Merge git://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf
[linux-2.6-microblaze.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
362         raw_write_seqcount_begin(&dentry->d_seq);
363         __d_clear_type_and_inode(dentry);
364         hlist_del_init(&dentry->d_u.d_alias);
365         raw_write_seqcount_end(&dentry->d_seq);
366         spin_unlock(&dentry->d_lock);
367         spin_unlock(&inode->i_lock);
368         if (!inode->i_nlink)
369                 fsnotify_inoderemove(inode);
370         if (dentry->d_op && dentry->d_op->d_iput)
371                 dentry->d_op->d_iput(dentry, inode);
372         else
373                 iput(inode);
374 }
375
376 /*
377  * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
378  * is in use - which includes both the "real" per-superblock
379  * LRU list _and_ the DCACHE_SHRINK_LIST use.
380  *
381  * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
382  * on the shrink list (ie not on the superblock LRU list).
383  *
384  * The per-cpu "nr_dentry_unused" counters are updated with
385  * the DCACHE_LRU_LIST bit.
386  *
387  * These helper functions make sure we always follow the
388  * rules. d_lock must be held by the caller.
389  */
390 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
391 static void d_lru_add(struct dentry *dentry)
392 {
393         D_FLAG_VERIFY(dentry, 0);
394         dentry->d_flags |= DCACHE_LRU_LIST;
395         this_cpu_inc(nr_dentry_unused);
396         WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
397 }
398
399 static void d_lru_del(struct dentry *dentry)
400 {
401         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
402         dentry->d_flags &= ~DCACHE_LRU_LIST;
403         this_cpu_dec(nr_dentry_unused);
404         WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
405 }
406
407 static void d_shrink_del(struct dentry *dentry)
408 {
409         D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
410         list_del_init(&dentry->d_lru);
411         dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
412         this_cpu_dec(nr_dentry_unused);
413 }
414
415 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
416 {
417         D_FLAG_VERIFY(dentry, 0);
418         list_add(&dentry->d_lru, list);
419         dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
420         this_cpu_inc(nr_dentry_unused);
421 }
422
423 /*
424  * These can only be called under the global LRU lock, ie during the
425  * callback for freeing the LRU list. "isolate" removes it from the
426  * LRU lists entirely, while shrink_move moves it to the indicated
427  * private list.
428  */
429 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
430 {
431         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
432         dentry->d_flags &= ~DCACHE_LRU_LIST;
433         this_cpu_dec(nr_dentry_unused);
434         list_lru_isolate(lru, &dentry->d_lru);
435 }
436
437 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
438                               struct list_head *list)
439 {
440         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
441         dentry->d_flags |= DCACHE_SHRINK_LIST;
442         list_lru_isolate_move(lru, &dentry->d_lru, list);
443 }
444
445 /**
446  * d_drop - drop a dentry
447  * @dentry: dentry to drop
448  *
449  * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
450  * be found through a VFS lookup any more. Note that this is different from
451  * deleting the dentry - d_delete will try to mark the dentry negative if
452  * possible, giving a successful _negative_ lookup, while d_drop will
453  * just make the cache lookup fail.
454  *
455  * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
456  * reason (NFS timeouts or autofs deletes).
457  *
458  * __d_drop requires dentry->d_lock
459  * ___d_drop doesn't mark dentry as "unhashed"
460  *   (dentry->d_hash.pprev will be LIST_POISON2, not NULL).
461  */
462 static void ___d_drop(struct dentry *dentry)
463 {
464         struct hlist_bl_head *b;
465         /*
466          * Hashed dentries are normally on the dentry hashtable,
467          * with the exception of those newly allocated by
468          * d_obtain_root, which are always IS_ROOT:
469          */
470         if (unlikely(IS_ROOT(dentry)))
471                 b = &dentry->d_sb->s_roots;
472         else
473                 b = d_hash(dentry->d_name.hash);
474
475         hlist_bl_lock(b);
476         __hlist_bl_del(&dentry->d_hash);
477         hlist_bl_unlock(b);
478 }
479
480 void __d_drop(struct dentry *dentry)
481 {
482         if (!d_unhashed(dentry)) {
483                 ___d_drop(dentry);
484                 dentry->d_hash.pprev = NULL;
485                 write_seqcount_invalidate(&dentry->d_seq);
486         }
487 }
488 EXPORT_SYMBOL(__d_drop);
489
490 void d_drop(struct dentry *dentry)
491 {
492         spin_lock(&dentry->d_lock);
493         __d_drop(dentry);
494         spin_unlock(&dentry->d_lock);
495 }
496 EXPORT_SYMBOL(d_drop);
497
498 static inline void dentry_unlist(struct dentry *dentry, struct dentry *parent)
499 {
500         struct dentry *next;
501         /*
502          * Inform d_walk() and shrink_dentry_list() that we are no longer
503          * attached to the dentry tree
504          */
505         dentry->d_flags |= DCACHE_DENTRY_KILLED;
506         if (unlikely(list_empty(&dentry->d_child)))
507                 return;
508         __list_del_entry(&dentry->d_child);
509         /*
510          * Cursors can move around the list of children.  While we'd been
511          * a normal list member, it didn't matter - ->d_child.next would've
512          * been updated.  However, from now on it won't be and for the
513          * things like d_walk() it might end up with a nasty surprise.
514          * Normally d_walk() doesn't care about cursors moving around -
515          * ->d_lock on parent prevents that and since a cursor has no children
516          * of its own, we get through it without ever unlocking the parent.
517          * There is one exception, though - if we ascend from a child that
518          * gets killed as soon as we unlock it, the next sibling is found
519          * using the value left in its ->d_child.next.  And if _that_
520          * pointed to a cursor, and cursor got moved (e.g. by lseek())
521          * before d_walk() regains parent->d_lock, we'll end up skipping
522          * everything the cursor had been moved past.
523          *
524          * Solution: make sure that the pointer left behind in ->d_child.next
525          * points to something that won't be moving around.  I.e. skip the
526          * cursors.
527          */
528         while (dentry->d_child.next != &parent->d_subdirs) {
529                 next = list_entry(dentry->d_child.next, struct dentry, d_child);
530                 if (likely(!(next->d_flags & DCACHE_DENTRY_CURSOR)))
531                         break;
532                 dentry->d_child.next = next->d_child.next;
533         }
534 }
535
536 static void __dentry_kill(struct dentry *dentry)
537 {
538         struct dentry *parent = NULL;
539         bool can_free = true;
540         if (!IS_ROOT(dentry))
541                 parent = dentry->d_parent;
542
543         /*
544          * The dentry is now unrecoverably dead to the world.
545          */
546         lockref_mark_dead(&dentry->d_lockref);
547
548         /*
549          * inform the fs via d_prune that this dentry is about to be
550          * unhashed and destroyed.
551          */
552         if (dentry->d_flags & DCACHE_OP_PRUNE)
553                 dentry->d_op->d_prune(dentry);
554
555         if (dentry->d_flags & DCACHE_LRU_LIST) {
556                 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
557                         d_lru_del(dentry);
558         }
559         /* if it was on the hash then remove it */
560         __d_drop(dentry);
561         dentry_unlist(dentry, parent);
562         if (parent)
563                 spin_unlock(&parent->d_lock);
564         if (dentry->d_inode)
565                 dentry_unlink_inode(dentry);
566         else
567                 spin_unlock(&dentry->d_lock);
568         this_cpu_dec(nr_dentry);
569         if (dentry->d_op && dentry->d_op->d_release)
570                 dentry->d_op->d_release(dentry);
571
572         spin_lock(&dentry->d_lock);
573         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
574                 dentry->d_flags |= DCACHE_MAY_FREE;
575                 can_free = false;
576         }
577         spin_unlock(&dentry->d_lock);
578         if (likely(can_free))
579                 dentry_free(dentry);
580         cond_resched();
581 }
582
583 static struct dentry *__lock_parent(struct dentry *dentry)
584 {
585         struct dentry *parent;
586         rcu_read_lock();
587         spin_unlock(&dentry->d_lock);
588 again:
589         parent = READ_ONCE(dentry->d_parent);
590         spin_lock(&parent->d_lock);
591         /*
592          * We can't blindly lock dentry until we are sure
593          * that we won't violate the locking order.
594          * Any changes of dentry->d_parent must have
595          * been done with parent->d_lock held, so
596          * spin_lock() above is enough of a barrier
597          * for checking if it's still our child.
598          */
599         if (unlikely(parent != dentry->d_parent)) {
600                 spin_unlock(&parent->d_lock);
601                 goto again;
602         }
603         rcu_read_unlock();
604         if (parent != dentry)
605                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
606         else
607                 parent = NULL;
608         return parent;
609 }
610
611 static inline struct dentry *lock_parent(struct dentry *dentry)
612 {
613         struct dentry *parent = dentry->d_parent;
614         if (IS_ROOT(dentry))
615                 return NULL;
616         if (likely(spin_trylock(&parent->d_lock)))
617                 return parent;
618         return __lock_parent(dentry);
619 }
620
621 static inline bool retain_dentry(struct dentry *dentry)
622 {
623         WARN_ON(d_in_lookup(dentry));
624
625         /* Unreachable? Get rid of it */
626         if (unlikely(d_unhashed(dentry)))
627                 return false;
628
629         if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
630                 return false;
631
632         if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
633                 if (dentry->d_op->d_delete(dentry))
634                         return false;
635         }
636         /* retain; LRU fodder */
637         dentry->d_lockref.count--;
638         if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
639                 d_lru_add(dentry);
640         else if (unlikely(!(dentry->d_flags & DCACHE_REFERENCED)))
641                 dentry->d_flags |= DCACHE_REFERENCED;
642         return true;
643 }
644
645 /*
646  * Finish off a dentry we've decided to kill.
647  * dentry->d_lock must be held, returns with it unlocked.
648  * Returns dentry requiring refcount drop, or NULL if we're done.
649  */
650 static struct dentry *dentry_kill(struct dentry *dentry)
651         __releases(dentry->d_lock)
652 {
653         struct inode *inode = dentry->d_inode;
654         struct dentry *parent = NULL;
655
656         if (inode && unlikely(!spin_trylock(&inode->i_lock)))
657                 goto slow_positive;
658
659         if (!IS_ROOT(dentry)) {
660                 parent = dentry->d_parent;
661                 if (unlikely(!spin_trylock(&parent->d_lock))) {
662                         parent = __lock_parent(dentry);
663                         if (likely(inode || !dentry->d_inode))
664                                 goto got_locks;
665                         /* negative that became positive */
666                         if (parent)
667                                 spin_unlock(&parent->d_lock);
668                         inode = dentry->d_inode;
669                         goto slow_positive;
670                 }
671         }
672         __dentry_kill(dentry);
673         return parent;
674
675 slow_positive:
676         spin_unlock(&dentry->d_lock);
677         spin_lock(&inode->i_lock);
678         spin_lock(&dentry->d_lock);
679         parent = lock_parent(dentry);
680 got_locks:
681         if (unlikely(dentry->d_lockref.count != 1)) {
682                 dentry->d_lockref.count--;
683         } else if (likely(!retain_dentry(dentry))) {
684                 __dentry_kill(dentry);
685                 return parent;
686         }
687         /* we are keeping it, after all */
688         if (inode)
689                 spin_unlock(&inode->i_lock);
690         if (parent)
691                 spin_unlock(&parent->d_lock);
692         spin_unlock(&dentry->d_lock);
693         return NULL;
694 }
695
696 /*
697  * Try to do a lockless dput(), and return whether that was successful.
698  *
699  * If unsuccessful, we return false, having already taken the dentry lock.
700  *
701  * The caller needs to hold the RCU read lock, so that the dentry is
702  * guaranteed to stay around even if the refcount goes down to zero!
703  */
704 static inline bool fast_dput(struct dentry *dentry)
705 {
706         int ret;
707         unsigned int d_flags;
708
709         /*
710          * If we have a d_op->d_delete() operation, we sould not
711          * let the dentry count go to zero, so use "put_or_lock".
