root dentries need RCU-delayed freeing
[muen/linux.git] / fs / dcache.c
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/bootmem.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  */
72 int sysctl_vfs_cache_pressure __read_mostly = 100;
73 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
74
75 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
76
77 EXPORT_SYMBOL(rename_lock);
78
79 static struct kmem_cache *dentry_cache __read_mostly;
80
81 const struct qstr empty_name = QSTR_INIT("", 0);
82 EXPORT_SYMBOL(empty_name);
83 const struct qstr slash_name = QSTR_INIT("/", 1);
84 EXPORT_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
95 static unsigned int d_hash_shift __read_mostly;
96
97 static struct hlist_bl_head *dentry_hashtable __read_mostly;
98
99 static 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
105 static struct hlist_bl_head in_lookup_hashtable[1 << IN_LOOKUP_SHIFT];
106
107 static 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. */
116 struct dentry_stat_t dentry_stat = {
117         .age_limit = 45,
118 };
119
120 static DEFINE_PER_CPU(long, nr_dentry);
121 static DEFINE_PER_CPU(long, nr_dentry_unused);
122
123 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
124
125 /*
126  * Here we resort to our own counters instead of using generic per-cpu counters
127  * for consistency with what the vfs inode code does. We are expected to harvest
128  * better code and performance by having our own specialized counters.
129  *
130  * Please note that the loop is done over all possible CPUs, not over all online
131  * CPUs. The reason for this is that we don't want to play games with CPUs going
132  * on and off. If one of them goes off, we will just keep their counters.
133  *
134  * glommer: See cffbc8a for details, and if you ever intend to change this,
135  * please update all vfs counters to match.
136  */
137 static long get_nr_dentry(void)
138 {
139         int i;
140         long sum = 0;
141         for_each_possible_cpu(i)
142                 sum += per_cpu(nr_dentry, i);
143         return sum < 0 ? 0 : sum;
144 }
145
146 static long get_nr_dentry_unused(void)
147 {
148         int i;
149         long sum = 0;
150         for_each_possible_cpu(i)
151                 sum += per_cpu(nr_dentry_unused, i);
152         return sum < 0 ? 0 : sum;
153 }
154
155 int proc_nr_dentry(struct ctl_table *table, int write, void __user *buffer,
156                    size_t *lenp, loff_t *ppos)
157 {
158         dentry_stat.nr_dentry = get_nr_dentry();
159         dentry_stat.nr_unused = get_nr_dentry_unused();
160         return proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
161 }
162 #endif
163
164 /*
165  * Compare 2 name strings, return 0 if they match, otherwise non-zero.
166  * The strings are both count bytes long, and count is non-zero.
167  */
168 #ifdef CONFIG_DCACHE_WORD_ACCESS
169
170 #include <asm/word-at-a-time.h>
171 /*
172  * NOTE! 'cs' and 'scount' come from a dentry, so it has a
173  * aligned allocation for this particular component. We don't
174  * strictly need the load_unaligned_zeropad() safety, but it
175  * doesn't hurt either.
176  *
177  * In contrast, 'ct' and 'tcount' can be from a pathname, and do
178  * need the careful unaligned handling.
179  */
180 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
181 {
182         unsigned long a,b,mask;
183
184         for (;;) {
185                 a = read_word_at_a_time(cs);
186                 b = load_unaligned_zeropad(ct);
187                 if (tcount < sizeof(unsigned long))
188                         break;
189                 if (unlikely(a != b))
190                         return 1;
191                 cs += sizeof(unsigned long);
192                 ct += sizeof(unsigned long);
193                 tcount -= sizeof(unsigned long);
194                 if (!tcount)
195                         return 0;
196         }
197         mask = bytemask_from_count(tcount);
198         return unlikely(!!((a ^ b) & mask));
199 }
200
201 #else
202
203 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
204 {
205         do {
206                 if (*cs != *ct)
207                         return 1;
208                 cs++;
209                 ct++;
210                 tcount--;
211         } while (tcount);
212         return 0;
213 }
214
215 #endif
216
217 static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount)
218 {
219         /*
220          * Be careful about RCU walk racing with rename:
221          * use 'READ_ONCE' to fetch the name pointer.
222          *
223          * NOTE! Even if a rename will mean that the length
224          * was not loaded atomically, we don't care. The
225          * RCU walk will check the sequence count eventually,
226          * and catch it. And we won't overrun the buffer,
227          * because we're reading the name pointer atomically,
228          * and a dentry name is guaranteed to be properly
229          * terminated with a NUL byte.
230          *
231          * End result: even if 'len' is wrong, we'll exit
232          * early because the data cannot match (there can
233          * be no NUL in the ct/tcount data)
234          */
235         const unsigned char *cs = READ_ONCE(dentry->d_name.name);
236
237         return dentry_string_cmp(cs, ct, tcount);
238 }
239
240 struct external_name {
241         union {
242                 atomic_t count;
243                 struct rcu_head head;
244         } u;
245         unsigned char name[];
246 };
247
248 static inline struct external_name *external_name(struct dentry *dentry)
249 {
250         return container_of(dentry->d_name.name, struct external_name, name[0]);
251 }
252
253 static void __d_free(struct rcu_head *head)
254 {
255         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
256
257         kmem_cache_free(dentry_cache, dentry); 
258 }
259
260 static void __d_free_external_name(struct rcu_head *head)
261 {
262         struct external_name *name = container_of(head, struct external_name,
263                                                   u.head);
264
265         mod_node_page_state(page_pgdat(virt_to_page(name)),
266                             NR_INDIRECTLY_RECLAIMABLE_BYTES,
267                             -ksize(name));
268
269         kfree(name);
270 }
271
272 static void __d_free_external(struct rcu_head *head)
273 {
274         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
275
276         __d_free_external_name(&external_name(dentry)->u.head);
277
278         kmem_cache_free(dentry_cache, dentry);
279 }
280
281 static inline int dname_external(const struct dentry *dentry)
282 {
283         return dentry->d_name.name != dentry->d_iname;
284 }
285
286 void take_dentry_name_snapshot(struct name_snapshot *name, struct dentry *dentry)
287 {
288         spin_lock(&dentry->d_lock);
289         if (unlikely(dname_external(dentry))) {
290                 struct external_name *p = external_name(dentry);
291                 atomic_inc(&p->u.count);
292                 spin_unlock(&dentry->d_lock);
293                 name->name = p->name;
294         } else {
295                 memcpy(name->inline_name, dentry->d_iname, DNAME_INLINE_LEN);
296                 spin_unlock(&dentry->d_lock);
297                 name->name = name->inline_name;
298         }
299 }
300 EXPORT_SYMBOL(take_dentry_name_snapshot);
301
302 void release_dentry_name_snapshot(struct name_snapshot *name)
303 {
304         if (unlikely(name->name != name->inline_name)) {
305                 struct external_name *p;
306                 p = container_of(name->name, struct external_name, name[0]);
307                 if (unlikely(atomic_dec_and_test(&p->u.count)))
308                         call_rcu(&p->u.head, __d_free_external_name);
309         }
310 }
311 EXPORT_SYMBOL(release_dentry_name_snapshot);
312
313 static inline void __d_set_inode_and_type(struct dentry *dentry,
314                                           struct inode *inode,
315                                           unsigned type_flags)
316 {
317         unsigned flags;
318
319         dentry->d_inode = inode;
320         flags = READ_ONCE(dentry->d_flags);
321         flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
322         flags |= type_flags;
323         WRITE_ONCE(dentry->d_flags, flags);
324 }
325
326 static inline void __d_clear_type_and_inode(struct dentry *dentry)
327 {
328         unsigned flags = READ_ONCE(dentry->d_flags);
329
330         flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
331         WRITE_ONCE(dentry->d_flags, flags);
332         dentry->d_inode = NULL;
333 }
334
335 static void dentry_free(struct dentry *dentry)
336 {
337         WARN_ON(!hlist_unhashed(&dentry->d_u.d_alias));
338         if (unlikely(dname_external(dentry))) {
339                 struct external_name *p = external_name(dentry);
340                 if (likely(atomic_dec_and_test(&p->u.count))) {
341                         call_rcu(&dentry->d_u.d_rcu, __d_free_external);
342                         return;
343                 }
344         }
345         /* if dentry was never visible to RCU, immediate free is OK */
346         if (!(dentry->d_flags & DCACHE_RCUACCESS))
347                 __d_free(&dentry->d_u.d_rcu);
348         else
349                 call_rcu(&dentry->d_u.d_rcu, __d_free);
350 }
351
352 /*
353  * Release the dentry's inode, using the filesystem
354  * d_iput() operation if defined.
355  */
356 static void dentry_unlink_inode(struct dentry * dentry)
357         __releases(dentry->d_lock)
358         __releases(dentry->d_inode->i_lock)
359 {
360         struct inode *inode = dentry->d_inode;
361         bool hashed = !d_unhashed(dentry);
362
363         if (hashed)
364                 raw_write_seqcount_begin(&dentry->d_seq);
365         __d_clear_type_and_inode(dentry);
366         hlist_del_init(&dentry->d_u.d_alias);
367         if (hashed)
368                 raw_write_seqcount_end(&dentry->d_seq);
369         spin_unlock(&dentry->d_lock);
370         spin_unlock(&inode->i_lock);
371         if (!inode->i_nlink)
372                 fsnotify_inoderemove(inode);
373         if (dentry->d_op && dentry->d_op->d_iput)
374                 dentry->d_op->d_iput(dentry, inode);
375         else
376                 iput(inode);
377 }
378
379 /*
380  * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
381  * is in use - which includes both the "real" per-superblock
382  * LRU list _and_ the DCACHE_SHRINK_LIST use.
383  *
384  * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
385  * on the shrink list (ie not on the superblock LRU list).
386  *
387  * The per-cpu "nr_dentry_unused" counters are updated with
388  * the DCACHE_LRU_LIST bit.
389  *
390  * These helper functions make sure we always follow the
391  * rules. d_lock must be held by the caller.
392  */
393 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
394 static void d_lru_add(struct dentry *dentry)
395 {
396         D_FLAG_VERIFY(dentry, 0);
397         dentry->d_flags |= DCACHE_LRU_LIST;
398         this_cpu_inc(nr_dentry_unused);
399         WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
400 }
401
402 static void d_lru_del(struct dentry *dentry)
403 {
404         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
405         dentry->d_flags &= ~DCACHE_LRU_LIST;
406         this_cpu_dec(nr_dentry_unused);
407         WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
408 }
409
410 static void d_shrink_del(struct dentry *dentry)
411 {
412         D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
413         list_del_init(&dentry->d_lru);
414         dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
415         this_cpu_dec(nr_dentry_unused);
416 }
417
418 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
419 {
420         D_FLAG_VERIFY(dentry, 0);
421         list_add(&dentry->d_lru, list);
422         dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
423         this_cpu_inc(nr_dentry_unused);
424 }
425
426 /*
427  * These can only be called under the global LRU lock, ie during the
428  * callback for freeing the LRU list. "isolate" removes it from the
429  * LRU lists entirely, while shrink_move moves it to the indicated
430  * private list.
431  */
432 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
433 {
434         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
435         dentry->d_flags &= ~DCACHE_LRU_LIST;
436         this_cpu_dec(nr_dentry_unused);
437         list_lru_isolate(lru, &dentry->d_lru);
438 }
439
440 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
441                               struct list_head *list)
442 {
443         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
444         dentry->d_flags |= DCACHE_SHRINK_LIST;
445         list_lru_isolate_move(lru, &dentry->d_lru, list);
446 }
447
448 /**
449  * d_drop - drop a dentry
450  * @dentry: dentry to drop
451  *
452  * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
453  * be found through a VFS lookup any more. Note that this is different from
454  * deleting the dentry - d_delete will try to mark the dentry negative if
455  * possible, giving a successful _negative_ lookup, while d_drop will
456  * just make the cache lookup fail.
457  *
458  * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
459  * reason (NFS timeouts or autofs deletes).
460  *
461  * __d_drop requires dentry->d_lock
462  * ___d_drop doesn't mark dentry as "unhashed"
463  *   (dentry->d_hash.pprev will be LIST_POISON2, not NULL).
464  */
465 static void ___d_drop(struct dentry *dentry)
466 {
467         struct hlist_bl_head *b;
468         /*
469          * Hashed dentries are normally on the dentry hashtable,
470          * with the exception of those newly allocated by
471          * d_obtain_root, which are always IS_ROOT:
472          */
473         if (unlikely(IS_ROOT(dentry)))
474                 b = &dentry->d_sb->s_roots;
475         else
476                 b = d_hash(dentry->d_name.hash);
477
478         hlist_bl_lock(b);
479         __hlist_bl_del(&dentry->d_hash);
480         hlist_bl_unlock(b);
481 }
482
483 void __d_drop(struct dentry *dentry)
484 {
485         if (!d_unhashed(dentry)) {
486                 ___d_drop(dentry);
487                 dentry->d_hash.pprev = NULL;
488                 write_seqcount_invalidate(&dentry->d_seq);
489         }
490 }
491 EXPORT_SYMBOL(__d_drop);
492
493 void d_drop(struct dentry *dentry)
494 {
495         spin_lock(&dentry->d_lock);
496         __d_drop(dentry);
497         spin_unlock(&dentry->d_lock);
498 }
499 EXPORT_SYMBOL(d_drop);
500
501 static inline void dentry_unlist(struct dentry *dentry, struct dentry *parent)
502 {
503         struct dentry *next;
504         /*
505          * Inform d_walk() and shrink_dentry_list() that we are no longer
506          * attached to the dentry tree
507          */
508         dentry->d_flags |= DCACHE_DENTRY_KILLED;
509         if (unlikely(list_empty(&dentry->d_child)))
510                 return;
511         __list_del_entry(&dentry->d_child);
512         /*
513          * Cursors can move around the list of children.  While we'd been
514          * a normal list member, it didn't matter - ->d_child.next would've
515          * been updated.  However, from now on it won't be and for the
516          * things like d_walk() it might end up with a nasty surprise.
