86d2de63461e1550efe205643d66800dd12f15ac
[muen/linux.git] / fs / dcache.c
1 /*
2  * fs/dcache.c
3  *
4  * Complete reimplementation
5  * (C) 1997 Thomas Schoebel-Theuer,
6  * with heavy changes by Linus Torvalds
7  */
8
9 /*
10  * Notes on the allocation strategy:
11  *
12  * The dcache is a master of the icache - whenever a dcache entry
13  * exists, the inode will always exist. "iput()" is done either when
14  * the dcache entry is deleted or garbage collected.
15  */
16
17 #include <linux/ratelimit.h>
18 #include <linux/string.h>
19 #include <linux/mm.h>
20 #include <linux/fs.h>
21 #include <linux/fsnotify.h>
22 #include <linux/slab.h>
23 #include <linux/init.h>
24 #include <linux/hash.h>
25 #include <linux/cache.h>
26 #include <linux/export.h>
27 #include <linux/security.h>
28 #include <linux/seqlock.h>
29 #include <linux/bootmem.h>
30 #include <linux/bit_spinlock.h>
31 #include <linux/rculist_bl.h>
32 #include <linux/list_lru.h>
33 #include "internal.h"
34 #include "mount.h"
35
36 /*
37  * Usage:
38  * dcache->d_inode->i_lock protects:
39  *   - i_dentry, d_u.d_alias, d_inode of aliases
40  * dcache_hash_bucket lock protects:
41  *   - the dcache hash table
42  * s_roots bl list spinlock protects:
43  *   - the s_roots list (see __d_drop)
44  * dentry->d_sb->s_dentry_lru_lock protects:
45  *   - the dcache lru lists and counters
46  * d_lock protects:
47  *   - d_flags
48  *   - d_name
49  *   - d_lru
50  *   - d_count
51  *   - d_unhashed()
52  *   - d_parent and d_subdirs
53  *   - childrens' d_child and d_parent
54  *   - d_u.d_alias, d_inode
55  *
56  * Ordering:
57  * dentry->d_inode->i_lock
58  *   dentry->d_lock
59  *     dentry->d_sb->s_dentry_lru_lock
60  *     dcache_hash_bucket lock
61  *     s_roots lock
62  *
63  * If there is an ancestor relationship:
64  * dentry->d_parent->...->d_parent->d_lock
65  *   ...
66  *     dentry->d_parent->d_lock
67  *       dentry->d_lock
68  *
69  * If no ancestor relationship:
70  * arbitrary, since it's serialized on rename_lock
71  */
72 int sysctl_vfs_cache_pressure __read_mostly = 100;
73 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
74
75 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
76
77 EXPORT_SYMBOL(rename_lock);
78
79 static struct kmem_cache *dentry_cache __read_mostly;
80
81 const struct qstr empty_name = QSTR_INIT("", 0);
82 EXPORT_SYMBOL(empty_name);
83 const struct qstr slash_name = QSTR_INIT("/", 1);
84 EXPORT_SYMBOL(slash_name);
85
86 /*
87  * This is the single most critical data structure when it comes
88  * to the dcache: the hashtable for lookups. Somebody should try
89  * to make this good - I've just made it work.
90  *
91  * This hash-function tries to avoid losing too many bits of hash
92  * information, yet avoid using a prime hash-size or similar.
93  */
94
95 static unsigned int d_hash_shift __read_mostly;
96
97 static struct hlist_bl_head *dentry_hashtable __read_mostly;
98
99 static inline struct hlist_bl_head *d_hash(unsigned int hash)
100 {
101         return dentry_hashtable + (hash >> d_hash_shift);
102 }
103
104 #define IN_LOOKUP_SHIFT 10
105 static struct hlist_bl_head in_lookup_hashtable[1 << IN_LOOKUP_SHIFT];
106
107 static inline struct hlist_bl_head *in_lookup_hash(const struct dentry *parent,
108                                         unsigned int hash)
109 {
110         hash += (unsigned long) parent / L1_CACHE_BYTES;
111         return in_lookup_hashtable + hash_32(hash, IN_LOOKUP_SHIFT);
112 }
113
114
115 /* Statistics gathering. */
116 struct dentry_stat_t dentry_stat = {
117         .age_limit = 45,
118 };
119
120 static DEFINE_PER_CPU(long, nr_dentry);
121 static DEFINE_PER_CPU(long, nr_dentry_unused);
122
123 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
124
125 /*
126  * Here we resort to our own counters instead of using generic per-cpu counters
127  * for consistency with what the vfs inode code does. We are expected to harvest
128  * better code and performance by having our own specialized counters.
129  *
130  * Please note that the loop is done over all possible CPUs, not over all online
131  * CPUs. The reason for this is that we don't want to play games with CPUs going
132  * on and off. If one of them goes off, we will just keep their counters.
133  *
134  * glommer: See cffbc8a for details, and if you ever intend to change this,
135  * please update all vfs counters to match.
136  */
137 static long get_nr_dentry(void)
138 {
139         int i;
140         long sum = 0;
141         for_each_possible_cpu(i)
142                 sum += per_cpu(nr_dentry, i);
143         return sum < 0 ? 0 : sum;
144 }
145
146 static long get_nr_dentry_unused(void)
147 {
148         int i;
149         long sum = 0;
150         for_each_possible_cpu(i)
151                 sum += per_cpu(nr_dentry_unused, i);
152         return sum < 0 ? 0 : sum;
153 }
154
155 int proc_nr_dentry(struct ctl_table *table, int write, void __user *buffer,
156                    size_t *lenp, loff_t *ppos)
157 {
158         dentry_stat.nr_dentry = get_nr_dentry();
159         dentry_stat.nr_unused = get_nr_dentry_unused();
160         return proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
161 }
162 #endif
163
164 /*
165  * Compare 2 name strings, return 0 if they match, otherwise non-zero.
166  * The strings are both count bytes long, and count is non-zero.
167  */
168 #ifdef CONFIG_DCACHE_WORD_ACCESS
169
170 #include <asm/word-at-a-time.h>
171 /*
172  * NOTE! 'cs' and 'scount' come from a dentry, so it has a
173  * aligned allocation for this particular component. We don't
174  * strictly need the load_unaligned_zeropad() safety, but it
175  * doesn't hurt either.
176  *
177  * In contrast, 'ct' and 'tcount' can be from a pathname, and do
178  * need the careful unaligned handling.
179  */
180 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
181 {
182         unsigned long a,b,mask;
183
184         for (;;) {
185                 a = read_word_at_a_time(cs);
186                 b = load_unaligned_zeropad(ct);
187                 if (tcount < sizeof(unsigned long))
188                         break;
189                 if (unlikely(a != b))
190                         return 1;
191                 cs += sizeof(unsigned long);
192                 ct += sizeof(unsigned long);
193                 tcount -= sizeof(unsigned long);
194                 if (!tcount)
195                         return 0;
196         }
197         mask = bytemask_from_count(tcount);
198         return unlikely(!!((a ^ b) & mask));
199 }
200
201 #else
202
203 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
204 {
205         do {
206                 if (*cs != *ct)
207                         return 1;
208                 cs++;
209                 ct++;
210                 tcount--;
211         } while (tcount);
212         return 0;
213 }
214
215 #endif
216
217 static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount)
218 {
219         /*
220          * Be careful about RCU walk racing with rename:
221          * use 'READ_ONCE' to fetch the name pointer.
222          *
223          * NOTE! Even if a rename will mean that the length
224          * was not loaded atomically, we don't care. The
225          * RCU walk will check the sequence count eventually,
226          * and catch it. And we won't overrun the buffer,
227          * because we're reading the name pointer atomically,
228          * and a dentry name is guaranteed to be properly
229          * terminated with a NUL byte.
230          *
231          * End result: even if 'len' is wrong, we'll exit
232          * early because the data cannot match (there can
233          * be no NUL in the ct/tcount data)
234          */
235         const unsigned char *cs = READ_ONCE(dentry->d_name.name);
236
237         return dentry_string_cmp(cs, ct, tcount);
238 }
239
240 struct external_name {
241         union {
242                 atomic_t count;
243                 struct rcu_head head;
244         } u;
245         unsigned char name[];
246 };
247
248 static inline struct external_name *external_name(struct dentry *dentry)
249 {
250         return container_of(dentry->d_name.name, struct external_name, name[0]);
251 }
252
253 static void __d_free(struct rcu_head *head)
254 {
255         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
256
257         kmem_cache_free(dentry_cache, dentry); 
258 }
259
260 static void __d_free_external_name(struct rcu_head *head)
261 {
262         struct external_name *name = container_of(head, struct external_name,
263                                                   u.head);
264
265         mod_node_page_state(page_pgdat(virt_to_page(name)),
266                             NR_INDIRECTLY_RECLAIMABLE_BYTES,
267                             -ksize(name));
268
269         kfree(name);
270 }
271
272 static void __d_free_external(struct rcu_head *head)
273 {
274         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
275
276         __d_free_external_name(&external_name(dentry)->u.head);
277
278         kmem_cache_free(dentry_cache, dentry);
279 }
280
281 static inline int dname_external(const struct dentry *dentry)
282 {
283         return dentry->d_name.name != dentry->d_iname;
284 }
285
286 void take_dentry_name_snapshot(struct name_snapshot *name, struct dentry *dentry)
287 {
288         spin_lock(&dentry->d_lock);
289         if (unlikely(dname_external(dentry))) {
290                 struct external_name *p = external_name(dentry);
291                 atomic_inc(&p->u.count);
292                 spin_unlock(&dentry->d_lock);
293                 name->name = p->name;
294         } else {
295                 memcpy(name->inline_name, dentry->d_iname, DNAME_INLINE_LEN);
296                 spin_unlock(&dentry->d_lock);
297                 name->name = name->inline_name;
298         }
299 }
300 EXPORT_SYMBOL(take_dentry_name_snapshot);
301
302 void release_dentry_name_snapshot(struct name_snapshot *name)
303 {
304         if (unlikely(name->name != name->inline_name)) {
305                 struct external_name *p;
306                 p = container_of(name->name, struct external_name, name[0]);
307                 if (unlikely(atomic_dec_and_test(&p->u.count)))
308                         call_rcu(&p->u.head, __d_free_external_name);
309         }
310 }
311 EXPORT_SYMBOL(release_dentry_name_snapshot);
312
313 static inline void __d_set_inode_and_type(struct dentry *dentry,
314                                           struct inode *inode,
315                                           unsigned type_flags)
316 {
317         unsigned flags;
318
319         dentry->d_inode = inode;
320         flags = READ_ONCE(dentry->d_flags);
321         flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
322         flags |= type_flags;
323         WRITE_ONCE(dentry->d_flags, flags);
324 }
325
326 static inline void __d_clear_type_and_inode(struct dentry *dentry)
327 {
328         unsigned flags = READ_ONCE(dentry->d_flags);
329
330         flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
331         WRITE_ONCE(dentry->d_flags, flags);
332         dentry->d_inode = NULL;
333 }
334
335 static void dentry_free(struct dentry *dentry)
336 {
337         WARN_ON(!hlist_unhashed(&dentry->d_u.d_alias));
338         if (unlikely(dname_external(dentry))) {
339                 struct external_name *p = external_name(dentry);
340                 if (likely(atomic_dec_and_test(&p->u.count))) {
341                         call_rcu(&dentry->d_u.d_rcu, __d_free_external);
342                         return;
343                 }
344         }
345         /* if dentry was never visible to RCU, immediate free is OK */
346         if (!(dentry->d_flags & DCACHE_RCUACCESS))
347                 __d_free(&dentry->d_u.d_rcu);
348         else
349                 call_rcu(&dentry->d_u.d_rcu, __d_free);
350 }
351
352 /*
353  * Release the dentry's inode, using the filesystem
354  * d_iput() operation if defined.