712          */
713         if (unlikely(dentry->d_flags & DCACHE_OP_DELETE))
714                 return lockref_put_or_lock(&dentry->d_lockref);
715
716         /*
717          * .. otherwise, we can try to just decrement the
718          * lockref optimistically.
719          */
720         ret = lockref_put_return(&dentry->d_lockref);
721
722         /*
723          * If the lockref_put_return() failed due to the lock being held
724          * by somebody else, the fast path has failed. We will need to
725          * get the lock, and then check the count again.
726          */
727         if (unlikely(ret < 0)) {
728                 spin_lock(&dentry->d_lock);
729                 if (dentry->d_lockref.count > 1) {
730                         dentry->d_lockref.count--;
731                         spin_unlock(&dentry->d_lock);
732                         return true;
733                 }
734                 return false;
735         }
736
737         /*
738          * If we weren't the last ref, we're done.
739          */
740         if (ret)
741                 return true;
742
743         /*
744          * Careful, careful. The reference count went down
745          * to zero, but we don't hold the dentry lock, so
746          * somebody else could get it again, and do another
747          * dput(), and we need to not race with that.
748          *
749          * However, there is a very special and common case
750          * where we don't care, because there is nothing to
751          * do: the dentry is still hashed, it does not have
752          * a 'delete' op, and it's referenced and already on
753          * the LRU list.
754          *
755          * NOTE! Since we aren't locked, these values are
756          * not "stable". However, it is sufficient that at
757          * some point after we dropped the reference the
758          * dentry was hashed and the flags had the proper
759          * value. Other dentry users may have re-gotten
760          * a reference to the dentry and change that, but
761          * our work is done - we can leave the dentry
762          * around with a zero refcount.
763          */
764         smp_rmb();
765         d_flags = READ_ONCE(dentry->d_flags);
766         d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST | DCACHE_DISCONNECTED;
767
768         /* Nothing to do? Dropping the reference was all we needed? */
769         if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry))
770                 return true;
771
772         /*
773          * Not the fast normal case? Get the lock. We've already decremented
774          * the refcount, but we'll need to re-check the situation after
775          * getting the lock.
776          */
777         spin_lock(&dentry->d_lock);
778
779         /*
780          * Did somebody else grab a reference to it in the meantime, and
781          * we're no longer the last user after all? Alternatively, somebody
782          * else could have killed it and marked it dead. Either way, we
783          * don't need to do anything else.
784          */
785         if (dentry->d_lockref.count) {
786                 spin_unlock(&dentry->d_lock);
787                 return true;
788         }
789
790         /*
791          * Re-get the reference we optimistically dropped. We hold the
792          * lock, and we just tested that it was zero, so we can just
793          * set it to 1.
794          */
795         dentry->d_lockref.count = 1;
796         return false;
797 }
798
799
800 /* 
801  * This is dput
802  *
803  * This is complicated by the fact that we do not want to put
804  * dentries that are no longer on any hash chain on the unused
805  * list: we'd much rather just get rid of them immediately.
806  *
807  * However, that implies that we have to traverse the dentry
808  * tree upwards to the parents which might _also_ now be
809  * scheduled for deletion (it may have been only waiting for
810  * its last child to go away).
811  *
812  * This tail recursion is done by hand as we don't want to depend
813  * on the compiler to always get this right (gcc generally doesn't).
814  * Real recursion would eat up our stack space.
815  */
816
817 /*
818  * dput - release a dentry
819  * @dentry: dentry to release 
820  *
821  * Release a dentry. This will drop the usage count and if appropriate
822  * call the dentry unlink method as well as removing it from the queues and
823  * releasing its resources. If the parent dentries were scheduled for release
824  * they too may now get deleted.
825  */
826 void dput(struct dentry *dentry)
827 {
828         while (dentry) {
829                 might_sleep();
830
831                 rcu_read_lock();
832                 if (likely(fast_dput(dentry))) {
833                         rcu_read_unlock();
834                         return;
835                 }
836
837                 /* Slow case: now with the dentry lock held */
838                 rcu_read_unlock();
839
840                 if (likely(retain_dentry(dentry))) {
841                         spin_unlock(&dentry->d_lock);
842                         return;
843                 }
844
845                 dentry = dentry_kill(dentry);
846         }
847 }
848 EXPORT_SYMBOL(dput);
849
850
851 /* This must be called with d_lock held */
852 static inline void __dget_dlock(struct dentry *dentry)
853 {
854         dentry->d_lockref.count++;
855 }
856
857 static inline void __dget(struct dentry *dentry)
858 {
859         lockref_get(&dentry->d_lockref);
860 }
861
862 struct dentry *dget_parent(struct dentry *dentry)
863 {
864         int gotref;
865         struct dentry *ret;
866
867         /*
868          * Do optimistic parent lookup without any
869          * locking.
870          */
871         rcu_read_lock();
872         ret = READ_ONCE(dentry->d_parent);
873         gotref = lockref_get_not_zero(&ret->d_lockref);
874         rcu_read_unlock();
875         if (likely(gotref)) {
876                 if (likely(ret == READ_ONCE(dentry->d_parent)))
877                         return ret;
878                 dput(ret);
879         }
880
881 repeat:
882         /*
883          * Don't need rcu_dereference because we re-check it was correct under
884          * the lock.
885          */
886         rcu_read_lock();
887         ret = dentry->d_parent;
888         spin_lock(&ret->d_lock);
889         if (unlikely(ret != dentry->d_parent)) {
890                 spin_unlock(&ret->d_lock);
891                 rcu_read_unlock();
892                 goto repeat;
893         }
894         rcu_read_unlock();
895         BUG_ON(!ret->d_lockref.count);
896         ret->d_lockref.count++;
897         spin_unlock(&ret->d_lock);
898         return ret;
899 }
900 EXPORT_SYMBOL(dget_parent);
901
902 static struct dentry * __d_find_any_alias(struct inode *inode)
903 {
904         struct dentry *alias;
905
906         if (hlist_empty(&inode->i_dentry))
907                 return NULL;
908         alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
909         __dget(alias);
910         return alias;
911 }
912
913 /**
914  * d_find_any_alias - find any alias for a given inode
915  * @inode: inode to find an alias for
916  *
917  * If any aliases exist for the given inode, take and return a
918  * reference for one of them.  If no aliases exist, return %NULL.
919  */
920 struct dentry *d_find_any_alias(struct inode *inode)
921 {
922         struct dentry *de;
923
924         spin_lock(&inode->i_lock);
925         de = __d_find_any_alias(inode);
926         spin_unlock(&inode->i_lock);
927         return de;
928 }
929 EXPORT_SYMBOL(d_find_any_alias);
930
931 /**
932  * d_find_alias - grab a hashed alias of inode
933  * @inode: inode in question
934  *
935  * If inode has a hashed alias, or is a directory and has any alias,
936  * acquire the reference to alias and return it. Otherwise return NULL.
937  * Notice that if inode is a directory there can be only one alias and
938  * it can be unhashed only if it has no children, or if it is the root
939  * of a filesystem, or if the directory was renamed and d_revalidate
940  * was the first vfs operation to notice.
941  *
942  * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
943  * any other hashed alias over that one.
944  */
945 static struct dentry *__d_find_alias(struct inode *inode)
946 {
947         struct dentry *alias;
948
949         if (S_ISDIR(inode->i_mode))
950                 return __d_find_any_alias(inode);
951
952         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
953                 spin_lock(&alias->d_lock);
954                 if (!d_unhashed(alias)) {
955                         __dget_dlock(alias);
956                         spin_unlock(&alias->d_lock);
957                         return alias;
958                 }
959                 spin_unlock(&alias->d_lock);
960         }
961         return NULL;
962 }
963
964 struct dentry *d_find_alias(struct inode *inode)
965 {
966         struct dentry *de = NULL;
967
968         if (!hlist_empty(&inode->i_dentry)) {
969                 spin_lock(&inode->i_lock);
970                 de = __d_find_alias(inode);
971                 spin_unlock(&inode->i_lock);
972         }
973         return de;
974 }
975 EXPORT_SYMBOL(d_find_alias);
976
977 /*
978  *      Try to kill dentries associated with this inode.
979  * WARNING: you must own a reference to inode.
980  */
981 void d_prune_aliases(struct inode *inode)
982 {
983         struct dentry *dentry;
984 restart:
985         spin_lock(&inode->i_lock);
986         hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
987                 spin_lock(&dentry->d_lock);
988                 if (!dentry->d_lockref.count) {
989                         struct dentry *parent = lock_parent(dentry);
990                         if (likely(!dentry->d_lockref.count)) {
991                                 __dentry_kill(dentry);
992                                 dput(parent);
993                                 goto restart;
994                         }
995                         if (parent)
996                                 spin_unlock(&parent->d_lock);
997                 }
998                 spin_unlock(&dentry->d_lock);
999         }
1000         spin_unlock(&inode->i_lock);
1001 }
1002 EXPORT_SYMBOL(d_prune_aliases);
1003
1004 /*
1005  * Lock a dentry from shrink list.
1006  * Called under rcu_read_lock() and dentry->d_lock; the former
1007  * guarantees that nothing we access will be freed under us.
1008  * Note that dentry is *not* protected from concurrent dentry_kill(),
1009  * d_delete(), etc.
1010  *
1011  * Return false if dentry has been disrupted or grabbed, leaving
1012  * the caller to kick it off-list.  Otherwise, return true and have
1013  * that dentry's inode and parent both locked.
1014  */
1015 static bool shrink_lock_dentry(struct dentry *dentry)
1016 {
1017         struct inode *inode;
1018         struct dentry *parent;
1019
1020         if (dentry->d_lockref.count)
1021                 return false;
1022
1023         inode = dentry->d_inode;
1024         if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
1025                 spin_unlock(&dentry->d_lock);
1026                 spin_lock(&inode->i_lock);
1027                 spin_lock(&dentry->d_lock);
1028                 if (unlikely(dentry->d_lockref.count))
1029                         goto out;
1030                 /* changed inode means that somebody had grabbed it */
1031                 if (unlikely(inode != dentry->d_inode))
1032                         goto out;
1033         }
1034
1035         parent = dentry->d_parent;
1036         if (IS_ROOT(dentry) || likely(spin_trylock(&parent->d_lock)))
1037                 return true;
1038
1039         spin_unlock(&dentry->d_lock);
1040         spin_lock(&parent->d_lock);
1041         if (unlikely(parent != dentry->d_parent)) {
1042                 spin_unlock(&parent->d_lock);
1043                 spin_lock(&dentry->d_lock);
1044                 goto out;
1045         }
1046         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1047         if (likely(!dentry->d_lockref.count))
1048                 return true;
1049         spin_unlock(&parent->d_lock);
1050 out:
1051         if (inode)
1052                 spin_unlock(&inode->i_lock);
1053         return false;
1054 }
1055
1056 static void shrink_dentry_list(struct list_head *list)
1057 {
1058         while (!list_empty(list)) {
1059                 struct dentry *dentry, *parent;
1060
1061                 dentry = list_entry(list->prev, struct dentry, d_lru);
1062                 spin_lock(&dentry->d_lock);
1063                 rcu_read_lock();
1064                 if (!shrink_lock_dentry(dentry)) {
1065                         bool can_free = false;
1066                         rcu_read_unlock();
1067                         d_shrink_del(dentry);
1068                         if (dentry->d_lockref.count < 0)
1069                                 can_free = dentry->d_flags & DCACHE_MAY_FREE;
1070                         spin_unlock(&dentry->d_lock);
1071                         if (can_free)
1072                                 dentry_free(dentry);
1073                         continue;
1074                 }
1075                 rcu_read_unlock();
1076                 d_shrink_del(dentry);
1077                 parent = dentry->d_parent;
1078                 __dentry_kill(dentry);
1079                 if (parent == dentry)
1080                         continue;
1081                 /*
1082                  * We need to prune ancestors too. This is necessary to prevent
1083                  * quadratic behavior of shrink_dcache_parent(), but is also
1084                  * expected to be beneficial in reducing dentry cache
1085                  * fragmentation.