517          * Normally d_walk() doesn't care about cursors moving around -
518          * ->d_lock on parent prevents that and since a cursor has no children
519          * of its own, we get through it without ever unlocking the parent.
520          * There is one exception, though - if we ascend from a child that
521          * gets killed as soon as we unlock it, the next sibling is found
522          * using the value left in its ->d_child.next.  And if _that_
523          * pointed to a cursor, and cursor got moved (e.g. by lseek())
524          * before d_walk() regains parent->d_lock, we'll end up skipping
525          * everything the cursor had been moved past.
526          *
527          * Solution: make sure that the pointer left behind in ->d_child.next
528          * points to something that won't be moving around.  I.e. skip the
529          * cursors.
530          */
531         while (dentry->d_child.next != &parent->d_subdirs) {
532                 next = list_entry(dentry->d_child.next, struct dentry, d_child);
533                 if (likely(!(next->d_flags & DCACHE_DENTRY_CURSOR)))
534                         break;
535                 dentry->d_child.next = next->d_child.next;
536         }
537 }
538
539 static void __dentry_kill(struct dentry *dentry)
540 {
541         struct dentry *parent = NULL;
542         bool can_free = true;
543         if (!IS_ROOT(dentry))
544                 parent = dentry->d_parent;
545
546         /*
547          * The dentry is now unrecoverably dead to the world.
548          */
549         lockref_mark_dead(&dentry->d_lockref);
550
551         /*
552          * inform the fs via d_prune that this dentry is about to be
553          * unhashed and destroyed.
554          */
555         if (dentry->d_flags & DCACHE_OP_PRUNE)
556                 dentry->d_op->d_prune(dentry);
557
558         if (dentry->d_flags & DCACHE_LRU_LIST) {
559                 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
560                         d_lru_del(dentry);
561         }
562         /* if it was on the hash then remove it */
563         __d_drop(dentry);
564         dentry_unlist(dentry, parent);
565         if (parent)
566                 spin_unlock(&parent->d_lock);
567         if (dentry->d_inode)
568                 dentry_unlink_inode(dentry);
569         else
570                 spin_unlock(&dentry->d_lock);
571         this_cpu_dec(nr_dentry);
572         if (dentry->d_op && dentry->d_op->d_release)
573                 dentry->d_op->d_release(dentry);
574
575         spin_lock(&dentry->d_lock);
576         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
577                 dentry->d_flags |= DCACHE_MAY_FREE;
578                 can_free = false;
579         }
580         spin_unlock(&dentry->d_lock);
581         if (likely(can_free))
582                 dentry_free(dentry);
583         cond_resched();
584 }
585
586 static struct dentry *__lock_parent(struct dentry *dentry)
587 {
588         struct dentry *parent;
589         rcu_read_lock();
590         spin_unlock(&dentry->d_lock);
591 again:
592         parent = READ_ONCE(dentry->d_parent);
593         spin_lock(&parent->d_lock);
594         /*
595          * We can't blindly lock dentry until we are sure
596          * that we won't violate the locking order.
597          * Any changes of dentry->d_parent must have
598          * been done with parent->d_lock held, so
599          * spin_lock() above is enough of a barrier
600          * for checking if it's still our child.
601          */
602         if (unlikely(parent != dentry->d_parent)) {
603                 spin_unlock(&parent->d_lock);
604                 goto again;
605         }
606         rcu_read_unlock();
607         if (parent != dentry)
608                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
609         else
610                 parent = NULL;
611         return parent;
612 }
613
614 static inline struct dentry *lock_parent(struct dentry *dentry)
615 {
616         struct dentry *parent = dentry->d_parent;
617         if (IS_ROOT(dentry))
618                 return NULL;
619         if (likely(spin_trylock(&parent->d_lock)))
620                 return parent;
621         return __lock_parent(dentry);
622 }
623
624 static inline bool retain_dentry(struct dentry *dentry)
625 {
626         WARN_ON(d_in_lookup(dentry));
627
628         /* Unreachable? Get rid of it */
629         if (unlikely(d_unhashed(dentry)))
630                 return false;
631
632         if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
633                 return false;
634
635         if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
636                 if (dentry->d_op->d_delete(dentry))
637                         return false;
638         }
639         /* retain; LRU fodder */
640         dentry->d_lockref.count--;
641         if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
642                 d_lru_add(dentry);
643         else if (unlikely(!(dentry->d_flags & DCACHE_REFERENCED)))
644                 dentry->d_flags |= DCACHE_REFERENCED;
645         return true;
646 }
647
648 /*
649  * Finish off a dentry we've decided to kill.
650  * dentry->d_lock must be held, returns with it unlocked.
651  * Returns dentry requiring refcount drop, or NULL if we're done.
652  */
653 static struct dentry *dentry_kill(struct dentry *dentry)
654         __releases(dentry->d_lock)
655 {
656         struct inode *inode = dentry->d_inode;
657         struct dentry *parent = NULL;
658
659         if (inode && unlikely(!spin_trylock(&inode->i_lock)))
660                 goto slow_positive;
661
662         if (!IS_ROOT(dentry)) {
663                 parent = dentry->d_parent;
664                 if (unlikely(!spin_trylock(&parent->d_lock))) {
665                         parent = __lock_parent(dentry);
666                         if (likely(inode || !dentry->d_inode))
667                                 goto got_locks;
668                         /* negative that became positive */
669                         if (parent)
670                                 spin_unlock(&parent->d_lock);
671                         inode = dentry->d_inode;
672                         goto slow_positive;
673                 }
674         }
675         __dentry_kill(dentry);
676         return parent;
677
678 slow_positive:
679         spin_unlock(&dentry->d_lock);
680         spin_lock(&inode->i_lock);
681         spin_lock(&dentry->d_lock);
682         parent = lock_parent(dentry);
683 got_locks:
684         if (unlikely(dentry->d_lockref.count != 1)) {
685                 dentry->d_lockref.count--;
686         } else if (likely(!retain_dentry(dentry))) {
687                 __dentry_kill(dentry);
688                 return parent;
689         }
690         /* we are keeping it, after all */
691         if (inode)
692                 spin_unlock(&inode->i_lock);
693         if (parent)
694                 spin_unlock(&parent->d_lock);
695         spin_unlock(&dentry->d_lock);
696         return NULL;
697 }
698
699 /*
700  * Try to do a lockless dput(), and return whether that was successful.
701  *
702  * If unsuccessful, we return false, having already taken the dentry lock.
703  *
704  * The caller needs to hold the RCU read lock, so that the dentry is
705  * guaranteed to stay around even if the refcount goes down to zero!
706  */
707 static inline bool fast_dput(struct dentry *dentry)
708 {
709         int ret;
710         unsigned int d_flags;
711
712         /*
713          * If we have a d_op->d_delete() operation, we sould not
714          * let the dentry count go to zero, so use "put_or_lock".
715          */
716         if (unlikely(dentry->d_flags & DCACHE_OP_DELETE))
717                 return lockref_put_or_lock(&dentry->d_lockref);
718
719         /*
720          * .. otherwise, we can try to just decrement the
721          * lockref optimistically.
722          */
723         ret = lockref_put_return(&dentry->d_lockref);
724
725         /*
726          * If the lockref_put_return() failed due to the lock being held
727          * by somebody else, the fast path has failed. We will need to
728          * get the lock, and then check the count again.
729          */
730         if (unlikely(ret < 0)) {
731                 spin_lock(&dentry->d_lock);
732                 if (dentry->d_lockref.count > 1) {
733                         dentry->d_lockref.count--;
734                         spin_unlock(&dentry->d_lock);
735                         return 1;
736                 }
737                 return 0;
738         }
739
740         /*
741          * If we weren't the last ref, we're done.
742          */
743         if (ret)
744                 return 1;
745
746         /*
747          * Careful, careful. The reference count went down
748          * to zero, but we don't hold the dentry lock, so
749          * somebody else could get it again, and do another
750          * dput(), and we need to not race with that.
751          *
752          * However, there is a very special and common case
753          * where we don't care, because there is nothing to
754          * do: the dentry is still hashed, it does not have
755          * a 'delete' op, and it's referenced and already on
756          * the LRU list.
757          *
758          * NOTE! Since we aren't locked, these values are
759          * not "stable". However, it is sufficient that at
760          * some point after we dropped the reference the
761          * dentry was hashed and the flags had the proper
762          * value. Other dentry users may have re-gotten
763          * a reference to the dentry and change that, but
764          * our work is done - we can leave the dentry
765          * around with a zero refcount.
766          */
767         smp_rmb();
768         d_flags = READ_ONCE(dentry->d_flags);
769         d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST | DCACHE_DISCONNECTED;
770
771         /* Nothing to do? Dropping the reference was all we needed? */
772         if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry))
773                 return 1;
774
775         /*
776          * Not the fast normal case? Get the lock. We've already decremented
777          * the refcount, but we'll need to re-check the situation after
778          * getting the lock.
779          */
780         spin_lock(&dentry->d_lock);
781
782         /*
783          * Did somebody else grab a reference to it in the meantime, and
784          * we're no longer the last user after all? Alternatively, somebody
785          * else could have killed it and marked it dead. Either way, we
786          * don't need to do anything else.
787          */
788         if (dentry->d_lockref.count) {
789                 spin_unlock(&dentry->d_lock);
790                 return 1;
791         }
792
793         /*
794          * Re-get the reference we optimistically dropped. We hold the
795          * lock, and we just tested that it was zero, so we can just
796          * set it to 1.
797          */
798         dentry->d_lockref.count = 1;
799         return 0;
800 }
801
802
803 /* 
804  * This is dput
805  *
806  * This is complicated by the fact that we do not want to put
807  * dentries that are no longer on any hash chain on the unused
808  * list: we'd much rather just get rid of them immediately.
809  *
810  * However, that implies that we have to traverse the dentry
811  * tree upwards to the parents which might _also_ now be
812  * scheduled for deletion (it may have been only waiting for
813  * its last child to go away).
814  *
815  * This tail recursion is done by hand as we don't want to depend
816  * on the compiler to always get this right (gcc generally doesn't).
817  * Real recursion would eat up our stack space.
818  */
819
820 /*
821  * dput - release a dentry
822  * @dentry: dentry to release 
823  *
824  * Release a dentry. This will drop the usage count and if appropriate
825  * call the dentry unlink method as well as removing it from the queues and
826  * releasing its resources. If the parent dentries were scheduled for release
827  * they too may now get deleted.
828  */
829 void dput(struct dentry *dentry)
830 {
831         while (dentry) {
832                 might_sleep();
833
834                 rcu_read_lock();
835                 if (likely(fast_dput(dentry))) {
836                         rcu_read_unlock();
837                         return;
838                 }
839
840                 /* Slow case: now with the dentry lock held */
841                 rcu_read_unlock();
842
843                 if (likely(retain_dentry(dentry))) {
844                         spin_unlock(&dentry->d_lock);
845                         return;
846                 }
847
848                 dentry = dentry_kill(dentry);
849         }
850 }
851 EXPORT_SYMBOL(dput);
852
853
854 /* This must be called with d_lock held */
855 static inline void __dget_dlock(struct dentry *dentry)
856 {
857         dentry->d_lockref.count++;
858 }
859
860 static inline void __dget(struct dentry *dentry)
861 {
862         lockref_get(&dentry->d_lockref);
863 }
864
865 struct dentry *dget_parent(struct dentry *dentry)
866 {
867         int gotref;
868         struct dentry *ret;
869
870         /*
871          * Do optimistic parent lookup without any
872          * locking.
873          */
874         rcu_read_lock();
875         ret = READ_ONCE(dentry->d_parent);
876         gotref = lockref_get_not_zero(&ret->d_lockref);
877         rcu_read_unlock();
878         if (likely(gotref)) {
879                 if (likely(ret == READ_ONCE(dentry->d_parent)))
880                         return ret;
881                 dput(ret);
882         }
883
884 repeat:
885         /*
886          * Don't need rcu_dereference because we re-check it was correct under
887          * the lock.
888          */
889         rcu_read_lock();
890         ret = dentry->d_parent;
891         spin_lock(&ret->d_lock);
892         if (unlikely(ret != dentry->d_parent)) {
893                 spin_unlock(&ret->d_lock);
894                 rcu_read_unlock();
895                 goto repeat;
896         }
897         rcu_read_unlock();
898         BUG_ON(!ret->d_lockref.count);
899         ret->d_lockref.count++;
900         spin_unlock(&ret->d_lock);
901         return ret;
902 }
903 EXPORT_SYMBOL(dget_parent);
904
905 static struct dentry * __d_find_any_alias(struct inode *inode)
906 {
907         struct dentry *alias;
908
909         if (hlist_empty(&inode->i_dentry))
910                 return NULL;
911         alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
912         __dget(alias);
913         return alias;
914 }
915
916 /**
917  * d_find_any_alias - find any alias for a given inode
918  * @inode: inode to find an alias for
919  *
920  * If any aliases exist for the given inode, take and return a
921  * reference for one of them.  If no aliases exist, return %NULL.
