make non-exchanging __d_move() copy ->d_parent rather than swap them
[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(struct rcu_head *head)
261 {
262         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
263         kfree(external_name(dentry));
264         kmem_cache_free(dentry_cache, dentry); 
265 }
266
267 static inline int dname_external(const struct dentry *dentry)
268 {
269         return dentry->d_name.name != dentry->d_iname;
270 }
271
272 void take_dentry_name_snapshot(struct name_snapshot *name, struct dentry *dentry)
273 {
274         spin_lock(&dentry->d_lock);
275         if (unlikely(dname_external(dentry))) {
276                 struct external_name *p = external_name(dentry);
277                 atomic_inc(&p->u.count);
278                 spin_unlock(&dentry->d_lock);
279                 name->name = p->name;
280         } else {
281                 memcpy(name->inline_name, dentry->d_iname, DNAME_INLINE_LEN);
282                 spin_unlock(&dentry->d_lock);
283                 name->name = name->inline_name;
284         }
285 }
286 EXPORT_SYMBOL(take_dentry_name_snapshot);
287
288 void release_dentry_name_snapshot(struct name_snapshot *name)
289 {
290         if (unlikely(name->name != name->inline_name)) {
291                 struct external_name *p;
292                 p = container_of(name->name, struct external_name, name[0]);
293                 if (unlikely(atomic_dec_and_test(&p->u.count)))
294                         kfree_rcu(p, u.head);
295         }
296 }
297 EXPORT_SYMBOL(release_dentry_name_snapshot);
298
299 static inline void __d_set_inode_and_type(struct dentry *dentry,
300                                           struct inode *inode,
301                                           unsigned type_flags)
302 {
303         unsigned flags;
304
305         dentry->d_inode = inode;
306         flags = READ_ONCE(dentry->d_flags);
307         flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
308         flags |= type_flags;
309         WRITE_ONCE(dentry->d_flags, flags);
310 }
311
312 static inline void __d_clear_type_and_inode(struct dentry *dentry)
313 {
314         unsigned flags = READ_ONCE(dentry->d_flags);
315
316         flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
317         WRITE_ONCE(dentry->d_flags, flags);
318         dentry->d_inode = NULL;
319 }
320
321 static void dentry_free(struct dentry *dentry)
322 {
323         WARN_ON(!hlist_unhashed(&dentry->d_u.d_alias));
324         if (unlikely(dname_external(dentry))) {
325                 struct external_name *p = external_name(dentry);
326                 if (likely(atomic_dec_and_test(&p->u.count))) {
327                         call_rcu(&dentry->d_u.d_rcu, __d_free_external);
328                         return;
329                 }
330         }
331         /* if dentry was never visible to RCU, immediate free is OK */
332         if (!(dentry->d_flags & DCACHE_RCUACCESS))
333                 __d_free(&dentry->d_u.d_rcu);
334         else
335                 call_rcu(&dentry->d_u.d_rcu, __d_free);
336 }
337
338 /*
339  * Release the dentry's inode, using the filesystem
340  * d_iput() operation if defined.
341  */
342 static void dentry_unlink_inode(struct dentry * dentry)
343         __releases(dentry->d_lock)
344         __releases(dentry->d_inode->i_lock)
345 {
346         struct inode *inode = dentry->d_inode;
347         bool hashed = !d_unhashed(dentry);
348
349         if (hashed)
350                 raw_write_seqcount_begin(&dentry->d_seq);
351         __d_clear_type_and_inode(dentry);
352         hlist_del_init(&dentry->d_u.d_alias);
353         if (hashed)
354                 raw_write_seqcount_end(&dentry->d_seq);
355         spin_unlock(&dentry->d_lock);
356         spin_unlock(&inode->i_lock);
357         if (!inode->i_nlink)
358                 fsnotify_inoderemove(inode);
359         if (dentry->d_op && dentry->d_op->d_iput)
360                 dentry->d_op->d_iput(dentry, inode);
361         else
362                 iput(inode);
363 }
364
365 /*
366  * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
367  * is in use - which includes both the "real" per-superblock
368  * LRU list _and_ the DCACHE_SHRINK_LIST use.
369  *
370  * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
371  * on the shrink list (ie not on the superblock LRU list).
372  *
373  * The per-cpu "nr_dentry_unused" counters are updated with
374  * the DCACHE_LRU_LIST bit.
375  *
376  * These helper functions make sure we always follow the
377  * rules. d_lock must be held by the caller.
378  */
379 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
380 static void d_lru_add(struct dentry *dentry)
381 {
382         D_FLAG_VERIFY(dentry, 0);
383         dentry->d_flags |= DCACHE_LRU_LIST;
384         this_cpu_inc(nr_dentry_unused);
385         WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
386 }
387
388 static void d_lru_del(struct dentry *dentry)
389 {
390         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
391         dentry->d_flags &= ~DCACHE_LRU_LIST;
392         this_cpu_dec(nr_dentry_unused);
393         WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
394 }
395
396 static void d_shrink_del(struct dentry *dentry)
397 {
398         D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
399         list_del_init(&dentry->d_lru);
400         dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
401         this_cpu_dec(nr_dentry_unused);
402 }
403
404 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
405 {
406         D_FLAG_VERIFY(dentry, 0);
407         list_add(&dentry->d_lru, list);
408         dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
409         this_cpu_inc(nr_dentry_unused);
410 }
411
412 /*
413  * These can only be called under the global LRU lock, ie during the
414  * callback for freeing the LRU list. "isolate" removes it from the
415  * LRU lists entirely, while shrink_move moves it to the indicated
416  * private list.
417  */
418 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
419 {
420         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
421         dentry->d_flags &= ~DCACHE_LRU_LIST;
422         this_cpu_dec(nr_dentry_unused);
423         list_lru_isolate(lru, &dentry->d_lru);
424 }
425
426 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
427                               struct list_head *list)
428 {
429         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
430         dentry->d_flags |= DCACHE_SHRINK_LIST;
431         list_lru_isolate_move(lru, &dentry->d_lru, list);
432 }
433
434 /**
435  * d_drop - drop a dentry
436  * @dentry: dentry to drop
437  *
438  * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
439  * be found through a VFS lookup any more. Note that this is different from
440  * deleting the dentry - d_delete will try to mark the dentry negative if
441  * possible, giving a successful _negative_ lookup, while d_drop will
442  * just make the cache lookup fail.
443  *
444  * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
445  * reason (NFS timeouts or autofs deletes).
446  *
447  * __d_drop requires dentry->d_lock
448  * ___d_drop doesn't mark dentry as "unhashed"
449  *   (dentry->d_hash.pprev will be LIST_POISON2, not NULL).
450  */
451 static void ___d_drop(struct dentry *dentry)
452 {
453         struct hlist_bl_head *b;
454         /*
455          * Hashed dentries are normally on the dentry hashtable,
456          * with the exception of those newly allocated by
457          * d_obtain_root, which are always IS_ROOT:
458          */
459         if (unlikely(IS_ROOT(dentry)))
460                 b = &dentry->d_sb->s_roots;
461         else
462                 b = d_hash(dentry->d_name.hash);
463
464         hlist_bl_lock(b);
465         __hlist_bl_del(&dentry->d_hash);
466         hlist_bl_unlock(b);
467 }
468
469 void __d_drop(struct dentry *dentry)
470 {
471         if (!d_unhashed(dentry)) {
472                 ___d_drop(dentry);
473                 dentry->d_hash.pprev = NULL;
474                 write_seqcount_invalidate(&dentry->d_seq);
475         }
476 }
477 EXPORT_SYMBOL(__d_drop);
478
479 void d_drop(struct dentry *dentry)
480 {
481         spin_lock(&dentry->d_lock);
482         __d_drop(dentry);
483         spin_unlock(&dentry->d_lock);
484 }
485 EXPORT_SYMBOL(d_drop);
486
487 static inline void dentry_unlist(struct dentry *dentry, struct dentry *parent)
488 {
489         struct dentry *next;
490         /*
491          * Inform d_walk() and shrink_dentry_list() that we are no longer
492          * attached to the dentry tree
493          */
494         dentry->d_flags |= DCACHE_DENTRY_KILLED;
495         if (unlikely(list_empty(&dentry->d_child)))
496                 return;
497         __list_del_entry(&dentry->d_child);
498         /*
499          * Cursors can move around the list of children.  While we'd been
500          * a normal list member, it didn't matter - ->d_child.next would've
501          * been updated.  However, from now on it won't be and for the
502          * things like d_walk() it might end up with a nasty surprise.
503          * Normally d_walk() doesn't care about cursors moving around -
504          * ->d_lock on parent prevents that and since a cursor has no children
505          * of its own, we get through it without ever unlocking the parent.
506          * There is one exception, though - if we ascend from a child that
507          * gets killed as soon as we unlock it, the next sibling is found
508          * using the value left in its ->d_child.next.  And if _that_
509          * pointed to a cursor, and cursor got moved (e.g. by lseek())
510          * before d_walk() regains parent->d_lock, we'll end up skipping
511          * everything the cursor had been moved past.
512          *
513          * Solution: make sure that the pointer left behind in ->d_child.next
514          * points to something that won't be moving around.  I.e. skip the
515          * cursors.
516          */
517         while (dentry->d_child.next != &parent->d_subdirs) {
518                 next = list_entry(dentry->d_child.next, struct dentry, d_child);
519                 if (likely(!(next->d_flags & DCACHE_DENTRY_CURSOR)))
520                         break;
521                 dentry->d_child.next = next->d_child.next;
522         }
523 }
524
525 static void __dentry_kill(struct dentry *dentry)
526 {
527         struct dentry *parent = NULL;
528         bool can_free = true;
529         if (!IS_ROOT(dentry))
530                 parent = dentry->d_parent;
531
532         /*
533          * The dentry is now unrecoverably dead to the world.
534          */
535         lockref_mark_dead(&dentry->d_lockref);
536
537         /*
538          * inform the fs via d_prune that this dentry is about to be
539          * unhashed and destroyed.
540          */
541         if (dentry->d_flags & DCACHE_OP_PRUNE)
542                 dentry->d_op->d_prune(dentry);
543
544         if (dentry->d_flags & DCACHE_LRU_LIST) {
545                 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
546                         d_lru_del(dentry);
547         }
548         /* if it was on the hash then remove it */
549         __d_drop(dentry);
550         dentry_unlist(dentry, parent);
551         if (parent)
552                 spin_unlock(&parent->d_lock);
553         if (dentry->d_inode)
554                 dentry_unlink_inode(dentry);
555         else
556                 spin_unlock(&dentry->d_lock);
557         this_cpu_dec(nr_dentry);
558         if (dentry->d_op && dentry->d_op->d_release)
559                 dentry->d_op->d_release(dentry);
560
561         spin_lock(&dentry->d_lock);
562         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
563                 dentry->d_flags |= DCACHE_MAY_FREE;
564                 can_free = false;
565         }
566         spin_unlock(&dentry->d_lock);
567         if (likely(can_free))
568                 dentry_free(dentry);
569 }
570
571 static struct dentry *__lock_parent(struct dentry *dentry)
572 {
573         struct dentry *parent;
574         rcu_read_lock();
575         spin_unlock(&dentry->d_lock);
576 again:
577         parent = READ_ONCE(dentry->d_parent);
578         spin_lock(&parent->d_lock);
579         /*
580          * We can't blindly lock dentry until we are sure
581          * that we won't violate the locking order.
582          * Any changes of dentry->d_parent must have
583          * been done with parent->d_lock held, so
584          * spin_lock() above is enough of a barrier
585          * for checking if it's still our child.
586          */
587         if (unlikely(parent != dentry->d_parent)) {
588                 spin_unlock(&parent->d_lock);
589                 goto again;
590         }
591         rcu_read_unlock();
592         if (parent != dentry)
593                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
594         else
595                 parent = NULL;
596         return parent;
597 }
598
599 static inline struct dentry *lock_parent(struct dentry *dentry)
600 {
601         struct dentry *parent = dentry->d_parent;
602         if (IS_ROOT(dentry))
603                 return NULL;
604         if (likely(spin_trylock(&parent->d_lock)))
605                 return parent;
606         return __lock_parent(dentry);
607 }
608
609 static inline bool retain_dentry(struct dentry *dentry)
610 {
611         WARN_ON(d_in_lookup(dentry));
612
613         /* Unreachable? Get rid of it */
614         if (unlikely(d_unhashed(dentry)))
615                 return false;
616
617         if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
618                 return false;
619
620         if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
621                 if (dentry->d_op->d_delete(dentry))
622                         return false;
623         }
624         /* retain; LRU fodder */
625         dentry->d_lockref.count--;
626         if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
627                 d_lru_add(dentry);
628         else if (unlikely(!(dentry->d_flags & DCACHE_REFERENCED)))
629                 dentry->d_flags |= DCACHE_REFERENCED;
630         return true;
631 }
632
633 /*
634  * Finish off a dentry we've decided to kill.