1086                  */
1087                 dentry = parent;
1088                 while (dentry && !lockref_put_or_lock(&dentry->d_lockref))
1089                         dentry = dentry_kill(dentry);
1090         }
1091 }
1092
1093 static enum lru_status dentry_lru_isolate(struct list_head *item,
1094                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1095 {
1096         struct list_head *freeable = arg;
1097         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1098
1099
1100         /*
1101          * we are inverting the lru lock/dentry->d_lock here,
1102          * so use a trylock. If we fail to get the lock, just skip
1103          * it
1104          */
1105         if (!spin_trylock(&dentry->d_lock))
1106                 return LRU_SKIP;
1107
1108         /*
1109          * Referenced dentries are still in use. If they have active
1110          * counts, just remove them from the LRU. Otherwise give them
1111          * another pass through the LRU.
1112          */
1113         if (dentry->d_lockref.count) {
1114                 d_lru_isolate(lru, dentry);
1115                 spin_unlock(&dentry->d_lock);
1116                 return LRU_REMOVED;
1117         }
1118
1119         if (dentry->d_flags & DCACHE_REFERENCED) {
1120                 dentry->d_flags &= ~DCACHE_REFERENCED;
1121                 spin_unlock(&dentry->d_lock);
1122
1123                 /*
1124                  * The list move itself will be made by the common LRU code. At
1125                  * this point, we've dropped the dentry->d_lock but keep the
1126                  * lru lock. This is safe to do, since every list movement is
1127                  * protected by the lru lock even if both locks are held.
1128                  *
1129                  * This is guaranteed by the fact that all LRU management
1130                  * functions are intermediated by the LRU API calls like
1131                  * list_lru_add and list_lru_del. List movement in this file
1132                  * only ever occur through this functions or through callbacks
1133                  * like this one, that are called from the LRU API.
1134                  *
1135                  * The only exceptions to this are functions like
1136                  * shrink_dentry_list, and code that first checks for the
1137                  * DCACHE_SHRINK_LIST flag.  Those are guaranteed to be
1138                  * operating only with stack provided lists after they are
1139                  * properly isolated from the main list.  It is thus, always a
1140                  * local access.
1141                  */
1142                 return LRU_ROTATE;
1143         }
1144
1145         d_lru_shrink_move(lru, dentry, freeable);
1146         spin_unlock(&dentry->d_lock);
1147
1148         return LRU_REMOVED;
1149 }
1150
1151 /**
1152  * prune_dcache_sb - shrink the dcache
1153  * @sb: superblock
1154  * @sc: shrink control, passed to list_lru_shrink_walk()
1155  *
1156  * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1157  * is done when we need more memory and called from the superblock shrinker
1158  * function.
1159  *
1160  * This function may fail to free any resources if all the dentries are in
1161  * use.
1162  */
1163 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1164 {
1165         LIST_HEAD(dispose);
1166         long freed;
1167
1168         freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1169                                      dentry_lru_isolate, &dispose);
1170         shrink_dentry_list(&dispose);
1171         return freed;
1172 }
1173
1174 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1175                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1176 {
1177         struct list_head *freeable = arg;
1178         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1179
1180         /*
1181          * we are inverting the lru lock/dentry->d_lock here,
1182          * so use a trylock. If we fail to get the lock, just skip
1183          * it
1184          */
1185         if (!spin_trylock(&dentry->d_lock))
1186                 return LRU_SKIP;
1187
1188         d_lru_shrink_move(lru, dentry, freeable);
1189         spin_unlock(&dentry->d_lock);
1190
1191         return LRU_REMOVED;
1192 }
1193
1194
1195 /**
1196  * shrink_dcache_sb - shrink dcache for a superblock
1197  * @sb: superblock
1198  *
1199  * Shrink the dcache for the specified super block. This is used to free
1200  * the dcache before unmounting a file system.
1201  */
1202 void shrink_dcache_sb(struct super_block *sb)
1203 {
1204         long freed;
1205
1206         do {
1207                 LIST_HEAD(dispose);
1208
1209                 freed = list_lru_walk(&sb->s_dentry_lru,
1210                         dentry_lru_isolate_shrink, &dispose, 1024);
1211
1212                 this_cpu_sub(nr_dentry_unused, freed);
1213                 shrink_dentry_list(&dispose);
1214         } while (list_lru_count(&sb->s_dentry_lru) > 0);
1215 }
1216 EXPORT_SYMBOL(shrink_dcache_sb);
1217
1218 /**
1219  * enum d_walk_ret - action to talke during tree walk
1220  * @D_WALK_CONTINUE:    contrinue walk
1221  * @D_WALK_QUIT:        quit walk
1222  * @D_WALK_NORETRY:     quit when retry is needed
1223  * @D_WALK_SKIP:        skip this dentry and its children
1224  */
1225 enum d_walk_ret {
1226         D_WALK_CONTINUE,
1227         D_WALK_QUIT,
1228         D_WALK_NORETRY,
1229         D_WALK_SKIP,
1230 };
1231
1232 /**
1233  * d_walk - walk the dentry tree
1234  * @parent:     start of walk
1235  * @data:       data passed to @enter() and @finish()
1236  * @enter:      callback when first entering the dentry
1237  *
1238  * The @enter() callbacks are called with d_lock held.
1239  */
1240 static void d_walk(struct dentry *parent, void *data,
1241                    enum d_walk_ret (*enter)(void *, struct dentry *))
1242 {
1243         struct dentry *this_parent;
1244         struct list_head *next;
1245         unsigned seq = 0;
1246         enum d_walk_ret ret;
1247         bool retry = true;
1248
1249 again:
1250         read_seqbegin_or_lock(&rename_lock, &seq);
1251         this_parent = parent;
1252         spin_lock(&this_parent->d_lock);
1253
1254         ret = enter(data, this_parent);
1255         switch (ret) {
1256         case D_WALK_CONTINUE:
1257                 break;
1258         case D_WALK_QUIT:
1259         case D_WALK_SKIP:
1260                 goto out_unlock;
1261         case D_WALK_NORETRY:
1262                 retry = false;
1263                 break;
1264         }
1265 repeat:
1266         next = this_parent->d_subdirs.next;
1267 resume:
1268         while (next != &this_parent->d_subdirs) {
1269                 struct list_head *tmp = next;
1270                 struct dentry *dentry = list_entry(tmp, struct dentry, d_child);
1271                 next = tmp->next;
1272
1273                 if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
1274                         continue;
1275
1276                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1277
1278                 ret = enter(data, dentry);
1279                 switch (ret) {
1280                 case D_WALK_CONTINUE:
1281                         break;
1282                 case D_WALK_QUIT:
1283                         spin_unlock(&dentry->d_lock);
1284                         goto out_unlock;
1285                 case D_WALK_NORETRY:
1286                         retry = false;
1287                         break;
1288                 case D_WALK_SKIP:
1289                         spin_unlock(&dentry->d_lock);
1290                         continue;
1291                 }
1292
1293                 if (!list_empty(&dentry->d_subdirs)) {
1294                         spin_unlock(&this_parent->d_lock);
1295                         spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1296                         this_parent = dentry;
1297                         spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1298                         goto repeat;
1299                 }
1300                 spin_unlock(&dentry->d_lock);
1301         }
1302         /*
1303          * All done at this level ... ascend and resume the search.
1304          */
1305         rcu_read_lock();
1306 ascend:
1307         if (this_parent != parent) {
1308                 struct dentry *child = this_parent;
1309                 this_parent = child->d_parent;
1310
1311                 spin_unlock(&child->d_lock);
1312                 spin_lock(&this_parent->d_lock);
1313
1314                 /* might go back up the wrong parent if we have had a rename. */
1315                 if (need_seqretry(&rename_lock, seq))
1316                         goto rename_retry;
1317                 /* go into the first sibling still alive */
1318                 do {
1319                         next = child->d_child.next;
1320                         if (next == &this_parent->d_subdirs)
1321                                 goto ascend;
1322                         child = list_entry(next, struct dentry, d_child);
1323                 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED));
1324                 rcu_read_unlock();
1325                 goto resume;
1326         }
1327         if (need_seqretry(&rename_lock, seq))
1328                 goto rename_retry;
1329         rcu_read_unlock();
1330
1331 out_unlock:
1332         spin_unlock(&this_parent->d_lock);
1333         done_seqretry(&rename_lock, seq);
1334         return;
1335
1336 rename_retry:
1337         spin_unlock(&this_parent->d_lock);
1338         rcu_read_unlock();
1339         BUG_ON(seq & 1);
1340         if (!retry)
1341                 return;
1342         seq = 1;
1343         goto again;
1344 }
1345
1346 struct check_mount {
1347         struct vfsmount *mnt;
1348         unsigned int mounted;
1349 };
1350
1351 static enum d_walk_ret path_check_mount(void *data, struct dentry *dentry)
1352 {
1353         struct check_mount *info = data;
1354         struct path path = { .mnt = info->mnt, .dentry = dentry };
1355
1356         if (likely(!d_mountpoint(dentry)))
1357                 return D_WALK_CONTINUE;
1358         if (__path_is_mountpoint(&path)) {
1359                 info->mounted = 1;
1360                 return D_WALK_QUIT;
1361         }
1362         return D_WALK_CONTINUE;
1363 }
1364
1365 /**
1366  * path_has_submounts - check for mounts over a dentry in the
1367  *                      current namespace.
1368  * @parent: path to check.
1369  *
1370  * Return true if the parent or its subdirectories contain
1371  * a mount point in the current namespace.
1372  */
1373 int path_has_submounts(const struct path *parent)
1374 {
1375         struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };
1376
1377         read_seqlock_excl(&mount_lock);
1378         d_walk(parent->dentry, &data, path_check_mount);
1379         read_sequnlock_excl(&mount_lock);
1380
1381         return data.mounted;
1382 }
1383 EXPORT_SYMBOL(path_has_submounts);
1384
1385 /*
1386  * Called by mount code to set a mountpoint and check if the mountpoint is
1387  * reachable (e.g. NFS can unhash a directory dentry and then the complete
1388  * subtree can become unreachable).
1389  *
1390  * Only one of d_invalidate() and d_set_mounted() must succeed.  For
1391  * this reason take rename_lock and d_lock on dentry and ancestors.
1392  */
1393 int d_set_mounted(struct dentry *dentry)
1394 {
1395         struct dentry *p;
1396         int ret = -ENOENT;
1397         write_seqlock(&rename_lock);
1398         for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1399                 /* Need exclusion wrt. d_invalidate() */
1400                 spin_lock(&p->d_lock);
1401                 if (unlikely(d_unhashed(p))) {
1402                         spin_unlock(&p->d_lock);
1403                         goto out;
1404                 }
1405                 spin_unlock(&p->d_lock);
1406         }
1407         spin_lock(&dentry->d_lock);
1408         if (!d_unlinked(dentry)) {
1409                 ret = -EBUSY;
1410                 if (!d_mountpoint(dentry)) {
1411                         dentry->d_flags |= DCACHE_MOUNTED;
1412                         ret = 0;
1413                 }
1414         }
1415         spin_unlock(&dentry->d_lock);
1416 out:
1417         write_sequnlock(&rename_lock);
1418         return ret;
1419 }
1420
1421 /*
1422  * Search the dentry child list of the specified parent,
1423  * and move any unused dentries to the end of the unused
1424  * list for prune_dcache(). We descend to the next level
1425  * whenever the d_subdirs list is non-empty and continue
1426  * searching.
1427  *
1428  * It returns zero iff there are no unused children,
1429  * otherwise  it returns the number of children moved to
1430  * the end of the unused list. This may not be the total
1431  * number of unused children, because select_parent can
1432  * drop the lock and return early due to latency
1433  * constraints.
