Linux 6.9-rc1
[linux-2.6-microblaze.git] / fs / dcache.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * fs/dcache.c
4  *
5  * Complete reimplementation
6  * (C) 1997 Thomas Schoebel-Theuer,
7  * with heavy changes by Linus Torvalds
8  */
9
10 /*
11  * Notes on the allocation strategy:
12  *
13  * The dcache is a master of the icache - whenever a dcache entry
14  * exists, the inode will always exist. "iput()" is done either when
15  * the dcache entry is deleted or garbage collected.
16  */
17
18 #include <linux/ratelimit.h>
19 #include <linux/string.h>
20 #include <linux/mm.h>
21 #include <linux/fs.h>
22 #include <linux/fscrypt.h>
23 #include <linux/fsnotify.h>
24 #include <linux/slab.h>
25 #include <linux/init.h>
26 #include <linux/hash.h>
27 #include <linux/cache.h>
28 #include <linux/export.h>
29 #include <linux/security.h>
30 #include <linux/seqlock.h>
31 #include <linux/memblock.h>
32 #include <linux/bit_spinlock.h>
33 #include <linux/rculist_bl.h>
34 #include <linux/list_lru.h>
35 #include "internal.h"
36 #include "mount.h"
37
38 /*
39  * Usage:
40  * dcache->d_inode->i_lock protects:
41  *   - i_dentry, d_u.d_alias, d_inode of aliases
42  * dcache_hash_bucket lock protects:
43  *   - the dcache hash table
44  * s_roots bl list spinlock protects:
45  *   - the s_roots list (see __d_drop)
46  * dentry->d_sb->s_dentry_lru_lock protects:
47  *   - the dcache lru lists and counters
48  * d_lock protects:
49  *   - d_flags
50  *   - d_name
51  *   - d_lru
52  *   - d_count
53  *   - d_unhashed()
54  *   - d_parent and d_chilren
55  *   - childrens' d_sib and d_parent
56  *   - d_u.d_alias, d_inode
57  *
58  * Ordering:
59  * dentry->d_inode->i_lock
60  *   dentry->d_lock
61  *     dentry->d_sb->s_dentry_lru_lock
62  *     dcache_hash_bucket lock
63  *     s_roots lock
64  *
65  * If there is an ancestor relationship:
66  * dentry->d_parent->...->d_parent->d_lock
67  *   ...
68  *     dentry->d_parent->d_lock
69  *       dentry->d_lock
70  *
71  * If no ancestor relationship:
72  * arbitrary, since it's serialized on rename_lock
73  */
74 int sysctl_vfs_cache_pressure __read_mostly = 100;
75 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
76
77 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
78
79 EXPORT_SYMBOL(rename_lock);
80
81 static struct kmem_cache *dentry_cache __ro_after_init;
82
83 const struct qstr empty_name = QSTR_INIT("", 0);
84 EXPORT_SYMBOL(empty_name);
85 const struct qstr slash_name = QSTR_INIT("/", 1);
86 EXPORT_SYMBOL(slash_name);
87 const struct qstr dotdot_name = QSTR_INIT("..", 2);
88 EXPORT_SYMBOL(dotdot_name);
89
90 /*
91  * This is the single most critical data structure when it comes
92  * to the dcache: the hashtable for lookups. Somebody should try
93  * to make this good - I've just made it work.
94  *
95  * This hash-function tries to avoid losing too many bits of hash
96  * information, yet avoid using a prime hash-size or similar.
97  */
98
99 static unsigned int d_hash_shift __ro_after_init;
100
101 static struct hlist_bl_head *dentry_hashtable __ro_after_init;
102
103 static inline struct hlist_bl_head *d_hash(unsigned int hash)
104 {
105         return dentry_hashtable + (hash >> d_hash_shift);
106 }
107
108 #define IN_LOOKUP_SHIFT 10
109 static struct hlist_bl_head in_lookup_hashtable[1 << IN_LOOKUP_SHIFT];
110
111 static inline struct hlist_bl_head *in_lookup_hash(const struct dentry *parent,
112                                         unsigned int hash)
113 {
114         hash += (unsigned long) parent / L1_CACHE_BYTES;
115         return in_lookup_hashtable + hash_32(hash, IN_LOOKUP_SHIFT);
116 }
117
118 struct dentry_stat_t {
119         long nr_dentry;
120         long nr_unused;
121         long age_limit;         /* age in seconds */
122         long want_pages;        /* pages requested by system */
123         long nr_negative;       /* # of unused negative dentries */
124         long dummy;             /* Reserved for future use */
125 };
126
127 static DEFINE_PER_CPU(long, nr_dentry);
128 static DEFINE_PER_CPU(long, nr_dentry_unused);
129 static DEFINE_PER_CPU(long, nr_dentry_negative);
130
131 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
132 /* Statistics gathering. */
133 static struct dentry_stat_t dentry_stat = {
134         .age_limit = 45,
135 };
136
137 /*
138  * Here we resort to our own counters instead of using generic per-cpu counters
139  * for consistency with what the vfs inode code does. We are expected to harvest
140  * better code and performance by having our own specialized counters.
141  *
142  * Please note that the loop is done over all possible CPUs, not over all online
143  * CPUs. The reason for this is that we don't want to play games with CPUs going
144  * on and off. If one of them goes off, we will just keep their counters.
145  *
146  * glommer: See cffbc8a for details, and if you ever intend to change this,
147  * please update all vfs counters to match.
148  */
149 static long get_nr_dentry(void)
150 {
151         int i;
152         long sum = 0;
153         for_each_possible_cpu(i)
154                 sum += per_cpu(nr_dentry, i);
155         return sum < 0 ? 0 : sum;
156 }
157
158 static long get_nr_dentry_unused(void)
159 {
160         int i;
161         long sum = 0;
162         for_each_possible_cpu(i)
163                 sum += per_cpu(nr_dentry_unused, i);
164         return sum < 0 ? 0 : sum;
165 }
166
167 static long get_nr_dentry_negative(void)
168 {
169         int i;
170         long sum = 0;
171
172         for_each_possible_cpu(i)
173                 sum += per_cpu(nr_dentry_negative, i);
174         return sum < 0 ? 0 : sum;
175 }
176
177 static int proc_nr_dentry(struct ctl_table *table, int write, void *buffer,
178                           size_t *lenp, loff_t *ppos)
179 {
180         dentry_stat.nr_dentry = get_nr_dentry();
181         dentry_stat.nr_unused = get_nr_dentry_unused();
182         dentry_stat.nr_negative = get_nr_dentry_negative();
183         return proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
184 }
185
186 static struct ctl_table fs_dcache_sysctls[] = {
187         {
188                 .procname       = "dentry-state",
189                 .data           = &dentry_stat,
190                 .maxlen         = 6*sizeof(long),
191                 .mode           = 0444,
192                 .proc_handler   = proc_nr_dentry,
193         },
194 };
195
196 static int __init init_fs_dcache_sysctls(void)
197 {
198         register_sysctl_init("fs", fs_dcache_sysctls);
199         return 0;
200 }
201 fs_initcall(init_fs_dcache_sysctls);
202 #endif
203
204 /*
205  * Compare 2 name strings, return 0 if they match, otherwise non-zero.
206  * The strings are both count bytes long, and count is non-zero.
207  */
208 #ifdef CONFIG_DCACHE_WORD_ACCESS
209
210 #include <asm/word-at-a-time.h>
211 /*
212  * NOTE! 'cs' and 'scount' come from a dentry, so it has a
213  * aligned allocation for this particular component. We don't
214  * strictly need the load_unaligned_zeropad() safety, but it
215  * doesn't hurt either.
216  *
217  * In contrast, 'ct' and 'tcount' can be from a pathname, and do
218  * need the careful unaligned handling.
219  */
220 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
221 {
222         unsigned long a,b,mask;
223
224         for (;;) {
225                 a = read_word_at_a_time(cs);
226                 b = load_unaligned_zeropad(ct);
227                 if (tcount < sizeof(unsigned long))
228                         break;
229                 if (unlikely(a != b))
230                         return 1;
231                 cs += sizeof(unsigned long);
232                 ct += sizeof(unsigned long);
233                 tcount -= sizeof(unsigned long);
234                 if (!tcount)
235                         return 0;
236         }
237         mask = bytemask_from_count(tcount);
238         return unlikely(!!((a ^ b) & mask));
239 }
240
241 #else
242
243 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
244 {
245         do {
246                 if (*cs != *ct)
247                         return 1;
248                 cs++;
249                 ct++;
250                 tcount--;
251         } while (tcount);
252         return 0;
253 }
254
255 #endif
256
257 static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount)
258 {
259         /*
260          * Be careful about RCU walk racing with rename:
261          * use 'READ_ONCE' to fetch the name pointer.
262          *
263          * NOTE! Even if a rename will mean that the length
264          * was not loaded atomically, we don't care. The
265          * RCU walk will check the sequence count eventually,
266          * and catch it. And we won't overrun the buffer,
267          * because we're reading the name pointer atomically,
268          * and a dentry name is guaranteed to be properly
269          * terminated with a NUL byte.
270          *
271          * End result: even if 'len' is wrong, we'll exit
272          * early because the data cannot match (there can
273          * be no NUL in the ct/tcount data)
274          */
275         const unsigned char *cs = READ_ONCE(dentry->d_name.name);
276
277         return dentry_string_cmp(cs, ct, tcount);
278 }
279
280 struct external_name {
281         union {
282                 atomic_t count;
283                 struct rcu_head head;
284         } u;
285         unsigned char name[];
286 };
287
288 static inline struct external_name *external_name(struct dentry *dentry)
289 {
290         return container_of(dentry->d_name.name, struct external_name, name[0]);
291 }
292
293 static void __d_free(struct rcu_head *head)
294 {
295         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
296
297         kmem_cache_free(dentry_cache, dentry); 
298 }
299
300 static void __d_free_external(struct rcu_head *head)
301 {
302         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
303         kfree(external_name(dentry));
304         kmem_cache_free(dentry_cache, dentry);
305 }
306
307 static inline int dname_external(const struct dentry *dentry)
308 {
309         return dentry->d_name.name != dentry->d_iname;
310 }
311
312 void take_dentry_name_snapshot(struct name_snapshot *name, struct dentry *dentry)
313 {
314         spin_lock(&dentry->d_lock);
315         name->name = dentry->d_name;
316         if (unlikely(dname_external(dentry))) {
317                 atomic_inc(&external_name(dentry)->u.count);
318         } else {
319                 memcpy(name->inline_name, dentry->d_iname,
320                        dentry->d_name.len + 1);
321                 name->name.name = name->inline_name;
322         }
323         spin_unlock(&dentry->d_lock);
324 }
325 EXPORT_SYMBOL(take_dentry_name_snapshot);
326
327 void release_dentry_name_snapshot(struct name_snapshot *name)
328 {
329         if (unlikely(name->name.name != name->inline_name)) {
330                 struct external_name *p;
331                 p = container_of(name->name.name, struct external_name, name[0]);
332                 if (unlikely(atomic_dec_and_test(&p->u.count)))
333                         kfree_rcu(p, u.head);
334         }
335 }
336 EXPORT_SYMBOL(release_dentry_name_snapshot);
337
338 static inline void __d_set_inode_and_type(struct dentry *dentry,
339                                           struct inode *inode,
340                                           unsigned type_flags)
341 {
342         unsigned flags;
343
344         dentry->d_inode = inode;
345         flags = READ_ONCE(dentry->d_flags);
346         flags &= ~DCACHE_ENTRY_TYPE;
347         flags |= type_flags;
348         smp_store_release(&dentry->d_flags, flags);
349 }
350
351 static inline void __d_clear_type_and_inode(struct dentry *dentry)
352 {
353         unsigned flags = READ_ONCE(dentry->d_flags);
354
355         flags &= ~DCACHE_ENTRY_TYPE;
356         WRITE_ONCE(dentry->d_flags, flags);
357         dentry->d_inode = NULL;
358         if (dentry->d_flags & DCACHE_LRU_LIST)
359                 this_cpu_inc(nr_dentry_negative);
360 }
361
362 static void dentry_free(struct dentry *dentry)
363 {
364         WARN_ON(!hlist_unhashed(&dentry->d_u.d_alias));
365         if (unlikely(dname_external(dentry))) {
366                 struct external_name *p = external_name(dentry);
367                 if (likely(atomic_dec_and_test(&p->u.count))) {
368                         call_rcu(&dentry->d_u.d_rcu, __d_free_external);
369                         return;
370                 }
371         }
372         /* if dentry was never visible to RCU, immediate free is OK */
373         if (dentry->d_flags & DCACHE_NORCU)
374                 __d_free(&dentry->d_u.d_rcu);
375         else
376                 call_rcu(&dentry->d_u.d_rcu, __d_free);
377 }
378
379 /*
380  * Release the dentry's inode, using the filesystem
381  * d_iput() operation if defined.
