
1874void d_instantiate(struct dentry *entry, struct inode * inode)
1875{
1876 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1877 if (inode) {
1878 security_d_instantiate(entry, inode);
1879 spin_lock(&inode->i_lock);
1880 __d_instantiate(entry, inode);
1881 spin_unlock(&inode->i_lock);
1882 }
1883}
1884EXPORT_SYMBOL(d_instantiate);
1885
1886/*
1887 * This should be equivalent to d_instantiate() + unlock_new_inode(),
1888 * with lockdep-related part of unlock_new_inode() done before
1889 * anything else. Use that instead of open-coding d_instantiate()/
1890 * unlock_new_inode() combinations.
1891 */
1892void d_instantiate_new(struct dentry *entry, struct inode *inode)
1893{
1894 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1895 BUG_ON(!inode);
1896 lockdep_annotate_inode_mutex_key(inode);
1897 security_d_instantiate(entry, inode);
1898 spin_lock(&inode->i_lock);
1899 __d_instantiate(entry, inode);
1900 WARN_ON(!(inode->i_state & I_NEW));
1901 inode->i_state &= ~I_NEW & ~I_CREATING;
1902 smp_mb();
1903 wake_up_bit(&inode->i_state, __I_NEW);
1904 spin_unlock(&inode->i_lock);
1905}
1906EXPORT_SYMBOL(d_instantiate_new);
1907
1908struct dentry *d_make_root(struct inode *root_inode)
1909{
1910 struct dentry *res = NULL;
1911
1912 if (root_inode) {
1913 res = d_alloc_anon(root_inode->i_sb);
1914 if (res) {
1915 res->d_flags |= DCACHE_RCUACCESS;
1916 d_instantiate(res, root_inode);
1917 } else {
1918 iput(root_inode);
1919 }
1920 }
1921 return res;
1922}
1923EXPORT_SYMBOL(d_make_root);
1924
1925static struct dentry *__d_instantiate_anon(struct dentry *dentry,
1926 struct inode *inode,
1927 bool disconnected)
1928{
1929 struct dentry *res;
1930 unsigned add_flags;
1931
1932 security_d_instantiate(dentry, inode);
1933 spin_lock(&inode->i_lock);
1934 res = __d_find_any_alias(inode);
1935 if (res) {
1936 spin_unlock(&inode->i_lock);
1937 dput(dentry);
1938 goto out_iput;
1939 }
1940
1941 /* attach a disconnected dentry */
1942 add_flags = d_flags_for_inode(inode);
1943
1944 if (disconnected)
1945 add_flags |= DCACHE_DISCONNECTED;
1946
1947 spin_lock(&dentry->d_lock);
1948 __d_set_inode_and_type(dentry, inode, add_flags);
1949 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1950 if (!disconnected) {
1951 hlist_bl_lock(&dentry->d_sb->s_roots);
1952 hlist_bl_add_head(&dentry->d_hash, &dentry->d_sb->s_roots);
1953 hlist_bl_unlock(&dentry->d_sb->s_roots);
1954 }
1955 spin_unlock(&dentry->d_lock);
1956 spin_unlock(&inode->i_lock);
1957
1958 return dentry;
1959
1960 out_iput:
1961 iput(inode);
1962 return res;
1963}
1964
1965struct dentry *d_instantiate_anon(struct dentry *dentry, struct inode *inode)
1966{
1967 return __d_instantiate_anon(dentry, inode, true);
1968}
1969EXPORT_SYMBOL(d_instantiate_anon);
1970
1971static struct dentry *__d_obtain_alias(struct inode *inode, bool disconnected)
1972{
1973 struct dentry *tmp;
1974 struct dentry *res;
1975
1976 if (!inode)
1977 return ERR_PTR(-ESTALE);
1978 if (IS_ERR(inode))
1979 return ERR_CAST(inode);
1980
1981 res = d_find_any_alias(inode);
1982 if (res)
1983 goto out_iput;
1984
1985 tmp = d_alloc_anon(inode->i_sb);
1986 if (!tmp) {
1987 res = ERR_PTR(-ENOMEM);
1988 goto out_iput;
1989 }
1990
1991 return __d_instantiate_anon(tmp, inode, disconnected);
1992
1993out_iput:
1994 iput(inode);
1995 return res;
1996}
1997
1998/**
1999 * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
2000 * @inode: inode to allocate the dentry for
2001 *
2002 * Obtain a dentry for an inode resulting from NFS filehandle conversion or
2003 * similar open by handle operations. The returned dentry may be anonymous,
2004 * or may have a full name (if the inode was already in the cache).
2005 *
2006 * When called on a directory inode, we must ensure that the inode only ever
2007 * has one dentry. If a dentry is found, that is returned instead of
2008 * allocating a new one.
2009 *
2010 * On successful return, the reference to the inode has been transferred
2011 * to the dentry. In case of an error the reference on the inode is released.
2012 * To make it easier to use in export operations a %NULL or IS_ERR inode may
2013 * be passed in and the error will be propagated to the return value,
2014 * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
2015 */
2016struct dentry *d_obtain_alias(struct inode *inode)
2017{
2018 return __d_obtain_alias(inode, true);
2019}
2020EXPORT_SYMBOL(d_obtain_alias);
2021
2022/**
2023 * d_obtain_root - find or allocate a dentry for a given inode
2024 * @inode: inode to allocate the dentry for
2025 *
2026 * Obtain an IS_ROOT dentry for the root of a filesystem.
