mm, migrate: remove reason argument from new_page_t
[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         bool hashed = !d_unhashed(dentry);
362
363         if (hashed)
364                 raw_write_seqcount_begin(&dentry->d_seq);
365         __d_clear_type_and_inode(dentry);
366         hlist_del_init(&dentry->d_u.d_alias);
367         if (hashed)
368                 raw_write_seqcount_end(&dentry->d_seq);
369         spin_unlock(&dentry->d_lock);
370         spin_unlock(&inode->i_lock);
371         if (!inode->i_nlink)
372                 fsnotify_inoderemove(inode);
373         if (dentry->d_op && dentry->d_op->d_iput)
374                 dentry->d_op->d_iput(dentry, inode);
375         else
376                 iput(inode);
377 }
378
379 /*
380  * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
381  * is in use - which includes both the "real" per-superblock
382  * LRU list _and_ the DCACHE_SHRINK_LIST use.
383  *
384  * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
385  * on the shrink list (ie not on the superblock LRU list).
386  *
387  * The per-cpu "nr_dentry_unused" counters are updated with
388  * the DCACHE_LRU_LIST bit.
389  *
390  * These helper functions make sure we always follow the
391  * rules. d_lock must be held by the caller.
392  */
393 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
394 static void d_lru_add(struct dentry *dentry)
395 {
396         D_FLAG_VERIFY(dentry, 0);
397         dentry->d_flags |= DCACHE_LRU_LIST;
398         this_cpu_inc(nr_dentry_unused);
399         WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
400 }
401
402 static void d_lru_del(struct dentry *dentry)
403 {
404         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
405         dentry->d_flags &= ~DCACHE_LRU_LIST;
406         this_cpu_dec(nr_dentry_unused);
407         WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
408 }
409
410 static void d_shrink_del(struct dentry *dentry)
411 {
412         D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
413         list_del_init(&dentry->d_lru);
414         dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
415         this_cpu_dec(nr_dentry_unused);
416 }
417
418 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
419 {
420         D_FLAG_VERIFY(dentry, 0);
421         list_add(&dentry->d_lru, list);
422         dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
423         this_cpu_inc(nr_dentry_unused);
424 }
425
426 /*
427  * These can only be called under the global LRU lock, ie during the
428  * callback for freeing the LRU list. "isolate" removes it from the
429  * LRU lists entirely, while shrink_move moves it to the indicated
430  * private list.
431  */
432 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
433 {
434         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
435         dentry->d_flags &= ~DCACHE_LRU_LIST;
436         this_cpu_dec(nr_dentry_unused);
437         list_lru_isolate(lru, &dentry->d_lru);
438 }
439
440 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
441                               struct list_head *list)
442 {
443         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
444         dentry->d_flags |= DCACHE_SHRINK_LIST;
445         list_lru_isolate_move(lru, &dentry->d_lru, list);
446 }
447
448 /**
449  * d_drop - drop a dentry
450  * @dentry: dentry to drop
451  *
452  * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
453  * be found through a VFS lookup any more. Note that this is different from
454  * deleting the dentry - d_delete will try to mark the dentry negative if
455  * possible, giving a successful _negative_ lookup, while d_drop will
456  * just make the cache lookup fail.
457  *
458  * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
459  * reason (NFS timeouts or autofs deletes).
460  *
461  * __d_drop requires dentry->d_lock
462  * ___d_drop doesn't mark dentry as "unhashed"
463  *   (dentry->d_hash.pprev will be LIST_POISON2, not NULL).
464  */
465 static void ___d_drop(struct dentry *dentry)
466 {
467         struct hlist_bl_head *b;
468         /*
469          * Hashed dentries are normally on the dentry hashtable,
470          * with the exception of those newly allocated by
471          * d_obtain_root, which are always IS_ROOT:
472          */
473         if (unlikely(IS_ROOT(dentry)))
474                 b = &dentry->d_sb->s_roots;
475         else
476                 b = d_hash(dentry->d_name.hash);
477
478         hlist_bl_lock(b);
479         __hlist_bl_del(&dentry->d_hash);
480         hlist_bl_unlock(b);
481 }
482
483 void __d_drop(struct dentry *dentry)
484 {
485         if (!d_unhashed(dentry)) {
486                 ___d_drop(dentry);
487                 dentry->d_hash.pprev = NULL;
488                 write_seqcount_invalidate(&dentry->d_seq);
489         }
490 }
491 EXPORT_SYMBOL(__d_drop);
492
493 void d_drop(struct dentry *dentry)
494 {
495         spin_lock(&dentry->d_lock);
496         __d_drop(dentry);
497         spin_unlock(&dentry->d_lock);
498 }
499 EXPORT_SYMBOL(d_drop);
500
501 static inline void dentry_unlist(struct dentry *dentry, struct dentry *parent)
502 {
503         struct dentry *next;
504         /*
505          * Inform d_walk() and shrink_dentry_list() that we are no longer
506          * attached to the dentry tree
507          */
508         dentry->d_flags |= DCACHE_DENTRY_KILLED;
509         if (unlikely(list_empty(&dentry->d_child)))
510                 return;
511         __list_del_entry(&dentry->d_child);
512         /*
513          * Cursors can move around the list of children.  While we'd been
514          * a normal list member, it didn't matter - ->d_child.next would've
515          * been updated.  However, from now on it won't be and for the
516          * things like d_walk() it might end up with a nasty surprise.
517          * Normally d_walk() doesn't care about cursors moving around -
518          * ->d_lock on parent prevents that and since a cursor has no children
519          * of its own, we get through it without ever unlocking the parent.
520          * There is one exception, though - if we ascend from a child that
521          * gets killed as soon as we unlock it, the next sibling is found
522          * using the value left in its ->d_child.next.  And if _that_
523          * pointed to a cursor, and cursor got moved (e.g. by lseek())
524          * before d_walk() regains parent->d_lock, we'll end up skipping
525          * everything the cursor had been moved past.
526          *
527          * Solution: make sure that the pointer left behind in ->d_child.next
528          * points to something that won't be moving around.  I.e. skip the
529          * cursors.
530          */
531         while (dentry->d_child.next != &parent->d_subdirs) {
532                 next = list_entry(dentry->d_child.next, struct dentry, d_child);
533                 if (likely(!(next->d_flags & DCACHE_DENTRY_CURSOR)))
534                         break;
535                 dentry->d_child.next = next->d_child.next;
536         }
537 }
538
539 static void __dentry_kill(struct dentry *dentry)
540 {
541         struct dentry *parent = NULL;
542         bool can_free = true;
543         if (!IS_ROOT(dentry))
544                 parent = dentry->d_parent;
545
546         /*
547          * The dentry is now unrecoverably dead to the world.
548          */
549         lockref_mark_dead(&dentry->d_lockref);
550
551         /*
552          * inform the fs via d_prune that this dentry is about to be
553          * unhashed and destroyed.
554          */
555         if (dentry->d_flags & DCACHE_OP_PRUNE)
556                 dentry->d_op->d_prune(dentry);
557
558         if (dentry->d_flags & DCACHE_LRU_LIST) {
559                 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
560                         d_lru_del(dentry);
561         }
562         /* if it was on the hash then remove it */
563         __d_drop(dentry);
564         dentry_unlist(dentry, parent);
565         if (parent)
566                 spin_unlock(&parent->d_lock);
567         if (dentry->d_inode)
568                 dentry_unlink_inode(dentry);
569         else
570                 spin_unlock(&dentry->d_lock);
571         this_cpu_dec(nr_dentry);
572         if (dentry->d_op && dentry->d_op->d_release)
573                 dentry->d_op->d_release(dentry);
574
575         spin_lock(&dentry->d_lock);
576         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
577                 dentry->d_flags |= DCACHE_MAY_FREE;
578                 can_free = false;
579         }
580         spin_unlock(&dentry->d_lock);
581         if (likely(can_free))
582                 dentry_free(dentry);
583 }
584
585 static struct dentry *__lock_parent(struct dentry *dentry)
586 {
587         struct dentry *parent;
588         rcu_read_lock();
589         spin_unlock(&dentry->d_lock);
590 again:
591         parent = READ_ONCE(dentry->d_parent);
592         spin_lock(&parent->d_lock);
593         /*
594          * We can't blindly lock dentry until we are sure
595          * that we won't violate the locking order.
596          * Any changes of dentry->d_parent must have
597          * been done with parent->d_lock held, so
598          * spin_lock() above is enough of a barrier
599          * for checking if it's still our child.
600          */
601         if (unlikely(parent != dentry->d_parent)) {
602                 spin_unlock(&parent->d_lock);
603                 goto again;
604         }
605         rcu_read_unlock();
606         if (parent != dentry)
607                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
608         else
609                 parent = NULL;
610         return parent;
611 }
612
613 static inline struct dentry *lock_parent(struct dentry *dentry)
614 {
615         struct dentry *parent = dentry->d_parent;
616         if (IS_ROOT(dentry))
617                 return NULL;
618         if (likely(spin_trylock(&parent->d_lock)))
619                 return parent;
620         return __lock_parent(dentry);
621 }
622
623 static inline bool retain_dentry(struct dentry *dentry)
624 {
625         WARN_ON(d_in_lookup(dentry));
626
627         /* Unreachable? Get rid of it */
628         if (unlikely(d_unhashed(dentry)))
629                 return false;
630
631         if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
632                 return false;
633
634         if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
635                 if (dentry->d_op->d_delete(dentry))
636                         return false;
637         }
638         /* retain; LRU fodder */
639         dentry->d_lockref.count--;
640         if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
641                 d_lru_add(dentry);
642         else if (unlikely(!(dentry->d_flags & DCACHE_REFERENCED)))
643                 dentry->d_flags |= DCACHE_REFERENCED;
644         return true;
645 }
646
647 /*
648  * Finish off a dentry we've decided to kill.
649  * dentry->d_lock must be held, returns with it unlocked.
650  * Returns dentry requiring refcount drop, or NULL if we're done.
651  */
652 static struct dentry *dentry_kill(struct dentry *dentry)
653         __releases(dentry->d_lock)
654 {
655         struct inode *inode = dentry->d_inode;
656         struct dentry *parent = NULL;
657
658         if (inode && unlikely(!spin_trylock(&inode->i_lock)))
659                 goto slow_positive;
660
661         if (!IS_ROOT(dentry)) {
662                 parent = dentry->d_parent;
663                 if (unlikely(!spin_trylock(&parent->d_lock))) {
664                         parent = __lock_parent(dentry);
665                         if (likely(inode || !dentry->d_inode))
666                                 goto got_locks;
667                         /* negative that became positive */
668                         if (parent)
669                                 spin_unlock(&parent->d_lock);
670                         inode = dentry->d_inode;
671                         goto slow_positive;
672                 }
673         }
674         __dentry_kill(dentry);
675         return parent;
676
677 slow_positive:
678         spin_unlock(&dentry->d_lock);
679         spin_lock(&inode->i_lock);
680         spin_lock(&dentry->d_lock);
681         parent = lock_parent(dentry);
682 got_locks:
683         if (unlikely(dentry->d_lockref.count != 1)) {
684                 dentry->d_lockref.count--;
685         } else if (likely(!retain_dentry(dentry))) {
686                 __dentry_kill(dentry);
687                 return parent;
688         }
689         /* we are keeping it, after all */
690         if (inode)
691                 spin_unlock(&inode->i_lock);
692         if (parent)
693                 spin_unlock(&parent->d_lock);
694         spin_unlock(&dentry->d_lock);
695         return NULL;
696 }
697
698 /*
699  * Try to do a lockless dput(), and return whether that was successful.
700  *
701  * If unsuccessful, we return false, having already taken the dentry lock.
702  *
703  * The caller needs to hold the RCU read lock, so that the dentry is
704  * guaranteed to stay around even if the refcount goes down to zero!
705  */
706 static inline bool fast_dput(struct dentry *dentry)
707 {
708         int ret;
709         unsigned int d_flags;
710
711         /*
712          * If we have a d_op->d_delete() operation, we sould not
713          * let the dentry count go to zero, so use "put_or_lock".
714          */
715         if (unlikely(dentry->d_flags & DCACHE_OP_DELETE))
716                 return lockref_put_or_lock(&dentry->d_lockref);
717
718         /*
719          * .. otherwise, we can try to just decrement the
720          * lockref optimistically.