382  */
383 static void dentry_unlink_inode(struct dentry * dentry)
384         __releases(dentry->d_lock)
385         __releases(dentry->d_inode->i_lock)
386 {
387         struct inode *inode = dentry->d_inode;
388
389         raw_write_seqcount_begin(&dentry->d_seq);
390         __d_clear_type_and_inode(dentry);
391         hlist_del_init(&dentry->d_u.d_alias);
392         raw_write_seqcount_end(&dentry->d_seq);
393         spin_unlock(&dentry->d_lock);
394         spin_unlock(&inode->i_lock);
395         if (!inode->i_nlink)
396                 fsnotify_inoderemove(inode);
397         if (dentry->d_op && dentry->d_op->d_iput)
398                 dentry->d_op->d_iput(dentry, inode);
399         else
400                 iput(inode);
401 }
402
403 /*
404  * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
405  * is in use - which includes both the "real" per-superblock
406  * LRU list _and_ the DCACHE_SHRINK_LIST use.
407  *
408  * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
409  * on the shrink list (ie not on the superblock LRU list).
410  *
411  * The per-cpu "nr_dentry_unused" counters are updated with
412  * the DCACHE_LRU_LIST bit.
413  *
414  * The per-cpu "nr_dentry_negative" counters are only updated
415  * when deleted from or added to the per-superblock LRU list, not
416  * from/to the shrink list. That is to avoid an unneeded dec/inc
417  * pair when moving from LRU to shrink list in select_collect().
418  *
419  * These helper functions make sure we always follow the
420  * rules. d_lock must be held by the caller.
421  */
422 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
423 static void d_lru_add(struct dentry *dentry)
424 {
425         D_FLAG_VERIFY(dentry, 0);
426         dentry->d_flags |= DCACHE_LRU_LIST;
427         this_cpu_inc(nr_dentry_unused);
428         if (d_is_negative(dentry))
429                 this_cpu_inc(nr_dentry_negative);
430         WARN_ON_ONCE(!list_lru_add_obj(
431                         &dentry->d_sb->s_dentry_lru, &dentry->d_lru));
432 }
433
434 static void d_lru_del(struct dentry *dentry)
435 {
436         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
437         dentry->d_flags &= ~DCACHE_LRU_LIST;
438         this_cpu_dec(nr_dentry_unused);
439         if (d_is_negative(dentry))
440                 this_cpu_dec(nr_dentry_negative);
441         WARN_ON_ONCE(!list_lru_del_obj(
442                         &dentry->d_sb->s_dentry_lru, &dentry->d_lru));
443 }
444
445 static void d_shrink_del(struct dentry *dentry)
446 {
447         D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
448         list_del_init(&dentry->d_lru);
449         dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
450         this_cpu_dec(nr_dentry_unused);
451 }
452
453 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
454 {
455         D_FLAG_VERIFY(dentry, 0);
456         list_add(&dentry->d_lru, list);
457         dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
458         this_cpu_inc(nr_dentry_unused);
459 }
460
461 /*
462  * These can only be called under the global LRU lock, ie during the
463  * callback for freeing the LRU list. "isolate" removes it from the
464  * LRU lists entirely, while shrink_move moves it to the indicated
465  * private list.
466  */
467 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
468 {
469         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
470         dentry->d_flags &= ~DCACHE_LRU_LIST;
471         this_cpu_dec(nr_dentry_unused);
472         if (d_is_negative(dentry))
473                 this_cpu_dec(nr_dentry_negative);
474         list_lru_isolate(lru, &dentry->d_lru);
475 }
476
477 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
478                               struct list_head *list)
479 {
480         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
481         dentry->d_flags |= DCACHE_SHRINK_LIST;
482         if (d_is_negative(dentry))
483                 this_cpu_dec(nr_dentry_negative);
484         list_lru_isolate_move(lru, &dentry->d_lru, list);
485 }
486
487 static void ___d_drop(struct dentry *dentry)
488 {
489         struct hlist_bl_head *b;
490         /*
491          * Hashed dentries are normally on the dentry hashtable,
492          * with the exception of those newly allocated by
493          * d_obtain_root, which are always IS_ROOT:
494          */
495         if (unlikely(IS_ROOT(dentry)))
496                 b = &dentry->d_sb->s_roots;
497         else
498                 b = d_hash(dentry->d_name.hash);
499
500         hlist_bl_lock(b);
501         __hlist_bl_del(&dentry->d_hash);
502         hlist_bl_unlock(b);
503 }
504
505 void __d_drop(struct dentry *dentry)
506 {
507         if (!d_unhashed(dentry)) {
508                 ___d_drop(dentry);
509                 dentry->d_hash.pprev = NULL;
510                 write_seqcount_invalidate(&dentry->d_seq);
511         }
512 }
513 EXPORT_SYMBOL(__d_drop);
514
515 /**
516  * d_drop - drop a dentry
517  * @dentry: dentry to drop
518  *
519  * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
520  * be found through a VFS lookup any more. Note that this is different from
521  * deleting the dentry - d_delete will try to mark the dentry negative if
522  * possible, giving a successful _negative_ lookup, while d_drop will
523  * just make the cache lookup fail.
524  *
525  * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
526  * reason (NFS timeouts or autofs deletes).
527  *
528  * __d_drop requires dentry->d_lock
529  *
530  * ___d_drop doesn't mark dentry as "unhashed"
531  * (dentry->d_hash.pprev will be LIST_POISON2, not NULL).
532  */
533 void d_drop(struct dentry *dentry)
534 {
535         spin_lock(&dentry->d_lock);
536         __d_drop(dentry);
537         spin_unlock(&dentry->d_lock);
538 }
539 EXPORT_SYMBOL(d_drop);
540
541 static inline void dentry_unlist(struct dentry *dentry)
542 {
543         struct dentry *next;
544         /*
545          * Inform d_walk() and shrink_dentry_list() that we are no longer
546          * attached to the dentry tree
547          */
548         dentry->d_flags |= DCACHE_DENTRY_KILLED;
549         if (unlikely(hlist_unhashed(&dentry->d_sib)))
550                 return;
551         __hlist_del(&dentry->d_sib);
552         /*
553          * Cursors can move around the list of children.  While we'd been
554          * a normal list member, it didn't matter - ->d_sib.next would've
555          * been updated.  However, from now on it won't be and for the
556          * things like d_walk() it might end up with a nasty surprise.
557          * Normally d_walk() doesn't care about cursors moving around -
558          * ->d_lock on parent prevents that and since a cursor has no children
559          * of its own, we get through it without ever unlocking the parent.
560          * There is one exception, though - if we ascend from a child that
561          * gets killed as soon as we unlock it, the next sibling is found
562          * using the value left in its ->d_sib.next.  And if _that_
563          * pointed to a cursor, and cursor got moved (e.g. by lseek())
564          * before d_walk() regains parent->d_lock, we'll end up skipping
565          * everything the cursor had been moved past.
566          *
567          * Solution: make sure that the pointer left behind in ->d_sib.next
568          * points to something that won't be moving around.  I.e. skip the
569          * cursors.
570          */
571         while (dentry->d_sib.next) {
572                 next = hlist_entry(dentry->d_sib.next, struct dentry, d_sib);
573                 if (likely(!(next->d_flags & DCACHE_DENTRY_CURSOR)))
574                         break;
575                 dentry->d_sib.next = next->d_sib.next;
576         }
577 }
578
579 static struct dentry *__dentry_kill(struct dentry *dentry)
580 {
581         struct dentry *parent = NULL;
582         bool can_free = true;
583
584         /*
585          * The dentry is now unrecoverably dead to the world.
586          */
587         lockref_mark_dead(&dentry->d_lockref);
588
589         /*
590          * inform the fs via d_prune that this dentry is about to be
591          * unhashed and destroyed.
592          */
593         if (dentry->d_flags & DCACHE_OP_PRUNE)
594                 dentry->d_op->d_prune(dentry);
595
596         if (dentry->d_flags & DCACHE_LRU_LIST) {
597                 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
598                         d_lru_del(dentry);
599         }
600         /* if it was on the hash then remove it */
601         __d_drop(dentry);
602         if (dentry->d_inode)
603                 dentry_unlink_inode(dentry);
604         else
605                 spin_unlock(&dentry->d_lock);
606         this_cpu_dec(nr_dentry);
607         if (dentry->d_op && dentry->d_op->d_release)
608                 dentry->d_op->d_release(dentry);
609
610         cond_resched();
611         /* now that it's negative, ->d_parent is stable */
612         if (!IS_ROOT(dentry)) {
613                 parent = dentry->d_parent;
614                 spin_lock(&parent->d_lock);
615         }
616         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
617         dentry_unlist(dentry);
618         if (dentry->d_flags & DCACHE_SHRINK_LIST)
619                 can_free = false;
620         spin_unlock(&dentry->d_lock);
621         if (likely(can_free))
622                 dentry_free(dentry);
623         if (parent && --parent->d_lockref.count) {
624                 spin_unlock(&parent->d_lock);
625                 return NULL;
626         }
627         return parent;
628 }
629
630 /*
631  * Lock a dentry for feeding it to __dentry_kill().
632  * Called under rcu_read_lock() and dentry->d_lock; the former
633  * guarantees that nothing we access will be freed under us.
634  * Note that dentry is *not* protected from concurrent dentry_kill(),
635  * d_delete(), etc.
636  *
637  * Return false if dentry is busy.  Otherwise, return true and have
638  * that dentry's inode locked.
639  */
640
641 static bool lock_for_kill(struct dentry *dentry)
642 {
643         struct inode *inode = dentry->d_inode;
644
645         if (unlikely(dentry->d_lockref.count))
646                 return false;
647
648         if (!inode || likely(spin_trylock(&inode->i_lock)))
649                 return true;
650
651         do {
652                 spin_unlock(&dentry->d_lock);
653                 spin_lock(&inode->i_lock);
654                 spin_lock(&dentry->d_lock);
655                 if (likely(inode == dentry->d_inode))
656                         break;
657                 spin_unlock(&inode->i_lock);
658                 inode = dentry->d_inode;
659         } while (inode);
660         if (likely(!dentry->d_lockref.count))
661                 return true;
662         if (inode)
663                 spin_unlock(&inode->i_lock);
664         return false;
665 }
666
667 /*
668  * Decide if dentry is worth retaining.  Usually this is called with dentry
669  * locked; if not locked, we are more limited and might not be able to tell
670  * without a lock.  False in this case means "punt to locked path and recheck".
671  *
672  * In case we aren't locked, these predicates are not "stable". However, it is
673  * sufficient that at some point after we dropped the reference the dentry was
674  * hashed and the flags had the proper value. Other dentry users may have
675  * re-gotten a reference to the dentry and change that, but our work is done -
676  * we can leave the dentry around with a zero refcount.