355  */
356 static void dentry_unlink_inode(struct dentry * dentry)
357         __releases(dentry->d_lock)
358         __releases(dentry->d_inode->i_lock)
359 {
360         struct inode *inode = dentry->d_inode;
361         bool hashed = !d_unhashed(dentry);
362
363         if (hashed)
364                 raw_write_seqcount_begin(&dentry->d_seq);
365         __d_clear_type_and_inode(dentry);
366         hlist_del_init(&dentry->d_u.d_alias);
367         if (hashed)
368                 raw_write_seqcount_end(&dentry->d_seq);
369         spin_unlock(&dentry->d_lock);
370         spin_unlock(&inode->i_lock);
371         if (!inode->i_nlink)
372                 fsnotify_inoderemove(inode);
373         if (dentry->d_op && dentry->d_op->d_iput)
374                 dentry->d_op->d_iput(dentry, inode);
375         else
376                 iput(inode);
377 }
378
379 /*
380  * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
381  * is in use - which includes both the "real" per-superblock
382  * LRU list _and_ the DCACHE_SHRINK_LIST use.
383  *
384  * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
385  * on the shrink list (ie not on the superblock LRU list).
386  *
387  * The per-cpu "nr_dentry_unused" counters are updated with
388  * the DCACHE_LRU_LIST bit.
389  *
390  * These helper functions make sure we always follow the
391  * rules. d_lock must be held by the caller.
392  */
393 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
394 static void d_lru_add(struct dentry *dentry)
395 {
396         D_FLAG_VERIFY(dentry, 0);
397         dentry->d_flags |= DCACHE_LRU_LIST;
398         this_cpu_inc(nr_dentry_unused);
399         WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
400 }
401
402 static void d_lru_del(struct dentry *dentry)
403 {
404         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
405         dentry->d_flags &= ~DCACHE_LRU_LIST;
406         this_cpu_dec(nr_dentry_unused);
407         WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
408 }
409
410 static void d_shrink_del(struct dentry *dentry)
411 {
412         D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
413         list_del_init(&dentry->d_lru);
414         dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
415         this_cpu_dec(nr_dentry_unused);
416 }
417
418 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
419 {
420         D_FLAG_VERIFY(dentry, 0);
421         list_add(&dentry->d_lru, list);
422         dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
423         this_cpu_inc(nr_dentry_unused);
424 }
425
426 /*
427  * These can only be called under the global LRU lock, ie during the
428  * callback for freeing the LRU list. "isolate" removes it from the
429  * LRU lists entirely, while shrink_move moves it to the indicated
430  * private list.
431  */
432 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
433 {
434         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
435         dentry->d_flags &= ~DCACHE_LRU_LIST;
436         this_cpu_dec(nr_dentry_unused);
437         list_lru_isolate(lru, &dentry->d_lru);
438 }
439
440 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
441                               struct list_head *list)
442 {
443         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
444         dentry->d_flags |= DCACHE_SHRINK_LIST;
445         list_lru_isolate_move(lru, &dentry->d_lru, list);
446 }
447
448 /**
449  * d_drop - drop a dentry
450  * @dentry: dentry to drop
451  *
452  * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
453  * be found through a VFS lookup any more. Note that this is different from
454  * deleting the dentry - d_delete will try to mark the dentry negative if
455  * possible, giving a successful _negative_ lookup, while d_drop will
456  * just make the cache lookup fail.
457  *
458  * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
459  * reason (NFS timeouts or autofs deletes).
460  *
461  * __d_drop requires dentry->d_lock
462  * ___d_drop doesn't mark dentry as "unhashed"
463  *   (dentry->d_hash.pprev will be LIST_POISON2, not NULL).
464  */
465 static void ___d_drop(struct dentry *dentry)
466 {
467         struct hlist_bl_head *b;
468         /*
469          * Hashed dentries are normally on the dentry hashtable,
470          * with the exception of those newly allocated by
471          * d_obtain_root, which are always IS_ROOT:
472          */
473         if (unlikely(IS_ROOT(dentry)))
474                 b = &dentry->d_sb->s_roots;
475         else
476                 b = d_hash(dentry->d_name.hash);
477
478         hlist_bl_lock(b);
479         __hlist_bl_del(&dentry->d_hash);
480         hlist_bl_unlock(b);
481 }
482
483 void __d_drop(struct dentry *dentry)
484 {
485         if (!d_unhashed(dentry)) {
486                 ___d_drop(dentry);
487                 dentry->d_hash.pprev = NULL;
488                 write_seqcount_invalidate(&dentry->d_seq);
489         }
490 }
491 EXPORT_SYMBOL(__d_drop);
492
493 void d_drop(struct dentry *dentry)
494 {
495         spin_lock(&dentry->d_lock);
496         __d_drop(dentry);
497         spin_unlock(&dentry->d_lock);
498 }
499 EXPORT_SYMBOL(d_drop);
500
501 static inline void dentry_unlist(struct dentry *dentry, struct dentry *parent)
502 {
503         struct dentry *next;
504         /*
505          * Inform d_walk() and shrink_dentry_list() that we are no longer
506          * attached to the dentry tree
507          */
508         dentry->d_flags |= DCACHE_DENTRY_KILLED;
509         if (unlikely(list_empty(&dentry->d_child)))
510                 return;
511         __list_del_entry(&dentry->d_child);
512         /*
513          * Cursors can move around the list of children.  While we'd been
514          * a normal list member, it didn't matter - ->d_child.next would've
515          * been updated.  However, from now on it won't be and for the
516          * things like d_walk() it might end up with a nasty surprise.
517          * Normally d_walk() doesn't care about cursors moving around -
518          * ->d_lock on parent prevents that and since a cursor has no children
519          * of its own, we get through it without ever unlocking the parent.
520          * There is one exception, though - if we ascend from a child that
521          * gets killed as soon as we unlock it, the next sibling is found
522          * using the value left in its ->d_child.next.  And if _that_
523          * pointed to a cursor, and cursor got moved (e.g. by lseek())
524          * before d_walk() regains parent->d_lock, we'll end up skipping
525          * everything the cursor had been moved past.
526          *
527          * Solution: make sure that the pointer left behind in ->d_child.next
528          * points to something that won't be moving around.  I.e. skip the
529          * cursors.
530          */
531         while (dentry->d_child.next != &parent->d_subdirs) {
532                 next = list_entry(dentry->d_child.next, struct dentry, d_child);
533                 if (likely(!(next->d_flags & DCACHE_DENTRY_CURSOR)))
534                         break;
535                 dentry->d_child.next = next->d_child.next;
536         }
537 }
538
539 static void __dentry_kill(struct dentry *dentry)
540 {
541         struct dentry *parent = NULL;
542         bool can_free = true;
543         if (!IS_ROOT(dentry))
544                 parent = dentry->d_parent;
545
546         /*
547          * The dentry is now unrecoverably dead to the world.
548          */
549         lockref_mark_dead(&dentry->d_lockref);
550
551         /*
552          * inform the fs via d_prune that this dentry is about to be
553          * unhashed and destroyed.
554          */
555         if (dentry->d_flags & DCACHE_OP_PRUNE)
556                 dentry->d_op->d_prune(dentry);
557
558         if (dentry->d_flags & DCACHE_LRU_LIST) {
559                 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
560                         d_lru_del(dentry);
561         }
562         /* if it was on the hash then remove it */
563         __d_drop(dentry);
564         dentry_unlist(dentry, parent);
565         if (parent)
566                 spin_unlock(&parent->d_lock);
567         if (dentry->d_inode)
568                 dentry_unlink_inode(dentry);
569         else
570                 spin_unlock(&dentry->d_lock);
571         this_cpu_dec(nr_dentry);
572         if (dentry->d_op && dentry->d_op->d_release)
573                 dentry->d_op->d_release(dentry);
574
575         spin_lock(&dentry->d_lock);
576         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
577                 dentry->d_flags |= DCACHE_MAY_FREE;
578                 can_free = false;
579         }
580         spin_unlock(&dentry->d_lock);
581         if (likely(can_free))
582                 dentry_free(dentry);
583 }
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                 cond_resched();
1056
1057                 dentry = list_entry(list->prev, struct dentry, d_lru);
1058                 spin_lock(&dentry->d_lock);
1059                 rcu_read_lock();
1060                 if (!shrink_lock_dentry(dentry)) {
1061                         bool can_free = false;
1062                         rcu_read_unlock();
1063                         d_shrink_del(dentry);
1064                         if (dentry->d_lockref.count < 0)
1065                                 can_free = dentry->d_flags & DCACHE_MAY_FREE;
1066                         spin_unlock(&dentry->d_lock);
1067                         if (can_free)
1068                                 dentry_free(dentry);
1069                         continue;
1070                 }
1071                 rcu_read_unlock();
1072                 d_shrink_del(dentry);
1073                 parent = dentry->d_parent;
1074                 __dentry_kill(dentry);
1075                 if (parent == dentry)
1076                         continue;
1077                 /*
1078                  * We need to prune ancestors too. This is necessary to prevent
1079                  * quadratic behavior of shrink_dcache_parent(), but is also
1080                  * expected to be beneficial in reducing dentry cache
1081                  * fragmentation.
1082                  */
1083                 dentry = parent;
1084                 while (dentry && !lockref_put_or_lock(&dentry->d_lockref))
1085                         dentry = dentry_kill(dentry);
1086         }
1087 }
1088
1089 static enum lru_status dentry_lru_isolate(struct list_head *item,
1090                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1091 {
1092         struct list_head *freeable = arg;
1093         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1094
1095
1096         /*
1097          * we are inverting the lru lock/dentry->d_lock here,
1098          * so use a trylock. If we fail to get the lock, just skip
1099          * it
1100          */
1101         if (!spin_trylock(&dentry->d_lock))
1102                 return LRU_SKIP;
1103
1104         /*
1105          * Referenced dentries are still in use. If they have active
1106          * counts, just remove them from the LRU. Otherwise give them
1107          * another pass through the LRU.
1108          */
1109         if (dentry->d_lockref.count) {
1110                 d_lru_isolate(lru, dentry);
1111                 spin_unlock(&dentry->d_lock);
1112                 return LRU_REMOVED;
1113         }
1114
1115         if (dentry->d_flags & DCACHE_REFERENCED) {
1116                 dentry->d_flags &= ~DCACHE_REFERENCED;
1117                 spin_unlock(&dentry->d_lock);
1118
1119                 /*
1120                  * The list move itself will be made by the common LRU code. At
1121                  * this point, we've dropped the dentry->d_lock but keep the
1122                  * lru lock. This is safe to do, since every list movement is
1123                  * protected by the lru lock even if both locks are held.
1124                  *
1125                  * This is guaranteed by the fact that all LRU management
1126                  * functions are intermediated by the LRU API calls like
1127                  * list_lru_add and list_lru_del. List movement in this file
1128                  * only ever occur through this functions or through callbacks
1129                  * like this one, that are called from the LRU API.