635  * dentry->d_lock must be held, returns with it unlocked.
636  * Returns dentry requiring refcount drop, or NULL if we're done.
637  */
638 static struct dentry *dentry_kill(struct dentry *dentry)
639         __releases(dentry->d_lock)
640 {
641         struct inode *inode = dentry->d_inode;
642         struct dentry *parent = NULL;
643
644         if (inode && unlikely(!spin_trylock(&inode->i_lock)))
645                 goto slow_positive;
646
647         if (!IS_ROOT(dentry)) {
648                 parent = dentry->d_parent;
649                 if (unlikely(!spin_trylock(&parent->d_lock))) {
650                         parent = __lock_parent(dentry);
651                         if (likely(inode || !dentry->d_inode))
652                                 goto got_locks;
653                         /* negative that became positive */
654                         if (parent)
655                                 spin_unlock(&parent->d_lock);
656                         inode = dentry->d_inode;
657                         goto slow_positive;
658                 }
659         }
660         __dentry_kill(dentry);
661         return parent;
662
663 slow_positive:
664         spin_unlock(&dentry->d_lock);
665         spin_lock(&inode->i_lock);
666         spin_lock(&dentry->d_lock);
667         parent = lock_parent(dentry);
668 got_locks:
669         if (unlikely(dentry->d_lockref.count != 1)) {
670                 dentry->d_lockref.count--;
671         } else if (likely(!retain_dentry(dentry))) {
672                 __dentry_kill(dentry);
673                 return parent;
674         }
675         /* we are keeping it, after all */
676         if (inode)
677                 spin_unlock(&inode->i_lock);
678         if (parent)
679                 spin_unlock(&parent->d_lock);
680         spin_unlock(&dentry->d_lock);
681         return NULL;
682 }
683
684 /*
685  * Try to do a lockless dput(), and return whether that was successful.
686  *
687  * If unsuccessful, we return false, having already taken the dentry lock.
688  *
689  * The caller needs to hold the RCU read lock, so that the dentry is
690  * guaranteed to stay around even if the refcount goes down to zero!
691  */
692 static inline bool fast_dput(struct dentry *dentry)
693 {
694         int ret;
695         unsigned int d_flags;
696
697         /*
698          * If we have a d_op->d_delete() operation, we sould not
699          * let the dentry count go to zero, so use "put_or_lock".
700          */
701         if (unlikely(dentry->d_flags & DCACHE_OP_DELETE))
702                 return lockref_put_or_lock(&dentry->d_lockref);
703
704         /*
705          * .. otherwise, we can try to just decrement the
706          * lockref optimistically.
707          */
708         ret = lockref_put_return(&dentry->d_lockref);
709
710         /*
711          * If the lockref_put_return() failed due to the lock being held
712          * by somebody else, the fast path has failed. We will need to
713          * get the lock, and then check the count again.
714          */
715         if (unlikely(ret < 0)) {
716                 spin_lock(&dentry->d_lock);
717                 if (dentry->d_lockref.count > 1) {
718                         dentry->d_lockref.count--;
719                         spin_unlock(&dentry->d_lock);
720                         return 1;
721                 }
722                 return 0;
723         }
724
725         /*
726          * If we weren't the last ref, we're done.
727          */
728         if (ret)
729                 return 1;
730
731         /*
732          * Careful, careful. The reference count went down
733          * to zero, but we don't hold the dentry lock, so
734          * somebody else could get it again, and do another
735          * dput(), and we need to not race with that.
736          *
737          * However, there is a very special and common case
738          * where we don't care, because there is nothing to
739          * do: the dentry is still hashed, it does not have
740          * a 'delete' op, and it's referenced and already on
741          * the LRU list.
742          *
743          * NOTE! Since we aren't locked, these values are
744          * not "stable". However, it is sufficient that at
745          * some point after we dropped the reference the
746          * dentry was hashed and the flags had the proper
747          * value. Other dentry users may have re-gotten
748          * a reference to the dentry and change that, but
749          * our work is done - we can leave the dentry
750          * around with a zero refcount.
751          */
752         smp_rmb();
753         d_flags = READ_ONCE(dentry->d_flags);
754         d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST | DCACHE_DISCONNECTED;
755
756         /* Nothing to do? Dropping the reference was all we needed? */
757         if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry))
758                 return 1;
759
760         /*
761          * Not the fast normal case? Get the lock. We've already decremented
762          * the refcount, but we'll need to re-check the situation after
763          * getting the lock.
764          */
765         spin_lock(&dentry->d_lock);
766
767         /*
768          * Did somebody else grab a reference to it in the meantime, and
769          * we're no longer the last user after all? Alternatively, somebody
770          * else could have killed it and marked it dead. Either way, we
771          * don't need to do anything else.
772          */
773         if (dentry->d_lockref.count) {
774                 spin_unlock(&dentry->d_lock);
775                 return 1;
776         }
777
778         /*
779          * Re-get the reference we optimistically dropped. We hold the
780          * lock, and we just tested that it was zero, so we can just
781          * set it to 1.
782          */
783         dentry->d_lockref.count = 1;
784         return 0;
785 }
786
787
788 /* 
789  * This is dput
790  *
791  * This is complicated by the fact that we do not want to put
792  * dentries that are no longer on any hash chain on the unused
793  * list: we'd much rather just get rid of them immediately.
794  *
795  * However, that implies that we have to traverse the dentry
796  * tree upwards to the parents which might _also_ now be
797  * scheduled for deletion (it may have been only waiting for
798  * its last child to go away).
799  *
800  * This tail recursion is done by hand as we don't want to depend
801  * on the compiler to always get this right (gcc generally doesn't).
802  * Real recursion would eat up our stack space.
803  */
804
805 /*
806  * dput - release a dentry
807  * @dentry: dentry to release 
808  *
809  * Release a dentry. This will drop the usage count and if appropriate
810  * call the dentry unlink method as well as removing it from the queues and
811  * releasing its resources. If the parent dentries were scheduled for release
812  * they too may now get deleted.
813  */
814 void dput(struct dentry *dentry)
815 {
816         if (unlikely(!dentry))
817                 return;
818
819 repeat:
820         might_sleep();
821
822         rcu_read_lock();
823         if (likely(fast_dput(dentry))) {
824                 rcu_read_unlock();
825                 return;
826         }
827
828         /* Slow case: now with the dentry lock held */
829         rcu_read_unlock();
830
831         if (likely(retain_dentry(dentry))) {
832                 spin_unlock(&dentry->d_lock);
833                 return;
834         }
835
836         dentry = dentry_kill(dentry);
837         if (dentry) {
838                 cond_resched();
839                 goto repeat;
840         }
841 }
842 EXPORT_SYMBOL(dput);
843
844
845 /* This must be called with d_lock held */
846 static inline void __dget_dlock(struct dentry *dentry)
847 {
848         dentry->d_lockref.count++;
849 }
850
851 static inline void __dget(struct dentry *dentry)
852 {
853         lockref_get(&dentry->d_lockref);
854 }
855
856 struct dentry *dget_parent(struct dentry *dentry)
857 {
858         int gotref;
859         struct dentry *ret;
860
861         /*
862          * Do optimistic parent lookup without any
863          * locking.
864          */
865         rcu_read_lock();
866         ret = READ_ONCE(dentry->d_parent);
867         gotref = lockref_get_not_zero(&ret->d_lockref);
868         rcu_read_unlock();
869         if (likely(gotref)) {
870                 if (likely(ret == READ_ONCE(dentry->d_parent)))
871                         return ret;
872                 dput(ret);
873         }
874
875 repeat:
876         /*
877          * Don't need rcu_dereference because we re-check it was correct under
878          * the lock.
879          */
880         rcu_read_lock();
881         ret = dentry->d_parent;
882         spin_lock(&ret->d_lock);
883         if (unlikely(ret != dentry->d_parent)) {
884                 spin_unlock(&ret->d_lock);
885                 rcu_read_unlock();
886                 goto repeat;
887         }
888         rcu_read_unlock();
889         BUG_ON(!ret->d_lockref.count);
890         ret->d_lockref.count++;
891         spin_unlock(&ret->d_lock);
892         return ret;
893 }
894 EXPORT_SYMBOL(dget_parent);
895
896 /**
897  * d_find_alias - grab a hashed alias of inode
898  * @inode: inode in question
899  *
900  * If inode has a hashed alias, or is a directory and has any alias,
901  * acquire the reference to alias and return it. Otherwise return NULL.
902  * Notice that if inode is a directory there can be only one alias and
903  * it can be unhashed only if it has no children, or if it is the root
904  * of a filesystem, or if the directory was renamed and d_revalidate
905  * was the first vfs operation to notice.
906  *
907  * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
908  * any other hashed alias over that one.
909  */
910 static struct dentry *__d_find_alias(struct inode *inode)
911 {
912         struct dentry *alias, *discon_alias;
913
914 again:
915         discon_alias = NULL;
916         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
917                 spin_lock(&alias->d_lock);
918                 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
919                         if (IS_ROOT(alias) &&
920                             (alias->d_flags & DCACHE_DISCONNECTED)) {
921                                 discon_alias = alias;
922                         } else {
923                                 __dget_dlock(alias);
924                                 spin_unlock(&alias->d_lock);
925                                 return alias;
926                         }
927                 }
928                 spin_unlock(&alias->d_lock);
929         }
930         if (discon_alias) {
931                 alias = discon_alias;
932                 spin_lock(&alias->d_lock);
933                 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
934                         __dget_dlock(alias);
935                         spin_unlock(&alias->d_lock);
936                         return alias;
937                 }
938                 spin_unlock(&alias->d_lock);
939                 goto again;
940         }
941         return NULL;
942 }
943
944 struct dentry *d_find_alias(struct inode *inode)
945 {
946         struct dentry *de = NULL;
947
948         if (!hlist_empty(&inode->i_dentry)) {
949                 spin_lock(&inode->i_lock);
950                 de = __d_find_alias(inode);
951                 spin_unlock(&inode->i_lock);
952         }
953         return de;
954 }
955 EXPORT_SYMBOL(d_find_alias);
956
957 /*
958  *      Try to kill dentries associated with this inode.
959  * WARNING: you must own a reference to inode.
960  */
961 void d_prune_aliases(struct inode *inode)
962 {
963         struct dentry *dentry;
964 restart:
965         spin_lock(&inode->i_lock);
966         hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
967                 spin_lock(&dentry->d_lock);
968                 if (!dentry->d_lockref.count) {
969                         struct dentry *parent = lock_parent(dentry);
970                         if (likely(!dentry->d_lockref.count)) {
971                                 __dentry_kill(dentry);
972                                 dput(parent);
973                                 goto restart;
974                         }
975                         if (parent)
976                                 spin_unlock(&parent->d_lock);
977                 }
978                 spin_unlock(&dentry->d_lock);
979         }
980         spin_unlock(&inode->i_lock);
981 }
982 EXPORT_SYMBOL(d_prune_aliases);
983
984 /*
985  * Lock a dentry from shrink list.
986  * Called under rcu_read_lock() and dentry->d_lock; the former
987  * guarantees that nothing we access will be freed under us.
988  * Note that dentry is *not* protected from concurrent dentry_kill(),
989  * d_delete(), etc.
990  *
991  * Return false if dentry has been disrupted or grabbed, leaving
992  * the caller to kick it off-list.  Otherwise, return true and have
993  * that dentry's inode and parent both locked.
