GNU Linux-libre 4.19.314-gnu1
[releases.git] / fs / dcache.c
1 /*
2  * fs/dcache.c
3  *
4  * Complete reimplementation
5  * (C) 1997 Thomas Schoebel-Theuer,
6  * with heavy changes by Linus Torvalds
7  */
8
9 /*
10  * Notes on the allocation strategy:
11  *
12  * The dcache is a master of the icache - whenever a dcache entry
13  * exists, the inode will always exist. "iput()" is done either when
14  * the dcache entry is deleted or garbage collected.
15  */
16
17 #include <linux/ratelimit.h>
18 #include <linux/string.h>
19 #include <linux/mm.h>
20 #include <linux/fs.h>
21 #include <linux/fsnotify.h>
22 #include <linux/slab.h>
23 #include <linux/init.h>
24 #include <linux/hash.h>
25 #include <linux/cache.h>
26 #include <linux/export.h>
27 #include <linux/security.h>
28 #include <linux/seqlock.h>
29 #include <linux/bootmem.h>
30 #include <linux/bit_spinlock.h>
31 #include <linux/rculist_bl.h>
32 #include <linux/list_lru.h>
33 #include "internal.h"
34 #include "mount.h"
35
36 /*
37  * Usage:
38  * dcache->d_inode->i_lock protects:
39  *   - i_dentry, d_u.d_alias, d_inode of aliases
40  * dcache_hash_bucket lock protects:
41  *   - the dcache hash table
42  * s_roots bl list spinlock protects:
43  *   - the s_roots list (see __d_drop)
44  * dentry->d_sb->s_dentry_lru_lock protects:
45  *   - the dcache lru lists and counters
46  * d_lock protects:
47  *   - d_flags
48  *   - d_name
49  *   - d_lru
50  *   - d_count
51  *   - d_unhashed()
52  *   - d_parent and d_subdirs
53  *   - childrens' d_child and d_parent
54  *   - d_u.d_alias, d_inode
55  *
56  * Ordering:
57  * dentry->d_inode->i_lock
58  *   dentry->d_lock
59  *     dentry->d_sb->s_dentry_lru_lock
60  *     dcache_hash_bucket lock
61  *     s_roots lock
62  *
63  * If there is an ancestor relationship:
64  * dentry->d_parent->...->d_parent->d_lock
65  *   ...
66  *     dentry->d_parent->d_lock
67  *       dentry->d_lock
68  *
69  * If no ancestor relationship:
70  * arbitrary, since it's serialized on rename_lock
71  */
72 int sysctl_vfs_cache_pressure __read_mostly = 100;
73 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
74
75 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
76
77 EXPORT_SYMBOL(rename_lock);
78
79 static struct kmem_cache *dentry_cache __read_mostly;
80
81 const struct qstr empty_name = QSTR_INIT("", 0);
82 EXPORT_SYMBOL(empty_name);
83 const struct qstr slash_name = QSTR_INIT("/", 1);
84 EXPORT_SYMBOL(slash_name);
85
86 /*
87  * This is the single most critical data structure when it comes
88  * to the dcache: the hashtable for lookups. Somebody should try
89  * to make this good - I've just made it work.
90  *
91  * This hash-function tries to avoid losing too many bits of hash
92  * information, yet avoid using a prime hash-size or similar.
93  */
94
95 static unsigned int d_hash_shift __read_mostly;
96
97 static struct hlist_bl_head *dentry_hashtable __read_mostly;
98
99 static inline struct hlist_bl_head *d_hash(unsigned int hash)
100 {
101         return dentry_hashtable + (hash >> d_hash_shift);
102 }
103
104 #define IN_LOOKUP_SHIFT 10
105 static struct hlist_bl_head in_lookup_hashtable[1 << IN_LOOKUP_SHIFT];
106
107 static inline struct hlist_bl_head *in_lookup_hash(const struct dentry *parent,
108                                         unsigned int hash)
109 {
110         hash += (unsigned long) parent / L1_CACHE_BYTES;
111         return in_lookup_hashtable + hash_32(hash, IN_LOOKUP_SHIFT);
112 }
113
114
115 /* Statistics gathering. */
116 struct dentry_stat_t dentry_stat = {
117         .age_limit = 45,
118 };
119
120 static DEFINE_PER_CPU(long, nr_dentry);
121 static DEFINE_PER_CPU(long, nr_dentry_unused);
122
123 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
124
125 /*
126  * Here we resort to our own counters instead of using generic per-cpu counters
127  * for consistency with what the vfs inode code does. We are expected to harvest
128  * better code and performance by having our own specialized counters.
129  *
130  * Please note that the loop is done over all possible CPUs, not over all online
131  * CPUs. The reason for this is that we don't want to play games with CPUs going
132  * on and off. If one of them goes off, we will just keep their counters.
133  *
134  * glommer: See cffbc8a for details, and if you ever intend to change this,
135  * please update all vfs counters to match.
136  */
137 static long get_nr_dentry(void)
138 {
139         int i;
140         long sum = 0;
141         for_each_possible_cpu(i)
142                 sum += per_cpu(nr_dentry, i);
143         return sum < 0 ? 0 : sum;
144 }
145
146 static long get_nr_dentry_unused(void)
147 {
148         int i;
149         long sum = 0;
150         for_each_possible_cpu(i)
151                 sum += per_cpu(nr_dentry_unused, i);
152         return sum < 0 ? 0 : sum;
153 }
154
155 int proc_nr_dentry(struct ctl_table *table, int write, void __user *buffer,
156                    size_t *lenp, loff_t *ppos)
157 {
158         dentry_stat.nr_dentry = get_nr_dentry();
159         dentry_stat.nr_unused = get_nr_dentry_unused();
160         return proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
161 }
162 #endif
163
164 /*
165  * Compare 2 name strings, return 0 if they match, otherwise non-zero.
166  * The strings are both count bytes long, and count is non-zero.
167  */
168 #ifdef CONFIG_DCACHE_WORD_ACCESS
169
170 #include <asm/word-at-a-time.h>
171 /*
172  * NOTE! 'cs' and 'scount' come from a dentry, so it has a
173  * aligned allocation for this particular component. We don't
174  * strictly need the load_unaligned_zeropad() safety, but it
175  * doesn't hurt either.
176  *
177  * In contrast, 'ct' and 'tcount' can be from a pathname, and do
178  * need the careful unaligned handling.
179  */
180 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
181 {
182         unsigned long a,b,mask;
183
184         for (;;) {
185                 a = read_word_at_a_time(cs);
186                 b = load_unaligned_zeropad(ct);
187                 if (tcount < sizeof(unsigned long))
188                         break;
189                 if (unlikely(a != b))
190                         return 1;
191                 cs += sizeof(unsigned long);
192                 ct += sizeof(unsigned long);
193                 tcount -= sizeof(unsigned long);
194                 if (!tcount)
195                         return 0;
196         }
197         mask = bytemask_from_count(tcount);
198         return unlikely(!!((a ^ b) & mask));
199 }
200
201 #else
202
203 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
204 {
205         do {
206                 if (*cs != *ct)
207                         return 1;
208                 cs++;
209                 ct++;
210                 tcount--;
211         } while (tcount);
212         return 0;
213 }
214
215 #endif
216
217 static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount)
218 {
219         /*
220          * Be careful about RCU walk racing with rename:
221          * use 'READ_ONCE' to fetch the name pointer.
222          *
223          * NOTE! Even if a rename will mean that the length
224          * was not loaded atomically, we don't care. The
225          * RCU walk will check the sequence count eventually,
226          * and catch it. And we won't overrun the buffer,
227          * because we're reading the name pointer atomically,
228          * and a dentry name is guaranteed to be properly
229          * terminated with a NUL byte.
230          *
231          * End result: even if 'len' is wrong, we'll exit
232          * early because the data cannot match (there can
233          * be no NUL in the ct/tcount data)
234          */
235         const unsigned char *cs = READ_ONCE(dentry->d_name.name);
236
237         return dentry_string_cmp(cs, ct, tcount);
238 }
239
240 struct external_name {
241         union {
242                 atomic_t count;
243                 struct rcu_head head;
244         } u;
245         unsigned char name[];
246 };
247
248 static inline struct external_name *external_name(struct dentry *dentry)
249 {
250         return container_of(dentry->d_name.name, struct external_name, name[0]);
251 }
252
253 static void __d_free(struct rcu_head *head)
254 {
255         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
256
257         kmem_cache_free(dentry_cache, dentry); 
258 }
259
260 static void __d_free_external_name(struct rcu_head *head)
261 {
262         struct external_name *name = container_of(head, struct external_name,
263                                                   u.head);
264
265         mod_node_page_state(page_pgdat(virt_to_page(name)),
266                             NR_INDIRECTLY_RECLAIMABLE_BYTES,
267                             -ksize(name));
268
269         kfree(name);
270 }
271
272 static void __d_free_external(struct rcu_head *head)
273 {
274         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
275
276         __d_free_external_name(&external_name(dentry)->u.head);
277
278         kmem_cache_free(dentry_cache, dentry);
279 }
280
281 static inline int dname_external(const struct dentry *dentry)
282 {
283         return dentry->d_name.name != dentry->d_iname;
284 }
285
286 void take_dentry_name_snapshot(struct name_snapshot *name, struct dentry *dentry)
287 {
288         spin_lock(&dentry->d_lock);
289         if (unlikely(dname_external(dentry))) {
290                 struct external_name *p = external_name(dentry);
291                 atomic_inc(&p->u.count);
292                 spin_unlock(&dentry->d_lock);
293                 name->name = p->name;
294         } else {
295                 memcpy(name->inline_name, dentry->d_iname,
296                        dentry->d_name.len + 1);
297                 spin_unlock(&dentry->d_lock);
298                 name->name = name->inline_name;
299         }
300 }
301 EXPORT_SYMBOL(take_dentry_name_snapshot);
302
303 void release_dentry_name_snapshot(struct name_snapshot *name)
304 {
305         if (unlikely(name->name != name->inline_name)) {
306                 struct external_name *p;
307                 p = container_of(name->name, struct external_name, name[0]);
308                 if (unlikely(atomic_dec_and_test(&p->u.count)))
309                         call_rcu(&p->u.head, __d_free_external_name);
310         }
311 }
312 EXPORT_SYMBOL(release_dentry_name_snapshot);
313
314 static inline void __d_set_inode_and_type(struct dentry *dentry,
315                                           struct inode *inode,
316                                           unsigned type_flags)
317 {
318         unsigned flags;
319
320         dentry->d_inode = inode;
321         flags = READ_ONCE(dentry->d_flags);
322         flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
323         flags |= type_flags;
324         WRITE_ONCE(dentry->d_flags, flags);
325 }
326
327 static inline void __d_clear_type_and_inode(struct dentry *dentry)
328 {
329         unsigned flags = READ_ONCE(dentry->d_flags);
330
331         flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
332         WRITE_ONCE(dentry->d_flags, flags);
333         dentry->d_inode = NULL;
334 }
335
336 static void dentry_free(struct dentry *dentry)
337 {
338         WARN_ON(!hlist_unhashed(&dentry->d_u.d_alias));
339         if (unlikely(dname_external(dentry))) {
340                 struct external_name *p = external_name(dentry);
341                 if (likely(atomic_dec_and_test(&p->u.count))) {
342                         call_rcu(&dentry->d_u.d_rcu, __d_free_external);
343                         return;
344                 }
345         }
346         /* if dentry was never visible to RCU, immediate free is OK */
347         if (dentry->d_flags & DCACHE_NORCU)
348                 __d_free(&dentry->d_u.d_rcu);
349         else
350                 call_rcu(&dentry->d_u.d_rcu, __d_free);
351 }
352
353 /*
354  * Release the dentry's inode, using the filesystem
355  * d_iput() operation if defined.
356  */
357 static void dentry_unlink_inode(struct dentry * dentry)
358         __releases(dentry->d_lock)
359         __releases(dentry->d_inode->i_lock)
360 {
361         struct inode *inode = dentry->d_inode;
362
363         raw_write_seqcount_begin(&dentry->d_seq);
364         __d_clear_type_and_inode(dentry);
365         hlist_del_init(&dentry->d_u.d_alias);
366         raw_write_seqcount_end(&dentry->d_seq);
367         spin_unlock(&dentry->d_lock);
368         spin_unlock(&inode->i_lock);
369         if (!inode->i_nlink)
370                 fsnotify_inoderemove(inode);
371         if (dentry->d_op && dentry->d_op->d_iput)
372                 dentry->d_op->d_iput(dentry, inode);
373         else
374                 iput(inode);
375 }
376
377 /*
378  * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
379  * is in use - which includes both the "real" per-superblock
380  * LRU list _and_ the DCACHE_SHRINK_LIST use.
381  *
382  * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
383  * on the shrink list (ie not on the superblock LRU list).
384  *
385  * The per-cpu "nr_dentry_unused" counters are updated with
386  * the DCACHE_LRU_LIST bit.
387  *
388  * These helper functions make sure we always follow the
389  * rules. d_lock must be held by the caller.