1130                  *
1131                  * The only exceptions to this are functions like
1132                  * shrink_dentry_list, and code that first checks for the
1133                  * DCACHE_SHRINK_LIST flag.  Those are guaranteed to be
1134                  * operating only with stack provided lists after they are
1135                  * properly isolated from the main list.  It is thus, always a
1136                  * local access.
1137                  */
1138                 return LRU_ROTATE;
1139         }
1140
1141         d_lru_shrink_move(lru, dentry, freeable);
1142         spin_unlock(&dentry->d_lock);
1143
1144         return LRU_REMOVED;
1145 }
1146
1147 /**
1148  * prune_dcache_sb - shrink the dcache
1149  * @sb: superblock
1150  * @sc: shrink control, passed to list_lru_shrink_walk()
1151  *
1152  * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1153  * is done when we need more memory and called from the superblock shrinker
1154  * function.
1155  *
1156  * This function may fail to free any resources if all the dentries are in
1157  * use.
1158  */
1159 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1160 {
1161         LIST_HEAD(dispose);
1162         long freed;
1163
1164         freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1165                                      dentry_lru_isolate, &dispose);
1166         shrink_dentry_list(&dispose);
1167         return freed;
1168 }
1169
1170 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1171                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1172 {
1173         struct list_head *freeable = arg;
1174         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1175
1176         /*
1177          * we are inverting the lru lock/dentry->d_lock here,
1178          * so use a trylock. If we fail to get the lock, just skip
1179          * it
1180          */
1181         if (!spin_trylock(&dentry->d_lock))
1182                 return LRU_SKIP;
1183
1184         d_lru_shrink_move(lru, dentry, freeable);
1185         spin_unlock(&dentry->d_lock);
1186
1187         return LRU_REMOVED;
1188 }
1189
1190
1191 /**
1192  * shrink_dcache_sb - shrink dcache for a superblock
1193  * @sb: superblock
1194  *
1195  * Shrink the dcache for the specified super block. This is used to free
1196  * the dcache before unmounting a file system.
1197  */
1198 void shrink_dcache_sb(struct super_block *sb)
1199 {
1200         long freed;
1201
1202         do {
1203                 LIST_HEAD(dispose);
1204
1205                 freed = list_lru_walk(&sb->s_dentry_lru,
1206                         dentry_lru_isolate_shrink, &dispose, 1024);
1207
1208                 this_cpu_sub(nr_dentry_unused, freed);
1209                 shrink_dentry_list(&dispose);
1210         } while (list_lru_count(&sb->s_dentry_lru) > 0);
1211 }
1212 EXPORT_SYMBOL(shrink_dcache_sb);
1213
1214 /**
1215  * enum d_walk_ret - action to talke during tree walk
1216  * @D_WALK_CONTINUE:    contrinue walk
1217  * @D_WALK_QUIT:        quit walk
1218  * @D_WALK_NORETRY:     quit when retry is needed
1219  * @D_WALK_SKIP:        skip this dentry and its children
1220  */
1221 enum d_walk_ret {
1222         D_WALK_CONTINUE,
1223         D_WALK_QUIT,
1224         D_WALK_NORETRY,
1225         D_WALK_SKIP,
1226 };
1227
1228 /**
1229  * d_walk - walk the dentry tree
1230  * @parent:     start of walk
1231  * @data:       data passed to @enter() and @finish()
1232  * @enter:      callback when first entering the dentry
1233  * @finish:     callback when successfully finished the walk
1234  *
1235  * The @enter() and @finish() callbacks are called with d_lock held.
1236  */
1237 static void d_walk(struct dentry *parent, void *data,
1238                    enum d_walk_ret (*enter)(void *, struct dentry *),
1239                    void (*finish)(void *))
1240 {
1241         struct dentry *this_parent;
1242         struct list_head *next;
1243         unsigned seq = 0;
1244         enum d_walk_ret ret;
1245         bool retry = true;
1246
1247 again:
1248         read_seqbegin_or_lock(&rename_lock, &seq);
1249         this_parent = parent;
1250         spin_lock(&this_parent->d_lock);
1251
1252         ret = enter(data, this_parent);
1253         switch (ret) {
1254         case D_WALK_CONTINUE:
1255                 break;
1256         case D_WALK_QUIT:
1257         case D_WALK_SKIP:
1258                 goto out_unlock;
1259         case D_WALK_NORETRY:
1260                 retry = false;
1261                 break;
1262         }
1263 repeat:
1264         next = this_parent->d_subdirs.next;
1265 resume:
1266         while (next != &this_parent->d_subdirs) {
1267                 struct list_head *tmp = next;
1268                 struct dentry *dentry = list_entry(tmp, struct dentry, d_child);
1269                 next = tmp->next;
1270
1271                 if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
1272                         continue;
1273
1274                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1275
1276                 ret = enter(data, dentry);
1277                 switch (ret) {
1278                 case D_WALK_CONTINUE:
1279                         break;
1280                 case D_WALK_QUIT:
1281                         spin_unlock(&dentry->d_lock);
1282                         goto out_unlock;
1283                 case D_WALK_NORETRY:
1284                         retry = false;
1285                         break;
1286                 case D_WALK_SKIP:
1287                         spin_unlock(&dentry->d_lock);
1288                         continue;
1289                 }
1290
1291                 if (!list_empty(&dentry->d_subdirs)) {
1292                         spin_unlock(&this_parent->d_lock);
1293                         spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1294                         this_parent = dentry;
1295                         spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1296                         goto repeat;
1297                 }
1298                 spin_unlock(&dentry->d_lock);
1299         }
1300         /*
1301          * All done at this level ... ascend and resume the search.
1302          */
1303         rcu_read_lock();
1304 ascend:
1305         if (this_parent != parent) {
1306                 struct dentry *child = this_parent;
1307                 this_parent = child->d_parent;
1308
1309                 spin_unlock(&child->d_lock);
1310                 spin_lock(&this_parent->d_lock);
1311
1312                 /* might go back up the wrong parent if we have had a rename. */
1313                 if (need_seqretry(&rename_lock, seq))
1314                         goto rename_retry;
1315                 /* go into the first sibling still alive */
1316                 do {
1317                         next = child->d_child.next;
1318                         if (next == &this_parent->d_subdirs)
1319                                 goto ascend;
1320                         child = list_entry(next, struct dentry, d_child);
1321                 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED));
1322                 rcu_read_unlock();
1323                 goto resume;
1324         }
1325         if (need_seqretry(&rename_lock, seq))
1326                 goto rename_retry;
1327         rcu_read_unlock();
1328         if (finish)
1329                 finish(data);
1330
1331 out_unlock:
1332         spin_unlock(&this_parent->d_lock);
1333         done_seqretry(&rename_lock, seq);
1334         return;
1335
1336 rename_retry:
1337         spin_unlock(&this_parent->d_lock);
1338         rcu_read_unlock();
1339         BUG_ON(seq & 1);
1340         if (!retry)
1341                 return;
1342         seq = 1;
1343         goto again;
1344 }
1345
1346 struct check_mount {
1347         struct vfsmount *mnt;
1348         unsigned int mounted;
1349 };
1350
1351 static enum d_walk_ret path_check_mount(void *data, struct dentry *dentry)
1352 {
1353         struct check_mount *info = data;
1354         struct path path = { .mnt = info->mnt, .dentry = dentry };
1355
1356         if (likely(!d_mountpoint(dentry)))
1357                 return D_WALK_CONTINUE;
1358         if (__path_is_mountpoint(&path)) {
1359                 info->mounted = 1;
1360                 return D_WALK_QUIT;
1361         }
1362         return D_WALK_CONTINUE;
1363 }
1364
1365 /**
1366  * path_has_submounts - check for mounts over a dentry in the
1367  *                      current namespace.
1368  * @parent: path to check.
1369  *
1370  * Return true if the parent or its subdirectories contain
1371  * a mount point in the current namespace.
1372  */
1373 int path_has_submounts(const struct path *parent)
1374 {
1375         struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };
1376
1377         read_seqlock_excl(&mount_lock);
1378         d_walk(parent->dentry, &data, path_check_mount, NULL);
1379         read_sequnlock_excl(&mount_lock);
1380
1381         return data.mounted;
1382 }
1383 EXPORT_SYMBOL(path_has_submounts);
1384
1385 /*
1386  * Called by mount code to set a mountpoint and check if the mountpoint is
1387  * reachable (e.g. NFS can unhash a directory dentry and then the complete
1388  * subtree can become unreachable).
1389  *
1390  * Only one of d_invalidate() and d_set_mounted() must succeed.  For
1391  * this reason take rename_lock and d_lock on dentry and ancestors.
1392  */
1393 int d_set_mounted(struct dentry *dentry)
1394 {
1395         struct dentry *p;
1396         int ret = -ENOENT;
1397         write_seqlock(&rename_lock);
1398         for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1399                 /* Need exclusion wrt. d_invalidate() */
1400                 spin_lock(&p->d_lock);
1401                 if (unlikely(d_unhashed(p))) {
1402                         spin_unlock(&p->d_lock);
1403                         goto out;
1404                 }
1405                 spin_unlock(&p->d_lock);
1406         }
1407         spin_lock(&dentry->d_lock);
1408         if (!d_unlinked(dentry)) {
1409                 ret = -EBUSY;
1410                 if (!d_mountpoint(dentry)) {
1411                         dentry->d_flags |= DCACHE_MOUNTED;
1412                         ret = 0;
1413                 }
1414         }
1415         spin_unlock(&dentry->d_lock);
1416 out:
1417         write_sequnlock(&rename_lock);
1418         return ret;
1419 }
1420
1421 /*
1422  * Search the dentry child list of the specified parent,
1423  * and move any unused dentries to the end of the unused
1424  * list for prune_dcache(). We descend to the next level
1425  * whenever the d_subdirs list is non-empty and continue
1426  * searching.
1427  *
1428  * It returns zero iff there are no unused children,
1429  * otherwise  it returns the number of children moved to
1430  * the end of the unused list. This may not be the total
1431  * number of unused children, because select_parent can
1432  * drop the lock and return early due to latency
1433  * constraints.
1434  */
1435
1436 struct select_data {
1437         struct dentry *start;
1438         struct list_head dispose;
1439         int found;
1440 };
1441
1442 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1443 {
1444         struct select_data *data = _data;
1445         enum d_walk_ret ret = D_WALK_CONTINUE;
1446
1447         if (data->start == dentry)
1448                 goto out;
1449
1450         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1451                 data->found++;
1452         } else {
1453                 if (dentry->d_flags & DCACHE_LRU_LIST)
1454                         d_lru_del(dentry);
1455                 if (!dentry->d_lockref.count) {
1456                         d_shrink_add(dentry, &data->dispose);
1457                         data->found++;
1458                 }
1459         }
1460         /*
1461          * We can return to the caller if we have found some (this
1462          * ensures forward progress). We'll be coming back to find
1463          * the rest.
1464          */
1465         if (!list_empty(&data->dispose))
1466                 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1467 out:
1468         return ret;
1469 }
1470
1471 /**
1472  * shrink_dcache_parent - prune dcache
1473  * @parent: parent of entries to prune
1474  *
1475  * Prune the dcache to remove unused children of the parent dentry.