994  */
995 static bool shrink_lock_dentry(struct dentry *dentry)
996 {
997         struct inode *inode;
998         struct dentry *parent;
999
1000         if (dentry->d_lockref.count)
1001                 return false;
1002
1003         inode = dentry->d_inode;
1004         if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
1005                 spin_unlock(&dentry->d_lock);
1006                 spin_lock(&inode->i_lock);
1007                 spin_lock(&dentry->d_lock);
1008                 if (unlikely(dentry->d_lockref.count))
1009                         goto out;
1010                 /* changed inode means that somebody had grabbed it */
1011                 if (unlikely(inode != dentry->d_inode))
1012                         goto out;
1013         }
1014
1015         parent = dentry->d_parent;
1016         if (IS_ROOT(dentry) || likely(spin_trylock(&parent->d_lock)))
1017                 return true;
1018
1019         spin_unlock(&dentry->d_lock);
1020         spin_lock(&parent->d_lock);
1021         if (unlikely(parent != dentry->d_parent)) {
1022                 spin_unlock(&parent->d_lock);
1023                 spin_lock(&dentry->d_lock);
1024                 goto out;
1025         }
1026         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1027         if (likely(!dentry->d_lockref.count))
1028                 return true;
1029         spin_unlock(&parent->d_lock);
1030 out:
1031         if (inode)
1032                 spin_unlock(&inode->i_lock);
1033         return false;
1034 }
1035
1036 static void shrink_dentry_list(struct list_head *list)
1037 {
1038         while (!list_empty(list)) {
1039                 struct dentry *dentry, *parent;
1040
1041                 dentry = list_entry(list->prev, struct dentry, d_lru);
1042                 spin_lock(&dentry->d_lock);
1043                 rcu_read_lock();
1044                 if (!shrink_lock_dentry(dentry)) {
1045                         bool can_free = false;
1046                         rcu_read_unlock();
1047                         d_shrink_del(dentry);
1048                         if (dentry->d_lockref.count < 0)
1049                                 can_free = dentry->d_flags & DCACHE_MAY_FREE;
1050                         spin_unlock(&dentry->d_lock);
1051                         if (can_free)
1052                                 dentry_free(dentry);
1053                         continue;
1054                 }
1055                 rcu_read_unlock();
1056                 d_shrink_del(dentry);
1057                 parent = dentry->d_parent;
1058                 __dentry_kill(dentry);
1059                 if (parent == dentry)
1060                         continue;
1061                 /*
1062                  * We need to prune ancestors too. This is necessary to prevent
1063                  * quadratic behavior of shrink_dcache_parent(), but is also
1064                  * expected to be beneficial in reducing dentry cache
1065                  * fragmentation.
1066                  */
1067                 dentry = parent;
1068                 while (dentry && !lockref_put_or_lock(&dentry->d_lockref))
1069                         dentry = dentry_kill(dentry);
1070         }
1071 }
1072
1073 static enum lru_status dentry_lru_isolate(struct list_head *item,
1074                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1075 {
1076         struct list_head *freeable = arg;
1077         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1078
1079
1080         /*
1081          * we are inverting the lru lock/dentry->d_lock here,
1082          * so use a trylock. If we fail to get the lock, just skip
1083          * it
1084          */
1085         if (!spin_trylock(&dentry->d_lock))
1086                 return LRU_SKIP;
1087
1088         /*
1089          * Referenced dentries are still in use. If they have active
1090          * counts, just remove them from the LRU. Otherwise give them
1091          * another pass through the LRU.
1092          */
1093         if (dentry->d_lockref.count) {
1094                 d_lru_isolate(lru, dentry);
1095                 spin_unlock(&dentry->d_lock);
1096                 return LRU_REMOVED;
1097         }
1098
1099         if (dentry->d_flags & DCACHE_REFERENCED) {
1100                 dentry->d_flags &= ~DCACHE_REFERENCED;
1101                 spin_unlock(&dentry->d_lock);
1102
1103                 /*
1104                  * The list move itself will be made by the common LRU code. At
1105                  * this point, we've dropped the dentry->d_lock but keep the
1106                  * lru lock. This is safe to do, since every list movement is
1107                  * protected by the lru lock even if both locks are held.
1108                  *
1109                  * This is guaranteed by the fact that all LRU management
1110                  * functions are intermediated by the LRU API calls like
1111                  * list_lru_add and list_lru_del. List movement in this file
1112                  * only ever occur through this functions or through callbacks
1113                  * like this one, that are called from the LRU API.
1114                  *
1115                  * The only exceptions to this are functions like
1116                  * shrink_dentry_list, and code that first checks for the
1117                  * DCACHE_SHRINK_LIST flag.  Those are guaranteed to be
1118                  * operating only with stack provided lists after they are
1119                  * properly isolated from the main list.  It is thus, always a
1120                  * local access.
1121                  */
1122                 return LRU_ROTATE;
1123         }
1124
1125         d_lru_shrink_move(lru, dentry, freeable);
1126         spin_unlock(&dentry->d_lock);
1127
1128         return LRU_REMOVED;
1129 }
1130
1131 /**
1132  * prune_dcache_sb - shrink the dcache
1133  * @sb: superblock
1134  * @sc: shrink control, passed to list_lru_shrink_walk()
1135  *
1136  * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1137  * is done when we need more memory and called from the superblock shrinker
1138  * function.
1139  *
1140  * This function may fail to free any resources if all the dentries are in
1141  * use.
1142  */
1143 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1144 {
1145         LIST_HEAD(dispose);
1146         long freed;
1147
1148         freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1149                                      dentry_lru_isolate, &dispose);
1150         shrink_dentry_list(&dispose);
1151         return freed;
1152 }
1153
1154 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1155                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1156 {
1157         struct list_head *freeable = arg;
1158         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1159
1160         /*
1161          * we are inverting the lru lock/dentry->d_lock here,
1162          * so use a trylock. If we fail to get the lock, just skip
1163          * it
1164          */
1165         if (!spin_trylock(&dentry->d_lock))
1166                 return LRU_SKIP;
1167
1168         d_lru_shrink_move(lru, dentry, freeable);
1169         spin_unlock(&dentry->d_lock);
1170
1171         return LRU_REMOVED;
1172 }
1173
1174
1175 /**
1176  * shrink_dcache_sb - shrink dcache for a superblock
1177  * @sb: superblock
1178  *
1179  * Shrink the dcache for the specified super block. This is used to free
1180  * the dcache before unmounting a file system.
1181  */
1182 void shrink_dcache_sb(struct super_block *sb)
1183 {
1184         long freed;
1185
1186         do {
1187                 LIST_HEAD(dispose);
1188
1189                 freed = list_lru_walk(&sb->s_dentry_lru,
1190                         dentry_lru_isolate_shrink, &dispose, 1024);
1191
1192                 this_cpu_sub(nr_dentry_unused, freed);
1193                 shrink_dentry_list(&dispose);
1194                 cond_resched();
1195         } while (list_lru_count(&sb->s_dentry_lru) > 0);
1196 }
1197 EXPORT_SYMBOL(shrink_dcache_sb);
1198
1199 /**
1200  * enum d_walk_ret - action to talke during tree walk
1201  * @D_WALK_CONTINUE:    contrinue walk
1202  * @D_WALK_QUIT:        quit walk
1203  * @D_WALK_NORETRY:     quit when retry is needed
1204  * @D_WALK_SKIP:        skip this dentry and its children
1205  */
1206 enum d_walk_ret {
1207         D_WALK_CONTINUE,
1208         D_WALK_QUIT,
1209         D_WALK_NORETRY,
1210         D_WALK_SKIP,
1211 };
1212
1213 /**
1214  * d_walk - walk the dentry tree
1215  * @parent:     start of walk
1216  * @data:       data passed to @enter() and @finish()
1217  * @enter:      callback when first entering the dentry
1218  * @finish:     callback when successfully finished the walk
1219  *
1220  * The @enter() and @finish() callbacks are called with d_lock held.
1221  */
1222 static void d_walk(struct dentry *parent, void *data,
1223                    enum d_walk_ret (*enter)(void *, struct dentry *),
1224                    void (*finish)(void *))
1225 {
1226         struct dentry *this_parent;
1227         struct list_head *next;
1228         unsigned seq = 0;
1229         enum d_walk_ret ret;
1230         bool retry = true;
1231
1232 again:
1233         read_seqbegin_or_lock(&rename_lock, &seq);
1234         this_parent = parent;
1235         spin_lock(&this_parent->d_lock);
1236
1237         ret = enter(data, this_parent);
1238         switch (ret) {
1239         case D_WALK_CONTINUE:
1240                 break;
1241         case D_WALK_QUIT:
1242         case D_WALK_SKIP:
1243                 goto out_unlock;
1244         case D_WALK_NORETRY:
1245                 retry = false;
1246                 break;
1247         }
1248 repeat:
1249         next = this_parent->d_subdirs.next;
1250 resume:
1251         while (next != &this_parent->d_subdirs) {
1252                 struct list_head *tmp = next;
1253                 struct dentry *dentry = list_entry(tmp, struct dentry, d_child);
1254                 next = tmp->next;
1255
1256                 if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
1257                         continue;
1258
1259                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1260
1261                 ret = enter(data, dentry);
1262                 switch (ret) {
1263                 case D_WALK_CONTINUE:
1264                         break;
1265                 case D_WALK_QUIT:
1266                         spin_unlock(&dentry->d_lock);
1267                         goto out_unlock;
1268                 case D_WALK_NORETRY:
1269                         retry = false;
1270                         break;
1271                 case D_WALK_SKIP:
1272                         spin_unlock(&dentry->d_lock);
1273                         continue;
1274                 }
1275
1276                 if (!list_empty(&dentry->d_subdirs)) {
1277                         spin_unlock(&this_parent->d_lock);
1278                         spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1279                         this_parent = dentry;
1280                         spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1281                         goto repeat;
1282                 }
1283                 spin_unlock(&dentry->d_lock);
1284         }
1285         /*
1286          * All done at this level ... ascend and resume the search.
1287          */
1288         rcu_read_lock();
1289 ascend:
1290         if (this_parent != parent) {
1291                 struct dentry *child = this_parent;
1292                 this_parent = child->d_parent;
1293
1294                 spin_unlock(&child->d_lock);
1295                 spin_lock(&this_parent->d_lock);
1296
1297                 /* might go back up the wrong parent if we have had a rename. */
1298                 if (need_seqretry(&rename_lock, seq))
1299                         goto rename_retry;
1300                 /* go into the first sibling still alive */
1301                 do {
1302                         next = child->d_child.next;
1303                         if (next == &this_parent->d_subdirs)
1304                                 goto ascend;
1305                         child = list_entry(next, struct dentry, d_child);
1306                 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED));
1307                 rcu_read_unlock();
1308                 goto resume;
1309         }
1310         if (need_seqretry(&rename_lock, seq))
1311                 goto rename_retry;
1312         rcu_read_unlock();
1313         if (finish)
1314                 finish(data);
1315
1316 out_unlock:
1317         spin_unlock(&this_parent->d_lock);
1318         done_seqretry(&rename_lock, seq);
1319         return;
1320
1321 rename_retry:
1322         spin_unlock(&this_parent->d_lock);
1323         rcu_read_unlock();
1324         BUG_ON(seq & 1);
1325         if (!retry)
1326                 return;
1327         seq = 1;
1328         goto again;
1329 }
1330
1331 struct check_mount {
1332         struct vfsmount *mnt;
1333         unsigned int mounted;
1334 };
1335
1336 static enum d_walk_ret path_check_mount(void *data, struct dentry *dentry)
1337 {
1338         struct check_mount *info = data;
1339         struct path path = { .mnt = info->mnt, .dentry = dentry };
1340
1341         if (likely(!d_mountpoint(dentry)))
1342                 return D_WALK_CONTINUE;
1343         if (__path_is_mountpoint(&path)) {
1344                 info->mounted = 1;
1345                 return D_WALK_QUIT;
1346         }
1347         return D_WALK_CONTINUE;
1348 }
1349
1350 /**
1351  * path_has_submounts - check for mounts over a dentry in the
1352  *                      current namespace.
1353  * @parent: path to check.
1354  *
1355  * Return true if the parent or its subdirectories contain
1356  * a mount point in the current namespace.
1357  */
1358 int path_has_submounts(const struct path *parent)
1359 {
1360         struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };
1361
1362         read_seqlock_excl(&mount_lock);
1363         d_walk(parent->dentry, &data, path_check_mount, NULL);
1364         read_sequnlock_excl(&mount_lock);
1365
1366         return data.mounted;
1367 }
1368 EXPORT_SYMBOL(path_has_submounts);
1369
1370 /*
1371  * Called by mount code to set a mountpoint and check if the mountpoint is
1372  * reachable (e.g. NFS can unhash a directory dentry and then the complete
1373  * subtree can become unreachable).
1374  *
1375  * Only one of d_invalidate() and d_set_mounted() must succeed.  For
1376  * this reason take rename_lock and d_lock on dentry and ancestors.
1377  */
1378 int d_set_mounted(struct dentry *dentry)
1379 {
1380         struct dentry *p;
1381         int ret = -ENOENT;
1382         write_seqlock(&rename_lock);
1383         for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1384                 /* Need exclusion wrt. d_invalidate() */
1385                 spin_lock(&p->d_lock);
1386                 if (unlikely(d_unhashed(p))) {
1387                         spin_unlock(&p->d_lock);
1388                         goto out;
1389                 }
1390                 spin_unlock(&p->d_lock);
1391         }
1392         spin_lock(&dentry->d_lock);
1393         if (!d_unlinked(dentry)) {
1394                 ret = -EBUSY;
1395                 if (!d_mountpoint(dentry)) {
1396                         dentry->d_flags |= DCACHE_MOUNTED;
1397                         ret = 0;
1398                 }
1399         }
1400         spin_unlock(&dentry->d_lock);
1401 out:
1402         write_sequnlock(&rename_lock);
1403         return ret;
1404 }
1405
1406 /*
1407  * Search the dentry child list of the specified parent,
1408  * and move any unused dentries to the end of the unused
1409  * list for prune_dcache(). We descend to the next level
1410  * whenever the d_subdirs list is non-empty and continue
1411  * searching.