721          */
722         ret = lockref_put_return(&dentry->d_lockref);
723
724         /*
725          * If the lockref_put_return() failed due to the lock being held
726          * by somebody else, the fast path has failed. We will need to
727          * get the lock, and then check the count again.
728          */
729         if (unlikely(ret < 0)) {
730                 spin_lock(&dentry->d_lock);
731                 if (dentry->d_lockref.count > 1) {
732                         dentry->d_lockref.count--;
733                         spin_unlock(&dentry->d_lock);
734                         return 1;
735                 }
736                 return 0;
737         }
738
739         /*
740          * If we weren't the last ref, we're done.
741          */
742         if (ret)
743                 return 1;
744
745         /*
746          * Careful, careful. The reference count went down
747          * to zero, but we don't hold the dentry lock, so
748          * somebody else could get it again, and do another
749          * dput(), and we need to not race with that.
750          *
751          * However, there is a very special and common case
752          * where we don't care, because there is nothing to
753          * do: the dentry is still hashed, it does not have
754          * a 'delete' op, and it's referenced and already on
755          * the LRU list.
756          *
757          * NOTE! Since we aren't locked, these values are
758          * not "stable". However, it is sufficient that at
759          * some point after we dropped the reference the
760          * dentry was hashed and the flags had the proper
761          * value. Other dentry users may have re-gotten
762          * a reference to the dentry and change that, but
763          * our work is done - we can leave the dentry
764          * around with a zero refcount.
765          */
766         smp_rmb();
767         d_flags = READ_ONCE(dentry->d_flags);
768         d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST | DCACHE_DISCONNECTED;
769
770         /* Nothing to do? Dropping the reference was all we needed? */
771         if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry))
772                 return 1;
773
774         /*
775          * Not the fast normal case? Get the lock. We've already decremented
776          * the refcount, but we'll need to re-check the situation after
777          * getting the lock.
778          */
779         spin_lock(&dentry->d_lock);
780
781         /*
782          * Did somebody else grab a reference to it in the meantime, and
783          * we're no longer the last user after all? Alternatively, somebody
784          * else could have killed it and marked it dead. Either way, we
785          * don't need to do anything else.
786          */
787         if (dentry->d_lockref.count) {
788                 spin_unlock(&dentry->d_lock);
789                 return 1;
790         }
791
792         /*
793          * Re-get the reference we optimistically dropped. We hold the
794          * lock, and we just tested that it was zero, so we can just
795          * set it to 1.
796          */
797         dentry->d_lockref.count = 1;
798         return 0;
799 }
800
801
802 /* 
803  * This is dput
804  *
805  * This is complicated by the fact that we do not want to put
806  * dentries that are no longer on any hash chain on the unused
807  * list: we'd much rather just get rid of them immediately.
808  *
809  * However, that implies that we have to traverse the dentry
810  * tree upwards to the parents which might _also_ now be
811  * scheduled for deletion (it may have been only waiting for
812  * its last child to go away).
813  *
814  * This tail recursion is done by hand as we don't want to depend
815  * on the compiler to always get this right (gcc generally doesn't).
816  * Real recursion would eat up our stack space.
817  */
818
819 /*
820  * dput - release a dentry
821  * @dentry: dentry to release 
822  *
823  * Release a dentry. This will drop the usage count and if appropriate
824  * call the dentry unlink method as well as removing it from the queues and
825  * releasing its resources. If the parent dentries were scheduled for release
826  * they too may now get deleted.
827  */
828 void dput(struct dentry *dentry)
829 {
830         if (unlikely(!dentry))
831                 return;
832
833 repeat:
834         might_sleep();
835
836         rcu_read_lock();
837         if (likely(fast_dput(dentry))) {
838                 rcu_read_unlock();
839                 return;
840         }
841
842         /* Slow case: now with the dentry lock held */
843         rcu_read_unlock();
844
845         if (likely(retain_dentry(dentry))) {
846                 spin_unlock(&dentry->d_lock);
847                 return;
848         }
849
850         dentry = dentry_kill(dentry);
851         if (dentry) {
852                 cond_resched();
853                 goto repeat;
854         }
855 }
856 EXPORT_SYMBOL(dput);
857
858
859 /* This must be called with d_lock held */
860 static inline void __dget_dlock(struct dentry *dentry)
861 {
862         dentry->d_lockref.count++;
863 }
864
865 static inline void __dget(struct dentry *dentry)
866 {
867         lockref_get(&dentry->d_lockref);
868 }
869
870 struct dentry *dget_parent(struct dentry *dentry)
871 {
872         int gotref;
873         struct dentry *ret;
874
875         /*
876          * Do optimistic parent lookup without any
877          * locking.
878          */
879         rcu_read_lock();
880         ret = READ_ONCE(dentry->d_parent);
881         gotref = lockref_get_not_zero(&ret->d_lockref);
882         rcu_read_unlock();
883         if (likely(gotref)) {
884                 if (likely(ret == READ_ONCE(dentry->d_parent)))
885                         return ret;
886                 dput(ret);
887         }
888
889 repeat:
890         /*
891          * Don't need rcu_dereference because we re-check it was correct under
892          * the lock.
893          */
894         rcu_read_lock();
895         ret = dentry->d_parent;
896         spin_lock(&ret->d_lock);
897         if (unlikely(ret != dentry->d_parent)) {
898                 spin_unlock(&ret->d_lock);
899                 rcu_read_unlock();
900                 goto repeat;
901         }
902         rcu_read_unlock();
903         BUG_ON(!ret->d_lockref.count);
904         ret->d_lockref.count++;
905         spin_unlock(&ret->d_lock);
906         return ret;
907 }
908 EXPORT_SYMBOL(dget_parent);
909
910 /**
911  * d_find_alias - grab a hashed alias of inode
912  * @inode: inode in question
913  *
914  * If inode has a hashed alias, or is a directory and has any alias,
915  * acquire the reference to alias and return it. Otherwise return NULL.
916  * Notice that if inode is a directory there can be only one alias and
917  * it can be unhashed only if it has no children, or if it is the root
918  * of a filesystem, or if the directory was renamed and d_revalidate
919  * was the first vfs operation to notice.
920  *
921  * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
922  * any other hashed alias over that one.
923  */
924 static struct dentry *__d_find_alias(struct inode *inode)
925 {
926         struct dentry *alias, *discon_alias;
927
928 again:
929         discon_alias = NULL;
930         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
931                 spin_lock(&alias->d_lock);
932                 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
933                         if (IS_ROOT(alias) &&
934                             (alias->d_flags & DCACHE_DISCONNECTED)) {
935                                 discon_alias = alias;
936                         } else {
937                                 __dget_dlock(alias);
938                                 spin_unlock(&alias->d_lock);
939                                 return alias;
940                         }
941                 }
942                 spin_unlock(&alias->d_lock);
943         }
944         if (discon_alias) {
945                 alias = discon_alias;
946                 spin_lock(&alias->d_lock);
947                 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
948                         __dget_dlock(alias);
949                         spin_unlock(&alias->d_lock);
950                         return alias;
951                 }
952                 spin_unlock(&alias->d_lock);
953                 goto again;
954         }
955         return NULL;
956 }
957
958 struct dentry *d_find_alias(struct inode *inode)
959 {
960         struct dentry *de = NULL;
961
962         if (!hlist_empty(&inode->i_dentry)) {
963                 spin_lock(&inode->i_lock);
964                 de = __d_find_alias(inode);
965                 spin_unlock(&inode->i_lock);
966         }
967         return de;
968 }
969 EXPORT_SYMBOL(d_find_alias);
970
971 /*
972  *      Try to kill dentries associated with this inode.
973  * WARNING: you must own a reference to inode.
974  */
975 void d_prune_aliases(struct inode *inode)
976 {
977         struct dentry *dentry;
978 restart:
979         spin_lock(&inode->i_lock);
980         hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
981                 spin_lock(&dentry->d_lock);
982                 if (!dentry->d_lockref.count) {
983                         struct dentry *parent = lock_parent(dentry);
984                         if (likely(!dentry->d_lockref.count)) {
985                                 __dentry_kill(dentry);
986                                 dput(parent);
987                                 goto restart;
988                         }
989                         if (parent)
990                                 spin_unlock(&parent->d_lock);
991                 }
992                 spin_unlock(&dentry->d_lock);
993         }
994         spin_unlock(&inode->i_lock);
995 }
996 EXPORT_SYMBOL(d_prune_aliases);
997
998 /*
999  * Lock a dentry from shrink list.
1000  * Called under rcu_read_lock() and dentry->d_lock; the former
1001  * guarantees that nothing we access will be freed under us.
1002  * Note that dentry is *not* protected from concurrent dentry_kill(),
1003  * d_delete(), etc.
1004  *
1005  * Return false if dentry has been disrupted or grabbed, leaving
1006  * the caller to kick it off-list.  Otherwise, return true and have
1007  * that dentry's inode and parent both locked.
1008  */
1009 static bool shrink_lock_dentry(struct dentry *dentry)
1010 {
1011         struct inode *inode;
1012         struct dentry *parent;
1013
1014         if (dentry->d_lockref.count)
1015                 return false;
1016
1017         inode = dentry->d_inode;
1018         if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
1019                 spin_unlock(&dentry->d_lock);
1020                 spin_lock(&inode->i_lock);
1021                 spin_lock(&dentry->d_lock);
1022                 if (unlikely(dentry->d_lockref.count))
1023                         goto out;
1024                 /* changed inode means that somebody had grabbed it */
1025                 if (unlikely(inode != dentry->d_inode))
1026                         goto out;
1027         }
1028
1029         parent = dentry->d_parent;
1030         if (IS_ROOT(dentry) || likely(spin_trylock(&parent->d_lock)))
1031                 return true;
1032
1033         spin_unlock(&dentry->d_lock);
1034         spin_lock(&parent->d_lock);
1035         if (unlikely(parent != dentry->d_parent)) {
1036                 spin_unlock(&parent->d_lock);
1037                 spin_lock(&dentry->d_lock);
1038                 goto out;
1039         }
1040         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1041         if (likely(!dentry->d_lockref.count))
1042                 return true;
1043         spin_unlock(&parent->d_lock);
1044 out:
1045         if (inode)
1046                 spin_unlock(&inode->i_lock);
1047         return false;
1048 }
1049
1050 static void shrink_dentry_list(struct list_head *list)
1051 {
1052         while (!list_empty(list)) {
1053                 struct dentry *dentry, *parent;
1054
1055                 dentry = list_entry(list->prev, struct dentry, d_lru);
1056                 spin_lock(&dentry->d_lock);
1057                 rcu_read_lock();
1058                 if (!shrink_lock_dentry(dentry)) {
1059                         bool can_free = false;
1060                         rcu_read_unlock();
1061                         d_shrink_del(dentry);
1062                         if (dentry->d_lockref.count < 0)
1063                                 can_free = dentry->d_flags & DCACHE_MAY_FREE;
1064                         spin_unlock(&dentry->d_lock);
1065                         if (can_free)
1066                                 dentry_free(dentry);
1067                         continue;
1068                 }
1069                 rcu_read_unlock();
1070                 d_shrink_del(dentry);
1071                 parent = dentry->d_parent;
1072                 __dentry_kill(dentry);
1073                 if (parent == dentry)
1074                         continue;
1075                 /*
1076                  * We need to prune ancestors too. This is necessary to prevent
1077                  * quadratic behavior of shrink_dcache_parent(), but is also
1078                  * expected to be beneficial in reducing dentry cache
1079                  * fragmentation.
1080                  */
1081                 dentry = parent;
1082                 while (dentry && !lockref_put_or_lock(&dentry->d_lockref))
1083                         dentry = dentry_kill(dentry);
1084         }
1085 }
1086
1087 static enum lru_status dentry_lru_isolate(struct list_head *item,
1088                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1089 {
1090         struct list_head *freeable = arg;
1091         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1092
1093
1094         /*
1095          * we are inverting the lru lock/dentry->d_lock here,
1096          * so use a trylock. If we fail to get the lock, just skip
1097          * it
1098          */
1099         if (!spin_trylock(&dentry->d_lock))
1100                 return LRU_SKIP;
1101
1102         /*
1103          * Referenced dentries are still in use. If they have active
1104          * counts, just remove them from the LRU. Otherwise give them
1105          * another pass through the LRU.