1434  */
1435
1436 struct select_data {
1437         struct dentry *start;
1438         struct list_head dispose;
1439         int found;
1440 };
1441
1442 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1443 {
1444         struct select_data *data = _data;
1445         enum d_walk_ret ret = D_WALK_CONTINUE;
1446
1447         if (data->start == dentry)
1448                 goto out;
1449
1450         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1451                 data->found++;
1452         } else {
1453                 if (dentry->d_flags & DCACHE_LRU_LIST)
1454                         d_lru_del(dentry);
1455                 if (!dentry->d_lockref.count) {
1456                         d_shrink_add(dentry, &data->dispose);
1457                         data->found++;
1458                 }
1459         }
1460         /*
1461          * We can return to the caller if we have found some (this
1462          * ensures forward progress). We'll be coming back to find
1463          * the rest.
1464          */
1465         if (!list_empty(&data->dispose))
1466                 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1467 out:
1468         return ret;
1469 }
1470
1471 /**
1472  * shrink_dcache_parent - prune dcache
1473  * @parent: parent of entries to prune
1474  *
1475  * Prune the dcache to remove unused children of the parent dentry.
1476  */
1477 void shrink_dcache_parent(struct dentry *parent)
1478 {
1479         for (;;) {
1480                 struct select_data data;
1481
1482                 INIT_LIST_HEAD(&data.dispose);
1483                 data.start = parent;
1484                 data.found = 0;
1485
1486                 d_walk(parent, &data, select_collect);
1487
1488                 if (!list_empty(&data.dispose)) {
1489                         shrink_dentry_list(&data.dispose);
1490                         continue;
1491                 }
1492
1493                 cond_resched();
1494                 if (!data.found)
1495                         break;
1496         }
1497 }
1498 EXPORT_SYMBOL(shrink_dcache_parent);
1499
1500 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1501 {
1502         /* it has busy descendents; complain about those instead */
1503         if (!list_empty(&dentry->d_subdirs))
1504                 return D_WALK_CONTINUE;
1505
1506         /* root with refcount 1 is fine */
1507         if (dentry == _data && dentry->d_lockref.count == 1)
1508                 return D_WALK_CONTINUE;
1509
1510         printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} "
1511                         " still in use (%d) [unmount of %s %s]\n",
1512                        dentry,
1513                        dentry->d_inode ?
1514                        dentry->d_inode->i_ino : 0UL,
1515                        dentry,
1516                        dentry->d_lockref.count,
1517                        dentry->d_sb->s_type->name,
1518                        dentry->d_sb->s_id);
1519         WARN_ON(1);
1520         return D_WALK_CONTINUE;
1521 }
1522
1523 static void do_one_tree(struct dentry *dentry)
1524 {
1525         shrink_dcache_parent(dentry);
1526         d_walk(dentry, dentry, umount_check);
1527         d_drop(dentry);
1528         dput(dentry);
1529 }
1530
1531 /*
1532  * destroy the dentries attached to a superblock on unmounting
1533  */
1534 void shrink_dcache_for_umount(struct super_block *sb)
1535 {
1536         struct dentry *dentry;
1537
1538         WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
1539
1540         dentry = sb->s_root;
1541         sb->s_root = NULL;
1542         do_one_tree(dentry);
1543
1544         while (!hlist_bl_empty(&sb->s_roots)) {
1545                 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_roots), struct dentry, d_hash));
1546                 do_one_tree(dentry);
1547         }
1548 }
1549
1550 static enum d_walk_ret find_submount(void *_data, struct dentry *dentry)
1551 {
1552         struct dentry **victim = _data;
1553         if (d_mountpoint(dentry)) {
1554                 __dget_dlock(dentry);
1555                 *victim = dentry;
1556                 return D_WALK_QUIT;
1557         }
1558         return D_WALK_CONTINUE;
1559 }
1560
1561 /**
1562  * d_invalidate - detach submounts, prune dcache, and drop
1563  * @dentry: dentry to invalidate (aka detach, prune and drop)
1564  */
1565 void d_invalidate(struct dentry *dentry)
1566 {
1567         bool had_submounts = false;
1568         spin_lock(&dentry->d_lock);
1569         if (d_unhashed(dentry)) {
1570                 spin_unlock(&dentry->d_lock);
1571                 return;
1572         }
1573         __d_drop(dentry);
1574         spin_unlock(&dentry->d_lock);
1575
1576         /* Negative dentries can be dropped without further checks */
1577         if (!dentry->d_inode)
1578                 return;
1579
1580         shrink_dcache_parent(dentry);
1581         for (;;) {
1582                 struct dentry *victim = NULL;
1583                 d_walk(dentry, &victim, find_submount);
1584                 if (!victim) {
1585                         if (had_submounts)
1586                                 shrink_dcache_parent(dentry);
1587                         return;
1588                 }
1589                 had_submounts = true;
1590                 detach_mounts(victim);
1591                 dput(victim);
1592         }
1593 }
1594 EXPORT_SYMBOL(d_invalidate);
1595
1596 /**
1597  * __d_alloc    -       allocate a dcache entry
1598  * @sb: filesystem it will belong to
1599  * @name: qstr of the name
1600  *
1601  * Allocates a dentry. It returns %NULL if there is insufficient memory
1602  * available. On a success the dentry is returned. The name passed in is
1603  * copied and the copy passed in may be reused after this call.
1604  */
1605  
1606 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1607 {
1608         struct external_name *ext = NULL;
1609         struct dentry *dentry;
1610         char *dname;
1611         int err;
1612
1613         dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1614         if (!dentry)
1615                 return NULL;
1616
1617         /*
1618          * We guarantee that the inline name is always NUL-terminated.
1619          * This way the memcpy() done by the name switching in rename
1620          * will still always have a NUL at the end, even if we might
1621          * be overwriting an internal NUL character
1622          */
1623         dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1624         if (unlikely(!name)) {
1625                 name = &slash_name;
1626                 dname = dentry->d_iname;
1627         } else if (name->len > DNAME_INLINE_LEN-1) {
1628                 size_t size = offsetof(struct external_name, name[1]);
1629
1630                 ext = kmalloc(size + name->len, GFP_KERNEL_ACCOUNT);
1631                 if (!ext) {
1632                         kmem_cache_free(dentry_cache, dentry); 
1633                         return NULL;
1634                 }
1635                 atomic_set(&ext->u.count, 1);
1636                 dname = ext->name;
1637         } else  {
1638                 dname = dentry->d_iname;
1639         }       
1640
1641         dentry->d_name.len = name->len;
1642         dentry->d_name.hash = name->hash;
1643         memcpy(dname, name->name, name->len);
1644         dname[name->len] = 0;
1645
1646         /* Make sure we always see the terminating NUL character */
1647         smp_store_release(&dentry->d_name.name, dname); /* ^^^ */
1648
1649         dentry->d_lockref.count = 1;
1650         dentry->d_flags = 0;
1651         spin_lock_init(&dentry->d_lock);
1652         seqcount_init(&dentry->d_seq);
1653         dentry->d_inode = NULL;
1654         dentry->d_parent = dentry;
1655         dentry->d_sb = sb;
1656         dentry->d_op = NULL;
1657         dentry->d_fsdata = NULL;
1658         INIT_HLIST_BL_NODE(&dentry->d_hash);
1659         INIT_LIST_HEAD(&dentry->d_lru);
1660         INIT_LIST_HEAD(&dentry->d_subdirs);
1661         INIT_HLIST_NODE(&dentry->d_u.d_alias);
1662         INIT_LIST_HEAD(&dentry->d_child);
1663         d_set_d_op(dentry, dentry->d_sb->s_d_op);
1664
1665         if (dentry->d_op && dentry->d_op->d_init) {
1666                 err = dentry->d_op->d_init(dentry);
1667                 if (err) {
1668                         if (dname_external(dentry))
1669                                 kfree(external_name(dentry));
1670                         kmem_cache_free(dentry_cache, dentry);
1671                         return NULL;
1672                 }
1673         }
1674
1675         if (unlikely(ext)) {
1676                 pg_data_t *pgdat = page_pgdat(virt_to_page(ext));
1677                 mod_node_page_state(pgdat, NR_INDIRECTLY_RECLAIMABLE_BYTES,
1678                                     ksize(ext));
1679         }
1680
1681         this_cpu_inc(nr_dentry);
1682
1683         return dentry;
1684 }
1685
1686 /**
1687  * d_alloc      -       allocate a dcache entry
1688  * @parent: parent of entry to allocate
1689  * @name: qstr of the name
1690  *
1691  * Allocates a dentry. It returns %NULL if there is insufficient memory
1692  * available. On a success the dentry is returned. The name passed in is
1693  * copied and the copy passed in may be reused after this call.
1694  */
1695 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1696 {
1697         struct dentry *dentry = __d_alloc(parent->d_sb, name);
1698         if (!dentry)
1699                 return NULL;
1700         dentry->d_flags |= DCACHE_RCUACCESS;
1701         spin_lock(&parent->d_lock);
1702         /*
1703          * don't need child lock because it is not subject
1704          * to concurrency here
1705          */
1706         __dget_dlock(parent);
1707         dentry->d_parent = parent;
1708         list_add(&dentry->d_child, &parent->d_subdirs);
1709         spin_unlock(&parent->d_lock);
1710
1711         return dentry;
1712 }
1713 EXPORT_SYMBOL(d_alloc);
1714
1715 struct dentry *d_alloc_anon(struct super_block *sb)
1716 {
1717         return __d_alloc(sb, NULL);
1718 }
1719 EXPORT_SYMBOL(d_alloc_anon);
1720
1721 struct dentry *d_alloc_cursor(struct dentry * parent)
1722 {
1723         struct dentry *dentry = d_alloc_anon(parent->d_sb);
1724         if (dentry) {
1725                 dentry->d_flags |= DCACHE_RCUACCESS | DCACHE_DENTRY_CURSOR;
1726                 dentry->d_parent = dget(parent);
1727         }
1728         return dentry;
1729 }
1730
1731 /**
1732  * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1733  * @sb: the superblock
1734  * @name: qstr of the name
1735  *
1736  * For a filesystem that just pins its dentries in memory and never
1737  * performs lookups at all, return an unhashed IS_ROOT dentry.
1738  */
1739 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1740 {
1741         return __d_alloc(sb, name);
1742 }
1743 EXPORT_SYMBOL(d_alloc_pseudo);
1744
1745 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1746 {
1747         struct qstr q;
1748
1749         q.name = name;
1750         q.hash_len = hashlen_string(parent, name);
1751         return d_alloc(parent, &q);
1752 }
1753 EXPORT_SYMBOL(d_alloc_name);
1754
1755 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1756 {
1757         WARN_ON_ONCE(dentry->d_op);
1758         WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH  |
1759                                 DCACHE_OP_COMPARE       |
1760                                 DCACHE_OP_REVALIDATE    |
1761                                 DCACHE_OP_WEAK_REVALIDATE       |
1762                                 DCACHE_OP_DELETE        |
1763                                 DCACHE_OP_REAL));
1764         dentry->d_op = op;
1765         if (!op)
1766                 return;
1767         if (op->d_hash)
1768                 dentry->d_flags |= DCACHE_OP_HASH;
1769         if (op->d_compare)
1770                 dentry->d_flags |= DCACHE_OP_COMPARE;
1771         if (op->d_revalidate)
1772                 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1773         if (op->d_weak_revalidate)
1774                 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1775         if (op->d_delete)
1776                 dentry->d_flags |= DCACHE_OP_DELETE;
1777         if (op->d_prune)
1778                 dentry->d_flags |= DCACHE_OP_PRUNE;
1779         if (op->d_real)
1780                 dentry->d_flags |= DCACHE_OP_REAL;
1781
1782 }
1783 EXPORT_SYMBOL(d_set_d_op);
1784
1785
1786 /*
1787  * d_set_fallthru - Mark a dentry as falling through to a lower layer
1788  * @dentry - The dentry to mark
1789  *
1790  * Mark a dentry as falling through to the lower layer (as set with
1791  * d_pin_lower()).  This flag may be recorded on the medium.