677  */
678 static inline bool retain_dentry(struct dentry *dentry, bool locked)
679 {
680         unsigned int d_flags;
681
682         smp_rmb();
683         d_flags = READ_ONCE(dentry->d_flags);
684
685         // Unreachable? Nobody would be able to look it up, no point retaining
686         if (unlikely(d_unhashed(dentry)))
687                 return false;
688
689         // Same if it's disconnected
690         if (unlikely(d_flags & DCACHE_DISCONNECTED))
691                 return false;
692
693         // ->d_delete() might tell us not to bother, but that requires
694         // ->d_lock; can't decide without it
695         if (unlikely(d_flags & DCACHE_OP_DELETE)) {
696                 if (!locked || dentry->d_op->d_delete(dentry))
697                         return false;
698         }
699
700         // Explicitly told not to bother
701         if (unlikely(d_flags & DCACHE_DONTCACHE))
702                 return false;
703
704         // At this point it looks like we ought to keep it.  We also might
705         // need to do something - put it on LRU if it wasn't there already
706         // and mark it referenced if it was on LRU, but not marked yet.
707         // Unfortunately, both actions require ->d_lock, so in lockless
708         // case we'd have to punt rather than doing those.
709         if (unlikely(!(d_flags & DCACHE_LRU_LIST))) {
710                 if (!locked)
711                         return false;
712                 d_lru_add(dentry);
713         } else if (unlikely(!(d_flags & DCACHE_REFERENCED))) {
714                 if (!locked)
715                         return false;
716                 dentry->d_flags |= DCACHE_REFERENCED;
717         }
718         return true;
719 }
720
721 void d_mark_dontcache(struct inode *inode)
722 {
723         struct dentry *de;
724
725         spin_lock(&inode->i_lock);
726         hlist_for_each_entry(de, &inode->i_dentry, d_u.d_alias) {
727                 spin_lock(&de->d_lock);
728                 de->d_flags |= DCACHE_DONTCACHE;
729                 spin_unlock(&de->d_lock);
730         }
731         inode->i_state |= I_DONTCACHE;
732         spin_unlock(&inode->i_lock);
733 }
734 EXPORT_SYMBOL(d_mark_dontcache);
735
736 /*
737  * Try to do a lockless dput(), and return whether that was successful.
738  *
739  * If unsuccessful, we return false, having already taken the dentry lock.
740  * In that case refcount is guaranteed to be zero and we have already
741  * decided that it's not worth keeping around.
742  *
743  * The caller needs to hold the RCU read lock, so that the dentry is
744  * guaranteed to stay around even if the refcount goes down to zero!
745  */
746 static inline bool fast_dput(struct dentry *dentry)
747 {
748         int ret;
749
750         /*
751          * try to decrement the lockref optimistically.
752          */
753         ret = lockref_put_return(&dentry->d_lockref);
754
755         /*
756          * If the lockref_put_return() failed due to the lock being held
757          * by somebody else, the fast path has failed. We will need to
758          * get the lock, and then check the count again.
759          */
760         if (unlikely(ret < 0)) {
761                 spin_lock(&dentry->d_lock);
762                 if (WARN_ON_ONCE(dentry->d_lockref.count <= 0)) {
763                         spin_unlock(&dentry->d_lock);
764                         return true;
765                 }
766                 dentry->d_lockref.count--;
767                 goto locked;
768         }
769
770         /*
771          * If we weren't the last ref, we're done.
772          */
773         if (ret)
774                 return true;
775
776         /*
777          * Can we decide that decrement of refcount is all we needed without
778          * taking the lock?  There's a very common case when it's all we need -
779          * dentry looks like it ought to be retained and there's nothing else
780          * to do.
781          */
782         if (retain_dentry(dentry, false))
783                 return true;
784
785         /*
786          * Either not worth retaining or we can't tell without the lock.
787          * Get the lock, then.  We've already decremented the refcount to 0,
788          * but we'll need to re-check the situation after getting the lock.
789          */
790         spin_lock(&dentry->d_lock);
791
792         /*
793          * Did somebody else grab a reference to it in the meantime, and
794          * we're no longer the last user after all? Alternatively, somebody
795          * else could have killed it and marked it dead. Either way, we
796          * don't need to do anything else.
797          */
798 locked:
799         if (dentry->d_lockref.count || retain_dentry(dentry, true)) {
800                 spin_unlock(&dentry->d_lock);
801                 return true;
802         }
803         return false;
804 }
805
806
807 /* 
808  * This is dput
809  *
810  * This is complicated by the fact that we do not want to put
811  * dentries that are no longer on any hash chain on the unused
812  * list: we'd much rather just get rid of them immediately.
813  *
814  * However, that implies that we have to traverse the dentry
815  * tree upwards to the parents which might _also_ now be
816  * scheduled for deletion (it may have been only waiting for
817  * its last child to go away).
818  *
819  * This tail recursion is done by hand as we don't want to depend
820  * on the compiler to always get this right (gcc generally doesn't).
821  * Real recursion would eat up our stack space.
822  */
823
824 /*
825  * dput - release a dentry
826  * @dentry: dentry to release 
827  *
828  * Release a dentry. This will drop the usage count and if appropriate
829  * call the dentry unlink method as well as removing it from the queues and
830  * releasing its resources. If the parent dentries were scheduled for release
831  * they too may now get deleted.
832  */
833 void dput(struct dentry *dentry)
834 {
835         if (!dentry)
836                 return;
837         might_sleep();
838         rcu_read_lock();
839         if (likely(fast_dput(dentry))) {
840                 rcu_read_unlock();
841                 return;
842         }
843         while (lock_for_kill(dentry)) {
844                 rcu_read_unlock();
845                 dentry = __dentry_kill(dentry);
846                 if (!dentry)
847                         return;
848                 if (retain_dentry(dentry, true)) {
849                         spin_unlock(&dentry->d_lock);
850                         return;
851                 }
852                 rcu_read_lock();
853         }
854         rcu_read_unlock();
855         spin_unlock(&dentry->d_lock);
856 }
857 EXPORT_SYMBOL(dput);
858
859 static void to_shrink_list(struct dentry *dentry, struct list_head *list)
860 __must_hold(&dentry->d_lock)
861 {
862         if (!(dentry->d_flags & DCACHE_SHRINK_LIST)) {
863                 if (dentry->d_flags & DCACHE_LRU_LIST)
864                         d_lru_del(dentry);
865                 d_shrink_add(dentry, list);
866         }
867 }
868
869 void dput_to_list(struct dentry *dentry, struct list_head *list)
870 {
871         rcu_read_lock();
872         if (likely(fast_dput(dentry))) {
873                 rcu_read_unlock();
874                 return;
875         }
876         rcu_read_unlock();
877         to_shrink_list(dentry, list);
878         spin_unlock(&dentry->d_lock);
879 }
880
881 struct dentry *dget_parent(struct dentry *dentry)
882 {
883         int gotref;
884         struct dentry *ret;
885         unsigned seq;
886
887         /*
888          * Do optimistic parent lookup without any
889          * locking.
890          */
891         rcu_read_lock();
892         seq = raw_seqcount_begin(&dentry->d_seq);
893         ret = READ_ONCE(dentry->d_parent);
894         gotref = lockref_get_not_zero(&ret->d_lockref);
895         rcu_read_unlock();
896         if (likely(gotref)) {
897                 if (!read_seqcount_retry(&dentry->d_seq, seq))
898                         return ret;
899                 dput(ret);
900         }
901
902 repeat:
903         /*
904          * Don't need rcu_dereference because we re-check it was correct under
905          * the lock.
906          */
907         rcu_read_lock();
908         ret = dentry->d_parent;
909         spin_lock(&ret->d_lock);
910         if (unlikely(ret != dentry->d_parent)) {
911                 spin_unlock(&ret->d_lock);
912                 rcu_read_unlock();
913                 goto repeat;
914         }
915         rcu_read_unlock();
916         BUG_ON(!ret->d_lockref.count);
917         ret->d_lockref.count++;
918         spin_unlock(&ret->d_lock);
919         return ret;
920 }
921 EXPORT_SYMBOL(dget_parent);
922
923 static struct dentry * __d_find_any_alias(struct inode *inode)
924 {
925         struct dentry *alias;
926
927         if (hlist_empty(&inode->i_dentry))
928                 return NULL;
929         alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
930         lockref_get(&alias->d_lockref);
931         return alias;
932 }
933
934 /**
935  * d_find_any_alias - find any alias for a given inode
936  * @inode: inode to find an alias for
937  *
938  * If any aliases exist for the given inode, take and return a
939  * reference for one of them.  If no aliases exist, return %NULL.
940  */
941 struct dentry *d_find_any_alias(struct inode *inode)
942 {
943         struct dentry *de;
944
945         spin_lock(&inode->i_lock);
946         de = __d_find_any_alias(inode);
947         spin_unlock(&inode->i_lock);
948         return de;
949 }
950 EXPORT_SYMBOL(d_find_any_alias);
951
952 static struct dentry *__d_find_alias(struct inode *inode)
953 {
954         struct dentry *alias;
955
956         if (S_ISDIR(inode->i_mode))
957                 return __d_find_any_alias(inode);
958
959         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
960                 spin_lock(&alias->d_lock);
961                 if (!d_unhashed(alias)) {
962                         dget_dlock(alias);
963                         spin_unlock(&alias->d_lock);
964                         return alias;
965                 }
966                 spin_unlock(&alias->d_lock);
967         }
968         return NULL;
969 }
970
971 /**
972  * d_find_alias - grab a hashed alias of inode
973  * @inode: inode in question
974  *
975  * If inode has a hashed alias, or is a directory and has any alias,
976  * acquire the reference to alias and return it. Otherwise return NULL.
977  * Notice that if inode is a directory there can be only one alias and
978  * it can be unhashed only if it has no children, or if it is the root
979  * of a filesystem, or if the directory was renamed and d_revalidate
980  * was the first vfs operation to notice.
981  *
982  * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
983  * any other hashed alias over that one.
984  */
985 struct dentry *d_find_alias(struct inode *inode)
986 {
987         struct dentry *de = NULL;
988
989         if (!hlist_empty(&inode->i_dentry)) {
990                 spin_lock(&inode->i_lock);
991                 de = __d_find_alias(inode);
992                 spin_unlock(&inode->i_lock);
993         }
994         return de;
995 }
996 EXPORT_SYMBOL(d_find_alias);
997
998 /*
999  *  Caller MUST be holding rcu_read_lock() and be guaranteed
1000  *  that inode won't get freed until rcu_read_unlock().
1001  */
1002 struct dentry *d_find_alias_rcu(struct inode *inode)
1003 {
1004         struct hlist_head *l = &inode->i_dentry;
1005         struct dentry *de = NULL;
1006
1007         spin_lock(&inode->i_lock);
1008         // ->i_dentry and ->i_rcu are colocated, but the latter won't be
1009         // used without having I_FREEING set, which means no aliases left
1010         if (likely(!(inode->i_state & I_FREEING) && !hlist_empty(l))) {
1011                 if (S_ISDIR(inode->i_mode)) {
1012                         de = hlist_entry(l->first, struct dentry, d_u.d_alias);
1013                 } else {
1014                         hlist_for_each_entry(de, l, d_u.d_alias)
1015                                 if (!d_unhashed(de))
1016                                         break;
1017                 }
1018         }
1019         spin_unlock(&inode->i_lock);
1020         return de;
1021 }
1022
1023 /*
1024  *      Try to kill dentries associated with this inode.
1025  * WARNING: you must own a reference to inode.