1106          */
1107         if (dentry->d_lockref.count) {
1108                 d_lru_isolate(lru, dentry);
1109                 spin_unlock(&dentry->d_lock);
1110                 return LRU_REMOVED;
1111         }
1112
1113         if (dentry->d_flags & DCACHE_REFERENCED) {
1114                 dentry->d_flags &= ~DCACHE_REFERENCED;
1115                 spin_unlock(&dentry->d_lock);
1116
1117                 /*
1118                  * The list move itself will be made by the common LRU code. At
1119                  * this point, we've dropped the dentry->d_lock but keep the
1120                  * lru lock. This is safe to do, since every list movement is
1121                  * protected by the lru lock even if both locks are held.
1122                  *
1123                  * This is guaranteed by the fact that all LRU management
1124                  * functions are intermediated by the LRU API calls like
1125                  * list_lru_add and list_lru_del. List movement in this file
1126                  * only ever occur through this functions or through callbacks
1127                  * like this one, that are called from the LRU API.
1128                  *
1129                  * The only exceptions to this are functions like
1130                  * shrink_dentry_list, and code that first checks for the
1131                  * DCACHE_SHRINK_LIST flag.  Those are guaranteed to be
1132                  * operating only with stack provided lists after they are
1133                  * properly isolated from the main list.  It is thus, always a
1134                  * local access.
1135                  */
1136                 return LRU_ROTATE;
1137         }
1138
1139         d_lru_shrink_move(lru, dentry, freeable);
1140         spin_unlock(&dentry->d_lock);
1141
1142         return LRU_REMOVED;
1143 }
1144
1145 /**
1146  * prune_dcache_sb - shrink the dcache
1147  * @sb: superblock
1148  * @sc: shrink control, passed to list_lru_shrink_walk()
1149  *
1150  * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1151  * is done when we need more memory and called from the superblock shrinker
1152  * function.
1153  *
1154  * This function may fail to free any resources if all the dentries are in
1155  * use.
1156  */
1157 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1158 {
1159         LIST_HEAD(dispose);
1160         long freed;
1161
1162         freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1163                                      dentry_lru_isolate, &dispose);
1164         shrink_dentry_list(&dispose);
1165         return freed;
1166 }
1167
1168 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1169                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1170 {
1171         struct list_head *freeable = arg;
1172         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1173
1174         /*
1175          * we are inverting the lru lock/dentry->d_lock here,
1176          * so use a trylock. If we fail to get the lock, just skip
1177          * it
1178          */
1179         if (!spin_trylock(&dentry->d_lock))
1180                 return LRU_SKIP;
1181
1182         d_lru_shrink_move(lru, dentry, freeable);
1183         spin_unlock(&dentry->d_lock);
1184
1185         return LRU_REMOVED;
1186 }
1187
1188
1189 /**
1190  * shrink_dcache_sb - shrink dcache for a superblock
1191  * @sb: superblock
1192  *
1193  * Shrink the dcache for the specified super block. This is used to free
1194  * the dcache before unmounting a file system.
1195  */
1196 void shrink_dcache_sb(struct super_block *sb)
1197 {
1198         long freed;
1199
1200         do {
1201                 LIST_HEAD(dispose);
1202
1203                 freed = list_lru_walk(&sb->s_dentry_lru,
1204                         dentry_lru_isolate_shrink, &dispose, 1024);
1205
1206                 this_cpu_sub(nr_dentry_unused, freed);
1207                 shrink_dentry_list(&dispose);
1208                 cond_resched();
1209         } while (list_lru_count(&sb->s_dentry_lru) > 0);
1210 }
1211 EXPORT_SYMBOL(shrink_dcache_sb);
1212
1213 /**
1214  * enum d_walk_ret - action to talke during tree walk
1215  * @D_WALK_CONTINUE:    contrinue walk
1216  * @D_WALK_QUIT:        quit walk
1217  * @D_WALK_NORETRY:     quit when retry is needed
1218  * @D_WALK_SKIP:        skip this dentry and its children
1219  */
1220 enum d_walk_ret {
1221         D_WALK_CONTINUE,
1222         D_WALK_QUIT,
1223         D_WALK_NORETRY,
1224         D_WALK_SKIP,
1225 };
1226
1227 /**
1228  * d_walk - walk the dentry tree
1229  * @parent:     start of walk
1230  * @data:       data passed to @enter() and @finish()
1231  * @enter:      callback when first entering the dentry
1232  * @finish:     callback when successfully finished the walk
1233  *
1234  * The @enter() and @finish() callbacks are called with d_lock held.
1235  */
1236 static void d_walk(struct dentry *parent, void *data,
1237                    enum d_walk_ret (*enter)(void *, struct dentry *),
1238                    void (*finish)(void *))
1239 {
1240         struct dentry *this_parent;
1241         struct list_head *next;
1242         unsigned seq = 0;
1243         enum d_walk_ret ret;
1244         bool retry = true;
1245
1246 again:
1247         read_seqbegin_or_lock(&rename_lock, &seq);
1248         this_parent = parent;
1249         spin_lock(&this_parent->d_lock);
1250
1251         ret = enter(data, this_parent);
1252         switch (ret) {
1253         case D_WALK_CONTINUE:
1254                 break;
1255         case D_WALK_QUIT:
1256         case D_WALK_SKIP:
1257                 goto out_unlock;
1258         case D_WALK_NORETRY:
1259                 retry = false;
1260                 break;
1261         }
1262 repeat:
1263         next = this_parent->d_subdirs.next;
1264 resume:
1265         while (next != &this_parent->d_subdirs) {
1266                 struct list_head *tmp = next;
1267                 struct dentry *dentry = list_entry(tmp, struct dentry, d_child);
1268                 next = tmp->next;
1269
1270                 if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
1271                         continue;
1272
1273                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1274
1275                 ret = enter(data, dentry);
1276                 switch (ret) {
1277                 case D_WALK_CONTINUE:
1278                         break;
1279                 case D_WALK_QUIT:
1280                         spin_unlock(&dentry->d_lock);
1281                         goto out_unlock;
1282                 case D_WALK_NORETRY:
1283                         retry = false;
1284                         break;
1285                 case D_WALK_SKIP:
1286                         spin_unlock(&dentry->d_lock);
1287                         continue;
1288                 }
1289
1290                 if (!list_empty(&dentry->d_subdirs)) {
1291                         spin_unlock(&this_parent->d_lock);
1292                         spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1293                         this_parent = dentry;
1294                         spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1295                         goto repeat;
1296                 }
1297                 spin_unlock(&dentry->d_lock);
1298         }
1299         /*
1300          * All done at this level ... ascend and resume the search.
1301          */
1302         rcu_read_lock();
1303 ascend:
1304         if (this_parent != parent) {
1305                 struct dentry *child = this_parent;
1306                 this_parent = child->d_parent;
1307
1308                 spin_unlock(&child->d_lock);
1309                 spin_lock(&this_parent->d_lock);
1310
1311                 /* might go back up the wrong parent if we have had a rename. */
1312                 if (need_seqretry(&rename_lock, seq))
1313                         goto rename_retry;
1314                 /* go into the first sibling still alive */
1315                 do {
1316                         next = child->d_child.next;
1317                         if (next == &this_parent->d_subdirs)
1318                                 goto ascend;
1319                         child = list_entry(next, struct dentry, d_child);
1320                 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED));
1321                 rcu_read_unlock();
1322                 goto resume;
1323         }
1324         if (need_seqretry(&rename_lock, seq))
1325                 goto rename_retry;
1326         rcu_read_unlock();
1327         if (finish)
1328                 finish(data);
1329
1330 out_unlock:
1331         spin_unlock(&this_parent->d_lock);
1332         done_seqretry(&rename_lock, seq);
1333         return;
1334
1335 rename_retry:
1336         spin_unlock(&this_parent->d_lock);
1337         rcu_read_unlock();
1338         BUG_ON(seq & 1);
1339         if (!retry)
1340                 return;
1341         seq = 1;
1342         goto again;
1343 }
1344
1345 struct check_mount {
1346         struct vfsmount *mnt;
1347         unsigned int mounted;
1348 };
1349
1350 static enum d_walk_ret path_check_mount(void *data, struct dentry *dentry)
1351 {
1352         struct check_mount *info = data;
1353         struct path path = { .mnt = info->mnt, .dentry = dentry };
1354
1355         if (likely(!d_mountpoint(dentry)))
1356                 return D_WALK_CONTINUE;
1357         if (__path_is_mountpoint(&path)) {
1358                 info->mounted = 1;
1359                 return D_WALK_QUIT;
1360         }
1361         return D_WALK_CONTINUE;
1362 }
1363
1364 /**
1365  * path_has_submounts - check for mounts over a dentry in the
1366  *                      current namespace.
1367  * @parent: path to check.
1368  *
1369  * Return true if the parent or its subdirectories contain
1370  * a mount point in the current namespace.
1371  */
1372 int path_has_submounts(const struct path *parent)
1373 {
1374         struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };
1375
1376         read_seqlock_excl(&mount_lock);
1377         d_walk(parent->dentry, &data, path_check_mount, NULL);
1378         read_sequnlock_excl(&mount_lock);
1379
1380         return data.mounted;
1381 }
1382 EXPORT_SYMBOL(path_has_submounts);
1383
1384 /*
1385  * Called by mount code to set a mountpoint and check if the mountpoint is
1386  * reachable (e.g. NFS can unhash a directory dentry and then the complete
1387  * subtree can become unreachable).
1388  *
1389  * Only one of d_invalidate() and d_set_mounted() must succeed.  For
1390  * this reason take rename_lock and d_lock on dentry and ancestors.
1391  */
1392 int d_set_mounted(struct dentry *dentry)
1393 {
1394         struct dentry *p;
1395         int ret = -ENOENT;
1396         write_seqlock(&rename_lock);
1397         for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1398                 /* Need exclusion wrt. d_invalidate() */
1399                 spin_lock(&p->d_lock);
1400                 if (unlikely(d_unhashed(p))) {
1401                         spin_unlock(&p->d_lock);
1402                         goto out;
1403                 }
1404                 spin_unlock(&p->d_lock);
1405         }
1406         spin_lock(&dentry->d_lock);
1407         if (!d_unlinked(dentry)) {
1408                 ret = -EBUSY;
1409                 if (!d_mountpoint(dentry)) {
1410                         dentry->d_flags |= DCACHE_MOUNTED;
1411                         ret = 0;
1412                 }
1413         }
1414         spin_unlock(&dentry->d_lock);
1415 out:
1416         write_sequnlock(&rename_lock);
1417         return ret;
1418 }
1419
1420 /*
1421  * Search the dentry child list of the specified parent,
1422  * and move any unused dentries to the end of the unused
1423  * list for prune_dcache(). We descend to the next level
1424  * whenever the d_subdirs list is non-empty and continue
1425  * searching.
1426  *
1427  * It returns zero iff there are no unused children,
1428  * otherwise  it returns the number of children moved to
1429  * the end of the unused list. This may not be the total
1430  * number of unused children, because select_parent can
1431  * drop the lock and return early due to latency
1432  * constraints.
1433  */
1434
1435 struct select_data {
1436         struct dentry *start;
1437         struct list_head dispose;
1438         int found;
1439 };
1440
1441 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1442 {
1443         struct select_data *data = _data;
1444         enum d_walk_ret ret = D_WALK_CONTINUE;
1445
1446         if (data->start == dentry)
1447                 goto out;
1448
1449         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1450                 data->found++;
1451         } else {
1452                 if (dentry->d_flags & DCACHE_LRU_LIST)
1453                         d_lru_del(dentry);
1454                 if (!dentry->d_lockref.count) {
1455                         d_shrink_add(dentry, &data->dispose);
1456                         data->found++;
1457                 }
1458         }
1459         /*
1460          * We can return to the caller if we have found some (this
1461          * ensures forward progress). We'll be coming back to find
1462          * the rest.
1463          */
1464         if (!list_empty(&data->dispose))
1465                 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1466 out:
1467         return ret;
1468 }
1469
1470 /**
1471  * shrink_dcache_parent - prune dcache
1472  * @parent: parent of entries to prune
1473  *
1474  * Prune the dcache to remove unused children of the parent dentry.