1476  */
1477 void shrink_dcache_parent(struct dentry *parent)
1478 {
1479         for (;;) {
1480                 struct select_data data;
1481
1482                 INIT_LIST_HEAD(&data.dispose);
1483                 data.start = parent;
1484                 data.found = 0;
1485
1486                 d_walk(parent, &data, select_collect, NULL);
1487                 if (!data.found)
1488                         break;
1489
1490                 shrink_dentry_list(&data.dispose);
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         }
1618 }
1619 EXPORT_SYMBOL(d_invalidate);
1620
1621 /**
1622  * __d_alloc    -       allocate a dcache entry
1623  * @sb: filesystem it will belong to
1624  * @name: qstr of the name
1625  *
1626  * Allocates a dentry. It returns %NULL if there is insufficient memory
1627  * available. On a success the dentry is returned. The name passed in is
1628  * copied and the copy passed in may be reused after this call.
1629  */
1630  
1631 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1632 {
1633         struct external_name *ext = NULL;
1634         struct dentry *dentry;
1635         char *dname;
1636         int err;
1637
1638         dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1639         if (!dentry)
1640                 return NULL;
1641
1642         /*
1643          * We guarantee that the inline name is always NUL-terminated.
1644          * This way the memcpy() done by the name switching in rename
1645          * will still always have a NUL at the end, even if we might
1646          * be overwriting an internal NUL character
1647          */
1648         dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1649         if (unlikely(!name)) {
1650                 name = &slash_name;
1651                 dname = dentry->d_iname;
1652         } else if (name->len > DNAME_INLINE_LEN-1) {
1653                 size_t size = offsetof(struct external_name, name[1]);
1654
1655                 ext = kmalloc(size + name->len, GFP_KERNEL_ACCOUNT);
1656                 if (!ext) {
1657                         kmem_cache_free(dentry_cache, dentry); 
1658                         return NULL;
1659                 }
1660                 atomic_set(&ext->u.count, 1);
1661                 dname = ext->name;
1662         } else  {
1663                 dname = dentry->d_iname;
1664         }       
1665
1666         dentry->d_name.len = name->len;
1667         dentry->d_name.hash = name->hash;
1668         memcpy(dname, name->name, name->len);
1669         dname[name->len] = 0;
1670
1671         /* Make sure we always see the terminating NUL character */
1672         smp_store_release(&dentry->d_name.name, dname); /* ^^^ */
1673
1674         dentry->d_lockref.count = 1;
1675         dentry->d_flags = 0;
1676         spin_lock_init(&dentry->d_lock);
1677         seqcount_init(&dentry->d_seq);
1678         dentry->d_inode = NULL;
1679         dentry->d_parent = dentry;
1680         dentry->d_sb = sb;
1681         dentry->d_op = NULL;
1682         dentry->d_fsdata = NULL;
1683         INIT_HLIST_BL_NODE(&dentry->d_hash);
1684         INIT_LIST_HEAD(&dentry->d_lru);
1685         INIT_LIST_HEAD(&dentry->d_subdirs);
1686         INIT_HLIST_NODE(&dentry->d_u.d_alias);
1687         INIT_LIST_HEAD(&dentry->d_child);
1688         d_set_d_op(dentry, dentry->d_sb->s_d_op);
1689
1690         if (dentry->d_op && dentry->d_op->d_init) {
1691                 err = dentry->d_op->d_init(dentry);
1692                 if (err) {
1693                         if (dname_external(dentry))
1694                                 kfree(external_name(dentry));
1695                         kmem_cache_free(dentry_cache, dentry);
1696                         return NULL;
1697                 }
1698         }
1699
1700         if (unlikely(ext)) {
1701                 pg_data_t *pgdat = page_pgdat(virt_to_page(ext));
1702                 mod_node_page_state(pgdat, NR_INDIRECTLY_RECLAIMABLE_BYTES,
1703                                     ksize(ext));
1704         }
1705
1706         this_cpu_inc(nr_dentry);
1707
1708         return dentry;
1709 }
1710
1711 /**
1712  * d_alloc      -       allocate a dcache entry
1713  * @parent: parent of entry to allocate
1714  * @name: qstr of the name
1715  *
1716  * Allocates a dentry. It returns %NULL if there is insufficient memory
1717  * available. On a success the dentry is returned. The name passed in is
1718  * copied and the copy passed in may be reused after this call.
1719  */
1720 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1721 {
1722         struct dentry *dentry = __d_alloc(parent->d_sb, name);
1723         if (!dentry)
1724                 return NULL;
1725         dentry->d_flags |= DCACHE_RCUACCESS;
1726         spin_lock(&parent->d_lock);
1727         /*
1728          * don't need child lock because it is not subject
1729          * to concurrency here
1730          */
1731         __dget_dlock(parent);
1732         dentry->d_parent = parent;
1733         list_add(&dentry->d_child, &parent->d_subdirs);
1734         spin_unlock(&parent->d_lock);
1735
1736         return dentry;
1737 }
1738 EXPORT_SYMBOL(d_alloc);
1739
1740 struct dentry *d_alloc_anon(struct super_block *sb)
1741 {
1742         return __d_alloc(sb, NULL);
1743 }
1744 EXPORT_SYMBOL(d_alloc_anon);
1745
1746 struct dentry *d_alloc_cursor(struct dentry * parent)
1747 {
1748         struct dentry *dentry = d_alloc_anon(parent->d_sb);
1749         if (dentry) {
1750                 dentry->d_flags |= DCACHE_RCUACCESS | DCACHE_DENTRY_CURSOR;
1751                 dentry->d_parent = dget(parent);
1752         }
1753         return dentry;
1754 }
1755
1756 /**
1757  * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1758  * @sb: the superblock
1759  * @name: qstr of the name
1760  *
1761  * For a filesystem that just pins its dentries in memory and never
1762  * performs lookups at all, return an unhashed IS_ROOT dentry.
1763  */
1764 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1765 {
1766         return __d_alloc(sb, name);
1767 }
1768 EXPORT_SYMBOL(d_alloc_pseudo);
1769
1770 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1771 {
1772         struct qstr q;
1773
1774         q.name = name;
1775         q.hash_len = hashlen_string(parent, name);
1776         return d_alloc(parent, &q);
1777 }
1778 EXPORT_SYMBOL(d_alloc_name);
1779
1780 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1781 {
1782         WARN_ON_ONCE(dentry->d_op);
1783         WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH  |
1784                                 DCACHE_OP_COMPARE       |
1785                                 DCACHE_OP_REVALIDATE    |
1786                                 DCACHE_OP_WEAK_REVALIDATE       |
1787                                 DCACHE_OP_DELETE        |
1788                                 DCACHE_OP_REAL));
1789         dentry->d_op = op;
1790         if (!op)
1791                 return;
1792         if (op->d_hash)
1793                 dentry->d_flags |= DCACHE_OP_HASH;
1794         if (op->d_compare)
1795                 dentry->d_flags |= DCACHE_OP_COMPARE;
1796         if (op->d_revalidate)
1797                 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1798         if (op->d_weak_revalidate)
1799                 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1800         if (op->d_delete)
1801                 dentry->d_flags |= DCACHE_OP_DELETE;
1802         if (op->d_prune)
1803                 dentry->d_flags |= DCACHE_OP_PRUNE;
1804         if (op->d_real)
1805                 dentry->d_flags |= DCACHE_OP_REAL;
1806
1807 }
1808 EXPORT_SYMBOL(d_set_d_op);
1809
1810
1811 /*
1812  * d_set_fallthru - Mark a dentry as falling through to a lower layer
1813  * @dentry - The dentry to mark
1814  *
1815  * Mark a dentry as falling through to the lower layer (as set with
1816  * d_pin_lower()).  This flag may be recorded on the medium.
1817  */
1818 void d_set_fallthru(struct dentry *dentry)
1819 {
1820         spin_lock(&dentry->d_lock);
1821         dentry->d_flags |= DCACHE_FALLTHRU;
1822         spin_unlock(&dentry->d_lock);
1823 }
1824 EXPORT_SYMBOL(d_set_fallthru);
1825
1826 static unsigned d_flags_for_inode(struct inode *inode)
1827 {
1828         unsigned add_flags = DCACHE_REGULAR_TYPE;
1829
1830         if (!inode)
1831                 return DCACHE_MISS_TYPE;
1832
1833         if (S_ISDIR(inode->i_mode)) {
1834                 add_flags = DCACHE_DIRECTORY_TYPE;
1835                 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1836                         if (unlikely(!inode->i_op->lookup))
1837                                 add_flags = DCACHE_AUTODIR_TYPE;
1838                         else
1839                                 inode->i_opflags |= IOP_LOOKUP;
1840                 }
1841                 goto type_determined;
1842         }
1843
1844         if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1845                 if (unlikely(inode->i_op->get_link)) {
1846                         add_flags = DCACHE_SYMLINK_TYPE;
1847                         goto type_determined;
1848                 }
1849                 inode->i_opflags |= IOP_NOFOLLOW;
1850         }
1851
1852         if (unlikely(!S_ISREG(inode->i_mode)))
1853                 add_flags = DCACHE_SPECIAL_TYPE;
1854
1855 type_determined:
1856         if (unlikely(IS_AUTOMOUNT(inode)))
1857                 add_flags |= DCACHE_NEED_AUTOMOUNT;
1858         return add_flags;
1859 }
1860
1861 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1862 {
1863         unsigned add_flags = d_flags_for_inode(inode);
1864         WARN_ON(d_in_lookup(dentry));
1865
1866         spin_lock(&dentry->d_lock);
1867         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1868         raw_write_seqcount_begin(&dentry->d_seq);
1869         __d_set_inode_and_type(dentry, inode, add_flags);
1870         raw_write_seqcount_end(&dentry->d_seq);
1871         fsnotify_update_flags(dentry);
1872         spin_unlock(&dentry->d_lock);
1873 }
1874
1875 /**
1876  * d_instantiate - fill in inode information for a dentry
1877  * @entry: dentry to complete
1878  * @inode: inode to attach to this dentry
1879  *
1880  * Fill in inode information in the entry.
1881  *
1882  * This turns negative dentries into productive full members
1883  * of society.
1884  *
1885  * NOTE! This assumes that the inode count has been incremented
1886  * (or otherwise set) by the caller to indicate that it is now
1887  * in use by the dcache.
1888  */
1889  
1890 void d_instantiate(struct dentry *entry, struct inode * inode)
1891 {
1892         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1893         if (inode) {
1894                 security_d_instantiate(entry, inode);
1895                 spin_lock(&inode->i_lock);
1896                 __d_instantiate(entry, inode);
1897                 spin_unlock(&inode->i_lock);
1898         }
1899 }
1900 EXPORT_SYMBOL(d_instantiate);
1901
1902 /**
1903  * d_instantiate_no_diralias - instantiate a non-aliased dentry
1904  * @entry: dentry to complete
1905  * @inode: inode to attach to this dentry
1906  *
1907  * Fill in inode information in the entry.  If a directory alias is found, then
1908  * return an error (and drop inode).  Together with d_materialise_unique() this
1909  * guarantees that a directory inode may never have more than one alias.