922  */
923 struct dentry *d_find_any_alias(struct inode *inode)
924 {
925         struct dentry *de;
926
927         spin_lock(&inode->i_lock);
928         de = __d_find_any_alias(inode);
929         spin_unlock(&inode->i_lock);
930         return de;
931 }
932 EXPORT_SYMBOL(d_find_any_alias);
933
934 /**
935  * d_find_alias - grab a hashed alias of inode
936  * @inode: inode in question
937  *
938  * If inode has a hashed alias, or is a directory and has any alias,
939  * acquire the reference to alias and return it. Otherwise return NULL.
940  * Notice that if inode is a directory there can be only one alias and
941  * it can be unhashed only if it has no children, or if it is the root
942  * of a filesystem, or if the directory was renamed and d_revalidate
943  * was the first vfs operation to notice.
944  *
945  * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
946  * any other hashed alias over that one.
947  */
948 static struct dentry *__d_find_alias(struct inode *inode)
949 {
950         struct dentry *alias;
951
952         if (S_ISDIR(inode->i_mode))
953                 return __d_find_any_alias(inode);
954
955         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
956                 spin_lock(&alias->d_lock);
957                 if (!d_unhashed(alias)) {
958                         __dget_dlock(alias);
959                         spin_unlock(&alias->d_lock);
960                         return alias;
961                 }
962                 spin_unlock(&alias->d_lock);
963         }
964         return NULL;
965 }
966
967 struct dentry *d_find_alias(struct inode *inode)
968 {
969         struct dentry *de = NULL;
970
971         if (!hlist_empty(&inode->i_dentry)) {
972                 spin_lock(&inode->i_lock);
973                 de = __d_find_alias(inode);
974                 spin_unlock(&inode->i_lock);
975         }
976         return de;
977 }
978 EXPORT_SYMBOL(d_find_alias);
979
980 /*
981  *      Try to kill dentries associated with this inode.
982  * WARNING: you must own a reference to inode.
983  */
984 void d_prune_aliases(struct inode *inode)
985 {
986         struct dentry *dentry;
987 restart:
988         spin_lock(&inode->i_lock);
989         hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
990                 spin_lock(&dentry->d_lock);
991                 if (!dentry->d_lockref.count) {
992                         struct dentry *parent = lock_parent(dentry);
993                         if (likely(!dentry->d_lockref.count)) {
994                                 __dentry_kill(dentry);
995                                 dput(parent);
996                                 goto restart;
997                         }
998                         if (parent)
999                                 spin_unlock(&parent->d_lock);
1000                 }
1001                 spin_unlock(&dentry->d_lock);
1002         }
1003         spin_unlock(&inode->i_lock);
1004 }
1005 EXPORT_SYMBOL(d_prune_aliases);
1006
1007 /*
1008  * Lock a dentry from shrink list.
1009  * Called under rcu_read_lock() and dentry->d_lock; the former
1010  * guarantees that nothing we access will be freed under us.
1011  * Note that dentry is *not* protected from concurrent dentry_kill(),
1012  * d_delete(), etc.
1013  *
1014  * Return false if dentry has been disrupted or grabbed, leaving
1015  * the caller to kick it off-list.  Otherwise, return true and have
1016  * that dentry's inode and parent both locked.
1017  */
1018 static bool shrink_lock_dentry(struct dentry *dentry)
1019 {
1020         struct inode *inode;
1021         struct dentry *parent;
1022
1023         if (dentry->d_lockref.count)
1024                 return false;
1025
1026         inode = dentry->d_inode;
1027         if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
1028                 spin_unlock(&dentry->d_lock);
1029                 spin_lock(&inode->i_lock);
1030                 spin_lock(&dentry->d_lock);
1031                 if (unlikely(dentry->d_lockref.count))
1032                         goto out;
1033                 /* changed inode means that somebody had grabbed it */
1034                 if (unlikely(inode != dentry->d_inode))
1035                         goto out;
1036         }
1037
1038         parent = dentry->d_parent;
1039         if (IS_ROOT(dentry) || likely(spin_trylock(&parent->d_lock)))
1040                 return true;
1041
1042         spin_unlock(&dentry->d_lock);
1043         spin_lock(&parent->d_lock);
1044         if (unlikely(parent != dentry->d_parent)) {
1045                 spin_unlock(&parent->d_lock);
1046                 spin_lock(&dentry->d_lock);
1047                 goto out;
1048         }
1049         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1050         if (likely(!dentry->d_lockref.count))
1051                 return true;
1052         spin_unlock(&parent->d_lock);
1053 out:
1054         if (inode)
1055                 spin_unlock(&inode->i_lock);
1056         return false;
1057 }
1058
1059 static void shrink_dentry_list(struct list_head *list)
1060 {
1061         while (!list_empty(list)) {
1062                 struct dentry *dentry, *parent;
1063
1064                 dentry = list_entry(list->prev, struct dentry, d_lru);
1065                 spin_lock(&dentry->d_lock);
1066                 rcu_read_lock();
1067                 if (!shrink_lock_dentry(dentry)) {
1068                         bool can_free = false;
1069                         rcu_read_unlock();
1070                         d_shrink_del(dentry);
1071                         if (dentry->d_lockref.count < 0)
1072                                 can_free = dentry->d_flags & DCACHE_MAY_FREE;
1073                         spin_unlock(&dentry->d_lock);
1074                         if (can_free)
1075                                 dentry_free(dentry);
1076                         continue;
1077                 }
1078                 rcu_read_unlock();
1079                 d_shrink_del(dentry);
1080                 parent = dentry->d_parent;
1081                 __dentry_kill(dentry);
1082                 if (parent == dentry)
1083                         continue;
1084                 /*
1085                  * We need to prune ancestors too. This is necessary to prevent
1086                  * quadratic behavior of shrink_dcache_parent(), but is also
1087                  * expected to be beneficial in reducing dentry cache
1088                  * fragmentation.
1089                  */
1090                 dentry = parent;
1091                 while (dentry && !lockref_put_or_lock(&dentry->d_lockref))
1092                         dentry = dentry_kill(dentry);
1093         }
1094 }
1095
1096 static enum lru_status dentry_lru_isolate(struct list_head *item,
1097                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1098 {
1099         struct list_head *freeable = arg;
1100         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1101
1102
1103         /*
1104          * we are inverting the lru lock/dentry->d_lock here,
1105          * so use a trylock. If we fail to get the lock, just skip
1106          * it
1107          */
1108         if (!spin_trylock(&dentry->d_lock))
1109                 return LRU_SKIP;
1110
1111         /*
1112          * Referenced dentries are still in use. If they have active
1113          * counts, just remove them from the LRU. Otherwise give them
1114          * another pass through the LRU.
1115          */
1116         if (dentry->d_lockref.count) {
1117                 d_lru_isolate(lru, dentry);
1118                 spin_unlock(&dentry->d_lock);
1119                 return LRU_REMOVED;
1120         }
1121
1122         if (dentry->d_flags & DCACHE_REFERENCED) {
1123                 dentry->d_flags &= ~DCACHE_REFERENCED;
1124                 spin_unlock(&dentry->d_lock);
1125
1126                 /*
1127                  * The list move itself will be made by the common LRU code. At
1128                  * this point, we've dropped the dentry->d_lock but keep the
1129                  * lru lock. This is safe to do, since every list movement is
1130                  * protected by the lru lock even if both locks are held.
1131                  *
1132                  * This is guaranteed by the fact that all LRU management
1133                  * functions are intermediated by the LRU API calls like
1134                  * list_lru_add and list_lru_del. List movement in this file
1135                  * only ever occur through this functions or through callbacks
1136                  * like this one, that are called from the LRU API.
1137                  *
1138                  * The only exceptions to this are functions like
1139                  * shrink_dentry_list, and code that first checks for the
1140                  * DCACHE_SHRINK_LIST flag.  Those are guaranteed to be
1141                  * operating only with stack provided lists after they are
1142                  * properly isolated from the main list.  It is thus, always a
1143                  * local access.
1144                  */
1145                 return LRU_ROTATE;
1146         }
1147
1148         d_lru_shrink_move(lru, dentry, freeable);
1149         spin_unlock(&dentry->d_lock);
1150
1151         return LRU_REMOVED;
1152 }
1153
1154 /**
1155  * prune_dcache_sb - shrink the dcache
1156  * @sb: superblock
1157  * @sc: shrink control, passed to list_lru_shrink_walk()
1158  *
1159  * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1160  * is done when we need more memory and called from the superblock shrinker
1161  * function.
1162  *
1163  * This function may fail to free any resources if all the dentries are in
1164  * use.
1165  */
1166 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1167 {
1168         LIST_HEAD(dispose);
1169         long freed;
1170
1171         freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1172                                      dentry_lru_isolate, &dispose);
1173         shrink_dentry_list(&dispose);
1174         return freed;
1175 }
1176
1177 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1178                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1179 {
1180         struct list_head *freeable = arg;
1181         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1182
1183         /*
1184          * we are inverting the lru lock/dentry->d_lock here,
1185          * so use a trylock. If we fail to get the lock, just skip
1186          * it
1187          */
1188         if (!spin_trylock(&dentry->d_lock))
1189                 return LRU_SKIP;
1190
1191         d_lru_shrink_move(lru, dentry, freeable);
1192         spin_unlock(&dentry->d_lock);
1193
1194         return LRU_REMOVED;
1195 }
1196
1197
1198 /**
1199  * shrink_dcache_sb - shrink dcache for a superblock
1200  * @sb: superblock
1201  *
1202  * Shrink the dcache for the specified super block. This is used to free
1203  * the dcache before unmounting a file system.
1204  */
1205 void shrink_dcache_sb(struct super_block *sb)
1206 {
1207         long freed;
1208
1209         do {
1210                 LIST_HEAD(dispose);
1211
1212                 freed = list_lru_walk(&sb->s_dentry_lru,
1213                         dentry_lru_isolate_shrink, &dispose, 1024);
1214
1215                 this_cpu_sub(nr_dentry_unused, freed);
1216                 shrink_dentry_list(&dispose);
1217         } while (list_lru_count(&sb->s_dentry_lru) > 0);
1218 }
1219 EXPORT_SYMBOL(shrink_dcache_sb);
1220
1221 /**
1222  * enum d_walk_ret - action to talke during tree walk
1223  * @D_WALK_CONTINUE:    contrinue walk
1224  * @D_WALK_QUIT:        quit walk
1225  * @D_WALK_NORETRY:     quit when retry is needed
1226  * @D_WALK_SKIP:        skip this dentry and its children
1227  */
1228 enum d_walk_ret {
1229         D_WALK_CONTINUE,
1230         D_WALK_QUIT,
1231         D_WALK_NORETRY,
1232         D_WALK_SKIP,
1233 };
1234
1235 /**
1236  * d_walk - walk the dentry tree
1237  * @parent:     start of walk
1238  * @data:       data passed to @enter() and @finish()
1239  * @enter:      callback when first entering the dentry
1240  *
1241  * The @enter() callbacks are called with d_lock held.
1242  */
1243 static void d_walk(struct dentry *parent, void *data,
1244                    enum d_walk_ret (*enter)(void *, struct dentry *))
1245 {
1246         struct dentry *this_parent;
1247         struct list_head *next;
1248         unsigned seq = 0;
1249         enum d_walk_ret ret;
1250         bool retry = true;
1251
1252 again:
1253         read_seqbegin_or_lock(&rename_lock, &seq);
1254         this_parent = parent;
1255         spin_lock(&this_parent->d_lock);
1256
1257         ret = enter(data, this_parent);
1258         switch (ret) {
1259         case D_WALK_CONTINUE:
1260                 break;
1261         case D_WALK_QUIT:
1262         case D_WALK_SKIP:
1263                 goto out_unlock;
1264         case D_WALK_NORETRY:
1265                 retry = false;
1266                 break;
1267         }
1268 repeat:
1269         next = this_parent->d_subdirs.next;
1270 resume:
1271         while (next != &this_parent->d_subdirs) {
1272                 struct list_head *tmp = next;
1273                 struct dentry *dentry = list_entry(tmp, struct dentry, d_child);
1274                 next = tmp->next;
1275
1276                 if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
1277                         continue;
1278
1279                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1280
1281                 ret = enter(data, dentry);
1282                 switch (ret) {
1283                 case D_WALK_CONTINUE:
1284                         break;
1285                 case D_WALK_QUIT:
1286                         spin_unlock(&dentry->d_lock);
1287                         goto out_unlock;
1288                 case D_WALK_NORETRY:
1289                         retry = false;
1290                         break;
1291                 case D_WALK_SKIP:
1292                         spin_unlock(&dentry->d_lock);
1293                         continue;
1294                 }
1295
1296                 if (!list_empty(&dentry->d_subdirs)) {
1297                         spin_unlock(&this_parent->d_lock);
1298                         spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1299                         this_parent = dentry;
1300                         spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1301                         goto repeat;
1302                 }
1303                 spin_unlock(&dentry->d_lock);
1304         }
1305         /*
1306          * All done at this level ... ascend and resume the search.
