restore cond_resched() in shrink_dcache_parent()
[muen/linux.git] / fs / dcache.c
1 /*
2  * fs/dcache.c
3  *
4  * Complete reimplementation
5  * (C) 1997 Thomas Schoebel-Theuer,
6  * with heavy changes by Linus Torvalds
7  */
8
9 /*
10  * Notes on the allocation strategy:
11  *
12  * The dcache is a master of the icache - whenever a dcache entry
13  * exists, the inode will always exist. "iput()" is done either when
14  * the dcache entry is deleted or garbage collected.
15  */
16
17 #include <linux/ratelimit.h>
18 #include <linux/string.h>
19 #include <linux/mm.h>
20 #include <linux/fs.h>
21 #include <linux/fsnotify.h>
22 #include <linux/slab.h>
23 #include <linux/init.h>
24 #include <linux/hash.h>
25 #include <linux/cache.h>
26 #include <linux/export.h>
27 #include <linux/security.h>
28 #include <linux/seqlock.h>
29 #include <linux/bootmem.h>
30 #include <linux/bit_spinlock.h>
31 #include <linux/rculist_bl.h>
32 #include <linux/list_lru.h>
33 #include "internal.h"
34 #include "mount.h"
35
36 /*
37  * Usage:
38  * dcache->d_inode->i_lock protects:
39  *   - i_dentry, d_u.d_alias, d_inode of aliases
40  * dcache_hash_bucket lock protects:
41  *   - the dcache hash table
42  * s_roots bl list spinlock protects:
43  *   - the s_roots list (see __d_drop)
44  * dentry->d_sb->s_dentry_lru_lock protects:
45  *   - the dcache lru lists and counters
46  * d_lock protects:
47  *   - d_flags
48  *   - d_name
49  *   - d_lru
50  *   - d_count
51  *   - d_unhashed()
52  *   - d_parent and d_subdirs
53  *   - childrens' d_child and d_parent
54  *   - d_u.d_alias, d_inode
55  *
56  * Ordering:
57  * dentry->d_inode->i_lock
58  *   dentry->d_lock
59  *     dentry->d_sb->s_dentry_lru_lock
60  *     dcache_hash_bucket lock
61  *     s_roots lock
62  *
63  * If there is an ancestor relationship:
64  * dentry->d_parent->...->d_parent->d_lock
65  *   ...
66  *     dentry->d_parent->d_lock
67  *       dentry->d_lock
68  *
69  * If no ancestor relationship:
70  * arbitrary, since it's serialized on rename_lock
71  */
72 int sysctl_vfs_cache_pressure __read_mostly = 100;
73 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
74
75 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
76
77 EXPORT_SYMBOL(rename_lock);
78
79 static struct kmem_cache *dentry_cache __read_mostly;
80
81 const struct qstr empty_name = QSTR_INIT("", 0);
82 EXPORT_SYMBOL(empty_name);
83 const struct qstr slash_name = QSTR_INIT("/", 1);
84 EXPORT_SYMBOL(slash_name);
85
86 /*
87  * This is the single most critical data structure when it comes
88  * to the dcache: the hashtable for lookups. Somebody should try
89  * to make this good - I've just made it work.
90  *
91  * This hash-function tries to avoid losing too many bits of hash
92  * information, yet avoid using a prime hash-size or similar.
93  */
94
95 static unsigned int d_hash_shift __read_mostly;
96
97 static struct hlist_bl_head *dentry_hashtable __read_mostly;
98
99 static inline struct hlist_bl_head *d_hash(unsigned int hash)
100 {
101         return dentry_hashtable + (hash >> d_hash_shift);
102 }
103
104 #define IN_LOOKUP_SHIFT 10
105 static struct hlist_bl_head in_lookup_hashtable[1 << IN_LOOKUP_SHIFT];
106
107 static inline struct hlist_bl_head *in_lookup_hash(const struct dentry *parent,
108                                         unsigned int hash)
109 {
110         hash += (unsigned long) parent / L1_CACHE_BYTES;
111         return in_lookup_hashtable + hash_32(hash, IN_LOOKUP_SHIFT);
112 }
113
114
115 /* Statistics gathering. */
116 struct dentry_stat_t dentry_stat = {
117         .age_limit = 45,
118 };
119
120 static DEFINE_PER_CPU(long, nr_dentry);
121 static DEFINE_PER_CPU(long, nr_dentry_unused);
122
123 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
124
125 /*
126  * Here we resort to our own counters instead of using generic per-cpu counters
127  * for consistency with what the vfs inode code does. We are expected to harvest
128  * better code and performance by having our own specialized counters.
129  *
130  * Please note that the loop is done over all possible CPUs, not over all online
131  * CPUs. The reason for this is that we don't want to play games with CPUs going
132  * on and off. If one of them goes off, we will just keep their counters.
133  *
134  * glommer: See cffbc8a for details, and if you ever intend to change this,
135  * please update all vfs counters to match.
136  */
137 static long get_nr_dentry(void)
138 {
139         int i;
140         long sum = 0;
141         for_each_possible_cpu(i)
142                 sum += per_cpu(nr_dentry, i);
143         return sum < 0 ? 0 : sum;
144 }
145
146 static long get_nr_dentry_unused(void)
147 {
148         int i;
149         long sum = 0;
150         for_each_possible_cpu(i)
151                 sum += per_cpu(nr_dentry_unused, i);
152         return sum < 0 ? 0 : sum;
153 }
154
155 int proc_nr_dentry(struct ctl_table *table, int write, void __user *buffer,
156                    size_t *lenp, loff_t *ppos)
157 {
158         dentry_stat.nr_dentry = get_nr_dentry();
159         dentry_stat.nr_unused = get_nr_dentry_unused();
160         return proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
161 }
162 #endif
163
164 /*
165  * Compare 2 name strings, return 0 if they match, otherwise non-zero.
166  * The strings are both count bytes long, and count is non-zero.
167  */
168 #ifdef CONFIG_DCACHE_WORD_ACCESS
169
170 #include <asm/word-at-a-time.h>
171 /*
172  * NOTE! 'cs' and 'scount' come from a dentry, so it has a
173  * aligned allocation for this particular component. We don't
174  * strictly need the load_unaligned_zeropad() safety, but it
175  * doesn't hurt either.
176  *
177  * In contrast, 'ct' and 'tcount' can be from a pathname, and do
178  * need the careful unaligned handling.
179  */
180 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
181 {
182         unsigned long a,b,mask;
183
184         for (;;) {
185                 a = read_word_at_a_time(cs);
186                 b = load_unaligned_zeropad(ct);
187                 if (tcount < sizeof(unsigned long))
188                         break;
189                 if (unlikely(a != b))
190                         return 1;
191                 cs += sizeof(unsigned long);
192                 ct += sizeof(unsigned long);
193                 tcount -= sizeof(unsigned long);
194                 if (!tcount)
195                         return 0;
196         }
197         mask = bytemask_from_count(tcount);
198         return unlikely(!!((a ^ b) & mask));
199 }
200
201 #else
202
203 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
204 {
205         do {
206                 if (*cs != *ct)
207                         return 1;
208                 cs++;
209                 ct++;
210                 tcount--;
211         } while (tcount);
212         return 0;
213 }
214
215 #endif
216
217 static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount)
218 {
219         /*
220          * Be careful about RCU walk racing with rename:
221          * use 'READ_ONCE' to fetch the name pointer.
222          *
223          * NOTE! Even if a rename will mean that the length
224          * was not loaded atomically, we don't care. The
225          * RCU walk will check the sequence count eventually,
226          * and catch it. And we won't overrun the buffer,
227          * because we're reading the name pointer atomically,
228          * and a dentry name is guaranteed to be properly
229          * terminated with a NUL byte.
230          *
231          * End result: even if 'len' is wrong, we'll exit
232          * early because the data cannot match (there can
233          * be no NUL in the ct/tcount data)
234          */
235         const unsigned char *cs = READ_ONCE(dentry->d_name.name);
236
237         return dentry_string_cmp(cs, ct, tcount);
238 }
239
240 struct external_name {
241         union {
242                 atomic_t count;
243                 struct rcu_head head;
244         } u;
245         unsigned char name[];
246 };
247
248 static inline struct external_name *external_name(struct dentry *dentry)
249 {
250         return container_of(dentry->d_name.name, struct external_name, name[0]);
251 }
252
253 static void __d_free(struct rcu_head *head)
254 {
255         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
256
257         kmem_cache_free(dentry_cache, dentry); 
258 }
259
260 static void __d_free_external_name(struct rcu_head *head)
261 {
262         struct external_name *name = container_of(head, struct external_name,
263                                                   u.head);
264
265         mod_node_page_state(page_pgdat(virt_to_page(name)),
266                             NR_INDIRECTLY_RECLAIMABLE_BYTES,
267                             -ksize(name));
268
269         kfree(name);
270 }
271
272 static void __d_free_external(struct rcu_head *head)
273 {
274         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
275
276         __d_free_external_name(&external_name(dentry)->u.head);
277
278         kmem_cache_free(dentry_cache, dentry);
279 }
280
281 static inline int dname_external(const struct dentry *dentry)
282 {
283         return dentry->d_name.name != dentry->d_iname;
284 }
285
286 void take_dentry_name_snapshot(struct name_snapshot *name, struct dentry *dentry)
287 {
288         spin_lock(&dentry->d_lock);
289         if (unlikely(dname_external(dentry))) {
290                 struct external_name *p = external_name(dentry);
291                 atomic_inc(&p->u.count);
292                 spin_unlock(&dentry->d_lock);
293                 name->name = p->name;
294         } else {
295                 memcpy(name->inline_name, dentry->d_iname, DNAME_INLINE_LEN);
296                 spin_unlock(&dentry->d_lock);
297                 name->name = name->inline_name;
298         }
299 }
300 EXPORT_SYMBOL(take_dentry_name_snapshot);
301
302 void release_dentry_name_snapshot(struct name_snapshot *name)
303 {
304         if (unlikely(name->name != name->inline_name)) {
305                 struct external_name *p;
306                 p = container_of(name->name, struct external_name, name[0]);
307                 if (unlikely(atomic_dec_and_test(&p->u.count)))
308                         call_rcu(&p->u.head, __d_free_external_name);
309         }
310 }
311 EXPORT_SYMBOL(release_dentry_name_snapshot);
312
313 static inline void __d_set_inode_and_type(struct dentry *dentry,
314                                           struct inode *inode,
315                                           unsigned type_flags)
316 {
317         unsigned flags;
318
319         dentry->d_inode = inode;
320         flags = READ_ONCE(dentry->d_flags);
321         flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
322         flags |= type_flags;
323         WRITE_ONCE(dentry->d_flags, flags);
324 }
325
326 static inline void __d_clear_type_and_inode(struct dentry *dentry)
327 {
328         unsigned flags = READ_ONCE(dentry->d_flags);
329
330         flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
331         WRITE_ONCE(dentry->d_flags, flags);
332         dentry->d_inode = NULL;
333 }
334
335 static void dentry_free(struct dentry *dentry)
336 {
337         WARN_ON(!hlist_unhashed(&dentry->d_u.d_alias));
338         if (unlikely(dname_external(dentry))) {
339                 struct external_name *p = external_name(dentry);
340                 if (likely(atomic_dec_and_test(&p->u.count))) {
341                         call_rcu(&dentry->d_u.d_rcu, __d_free_external);
342                         return;
343                 }
344         }
345         /* if dentry was never visible to RCU, immediate free is OK */
346         if (!(dentry->d_flags & DCACHE_RCUACCESS))
347                 __d_free(&dentry->d_u.d_rcu);
348         else
349                 call_rcu(&dentry->d_u.d_rcu, __d_free);
350 }
351
352 /*
353  * Release the dentry's inode, using the filesystem
354  * d_iput() operation if defined.
355  */
356 static void dentry_unlink_inode(struct dentry * dentry)
357         __releases(dentry->d_lock)
358         __releases(dentry->d_inode->i_lock)
359 {
360         struct inode *inode = dentry->d_inode;
361         bool hashed = !d_unhashed(dentry);
362
363         if (hashed)
364                 raw_write_seqcount_begin(&dentry->d_seq);
365         __d_clear_type_and_inode(dentry);
366         hlist_del_init(&dentry->d_u.d_alias);
367         if (hashed)
368                 raw_write_seqcount_end(&dentry->d_seq);
369         spin_unlock(&dentry->d_lock);
370         spin_unlock(&inode->i_lock);
371         if (!inode->i_nlink)
372                 fsnotify_inoderemove(inode);
373         if (dentry->d_op && dentry->d_op->d_iput)
374                 dentry->d_op->d_iput(dentry, inode);
375         else
376                 iput(inode);
377 }
378
379 /*
380  * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
381  * is in use - which includes both the "real" per-superblock
382  * LRU list _and_ the DCACHE_SHRINK_LIST use.
383  *
384  * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
385  * on the shrink list (ie not on the superblock LRU list).
386  *
387  * The per-cpu "nr_dentry_unused" counters are updated with
388  * the DCACHE_LRU_LIST bit.
389  *
390  * These helper functions make sure we always follow the
391  * rules. d_lock must be held by the caller.