1026  */
1027 void d_prune_aliases(struct inode *inode)
1028 {
1029         LIST_HEAD(dispose);
1030         struct dentry *dentry;
1031
1032         spin_lock(&inode->i_lock);
1033         hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
1034                 spin_lock(&dentry->d_lock);
1035                 if (!dentry->d_lockref.count)
1036                         to_shrink_list(dentry, &dispose);
1037                 spin_unlock(&dentry->d_lock);
1038         }
1039         spin_unlock(&inode->i_lock);
1040         shrink_dentry_list(&dispose);
1041 }
1042 EXPORT_SYMBOL(d_prune_aliases);
1043
1044 static inline void shrink_kill(struct dentry *victim)
1045 {
1046         do {
1047                 rcu_read_unlock();
1048                 victim = __dentry_kill(victim);
1049                 rcu_read_lock();
1050         } while (victim && lock_for_kill(victim));
1051         rcu_read_unlock();
1052         if (victim)
1053                 spin_unlock(&victim->d_lock);
1054 }
1055
1056 void shrink_dentry_list(struct list_head *list)
1057 {
1058         while (!list_empty(list)) {
1059                 struct dentry *dentry;
1060
1061                 dentry = list_entry(list->prev, struct dentry, d_lru);
1062                 spin_lock(&dentry->d_lock);
1063                 rcu_read_lock();
1064                 if (!lock_for_kill(dentry)) {
1065                         bool can_free;
1066                         rcu_read_unlock();
1067                         d_shrink_del(dentry);
1068                         can_free = dentry->d_flags & DCACHE_DENTRY_KILLED;
1069                         spin_unlock(&dentry->d_lock);
1070                         if (can_free)
1071                                 dentry_free(dentry);
1072                         continue;
1073                 }
1074                 d_shrink_del(dentry);
1075                 shrink_kill(dentry);
1076         }
1077 }
1078
1079 static enum lru_status dentry_lru_isolate(struct list_head *item,
1080                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1081 {
1082         struct list_head *freeable = arg;
1083         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1084
1085
1086         /*
1087          * we are inverting the lru lock/dentry->d_lock here,
1088          * so use a trylock. If we fail to get the lock, just skip
1089          * it
1090          */
1091         if (!spin_trylock(&dentry->d_lock))
1092                 return LRU_SKIP;
1093
1094         /*
1095          * Referenced dentries are still in use. If they have active
1096          * counts, just remove them from the LRU. Otherwise give them
1097          * another pass through the LRU.
1098          */
1099         if (dentry->d_lockref.count) {
1100                 d_lru_isolate(lru, dentry);
1101                 spin_unlock(&dentry->d_lock);
1102                 return LRU_REMOVED;
1103         }
1104
1105         if (dentry->d_flags & DCACHE_REFERENCED) {
1106                 dentry->d_flags &= ~DCACHE_REFERENCED;
1107                 spin_unlock(&dentry->d_lock);
1108
1109                 /*
1110                  * The list move itself will be made by the common LRU code. At
1111                  * this point, we've dropped the dentry->d_lock but keep the
1112                  * lru lock. This is safe to do, since every list movement is
1113                  * protected by the lru lock even if both locks are held.
1114                  *
1115                  * This is guaranteed by the fact that all LRU management
1116                  * functions are intermediated by the LRU API calls like
1117                  * list_lru_add_obj and list_lru_del_obj. List movement in this file
1118                  * only ever occur through this functions or through callbacks
1119                  * like this one, that are called from the LRU API.
1120                  *
1121                  * The only exceptions to this are functions like
1122                  * shrink_dentry_list, and code that first checks for the
1123                  * DCACHE_SHRINK_LIST flag.  Those are guaranteed to be
1124                  * operating only with stack provided lists after they are
1125                  * properly isolated from the main list.  It is thus, always a
1126                  * local access.
1127                  */
1128                 return LRU_ROTATE;
1129         }
1130
1131         d_lru_shrink_move(lru, dentry, freeable);
1132         spin_unlock(&dentry->d_lock);
1133
1134         return LRU_REMOVED;
1135 }
1136
1137 /**
1138  * prune_dcache_sb - shrink the dcache
1139  * @sb: superblock
1140  * @sc: shrink control, passed to list_lru_shrink_walk()
1141  *
1142  * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1143  * is done when we need more memory and called from the superblock shrinker
1144  * function.
1145  *
1146  * This function may fail to free any resources if all the dentries are in
1147  * use.
1148  */
1149 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1150 {
1151         LIST_HEAD(dispose);
1152         long freed;
1153
1154         freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1155                                      dentry_lru_isolate, &dispose);
1156         shrink_dentry_list(&dispose);
1157         return freed;
1158 }
1159
1160 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1161                 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1162 {
1163         struct list_head *freeable = arg;
1164         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1165
1166         /*
1167          * we are inverting the lru lock/dentry->d_lock here,
1168          * so use a trylock. If we fail to get the lock, just skip
1169          * it
1170          */
1171         if (!spin_trylock(&dentry->d_lock))
1172                 return LRU_SKIP;
1173
1174         d_lru_shrink_move(lru, dentry, freeable);
1175         spin_unlock(&dentry->d_lock);
1176
1177         return LRU_REMOVED;
1178 }
1179
1180
1181 /**
1182  * shrink_dcache_sb - shrink dcache for a superblock
1183  * @sb: superblock
1184  *
1185  * Shrink the dcache for the specified super block. This is used to free
1186  * the dcache before unmounting a file system.
1187  */
1188 void shrink_dcache_sb(struct super_block *sb)
1189 {
1190         do {
1191                 LIST_HEAD(dispose);
1192
1193                 list_lru_walk(&sb->s_dentry_lru,
1194                         dentry_lru_isolate_shrink, &dispose, 1024);
1195                 shrink_dentry_list(&dispose);
1196         } while (list_lru_count(&sb->s_dentry_lru) > 0);
1197 }
1198 EXPORT_SYMBOL(shrink_dcache_sb);
1199
1200 /**
1201  * enum d_walk_ret - action to talke during tree walk
1202  * @D_WALK_CONTINUE:    contrinue walk
1203  * @D_WALK_QUIT:        quit walk
1204  * @D_WALK_NORETRY:     quit when retry is needed
1205  * @D_WALK_SKIP:        skip this dentry and its children
1206  */
1207 enum d_walk_ret {
1208         D_WALK_CONTINUE,
1209         D_WALK_QUIT,
1210         D_WALK_NORETRY,
1211         D_WALK_SKIP,
1212 };
1213
1214 /**
1215  * d_walk - walk the dentry tree
1216  * @parent:     start of walk
1217  * @data:       data passed to @enter() and @finish()
1218  * @enter:      callback when first entering the dentry
1219  *
1220  * The @enter() callbacks are called with d_lock held.
1221  */
1222 static void d_walk(struct dentry *parent, void *data,
1223                    enum d_walk_ret (*enter)(void *, struct dentry *))
1224 {
1225         struct dentry *this_parent, *dentry;
1226         unsigned seq = 0;
1227         enum d_walk_ret ret;
1228         bool retry = true;
1229
1230 again:
1231         read_seqbegin_or_lock(&rename_lock, &seq);
1232         this_parent = parent;
1233         spin_lock(&this_parent->d_lock);
1234
1235         ret = enter(data, this_parent);
1236         switch (ret) {
1237         case D_WALK_CONTINUE:
1238                 break;
1239         case D_WALK_QUIT:
1240         case D_WALK_SKIP:
1241                 goto out_unlock;
1242         case D_WALK_NORETRY:
1243                 retry = false;
1244                 break;
1245         }
1246 repeat:
1247         dentry = d_first_child(this_parent);
1248 resume:
1249         hlist_for_each_entry_from(dentry, d_sib) {
1250                 if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
1251                         continue;
1252
1253                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1254
1255                 ret = enter(data, dentry);
1256                 switch (ret) {
1257                 case D_WALK_CONTINUE:
1258                         break;
1259                 case D_WALK_QUIT:
1260                         spin_unlock(&dentry->d_lock);
1261                         goto out_unlock;
1262                 case D_WALK_NORETRY:
1263                         retry = false;
1264                         break;
1265                 case D_WALK_SKIP:
1266                         spin_unlock(&dentry->d_lock);
1267                         continue;
1268                 }
1269
1270                 if (!hlist_empty(&dentry->d_children)) {
1271                         spin_unlock(&this_parent->d_lock);
1272                         spin_release(&dentry->d_lock.dep_map, _RET_IP_);
1273                         this_parent = dentry;
1274                         spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1275                         goto repeat;
1276                 }
1277                 spin_unlock(&dentry->d_lock);
1278         }
1279         /*
1280          * All done at this level ... ascend and resume the search.
1281          */
1282         rcu_read_lock();
1283 ascend:
1284         if (this_parent != parent) {
1285                 dentry = this_parent;
1286                 this_parent = dentry->d_parent;
1287
1288                 spin_unlock(&dentry->d_lock);
1289                 spin_lock(&this_parent->d_lock);
1290
1291                 /* might go back up the wrong parent if we have had a rename. */
1292                 if (need_seqretry(&rename_lock, seq))
1293                         goto rename_retry;
1294                 /* go into the first sibling still alive */
1295                 hlist_for_each_entry_continue(dentry, d_sib) {
1296                         if (likely(!(dentry->d_flags & DCACHE_DENTRY_KILLED))) {
1297                                 rcu_read_unlock();
1298                                 goto resume;
1299                         }
1300                 }
1301                 goto ascend;
1302         }
1303         if (need_seqretry(&rename_lock, seq))
1304                 goto rename_retry;
1305         rcu_read_unlock();
1306
1307 out_unlock:
1308         spin_unlock(&this_parent->d_lock);
1309         done_seqretry(&rename_lock, seq);
1310         return;
1311
1312 rename_retry:
1313         spin_unlock(&this_parent->d_lock);
1314         rcu_read_unlock();
1315         BUG_ON(seq & 1);
1316         if (!retry)
1317                 return;
1318         seq = 1;
1319         goto again;
1320 }
1321
1322 struct check_mount {
1323         struct vfsmount *mnt;
1324         unsigned int mounted;
1325 };
1326
1327 static enum d_walk_ret path_check_mount(void *data, struct dentry *dentry)
1328 {
1329         struct check_mount *info = data;
1330         struct path path = { .mnt = info->mnt, .dentry = dentry };
1331
1332         if (likely(!d_mountpoint(dentry)))
1333                 return D_WALK_CONTINUE;
1334         if (__path_is_mountpoint(&path)) {
1335                 info->mounted = 1;
1336                 return D_WALK_QUIT;
1337         }
1338         return D_WALK_CONTINUE;
1339 }
1340
1341 /**
1342  * path_has_submounts - check for mounts over a dentry in the
1343  *                      current namespace.
1344  * @parent: path to check.
1345  *
1346  * Return true if the parent or its subdirectories contain
1347  * a mount point in the current namespace.
1348  */
1349 int path_has_submounts(const struct path *parent)
1350 {
1351         struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };
1352
1353         read_seqlock_excl(&mount_lock);
1354         d_walk(parent->dentry, &data, path_check_mount);
1355         read_sequnlock_excl(&mount_lock);
1356
1357         return data.mounted;
1358 }
1359 EXPORT_SYMBOL(path_has_submounts);
1360
1361 /*
1362  * Called by mount code to set a mountpoint and check if the mountpoint is
1363  * reachable (e.g. NFS can unhash a directory dentry and then the complete
1364  * subtree can become unreachable).
1365  *
1366  * Only one of d_invalidate() and d_set_mounted() must succeed.  For
1367  * this reason take rename_lock and d_lock on dentry and ancestors.
1368  */
1369 int d_set_mounted(struct dentry *dentry)
1370 {
1371         struct dentry *p;
1372         int ret = -ENOENT;
1373         write_seqlock(&rename_lock);
1374         for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1375                 /* Need exclusion wrt. d_invalidate() */
1376                 spin_lock(&p->d_lock);
1377                 if (unlikely(d_unhashed(p))) {
1378                         spin_unlock(&p->d_lock);
1379                         goto out;
1380                 }
1381                 spin_unlock(&p->d_lock);
1382         }
1383         spin_lock(&dentry->d_lock);
1384         if (!d_unlinked(dentry)) {
1385                 ret = -EBUSY;
1386                 if (!d_mountpoint(dentry)) {
1387                         dentry->d_flags |= DCACHE_MOUNTED;
1388                         ret = 0;
1389                 }
1390         }
1391         spin_unlock(&dentry->d_lock);
1392 out:
1393         write_sequnlock(&rename_lock);
1394         return ret;
1395 }
1396
1397 /*
1398  * Search the dentry child list of the specified parent,
1399  * and move any unused dentries to the end of the unused
1400  * list for prune_dcache(). We descend to the next level
1401  * whenever the d_children list is non-empty and continue
1402  * searching.
1403  *
1404  * It returns zero iff there are no unused children,
1405  * otherwise  it returns the number of children moved to
1406  * the end of the unused list. This may not be the total
1407  * number of unused children, because select_parent can
1408  * drop the lock and return early due to latency
1409  * constraints.