1307          */
1308         rcu_read_lock();
1309 ascend:
1310         if (this_parent != parent) {
1311                 struct dentry *child = this_parent;
1312                 this_parent = child->d_parent;
1313
1314                 spin_unlock(&child->d_lock);
1315                 spin_lock(&this_parent->d_lock);
1316
1317                 /* might go back up the wrong parent if we have had a rename. */
1318                 if (need_seqretry(&rename_lock, seq))
1319                         goto rename_retry;
1320                 /* go into the first sibling still alive */
1321                 do {
1322                         next = child->d_child.next;
1323                         if (next == &this_parent->d_subdirs)
1324                                 goto ascend;
1325                         child = list_entry(next, struct dentry, d_child);
1326                 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED));
1327                 rcu_read_unlock();
1328                 goto resume;
1329         }
1330         if (need_seqretry(&rename_lock, seq))
1331                 goto rename_retry;
1332         rcu_read_unlock();
1333
1334 out_unlock:
1335         spin_unlock(&this_parent->d_lock);
1336         done_seqretry(&rename_lock, seq);
1337         return;
1338
1339 rename_retry:
1340         spin_unlock(&this_parent->d_lock);
1341         rcu_read_unlock();
1342         BUG_ON(seq & 1);
1343         if (!retry)
1344                 return;
1345         seq = 1;
1346         goto again;
1347 }
1348
1349 struct check_mount {
1350         struct vfsmount *mnt;
1351         unsigned int mounted;
1352 };
1353
1354 static enum d_walk_ret path_check_mount(void *data, struct dentry *dentry)
1355 {
1356         struct check_mount *info = data;
1357         struct path path = { .mnt = info->mnt, .dentry = dentry };
1358
1359         if (likely(!d_mountpoint(dentry)))
1360                 return D_WALK_CONTINUE;
1361         if (__path_is_mountpoint(&path)) {
1362                 info->mounted = 1;
1363                 return D_WALK_QUIT;
1364         }
1365         return D_WALK_CONTINUE;
1366 }
1367
1368 /**
1369  * path_has_submounts - check for mounts over a dentry in the
1370  *                      current namespace.
1371  * @parent: path to check.
1372  *
1373  * Return true if the parent or its subdirectories contain
1374  * a mount point in the current namespace.
1375  */
1376 int path_has_submounts(const struct path *parent)
1377 {
1378         struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };
1379
1380         read_seqlock_excl(&mount_lock);
1381         d_walk(parent->dentry, &data, path_check_mount);
1382         read_sequnlock_excl(&mount_lock);
1383
1384         return data.mounted;
1385 }
1386 EXPORT_SYMBOL(path_has_submounts);
1387
1388 /*
1389  * Called by mount code to set a mountpoint and check if the mountpoint is
1390  * reachable (e.g. NFS can unhash a directory dentry and then the complete
1391  * subtree can become unreachable).
1392  *
1393  * Only one of d_invalidate() and d_set_mounted() must succeed.  For
1394  * this reason take rename_lock and d_lock on dentry and ancestors.
1395  */
1396 int d_set_mounted(struct dentry *dentry)
1397 {
1398         struct dentry *p;
1399         int ret = -ENOENT;
1400         write_seqlock(&rename_lock);
1401         for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1402                 /* Need exclusion wrt. d_invalidate() */
1403                 spin_lock(&p->d_lock);
1404                 if (unlikely(d_unhashed(p))) {
1405                         spin_unlock(&p->d_lock);
1406                         goto out;
1407                 }
1408                 spin_unlock(&p->d_lock);
1409         }
1410         spin_lock(&dentry->d_lock);
1411         if (!d_unlinked(dentry)) {
1412                 ret = -EBUSY;
1413                 if (!d_mountpoint(dentry)) {
1414                         dentry->d_flags |= DCACHE_MOUNTED;
1415                         ret = 0;
1416                 }
1417         }
1418         spin_unlock(&dentry->d_lock);
1419 out:
1420         write_sequnlock(&rename_lock);
1421         return ret;
1422 }
1423
1424 /*
1425  * Search the dentry child list of the specified parent,
1426  * and move any unused dentries to the end of the unused
1427  * list for prune_dcache(). We descend to the next level
1428  * whenever the d_subdirs list is non-empty and continue
1429  * searching.
1430  *
1431  * It returns zero iff there are no unused children,
1432  * otherwise  it returns the number of children moved to
1433  * the end of the unused list. This may not be the total
1434  * number of unused children, because select_parent can
1435  * drop the lock and return early due to latency
1436  * constraints.
1437  */
1438
1439 struct select_data {
1440         struct dentry *start;
1441         struct list_head dispose;
1442         int found;
1443 };
1444
1445 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1446 {
1447         struct select_data *data = _data;
1448         enum d_walk_ret ret = D_WALK_CONTINUE;
1449
1450         if (data->start == dentry)
1451                 goto out;
1452
1453         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1454                 data->found++;
1455         } else {
1456                 if (dentry->d_flags & DCACHE_LRU_LIST)
1457                         d_lru_del(dentry);
1458                 if (!dentry->d_lockref.count) {
1459                         d_shrink_add(dentry, &data->dispose);
1460                         data->found++;
1461                 }
1462         }
1463         /*
1464          * We can return to the caller if we have found some (this
1465          * ensures forward progress). We'll be coming back to find
1466          * the rest.
1467          */
1468         if (!list_empty(&data->dispose))
1469                 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1470 out:
1471         return ret;
1472 }
1473
1474 /**
1475  * shrink_dcache_parent - prune dcache
1476  * @parent: parent of entries to prune
1477  *
1478  * Prune the dcache to remove unused children of the parent dentry.
1479  */
1480 void shrink_dcache_parent(struct dentry *parent)
1481 {
1482         for (;;) {
1483                 struct select_data data;
1484
1485                 INIT_LIST_HEAD(&data.dispose);
1486                 data.start = parent;
1487                 data.found = 0;
1488
1489                 d_walk(parent, &data, select_collect);
1490
1491                 if (!list_empty(&data.dispose)) {
1492                         shrink_dentry_list(&data.dispose);
1493                         continue;
1494                 }
1495
1496                 cond_resched();
1497                 if (!data.found)
1498                         break;
1499         }
1500 }
1501 EXPORT_SYMBOL(shrink_dcache_parent);
1502
1503 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1504 {
1505         /* it has busy descendents; complain about those instead */
1506         if (!list_empty(&dentry->d_subdirs))
1507                 return D_WALK_CONTINUE;
1508
1509         /* root with refcount 1 is fine */
1510         if (dentry == _data && dentry->d_lockref.count == 1)
1511                 return D_WALK_CONTINUE;
1512
1513         printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} "
1514                         " still in use (%d) [unmount of %s %s]\n",
1515                        dentry,
1516                        dentry->d_inode ?
1517                        dentry->d_inode->i_ino : 0UL,
1518                        dentry,
1519                        dentry->d_lockref.count,
1520                        dentry->d_sb->s_type->name,
1521                        dentry->d_sb->s_id);
1522         WARN_ON(1);
1523         return D_WALK_CONTINUE;
1524 }
1525
1526 static void do_one_tree(struct dentry *dentry)
1527 {
1528         shrink_dcache_parent(dentry);
1529         d_walk(dentry, dentry, umount_check);
1530         d_drop(dentry);
1531         dput(dentry);
1532 }
1533
1534 /*
1535  * destroy the dentries attached to a superblock on unmounting
1536  */
1537 void shrink_dcache_for_umount(struct super_block *sb)
1538 {
1539         struct dentry *dentry;
1540
1541         WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
1542
1543         dentry = sb->s_root;
1544         sb->s_root = NULL;
1545         do_one_tree(dentry);
1546
1547         while (!hlist_bl_empty(&sb->s_roots)) {
1548                 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_roots), struct dentry, d_hash));
1549                 do_one_tree(dentry);
1550         }
1551 }
1552
1553 static enum d_walk_ret find_submount(void *_data, struct dentry *dentry)
1554 {
1555         struct dentry **victim = _data;
1556         if (d_mountpoint(dentry)) {
1557                 __dget_dlock(dentry);
1558                 *victim = dentry;
1559                 return D_WALK_QUIT;
1560         }
1561         return D_WALK_CONTINUE;
1562 }
1563
1564 /**
1565  * d_invalidate - detach submounts, prune dcache, and drop
1566  * @dentry: dentry to invalidate (aka detach, prune and drop)
1567  */
1568 void d_invalidate(struct dentry *dentry)
1569 {
1570         bool had_submounts = false;
1571         spin_lock(&dentry->d_lock);
1572         if (d_unhashed(dentry)) {
1573                 spin_unlock(&dentry->d_lock);
1574                 return;
1575         }
1576         __d_drop(dentry);
1577         spin_unlock(&dentry->d_lock);
1578
1579         /* Negative dentries can be dropped without further checks */
1580         if (!dentry->d_inode)
1581                 return;
1582
1583         shrink_dcache_parent(dentry);
1584         for (;;) {
1585                 struct dentry *victim = NULL;
1586                 d_walk(dentry, &victim, find_submount);
1587                 if (!victim) {
1588                         if (had_submounts)
1589                                 shrink_dcache_parent(dentry);
1590                         return;
1591                 }
1592                 had_submounts = true;
1593                 detach_mounts(victim);
1594                 dput(victim);
1595         }
1596 }
1597 EXPORT_SYMBOL(d_invalidate);
1598
1599 /**
1600  * __d_alloc    -       allocate a dcache entry
1601  * @sb: filesystem it will belong to
1602  * @name: qstr of the name
1603  *
1604  * Allocates a dentry. It returns %NULL if there is insufficient memory
1605  * available. On a success the dentry is returned. The name passed in is
1606  * copied and the copy passed in may be reused after this call.
1607  */
1608  
1609 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1610 {
1611         struct external_name *ext = NULL;
1612         struct dentry *dentry;
1613         char *dname;
1614         int err;
1615
1616         dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1617         if (!dentry)
1618                 return NULL;
1619
1620         /*
1621          * We guarantee that the inline name is always NUL-terminated.
1622          * This way the memcpy() done by the name switching in rename
1623          * will still always have a NUL at the end, even if we might
1624          * be overwriting an internal NUL character
1625          */
1626         dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1627         if (unlikely(!name)) {
1628                 name = &slash_name;
1629                 dname = dentry->d_iname;
1630         } else if (name->len > DNAME_INLINE_LEN-1) {
1631                 size_t size = offsetof(struct external_name, name[1]);
1632
1633                 ext = kmalloc(size + name->len, GFP_KERNEL_ACCOUNT);
1634                 if (!ext) {
1635                         kmem_cache_free(dentry_cache, dentry); 
1636                         return NULL;
1637                 }
1638                 atomic_set(&ext->u.count, 1);
1639                 dname = ext->name;
1640         } else  {
1641                 dname = dentry->d_iname;
1642         }       
1643
1644         dentry->d_name.len = name->len;
1645         dentry->d_name.hash = name->hash;
1646         memcpy(dname, name->name, name->len);
1647         dname[name->len] = 0;
1648
1649         /* Make sure we always see the terminating NUL character */
1650         smp_store_release(&dentry->d_name.name, dname); /* ^^^ */
1651
1652         dentry->d_lockref.count = 1;
1653         dentry->d_flags = 0;
1654         spin_lock_init(&dentry->d_lock);
1655         seqcount_init(&dentry->d_seq);
1656         dentry->d_inode = NULL;
1657         dentry->d_parent = dentry;
1658         dentry->d_sb = sb;
1659         dentry->d_op = NULL;
1660         dentry->d_fsdata = NULL;
1661         INIT_HLIST_BL_NODE(&dentry->d_hash);
1662         INIT_LIST_HEAD(&dentry->d_lru);
1663         INIT_LIST_HEAD(&dentry->d_subdirs);
1664         INIT_HLIST_NODE(&dentry->d_u.d_alias);
1665         INIT_LIST_HEAD(&dentry->d_child);
1666         d_set_d_op(dentry, dentry->d_sb->s_d_op);
1667
1668         if (dentry->d_op && dentry->d_op->d_init) {
1669                 err = dentry->d_op->d_init(dentry);
1670                 if (err) {
1671                         if (dname_external(dentry))
1672                                 kfree(external_name(dentry));
1673                         kmem_cache_free(dentry_cache, dentry);
1674                         return NULL;
1675                 }
1676         }
1677
1678         if (unlikely(ext)) {
1679                 pg_data_t *pgdat = page_pgdat(virt_to_page(ext));
1680                 mod_node_page_state(pgdat, NR_INDIRECTLY_RECLAIMABLE_BYTES,
1681                                     ksize(ext));
1682         }
1683
1684         this_cpu_inc(nr_dentry);
1685
1686         return dentry;
1687 }
1688
1689 /**
1690  * d_alloc      -       allocate a dcache entry
1691  * @parent: parent of entry to allocate
1692  * @name: qstr of the name
1693  *
1694  * Allocates a dentry. It returns %NULL if there is insufficient memory
1695  * available. On a success the dentry is returned. The name passed in is
1696  * copied and the copy passed in may be reused after this call.