2027 *
2028 * We must ensure that directory inodes only ever have one dentry. If a
2029 * dentry is found, that is returned instead of allocating a new one.
2030 *
2031 * On successful return, the reference to the inode has been transferred
2032 * to the dentry. In case of an error the reference on the inode is
2033 * released. A %NULL or IS_ERR inode may be passed in and will be the
2034 * error will be propagate to the return value, with a %NULL @inode
2035 * replaced by ERR_PTR(-ESTALE).
2036 */
2037struct dentry *d_obtain_root(struct inode *inode)
2038{
2039 return __d_obtain_alias(inode, false);
2040}
2041EXPORT_SYMBOL(d_obtain_root);
2042
2043/**
2044 * d_add_ci - lookup or allocate new dentry with case-exact name
2045 * @inode: the inode case-insensitive lookup has found
2046 * @dentry: the negative dentry that was passed to the parent's lookup func
2047 * @name: the case-exact name to be associated with the returned dentry
2048 *
2049 * This is to avoid filling the dcache with case-insensitive names to the
2050 * same inode, only the actual correct case is stored in the dcache for
2051 * case-insensitive filesystems.
2052 *
2053 * For a case-insensitive lookup match and if the the case-exact dentry
2054 * already exists in in the dcache, use it and return it.
2055 *
2056 * If no entry exists with the exact case name, allocate new dentry with
2057 * the exact case, and return the spliced entry.
2058 */
2059struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2060 struct qstr *name)
2061{
2062 struct dentry *found, *res;
2063
2064 /*
2065 * First check if a dentry matching the name already exists,
2066 * if not go ahead and create it now.
2067 */
2068 found = d_hash_and_lookup(dentry->d_parent, name);
2069 if (found) {
2070 iput(inode);
2071 return found;
2072 }
2073 if (d_in_lookup(dentry)) {
2074 found = d_alloc_parallel(dentry->d_parent, name,
2075 dentry->d_wait);
2076 if (IS_ERR(found) || !d_in_lookup(found)) {
2077 iput(inode);
2078 return found;
2079 }
2080 } else {
2081 found = d_alloc(dentry->d_parent, name);
2082 if (!found) {
2083 iput(inode);
2084 return ERR_PTR(-ENOMEM);
2085 }
2086 }
2087 res = d_splice_alias(inode, found);
2088 if (res) {
2089 dput(found);
2090 return res;
2091 }
2092 return found;
2093}
2094EXPORT_SYMBOL(d_add_ci);
2095
2096
2097static inline bool d_same_name(const struct dentry *dentry,
2098 const struct dentry *parent,
2099 const struct qstr *name)
2100{
2101 if (likely(!(parent->d_flags & DCACHE_OP_COMPARE))) {
2102 if (dentry->d_name.len != name->len)
2103 return false;
2104 return dentry_cmp(dentry, name->name, name->len) == 0;
2105 }
2106 return parent->d_op->d_compare(dentry,
2107 dentry->d_name.len, dentry->d_name.name,
2108 name) == 0;
2109}
2110
2111/**
2112 * __d_lookup_rcu - search for a dentry (racy, store-free)
2113 * @parent: parent dentry
2114 * @name: qstr of name we wish to find
2115 * @seqp: returns d_seq value at the point where the dentry was found
2116 * Returns: dentry, or NULL
2117 *
2118 * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2119 * resolution (store-free path walking) design described in
2120 * Documentation/filesystems/path-lookup.txt.
2121 *
2122 * This is not to be used outside core vfs.
2123 *
2124 * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2125 * held, and rcu_read_lock held. The returned dentry must not be stored into
2126 * without taking d_lock and checking d_seq sequence count against @seq
2127 * returned here.
2128 *
2129 * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2130 * function.
2131 *
2132 * Alternatively, __d_lookup_rcu may be called again to look up the child of
2133 * the returned dentry, so long as its parent's seqlock is checked after the
2134 * child is looked up. Thus, an interlocking stepping of sequence lock checks
2135 * is formed, giving integrity down the path walk.
2136 *
2137 * NOTE! The caller *has* to check the resulting dentry against the sequence
2138 * number we've returned before using any of the resulting dentry state!
2139 */
2140struct dentry *__d_lookup_rcu(const struct dentry *parent,
2141 const struct qstr *name,
2142 unsigned *seqp)
2143{
2144 u64 hashlen = name->hash_len;
2145 const unsigned char *str = name->name;
2146 struct hlist_bl_head *b = d_hash(hashlen_hash(hashlen));
2147 struct hlist_bl_node *node;
2148 struct dentry *dentry;
2149
2150 /*
2151 * Note: There is significant duplication with __d_lookup_rcu which is
2152 * required to prevent single threaded performance regressions
2153 * especially on architectures where smp_rmb (in seqcounts) are costly.
2154 * Keep the two functions in sync.
2155 */
2156
2157 /*
2158 * The hash list is protected using RCU.
2159 *
2160 * Carefully use d_seq when comparing a candidate dentry, to avoid
2161 * races with d_move().
2162 *
2163 * It is possible that concurrent renames can mess up our list
2164 * walk here and result in missing our dentry, resulting in the
2165 * false-negative result. d_lookup() protects against concurrent
2166 * renames using rename_lock seqlock.
2167 *
2168 * See Documentation/filesystems/path-lookup.txt for more details.
2169 */
2170 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2171 unsigned seq;
2172
2173seqretry:
2174 /*
2175 * The dentry sequence count protects us from concurrent
2176 * renames, and thus protects parent and name fields.