1412  *
1413  * It returns zero iff there are no unused children,
1414  * otherwise  it returns the number of children moved to
1415  * the end of the unused list. This may not be the total
1416  * number of unused children, because select_parent can
1417  * drop the lock and return early due to latency
1418  * constraints.
1419  */
1420
1421 struct select_data {
1422         struct dentry *start;
1423         struct list_head dispose;
1424         int found;
1425 };
1426
1427 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1428 {
1429         struct select_data *data = _data;
1430         enum d_walk_ret ret = D_WALK_CONTINUE;
1431
1432         if (data->start == dentry)
1433                 goto out;
1434
1435         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1436                 data->found++;
1437         } else {
1438                 if (dentry->d_flags & DCACHE_LRU_LIST)
1439                         d_lru_del(dentry);
1440                 if (!dentry->d_lockref.count) {
1441                         d_shrink_add(dentry, &data->dispose);
1442                         data->found++;
1443                 }
1444         }
1445         /*
1446          * We can return to the caller if we have found some (this
1447          * ensures forward progress). We'll be coming back to find
1448          * the rest.
1449          */
1450         if (!list_empty(&data->dispose))
1451                 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1452 out:
1453         return ret;
1454 }
1455
1456 /**
1457  * shrink_dcache_parent - prune dcache
1458  * @parent: parent of entries to prune
1459  *
1460  * Prune the dcache to remove unused children of the parent dentry.
1461  */
1462 void shrink_dcache_parent(struct dentry *parent)
1463 {
1464         for (;;) {
1465                 struct select_data data;
1466
1467                 INIT_LIST_HEAD(&data.dispose);
1468                 data.start = parent;
1469                 data.found = 0;
1470
1471                 d_walk(parent, &data, select_collect, NULL);
1472                 if (!data.found)
1473                         break;
1474
1475                 shrink_dentry_list(&data.dispose);
1476                 cond_resched();
1477         }
1478 }
1479 EXPORT_SYMBOL(shrink_dcache_parent);
1480
1481 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1482 {
1483         /* it has busy descendents; complain about those instead */
1484         if (!list_empty(&dentry->d_subdirs))
1485                 return D_WALK_CONTINUE;
1486
1487         /* root with refcount 1 is fine */
1488         if (dentry == _data && dentry->d_lockref.count == 1)
1489                 return D_WALK_CONTINUE;
1490
1491         printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} "
1492                         " still in use (%d) [unmount of %s %s]\n",
1493                        dentry,
1494                        dentry->d_inode ?
1495                        dentry->d_inode->i_ino : 0UL,
1496                        dentry,
1497                        dentry->d_lockref.count,
1498                        dentry->d_sb->s_type->name,
1499                        dentry->d_sb->s_id);
1500         WARN_ON(1);
1501         return D_WALK_CONTINUE;
1502 }
1503
1504 static void do_one_tree(struct dentry *dentry)
1505 {
1506         shrink_dcache_parent(dentry);
1507         d_walk(dentry, dentry, umount_check, NULL);
1508         d_drop(dentry);
1509         dput(dentry);
1510 }
1511
1512 /*
1513  * destroy the dentries attached to a superblock on unmounting
1514  */
1515 void shrink_dcache_for_umount(struct super_block *sb)
1516 {
1517         struct dentry *dentry;
1518
1519         WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
1520
1521         dentry = sb->s_root;
1522         sb->s_root = NULL;
1523         do_one_tree(dentry);
1524
1525         while (!hlist_bl_empty(&sb->s_roots)) {
1526                 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_roots), struct dentry, d_hash));
1527                 do_one_tree(dentry);
1528         }
1529 }
1530
1531 struct detach_data {
1532         struct select_data select;
1533         struct dentry *mountpoint;
1534 };
1535 static enum d_walk_ret detach_and_collect(void *_data, struct dentry *dentry)
1536 {
1537         struct detach_data *data = _data;
1538
1539         if (d_mountpoint(dentry)) {
1540                 __dget_dlock(dentry);
1541                 data->mountpoint = dentry;
1542                 return D_WALK_QUIT;
1543         }
1544
1545         return select_collect(&data->select, dentry);
1546 }
1547
1548 static void check_and_drop(void *_data)
1549 {
1550         struct detach_data *data = _data;
1551
1552         if (!data->mountpoint && list_empty(&data->select.dispose))
1553                 __d_drop(data->select.start);
1554 }
1555
1556 /**
1557  * d_invalidate - detach submounts, prune dcache, and drop
1558  * @dentry: dentry to invalidate (aka detach, prune and drop)
1559  *
1560  * no dcache lock.
1561  *
1562  * The final d_drop is done as an atomic operation relative to
1563  * rename_lock ensuring there are no races with d_set_mounted.  This
1564  * ensures there are no unhashed dentries on the path to a mountpoint.
1565  */
1566 void d_invalidate(struct dentry *dentry)
1567 {
1568         /*
1569          * If it's already been dropped, return OK.
1570          */
1571         spin_lock(&dentry->d_lock);
1572         if (d_unhashed(dentry)) {
1573                 spin_unlock(&dentry->d_lock);
1574                 return;
1575         }
1576         spin_unlock(&dentry->d_lock);
1577
1578         /* Negative dentries can be dropped without further checks */
1579         if (!dentry->d_inode) {
1580                 d_drop(dentry);
1581                 return;
1582         }
1583
1584         for (;;) {
1585                 struct detach_data data;
1586
1587                 data.mountpoint = NULL;
1588                 INIT_LIST_HEAD(&data.select.dispose);
1589                 data.select.start = dentry;
1590                 data.select.found = 0;
1591
1592                 d_walk(dentry, &data, detach_and_collect, check_and_drop);
1593
1594                 if (!list_empty(&data.select.dispose))
1595                         shrink_dentry_list(&data.select.dispose);
1596                 else if (!data.mountpoint)
1597                         return;
1598
1599                 if (data.mountpoint) {
1600                         detach_mounts(data.mountpoint);
1601                         dput(data.mountpoint);
1602                 }
1603                 cond_resched();
1604         }
1605 }
1606 EXPORT_SYMBOL(d_invalidate);
1607
1608 /**
1609  * __d_alloc    -       allocate a dcache entry
1610  * @sb: filesystem it will belong to
1611  * @name: qstr of the name
1612  *
1613  * Allocates a dentry. It returns %NULL if there is insufficient memory
1614  * available. On a success the dentry is returned. The name passed in is
1615  * copied and the copy passed in may be reused after this call.
1616  */
1617  
1618 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1619 {
1620         struct dentry *dentry;
1621         char *dname;
1622         int err;
1623
1624         dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1625         if (!dentry)
1626                 return NULL;
1627
1628         /*
1629          * We guarantee that the inline name is always NUL-terminated.
1630          * This way the memcpy() done by the name switching in rename
1631          * will still always have a NUL at the end, even if we might
1632          * be overwriting an internal NUL character
1633          */
1634         dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1635         if (unlikely(!name)) {
1636                 name = &slash_name;
1637                 dname = dentry->d_iname;
1638         } else if (name->len > DNAME_INLINE_LEN-1) {
1639                 size_t size = offsetof(struct external_name, name[1]);
1640                 struct external_name *p = kmalloc(size + name->len,
1641                                                   GFP_KERNEL_ACCOUNT);
1642                 if (!p) {
1643                         kmem_cache_free(dentry_cache, dentry); 
1644                         return NULL;
1645                 }
1646                 atomic_set(&p->u.count, 1);
1647                 dname = p->name;
1648         } else  {
1649                 dname = dentry->d_iname;
1650         }       
1651
1652         dentry->d_name.len = name->len;
1653         dentry->d_name.hash = name->hash;
1654         memcpy(dname, name->name, name->len);
1655         dname[name->len] = 0;
1656
1657         /* Make sure we always see the terminating NUL character */
1658         smp_store_release(&dentry->d_name.name, dname); /* ^^^ */
1659
1660         dentry->d_lockref.count = 1;
1661         dentry->d_flags = 0;
1662         spin_lock_init(&dentry->d_lock);
1663         seqcount_init(&dentry->d_seq);
1664         dentry->d_inode = NULL;
1665         dentry->d_parent = dentry;
1666         dentry->d_sb = sb;
1667         dentry->d_op = NULL;
1668         dentry->d_fsdata = NULL;
1669         INIT_HLIST_BL_NODE(&dentry->d_hash);
1670         INIT_LIST_HEAD(&dentry->d_lru);
1671         INIT_LIST_HEAD(&dentry->d_subdirs);
1672         INIT_HLIST_NODE(&dentry->d_u.d_alias);
1673         INIT_LIST_HEAD(&dentry->d_child);
1674         d_set_d_op(dentry, dentry->d_sb->s_d_op);
1675
1676         if (dentry->d_op && dentry->d_op->d_init) {
1677                 err = dentry->d_op->d_init(dentry);
1678                 if (err) {
1679                         if (dname_external(dentry))
1680                                 kfree(external_name(dentry));
1681                         kmem_cache_free(dentry_cache, dentry);
1682                         return NULL;
1683                 }
1684         }
1685
1686         this_cpu_inc(nr_dentry);
1687
1688         return dentry;
1689 }
1690
1691 /**
1692  * d_alloc      -       allocate a dcache entry
1693  * @parent: parent of entry to allocate
1694  * @name: qstr of the name
1695  *
1696  * Allocates a dentry. It returns %NULL if there is insufficient memory
1697  * available. On a success the dentry is returned. The name passed in is
1698  * copied and the copy passed in may be reused after this call.
1699  */
1700 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1701 {
1702         struct dentry *dentry = __d_alloc(parent->d_sb, name);
1703         if (!dentry)
1704                 return NULL;
1705         dentry->d_flags |= DCACHE_RCUACCESS;
1706         spin_lock(&parent->d_lock);
1707         /*
1708          * don't need child lock because it is not subject
1709          * to concurrency here
1710          */
1711         __dget_dlock(parent);
1712         dentry->d_parent = parent;
1713         list_add(&dentry->d_child, &parent->d_subdirs);
1714         spin_unlock(&parent->d_lock);
1715
1716         return dentry;
1717 }
1718 EXPORT_SYMBOL(d_alloc);
1719
1720 struct dentry *d_alloc_anon(struct super_block *sb)
1721 {
1722         return __d_alloc(sb, NULL);
1723 }
1724 EXPORT_SYMBOL(d_alloc_anon);
1725
1726 struct dentry *d_alloc_cursor(struct dentry * parent)
1727 {
1728         struct dentry *dentry = d_alloc_anon(parent->d_sb);
1729         if (dentry) {
1730                 dentry->d_flags |= DCACHE_RCUACCESS | DCACHE_DENTRY_CURSOR;
1731                 dentry->d_parent = dget(parent);
1732         }
1733         return dentry;
1734 }
1735
1736 /**
1737  * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1738  * @sb: the superblock
1739  * @name: qstr of the name
1740  *
1741  * For a filesystem that just pins its dentries in memory and never
1742  * performs lookups at all, return an unhashed IS_ROOT dentry.
1743  */
1744 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1745 {
1746         return __d_alloc(sb, name);
1747 }
1748 EXPORT_SYMBOL(d_alloc_pseudo);
1749
1750 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1751 {
1752         struct qstr q;
1753
1754         q.name = name;
1755         q.hash_len = hashlen_string(parent, name);
1756         return d_alloc(parent, &q);
1757 }
1758 EXPORT_SYMBOL(d_alloc_name);
1759
1760 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1761 {
1762         WARN_ON_ONCE(dentry->d_op);
1763         WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH  |
1764                                 DCACHE_OP_COMPARE       |
1765                                 DCACHE_OP_REVALIDATE    |
1766                                 DCACHE_OP_WEAK_REVALIDATE       |
1767                                 DCACHE_OP_DELETE        |
1768                                 DCACHE_OP_REAL));
1769         dentry->d_op = op;
1770         if (!op)
1771                 return;
1772         if (op->d_hash)
1773                 dentry->d_flags |= DCACHE_OP_HASH;
1774         if (op->d_compare)
1775                 dentry->d_flags |= DCACHE_OP_COMPARE;
1776         if (op->d_revalidate)
1777                 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1778         if (op->d_weak_revalidate)
1779                 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1780         if (op->d_delete)
1781                 dentry->d_flags |= DCACHE_OP_DELETE;
1782         if (op->d_prune)
1783                 dentry->d_flags |= DCACHE_OP_PRUNE;
1784         if (op->d_real)
1785                 dentry->d_flags |= DCACHE_OP_REAL;
1786
1787 }
1788 EXPORT_SYMBOL(d_set_d_op);
1789
1790
1791 /*
1792  * d_set_fallthru - Mark a dentry as falling through to a lower layer
1793  * @dentry - The dentry to mark
1794  *
1795  * Mark a dentry as falling through to the lower layer (as set with
1796  * d_pin_lower()).  This flag may be recorded on the medium.