1410  */
1411
1412 struct select_data {
1413         struct dentry *start;
1414         union {
1415                 long found;
1416                 struct dentry *victim;
1417         };
1418         struct list_head dispose;
1419 };
1420
1421 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1422 {
1423         struct select_data *data = _data;
1424         enum d_walk_ret ret = D_WALK_CONTINUE;
1425
1426         if (data->start == dentry)
1427                 goto out;
1428
1429         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1430                 data->found++;
1431         } else if (!dentry->d_lockref.count) {
1432                 to_shrink_list(dentry, &data->dispose);
1433                 data->found++;
1434         } else if (dentry->d_lockref.count < 0) {
1435                 data->found++;
1436         }
1437         /*
1438          * We can return to the caller if we have found some (this
1439          * ensures forward progress). We'll be coming back to find
1440          * the rest.
1441          */
1442         if (!list_empty(&data->dispose))
1443                 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1444 out:
1445         return ret;
1446 }
1447
1448 static enum d_walk_ret select_collect2(void *_data, struct dentry *dentry)
1449 {
1450         struct select_data *data = _data;
1451         enum d_walk_ret ret = D_WALK_CONTINUE;
1452
1453         if (data->start == dentry)
1454                 goto out;
1455
1456         if (!dentry->d_lockref.count) {
1457                 if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1458                         rcu_read_lock();
1459                         data->victim = dentry;
1460                         return D_WALK_QUIT;
1461                 }
1462                 to_shrink_list(dentry, &data->dispose);
1463         }
1464         /*
1465          * We can return to the caller if we have found some (this
1466          * ensures forward progress). We'll be coming back to find
1467          * the rest.
1468          */
1469         if (!list_empty(&data->dispose))
1470                 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1471 out:
1472         return ret;
1473 }
1474
1475 /**
1476  * shrink_dcache_parent - prune dcache
1477  * @parent: parent of entries to prune
1478  *
1479  * Prune the dcache to remove unused children of the parent dentry.
1480  */
1481 void shrink_dcache_parent(struct dentry *parent)
1482 {
1483         for (;;) {
1484                 struct select_data data = {.start = parent};
1485
1486                 INIT_LIST_HEAD(&data.dispose);
1487                 d_walk(parent, &data, select_collect);
1488
1489                 if (!list_empty(&data.dispose)) {
1490                         shrink_dentry_list(&data.dispose);
1491                         continue;
1492                 }
1493
1494                 cond_resched();
1495                 if (!data.found)
1496                         break;
1497                 data.victim = NULL;
1498                 d_walk(parent, &data, select_collect2);
1499                 if (data.victim) {
1500                         spin_lock(&data.victim->d_lock);
1501                         if (!lock_for_kill(data.victim)) {
1502                                 spin_unlock(&data.victim->d_lock);
1503                                 rcu_read_unlock();
1504                         } else {
1505                                 shrink_kill(data.victim);
1506                         }
1507                 }
1508                 if (!list_empty(&data.dispose))
1509                         shrink_dentry_list(&data.dispose);
1510         }
1511 }
1512 EXPORT_SYMBOL(shrink_dcache_parent);
1513
1514 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1515 {
1516         /* it has busy descendents; complain about those instead */
1517         if (!hlist_empty(&dentry->d_children))
1518                 return D_WALK_CONTINUE;
1519
1520         /* root with refcount 1 is fine */
1521         if (dentry == _data && dentry->d_lockref.count == 1)
1522                 return D_WALK_CONTINUE;
1523
1524         WARN(1, "BUG: Dentry %p{i=%lx,n=%pd} "
1525                         " still in use (%d) [unmount of %s %s]\n",
1526                        dentry,
1527                        dentry->d_inode ?
1528                        dentry->d_inode->i_ino : 0UL,
1529                        dentry,
1530                        dentry->d_lockref.count,
1531                        dentry->d_sb->s_type->name,
1532                        dentry->d_sb->s_id);
1533         return D_WALK_CONTINUE;
1534 }
1535
1536 static void do_one_tree(struct dentry *dentry)
1537 {
1538         shrink_dcache_parent(dentry);
1539         d_walk(dentry, dentry, umount_check);
1540         d_drop(dentry);
1541         dput(dentry);
1542 }
1543
1544 /*
1545  * destroy the dentries attached to a superblock on unmounting
1546  */
1547 void shrink_dcache_for_umount(struct super_block *sb)
1548 {
1549         struct dentry *dentry;
1550
1551         WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
1552
1553         dentry = sb->s_root;
1554         sb->s_root = NULL;
1555         do_one_tree(dentry);
1556
1557         while (!hlist_bl_empty(&sb->s_roots)) {
1558                 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_roots), struct dentry, d_hash));
1559                 do_one_tree(dentry);
1560         }
1561 }
1562
1563 static enum d_walk_ret find_submount(void *_data, struct dentry *dentry)
1564 {
1565         struct dentry **victim = _data;
1566         if (d_mountpoint(dentry)) {
1567                 *victim = dget_dlock(dentry);
1568                 return D_WALK_QUIT;
1569         }
1570         return D_WALK_CONTINUE;
1571 }
1572
1573 /**
1574  * d_invalidate - detach submounts, prune dcache, and drop
1575  * @dentry: dentry to invalidate (aka detach, prune and drop)
1576  */
1577 void d_invalidate(struct dentry *dentry)
1578 {
1579         bool had_submounts = false;
1580         spin_lock(&dentry->d_lock);
1581         if (d_unhashed(dentry)) {
1582                 spin_unlock(&dentry->d_lock);
1583                 return;
1584         }
1585         __d_drop(dentry);
1586         spin_unlock(&dentry->d_lock);
1587
1588         /* Negative dentries can be dropped without further checks */
1589         if (!dentry->d_inode)
1590                 return;
1591
1592         shrink_dcache_parent(dentry);
1593         for (;;) {
1594                 struct dentry *victim = NULL;
1595                 d_walk(dentry, &victim, find_submount);
1596                 if (!victim) {
1597                         if (had_submounts)
1598                                 shrink_dcache_parent(dentry);
1599                         return;
1600                 }
1601                 had_submounts = true;
1602                 detach_mounts(victim);
1603                 dput(victim);
1604         }
1605 }
1606 EXPORT_SYMBOL(d_invalidate);
1607
1608 /**
1609  * __d_alloc    -       allocate a dcache entry
1610  * @sb: filesystem it will belong to
1611  * @name: qstr of the name
1612  *
1613  * Allocates a dentry. It returns %NULL if there is insufficient memory
1614  * available. On a success the dentry is returned. The name passed in is
1615  * copied and the copy passed in may be reused after this call.
1616  */
1617  
1618 static struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1619 {
1620         struct dentry *dentry;
1621         char *dname;
1622         int err;
1623
1624         dentry = kmem_cache_alloc_lru(dentry_cache, &sb->s_dentry_lru,
1625                                       GFP_KERNEL);
1626         if (!dentry)
1627                 return NULL;
1628
1629         /*
1630          * We guarantee that the inline name is always NUL-terminated.
1631          * This way the memcpy() done by the name switching in rename
1632          * will still always have a NUL at the end, even if we might
1633          * be overwriting an internal NUL character
1634          */
1635         dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1636         if (unlikely(!name)) {
1637                 name = &slash_name;
1638                 dname = dentry->d_iname;
1639         } else if (name->len > DNAME_INLINE_LEN-1) {
1640                 size_t size = offsetof(struct external_name, name[1]);
1641                 struct external_name *p = kmalloc(size + name->len,
1642                                                   GFP_KERNEL_ACCOUNT |
1643                                                   __GFP_RECLAIMABLE);
1644                 if (!p) {
1645                         kmem_cache_free(dentry_cache, dentry); 
1646                         return NULL;
1647                 }
1648                 atomic_set(&p->u.count, 1);
1649                 dname = p->name;
1650         } else  {
1651                 dname = dentry->d_iname;
1652         }       
1653
1654         dentry->d_name.len = name->len;
1655         dentry->d_name.hash = name->hash;
1656         memcpy(dname, name->name, name->len);
1657         dname[name->len] = 0;
1658
1659         /* Make sure we always see the terminating NUL character */
1660         smp_store_release(&dentry->d_name.name, dname); /* ^^^ */
1661
1662         dentry->d_lockref.count = 1;
1663         dentry->d_flags = 0;
1664         spin_lock_init(&dentry->d_lock);
1665         seqcount_spinlock_init(&dentry->d_seq, &dentry->d_lock);
1666         dentry->d_inode = NULL;
1667         dentry->d_parent = dentry;
1668         dentry->d_sb = sb;
1669         dentry->d_op = NULL;
1670         dentry->d_fsdata = NULL;
1671         INIT_HLIST_BL_NODE(&dentry->d_hash);
1672         INIT_LIST_HEAD(&dentry->d_lru);
1673         INIT_HLIST_HEAD(&dentry->d_children);
1674         INIT_HLIST_NODE(&dentry->d_u.d_alias);
1675         INIT_HLIST_NODE(&dentry->d_sib);
1676         d_set_d_op(dentry, dentry->d_sb->s_d_op);
1677
1678         if (dentry->d_op && dentry->d_op->d_init) {
1679                 err = dentry->d_op->d_init(dentry);
1680                 if (err) {
1681                         if (dname_external(dentry))
1682                                 kfree(external_name(dentry));
1683                         kmem_cache_free(dentry_cache, dentry);
1684                         return NULL;
1685                 }
1686         }
1687
1688         this_cpu_inc(nr_dentry);
1689
1690         return dentry;
1691 }
1692
1693 /**
1694  * d_alloc      -       allocate a dcache entry
1695  * @parent: parent of entry to allocate
1696  * @name: qstr of the name
1697  *
1698  * Allocates a dentry. It returns %NULL if there is insufficient memory
1699  * available. On a success the dentry is returned. The name passed in is
1700  * copied and the copy passed in may be reused after this call.
1701  */
1702 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1703 {
1704         struct dentry *dentry = __d_alloc(parent->d_sb, name);
1705         if (!dentry)
1706                 return NULL;
1707         spin_lock(&parent->d_lock);
1708         /*
1709          * don't need child lock because it is not subject
1710          * to concurrency here
1711          */
1712         dentry->d_parent = dget_dlock(parent);
1713         hlist_add_head(&dentry->d_sib, &parent->d_children);
1714         spin_unlock(&parent->d_lock);
1715
1716         return dentry;
1717 }
1718 EXPORT_SYMBOL(d_alloc);
1719
1720 struct dentry *d_alloc_anon(struct super_block *sb)
1721 {
1722         return __d_alloc(sb, NULL);
1723 }
1724 EXPORT_SYMBOL(d_alloc_anon);
1725
1726 struct dentry *d_alloc_cursor(struct dentry * parent)
1727 {
1728         struct dentry *dentry = d_alloc_anon(parent->d_sb);
1729         if (dentry) {
1730                 dentry->d_flags |= DCACHE_DENTRY_CURSOR;
1731                 dentry->d_parent = dget(parent);
1732         }
1733         return dentry;
1734 }
1735
1736 /**
1737  * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1738  * @sb: the superblock
1739  * @name: qstr of the name
1740  *
1741  * For a filesystem that just pins its dentries in memory and never
1742  * performs lookups at all, return an unhashed IS_ROOT dentry.
1743  * This is used for pipes, sockets et.al. - the stuff that should
1744  * never be anyone's children or parents.  Unlike all other
1745  * dentries, these will not have RCU delay between dropping the
1746  * last reference and freeing them.
1747  *
1748  * The only user is alloc_file_pseudo() and that's what should
1749  * be considered a public interface.  Don't use directly.