390  */
391 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
392 static void d_lru_add(struct dentry *dentry)
393 {
394         D_FLAG_VERIFY(dentry, 0);
395         dentry->d_flags |= DCACHE_LRU_LIST;
396         this_cpu_inc(nr_dentry_unused);
397         WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
398 }
399
400 static void d_lru_del(struct dentry *dentry)
401 {
402         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
403         dentry->d_flags &= ~DCACHE_LRU_LIST;
404         this_cpu_dec(nr_dentry_unused);
405         WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
406 }
407
408 static void d_shrink_del(struct dentry *dentry)
409 {
410         D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
411         list_del_init(&dentry->d_lru);
412         dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
413         this_cpu_dec(nr_dentry_unused);
414 }
415
416 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
417 {
418         D_FLAG_VERIFY(dentry, 0);
419         list_add(&dentry->d_lru, list);
420         dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
421         this_cpu_inc(nr_dentry_unused);
422 }
423
424 /*
425  * These can only be called under the global LRU lock, ie during the
426  * callback for freeing the LRU list. "isolate" removes it from the
427  * LRU lists entirely, while shrink_move moves it to the indicated
428  * private list.
429  */
430 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
431 {
432         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
433         dentry->d_flags &= ~DCACHE_LRU_LIST;
434         this_cpu_dec(nr_dentry_unused);
435         list_lru_isolate(lru, &dentry->d_lru);
436 }
437
438 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
439                               struct list_head *list)
440 {
441         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
442         dentry->d_flags |= DCACHE_SHRINK_LIST;
443         list_lru_isolate_move(lru, &dentry->d_lru, list);
444 }
445
446 /**
447  * d_drop - drop a dentry
448  * @dentry: dentry to drop
449  *
450  * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
451  * be found through a VFS lookup any more. Note that this is different from
452  * deleting the dentry - d_delete will try to mark the dentry negative if
453  * possible, giving a successful _negative_ lookup, while d_drop will
454  * just make the cache lookup fail.
455  *
456  * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
457  * reason (NFS timeouts or autofs deletes).
458  *
459  * __d_drop requires dentry->d_lock
460  * ___d_drop doesn't mark dentry as "unhashed"
461  *   (dentry->d_hash.pprev will be LIST_POISON2, not NULL).
462  */
463 static void ___d_drop(struct dentry *dentry)
464 {
465         struct hlist_bl_head *b;
466         /*
467          * Hashed dentries are normally on the dentry hashtable,
468          * with the exception of those newly allocated by
469          * d_obtain_root, which are always IS_ROOT:
470          */
471         if (unlikely(IS_ROOT(dentry)))
472                 b = &dentry->d_sb->s_roots;
473         else
474                 b = d_hash(dentry->d_name.hash);
475
476         hlist_bl_lock(b);
477         __hlist_bl_del(&dentry->d_hash);
478         hlist_bl_unlock(b);
479 }
480
481 void __d_drop(struct dentry *dentry)
482 {
483         if (!d_unhashed(dentry)) {
484                 ___d_drop(dentry);
485                 dentry->d_hash.pprev = NULL;
486                 write_seqcount_invalidate(&dentry->d_seq);
487         }
488 }
489 EXPORT_SYMBOL(__d_drop);
490
491 void d_drop(struct dentry *dentry)
492 {
493         spin_lock(&dentry->d_lock);
494         __d_drop(dentry);
495         spin_unlock(&dentry->d_lock);
496 }
497 EXPORT_SYMBOL(d_drop);
498
499 static inline void dentry_unlist(struct dentry *dentry, struct dentry *parent)
500 {
501         struct dentry *next;
502         /*
503          * Inform d_walk() and shrink_dentry_list() that we are no longer
504          * attached to the dentry tree
505          */
506         dentry->d_flags |= DCACHE_DENTRY_KILLED;
507         if (unlikely(list_empty(&dentry->d_child)))
508                 return;
509         __list_del_entry(&dentry->d_child);
510         /*
511          * Cursors can move around the list of children.  While we'd been
512          * a normal list member, it didn't matter - ->d_child.next would've
513          * been updated.  However, from now on it won't be and for the
514          * things like d_walk() it might end up with a nasty surprise.
515          * Normally d_walk() doesn't care about cursors moving around -
516          * ->d_lock on parent prevents that and since a cursor has no children
517          * of its own, we get through it without ever unlocking the parent.
518          * There is one exception, though - if we ascend from a child that
519          * gets killed as soon as we unlock it, the next sibling is found
520          * using the value left in its ->d_child.next.  And if _that_
521          * pointed to a cursor, and cursor got moved (e.g. by lseek())
522          * before d_walk() regains parent->d_lock, we'll end up skipping
523          * everything the cursor had been moved past.
524          *
525          * Solution: make sure that the pointer left behind in ->d_child.next
526          * points to something that won't be moving around.  I.e. skip the
527          * cursors.
528          */
529         while (dentry->d_child.next != &parent->d_subdirs) {
530                 next = list_entry(dentry->d_child.next, struct dentry, d_child);
531                 if (likely(!(next->d_flags & DCACHE_DENTRY_CURSOR)))
532                         break;
533                 dentry->d_child.next = next->d_child.next;
534         }
535 }
536
537 static void __dentry_kill(struct dentry *dentry)
538 {
539         struct dentry *parent = NULL;
540         bool can_free = true;
541         if (!IS_ROOT(dentry))
542                 parent = dentry->d_parent;
543
544         /*
545          * The dentry is now unrecoverably dead to the world.
546          */
547         lockref_mark_dead(&dentry->d_lockref);
548
549         /*
550          * inform the fs via d_prune that this dentry is about to be
551          * unhashed and destroyed.
552          */
553         if (dentry->d_flags & DCACHE_OP_PRUNE)
554                 dentry->d_op->d_prune(dentry);
555
556         if (dentry->d_flags & DCACHE_LRU_LIST) {
557                 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
558                         d_lru_del(dentry);
559         }
560         /* if it was on the hash then remove it */
561         __d_drop(dentry);
562         dentry_unlist(dentry, parent);
563         if (parent)
564                 spin_unlock(&parent->d_lock);
565         if (dentry->d_inode)
566                 dentry_unlink_inode(dentry);
567         else
568                 spin_unlock(&dentry->d_lock);
569         this_cpu_dec(nr_dentry);
570         if (dentry->d_op && dentry->d_op->d_release)
571                 dentry->d_op->d_release(dentry);
572
573         spin_lock(&dentry->d_lock);
574         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
575                 dentry->d_flags |= DCACHE_MAY_FREE;
576                 can_free = false;
577         }
578         spin_unlock(&dentry->d_lock);
579         if (likely(can_free))
580                 dentry_free(dentry);
581         cond_resched();
582 }
583
584 static struct dentry *__lock_parent(struct dentry *dentry)
585 {
586         struct dentry *parent;
587         rcu_read_lock();
588         spin_unlock(&dentry->d_lock);
589 again:
590         parent = READ_ONCE(dentry->d_parent);
591         spin_lock(&parent->d_lock);
592         /*
593          * We can't blindly lock dentry until we are sure
594          * that we won't violate the locking order.
595          * Any changes of dentry->d_parent must have
596          * been done with parent->d_lock held, so
597          * spin_lock() above is enough of a barrier
598          * for checking if it's still our child.
599          */
600         if (unlikely(parent != dentry->d_parent)) {
601                 spin_unlock(&parent->d_lock);
602                 goto again;
603         }
604         rcu_read_unlock();
605         if (parent != dentry)
606                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
607         else
608                 parent = NULL;
609         return parent;
610 }
611
612 static inline struct dentry *lock_parent(struct dentry *dentry)
613 {
614         struct dentry *parent = dentry->d_parent;
615         if (IS_ROOT(dentry))
616                 return NULL;
617         if (likely(spin_trylock(&parent->d_lock)))
618                 return parent;
619         return __lock_parent(dentry);
620 }
621
622 static inline bool retain_dentry(struct dentry *dentry)
623 {
624         WARN_ON(d_in_lookup(dentry));
625
626         /* Unreachable? Get rid of it */
627         if (unlikely(d_unhashed(dentry)))
628                 return false;
629
630         if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
631                 return false;
632
633         if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
634                 if (dentry->d_op->d_delete(dentry))
635                         return false;
636         }
637         /* retain; LRU fodder */
638         dentry->d_lockref.count--;
639         if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
640                 d_lru_add(dentry);
641         else if (unlikely(!(dentry->d_flags & DCACHE_REFERENCED)))
642                 dentry->d_flags |= DCACHE_REFERENCED;
643         return true;
644 }
645
646 /*
647  * Finish off a dentry we've decided to kill.
648  * dentry->d_lock must be held, returns with it unlocked.
649  * Returns dentry requiring refcount drop, or NULL if we're done.
650  */
651 static struct dentry *dentry_kill(struct dentry *dentry)
652         __releases(dentry->d_lock)
653 {
654         struct inode *inode = dentry->d_inode;
655         struct dentry *parent = NULL;
656
657         if (inode && unlikely(!spin_trylock(&inode->i_lock)))
658                 goto slow_positive;
659
660         if (!IS_ROOT(dentry)) {
661                 parent = dentry->d_parent;
662                 if (unlikely(!spin_trylock(&parent->d_lock))) {
663                         parent = __lock_parent(dentry);
664                         if (likely(inode || !dentry->d_inode))
665                                 goto got_locks;
666                         /* negative that became positive */
667                         if (parent)
668                                 spin_unlock(&parent->d_lock);
669                         inode = dentry->d_inode;
670                         goto slow_positive;
671                 }
672         }
673         __dentry_kill(dentry);
674         return parent;
675
676 slow_positive:
677         spin_unlock(&dentry->d_lock);
678         spin_lock(&inode->i_lock);
679         spin_lock(&dentry->d_lock);
680         parent = lock_parent(dentry);
681 got_locks:
682         if (unlikely(dentry->d_lockref.count != 1)) {
683                 dentry->d_lockref.count--;
684         } else if (likely(!retain_dentry(dentry))) {
685                 __dentry_kill(dentry);
686                 return parent;
687         }
688         /* we are keeping it, after all */
689         if (inode)
690                 spin_unlock(&inode->i_lock);
691         if (parent)
692                 spin_unlock(&parent->d_lock);
693         spin_unlock(&dentry->d_lock);
694         return NULL;
695 }
696
697 /*
698  * Try to do a lockless dput(), and return whether that was successful.
699  *
700  * If unsuccessful, we return false, having already taken the dentry lock.
701  *
702  * The caller needs to hold the RCU read lock, so that the dentry is
703  * guaranteed to stay around even if the refcount goes down to zero!
704  */
705 static inline bool fast_dput(struct dentry *dentry)
706 {
707         int ret;
708         unsigned int d_flags;
709
710         /*
711          * If we have a d_op->d_delete() operation, we sould not
712          * let the dentry count go to zero, so use "put_or_lock".
713          */
714         if (unlikely(dentry->d_flags & DCACHE_OP_DELETE))
715                 return lockref_put_or_lock(&dentry->d_lockref);
716
717         /*
718          * .. otherwise, we can try to just decrement the
719          * lockref optimistically.
720          */
721         ret = lockref_put_return(&dentry->d_lockref);
722
723         /*
724          * If the lockref_put_return() failed due to the lock being held
725          * by somebody else, the fast path has failed. We will need to
726          * get the lock, and then check the count again.
727          */
728         if (unlikely(ret < 0)) {
729                 spin_lock(&dentry->d_lock);
730                 if (WARN_ON_ONCE(dentry->d_lockref.count <= 0)) {
731                         spin_unlock(&dentry->d_lock);
732                         return true;
733                 }
734                 dentry->d_lockref.count--;
735                 goto locked;
736         }
737
738         /*
739          * If we weren't the last ref, we're done.
740          */
741         if (ret)
742                 return true;
743
744         /*
745          * Careful, careful. The reference count went down
746          * to zero, but we don't hold the dentry lock, so
747          * somebody else could get it again, and do another
748          * dput(), and we need to not race with that.
749          *
750          * However, there is a very special and common case
751          * where we don't care, because there is nothing to
752          * do: the dentry is still hashed, it does not have
753          * a 'delete' op, and it's referenced and already on
754          * the LRU list.
755          *
756          * NOTE! Since we aren't locked, these values are
757          * not "stable". However, it is sufficient that at
758          * some point after we dropped the reference the
759          * dentry was hashed and the flags had the proper
760          * value. Other dentry users may have re-gotten
761          * a reference to the dentry and change that, but
762          * our work is done - we can leave the dentry
763          * around with a zero refcount.
764          */
765         smp_rmb();
766         d_flags = READ_ONCE(dentry->d_flags);
767         d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST | DCACHE_DISCONNECTED;
768
769         /* Nothing to do? Dropping the reference was all we needed? */
770         if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry))
771                 return true;
772
773         /*
774          * Not the fast normal case? Get the lock. We've already decremented
775          * the refcount, but we'll need to re-check the situation after
776          * getting the lock.
777          */
778         spin_lock(&dentry->d_lock);
779
780         /*
781          * Did somebody else grab a reference to it in the meantime, and
782          * we're no longer the last user after all? Alternatively, somebody
783          * else could have killed it and marked it dead. Either way, we
784          * don't need to do anything else.
785          */
786 locked:
787         if (dentry->d_lockref.count) {
788                 spin_unlock(&dentry->d_lock);
789                 return true;
790         }
791
792         /*
793          * Re-get the reference we optimistically dropped. We hold the
794          * lock, and we just tested that it was zero, so we can just
795          * set it to 1.
796          */
797         dentry->d_lockref.count = 1;
798         return false;
799 }
800
801
802 /* 
803  * This is dput
804  *
805  * This is complicated by the fact that we do not want to put
806  * dentries that are no longer on any hash chain on the unused
807  * list: we'd much rather just get rid of them immediately.
808  *
809  * However, that implies that we have to traverse the dentry
810  * tree upwards to the parents which might _also_ now be
811  * scheduled for deletion (it may have been only waiting for
812  * its last child to go away).