1792  */
1793 void d_set_fallthru(struct dentry *dentry)
1794 {
1795         spin_lock(&dentry->d_lock);
1796         dentry->d_flags |= DCACHE_FALLTHRU;
1797         spin_unlock(&dentry->d_lock);
1798 }
1799 EXPORT_SYMBOL(d_set_fallthru);
1800
1801 static unsigned d_flags_for_inode(struct inode *inode)
1802 {
1803         unsigned add_flags = DCACHE_REGULAR_TYPE;
1804
1805         if (!inode)
1806                 return DCACHE_MISS_TYPE;
1807
1808         if (S_ISDIR(inode->i_mode)) {
1809                 add_flags = DCACHE_DIRECTORY_TYPE;
1810                 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1811                         if (unlikely(!inode->i_op->lookup))
1812                                 add_flags = DCACHE_AUTODIR_TYPE;
1813                         else
1814                                 inode->i_opflags |= IOP_LOOKUP;
1815                 }
1816                 goto type_determined;
1817         }
1818
1819         if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1820                 if (unlikely(inode->i_op->get_link)) {
1821                         add_flags = DCACHE_SYMLINK_TYPE;
1822                         goto type_determined;
1823                 }
1824                 inode->i_opflags |= IOP_NOFOLLOW;
1825         }
1826
1827         if (unlikely(!S_ISREG(inode->i_mode)))
1828                 add_flags = DCACHE_SPECIAL_TYPE;
1829
1830 type_determined:
1831         if (unlikely(IS_AUTOMOUNT(inode)))
1832                 add_flags |= DCACHE_NEED_AUTOMOUNT;
1833         return add_flags;
1834 }
1835
1836 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1837 {
1838         unsigned add_flags = d_flags_for_inode(inode);
1839         WARN_ON(d_in_lookup(dentry));
1840
1841         spin_lock(&dentry->d_lock);
1842         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1843         raw_write_seqcount_begin(&dentry->d_seq);
1844         __d_set_inode_and_type(dentry, inode, add_flags);
1845         raw_write_seqcount_end(&dentry->d_seq);
1846         fsnotify_update_flags(dentry);
1847         spin_unlock(&dentry->d_lock);
1848 }
1849
1850 /**
1851  * d_instantiate - fill in inode information for a dentry
1852  * @entry: dentry to complete
1853  * @inode: inode to attach to this dentry
1854  *
1855  * Fill in inode information in the entry.
1856  *
1857  * This turns negative dentries into productive full members
1858  * of society.
1859  *
1860  * NOTE! This assumes that the inode count has been incremented
1861  * (or otherwise set) by the caller to indicate that it is now
1862  * in use by the dcache.
1863  */
1864  
1865 void d_instantiate(struct dentry *entry, struct inode * inode)
1866 {
1867         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1868         if (inode) {
1869                 security_d_instantiate(entry, inode);
1870                 spin_lock(&inode->i_lock);
1871                 __d_instantiate(entry, inode);
1872                 spin_unlock(&inode->i_lock);
1873         }
1874 }
1875 EXPORT_SYMBOL(d_instantiate);
1876
1877 /*
1878  * This should be equivalent to d_instantiate() + unlock_new_inode(),
1879  * with lockdep-related part of unlock_new_inode() done before
1880  * anything else.  Use that instead of open-coding d_instantiate()/
1881  * unlock_new_inode() combinations.
1882  */
1883 void d_instantiate_new(struct dentry *entry, struct inode *inode)
1884 {
1885         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1886         BUG_ON(!inode);
1887         lockdep_annotate_inode_mutex_key(inode);
1888         security_d_instantiate(entry, inode);
1889         spin_lock(&inode->i_lock);
1890         __d_instantiate(entry, inode);
1891         WARN_ON(!(inode->i_state & I_NEW));
1892         inode->i_state &= ~I_NEW & ~I_CREATING;
1893         smp_mb();
1894         wake_up_bit(&inode->i_state, __I_NEW);
1895         spin_unlock(&inode->i_lock);
1896 }
1897 EXPORT_SYMBOL(d_instantiate_new);
1898
1899 struct dentry *d_make_root(struct inode *root_inode)
1900 {
1901         struct dentry *res = NULL;
1902
1903         if (root_inode) {
1904                 res = d_alloc_anon(root_inode->i_sb);
1905                 if (res) {
1906                         res->d_flags |= DCACHE_RCUACCESS;
1907                         d_instantiate(res, root_inode);
1908                 } else {
1909                         iput(root_inode);
1910                 }
1911         }
1912         return res;
1913 }
1914 EXPORT_SYMBOL(d_make_root);
1915
1916 static struct dentry *__d_instantiate_anon(struct dentry *dentry,
1917                                            struct inode *inode,
1918                                            bool disconnected)
1919 {
1920         struct dentry *res;
1921         unsigned add_flags;
1922
1923         security_d_instantiate(dentry, inode);
1924         spin_lock(&inode->i_lock);
1925         res = __d_find_any_alias(inode);
1926         if (res) {
1927                 spin_unlock(&inode->i_lock);
1928                 dput(dentry);
1929                 goto out_iput;
1930         }
1931
1932         /* attach a disconnected dentry */
1933         add_flags = d_flags_for_inode(inode);
1934
1935         if (disconnected)
1936                 add_flags |= DCACHE_DISCONNECTED;
1937
1938         spin_lock(&dentry->d_lock);
1939         __d_set_inode_and_type(dentry, inode, add_flags);
1940         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1941         if (!disconnected) {
1942                 hlist_bl_lock(&dentry->d_sb->s_roots);
1943                 hlist_bl_add_head(&dentry->d_hash, &dentry->d_sb->s_roots);
1944                 hlist_bl_unlock(&dentry->d_sb->s_roots);
1945         }
1946         spin_unlock(&dentry->d_lock);
1947         spin_unlock(&inode->i_lock);
1948
1949         return dentry;
1950
1951  out_iput:
1952         iput(inode);
1953         return res;
1954 }
1955
1956 struct dentry *d_instantiate_anon(struct dentry *dentry, struct inode *inode)
1957 {
1958         return __d_instantiate_anon(dentry, inode, true);
1959 }
1960 EXPORT_SYMBOL(d_instantiate_anon);
1961
1962 static struct dentry *__d_obtain_alias(struct inode *inode, bool disconnected)
1963 {
1964         struct dentry *tmp;
1965         struct dentry *res;
1966
1967         if (!inode)
1968                 return ERR_PTR(-ESTALE);
1969         if (IS_ERR(inode))
1970                 return ERR_CAST(inode);
1971
1972         res = d_find_any_alias(inode);
1973         if (res)
1974                 goto out_iput;
1975
1976         tmp = d_alloc_anon(inode->i_sb);
1977         if (!tmp) {
1978                 res = ERR_PTR(-ENOMEM);
1979                 goto out_iput;
1980         }
1981
1982         return __d_instantiate_anon(tmp, inode, disconnected);
1983
1984 out_iput:
1985         iput(inode);
1986         return res;
1987 }
1988
1989 /**
1990  * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
1991  * @inode: inode to allocate the dentry for
1992  *
1993  * Obtain a dentry for an inode resulting from NFS filehandle conversion or
1994  * similar open by handle operations.  The returned dentry may be anonymous,
1995  * or may have a full name (if the inode was already in the cache).
1996  *
1997  * When called on a directory inode, we must ensure that the inode only ever
1998  * has one dentry.  If a dentry is found, that is returned instead of
1999  * allocating a new one.
2000  *
2001  * On successful return, the reference to the inode has been transferred
2002  * to the dentry.  In case of an error the reference on the inode is released.
2003  * To make it easier to use in export operations a %NULL or IS_ERR inode may
2004  * be passed in and the error will be propagated to the return value,
2005  * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
2006  */
2007 struct dentry *d_obtain_alias(struct inode *inode)
2008 {
2009         return __d_obtain_alias(inode, true);
2010 }
2011 EXPORT_SYMBOL(d_obtain_alias);
2012
2013 /**
2014  * d_obtain_root - find or allocate a dentry for a given inode
2015  * @inode: inode to allocate the dentry for
2016  *
2017  * Obtain an IS_ROOT dentry for the root of a filesystem.
2018  *
2019  * We must ensure that directory inodes only ever have one dentry.  If a
2020  * dentry is found, that is returned instead of allocating a new one.
2021  *
2022  * On successful return, the reference to the inode has been transferred
2023  * to the dentry.  In case of an error the reference on the inode is
2024  * released.  A %NULL or IS_ERR inode may be passed in and will be the
2025  * error will be propagate to the return value, with a %NULL @inode
2026  * replaced by ERR_PTR(-ESTALE).
2027  */
2028 struct dentry *d_obtain_root(struct inode *inode)
2029 {
2030         return __d_obtain_alias(inode, false);
2031 }
2032 EXPORT_SYMBOL(d_obtain_root);
2033
2034 /**
2035  * d_add_ci - lookup or allocate new dentry with case-exact name
2036  * @inode:  the inode case-insensitive lookup has found
2037  * @dentry: the negative dentry that was passed to the parent's lookup func
2038  * @name:   the case-exact name to be associated with the returned dentry
2039  *
2040  * This is to avoid filling the dcache with case-insensitive names to the
2041  * same inode, only the actual correct case is stored in the dcache for
2042  * case-insensitive filesystems.
2043  *
2044  * For a case-insensitive lookup match and if the the case-exact dentry
2045  * already exists in in the dcache, use it and return it.
2046  *
2047  * If no entry exists with the exact case name, allocate new dentry with
2048  * the exact case, and return the spliced entry.
2049  */
2050 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2051                         struct qstr *name)
2052 {
2053         struct dentry *found, *res;
2054
2055         /*
2056          * First check if a dentry matching the name already exists,
2057          * if not go ahead and create it now.
2058          */
2059         found = d_hash_and_lookup(dentry->d_parent, name);
2060         if (found) {
2061                 iput(inode);
2062                 return found;
2063         }
2064         if (d_in_lookup(dentry)) {
2065                 found = d_alloc_parallel(dentry->d_parent, name,
2066                                         dentry->d_wait);
2067                 if (IS_ERR(found) || !d_in_lookup(found)) {
2068                         iput(inode);
2069                         return found;
2070                 }
2071         } else {
2072                 found = d_alloc(dentry->d_parent, name);
2073                 if (!found) {
2074                         iput(inode);
2075                         return ERR_PTR(-ENOMEM);
2076                 } 
2077         }
2078         res = d_splice_alias(inode, found);
2079         if (res) {
2080                 dput(found);
2081                 return res;
2082         }
2083         return found;
2084 }
2085 EXPORT_SYMBOL(d_add_ci);
2086
2087
2088 static inline bool d_same_name(const struct dentry *dentry,
2089                                 const struct dentry *parent,
2090                                 const struct qstr *name)
2091 {
2092         if (likely(!(parent->d_flags & DCACHE_OP_COMPARE))) {
2093                 if (dentry->d_name.len != name->len)
2094                         return false;
2095                 return dentry_cmp(dentry, name->name, name->len) == 0;
2096         }
2097         return parent->d_op->d_compare(dentry,
2098                                        dentry->d_name.len, dentry->d_name.name,
2099                                        name) == 0;
2100 }
2101
2102 /**
2103  * __d_lookup_rcu - search for a dentry (racy, store-free)
2104  * @parent: parent dentry
2105  * @name: qstr of name we wish to find
2106  * @seqp: returns d_seq value at the point where the dentry was found
2107  * Returns: dentry, or NULL
2108  *
2109  * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2110  * resolution (store-free path walking) design described in
2111  * Documentation/filesystems/path-lookup.txt.
2112  *
2113  * This is not to be used outside core vfs.
2114  *
2115  * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2116  * held, and rcu_read_lock held. The returned dentry must not be stored into
2117  * without taking d_lock and checking d_seq sequence count against @seq
2118  * returned here.