392  */
393 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
394 static void d_lru_add(struct dentry *dentry)
395 {
396         D_FLAG_VERIFY(dentry, 0);
397         dentry->d_flags |= DCACHE_LRU_LIST;
398         this_cpu_inc(nr_dentry_unused);
399         WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
400 }
401
402 static void d_lru_del(struct dentry *dentry)
403 {
404         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
405         dentry->d_flags &= ~DCACHE_LRU_LIST;
406         this_cpu_dec(nr_dentry_unused);
407         WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
408 }
409
410 static void d_shrink_del(struct dentry *dentry)
411 {
412         D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
413         list_del_init(&dentry->d_lru);
414         dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
415         this_cpu_dec(nr_dentry_unused);
416 }
417
418 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
419 {
420         D_FLAG_VERIFY(dentry, 0);
421         list_add(&dentry->d_lru, list);
422         dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
423         this_cpu_inc(nr_dentry_unused);
424 }
425
426 /*
427  * These can only be called under the global LRU lock, ie during the
428  * callback for freeing the LRU list. "isolate" removes it from the
429  * LRU lists entirely, while shrink_move moves it to the indicated
430  * private list.
431  */
432 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
433 {
434         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
435         dentry->d_flags &= ~DCACHE_LRU_LIST;
436         this_cpu_dec(nr_dentry_unused);
437         list_lru_isolate(lru, &dentry->d_lru);
438 }
439
440 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
441                               struct list_head *list)
442 {
443         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
444         dentry->d_flags |= DCACHE_SHRINK_LIST;
445         list_lru_isolate_move(lru, &dentry->d_lru, list);
446 }
447
448 /**
449  * d_drop - drop a dentry
450  * @dentry: dentry to drop
451  *
452  * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
453  * be found through a VFS lookup any more. Note that this is different from
454  * deleting the dentry - d_delete will try to mark the dentry negative if
455  * possible, giving a successful _negative_ lookup, while d_drop will
456  * just make the cache lookup fail.
457  *
458  * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
459  * reason (NFS timeouts or autofs deletes).
460  *
461  * __d_drop requires dentry->d_lock
462  * ___d_drop doesn't mark dentry as "unhashed"
463  *   (dentry->d_hash.pprev will be LIST_POISON2, not NULL).
464  */
465 static void ___d_drop(struct dentry *dentry)
466 {
467         struct hlist_bl_head *b;
468         /*
469          * Hashed dentries are normally on the dentry hashtable,
470          * with the exception of those newly allocated by
471          * d_obtain_root, which are always IS_ROOT:
472          */
473         if (unlikely(IS_ROOT(dentry)))
474                 b = &dentry->d_sb->s_roots;
475         else
476                 b = d_hash(dentry->d_name.hash);
477
478         hlist_bl_lock(b);
479         __hlist_bl_del(&dentry->d_hash);
480         hlist_bl_unlock(b);
481 }
482
483 void __d_drop(struct dentry *dentry)
484 {
485         if (!d_unhashed(dentry)) {
486                 ___d_drop(dentry);
487                 dentry->d_hash.pprev = NULL;
488                 write_seqcount_invalidate(&dentry->d_seq);
489         }
490 }
491 EXPORT_SYMBOL(__d_drop);
492
493 void d_drop(struct dentry *dentry)
494 {
495         spin_lock(&dentry->d_lock);
496         __d_drop(dentry);
497         spin_unlock(&dentry->d_lock);
498 }
499 EXPORT_SYMBOL(d_drop);
500
501 static inline void dentry_unlist(struct dentry *dentry, struct dentry *parent)
502 {
503         struct dentry *next;
504         /*
505          * Inform d_walk() and shrink_dentry_list() that we are no longer
506          * attached to the dentry tree
507          */
508         dentry->d_flags |= DCACHE_DENTRY_KILLED;
509         if (unlikely(list_empty(&dentry->d_child)))
510                 return;
511         __list_del_entry(&dentry->d_child);
512         /*
513          * Cursors can move around the list of children.  While we'd been
514          * a normal list member, it didn't matter - ->d_child.next would've
515          * been updated.  However, from now on it won't be and for the
516          * things like d_walk() it might end up with a nasty surprise.
517          * Normally d_walk() doesn't care about cursors moving around -
518          * ->d_lock on parent prevents that and since a cursor has no children
519          * of its own, we get through it without ever unlocking the parent.
520          * There is one exception, though - if we ascend from a child that
521          * gets killed as soon as we unlock it, the next sibling is found
522          * using the value left in its ->d_child.next.  And if _that_
523          * pointed to a cursor, and cursor got moved (e.g. by lseek())
524          * before d_walk() regains parent->d_lock, we'll end up skipping
525          * everything the cursor had been moved past.
526          *
527          * Solution: make sure that the pointer left behind in ->d_child.next
528          * points to something that won't be moving around.  I.e. skip the
529          * cursors.
530          */
531         while (dentry->d_child.next != &parent->d_subdirs) {
532                 next = list_entry(dentry->d_child.next, struct dentry, d_child);
533                 if (likely(!(next->d_flags & DCACHE_DENTRY_CURSOR)))
534                         break;
535                 dentry->d_child.next = next->d_child.next;
536         }
537 }
538
539 static void __dentry_kill(struct dentry *dentry)
540 {
541         struct dentry *parent = NULL;
542         bool can_free = true;
543         if (!IS_ROOT(dentry))
544                 parent = dentry->d_parent;
545
546         /*
547          * The dentry is now unrecoverably dead to the world.
548          */
549         lockref_mark_dead(&dentry->d_lockref);
550
551         /*
552          * inform the fs via d_prune that this dentry is about to be
553          * unhashed and destroyed.
554          */
555         if (dentry->d_flags & DCACHE_OP_PRUNE)
556                 dentry->d_op->d_prune(dentry);
557
558         if (dentry->d_flags & DCACHE_LRU_LIST) {
559                 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
560                         d_lru_del(dentry);
561         }
562         /* if it was on the hash then remove it */
563         __d_drop(dentry);
564         dentry_unlist(dentry, parent);
565         if (parent)
566                 spin_unlock(&parent->d_lock);
567         if (dentry->d_inode)
568                 dentry_unlink_inode(dentry);
569         else
570                 spin_unlock(&dentry->d_lock);
571         this_cpu_dec(nr_dentry);
572         if (dentry->d_op && dentry->d_op->d_release)
573                 dentry->d_op->d_release(dentry);
574
575         spin_lock(&dentry->d_lock);
576         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
577                 dentry->d_flags |= DCACHE_MAY_FREE;
578                 can_free = false;
579         }
580         spin_unlock(&dentry->d_lock);
581         if (likely(can_free))
582                 dentry_free(dentry);
583         cond_resched();
584 }
585
586 static struct dentry *__lock_parent(struct dentry *dentry)
587 {
588         struct dentry *parent;
589         rcu_read_lock();
590         spin_unlock(&dentry->d_lock);
591 again:
592         parent = READ_ONCE(dentry->d_parent);
593         spin_lock(&parent->d_lock);
594         /*
595          * We can't blindly lock dentry until we are sure
596          * that we won't violate the locking order.
597          * Any changes of dentry->d_parent must have
598          * been done with parent->d_lock held, so
599          * spin_lock() above is enough of a barrier
600          * for checking if it's still our child.
601          */
602         if (unlikely(parent != dentry->d_parent)) {
603                 spin_unlock(&parent->d_lock);
604                 goto again;
605         }
606         rcu_read_unlock();
607         if (parent != dentry)
608                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
609         else
610                 parent = NULL;
611         return parent;
612 }
613
614 static inline struct dentry *lock_parent(struct dentry *dentry)
615 {
616         struct dentry *parent = dentry->d_parent;
617         if (IS_ROOT(dentry))
618                 return NULL;
619         if (likely(spin_trylock(&parent->d_lock)))
620                 return parent;
621         return __lock_parent(dentry);
622 }
623
624 static inline bool retain_dentry(struct dentry *dentry)
625 {
626         WARN_ON(d_in_lookup(dentry));
627
628         /* Unreachable? Get rid of it */
629         if (unlikely(d_unhashed(dentry)))
630                 return false;
631
632         if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
633                 return false;
634
635         if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
636                 if (dentry->d_op->d_delete(dentry))
637                         return false;
638         }
639         /* retain; LRU fodder */
640         dentry->d_lockref.count--;
641         if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
642                 d_lru_add(dentry);
643         else if (unlikely(!(dentry->d_flags & DCACHE_REFERENCED)))
644                 dentry->d_flags |= DCACHE_REFERENCED;
645         return true;
646 }
647
648 /*
649  * Finish off a dentry we've decided to kill.
650  * dentry->d_lock must be held, returns with it unlocked.
651  * Returns dentry requiring refcount drop, or NULL if we're done.
652  */
653 static struct dentry *dentry_kill(struct dentry *dentry)
654         __releases(dentry->d_lock)
655 {
656         struct inode *inode = dentry->d_inode;
657         struct dentry *parent = NULL;
658
659         if (inode && unlikely(!spin_trylock(&inode->i_lock)))
660                 goto slow_positive;
661
662         if (!IS_ROOT(dentry)) {
663                 parent = dentry->d_parent;
664                 if (unlikely(!spin_trylock(&parent->d_lock))) {
665                         parent = __lock_parent(dentry);
666                         if (likely(inode || !dentry->d_inode))
667                                 goto got_locks;
668                         /* negative that became positive */
669                         if (parent)
670                                 spin_unlock(&parent->d_lock);
671                         inode = dentry->d_inode;
672                         goto slow_positive;
673                 }
674         }
675         __dentry_kill(dentry);
676         return parent;
677
678 slow_positive:
679         spin_unlock(&dentry->d_lock);
680         spin_lock(&inode->i_lock);
681         spin_lock(&dentry->d_lock);
682         parent = lock_parent(dentry);
683 got_locks:
684         if (unlikely(dentry->d_lockref.count != 1)) {
685                 dentry->d_lockref.count--;
686         } else if (likely(!retain_dentry(dentry))) {
687                 __dentry_kill(dentry);
688                 return parent;
689         }
690         /* we are keeping it, after all */
691         if (inode)
692                 spin_unlock(&inode->i_lock);
693         if (parent)
694                 spin_unlock(&parent->d_lock);
695         spin_unlock(&dentry->d_lock);
696         return NULL;
697 }
698
699 /*
700  * Try to do a lockless dput(), and return whether that was successful.
701  *
702  * If unsuccessful, we return false, having already taken the dentry lock.
703  *
704  * The caller needs to hold the RCU read lock, so that the dentry is
705  * guaranteed to stay around even if the refcount goes down to zero!
706  */
707 static inline bool fast_dput(struct dentry *dentry)
708 {
709         int ret;
710         unsigned int d_flags;
711
712         /*
713          * If we have a d_op->d_delete() operation, we sould not
714          * let the dentry count go to zero, so use "put_or_lock".
715          */
716         if (unlikely(dentry->d_flags & DCACHE_OP_DELETE))
717                 return lockref_put_or_lock(&dentry->d_lockref);
718
719         /*
720          * .. otherwise, we can try to just decrement the
721          * lockref optimistically.
722          */
723         ret = lockref_put_return(&dentry->d_lockref);
724
725         /*
726          * If the lockref_put_return() failed due to the lock being held
727          * by somebody else, the fast path has failed. We will need to
728          * get the lock, and then check the count again.
729          */
730         if (unlikely(ret < 0)) {
731                 spin_lock(&dentry->d_lock);
732                 if (dentry->d_lockref.count > 1) {
733                         dentry->d_lockref.count--;
734                         spin_unlock(&dentry->d_lock);
735                         return 1;
736                 }
737                 return 0;
738         }
739
740         /*
741          * If we weren't the last ref, we're done.
742          */
743         if (ret)
744                 return 1;
745
746         /*
747          * Careful, careful. The reference count went down
748          * to zero, but we don't hold the dentry lock, so
749          * somebody else could get it again, and do another
750          * dput(), and we need to not race with that.
751          *
752          * However, there is a very special and common case
753          * where we don't care, because there is nothing to
754          * do: the dentry is still hashed, it does not have
755          * a 'delete' op, and it's referenced and already on
756          * the LRU list.
757          *
758          * NOTE! Since we aren't locked, these values are
759          * not "stable". However, it is sufficient that at
760          * some point after we dropped the reference the
761          * dentry was hashed and the flags had the proper
762          * value. Other dentry users may have re-gotten
763          * a reference to the dentry and change that, but
764          * our work is done - we can leave the dentry
765          * around with a zero refcount.
766          */
767         smp_rmb();
768         d_flags = READ_ONCE(dentry->d_flags);
769         d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST | DCACHE_DISCONNECTED;
770
771         /* Nothing to do? Dropping the reference was all we needed? */
772         if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry))
773                 return 1;
774
775         /*
776          * Not the fast normal case? Get the lock. We've already decremented
777          * the refcount, but we'll need to re-check the situation after
778          * getting the lock.
779          */
780         spin_lock(&dentry->d_lock);
781
782         /*
783          * Did somebody else grab a reference to it in the meantime, and
784          * we're no longer the last user after all? Alternatively, somebody
785          * else could have killed it and marked it dead. Either way, we
786          * don't need to do anything else.
787          */
788         if (dentry->d_lockref.count) {
789                 spin_unlock(&dentry->d_lock);
790                 return 1;
791         }
792
793         /*
794          * Re-get the reference we optimistically dropped. We hold the
795          * lock, and we just tested that it was zero, so we can just
796          * set it to 1.
797          */
798         dentry->d_lockref.count = 1;
799         return 0;
800 }
801
802
803 /* 
804  * This is dput
805  *
806  * This is complicated by the fact that we do not want to put
807  * dentries that are no longer on any hash chain on the unused
808  * list: we'd much rather just get rid of them immediately.
809  *
810  * However, that implies that we have to traverse the dentry
811  * tree upwards to the parents which might _also_ now be
812  * scheduled for deletion (it may have been only waiting for
813  * its last child to go away).
814  *
815  * This tail recursion is done by hand as we don't want to depend
816  * on the compiler to always get this right (gcc generally doesn't).