813  *
814  * This tail recursion is done by hand as we don't want to depend
815  * on the compiler to always get this right (gcc generally doesn't).
816  * Real recursion would eat up our stack space.
817  */
818
819 /*
820  * dput - release a dentry
821  * @dentry: dentry to release 
822  *
823  * Release a dentry. This will drop the usage count and if appropriate
824  * call the dentry unlink method as well as removing it from the queues and
825  * releasing its resources. If the parent dentries were scheduled for release
826  * they too may now get deleted.
827  */
828 void dput(struct dentry *dentry)
829 {
830         while (dentry) {
831                 might_sleep();
832
833                 rcu_read_lock();
834                 if (likely(fast_dput(dentry))) {
835                         rcu_read_unlock();
836                         return;
837                 }
838
839                 /* Slow case: now with the dentry lock held */
840                 rcu_read_unlock();
841
842                 if (likely(retain_dentry(dentry))) {
843                         spin_unlock(&dentry->d_lock);
844                         return;
845                 }
846
847                 dentry = dentry_kill(dentry);
848         }
849 }
850 EXPORT_SYMBOL(dput);
851
852
853 /* This must be called with d_lock held */
854 static inline void __dget_dlock(struct dentry *dentry)
855 {
856         dentry->d_lockref.count++;
857 }
858
859 static inline void __dget(struct dentry *dentry)
860 {
861         lockref_get(&dentry->d_lockref);
862 }
863
864 struct dentry *dget_parent(struct dentry *dentry)
865 {
866         int gotref;
867         struct dentry *ret;
868         unsigned seq;
869
870         /*
871          * Do optimistic parent lookup without any
872          * locking.
873          */
874         rcu_read_lock();
875         seq = raw_seqcount_begin(&dentry->d_seq);
876         ret = READ_ONCE(dentry->d_parent);
877         gotref = lockref_get_not_zero(&ret->d_lockref);
878         rcu_read_unlock();
879         if (likely(gotref)) {
880                 if (!read_seqcount_retry(&dentry->d_seq, seq))
881                         return ret;
882                 dput(ret);
883         }
884
885 repeat:
886         /*
887          * Don't need rcu_dereference because we re-check it was correct under
888          * the lock.
889          */
890         rcu_read_lock();
891         ret = dentry->d_parent;
892         spin_lock(&ret->d_lock);
893         if (unlikely(ret != dentry->d_parent)) {
894                 spin_unlock(&ret->d_lock);
895                 rcu_read_unlock();
896                 goto repeat;
897         }
898         rcu_read_unlock();
899         BUG_ON(!ret->d_lockref.count);
900         ret->d_lockref.count++;
901         spin_unlock(&ret->d_lock);
902         return ret;
903 }
904 EXPORT_SYMBOL(dget_parent);
905
906 static struct dentry * __d_find_any_alias(struct inode *inode)
907 {
908         struct dentry *alias;
909
910         if (hlist_empty(&inode->i_dentry))
911                 return NULL;
912         alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
913         __dget(alias);
914         return alias;
915 }
916
917 /**
918  * d_find_any_alias - find any alias for a given inode
919  * @inode: inode to find an alias for
920  *
921  * If any aliases exist for the given inode, take and return a
922  * reference for one of them.  If no aliases exist, return %NULL.
923  */
924 struct dentry *d_find_any_alias(struct inode *inode)
925 {
926         struct dentry *de;
927
928         spin_lock(&inode->i_lock);
929         de = __d_find_any_alias(inode);
930         spin_unlock(&inode->i_lock);
931         return de;
932 }
933 EXPORT_SYMBOL(d_find_any_alias);
934
935 /**
936  * d_find_alias - grab a hashed alias of inode
937  * @inode: inode in question
938  *
939  * If inode has a hashed alias, or is a directory and has any alias,
940  * acquire the reference to alias and return it. Otherwise return NULL.
941  * Notice that if inode is a directory there can be only one alias and
942  * it can be unhashed only if it has no children, or if it is the root
943  * of a filesystem, or if the directory was renamed and d_revalidate
944  * was the first vfs operation to notice.
945  *
946  * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
947  * any other hashed alias over that one.
948  */
949 static struct dentry *__d_find_alias(struct inode *inode)
950 {
951         struct dentry *alias;
952
953         if (S_ISDIR(inode->i_mode))
954                 return __d_find_any_alias(inode);
955
956         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
957                 spin_lock(&alias->d_lock);
958                 if (!d_unhashed(alias)) {
959                         __dget_dlock(alias);
960                         spin_unlock(&alias->d_lock);
961                         return alias;
962                 }
963                 spin_unlock(&alias->d_lock);
964         }
965         return NULL;
966 }
967
968 struct dentry *d_find_alias(struct inode *inode)
969 {
970         struct dentry *de = NULL;
971
972         if (!hlist_empty(&inode->i_dentry)) {
973                 spin_lock(&inode->i_lock);
974                 de = __d_find_alias(inode);
975                 spin_unlock(&inode->i_lock);
976         }
977         return de;
978 }
979 EXPORT_SYMBOL(d_find_alias);
980
981 /*
982  *      Try to kill dentries associated with this inode.
983  * WARNING: you must own a reference to inode.
984  */
985 void d_prune_aliases(struct inode *inode)
986 {
987         struct dentry *dentry;
988 restart:
989         spin_lock(&inode->i_lock);
990         hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
991                 spin_lock(&dentry->d_lock);
992                 if (!dentry->d_lockref.count) {
993                         struct dentry *parent = lock_parent(dentry);
994                         if (likely(!dentry->d_lockref.count)) {
995                                 __dentry_kill(dentry);
996                                 dput(parent);
997                                 goto restart;
998                         }
999                         if (parent)
1000                                 spin_unlock(&parent->d_lock);
1001                 }
1002                 spin_unlock(&dentry->d_lock);
1003         }
1004         spin_unlock(&inode->i_lock);
1005 }
1006 EXPORT_SYMBOL(d_prune_aliases);
1007
1008 /*
1009  * Lock a dentry from shrink list.
1010  * Called under rcu_read_lock() and dentry->d_lock; the former
1011  * guarantees that nothing we access will be freed under us.
1012  * Note that dentry is *not* protected from concurrent dentry_kill(),
1013  * d_delete(), etc.
1014  *
1015  * Return false if dentry has been disrupted or grabbed, leaving
1016  * the caller to kick it off-list.  Otherwise, return true and have
1017  * that dentry's inode and parent both locked.
1018  */
1019 static bool shrink_lock_dentry(struct dentry *dentry)
1020 {
1021         struct inode *inode;
1022         struct dentry *parent;
1023
1024         if (dentry->d_lockref.count)
1025                 return false;
1026
1027         inode = dentry->d_inode;
1028         if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
1029                 spin_unlock(&dentry->d_lock);
1030                 spin_lock(&inode->i_lock);
1031                 spin_lock(&dentry->d_lock);
1032                 if (unlikely(dentry->d_lockref.count))
1033                         goto out;
1034                 /* changed inode means that somebody had grabbed it */
1035                 if (unlikely(inode != dentry->d_inode))
1036                         goto out;
1037         }
1038
1039         parent = dentry->d_parent;
1040         if (IS_ROOT(dentry) || likely(spin_trylock(&parent->d_lock)))
1041                 return true;
1042
1043         spin_unlock(&dentry->d_lock);
1044         spin_lock(&parent->d_lock);
1045         if (unlikely(parent != dentry->d_parent)) {
1046                 spin_unlock(&parent->d_lock);
1047                 spin_lock(&dentry->d_lock);
1048                 goto out;
1049         }
1050         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1051         if (likely(!dentry->d_lockref.count))
1052                 return true;
1053         spin_unlock(&parent->d_lock);
1054 out:
1055         if (inode)
1056                 spin_unlock(&inode->i_lock);
1057         return false;
1058 }
1059
1060 static void shrink_dentry_list(struct list_head *list)
1061 {
1062         while (!list_empty(list)) {
1063                 struct dentry *dentry, *parent;
1064
1065                 dentry = list_entry(list->prev, struct dentry, d_lru);
1066                 spin_lock(&dentry->d_lock);
1067                 rcu_read_lock();
1068                 if (!shrink_lock_dentry(dentry)) {
1069                         bool can_free = false;
1070                         rcu_read_unlock();
1071                         d_shrink_del(dentry);
1072                         if (dentry->d_lockref.count < 0)
1073                                 can_free = dentry->d_flags & DCACHE_MAY_FREE;
1074                         spin_unlock(&dentry->d_lock);
1075                         if (can_free)
1076                                 dentry_free(dentry);
1077                         continue;
1078                 }
1079                 rcu_read_unlock();
1080                 d_shrink_del(dentry);
1081                 parent = dentry->d_parent;
1082                 __dentry_kill(dentry);
1083                 if (parent == dentry)
1084                         continue;
1085                 /*
1086                  * We need to prune ancestors too. This is necessary to prevent
1087                  * quadratic behavior of shrink_dcache_parent(), but is also
1088                  * expected to be beneficial in reducing dentry cache
1089                  * fragmentation.
1090                  */
1091                 dentry = parent;
1092                 while (dentry && !lockref_put_or_lock(&dentry->d_lockref))
1093                         dentry = dentry_kill(dentry);
1094         }
1095 }
1096
1097 static enum lru_status dentry_lru_isolate(struct list_head *item,
1098                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1099 {
1100         struct list_head *freeable = arg;
1101         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1102
1103
1104         /*
1105          * we are inverting the lru lock/dentry->d_lock here,
1106          * so use a trylock. If we fail to get the lock, just skip
1107          * it
1108          */
1109         if (!spin_trylock(&dentry->d_lock))
1110                 return LRU_SKIP;
1111
1112         /*
1113          * Referenced dentries are still in use. If they have active
1114          * counts, just remove them from the LRU. Otherwise give them
1115          * another pass through the LRU.
1116          */
1117         if (dentry->d_lockref.count) {
1118                 d_lru_isolate(lru, dentry);
1119                 spin_unlock(&dentry->d_lock);
1120                 return LRU_REMOVED;
1121         }
1122
1123         if (dentry->d_flags & DCACHE_REFERENCED) {
1124                 dentry->d_flags &= ~DCACHE_REFERENCED;
1125                 spin_unlock(&dentry->d_lock);
1126
1127                 /*
1128                  * The list move itself will be made by the common LRU code. At
1129                  * this point, we've dropped the dentry->d_lock but keep the
1130                  * lru lock. This is safe to do, since every list movement is
1131                  * protected by the lru lock even if both locks are held.
1132                  *
1133                  * This is guaranteed by the fact that all LRU management
1134                  * functions are intermediated by the LRU API calls like
1135                  * list_lru_add and list_lru_del. List movement in this file
1136                  * only ever occur through this functions or through callbacks
1137                  * like this one, that are called from the LRU API.
1138                  *
1139                  * The only exceptions to this are functions like
1140                  * shrink_dentry_list, and code that first checks for the
1141                  * DCACHE_SHRINK_LIST flag.  Those are guaranteed to be
1142                  * operating only with stack provided lists after they are
1143                  * properly isolated from the main list.  It is thus, always a
1144                  * local access.
1145                  */
1146                 return LRU_ROTATE;
1147         }
1148
1149         d_lru_shrink_move(lru, dentry, freeable);
1150         spin_unlock(&dentry->d_lock);
1151
1152         return LRU_REMOVED;
1153 }
1154
1155 /**
1156  * prune_dcache_sb - shrink the dcache
1157  * @sb: superblock
1158  * @sc: shrink control, passed to list_lru_shrink_walk()
1159  *
1160  * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1161  * is done when we need more memory and called from the superblock shrinker
1162  * function.
1163  *
1164  * This function may fail to free any resources if all the dentries are in
1165  * use.
1166  */
1167 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1168 {
1169         LIST_HEAD(dispose);
1170         long freed;
1171
1172         freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1173                                      dentry_lru_isolate, &dispose);
1174         shrink_dentry_list(&dispose);
1175         return freed;
1176 }
1177
1178 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1179                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1180 {
1181         struct list_head *freeable = arg;
1182         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1183
1184         /*
1185          * we are inverting the lru lock/dentry->d_lock here,
1186          * so use a trylock. If we fail to get the lock, just skip
1187          * it
1188          */
1189         if (!spin_trylock(&dentry->d_lock))
1190                 return LRU_SKIP;
1191
1192         d_lru_shrink_move(lru, dentry, freeable);
1193         spin_unlock(&dentry->d_lock);
1194
1195         return LRU_REMOVED;
1196 }
1197
1198
1199 /**
1200  * shrink_dcache_sb - shrink dcache for a superblock
1201  * @sb: superblock
1202  *
1203  * Shrink the dcache for the specified super block. This is used to free
1204  * the dcache before unmounting a file system.
1205  */
1206 void shrink_dcache_sb(struct super_block *sb)
1207 {
1208         do {
1209                 LIST_HEAD(dispose);
1210
1211                 list_lru_walk(&sb->s_dentry_lru,
1212                         dentry_lru_isolate_shrink, &dispose, 1024);
1213                 shrink_dentry_list(&dispose);
1214         } while (list_lru_count(&sb->s_dentry_lru) > 0);
1215 }
1216 EXPORT_SYMBOL(shrink_dcache_sb);
1217
1218 /**
1219  * enum d_walk_ret - action to talke during tree walk
1220  * @D_WALK_CONTINUE:    contrinue walk
1221  * @D_WALK_QUIT:        quit walk
1222  * @D_WALK_NORETRY:     quit when retry is needed
1223  * @D_WALK_SKIP:        skip this dentry and its children
1224  */
1225 enum d_walk_ret {
1226         D_WALK_CONTINUE,
1227         D_WALK_QUIT,
1228         D_WALK_NORETRY,
1229         D_WALK_SKIP,
1230 };
1231
1232 /**
1233  * d_walk - walk the dentry tree
1234  * @parent:     start of walk
1235  * @data:       data passed to @enter() and @finish()
1236  * @enter:      callback when first entering the dentry
1237  *
1238  * The @enter() callbacks are called with d_lock held.