1697  */
1698 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1699 {
1700         struct dentry *dentry = __d_alloc(parent->d_sb, name);
1701         if (!dentry)
1702                 return NULL;
1703         dentry->d_flags |= DCACHE_RCUACCESS;
1704         spin_lock(&parent->d_lock);
1705         /*
1706          * don't need child lock because it is not subject
1707          * to concurrency here
1708          */
1709         __dget_dlock(parent);
1710         dentry->d_parent = parent;
1711         list_add(&dentry->d_child, &parent->d_subdirs);
1712         spin_unlock(&parent->d_lock);
1713
1714         return dentry;
1715 }
1716 EXPORT_SYMBOL(d_alloc);
1717
1718 struct dentry *d_alloc_anon(struct super_block *sb)
1719 {
1720         return __d_alloc(sb, NULL);
1721 }
1722 EXPORT_SYMBOL(d_alloc_anon);
1723
1724 struct dentry *d_alloc_cursor(struct dentry * parent)
1725 {
1726         struct dentry *dentry = d_alloc_anon(parent->d_sb);
1727         if (dentry) {
1728                 dentry->d_flags |= DCACHE_RCUACCESS | DCACHE_DENTRY_CURSOR;
1729                 dentry->d_parent = dget(parent);
1730         }
1731         return dentry;
1732 }
1733
1734 /**
1735  * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1736  * @sb: the superblock
1737  * @name: qstr of the name
1738  *
1739  * For a filesystem that just pins its dentries in memory and never
1740  * performs lookups at all, return an unhashed IS_ROOT dentry.
1741  */
1742 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1743 {
1744         return __d_alloc(sb, name);
1745 }
1746 EXPORT_SYMBOL(d_alloc_pseudo);
1747
1748 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1749 {
1750         struct qstr q;
1751
1752         q.name = name;
1753         q.hash_len = hashlen_string(parent, name);
1754         return d_alloc(parent, &q);
1755 }
1756 EXPORT_SYMBOL(d_alloc_name);
1757
1758 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1759 {
1760         WARN_ON_ONCE(dentry->d_op);
1761         WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH  |
1762                                 DCACHE_OP_COMPARE       |
1763                                 DCACHE_OP_REVALIDATE    |
1764                                 DCACHE_OP_WEAK_REVALIDATE       |
1765                                 DCACHE_OP_DELETE        |
1766                                 DCACHE_OP_REAL));
1767         dentry->d_op = op;
1768         if (!op)
1769                 return;
1770         if (op->d_hash)
1771                 dentry->d_flags |= DCACHE_OP_HASH;
1772         if (op->d_compare)
1773                 dentry->d_flags |= DCACHE_OP_COMPARE;
1774         if (op->d_revalidate)
1775                 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1776         if (op->d_weak_revalidate)
1777                 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1778         if (op->d_delete)
1779                 dentry->d_flags |= DCACHE_OP_DELETE;
1780         if (op->d_prune)
1781                 dentry->d_flags |= DCACHE_OP_PRUNE;
1782         if (op->d_real)
1783                 dentry->d_flags |= DCACHE_OP_REAL;
1784
1785 }
1786 EXPORT_SYMBOL(d_set_d_op);
1787
1788
1789 /*
1790  * d_set_fallthru - Mark a dentry as falling through to a lower layer
1791  * @dentry - The dentry to mark
1792  *
1793  * Mark a dentry as falling through to the lower layer (as set with
1794  * d_pin_lower()).  This flag may be recorded on the medium.
1795  */
1796 void d_set_fallthru(struct dentry *dentry)
1797 {
1798         spin_lock(&dentry->d_lock);
1799         dentry->d_flags |= DCACHE_FALLTHRU;
1800         spin_unlock(&dentry->d_lock);
1801 }
1802 EXPORT_SYMBOL(d_set_fallthru);
1803
1804 static unsigned d_flags_for_inode(struct inode *inode)
1805 {
1806         unsigned add_flags = DCACHE_REGULAR_TYPE;
1807
1808         if (!inode)
1809                 return DCACHE_MISS_TYPE;
1810
1811         if (S_ISDIR(inode->i_mode)) {
1812                 add_flags = DCACHE_DIRECTORY_TYPE;
1813                 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1814                         if (unlikely(!inode->i_op->lookup))
1815                                 add_flags = DCACHE_AUTODIR_TYPE;
1816                         else
1817                                 inode->i_opflags |= IOP_LOOKUP;
1818                 }
1819                 goto type_determined;
1820         }
1821
1822         if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1823                 if (unlikely(inode->i_op->get_link)) {
1824                         add_flags = DCACHE_SYMLINK_TYPE;
1825                         goto type_determined;
1826                 }
1827                 inode->i_opflags |= IOP_NOFOLLOW;
1828         }
1829
1830         if (unlikely(!S_ISREG(inode->i_mode)))
1831                 add_flags = DCACHE_SPECIAL_TYPE;
1832
1833 type_determined:
1834         if (unlikely(IS_AUTOMOUNT(inode)))
1835                 add_flags |= DCACHE_NEED_AUTOMOUNT;
1836         return add_flags;
1837 }
1838
1839 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1840 {
1841         unsigned add_flags = d_flags_for_inode(inode);
1842         WARN_ON(d_in_lookup(dentry));
1843
1844         spin_lock(&dentry->d_lock);
1845         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1846         raw_write_seqcount_begin(&dentry->d_seq);
1847         __d_set_inode_and_type(dentry, inode, add_flags);
1848         raw_write_seqcount_end(&dentry->d_seq);
1849         fsnotify_update_flags(dentry);
1850         spin_unlock(&dentry->d_lock);
1851 }
1852
1853 /**
1854  * d_instantiate - fill in inode information for a dentry
1855  * @entry: dentry to complete
1856  * @inode: inode to attach to this dentry
1857  *
1858  * Fill in inode information in the entry.
1859  *
1860  * This turns negative dentries into productive full members
1861  * of society.
1862  *
1863  * NOTE! This assumes that the inode count has been incremented
1864  * (or otherwise set) by the caller to indicate that it is now
1865  * in use by the dcache.
1866  */
1867  
1868 void d_instantiate(struct dentry *entry, struct inode * inode)
1869 {
1870         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1871         if (inode) {
1872                 security_d_instantiate(entry, inode);
1873                 spin_lock(&inode->i_lock);
1874                 __d_instantiate(entry, inode);
1875                 spin_unlock(&inode->i_lock);
1876         }
1877 }
1878 EXPORT_SYMBOL(d_instantiate);
1879
1880 /*
1881  * This should be equivalent to d_instantiate() + unlock_new_inode(),
1882  * with lockdep-related part of unlock_new_inode() done before
1883  * anything else.  Use that instead of open-coding d_instantiate()/
1884  * unlock_new_inode() combinations.
1885  */
1886 void d_instantiate_new(struct dentry *entry, struct inode *inode)
1887 {
1888         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1889         BUG_ON(!inode);
1890         lockdep_annotate_inode_mutex_key(inode);
1891         security_d_instantiate(entry, inode);
1892         spin_lock(&inode->i_lock);
1893         __d_instantiate(entry, inode);
1894         WARN_ON(!(inode->i_state & I_NEW));
1895         inode->i_state &= ~I_NEW;
1896         smp_mb();
1897         wake_up_bit(&inode->i_state, __I_NEW);
1898         spin_unlock(&inode->i_lock);
1899 }
1900 EXPORT_SYMBOL(d_instantiate_new);
1901
1902 /**
1903  * d_instantiate_no_diralias - instantiate a non-aliased dentry
1904  * @entry: dentry to complete
1905  * @inode: inode to attach to this dentry
1906  *
1907  * Fill in inode information in the entry.  If a directory alias is found, then
1908  * return an error (and drop inode).  Together with d_materialise_unique() this
1909  * guarantees that a directory inode may never have more than one alias.
1910  */
1911 int d_instantiate_no_diralias(struct dentry *entry, struct inode *inode)
1912 {
1913         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1914
1915         security_d_instantiate(entry, inode);
1916         spin_lock(&inode->i_lock);
1917         if (S_ISDIR(inode->i_mode) && !hlist_empty(&inode->i_dentry)) {
1918                 spin_unlock(&inode->i_lock);
1919                 iput(inode);
1920                 return -EBUSY;
1921         }
1922         __d_instantiate(entry, inode);
1923         spin_unlock(&inode->i_lock);
1924
1925         return 0;
1926 }
1927 EXPORT_SYMBOL(d_instantiate_no_diralias);
1928
1929 struct dentry *d_make_root(struct inode *root_inode)
1930 {
1931         struct dentry *res = NULL;
1932
1933         if (root_inode) {
1934                 res = d_alloc_anon(root_inode->i_sb);
1935                 if (res) {
1936                         res->d_flags |= DCACHE_RCUACCESS;
1937                         d_instantiate(res, root_inode);
1938                 } else {
1939                         iput(root_inode);
1940                 }
1941         }
1942         return res;
1943 }
1944 EXPORT_SYMBOL(d_make_root);
1945
1946 static struct dentry *__d_instantiate_anon(struct dentry *dentry,
1947                                            struct inode *inode,
1948                                            bool disconnected)
1949 {
1950         struct dentry *res;
1951         unsigned add_flags;
1952
1953         security_d_instantiate(dentry, inode);
1954         spin_lock(&inode->i_lock);
1955         res = __d_find_any_alias(inode);
1956         if (res) {
1957                 spin_unlock(&inode->i_lock);
1958                 dput(dentry);
1959                 goto out_iput;
1960         }
1961
1962         /* attach a disconnected dentry */
1963         add_flags = d_flags_for_inode(inode);
1964
1965         if (disconnected)
1966                 add_flags |= DCACHE_DISCONNECTED;
1967
1968         spin_lock(&dentry->d_lock);
1969         __d_set_inode_and_type(dentry, inode, add_flags);
1970         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1971         if (!disconnected) {
1972                 hlist_bl_lock(&dentry->d_sb->s_roots);
1973                 hlist_bl_add_head(&dentry->d_hash, &dentry->d_sb->s_roots);
1974                 hlist_bl_unlock(&dentry->d_sb->s_roots);
1975         }
1976         spin_unlock(&dentry->d_lock);
1977         spin_unlock(&inode->i_lock);
1978
1979         return dentry;
1980
1981  out_iput:
1982         iput(inode);
1983         return res;
1984 }
1985
1986 struct dentry *d_instantiate_anon(struct dentry *dentry, struct inode *inode)
1987 {
1988         return __d_instantiate_anon(dentry, inode, true);
1989 }
1990 EXPORT_SYMBOL(d_instantiate_anon);
1991
1992 static struct dentry *__d_obtain_alias(struct inode *inode, bool disconnected)
1993 {
1994         struct dentry *tmp;
1995         struct dentry *res;
1996
1997         if (!inode)
1998                 return ERR_PTR(-ESTALE);
1999         if (IS_ERR(inode))
2000                 return ERR_CAST(inode);
2001
2002         res = d_find_any_alias(inode);
2003         if (res)
2004                 goto out_iput;
2005
2006         tmp = d_alloc_anon(inode->i_sb);
2007         if (!tmp) {
2008                 res = ERR_PTR(-ENOMEM);
2009                 goto out_iput;
2010         }
2011
2012         return __d_instantiate_anon(tmp, inode, disconnected);
2013
2014 out_iput:
2015         iput(inode);
2016         return res;
2017 }
2018
2019 /**
2020  * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
2021  * @inode: inode to allocate the dentry for
2022  *
2023  * Obtain a dentry for an inode resulting from NFS filehandle conversion or
2024  * similar open by handle operations.  The returned dentry may be anonymous,
2025  * or may have a full name (if the inode was already in the cache).
2026  *
2027  * When called on a directory inode, we must ensure that the inode only ever
2028  * has one dentry.  If a dentry is found, that is returned instead of
2029  * 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 released.
2033  * To make it easier to use in export operations a %NULL or IS_ERR inode may
2034  * be passed in and the error will be propagated to the return value,
2035  * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
2036  */
2037 struct dentry *d_obtain_alias(struct inode *inode)
2038 {
2039         return __d_obtain_alias(inode, true);
2040 }
2041 EXPORT_SYMBOL(d_obtain_alias);
2042
2043 /**
2044  * d_obtain_root - find or allocate a dentry for a given inode
2045  * @inode: inode to allocate the dentry for
2046  *
2047  * Obtain an IS_ROOT dentry for the root of a filesystem.
2048  *
2049  * We must ensure that directory inodes only ever have one dentry.  If a
2050  * dentry is found, that is returned instead of allocating a new one.
2051  *
2052  * On successful return, the reference to the inode has been transferred
2053  * to the dentry.  In case of an error the reference on the inode is
2054  * released.  A %NULL or IS_ERR inode may be passed in and will be the
2055  * error will be propagate to the return value, with a %NULL @inode
2056  * replaced by ERR_PTR(-ESTALE).