1910  */
1911 int d_instantiate_no_diralias(struct dentry *entry, struct inode *inode)
1912 {
1913         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1914
1915         security_d_instantiate(entry, inode);
1916         spin_lock(&inode->i_lock);
1917         if (S_ISDIR(inode->i_mode) && !hlist_empty(&inode->i_dentry)) {
1918                 spin_unlock(&inode->i_lock);
1919                 iput(inode);
1920                 return -EBUSY;
1921         }
1922         __d_instantiate(entry, inode);
1923         spin_unlock(&inode->i_lock);
1924
1925         return 0;
1926 }
1927 EXPORT_SYMBOL(d_instantiate_no_diralias);
1928
1929 struct dentry *d_make_root(struct inode *root_inode)
1930 {
1931         struct dentry *res = NULL;
1932
1933         if (root_inode) {
1934                 res = d_alloc_anon(root_inode->i_sb);
1935                 if (res)
1936                         d_instantiate(res, root_inode);
1937                 else
1938                         iput(root_inode);
1939         }
1940         return res;
1941 }
1942 EXPORT_SYMBOL(d_make_root);
1943
1944 static struct dentry * __d_find_any_alias(struct inode *inode)
1945 {
1946         struct dentry *alias;
1947
1948         if (hlist_empty(&inode->i_dentry))
1949                 return NULL;
1950         alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
1951         __dget(alias);
1952         return alias;
1953 }
1954
1955 /**
1956  * d_find_any_alias - find any alias for a given inode
1957  * @inode: inode to find an alias for
1958  *
1959  * If any aliases exist for the given inode, take and return a
1960  * reference for one of them.  If no aliases exist, return %NULL.
1961  */
1962 struct dentry *d_find_any_alias(struct inode *inode)
1963 {
1964         struct dentry *de;
1965
1966         spin_lock(&inode->i_lock);
1967         de = __d_find_any_alias(inode);
1968         spin_unlock(&inode->i_lock);
1969         return de;
1970 }
1971 EXPORT_SYMBOL(d_find_any_alias);
1972
1973 static struct dentry *__d_instantiate_anon(struct dentry *dentry,
1974                                            struct inode *inode,
1975                                            bool disconnected)
1976 {
1977         struct dentry *res;
1978         unsigned add_flags;
1979
1980         security_d_instantiate(dentry, inode);
1981         spin_lock(&inode->i_lock);
1982         res = __d_find_any_alias(inode);
1983         if (res) {
1984                 spin_unlock(&inode->i_lock);
1985                 dput(dentry);
1986                 goto out_iput;
1987         }
1988
1989         /* attach a disconnected dentry */
1990         add_flags = d_flags_for_inode(inode);
1991
1992         if (disconnected)
1993                 add_flags |= DCACHE_DISCONNECTED;
1994
1995         spin_lock(&dentry->d_lock);
1996         __d_set_inode_and_type(dentry, inode, add_flags);
1997         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1998         if (!disconnected) {
1999                 hlist_bl_lock(&dentry->d_sb->s_roots);
2000                 hlist_bl_add_head(&dentry->d_hash, &dentry->d_sb->s_roots);
2001                 hlist_bl_unlock(&dentry->d_sb->s_roots);
2002         }
2003         spin_unlock(&dentry->d_lock);
2004         spin_unlock(&inode->i_lock);
2005
2006         return dentry;
2007
2008  out_iput:
2009         iput(inode);
2010         return res;
2011 }
2012
2013 struct dentry *d_instantiate_anon(struct dentry *dentry, struct inode *inode)
2014 {
2015         return __d_instantiate_anon(dentry, inode, true);
2016 }
2017 EXPORT_SYMBOL(d_instantiate_anon);
2018
2019 static struct dentry *__d_obtain_alias(struct inode *inode, bool disconnected)
2020 {
2021         struct dentry *tmp;
2022         struct dentry *res;
2023
2024         if (!inode)
2025                 return ERR_PTR(-ESTALE);
2026         if (IS_ERR(inode))
2027                 return ERR_CAST(inode);
2028
2029         res = d_find_any_alias(inode);
2030         if (res)
2031                 goto out_iput;
2032
2033         tmp = d_alloc_anon(inode->i_sb);
2034         if (!tmp) {
2035                 res = ERR_PTR(-ENOMEM);
2036                 goto out_iput;
2037         }
2038
2039         return __d_instantiate_anon(tmp, inode, disconnected);
2040
2041 out_iput:
2042         iput(inode);
2043         return res;
2044 }
2045
2046 /**
2047  * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
2048  * @inode: inode to allocate the dentry for
2049  *
2050  * Obtain a dentry for an inode resulting from NFS filehandle conversion or
2051  * similar open by handle operations.  The returned dentry may be anonymous,
2052  * or may have a full name (if the inode was already in the cache).
2053  *
2054  * When called on a directory inode, we must ensure that the inode only ever
2055  * has one dentry.  If a dentry is found, that is returned instead of
2056  * allocating a new one.
2057  *
2058  * On successful return, the reference to the inode has been transferred
2059  * to the dentry.  In case of an error the reference on the inode is released.
2060  * To make it easier to use in export operations a %NULL or IS_ERR inode may
2061  * be passed in and the error will be propagated to the return value,
2062  * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
2063  */
2064 struct dentry *d_obtain_alias(struct inode *inode)
2065 {
2066         return __d_obtain_alias(inode, true);
2067 }
2068 EXPORT_SYMBOL(d_obtain_alias);
2069
2070 /**
2071  * d_obtain_root - find or allocate a dentry for a given inode
2072  * @inode: inode to allocate the dentry for
2073  *
2074  * Obtain an IS_ROOT dentry for the root of a filesystem.
2075  *
2076  * We must ensure that directory inodes only ever have one dentry.  If a
2077  * dentry is found, that is returned instead of allocating a new one.
2078  *
2079  * On successful return, the reference to the inode has been transferred
2080  * to the dentry.  In case of an error the reference on the inode is
2081  * released.  A %NULL or IS_ERR inode may be passed in and will be the
2082  * error will be propagate to the return value, with a %NULL @inode
2083  * replaced by ERR_PTR(-ESTALE).
2084  */
2085 struct dentry *d_obtain_root(struct inode *inode)
2086 {
2087         return __d_obtain_alias(inode, false);
2088 }
2089 EXPORT_SYMBOL(d_obtain_root);
2090
2091 /**
2092  * d_add_ci - lookup or allocate new dentry with case-exact name
2093  * @inode:  the inode case-insensitive lookup has found
2094  * @dentry: the negative dentry that was passed to the parent's lookup func
2095  * @name:   the case-exact name to be associated with the returned dentry
2096  *
2097  * This is to avoid filling the dcache with case-insensitive names to the
2098  * same inode, only the actual correct case is stored in the dcache for
2099  * case-insensitive filesystems.
2100  *
2101  * For a case-insensitive lookup match and if the the case-exact dentry
2102  * already exists in in the dcache, use it and return it.
2103  *
2104  * If no entry exists with the exact case name, allocate new dentry with
2105  * the exact case, and return the spliced entry.
2106  */
2107 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2108                         struct qstr *name)
2109 {
2110         struct dentry *found, *res;
2111
2112         /*
2113          * First check if a dentry matching the name already exists,
2114          * if not go ahead and create it now.
2115          */
2116         found = d_hash_and_lookup(dentry->d_parent, name);
2117         if (found) {
2118                 iput(inode);
2119                 return found;
2120         }
2121         if (d_in_lookup(dentry)) {
2122                 found = d_alloc_parallel(dentry->d_parent, name,
2123                                         dentry->d_wait);
2124                 if (IS_ERR(found) || !d_in_lookup(found)) {
2125                         iput(inode);
2126                         return found;
2127                 }
2128         } else {
2129                 found = d_alloc(dentry->d_parent, name);
2130                 if (!found) {
2131                         iput(inode);
2132                         return ERR_PTR(-ENOMEM);
2133                 } 
2134         }
2135         res = d_splice_alias(inode, found);
2136         if (res) {
2137                 dput(found);
2138                 return res;
2139         }
2140         return found;
2141 }
2142 EXPORT_SYMBOL(d_add_ci);
2143
2144
2145 static inline bool d_same_name(const struct dentry *dentry,
2146                                 const struct dentry *parent,
2147                                 const struct qstr *name)
2148 {
2149         if (likely(!(parent->d_flags & DCACHE_OP_COMPARE))) {
2150                 if (dentry->d_name.len != name->len)
2151                         return false;
2152                 return dentry_cmp(dentry, name->name, name->len) == 0;
2153         }
2154         return parent->d_op->d_compare(dentry,
2155                                        dentry->d_name.len, dentry->d_name.name,
2156                                        name) == 0;
2157 }
2158
2159 /**
2160  * __d_lookup_rcu - search for a dentry (racy, store-free)
2161  * @parent: parent dentry
2162  * @name: qstr of name we wish to find
2163  * @seqp: returns d_seq value at the point where the dentry was found
2164  * Returns: dentry, or NULL
2165  *
2166  * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2167  * resolution (store-free path walking) design described in
2168  * Documentation/filesystems/path-lookup.txt.
2169  *
2170  * This is not to be used outside core vfs.
2171  *
2172  * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2173  * held, and rcu_read_lock held. The returned dentry must not be stored into
2174  * without taking d_lock and checking d_seq sequence count against @seq
2175  * returned here.
2176  *
2177  * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2178  * function.
2179  *
2180  * Alternatively, __d_lookup_rcu may be called again to look up the child of
2181  * the returned dentry, so long as its parent's seqlock is checked after the
2182  * child is looked up. Thus, an interlocking stepping of sequence lock checks
2183  * is formed, giving integrity down the path walk.
2184  *
2185  * NOTE! The caller *has* to check the resulting dentry against the sequence
2186  * number we've returned before using any of the resulting dentry state!
2187  */
2188 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2189                                 const struct qstr *name,
2190                                 unsigned *seqp)
2191 {
2192         u64 hashlen = name->hash_len;
2193         const unsigned char *str = name->name;
2194         struct hlist_bl_head *b = d_hash(hashlen_hash(hashlen));
2195         struct hlist_bl_node *node;
2196         struct dentry *dentry;
2197
2198         /*
2199          * Note: There is significant duplication with __d_lookup_rcu which is
2200          * required to prevent single threaded performance regressions
2201          * especially on architectures where smp_rmb (in seqcounts) are costly.
2202          * Keep the two functions in sync.
2203          */
2204
2205         /*
2206          * The hash list is protected using RCU.
2207          *
2208          * Carefully use d_seq when comparing a candidate dentry, to avoid
2209          * races with d_move().
2210          *
2211          * It is possible that concurrent renames can mess up our list
2212          * walk here and result in missing our dentry, resulting in the
2213          * false-negative result. d_lookup() protects against concurrent
2214          * renames using rename_lock seqlock.
2215          *
2216          * See Documentation/filesystems/path-lookup.txt for more details.
2217          */
2218         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2219                 unsigned seq;
2220
2221 seqretry:
2222                 /*
2223                  * The dentry sequence count protects us from concurrent
2224                  * renames, and thus protects parent and name fields.
2225                  *
2226                  * The caller must perform a seqcount check in order
2227                  * to do anything useful with the returned dentry.
2228                  *
2229                  * NOTE! We do a "raw" seqcount_begin here. That means that
2230                  * we don't wait for the sequence count to stabilize if it
2231                  * is in the middle of a sequence change. If we do the slow
2232                  * dentry compare, we will do seqretries until it is stable,
2233                  * and if we end up with a successful lookup, we actually
2234                  * want to exit RCU lookup anyway.
2235                  *
2236                  * Note that raw_seqcount_begin still *does* smp_rmb(), so
2237                  * we are still guaranteed NUL-termination of ->d_name.name.