1239  */
1240 static void d_walk(struct dentry *parent, void *data,
1241                    enum d_walk_ret (*enter)(void *, struct dentry *))
1242 {
1243         struct dentry *this_parent;
1244         struct list_head *next;
1245         unsigned seq = 0;
1246         enum d_walk_ret ret;
1247         bool retry = true;
1248
1249 again:
1250         read_seqbegin_or_lock(&rename_lock, &seq);
1251         this_parent = parent;
1252         spin_lock(&this_parent->d_lock);
1253
1254         ret = enter(data, this_parent);
1255         switch (ret) {
1256         case D_WALK_CONTINUE:
1257                 break;
1258         case D_WALK_QUIT:
1259         case D_WALK_SKIP:
1260                 goto out_unlock;
1261         case D_WALK_NORETRY:
1262                 retry = false;
1263                 break;
1264         }
1265 repeat:
1266         next = this_parent->d_subdirs.next;
1267 resume:
1268         while (next != &this_parent->d_subdirs) {
1269                 struct list_head *tmp = next;
1270                 struct dentry *dentry = list_entry(tmp, struct dentry, d_child);
1271                 next = tmp->next;
1272
1273                 if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
1274                         continue;
1275
1276                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1277
1278                 ret = enter(data, dentry);
1279                 switch (ret) {
1280                 case D_WALK_CONTINUE:
1281                         break;
1282                 case D_WALK_QUIT:
1283                         spin_unlock(&dentry->d_lock);
1284                         goto out_unlock;
1285                 case D_WALK_NORETRY:
1286                         retry = false;
1287                         break;
1288                 case D_WALK_SKIP:
1289                         spin_unlock(&dentry->d_lock);
1290                         continue;
1291                 }
1292
1293                 if (!list_empty(&dentry->d_subdirs)) {
1294                         spin_unlock(&this_parent->d_lock);
1295                         spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1296                         this_parent = dentry;
1297                         spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1298                         goto repeat;
1299                 }
1300                 spin_unlock(&dentry->d_lock);
1301         }
1302         /*
1303          * All done at this level ... ascend and resume the search.
1304          */
1305         rcu_read_lock();
1306 ascend:
1307         if (this_parent != parent) {
1308                 struct dentry *child = this_parent;
1309                 this_parent = child->d_parent;
1310
1311                 spin_unlock(&child->d_lock);
1312                 spin_lock(&this_parent->d_lock);
1313
1314                 /* might go back up the wrong parent if we have had a rename. */
1315                 if (need_seqretry(&rename_lock, seq))
1316                         goto rename_retry;
1317                 /* go into the first sibling still alive */
1318                 do {
1319                         next = child->d_child.next;
1320                         if (next == &this_parent->d_subdirs)
1321                                 goto ascend;
1322                         child = list_entry(next, struct dentry, d_child);
1323                 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED));
1324                 rcu_read_unlock();
1325                 goto resume;
1326         }
1327         if (need_seqretry(&rename_lock, seq))
1328                 goto rename_retry;
1329         rcu_read_unlock();
1330
1331 out_unlock:
1332         spin_unlock(&this_parent->d_lock);
1333         done_seqretry(&rename_lock, seq);
1334         return;
1335
1336 rename_retry:
1337         spin_unlock(&this_parent->d_lock);
1338         rcu_read_unlock();
1339         BUG_ON(seq & 1);
1340         if (!retry)
1341                 return;
1342         seq = 1;
1343         goto again;
1344 }
1345
1346 struct check_mount {
1347         struct vfsmount *mnt;
1348         unsigned int mounted;
1349 };
1350
1351 static enum d_walk_ret path_check_mount(void *data, struct dentry *dentry)
1352 {
1353         struct check_mount *info = data;
1354         struct path path = { .mnt = info->mnt, .dentry = dentry };
1355
1356         if (likely(!d_mountpoint(dentry)))
1357                 return D_WALK_CONTINUE;
1358         if (__path_is_mountpoint(&path)) {
1359                 info->mounted = 1;
1360                 return D_WALK_QUIT;
1361         }
1362         return D_WALK_CONTINUE;
1363 }
1364
1365 /**
1366  * path_has_submounts - check for mounts over a dentry in the
1367  *                      current namespace.
1368  * @parent: path to check.
1369  *
1370  * Return true if the parent or its subdirectories contain
1371  * a mount point in the current namespace.
1372  */
1373 int path_has_submounts(const struct path *parent)
1374 {
1375         struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };
1376
1377         read_seqlock_excl(&mount_lock);
1378         d_walk(parent->dentry, &data, path_check_mount);
1379         read_sequnlock_excl(&mount_lock);
1380
1381         return data.mounted;
1382 }
1383 EXPORT_SYMBOL(path_has_submounts);
1384
1385 /*
1386  * Called by mount code to set a mountpoint and check if the mountpoint is
1387  * reachable (e.g. NFS can unhash a directory dentry and then the complete
1388  * subtree can become unreachable).
1389  *
1390  * Only one of d_invalidate() and d_set_mounted() must succeed.  For
1391  * this reason take rename_lock and d_lock on dentry and ancestors.
1392  */
1393 int d_set_mounted(struct dentry *dentry)
1394 {
1395         struct dentry *p;
1396         int ret = -ENOENT;
1397         write_seqlock(&rename_lock);
1398         for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1399                 /* Need exclusion wrt. d_invalidate() */
1400                 spin_lock(&p->d_lock);
1401                 if (unlikely(d_unhashed(p))) {
1402                         spin_unlock(&p->d_lock);
1403                         goto out;
1404                 }
1405                 spin_unlock(&p->d_lock);
1406         }
1407         spin_lock(&dentry->d_lock);
1408         if (!d_unlinked(dentry)) {
1409                 ret = -EBUSY;
1410                 if (!d_mountpoint(dentry)) {
1411                         dentry->d_flags |= DCACHE_MOUNTED;
1412                         ret = 0;
1413                 }
1414         }
1415         spin_unlock(&dentry->d_lock);
1416 out:
1417         write_sequnlock(&rename_lock);
1418         return ret;
1419 }
1420
1421 /*
1422  * Search the dentry child list of the specified parent,
1423  * and move any unused dentries to the end of the unused
1424  * list for prune_dcache(). We descend to the next level
1425  * whenever the d_subdirs list is non-empty and continue
1426  * searching.
1427  *
1428  * It returns zero iff there are no unused children,
1429  * otherwise  it returns the number of children moved to
1430  * the end of the unused list. This may not be the total
1431  * number of unused children, because select_parent can
1432  * drop the lock and return early due to latency
1433  * constraints.
1434  */
1435
1436 struct select_data {
1437         struct dentry *start;
1438         struct list_head dispose;
1439         int found;
1440 };
1441
1442 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1443 {
1444         struct select_data *data = _data;
1445         enum d_walk_ret ret = D_WALK_CONTINUE;
1446
1447         if (data->start == dentry)
1448                 goto out;
1449
1450         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1451                 data->found++;
1452         } else {
1453                 if (dentry->d_flags & DCACHE_LRU_LIST)
1454                         d_lru_del(dentry);
1455                 if (!dentry->d_lockref.count) {
1456                         d_shrink_add(dentry, &data->dispose);
1457                         data->found++;
1458                 }
1459         }
1460         /*
1461          * We can return to the caller if we have found some (this
1462          * ensures forward progress). We'll be coming back to find
1463          * the rest.
1464          */
1465         if (!list_empty(&data->dispose))
1466                 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1467 out:
1468         return ret;
1469 }
1470
1471 /**
1472  * shrink_dcache_parent - prune dcache
1473  * @parent: parent of entries to prune
1474  *
1475  * Prune the dcache to remove unused children of the parent dentry.
1476  */
1477 void shrink_dcache_parent(struct dentry *parent)
1478 {
1479         for (;;) {
1480                 struct select_data data;
1481
1482                 INIT_LIST_HEAD(&data.dispose);
1483                 data.start = parent;
1484                 data.found = 0;
1485
1486                 d_walk(parent, &data, select_collect);
1487
1488                 if (!list_empty(&data.dispose)) {
1489                         shrink_dentry_list(&data.dispose);
1490                         continue;
1491                 }
1492
1493                 cond_resched();
1494                 if (!data.found)
1495                         break;
1496         }
1497 }
1498 EXPORT_SYMBOL(shrink_dcache_parent);
1499
1500 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1501 {
1502         /* it has busy descendents; complain about those instead */
1503         if (!list_empty(&dentry->d_subdirs))
1504                 return D_WALK_CONTINUE;
1505
1506         /* root with refcount 1 is fine */
1507         if (dentry == _data && dentry->d_lockref.count == 1)
1508                 return D_WALK_CONTINUE;
1509
1510         printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} "
1511                         " still in use (%d) [unmount of %s %s]\n",
1512                        dentry,
1513                        dentry->d_inode ?
1514                        dentry->d_inode->i_ino : 0UL,
1515                        dentry,
1516                        dentry->d_lockref.count,
1517                        dentry->d_sb->s_type->name,
1518                        dentry->d_sb->s_id);
1519         WARN_ON(1);
1520         return D_WALK_CONTINUE;
1521 }
1522
1523 static void do_one_tree(struct dentry *dentry)
1524 {
1525         shrink_dcache_parent(dentry);
1526         d_walk(dentry, dentry, umount_check);
1527         d_drop(dentry);
1528         dput(dentry);
1529 }
1530
1531 /*
1532  * destroy the dentries attached to a superblock on unmounting
1533  */
1534 void shrink_dcache_for_umount(struct super_block *sb)
1535 {
1536         struct dentry *dentry;
1537
1538         WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
1539
1540         dentry = sb->s_root;
1541         sb->s_root = NULL;
1542         do_one_tree(dentry);
1543
1544         while (!hlist_bl_empty(&sb->s_roots)) {
1545                 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_roots), struct dentry, d_hash));
1546                 do_one_tree(dentry);
1547         }
1548 }
1549
1550 static enum d_walk_ret find_submount(void *_data, struct dentry *dentry)
1551 {
1552         struct dentry **victim = _data;
1553         if (d_mountpoint(dentry)) {
1554                 __dget_dlock(dentry);
1555                 *victim = dentry;
1556                 return D_WALK_QUIT;
1557         }
1558         return D_WALK_CONTINUE;
1559 }
1560
1561 /**
1562  * d_invalidate - detach submounts, prune dcache, and drop
1563  * @dentry: dentry to invalidate (aka detach, prune and drop)
1564  */
1565 void d_invalidate(struct dentry *dentry)
1566 {
1567         bool had_submounts = false;
1568         spin_lock(&dentry->d_lock);
1569         if (d_unhashed(dentry)) {
1570                 spin_unlock(&dentry->d_lock);
1571                 return;
1572         }
1573         __d_drop(dentry);
1574         spin_unlock(&dentry->d_lock);
1575
1576         /* Negative dentries can be dropped without further checks */
1577         if (!dentry->d_inode)
1578                 return;
1579
1580         shrink_dcache_parent(dentry);
1581         for (;;) {
1582                 struct dentry *victim = NULL;
1583                 d_walk(dentry, &victim, find_submount);
1584                 if (!victim) {
1585                         if (had_submounts)
1586                                 shrink_dcache_parent(dentry);
1587                         return;
1588                 }
1589                 had_submounts = true;
1590                 detach_mounts(victim);
1591                 dput(victim);
1592         }
1593 }
1594 EXPORT_SYMBOL(d_invalidate);
1595
1596 /**
1597  * __d_alloc    -       allocate a dcache entry
1598  * @sb: filesystem it will belong to
1599  * @name: qstr of the name
1600  *
1601  * Allocates a dentry. It returns %NULL if there is insufficient memory
1602  * available. On a success the dentry is returned. The name passed in is
1603  * copied and the copy passed in may be reused after this call.
1604  */
1605  
1606 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1607 {
1608         struct external_name *ext = NULL;
1609         struct dentry *dentry;
1610         char *dname;
1611         int err;
1612
1613         dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1614         if (!dentry)
1615                 return NULL;
1616
1617         /*
1618          * We guarantee that the inline name is always NUL-terminated.