1797  */
1798 void d_set_fallthru(struct dentry *dentry)
1799 {
1800         spin_lock(&dentry->d_lock);
1801         dentry->d_flags |= DCACHE_FALLTHRU;
1802         spin_unlock(&dentry->d_lock);
1803 }
1804 EXPORT_SYMBOL(d_set_fallthru);
1805
1806 static unsigned d_flags_for_inode(struct inode *inode)
1807 {
1808         unsigned add_flags = DCACHE_REGULAR_TYPE;
1809
1810         if (!inode)
1811                 return DCACHE_MISS_TYPE;
1812
1813         if (S_ISDIR(inode->i_mode)) {
1814                 add_flags = DCACHE_DIRECTORY_TYPE;
1815                 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1816                         if (unlikely(!inode->i_op->lookup))
1817                                 add_flags = DCACHE_AUTODIR_TYPE;
1818                         else
1819                                 inode->i_opflags |= IOP_LOOKUP;
1820                 }
1821                 goto type_determined;
1822         }
1823
1824         if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1825                 if (unlikely(inode->i_op->get_link)) {
1826                         add_flags = DCACHE_SYMLINK_TYPE;
1827                         goto type_determined;
1828                 }
1829                 inode->i_opflags |= IOP_NOFOLLOW;
1830         }
1831
1832         if (unlikely(!S_ISREG(inode->i_mode)))
1833                 add_flags = DCACHE_SPECIAL_TYPE;
1834
1835 type_determined:
1836         if (unlikely(IS_AUTOMOUNT(inode)))
1837                 add_flags |= DCACHE_NEED_AUTOMOUNT;
1838         return add_flags;
1839 }
1840
1841 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1842 {
1843         unsigned add_flags = d_flags_for_inode(inode);
1844         WARN_ON(d_in_lookup(dentry));
1845
1846         spin_lock(&dentry->d_lock);
1847         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1848         raw_write_seqcount_begin(&dentry->d_seq);
1849         __d_set_inode_and_type(dentry, inode, add_flags);
1850         raw_write_seqcount_end(&dentry->d_seq);
1851         fsnotify_update_flags(dentry);
1852         spin_unlock(&dentry->d_lock);
1853 }
1854
1855 /**
1856  * d_instantiate - fill in inode information for a dentry
1857  * @entry: dentry to complete
1858  * @inode: inode to attach to this dentry
1859  *
1860  * Fill in inode information in the entry.
1861  *
1862  * This turns negative dentries into productive full members
1863  * of society.
1864  *
1865  * NOTE! This assumes that the inode count has been incremented
1866  * (or otherwise set) by the caller to indicate that it is now
1867  * in use by the dcache.
1868  */
1869  
1870 void d_instantiate(struct dentry *entry, struct inode * inode)
1871 {
1872         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1873         if (inode) {
1874                 security_d_instantiate(entry, inode);
1875                 spin_lock(&inode->i_lock);
1876                 __d_instantiate(entry, inode);
1877                 spin_unlock(&inode->i_lock);
1878         }
1879 }
1880 EXPORT_SYMBOL(d_instantiate);
1881
1882 /**
1883  * d_instantiate_no_diralias - instantiate a non-aliased dentry
1884  * @entry: dentry to complete
1885  * @inode: inode to attach to this dentry
1886  *
1887  * Fill in inode information in the entry.  If a directory alias is found, then
1888  * return an error (and drop inode).  Together with d_materialise_unique() this
1889  * guarantees that a directory inode may never have more than one alias.
1890  */
1891 int d_instantiate_no_diralias(struct dentry *entry, struct inode *inode)
1892 {
1893         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1894
1895         security_d_instantiate(entry, inode);
1896         spin_lock(&inode->i_lock);
1897         if (S_ISDIR(inode->i_mode) && !hlist_empty(&inode->i_dentry)) {
1898                 spin_unlock(&inode->i_lock);
1899                 iput(inode);
1900                 return -EBUSY;
1901         }
1902         __d_instantiate(entry, inode);
1903         spin_unlock(&inode->i_lock);
1904
1905         return 0;
1906 }
1907 EXPORT_SYMBOL(d_instantiate_no_diralias);
1908
1909 struct dentry *d_make_root(struct inode *root_inode)
1910 {
1911         struct dentry *res = NULL;
1912
1913         if (root_inode) {
1914                 res = d_alloc_anon(root_inode->i_sb);
1915                 if (res)
1916                         d_instantiate(res, root_inode);
1917                 else
1918                         iput(root_inode);
1919         }
1920         return res;
1921 }
1922 EXPORT_SYMBOL(d_make_root);
1923
1924 static struct dentry * __d_find_any_alias(struct inode *inode)
1925 {
1926         struct dentry *alias;
1927
1928         if (hlist_empty(&inode->i_dentry))
1929                 return NULL;
1930         alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
1931         __dget(alias);
1932         return alias;
1933 }
1934
1935 /**
1936  * d_find_any_alias - find any alias for a given inode
1937  * @inode: inode to find an alias for
1938  *
1939  * If any aliases exist for the given inode, take and return a
1940  * reference for one of them.  If no aliases exist, return %NULL.
1941  */
1942 struct dentry *d_find_any_alias(struct inode *inode)
1943 {
1944         struct dentry *de;
1945
1946         spin_lock(&inode->i_lock);
1947         de = __d_find_any_alias(inode);
1948         spin_unlock(&inode->i_lock);
1949         return de;
1950 }
1951 EXPORT_SYMBOL(d_find_any_alias);
1952
1953 static struct dentry *__d_instantiate_anon(struct dentry *dentry,
1954                                            struct inode *inode,
1955                                            bool disconnected)
1956 {
1957         struct dentry *res;
1958         unsigned add_flags;
1959
1960         security_d_instantiate(dentry, inode);
1961         spin_lock(&inode->i_lock);
1962         res = __d_find_any_alias(inode);
1963         if (res) {
1964                 spin_unlock(&inode->i_lock);
1965                 dput(dentry);
1966                 goto out_iput;
1967         }
1968
1969         /* attach a disconnected dentry */
1970         add_flags = d_flags_for_inode(inode);
1971
1972         if (disconnected)
1973                 add_flags |= DCACHE_DISCONNECTED;
1974
1975         spin_lock(&dentry->d_lock);
1976         __d_set_inode_and_type(dentry, inode, add_flags);
1977         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1978         if (!disconnected) {
1979                 hlist_bl_lock(&dentry->d_sb->s_roots);
1980                 hlist_bl_add_head(&dentry->d_hash, &dentry->d_sb->s_roots);
1981                 hlist_bl_unlock(&dentry->d_sb->s_roots);
1982         }
1983         spin_unlock(&dentry->d_lock);
1984         spin_unlock(&inode->i_lock);
1985
1986         return dentry;
1987
1988  out_iput:
1989         iput(inode);
1990         return res;
1991 }
1992
1993 struct dentry *d_instantiate_anon(struct dentry *dentry, struct inode *inode)
1994 {
1995         return __d_instantiate_anon(dentry, inode, true);
1996 }
1997 EXPORT_SYMBOL(d_instantiate_anon);
1998
1999 static struct dentry *__d_obtain_alias(struct inode *inode, bool disconnected)
2000 {
2001         struct dentry *tmp;
2002         struct dentry *res;
2003
2004         if (!inode)
2005                 return ERR_PTR(-ESTALE);
2006         if (IS_ERR(inode))
2007                 return ERR_CAST(inode);
2008
2009         res = d_find_any_alias(inode);
2010         if (res)
2011                 goto out_iput;
2012
2013         tmp = d_alloc_anon(inode->i_sb);
2014         if (!tmp) {
2015                 res = ERR_PTR(-ENOMEM);
2016                 goto out_iput;
2017         }
2018
2019         return __d_instantiate_anon(tmp, inode, disconnected);
2020
2021 out_iput:
2022         iput(inode);
2023         return res;
2024 }
2025
2026 /**
2027  * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
2028  * @inode: inode to allocate the dentry for
2029  *
2030  * Obtain a dentry for an inode resulting from NFS filehandle conversion or
2031  * similar open by handle operations.  The returned dentry may be anonymous,
2032  * or may have a full name (if the inode was already in the cache).
2033  *
2034  * When called on a directory inode, we must ensure that the inode only ever
2035  * has one dentry.  If a dentry is found, that is returned instead of
2036  * allocating a new one.
2037  *
2038  * On successful return, the reference to the inode has been transferred
2039  * to the dentry.  In case of an error the reference on the inode is released.
2040  * To make it easier to use in export operations a %NULL or IS_ERR inode may
2041  * be passed in and the error will be propagated to the return value,
2042  * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
2043  */
2044 struct dentry *d_obtain_alias(struct inode *inode)
2045 {
2046         return __d_obtain_alias(inode, true);
2047 }
2048 EXPORT_SYMBOL(d_obtain_alias);
2049
2050 /**
2051  * d_obtain_root - find or allocate a dentry for a given inode
2052  * @inode: inode to allocate the dentry for
2053  *
2054  * Obtain an IS_ROOT dentry for the root of a filesystem.
2055  *
2056  * We must ensure that directory inodes only ever have one dentry.  If a
2057  * dentry is found, that is returned instead of allocating a new one.
2058  *
2059  * On successful return, the reference to the inode has been transferred
2060  * to the dentry.  In case of an error the reference on the inode is
2061  * released.  A %NULL or IS_ERR inode may be passed in and will be the
2062  * error will be propagate to the return value, with a %NULL @inode
2063  * replaced by ERR_PTR(-ESTALE).
2064  */
2065 struct dentry *d_obtain_root(struct inode *inode)
2066 {
2067         return __d_obtain_alias(inode, false);
2068 }
2069 EXPORT_SYMBOL(d_obtain_root);
2070
2071 /**
2072  * d_add_ci - lookup or allocate new dentry with case-exact name
2073  * @inode:  the inode case-insensitive lookup has found
2074  * @dentry: the negative dentry that was passed to the parent's lookup func
2075  * @name:   the case-exact name to be associated with the returned dentry
2076  *
2077  * This is to avoid filling the dcache with case-insensitive names to the
2078  * same inode, only the actual correct case is stored in the dcache for
2079  * case-insensitive filesystems.
2080  *
2081  * For a case-insensitive lookup match and if the the case-exact dentry
2082  * already exists in in the dcache, use it and return it.
2083  *
2084  * If no entry exists with the exact case name, allocate new dentry with
2085  * the exact case, and return the spliced entry.
2086  */
2087 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2088                         struct qstr *name)
2089 {
2090         struct dentry *found, *res;
2091
2092         /*
2093          * First check if a dentry matching the name already exists,
2094          * if not go ahead and create it now.
2095          */
2096         found = d_hash_and_lookup(dentry->d_parent, name);
2097         if (found) {
2098                 iput(inode);
2099                 return found;
2100         }
2101         if (d_in_lookup(dentry)) {
2102                 found = d_alloc_parallel(dentry->d_parent, name,
2103                                         dentry->d_wait);
2104                 if (IS_ERR(found) || !d_in_lookup(found)) {
2105                         iput(inode);
2106                         return found;
2107                 }
2108         } else {
2109                 found = d_alloc(dentry->d_parent, name);
2110                 if (!found) {
2111                         iput(inode);
2112                         return ERR_PTR(-ENOMEM);
2113                 } 
2114         }
2115         res = d_splice_alias(inode, found);
2116         if (res) {
2117                 dput(found);
2118                 return res;
2119         }
2120         return found;
2121 }
2122 EXPORT_SYMBOL(d_add_ci);
2123
2124
2125 static inline bool d_same_name(const struct dentry *dentry,
2126                                 const struct dentry *parent,
2127                                 const struct qstr *name)
2128 {
2129         if (likely(!(parent->d_flags & DCACHE_OP_COMPARE))) {
2130                 if (dentry->d_name.len != name->len)
2131                         return false;
2132                 return dentry_cmp(dentry, name->name, name->len) == 0;
2133         }
2134         return parent->d_op->d_compare(dentry,
2135                                        dentry->d_name.len, dentry->d_name.name,
2136                                        name) == 0;
2137 }
2138
2139 /**
2140  * __d_lookup_rcu - search for a dentry (racy, store-free)
2141  * @parent: parent dentry
2142  * @name: qstr of name we wish to find
2143  * @seqp: returns d_seq value at the point where the dentry was found
2144  * Returns: dentry, or NULL
2145  *
2146  * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2147  * resolution (store-free path walking) design described in
2148  * Documentation/filesystems/path-lookup.txt.