817  * Real recursion would eat up our stack space.
818  */
819
820 /*
821  * dput - release a dentry
822  * @dentry: dentry to release 
823  *
824  * Release a dentry. This will drop the usage count and if appropriate
825  * call the dentry unlink method as well as removing it from the queues and
826  * releasing its resources. If the parent dentries were scheduled for release
827  * they too may now get deleted.
828  */
829 void dput(struct dentry *dentry)
830 {
831         while (dentry) {
832                 might_sleep();
833
834                 rcu_read_lock();
835                 if (likely(fast_dput(dentry))) {
836                         rcu_read_unlock();
837                         return;
838                 }
839
840                 /* Slow case: now with the dentry lock held */
841                 rcu_read_unlock();
842
843                 if (likely(retain_dentry(dentry))) {
844                         spin_unlock(&dentry->d_lock);
845                         return;
846                 }
847
848                 dentry = dentry_kill(dentry);
849         }
850 }
851 EXPORT_SYMBOL(dput);
852
853
854 /* This must be called with d_lock held */
855 static inline void __dget_dlock(struct dentry *dentry)
856 {
857         dentry->d_lockref.count++;
858 }
859
860 static inline void __dget(struct dentry *dentry)
861 {
862         lockref_get(&dentry->d_lockref);
863 }
864
865 struct dentry *dget_parent(struct dentry *dentry)
866 {
867         int gotref;
868         struct dentry *ret;
869
870         /*
871          * Do optimistic parent lookup without any
872          * locking.
873          */
874         rcu_read_lock();
875         ret = READ_ONCE(dentry->d_parent);
876         gotref = lockref_get_not_zero(&ret->d_lockref);
877         rcu_read_unlock();
878         if (likely(gotref)) {
879                 if (likely(ret == READ_ONCE(dentry->d_parent)))
880                         return ret;
881                 dput(ret);
882         }
883
884 repeat:
885         /*
886          * Don't need rcu_dereference because we re-check it was correct under
887          * the lock.
888          */
889         rcu_read_lock();
890         ret = dentry->d_parent;
891         spin_lock(&ret->d_lock);
892         if (unlikely(ret != dentry->d_parent)) {
893                 spin_unlock(&ret->d_lock);
894                 rcu_read_unlock();
895                 goto repeat;
896         }
897         rcu_read_unlock();
898         BUG_ON(!ret->d_lockref.count);
899         ret->d_lockref.count++;
900         spin_unlock(&ret->d_lock);
901         return ret;
902 }
903 EXPORT_SYMBOL(dget_parent);
904
905 /**
906  * d_find_alias - grab a hashed alias of inode
907  * @inode: inode in question
908  *
909  * If inode has a hashed alias, or is a directory and has any alias,
910  * acquire the reference to alias and return it. Otherwise return NULL.
911  * Notice that if inode is a directory there can be only one alias and
912  * it can be unhashed only if it has no children, or if it is the root
913  * of a filesystem, or if the directory was renamed and d_revalidate
914  * was the first vfs operation to notice.
915  *
916  * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
917  * any other hashed alias over that one.
918  */
919 static struct dentry *__d_find_alias(struct inode *inode)
920 {
921         struct dentry *alias, *discon_alias;
922
923 again:
924         discon_alias = NULL;
925         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
926                 spin_lock(&alias->d_lock);
927                 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
928                         if (IS_ROOT(alias) &&
929                             (alias->d_flags & DCACHE_DISCONNECTED)) {
930                                 discon_alias = alias;
931                         } else {
932                                 __dget_dlock(alias);
933                                 spin_unlock(&alias->d_lock);
934                                 return alias;
935                         }
936                 }
937                 spin_unlock(&alias->d_lock);
938         }
939         if (discon_alias) {
940                 alias = discon_alias;
941                 spin_lock(&alias->d_lock);
942                 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
943                         __dget_dlock(alias);
944                         spin_unlock(&alias->d_lock);
945                         return alias;
946                 }
947                 spin_unlock(&alias->d_lock);
948                 goto again;
949         }
950         return NULL;
951 }
952
953 struct dentry *d_find_alias(struct inode *inode)
954 {
955         struct dentry *de = NULL;
956
957         if (!hlist_empty(&inode->i_dentry)) {
958                 spin_lock(&inode->i_lock);
959                 de = __d_find_alias(inode);
960                 spin_unlock(&inode->i_lock);
961         }
962         return de;
963 }
964 EXPORT_SYMBOL(d_find_alias);
965
966 /*
967  *      Try to kill dentries associated with this inode.
968  * WARNING: you must own a reference to inode.
969  */
970 void d_prune_aliases(struct inode *inode)
971 {
972         struct dentry *dentry;
973 restart:
974         spin_lock(&inode->i_lock);
975         hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
976                 spin_lock(&dentry->d_lock);
977                 if (!dentry->d_lockref.count) {
978                         struct dentry *parent = lock_parent(dentry);
979                         if (likely(!dentry->d_lockref.count)) {
980                                 __dentry_kill(dentry);
981                                 dput(parent);
982                                 goto restart;
983                         }
984                         if (parent)
985                                 spin_unlock(&parent->d_lock);
986                 }
987                 spin_unlock(&dentry->d_lock);
988         }
989         spin_unlock(&inode->i_lock);
990 }
991 EXPORT_SYMBOL(d_prune_aliases);
992
993 /*
994  * Lock a dentry from shrink list.
995  * Called under rcu_read_lock() and dentry->d_lock; the former
996  * guarantees that nothing we access will be freed under us.
997  * Note that dentry is *not* protected from concurrent dentry_kill(),
998  * d_delete(), etc.
999  *
1000  * Return false if dentry has been disrupted or grabbed, leaving
1001  * the caller to kick it off-list.  Otherwise, return true and have
1002  * that dentry's inode and parent both locked.
1003  */
1004 static bool shrink_lock_dentry(struct dentry *dentry)
1005 {
1006         struct inode *inode;
1007         struct dentry *parent;
1008
1009         if (dentry->d_lockref.count)
1010                 return false;
1011
1012         inode = dentry->d_inode;
1013         if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
1014                 spin_unlock(&dentry->d_lock);
1015                 spin_lock(&inode->i_lock);
1016                 spin_lock(&dentry->d_lock);
1017                 if (unlikely(dentry->d_lockref.count))
1018                         goto out;
1019                 /* changed inode means that somebody had grabbed it */
1020                 if (unlikely(inode != dentry->d_inode))
1021                         goto out;
1022         }
1023
1024         parent = dentry->d_parent;
1025         if (IS_ROOT(dentry) || likely(spin_trylock(&parent->d_lock)))
1026                 return true;
1027
1028         spin_unlock(&dentry->d_lock);
1029         spin_lock(&parent->d_lock);
1030         if (unlikely(parent != dentry->d_parent)) {
1031                 spin_unlock(&parent->d_lock);
1032                 spin_lock(&dentry->d_lock);
1033                 goto out;
1034         }
1035         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1036         if (likely(!dentry->d_lockref.count))
1037                 return true;
1038         spin_unlock(&parent->d_lock);
1039 out:
1040         if (inode)
1041                 spin_unlock(&inode->i_lock);
1042         return false;
1043 }
1044
1045 static void shrink_dentry_list(struct list_head *list)
1046 {
1047         while (!list_empty(list)) {
1048                 struct dentry *dentry, *parent;
1049
1050                 dentry = list_entry(list->prev, struct dentry, d_lru);
1051                 spin_lock(&dentry->d_lock);
1052                 rcu_read_lock();
1053                 if (!shrink_lock_dentry(dentry)) {
1054                         bool can_free = false;
1055                         rcu_read_unlock();
1056                         d_shrink_del(dentry);
1057                         if (dentry->d_lockref.count < 0)
1058                                 can_free = dentry->d_flags & DCACHE_MAY_FREE;
1059                         spin_unlock(&dentry->d_lock);
1060                         if (can_free)
1061                                 dentry_free(dentry);
1062                         continue;
1063                 }
1064                 rcu_read_unlock();
1065                 d_shrink_del(dentry);
1066                 parent = dentry->d_parent;
1067                 __dentry_kill(dentry);
1068                 if (parent == dentry)
1069                         continue;
1070                 /*
1071                  * We need to prune ancestors too. This is necessary to prevent
1072                  * quadratic behavior of shrink_dcache_parent(), but is also
1073                  * expected to be beneficial in reducing dentry cache
1074                  * fragmentation.
1075                  */
1076                 dentry = parent;
1077                 while (dentry && !lockref_put_or_lock(&dentry->d_lockref))
1078                         dentry = dentry_kill(dentry);
1079         }
1080 }
1081
1082 static enum lru_status dentry_lru_isolate(struct list_head *item,
1083                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1084 {
1085         struct list_head *freeable = arg;
1086         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1087
1088
1089         /*
1090          * we are inverting the lru lock/dentry->d_lock here,
1091          * so use a trylock. If we fail to get the lock, just skip
1092          * it
1093          */
1094         if (!spin_trylock(&dentry->d_lock))
1095                 return LRU_SKIP;
1096
1097         /*
1098          * Referenced dentries are still in use. If they have active
1099          * counts, just remove them from the LRU. Otherwise give them
1100          * another pass through the LRU.
1101          */
1102         if (dentry->d_lockref.count) {
1103                 d_lru_isolate(lru, dentry);
1104                 spin_unlock(&dentry->d_lock);
1105                 return LRU_REMOVED;
1106         }
1107
1108         if (dentry->d_flags & DCACHE_REFERENCED) {
1109                 dentry->d_flags &= ~DCACHE_REFERENCED;
1110                 spin_unlock(&dentry->d_lock);
1111
1112                 /*
1113                  * The list move itself will be made by the common LRU code. At
1114                  * this point, we've dropped the dentry->d_lock but keep the
1115                  * lru lock. This is safe to do, since every list movement is
1116                  * protected by the lru lock even if both locks are held.
1117                  *
1118                  * This is guaranteed by the fact that all LRU management
1119                  * functions are intermediated by the LRU API calls like
1120                  * list_lru_add and list_lru_del. List movement in this file
1121                  * only ever occur through this functions or through callbacks
1122                  * like this one, that are called from the LRU API.
1123                  *
1124                  * The only exceptions to this are functions like
1125                  * shrink_dentry_list, and code that first checks for the
1126                  * DCACHE_SHRINK_LIST flag.  Those are guaranteed to be
1127                  * operating only with stack provided lists after they are
1128                  * properly isolated from the main list.  It is thus, always a
1129                  * local access.
1130                  */
1131                 return LRU_ROTATE;
1132         }
1133
1134         d_lru_shrink_move(lru, dentry, freeable);
1135         spin_unlock(&dentry->d_lock);
1136
1137         return LRU_REMOVED;
1138 }
1139
1140 /**
1141  * prune_dcache_sb - shrink the dcache
1142  * @sb: superblock
1143  * @sc: shrink control, passed to list_lru_shrink_walk()
1144  *
1145  * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1146  * is done when we need more memory and called from the superblock shrinker
1147  * function.
1148  *
1149  * This function may fail to free any resources if all the dentries are in
1150  * use.
1151  */
1152 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1153 {
1154         LIST_HEAD(dispose);
1155         long freed;
1156
1157         freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1158                                      dentry_lru_isolate, &dispose);
1159         shrink_dentry_list(&dispose);
1160         return freed;
1161 }
1162
1163 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1164                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1165 {
1166         struct list_head *freeable = arg;
1167         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1168
1169         /*
1170          * we are inverting the lru lock/dentry->d_lock here,
1171          * so use a trylock. If we fail to get the lock, just skip
1172          * it
1173          */
1174         if (!spin_trylock(&dentry->d_lock))
1175                 return LRU_SKIP;
1176
1177         d_lru_shrink_move(lru, dentry, freeable);
1178         spin_unlock(&dentry->d_lock);
1179
1180         return LRU_REMOVED;
1181 }
1182
1183
1184 /**
1185  * shrink_dcache_sb - shrink dcache for a superblock
1186  * @sb: superblock
1187  *
1188  * Shrink the dcache for the specified super block. This is used to free
1189  * the dcache before unmounting a file system.
1190  */
1191 void shrink_dcache_sb(struct super_block *sb)
1192 {
1193         long freed;
1194
1195         do {
1196                 LIST_HEAD(dispose);
1197
1198                 freed = list_lru_walk(&sb->s_dentry_lru,
1199                         dentry_lru_isolate_shrink, &dispose, 1024);
1200
1201                 this_cpu_sub(nr_dentry_unused, freed);
1202                 shrink_dentry_list(&dispose);
1203         } while (list_lru_count(&sb->s_dentry_lru) > 0);
1204 }
1205 EXPORT_SYMBOL(shrink_dcache_sb);
1206
1207 /**
1208  * enum d_walk_ret - action to talke during tree walk
1209  * @D_WALK_CONTINUE:    contrinue walk
1210  * @D_WALK_QUIT:        quit walk
1211  * @D_WALK_NORETRY:     quit when retry is needed
1212  * @D_WALK_SKIP:        skip this dentry and its children
1213  */
1214 enum d_walk_ret {
1215         D_WALK_CONTINUE,
1216         D_WALK_QUIT,
1217         D_WALK_NORETRY,
1218         D_WALK_SKIP,
1219 };
1220
1221 /**
1222  * d_walk - walk the dentry tree
1223  * @parent:     start of walk
1224  * @data:       data passed to @enter() and @finish()
1225  * @enter:      callback when first entering the dentry
1226  *
1227  * The @enter() callbacks are called with d_lock held.