1475  */
1476 void shrink_dcache_parent(struct dentry *parent)
1477 {
1478         for (;;) {
1479                 struct select_data data;
1480
1481                 INIT_LIST_HEAD(&data.dispose);
1482                 data.start = parent;
1483                 data.found = 0;
1484
1485                 d_walk(parent, &data, select_collect, NULL);
1486                 if (!data.found)
1487                         break;
1488
1489                 shrink_dentry_list(&data.dispose);
1490                 cond_resched();
1491         }
1492 }
1493 EXPORT_SYMBOL(shrink_dcache_parent);
1494
1495 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1496 {
1497         /* it has busy descendents; complain about those instead */
1498         if (!list_empty(&dentry->d_subdirs))
1499                 return D_WALK_CONTINUE;
1500
1501         /* root with refcount 1 is fine */
1502         if (dentry == _data && dentry->d_lockref.count == 1)
1503                 return D_WALK_CONTINUE;
1504
1505         printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} "
1506                         " still in use (%d) [unmount of %s %s]\n",
1507                        dentry,
1508                        dentry->d_inode ?
1509                        dentry->d_inode->i_ino : 0UL,
1510                        dentry,
1511                        dentry->d_lockref.count,
1512                        dentry->d_sb->s_type->name,
1513                        dentry->d_sb->s_id);
1514         WARN_ON(1);
1515         return D_WALK_CONTINUE;
1516 }
1517
1518 static void do_one_tree(struct dentry *dentry)
1519 {
1520         shrink_dcache_parent(dentry);
1521         d_walk(dentry, dentry, umount_check, NULL);
1522         d_drop(dentry);
1523         dput(dentry);
1524 }
1525
1526 /*
1527  * destroy the dentries attached to a superblock on unmounting
1528  */
1529 void shrink_dcache_for_umount(struct super_block *sb)
1530 {
1531         struct dentry *dentry;
1532
1533         WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
1534
1535         dentry = sb->s_root;
1536         sb->s_root = NULL;
1537         do_one_tree(dentry);
1538
1539         while (!hlist_bl_empty(&sb->s_roots)) {
1540                 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_roots), struct dentry, d_hash));
1541                 do_one_tree(dentry);
1542         }
1543 }
1544
1545 struct detach_data {
1546         struct select_data select;
1547         struct dentry *mountpoint;
1548 };
1549 static enum d_walk_ret detach_and_collect(void *_data, struct dentry *dentry)
1550 {
1551         struct detach_data *data = _data;
1552
1553         if (d_mountpoint(dentry)) {
1554                 __dget_dlock(dentry);
1555                 data->mountpoint = dentry;
1556                 return D_WALK_QUIT;
1557         }
1558
1559         return select_collect(&data->select, dentry);
1560 }
1561
1562 static void check_and_drop(void *_data)
1563 {
1564         struct detach_data *data = _data;
1565
1566         if (!data->mountpoint && list_empty(&data->select.dispose))
1567                 __d_drop(data->select.start);
1568 }
1569
1570 /**
1571  * d_invalidate - detach submounts, prune dcache, and drop
1572  * @dentry: dentry to invalidate (aka detach, prune and drop)
1573  *
1574  * no dcache lock.
1575  *
1576  * The final d_drop is done as an atomic operation relative to
1577  * rename_lock ensuring there are no races with d_set_mounted.  This
1578  * ensures there are no unhashed dentries on the path to a mountpoint.
1579  */
1580 void d_invalidate(struct dentry *dentry)
1581 {
1582         /*
1583          * If it's already been dropped, return OK.
1584          */
1585         spin_lock(&dentry->d_lock);
1586         if (d_unhashed(dentry)) {
1587                 spin_unlock(&dentry->d_lock);
1588                 return;
1589         }
1590         spin_unlock(&dentry->d_lock);
1591
1592         /* Negative dentries can be dropped without further checks */
1593         if (!dentry->d_inode) {
1594                 d_drop(dentry);
1595                 return;
1596         }
1597
1598         for (;;) {
1599                 struct detach_data data;
1600
1601                 data.mountpoint = NULL;
1602                 INIT_LIST_HEAD(&data.select.dispose);
1603                 data.select.start = dentry;
1604                 data.select.found = 0;
1605
1606                 d_walk(dentry, &data, detach_and_collect, check_and_drop);
1607
1608                 if (!list_empty(&data.select.dispose))
1609                         shrink_dentry_list(&data.select.dispose);
1610                 else if (!data.mountpoint)
1611                         return;
1612
1613                 if (data.mountpoint) {
1614                         detach_mounts(data.mountpoint);
1615                         dput(data.mountpoint);
1616                 }
1617                 cond_resched();
1618         }
1619 }
1620 EXPORT_SYMBOL(d_invalidate);
1621
1622 /**
1623  * __d_alloc    -       allocate a dcache entry
1624  * @sb: filesystem it will belong to
1625  * @name: qstr of the name
1626  *
1627  * Allocates a dentry. It returns %NULL if there is insufficient memory
1628  * available. On a success the dentry is returned. The name passed in is
1629  * copied and the copy passed in may be reused after this call.
1630  */
1631  
1632 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1633 {
1634         struct external_name *ext = NULL;
1635         struct dentry *dentry;
1636         char *dname;
1637         int err;
1638
1639         dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1640         if (!dentry)
1641                 return NULL;
1642
1643         /*
1644          * We guarantee that the inline name is always NUL-terminated.
1645          * This way the memcpy() done by the name switching in rename
1646          * will still always have a NUL at the end, even if we might
1647          * be overwriting an internal NUL character
1648          */
1649         dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1650         if (unlikely(!name)) {
1651                 name = &slash_name;
1652                 dname = dentry->d_iname;
1653         } else if (name->len > DNAME_INLINE_LEN-1) {
1654                 size_t size = offsetof(struct external_name, name[1]);
1655
1656                 ext = kmalloc(size + name->len, GFP_KERNEL_ACCOUNT);
1657                 if (!ext) {
1658                         kmem_cache_free(dentry_cache, dentry); 
1659                         return NULL;
1660                 }
1661                 atomic_set(&ext->u.count, 1);
1662                 dname = ext->name;
1663         } else  {
1664                 dname = dentry->d_iname;
1665         }       
1666
1667         dentry->d_name.len = name->len;
1668         dentry->d_name.hash = name->hash;
1669         memcpy(dname, name->name, name->len);
1670         dname[name->len] = 0;
1671
1672         /* Make sure we always see the terminating NUL character */
1673         smp_store_release(&dentry->d_name.name, dname); /* ^^^ */
1674
1675         dentry->d_lockref.count = 1;
1676         dentry->d_flags = 0;
1677         spin_lock_init(&dentry->d_lock);
1678         seqcount_init(&dentry->d_seq);
1679         dentry->d_inode = NULL;
1680         dentry->d_parent = dentry;
1681         dentry->d_sb = sb;
1682         dentry->d_op = NULL;
1683         dentry->d_fsdata = NULL;
1684         INIT_HLIST_BL_NODE(&dentry->d_hash);
1685         INIT_LIST_HEAD(&dentry->d_lru);
1686         INIT_LIST_HEAD(&dentry->d_subdirs);
1687         INIT_HLIST_NODE(&dentry->d_u.d_alias);
1688         INIT_LIST_HEAD(&dentry->d_child);
1689         d_set_d_op(dentry, dentry->d_sb->s_d_op);
1690
1691         if (dentry->d_op && dentry->d_op->d_init) {
1692                 err = dentry->d_op->d_init(dentry);
1693                 if (err) {
1694                         if (dname_external(dentry))
1695                                 kfree(external_name(dentry));
1696                         kmem_cache_free(dentry_cache, dentry);
1697                         return NULL;
1698                 }
1699         }
1700
1701         if (unlikely(ext)) {
1702                 pg_data_t *pgdat = page_pgdat(virt_to_page(ext));
1703                 mod_node_page_state(pgdat, NR_INDIRECTLY_RECLAIMABLE_BYTES,
1704                                     ksize(ext));
1705         }
1706
1707         this_cpu_inc(nr_dentry);
1708
1709         return dentry;
1710 }
1711
1712 /**
1713  * d_alloc      -       allocate a dcache entry
1714  * @parent: parent of entry to allocate
1715  * @name: qstr of the name
1716  *
1717  * Allocates a dentry. It returns %NULL if there is insufficient memory
1718  * available. On a success the dentry is returned. The name passed in is
1719  * copied and the copy passed in may be reused after this call.
1720  */
1721 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1722 {
1723         struct dentry *dentry = __d_alloc(parent->d_sb, name);
1724         if (!dentry)
1725                 return NULL;
1726         dentry->d_flags |= DCACHE_RCUACCESS;
1727         spin_lock(&parent->d_lock);
1728         /*
1729          * don't need child lock because it is not subject
1730          * to concurrency here
1731          */
1732         __dget_dlock(parent);
1733         dentry->d_parent = parent;
1734         list_add(&dentry->d_child, &parent->d_subdirs);
1735         spin_unlock(&parent->d_lock);
1736
1737         return dentry;
1738 }
1739 EXPORT_SYMBOL(d_alloc);
1740
1741 struct dentry *d_alloc_anon(struct super_block *sb)
1742 {
1743         return __d_alloc(sb, NULL);
1744 }
1745 EXPORT_SYMBOL(d_alloc_anon);
1746
1747 struct dentry *d_alloc_cursor(struct dentry * parent)
1748 {
1749         struct dentry *dentry = d_alloc_anon(parent->d_sb);
1750         if (dentry) {
1751                 dentry->d_flags |= DCACHE_RCUACCESS | DCACHE_DENTRY_CURSOR;
1752                 dentry->d_parent = dget(parent);
1753         }
1754         return dentry;
1755 }
1756
1757 /**
1758  * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1759  * @sb: the superblock
1760  * @name: qstr of the name
1761  *
1762  * For a filesystem that just pins its dentries in memory and never
1763  * performs lookups at all, return an unhashed IS_ROOT dentry.
1764  */
1765 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1766 {
1767         return __d_alloc(sb, name);
1768 }
1769 EXPORT_SYMBOL(d_alloc_pseudo);
1770
1771 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1772 {
1773         struct qstr q;
1774
1775         q.name = name;
1776         q.hash_len = hashlen_string(parent, name);
1777         return d_alloc(parent, &q);
1778 }
1779 EXPORT_SYMBOL(d_alloc_name);
1780
1781 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1782 {
1783         WARN_ON_ONCE(dentry->d_op);
1784         WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH  |
1785                                 DCACHE_OP_COMPARE       |
1786                                 DCACHE_OP_REVALIDATE    |
1787                                 DCACHE_OP_WEAK_REVALIDATE       |
1788                                 DCACHE_OP_DELETE        |
1789                                 DCACHE_OP_REAL));
1790         dentry->d_op = op;
1791         if (!op)
1792                 return;
1793         if (op->d_hash)
1794                 dentry->d_flags |= DCACHE_OP_HASH;
1795         if (op->d_compare)
1796                 dentry->d_flags |= DCACHE_OP_COMPARE;
1797         if (op->d_revalidate)
1798                 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1799         if (op->d_weak_revalidate)
1800                 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1801         if (op->d_delete)
1802                 dentry->d_flags |= DCACHE_OP_DELETE;
1803         if (op->d_prune)
1804                 dentry->d_flags |= DCACHE_OP_PRUNE;
1805         if (op->d_real)
1806                 dentry->d_flags |= DCACHE_OP_REAL;
1807
1808 }
1809 EXPORT_SYMBOL(d_set_d_op);
1810
1811
1812 /*
1813  * d_set_fallthru - Mark a dentry as falling through to a lower layer
1814  * @dentry - The dentry to mark
1815  *
1816  * Mark a dentry as falling through to the lower layer (as set with
1817  * d_pin_lower()).  This flag may be recorded on the medium.