2238                  */
2239                 seq = raw_seqcount_begin(&dentry->d_seq);
2240                 if (dentry->d_parent != parent)
2241                         continue;
2242                 if (d_unhashed(dentry))
2243                         continue;
2244
2245                 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) {
2246                         int tlen;
2247                         const char *tname;
2248                         if (dentry->d_name.hash != hashlen_hash(hashlen))
2249                                 continue;
2250                         tlen = dentry->d_name.len;
2251                         tname = dentry->d_name.name;
2252                         /* we want a consistent (name,len) pair */
2253                         if (read_seqcount_retry(&dentry->d_seq, seq)) {
2254                                 cpu_relax();
2255                                 goto seqretry;
2256                         }
2257                         if (parent->d_op->d_compare(dentry,
2258                                                     tlen, tname, name) != 0)
2259                                 continue;
2260                 } else {
2261                         if (dentry->d_name.hash_len != hashlen)
2262                                 continue;
2263                         if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
2264                                 continue;
2265                 }
2266                 *seqp = seq;
2267                 return dentry;
2268         }
2269         return NULL;
2270 }
2271
2272 /**
2273  * d_lookup - search for a dentry
2274  * @parent: parent dentry
2275  * @name: qstr of name we wish to find
2276  * Returns: dentry, or NULL
2277  *
2278  * d_lookup searches the children of the parent dentry for the name in
2279  * question. If the dentry is found its reference count is incremented and the
2280  * dentry is returned. The caller must use dput to free the entry when it has
2281  * finished using it. %NULL is returned if the dentry does not exist.
2282  */
2283 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2284 {
2285         struct dentry *dentry;
2286         unsigned seq;
2287
2288         do {
2289                 seq = read_seqbegin(&rename_lock);
2290                 dentry = __d_lookup(parent, name);
2291                 if (dentry)
2292                         break;
2293         } while (read_seqretry(&rename_lock, seq));
2294         return dentry;
2295 }
2296 EXPORT_SYMBOL(d_lookup);
2297
2298 /**
2299  * __d_lookup - search for a dentry (racy)
2300  * @parent: parent dentry
2301  * @name: qstr of name we wish to find
2302  * Returns: dentry, or NULL
2303  *
2304  * __d_lookup is like d_lookup, however it may (rarely) return a
2305  * false-negative result due to unrelated rename activity.
2306  *
2307  * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2308  * however it must be used carefully, eg. with a following d_lookup in
2309  * the case of failure.
2310  *
2311  * __d_lookup callers must be commented.
2312  */
2313 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2314 {
2315         unsigned int hash = name->hash;
2316         struct hlist_bl_head *b = d_hash(hash);
2317         struct hlist_bl_node *node;
2318         struct dentry *found = NULL;
2319         struct dentry *dentry;
2320
2321         /*
2322          * Note: There is significant duplication with __d_lookup_rcu which is
2323          * required to prevent single threaded performance regressions
2324          * especially on architectures where smp_rmb (in seqcounts) are costly.
2325          * Keep the two functions in sync.
2326          */
2327
2328         /*
2329          * The hash list is protected using RCU.
2330          *
2331          * Take d_lock when comparing a candidate dentry, to avoid races
2332          * with d_move().
2333          *
2334          * It is possible that concurrent renames can mess up our list
2335          * walk here and result in missing our dentry, resulting in the
2336          * false-negative result. d_lookup() protects against concurrent
2337          * renames using rename_lock seqlock.
2338          *
2339          * See Documentation/filesystems/path-lookup.txt for more details.
2340          */
2341         rcu_read_lock();
2342         
2343         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2344
2345                 if (dentry->d_name.hash != hash)
2346                         continue;
2347
2348                 spin_lock(&dentry->d_lock);
2349                 if (dentry->d_parent != parent)
2350                         goto next;
2351                 if (d_unhashed(dentry))
2352                         goto next;
2353
2354                 if (!d_same_name(dentry, parent, name))
2355                         goto next;
2356
2357                 dentry->d_lockref.count++;
2358                 found = dentry;
2359                 spin_unlock(&dentry->d_lock);
2360                 break;
2361 next:
2362                 spin_unlock(&dentry->d_lock);
2363         }
2364         rcu_read_unlock();
2365
2366         return found;
2367 }
2368
2369 /**
2370  * d_hash_and_lookup - hash the qstr then search for a dentry
2371  * @dir: Directory to search in
2372  * @name: qstr of name we wish to find
2373  *
2374  * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2375  */
2376 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2377 {
2378         /*
2379          * Check for a fs-specific hash function. Note that we must
2380          * calculate the standard hash first, as the d_op->d_hash()
2381          * routine may choose to leave the hash value unchanged.
2382          */
2383         name->hash = full_name_hash(dir, name->name, name->len);
2384         if (dir->d_flags & DCACHE_OP_HASH) {
2385                 int err = dir->d_op->d_hash(dir, name);
2386                 if (unlikely(err < 0))
2387                         return ERR_PTR(err);
2388         }
2389         return d_lookup(dir, name);
2390 }
2391 EXPORT_SYMBOL(d_hash_and_lookup);
2392
2393 /*
2394  * When a file is deleted, we have two options:
2395  * - turn this dentry into a negative dentry
2396  * - unhash this dentry and free it.
2397  *
2398  * Usually, we want to just turn this into
2399  * a negative dentry, but if anybody else is
2400  * currently using the dentry or the inode
2401  * we can't do that and we fall back on removing
2402  * it from the hash queues and waiting for
2403  * it to be deleted later when it has no users
2404  */
2405  
2406 /**
2407  * d_delete - delete a dentry
2408  * @dentry: The dentry to delete
2409  *
2410  * Turn the dentry into a negative dentry if possible, otherwise
2411  * remove it from the hash queues so it can be deleted later
2412  */
2413  
2414 void d_delete(struct dentry * dentry)
2415 {
2416         struct inode *inode = dentry->d_inode;
2417         int isdir = d_is_dir(dentry);
2418
2419         spin_lock(&inode->i_lock);
2420         spin_lock(&dentry->d_lock);
2421         /*
2422          * Are we the only user?
2423          */
2424         if (dentry->d_lockref.count == 1) {
2425                 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2426                 dentry_unlink_inode(dentry);
2427         } else {
2428                 __d_drop(dentry);
2429                 spin_unlock(&dentry->d_lock);
2430                 spin_unlock(&inode->i_lock);
2431         }
2432         fsnotify_nameremove(dentry, isdir);
2433 }
2434 EXPORT_SYMBOL(d_delete);
2435
2436 static void __d_rehash(struct dentry *entry)
2437 {
2438         struct hlist_bl_head *b = d_hash(entry->d_name.hash);
2439
2440         hlist_bl_lock(b);
2441         hlist_bl_add_head_rcu(&entry->d_hash, b);
2442         hlist_bl_unlock(b);
2443 }
2444
2445 /**
2446  * d_rehash     - add an entry back to the hash
2447  * @entry: dentry to add to the hash
2448  *
2449  * Adds a dentry to the hash according to its name.
2450  */
2451  
2452 void d_rehash(struct dentry * entry)
2453 {
2454         spin_lock(&entry->d_lock);
2455         __d_rehash(entry);
2456         spin_unlock(&entry->d_lock);
2457 }
2458 EXPORT_SYMBOL(d_rehash);
2459
2460 static inline unsigned start_dir_add(struct inode *dir)
2461 {
2462
2463         for (;;) {
2464                 unsigned n = dir->i_dir_seq;
2465                 if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
2466                         return n;
2467                 cpu_relax();
2468         }
2469 }
2470
2471 static inline void end_dir_add(struct inode *dir, unsigned n)
2472 {
2473         smp_store_release(&dir->i_dir_seq, n + 2);
2474 }
2475
2476 static void d_wait_lookup(struct dentry *dentry)
2477 {
2478         if (d_in_lookup(dentry)) {
2479                 DECLARE_WAITQUEUE(wait, current);
2480                 add_wait_queue(dentry->d_wait, &wait);
2481                 do {
2482                         set_current_state(TASK_UNINTERRUPTIBLE);
2483                         spin_unlock(&dentry->d_lock);
2484                         schedule();
2485                         spin_lock(&dentry->d_lock);
2486                 } while (d_in_lookup(dentry));
2487         }
2488 }
2489
2490 struct dentry *d_alloc_parallel(struct dentry *parent,
2491                                 const struct qstr *name,
2492                                 wait_queue_head_t *wq)
2493 {
2494         unsigned int hash = name->hash;
2495         struct hlist_bl_head *b = in_lookup_hash(parent, hash);
2496         struct hlist_bl_node *node;
2497         struct dentry *new = d_alloc(parent, name);
2498         struct dentry *dentry;
2499         unsigned seq, r_seq, d_seq;
2500
2501         if (unlikely(!new))
2502                 return ERR_PTR(-ENOMEM);
2503
2504 retry:
2505         rcu_read_lock();
2506         seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
2507         r_seq = read_seqbegin(&rename_lock);
2508         dentry = __d_lookup_rcu(parent, name, &d_seq);
2509         if (unlikely(dentry)) {
2510                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2511                         rcu_read_unlock();
2512                         goto retry;
2513                 }
2514                 if (read_seqcount_retry(&dentry->d_seq, d_seq)) {
2515                         rcu_read_unlock();
2516                         dput(dentry);
2517                         goto retry;
2518                 }
2519                 rcu_read_unlock();
2520                 dput(new);
2521                 return dentry;
2522         }
2523         if (unlikely(read_seqretry(&rename_lock, r_seq))) {
2524                 rcu_read_unlock();
2525                 goto retry;
2526         }
2527
2528         if (unlikely(seq & 1)) {
2529                 rcu_read_unlock();
2530                 goto retry;
2531         }
2532
2533         hlist_bl_lock(b);
2534         if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {
2535                 hlist_bl_unlock(b);
2536                 rcu_read_unlock();
2537                 goto retry;
2538         }
2539         /*
2540          * No changes for the parent since the beginning of d_lookup().
2541          * Since all removals from the chain happen with hlist_bl_lock(),
2542          * any potential in-lookup matches are going to stay here until
2543          * we unlock the chain.  All fields are stable in everything
2544          * we encounter.
2545          */
2546         hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
2547                 if (dentry->d_name.hash != hash)
2548                         continue;
2549                 if (dentry->d_parent != parent)
2550                         continue;
2551                 if (!d_same_name(dentry, parent, name))
2552                         continue;
2553                 hlist_bl_unlock(b);
2554                 /* now we can try to grab a reference */
2555                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2556                         rcu_read_unlock();
2557                         goto retry;
2558                 }
2559
2560                 rcu_read_unlock();
2561                 /*
2562                  * somebody is likely to be still doing lookup for it;
2563                  * wait for them to finish
2564                  */
2565                 spin_lock(&dentry->d_lock);
2566                 d_wait_lookup(dentry);
2567                 /*
2568                  * it's not in-lookup anymore; in principle we should repeat
2569                  * everything from dcache lookup, but it's likely to be what
2570                  * d_lookup() would've found anyway.  If it is, just return it;
2571                  * otherwise we really have to repeat the whole thing.