1228  */
1229 static void d_walk(struct dentry *parent, void *data,
1230                    enum d_walk_ret (*enter)(void *, struct dentry *))
1231 {
1232         struct dentry *this_parent;
1233         struct list_head *next;
1234         unsigned seq = 0;
1235         enum d_walk_ret ret;
1236         bool retry = true;
1237
1238 again:
1239         read_seqbegin_or_lock(&rename_lock, &seq);
1240         this_parent = parent;
1241         spin_lock(&this_parent->d_lock);
1242
1243         ret = enter(data, this_parent);
1244         switch (ret) {
1245         case D_WALK_CONTINUE:
1246                 break;
1247         case D_WALK_QUIT:
1248         case D_WALK_SKIP:
1249                 goto out_unlock;
1250         case D_WALK_NORETRY:
1251                 retry = false;
1252                 break;
1253         }
1254 repeat:
1255         next = this_parent->d_subdirs.next;
1256 resume:
1257         while (next != &this_parent->d_subdirs) {
1258                 struct list_head *tmp = next;
1259                 struct dentry *dentry = list_entry(tmp, struct dentry, d_child);
1260                 next = tmp->next;
1261
1262                 if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
1263                         continue;
1264
1265                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1266
1267                 ret = enter(data, dentry);
1268                 switch (ret) {
1269                 case D_WALK_CONTINUE:
1270                         break;
1271                 case D_WALK_QUIT:
1272                         spin_unlock(&dentry->d_lock);
1273                         goto out_unlock;
1274                 case D_WALK_NORETRY:
1275                         retry = false;
1276                         break;
1277                 case D_WALK_SKIP:
1278                         spin_unlock(&dentry->d_lock);
1279                         continue;
1280                 }
1281
1282                 if (!list_empty(&dentry->d_subdirs)) {
1283                         spin_unlock(&this_parent->d_lock);
1284                         spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1285                         this_parent = dentry;
1286                         spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1287                         goto repeat;
1288                 }
1289                 spin_unlock(&dentry->d_lock);
1290         }
1291         /*
1292          * All done at this level ... ascend and resume the search.
1293          */
1294         rcu_read_lock();
1295 ascend:
1296         if (this_parent != parent) {
1297                 struct dentry *child = this_parent;
1298                 this_parent = child->d_parent;
1299
1300                 spin_unlock(&child->d_lock);
1301                 spin_lock(&this_parent->d_lock);
1302
1303                 /* might go back up the wrong parent if we have had a rename. */
1304                 if (need_seqretry(&rename_lock, seq))
1305                         goto rename_retry;
1306                 /* go into the first sibling still alive */
1307                 do {
1308                         next = child->d_child.next;
1309                         if (next == &this_parent->d_subdirs)
1310                                 goto ascend;
1311                         child = list_entry(next, struct dentry, d_child);
1312                 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED));
1313                 rcu_read_unlock();
1314                 goto resume;
1315         }
1316         if (need_seqretry(&rename_lock, seq))
1317                 goto rename_retry;
1318         rcu_read_unlock();
1319
1320 out_unlock:
1321         spin_unlock(&this_parent->d_lock);
1322         done_seqretry(&rename_lock, seq);
1323         return;
1324
1325 rename_retry:
1326         spin_unlock(&this_parent->d_lock);
1327         rcu_read_unlock();
1328         BUG_ON(seq & 1);
1329         if (!retry)
1330                 return;
1331         seq = 1;
1332         goto again;
1333 }
1334
1335 struct check_mount {
1336         struct vfsmount *mnt;
1337         unsigned int mounted;
1338 };
1339
1340 static enum d_walk_ret path_check_mount(void *data, struct dentry *dentry)
1341 {
1342         struct check_mount *info = data;
1343         struct path path = { .mnt = info->mnt, .dentry = dentry };
1344
1345         if (likely(!d_mountpoint(dentry)))
1346                 return D_WALK_CONTINUE;
1347         if (__path_is_mountpoint(&path)) {
1348                 info->mounted = 1;
1349                 return D_WALK_QUIT;
1350         }
1351         return D_WALK_CONTINUE;
1352 }
1353
1354 /**
1355  * path_has_submounts - check for mounts over a dentry in the
1356  *                      current namespace.
1357  * @parent: path to check.
1358  *
1359  * Return true if the parent or its subdirectories contain
1360  * a mount point in the current namespace.
1361  */
1362 int path_has_submounts(const struct path *parent)
1363 {
1364         struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };
1365
1366         read_seqlock_excl(&mount_lock);
1367         d_walk(parent->dentry, &data, path_check_mount);
1368         read_sequnlock_excl(&mount_lock);
1369
1370         return data.mounted;
1371 }
1372 EXPORT_SYMBOL(path_has_submounts);
1373
1374 /*
1375  * Called by mount code to set a mountpoint and check if the mountpoint is
1376  * reachable (e.g. NFS can unhash a directory dentry and then the complete
1377  * subtree can become unreachable).
1378  *
1379  * Only one of d_invalidate() and d_set_mounted() must succeed.  For
1380  * this reason take rename_lock and d_lock on dentry and ancestors.
1381  */
1382 int d_set_mounted(struct dentry *dentry)
1383 {
1384         struct dentry *p;
1385         int ret = -ENOENT;
1386         write_seqlock(&rename_lock);
1387         for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1388                 /* Need exclusion wrt. d_invalidate() */
1389                 spin_lock(&p->d_lock);
1390                 if (unlikely(d_unhashed(p))) {
1391                         spin_unlock(&p->d_lock);
1392                         goto out;
1393                 }
1394                 spin_unlock(&p->d_lock);
1395         }
1396         spin_lock(&dentry->d_lock);
1397         if (!d_unlinked(dentry)) {
1398                 ret = -EBUSY;
1399                 if (!d_mountpoint(dentry)) {
1400                         dentry->d_flags |= DCACHE_MOUNTED;
1401                         ret = 0;
1402                 }
1403         }
1404         spin_unlock(&dentry->d_lock);
1405 out:
1406         write_sequnlock(&rename_lock);
1407         return ret;
1408 }
1409
1410 /*
1411  * Search the dentry child list of the specified parent,
1412  * and move any unused dentries to the end of the unused
1413  * list for prune_dcache(). We descend to the next level
1414  * whenever the d_subdirs list is non-empty and continue
1415  * searching.
1416  *
1417  * It returns zero iff there are no unused children,
1418  * otherwise  it returns the number of children moved to
1419  * the end of the unused list. This may not be the total
1420  * number of unused children, because select_parent can
1421  * drop the lock and return early due to latency
1422  * constraints.
1423  */
1424
1425 struct select_data {
1426         struct dentry *start;
1427         struct list_head dispose;
1428         int found;
1429 };
1430
1431 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1432 {
1433         struct select_data *data = _data;
1434         enum d_walk_ret ret = D_WALK_CONTINUE;
1435
1436         if (data->start == dentry)
1437                 goto out;
1438
1439         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1440                 data->found++;
1441         } else {
1442                 if (dentry->d_flags & DCACHE_LRU_LIST)
1443                         d_lru_del(dentry);
1444                 if (!dentry->d_lockref.count) {
1445                         d_shrink_add(dentry, &data->dispose);
1446                         data->found++;
1447                 }
1448         }
1449         /*
1450          * We can return to the caller if we have found some (this
1451          * ensures forward progress). We'll be coming back to find
1452          * the rest.
1453          */
1454         if (!list_empty(&data->dispose))
1455                 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1456 out:
1457         return ret;
1458 }
1459
1460 /**
1461  * shrink_dcache_parent - prune dcache
1462  * @parent: parent of entries to prune
1463  *
1464  * Prune the dcache to remove unused children of the parent dentry.
1465  */
1466 void shrink_dcache_parent(struct dentry *parent)
1467 {
1468         for (;;) {
1469                 struct select_data data;
1470
1471                 INIT_LIST_HEAD(&data.dispose);
1472                 data.start = parent;
1473                 data.found = 0;
1474
1475                 d_walk(parent, &data, select_collect);
1476
1477                 if (!list_empty(&data.dispose)) {
1478                         shrink_dentry_list(&data.dispose);
1479                         continue;
1480                 }
1481
1482                 cond_resched();
1483                 if (!data.found)
1484                         break;
1485         }
1486 }
1487 EXPORT_SYMBOL(shrink_dcache_parent);
1488
1489 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1490 {
1491         /* it has busy descendents; complain about those instead */
1492         if (!list_empty(&dentry->d_subdirs))
1493                 return D_WALK_CONTINUE;
1494
1495         /* root with refcount 1 is fine */
1496         if (dentry == _data && dentry->d_lockref.count == 1)
1497                 return D_WALK_CONTINUE;
1498
1499         printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} "
1500                         " still in use (%d) [unmount of %s %s]\n",
1501                        dentry,
1502                        dentry->d_inode ?
1503                        dentry->d_inode->i_ino : 0UL,
1504                        dentry,
1505                        dentry->d_lockref.count,
1506                        dentry->d_sb->s_type->name,
1507                        dentry->d_sb->s_id);
1508         WARN_ON(1);
1509         return D_WALK_CONTINUE;
1510 }
1511
1512 static void do_one_tree(struct dentry *dentry)
1513 {
1514         shrink_dcache_parent(dentry);
1515         d_walk(dentry, dentry, umount_check);
1516         d_drop(dentry);
1517         dput(dentry);
1518 }
1519
1520 /*
1521  * destroy the dentries attached to a superblock on unmounting
1522  */
1523 void shrink_dcache_for_umount(struct super_block *sb)
1524 {
1525         struct dentry *dentry;
1526
1527         WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
1528
1529         dentry = sb->s_root;
1530         sb->s_root = NULL;
1531         do_one_tree(dentry);
1532
1533         while (!hlist_bl_empty(&sb->s_roots)) {
1534                 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_roots), struct dentry, d_hash));
1535                 do_one_tree(dentry);
1536         }
1537 }
1538
1539 static enum d_walk_ret find_submount(void *_data, struct dentry *dentry)
1540 {
1541         struct dentry **victim = _data;
1542         if (d_mountpoint(dentry)) {
1543                 __dget_dlock(dentry);
1544                 *victim = dentry;
1545                 return D_WALK_QUIT;
1546         }
1547         return D_WALK_CONTINUE;
1548 }
1549
1550 /**
1551  * d_invalidate - detach submounts, prune dcache, and drop
1552  * @dentry: dentry to invalidate (aka detach, prune and drop)
1553  */
1554 void d_invalidate(struct dentry *dentry)
1555 {
1556         bool had_submounts = false;
1557         spin_lock(&dentry->d_lock);
1558         if (d_unhashed(dentry)) {
1559                 spin_unlock(&dentry->d_lock);
1560                 return;
1561         }
1562         __d_drop(dentry);
1563         spin_unlock(&dentry->d_lock);
1564
1565         /* Negative dentries can be dropped without further checks */
1566         if (!dentry->d_inode)
1567                 return;
1568
1569         shrink_dcache_parent(dentry);
1570         for (;;) {
1571                 struct dentry *victim = NULL;
1572                 d_walk(dentry, &victim, find_submount);
1573                 if (!victim) {
1574                         if (had_submounts)
1575                                 shrink_dcache_parent(dentry);
1576                         return;
1577                 }
1578                 had_submounts = true;
1579                 detach_mounts(victim);
1580                 dput(victim);
1581         }
1582 }
1583 EXPORT_SYMBOL(d_invalidate);
1584
1585 /**
1586  * __d_alloc    -       allocate a dcache entry
1587  * @sb: filesystem it will belong to
1588  * @name: qstr of the name
1589  *
1590  * Allocates a dentry. It returns %NULL if there is insufficient memory
1591  * available. On a success the dentry is returned. The name passed in is
1592  * copied and the copy passed in may be reused after this call.
1593  */
1594  
1595 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1596 {
1597         struct external_name *ext = NULL;
1598         struct dentry *dentry;
1599         char *dname;
1600         int err;
1601
1602         dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1603         if (!dentry)
1604                 return NULL;
1605
1606         /*
1607          * We guarantee that the inline name is always NUL-terminated.
1608          * This way the memcpy() done by the name switching in rename
1609          * will still always have a NUL at the end, even if we might
1610          * be overwriting an internal NUL character
1611          */
1612         dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1613         if (unlikely(!name)) {
1614                 name = &slash_name;
1615                 dname = dentry->d_iname;
1616         } else if (name->len > DNAME_INLINE_LEN-1) {
1617                 size_t size = offsetof(struct external_name, name[1]);
1618
1619                 ext = kmalloc(size + name->len, GFP_KERNEL_ACCOUNT);
1620                 if (!ext) {
1621                         kmem_cache_free(dentry_cache, dentry); 
1622                         return NULL;
1623                 }
1624                 atomic_set(&ext->u.count, 1);
1625                 dname = ext->name;
1626         } else  {
1627                 dname = dentry->d_iname;
1628         }       
1629
1630         dentry->d_name.len = name->len;
1631         dentry->d_name.hash = name->hash;
1632         memcpy(dname, name->name, name->len);
1633         dname[name->len] = 0;
1634
1635         /* Make sure we always see the terminating NUL character */
1636         smp_store_release(&dentry->d_name.name, dname); /* ^^^ */
1637
1638         dentry->d_lockref.count = 1;
1639         dentry->d_flags = 0;
1640         spin_lock_init(&dentry->d_lock);
1641         seqcount_init(&dentry->d_seq);
1642         dentry->d_inode = NULL;
1643         dentry->d_parent = dentry;
1644         dentry->d_sb = sb;
1645         dentry->d_op = NULL;
1646         dentry->d_fsdata = NULL;
1647         INIT_HLIST_BL_NODE(&dentry->d_hash);
1648         INIT_LIST_HEAD(&dentry->d_lru);
1649         INIT_LIST_HEAD(&dentry->d_subdirs);
1650         INIT_HLIST_NODE(&dentry->d_u.d_alias);
1651         INIT_LIST_HEAD(&dentry->d_child);
1652         d_set_d_op(dentry, dentry->d_sb->s_d_op);
1653
1654         if (dentry->d_op && dentry->d_op->d_init) {
1655                 err = dentry->d_op->d_init(dentry);
1656                 if (err) {
1657                         if (dname_external(dentry))
1658                                 kfree(external_name(dentry));
1659                         kmem_cache_free(dentry_cache, dentry);
1660                         return NULL;
1661                 }
1662         }
1663
1664         if (unlikely(ext)) {
1665                 pg_data_t *pgdat = page_pgdat(virt_to_page(ext));
1666                 mod_node_page_state(pgdat, NR_INDIRECTLY_RECLAIMABLE_BYTES,
1667                                     ksize(ext));
1668         }
1669
1670         this_cpu_inc(nr_dentry);
1671
1672         return dentry;
1673 }
1674
1675 /**
1676  * d_alloc      -       allocate a dcache entry
1677  * @parent: parent of entry to allocate
1678  * @name: qstr of the name
1679  *
1680  * Allocates a dentry. It returns %NULL if there is insufficient memory
1681  * available. On a success the dentry is returned. The name passed in is
1682  * copied and the copy passed in may be reused after this call.