1818  */
1819 void d_set_fallthru(struct dentry *dentry)
1820 {
1821         spin_lock(&dentry->d_lock);
1822         dentry->d_flags |= DCACHE_FALLTHRU;
1823         spin_unlock(&dentry->d_lock);
1824 }
1825 EXPORT_SYMBOL(d_set_fallthru);
1826
1827 static unsigned d_flags_for_inode(struct inode *inode)
1828 {
1829         unsigned add_flags = DCACHE_REGULAR_TYPE;
1830
1831         if (!inode)
1832                 return DCACHE_MISS_TYPE;
1833
1834         if (S_ISDIR(inode->i_mode)) {
1835                 add_flags = DCACHE_DIRECTORY_TYPE;
1836                 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1837                         if (unlikely(!inode->i_op->lookup))
1838                                 add_flags = DCACHE_AUTODIR_TYPE;
1839                         else
1840                                 inode->i_opflags |= IOP_LOOKUP;
1841                 }
1842                 goto type_determined;
1843         }
1844
1845         if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1846                 if (unlikely(inode->i_op->get_link)) {
1847                         add_flags = DCACHE_SYMLINK_TYPE;
1848                         goto type_determined;
1849                 }
1850                 inode->i_opflags |= IOP_NOFOLLOW;
1851         }
1852
1853         if (unlikely(!S_ISREG(inode->i_mode)))
1854                 add_flags = DCACHE_SPECIAL_TYPE;
1855
1856 type_determined:
1857         if (unlikely(IS_AUTOMOUNT(inode)))
1858                 add_flags |= DCACHE_NEED_AUTOMOUNT;
1859         return add_flags;
1860 }
1861
1862 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1863 {
1864         unsigned add_flags = d_flags_for_inode(inode);
1865         WARN_ON(d_in_lookup(dentry));
1866
1867         spin_lock(&dentry->d_lock);
1868         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1869         raw_write_seqcount_begin(&dentry->d_seq);
1870         __d_set_inode_and_type(dentry, inode, add_flags);
1871         raw_write_seqcount_end(&dentry->d_seq);
1872         fsnotify_update_flags(dentry);
1873         spin_unlock(&dentry->d_lock);
1874 }
1875
1876 /**
1877  * d_instantiate - fill in inode information for a dentry
1878  * @entry: dentry to complete
1879  * @inode: inode to attach to this dentry
1880  *
1881  * Fill in inode information in the entry.
1882  *
1883  * This turns negative dentries into productive full members
1884  * of society.
1885  *
1886  * NOTE! This assumes that the inode count has been incremented
1887  * (or otherwise set) by the caller to indicate that it is now
1888  * in use by the dcache.
1889  */
1890  
1891 void d_instantiate(struct dentry *entry, struct inode * inode)
1892 {
1893         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1894         if (inode) {
1895                 security_d_instantiate(entry, inode);
1896                 spin_lock(&inode->i_lock);
1897                 __d_instantiate(entry, inode);
1898                 spin_unlock(&inode->i_lock);
1899         }
1900 }
1901 EXPORT_SYMBOL(d_instantiate);
1902
1903 /**
1904  * d_instantiate_no_diralias - instantiate a non-aliased dentry
1905  * @entry: dentry to complete
1906  * @inode: inode to attach to this dentry
1907  *
1908  * Fill in inode information in the entry.  If a directory alias is found, then
1909  * return an error (and drop inode).  Together with d_materialise_unique() this
1910  * guarantees that a directory inode may never have more than one alias.
1911  */
1912 int d_instantiate_no_diralias(struct dentry *entry, struct inode *inode)
1913 {
1914         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1915
1916         security_d_instantiate(entry, inode);
1917         spin_lock(&inode->i_lock);
1918         if (S_ISDIR(inode->i_mode) && !hlist_empty(&inode->i_dentry)) {
1919                 spin_unlock(&inode->i_lock);
1920                 iput(inode);
1921                 return -EBUSY;
1922         }
1923         __d_instantiate(entry, inode);
1924         spin_unlock(&inode->i_lock);
1925
1926         return 0;
1927 }
1928 EXPORT_SYMBOL(d_instantiate_no_diralias);
1929
1930 struct dentry *d_make_root(struct inode *root_inode)
1931 {
1932         struct dentry *res = NULL;
1933
1934         if (root_inode) {
1935                 res = d_alloc_anon(root_inode->i_sb);
1936                 if (res)
1937                         d_instantiate(res, root_inode);
1938                 else
1939                         iput(root_inode);
1940         }
1941         return res;
1942 }
1943 EXPORT_SYMBOL(d_make_root);
1944
1945 static struct dentry * __d_find_any_alias(struct inode *inode)
1946 {
1947         struct dentry *alias;
1948
1949         if (hlist_empty(&inode->i_dentry))
1950                 return NULL;
1951         alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
1952         __dget(alias);
1953         return alias;
1954 }
1955
1956 /**
1957  * d_find_any_alias - find any alias for a given inode
1958  * @inode: inode to find an alias for
1959  *
1960  * If any aliases exist for the given inode, take and return a
1961  * reference for one of them.  If no aliases exist, return %NULL.
1962  */
1963 struct dentry *d_find_any_alias(struct inode *inode)
1964 {
1965         struct dentry *de;
1966
1967         spin_lock(&inode->i_lock);
1968         de = __d_find_any_alias(inode);
1969         spin_unlock(&inode->i_lock);
1970         return de;
1971 }
1972 EXPORT_SYMBOL(d_find_any_alias);
1973
1974 static struct dentry *__d_instantiate_anon(struct dentry *dentry,
1975                                            struct inode *inode,
1976                                            bool disconnected)
1977 {
1978         struct dentry *res;
1979         unsigned add_flags;
1980
1981         security_d_instantiate(dentry, inode);
1982         spin_lock(&inode->i_lock);
1983         res = __d_find_any_alias(inode);
1984         if (res) {
1985                 spin_unlock(&inode->i_lock);
1986                 dput(dentry);
1987                 goto out_iput;
1988         }
1989
1990         /* attach a disconnected dentry */
1991         add_flags = d_flags_for_inode(inode);
1992
1993         if (disconnected)
1994                 add_flags |= DCACHE_DISCONNECTED;
1995
1996         spin_lock(&dentry->d_lock);
1997         __d_set_inode_and_type(dentry, inode, add_flags);
1998         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1999         if (!disconnected) {
2000                 hlist_bl_lock(&dentry->d_sb->s_roots);
2001                 hlist_bl_add_head(&dentry->d_hash, &dentry->d_sb->s_roots);
2002                 hlist_bl_unlock(&dentry->d_sb->s_roots);
2003         }
2004         spin_unlock(&dentry->d_lock);
2005         spin_unlock(&inode->i_lock);
2006
2007         return dentry;
2008
2009  out_iput:
2010         iput(inode);
2011         return res;
2012 }
2013
2014 struct dentry *d_instantiate_anon(struct dentry *dentry, struct inode *inode)
2015 {
2016         return __d_instantiate_anon(dentry, inode, true);
2017 }
2018 EXPORT_SYMBOL(d_instantiate_anon);
2019
2020 static struct dentry *__d_obtain_alias(struct inode *inode, bool disconnected)
2021 {
2022         struct dentry *tmp;
2023         struct dentry *res;
2024
2025         if (!inode)
2026                 return ERR_PTR(-ESTALE);
2027         if (IS_ERR(inode))
2028                 return ERR_CAST(inode);
2029
2030         res = d_find_any_alias(inode);
2031         if (res)
2032                 goto out_iput;
2033
2034         tmp = d_alloc_anon(inode->i_sb);
2035         if (!tmp) {
2036                 res = ERR_PTR(-ENOMEM);
2037                 goto out_iput;
2038         }
2039
2040         return __d_instantiate_anon(tmp, inode, disconnected);
2041
2042 out_iput:
2043         iput(inode);
2044         return res;
2045 }
2046
2047 /**
2048  * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
2049  * @inode: inode to allocate the dentry for
2050  *
2051  * Obtain a dentry for an inode resulting from NFS filehandle conversion or
2052  * similar open by handle operations.  The returned dentry may be anonymous,
2053  * or may have a full name (if the inode was already in the cache).
2054  *
2055  * When called on a directory inode, we must ensure that the inode only ever
2056  * has one dentry.  If a dentry is found, that is returned instead of
2057  * allocating a new one.
2058  *
2059  * On successful return, the reference to the inode has been transferred
2060  * to the dentry.  In case of an error the reference on the inode is released.
2061  * To make it easier to use in export operations a %NULL or IS_ERR inode may
2062  * be passed in and the error will be propagated to the return value,
2063  * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
2064  */
2065 struct dentry *d_obtain_alias(struct inode *inode)
2066 {
2067         return __d_obtain_alias(inode, true);
2068 }
2069 EXPORT_SYMBOL(d_obtain_alias);
2070
2071 /**
2072  * d_obtain_root - find or allocate a dentry for a given inode
2073  * @inode: inode to allocate the dentry for
2074  *
2075  * Obtain an IS_ROOT dentry for the root of a filesystem.
2076  *
2077  * We must ensure that directory inodes only ever have one dentry.  If a
2078  * dentry is found, that is returned instead of allocating a new one.
2079  *
2080  * On successful return, the reference to the inode has been transferred
2081  * to the dentry.  In case of an error the reference on the inode is
2082  * released.  A %NULL or IS_ERR inode may be passed in and will be the
2083  * error will be propagate to the return value, with a %NULL @inode
2084  * replaced by ERR_PTR(-ESTALE).
2085  */
2086 struct dentry *d_obtain_root(struct inode *inode)
2087 {
2088         return __d_obtain_alias(inode, false);
2089 }
2090 EXPORT_SYMBOL(d_obtain_root);
2091
2092 /**
2093  * d_add_ci - lookup or allocate new dentry with case-exact name
2094  * @inode:  the inode case-insensitive lookup has found
2095  * @dentry: the negative dentry that was passed to the parent's lookup func
2096  * @name:   the case-exact name to be associated with the returned dentry
2097  *
2098  * This is to avoid filling the dcache with case-insensitive names to the
2099  * same inode, only the actual correct case is stored in the dcache for
2100  * case-insensitive filesystems.
2101  *
2102  * For a case-insensitive lookup match and if the the case-exact dentry
2103  * already exists in in the dcache, use it and return it.
2104  *
2105  * If no entry exists with the exact case name, allocate new dentry with
2106  * the exact case, and return the spliced entry.
2107  */
2108 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2109                         struct qstr *name)
2110 {
2111         struct dentry *found, *res;
2112
2113         /*
2114          * First check if a dentry matching the name already exists,
2115          * if not go ahead and create it now.
2116          */
2117         found = d_hash_and_lookup(dentry->d_parent, name);
2118         if (found) {
2119                 iput(inode);
2120                 return found;
2121         }
2122         if (d_in_lookup(dentry)) {
2123                 found = d_alloc_parallel(dentry->d_parent, name,
2124                                         dentry->d_wait);
2125                 if (IS_ERR(found) || !d_in_lookup(found)) {
2126                         iput(inode);
2127                         return found;
2128                 }
2129         } else {
2130                 found = d_alloc(dentry->d_parent, name);
2131                 if (!found) {
2132                         iput(inode);
2133                         return ERR_PTR(-ENOMEM);
2134                 } 
2135         }
2136         res = d_splice_alias(inode, found);
2137         if (res) {
2138                 dput(found);
2139                 return res;
2140         }
2141         return found;
2142 }
2143 EXPORT_SYMBOL(d_add_ci);
2144
2145
2146 static inline bool d_same_name(const struct dentry *dentry,
2147                                 const struct dentry *parent,
2148                                 const struct qstr *name)
2149 {
2150         if (likely(!(parent->d_flags & DCACHE_OP_COMPARE))) {
2151                 if (dentry->d_name.len != name->len)
2152                         return false;
2153                 return dentry_cmp(dentry, name->name, name->len) == 0;
2154         }
2155         return parent->d_op->d_compare(dentry,
2156                                        dentry->d_name.len, dentry->d_name.name,
2157                                        name) == 0;
2158 }
2159
2160 /**
2161  * __d_lookup_rcu - search for a dentry (racy, store-free)
2162  * @parent: parent dentry
2163  * @name: qstr of name we wish to find
2164  * @seqp: returns d_seq value at the point where the dentry was found
2165  * Returns: dentry, or NULL
2166  *
2167  * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2168  * resolution (store-free path walking) design described in
2169  * Documentation/filesystems/path-lookup.txt.
2170  *
2171  * This is not to be used outside core vfs.
2172  *
2173  * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2174  * held, and rcu_read_lock held. The returned dentry must not be stored into
2175  * without taking d_lock and checking d_seq sequence count against @seq
2176  * returned here.
2177  *
2178  * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2179  * function.