2177 *
2178 * The caller must perform a seqcount check in order
2179 * to do anything useful with the returned dentry.
2180 *
2181 * NOTE! We do a "raw" seqcount_begin here. That means that
2182 * we don't wait for the sequence count to stabilize if it
2183 * is in the middle of a sequence change. If we do the slow
2184 * dentry compare, we will do seqretries until it is stable,
2185 * and if we end up with a successful lookup, we actually
2186 * want to exit RCU lookup anyway.
2187 *
2188 * Note that raw_seqcount_begin still *does* smp_rmb(), so
2189 * we are still guaranteed NUL-termination of ->d_name.name.
2190 */
2191 seq = raw_seqcount_begin(&dentry->d_seq);
2192 if (dentry->d_parent != parent)
2193 continue;
2194 if (d_unhashed(dentry))
2195 continue;
2196
2197 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) {
2198 int tlen;
2199 const char *tname;
2200 if (dentry->d_name.hash != hashlen_hash(hashlen))
2201 continue;
2202 tlen = dentry->d_name.len;
2203 tname = dentry->d_name.name;
2204 /* we want a consistent (name,len) pair */
2205 if (read_seqcount_retry(&dentry->d_seq, seq)) {
2206 cpu_relax();
2207 goto seqretry;
2208 }
2209 if (parent->d_op->d_compare(dentry,
2210 tlen, tname, name) != 0)
2211 continue;
2212 } else {
2213 if (dentry->d_name.hash_len != hashlen)
2214 continue;
2215 if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
2216 continue;
2217 }
2218 *seqp = seq;
2219 return dentry;
2220 }
2221 return NULL;
2222}
2223
2224/**
2225 * d_lookup - search for a dentry
2226 * @parent: parent dentry
2227 * @name: qstr of name we wish to find
2228 * Returns: dentry, or NULL
2229 *
2230 * d_lookup searches the children of the parent dentry for the name in
2231 * question. If the dentry is found its reference count is incremented and the
2232 * dentry is returned. The caller must use dput to free the entry when it has
2233 * finished using it. %NULL is returned if the dentry does not exist.
2234 */
2235struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2236{
2237 struct dentry *dentry;
2238 unsigned seq;
2239
2240 do {
2241 seq = read_seqbegin(&rename_lock);
2242 dentry = __d_lookup(parent, name);
2243 if (dentry)
2244 break;
2245 } while (read_seqretry(&rename_lock, seq));
2246 return dentry;
2247}
2248EXPORT_SYMBOL(d_lookup);
2249
2250/**
2251 * __d_lookup - search for a dentry (racy)
2252 * @parent: parent dentry
2253 * @name: qstr of name we wish to find
2254 * Returns: dentry, or NULL
2255 *
2256 * __d_lookup is like d_lookup, however it may (rarely) return a
2257 * false-negative result due to unrelated rename activity.
2258 *
2259 * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2260 * however it must be used carefully, eg. with a following d_lookup in
2261 * the case of failure.
2262 *
2263 * __d_lookup callers must be commented.
2264 */
2265struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2266{
2267 unsigned int hash = name->hash;
2268 struct hlist_bl_head *b = d_hash(hash);
2269 struct hlist_bl_node *node;
2270 struct dentry *found = NULL;
2271 struct dentry *dentry;
2272
2273 /*
2274 * Note: There is significant duplication with __d_lookup_rcu which is
2275 * required to prevent single threaded performance regressions
2276 * especially on architectures where smp_rmb (in seqcounts) are costly.
2277 * Keep the two functions in sync.
2278 */
2279
2280 /*
2281 * The hash list is protected using RCU.
2282 *
2283 * Take d_lock when comparing a candidate dentry, to avoid races
2284 * with d_move().
2285 *
2286 * It is possible that concurrent renames can mess up our list
2287 * walk here and result in missing our dentry, resulting in the
2288 * false-negative result. d_lookup() protects against concurrent
2289 * renames using rename_lock seqlock.
2290 *
2291 * See Documentation/filesystems/path-lookup.txt for more details.
2292 */
2293 rcu_read_lock();
2294
2295 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2296
2297 if (dentry->d_name.hash != hash)
2298 continue;
2299
2300 spin_lock(&dentry->d_lock);
2301 if (dentry->d_parent != parent)
2302 goto next;
2303 if (d_unhashed(dentry))
2304 goto next;
2305
2306 if (!d_same_name(dentry, parent, name))
2307 goto next;
2308
2309 dentry->d_lockref.count++;
2310 found = dentry;
2311 spin_unlock(&dentry->d_lock);
2312 break;
2313next:
2314 spin_unlock(&dentry->d_lock);
2315 }
2316 rcu_read_unlock();
2317
2318 return found;
2319}
2320
2321/**
2322 * d_hash_and_lookup - hash the qstr then search for a dentry
2323 * @dir: Directory to search in
2324 * @name: qstr of name we wish to find
2325 *
2326 * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2327 */
2328struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2329{
2330 /*
2331 * Check for a fs-specific hash function. Note that we must
2332 * calculate the standard hash first, as the d_op->d_hash()
2333 * routine may choose to leave the hash value unchanged.
2334 */
2335 name->hash = full_name_hash(dir, name->name, name->len);
2336 if (dir->d_flags & DCACHE_OP_HASH) {
2337 int err = dir->d_op->d_hash(dir, name);
2338 if (unlikely(err < 0))
2339 return ERR_PTR(err);
2340 }
2341 return d_lookup(dir, name);
2342}
2343EXPORT_SYMBOL(d_hash_and_lookup);
2344
2345/*
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.