1619          * This way the memcpy() done by the name switching in rename
1620          * will still always have a NUL at the end, even if we might
1621          * be overwriting an internal NUL character
1622          */
1623         dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1624         if (unlikely(!name)) {
1625                 name = &slash_name;
1626                 dname = dentry->d_iname;
1627         } else if (name->len > DNAME_INLINE_LEN-1) {
1628                 size_t size = offsetof(struct external_name, name[1]);
1629
1630                 ext = kmalloc(size + name->len, GFP_KERNEL_ACCOUNT);
1631                 if (!ext) {
1632                         kmem_cache_free(dentry_cache, dentry); 
1633                         return NULL;
1634                 }
1635                 atomic_set(&ext->u.count, 1);
1636                 dname = ext->name;
1637         } else  {
1638                 dname = dentry->d_iname;
1639         }       
1640
1641         dentry->d_name.len = name->len;
1642         dentry->d_name.hash = name->hash;
1643         memcpy(dname, name->name, name->len);
1644         dname[name->len] = 0;
1645
1646         /* Make sure we always see the terminating NUL character */
1647         smp_store_release(&dentry->d_name.name, dname); /* ^^^ */
1648
1649         dentry->d_lockref.count = 1;
1650         dentry->d_flags = 0;
1651         spin_lock_init(&dentry->d_lock);
1652         seqcount_init(&dentry->d_seq);
1653         dentry->d_inode = NULL;
1654         dentry->d_parent = dentry;
1655         dentry->d_sb = sb;
1656         dentry->d_op = NULL;
1657         dentry->d_fsdata = NULL;
1658         INIT_HLIST_BL_NODE(&dentry->d_hash);
1659         INIT_LIST_HEAD(&dentry->d_lru);
1660         INIT_LIST_HEAD(&dentry->d_subdirs);
1661         INIT_HLIST_NODE(&dentry->d_u.d_alias);
1662         INIT_LIST_HEAD(&dentry->d_child);
1663         d_set_d_op(dentry, dentry->d_sb->s_d_op);
1664
1665         if (dentry->d_op && dentry->d_op->d_init) {
1666                 err = dentry->d_op->d_init(dentry);
1667                 if (err) {
1668                         if (dname_external(dentry))
1669                                 kfree(external_name(dentry));
1670                         kmem_cache_free(dentry_cache, dentry);
1671                         return NULL;
1672                 }
1673         }
1674
1675         if (unlikely(ext)) {
1676                 pg_data_t *pgdat = page_pgdat(virt_to_page(ext));
1677                 mod_node_page_state(pgdat, NR_INDIRECTLY_RECLAIMABLE_BYTES,
1678                                     ksize(ext));
1679         }
1680
1681         this_cpu_inc(nr_dentry);
1682
1683         return dentry;
1684 }
1685
1686 /**
1687  * d_alloc      -       allocate a dcache entry
1688  * @parent: parent of entry to allocate
1689  * @name: qstr of the name
1690  *
1691  * Allocates a dentry. It returns %NULL if there is insufficient memory
1692  * available. On a success the dentry is returned. The name passed in is
1693  * copied and the copy passed in may be reused after this call.
1694  */
1695 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1696 {
1697         struct dentry *dentry = __d_alloc(parent->d_sb, name);
1698         if (!dentry)
1699                 return NULL;
1700         spin_lock(&parent->d_lock);
1701         /*
1702          * don't need child lock because it is not subject
1703          * to concurrency here
1704          */
1705         __dget_dlock(parent);
1706         dentry->d_parent = parent;
1707         list_add(&dentry->d_child, &parent->d_subdirs);
1708         spin_unlock(&parent->d_lock);
1709
1710         return dentry;
1711 }
1712 EXPORT_SYMBOL(d_alloc);
1713
1714 struct dentry *d_alloc_anon(struct super_block *sb)
1715 {
1716         return __d_alloc(sb, NULL);
1717 }
1718 EXPORT_SYMBOL(d_alloc_anon);
1719
1720 struct dentry *d_alloc_cursor(struct dentry * parent)
1721 {
1722         struct dentry *dentry = d_alloc_anon(parent->d_sb);
1723         if (dentry) {
1724                 dentry->d_flags |= DCACHE_DENTRY_CURSOR;
1725                 dentry->d_parent = dget(parent);
1726         }
1727         return dentry;
1728 }
1729
1730 /**
1731  * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1732  * @sb: the superblock
1733  * @name: qstr of the name
1734  *
1735  * For a filesystem that just pins its dentries in memory and never
1736  * performs lookups at all, return an unhashed IS_ROOT dentry.
1737  * This is used for pipes, sockets et.al. - the stuff that should
1738  * never be anyone's children or parents.  Unlike all other
1739  * dentries, these will not have RCU delay between dropping the
1740  * last reference and freeing them.
1741  */
1742 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1743 {
1744         struct dentry *dentry = __d_alloc(sb, name);
1745         if (likely(dentry))
1746                 dentry->d_flags |= DCACHE_NORCU;
1747         return dentry;
1748 }
1749 EXPORT_SYMBOL(d_alloc_pseudo);
1750
1751 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1752 {
1753         struct qstr q;
1754
1755         q.name = name;
1756         q.hash_len = hashlen_string(parent, name);
1757         return d_alloc(parent, &q);
1758 }
1759 EXPORT_SYMBOL(d_alloc_name);
1760
1761 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1762 {
1763         WARN_ON_ONCE(dentry->d_op);
1764         WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH  |
1765                                 DCACHE_OP_COMPARE       |
1766                                 DCACHE_OP_REVALIDATE    |
1767                                 DCACHE_OP_WEAK_REVALIDATE       |
1768                                 DCACHE_OP_DELETE        |
1769                                 DCACHE_OP_REAL));
1770         dentry->d_op = op;
1771         if (!op)
1772                 return;
1773         if (op->d_hash)
1774                 dentry->d_flags |= DCACHE_OP_HASH;
1775         if (op->d_compare)
1776                 dentry->d_flags |= DCACHE_OP_COMPARE;
1777         if (op->d_revalidate)
1778                 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1779         if (op->d_weak_revalidate)
1780                 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1781         if (op->d_delete)
1782                 dentry->d_flags |= DCACHE_OP_DELETE;
1783         if (op->d_prune)
1784                 dentry->d_flags |= DCACHE_OP_PRUNE;
1785         if (op->d_real)
1786                 dentry->d_flags |= DCACHE_OP_REAL;
1787
1788 }
1789 EXPORT_SYMBOL(d_set_d_op);
1790
1791
1792 /*
1793  * d_set_fallthru - Mark a dentry as falling through to a lower layer
1794  * @dentry - The dentry to mark
1795  *
1796  * Mark a dentry as falling through to the lower layer (as set with
1797  * d_pin_lower()).  This flag may be recorded on the medium.
1798  */
1799 void d_set_fallthru(struct dentry *dentry)
1800 {
1801         spin_lock(&dentry->d_lock);
1802         dentry->d_flags |= DCACHE_FALLTHRU;
1803         spin_unlock(&dentry->d_lock);
1804 }
1805 EXPORT_SYMBOL(d_set_fallthru);
1806
1807 static unsigned d_flags_for_inode(struct inode *inode)
1808 {
1809         unsigned add_flags = DCACHE_REGULAR_TYPE;
1810
1811         if (!inode)
1812                 return DCACHE_MISS_TYPE;
1813
1814         if (S_ISDIR(inode->i_mode)) {
1815                 add_flags = DCACHE_DIRECTORY_TYPE;
1816                 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1817                         if (unlikely(!inode->i_op->lookup))
1818                                 add_flags = DCACHE_AUTODIR_TYPE;
1819                         else
1820                                 inode->i_opflags |= IOP_LOOKUP;
1821                 }
1822                 goto type_determined;
1823         }
1824
1825         if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1826                 if (unlikely(inode->i_op->get_link)) {
1827                         add_flags = DCACHE_SYMLINK_TYPE;
1828                         goto type_determined;
1829                 }
1830                 inode->i_opflags |= IOP_NOFOLLOW;
1831         }
1832
1833         if (unlikely(!S_ISREG(inode->i_mode)))
1834                 add_flags = DCACHE_SPECIAL_TYPE;
1835
1836 type_determined:
1837         if (unlikely(IS_AUTOMOUNT(inode)))
1838                 add_flags |= DCACHE_NEED_AUTOMOUNT;
1839         return add_flags;
1840 }
1841
1842 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1843 {
1844         unsigned add_flags = d_flags_for_inode(inode);
1845         WARN_ON(d_in_lookup(dentry));
1846
1847         spin_lock(&dentry->d_lock);
1848         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1849         raw_write_seqcount_begin(&dentry->d_seq);
1850         __d_set_inode_and_type(dentry, inode, add_flags);
1851         raw_write_seqcount_end(&dentry->d_seq);
1852         fsnotify_update_flags(dentry);
1853         spin_unlock(&dentry->d_lock);
1854 }
1855
1856 /**
1857  * d_instantiate - fill in inode information for a dentry
1858  * @entry: dentry to complete
1859  * @inode: inode to attach to this dentry
1860  *
1861  * Fill in inode information in the entry.
1862  *
1863  * This turns negative dentries into productive full members
1864  * of society.
1865  *
1866  * NOTE! This assumes that the inode count has been incremented
1867  * (or otherwise set) by the caller to indicate that it is now
1868  * in use by the dcache.
1869  */
1870  
1871 void d_instantiate(struct dentry *entry, struct inode * inode)
1872 {
1873         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1874         if (inode) {
1875                 security_d_instantiate(entry, inode);
1876                 spin_lock(&inode->i_lock);
1877                 __d_instantiate(entry, inode);
1878                 spin_unlock(&inode->i_lock);
1879         }
1880 }
1881 EXPORT_SYMBOL(d_instantiate);
1882
1883 /*
1884  * This should be equivalent to d_instantiate() + unlock_new_inode(),
1885  * with lockdep-related part of unlock_new_inode() done before
1886  * anything else.  Use that instead of open-coding d_instantiate()/
1887  * unlock_new_inode() combinations.
1888  */
1889 void d_instantiate_new(struct dentry *entry, struct inode *inode)
1890 {
1891         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1892         BUG_ON(!inode);
1893         lockdep_annotate_inode_mutex_key(inode);
1894         security_d_instantiate(entry, inode);
1895         spin_lock(&inode->i_lock);
1896         __d_instantiate(entry, inode);
1897         WARN_ON(!(inode->i_state & I_NEW));
1898         inode->i_state &= ~I_NEW & ~I_CREATING;
1899         smp_mb();
1900         wake_up_bit(&inode->i_state, __I_NEW);
1901         spin_unlock(&inode->i_lock);
1902 }
1903 EXPORT_SYMBOL(d_instantiate_new);
1904
1905 struct dentry *d_make_root(struct inode *root_inode)
1906 {
1907         struct dentry *res = NULL;
1908
1909         if (root_inode) {
1910                 res = d_alloc_anon(root_inode->i_sb);
1911                 if (res)
1912                         d_instantiate(res, root_inode);
1913                 else
1914                         iput(root_inode);
1915         }
1916         return res;
1917 }
1918 EXPORT_SYMBOL(d_make_root);
1919
1920 static struct dentry *__d_instantiate_anon(struct dentry *dentry,
1921                                            struct inode *inode,
1922                                            bool disconnected)
1923 {
1924         struct dentry *res;
1925         unsigned add_flags;
1926
1927         security_d_instantiate(dentry, inode);
1928         spin_lock(&inode->i_lock);
1929         res = __d_find_any_alias(inode);
1930         if (res) {
1931                 spin_unlock(&inode->i_lock);
1932                 dput(dentry);
1933                 goto out_iput;
1934         }
1935
1936         /* attach a disconnected dentry */
1937         add_flags = d_flags_for_inode(inode);
1938
1939         if (disconnected)
1940                 add_flags |= DCACHE_DISCONNECTED;
1941
1942         spin_lock(&dentry->d_lock);
1943         __d_set_inode_and_type(dentry, inode, add_flags);
1944         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1945         if (!disconnected) {
1946                 hlist_bl_lock(&dentry->d_sb->s_roots);
1947                 hlist_bl_add_head(&dentry->d_hash, &dentry->d_sb->s_roots);
1948                 hlist_bl_unlock(&dentry->d_sb->s_roots);
1949         }
1950         spin_unlock(&dentry->d_lock);
1951         spin_unlock(&inode->i_lock);
1952
1953         return dentry;
1954
1955  out_iput:
1956         iput(inode);
1957         return res;
1958 }
1959
1960 struct dentry *d_instantiate_anon(struct dentry *dentry, struct inode *inode)
1961 {
1962         return __d_instantiate_anon(dentry, inode, true);
1963 }
1964 EXPORT_SYMBOL(d_instantiate_anon);
1965
1966 static struct dentry *__d_obtain_alias(struct inode *inode, bool disconnected)
1967 {
1968         struct dentry *tmp;
1969         struct dentry *res;
1970
1971         if (!inode)
1972                 return ERR_PTR(-ESTALE);
1973         if (IS_ERR(inode))
1974                 return ERR_CAST(inode);
1975
1976         res = d_find_any_alias(inode);
1977         if (res)
1978                 goto out_iput;
1979
1980         tmp = d_alloc_anon(inode->i_sb);
1981         if (!tmp) {
1982                 res = ERR_PTR(-ENOMEM);
1983                 goto out_iput;
1984         }
1985
1986         return __d_instantiate_anon(tmp, inode, disconnected);
1987
1988 out_iput:
1989         iput(inode);
1990         return res;
1991 }
1992
1993 /**
1994  * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
1995  * @inode: inode to allocate the dentry for
1996  *
1997  * Obtain a dentry for an inode resulting from NFS filehandle conversion or
1998  * similar open by handle operations.  The returned dentry may be anonymous,
1999  * or may have a full name (if the inode was already in the cache).
2000  *
2001  * When called on a directory inode, we must ensure that the inode only ever
2002  * has one dentry.  If a dentry is found, that is returned instead of
2003  * allocating a new one.
2004  *
2005  * On successful return, the reference to the inode has been transferred
2006  * to the dentry.  In case of an error the reference on the inode is released.