2572                  */
2573                 if (unlikely(dentry->d_name.hash != hash))
2574                         goto mismatch;
2575                 if (unlikely(dentry->d_parent != parent))
2576                         goto mismatch;
2577                 if (unlikely(d_unhashed(dentry)))
2578                         goto mismatch;
2579                 if (unlikely(!d_same_name(dentry, parent, name)))
2580                         goto mismatch;
2581                 /* OK, it *is* a hashed match; return it */
2582                 spin_unlock(&dentry->d_lock);
2583                 dput(new);
2584                 return dentry;
2585         }
2586         rcu_read_unlock();
2587         /* we can't take ->d_lock here; it's OK, though. */
2588         new->d_flags |= DCACHE_PAR_LOOKUP;
2589         new->d_wait = wq;
2590         hlist_bl_add_head_rcu(&new->d_u.d_in_lookup_hash, b);
2591         hlist_bl_unlock(b);
2592         return new;
2593 mismatch:
2594         spin_unlock(&dentry->d_lock);
2595         dput(dentry);
2596         goto retry;
2597 }
2598 EXPORT_SYMBOL(d_alloc_parallel);
2599
2600 void __d_lookup_done(struct dentry *dentry)
2601 {
2602         struct hlist_bl_head *b = in_lookup_hash(dentry->d_parent,
2603                                                  dentry->d_name.hash);
2604         hlist_bl_lock(b);
2605         dentry->d_flags &= ~DCACHE_PAR_LOOKUP;
2606         __hlist_bl_del(&dentry->d_u.d_in_lookup_hash);
2607         wake_up_all(dentry->d_wait);
2608         dentry->d_wait = NULL;
2609         hlist_bl_unlock(b);
2610         INIT_HLIST_NODE(&dentry->d_u.d_alias);
2611         INIT_LIST_HEAD(&dentry->d_lru);
2612 }
2613 EXPORT_SYMBOL(__d_lookup_done);
2614
2615 /* inode->i_lock held if inode is non-NULL */
2616
2617 static inline void __d_add(struct dentry *dentry, struct inode *inode)
2618 {
2619         struct inode *dir = NULL;
2620         unsigned n;
2621         spin_lock(&dentry->d_lock);
2622         if (unlikely(d_in_lookup(dentry))) {
2623                 dir = dentry->d_parent->d_inode;
2624                 n = start_dir_add(dir);
2625                 __d_lookup_done(dentry);
2626         }
2627         if (inode) {
2628                 unsigned add_flags = d_flags_for_inode(inode);
2629                 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2630                 raw_write_seqcount_begin(&dentry->d_seq);
2631                 __d_set_inode_and_type(dentry, inode, add_flags);
2632                 raw_write_seqcount_end(&dentry->d_seq);
2633                 fsnotify_update_flags(dentry);
2634         }
2635         __d_rehash(dentry);
2636         if (dir)
2637                 end_dir_add(dir, n);
2638         spin_unlock(&dentry->d_lock);
2639         if (inode)
2640                 spin_unlock(&inode->i_lock);
2641 }
2642
2643 /**
2644  * d_add - add dentry to hash queues
2645  * @entry: dentry to add
2646  * @inode: The inode to attach to this dentry
2647  *
2648  * This adds the entry to the hash queues and initializes @inode.
2649  * The entry was actually filled in earlier during d_alloc().
2650  */
2651
2652 void d_add(struct dentry *entry, struct inode *inode)
2653 {
2654         if (inode) {
2655                 security_d_instantiate(entry, inode);
2656                 spin_lock(&inode->i_lock);
2657         }
2658         __d_add(entry, inode);
2659 }
2660 EXPORT_SYMBOL(d_add);
2661
2662 /**
2663  * d_exact_alias - find and hash an exact unhashed alias
2664  * @entry: dentry to add
2665  * @inode: The inode to go with this dentry
2666  *
2667  * If an unhashed dentry with the same name/parent and desired
2668  * inode already exists, hash and return it.  Otherwise, return
2669  * NULL.
2670  *
2671  * Parent directory should be locked.
2672  */
2673 struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2674 {
2675         struct dentry *alias;
2676         unsigned int hash = entry->d_name.hash;
2677
2678         spin_lock(&inode->i_lock);
2679         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2680                 /*
2681                  * Don't need alias->d_lock here, because aliases with
2682                  * d_parent == entry->d_parent are not subject to name or
2683                  * parent changes, because the parent inode i_mutex is held.
2684                  */
2685                 if (alias->d_name.hash != hash)
2686                         continue;
2687                 if (alias->d_parent != entry->d_parent)
2688                         continue;
2689                 if (!d_same_name(alias, entry->d_parent, &entry->d_name))
2690                         continue;
2691                 spin_lock(&alias->d_lock);
2692                 if (!d_unhashed(alias)) {
2693                         spin_unlock(&alias->d_lock);
2694                         alias = NULL;
2695                 } else {
2696                         __dget_dlock(alias);
2697                         __d_rehash(alias);
2698                         spin_unlock(&alias->d_lock);
2699                 }
2700                 spin_unlock(&inode->i_lock);
2701                 return alias;
2702         }
2703         spin_unlock(&inode->i_lock);
2704         return NULL;
2705 }
2706 EXPORT_SYMBOL(d_exact_alias);
2707
2708 /**
2709  * dentry_update_name_case - update case insensitive dentry with a new name
2710  * @dentry: dentry to be updated
2711  * @name: new name
2712  *
2713  * Update a case insensitive dentry with new case of name.
2714  *
2715  * dentry must have been returned by d_lookup with name @name. Old and new
2716  * name lengths must match (ie. no d_compare which allows mismatched name
2717  * lengths).
2718  *
2719  * Parent inode i_mutex must be held over d_lookup and into this call (to
2720  * keep renames and concurrent inserts, and readdir(2) away).
2721  */
2722 void dentry_update_name_case(struct dentry *dentry, const struct qstr *name)
2723 {
2724         BUG_ON(!inode_is_locked(dentry->d_parent->d_inode));
2725         BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */
2726
2727         spin_lock(&dentry->d_lock);
2728         write_seqcount_begin(&dentry->d_seq);
2729         memcpy((unsigned char *)dentry->d_name.name, name->name, name->len);
2730         write_seqcount_end(&dentry->d_seq);
2731         spin_unlock(&dentry->d_lock);
2732 }
2733 EXPORT_SYMBOL(dentry_update_name_case);
2734
2735 static void swap_names(struct dentry *dentry, struct dentry *target)
2736 {
2737         if (unlikely(dname_external(target))) {
2738                 if (unlikely(dname_external(dentry))) {
2739                         /*
2740                          * Both external: swap the pointers
2741                          */
2742                         swap(target->d_name.name, dentry->d_name.name);
2743                 } else {
2744                         /*
2745                          * dentry:internal, target:external.  Steal target's
2746                          * storage and make target internal.
2747                          */
2748                         memcpy(target->d_iname, dentry->d_name.name,
2749                                         dentry->d_name.len + 1);
2750                         dentry->d_name.name = target->d_name.name;
2751                         target->d_name.name = target->d_iname;
2752                 }
2753         } else {
2754                 if (unlikely(dname_external(dentry))) {
2755                         /*
2756                          * dentry:external, target:internal.  Give dentry's
2757                          * storage to target and make dentry internal
2758                          */
2759                         memcpy(dentry->d_iname, target->d_name.name,
2760                                         target->d_name.len + 1);
2761                         target->d_name.name = dentry->d_name.name;
2762                         dentry->d_name.name = dentry->d_iname;
2763                 } else {
2764                         /*
2765                          * Both are internal.
2766                          */
2767                         unsigned int i;
2768                         BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2769                         for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2770                                 swap(((long *) &dentry->d_iname)[i],
2771                                      ((long *) &target->d_iname)[i]);
2772                         }
2773                 }
2774         }
2775         swap(dentry->d_name.hash_len, target->d_name.hash_len);
2776 }
2777
2778 static void copy_name(struct dentry *dentry, struct dentry *target)
2779 {
2780         struct external_name *old_name = NULL;
2781         if (unlikely(dname_external(dentry)))
2782                 old_name = external_name(dentry);
2783         if (unlikely(dname_external(target))) {
2784                 atomic_inc(&external_name(target)->u.count);
2785                 dentry->d_name = target->d_name;
2786         } else {
2787                 memcpy(dentry->d_iname, target->d_name.name,
2788                                 target->d_name.len + 1);
2789                 dentry->d_name.name = dentry->d_iname;
2790                 dentry->d_name.hash_len = target->d_name.hash_len;
2791         }
2792         if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2793                 call_rcu(&old_name->u.head, __d_free_external_name);
2794 }
2795
2796 /*
2797  * __d_move - move a dentry
2798  * @dentry: entry to move
2799  * @target: new dentry
2800  * @exchange: exchange the two dentries
2801  *
2802  * Update the dcache to reflect the move of a file name. Negative
2803  * dcache entries should not be moved in this way. Caller must hold
2804  * rename_lock, the i_mutex of the source and target directories,
2805  * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2806  */
2807 static void __d_move(struct dentry *dentry, struct dentry *target,
2808                      bool exchange)
2809 {
2810         struct dentry *old_parent, *p;
2811         struct inode *dir = NULL;
2812         unsigned n;
2813
2814         WARN_ON(!dentry->d_inode);
2815         if (WARN_ON(dentry == target))
2816                 return;
2817
2818         BUG_ON(d_ancestor(target, dentry));
2819         old_parent = dentry->d_parent;
2820         p = d_ancestor(old_parent, target);
2821         if (IS_ROOT(dentry)) {
2822                 BUG_ON(p);
2823                 spin_lock(&target->d_parent->d_lock);
2824         } else if (!p) {
2825                 /* target is not a descendent of dentry->d_parent */
2826                 spin_lock(&target->d_parent->d_lock);
2827                 spin_lock_nested(&old_parent->d_lock, DENTRY_D_LOCK_NESTED);
2828         } else {
2829                 BUG_ON(p == dentry);
2830                 spin_lock(&old_parent->d_lock);
2831                 if (p != target)
2832                         spin_lock_nested(&target->d_parent->d_lock,
2833                                         DENTRY_D_LOCK_NESTED);
2834         }
2835         spin_lock_nested(&dentry->d_lock, 2);
2836         spin_lock_nested(&target->d_lock, 3);
2837
2838         if (unlikely(d_in_lookup(target))) {
2839                 dir = target->d_parent->d_inode;
2840                 n = start_dir_add(dir);
2841                 __d_lookup_done(target);
2842         }
2843
2844         write_seqcount_begin(&dentry->d_seq);
2845         write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2846
2847         /* unhash both */
2848         if (!d_unhashed(dentry))
2849                 ___d_drop(dentry);
2850         if (!d_unhashed(target))
2851                 ___d_drop(target);
2852
2853         /* ... and switch them in the tree */
2854         dentry->d_parent = target->d_parent;
2855         if (!exchange) {
2856                 copy_name(dentry, target);
2857                 target->d_hash.pprev = NULL;
2858                 dentry->d_parent->d_lockref.count++;
2859                 if (dentry == old_parent)
2860                         dentry->d_flags |= DCACHE_RCUACCESS;
2861                 else
2862                         WARN_ON(!--old_parent->d_lockref.count);
2863         } else {
2864                 target->d_parent = old_parent;
2865                 swap_names(dentry, target);
2866                 list_move(&target->d_child, &target->d_parent->d_subdirs);
2867                 __d_rehash(target);
2868                 fsnotify_update_flags(target);
2869         }
2870         list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2871         __d_rehash(dentry);
2872         fsnotify_update_flags(dentry);
2873
2874         write_seqcount_end(&target->d_seq);
2875         write_seqcount_end(&dentry->d_seq);
2876
2877         if (dir)
2878                 end_dir_add(dir, n);
2879
2880         if (dentry->d_parent != old_parent)
2881                 spin_unlock(&dentry->d_parent->d_lock);
2882         if (dentry != old_parent)
2883                 spin_unlock(&old_parent->d_lock);
2884         spin_unlock(&target->d_lock);
2885         spin_unlock(&dentry->d_lock);
2886 }
2887
2888 /*
2889  * d_move - move a dentry
2890  * @dentry: entry to move
2891  * @target: new dentry
2892  *
2893  * Update the dcache to reflect the move of a file name. Negative
2894  * dcache entries should not be moved in this way. See the locking
2895  * requirements for __d_move.