2149  *
2150  * This is not to be used outside core vfs.
2151  *
2152  * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2153  * held, and rcu_read_lock held. The returned dentry must not be stored into
2154  * without taking d_lock and checking d_seq sequence count against @seq
2155  * returned here.
2156  *
2157  * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2158  * function.
2159  *
2160  * Alternatively, __d_lookup_rcu may be called again to look up the child of
2161  * the returned dentry, so long as its parent's seqlock is checked after the
2162  * child is looked up. Thus, an interlocking stepping of sequence lock checks
2163  * is formed, giving integrity down the path walk.
2164  *
2165  * NOTE! The caller *has* to check the resulting dentry against the sequence
2166  * number we've returned before using any of the resulting dentry state!
2167  */
2168 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2169                                 const struct qstr *name,
2170                                 unsigned *seqp)
2171 {
2172         u64 hashlen = name->hash_len;
2173         const unsigned char *str = name->name;
2174         struct hlist_bl_head *b = d_hash(hashlen_hash(hashlen));
2175         struct hlist_bl_node *node;
2176         struct dentry *dentry;
2177
2178         /*
2179          * Note: There is significant duplication with __d_lookup_rcu which is
2180          * required to prevent single threaded performance regressions
2181          * especially on architectures where smp_rmb (in seqcounts) are costly.
2182          * Keep the two functions in sync.
2183          */
2184
2185         /*
2186          * The hash list is protected using RCU.
2187          *
2188          * Carefully use d_seq when comparing a candidate dentry, to avoid
2189          * races with d_move().
2190          *
2191          * It is possible that concurrent renames can mess up our list
2192          * walk here and result in missing our dentry, resulting in the
2193          * false-negative result. d_lookup() protects against concurrent
2194          * renames using rename_lock seqlock.
2195          *
2196          * See Documentation/filesystems/path-lookup.txt for more details.
2197          */
2198         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2199                 unsigned seq;
2200
2201 seqretry:
2202                 /*
2203                  * The dentry sequence count protects us from concurrent
2204                  * renames, and thus protects parent and name fields.
2205                  *
2206                  * The caller must perform a seqcount check in order
2207                  * to do anything useful with the returned dentry.
2208                  *
2209                  * NOTE! We do a "raw" seqcount_begin here. That means that
2210                  * we don't wait for the sequence count to stabilize if it
2211                  * is in the middle of a sequence change. If we do the slow
2212                  * dentry compare, we will do seqretries until it is stable,
2213                  * and if we end up with a successful lookup, we actually
2214                  * want to exit RCU lookup anyway.
2215                  *
2216                  * Note that raw_seqcount_begin still *does* smp_rmb(), so
2217                  * we are still guaranteed NUL-termination of ->d_name.name.
2218                  */
2219                 seq = raw_seqcount_begin(&dentry->d_seq);
2220                 if (dentry->d_parent != parent)
2221                         continue;
2222                 if (d_unhashed(dentry))
2223                         continue;
2224
2225                 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) {
2226                         int tlen;
2227                         const char *tname;
2228                         if (dentry->d_name.hash != hashlen_hash(hashlen))
2229                                 continue;
2230                         tlen = dentry->d_name.len;
2231                         tname = dentry->d_name.name;
2232                         /* we want a consistent (name,len) pair */
2233                         if (read_seqcount_retry(&dentry->d_seq, seq)) {
2234                                 cpu_relax();
2235                                 goto seqretry;
2236                         }
2237                         if (parent->d_op->d_compare(dentry,
2238                                                     tlen, tname, name) != 0)
2239                                 continue;
2240                 } else {
2241                         if (dentry->d_name.hash_len != hashlen)
2242                                 continue;
2243                         if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
2244                                 continue;
2245                 }
2246                 *seqp = seq;
2247                 return dentry;
2248         }
2249         return NULL;
2250 }
2251
2252 /**
2253  * d_lookup - search for a dentry
2254  * @parent: parent dentry
2255  * @name: qstr of name we wish to find
2256  * Returns: dentry, or NULL
2257  *
2258  * d_lookup searches the children of the parent dentry for the name in
2259  * question. If the dentry is found its reference count is incremented and the
2260  * dentry is returned. The caller must use dput to free the entry when it has
2261  * finished using it. %NULL is returned if the dentry does not exist.
2262  */
2263 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2264 {
2265         struct dentry *dentry;
2266         unsigned seq;
2267
2268         do {
2269                 seq = read_seqbegin(&rename_lock);
2270                 dentry = __d_lookup(parent, name);
2271                 if (dentry)
2272                         break;
2273         } while (read_seqretry(&rename_lock, seq));
2274         return dentry;
2275 }
2276 EXPORT_SYMBOL(d_lookup);
2277
2278 /**
2279  * __d_lookup - search for a dentry (racy)
2280  * @parent: parent dentry
2281  * @name: qstr of name we wish to find
2282  * Returns: dentry, or NULL
2283  *
2284  * __d_lookup is like d_lookup, however it may (rarely) return a
2285  * false-negative result due to unrelated rename activity.
2286  *
2287  * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2288  * however it must be used carefully, eg. with a following d_lookup in
2289  * the case of failure.
2290  *
2291  * __d_lookup callers must be commented.
2292  */
2293 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2294 {
2295         unsigned int hash = name->hash;
2296         struct hlist_bl_head *b = d_hash(hash);
2297         struct hlist_bl_node *node;
2298         struct dentry *found = NULL;
2299         struct dentry *dentry;
2300
2301         /*
2302          * Note: There is significant duplication with __d_lookup_rcu which is
2303          * required to prevent single threaded performance regressions
2304          * especially on architectures where smp_rmb (in seqcounts) are costly.
2305          * Keep the two functions in sync.
2306          */
2307
2308         /*
2309          * The hash list is protected using RCU.
2310          *
2311          * Take d_lock when comparing a candidate dentry, to avoid races
2312          * with d_move().
2313          *
2314          * It is possible that concurrent renames can mess up our list
2315          * walk here and result in missing our dentry, resulting in the
2316          * false-negative result. d_lookup() protects against concurrent
2317          * renames using rename_lock seqlock.
2318          *
2319          * See Documentation/filesystems/path-lookup.txt for more details.
2320          */
2321         rcu_read_lock();
2322         
2323         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2324
2325                 if (dentry->d_name.hash != hash)
2326                         continue;
2327
2328                 spin_lock(&dentry->d_lock);
2329                 if (dentry->d_parent != parent)
2330                         goto next;
2331                 if (d_unhashed(dentry))
2332                         goto next;
2333
2334                 if (!d_same_name(dentry, parent, name))
2335                         goto next;
2336
2337                 dentry->d_lockref.count++;
2338                 found = dentry;
2339                 spin_unlock(&dentry->d_lock);
2340                 break;
2341 next:
2342                 spin_unlock(&dentry->d_lock);
2343         }
2344         rcu_read_unlock();
2345
2346         return found;
2347 }
2348
2349 /**
2350  * d_hash_and_lookup - hash the qstr then search for a dentry
2351  * @dir: Directory to search in
2352  * @name: qstr of name we wish to find
2353  *
2354  * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2355  */
2356 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2357 {
2358         /*
2359          * Check for a fs-specific hash function. Note that we must
2360          * calculate the standard hash first, as the d_op->d_hash()
2361          * routine may choose to leave the hash value unchanged.
2362          */
2363         name->hash = full_name_hash(dir, name->name, name->len);
2364         if (dir->d_flags & DCACHE_OP_HASH) {
2365                 int err = dir->d_op->d_hash(dir, name);
2366                 if (unlikely(err < 0))
2367                         return ERR_PTR(err);
2368         }
2369         return d_lookup(dir, name);
2370 }
2371 EXPORT_SYMBOL(d_hash_and_lookup);
2372
2373 /*
2374  * When a file is deleted, we have two options:
2375  * - turn this dentry into a negative dentry
2376  * - unhash this dentry and free it.
2377  *
2378  * Usually, we want to just turn this into
2379  * a negative dentry, but if anybody else is
2380  * currently using the dentry or the inode
2381  * we can't do that and we fall back on removing
2382  * it from the hash queues and waiting for
2383  * it to be deleted later when it has no users
2384  */
2385  
2386 /**
2387  * d_delete - delete a dentry
2388  * @dentry: The dentry to delete
2389  *
2390  * Turn the dentry into a negative dentry if possible, otherwise
2391  * remove it from the hash queues so it can be deleted later
2392  */
2393  
2394 void d_delete(struct dentry * dentry)
2395 {
2396         struct inode *inode = dentry->d_inode;
2397         int isdir = d_is_dir(dentry);
2398
2399         spin_lock(&inode->i_lock);
2400         spin_lock(&dentry->d_lock);
2401         /*
2402          * Are we the only user?
2403          */
2404         if (dentry->d_lockref.count == 1) {
2405                 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2406                 dentry_unlink_inode(dentry);
2407         } else {
2408                 __d_drop(dentry);
2409                 spin_unlock(&dentry->d_lock);
2410                 spin_unlock(&inode->i_lock);
2411         }
2412         fsnotify_nameremove(dentry, isdir);
2413 }
2414 EXPORT_SYMBOL(d_delete);
2415
2416 static void __d_rehash(struct dentry *entry)
2417 {
2418         struct hlist_bl_head *b = d_hash(entry->d_name.hash);
2419
2420         hlist_bl_lock(b);
2421         hlist_bl_add_head_rcu(&entry->d_hash, b);
2422         hlist_bl_unlock(b);
2423 }
2424
2425 /**
2426  * d_rehash     - add an entry back to the hash
2427  * @entry: dentry to add to the hash
2428  *
2429  * Adds a dentry to the hash according to its name.
2430  */
2431  
2432 void d_rehash(struct dentry * entry)
2433 {
2434         spin_lock(&entry->d_lock);
2435         __d_rehash(entry);
2436         spin_unlock(&entry->d_lock);
2437 }
2438 EXPORT_SYMBOL(d_rehash);
2439
2440 static inline unsigned start_dir_add(struct inode *dir)
2441 {
2442
2443         for (;;) {
2444                 unsigned n = dir->i_dir_seq;
2445                 if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
2446                         return n;
2447                 cpu_relax();
2448         }
2449 }
2450
2451 static inline void end_dir_add(struct inode *dir, unsigned n)
2452 {
2453         smp_store_release(&dir->i_dir_seq, n + 2);
2454 }
2455
2456 static void d_wait_lookup(struct dentry *dentry)
2457 {
2458         if (d_in_lookup(dentry)) {
2459                 DECLARE_WAITQUEUE(wait, current);
2460                 add_wait_queue(dentry->d_wait, &wait);
2461                 do {
2462                         set_current_state(TASK_UNINTERRUPTIBLE);
2463                         spin_unlock(&dentry->d_lock);
2464                         schedule();
2465                         spin_lock(&dentry->d_lock);
2466                 } while (d_in_lookup(dentry));
2467         }
2468 }
2469
2470 struct dentry *d_alloc_parallel(struct dentry *parent,
2471                                 const struct qstr *name,
2472                                 wait_queue_head_t *wq)
2473 {
2474         unsigned int hash = name->hash;
2475         struct hlist_bl_head *b = in_lookup_hash(parent, hash);
2476         struct hlist_bl_node *node;
2477         struct dentry *new = d_alloc(parent, name);
2478         struct dentry *dentry;
2479         unsigned seq, r_seq, d_seq;
2480
2481         if (unlikely(!new))
2482                 return ERR_PTR(-ENOMEM);
2483
2484 retry:
2485         rcu_read_lock();
2486         seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
2487         r_seq = read_seqbegin(&rename_lock);
2488         dentry = __d_lookup_rcu(parent, name, &d_seq);
2489         if (unlikely(dentry)) {
2490                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2491                         rcu_read_unlock();
2492                         goto retry;
2493                 }
2494                 if (read_seqcount_retry(&dentry->d_seq, d_seq)) {
2495                         rcu_read_unlock();
2496                         dput(dentry);
2497                         goto retry;
2498                 }
2499                 rcu_read_unlock();
2500                 dput(new);
2501                 return dentry;
2502         }
2503         if (unlikely(read_seqretry(&rename_lock, r_seq))) {
2504                 rcu_read_unlock();
2505                 goto retry;
2506         }
2507
2508         if (unlikely(seq & 1)) {
2509                 rcu_read_unlock();
2510                 goto retry;
2511         }
2512
2513         hlist_bl_lock(b);
2514         if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {
2515                 hlist_bl_unlock(b);
2516                 rcu_read_unlock();
2517                 goto retry;
2518         }
2519         /*
2520          * No changes for the parent since the beginning of d_lookup().