1683  */
1684 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1685 {
1686         struct dentry *dentry = __d_alloc(parent->d_sb, name);
1687         if (!dentry)
1688                 return NULL;
1689         dentry->d_flags |= DCACHE_RCUACCESS;
1690         spin_lock(&parent->d_lock);
1691         /*
1692          * don't need child lock because it is not subject
1693          * to concurrency here
1694          */
1695         __dget_dlock(parent);
1696         dentry->d_parent = parent;
1697         list_add(&dentry->d_child, &parent->d_subdirs);
1698         spin_unlock(&parent->d_lock);
1699
1700         return dentry;
1701 }
1702 EXPORT_SYMBOL(d_alloc);
1703
1704 struct dentry *d_alloc_anon(struct super_block *sb)
1705 {
1706         return __d_alloc(sb, NULL);
1707 }
1708 EXPORT_SYMBOL(d_alloc_anon);
1709
1710 struct dentry *d_alloc_cursor(struct dentry * parent)
1711 {
1712         struct dentry *dentry = d_alloc_anon(parent->d_sb);
1713         if (dentry) {
1714                 dentry->d_flags |= DCACHE_RCUACCESS | DCACHE_DENTRY_CURSOR;
1715                 dentry->d_parent = dget(parent);
1716         }
1717         return dentry;
1718 }
1719
1720 /**
1721  * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1722  * @sb: the superblock
1723  * @name: qstr of the name
1724  *
1725  * For a filesystem that just pins its dentries in memory and never
1726  * performs lookups at all, return an unhashed IS_ROOT dentry.
1727  */
1728 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1729 {
1730         return __d_alloc(sb, name);
1731 }
1732 EXPORT_SYMBOL(d_alloc_pseudo);
1733
1734 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1735 {
1736         struct qstr q;
1737
1738         q.name = name;
1739         q.hash_len = hashlen_string(parent, name);
1740         return d_alloc(parent, &q);
1741 }
1742 EXPORT_SYMBOL(d_alloc_name);
1743
1744 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1745 {
1746         WARN_ON_ONCE(dentry->d_op);
1747         WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH  |
1748                                 DCACHE_OP_COMPARE       |
1749                                 DCACHE_OP_REVALIDATE    |
1750                                 DCACHE_OP_WEAK_REVALIDATE       |
1751                                 DCACHE_OP_DELETE        |
1752                                 DCACHE_OP_REAL));
1753         dentry->d_op = op;
1754         if (!op)
1755                 return;
1756         if (op->d_hash)
1757                 dentry->d_flags |= DCACHE_OP_HASH;
1758         if (op->d_compare)
1759                 dentry->d_flags |= DCACHE_OP_COMPARE;
1760         if (op->d_revalidate)
1761                 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1762         if (op->d_weak_revalidate)
1763                 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1764         if (op->d_delete)
1765                 dentry->d_flags |= DCACHE_OP_DELETE;
1766         if (op->d_prune)
1767                 dentry->d_flags |= DCACHE_OP_PRUNE;
1768         if (op->d_real)
1769                 dentry->d_flags |= DCACHE_OP_REAL;
1770
1771 }
1772 EXPORT_SYMBOL(d_set_d_op);
1773
1774
1775 /*
1776  * d_set_fallthru - Mark a dentry as falling through to a lower layer
1777  * @dentry - The dentry to mark
1778  *
1779  * Mark a dentry as falling through to the lower layer (as set with
1780  * d_pin_lower()).  This flag may be recorded on the medium.
1781  */
1782 void d_set_fallthru(struct dentry *dentry)
1783 {
1784         spin_lock(&dentry->d_lock);
1785         dentry->d_flags |= DCACHE_FALLTHRU;
1786         spin_unlock(&dentry->d_lock);
1787 }
1788 EXPORT_SYMBOL(d_set_fallthru);
1789
1790 static unsigned d_flags_for_inode(struct inode *inode)
1791 {
1792         unsigned add_flags = DCACHE_REGULAR_TYPE;
1793
1794         if (!inode)
1795                 return DCACHE_MISS_TYPE;
1796
1797         if (S_ISDIR(inode->i_mode)) {
1798                 add_flags = DCACHE_DIRECTORY_TYPE;
1799                 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1800                         if (unlikely(!inode->i_op->lookup))
1801                                 add_flags = DCACHE_AUTODIR_TYPE;
1802                         else
1803                                 inode->i_opflags |= IOP_LOOKUP;
1804                 }
1805                 goto type_determined;
1806         }
1807
1808         if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1809                 if (unlikely(inode->i_op->get_link)) {
1810                         add_flags = DCACHE_SYMLINK_TYPE;
1811                         goto type_determined;
1812                 }
1813                 inode->i_opflags |= IOP_NOFOLLOW;
1814         }
1815
1816         if (unlikely(!S_ISREG(inode->i_mode)))
1817                 add_flags = DCACHE_SPECIAL_TYPE;
1818
1819 type_determined:
1820         if (unlikely(IS_AUTOMOUNT(inode)))
1821                 add_flags |= DCACHE_NEED_AUTOMOUNT;
1822         return add_flags;
1823 }
1824
1825 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1826 {
1827         unsigned add_flags = d_flags_for_inode(inode);
1828         WARN_ON(d_in_lookup(dentry));
1829
1830         spin_lock(&dentry->d_lock);
1831         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1832         raw_write_seqcount_begin(&dentry->d_seq);
1833         __d_set_inode_and_type(dentry, inode, add_flags);
1834         raw_write_seqcount_end(&dentry->d_seq);
1835         fsnotify_update_flags(dentry);
1836         spin_unlock(&dentry->d_lock);
1837 }
1838
1839 /**
1840  * d_instantiate - fill in inode information for a dentry
1841  * @entry: dentry to complete
1842  * @inode: inode to attach to this dentry
1843  *
1844  * Fill in inode information in the entry.
1845  *
1846  * This turns negative dentries into productive full members
1847  * of society.
1848  *
1849  * NOTE! This assumes that the inode count has been incremented
1850  * (or otherwise set) by the caller to indicate that it is now
1851  * in use by the dcache.
1852  */
1853  
1854 void d_instantiate(struct dentry *entry, struct inode * inode)
1855 {
1856         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1857         if (inode) {
1858                 security_d_instantiate(entry, inode);
1859                 spin_lock(&inode->i_lock);
1860                 __d_instantiate(entry, inode);
1861                 spin_unlock(&inode->i_lock);
1862         }
1863 }
1864 EXPORT_SYMBOL(d_instantiate);
1865
1866 /**
1867  * d_instantiate_no_diralias - instantiate a non-aliased dentry
1868  * @entry: dentry to complete
1869  * @inode: inode to attach to this dentry
1870  *
1871  * Fill in inode information in the entry.  If a directory alias is found, then
1872  * return an error (and drop inode).  Together with d_materialise_unique() this
1873  * guarantees that a directory inode may never have more than one alias.
1874  */
1875 int d_instantiate_no_diralias(struct dentry *entry, struct inode *inode)
1876 {
1877         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1878
1879         security_d_instantiate(entry, inode);
1880         spin_lock(&inode->i_lock);
1881         if (S_ISDIR(inode->i_mode) && !hlist_empty(&inode->i_dentry)) {
1882                 spin_unlock(&inode->i_lock);
1883                 iput(inode);
1884                 return -EBUSY;
1885         }
1886         __d_instantiate(entry, inode);
1887         spin_unlock(&inode->i_lock);
1888
1889         return 0;
1890 }
1891 EXPORT_SYMBOL(d_instantiate_no_diralias);
1892
1893 struct dentry *d_make_root(struct inode *root_inode)
1894 {
1895         struct dentry *res = NULL;
1896
1897         if (root_inode) {
1898                 res = d_alloc_anon(root_inode->i_sb);
1899                 if (res)
1900                         d_instantiate(res, root_inode);
1901                 else
1902                         iput(root_inode);
1903         }
1904         return res;
1905 }
1906 EXPORT_SYMBOL(d_make_root);
1907
1908 static struct dentry * __d_find_any_alias(struct inode *inode)
1909 {
1910         struct dentry *alias;
1911
1912         if (hlist_empty(&inode->i_dentry))
1913                 return NULL;
1914         alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
1915         __dget(alias);
1916         return alias;
1917 }
1918
1919 /**
1920  * d_find_any_alias - find any alias for a given inode
1921  * @inode: inode to find an alias for
1922  *
1923  * If any aliases exist for the given inode, take and return a
1924  * reference for one of them.  If no aliases exist, return %NULL.
1925  */
1926 struct dentry *d_find_any_alias(struct inode *inode)
1927 {
1928         struct dentry *de;
1929
1930         spin_lock(&inode->i_lock);
1931         de = __d_find_any_alias(inode);
1932         spin_unlock(&inode->i_lock);
1933         return de;
1934 }
1935 EXPORT_SYMBOL(d_find_any_alias);
1936
1937 static struct dentry *__d_instantiate_anon(struct dentry *dentry,
1938                                            struct inode *inode,
1939                                            bool disconnected)
1940 {
1941         struct dentry *res;
1942         unsigned add_flags;
1943
1944         security_d_instantiate(dentry, inode);
1945         spin_lock(&inode->i_lock);
1946         res = __d_find_any_alias(inode);
1947         if (res) {
1948                 spin_unlock(&inode->i_lock);
1949                 dput(dentry);
1950                 goto out_iput;
1951         }
1952
1953         /* attach a disconnected dentry */
1954         add_flags = d_flags_for_inode(inode);
1955
1956         if (disconnected)
1957                 add_flags |= DCACHE_DISCONNECTED;
1958
1959         spin_lock(&dentry->d_lock);
1960         __d_set_inode_and_type(dentry, inode, add_flags);
1961         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1962         if (!disconnected) {
1963                 hlist_bl_lock(&dentry->d_sb->s_roots);
1964                 hlist_bl_add_head(&dentry->d_hash, &dentry->d_sb->s_roots);
1965                 hlist_bl_unlock(&dentry->d_sb->s_roots);
1966         }
1967         spin_unlock(&dentry->d_lock);
1968         spin_unlock(&inode->i_lock);
1969
1970         return dentry;
1971
1972  out_iput:
1973         iput(inode);
1974         return res;
1975 }
1976
1977 struct dentry *d_instantiate_anon(struct dentry *dentry, struct inode *inode)
1978 {
1979         return __d_instantiate_anon(dentry, inode, true);
1980 }
1981 EXPORT_SYMBOL(d_instantiate_anon);
1982
1983 static struct dentry *__d_obtain_alias(struct inode *inode, bool disconnected)
1984 {
1985         struct dentry *tmp;
1986         struct dentry *res;
1987
1988         if (!inode)
1989                 return ERR_PTR(-ESTALE);
1990         if (IS_ERR(inode))
1991                 return ERR_CAST(inode);
1992
1993         res = d_find_any_alias(inode);
1994         if (res)
1995                 goto out_iput;
1996
1997         tmp = d_alloc_anon(inode->i_sb);
1998         if (!tmp) {
1999                 res = ERR_PTR(-ENOMEM);
2000                 goto out_iput;
2001         }
2002
2003         return __d_instantiate_anon(tmp, inode, disconnected);
2004
2005 out_iput:
2006         iput(inode);
2007         return res;
2008 }
2009
2010 /**
2011  * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
2012  * @inode: inode to allocate the dentry for
2013  *
2014  * Obtain a dentry for an inode resulting from NFS filehandle conversion or
2015  * similar open by handle operations.  The returned dentry may be anonymous,
2016  * or may have a full name (if the inode was already in the cache).
2017  *
2018  * When called on a directory inode, we must ensure that the inode only ever
2019  * has one dentry.  If a dentry is found, that is returned instead of
2020  * allocating a new one.
2021  *
2022  * On successful return, the reference to the inode has been transferred
2023  * to the dentry.  In case of an error the reference on the inode is released.
2024  * To make it easier to use in export operations a %NULL or IS_ERR inode may
2025  * be passed in and the error will be propagated to the return value,
2026  * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
2027  */
2028 struct dentry *d_obtain_alias(struct inode *inode)
2029 {
2030         return __d_obtain_alias(inode, true);
2031 }
2032 EXPORT_SYMBOL(d_obtain_alias);
2033
2034 /**
2035  * d_obtain_root - find or allocate a dentry for a given inode
2036  * @inode: inode to allocate the dentry for
2037  *
2038  * Obtain an IS_ROOT dentry for the root of a filesystem.