2349 *
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
2356 */
2357
2358/**
2359 * d_delete - delete a dentry
2360 * @dentry: The dentry to delete
2361 *
2362 * Turn the dentry into a negative dentry if possible, otherwise
2363 * remove it from the hash queues so it can be deleted later
2364 */
2365
2366void d_delete(struct dentry * dentry)
2367{
2368 struct inode *inode = dentry->d_inode;
2369 int isdir = d_is_dir(dentry);
2370
2371 spin_lock(&inode->i_lock);
2372 spin_lock(&dentry->d_lock);
2373 /*
2374 * Are we the only user?
2375 */
2376 if (dentry->d_lockref.count == 1) {
2377 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2378 dentry_unlink_inode(dentry);
2379 } else {
2380 __d_drop(dentry);
2381 spin_unlock(&dentry->d_lock);
2382 spin_unlock(&inode->i_lock);
2383 }
2384 fsnotify_nameremove(dentry, isdir);
2385}
2386EXPORT_SYMBOL(d_delete);
2387
2388static void __d_rehash(struct dentry *entry)
2389{
2390 struct hlist_bl_head *b = d_hash(entry->d_name.hash);
2391
2392 hlist_bl_lock(b);
2393 hlist_bl_add_head_rcu(&entry->d_hash, b);
2394 hlist_bl_unlock(b);
2395}
2396
2397/**
2398 * d_rehash - add an entry back to the hash
2399 * @entry: dentry to add to the hash
2400 *
2401 * Adds a dentry to the hash according to its name.
2402 */
2403
2404void d_rehash(struct dentry * entry)
2405{
2406 spin_lock(&entry->d_lock);
2407 __d_rehash(entry);
2408 spin_unlock(&entry->d_lock);
2409}
2410EXPORT_SYMBOL(d_rehash);
2411
2412static inline unsigned start_dir_add(struct inode *dir)
2413{
2414
2415 for (;;) {
2416 unsigned n = dir->i_dir_seq;
2417 if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
2418 return n;
2419 cpu_relax();
2420 }
2421}
2422
2423static inline void end_dir_add(struct inode *dir, unsigned n)
2424{
2425 smp_store_release(&dir->i_dir_seq, n + 2);
2426}
2427
2428static void d_wait_lookup(struct dentry *dentry)
2429{
2430 if (d_in_lookup(dentry)) {
2431 DECLARE_WAITQUEUE(wait, current);
2432 add_wait_queue(dentry->d_wait, &wait);
2433 do {
2434 set_current_state(TASK_UNINTERRUPTIBLE);
2435 spin_unlock(&dentry->d_lock);
2436 schedule();
2437 spin_lock(&dentry->d_lock);
2438 } while (d_in_lookup(dentry));
2439 }
2440}
2441
2442struct dentry *d_alloc_parallel(struct dentry *parent,
2443 const struct qstr *name,
2444 wait_queue_head_t *wq)
2445{
2446 unsigned int hash = name->hash;
2447 struct hlist_bl_head *b = in_lookup_hash(parent, hash);
2448 struct hlist_bl_node *node;
2449 struct dentry *new = d_alloc(parent, name);
2450 struct dentry *dentry;
2451 unsigned seq, r_seq, d_seq;
2452
2453 if (unlikely(!new))
2454 return ERR_PTR(-ENOMEM);
2455
2456retry:
2457 rcu_read_lock();
2458 seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
2459 r_seq = read_seqbegin(&rename_lock);
2460 dentry = __d_lookup_rcu(parent, name, &d_seq);
2461 if (unlikely(dentry)) {
2462 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2463 rcu_read_unlock();
2464 goto retry;
2465 }
2466 if (read_seqcount_retry(&dentry->d_seq, d_seq)) {
2467 rcu_read_unlock();
2468 dput(dentry);
2469 goto retry;
2470 }
2471 rcu_read_unlock();
2472 dput(new);
2473 return dentry;
2474 }
2475 if (unlikely(read_seqretry(&rename_lock, r_seq))) {
2476 rcu_read_unlock();
2477 goto retry;
2478 }
2479
2480 if (unlikely(seq & 1)) {
2481 rcu_read_unlock();
2482 goto retry;
2483 }
2484
2485 hlist_bl_lock(b);
2486 if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {
2487 hlist_bl_unlock(b);
2488 rcu_read_unlock();
2489 goto retry;
2490 }
2491 /*
2492 * No changes for the parent since the beginning of d_lookup().
2493 * Since all removals from the chain happen with hlist_bl_lock(),
2494 * any potential in-lookup matches are going to stay here until
2495 * we unlock the chain. All fields are stable in everything
2496 * we encounter.
2497 */
2498 hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
2499 if (dentry->d_name.hash != hash)
2500 continue;
2501 if (dentry->d_parent != parent)
2502 continue;
2503 if (!d_same_name(dentry, parent, name))
2504 continue;
2505 hlist_bl_unlock(b);
2506 /* now we can try to grab a reference */
2507 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2508 rcu_read_unlock();
2509 goto retry;
2510 }
2511
2512 rcu_read_unlock();
2513 /*
2514 * somebody is likely to be still doing lookup for it;
2515 * wait for them to finish
2516 */
2517 spin_lock(&dentry->d_lock);
2518 d_wait_lookup(dentry);
2519 /*
2520 * it's not in-lookup anymore; in principle we should repeat
2521 * everything from dcache lookup, but it's likely to be what
2522 * d_lookup() would've found anyway. If it is, just return it;
2523 * otherwise we really have to repeat the whole thing.