1750  */
1751 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1752 {
1753         static const struct dentry_operations anon_ops = {
1754                 .d_dname = simple_dname
1755         };
1756         struct dentry *dentry = __d_alloc(sb, name);
1757         if (likely(dentry)) {
1758                 dentry->d_flags |= DCACHE_NORCU;
1759                 if (!sb->s_d_op)
1760                         d_set_d_op(dentry, &anon_ops);
1761         }
1762         return dentry;
1763 }
1764
1765 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1766 {
1767         struct qstr q;
1768
1769         q.name = name;
1770         q.hash_len = hashlen_string(parent, name);
1771         return d_alloc(parent, &q);
1772 }
1773 EXPORT_SYMBOL(d_alloc_name);
1774
1775 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1776 {
1777         WARN_ON_ONCE(dentry->d_op);
1778         WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH  |
1779                                 DCACHE_OP_COMPARE       |
1780                                 DCACHE_OP_REVALIDATE    |
1781                                 DCACHE_OP_WEAK_REVALIDATE       |
1782                                 DCACHE_OP_DELETE        |
1783                                 DCACHE_OP_REAL));
1784         dentry->d_op = op;
1785         if (!op)
1786                 return;
1787         if (op->d_hash)
1788                 dentry->d_flags |= DCACHE_OP_HASH;
1789         if (op->d_compare)
1790                 dentry->d_flags |= DCACHE_OP_COMPARE;
1791         if (op->d_revalidate)
1792                 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1793         if (op->d_weak_revalidate)
1794                 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1795         if (op->d_delete)
1796                 dentry->d_flags |= DCACHE_OP_DELETE;
1797         if (op->d_prune)
1798                 dentry->d_flags |= DCACHE_OP_PRUNE;
1799         if (op->d_real)
1800                 dentry->d_flags |= DCACHE_OP_REAL;
1801
1802 }
1803 EXPORT_SYMBOL(d_set_d_op);
1804
1805 static unsigned d_flags_for_inode(struct inode *inode)
1806 {
1807         unsigned add_flags = DCACHE_REGULAR_TYPE;
1808
1809         if (!inode)
1810                 return DCACHE_MISS_TYPE;
1811
1812         if (S_ISDIR(inode->i_mode)) {
1813                 add_flags = DCACHE_DIRECTORY_TYPE;
1814                 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1815                         if (unlikely(!inode->i_op->lookup))
1816                                 add_flags = DCACHE_AUTODIR_TYPE;
1817                         else
1818                                 inode->i_opflags |= IOP_LOOKUP;
1819                 }
1820                 goto type_determined;
1821         }
1822
1823         if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1824                 if (unlikely(inode->i_op->get_link)) {
1825                         add_flags = DCACHE_SYMLINK_TYPE;
1826                         goto type_determined;
1827                 }
1828                 inode->i_opflags |= IOP_NOFOLLOW;
1829         }
1830
1831         if (unlikely(!S_ISREG(inode->i_mode)))
1832                 add_flags = DCACHE_SPECIAL_TYPE;
1833
1834 type_determined:
1835         if (unlikely(IS_AUTOMOUNT(inode)))
1836                 add_flags |= DCACHE_NEED_AUTOMOUNT;
1837         return add_flags;
1838 }
1839
1840 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1841 {
1842         unsigned add_flags = d_flags_for_inode(inode);
1843         WARN_ON(d_in_lookup(dentry));
1844
1845         spin_lock(&dentry->d_lock);
1846         /*
1847          * Decrement negative dentry count if it was in the LRU list.
1848          */
1849         if (dentry->d_flags & DCACHE_LRU_LIST)
1850                 this_cpu_dec(nr_dentry_negative);
1851         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1852         raw_write_seqcount_begin(&dentry->d_seq);
1853         __d_set_inode_and_type(dentry, inode, add_flags);
1854         raw_write_seqcount_end(&dentry->d_seq);
1855         fsnotify_update_flags(dentry);
1856         spin_unlock(&dentry->d_lock);
1857 }
1858
1859 /**
1860  * d_instantiate - fill in inode information for a dentry
1861  * @entry: dentry to complete
1862  * @inode: inode to attach to this dentry
1863  *
1864  * Fill in inode information in the entry.
1865  *
1866  * This turns negative dentries into productive full members
1867  * of society.
1868  *
1869  * NOTE! This assumes that the inode count has been incremented
1870  * (or otherwise set) by the caller to indicate that it is now
1871  * in use by the dcache.
1872  */
1873  
1874 void d_instantiate(struct dentry *entry, struct inode * inode)
1875 {
1876         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1877         if (inode) {
1878                 security_d_instantiate(entry, inode);
1879                 spin_lock(&inode->i_lock);
1880                 __d_instantiate(entry, inode);
1881                 spin_unlock(&inode->i_lock);
1882         }
1883 }
1884 EXPORT_SYMBOL(d_instantiate);
1885
1886 /*
1887  * This should be equivalent to d_instantiate() + unlock_new_inode(),
1888  * with lockdep-related part of unlock_new_inode() done before
1889  * anything else.  Use that instead of open-coding d_instantiate()/
1890  * unlock_new_inode() combinations.
1891  */
1892 void d_instantiate_new(struct dentry *entry, struct inode *inode)
1893 {
1894         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1895         BUG_ON(!inode);
1896         lockdep_annotate_inode_mutex_key(inode);
1897         security_d_instantiate(entry, inode);
1898         spin_lock(&inode->i_lock);
1899         __d_instantiate(entry, inode);
1900         WARN_ON(!(inode->i_state & I_NEW));
1901         inode->i_state &= ~I_NEW & ~I_CREATING;
1902         smp_mb();
1903         wake_up_bit(&inode->i_state, __I_NEW);
1904         spin_unlock(&inode->i_lock);
1905 }
1906 EXPORT_SYMBOL(d_instantiate_new);
1907
1908 struct dentry *d_make_root(struct inode *root_inode)
1909 {
1910         struct dentry *res = NULL;
1911
1912         if (root_inode) {
1913                 res = d_alloc_anon(root_inode->i_sb);
1914                 if (res)
1915                         d_instantiate(res, root_inode);
1916                 else
1917                         iput(root_inode);
1918         }
1919         return res;
1920 }
1921 EXPORT_SYMBOL(d_make_root);
1922
1923 static struct dentry *__d_obtain_alias(struct inode *inode, bool disconnected)
1924 {
1925         struct super_block *sb;
1926         struct dentry *new, *res;
1927
1928         if (!inode)
1929                 return ERR_PTR(-ESTALE);
1930         if (IS_ERR(inode))
1931                 return ERR_CAST(inode);
1932
1933         sb = inode->i_sb;
1934
1935         res = d_find_any_alias(inode); /* existing alias? */
1936         if (res)
1937                 goto out;
1938
1939         new = d_alloc_anon(sb);
1940         if (!new) {
1941                 res = ERR_PTR(-ENOMEM);
1942                 goto out;
1943         }
1944
1945         security_d_instantiate(new, inode);
1946         spin_lock(&inode->i_lock);
1947         res = __d_find_any_alias(inode); /* recheck under lock */
1948         if (likely(!res)) { /* still no alias, attach a disconnected dentry */
1949                 unsigned add_flags = d_flags_for_inode(inode);
1950
1951                 if (disconnected)
1952                         add_flags |= DCACHE_DISCONNECTED;
1953
1954                 spin_lock(&new->d_lock);
1955                 __d_set_inode_and_type(new, inode, add_flags);
1956                 hlist_add_head(&new->d_u.d_alias, &inode->i_dentry);
1957                 if (!disconnected) {
1958                         hlist_bl_lock(&sb->s_roots);
1959                         hlist_bl_add_head(&new->d_hash, &sb->s_roots);
1960                         hlist_bl_unlock(&sb->s_roots);
1961                 }
1962                 spin_unlock(&new->d_lock);
1963                 spin_unlock(&inode->i_lock);
1964                 inode = NULL; /* consumed by new->d_inode */
1965                 res = new;
1966         } else {
1967                 spin_unlock(&inode->i_lock);
1968                 dput(new);
1969         }
1970
1971  out:
1972         iput(inode);
1973         return res;
1974 }
1975
1976 /**
1977  * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
1978  * @inode: inode to allocate the dentry for
1979  *
1980  * Obtain a dentry for an inode resulting from NFS filehandle conversion or
1981  * similar open by handle operations.  The returned dentry may be anonymous,
1982  * or may have a full name (if the inode was already in the cache).
1983  *
1984  * When called on a directory inode, we must ensure that the inode only ever
1985  * has one dentry.  If a dentry is found, that is returned instead of
1986  * allocating a new one.
1987  *
1988  * On successful return, the reference to the inode has been transferred
1989  * to the dentry.  In case of an error the reference on the inode is released.
1990  * To make it easier to use in export operations a %NULL or IS_ERR inode may
1991  * be passed in and the error will be propagated to the return value,
1992  * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
1993  */
1994 struct dentry *d_obtain_alias(struct inode *inode)
1995 {
1996         return __d_obtain_alias(inode, true);
1997 }
1998 EXPORT_SYMBOL(d_obtain_alias);
1999
2000 /**
2001  * d_obtain_root - find or allocate a dentry for a given inode
2002  * @inode: inode to allocate the dentry for
2003  *
2004  * Obtain an IS_ROOT dentry for the root of a filesystem.
2005  *
2006  * We must ensure that directory inodes only ever have one dentry.  If a
2007  * dentry is found, that is returned instead of allocating a new one.
2008  *
2009  * On successful return, the reference to the inode has been transferred
2010  * to the dentry.  In case of an error the reference on the inode is
2011  * released.  A %NULL or IS_ERR inode may be passed in and will be the
2012  * error will be propagate to the return value, with a %NULL @inode
2013  * replaced by ERR_PTR(-ESTALE).
2014  */
2015 struct dentry *d_obtain_root(struct inode *inode)
2016 {
2017         return __d_obtain_alias(inode, false);
2018 }
2019 EXPORT_SYMBOL(d_obtain_root);
2020
2021 /**
2022  * d_add_ci - lookup or allocate new dentry with case-exact name
2023  * @inode:  the inode case-insensitive lookup has found
2024  * @dentry: the negative dentry that was passed to the parent's lookup func
2025  * @name:   the case-exact name to be associated with the returned dentry
2026  *
2027  * This is to avoid filling the dcache with case-insensitive names to the
2028  * same inode, only the actual correct case is stored in the dcache for
2029  * case-insensitive filesystems.
2030  *
2031  * For a case-insensitive lookup match and if the case-exact dentry
2032  * already exists in the dcache, use it and return it.
2033  *
2034  * If no entry exists with the exact case name, allocate new dentry with
2035  * the exact case, and return the spliced entry.
2036  */
2037 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2038                         struct qstr *name)
2039 {
2040         struct dentry *found, *res;
2041
2042         /*
2043          * First check if a dentry matching the name already exists,
2044          * if not go ahead and create it now.
2045          */
2046         found = d_hash_and_lookup(dentry->d_parent, name);
2047         if (found) {
2048                 iput(inode);
2049                 return found;
2050         }
2051         if (d_in_lookup(dentry)) {
2052                 found = d_alloc_parallel(dentry->d_parent, name,
2053                                         dentry->d_wait);
2054                 if (IS_ERR(found) || !d_in_lookup(found)) {
2055                         iput(inode);
2056                         return found;
2057                 }
2058         } else {
2059                 found = d_alloc(dentry->d_parent, name);
2060                 if (!found) {
2061                         iput(inode);
2062                         return ERR_PTR(-ENOMEM);
2063                 } 
2064         }
2065         res = d_splice_alias(inode, found);
2066         if (res) {
2067                 d_lookup_done(found);
2068                 dput(found);
2069                 return res;
2070         }
2071         return found;
2072 }
2073 EXPORT_SYMBOL(d_add_ci);
2074
2075 /**
2076  * d_same_name - compare dentry name with case-exact name
2077  * @parent: parent dentry
2078  * @dentry: the negative dentry that was passed to the parent's lookup func
2079  * @name:   the case-exact name to be associated with the returned dentry
2080  *
2081  * Return: true if names are same, or false
2082  */
2083 bool d_same_name(const struct dentry *dentry, const struct dentry *parent,
2084                  const struct qstr *name)
2085 {
2086         if (likely(!(parent->d_flags & DCACHE_OP_COMPARE))) {
2087                 if (dentry->d_name.len != name->len)
2088                         return false;
2089                 return dentry_cmp(dentry, name->name, name->len) == 0;
2090         }
2091         return parent->d_op->d_compare(dentry,
2092                                        dentry->d_name.len, dentry->d_name.name,
2093                                        name) == 0;
2094 }
2095 EXPORT_SYMBOL_GPL(d_same_name);
2096
2097 /*
2098  * This is __d_lookup_rcu() when the parent dentry has
2099  * DCACHE_OP_COMPARE, which makes things much nastier.