2039  *
2040  * We must ensure that directory inodes only ever have one dentry.  If a
2041  * dentry is found, that is returned instead of allocating a new one.
2042  *
2043  * On successful return, the reference to the inode has been transferred
2044  * to the dentry.  In case of an error the reference on the inode is
2045  * released.  A %NULL or IS_ERR inode may be passed in and will be the
2046  * error will be propagate to the return value, with a %NULL @inode
2047  * replaced by ERR_PTR(-ESTALE).
2048  */
2049 struct dentry *d_obtain_root(struct inode *inode)
2050 {
2051         return __d_obtain_alias(inode, false);
2052 }
2053 EXPORT_SYMBOL(d_obtain_root);
2054
2055 /**
2056  * d_add_ci - lookup or allocate new dentry with case-exact name
2057  * @inode:  the inode case-insensitive lookup has found
2058  * @dentry: the negative dentry that was passed to the parent's lookup func
2059  * @name:   the case-exact name to be associated with the returned dentry
2060  *
2061  * This is to avoid filling the dcache with case-insensitive names to the
2062  * same inode, only the actual correct case is stored in the dcache for
2063  * case-insensitive filesystems.
2064  *
2065  * For a case-insensitive lookup match and if the the case-exact dentry
2066  * already exists in in the dcache, use it and return it.
2067  *
2068  * If no entry exists with the exact case name, allocate new dentry with
2069  * the exact case, and return the spliced entry.
2070  */
2071 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2072                         struct qstr *name)
2073 {
2074         struct dentry *found, *res;
2075
2076         /*
2077          * First check if a dentry matching the name already exists,
2078          * if not go ahead and create it now.
2079          */
2080         found = d_hash_and_lookup(dentry->d_parent, name);
2081         if (found) {
2082                 iput(inode);
2083                 return found;
2084         }
2085         if (d_in_lookup(dentry)) {
2086                 found = d_alloc_parallel(dentry->d_parent, name,
2087                                         dentry->d_wait);
2088                 if (IS_ERR(found) || !d_in_lookup(found)) {
2089                         iput(inode);
2090                         return found;
2091                 }
2092         } else {
2093                 found = d_alloc(dentry->d_parent, name);
2094                 if (!found) {
2095                         iput(inode);
2096                         return ERR_PTR(-ENOMEM);
2097                 } 
2098         }
2099         res = d_splice_alias(inode, found);
2100         if (res) {
2101                 dput(found);
2102                 return res;
2103         }
2104         return found;
2105 }
2106 EXPORT_SYMBOL(d_add_ci);
2107
2108
2109 static inline bool d_same_name(const struct dentry *dentry,
2110                                 const struct dentry *parent,
2111                                 const struct qstr *name)
2112 {
2113         if (likely(!(parent->d_flags & DCACHE_OP_COMPARE))) {
2114                 if (dentry->d_name.len != name->len)
2115                         return false;
2116                 return dentry_cmp(dentry, name->name, name->len) == 0;
2117         }
2118         return parent->d_op->d_compare(dentry,
2119                                        dentry->d_name.len, dentry->d_name.name,
2120                                        name) == 0;
2121 }
2122
2123 /**
2124  * __d_lookup_rcu - search for a dentry (racy, store-free)
2125  * @parent: parent dentry
2126  * @name: qstr of name we wish to find
2127  * @seqp: returns d_seq value at the point where the dentry was found
2128  * Returns: dentry, or NULL
2129  *
2130  * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2131  * resolution (store-free path walking) design described in
2132  * Documentation/filesystems/path-lookup.txt.
2133  *
2134  * This is not to be used outside core vfs.
2135  *
2136  * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2137  * held, and rcu_read_lock held. The returned dentry must not be stored into
2138  * without taking d_lock and checking d_seq sequence count against @seq
2139  * returned here.
2140  *
2141  * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2142  * function.
2143  *
2144  * Alternatively, __d_lookup_rcu may be called again to look up the child of
2145  * the returned dentry, so long as its parent's seqlock is checked after the
2146  * child is looked up. Thus, an interlocking stepping of sequence lock checks
2147  * is formed, giving integrity down the path walk.
2148  *
2149  * NOTE! The caller *has* to check the resulting dentry against the sequence
2150  * number we've returned before using any of the resulting dentry state!
2151  */
2152 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2153                                 const struct qstr *name,
2154                                 unsigned *seqp)
2155 {
2156         u64 hashlen = name->hash_len;
2157         const unsigned char *str = name->name;
2158         struct hlist_bl_head *b = d_hash(hashlen_hash(hashlen));
2159         struct hlist_bl_node *node;
2160         struct dentry *dentry;
2161
2162         /*
2163          * Note: There is significant duplication with __d_lookup_rcu which is
2164          * required to prevent single threaded performance regressions
2165          * especially on architectures where smp_rmb (in seqcounts) are costly.
2166          * Keep the two functions in sync.
2167          */
2168
2169         /*
2170          * The hash list is protected using RCU.
2171          *
2172          * Carefully use d_seq when comparing a candidate dentry, to avoid
2173          * races with d_move().
2174          *
2175          * It is possible that concurrent renames can mess up our list
2176          * walk here and result in missing our dentry, resulting in the
2177          * false-negative result. d_lookup() protects against concurrent
2178          * renames using rename_lock seqlock.
2179          *
2180          * See Documentation/filesystems/path-lookup.txt for more details.
2181          */
2182         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2183                 unsigned seq;
2184
2185 seqretry:
2186                 /*
2187                  * The dentry sequence count protects us from concurrent
2188                  * renames, and thus protects parent and name fields.
2189                  *
2190                  * The caller must perform a seqcount check in order
2191                  * to do anything useful with the returned dentry.
2192                  *
2193                  * NOTE! We do a "raw" seqcount_begin here. That means that
2194                  * we don't wait for the sequence count to stabilize if it
2195                  * is in the middle of a sequence change. If we do the slow
2196                  * dentry compare, we will do seqretries until it is stable,
2197                  * and if we end up with a successful lookup, we actually
2198                  * want to exit RCU lookup anyway.
2199                  *
2200                  * Note that raw_seqcount_begin still *does* smp_rmb(), so
2201                  * we are still guaranteed NUL-termination of ->d_name.name.
2202                  */
2203                 seq = raw_seqcount_begin(&dentry->d_seq);
2204                 if (dentry->d_parent != parent)
2205                         continue;
2206                 if (d_unhashed(dentry))
2207                         continue;
2208
2209                 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) {
2210                         int tlen;
2211                         const char *tname;
2212                         if (dentry->d_name.hash != hashlen_hash(hashlen))
2213                                 continue;
2214                         tlen = dentry->d_name.len;
2215                         tname = dentry->d_name.name;
2216                         /* we want a consistent (name,len) pair */
2217                         if (read_seqcount_retry(&dentry->d_seq, seq)) {
2218                                 cpu_relax();
2219                                 goto seqretry;
2220                         }
2221                         if (parent->d_op->d_compare(dentry,
2222                                                     tlen, tname, name) != 0)
2223                                 continue;
2224                 } else {
2225                         if (dentry->d_name.hash_len != hashlen)
2226                                 continue;
2227                         if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
2228                                 continue;
2229                 }
2230                 *seqp = seq;
2231                 return dentry;
2232         }
2233         return NULL;
2234 }
2235
2236 /**
2237  * d_lookup - search for a dentry
2238  * @parent: parent dentry
2239  * @name: qstr of name we wish to find
2240  * Returns: dentry, or NULL
2241  *
2242  * d_lookup searches the children of the parent dentry for the name in
2243  * question. If the dentry is found its reference count is incremented and the
2244  * dentry is returned. The caller must use dput to free the entry when it has
2245  * finished using it. %NULL is returned if the dentry does not exist.
2246  */
2247 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2248 {
2249         struct dentry *dentry;
2250         unsigned seq;
2251
2252         do {
2253                 seq = read_seqbegin(&rename_lock);
2254                 dentry = __d_lookup(parent, name);
2255                 if (dentry)
2256                         break;
2257         } while (read_seqretry(&rename_lock, seq));
2258         return dentry;
2259 }
2260 EXPORT_SYMBOL(d_lookup);
2261
2262 /**
2263  * __d_lookup - search for a dentry (racy)
2264  * @parent: parent dentry
2265  * @name: qstr of name we wish to find
2266  * Returns: dentry, or NULL
2267  *
2268  * __d_lookup is like d_lookup, however it may (rarely) return a
2269  * false-negative result due to unrelated rename activity.
2270  *
2271  * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2272  * however it must be used carefully, eg. with a following d_lookup in
2273  * the case of failure.
2274  *
2275  * __d_lookup callers must be commented.
2276  */
2277 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2278 {
2279         unsigned int hash = name->hash;
2280         struct hlist_bl_head *b = d_hash(hash);
2281         struct hlist_bl_node *node;
2282         struct dentry *found = NULL;
2283         struct dentry *dentry;
2284
2285         /*
2286          * Note: There is significant duplication with __d_lookup_rcu which is
2287          * required to prevent single threaded performance regressions
2288          * especially on architectures where smp_rmb (in seqcounts) are costly.
2289          * Keep the two functions in sync.
2290          */
2291
2292         /*
2293          * The hash list is protected using RCU.
2294          *
2295          * Take d_lock when comparing a candidate dentry, to avoid races
2296          * with d_move().
2297          *
2298          * It is possible that concurrent renames can mess up our list
2299          * walk here and result in missing our dentry, resulting in the
2300          * false-negative result. d_lookup() protects against concurrent
2301          * renames using rename_lock seqlock.
2302          *
2303          * See Documentation/filesystems/path-lookup.txt for more details.
2304          */
2305         rcu_read_lock();
2306         
2307         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2308
2309                 if (dentry->d_name.hash != hash)
2310                         continue;
2311
2312                 spin_lock(&dentry->d_lock);
2313                 if (dentry->d_parent != parent)
2314                         goto next;
2315                 if (d_unhashed(dentry))
2316                         goto next;
2317
2318                 if (!d_same_name(dentry, parent, name))
2319                         goto next;
2320
2321                 dentry->d_lockref.count++;
2322                 found = dentry;
2323                 spin_unlock(&dentry->d_lock);
2324                 break;
2325 next:
2326                 spin_unlock(&dentry->d_lock);
2327         }
2328         rcu_read_unlock();
2329
2330         return found;
2331 }
2332
2333 /**
2334  * d_hash_and_lookup - hash the qstr then search for a dentry
2335  * @dir: Directory to search in
2336  * @name: qstr of name we wish to find
2337  *
2338  * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2339  */
2340 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2341 {
2342         /*
2343          * Check for a fs-specific hash function. Note that we must
2344          * calculate the standard hash first, as the d_op->d_hash()
2345          * routine may choose to leave the hash value unchanged.
2346          */
2347         name->hash = full_name_hash(dir, name->name, name->len);
2348         if (dir->d_flags & DCACHE_OP_HASH) {
2349                 int err = dir->d_op->d_hash(dir, name);
2350                 if (unlikely(err < 0))
2351                         return ERR_PTR(err);
2352         }
2353         return d_lookup(dir, name);
2354 }
2355 EXPORT_SYMBOL(d_hash_and_lookup);
2356
2357 /*
2358  * When a file is deleted, we have two options:
2359  * - turn this dentry into a negative dentry
2360  * - unhash this dentry and free it.
2361  *
2362  * Usually, we want to just turn this into
2363  * a negative dentry, but if anybody else is
2364  * currently using the dentry or the inode
2365  * we can't do that and we fall back on removing
2366  * it from the hash queues and waiting for
2367  * it to be deleted later when it has no users
2368  */
2369  
2370 /**
2371  * d_delete - delete a dentry
2372  * @dentry: The dentry to delete
2373  *
2374  * Turn the dentry into a negative dentry if possible, otherwise
2375  * remove it from the hash queues so it can be deleted later
2376  */
2377  
2378 void d_delete(struct dentry * dentry)
2379 {
2380         struct inode *inode = dentry->d_inode;
2381         int isdir = d_is_dir(dentry);
2382
2383         spin_lock(&inode->i_lock);
2384         spin_lock(&dentry->d_lock);
2385         /*
2386          * Are we the only user?
2387          */
2388         if (dentry->d_lockref.count == 1) {
2389                 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2390                 dentry_unlink_inode(dentry);
2391         } else {
2392                 __d_drop(dentry);
2393                 spin_unlock(&dentry->d_lock);
2394                 spin_unlock(&inode->i_lock);
2395         }
2396         fsnotify_nameremove(dentry, isdir);
2397 }
2398 EXPORT_SYMBOL(d_delete);
2399
2400 static void __d_rehash(struct dentry *entry)
2401 {
2402         struct hlist_bl_head *b = d_hash(entry->d_name.hash);
2403
2404         hlist_bl_lock(b);
2405         hlist_bl_add_head_rcu(&entry->d_hash, b);
2406         hlist_bl_unlock(b);
2407 }
2408
2409 /**
2410  * d_rehash     - add an entry back to the hash
2411  * @entry: dentry to add to the hash
2412  *
2413  * Adds a dentry to the hash according to its name.