2057  */
2058 struct dentry *d_obtain_root(struct inode *inode)
2059 {
2060         return __d_obtain_alias(inode, false);
2061 }
2062 EXPORT_SYMBOL(d_obtain_root);
2063
2064 /**
2065  * d_add_ci - lookup or allocate new dentry with case-exact name
2066  * @inode:  the inode case-insensitive lookup has found
2067  * @dentry: the negative dentry that was passed to the parent's lookup func
2068  * @name:   the case-exact name to be associated with the returned dentry
2069  *
2070  * This is to avoid filling the dcache with case-insensitive names to the
2071  * same inode, only the actual correct case is stored in the dcache for
2072  * case-insensitive filesystems.
2073  *
2074  * For a case-insensitive lookup match and if the the case-exact dentry
2075  * already exists in in the dcache, use it and return it.
2076  *
2077  * If no entry exists with the exact case name, allocate new dentry with
2078  * the exact case, and return the spliced entry.
2079  */
2080 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2081                         struct qstr *name)
2082 {
2083         struct dentry *found, *res;
2084
2085         /*
2086          * First check if a dentry matching the name already exists,
2087          * if not go ahead and create it now.
2088          */
2089         found = d_hash_and_lookup(dentry->d_parent, name);
2090         if (found) {
2091                 iput(inode);
2092                 return found;
2093         }
2094         if (d_in_lookup(dentry)) {
2095                 found = d_alloc_parallel(dentry->d_parent, name,
2096                                         dentry->d_wait);
2097                 if (IS_ERR(found) || !d_in_lookup(found)) {
2098                         iput(inode);
2099                         return found;
2100                 }
2101         } else {
2102                 found = d_alloc(dentry->d_parent, name);
2103                 if (!found) {
2104                         iput(inode);
2105                         return ERR_PTR(-ENOMEM);
2106                 } 
2107         }
2108         res = d_splice_alias(inode, found);
2109         if (res) {
2110                 dput(found);
2111                 return res;
2112         }
2113         return found;
2114 }
2115 EXPORT_SYMBOL(d_add_ci);
2116
2117
2118 static inline bool d_same_name(const struct dentry *dentry,
2119                                 const struct dentry *parent,
2120                                 const struct qstr *name)
2121 {
2122         if (likely(!(parent->d_flags & DCACHE_OP_COMPARE))) {
2123                 if (dentry->d_name.len != name->len)
2124                         return false;
2125                 return dentry_cmp(dentry, name->name, name->len) == 0;
2126         }
2127         return parent->d_op->d_compare(dentry,
2128                                        dentry->d_name.len, dentry->d_name.name,
2129                                        name) == 0;
2130 }
2131
2132 /**
2133  * __d_lookup_rcu - search for a dentry (racy, store-free)
2134  * @parent: parent dentry
2135  * @name: qstr of name we wish to find
2136  * @seqp: returns d_seq value at the point where the dentry was found
2137  * Returns: dentry, or NULL
2138  *
2139  * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2140  * resolution (store-free path walking) design described in
2141  * Documentation/filesystems/path-lookup.txt.
2142  *
2143  * This is not to be used outside core vfs.
2144  *
2145  * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2146  * held, and rcu_read_lock held. The returned dentry must not be stored into
2147  * without taking d_lock and checking d_seq sequence count against @seq
2148  * returned here.
2149  *
2150  * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2151  * function.
2152  *
2153  * Alternatively, __d_lookup_rcu may be called again to look up the child of
2154  * the returned dentry, so long as its parent's seqlock is checked after the
2155  * child is looked up. Thus, an interlocking stepping of sequence lock checks
2156  * is formed, giving integrity down the path walk.
2157  *
2158  * NOTE! The caller *has* to check the resulting dentry against the sequence
2159  * number we've returned before using any of the resulting dentry state!
2160  */
2161 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2162                                 const struct qstr *name,
2163                                 unsigned *seqp)
2164 {
2165         u64 hashlen = name->hash_len;
2166         const unsigned char *str = name->name;
2167         struct hlist_bl_head *b = d_hash(hashlen_hash(hashlen));
2168         struct hlist_bl_node *node;
2169         struct dentry *dentry;
2170
2171         /*
2172          * Note: There is significant duplication with __d_lookup_rcu which is
2173          * required to prevent single threaded performance regressions
2174          * especially on architectures where smp_rmb (in seqcounts) are costly.
2175          * Keep the two functions in sync.
2176          */
2177
2178         /*
2179          * The hash list is protected using RCU.
2180          *
2181          * Carefully use d_seq when comparing a candidate dentry, to avoid
2182          * races with d_move().
2183          *
2184          * It is possible that concurrent renames can mess up our list
2185          * walk here and result in missing our dentry, resulting in the
2186          * false-negative result. d_lookup() protects against concurrent
2187          * renames using rename_lock seqlock.
2188          *
2189          * See Documentation/filesystems/path-lookup.txt for more details.
2190          */
2191         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2192                 unsigned seq;
2193
2194 seqretry:
2195                 /*
2196                  * The dentry sequence count protects us from concurrent
2197                  * renames, and thus protects parent and name fields.
2198                  *
2199                  * The caller must perform a seqcount check in order
2200                  * to do anything useful with the returned dentry.
2201                  *
2202                  * NOTE! We do a "raw" seqcount_begin here. That means that
2203                  * we don't wait for the sequence count to stabilize if it
2204                  * is in the middle of a sequence change. If we do the slow
2205                  * dentry compare, we will do seqretries until it is stable,
2206                  * and if we end up with a successful lookup, we actually
2207                  * want to exit RCU lookup anyway.
2208                  *
2209                  * Note that raw_seqcount_begin still *does* smp_rmb(), so
2210                  * we are still guaranteed NUL-termination of ->d_name.name.
2211                  */
2212                 seq = raw_seqcount_begin(&dentry->d_seq);
2213                 if (dentry->d_parent != parent)
2214                         continue;
2215                 if (d_unhashed(dentry))
2216                         continue;
2217
2218                 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) {
2219                         int tlen;
2220                         const char *tname;
2221                         if (dentry->d_name.hash != hashlen_hash(hashlen))
2222                                 continue;
2223                         tlen = dentry->d_name.len;
2224                         tname = dentry->d_name.name;
2225                         /* we want a consistent (name,len) pair */
2226                         if (read_seqcount_retry(&dentry->d_seq, seq)) {
2227                                 cpu_relax();
2228                                 goto seqretry;
2229                         }
2230                         if (parent->d_op->d_compare(dentry,
2231                                                     tlen, tname, name) != 0)
2232                                 continue;
2233                 } else {
2234                         if (dentry->d_name.hash_len != hashlen)
2235                                 continue;
2236                         if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
2237                                 continue;
2238                 }
2239                 *seqp = seq;
2240                 return dentry;
2241         }
2242         return NULL;
2243 }
2244
2245 /**
2246  * d_lookup - search for a dentry
2247  * @parent: parent dentry
2248  * @name: qstr of name we wish to find
2249  * Returns: dentry, or NULL
2250  *
2251  * d_lookup searches the children of the parent dentry for the name in
2252  * question. If the dentry is found its reference count is incremented and the
2253  * dentry is returned. The caller must use dput to free the entry when it has
2254  * finished using it. %NULL is returned if the dentry does not exist.
2255  */
2256 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2257 {
2258         struct dentry *dentry;
2259         unsigned seq;
2260
2261         do {
2262                 seq = read_seqbegin(&rename_lock);
2263                 dentry = __d_lookup(parent, name);
2264                 if (dentry)
2265                         break;
2266         } while (read_seqretry(&rename_lock, seq));
2267         return dentry;
2268 }
2269 EXPORT_SYMBOL(d_lookup);
2270
2271 /**
2272  * __d_lookup - search for a dentry (racy)
2273  * @parent: parent dentry
2274  * @name: qstr of name we wish to find
2275  * Returns: dentry, or NULL
2276  *
2277  * __d_lookup is like d_lookup, however it may (rarely) return a
2278  * false-negative result due to unrelated rename activity.
2279  *
2280  * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2281  * however it must be used carefully, eg. with a following d_lookup in
2282  * the case of failure.
2283  *
2284  * __d_lookup callers must be commented.
2285  */
2286 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2287 {
2288         unsigned int hash = name->hash;
2289         struct hlist_bl_head *b = d_hash(hash);
2290         struct hlist_bl_node *node;
2291         struct dentry *found = NULL;
2292         struct dentry *dentry;
2293
2294         /*
2295          * Note: There is significant duplication with __d_lookup_rcu which is
2296          * required to prevent single threaded performance regressions
2297          * especially on architectures where smp_rmb (in seqcounts) are costly.
2298          * Keep the two functions in sync.
2299          */
2300
2301         /*
2302          * The hash list is protected using RCU.
2303          *
2304          * Take d_lock when comparing a candidate dentry, to avoid races
2305          * with d_move().
2306          *
2307          * It is possible that concurrent renames can mess up our list
2308          * walk here and result in missing our dentry, resulting in the
2309          * false-negative result. d_lookup() protects against concurrent
2310          * renames using rename_lock seqlock.
2311          *
2312          * See Documentation/filesystems/path-lookup.txt for more details.
2313          */
2314         rcu_read_lock();
2315         
2316         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2317
2318                 if (dentry->d_name.hash != hash)
2319                         continue;
2320
2321                 spin_lock(&dentry->d_lock);
2322                 if (dentry->d_parent != parent)
2323                         goto next;
2324                 if (d_unhashed(dentry))
2325                         goto next;
2326
2327                 if (!d_same_name(dentry, parent, name))
2328                         goto next;
2329
2330                 dentry->d_lockref.count++;
2331                 found = dentry;
2332                 spin_unlock(&dentry->d_lock);
2333                 break;
2334 next:
2335                 spin_unlock(&dentry->d_lock);
2336         }
2337         rcu_read_unlock();
2338
2339         return found;
2340 }
2341
2342 /**
2343  * d_hash_and_lookup - hash the qstr then search for a dentry
2344  * @dir: Directory to search in
2345  * @name: qstr of name we wish to find
2346  *
2347  * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2348  */
2349 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2350 {
2351         /*
2352          * Check for a fs-specific hash function. Note that we must
2353          * calculate the standard hash first, as the d_op->d_hash()
2354          * routine may choose to leave the hash value unchanged.
2355          */
2356         name->hash = full_name_hash(dir, name->name, name->len);
2357         if (dir->d_flags & DCACHE_OP_HASH) {
2358                 int err = dir->d_op->d_hash(dir, name);
2359                 if (unlikely(err < 0))
2360                         return ERR_PTR(err);
2361         }
2362         return d_lookup(dir, name);
2363 }
2364 EXPORT_SYMBOL(d_hash_and_lookup);
2365
2366 /*
2367  * When a file is deleted, we have two options:
2368  * - turn this dentry into a negative dentry
2369  * - unhash this dentry and free it.
2370  *
2371  * Usually, we want to just turn this into
2372  * a negative dentry, but if anybody else is
2373  * currently using the dentry or the inode
2374  * we can't do that and we fall back on removing
2375  * it from the hash queues and waiting for
2376  * it to be deleted later when it has no users
2377  */
2378  
2379 /**
2380  * d_delete - delete a dentry
2381  * @dentry: The dentry to delete
2382  *
2383  * Turn the dentry into a negative dentry if possible, otherwise
2384  * remove it from the hash queues so it can be deleted later
2385  */
2386  
2387 void d_delete(struct dentry * dentry)
2388 {
2389         struct inode *inode = dentry->d_inode;
2390         int isdir = d_is_dir(dentry);
2391
2392         spin_lock(&inode->i_lock);
2393         spin_lock(&dentry->d_lock);
2394         /*
2395          * Are we the only user?
2396          */
2397         if (dentry->d_lockref.count == 1) {
2398                 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2399                 dentry_unlink_inode(dentry);
2400         } else {
2401                 __d_drop(dentry);
2402                 spin_unlock(&dentry->d_lock);
2403                 spin_unlock(&inode->i_lock);
2404         }
2405         fsnotify_nameremove(dentry, isdir);
2406 }
2407 EXPORT_SYMBOL(d_delete);
2408
2409 static void __d_rehash(struct dentry *entry)
2410 {
2411         struct hlist_bl_head *b = d_hash(entry->d_name.hash);
2412
2413         hlist_bl_lock(b);
2414         hlist_bl_add_head_rcu(&entry->d_hash, b);
2415         hlist_bl_unlock(b);
2416 }
2417
2418 /**
2419  * d_rehash     - add an entry back to the hash
2420  * @entry: dentry to add to the hash
2421  *
2422  * Adds a dentry to the hash according to its name.