2524 */
2525 if (unlikely(dentry->d_name.hash != hash))
2526 goto mismatch;
2527 if (unlikely(dentry->d_parent != parent))
2528 goto mismatch;
2529 if (unlikely(d_unhashed(dentry)))
2530 goto mismatch;
2531 if (unlikely(!d_same_name(dentry, parent, name)))
2532 goto mismatch;
2533 /* OK, it *is* a hashed match; return it */
2534 spin_unlock(&dentry->d_lock);
2535 dput(new);
2536 return dentry;
2537 }
2538 rcu_read_unlock();
2539 /* we can't take ->d_lock here; it's OK, though. */
2540 new->d_flags |= DCACHE_PAR_LOOKUP;
2541 new->d_wait = wq;
2542 hlist_bl_add_head_rcu(&new->d_u.d_in_lookup_hash, b);
2543 hlist_bl_unlock(b);
2544 return new;
2545mismatch:
2546 spin_unlock(&dentry->d_lock);
2547 dput(dentry);
2548 goto retry;
2549}
2550EXPORT_SYMBOL(d_alloc_parallel);
2551
2552void __d_lookup_done(struct dentry *dentry)
2553{
2554 struct hlist_bl_head *b = in_lookup_hash(dentry->d_parent,
2555 dentry->d_name.hash);
2556 hlist_bl_lock(b);
2557 dentry->d_flags &= ~DCACHE_PAR_LOOKUP;
2558 __hlist_bl_del(&dentry->d_u.d_in_lookup_hash);
2559 wake_up_all(dentry->d_wait);
2560 dentry->d_wait = NULL;
2561 hlist_bl_unlock(b);
2562 INIT_HLIST_NODE(&dentry->d_u.d_alias);
2563 INIT_LIST_HEAD(&dentry->d_lru);
2564}
2565EXPORT_SYMBOL(__d_lookup_done);
2566
2567/* inode->i_lock held if inode is non-NULL */
2568
2569static inline void __d_add(struct dentry *dentry, struct inode *inode)
2570{
2571 struct inode *dir = NULL;
2572 unsigned n;
2573 spin_lock(&dentry->d_lock);
2574 if (unlikely(d_in_lookup(dentry))) {
2575 dir = dentry->d_parent->d_inode;
2576 n = start_dir_add(dir);
2577 __d_lookup_done(dentry);
2578 }
2579 if (inode) {
2580 unsigned add_flags = d_flags_for_inode(inode);
2581 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2582 raw_write_seqcount_begin(&dentry->d_seq);
2583 __d_set_inode_and_type(dentry, inode, add_flags);
2584 raw_write_seqcount_end(&dentry->d_seq);
2585 fsnotify_update_flags(dentry);
2586 }
2587 __d_rehash(dentry);
2588 if (dir)
2589 end_dir_add(dir, n);
2590 spin_unlock(&dentry->d_lock);
2591 if (inode)
2592 spin_unlock(&inode->i_lock);
2593}
2594
2595/**
2596 * d_add - add dentry to hash queues
2597 * @entry: dentry to add
2598 * @inode: The inode to attach to this dentry
2599 *
2600 * This adds the entry to the hash queues and initializes @inode.
2601 * The entry was actually filled in earlier during d_alloc().
2602 */
2603
2604void d_add(struct dentry *entry, struct inode *inode)
2605{
2606 if (inode) {
2607 security_d_instantiate(entry, inode);
2608 spin_lock(&inode->i_lock);
2609 }
2610 __d_add(entry, inode);
2611}
2612EXPORT_SYMBOL(d_add);
2613
2614/**
2615 * d_exact_alias - find and hash an exact unhashed alias
2616 * @entry: dentry to add
2617 * @inode: The inode to go with this dentry
2618 *
2619 * If an unhashed dentry with the same name/parent and desired
2620 * inode already exists, hash and return it. Otherwise, return
2621 * NULL.
2622 *
2623 * Parent directory should be locked.
2624 */
2625struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2626{
2627 struct dentry *alias;
2628 unsigned int hash = entry->d_name.hash;
2629
2630 spin_lock(&inode->i_lock);
2631 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2632 /*
2633 * Don't need alias->d_lock here, because aliases with
2634 * d_parent == entry->d_parent are not subject to name or
2635 * parent changes, because the parent inode i_mutex is held.
2636 */
2637 if (alias->d_name.hash != hash)
2638 continue;
2639 if (alias->d_parent != entry->d_parent)
2640 continue;
2641 if (!d_same_name(alias, entry->d_parent, &entry->d_name))
2642 continue;
2643 spin_lock(&alias->d_lock);
2644 if (!d_unhashed(alias)) {
2645 spin_unlock(&alias->d_lock);
2646 alias = NULL;
2647 } else {
2648 __dget_dlock(alias);
2649 __d_rehash(alias);
2650 spin_unlock(&alias->d_lock);
2651 }
2652 spin_unlock(&inode->i_lock);
2653 return alias;
2654 }
2655 spin_unlock(&inode->i_lock);
2656 return NULL;
2657}
2658EXPORT_SYMBOL(d_exact_alias);
2659
2660static void swap_names(struct dentry *dentry, struct dentry *target)
2661{
2662 if (unlikely(dname_external(target))) {
2663 if (unlikely(dname_external(dentry))) {
2664 /*
2665 * Both external: swap the pointers
2666 */
2667 swap(target->d_name.name, dentry->d_name.name);
2668 } else {
2669 /*
2670 * dentry:internal, target:external. Steal target's
2671 * storage and make target internal.