2414  */
2415  
2416 void d_rehash(struct dentry * entry)
2417 {
2418         spin_lock(&entry->d_lock);
2419         __d_rehash(entry);
2420         spin_unlock(&entry->d_lock);
2421 }
2422 EXPORT_SYMBOL(d_rehash);
2423
2424 static inline unsigned start_dir_add(struct inode *dir)
2425 {
2426
2427         for (;;) {
2428                 unsigned n = dir->i_dir_seq;
2429                 if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
2430                         return n;
2431                 cpu_relax();
2432         }
2433 }
2434
2435 static inline void end_dir_add(struct inode *dir, unsigned n)
2436 {
2437         smp_store_release(&dir->i_dir_seq, n + 2);
2438 }
2439
2440 static void d_wait_lookup(struct dentry *dentry)
2441 {
2442         if (d_in_lookup(dentry)) {
2443                 DECLARE_WAITQUEUE(wait, current);
2444                 add_wait_queue(dentry->d_wait, &wait);
2445                 do {
2446                         set_current_state(TASK_UNINTERRUPTIBLE);
2447                         spin_unlock(&dentry->d_lock);
2448                         schedule();
2449                         spin_lock(&dentry->d_lock);
2450                 } while (d_in_lookup(dentry));
2451         }
2452 }
2453
2454 struct dentry *d_alloc_parallel(struct dentry *parent,
2455                                 const struct qstr *name,
2456                                 wait_queue_head_t *wq)
2457 {
2458         unsigned int hash = name->hash;
2459         struct hlist_bl_head *b = in_lookup_hash(parent, hash);
2460         struct hlist_bl_node *node;
2461         struct dentry *new = d_alloc(parent, name);
2462         struct dentry *dentry;
2463         unsigned seq, r_seq, d_seq;
2464
2465         if (unlikely(!new))
2466                 return ERR_PTR(-ENOMEM);
2467
2468 retry:
2469         rcu_read_lock();
2470         seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
2471         r_seq = read_seqbegin(&rename_lock);
2472         dentry = __d_lookup_rcu(parent, name, &d_seq);
2473         if (unlikely(dentry)) {
2474                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2475                         rcu_read_unlock();
2476                         goto retry;
2477                 }
2478                 if (read_seqcount_retry(&dentry->d_seq, d_seq)) {
2479                         rcu_read_unlock();
2480                         dput(dentry);
2481                         goto retry;
2482                 }
2483                 rcu_read_unlock();
2484                 dput(new);
2485                 return dentry;
2486         }
2487         if (unlikely(read_seqretry(&rename_lock, r_seq))) {
2488                 rcu_read_unlock();
2489                 goto retry;
2490         }
2491
2492         if (unlikely(seq & 1)) {
2493                 rcu_read_unlock();
2494                 goto retry;
2495         }
2496
2497         hlist_bl_lock(b);
2498         if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {
2499                 hlist_bl_unlock(b);
2500                 rcu_read_unlock();
2501                 goto retry;
2502         }
2503         /*
2504          * No changes for the parent since the beginning of d_lookup().
2505          * Since all removals from the chain happen with hlist_bl_lock(),
2506          * any potential in-lookup matches are going to stay here until
2507          * we unlock the chain.  All fields are stable in everything
2508          * we encounter.
2509          */
2510         hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
2511                 if (dentry->d_name.hash != hash)
2512                         continue;
2513                 if (dentry->d_parent != parent)
2514                         continue;
2515                 if (!d_same_name(dentry, parent, name))
2516                         continue;
2517                 hlist_bl_unlock(b);
2518                 /* now we can try to grab a reference */
2519                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2520                         rcu_read_unlock();
2521                         goto retry;
2522                 }
2523
2524                 rcu_read_unlock();
2525                 /*
2526                  * somebody is likely to be still doing lookup for it;
2527                  * wait for them to finish
2528                  */
2529                 spin_lock(&dentry->d_lock);
2530                 d_wait_lookup(dentry);
2531                 /*
2532                  * it's not in-lookup anymore; in principle we should repeat
2533                  * everything from dcache lookup, but it's likely to be what
2534                  * d_lookup() would've found anyway.  If it is, just return it;
2535                  * otherwise we really have to repeat the whole thing.
2536                  */
2537                 if (unlikely(dentry->d_name.hash != hash))
2538                         goto mismatch;
2539                 if (unlikely(dentry->d_parent != parent))
2540                         goto mismatch;
2541                 if (unlikely(d_unhashed(dentry)))
2542                         goto mismatch;
2543                 if (unlikely(!d_same_name(dentry, parent, name)))
2544                         goto mismatch;
2545                 /* OK, it *is* a hashed match; return it */
2546                 spin_unlock(&dentry->d_lock);
2547                 dput(new);
2548                 return dentry;
2549         }
2550         rcu_read_unlock();
2551         /* we can't take ->d_lock here; it's OK, though. */
2552         new->d_flags |= DCACHE_PAR_LOOKUP;
2553         new->d_wait = wq;
2554         hlist_bl_add_head_rcu(&new->d_u.d_in_lookup_hash, b);
2555         hlist_bl_unlock(b);
2556         return new;
2557 mismatch:
2558         spin_unlock(&dentry->d_lock);
2559         dput(dentry);
2560         goto retry;
2561 }
2562 EXPORT_SYMBOL(d_alloc_parallel);
2563
2564 void __d_lookup_done(struct dentry *dentry)
2565 {
2566         struct hlist_bl_head *b = in_lookup_hash(dentry->d_parent,
2567                                                  dentry->d_name.hash);
2568         hlist_bl_lock(b);
2569         dentry->d_flags &= ~DCACHE_PAR_LOOKUP;
2570         __hlist_bl_del(&dentry->d_u.d_in_lookup_hash);
2571         wake_up_all(dentry->d_wait);
2572         dentry->d_wait = NULL;
2573         hlist_bl_unlock(b);
2574         INIT_HLIST_NODE(&dentry->d_u.d_alias);
2575         INIT_LIST_HEAD(&dentry->d_lru);
2576 }
2577 EXPORT_SYMBOL(__d_lookup_done);
2578
2579 /* inode->i_lock held if inode is non-NULL */
2580
2581 static inline void __d_add(struct dentry *dentry, struct inode *inode)
2582 {
2583         struct inode *dir = NULL;
2584         unsigned n;
2585         spin_lock(&dentry->d_lock);
2586         if (unlikely(d_in_lookup(dentry))) {
2587                 dir = dentry->d_parent->d_inode;
2588                 n = start_dir_add(dir);
2589                 __d_lookup_done(dentry);
2590         }
2591         if (inode) {
2592                 unsigned add_flags = d_flags_for_inode(inode);
2593                 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2594                 raw_write_seqcount_begin(&dentry->d_seq);
2595                 __d_set_inode_and_type(dentry, inode, add_flags);
2596                 raw_write_seqcount_end(&dentry->d_seq);
2597                 fsnotify_update_flags(dentry);
2598         }
2599         __d_rehash(dentry);
2600         if (dir)
2601                 end_dir_add(dir, n);
2602         spin_unlock(&dentry->d_lock);
2603         if (inode)
2604                 spin_unlock(&inode->i_lock);
2605 }
2606
2607 /**
2608  * d_add - add dentry to hash queues
2609  * @entry: dentry to add
2610  * @inode: The inode to attach to this dentry
2611  *
2612  * This adds the entry to the hash queues and initializes @inode.
2613  * The entry was actually filled in earlier during d_alloc().
2614  */
2615
2616 void d_add(struct dentry *entry, struct inode *inode)
2617 {
2618         if (inode) {
2619                 security_d_instantiate(entry, inode);
2620                 spin_lock(&inode->i_lock);
2621         }
2622         __d_add(entry, inode);
2623 }
2624 EXPORT_SYMBOL(d_add);
2625
2626 /**
2627  * d_exact_alias - find and hash an exact unhashed alias
2628  * @entry: dentry to add
2629  * @inode: The inode to go with this dentry
2630  *
2631  * If an unhashed dentry with the same name/parent and desired
2632  * inode already exists, hash and return it.  Otherwise, return
2633  * NULL.
2634  *
2635  * Parent directory should be locked.
2636  */
2637 struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2638 {
2639         struct dentry *alias;
2640         unsigned int hash = entry->d_name.hash;
2641
2642         spin_lock(&inode->i_lock);
2643         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2644                 /*
2645                  * Don't need alias->d_lock here, because aliases with
2646                  * d_parent == entry->d_parent are not subject to name or
2647                  * parent changes, because the parent inode i_mutex is held.
2648                  */
2649                 if (alias->d_name.hash != hash)
2650                         continue;
2651                 if (alias->d_parent != entry->d_parent)
2652                         continue;
2653                 if (!d_same_name(alias, entry->d_parent, &entry->d_name))
2654                         continue;
2655                 spin_lock(&alias->d_lock);
2656                 if (!d_unhashed(alias)) {
2657                         spin_unlock(&alias->d_lock);
2658                         alias = NULL;
2659                 } else {
2660                         __dget_dlock(alias);
2661                         __d_rehash(alias);
2662                         spin_unlock(&alias->d_lock);
2663                 }
2664                 spin_unlock(&inode->i_lock);
2665                 return alias;
2666         }
2667         spin_unlock(&inode->i_lock);
2668         return NULL;
2669 }
2670 EXPORT_SYMBOL(d_exact_alias);
2671
2672 /**
2673  * dentry_update_name_case - update case insensitive dentry with a new name
2674  * @dentry: dentry to be updated
2675  * @name: new name
2676  *
2677  * Update a case insensitive dentry with new case of name.
2678  *
2679  * dentry must have been returned by d_lookup with name @name. Old and new
2680  * name lengths must match (ie. no d_compare which allows mismatched name
2681  * lengths).
2682  *
2683  * Parent inode i_mutex must be held over d_lookup and into this call (to
2684  * keep renames and concurrent inserts, and readdir(2) away).
2685  */
2686 void dentry_update_name_case(struct dentry *dentry, const struct qstr *name)
2687 {
2688         BUG_ON(!inode_is_locked(dentry->d_parent->d_inode));
2689         BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */
2690
2691         spin_lock(&dentry->d_lock);
2692         write_seqcount_begin(&dentry->d_seq);
2693         memcpy((unsigned char *)dentry->d_name.name, name->name, name->len);
2694         write_seqcount_end(&dentry->d_seq);
2695         spin_unlock(&dentry->d_lock);
2696 }
2697 EXPORT_SYMBOL(dentry_update_name_case);
2698
2699 static void swap_names(struct dentry *dentry, struct dentry *target)
2700 {
2701         if (unlikely(dname_external(target))) {
2702                 if (unlikely(dname_external(dentry))) {
2703                         /*
2704                          * Both external: swap the pointers
2705                          */
2706                         swap(target->d_name.name, dentry->d_name.name);
2707                 } else {
2708                         /*
2709                          * dentry:internal, target:external.  Steal target's
2710                          * storage and make target internal.
2711                          */
2712                         memcpy(target->d_iname, dentry->d_name.name,
2713                                         dentry->d_name.len + 1);
2714                         dentry->d_name.name = target->d_name.name;
2715                         target->d_name.name = target->d_iname;
2716                 }
2717         } else {
2718                 if (unlikely(dname_external(dentry))) {
2719                         /*
2720                          * dentry:external, target:internal.  Give dentry's
2721                          * storage to target and make dentry internal
2722                          */
2723                         memcpy(dentry->d_iname, target->d_name.name,
2724                                         target->d_name.len + 1);
2725                         target->d_name.name = dentry->d_name.name;
2726                         dentry->d_name.name = dentry->d_iname;
2727                 } else {
2728                         /*
2729                          * Both are internal.