2180  *
2181  * Alternatively, __d_lookup_rcu may be called again to look up the child of
2182  * the returned dentry, so long as its parent's seqlock is checked after the
2183  * child is looked up. Thus, an interlocking stepping of sequence lock checks
2184  * is formed, giving integrity down the path walk.
2185  *
2186  * NOTE! The caller *has* to check the resulting dentry against the sequence
2187  * number we've returned before using any of the resulting dentry state!
2188  */
2189 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2190                                 const struct qstr *name,
2191                                 unsigned *seqp)
2192 {
2193         u64 hashlen = name->hash_len;
2194         const unsigned char *str = name->name;
2195         struct hlist_bl_head *b = d_hash(hashlen_hash(hashlen));
2196         struct hlist_bl_node *node;
2197         struct dentry *dentry;
2198
2199         /*
2200          * Note: There is significant duplication with __d_lookup_rcu which is
2201          * required to prevent single threaded performance regressions
2202          * especially on architectures where smp_rmb (in seqcounts) are costly.
2203          * Keep the two functions in sync.
2204          */
2205
2206         /*
2207          * The hash list is protected using RCU.
2208          *
2209          * Carefully use d_seq when comparing a candidate dentry, to avoid
2210          * races with d_move().
2211          *
2212          * It is possible that concurrent renames can mess up our list
2213          * walk here and result in missing our dentry, resulting in the
2214          * false-negative result. d_lookup() protects against concurrent
2215          * renames using rename_lock seqlock.
2216          *
2217          * See Documentation/filesystems/path-lookup.txt for more details.
2218          */
2219         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2220                 unsigned seq;
2221
2222 seqretry:
2223                 /*
2224                  * The dentry sequence count protects us from concurrent
2225                  * renames, and thus protects parent and name fields.
2226                  *
2227                  * The caller must perform a seqcount check in order
2228                  * to do anything useful with the returned dentry.
2229                  *
2230                  * NOTE! We do a "raw" seqcount_begin here. That means that
2231                  * we don't wait for the sequence count to stabilize if it
2232                  * is in the middle of a sequence change. If we do the slow
2233                  * dentry compare, we will do seqretries until it is stable,
2234                  * and if we end up with a successful lookup, we actually
2235                  * want to exit RCU lookup anyway.
2236                  *
2237                  * Note that raw_seqcount_begin still *does* smp_rmb(), so
2238                  * we are still guaranteed NUL-termination of ->d_name.name.
2239                  */
2240                 seq = raw_seqcount_begin(&dentry->d_seq);
2241                 if (dentry->d_parent != parent)
2242                         continue;
2243                 if (d_unhashed(dentry))
2244                         continue;
2245
2246                 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) {
2247                         int tlen;
2248                         const char *tname;
2249                         if (dentry->d_name.hash != hashlen_hash(hashlen))
2250                                 continue;
2251                         tlen = dentry->d_name.len;
2252                         tname = dentry->d_name.name;
2253                         /* we want a consistent (name,len) pair */
2254                         if (read_seqcount_retry(&dentry->d_seq, seq)) {
2255                                 cpu_relax();
2256                                 goto seqretry;
2257                         }
2258                         if (parent->d_op->d_compare(dentry,
2259                                                     tlen, tname, name) != 0)
2260                                 continue;
2261                 } else {
2262                         if (dentry->d_name.hash_len != hashlen)
2263                                 continue;
2264                         if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
2265                                 continue;
2266                 }
2267                 *seqp = seq;
2268                 return dentry;
2269         }
2270         return NULL;
2271 }
2272
2273 /**
2274  * d_lookup - search for a dentry
2275  * @parent: parent dentry
2276  * @name: qstr of name we wish to find
2277  * Returns: dentry, or NULL
2278  *
2279  * d_lookup searches the children of the parent dentry for the name in
2280  * question. If the dentry is found its reference count is incremented and the
2281  * dentry is returned. The caller must use dput to free the entry when it has
2282  * finished using it. %NULL is returned if the dentry does not exist.
2283  */
2284 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2285 {
2286         struct dentry *dentry;
2287         unsigned seq;
2288
2289         do {
2290                 seq = read_seqbegin(&rename_lock);
2291                 dentry = __d_lookup(parent, name);
2292                 if (dentry)
2293                         break;
2294         } while (read_seqretry(&rename_lock, seq));
2295         return dentry;
2296 }
2297 EXPORT_SYMBOL(d_lookup);
2298
2299 /**
2300  * __d_lookup - search for a dentry (racy)
2301  * @parent: parent dentry
2302  * @name: qstr of name we wish to find
2303  * Returns: dentry, or NULL
2304  *
2305  * __d_lookup is like d_lookup, however it may (rarely) return a
2306  * false-negative result due to unrelated rename activity.
2307  *
2308  * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2309  * however it must be used carefully, eg. with a following d_lookup in
2310  * the case of failure.
2311  *
2312  * __d_lookup callers must be commented.
2313  */
2314 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2315 {
2316         unsigned int hash = name->hash;
2317         struct hlist_bl_head *b = d_hash(hash);
2318         struct hlist_bl_node *node;
2319         struct dentry *found = NULL;
2320         struct dentry *dentry;
2321
2322         /*
2323          * Note: There is significant duplication with __d_lookup_rcu which is
2324          * required to prevent single threaded performance regressions
2325          * especially on architectures where smp_rmb (in seqcounts) are costly.
2326          * Keep the two functions in sync.
2327          */
2328
2329         /*
2330          * The hash list is protected using RCU.
2331          *
2332          * Take d_lock when comparing a candidate dentry, to avoid races
2333          * with d_move().
2334          *
2335          * It is possible that concurrent renames can mess up our list
2336          * walk here and result in missing our dentry, resulting in the
2337          * false-negative result. d_lookup() protects against concurrent
2338          * renames using rename_lock seqlock.
2339          *
2340          * See Documentation/filesystems/path-lookup.txt for more details.
2341          */
2342         rcu_read_lock();
2343         
2344         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2345
2346                 if (dentry->d_name.hash != hash)
2347                         continue;
2348
2349                 spin_lock(&dentry->d_lock);
2350                 if (dentry->d_parent != parent)
2351                         goto next;
2352                 if (d_unhashed(dentry))
2353                         goto next;
2354
2355                 if (!d_same_name(dentry, parent, name))
2356                         goto next;
2357
2358                 dentry->d_lockref.count++;
2359                 found = dentry;
2360                 spin_unlock(&dentry->d_lock);
2361                 break;
2362 next:
2363                 spin_unlock(&dentry->d_lock);
2364         }
2365         rcu_read_unlock();
2366
2367         return found;
2368 }
2369
2370 /**
2371  * d_hash_and_lookup - hash the qstr then search for a dentry
2372  * @dir: Directory to search in
2373  * @name: qstr of name we wish to find
2374  *
2375  * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2376  */
2377 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2378 {
2379         /*
2380          * Check for a fs-specific hash function. Note that we must
2381          * calculate the standard hash first, as the d_op->d_hash()
2382          * routine may choose to leave the hash value unchanged.
2383          */
2384         name->hash = full_name_hash(dir, name->name, name->len);
2385         if (dir->d_flags & DCACHE_OP_HASH) {
2386                 int err = dir->d_op->d_hash(dir, name);
2387                 if (unlikely(err < 0))
2388                         return ERR_PTR(err);
2389         }
2390         return d_lookup(dir, name);
2391 }
2392 EXPORT_SYMBOL(d_hash_and_lookup);
2393
2394 /*
2395  * When a file is deleted, we have two options:
2396  * - turn this dentry into a negative dentry
2397  * - unhash this dentry and free it.
2398  *
2399  * Usually, we want to just turn this into
2400  * a negative dentry, but if anybody else is
2401  * currently using the dentry or the inode
2402  * we can't do that and we fall back on removing
2403  * it from the hash queues and waiting for
2404  * it to be deleted later when it has no users
2405  */
2406  
2407 /**
2408  * d_delete - delete a dentry
2409  * @dentry: The dentry to delete
2410  *
2411  * Turn the dentry into a negative dentry if possible, otherwise
2412  * remove it from the hash queues so it can be deleted later
2413  */
2414  
2415 void d_delete(struct dentry * dentry)
2416 {
2417         struct inode *inode = dentry->d_inode;
2418         int isdir = d_is_dir(dentry);
2419
2420         spin_lock(&inode->i_lock);
2421         spin_lock(&dentry->d_lock);
2422         /*
2423          * Are we the only user?
2424          */
2425         if (dentry->d_lockref.count == 1) {
2426                 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2427                 dentry_unlink_inode(dentry);
2428         } else {
2429                 __d_drop(dentry);
2430                 spin_unlock(&dentry->d_lock);
2431                 spin_unlock(&inode->i_lock);
2432         }
2433         fsnotify_nameremove(dentry, isdir);
2434 }
2435 EXPORT_SYMBOL(d_delete);
2436
2437 static void __d_rehash(struct dentry *entry)
2438 {
2439         struct hlist_bl_head *b = d_hash(entry->d_name.hash);
2440
2441         hlist_bl_lock(b);
2442         hlist_bl_add_head_rcu(&entry->d_hash, b);
2443         hlist_bl_unlock(b);
2444 }
2445
2446 /**
2447  * d_rehash     - add an entry back to the hash
2448  * @entry: dentry to add to the hash
2449  *
2450  * Adds a dentry to the hash according to its name.
2451  */
2452  
2453 void d_rehash(struct dentry * entry)
2454 {
2455         spin_lock(&entry->d_lock);
2456         __d_rehash(entry);
2457         spin_unlock(&entry->d_lock);
2458 }
2459 EXPORT_SYMBOL(d_rehash);
2460
2461 static inline unsigned start_dir_add(struct inode *dir)
2462 {
2463
2464         for (;;) {
2465                 unsigned n = dir->i_dir_seq;
2466                 if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
2467                         return n;
2468                 cpu_relax();
2469         }
2470 }
2471
2472 static inline void end_dir_add(struct inode *dir, unsigned n)
2473 {
2474         smp_store_release(&dir->i_dir_seq, n + 2);
2475 }
2476
2477 static void d_wait_lookup(struct dentry *dentry)
2478 {
2479         if (d_in_lookup(dentry)) {
2480                 DECLARE_WAITQUEUE(wait, current);
2481                 add_wait_queue(dentry->d_wait, &wait);
2482                 do {
2483                         set_current_state(TASK_UNINTERRUPTIBLE);
2484                         spin_unlock(&dentry->d_lock);
2485                         schedule();
2486                         spin_lock(&dentry->d_lock);
2487                 } while (d_in_lookup(dentry));
2488         }
2489 }
2490
2491 struct dentry *d_alloc_parallel(struct dentry *parent,
2492                                 const struct qstr *name,
2493                                 wait_queue_head_t *wq)
2494 {
2495         unsigned int hash = name->hash;
2496         struct hlist_bl_head *b = in_lookup_hash(parent, hash);
2497         struct hlist_bl_node *node;
2498         struct dentry *new = d_alloc(parent, name);
2499         struct dentry *dentry;
2500         unsigned seq, r_seq, d_seq;
2501
2502         if (unlikely(!new))
2503                 return ERR_PTR(-ENOMEM);
2504
2505 retry:
2506         rcu_read_lock();
2507         seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
2508         r_seq = read_seqbegin(&rename_lock);
2509         dentry = __d_lookup_rcu(parent, name, &d_seq);
2510         if (unlikely(dentry)) {
2511                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2512                         rcu_read_unlock();
2513                         goto retry;
2514                 }
2515                 if (read_seqcount_retry(&dentry->d_seq, d_seq)) {
2516                         rcu_read_unlock();
2517                         dput(dentry);
2518                         goto retry;
2519                 }
2520                 rcu_read_unlock();
2521                 dput(new);
2522                 return dentry;
2523         }
2524         if (unlikely(read_seqretry(&rename_lock, r_seq))) {
2525                 rcu_read_unlock();
2526                 goto retry;
2527         }
2528
2529         if (unlikely(seq & 1)) {
2530                 rcu_read_unlock();
2531                 goto retry;
2532         }
2533
2534         hlist_bl_lock(b);
2535         if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {
2536                 hlist_bl_unlock(b);
2537                 rcu_read_unlock();
2538                 goto retry;
2539         }
2540         /*
2541          * No changes for the parent since the beginning of d_lookup().