2521          * Since all removals from the chain happen with hlist_bl_lock(),
2522          * any potential in-lookup matches are going to stay here until
2523          * we unlock the chain.  All fields are stable in everything
2524          * we encounter.
2525          */
2526         hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
2527                 if (dentry->d_name.hash != hash)
2528                         continue;
2529                 if (dentry->d_parent != parent)
2530                         continue;
2531                 if (!d_same_name(dentry, parent, name))
2532                         continue;
2533                 hlist_bl_unlock(b);
2534                 /* now we can try to grab a reference */
2535                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2536                         rcu_read_unlock();
2537                         goto retry;
2538                 }
2539
2540                 rcu_read_unlock();
2541                 /*
2542                  * somebody is likely to be still doing lookup for it;
2543                  * wait for them to finish
2544                  */
2545                 spin_lock(&dentry->d_lock);
2546                 d_wait_lookup(dentry);
2547                 /*
2548                  * it's not in-lookup anymore; in principle we should repeat
2549                  * everything from dcache lookup, but it's likely to be what
2550                  * d_lookup() would've found anyway.  If it is, just return it;
2551                  * otherwise we really have to repeat the whole thing.
2552                  */
2553                 if (unlikely(dentry->d_name.hash != hash))
2554                         goto mismatch;
2555                 if (unlikely(dentry->d_parent != parent))
2556                         goto mismatch;
2557                 if (unlikely(d_unhashed(dentry)))
2558                         goto mismatch;
2559                 if (unlikely(!d_same_name(dentry, parent, name)))
2560                         goto mismatch;
2561                 /* OK, it *is* a hashed match; return it */
2562                 spin_unlock(&dentry->d_lock);
2563                 dput(new);
2564                 return dentry;
2565         }
2566         rcu_read_unlock();
2567         /* we can't take ->d_lock here; it's OK, though. */
2568         new->d_flags |= DCACHE_PAR_LOOKUP;
2569         new->d_wait = wq;
2570         hlist_bl_add_head_rcu(&new->d_u.d_in_lookup_hash, b);
2571         hlist_bl_unlock(b);
2572         return new;
2573 mismatch:
2574         spin_unlock(&dentry->d_lock);
2575         dput(dentry);
2576         goto retry;
2577 }
2578 EXPORT_SYMBOL(d_alloc_parallel);
2579
2580 void __d_lookup_done(struct dentry *dentry)
2581 {
2582         struct hlist_bl_head *b = in_lookup_hash(dentry->d_parent,
2583                                                  dentry->d_name.hash);
2584         hlist_bl_lock(b);
2585         dentry->d_flags &= ~DCACHE_PAR_LOOKUP;
2586         __hlist_bl_del(&dentry->d_u.d_in_lookup_hash);
2587         wake_up_all(dentry->d_wait);
2588         dentry->d_wait = NULL;
2589         hlist_bl_unlock(b);
2590         INIT_HLIST_NODE(&dentry->d_u.d_alias);
2591         INIT_LIST_HEAD(&dentry->d_lru);
2592 }
2593 EXPORT_SYMBOL(__d_lookup_done);
2594
2595 /* inode->i_lock held if inode is non-NULL */
2596
2597 static inline void __d_add(struct dentry *dentry, struct inode *inode)
2598 {
2599         struct inode *dir = NULL;
2600         unsigned n;
2601         spin_lock(&dentry->d_lock);
2602         if (unlikely(d_in_lookup(dentry))) {
2603                 dir = dentry->d_parent->d_inode;
2604                 n = start_dir_add(dir);
2605                 __d_lookup_done(dentry);
2606         }
2607         if (inode) {
2608                 unsigned add_flags = d_flags_for_inode(inode);
2609                 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2610                 raw_write_seqcount_begin(&dentry->d_seq);
2611                 __d_set_inode_and_type(dentry, inode, add_flags);
2612                 raw_write_seqcount_end(&dentry->d_seq);
2613                 fsnotify_update_flags(dentry);
2614         }
2615         __d_rehash(dentry);
2616         if (dir)
2617                 end_dir_add(dir, n);
2618         spin_unlock(&dentry->d_lock);
2619         if (inode)
2620                 spin_unlock(&inode->i_lock);
2621 }
2622
2623 /**
2624  * d_add - add dentry to hash queues
2625  * @entry: dentry to add
2626  * @inode: The inode to attach to this dentry
2627  *
2628  * This adds the entry to the hash queues and initializes @inode.
2629  * The entry was actually filled in earlier during d_alloc().
2630  */
2631
2632 void d_add(struct dentry *entry, struct inode *inode)
2633 {
2634         if (inode) {
2635                 security_d_instantiate(entry, inode);
2636                 spin_lock(&inode->i_lock);
2637         }
2638         __d_add(entry, inode);
2639 }
2640 EXPORT_SYMBOL(d_add);
2641
2642 /**
2643  * d_exact_alias - find and hash an exact unhashed alias
2644  * @entry: dentry to add
2645  * @inode: The inode to go with this dentry
2646  *
2647  * If an unhashed dentry with the same name/parent and desired
2648  * inode already exists, hash and return it.  Otherwise, return
2649  * NULL.
2650  *
2651  * Parent directory should be locked.
2652  */
2653 struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2654 {
2655         struct dentry *alias;
2656         unsigned int hash = entry->d_name.hash;
2657
2658         spin_lock(&inode->i_lock);
2659         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2660                 /*
2661                  * Don't need alias->d_lock here, because aliases with
2662                  * d_parent == entry->d_parent are not subject to name or
2663                  * parent changes, because the parent inode i_mutex is held.
2664                  */
2665                 if (alias->d_name.hash != hash)
2666                         continue;
2667                 if (alias->d_parent != entry->d_parent)
2668                         continue;
2669                 if (!d_same_name(alias, entry->d_parent, &entry->d_name))
2670                         continue;
2671                 spin_lock(&alias->d_lock);
2672                 if (!d_unhashed(alias)) {
2673                         spin_unlock(&alias->d_lock);
2674                         alias = NULL;
2675                 } else {
2676                         __dget_dlock(alias);
2677                         __d_rehash(alias);
2678                         spin_unlock(&alias->d_lock);
2679                 }
2680                 spin_unlock(&inode->i_lock);
2681                 return alias;
2682         }
2683         spin_unlock(&inode->i_lock);
2684         return NULL;
2685 }
2686 EXPORT_SYMBOL(d_exact_alias);
2687
2688 /**
2689  * dentry_update_name_case - update case insensitive dentry with a new name
2690  * @dentry: dentry to be updated
2691  * @name: new name
2692  *
2693  * Update a case insensitive dentry with new case of name.
2694  *
2695  * dentry must have been returned by d_lookup with name @name. Old and new
2696  * name lengths must match (ie. no d_compare which allows mismatched name
2697  * lengths).
2698  *
2699  * Parent inode i_mutex must be held over d_lookup and into this call (to
2700  * keep renames and concurrent inserts, and readdir(2) away).
2701  */
2702 void dentry_update_name_case(struct dentry *dentry, const struct qstr *name)
2703 {
2704         BUG_ON(!inode_is_locked(dentry->d_parent->d_inode));
2705         BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */
2706
2707         spin_lock(&dentry->d_lock);
2708         write_seqcount_begin(&dentry->d_seq);
2709         memcpy((unsigned char *)dentry->d_name.name, name->name, name->len);
2710         write_seqcount_end(&dentry->d_seq);
2711         spin_unlock(&dentry->d_lock);
2712 }
2713 EXPORT_SYMBOL(dentry_update_name_case);
2714
2715 static void swap_names(struct dentry *dentry, struct dentry *target)
2716 {
2717         if (unlikely(dname_external(target))) {
2718                 if (unlikely(dname_external(dentry))) {
2719                         /*
2720                          * Both external: swap the pointers
2721                          */
2722                         swap(target->d_name.name, dentry->d_name.name);
2723                 } else {
2724                         /*
2725                          * dentry:internal, target:external.  Steal target's
2726                          * storage and make target internal.
2727                          */
2728                         memcpy(target->d_iname, dentry->d_name.name,
2729                                         dentry->d_name.len + 1);
2730                         dentry->d_name.name = target->d_name.name;
2731                         target->d_name.name = target->d_iname;
2732                 }
2733         } else {
2734                 if (unlikely(dname_external(dentry))) {
2735                         /*
2736                          * dentry:external, target:internal.  Give dentry's
2737                          * storage to target and make dentry internal
2738                          */
2739                         memcpy(dentry->d_iname, target->d_name.name,
2740                                         target->d_name.len + 1);
2741                         target->d_name.name = dentry->d_name.name;
2742                         dentry->d_name.name = dentry->d_iname;
2743                 } else {
2744                         /*
2745                          * Both are internal.
2746                          */
2747                         unsigned int i;
2748                         BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2749                         for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2750                                 swap(((long *) &dentry->d_iname)[i],
2751                                      ((long *) &target->d_iname)[i]);
2752                         }
2753                 }
2754         }
2755         swap(dentry->d_name.hash_len, target->d_name.hash_len);
2756 }
2757
2758 static void copy_name(struct dentry *dentry, struct dentry *target)
2759 {
2760         struct external_name *old_name = NULL;
2761         if (unlikely(dname_external(dentry)))
2762                 old_name = external_name(dentry);
2763         if (unlikely(dname_external(target))) {
2764                 atomic_inc(&external_name(target)->u.count);
2765                 dentry->d_name = target->d_name;
2766         } else {
2767                 memcpy(dentry->d_iname, target->d_name.name,
2768                                 target->d_name.len + 1);
2769                 dentry->d_name.name = dentry->d_iname;
2770                 dentry->d_name.hash_len = target->d_name.hash_len;
2771         }
2772         if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2773                 kfree_rcu(old_name, u.head);
2774 }
2775
2776 static void dentry_lock_for_move(struct dentry *dentry, struct dentry *target)
2777 {
2778         if (IS_ROOT(dentry) || dentry->d_parent == target->d_parent)
2779                 spin_lock(&target->d_parent->d_lock);
2780         else {
2781                 if (d_ancestor(dentry->d_parent, target->d_parent)) {
2782                         spin_lock(&dentry->d_parent->d_lock);
2783                         spin_lock_nested(&target->d_parent->d_lock,
2784                                                 DENTRY_D_LOCK_NESTED);
2785                 } else {
2786                         spin_lock(&target->d_parent->d_lock);
2787                         spin_lock_nested(&dentry->d_parent->d_lock,
2788                                                 DENTRY_D_LOCK_NESTED);
2789                 }
2790         }
2791         spin_lock_nested(&dentry->d_lock, 2);
2792         spin_lock_nested(&target->d_lock, 3);
2793 }
2794
2795 /*
2796  * __d_move - move a dentry
2797  * @dentry: entry to move
2798  * @target: new dentry
2799  * @exchange: exchange the two dentries
2800  *
2801  * Update the dcache to reflect the move of a file name. Negative
2802  * dcache entries should not be moved in this way. Caller must hold
2803  * rename_lock, the i_mutex of the source and target directories,
2804  * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2805  */
2806 static void __d_move(struct dentry *dentry, struct dentry *target,
2807                      bool exchange)
2808 {
2809         struct dentry *old_parent;
2810         struct inode *dir = NULL;
2811         unsigned n;
2812         if (!dentry->d_inode)
2813                 printk(KERN_WARNING "VFS: moving negative dcache entry\n");
2814
2815         BUG_ON(d_ancestor(dentry, target));
2816         BUG_ON(d_ancestor(target, dentry));
2817
2818         dentry_lock_for_move(dentry, target);
2819         if (unlikely(d_in_lookup(target))) {
2820                 dir = target->d_parent->d_inode;
2821                 n = start_dir_add(dir);
2822                 __d_lookup_done(target);
2823         }
2824
2825         write_seqcount_begin(&dentry->d_seq);
2826         write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2827
2828         old_parent = dentry->d_parent;
2829
2830         /* unhash both */
2831         if (!d_unhashed(dentry))
2832                 ___d_drop(dentry);
2833         if (!d_unhashed(target))
2834                 ___d_drop(target);
2835
2836         /* ... and switch them in the tree */
2837         dentry->d_parent = target->d_parent;
2838         if (!exchange) {
2839                 copy_name(dentry, target);
2840                 target->d_hash.pprev = NULL;
2841                 dentry->d_parent->d_lockref.count++;
2842                 if (dentry == old_parent)
2843                         dentry->d_flags |= DCACHE_RCUACCESS;
2844                 else
2845                         WARN_ON(!--old_parent->d_lockref.count);
2846         } else {
2847                 target->d_parent = old_parent;
2848                 swap_names(dentry, target);
2849                 list_move(&target->d_child, &target->d_parent->d_subdirs);
2850                 __d_rehash(target);
2851                 fsnotify_update_flags(target);
2852         }
2853         list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2854         __d_rehash(dentry);
2855         fsnotify_update_flags(dentry);
2856
2857         write_seqcount_end(&target->d_seq);
2858         write_seqcount_end(&dentry->d_seq);
2859
2860         if (dir)
2861                 end_dir_add(dir, n);
2862
2863         if (dentry->d_parent != old_parent)
2864                 spin_unlock(&dentry->d_parent->d_lock);
2865         if (dentry != old_parent)
2866                 spin_unlock(&old_parent->d_lock);
2867         spin_unlock(&target->d_lock);
2868         spin_unlock(&dentry->d_lock);
2869 }
2870
2871 /*
2872  * d_move - move a dentry
2873  * @dentry: entry to move
2874  * @target: new dentry
2875  *
2876  * Update the dcache to reflect the move of a file name. Negative
2877  * dcache entries should not be moved in this way. See the locking
2878  * requirements for __d_move.