2007  * To make it easier to use in export operations a %NULL or IS_ERR inode may
2008  * be passed in and the error will be propagated to the return value,
2009  * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
2010  */
2011 struct dentry *d_obtain_alias(struct inode *inode)
2012 {
2013         return __d_obtain_alias(inode, true);
2014 }
2015 EXPORT_SYMBOL(d_obtain_alias);
2016
2017 /**
2018  * d_obtain_root - find or allocate a dentry for a given inode
2019  * @inode: inode to allocate the dentry for
2020  *
2021  * Obtain an IS_ROOT dentry for the root of a filesystem.
2022  *
2023  * We must ensure that directory inodes only ever have one dentry.  If a
2024  * dentry is found, that is returned instead of allocating a new one.
2025  *
2026  * On successful return, the reference to the inode has been transferred
2027  * to the dentry.  In case of an error the reference on the inode is
2028  * released.  A %NULL or IS_ERR inode may be passed in and will be the
2029  * error will be propagate to the return value, with a %NULL @inode
2030  * replaced by ERR_PTR(-ESTALE).
2031  */
2032 struct dentry *d_obtain_root(struct inode *inode)
2033 {
2034         return __d_obtain_alias(inode, false);
2035 }
2036 EXPORT_SYMBOL(d_obtain_root);
2037
2038 /**
2039  * d_add_ci - lookup or allocate new dentry with case-exact name
2040  * @inode:  the inode case-insensitive lookup has found
2041  * @dentry: the negative dentry that was passed to the parent's lookup func
2042  * @name:   the case-exact name to be associated with the returned dentry
2043  *
2044  * This is to avoid filling the dcache with case-insensitive names to the
2045  * same inode, only the actual correct case is stored in the dcache for
2046  * case-insensitive filesystems.
2047  *
2048  * For a case-insensitive lookup match and if the the case-exact dentry
2049  * already exists in in the dcache, use it and return it.
2050  *
2051  * If no entry exists with the exact case name, allocate new dentry with
2052  * the exact case, and return the spliced entry.
2053  */
2054 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2055                         struct qstr *name)
2056 {
2057         struct dentry *found, *res;
2058
2059         /*
2060          * First check if a dentry matching the name already exists,
2061          * if not go ahead and create it now.
2062          */
2063         found = d_hash_and_lookup(dentry->d_parent, name);
2064         if (found) {
2065                 iput(inode);
2066                 return found;
2067         }
2068         if (d_in_lookup(dentry)) {
2069                 found = d_alloc_parallel(dentry->d_parent, name,
2070                                         dentry->d_wait);
2071                 if (IS_ERR(found) || !d_in_lookup(found)) {
2072                         iput(inode);
2073                         return found;
2074                 }
2075         } else {
2076                 found = d_alloc(dentry->d_parent, name);
2077                 if (!found) {
2078                         iput(inode);
2079                         return ERR_PTR(-ENOMEM);
2080                 } 
2081         }
2082         res = d_splice_alias(inode, found);
2083         if (res) {
2084                 dput(found);
2085                 return res;
2086         }
2087         return found;
2088 }
2089 EXPORT_SYMBOL(d_add_ci);
2090
2091
2092 static inline bool d_same_name(const struct dentry *dentry,
2093                                 const struct dentry *parent,
2094                                 const struct qstr *name)
2095 {
2096         if (likely(!(parent->d_flags & DCACHE_OP_COMPARE))) {
2097                 if (dentry->d_name.len != name->len)
2098                         return false;
2099                 return dentry_cmp(dentry, name->name, name->len) == 0;
2100         }
2101         return parent->d_op->d_compare(dentry,
2102                                        dentry->d_name.len, dentry->d_name.name,
2103                                        name) == 0;
2104 }
2105
2106 /**
2107  * __d_lookup_rcu - search for a dentry (racy, store-free)
2108  * @parent: parent dentry
2109  * @name: qstr of name we wish to find
2110  * @seqp: returns d_seq value at the point where the dentry was found
2111  * Returns: dentry, or NULL
2112  *
2113  * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2114  * resolution (store-free path walking) design described in
2115  * Documentation/filesystems/path-lookup.txt.
2116  *
2117  * This is not to be used outside core vfs.
2118  *
2119  * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2120  * held, and rcu_read_lock held. The returned dentry must not be stored into
2121  * without taking d_lock and checking d_seq sequence count against @seq
2122  * returned here.
2123  *
2124  * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2125  * function.
2126  *
2127  * Alternatively, __d_lookup_rcu may be called again to look up the child of
2128  * the returned dentry, so long as its parent's seqlock is checked after the
2129  * child is looked up. Thus, an interlocking stepping of sequence lock checks
2130  * is formed, giving integrity down the path walk.
2131  *
2132  * NOTE! The caller *has* to check the resulting dentry against the sequence
2133  * number we've returned before using any of the resulting dentry state!
2134  */
2135 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2136                                 const struct qstr *name,
2137                                 unsigned *seqp)
2138 {
2139         u64 hashlen = name->hash_len;
2140         const unsigned char *str = name->name;
2141         struct hlist_bl_head *b = d_hash(hashlen_hash(hashlen));
2142         struct hlist_bl_node *node;
2143         struct dentry *dentry;
2144
2145         /*
2146          * Note: There is significant duplication with __d_lookup_rcu which is
2147          * required to prevent single threaded performance regressions
2148          * especially on architectures where smp_rmb (in seqcounts) are costly.
2149          * Keep the two functions in sync.
2150          */
2151
2152         /*
2153          * The hash list is protected using RCU.
2154          *
2155          * Carefully use d_seq when comparing a candidate dentry, to avoid
2156          * races with d_move().
2157          *
2158          * It is possible that concurrent renames can mess up our list
2159          * walk here and result in missing our dentry, resulting in the
2160          * false-negative result. d_lookup() protects against concurrent
2161          * renames using rename_lock seqlock.
2162          *
2163          * See Documentation/filesystems/path-lookup.txt for more details.
2164          */
2165         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2166                 unsigned seq;
2167
2168 seqretry:
2169                 /*
2170                  * The dentry sequence count protects us from concurrent
2171                  * renames, and thus protects parent and name fields.
2172                  *
2173                  * The caller must perform a seqcount check in order
2174                  * to do anything useful with the returned dentry.
2175                  *
2176                  * NOTE! We do a "raw" seqcount_begin here. That means that
2177                  * we don't wait for the sequence count to stabilize if it
2178                  * is in the middle of a sequence change. If we do the slow
2179                  * dentry compare, we will do seqretries until it is stable,
2180                  * and if we end up with a successful lookup, we actually
2181                  * want to exit RCU lookup anyway.
2182                  *
2183                  * Note that raw_seqcount_begin still *does* smp_rmb(), so
2184                  * we are still guaranteed NUL-termination of ->d_name.name.
2185                  */
2186                 seq = raw_seqcount_begin(&dentry->d_seq);
2187                 if (dentry->d_parent != parent)
2188                         continue;
2189                 if (d_unhashed(dentry))
2190                         continue;
2191
2192                 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) {
2193                         int tlen;
2194                         const char *tname;
2195                         if (dentry->d_name.hash != hashlen_hash(hashlen))
2196                                 continue;
2197                         tlen = dentry->d_name.len;
2198                         tname = dentry->d_name.name;
2199                         /* we want a consistent (name,len) pair */
2200                         if (read_seqcount_retry(&dentry->d_seq, seq)) {
2201                                 cpu_relax();
2202                                 goto seqretry;
2203                         }
2204                         if (parent->d_op->d_compare(dentry,
2205                                                     tlen, tname, name) != 0)
2206                                 continue;
2207                 } else {
2208                         if (dentry->d_name.hash_len != hashlen)
2209                                 continue;
2210                         if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
2211                                 continue;
2212                 }
2213                 *seqp = seq;
2214                 return dentry;
2215         }
2216         return NULL;
2217 }
2218
2219 /**
2220  * d_lookup - search for a dentry
2221  * @parent: parent dentry
2222  * @name: qstr of name we wish to find
2223  * Returns: dentry, or NULL
2224  *
2225  * d_lookup searches the children of the parent dentry for the name in
2226  * question. If the dentry is found its reference count is incremented and the
2227  * dentry is returned. The caller must use dput to free the entry when it has
2228  * finished using it. %NULL is returned if the dentry does not exist.
2229  */
2230 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2231 {
2232         struct dentry *dentry;
2233         unsigned seq;
2234
2235         do {
2236                 seq = read_seqbegin(&rename_lock);
2237                 dentry = __d_lookup(parent, name);
2238                 if (dentry)
2239                         break;
2240         } while (read_seqretry(&rename_lock, seq));
2241         return dentry;
2242 }
2243 EXPORT_SYMBOL(d_lookup);
2244
2245 /**
2246  * __d_lookup - search for a dentry (racy)
2247  * @parent: parent dentry
2248  * @name: qstr of name we wish to find
2249  * Returns: dentry, or NULL
2250  *
2251  * __d_lookup is like d_lookup, however it may (rarely) return a
2252  * false-negative result due to unrelated rename activity.
2253  *
2254  * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2255  * however it must be used carefully, eg. with a following d_lookup in
2256  * the case of failure.
2257  *
2258  * __d_lookup callers must be commented.
2259  */
2260 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2261 {
2262         unsigned int hash = name->hash;
2263         struct hlist_bl_head *b = d_hash(hash);
2264         struct hlist_bl_node *node;
2265         struct dentry *found = NULL;
2266         struct dentry *dentry;
2267
2268         /*
2269          * Note: There is significant duplication with __d_lookup_rcu which is
2270          * required to prevent single threaded performance regressions
2271          * especially on architectures where smp_rmb (in seqcounts) are costly.
2272          * Keep the two functions in sync.
2273          */
2274
2275         /*
2276          * The hash list is protected using RCU.
2277          *
2278          * Take d_lock when comparing a candidate dentry, to avoid races
2279          * with d_move().
2280          *
2281          * It is possible that concurrent renames can mess up our list
2282          * walk here and result in missing our dentry, resulting in the
2283          * false-negative result. d_lookup() protects against concurrent
2284          * renames using rename_lock seqlock.
2285          *
2286          * See Documentation/filesystems/path-lookup.txt for more details.
2287          */
2288         rcu_read_lock();
2289         
2290         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2291
2292                 if (dentry->d_name.hash != hash)
2293                         continue;
2294
2295                 spin_lock(&dentry->d_lock);
2296                 if (dentry->d_parent != parent)
2297                         goto next;
2298                 if (d_unhashed(dentry))
2299                         goto next;
2300
2301                 if (!d_same_name(dentry, parent, name))
2302                         goto next;
2303
2304                 dentry->d_lockref.count++;
2305                 found = dentry;
2306                 spin_unlock(&dentry->d_lock);
2307                 break;
2308 next:
2309                 spin_unlock(&dentry->d_lock);
2310         }
2311         rcu_read_unlock();
2312
2313         return found;
2314 }
2315
2316 /**
2317  * d_hash_and_lookup - hash the qstr then search for a dentry
2318  * @dir: Directory to search in
2319  * @name: qstr of name we wish to find
2320  *
2321  * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2322  */
2323 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2324 {
2325         /*
2326          * Check for a fs-specific hash function. Note that we must
2327          * calculate the standard hash first, as the d_op->d_hash()
2328          * routine may choose to leave the hash value unchanged.
2329          */
2330         name->hash = full_name_hash(dir, name->name, name->len);
2331         if (dir->d_flags & DCACHE_OP_HASH) {
2332                 int err = dir->d_op->d_hash(dir, name);
2333                 if (unlikely(err < 0))
2334                         return ERR_PTR(err);
2335         }
2336         return d_lookup(dir, name);
2337 }
2338 EXPORT_SYMBOL(d_hash_and_lookup);
2339
2340 /*
2341  * When a file is deleted, we have two options:
2342  * - turn this dentry into a negative dentry
2343  * - unhash this dentry and free it.
2344  *
2345  * Usually, we want to just turn this into
2346  * a negative dentry, but if anybody else is
2347  * currently using the dentry or the inode
2348  * we can't do that and we fall back on removing
2349  * it from the hash queues and waiting for
2350  * it to be deleted later when it has no users
2351  */
2352  
2353 /**
2354  * d_delete - delete a dentry
2355  * @dentry: The dentry to delete
2356  *
2357  * Turn the dentry into a negative dentry if possible, otherwise
2358  * remove it from the hash queues so it can be deleted later
2359  */
2360  
2361 void d_delete(struct dentry * dentry)
2362 {
2363         struct inode *inode = dentry->d_inode;
2364         int isdir = d_is_dir(dentry);
2365
2366         spin_lock(&inode->i_lock);
2367         spin_lock(&dentry->d_lock);
2368         /*
2369          * Are we the only user?
2370          */
2371         if (dentry->d_lockref.count == 1) {
2372                 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2373                 dentry_unlink_inode(dentry);
2374         } else {
2375                 __d_drop(dentry);
2376                 spin_unlock(&dentry->d_lock);
2377                 spin_unlock(&inode->i_lock);
2378         }
2379         fsnotify_nameremove(dentry, isdir);
2380 }
2381 EXPORT_SYMBOL(d_delete);
2382
2383 static void __d_rehash(struct dentry *entry)
2384 {
2385         struct hlist_bl_head *b = d_hash(entry->d_name.hash);
2386
2387         hlist_bl_lock(b);
2388         hlist_bl_add_head_rcu(&entry->d_hash, b);
2389         hlist_bl_unlock(b);
2390 }
2391
2392 /**
2393  * d_rehash     - add an entry back to the hash
2394  * @entry: dentry to add to the hash
2395  *
2396  * Adds a dentry to the hash according to its name.