2119  *
2120  * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2121  * function.
2122  *
2123  * Alternatively, __d_lookup_rcu may be called again to look up the child of
2124  * the returned dentry, so long as its parent's seqlock is checked after the
2125  * child is looked up. Thus, an interlocking stepping of sequence lock checks
2126  * is formed, giving integrity down the path walk.
2127  *
2128  * NOTE! The caller *has* to check the resulting dentry against the sequence
2129  * number we've returned before using any of the resulting dentry state!
2130  */
2131 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2132                                 const struct qstr *name,
2133                                 unsigned *seqp)
2134 {
2135         u64 hashlen = name->hash_len;
2136         const unsigned char *str = name->name;
2137         struct hlist_bl_head *b = d_hash(hashlen_hash(hashlen));
2138         struct hlist_bl_node *node;
2139         struct dentry *dentry;
2140
2141         /*
2142          * Note: There is significant duplication with __d_lookup_rcu which is
2143          * required to prevent single threaded performance regressions
2144          * especially on architectures where smp_rmb (in seqcounts) are costly.
2145          * Keep the two functions in sync.
2146          */
2147
2148         /*
2149          * The hash list is protected using RCU.
2150          *
2151          * Carefully use d_seq when comparing a candidate dentry, to avoid
2152          * races with d_move().
2153          *
2154          * It is possible that concurrent renames can mess up our list
2155          * walk here and result in missing our dentry, resulting in the
2156          * false-negative result. d_lookup() protects against concurrent
2157          * renames using rename_lock seqlock.
2158          *
2159          * See Documentation/filesystems/path-lookup.txt for more details.
2160          */
2161         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2162                 unsigned seq;
2163
2164 seqretry:
2165                 /*
2166                  * The dentry sequence count protects us from concurrent
2167                  * renames, and thus protects parent and name fields.
2168                  *
2169                  * The caller must perform a seqcount check in order
2170                  * to do anything useful with the returned dentry.
2171                  *
2172                  * NOTE! We do a "raw" seqcount_begin here. That means that
2173                  * we don't wait for the sequence count to stabilize if it
2174                  * is in the middle of a sequence change. If we do the slow
2175                  * dentry compare, we will do seqretries until it is stable,
2176                  * and if we end up with a successful lookup, we actually
2177                  * want to exit RCU lookup anyway.
2178                  *
2179                  * Note that raw_seqcount_begin still *does* smp_rmb(), so
2180                  * we are still guaranteed NUL-termination of ->d_name.name.
2181                  */
2182                 seq = raw_seqcount_begin(&dentry->d_seq);
2183                 if (dentry->d_parent != parent)
2184                         continue;
2185                 if (d_unhashed(dentry))
2186                         continue;
2187
2188                 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) {
2189                         int tlen;
2190                         const char *tname;
2191                         if (dentry->d_name.hash != hashlen_hash(hashlen))
2192                                 continue;
2193                         tlen = dentry->d_name.len;
2194                         tname = dentry->d_name.name;
2195                         /* we want a consistent (name,len) pair */
2196                         if (read_seqcount_retry(&dentry->d_seq, seq)) {
2197                                 cpu_relax();
2198                                 goto seqretry;
2199                         }
2200                         if (parent->d_op->d_compare(dentry,
2201                                                     tlen, tname, name) != 0)
2202                                 continue;
2203                 } else {
2204                         if (dentry->d_name.hash_len != hashlen)
2205                                 continue;
2206                         if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
2207                                 continue;
2208                 }
2209                 *seqp = seq;
2210                 return dentry;
2211         }
2212         return NULL;
2213 }
2214
2215 /**
2216  * d_lookup - search for a dentry
2217  * @parent: parent dentry
2218  * @name: qstr of name we wish to find
2219  * Returns: dentry, or NULL
2220  *
2221  * d_lookup searches the children of the parent dentry for the name in
2222  * question. If the dentry is found its reference count is incremented and the
2223  * dentry is returned. The caller must use dput to free the entry when it has
2224  * finished using it. %NULL is returned if the dentry does not exist.
2225  */
2226 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2227 {
2228         struct dentry *dentry;
2229         unsigned seq;
2230
2231         do {
2232                 seq = read_seqbegin(&rename_lock);
2233                 dentry = __d_lookup(parent, name);
2234                 if (dentry)
2235                         break;
2236         } while (read_seqretry(&rename_lock, seq));
2237         return dentry;
2238 }
2239 EXPORT_SYMBOL(d_lookup);
2240
2241 /**
2242  * __d_lookup - search for a dentry (racy)
2243  * @parent: parent dentry
2244  * @name: qstr of name we wish to find
2245  * Returns: dentry, or NULL
2246  *
2247  * __d_lookup is like d_lookup, however it may (rarely) return a
2248  * false-negative result due to unrelated rename activity.
2249  *
2250  * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2251  * however it must be used carefully, eg. with a following d_lookup in
2252  * the case of failure.
2253  *
2254  * __d_lookup callers must be commented.
2255  */
2256 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2257 {
2258         unsigned int hash = name->hash;
2259         struct hlist_bl_head *b = d_hash(hash);
2260         struct hlist_bl_node *node;
2261         struct dentry *found = NULL;
2262         struct dentry *dentry;
2263
2264         /*
2265          * Note: There is significant duplication with __d_lookup_rcu which is
2266          * required to prevent single threaded performance regressions
2267          * especially on architectures where smp_rmb (in seqcounts) are costly.
2268          * Keep the two functions in sync.
2269          */
2270
2271         /*
2272          * The hash list is protected using RCU.
2273          *
2274          * Take d_lock when comparing a candidate dentry, to avoid races
2275          * with d_move().
2276          *
2277          * It is possible that concurrent renames can mess up our list
2278          * walk here and result in missing our dentry, resulting in the
2279          * false-negative result. d_lookup() protects against concurrent
2280          * renames using rename_lock seqlock.
2281          *
2282          * See Documentation/filesystems/path-lookup.txt for more details.
2283          */
2284         rcu_read_lock();
2285         
2286         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2287
2288                 if (dentry->d_name.hash != hash)
2289                         continue;
2290
2291                 spin_lock(&dentry->d_lock);
2292                 if (dentry->d_parent != parent)
2293                         goto next;
2294                 if (d_unhashed(dentry))
2295                         goto next;
2296
2297                 if (!d_same_name(dentry, parent, name))
2298                         goto next;
2299
2300                 dentry->d_lockref.count++;
2301                 found = dentry;
2302                 spin_unlock(&dentry->d_lock);
2303                 break;
2304 next:
2305                 spin_unlock(&dentry->d_lock);
2306         }
2307         rcu_read_unlock();
2308
2309         return found;
2310 }
2311
2312 /**
2313  * d_hash_and_lookup - hash the qstr then search for a dentry
2314  * @dir: Directory to search in
2315  * @name: qstr of name we wish to find
2316  *
2317  * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2318  */
2319 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2320 {
2321         /*
2322          * Check for a fs-specific hash function. Note that we must
2323          * calculate the standard hash first, as the d_op->d_hash()
2324          * routine may choose to leave the hash value unchanged.
2325          */
2326         name->hash = full_name_hash(dir, name->name, name->len);
2327         if (dir->d_flags & DCACHE_OP_HASH) {
2328                 int err = dir->d_op->d_hash(dir, name);
2329                 if (unlikely(err < 0))
2330                         return ERR_PTR(err);
2331         }
2332         return d_lookup(dir, name);
2333 }
2334 EXPORT_SYMBOL(d_hash_and_lookup);
2335
2336 /*
2337  * When a file is deleted, we have two options:
2338  * - turn this dentry into a negative dentry
2339  * - unhash this dentry and free it.
2340  *
2341  * Usually, we want to just turn this into
2342  * a negative dentry, but if anybody else is
2343  * currently using the dentry or the inode
2344  * we can't do that and we fall back on removing
2345  * it from the hash queues and waiting for
2346  * it to be deleted later when it has no users
2347  */
2348  
2349 /**
2350  * d_delete - delete a dentry
2351  * @dentry: The dentry to delete
2352  *
2353  * Turn the dentry into a negative dentry if possible, otherwise
2354  * remove it from the hash queues so it can be deleted later
2355  */
2356  
2357 void d_delete(struct dentry * dentry)
2358 {
2359         struct inode *inode = dentry->d_inode;
2360         int isdir = d_is_dir(dentry);
2361
2362         spin_lock(&inode->i_lock);
2363         spin_lock(&dentry->d_lock);
2364         /*
2365          * Are we the only user?
2366          */
2367         if (dentry->d_lockref.count == 1) {
2368                 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2369                 dentry_unlink_inode(dentry);
2370         } else {
2371                 __d_drop(dentry);
2372                 spin_unlock(&dentry->d_lock);
2373                 spin_unlock(&inode->i_lock);
2374         }
2375         fsnotify_nameremove(dentry, isdir);
2376 }
2377 EXPORT_SYMBOL(d_delete);
2378
2379 static void __d_rehash(struct dentry *entry)
2380 {
2381         struct hlist_bl_head *b = d_hash(entry->d_name.hash);
2382
2383         hlist_bl_lock(b);
2384         hlist_bl_add_head_rcu(&entry->d_hash, b);
2385         hlist_bl_unlock(b);
2386 }
2387
2388 /**
2389  * d_rehash     - add an entry back to the hash
2390  * @entry: dentry to add to the hash
2391  *
2392  * Adds a dentry to the hash according to its name.
2393  */
2394  
2395 void d_rehash(struct dentry * entry)
2396 {
2397         spin_lock(&entry->d_lock);
2398         __d_rehash(entry);
2399         spin_unlock(&entry->d_lock);
2400 }
2401 EXPORT_SYMBOL(d_rehash);
2402
2403 static inline unsigned start_dir_add(struct inode *dir)
2404 {
2405
2406         for (;;) {
2407                 unsigned n = dir->i_dir_seq;
2408                 if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
2409                         return n;
2410                 cpu_relax();
2411         }
2412 }
2413
2414 static inline void end_dir_add(struct inode *dir, unsigned n)
2415 {
2416         smp_store_release(&dir->i_dir_seq, n + 2);
2417 }
2418
2419 static void d_wait_lookup(struct dentry *dentry)
2420 {
2421         if (d_in_lookup(dentry)) {
2422                 DECLARE_WAITQUEUE(wait, current);
2423                 add_wait_queue(dentry->d_wait, &wait);
2424                 do {
2425                         set_current_state(TASK_UNINTERRUPTIBLE);
2426                         spin_unlock(&dentry->d_lock);
2427                         schedule();
2428                         spin_lock(&dentry->d_lock);
2429                 } while (d_in_lookup(dentry));
2430         }
2431 }
2432
2433 struct dentry *d_alloc_parallel(struct dentry *parent,
2434                                 const struct qstr *name,
2435                                 wait_queue_head_t *wq)
2436 {
2437         unsigned int hash = name->hash;
2438         struct hlist_bl_head *b = in_lookup_hash(parent, hash);
2439         struct hlist_bl_node *node;
2440         struct dentry *new = d_alloc(parent, name);
2441         struct dentry *dentry;
2442         unsigned seq, r_seq, d_seq;
2443
2444         if (unlikely(!new))
2445                 return ERR_PTR(-ENOMEM);
2446
2447 retry:
2448         rcu_read_lock();
2449         seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
2450         r_seq = read_seqbegin(&rename_lock);
2451         dentry = __d_lookup_rcu(parent, name, &d_seq);
2452         if (unlikely(dentry)) {
2453                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2454                         rcu_read_unlock();
2455                         goto retry;
2456                 }
2457                 if (read_seqcount_retry(&dentry->d_seq, d_seq)) {
2458                         rcu_read_unlock();
2459                         dput(dentry);
2460                         goto retry;
2461                 }
2462                 rcu_read_unlock();
2463                 dput(new);
2464                 return dentry;
2465         }
2466         if (unlikely(read_seqretry(&rename_lock, r_seq))) {
2467                 rcu_read_unlock();
2468                 goto retry;
2469         }
2470
2471         if (unlikely(seq & 1)) {
2472                 rcu_read_unlock();
2473                 goto retry;
2474         }
2475
2476         hlist_bl_lock(b);
2477         if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {
2478                 hlist_bl_unlock(b);
2479                 rcu_read_unlock();
2480                 goto retry;
2481         }
2482         /*
2483          * No changes for the parent since the beginning of d_lookup().