2542          * Since all removals from the chain happen with hlist_bl_lock(),
2543          * any potential in-lookup matches are going to stay here until
2544          * we unlock the chain.  All fields are stable in everything
2545          * we encounter.
2546          */
2547         hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
2548                 if (dentry->d_name.hash != hash)
2549                         continue;
2550                 if (dentry->d_parent != parent)
2551                         continue;
2552                 if (!d_same_name(dentry, parent, name))
2553                         continue;
2554                 hlist_bl_unlock(b);
2555                 /* now we can try to grab a reference */
2556                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2557                         rcu_read_unlock();
2558                         goto retry;
2559                 }
2560
2561                 rcu_read_unlock();
2562                 /*
2563                  * somebody is likely to be still doing lookup for it;
2564                  * wait for them to finish
2565                  */
2566                 spin_lock(&dentry->d_lock);
2567                 d_wait_lookup(dentry);
2568                 /*
2569                  * it's not in-lookup anymore; in principle we should repeat
2570                  * everything from dcache lookup, but it's likely to be what
2571                  * d_lookup() would've found anyway.  If it is, just return it;
2572                  * otherwise we really have to repeat the whole thing.
2573                  */
2574                 if (unlikely(dentry->d_name.hash != hash))
2575                         goto mismatch;
2576                 if (unlikely(dentry->d_parent != parent))
2577                         goto mismatch;
2578                 if (unlikely(d_unhashed(dentry)))
2579                         goto mismatch;
2580                 if (unlikely(!d_same_name(dentry, parent, name)))
2581                         goto mismatch;
2582                 /* OK, it *is* a hashed match; return it */
2583                 spin_unlock(&dentry->d_lock);
2584                 dput(new);
2585                 return dentry;
2586         }
2587         rcu_read_unlock();
2588         /* we can't take ->d_lock here; it's OK, though. */
2589         new->d_flags |= DCACHE_PAR_LOOKUP;
2590         new->d_wait = wq;
2591         hlist_bl_add_head_rcu(&new->d_u.d_in_lookup_hash, b);
2592         hlist_bl_unlock(b);
2593         return new;
2594 mismatch:
2595         spin_unlock(&dentry->d_lock);
2596         dput(dentry);
2597         goto retry;
2598 }
2599 EXPORT_SYMBOL(d_alloc_parallel);
2600
2601 void __d_lookup_done(struct dentry *dentry)
2602 {
2603         struct hlist_bl_head *b = in_lookup_hash(dentry->d_parent,
2604                                                  dentry->d_name.hash);
2605         hlist_bl_lock(b);
2606         dentry->d_flags &= ~DCACHE_PAR_LOOKUP;
2607         __hlist_bl_del(&dentry->d_u.d_in_lookup_hash);
2608         wake_up_all(dentry->d_wait);
2609         dentry->d_wait = NULL;
2610         hlist_bl_unlock(b);
2611         INIT_HLIST_NODE(&dentry->d_u.d_alias);
2612         INIT_LIST_HEAD(&dentry->d_lru);
2613 }
2614 EXPORT_SYMBOL(__d_lookup_done);
2615
2616 /* inode->i_lock held if inode is non-NULL */
2617
2618 static inline void __d_add(struct dentry *dentry, struct inode *inode)
2619 {
2620         struct inode *dir = NULL;
2621         unsigned n;
2622         spin_lock(&dentry->d_lock);
2623         if (unlikely(d_in_lookup(dentry))) {
2624                 dir = dentry->d_parent->d_inode;
2625                 n = start_dir_add(dir);
2626                 __d_lookup_done(dentry);
2627         }
2628         if (inode) {
2629                 unsigned add_flags = d_flags_for_inode(inode);
2630                 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2631                 raw_write_seqcount_begin(&dentry->d_seq);
2632                 __d_set_inode_and_type(dentry, inode, add_flags);
2633                 raw_write_seqcount_end(&dentry->d_seq);
2634                 fsnotify_update_flags(dentry);
2635         }
2636         __d_rehash(dentry);
2637         if (dir)
2638                 end_dir_add(dir, n);
2639         spin_unlock(&dentry->d_lock);
2640         if (inode)
2641                 spin_unlock(&inode->i_lock);
2642 }
2643
2644 /**
2645  * d_add - add dentry to hash queues
2646  * @entry: dentry to add
2647  * @inode: The inode to attach to this dentry
2648  *
2649  * This adds the entry to the hash queues and initializes @inode.
2650  * The entry was actually filled in earlier during d_alloc().
2651  */
2652
2653 void d_add(struct dentry *entry, struct inode *inode)
2654 {
2655         if (inode) {
2656                 security_d_instantiate(entry, inode);
2657                 spin_lock(&inode->i_lock);
2658         }
2659         __d_add(entry, inode);
2660 }
2661 EXPORT_SYMBOL(d_add);
2662
2663 /**
2664  * d_exact_alias - find and hash an exact unhashed alias
2665  * @entry: dentry to add
2666  * @inode: The inode to go with this dentry
2667  *
2668  * If an unhashed dentry with the same name/parent and desired
2669  * inode already exists, hash and return it.  Otherwise, return
2670  * NULL.
2671  *
2672  * Parent directory should be locked.
2673  */
2674 struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2675 {
2676         struct dentry *alias;
2677         unsigned int hash = entry->d_name.hash;
2678
2679         spin_lock(&inode->i_lock);
2680         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2681                 /*
2682                  * Don't need alias->d_lock here, because aliases with
2683                  * d_parent == entry->d_parent are not subject to name or
2684                  * parent changes, because the parent inode i_mutex is held.
2685                  */
2686                 if (alias->d_name.hash != hash)
2687                         continue;
2688                 if (alias->d_parent != entry->d_parent)
2689                         continue;
2690                 if (!d_same_name(alias, entry->d_parent, &entry->d_name))
2691                         continue;
2692                 spin_lock(&alias->d_lock);
2693                 if (!d_unhashed(alias)) {
2694                         spin_unlock(&alias->d_lock);
2695                         alias = NULL;
2696                 } else {
2697                         __dget_dlock(alias);
2698                         __d_rehash(alias);
2699                         spin_unlock(&alias->d_lock);
2700                 }
2701                 spin_unlock(&inode->i_lock);
2702                 return alias;
2703         }
2704         spin_unlock(&inode->i_lock);
2705         return NULL;
2706 }
2707 EXPORT_SYMBOL(d_exact_alias);
2708
2709 /**
2710  * dentry_update_name_case - update case insensitive dentry with a new name
2711  * @dentry: dentry to be updated
2712  * @name: new name
2713  *
2714  * Update a case insensitive dentry with new case of name.
2715  *
2716  * dentry must have been returned by d_lookup with name @name. Old and new
2717  * name lengths must match (ie. no d_compare which allows mismatched name
2718  * lengths).
2719  *
2720  * Parent inode i_mutex must be held over d_lookup and into this call (to
2721  * keep renames and concurrent inserts, and readdir(2) away).
2722  */
2723 void dentry_update_name_case(struct dentry *dentry, const struct qstr *name)
2724 {
2725         BUG_ON(!inode_is_locked(dentry->d_parent->d_inode));
2726         BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */
2727
2728         spin_lock(&dentry->d_lock);
2729         write_seqcount_begin(&dentry->d_seq);
2730         memcpy((unsigned char *)dentry->d_name.name, name->name, name->len);
2731         write_seqcount_end(&dentry->d_seq);
2732         spin_unlock(&dentry->d_lock);
2733 }
2734 EXPORT_SYMBOL(dentry_update_name_case);
2735
2736 static void swap_names(struct dentry *dentry, struct dentry *target)
2737 {
2738         if (unlikely(dname_external(target))) {
2739                 if (unlikely(dname_external(dentry))) {
2740                         /*
2741                          * Both external: swap the pointers
2742                          */
2743                         swap(target->d_name.name, dentry->d_name.name);
2744                 } else {
2745                         /*
2746                          * dentry:internal, target:external.  Steal target's
2747                          * storage and make target internal.
2748                          */
2749                         memcpy(target->d_iname, dentry->d_name.name,
2750                                         dentry->d_name.len + 1);
2751                         dentry->d_name.name = target->d_name.name;
2752                         target->d_name.name = target->d_iname;
2753                 }
2754         } else {
2755                 if (unlikely(dname_external(dentry))) {
2756                         /*
2757                          * dentry:external, target:internal.  Give dentry's
2758                          * storage to target and make dentry internal
2759                          */
2760                         memcpy(dentry->d_iname, target->d_name.name,
2761                                         target->d_name.len + 1);
2762                         target->d_name.name = dentry->d_name.name;
2763                         dentry->d_name.name = dentry->d_iname;
2764                 } else {
2765                         /*
2766                          * Both are internal.
2767                          */
2768                         unsigned int i;
2769                         BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2770                         for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2771                                 swap(((long *) &dentry->d_iname)[i],
2772                                      ((long *) &target->d_iname)[i]);
2773                         }
2774                 }
2775         }
2776         swap(dentry->d_name.hash_len, target->d_name.hash_len);
2777 }
2778
2779 static void copy_name(struct dentry *dentry, struct dentry *target)
2780 {
2781         struct external_name *old_name = NULL;
2782         if (unlikely(dname_external(dentry)))
2783                 old_name = external_name(dentry);
2784         if (unlikely(dname_external(target))) {
2785                 atomic_inc(&external_name(target)->u.count);
2786                 dentry->d_name = target->d_name;
2787         } else {
2788                 memcpy(dentry->d_iname, target->d_name.name,
2789                                 target->d_name.len + 1);
2790                 dentry->d_name.name = dentry->d_iname;
2791                 dentry->d_name.hash_len = target->d_name.hash_len;
2792         }
2793         if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2794                 call_rcu(&old_name->u.head, __d_free_external_name);
2795 }
2796
2797 /*
2798  * __d_move - move a dentry
2799  * @dentry: entry to move
2800  * @target: new dentry
2801  * @exchange: exchange the two dentries
2802  *
2803  * Update the dcache to reflect the move of a file name. Negative
2804  * dcache entries should not be moved in this way. Caller must hold
2805  * rename_lock, the i_mutex of the source and target directories,
2806  * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2807  */
2808 static void __d_move(struct dentry *dentry, struct dentry *target,
2809                      bool exchange)
2810 {
2811         struct dentry *old_parent, *p;
2812         struct inode *dir = NULL;
2813         unsigned n;
2814
2815         WARN_ON(!dentry->d_inode);
2816         if (WARN_ON(dentry == target))
2817                 return;
2818
2819         BUG_ON(d_ancestor(target, dentry));
2820         old_parent = dentry->d_parent;
2821         p = d_ancestor(old_parent, target);
2822         if (IS_ROOT(dentry)) {
2823                 BUG_ON(p);
2824                 spin_lock(&target->d_parent->d_lock);
2825         } else if (!p) {
2826                 /* target is not a descendent of dentry->d_parent */
2827                 spin_lock(&target->d_parent->d_lock);
2828                 spin_lock_nested(&old_parent->d_lock, DENTRY_D_LOCK_NESTED);
2829         } else {
2830                 BUG_ON(p == dentry);
2831                 spin_lock(&old_parent->d_lock);
2832                 if (p != target)
2833                         spin_lock_nested(&target->d_parent->d_lock,
2834                                         DENTRY_D_LOCK_NESTED);
2835         }
2836         spin_lock_nested(&dentry->d_lock, 2);
2837         spin_lock_nested(&target->d_lock, 3);
2838
2839         if (unlikely(d_in_lookup(target))) {
2840                 dir = target->d_parent->d_inode;
2841                 n = start_dir_add(dir);
2842                 __d_lookup_done(target);
2843         }
2844
2845         write_seqcount_begin(&dentry->d_seq);
2846         write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2847
2848         /* unhash both */
2849         if (!d_unhashed(dentry))
2850                 ___d_drop(dentry);
2851         if (!d_unhashed(target))
2852                 ___d_drop(target);
2853
2854         /* ... and switch them in the tree */
2855         dentry->d_parent = target->d_parent;
2856         if (!exchange) {
2857                 copy_name(dentry, target);
2858                 target->d_hash.pprev = NULL;
2859                 dentry->d_parent->d_lockref.count++;
2860                 if (dentry == old_parent)
2861                         dentry->d_flags |= DCACHE_RCUACCESS;
2862                 else
2863                         WARN_ON(!--old_parent->d_lockref.count);
2864         } else {
2865                 target->d_parent = old_parent;
2866                 swap_names(dentry, target);
2867                 list_move(&target->d_child, &target->d_parent->d_subdirs);
2868                 __d_rehash(target);
2869                 fsnotify_update_flags(target);
2870         }
2871         list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2872         __d_rehash(dentry);
2873         fsnotify_update_flags(dentry);
2874
2875         write_seqcount_end(&target->d_seq);
2876         write_seqcount_end(&dentry->d_seq);
2877
2878         if (dir)
2879                 end_dir_add(dir, n);
2880
2881         if (dentry->d_parent != old_parent)
2882                 spin_unlock(&dentry->d_parent->d_lock);
2883         if (dentry != old_parent)
2884                 spin_unlock(&old_parent->d_lock);
2885         spin_unlock(&target->d_lock);
2886         spin_unlock(&dentry->d_lock);
2887 }
2888
2889 /*
2890  * d_move - move a dentry
2891  * @dentry: entry to move
2892  * @target: new dentry
2893  *
2894  * Update the dcache to reflect the move of a file name. Negative
2895  * dcache entries should not be moved in this way. See the locking
2896  * requirements for __d_move.