2397  */
2398  
2399 void d_rehash(struct dentry * entry)
2400 {
2401         spin_lock(&entry->d_lock);
2402         __d_rehash(entry);
2403         spin_unlock(&entry->d_lock);
2404 }
2405 EXPORT_SYMBOL(d_rehash);
2406
2407 static inline unsigned start_dir_add(struct inode *dir)
2408 {
2409
2410         for (;;) {
2411                 unsigned n = dir->i_dir_seq;
2412                 if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
2413                         return n;
2414                 cpu_relax();
2415         }
2416 }
2417
2418 static inline void end_dir_add(struct inode *dir, unsigned n)
2419 {
2420         smp_store_release(&dir->i_dir_seq, n + 2);
2421 }
2422
2423 static void d_wait_lookup(struct dentry *dentry)
2424 {
2425         if (d_in_lookup(dentry)) {
2426                 DECLARE_WAITQUEUE(wait, current);
2427                 add_wait_queue(dentry->d_wait, &wait);
2428                 do {
2429                         set_current_state(TASK_UNINTERRUPTIBLE);
2430                         spin_unlock(&dentry->d_lock);
2431                         schedule();
2432                         spin_lock(&dentry->d_lock);
2433                 } while (d_in_lookup(dentry));
2434         }
2435 }
2436
2437 struct dentry *d_alloc_parallel(struct dentry *parent,
2438                                 const struct qstr *name,
2439                                 wait_queue_head_t *wq)
2440 {
2441         unsigned int hash = name->hash;
2442         struct hlist_bl_head *b = in_lookup_hash(parent, hash);
2443         struct hlist_bl_node *node;
2444         struct dentry *new = d_alloc(parent, name);
2445         struct dentry *dentry;
2446         unsigned seq, r_seq, d_seq;
2447
2448         if (unlikely(!new))
2449                 return ERR_PTR(-ENOMEM);
2450
2451 retry:
2452         rcu_read_lock();
2453         seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
2454         r_seq = read_seqbegin(&rename_lock);
2455         dentry = __d_lookup_rcu(parent, name, &d_seq);
2456         if (unlikely(dentry)) {
2457                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2458                         rcu_read_unlock();
2459                         goto retry;
2460                 }
2461                 if (read_seqcount_retry(&dentry->d_seq, d_seq)) {
2462                         rcu_read_unlock();
2463                         dput(dentry);
2464                         goto retry;
2465                 }
2466                 rcu_read_unlock();
2467                 dput(new);
2468                 return dentry;
2469         }
2470         if (unlikely(read_seqretry(&rename_lock, r_seq))) {
2471                 rcu_read_unlock();
2472                 goto retry;
2473         }
2474
2475         if (unlikely(seq & 1)) {
2476                 rcu_read_unlock();
2477                 goto retry;
2478         }
2479
2480         hlist_bl_lock(b);
2481         if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {
2482                 hlist_bl_unlock(b);
2483                 rcu_read_unlock();
2484                 goto retry;
2485         }
2486         /*
2487          * No changes for the parent since the beginning of d_lookup().
2488          * Since all removals from the chain happen with hlist_bl_lock(),
2489          * any potential in-lookup matches are going to stay here until
2490          * we unlock the chain.  All fields are stable in everything
2491          * we encounter.
2492          */
2493         hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
2494                 if (dentry->d_name.hash != hash)
2495                         continue;
2496                 if (dentry->d_parent != parent)
2497                         continue;
2498                 if (!d_same_name(dentry, parent, name))
2499                         continue;
2500                 hlist_bl_unlock(b);
2501                 /* now we can try to grab a reference */
2502                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2503                         rcu_read_unlock();
2504                         goto retry;
2505                 }
2506
2507                 rcu_read_unlock();
2508                 /*
2509                  * somebody is likely to be still doing lookup for it;
2510                  * wait for them to finish
2511                  */
2512                 spin_lock(&dentry->d_lock);
2513                 d_wait_lookup(dentry);
2514                 /*
2515                  * it's not in-lookup anymore; in principle we should repeat
2516                  * everything from dcache lookup, but it's likely to be what
2517                  * d_lookup() would've found anyway.  If it is, just return it;
2518                  * otherwise we really have to repeat the whole thing.
2519                  */
2520                 if (unlikely(dentry->d_name.hash != hash))
2521                         goto mismatch;
2522                 if (unlikely(dentry->d_parent != parent))
2523                         goto mismatch;
2524                 if (unlikely(d_unhashed(dentry)))
2525                         goto mismatch;
2526                 if (unlikely(!d_same_name(dentry, parent, name)))
2527                         goto mismatch;
2528                 /* OK, it *is* a hashed match; return it */
2529                 spin_unlock(&dentry->d_lock);
2530                 dput(new);
2531                 return dentry;
2532         }
2533         rcu_read_unlock();
2534         /* we can't take ->d_lock here; it's OK, though. */
2535         new->d_flags |= DCACHE_PAR_LOOKUP;
2536         new->d_wait = wq;
2537         hlist_bl_add_head_rcu(&new->d_u.d_in_lookup_hash, b);
2538         hlist_bl_unlock(b);
2539         return new;
2540 mismatch:
2541         spin_unlock(&dentry->d_lock);
2542         dput(dentry);
2543         goto retry;
2544 }
2545 EXPORT_SYMBOL(d_alloc_parallel);
2546
2547 void __d_lookup_done(struct dentry *dentry)
2548 {
2549         struct hlist_bl_head *b = in_lookup_hash(dentry->d_parent,
2550                                                  dentry->d_name.hash);
2551         hlist_bl_lock(b);
2552         dentry->d_flags &= ~DCACHE_PAR_LOOKUP;
2553         __hlist_bl_del(&dentry->d_u.d_in_lookup_hash);
2554         wake_up_all(dentry->d_wait);
2555         dentry->d_wait = NULL;
2556         hlist_bl_unlock(b);
2557         INIT_HLIST_NODE(&dentry->d_u.d_alias);
2558         INIT_LIST_HEAD(&dentry->d_lru);
2559 }
2560 EXPORT_SYMBOL(__d_lookup_done);
2561
2562 /* inode->i_lock held if inode is non-NULL */
2563
2564 static inline void __d_add(struct dentry *dentry, struct inode *inode)
2565 {
2566         struct inode *dir = NULL;
2567         unsigned n;
2568         spin_lock(&dentry->d_lock);
2569         if (unlikely(d_in_lookup(dentry))) {
2570                 dir = dentry->d_parent->d_inode;
2571                 n = start_dir_add(dir);
2572                 __d_lookup_done(dentry);
2573         }
2574         if (inode) {
2575                 unsigned add_flags = d_flags_for_inode(inode);
2576                 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2577                 raw_write_seqcount_begin(&dentry->d_seq);
2578                 __d_set_inode_and_type(dentry, inode, add_flags);
2579                 raw_write_seqcount_end(&dentry->d_seq);
2580                 fsnotify_update_flags(dentry);
2581         }
2582         __d_rehash(dentry);
2583         if (dir)
2584                 end_dir_add(dir, n);
2585         spin_unlock(&dentry->d_lock);
2586         if (inode)
2587                 spin_unlock(&inode->i_lock);
2588 }
2589
2590 /**
2591  * d_add - add dentry to hash queues
2592  * @entry: dentry to add
2593  * @inode: The inode to attach to this dentry
2594  *
2595  * This adds the entry to the hash queues and initializes @inode.
2596  * The entry was actually filled in earlier during d_alloc().
2597  */
2598
2599 void d_add(struct dentry *entry, struct inode *inode)
2600 {
2601         if (inode) {
2602                 security_d_instantiate(entry, inode);
2603                 spin_lock(&inode->i_lock);
2604         }
2605         __d_add(entry, inode);
2606 }
2607 EXPORT_SYMBOL(d_add);
2608
2609 /**
2610  * d_exact_alias - find and hash an exact unhashed alias
2611  * @entry: dentry to add
2612  * @inode: The inode to go with this dentry
2613  *
2614  * If an unhashed dentry with the same name/parent and desired
2615  * inode already exists, hash and return it.  Otherwise, return
2616  * NULL.
2617  *
2618  * Parent directory should be locked.
2619  */
2620 struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2621 {
2622         struct dentry *alias;
2623         unsigned int hash = entry->d_name.hash;
2624
2625         spin_lock(&inode->i_lock);
2626         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2627                 /*
2628                  * Don't need alias->d_lock here, because aliases with
2629                  * d_parent == entry->d_parent are not subject to name or
2630                  * parent changes, because the parent inode i_mutex is held.
2631                  */
2632                 if (alias->d_name.hash != hash)
2633                         continue;
2634                 if (alias->d_parent != entry->d_parent)
2635                         continue;
2636                 if (!d_same_name(alias, entry->d_parent, &entry->d_name))
2637                         continue;
2638                 spin_lock(&alias->d_lock);
2639                 if (!d_unhashed(alias)) {
2640                         spin_unlock(&alias->d_lock);
2641                         alias = NULL;
2642                 } else {
2643                         __dget_dlock(alias);
2644                         __d_rehash(alias);
2645                         spin_unlock(&alias->d_lock);
2646                 }
2647                 spin_unlock(&inode->i_lock);
2648                 return alias;
2649         }
2650         spin_unlock(&inode->i_lock);
2651         return NULL;
2652 }
2653 EXPORT_SYMBOL(d_exact_alias);
2654
2655 static void swap_names(struct dentry *dentry, struct dentry *target)
2656 {
2657         if (unlikely(dname_external(target))) {
2658                 if (unlikely(dname_external(dentry))) {
2659                         /*
2660                          * Both external: swap the pointers
2661                          */
2662                         swap(target->d_name.name, dentry->d_name.name);
2663                 } else {
2664                         /*
2665                          * dentry:internal, target:external.  Steal target's
2666                          * storage and make target internal.
2667                          */
2668                         memcpy(target->d_iname, dentry->d_name.name,
2669                                         dentry->d_name.len + 1);
2670                         dentry->d_name.name = target->d_name.name;
2671                         target->d_name.name = target->d_iname;
2672                 }
2673         } else {
2674                 if (unlikely(dname_external(dentry))) {
2675                         /*
2676                          * dentry:external, target:internal.  Give dentry's
2677                          * storage to target and make dentry internal
2678                          */
2679                         memcpy(dentry->d_iname, target->d_name.name,
2680                                         target->d_name.len + 1);
2681                         target->d_name.name = dentry->d_name.name;
2682                         dentry->d_name.name = dentry->d_iname;
2683                 } else {
2684                         /*
2685                          * Both are internal.
2686                          */
2687                         unsigned int i;
2688                         BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2689                         for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2690                                 swap(((long *) &dentry->d_iname)[i],
2691                                      ((long *) &target->d_iname)[i]);
2692                         }
2693                 }
2694         }
2695         swap(dentry->d_name.hash_len, target->d_name.hash_len);
2696 }
2697
2698 static void copy_name(struct dentry *dentry, struct dentry *target)
2699 {
2700         struct external_name *old_name = NULL;
2701         if (unlikely(dname_external(dentry)))
2702                 old_name = external_name(dentry);
2703         if (unlikely(dname_external(target))) {
2704                 atomic_inc(&external_name(target)->u.count);
2705                 dentry->d_name = target->d_name;
2706         } else {
2707                 memcpy(dentry->d_iname, target->d_name.name,
2708                                 target->d_name.len + 1);
2709                 dentry->d_name.name = dentry->d_iname;
2710                 dentry->d_name.hash_len = target->d_name.hash_len;
2711         }
2712         if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2713                 call_rcu(&old_name->u.head, __d_free_external_name);
2714 }
2715
2716 /*
2717  * When d_splice_alias() moves a directory's encrypted alias to its decrypted
2718  * alias as a result of the encryption key being added, DCACHE_ENCRYPTED_NAME
2719  * must be cleared.  Note that we don't have to support arbitrary moves of this
2720  * flag because fscrypt doesn't allow encrypted aliases to be the source or
2721  * target of a rename().