2484          * Since all removals from the chain happen with hlist_bl_lock(),
2485          * any potential in-lookup matches are going to stay here until
2486          * we unlock the chain.  All fields are stable in everything
2487          * we encounter.
2488          */
2489         hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
2490                 if (dentry->d_name.hash != hash)
2491                         continue;
2492                 if (dentry->d_parent != parent)
2493                         continue;
2494                 if (!d_same_name(dentry, parent, name))
2495                         continue;
2496                 hlist_bl_unlock(b);
2497                 /* now we can try to grab a reference */
2498                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2499                         rcu_read_unlock();
2500                         goto retry;
2501                 }
2502
2503                 rcu_read_unlock();
2504                 /*
2505                  * somebody is likely to be still doing lookup for it;
2506                  * wait for them to finish
2507                  */
2508                 spin_lock(&dentry->d_lock);
2509                 d_wait_lookup(dentry);
2510                 /*
2511                  * it's not in-lookup anymore; in principle we should repeat
2512                  * everything from dcache lookup, but it's likely to be what
2513                  * d_lookup() would've found anyway.  If it is, just return it;
2514                  * otherwise we really have to repeat the whole thing.
2515                  */
2516                 if (unlikely(dentry->d_name.hash != hash))
2517                         goto mismatch;
2518                 if (unlikely(dentry->d_parent != parent))
2519                         goto mismatch;
2520                 if (unlikely(d_unhashed(dentry)))
2521                         goto mismatch;
2522                 if (unlikely(!d_same_name(dentry, parent, name)))
2523                         goto mismatch;
2524                 /* OK, it *is* a hashed match; return it */
2525                 spin_unlock(&dentry->d_lock);
2526                 dput(new);
2527                 return dentry;
2528         }
2529         rcu_read_unlock();
2530         /* we can't take ->d_lock here; it's OK, though. */
2531         new->d_flags |= DCACHE_PAR_LOOKUP;
2532         new->d_wait = wq;
2533         hlist_bl_add_head_rcu(&new->d_u.d_in_lookup_hash, b);
2534         hlist_bl_unlock(b);
2535         return new;
2536 mismatch:
2537         spin_unlock(&dentry->d_lock);
2538         dput(dentry);
2539         goto retry;
2540 }
2541 EXPORT_SYMBOL(d_alloc_parallel);
2542
2543 void __d_lookup_done(struct dentry *dentry)
2544 {
2545         struct hlist_bl_head *b = in_lookup_hash(dentry->d_parent,
2546                                                  dentry->d_name.hash);
2547         hlist_bl_lock(b);
2548         dentry->d_flags &= ~DCACHE_PAR_LOOKUP;
2549         __hlist_bl_del(&dentry->d_u.d_in_lookup_hash);
2550         wake_up_all(dentry->d_wait);
2551         dentry->d_wait = NULL;
2552         hlist_bl_unlock(b);
2553         INIT_HLIST_NODE(&dentry->d_u.d_alias);
2554         INIT_LIST_HEAD(&dentry->d_lru);
2555 }
2556 EXPORT_SYMBOL(__d_lookup_done);
2557
2558 /* inode->i_lock held if inode is non-NULL */
2559
2560 static inline void __d_add(struct dentry *dentry, struct inode *inode)
2561 {
2562         struct inode *dir = NULL;
2563         unsigned n;
2564         spin_lock(&dentry->d_lock);
2565         if (unlikely(d_in_lookup(dentry))) {
2566                 dir = dentry->d_parent->d_inode;
2567                 n = start_dir_add(dir);
2568                 __d_lookup_done(dentry);
2569         }
2570         if (inode) {
2571                 unsigned add_flags = d_flags_for_inode(inode);
2572                 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2573                 raw_write_seqcount_begin(&dentry->d_seq);
2574                 __d_set_inode_and_type(dentry, inode, add_flags);
2575                 raw_write_seqcount_end(&dentry->d_seq);
2576                 fsnotify_update_flags(dentry);
2577         }
2578         __d_rehash(dentry);
2579         if (dir)
2580                 end_dir_add(dir, n);
2581         spin_unlock(&dentry->d_lock);
2582         if (inode)
2583                 spin_unlock(&inode->i_lock);
2584 }
2585
2586 /**
2587  * d_add - add dentry to hash queues
2588  * @entry: dentry to add
2589  * @inode: The inode to attach to this dentry
2590  *
2591  * This adds the entry to the hash queues and initializes @inode.
2592  * The entry was actually filled in earlier during d_alloc().
2593  */
2594
2595 void d_add(struct dentry *entry, struct inode *inode)
2596 {
2597         if (inode) {
2598                 security_d_instantiate(entry, inode);
2599                 spin_lock(&inode->i_lock);
2600         }
2601         __d_add(entry, inode);
2602 }
2603 EXPORT_SYMBOL(d_add);
2604
2605 /**
2606  * d_exact_alias - find and hash an exact unhashed alias
2607  * @entry: dentry to add
2608  * @inode: The inode to go with this dentry
2609  *
2610  * If an unhashed dentry with the same name/parent and desired
2611  * inode already exists, hash and return it.  Otherwise, return
2612  * NULL.
2613  *
2614  * Parent directory should be locked.
2615  */
2616 struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2617 {
2618         struct dentry *alias;
2619         unsigned int hash = entry->d_name.hash;
2620
2621         spin_lock(&inode->i_lock);
2622         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2623                 /*
2624                  * Don't need alias->d_lock here, because aliases with
2625                  * d_parent == entry->d_parent are not subject to name or
2626                  * parent changes, because the parent inode i_mutex is held.
2627                  */
2628                 if (alias->d_name.hash != hash)
2629                         continue;
2630                 if (alias->d_parent != entry->d_parent)
2631                         continue;
2632                 if (!d_same_name(alias, entry->d_parent, &entry->d_name))
2633                         continue;
2634                 spin_lock(&alias->d_lock);
2635                 if (!d_unhashed(alias)) {
2636                         spin_unlock(&alias->d_lock);
2637                         alias = NULL;
2638                 } else {
2639                         __dget_dlock(alias);
2640                         __d_rehash(alias);
2641                         spin_unlock(&alias->d_lock);
2642                 }
2643                 spin_unlock(&inode->i_lock);
2644                 return alias;
2645         }
2646         spin_unlock(&inode->i_lock);
2647         return NULL;
2648 }
2649 EXPORT_SYMBOL(d_exact_alias);
2650
2651 static void swap_names(struct dentry *dentry, struct dentry *target)
2652 {
2653         if (unlikely(dname_external(target))) {
2654                 if (unlikely(dname_external(dentry))) {
2655                         /*
2656                          * Both external: swap the pointers
2657                          */
2658                         swap(target->d_name.name, dentry->d_name.name);
2659                 } else {
2660                         /*
2661                          * dentry:internal, target:external.  Steal target's
2662                          * storage and make target internal.
2663                          */
2664                         memcpy(target->d_iname, dentry->d_name.name,
2665                                         dentry->d_name.len + 1);
2666                         dentry->d_name.name = target->d_name.name;
2667                         target->d_name.name = target->d_iname;
2668                 }
2669         } else {
2670                 if (unlikely(dname_external(dentry))) {
2671                         /*
2672                          * dentry:external, target:internal.  Give dentry's
2673                          * storage to target and make dentry internal
2674                          */
2675                         memcpy(dentry->d_iname, target->d_name.name,
2676                                         target->d_name.len + 1);
2677                         target->d_name.name = dentry->d_name.name;
2678                         dentry->d_name.name = dentry->d_iname;
2679                 } else {
2680                         /*
2681                          * Both are internal.
2682                          */
2683                         unsigned int i;
2684                         BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2685                         for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2686                                 swap(((long *) &dentry->d_iname)[i],
2687                                      ((long *) &target->d_iname)[i]);
2688                         }
2689                 }
2690         }
2691         swap(dentry->d_name.hash_len, target->d_name.hash_len);
2692 }
2693
2694 static void copy_name(struct dentry *dentry, struct dentry *target)
2695 {
2696         struct external_name *old_name = NULL;
2697         if (unlikely(dname_external(dentry)))
2698                 old_name = external_name(dentry);
2699         if (unlikely(dname_external(target))) {
2700                 atomic_inc(&external_name(target)->u.count);
2701                 dentry->d_name = target->d_name;
2702         } else {
2703                 memcpy(dentry->d_iname, target->d_name.name,
2704                                 target->d_name.len + 1);
2705                 dentry->d_name.name = dentry->d_iname;
2706                 dentry->d_name.hash_len = target->d_name.hash_len;
2707         }
2708         if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2709                 call_rcu(&old_name->u.head, __d_free_external_name);
2710 }
2711
2712 /*
2713  * __d_move - move a dentry
2714  * @dentry: entry to move
2715  * @target: new dentry
2716  * @exchange: exchange the two dentries
2717  *
2718  * Update the dcache to reflect the move of a file name. Negative
2719  * dcache entries should not be moved in this way. Caller must hold
2720  * rename_lock, the i_mutex of the source and target directories,
2721  * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2722  */
2723 static void __d_move(struct dentry *dentry, struct dentry *target,
2724                      bool exchange)
2725 {
2726         struct dentry *old_parent, *p;
2727         struct inode *dir = NULL;
2728         unsigned n;
2729
2730         WARN_ON(!dentry->d_inode);
2731         if (WARN_ON(dentry == target))
2732                 return;
2733
2734         BUG_ON(d_ancestor(target, dentry));
2735         old_parent = dentry->d_parent;
2736         p = d_ancestor(old_parent, target);
2737         if (IS_ROOT(dentry)) {
2738                 BUG_ON(p);
2739                 spin_lock(&target->d_parent->d_lock);
2740         } else if (!p) {
2741                 /* target is not a descendent of dentry->d_parent */
2742                 spin_lock(&target->d_parent->d_lock);
2743                 spin_lock_nested(&old_parent->d_lock, DENTRY_D_LOCK_NESTED);
2744         } else {
2745                 BUG_ON(p == dentry);
2746                 spin_lock(&old_parent->d_lock);
2747                 if (p != target)
2748                         spin_lock_nested(&target->d_parent->d_lock,
2749                                         DENTRY_D_LOCK_NESTED);
2750         }
2751         spin_lock_nested(&dentry->d_lock, 2);
2752         spin_lock_nested(&target->d_lock, 3);
2753
2754         if (unlikely(d_in_lookup(target))) {
2755                 dir = target->d_parent->d_inode;
2756                 n = start_dir_add(dir);
2757                 __d_lookup_done(target);
2758         }
2759
2760         write_seqcount_begin(&dentry->d_seq);
2761         write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2762
2763         /* unhash both */
2764         if (!d_unhashed(dentry))
2765                 ___d_drop(dentry);
2766         if (!d_unhashed(target))
2767                 ___d_drop(target);
2768
2769         /* ... and switch them in the tree */
2770         dentry->d_parent = target->d_parent;
2771         if (!exchange) {
2772                 copy_name(dentry, target);
2773                 target->d_hash.pprev = NULL;
2774                 dentry->d_parent->d_lockref.count++;
2775                 if (dentry == old_parent)
2776                         dentry->d_flags |= DCACHE_RCUACCESS;
2777                 else
2778                         WARN_ON(!--old_parent->d_lockref.count);
2779         } else {
2780                 target->d_parent = old_parent;
2781                 swap_names(dentry, target);
2782                 list_move(&target->d_child, &target->d_parent->d_subdirs);
2783                 __d_rehash(target);
2784                 fsnotify_update_flags(target);
2785         }
2786         list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2787         __d_rehash(dentry);
2788         fsnotify_update_flags(dentry);
2789
2790         write_seqcount_end(&target->d_seq);
2791         write_seqcount_end(&dentry->d_seq);
2792
2793         if (dir)
2794                 end_dir_add(dir, n);
2795
2796         if (dentry->d_parent != old_parent)
2797                 spin_unlock(&dentry->d_parent->d_lock);
2798         if (dentry != old_parent)
2799                 spin_unlock(&old_parent->d_lock);
2800         spin_unlock(&target->d_lock);
2801         spin_unlock(&dentry->d_lock);
2802 }
2803
2804 /*
2805  * d_move - move a dentry
2806  * @dentry: entry to move
2807  * @target: new dentry
2808  *
2809  * Update the dcache to reflect the move of a file name. Negative
2810  * dcache entries should not be moved in this way. See the locking
2811  * requirements for __d_move.