2897  */
2898 void d_move(struct dentry *dentry, struct dentry *target)
2899 {
2900         write_seqlock(&rename_lock);
2901         __d_move(dentry, target, false);
2902         write_sequnlock(&rename_lock);
2903 }
2904 EXPORT_SYMBOL(d_move);
2905
2906 /*
2907  * d_exchange - exchange two dentries
2908  * @dentry1: first dentry
2909  * @dentry2: second dentry
2910  */
2911 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2912 {
2913         write_seqlock(&rename_lock);
2914
2915         WARN_ON(!dentry1->d_inode);
2916         WARN_ON(!dentry2->d_inode);
2917         WARN_ON(IS_ROOT(dentry1));
2918         WARN_ON(IS_ROOT(dentry2));
2919
2920         __d_move(dentry1, dentry2, true);
2921
2922         write_sequnlock(&rename_lock);
2923 }
2924
2925 /**
2926  * d_ancestor - search for an ancestor
2927  * @p1: ancestor dentry
2928  * @p2: child dentry
2929  *
2930  * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2931  * an ancestor of p2, else NULL.
2932  */
2933 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2934 {
2935         struct dentry *p;
2936
2937         for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2938                 if (p->d_parent == p1)
2939                         return p;
2940         }
2941         return NULL;
2942 }
2943
2944 /*
2945  * This helper attempts to cope with remotely renamed directories
2946  *
2947  * It assumes that the caller is already holding
2948  * dentry->d_parent->d_inode->i_mutex, and rename_lock
2949  *
2950  * Note: If ever the locking in lock_rename() changes, then please
2951  * remember to update this too...
2952  */
2953 static int __d_unalias(struct inode *inode,
2954                 struct dentry *dentry, struct dentry *alias)
2955 {
2956         struct mutex *m1 = NULL;
2957         struct rw_semaphore *m2 = NULL;
2958         int ret = -ESTALE;
2959
2960         /* If alias and dentry share a parent, then no extra locks required */
2961         if (alias->d_parent == dentry->d_parent)
2962                 goto out_unalias;
2963
2964         /* See lock_rename() */
2965         if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2966                 goto out_err;
2967         m1 = &dentry->d_sb->s_vfs_rename_mutex;
2968         if (!inode_trylock_shared(alias->d_parent->d_inode))
2969                 goto out_err;
2970         m2 = &alias->d_parent->d_inode->i_rwsem;
2971 out_unalias:
2972         __d_move(alias, dentry, false);
2973         ret = 0;
2974 out_err:
2975         if (m2)
2976                 up_read(m2);
2977         if (m1)
2978                 mutex_unlock(m1);
2979         return ret;
2980 }
2981
2982 /**
2983  * d_splice_alias - splice a disconnected dentry into the tree if one exists
2984  * @inode:  the inode which may have a disconnected dentry
2985  * @dentry: a negative dentry which we want to point to the inode.
2986  *
2987  * If inode is a directory and has an IS_ROOT alias, then d_move that in
2988  * place of the given dentry and return it, else simply d_add the inode
2989  * to the dentry and return NULL.
2990  *
2991  * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2992  * we should error out: directories can't have multiple aliases.
2993  *
2994  * This is needed in the lookup routine of any filesystem that is exportable
2995  * (via knfsd) so that we can build dcache paths to directories effectively.
2996  *
2997  * If a dentry was found and moved, then it is returned.  Otherwise NULL
2998  * is returned.  This matches the expected return value of ->lookup.
2999  *
3000  * Cluster filesystems may call this function with a negative, hashed dentry.
3001  * In that case, we know that the inode will be a regular file, and also this
3002  * will only occur during atomic_open. So we need to check for the dentry
3003  * being already hashed only in the final case.
3004  */
3005 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
3006 {
3007         if (IS_ERR(inode))
3008                 return ERR_CAST(inode);
3009
3010         BUG_ON(!d_unhashed(dentry));
3011
3012         if (!inode)
3013                 goto out;
3014
3015         security_d_instantiate(dentry, inode);
3016         spin_lock(&inode->i_lock);
3017         if (S_ISDIR(inode->i_mode)) {
3018                 struct dentry *new = __d_find_any_alias(inode);
3019                 if (unlikely(new)) {
3020                         /* The reference to new ensures it remains an alias */
3021                         spin_unlock(&inode->i_lock);
3022                         write_seqlock(&rename_lock);
3023                         if (unlikely(d_ancestor(new, dentry))) {
3024                                 write_sequnlock(&rename_lock);
3025                                 dput(new);
3026                                 new = ERR_PTR(-ELOOP);
3027                                 pr_warn_ratelimited(
3028                                         "VFS: Lookup of '%s' in %s %s"
3029                                         " would have caused loop\n",
3030                                         dentry->d_name.name,
3031                                         inode->i_sb->s_type->name,
3032                                         inode->i_sb->s_id);
3033                         } else if (!IS_ROOT(new)) {
3034                                 struct dentry *old_parent = dget(new->d_parent);
3035                                 int err = __d_unalias(inode, dentry, new);
3036                                 write_sequnlock(&rename_lock);
3037                                 if (err) {
3038                                         dput(new);
3039                                         new = ERR_PTR(err);
3040                                 }
3041                                 dput(old_parent);
3042                         } else {
3043                                 __d_move(new, dentry, false);
3044                                 write_sequnlock(&rename_lock);
3045                         }
3046                         iput(inode);
3047                         return new;
3048                 }
3049         }
3050 out:
3051         __d_add(dentry, inode);
3052         return NULL;
3053 }
3054 EXPORT_SYMBOL(d_splice_alias);
3055
3056 /*
3057  * Test whether new_dentry is a subdirectory of old_dentry.
3058  *
3059  * Trivially implemented using the dcache structure
3060  */
3061
3062 /**
3063  * is_subdir - is new dentry a subdirectory of old_dentry
3064  * @new_dentry: new dentry
3065  * @old_dentry: old dentry
3066  *
3067  * Returns true if new_dentry is a subdirectory of the parent (at any depth).
3068  * Returns false otherwise.
3069  * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3070  */
3071   
3072 bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3073 {
3074         bool result;
3075         unsigned seq;
3076
3077         if (new_dentry == old_dentry)
3078                 return true;
3079
3080         do {
3081                 /* for restarting inner loop in case of seq retry */
3082                 seq = read_seqbegin(&rename_lock);
3083                 /*
3084                  * Need rcu_readlock to protect against the d_parent trashing
3085                  * due to d_move
3086                  */
3087                 rcu_read_lock();
3088                 if (d_ancestor(old_dentry, new_dentry))
3089                         result = true;
3090                 else
3091                         result = false;
3092                 rcu_read_unlock();
3093         } while (read_seqretry(&rename_lock, seq));
3094
3095         return result;
3096 }
3097 EXPORT_SYMBOL(is_subdir);
3098
3099 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3100 {
3101         struct dentry *root = data;
3102         if (dentry != root) {
3103                 if (d_unhashed(dentry) || !dentry->d_inode)
3104                         return D_WALK_SKIP;
3105
3106                 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3107                         dentry->d_flags |= DCACHE_GENOCIDE;
3108                         dentry->d_lockref.count--;
3109                 }
3110         }
3111         return D_WALK_CONTINUE;
3112 }
3113
3114 void d_genocide(struct dentry *parent)
3115 {
3116         d_walk(parent, parent, d_genocide_kill, NULL);
3117 }
3118
3119 EXPORT_SYMBOL(d_genocide);
3120
3121 void d_tmpfile(struct dentry *dentry, struct inode *inode)
3122 {
3123         inode_dec_link_count(inode);
3124         BUG_ON(dentry->d_name.name != dentry->d_iname ||
3125                 !hlist_unhashed(&dentry->d_u.d_alias) ||
3126                 !d_unlinked(dentry));
3127         spin_lock(&dentry->d_parent->d_lock);
3128         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3129         dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3130                                 (unsigned long long)inode->i_ino);
3131         spin_unlock(&dentry->d_lock);
3132         spin_unlock(&dentry->d_parent->d_lock);
3133         d_instantiate(dentry, inode);
3134 }
3135 EXPORT_SYMBOL(d_tmpfile);
3136
3137 static __initdata unsigned long dhash_entries;
3138 static int __init set_dhash_entries(char *str)
3139 {
3140         if (!str)
3141                 return 0;
3142         dhash_entries = simple_strtoul(str, &str, 0);
3143         return 1;
3144 }
3145 __setup("dhash_entries=", set_dhash_entries);
3146
3147 static void __init dcache_init_early(void)
3148 {
3149         /* If hashes are distributed across NUMA nodes, defer
3150          * hash allocation until vmalloc space is available.
3151          */
3152         if (hashdist)
3153                 return;
3154
3155         dentry_hashtable =
3156                 alloc_large_system_hash("Dentry cache",
3157                                         sizeof(struct hlist_bl_head),
3158                                         dhash_entries,
3159                                         13,
3160                                         HASH_EARLY | HASH_ZERO,
3161                                         &d_hash_shift,
3162                                         NULL,
3163                                         0,
3164                                         0);
3165         d_hash_shift = 32 - d_hash_shift;
3166 }
3167
3168 static void __init dcache_init(void)
3169 {
3170         /*
3171          * A constructor could be added for stable state like the lists,
3172          * but it is probably not worth it because of the cache nature
3173          * of the dcache.
3174          */
3175         dentry_cache = KMEM_CACHE_USERCOPY(dentry,
3176                 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD|SLAB_ACCOUNT,
3177                 d_iname);
3178
3179         /* Hash may have been set up in dcache_init_early */
3180         if (!hashdist)
3181                 return;
3182
3183         dentry_hashtable =
3184                 alloc_large_system_hash("Dentry cache",
3185                                         sizeof(struct hlist_bl_head),
3186                                         dhash_entries,
3187                                         13,
3188                                         HASH_ZERO,
3189                                         &d_hash_shift,
3190                                         NULL,
3191                                         0,
3192                                         0);
3193         d_hash_shift = 32 - d_hash_shift;
3194 }
3195
3196 /* SLAB cache for __getname() consumers */
3197 struct kmem_cache *names_cachep __read_mostly;
3198 EXPORT_SYMBOL(names_cachep);
3199
3200 void __init vfs_caches_init_early(void)
3201 {
3202         int i;
3203
3204         for (i = 0; i < ARRAY_SIZE(in_lookup_hashtable); i++)
3205                 INIT_HLIST_BL_HEAD(&in_lookup_hashtable[i]);
3206
3207         dcache_init_early();
3208         inode_init_early();
3209 }
3210
3211 void __init vfs_caches_init(void)
3212 {
3213         names_cachep = kmem_cache_create_usercopy("names_cache", PATH_MAX, 0,
3214                         SLAB_HWCACHE_ALIGN|SLAB_PANIC, 0, PATH_MAX, NULL);
3215
3216         dcache_init();
3217         inode_init();
3218         files_init();
3219         files_maxfiles_init();
3220         mnt_init();
3221         bdev_cache_init();
3222         chrdev_init();
3223 }