2423  */
2424  
2425 void d_rehash(struct dentry * entry)
2426 {
2427         spin_lock(&entry->d_lock);
2428         __d_rehash(entry);
2429         spin_unlock(&entry->d_lock);
2430 }
2431 EXPORT_SYMBOL(d_rehash);
2432
2433 static inline unsigned start_dir_add(struct inode *dir)
2434 {
2435
2436         for (;;) {
2437                 unsigned n = dir->i_dir_seq;
2438                 if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
2439                         return n;
2440                 cpu_relax();
2441         }
2442 }
2443
2444 static inline void end_dir_add(struct inode *dir, unsigned n)
2445 {
2446         smp_store_release(&dir->i_dir_seq, n + 2);
2447 }
2448
2449 static void d_wait_lookup(struct dentry *dentry)
2450 {
2451         if (d_in_lookup(dentry)) {
2452                 DECLARE_WAITQUEUE(wait, current);
2453                 add_wait_queue(dentry->d_wait, &wait);
2454                 do {
2455                         set_current_state(TASK_UNINTERRUPTIBLE);
2456                         spin_unlock(&dentry->d_lock);
2457                         schedule();
2458                         spin_lock(&dentry->d_lock);
2459                 } while (d_in_lookup(dentry));
2460         }
2461 }
2462
2463 struct dentry *d_alloc_parallel(struct dentry *parent,
2464                                 const struct qstr *name,
2465                                 wait_queue_head_t *wq)
2466 {
2467         unsigned int hash = name->hash;
2468         struct hlist_bl_head *b = in_lookup_hash(parent, hash);
2469         struct hlist_bl_node *node;
2470         struct dentry *new = d_alloc(parent, name);
2471         struct dentry *dentry;
2472         unsigned seq, r_seq, d_seq;
2473
2474         if (unlikely(!new))
2475                 return ERR_PTR(-ENOMEM);
2476
2477 retry:
2478         rcu_read_lock();
2479         seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
2480         r_seq = read_seqbegin(&rename_lock);
2481         dentry = __d_lookup_rcu(parent, name, &d_seq);
2482         if (unlikely(dentry)) {
2483                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2484                         rcu_read_unlock();
2485                         goto retry;
2486                 }
2487                 if (read_seqcount_retry(&dentry->d_seq, d_seq)) {
2488                         rcu_read_unlock();
2489                         dput(dentry);
2490                         goto retry;
2491                 }
2492                 rcu_read_unlock();
2493                 dput(new);
2494                 return dentry;
2495         }
2496         if (unlikely(read_seqretry(&rename_lock, r_seq))) {
2497                 rcu_read_unlock();
2498                 goto retry;
2499         }
2500
2501         if (unlikely(seq & 1)) {
2502                 rcu_read_unlock();
2503                 goto retry;
2504         }
2505
2506         hlist_bl_lock(b);
2507         if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {
2508                 hlist_bl_unlock(b);
2509                 rcu_read_unlock();
2510                 goto retry;
2511         }
2512         /*
2513          * No changes for the parent since the beginning of d_lookup().
2514          * Since all removals from the chain happen with hlist_bl_lock(),
2515          * any potential in-lookup matches are going to stay here until
2516          * we unlock the chain.  All fields are stable in everything
2517          * we encounter.
2518          */
2519         hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
2520                 if (dentry->d_name.hash != hash)
2521                         continue;
2522                 if (dentry->d_parent != parent)
2523                         continue;
2524                 if (!d_same_name(dentry, parent, name))
2525                         continue;
2526                 hlist_bl_unlock(b);
2527                 /* now we can try to grab a reference */
2528                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2529                         rcu_read_unlock();
2530                         goto retry;
2531                 }
2532
2533                 rcu_read_unlock();
2534                 /*
2535                  * somebody is likely to be still doing lookup for it;
2536                  * wait for them to finish
2537                  */
2538                 spin_lock(&dentry->d_lock);
2539                 d_wait_lookup(dentry);
2540                 /*
2541                  * it's not in-lookup anymore; in principle we should repeat
2542                  * everything from dcache lookup, but it's likely to be what
2543                  * d_lookup() would've found anyway.  If it is, just return it;
2544                  * otherwise we really have to repeat the whole thing.
2545                  */
2546                 if (unlikely(dentry->d_name.hash != hash))
2547                         goto mismatch;
2548                 if (unlikely(dentry->d_parent != parent))
2549                         goto mismatch;
2550                 if (unlikely(d_unhashed(dentry)))
2551                         goto mismatch;
2552                 if (unlikely(!d_same_name(dentry, parent, name)))
2553                         goto mismatch;
2554                 /* OK, it *is* a hashed match; return it */
2555                 spin_unlock(&dentry->d_lock);
2556                 dput(new);
2557                 return dentry;
2558         }
2559         rcu_read_unlock();
2560         /* we can't take ->d_lock here; it's OK, though. */
2561         new->d_flags |= DCACHE_PAR_LOOKUP;
2562         new->d_wait = wq;
2563         hlist_bl_add_head_rcu(&new->d_u.d_in_lookup_hash, b);
2564         hlist_bl_unlock(b);
2565         return new;
2566 mismatch:
2567         spin_unlock(&dentry->d_lock);
2568         dput(dentry);
2569         goto retry;
2570 }
2571 EXPORT_SYMBOL(d_alloc_parallel);
2572
2573 void __d_lookup_done(struct dentry *dentry)
2574 {
2575         struct hlist_bl_head *b = in_lookup_hash(dentry->d_parent,
2576                                                  dentry->d_name.hash);
2577         hlist_bl_lock(b);
2578         dentry->d_flags &= ~DCACHE_PAR_LOOKUP;
2579         __hlist_bl_del(&dentry->d_u.d_in_lookup_hash);
2580         wake_up_all(dentry->d_wait);
2581         dentry->d_wait = NULL;
2582         hlist_bl_unlock(b);
2583         INIT_HLIST_NODE(&dentry->d_u.d_alias);
2584         INIT_LIST_HEAD(&dentry->d_lru);
2585 }
2586 EXPORT_SYMBOL(__d_lookup_done);
2587
2588 /* inode->i_lock held if inode is non-NULL */
2589
2590 static inline void __d_add(struct dentry *dentry, struct inode *inode)
2591 {
2592         struct inode *dir = NULL;
2593         unsigned n;
2594         spin_lock(&dentry->d_lock);
2595         if (unlikely(d_in_lookup(dentry))) {
2596                 dir = dentry->d_parent->d_inode;
2597                 n = start_dir_add(dir);
2598                 __d_lookup_done(dentry);
2599         }
2600         if (inode) {
2601                 unsigned add_flags = d_flags_for_inode(inode);
2602                 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2603                 raw_write_seqcount_begin(&dentry->d_seq);
2604                 __d_set_inode_and_type(dentry, inode, add_flags);
2605                 raw_write_seqcount_end(&dentry->d_seq);
2606                 fsnotify_update_flags(dentry);
2607         }
2608         __d_rehash(dentry);
2609         if (dir)
2610                 end_dir_add(dir, n);
2611         spin_unlock(&dentry->d_lock);
2612         if (inode)
2613                 spin_unlock(&inode->i_lock);
2614 }
2615
2616 /**
2617  * d_add - add dentry to hash queues
2618  * @entry: dentry to add
2619  * @inode: The inode to attach to this dentry
2620  *
2621  * This adds the entry to the hash queues and initializes @inode.
2622  * The entry was actually filled in earlier during d_alloc().
2623  */
2624
2625 void d_add(struct dentry *entry, struct inode *inode)
2626 {
2627         if (inode) {
2628                 security_d_instantiate(entry, inode);
2629                 spin_lock(&inode->i_lock);
2630         }
2631         __d_add(entry, inode);
2632 }
2633 EXPORT_SYMBOL(d_add);
2634
2635 /**
2636  * d_exact_alias - find and hash an exact unhashed alias
2637  * @entry: dentry to add
2638  * @inode: The inode to go with this dentry
2639  *
2640  * If an unhashed dentry with the same name/parent and desired
2641  * inode already exists, hash and return it.  Otherwise, return
2642  * NULL.
2643  *
2644  * Parent directory should be locked.
2645  */
2646 struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2647 {
2648         struct dentry *alias;
2649         unsigned int hash = entry->d_name.hash;
2650
2651         spin_lock(&inode->i_lock);
2652         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2653                 /*
2654                  * Don't need alias->d_lock here, because aliases with
2655                  * d_parent == entry->d_parent are not subject to name or
2656                  * parent changes, because the parent inode i_mutex is held.
2657                  */
2658                 if (alias->d_name.hash != hash)
2659                         continue;
2660                 if (alias->d_parent != entry->d_parent)
2661                         continue;
2662                 if (!d_same_name(alias, entry->d_parent, &entry->d_name))
2663                         continue;
2664                 spin_lock(&alias->d_lock);
2665                 if (!d_unhashed(alias)) {
2666                         spin_unlock(&alias->d_lock);
2667                         alias = NULL;
2668                 } else {
2669                         __dget_dlock(alias);
2670                         __d_rehash(alias);
2671                         spin_unlock(&alias->d_lock);
2672                 }
2673                 spin_unlock(&inode->i_lock);
2674                 return alias;
2675         }
2676         spin_unlock(&inode->i_lock);
2677         return NULL;
2678 }
2679 EXPORT_SYMBOL(d_exact_alias);
2680
2681 /**
2682  * dentry_update_name_case - update case insensitive dentry with a new name
2683  * @dentry: dentry to be updated
2684  * @name: new name
2685  *
2686  * Update a case insensitive dentry with new case of name.
2687  *
2688  * dentry must have been returned by d_lookup with name @name. Old and new
2689  * name lengths must match (ie. no d_compare which allows mismatched name
2690  * lengths).
2691  *
2692  * Parent inode i_mutex must be held over d_lookup and into this call (to
2693  * keep renames and concurrent inserts, and readdir(2) away).
2694  */
2695 void dentry_update_name_case(struct dentry *dentry, const struct qstr *name)
2696 {
2697         BUG_ON(!inode_is_locked(dentry->d_parent->d_inode));
2698         BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */
2699
2700         spin_lock(&dentry->d_lock);
2701         write_seqcount_begin(&dentry->d_seq);
2702         memcpy((unsigned char *)dentry->d_name.name, name->name, name->len);
2703         write_seqcount_end(&dentry->d_seq);
2704         spin_unlock(&dentry->d_lock);
2705 }
2706 EXPORT_SYMBOL(dentry_update_name_case);
2707
2708 static void swap_names(struct dentry *dentry, struct dentry *target)
2709 {
2710         if (unlikely(dname_external(target))) {
2711                 if (unlikely(dname_external(dentry))) {
2712                         /*
2713                          * Both external: swap the pointers
2714                          */
2715                         swap(target->d_name.name, dentry->d_name.name);
2716                 } else {
2717                         /*
2718                          * dentry:internal, target:external.  Steal target's
2719                          * storage and make target internal.
2720                          */
2721                         memcpy(target->d_iname, dentry->d_name.name,
2722                                         dentry->d_name.len + 1);
2723                         dentry->d_name.name = target->d_name.name;
2724                         target->d_name.name = target->d_iname;
2725                 }
2726         } else {
2727                 if (unlikely(dname_external(dentry))) {
2728                         /*
2729                          * dentry:external, target:internal.  Give dentry's
2730                          * storage to target and make dentry internal
2731                          */
2732                         memcpy(dentry->d_iname, target->d_name.name,
2733                                         target->d_name.len + 1);
2734                         target->d_name.name = dentry->d_name.name;
2735                         dentry->d_name.name = dentry->d_iname;
2736                 } else {
2737                         /*
2738                          * Both are internal.
2739                          */
2740                         unsigned int i;
2741                         BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2742                         for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2743                                 swap(((long *) &dentry->d_iname)[i],
2744                                      ((long *) &target->d_iname)[i]);
2745                         }
2746                 }
2747         }
2748         swap(dentry->d_name.hash_len, target->d_name.hash_len);
2749 }
2750
2751 static void copy_name(struct dentry *dentry, struct dentry *target)
2752 {
2753         struct external_name *old_name = NULL;
2754         if (unlikely(dname_external(dentry)))
2755                 old_name = external_name(dentry);
2756         if (unlikely(dname_external(target))) {
2757                 atomic_inc(&external_name(target)->u.count);
2758                 dentry->d_name = target->d_name;
2759         } else {
2760                 memcpy(dentry->d_iname, target->d_name.name,
2761                                 target->d_name.len + 1);
2762                 dentry->d_name.name = dentry->d_iname;
2763                 dentry->d_name.hash_len = target->d_name.hash_len;
2764         }
2765         if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2766                 call_rcu(&old_name->u.head, __d_free_external_name);
2767 }
2768
2769 /*
2770  * __d_move - move a dentry
2771  * @dentry: entry to move
2772  * @target: new dentry
2773  * @exchange: exchange the two dentries
2774  *
2775  * Update the dcache to reflect the move of a file name. Negative
2776  * dcache entries should not be moved in this way. Caller must hold
2777  * rename_lock, the i_mutex of the source and target directories,
2778  * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2779  */
2780 static void __d_move(struct dentry *dentry, struct dentry *target,
2781                      bool exchange)
2782 {
2783         struct dentry *old_parent, *p;
2784         struct inode *dir = NULL;
2785         unsigned n;
2786
2787         WARN_ON(!dentry->d_inode);
2788         if (WARN_ON(dentry == target))
2789                 return;
2790
2791         BUG_ON(d_ancestor(target, dentry));
2792         old_parent = dentry->d_parent;
2793         p = d_ancestor(old_parent, target);
2794         if (IS_ROOT(dentry)) {
2795                 BUG_ON(p);
2796                 spin_lock(&target->d_parent->d_lock);
2797         } else if (!p) {
2798                 /* target is not a descendent of dentry->d_parent */
2799                 spin_lock(&target->d_parent->d_lock);
2800                 spin_lock_nested(&old_parent->d_lock, DENTRY_D_LOCK_NESTED);
2801         } else {
2802                 BUG_ON(p == dentry);
2803                 spin_lock(&old_parent->d_lock);
2804                 if (p != target)
2805                         spin_lock_nested(&target->d_parent->d_lock,
2806                                         DENTRY_D_LOCK_NESTED);
2807         }
2808         spin_lock_nested(&dentry->d_lock, 2);
2809         spin_lock_nested(&target->d_lock, 3);
2810
2811         if (unlikely(d_in_lookup(target))) {
2812                 dir = target->d_parent->d_inode;
2813                 n = start_dir_add(dir);
2814                 __d_lookup_done(target);
2815         }
2816
2817         write_seqcount_begin(&dentry->d_seq);
2818         write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2819
2820         /* unhash both */
2821         if (!d_unhashed(dentry))
2822                 ___d_drop(dentry);
2823         if (!d_unhashed(target))
2824                 ___d_drop(target);
2825
2826         /* ... and switch them in the tree */
2827         dentry->d_parent = target->d_parent;
2828         if (!exchange) {
2829                 copy_name(dentry, target);
2830                 target->d_hash.pprev = NULL;
2831                 dentry->d_parent->d_lockref.count++;
2832                 if (dentry == old_parent)
2833                         dentry->d_flags |= DCACHE_RCUACCESS;
2834                 else
2835                         WARN_ON(!--old_parent->d_lockref.count);
2836         } else {
2837                 target->d_parent = old_parent;
2838                 swap_names(dentry, target);
2839                 list_move(&target->d_child, &target->d_parent->d_subdirs);
2840                 __d_rehash(target);
2841                 fsnotify_update_flags(target);
2842         }
2843         list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2844         __d_rehash(dentry);
2845         fsnotify_update_flags(dentry);
2846
2847         write_seqcount_end(&target->d_seq);
2848         write_seqcount_end(&dentry->d_seq);
2849
2850         if (dir)
2851                 end_dir_add(dir, n);
2852
2853         if (dentry->d_parent != old_parent)
2854                 spin_unlock(&dentry->d_parent->d_lock);
2855         if (dentry != old_parent)
2856                 spin_unlock(&old_parent->d_lock);
2857         spin_unlock(&target->d_lock);
2858         spin_unlock(&dentry->d_lock);
2859 }
2860
2861 /*
2862  * d_move - move a dentry
2863  * @dentry: entry to move
2864  * @target: new dentry
2865  *
2866  * Update the dcache to reflect the move of a file name. Negative
2867  * dcache entries should not be moved in this way. See the locking
2868  * requirements for __d_move.