2812  */
2813 void d_move(struct dentry *dentry, struct dentry *target)
2814 {
2815         write_seqlock(&rename_lock);
2816         __d_move(dentry, target, false);
2817         write_sequnlock(&rename_lock);
2818 }
2819 EXPORT_SYMBOL(d_move);
2820
2821 /*
2822  * d_exchange - exchange two dentries
2823  * @dentry1: first dentry
2824  * @dentry2: second dentry
2825  */
2826 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2827 {
2828         write_seqlock(&rename_lock);
2829
2830         WARN_ON(!dentry1->d_inode);
2831         WARN_ON(!dentry2->d_inode);
2832         WARN_ON(IS_ROOT(dentry1));
2833         WARN_ON(IS_ROOT(dentry2));
2834
2835         __d_move(dentry1, dentry2, true);
2836
2837         write_sequnlock(&rename_lock);
2838 }
2839
2840 /**
2841  * d_ancestor - search for an ancestor
2842  * @p1: ancestor dentry
2843  * @p2: child dentry
2844  *
2845  * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2846  * an ancestor of p2, else NULL.
2847  */
2848 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2849 {
2850         struct dentry *p;
2851
2852         for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2853                 if (p->d_parent == p1)
2854                         return p;
2855         }
2856         return NULL;
2857 }
2858
2859 /*
2860  * This helper attempts to cope with remotely renamed directories
2861  *
2862  * It assumes that the caller is already holding
2863  * dentry->d_parent->d_inode->i_mutex, and rename_lock
2864  *
2865  * Note: If ever the locking in lock_rename() changes, then please
2866  * remember to update this too...
2867  */
2868 static int __d_unalias(struct inode *inode,
2869                 struct dentry *dentry, struct dentry *alias)
2870 {
2871         struct mutex *m1 = NULL;
2872         struct rw_semaphore *m2 = NULL;
2873         int ret = -ESTALE;
2874
2875         /* If alias and dentry share a parent, then no extra locks required */
2876         if (alias->d_parent == dentry->d_parent)
2877                 goto out_unalias;
2878
2879         /* See lock_rename() */
2880         if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2881                 goto out_err;
2882         m1 = &dentry->d_sb->s_vfs_rename_mutex;
2883         if (!inode_trylock_shared(alias->d_parent->d_inode))
2884                 goto out_err;
2885         m2 = &alias->d_parent->d_inode->i_rwsem;
2886 out_unalias:
2887         __d_move(alias, dentry, false);
2888         ret = 0;
2889 out_err:
2890         if (m2)
2891                 up_read(m2);
2892         if (m1)
2893                 mutex_unlock(m1);
2894         return ret;
2895 }
2896
2897 /**
2898  * d_splice_alias - splice a disconnected dentry into the tree if one exists
2899  * @inode:  the inode which may have a disconnected dentry
2900  * @dentry: a negative dentry which we want to point to the inode.
2901  *
2902  * If inode is a directory and has an IS_ROOT alias, then d_move that in
2903  * place of the given dentry and return it, else simply d_add the inode
2904  * to the dentry and return NULL.
2905  *
2906  * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2907  * we should error out: directories can't have multiple aliases.
2908  *
2909  * This is needed in the lookup routine of any filesystem that is exportable
2910  * (via knfsd) so that we can build dcache paths to directories effectively.
2911  *
2912  * If a dentry was found and moved, then it is returned.  Otherwise NULL
2913  * is returned.  This matches the expected return value of ->lookup.
2914  *
2915  * Cluster filesystems may call this function with a negative, hashed dentry.
2916  * In that case, we know that the inode will be a regular file, and also this
2917  * will only occur during atomic_open. So we need to check for the dentry
2918  * being already hashed only in the final case.
2919  */
2920 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
2921 {
2922         if (IS_ERR(inode))
2923                 return ERR_CAST(inode);
2924
2925         BUG_ON(!d_unhashed(dentry));
2926
2927         if (!inode)
2928                 goto out;
2929
2930         security_d_instantiate(dentry, inode);
2931         spin_lock(&inode->i_lock);
2932         if (S_ISDIR(inode->i_mode)) {
2933                 struct dentry *new = __d_find_any_alias(inode);
2934                 if (unlikely(new)) {
2935                         /* The reference to new ensures it remains an alias */
2936                         spin_unlock(&inode->i_lock);
2937                         write_seqlock(&rename_lock);
2938                         if (unlikely(d_ancestor(new, dentry))) {
2939                                 write_sequnlock(&rename_lock);
2940                                 dput(new);
2941                                 new = ERR_PTR(-ELOOP);
2942                                 pr_warn_ratelimited(
2943                                         "VFS: Lookup of '%s' in %s %s"
2944                                         " would have caused loop\n",
2945                                         dentry->d_name.name,
2946                                         inode->i_sb->s_type->name,
2947                                         inode->i_sb->s_id);
2948                         } else if (!IS_ROOT(new)) {
2949                                 struct dentry *old_parent = dget(new->d_parent);
2950                                 int err = __d_unalias(inode, dentry, new);
2951                                 write_sequnlock(&rename_lock);
2952                                 if (err) {
2953                                         dput(new);
2954                                         new = ERR_PTR(err);
2955                                 }
2956                                 dput(old_parent);
2957                         } else {
2958                                 __d_move(new, dentry, false);
2959                                 write_sequnlock(&rename_lock);
2960                         }
2961                         iput(inode);
2962                         return new;
2963                 }
2964         }
2965 out:
2966         __d_add(dentry, inode);
2967         return NULL;
2968 }
2969 EXPORT_SYMBOL(d_splice_alias);
2970
2971 /*
2972  * Test whether new_dentry is a subdirectory of old_dentry.
2973  *
2974  * Trivially implemented using the dcache structure
2975  */
2976
2977 /**
2978  * is_subdir - is new dentry a subdirectory of old_dentry
2979  * @new_dentry: new dentry
2980  * @old_dentry: old dentry
2981  *
2982  * Returns true if new_dentry is a subdirectory of the parent (at any depth).
2983  * Returns false otherwise.
2984  * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
2985  */
2986   
2987 bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
2988 {
2989         bool result;
2990         unsigned seq;
2991
2992         if (new_dentry == old_dentry)
2993                 return true;
2994
2995         do {
2996                 /* for restarting inner loop in case of seq retry */
2997                 seq = read_seqbegin(&rename_lock);
2998                 /*
2999                  * Need rcu_readlock to protect against the d_parent trashing
3000                  * due to d_move
3001                  */
3002                 rcu_read_lock();
3003                 if (d_ancestor(old_dentry, new_dentry))
3004                         result = true;
3005                 else
3006                         result = false;
3007                 rcu_read_unlock();
3008         } while (read_seqretry(&rename_lock, seq));
3009
3010         return result;
3011 }
3012 EXPORT_SYMBOL(is_subdir);
3013
3014 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3015 {
3016         struct dentry *root = data;
3017         if (dentry != root) {
3018                 if (d_unhashed(dentry) || !dentry->d_inode)
3019                         return D_WALK_SKIP;
3020
3021                 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3022                         dentry->d_flags |= DCACHE_GENOCIDE;
3023                         dentry->d_lockref.count--;
3024                 }
3025         }
3026         return D_WALK_CONTINUE;
3027 }
3028
3029 void d_genocide(struct dentry *parent)
3030 {
3031         d_walk(parent, parent, d_genocide_kill);
3032 }
3033
3034 EXPORT_SYMBOL(d_genocide);
3035
3036 void d_tmpfile(struct dentry *dentry, struct inode *inode)
3037 {
3038         inode_dec_link_count(inode);
3039         BUG_ON(dentry->d_name.name != dentry->d_iname ||
3040                 !hlist_unhashed(&dentry->d_u.d_alias) ||
3041                 !d_unlinked(dentry));
3042         spin_lock(&dentry->d_parent->d_lock);
3043         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3044         dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3045                                 (unsigned long long)inode->i_ino);
3046         spin_unlock(&dentry->d_lock);
3047         spin_unlock(&dentry->d_parent->d_lock);
3048         d_instantiate(dentry, inode);
3049 }
3050 EXPORT_SYMBOL(d_tmpfile);
3051
3052 static __initdata unsigned long dhash_entries;
3053 static int __init set_dhash_entries(char *str)
3054 {
3055         if (!str)
3056                 return 0;
3057         dhash_entries = simple_strtoul(str, &str, 0);
3058         return 1;
3059 }
3060 __setup("dhash_entries=", set_dhash_entries);
3061
3062 static void __init dcache_init_early(void)
3063 {
3064         /* If hashes are distributed across NUMA nodes, defer
3065          * hash allocation until vmalloc space is available.
3066          */
3067         if (hashdist)
3068                 return;
3069
3070         dentry_hashtable =
3071                 alloc_large_system_hash("Dentry cache",
3072                                         sizeof(struct hlist_bl_head),
3073                                         dhash_entries,
3074                                         13,
3075                                         HASH_EARLY | HASH_ZERO,
3076                                         &d_hash_shift,
3077                                         NULL,
3078                                         0,
3079                                         0);
3080         d_hash_shift = 32 - d_hash_shift;
3081 }
3082
3083 static void __init dcache_init(void)
3084 {
3085         /*
3086          * A constructor could be added for stable state like the lists,
3087          * but it is probably not worth it because of the cache nature
3088          * of the dcache.
3089          */
3090         dentry_cache = KMEM_CACHE_USERCOPY(dentry,
3091                 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD|SLAB_ACCOUNT,
3092                 d_iname);
3093
3094         /* Hash may have been set up in dcache_init_early */
3095         if (!hashdist)
3096                 return;
3097
3098         dentry_hashtable =
3099                 alloc_large_system_hash("Dentry cache",
3100                                         sizeof(struct hlist_bl_head),
3101                                         dhash_entries,
3102                                         13,
3103                                         HASH_ZERO,
3104                                         &d_hash_shift,
3105                                         NULL,
3106                                         0,
3107                                         0);
3108         d_hash_shift = 32 - d_hash_shift;
3109 }
3110
3111 /* SLAB cache for __getname() consumers */
3112 struct kmem_cache *names_cachep __read_mostly;
3113 EXPORT_SYMBOL(names_cachep);
3114
3115 void __init vfs_caches_init_early(void)
3116 {
3117         int i;
3118
3119         for (i = 0; i < ARRAY_SIZE(in_lookup_hashtable); i++)
3120                 INIT_HLIST_BL_HEAD(&in_lookup_hashtable[i]);
3121
3122         dcache_init_early();
3123         inode_init_early();
3124 }
3125
3126 void __init vfs_caches_init(void)
3127 {
3128         names_cachep = kmem_cache_create_usercopy("names_cache", PATH_MAX, 0,
3129                         SLAB_HWCACHE_ALIGN|SLAB_PANIC, 0, PATH_MAX, NULL);
3130
3131         dcache_init();
3132         inode_init();
3133         files_init();
3134         files_maxfiles_init();
3135         mnt_init();
3136         bdev_cache_init();
3137         chrdev_init();
3138 }