2722  */
2723 static inline void fscrypt_handle_d_move(struct dentry *dentry)
2724 {
2725 #if IS_ENABLED(CONFIG_FS_ENCRYPTION)
2726         dentry->d_flags &= ~DCACHE_ENCRYPTED_NAME;
2727 #endif
2728 }
2729
2730 /*
2731  * __d_move - move a dentry
2732  * @dentry: entry to move
2733  * @target: new dentry
2734  * @exchange: exchange the two dentries
2735  *
2736  * Update the dcache to reflect the move of a file name. Negative
2737  * dcache entries should not be moved in this way. Caller must hold
2738  * rename_lock, the i_mutex of the source and target directories,
2739  * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2740  */
2741 static void __d_move(struct dentry *dentry, struct dentry *target,
2742                      bool exchange)
2743 {
2744         struct dentry *old_parent, *p;
2745         struct inode *dir = NULL;
2746         unsigned n;
2747
2748         WARN_ON(!dentry->d_inode);
2749         if (WARN_ON(dentry == target))
2750                 return;
2751
2752         BUG_ON(d_ancestor(target, dentry));
2753         old_parent = dentry->d_parent;
2754         p = d_ancestor(old_parent, target);
2755         if (IS_ROOT(dentry)) {
2756                 BUG_ON(p);
2757                 spin_lock(&target->d_parent->d_lock);
2758         } else if (!p) {
2759                 /* target is not a descendent of dentry->d_parent */
2760                 spin_lock(&target->d_parent->d_lock);
2761                 spin_lock_nested(&old_parent->d_lock, DENTRY_D_LOCK_NESTED);
2762         } else {
2763                 BUG_ON(p == dentry);
2764                 spin_lock(&old_parent->d_lock);
2765                 if (p != target)
2766                         spin_lock_nested(&target->d_parent->d_lock,
2767                                         DENTRY_D_LOCK_NESTED);
2768         }
2769         spin_lock_nested(&dentry->d_lock, 2);
2770         spin_lock_nested(&target->d_lock, 3);
2771
2772         if (unlikely(d_in_lookup(target))) {
2773                 dir = target->d_parent->d_inode;
2774                 n = start_dir_add(dir);
2775                 __d_lookup_done(target);
2776         }
2777
2778         write_seqcount_begin(&dentry->d_seq);
2779         write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2780
2781         /* unhash both */
2782         if (!d_unhashed(dentry))
2783                 ___d_drop(dentry);
2784         if (!d_unhashed(target))
2785                 ___d_drop(target);
2786
2787         /* ... and switch them in the tree */
2788         dentry->d_parent = target->d_parent;
2789         if (!exchange) {
2790                 copy_name(dentry, target);
2791                 target->d_hash.pprev = NULL;
2792                 dentry->d_parent->d_lockref.count++;
2793                 if (dentry != old_parent) /* wasn't IS_ROOT */
2794                         WARN_ON(!--old_parent->d_lockref.count);
2795         } else {
2796                 target->d_parent = old_parent;
2797                 swap_names(dentry, target);
2798                 list_move(&target->d_child, &target->d_parent->d_subdirs);
2799                 __d_rehash(target);
2800                 fsnotify_update_flags(target);
2801         }
2802         list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2803         __d_rehash(dentry);
2804         fsnotify_update_flags(dentry);
2805         fscrypt_handle_d_move(dentry);
2806
2807         write_seqcount_end(&target->d_seq);
2808         write_seqcount_end(&dentry->d_seq);
2809
2810         if (dir)
2811                 end_dir_add(dir, n);
2812
2813         if (dentry->d_parent != old_parent)
2814                 spin_unlock(&dentry->d_parent->d_lock);
2815         if (dentry != old_parent)
2816                 spin_unlock(&old_parent->d_lock);
2817         spin_unlock(&target->d_lock);
2818         spin_unlock(&dentry->d_lock);
2819 }
2820
2821 /*
2822  * d_move - move a dentry
2823  * @dentry: entry to move
2824  * @target: new dentry
2825  *
2826  * Update the dcache to reflect the move of a file name. Negative
2827  * dcache entries should not be moved in this way. See the locking
2828  * requirements for __d_move.
2829  */
2830 void d_move(struct dentry *dentry, struct dentry *target)
2831 {
2832         write_seqlock(&rename_lock);
2833         __d_move(dentry, target, false);
2834         write_sequnlock(&rename_lock);
2835 }
2836 EXPORT_SYMBOL(d_move);
2837
2838 /*
2839  * d_exchange - exchange two dentries
2840  * @dentry1: first dentry
2841  * @dentry2: second dentry
2842  */
2843 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2844 {
2845         write_seqlock(&rename_lock);
2846
2847         WARN_ON(!dentry1->d_inode);
2848         WARN_ON(!dentry2->d_inode);
2849         WARN_ON(IS_ROOT(dentry1));
2850         WARN_ON(IS_ROOT(dentry2));
2851
2852         __d_move(dentry1, dentry2, true);
2853
2854         write_sequnlock(&rename_lock);
2855 }
2856
2857 /**
2858  * d_ancestor - search for an ancestor
2859  * @p1: ancestor dentry
2860  * @p2: child dentry
2861  *
2862  * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2863  * an ancestor of p2, else NULL.
2864  */
2865 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2866 {
2867         struct dentry *p;
2868
2869         for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2870                 if (p->d_parent == p1)
2871                         return p;
2872         }
2873         return NULL;
2874 }
2875
2876 /*
2877  * This helper attempts to cope with remotely renamed directories
2878  *
2879  * It assumes that the caller is already holding
2880  * dentry->d_parent->d_inode->i_mutex, and rename_lock
2881  *
2882  * Note: If ever the locking in lock_rename() changes, then please
2883  * remember to update this too...
2884  */
2885 static int __d_unalias(struct inode *inode,
2886                 struct dentry *dentry, struct dentry *alias)
2887 {
2888         struct mutex *m1 = NULL;
2889         struct rw_semaphore *m2 = NULL;
2890         int ret = -ESTALE;
2891
2892         /* If alias and dentry share a parent, then no extra locks required */
2893         if (alias->d_parent == dentry->d_parent)
2894                 goto out_unalias;
2895
2896         /* See lock_rename() */
2897         if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2898                 goto out_err;
2899         m1 = &dentry->d_sb->s_vfs_rename_mutex;
2900         if (!inode_trylock_shared(alias->d_parent->d_inode))
2901                 goto out_err;
2902         m2 = &alias->d_parent->d_inode->i_rwsem;
2903 out_unalias:
2904         __d_move(alias, dentry, false);
2905         ret = 0;
2906 out_err:
2907         if (m2)
2908                 up_read(m2);
2909         if (m1)
2910                 mutex_unlock(m1);
2911         return ret;
2912 }
2913
2914 /**
2915  * d_splice_alias - splice a disconnected dentry into the tree if one exists
2916  * @inode:  the inode which may have a disconnected dentry
2917  * @dentry: a negative dentry which we want to point to the inode.
2918  *
2919  * If inode is a directory and has an IS_ROOT alias, then d_move that in
2920  * place of the given dentry and return it, else simply d_add the inode
2921  * to the dentry and return NULL.
2922  *
2923  * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2924  * we should error out: directories can't have multiple aliases.
2925  *
2926  * This is needed in the lookup routine of any filesystem that is exportable
2927  * (via knfsd) so that we can build dcache paths to directories effectively.
2928  *
2929  * If a dentry was found and moved, then it is returned.  Otherwise NULL
2930  * is returned.  This matches the expected return value of ->lookup.
2931  *
2932  * Cluster filesystems may call this function with a negative, hashed dentry.
2933  * In that case, we know that the inode will be a regular file, and also this
2934  * will only occur during atomic_open. So we need to check for the dentry
2935  * being already hashed only in the final case.
2936  */
2937 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
2938 {
2939         if (IS_ERR(inode))
2940                 return ERR_CAST(inode);
2941
2942         BUG_ON(!d_unhashed(dentry));
2943
2944         if (!inode)
2945                 goto out;
2946
2947         security_d_instantiate(dentry, inode);
2948         spin_lock(&inode->i_lock);
2949         if (S_ISDIR(inode->i_mode)) {
2950                 struct dentry *new = __d_find_any_alias(inode);
2951                 if (unlikely(new)) {
2952                         /* The reference to new ensures it remains an alias */
2953                         spin_unlock(&inode->i_lock);
2954                         write_seqlock(&rename_lock);
2955                         if (unlikely(d_ancestor(new, dentry))) {
2956                                 write_sequnlock(&rename_lock);
2957                                 dput(new);
2958                                 new = ERR_PTR(-ELOOP);
2959                                 pr_warn_ratelimited(
2960                                         "VFS: Lookup of '%s' in %s %s"
2961                                         " would have caused loop\n",
2962                                         dentry->d_name.name,
2963                                         inode->i_sb->s_type->name,
2964                                         inode->i_sb->s_id);
2965                         } else if (!IS_ROOT(new)) {
2966                                 struct dentry *old_parent = dget(new->d_parent);
2967                                 int err = __d_unalias(inode, dentry, new);
2968                                 write_sequnlock(&rename_lock);
2969                                 if (err) {
2970                                         dput(new);
2971                                         new = ERR_PTR(err);
2972                                 }
2973                                 dput(old_parent);
2974                         } else {
2975                                 __d_move(new, dentry, false);
2976                                 write_sequnlock(&rename_lock);
2977                         }
2978                         iput(inode);
2979                         return new;
2980                 }
2981         }
2982 out:
2983         __d_add(dentry, inode);
2984         return NULL;
2985 }
2986 EXPORT_SYMBOL(d_splice_alias);
2987
2988 /*
2989  * Test whether new_dentry is a subdirectory of old_dentry.
2990  *
2991  * Trivially implemented using the dcache structure
2992  */
2993
2994 /**
2995  * is_subdir - is new dentry a subdirectory of old_dentry
2996  * @new_dentry: new dentry
2997  * @old_dentry: old dentry
2998  *
2999  * Returns true if new_dentry is a subdirectory of the parent (at any depth).
3000  * Returns false otherwise.
3001  * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3002  */
3003   
3004 bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3005 {
3006         bool result;
3007         unsigned seq;
3008
3009         if (new_dentry == old_dentry)
3010                 return true;
3011
3012         do {
3013                 /* for restarting inner loop in case of seq retry */
3014                 seq = read_seqbegin(&rename_lock);
3015                 /*
3016                  * Need rcu_readlock to protect against the d_parent trashing
3017                  * due to d_move
3018                  */
3019                 rcu_read_lock();
3020                 if (d_ancestor(old_dentry, new_dentry))
3021                         result = true;
3022                 else
3023                         result = false;
3024                 rcu_read_unlock();
3025         } while (read_seqretry(&rename_lock, seq));
3026
3027         return result;
3028 }
3029 EXPORT_SYMBOL(is_subdir);
3030
3031 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3032 {
3033         struct dentry *root = data;
3034         if (dentry != root) {
3035                 if (d_unhashed(dentry) || !dentry->d_inode)
3036                         return D_WALK_SKIP;
3037
3038                 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3039                         dentry->d_flags |= DCACHE_GENOCIDE;
3040                         dentry->d_lockref.count--;
3041                 }
3042         }
3043         return D_WALK_CONTINUE;
3044 }
3045
3046 void d_genocide(struct dentry *parent)
3047 {
3048         d_walk(parent, parent, d_genocide_kill);
3049 }
3050
3051 EXPORT_SYMBOL(d_genocide);
3052
3053 void d_tmpfile(struct dentry *dentry, struct inode *inode)
3054 {
3055         inode_dec_link_count(inode);
3056         BUG_ON(dentry->d_name.name != dentry->d_iname ||
3057                 !hlist_unhashed(&dentry->d_u.d_alias) ||
3058                 !d_unlinked(dentry));
3059         spin_lock(&dentry->d_parent->d_lock);
3060         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3061         dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3062                                 (unsigned long long)inode->i_ino);
3063         spin_unlock(&dentry->d_lock);
3064         spin_unlock(&dentry->d_parent->d_lock);
3065         d_instantiate(dentry, inode);
3066 }
3067 EXPORT_SYMBOL(d_tmpfile);
3068
3069 static __initdata unsigned long dhash_entries;
3070 static int __init set_dhash_entries(char *str)
3071 {
3072         if (!str)
3073                 return 0;
3074         dhash_entries = simple_strtoul(str, &str, 0);
3075         return 1;
3076 }
3077 __setup("dhash_entries=", set_dhash_entries);
3078
3079 static void __init dcache_init_early(void)
3080 {
3081         /* If hashes are distributed across NUMA nodes, defer
3082          * hash allocation until vmalloc space is available.
3083          */
3084         if (hashdist)
3085                 return;
3086
3087         dentry_hashtable =
3088                 alloc_large_system_hash("Dentry cache",
3089                                         sizeof(struct hlist_bl_head),
3090                                         dhash_entries,
3091                                         13,
3092                                         HASH_EARLY | HASH_ZERO,
3093                                         &d_hash_shift,
3094                                         NULL,
3095                                         0,
3096                                         0);
3097         d_hash_shift = 32 - d_hash_shift;
3098 }
3099
3100 static void __init dcache_init(void)
3101 {
3102         /*
3103          * A constructor could be added for stable state like the lists,
3104          * but it is probably not worth it because of the cache nature
3105          * of the dcache.
3106          */
3107         dentry_cache = KMEM_CACHE_USERCOPY(dentry,
3108                 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD|SLAB_ACCOUNT,
3109                 d_iname);
3110
3111         /* Hash may have been set up in dcache_init_early */
3112         if (!hashdist)
3113                 return;
3114
3115         dentry_hashtable =
3116                 alloc_large_system_hash("Dentry cache",
3117                                         sizeof(struct hlist_bl_head),
3118                                         dhash_entries,
3119                                         13,
3120                                         HASH_ZERO,
3121                                         &d_hash_shift,
3122                                         NULL,
3123                                         0,
3124                                         0);
3125         d_hash_shift = 32 - d_hash_shift;
3126 }
3127
3128 /* SLAB cache for __getname() consumers */
3129 struct kmem_cache *names_cachep __read_mostly;
3130 EXPORT_SYMBOL(names_cachep);
3131
3132 void __init vfs_caches_init_early(void)
3133 {
3134         int i;
3135
3136         for (i = 0; i < ARRAY_SIZE(in_lookup_hashtable); i++)
3137                 INIT_HLIST_BL_HEAD(&in_lookup_hashtable[i]);
3138
3139         dcache_init_early();
3140         inode_init_early();
3141 }
3142
3143 void __init vfs_caches_init(void)
3144 {
3145         names_cachep = kmem_cache_create_usercopy("names_cache", PATH_MAX, 0,
3146                         SLAB_HWCACHE_ALIGN|SLAB_PANIC, 0, PATH_MAX, NULL);
3147
3148         dcache_init();
3149         inode_init();
3150         files_init();
3151         files_maxfiles_init();
3152         mnt_init();
3153         bdev_cache_init();
3154         chrdev_init();
3155 }