2672 */
2673 memcpy(target->d_iname, dentry->d_name.name,
2674 dentry->d_name.len + 1);
2675 dentry->d_name.name = target->d_name.name;
2676 target->d_name.name = target->d_iname;
2677 }
2678 } else {
2679 if (unlikely(dname_external(dentry))) {
2680 /*
2681 * dentry:external, target:internal. Give dentry's
2682 * storage to target and make dentry internal
2683 */
2684 memcpy(dentry->d_iname, target->d_name.name,
2685 target->d_name.len + 1);
2686 target->d_name.name = dentry->d_name.name;
2687 dentry->d_name.name = dentry->d_iname;
2688 } else {
2689 /*
2690 * Both are internal.
2691 */
2692 unsigned int i;
2693 BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2694 for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2695 swap(((long *) &dentry->d_iname)[i],
2696 ((long *) &target->d_iname)[i]);
2697 }
2698 }
2699 }
2700 swap(dentry->d_name.hash_len, target->d_name.hash_len);
2701}
2702
2703static void copy_name(struct dentry *dentry, struct dentry *target)
2704{
2705 struct external_name *old_name = NULL;
2706 if (unlikely(dname_external(dentry)))
2707 old_name = external_name(dentry);
2708 if (unlikely(dname_external(target))) {
2709 atomic_inc(&external_name(target)->u.count);
2710 dentry->d_name = target->d_name;
2711 } else {
2712 memcpy(dentry->d_iname, target->d_name.name,
2713 target->d_name.len + 1);
2714 dentry->d_name.name = dentry->d_iname;
2715 dentry->d_name.hash_len = target->d_name.hash_len;
2716 }
2717 if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2718 kfree_rcu(old_name, u.head);
2719}
2720
2721/*
2722 * __d_move - move a dentry
2723 * @dentry: entry to move
2724 * @target: new dentry
2725 * @exchange: exchange the two dentries
2726 *
2727 * Update the dcache to reflect the move of a file name. Negative
2728 * dcache entries should not be moved in this way. Caller must hold
2729 * rename_lock, the i_mutex of the source and target directories,
2730 * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2731 */
2732static void __d_move(struct dentry *dentry, struct dentry *target,
2733 bool exchange)
2734{
2735 struct dentry *old_parent, *p;
2736 struct inode *dir = NULL;
2737 unsigned n;
2738
2739 WARN_ON(!dentry->d_inode);
2740 if (WARN_ON(dentry == target))
2741 return;
2742
2743 BUG_ON(d_ancestor(target, dentry));
2744 old_parent = dentry->d_parent;
2745 p = d_ancestor(old_parent, target);
2746 if (IS_ROOT(dentry)) {
2747 BUG_ON(p);
2748 spin_lock(&target->d_parent->d_lock);
2749 } else if (!p) {
2750 /* target is not a descendent of dentry->d_parent */
2751 spin_lock(&target->d_parent->d_lock);
2752 spin_lock_nested(&old_parent->d_lock, DENTRY_D_LOCK_NESTED);
2753 } else {
2754 BUG_ON(p == dentry);
2755 spin_lock(&old_parent->d_lock);
2756 if (p != target)
2757 spin_lock_nested(&target->d_parent->d_lock,
2758 DENTRY_D_LOCK_NESTED);
2759 }
2760 spin_lock_nested(&dentry->d_lock, 2);
2761 spin_lock_nested(&target->d_lock, 3);
2762
2763 if (unlikely(d_in_lookup(target))) {
2764 dir = target->d_parent->d_inode;
2765 n = start_dir_add(dir);
2766 __d_lookup_done(target);
2767 }
2768
2769 write_seqcount_begin(&dentry->d_seq);
2770 write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2771
2772 /* unhash both */
2773 if (!d_unhashed(dentry))
2774 ___d_drop(dentry);
2775 if (!d_unhashed(target))
2776 ___d_drop(target);
2777
2778 /* ... and switch them in the tree */
2779 dentry->d_parent = target->d_parent;
2780 if (!exchange) {
2781 copy_name(dentry, target);
2782 target->d_hash.pprev = NULL;
2783 dentry->d_parent->d_lockref.count++;
2784 if (dentry == old_parent)
2785 dentry->d_flags |= DCACHE_RCUACCESS;
2786 else
2787 WARN_ON(!--old_parent->d_lockref.count);
2788 } else {
2789 target->d_parent = old_parent;
2790 swap_names(dentry, target);
2791 list_move(&target->d_child, &target->d_parent->d_subdirs);
2792 __d_rehash(target);
2793 fsnotify_update_flags(target);
2794 }
2795 list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2796 __d_rehash(dentry);
2797 fsnotify_update_flags(dentry);
2798
2799 write_seqcount_end(&target->d_seq);
2800 write_seqcount_end(&dentry->d_seq);
2801
2802 if (dir)
2803 end_dir_add(dir, n);
2804
2805 if (dentry->d_parent != old_parent)
2806 spin_unlock(&dentry->d_parent->d_lock);
2807 if (dentry != old_parent)
2808 spin_unlock(&old_parent->d_lock);
2809 spin_unlock(&target->d_lock);
2810 spin_unlock(&dentry->d_lock);
2811}
2812
2813/*
2814 * d_move - move a dentry
2815 * @dentry: entry to move
2816 * @target: new dentry
2817 *
2818 * Update the dcache to reflect the move of a file name. Negative
2819 * dcache entries should not be moved in this way. See the locking
2820 * requirements for __d_move.