2730                          */
2731                         unsigned int i;
2732                         BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2733                         for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2734                                 swap(((long *) &dentry->d_iname)[i],
2735                                      ((long *) &target->d_iname)[i]);
2736                         }
2737                 }
2738         }
2739         swap(dentry->d_name.hash_len, target->d_name.hash_len);
2740 }
2741
2742 static void copy_name(struct dentry *dentry, struct dentry *target)
2743 {
2744         struct external_name *old_name = NULL;
2745         if (unlikely(dname_external(dentry)))
2746                 old_name = external_name(dentry);
2747         if (unlikely(dname_external(target))) {
2748                 atomic_inc(&external_name(target)->u.count);
2749                 dentry->d_name = target->d_name;
2750         } else {
2751                 memcpy(dentry->d_iname, target->d_name.name,
2752                                 target->d_name.len + 1);
2753                 dentry->d_name.name = dentry->d_iname;
2754                 dentry->d_name.hash_len = target->d_name.hash_len;
2755         }
2756         if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2757                 call_rcu(&old_name->u.head, __d_free_external_name);
2758 }
2759
2760 /*
2761  * __d_move - move a dentry
2762  * @dentry: entry to move
2763  * @target: new dentry
2764  * @exchange: exchange the two dentries
2765  *
2766  * Update the dcache to reflect the move of a file name. Negative
2767  * dcache entries should not be moved in this way. Caller must hold
2768  * rename_lock, the i_mutex of the source and target directories,
2769  * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2770  */
2771 static void __d_move(struct dentry *dentry, struct dentry *target,
2772                      bool exchange)
2773 {
2774         struct dentry *old_parent, *p;
2775         struct inode *dir = NULL;
2776         unsigned n;
2777
2778         WARN_ON(!dentry->d_inode);
2779         if (WARN_ON(dentry == target))
2780                 return;
2781
2782         BUG_ON(d_ancestor(target, dentry));
2783         old_parent = dentry->d_parent;
2784         p = d_ancestor(old_parent, target);
2785         if (IS_ROOT(dentry)) {
2786                 BUG_ON(p);
2787                 spin_lock(&target->d_parent->d_lock);
2788         } else if (!p) {
2789                 /* target is not a descendent of dentry->d_parent */
2790                 spin_lock(&target->d_parent->d_lock);
2791                 spin_lock_nested(&old_parent->d_lock, DENTRY_D_LOCK_NESTED);
2792         } else {
2793                 BUG_ON(p == dentry);
2794                 spin_lock(&old_parent->d_lock);
2795                 if (p != target)
2796                         spin_lock_nested(&target->d_parent->d_lock,
2797                                         DENTRY_D_LOCK_NESTED);
2798         }
2799         spin_lock_nested(&dentry->d_lock, 2);
2800         spin_lock_nested(&target->d_lock, 3);
2801
2802         if (unlikely(d_in_lookup(target))) {
2803                 dir = target->d_parent->d_inode;
2804                 n = start_dir_add(dir);
2805                 __d_lookup_done(target);
2806         }
2807
2808         write_seqcount_begin(&dentry->d_seq);
2809         write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2810
2811         /* unhash both */
2812         if (!d_unhashed(dentry))
2813                 ___d_drop(dentry);
2814         if (!d_unhashed(target))
2815                 ___d_drop(target);
2816
2817         /* ... and switch them in the tree */
2818         dentry->d_parent = target->d_parent;
2819         if (!exchange) {
2820                 copy_name(dentry, target);
2821                 target->d_hash.pprev = NULL;
2822                 dentry->d_parent->d_lockref.count++;
2823                 if (dentry == old_parent)
2824                         dentry->d_flags |= DCACHE_RCUACCESS;
2825                 else
2826                         WARN_ON(!--old_parent->d_lockref.count);
2827         } else {
2828                 target->d_parent = old_parent;
2829                 swap_names(dentry, target);
2830                 list_move(&target->d_child, &target->d_parent->d_subdirs);
2831                 __d_rehash(target);
2832                 fsnotify_update_flags(target);
2833         }
2834         list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2835         __d_rehash(dentry);
2836         fsnotify_update_flags(dentry);
2837
2838         write_seqcount_end(&target->d_seq);
2839         write_seqcount_end(&dentry->d_seq);
2840
2841         if (dir)
2842                 end_dir_add(dir, n);
2843
2844         if (dentry->d_parent != old_parent)
2845                 spin_unlock(&dentry->d_parent->d_lock);
2846         if (dentry != old_parent)
2847                 spin_unlock(&old_parent->d_lock);
2848         spin_unlock(&target->d_lock);
2849         spin_unlock(&dentry->d_lock);
2850 }
2851
2852 /*
2853  * d_move - move a dentry
2854  * @dentry: entry to move
2855  * @target: new dentry
2856  *
2857  * Update the dcache to reflect the move of a file name. Negative
2858  * dcache entries should not be moved in this way. See the locking
2859  * requirements for __d_move.
2860  */
2861 void d_move(struct dentry *dentry, struct dentry *target)
2862 {
2863         write_seqlock(&rename_lock);
2864         __d_move(dentry, target, false);
2865         write_sequnlock(&rename_lock);
2866 }
2867 EXPORT_SYMBOL(d_move);
2868
2869 /*
2870  * d_exchange - exchange two dentries
2871  * @dentry1: first dentry
2872  * @dentry2: second dentry
2873  */
2874 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2875 {
2876         write_seqlock(&rename_lock);
2877
2878         WARN_ON(!dentry1->d_inode);
2879         WARN_ON(!dentry2->d_inode);
2880         WARN_ON(IS_ROOT(dentry1));
2881         WARN_ON(IS_ROOT(dentry2));
2882
2883         __d_move(dentry1, dentry2, true);
2884
2885         write_sequnlock(&rename_lock);
2886 }
2887
2888 /**
2889  * d_ancestor - search for an ancestor
2890  * @p1: ancestor dentry
2891  * @p2: child dentry
2892  *
2893  * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2894  * an ancestor of p2, else NULL.
2895  */
2896 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2897 {
2898         struct dentry *p;
2899
2900         for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2901                 if (p->d_parent == p1)
2902                         return p;
2903         }
2904         return NULL;
2905 }
2906
2907 /*
2908  * This helper attempts to cope with remotely renamed directories
2909  *
2910  * It assumes that the caller is already holding
2911  * dentry->d_parent->d_inode->i_mutex, and rename_lock
2912  *
2913  * Note: If ever the locking in lock_rename() changes, then please
2914  * remember to update this too...
2915  */
2916 static int __d_unalias(struct inode *inode,
2917                 struct dentry *dentry, struct dentry *alias)
2918 {
2919         struct mutex *m1 = NULL;
2920         struct rw_semaphore *m2 = NULL;
2921         int ret = -ESTALE;
2922
2923         /* If alias and dentry share a parent, then no extra locks required */
2924         if (alias->d_parent == dentry->d_parent)
2925                 goto out_unalias;
2926
2927         /* See lock_rename() */
2928         if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2929                 goto out_err;
2930         m1 = &dentry->d_sb->s_vfs_rename_mutex;
2931         if (!inode_trylock_shared(alias->d_parent->d_inode))
2932                 goto out_err;
2933         m2 = &alias->d_parent->d_inode->i_rwsem;
2934 out_unalias:
2935         __d_move(alias, dentry, false);
2936         ret = 0;
2937 out_err:
2938         if (m2)
2939                 up_read(m2);
2940         if (m1)
2941                 mutex_unlock(m1);
2942         return ret;
2943 }
2944
2945 /**
2946  * d_splice_alias - splice a disconnected dentry into the tree if one exists
2947  * @inode:  the inode which may have a disconnected dentry
2948  * @dentry: a negative dentry which we want to point to the inode.
2949  *
2950  * If inode is a directory and has an IS_ROOT alias, then d_move that in
2951  * place of the given dentry and return it, else simply d_add the inode
2952  * to the dentry and return NULL.
2953  *
2954  * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2955  * we should error out: directories can't have multiple aliases.
2956  *
2957  * This is needed in the lookup routine of any filesystem that is exportable
2958  * (via knfsd) so that we can build dcache paths to directories effectively.
2959  *
2960  * If a dentry was found and moved, then it is returned.  Otherwise NULL
2961  * is returned.  This matches the expected return value of ->lookup.
2962  *
2963  * Cluster filesystems may call this function with a negative, hashed dentry.
2964  * In that case, we know that the inode will be a regular file, and also this
2965  * will only occur during atomic_open. So we need to check for the dentry
2966  * being already hashed only in the final case.
2967  */
2968 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
2969 {
2970         if (IS_ERR(inode))
2971                 return ERR_CAST(inode);
2972
2973         BUG_ON(!d_unhashed(dentry));
2974
2975         if (!inode)
2976                 goto out;
2977
2978         security_d_instantiate(dentry, inode);
2979         spin_lock(&inode->i_lock);
2980         if (S_ISDIR(inode->i_mode)) {
2981                 struct dentry *new = __d_find_any_alias(inode);
2982                 if (unlikely(new)) {
2983                         /* The reference to new ensures it remains an alias */
2984                         spin_unlock(&inode->i_lock);
2985                         write_seqlock(&rename_lock);
2986                         if (unlikely(d_ancestor(new, dentry))) {
2987                                 write_sequnlock(&rename_lock);
2988                                 dput(new);
2989                                 new = ERR_PTR(-ELOOP);
2990                                 pr_warn_ratelimited(
2991                                         "VFS: Lookup of '%s' in %s %s"
2992                                         " would have caused loop\n",
2993                                         dentry->d_name.name,
2994                                         inode->i_sb->s_type->name,
2995                                         inode->i_sb->s_id);
2996                         } else if (!IS_ROOT(new)) {
2997                                 struct dentry *old_parent = dget(new->d_parent);
2998                                 int err = __d_unalias(inode, dentry, new);
2999                                 write_sequnlock(&rename_lock);
3000                                 if (err) {
3001                                         dput(new);
3002                                         new = ERR_PTR(err);
3003                                 }
3004                                 dput(old_parent);
3005                         } else {
3006                                 __d_move(new, dentry, false);
3007                                 write_sequnlock(&rename_lock);
3008                         }
3009                         iput(inode);
3010                         return new;
3011                 }
3012         }
3013 out:
3014         __d_add(dentry, inode);
3015         return NULL;
3016 }
3017 EXPORT_SYMBOL(d_splice_alias);
3018
3019 /*
3020  * Test whether new_dentry is a subdirectory of old_dentry.
3021  *
3022  * Trivially implemented using the dcache structure
3023  */
3024
3025 /**
3026  * is_subdir - is new dentry a subdirectory of old_dentry
3027  * @new_dentry: new dentry
3028  * @old_dentry: old dentry
3029  *
3030  * Returns true if new_dentry is a subdirectory of the parent (at any depth).
3031  * Returns false otherwise.
3032  * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3033  */
3034   
3035 bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3036 {
3037         bool result;
3038         unsigned seq;
3039
3040         if (new_dentry == old_dentry)
3041                 return true;
3042
3043         do {
3044                 /* for restarting inner loop in case of seq retry */
3045                 seq = read_seqbegin(&rename_lock);
3046                 /*
3047                  * Need rcu_readlock to protect against the d_parent trashing
3048                  * due to d_move
3049                  */
3050                 rcu_read_lock();
3051                 if (d_ancestor(old_dentry, new_dentry))
3052                         result = true;
3053                 else
3054                         result = false;
3055                 rcu_read_unlock();
3056         } while (read_seqretry(&rename_lock, seq));
3057
3058         return result;
3059 }
3060 EXPORT_SYMBOL(is_subdir);
3061
3062 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3063 {
3064         struct dentry *root = data;
3065         if (dentry != root) {
3066                 if (d_unhashed(dentry) || !dentry->d_inode)
3067                         return D_WALK_SKIP;
3068
3069                 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3070                         dentry->d_flags |= DCACHE_GENOCIDE;
3071                         dentry->d_lockref.count--;
3072                 }
3073         }
3074         return D_WALK_CONTINUE;
3075 }
3076
3077 void d_genocide(struct dentry *parent)
3078 {
3079         d_walk(parent, parent, d_genocide_kill);
3080 }
3081
3082 EXPORT_SYMBOL(d_genocide);
3083
3084 void d_tmpfile(struct dentry *dentry, struct inode *inode)
3085 {
3086         inode_dec_link_count(inode);
3087         BUG_ON(dentry->d_name.name != dentry->d_iname ||
3088                 !hlist_unhashed(&dentry->d_u.d_alias) ||
3089                 !d_unlinked(dentry));
3090         spin_lock(&dentry->d_parent->d_lock);
3091         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3092         dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3093                                 (unsigned long long)inode->i_ino);
3094         spin_unlock(&dentry->d_lock);
3095         spin_unlock(&dentry->d_parent->d_lock);
3096         d_instantiate(dentry, inode);
3097 }
3098 EXPORT_SYMBOL(d_tmpfile);
3099
3100 static __initdata unsigned long dhash_entries;
3101 static int __init set_dhash_entries(char *str)
3102 {
3103         if (!str)
3104                 return 0;
3105         dhash_entries = simple_strtoul(str, &str, 0);
3106         return 1;
3107 }
3108 __setup("dhash_entries=", set_dhash_entries);
3109
3110 static void __init dcache_init_early(void)
3111 {
3112         /* If hashes are distributed across NUMA nodes, defer
3113          * hash allocation until vmalloc space is available.
3114          */
3115         if (hashdist)
3116                 return;
3117
3118         dentry_hashtable =
3119                 alloc_large_system_hash("Dentry cache",
3120                                         sizeof(struct hlist_bl_head),
3121                                         dhash_entries,
3122                                         13,
3123                                         HASH_EARLY | HASH_ZERO,
3124                                         &d_hash_shift,
3125                                         NULL,
3126                                         0,
3127                                         0);
3128         d_hash_shift = 32 - d_hash_shift;
3129 }
3130
3131 static void __init dcache_init(void)
3132 {
3133         /*
3134          * A constructor could be added for stable state like the lists,
3135          * but it is probably not worth it because of the cache nature
3136          * of the dcache.
3137          */
3138         dentry_cache = KMEM_CACHE_USERCOPY(dentry,
3139                 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD|SLAB_ACCOUNT,
3140                 d_iname);
3141
3142         /* Hash may have been set up in dcache_init_early */
3143         if (!hashdist)
3144                 return;
3145
3146         dentry_hashtable =
3147                 alloc_large_system_hash("Dentry cache",
3148                                         sizeof(struct hlist_bl_head),
3149                                         dhash_entries,
3150                                         13,
3151                                         HASH_ZERO,
3152                                         &d_hash_shift,
3153                                         NULL,
3154                                         0,
3155                                         0);
3156         d_hash_shift = 32 - d_hash_shift;
3157 }
3158
3159 /* SLAB cache for __getname() consumers */
3160 struct kmem_cache *names_cachep __read_mostly;
3161 EXPORT_SYMBOL(names_cachep);
3162
3163 void __init vfs_caches_init_early(void)
3164 {
3165         int i;
3166
3167         for (i = 0; i < ARRAY_SIZE(in_lookup_hashtable); i++)
3168                 INIT_HLIST_BL_HEAD(&in_lookup_hashtable[i]);
3169
3170         dcache_init_early();
3171         inode_init_early();
3172 }
3173
3174 void __init vfs_caches_init(void)
3175 {
3176         names_cachep = kmem_cache_create_usercopy("names_cache", PATH_MAX, 0,
3177                         SLAB_HWCACHE_ALIGN|SLAB_PANIC, 0, PATH_MAX, NULL);
3178
3179         dcache_init();
3180         inode_init();
3181         files_init();
3182         files_maxfiles_init();
3183         mnt_init();
3184         bdev_cache_init();
3185         chrdev_init();
3186 }