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