2821 */
2822void d_move(struct dentry *dentry, struct dentry *target)
2823{
2824 write_seqlock(&rename_lock);
2825 __d_move(dentry, target, false);
2826 write_sequnlock(&rename_lock);
2827}
2828EXPORT_SYMBOL(d_move);
2829
2830/*
2831 * d_exchange - exchange two dentries
2832 * @dentry1: first dentry
2833 * @dentry2: second dentry
2834 */
2835void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2836{
2837 write_seqlock(&rename_lock);
2838
2839 WARN_ON(!dentry1->d_inode);
2840 WARN_ON(!dentry2->d_inode);
2841 WARN_ON(IS_ROOT(dentry1));
2842 WARN_ON(IS_ROOT(dentry2));
2843
2844 __d_move(dentry1, dentry2, true);
2845
2846 write_sequnlock(&rename_lock);
2847}
2848
2849/**
2850 * d_ancestor - search for an ancestor
2851 * @p1: ancestor dentry
2852 * @p2: child dentry
2853 *
2854 * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2855 * an ancestor of p2, else NULL.
2856 */
2857struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2858{
2859 struct dentry *p;
2860
2861 for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2862 if (p->d_parent == p1)
2863 return p;
2864 }
2865 return NULL;
2866}
2867
2868/*
2869 * This helper attempts to cope with remotely renamed directories
2870 *
2871 * It assumes that the caller is already holding
2872 * dentry->d_parent->d_inode->i_mutex, and rename_lock
2873 *
2874 * Note: If ever the locking in lock_rename() changes, then please
2875 * remember to update this too...
2876 */
2877static int __d_unalias(struct inode *inode,
2878 struct dentry *dentry, struct dentry *alias)
2879{
2880 struct mutex *m1 = NULL;
2881 struct rw_semaphore *m2 = NULL;
2882 int ret = -ESTALE;
2883
2884 /* If alias and dentry share a parent, then no extra locks required */
2885 if (alias->d_parent == dentry->d_parent)
2886 goto out_unalias;
2887
2888 /* See lock_rename() */
2889 if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2890 goto out_err;
2891 m1 = &dentry->d_sb->s_vfs_rename_mutex;
2892 if (!inode_trylock_shared(alias->d_parent->d_inode))
2893 goto out_err;
2894 m2 = &alias->d_parent->d_inode->i_rwsem;
2895out_unalias:
2896 __d_move(alias, dentry, false);
2897 ret = 0;
2898out_err:
2899 if (m2)
2900 up_read(m2);
2901 if (m1)
2902 mutex_unlock(m1);
2903 return ret;
2904}
2905
2906/**
2907 * d_splice_alias - splice a disconnected dentry into the tree if one exists
2908 * @inode: the inode which may have a disconnected dentry
2909 * @dentry: a negative dentry which we want to point to the inode.
2910 *
2911 * If inode is a directory and has an IS_ROOT alias, then d_move that in
2912 * place of the given dentry and return it, else simply d_add the inode
2913 * to the dentry and return NULL.
2914 *
2915 * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2916 * we should error out: directories can't have multiple aliases.
2917 *
2918 * This is needed in the lookup routine of any filesystem that is exportable
2919 * (via knfsd) so that we can build dcache paths to directories effectively.
2920 *
2921 * If a dentry was found and moved, then it is returned. Otherwise NULL
2922 * is returned. This matches the expected return value of ->lookup.
2923 *
2924 * Cluster filesystems may call this function with a negative, hashed dentry.
2925 * In that case, we know that the inode will be a regular file, and also this
2926 * will only occur during atomic_open. So we need to check for the dentry
2927 * being already hashed only in the final case.
2928 */
2929struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
2930{
2931 if (IS_ERR(inode))
2932 return ERR_CAST(inode);
2933
2934 BUG_ON(!d_unhashed(dentry));
2935
2936 if (!inode)
2937 goto out;
2938
2939 security_d_instantiate(dentry, inode);
2940 spin_lock(&inode->i_lock);
2941 if (S_ISDIR(inode->i_mode)) {
2942 struct dentry *new = __d_find_any_alias(inode);
2943 if (unlikely(new)) {
2944 /* The reference to new ensures it remains an alias */
2945 spin_unlock(&inode->i_lock);
2946 write_seqlock(&rename_lock);
2947 if (unlikely(d_ancestor(new, dentry))) {
2948 write_sequnlock(&rename_lock);
2949 dput(new);
2950 new = ERR_PTR(-ELOOP);
2951 pr_warn_ratelimited(
2952 "VFS: Lookup of '%s' in %s %s"
2953 " would have caused loop\n",
2954 dentry->d_name.name,
2955 inode->i_sb->s_type->name,
2956 inode->i_sb->s_id);
2957 } else if (!IS_ROOT(new)) {
2958 struct dentry *old_parent = dget(new->d_parent);
2959 int err = __d_unalias(inode, dentry, new);
2960 write_sequnlock(&rename_lock);
2961 if (err) {
2962 dput(new);
2963 new = ERR_PTR(err);
2964 }
2965 dput(old_parent);
2966 } else {
2967 __d_move(new, dentry, false);
2968 write_sequnlock(&rename_lock);
2969 }
2970 iput(inode);
2971 return new;
2972 }
2973 }
2974out:
2975 __d_add(dentry, inode);
2976 return NULL;
2977}
2978EXPORT_SYMBOL(d_splice_alias);
2979
2980/*
2981 * Test whether new_dentry is a subdirectory of old_dentry.