2896  */
2897 void d_move(struct dentry *dentry, struct dentry *target)
2898 {
2899         write_seqlock(&rename_lock);
2900         __d_move(dentry, target, false);
2901         write_sequnlock(&rename_lock);
2902 }
2903 EXPORT_SYMBOL(d_move);
2904
2905 /*
2906  * d_exchange - exchange two dentries
2907  * @dentry1: first dentry
2908  * @dentry2: second dentry
2909  */
2910 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2911 {
2912         write_seqlock(&rename_lock);
2913
2914         WARN_ON(!dentry1->d_inode);
2915         WARN_ON(!dentry2->d_inode);
2916         WARN_ON(IS_ROOT(dentry1));
2917         WARN_ON(IS_ROOT(dentry2));
2918
2919         __d_move(dentry1, dentry2, true);
2920
2921         write_sequnlock(&rename_lock);
2922 }
2923
2924 /**
2925  * d_ancestor - search for an ancestor
2926  * @p1: ancestor dentry
2927  * @p2: child dentry
2928  *
2929  * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2930  * an ancestor of p2, else NULL.
2931  */
2932 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2933 {
2934         struct dentry *p;
2935
2936         for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2937                 if (p->d_parent == p1)
2938                         return p;
2939         }
2940         return NULL;
2941 }
2942
2943 /*
2944  * This helper attempts to cope with remotely renamed directories
2945  *
2946  * It assumes that the caller is already holding
2947  * dentry->d_parent->d_inode->i_mutex, and rename_lock
2948  *
2949  * Note: If ever the locking in lock_rename() changes, then please
2950  * remember to update this too...
2951  */
2952 static int __d_unalias(struct inode *inode,
2953                 struct dentry *dentry, struct dentry *alias)
2954 {
2955         struct mutex *m1 = NULL;
2956         struct rw_semaphore *m2 = NULL;
2957         int ret = -ESTALE;
2958
2959         /* If alias and dentry share a parent, then no extra locks required */
2960         if (alias->d_parent == dentry->d_parent)
2961                 goto out_unalias;
2962
2963         /* See lock_rename() */
2964         if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2965                 goto out_err;
2966         m1 = &dentry->d_sb->s_vfs_rename_mutex;
2967         if (!inode_trylock_shared(alias->d_parent->d_inode))
2968                 goto out_err;
2969         m2 = &alias->d_parent->d_inode->i_rwsem;
2970 out_unalias:
2971         __d_move(alias, dentry, false);
2972         ret = 0;
2973 out_err:
2974         if (m2)
2975                 up_read(m2);
2976         if (m1)
2977                 mutex_unlock(m1);
2978         return ret;
2979 }
2980
2981 /**
2982  * d_splice_alias - splice a disconnected dentry into the tree if one exists
2983  * @inode:  the inode which may have a disconnected dentry
2984  * @dentry: a negative dentry which we want to point to the inode.
2985  *
2986  * If inode is a directory and has an IS_ROOT alias, then d_move that in
2987  * place of the given dentry and return it, else simply d_add the inode
2988  * to the dentry and return NULL.
2989  *
2990  * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2991  * we should error out: directories can't have multiple aliases.
2992  *
2993  * This is needed in the lookup routine of any filesystem that is exportable
2994  * (via knfsd) so that we can build dcache paths to directories effectively.
2995  *
2996  * If a dentry was found and moved, then it is returned.  Otherwise NULL
2997  * is returned.  This matches the expected return value of ->lookup.
2998  *
2999  * Cluster filesystems may call this function with a negative, hashed dentry.
3000  * In that case, we know that the inode will be a regular file, and also this
3001  * will only occur during atomic_open. So we need to check for the dentry
3002  * being already hashed only in the final case.
3003  */
3004 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
3005 {
3006         if (IS_ERR(inode))
3007                 return ERR_CAST(inode);
3008
3009         BUG_ON(!d_unhashed(dentry));
3010
3011         if (!inode)
3012                 goto out;
3013
3014         security_d_instantiate(dentry, inode);
3015         spin_lock(&inode->i_lock);
3016         if (S_ISDIR(inode->i_mode)) {
3017                 struct dentry *new = __d_find_any_alias(inode);
3018                 if (unlikely(new)) {
3019                         /* The reference to new ensures it remains an alias */
3020                         spin_unlock(&inode->i_lock);
3021                         write_seqlock(&rename_lock);
3022                         if (unlikely(d_ancestor(new, dentry))) {
3023                                 write_sequnlock(&rename_lock);
3024                                 dput(new);
3025                                 new = ERR_PTR(-ELOOP);
3026                                 pr_warn_ratelimited(
3027                                         "VFS: Lookup of '%s' in %s %s"
3028                                         " would have caused loop\n",
3029                                         dentry->d_name.name,
3030                                         inode->i_sb->s_type->name,
3031                                         inode->i_sb->s_id);
3032                         } else if (!IS_ROOT(new)) {
3033                                 struct dentry *old_parent = dget(new->d_parent);
3034                                 int err = __d_unalias(inode, dentry, new);
3035                                 write_sequnlock(&rename_lock);
3036                                 if (err) {
3037                                         dput(new);
3038                                         new = ERR_PTR(err);
3039                                 }
3040                                 dput(old_parent);
3041                         } else {
3042                                 __d_move(new, dentry, false);
3043                                 write_sequnlock(&rename_lock);
3044                         }
3045                         iput(inode);
3046                         return new;
3047                 }
3048         }
3049 out:
3050         __d_add(dentry, inode);
3051         return NULL;
3052 }
3053 EXPORT_SYMBOL(d_splice_alias);
3054
3055 /*
3056  * Test whether new_dentry is a subdirectory of old_dentry.
3057  *
3058  * Trivially implemented using the dcache structure
3059  */
3060
3061 /**
3062  * is_subdir - is new dentry a subdirectory of old_dentry
3063  * @new_dentry: new dentry
3064  * @old_dentry: old dentry
3065  *
3066  * Returns true if new_dentry is a subdirectory of the parent (at any depth).
3067  * Returns false otherwise.
3068  * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3069  */
3070   
3071 bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3072 {
3073         bool result;
3074         unsigned seq;
3075
3076         if (new_dentry == old_dentry)
3077                 return true;
3078
3079         do {
3080                 /* for restarting inner loop in case of seq retry */
3081                 seq = read_seqbegin(&rename_lock);
3082                 /*
3083                  * Need rcu_readlock to protect against the d_parent trashing
3084                  * due to d_move
3085                  */
3086                 rcu_read_lock();
3087                 if (d_ancestor(old_dentry, new_dentry))
3088                         result = true;
3089                 else
3090                         result = false;
3091                 rcu_read_unlock();
3092         } while (read_seqretry(&rename_lock, seq));
3093
3094         return result;
3095 }
3096 EXPORT_SYMBOL(is_subdir);
3097
3098 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3099 {
3100         struct dentry *root = data;
3101         if (dentry != root) {
3102                 if (d_unhashed(dentry) || !dentry->d_inode)
3103                         return D_WALK_SKIP;
3104
3105                 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3106                         dentry->d_flags |= DCACHE_GENOCIDE;
3107                         dentry->d_lockref.count--;
3108                 }
3109         }
3110         return D_WALK_CONTINUE;
3111 }
3112
3113 void d_genocide(struct dentry *parent)
3114 {
3115         d_walk(parent, parent, d_genocide_kill, NULL);
3116 }
3117
3118 EXPORT_SYMBOL(d_genocide);
3119
3120 void d_tmpfile(struct dentry *dentry, struct inode *inode)
3121 {
3122         inode_dec_link_count(inode);
3123         BUG_ON(dentry->d_name.name != dentry->d_iname ||
3124                 !hlist_unhashed(&dentry->d_u.d_alias) ||
3125                 !d_unlinked(dentry));
3126         spin_lock(&dentry->d_parent->d_lock);
3127         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3128         dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3129                                 (unsigned long long)inode->i_ino);
3130         spin_unlock(&dentry->d_lock);
3131         spin_unlock(&dentry->d_parent->d_lock);
3132         d_instantiate(dentry, inode);
3133 }
3134 EXPORT_SYMBOL(d_tmpfile);
3135
3136 static __initdata unsigned long dhash_entries;
3137 static int __init set_dhash_entries(char *str)
3138 {
3139         if (!str)
3140                 return 0;
3141         dhash_entries = simple_strtoul(str, &str, 0);
3142         return 1;
3143 }
3144 __setup("dhash_entries=", set_dhash_entries);
3145
3146 static void __init dcache_init_early(void)
3147 {
3148         /* If hashes are distributed across NUMA nodes, defer
3149          * hash allocation until vmalloc space is available.
3150          */
3151         if (hashdist)
3152                 return;
3153
3154         dentry_hashtable =
3155                 alloc_large_system_hash("Dentry cache",
3156                                         sizeof(struct hlist_bl_head),
3157                                         dhash_entries,
3158                                         13,
3159                                         HASH_EARLY | HASH_ZERO,
3160                                         &d_hash_shift,
3161                                         NULL,
3162                                         0,
3163                                         0);
3164         d_hash_shift = 32 - d_hash_shift;
3165 }
3166
3167 static void __init dcache_init(void)
3168 {
3169         /*
3170          * A constructor could be added for stable state like the lists,
3171          * but it is probably not worth it because of the cache nature
3172          * of the dcache.
3173          */
3174         dentry_cache = KMEM_CACHE_USERCOPY(dentry,
3175                 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD|SLAB_ACCOUNT,
3176                 d_iname);
3177
3178         /* Hash may have been set up in dcache_init_early */
3179         if (!hashdist)
3180                 return;
3181
3182         dentry_hashtable =
3183                 alloc_large_system_hash("Dentry cache",
3184                                         sizeof(struct hlist_bl_head),
3185                                         dhash_entries,
3186                                         13,
3187                                         HASH_ZERO,
3188                                         &d_hash_shift,
3189                                         NULL,
3190                                         0,
3191                                         0);
3192         d_hash_shift = 32 - d_hash_shift;
3193 }
3194
3195 /* SLAB cache for __getname() consumers */
3196 struct kmem_cache *names_cachep __read_mostly;
3197 EXPORT_SYMBOL(names_cachep);
3198
3199 void __init vfs_caches_init_early(void)
3200 {
3201         int i;
3202
3203         for (i = 0; i < ARRAY_SIZE(in_lookup_hashtable); i++)
3204                 INIT_HLIST_BL_HEAD(&in_lookup_hashtable[i]);
3205
3206         dcache_init_early();
3207         inode_init_early();
3208 }
3209
3210 void __init vfs_caches_init(void)
3211 {
3212         names_cachep = kmem_cache_create_usercopy("names_cache", PATH_MAX, 0,
3213                         SLAB_HWCACHE_ALIGN|SLAB_PANIC, 0, PATH_MAX, NULL);
3214
3215         dcache_init();
3216         inode_init();
3217         files_init();
3218         files_maxfiles_init();
3219         mnt_init();
3220         bdev_cache_init();
3221         chrdev_init();
3222 }