2869  */
2870 void d_move(struct dentry *dentry, struct dentry *target)
2871 {
2872         write_seqlock(&rename_lock);
2873         __d_move(dentry, target, false);
2874         write_sequnlock(&rename_lock);
2875 }
2876 EXPORT_SYMBOL(d_move);
2877
2878 /*
2879  * d_exchange - exchange two dentries
2880  * @dentry1: first dentry
2881  * @dentry2: second dentry
2882  */
2883 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2884 {
2885         write_seqlock(&rename_lock);
2886
2887         WARN_ON(!dentry1->d_inode);
2888         WARN_ON(!dentry2->d_inode);
2889         WARN_ON(IS_ROOT(dentry1));
2890         WARN_ON(IS_ROOT(dentry2));
2891
2892         __d_move(dentry1, dentry2, true);
2893
2894         write_sequnlock(&rename_lock);
2895 }
2896
2897 /**
2898  * d_ancestor - search for an ancestor
2899  * @p1: ancestor dentry
2900  * @p2: child dentry
2901  *
2902  * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2903  * an ancestor of p2, else NULL.
2904  */
2905 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2906 {
2907         struct dentry *p;
2908
2909         for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2910                 if (p->d_parent == p1)
2911                         return p;
2912         }
2913         return NULL;
2914 }
2915
2916 /*
2917  * This helper attempts to cope with remotely renamed directories
2918  *
2919  * It assumes that the caller is already holding
2920  * dentry->d_parent->d_inode->i_mutex, and rename_lock
2921  *
2922  * Note: If ever the locking in lock_rename() changes, then please
2923  * remember to update this too...
2924  */
2925 static int __d_unalias(struct inode *inode,
2926                 struct dentry *dentry, struct dentry *alias)
2927 {
2928         struct mutex *m1 = NULL;
2929         struct rw_semaphore *m2 = NULL;
2930         int ret = -ESTALE;
2931
2932         /* If alias and dentry share a parent, then no extra locks required */
2933         if (alias->d_parent == dentry->d_parent)
2934                 goto out_unalias;
2935
2936         /* See lock_rename() */
2937         if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2938                 goto out_err;
2939         m1 = &dentry->d_sb->s_vfs_rename_mutex;
2940         if (!inode_trylock_shared(alias->d_parent->d_inode))
2941                 goto out_err;
2942         m2 = &alias->d_parent->d_inode->i_rwsem;
2943 out_unalias:
2944         __d_move(alias, dentry, false);
2945         ret = 0;
2946 out_err:
2947         if (m2)
2948                 up_read(m2);
2949         if (m1)
2950                 mutex_unlock(m1);
2951         return ret;
2952 }
2953
2954 /**
2955  * d_splice_alias - splice a disconnected dentry into the tree if one exists
2956  * @inode:  the inode which may have a disconnected dentry
2957  * @dentry: a negative dentry which we want to point to the inode.
2958  *
2959  * If inode is a directory and has an IS_ROOT alias, then d_move that in
2960  * place of the given dentry and return it, else simply d_add the inode
2961  * to the dentry and return NULL.
2962  *
2963  * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2964  * we should error out: directories can't have multiple aliases.
2965  *
2966  * This is needed in the lookup routine of any filesystem that is exportable
2967  * (via knfsd) so that we can build dcache paths to directories effectively.
2968  *
2969  * If a dentry was found and moved, then it is returned.  Otherwise NULL
2970  * is returned.  This matches the expected return value of ->lookup.
2971  *
2972  * Cluster filesystems may call this function with a negative, hashed dentry.
2973  * In that case, we know that the inode will be a regular file, and also this
2974  * will only occur during atomic_open. So we need to check for the dentry
2975  * being already hashed only in the final case.
2976  */
2977 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
2978 {
2979         if (IS_ERR(inode))
2980                 return ERR_CAST(inode);
2981
2982         BUG_ON(!d_unhashed(dentry));
2983
2984         if (!inode)
2985                 goto out;
2986
2987         security_d_instantiate(dentry, inode);
2988         spin_lock(&inode->i_lock);
2989         if (S_ISDIR(inode->i_mode)) {
2990                 struct dentry *new = __d_find_any_alias(inode);
2991                 if (unlikely(new)) {
2992                         /* The reference to new ensures it remains an alias */
2993                         spin_unlock(&inode->i_lock);
2994                         write_seqlock(&rename_lock);
2995                         if (unlikely(d_ancestor(new, dentry))) {
2996                                 write_sequnlock(&rename_lock);
2997                                 dput(new);
2998                                 new = ERR_PTR(-ELOOP);
2999                                 pr_warn_ratelimited(
3000                                         "VFS: Lookup of '%s' in %s %s"
3001                                         " would have caused loop\n",
3002                                         dentry->d_name.name,
3003                                         inode->i_sb->s_type->name,
3004                                         inode->i_sb->s_id);
3005                         } else if (!IS_ROOT(new)) {
3006                                 struct dentry *old_parent = dget(new->d_parent);
3007                                 int err = __d_unalias(inode, dentry, new);
3008                                 write_sequnlock(&rename_lock);
3009                                 if (err) {
3010                                         dput(new);
3011                                         new = ERR_PTR(err);
3012                                 }
3013                                 dput(old_parent);
3014                         } else {
3015                                 __d_move(new, dentry, false);
3016                                 write_sequnlock(&rename_lock);
3017                         }
3018                         iput(inode);
3019                         return new;
3020                 }
3021         }
3022 out:
3023         __d_add(dentry, inode);
3024         return NULL;
3025 }
3026 EXPORT_SYMBOL(d_splice_alias);
3027
3028 /*
3029  * Test whether new_dentry is a subdirectory of old_dentry.
3030  *
3031  * Trivially implemented using the dcache structure
3032  */
3033
3034 /**
3035  * is_subdir - is new dentry a subdirectory of old_dentry
3036  * @new_dentry: new dentry
3037  * @old_dentry: old dentry
3038  *
3039  * Returns true if new_dentry is a subdirectory of the parent (at any depth).
3040  * Returns false otherwise.
3041  * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3042  */
3043   
3044 bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3045 {
3046         bool result;
3047         unsigned seq;
3048
3049         if (new_dentry == old_dentry)
3050                 return true;
3051
3052         do {
3053                 /* for restarting inner loop in case of seq retry */
3054                 seq = read_seqbegin(&rename_lock);
3055                 /*
3056                  * Need rcu_readlock to protect against the d_parent trashing
3057                  * due to d_move
3058                  */
3059                 rcu_read_lock();
3060                 if (d_ancestor(old_dentry, new_dentry))
3061                         result = true;
3062                 else
3063                         result = false;
3064                 rcu_read_unlock();
3065         } while (read_seqretry(&rename_lock, seq));
3066
3067         return result;
3068 }
3069 EXPORT_SYMBOL(is_subdir);
3070
3071 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3072 {
3073         struct dentry *root = data;
3074         if (dentry != root) {
3075                 if (d_unhashed(dentry) || !dentry->d_inode)
3076                         return D_WALK_SKIP;
3077
3078                 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3079                         dentry->d_flags |= DCACHE_GENOCIDE;
3080                         dentry->d_lockref.count--;
3081                 }
3082         }
3083         return D_WALK_CONTINUE;
3084 }
3085
3086 void d_genocide(struct dentry *parent)
3087 {
3088         d_walk(parent, parent, d_genocide_kill);
3089 }
3090
3091 EXPORT_SYMBOL(d_genocide);
3092
3093 void d_tmpfile(struct dentry *dentry, struct inode *inode)
3094 {
3095         inode_dec_link_count(inode);
3096         BUG_ON(dentry->d_name.name != dentry->d_iname ||
3097                 !hlist_unhashed(&dentry->d_u.d_alias) ||
3098                 !d_unlinked(dentry));
3099         spin_lock(&dentry->d_parent->d_lock);
3100         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3101         dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3102                                 (unsigned long long)inode->i_ino);
3103         spin_unlock(&dentry->d_lock);
3104         spin_unlock(&dentry->d_parent->d_lock);
3105         d_instantiate(dentry, inode);
3106 }
3107 EXPORT_SYMBOL(d_tmpfile);
3108
3109 static __initdata unsigned long dhash_entries;
3110 static int __init set_dhash_entries(char *str)
3111 {
3112         if (!str)
3113                 return 0;
3114         dhash_entries = simple_strtoul(str, &str, 0);
3115         return 1;
3116 }
3117 __setup("dhash_entries=", set_dhash_entries);
3118
3119 static void __init dcache_init_early(void)
3120 {
3121         /* If hashes are distributed across NUMA nodes, defer
3122          * hash allocation until vmalloc space is available.
3123          */
3124         if (hashdist)
3125                 return;
3126
3127         dentry_hashtable =
3128                 alloc_large_system_hash("Dentry cache",
3129                                         sizeof(struct hlist_bl_head),
3130                                         dhash_entries,
3131                                         13,
3132                                         HASH_EARLY | HASH_ZERO,
3133                                         &d_hash_shift,
3134                                         NULL,
3135                                         0,
3136                                         0);
3137         d_hash_shift = 32 - d_hash_shift;
3138 }
3139
3140 static void __init dcache_init(void)
3141 {
3142         /*
3143          * A constructor could be added for stable state like the lists,
3144          * but it is probably not worth it because of the cache nature
3145          * of the dcache.
3146          */
3147         dentry_cache = KMEM_CACHE_USERCOPY(dentry,
3148                 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD|SLAB_ACCOUNT,
3149                 d_iname);
3150
3151         /* Hash may have been set up in dcache_init_early */
3152         if (!hashdist)
3153                 return;
3154
3155         dentry_hashtable =
3156                 alloc_large_system_hash("Dentry cache",
3157                                         sizeof(struct hlist_bl_head),
3158                                         dhash_entries,
3159                                         13,
3160                                         HASH_ZERO,
3161                                         &d_hash_shift,
3162                                         NULL,
3163                                         0,
3164                                         0);
3165         d_hash_shift = 32 - d_hash_shift;
3166 }
3167
3168 /* SLAB cache for __getname() consumers */
3169 struct kmem_cache *names_cachep __read_mostly;
3170 EXPORT_SYMBOL(names_cachep);
3171
3172 void __init vfs_caches_init_early(void)
3173 {
3174         int i;
3175
3176         for (i = 0; i < ARRAY_SIZE(in_lookup_hashtable); i++)
3177                 INIT_HLIST_BL_HEAD(&in_lookup_hashtable[i]);
3178
3179         dcache_init_early();
3180         inode_init_early();
3181 }
3182
3183 void __init vfs_caches_init(void)
3184 {
3185         names_cachep = kmem_cache_create_usercopy("names_cache", PATH_MAX, 0,
3186                         SLAB_HWCACHE_ALIGN|SLAB_PANIC, 0, PATH_MAX, NULL);
3187
3188         dcache_init();
3189         inode_init();
3190         files_init();
3191         files_maxfiles_init();
3192         mnt_init();
3193         bdev_cache_init();
3194         chrdev_init();
3195 }