2982 *
2983 * Trivially implemented using the dcache structure
2984 */
2985
2986/**
2987 * is_subdir - is new dentry a subdirectory of old_dentry
2988 * @new_dentry: new dentry
2989 * @old_dentry: old dentry
2990 *
2991 * Returns true if new_dentry is a subdirectory of the parent (at any depth).
2992 * Returns false otherwise.
2993 * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
2994 */
2995
2996bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
2997{
2998 bool result;
2999 unsigned seq;
3000
3001 if (new_dentry == old_dentry)
3002 return true;
3003
3004 do {
3005 /* for restarting inner loop in case of seq retry */
3006 seq = read_seqbegin(&rename_lock);
3007 /*
3008 * Need rcu_readlock to protect against the d_parent trashing
3009 * due to d_move
3010 */
3011 rcu_read_lock();
3012 if (d_ancestor(old_dentry, new_dentry))
3013 result = true;
3014 else
3015 result = false;
3016 rcu_read_unlock();
3017 } while (read_seqretry(&rename_lock, seq));
3018
3019 return result;
3020}
3021EXPORT_SYMBOL(is_subdir);
3022
3023static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3024{
3025 struct dentry *root = data;
3026 if (dentry != root) {
3027 if (d_unhashed(dentry) || !dentry->d_inode)
3028 return D_WALK_SKIP;
3029
3030 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3031 dentry->d_flags |= DCACHE_GENOCIDE;
3032 dentry->d_lockref.count--;
3033 }
3034 }
3035 return D_WALK_CONTINUE;
3036}
3037
3038void d_genocide(struct dentry *parent)
3039{
3040 d_walk(parent, parent, d_genocide_kill);
3041}
3042
3043EXPORT_SYMBOL(d_genocide);
3044
3045void d_tmpfile(struct dentry *dentry, struct inode *inode)
3046{
3047 inode_dec_link_count(inode);
3048 BUG_ON(dentry->d_name.name != dentry->d_iname ||
3049 !hlist_unhashed(&dentry->d_u.d_alias) ||
3050 !d_unlinked(dentry));
3051 spin_lock(&dentry->d_parent->d_lock);
3052 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3053 dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3054 (unsigned long long)inode->i_ino);
3055 spin_unlock(&dentry->d_lock);
3056 spin_unlock(&dentry->d_parent->d_lock);
3057 d_instantiate(dentry, inode);
3058}
3059EXPORT_SYMBOL(d_tmpfile);
3060
3061static __initdata unsigned long dhash_entries;
3062static int __init set_dhash_entries(char *str)
3063{
3064 if (!str)
3065 return 0;
3066 dhash_entries = simple_strtoul(str, &str, 0);
3067 return 1;
3068}
3069__setup("dhash_entries=", set_dhash_entries);
3070
3071static void __init dcache_init_early(void)
3072{
3073 /* If hashes are distributed across NUMA nodes, defer
3074 * hash allocation until vmalloc space is available.
3075 */
3076 if (hashdist)
3077 return;
3078
3079 dentry_hashtable =
3080 alloc_large_system_hash("Dentry cache",
3081 sizeof(struct hlist_bl_head),
3082 dhash_entries,
3083 13,
3084 HASH_EARLY | HASH_ZERO,
3085 &d_hash_shift,
3086 NULL,
3087 0,
3088 0);
3089 d_hash_shift = 32 - d_hash_shift;
3090}
3091
3092static void __init dcache_init(void)
3093{
3094 /*
3095 * A constructor could be added for stable state like the lists,
3096 * but it is probably not worth it because of the cache nature
3097 * of the dcache.
3098 */
3099 dentry_cache = KMEM_CACHE_USERCOPY(dentry,
3100 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD|SLAB_ACCOUNT,
3101 d_iname);
3102
3103 /* Hash may have been set up in dcache_init_early */
3104 if (!hashdist)
3105 return;
3106
3107 dentry_hashtable =
3108 alloc_large_system_hash("Dentry cache",
3109 sizeof(struct hlist_bl_head),
3110 dhash_entries,
3111 13,
3112 HASH_ZERO,
3113 &d_hash_shift,
3114 NULL,
3115 0,
3116 0);
3117 d_hash_shift = 32 - d_hash_shift;
3118}
3119
3120/* SLAB cache for __getname() consumers */
3121struct kmem_cache *names_cachep __read_mostly;
3122EXPORT_SYMBOL(names_cachep);
3123
3124void __init vfs_caches_init_early(void)
3125{
3126 int i;
3127
3128 for (i = 0; i < ARRAY_SIZE(in_lookup_hashtable); i++)
3129 INIT_HLIST_BL_HEAD(&in_lookup_hashtable[i]);
3130
3131 dcache_init_early();
3132 inode_init_early();
3133}
3134
3135void __init vfs_caches_init(void)
3136{
3137 names_cachep = kmem_cache_create_usercopy("names_cache", PATH_MAX, 0,
3138 SLAB_HWCACHE_ALIGN|SLAB_PANIC, 0, PATH_MAX, NULL);
3139
3140 dcache_init();
3141 inode_init();
3142 files_init();
3143 files_maxfiles_init();
3144 mnt_init();
3145 bdev_cache_init();
3146 chrdev_init();
3147}
3148