2879  */
2880 void d_move(struct dentry *dentry, struct dentry *target)
2881 {
2882         write_seqlock(&rename_lock);
2883         __d_move(dentry, target, false);
2884         write_sequnlock(&rename_lock);
2885 }
2886 EXPORT_SYMBOL(d_move);
2887
2888 /*
2889  * d_exchange - exchange two dentries
2890  * @dentry1: first dentry
2891  * @dentry2: second dentry
2892  */
2893 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2894 {
2895         write_seqlock(&rename_lock);
2896
2897         WARN_ON(!dentry1->d_inode);
2898         WARN_ON(!dentry2->d_inode);
2899         WARN_ON(IS_ROOT(dentry1));
2900         WARN_ON(IS_ROOT(dentry2));
2901
2902         __d_move(dentry1, dentry2, true);
2903
2904         write_sequnlock(&rename_lock);
2905 }
2906
2907 /**
2908  * d_ancestor - search for an ancestor
2909  * @p1: ancestor dentry
2910  * @p2: child dentry
2911  *
2912  * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2913  * an ancestor of p2, else NULL.
2914  */
2915 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2916 {
2917         struct dentry *p;
2918
2919         for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2920                 if (p->d_parent == p1)
2921                         return p;
2922         }
2923         return NULL;
2924 }
2925
2926 /*
2927  * This helper attempts to cope with remotely renamed directories
2928  *
2929  * It assumes that the caller is already holding
2930  * dentry->d_parent->d_inode->i_mutex, and rename_lock
2931  *
2932  * Note: If ever the locking in lock_rename() changes, then please
2933  * remember to update this too...
2934  */
2935 static int __d_unalias(struct inode *inode,
2936                 struct dentry *dentry, struct dentry *alias)
2937 {
2938         struct mutex *m1 = NULL;
2939         struct rw_semaphore *m2 = NULL;
2940         int ret = -ESTALE;
2941
2942         /* If alias and dentry share a parent, then no extra locks required */
2943         if (alias->d_parent == dentry->d_parent)
2944                 goto out_unalias;
2945
2946         /* See lock_rename() */
2947         if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2948                 goto out_err;
2949         m1 = &dentry->d_sb->s_vfs_rename_mutex;
2950         if (!inode_trylock_shared(alias->d_parent->d_inode))
2951                 goto out_err;
2952         m2 = &alias->d_parent->d_inode->i_rwsem;
2953 out_unalias:
2954         __d_move(alias, dentry, false);
2955         ret = 0;
2956 out_err:
2957         if (m2)
2958                 up_read(m2);
2959         if (m1)
2960                 mutex_unlock(m1);
2961         return ret;
2962 }
2963
2964 /**
2965  * d_splice_alias - splice a disconnected dentry into the tree if one exists
2966  * @inode:  the inode which may have a disconnected dentry
2967  * @dentry: a negative dentry which we want to point to the inode.
2968  *
2969  * If inode is a directory and has an IS_ROOT alias, then d_move that in
2970  * place of the given dentry and return it, else simply d_add the inode
2971  * to the dentry and return NULL.
2972  *
2973  * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2974  * we should error out: directories can't have multiple aliases.
2975  *
2976  * This is needed in the lookup routine of any filesystem that is exportable
2977  * (via knfsd) so that we can build dcache paths to directories effectively.
2978  *
2979  * If a dentry was found and moved, then it is returned.  Otherwise NULL
2980  * is returned.  This matches the expected return value of ->lookup.
2981  *
2982  * Cluster filesystems may call this function with a negative, hashed dentry.
2983  * In that case, we know that the inode will be a regular file, and also this
2984  * will only occur during atomic_open. So we need to check for the dentry
2985  * being already hashed only in the final case.
2986  */
2987 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
2988 {
2989         if (IS_ERR(inode))
2990                 return ERR_CAST(inode);
2991
2992         BUG_ON(!d_unhashed(dentry));
2993
2994         if (!inode)
2995                 goto out;
2996
2997         security_d_instantiate(dentry, inode);
2998         spin_lock(&inode->i_lock);
2999         if (S_ISDIR(inode->i_mode)) {
3000                 struct dentry *new = __d_find_any_alias(inode);
3001                 if (unlikely(new)) {
3002                         /* The reference to new ensures it remains an alias */
3003                         spin_unlock(&inode->i_lock);
3004                         write_seqlock(&rename_lock);
3005                         if (unlikely(d_ancestor(new, dentry))) {
3006                                 write_sequnlock(&rename_lock);
3007                                 dput(new);
3008                                 new = ERR_PTR(-ELOOP);
3009                                 pr_warn_ratelimited(
3010                                         "VFS: Lookup of '%s' in %s %s"
3011                                         " would have caused loop\n",
3012                                         dentry->d_name.name,
3013                                         inode->i_sb->s_type->name,
3014                                         inode->i_sb->s_id);
3015                         } else if (!IS_ROOT(new)) {
3016                                 struct dentry *old_parent = dget(new->d_parent);
3017                                 int err = __d_unalias(inode, dentry, new);
3018                                 write_sequnlock(&rename_lock);
3019                                 if (err) {
3020                                         dput(new);
3021                                         new = ERR_PTR(err);
3022                                 }
3023                                 dput(old_parent);
3024                         } else {
3025                                 __d_move(new, dentry, false);
3026                                 write_sequnlock(&rename_lock);
3027                         }
3028                         iput(inode);
3029                         return new;
3030                 }
3031         }
3032 out:
3033         __d_add(dentry, inode);
3034         return NULL;
3035 }
3036 EXPORT_SYMBOL(d_splice_alias);
3037
3038 /*
3039  * Test whether new_dentry is a subdirectory of old_dentry.
3040  *
3041  * Trivially implemented using the dcache structure
3042  */
3043
3044 /**
3045  * is_subdir - is new dentry a subdirectory of old_dentry
3046  * @new_dentry: new dentry
3047  * @old_dentry: old dentry
3048  *
3049  * Returns true if new_dentry is a subdirectory of the parent (at any depth).
3050  * Returns false otherwise.
3051  * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3052  */
3053   
3054 bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3055 {
3056         bool result;
3057         unsigned seq;
3058
3059         if (new_dentry == old_dentry)
3060                 return true;
3061
3062         do {
3063                 /* for restarting inner loop in case of seq retry */
3064                 seq = read_seqbegin(&rename_lock);
3065                 /*
3066                  * Need rcu_readlock to protect against the d_parent trashing
3067                  * due to d_move
3068                  */
3069                 rcu_read_lock();
3070                 if (d_ancestor(old_dentry, new_dentry))
3071                         result = true;
3072                 else
3073                         result = false;
3074                 rcu_read_unlock();
3075         } while (read_seqretry(&rename_lock, seq));
3076
3077         return result;
3078 }
3079 EXPORT_SYMBOL(is_subdir);
3080
3081 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3082 {
3083         struct dentry *root = data;
3084         if (dentry != root) {
3085                 if (d_unhashed(dentry) || !dentry->d_inode)
3086                         return D_WALK_SKIP;
3087
3088                 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3089                         dentry->d_flags |= DCACHE_GENOCIDE;
3090                         dentry->d_lockref.count--;
3091                 }
3092         }
3093         return D_WALK_CONTINUE;
3094 }
3095
3096 void d_genocide(struct dentry *parent)
3097 {
3098         d_walk(parent, parent, d_genocide_kill, NULL);
3099 }
3100
3101 void d_tmpfile(struct dentry *dentry, struct inode *inode)
3102 {
3103         inode_dec_link_count(inode);
3104         BUG_ON(dentry->d_name.name != dentry->d_iname ||
3105                 !hlist_unhashed(&dentry->d_u.d_alias) ||
3106                 !d_unlinked(dentry));
3107         spin_lock(&dentry->d_parent->d_lock);
3108         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3109         dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3110                                 (unsigned long long)inode->i_ino);
3111         spin_unlock(&dentry->d_lock);
3112         spin_unlock(&dentry->d_parent->d_lock);
3113         d_instantiate(dentry, inode);
3114 }
3115 EXPORT_SYMBOL(d_tmpfile);
3116
3117 static __initdata unsigned long dhash_entries;
3118 static int __init set_dhash_entries(char *str)
3119 {
3120         if (!str)
3121                 return 0;
3122         dhash_entries = simple_strtoul(str, &str, 0);
3123         return 1;
3124 }
3125 __setup("dhash_entries=", set_dhash_entries);
3126
3127 static void __init dcache_init_early(void)
3128 {
3129         /* If hashes are distributed across NUMA nodes, defer
3130          * hash allocation until vmalloc space is available.
3131          */
3132         if (hashdist)
3133                 return;
3134
3135         dentry_hashtable =
3136                 alloc_large_system_hash("Dentry cache",
3137                                         sizeof(struct hlist_bl_head),
3138                                         dhash_entries,
3139                                         13,
3140                                         HASH_EARLY | HASH_ZERO,
3141                                         &d_hash_shift,
3142                                         NULL,
3143                                         0,
3144                                         0);
3145         d_hash_shift = 32 - d_hash_shift;
3146 }
3147
3148 static void __init dcache_init(void)
3149 {
3150         /*
3151          * A constructor could be added for stable state like the lists,
3152          * but it is probably not worth it because of the cache nature
3153          * of the dcache.
3154          */
3155         dentry_cache = KMEM_CACHE_USERCOPY(dentry,
3156                 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD|SLAB_ACCOUNT,
3157                 d_iname);
3158
3159         /* Hash may have been set up in dcache_init_early */
3160         if (!hashdist)
3161                 return;
3162
3163         dentry_hashtable =
3164                 alloc_large_system_hash("Dentry cache",
3165                                         sizeof(struct hlist_bl_head),
3166                                         dhash_entries,
3167                                         13,
3168                                         HASH_ZERO,
3169                                         &d_hash_shift,
3170                                         NULL,
3171                                         0,
3172                                         0);
3173         d_hash_shift = 32 - d_hash_shift;
3174 }
3175
3176 /* SLAB cache for __getname() consumers */
3177 struct kmem_cache *names_cachep __read_mostly;
3178 EXPORT_SYMBOL(names_cachep);
3179
3180 EXPORT_SYMBOL(d_genocide);
3181
3182 void __init vfs_caches_init_early(void)
3183 {
3184         int i;
3185
3186         for (i = 0; i < ARRAY_SIZE(in_lookup_hashtable); i++)
3187                 INIT_HLIST_BL_HEAD(&in_lookup_hashtable[i]);
3188
3189         dcache_init_early();
3190         inode_init_early();
3191 }
3192
3193 void __init vfs_caches_init(void)
3194 {
3195         names_cachep = kmem_cache_create_usercopy("names_cache", PATH_MAX, 0,
3196                         SLAB_HWCACHE_ALIGN|SLAB_PANIC, 0, PATH_MAX, NULL);
3197
3198         dcache_init();
3199         inode_init();
3200         files_init();
3201         files_maxfiles_init();
3202         mnt_init();
3203         bdev_cache_init();
3204         chrdev_init();
3205 }