2100  */
2101 static noinline struct dentry *__d_lookup_rcu_op_compare(
2102         const struct dentry *parent,
2103         const struct qstr *name,
2104         unsigned *seqp)
2105 {
2106         u64 hashlen = name->hash_len;
2107         struct hlist_bl_head *b = d_hash(hashlen_hash(hashlen));
2108         struct hlist_bl_node *node;
2109         struct dentry *dentry;
2110
2111         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2112                 int tlen;
2113                 const char *tname;
2114                 unsigned seq;
2115
2116 seqretry:
2117                 seq = raw_seqcount_begin(&dentry->d_seq);
2118                 if (dentry->d_parent != parent)
2119                         continue;
2120                 if (d_unhashed(dentry))
2121                         continue;
2122                 if (dentry->d_name.hash != hashlen_hash(hashlen))
2123                         continue;
2124                 tlen = dentry->d_name.len;
2125                 tname = dentry->d_name.name;
2126                 /* we want a consistent (name,len) pair */
2127                 if (read_seqcount_retry(&dentry->d_seq, seq)) {
2128                         cpu_relax();
2129                         goto seqretry;
2130                 }
2131                 if (parent->d_op->d_compare(dentry, tlen, tname, name) != 0)
2132                         continue;
2133                 *seqp = seq;
2134                 return dentry;
2135         }
2136         return NULL;
2137 }
2138
2139 /**
2140  * __d_lookup_rcu - search for a dentry (racy, store-free)
2141  * @parent: parent dentry
2142  * @name: qstr of name we wish to find
2143  * @seqp: returns d_seq value at the point where the dentry was found
2144  * Returns: dentry, or NULL
2145  *
2146  * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2147  * resolution (store-free path walking) design described in
2148  * Documentation/filesystems/path-lookup.txt.
2149  *
2150  * This is not to be used outside core vfs.
2151  *
2152  * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2153  * held, and rcu_read_lock held. The returned dentry must not be stored into
2154  * without taking d_lock and checking d_seq sequence count against @seq
2155  * returned here.
2156  *
2157  * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2158  * function.
2159  *
2160  * Alternatively, __d_lookup_rcu may be called again to look up the child of
2161  * the returned dentry, so long as its parent's seqlock is checked after the
2162  * child is looked up. Thus, an interlocking stepping of sequence lock checks
2163  * is formed, giving integrity down the path walk.
2164  *
2165  * NOTE! The caller *has* to check the resulting dentry against the sequence
2166  * number we've returned before using any of the resulting dentry state!
2167  */
2168 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2169                                 const struct qstr *name,
2170                                 unsigned *seqp)
2171 {
2172         u64 hashlen = name->hash_len;
2173         const unsigned char *str = name->name;
2174         struct hlist_bl_head *b = d_hash(hashlen_hash(hashlen));
2175         struct hlist_bl_node *node;
2176         struct dentry *dentry;
2177
2178         /*
2179          * Note: There is significant duplication with __d_lookup_rcu which is
2180          * required to prevent single threaded performance regressions
2181          * especially on architectures where smp_rmb (in seqcounts) are costly.
2182          * Keep the two functions in sync.
2183          */
2184
2185         if (unlikely(parent->d_flags & DCACHE_OP_COMPARE))
2186                 return __d_lookup_rcu_op_compare(parent, name, seqp);
2187
2188         /*
2189          * The hash list is protected using RCU.
2190          *
2191          * Carefully use d_seq when comparing a candidate dentry, to avoid
2192          * races with d_move().
2193          *
2194          * It is possible that concurrent renames can mess up our list
2195          * walk here and result in missing our dentry, resulting in the
2196          * false-negative result. d_lookup() protects against concurrent
2197          * renames using rename_lock seqlock.
2198          *
2199          * See Documentation/filesystems/path-lookup.txt for more details.
2200          */
2201         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2202                 unsigned seq;
2203
2204                 /*
2205                  * The dentry sequence count protects us from concurrent
2206                  * renames, and thus protects parent and name fields.
2207                  *
2208                  * The caller must perform a seqcount check in order
2209                  * to do anything useful with the returned dentry.
2210                  *
2211                  * NOTE! We do a "raw" seqcount_begin here. That means that
2212                  * we don't wait for the sequence count to stabilize if it
2213                  * is in the middle of a sequence change. If we do the slow
2214                  * dentry compare, we will do seqretries until it is stable,
2215                  * and if we end up with a successful lookup, we actually
2216                  * want to exit RCU lookup anyway.
2217                  *
2218                  * Note that raw_seqcount_begin still *does* smp_rmb(), so
2219                  * we are still guaranteed NUL-termination of ->d_name.name.
2220                  */
2221                 seq = raw_seqcount_begin(&dentry->d_seq);
2222                 if (dentry->d_parent != parent)
2223                         continue;
2224                 if (d_unhashed(dentry))
2225                         continue;
2226                 if (dentry->d_name.hash_len != hashlen)
2227                         continue;
2228                 if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
2229                         continue;
2230                 *seqp = seq;
2231                 return dentry;
2232         }
2233         return NULL;
2234 }
2235
2236 /**
2237  * d_lookup - search for a dentry
2238  * @parent: parent dentry
2239  * @name: qstr of name we wish to find
2240  * Returns: dentry, or NULL
2241  *
2242  * d_lookup searches the children of the parent dentry for the name in
2243  * question. If the dentry is found its reference count is incremented and the
2244  * dentry is returned. The caller must use dput to free the entry when it has
2245  * finished using it. %NULL is returned if the dentry does not exist.
2246  */
2247 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2248 {
2249         struct dentry *dentry;
2250         unsigned seq;
2251
2252         do {
2253                 seq = read_seqbegin(&rename_lock);
2254                 dentry = __d_lookup(parent, name);
2255                 if (dentry)
2256                         break;
2257         } while (read_seqretry(&rename_lock, seq));
2258         return dentry;
2259 }
2260 EXPORT_SYMBOL(d_lookup);
2261
2262 /**
2263  * __d_lookup - search for a dentry (racy)
2264  * @parent: parent dentry
2265  * @name: qstr of name we wish to find
2266  * Returns: dentry, or NULL
2267  *
2268  * __d_lookup is like d_lookup, however it may (rarely) return a
2269  * false-negative result due to unrelated rename activity.
2270  *
2271  * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2272  * however it must be used carefully, eg. with a following d_lookup in
2273  * the case of failure.
2274  *
2275  * __d_lookup callers must be commented.
2276  */
2277 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2278 {
2279         unsigned int hash = name->hash;
2280         struct hlist_bl_head *b = d_hash(hash);
2281         struct hlist_bl_node *node;
2282         struct dentry *found = NULL;
2283         struct dentry *dentry;
2284
2285         /*
2286          * Note: There is significant duplication with __d_lookup_rcu which is
2287          * required to prevent single threaded performance regressions
2288          * especially on architectures where smp_rmb (in seqcounts) are costly.
2289          * Keep the two functions in sync.
2290          */
2291
2292         /*
2293          * The hash list is protected using RCU.
2294          *
2295          * Take d_lock when comparing a candidate dentry, to avoid races
2296          * with d_move().
2297          *
2298          * It is possible that concurrent renames can mess up our list
2299          * walk here and result in missing our dentry, resulting in the
2300          * false-negative result. d_lookup() protects against concurrent
2301          * renames using rename_lock seqlock.
2302          *
2303          * See Documentation/filesystems/path-lookup.txt for more details.
2304          */
2305         rcu_read_lock();
2306         
2307         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2308
2309                 if (dentry->d_name.hash != hash)
2310                         continue;
2311
2312                 spin_lock(&dentry->d_lock);
2313                 if (dentry->d_parent != parent)
2314                         goto next;
2315                 if (d_unhashed(dentry))
2316                         goto next;
2317
2318                 if (!d_same_name(dentry, parent, name))
2319                         goto next;
2320
2321                 dentry->d_lockref.count++;
2322                 found = dentry;
2323                 spin_unlock(&dentry->d_lock);
2324                 break;
2325 next:
2326                 spin_unlock(&dentry->d_lock);
2327         }
2328         rcu_read_unlock();
2329
2330         return found;
2331 }
2332
2333 /**
2334  * d_hash_and_lookup - hash the qstr then search for a dentry
2335  * @dir: Directory to search in
2336  * @name: qstr of name we wish to find
2337  *
2338  * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2339  */
2340 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2341 {
2342         /*
2343          * Check for a fs-specific hash function. Note that we must
2344          * calculate the standard hash first, as the d_op->d_hash()
2345          * routine may choose to leave the hash value unchanged.
2346          */
2347         name->hash = full_name_hash(dir, name->name, name->len);
2348         if (dir->d_flags & DCACHE_OP_HASH) {
2349                 int err = dir->d_op->d_hash(dir, name);
2350                 if (unlikely(err < 0))
2351                         return ERR_PTR(err);
2352         }
2353         return d_lookup(dir, name);
2354 }
2355 EXPORT_SYMBOL(d_hash_and_lookup);
2356
2357 /*
2358  * When a file is deleted, we have two options:
2359  * - turn this dentry into a negative dentry
2360  * - unhash this dentry and free it.
2361  *
2362  * Usually, we want to just turn this into
2363  * a negative dentry, but if anybody else is
2364  * currently using the dentry or the inode
2365  * we can't do that and we fall back on removing
2366  * it from the hash queues and waiting for
2367  * it to be deleted later when it has no users
2368  */
2369  
2370 /**
2371  * d_delete - delete a dentry
2372  * @dentry: The dentry to delete
2373  *
2374  * Turn the dentry into a negative dentry if possible, otherwise
2375  * remove it from the hash queues so it can be deleted later
2376  */
2377  
2378 void d_delete(struct dentry * dentry)
2379 {
2380         struct inode *inode = dentry->d_inode;
2381
2382         spin_lock(&inode->i_lock);
2383         spin_lock(&dentry->d_lock);
2384         /*
2385          * Are we the only user?
2386          */
2387         if (dentry->d_lockref.count == 1) {
2388                 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2389                 dentry_unlink_inode(dentry);
2390         } else {
2391                 __d_drop(dentry);
2392                 spin_unlock(&dentry->d_lock);
2393                 spin_unlock(&inode->i_lock);
2394         }
2395 }
2396 EXPORT_SYMBOL(d_delete);
2397
2398 static void __d_rehash(struct dentry *entry)
2399 {
2400         struct hlist_bl_head *b = d_hash(entry->d_name.hash);
2401
2402         hlist_bl_lock(b);
2403         hlist_bl_add_head_rcu(&entry->d_hash, b);
2404         hlist_bl_unlock(b);
2405 }
2406
2407 /**
2408  * d_rehash     - add an entry back to the hash
2409  * @entry: dentry to add to the hash
2410  *
2411  * Adds a dentry to the hash according to its name.
2412  */
2413  
2414 void d_rehash(struct dentry * entry)
2415 {
2416         spin_lock(&entry->d_lock);
2417         __d_rehash(entry);
2418         spin_unlock(&entry->d_lock);
2419 }
2420 EXPORT_SYMBOL(d_rehash);
2421
2422 static inline unsigned start_dir_add(struct inode *dir)
2423 {
2424         preempt_disable_nested();
2425         for (;;) {
2426                 unsigned n = dir->i_dir_seq;
2427                 if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
2428                         return n;
2429                 cpu_relax();
2430         }
2431 }
2432
2433 static inline void end_dir_add(struct inode *dir, unsigned int n,
2434                                wait_queue_head_t *d_wait)
2435 {
2436         smp_store_release(&dir->i_dir_seq, n + 2);
2437         preempt_enable_nested();
2438         wake_up_all(d_wait);
2439 }
2440
2441 static void d_wait_lookup(struct dentry *dentry)
2442 {
2443         if (d_in_lookup(dentry)) {
2444                 DECLARE_WAITQUEUE(wait, current);
2445                 add_wait_queue(dentry->d_wait, &wait);
2446                 do {
2447                         set_current_state(TASK_UNINTERRUPTIBLE);
2448                         spin_unlock(&dentry->d_lock);
2449                         schedule();
2450                         spin_lock(&dentry->d_lock);
2451                 } while (d_in_lookup(dentry));
2452         }
2453 }
2454
2455 struct dentry *d_alloc_parallel(struct dentry *parent,
2456                                 const struct qstr *name,
2457                                 wait_queue_head_t *wq)
2458 {
2459         unsigned int hash = name->hash;
2460         struct hlist_bl_head *b = in_lookup_hash(parent, hash);
2461         struct hlist_bl_node *node;
2462         struct dentry *new = d_alloc(parent, name);
2463         struct dentry *dentry;
2464         unsigned seq, r_seq, d_seq;
2465
2466         if (unlikely(!new))
2467                 return ERR_PTR(-ENOMEM);
2468
2469 retry:
2470         rcu_read_lock();
2471         seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
2472         r_seq = read_seqbegin(&rename_lock);
2473         dentry = __d_lookup_rcu(parent, name, &d_seq);
2474         if (unlikely(dentry)) {
2475                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2476                         rcu_read_unlock();
2477                         goto retry;
2478                 }
2479                 if (read_seqcount_retry(&dentry->d_seq, d_seq)) {
2480                         rcu_read_unlock();
2481                         dput(dentry);
2482                         goto retry;
2483                 }
2484                 rcu_read_unlock();
2485                 dput(new);
2486                 return dentry;
2487         }
2488         if (unlikely(read_seqretry(&rename_lock, r_seq))) {
2489                 rcu_read_unlock();
2490                 goto retry;
2491         }
2492
2493         if (unlikely(seq & 1)) {
2494                 rcu_read_unlock();
2495                 goto retry;
2496         }
2497
2498         hlist_bl_lock(b);
2499         if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {
2500                 hlist_bl_unlock(b);
2501                 rcu_read_unlock();
2502                 goto retry;
2503         }
2504         /*
2505          * No changes for the parent since the beginning of d_lookup().
2506          * Since all removals from the chain happen with hlist_bl_lock(),
2507          * any potential in-lookup matches are going to stay here until
2508          * we unlock the chain.  All fields are stable in everything
2509          * we encounter.
2510          */
2511         hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
2512                 if (dentry->d_name.hash != hash)
2513                         continue;
2514                 if (dentry->d_parent != parent)
2515                         continue;
2516                 if (!d_same_name(dentry, parent, name))
2517                         continue;
2518                 hlist_bl_unlock(b);
2519                 /* now we can try to grab a reference */
2520                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2521                         rcu_read_unlock();
2522                         goto retry;
2523                 }
2524
2525                 rcu_read_unlock();
2526                 /*
2527                  * somebody is likely to be still doing lookup for it;
2528                  * wait for them to finish
2529                  */
2530                 spin_lock(&dentry->d_lock);
2531                 d_wait_lookup(dentry);
2532                 /*
2533                  * it's not in-lookup anymore; in principle we should repeat
2534                  * everything from dcache lookup, but it's likely to be what
2535                  * d_lookup() would've found anyway.  If it is, just return it;
2536                  * otherwise we really have to repeat the whole thing.
2537                  */
2538                 if (unlikely(dentry->d_name.hash != hash))
2539                         goto mismatch;
2540                 if (unlikely(dentry->d_parent != parent))
2541                         goto mismatch;
2542                 if (unlikely(d_unhashed(dentry)))
2543                         goto mismatch;
2544                 if (unlikely(!d_same_name(dentry, parent, name)))
2545                         goto mismatch;
2546                 /* OK, it *is* a hashed match; return it */
2547                 spin_unlock(&dentry->d_lock);
2548                 dput(new);
2549                 return dentry;
2550         }
2551         rcu_read_unlock();
2552         /* we can't take ->d_lock here; it's OK, though. */
2553         new->d_flags |= DCACHE_PAR_LOOKUP;
2554         new->d_wait = wq;
2555         hlist_bl_add_head(&new->d_u.d_in_lookup_hash, b);
2556         hlist_bl_unlock(b);
2557         return new;
2558 mismatch:
2559         spin_unlock(&dentry->d_lock);
2560         dput(dentry);
2561         goto retry;
2562 }
2563 EXPORT_SYMBOL(d_alloc_parallel);
2564
2565 /*
2566  * - Unhash the dentry
2567  * - Retrieve and clear the waitqueue head in dentry
2568  * - Return the waitqueue head
2569  */
2570 static wait_queue_head_t *__d_lookup_unhash(struct dentry *dentry)
2571 {
2572         wait_queue_head_t *d_wait;
2573         struct hlist_bl_head *b;
2574
2575         lockdep_assert_held(&dentry->d_lock);
2576
2577         b = in_lookup_hash(dentry->d_parent, dentry->d_name.hash);
2578         hlist_bl_lock(b);
2579         dentry->d_flags &= ~DCACHE_PAR_LOOKUP;
2580         __hlist_bl_del(&dentry->d_u.d_in_lookup_hash);
2581         d_wait = dentry->d_wait;
2582         dentry->d_wait = NULL;
2583         hlist_bl_unlock(b);
2584         INIT_HLIST_NODE(&dentry->d_u.d_alias);
2585         INIT_LIST_HEAD(&dentry->d_lru);
2586         return d_wait;
2587 }
2588
2589 void __d_lookup_unhash_wake(struct dentry *dentry)
2590 {
2591         spin_lock(&dentry->d_lock);
2592         wake_up_all(__d_lookup_unhash(dentry));
2593         spin_unlock(&dentry->d_lock);
2594 }
2595 EXPORT_SYMBOL(__d_lookup_unhash_wake);
2596
2597 /* inode->i_lock held if inode is non-NULL */
2598
2599 static inline void __d_add(struct dentry *dentry, struct inode *inode)
2600 {
2601         wait_queue_head_t *d_wait;
2602         struct inode *dir = NULL;
2603         unsigned n;
2604         spin_lock(&dentry->d_lock);
2605         if (unlikely(d_in_lookup(dentry))) {
2606                 dir = dentry->d_parent->d_inode;
2607                 n = start_dir_add(dir);
2608                 d_wait = __d_lookup_unhash(dentry);
2609         }
2610         if (inode) {
2611                 unsigned add_flags = d_flags_for_inode(inode);
2612                 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2613                 raw_write_seqcount_begin(&dentry->d_seq);
2614                 __d_set_inode_and_type(dentry, inode, add_flags);
2615                 raw_write_seqcount_end(&dentry->d_seq);
2616                 fsnotify_update_flags(dentry);
2617         }
2618         __d_rehash(dentry);
2619         if (dir)
2620                 end_dir_add(dir, n, d_wait);
2621         spin_unlock(&dentry->d_lock);
2622         if (inode)
2623                 spin_unlock(&inode->i_lock);
2624 }
2625
2626 /**
2627  * d_add - add dentry to hash queues
2628  * @entry: dentry to add
2629  * @inode: The inode to attach to this dentry
2630  *
2631  * This adds the entry to the hash queues and initializes @inode.
2632  * The entry was actually filled in earlier during d_alloc().
2633  */
2634
2635 void d_add(struct dentry *entry, struct inode *inode)
2636 {
2637         if (inode) {
2638                 security_d_instantiate(entry, inode);
2639                 spin_lock(&inode->i_lock);
2640         }
2641         __d_add(entry, inode);
2642 }
2643 EXPORT_SYMBOL(d_add);
2644
2645 /**
2646  * d_exact_alias - find and hash an exact unhashed alias
2647  * @entry: dentry to add
2648  * @inode: The inode to go with this dentry
2649  *
2650  * If an unhashed dentry with the same name/parent and desired
2651  * inode already exists, hash and return it.  Otherwise, return
2652  * NULL.
2653  *
2654  * Parent directory should be locked.
2655  */
2656 struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2657 {
2658         struct dentry *alias;
2659         unsigned int hash = entry->d_name.hash;
2660
2661         spin_lock(&inode->i_lock);
2662         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2663                 /*
2664                  * Don't need alias->d_lock here, because aliases with
2665                  * d_parent == entry->d_parent are not subject to name or
2666                  * parent changes, because the parent inode i_mutex is held.
2667                  */
2668                 if (alias->d_name.hash != hash)
2669                         continue;
2670                 if (alias->d_parent != entry->d_parent)
2671                         continue;
2672                 if (!d_same_name(alias, entry->d_parent, &entry->d_name))
2673                         continue;
2674                 spin_lock(&alias->d_lock);
2675                 if (!d_unhashed(alias)) {
2676                         spin_unlock(&alias->d_lock);
2677                         alias = NULL;
2678                 } else {
2679                         dget_dlock(alias);
2680                         __d_rehash(alias);
2681                         spin_unlock(&alias->d_lock);
2682                 }
2683                 spin_unlock(&inode->i_lock);
2684                 return alias;
2685         }
2686         spin_unlock(&inode->i_lock);
2687         return NULL;
2688 }
2689 EXPORT_SYMBOL(d_exact_alias);
2690
2691 static void swap_names(struct dentry *dentry, struct dentry *target)
2692 {
2693         if (unlikely(dname_external(target))) {
2694                 if (unlikely(dname_external(dentry))) {
2695                         /*
2696                          * Both external: swap the pointers
2697                          */
2698                         swap(target->d_name.name, dentry->d_name.name);
2699                 } else {
2700                         /*
2701                          * dentry:internal, target:external.  Steal target's
2702                          * storage and make target internal.
2703                          */
2704                         memcpy(target->d_iname, dentry->d_name.name,
2705                                         dentry->d_name.len + 1);
2706                         dentry->d_name.name = target->d_name.name;
2707                         target->d_name.name = target->d_iname;
2708                 }
2709         } else {
2710                 if (unlikely(dname_external(dentry))) {
2711                         /*
2712                          * dentry:external, target:internal.  Give dentry's
2713                          * storage to target and make dentry internal
2714                          */
2715                         memcpy(dentry->d_iname, target->d_name.name,
2716                                         target->d_name.len + 1);
2717                         target->d_name.name = dentry->d_name.name;
2718                         dentry->d_name.name = dentry->d_iname;
2719                 } else {
2720                         /*
2721                          * Both are internal.
2722                          */
2723                         unsigned int i;
2724                         BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2725                         for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2726                                 swap(((long *) &dentry->d_iname)[i],
2727                                      ((long *) &target->d_iname)[i]);
2728                         }
2729                 }
2730         }
2731         swap(dentry->d_name.hash_len, target->d_name.hash_len);
2732 }
2733
2734 static void copy_name(struct dentry *dentry, struct dentry *target)
2735 {
2736         struct external_name *old_name = NULL;
2737         if (unlikely(dname_external(dentry)))
2738                 old_name = external_name(dentry);
2739         if (unlikely(dname_external(target))) {
2740                 atomic_inc(&external_name(target)->u.count);
2741                 dentry->d_name = target->d_name;
2742         } else {
2743                 memcpy(dentry->d_iname, target->d_name.name,
2744                                 target->d_name.len + 1);
2745                 dentry->d_name.name = dentry->d_iname;
2746                 dentry->d_name.hash_len = target->d_name.hash_len;
2747         }
2748         if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2749                 kfree_rcu(old_name, u.head);
2750 }
2751
2752 /*
2753  * __d_move - move a dentry
2754  * @dentry: entry to move
2755  * @target: new dentry
2756  * @exchange: exchange the two dentries
2757  *
2758  * Update the dcache to reflect the move of a file name. Negative
2759  * dcache entries should not be moved in this way. Caller must hold
2760  * rename_lock, the i_mutex of the source and target directories,
2761  * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2762  */
2763 static void __d_move(struct dentry *dentry, struct dentry *target,
2764                      bool exchange)
2765 {
2766         struct dentry *old_parent, *p;
2767         wait_queue_head_t *d_wait;
2768         struct inode *dir = NULL;
2769         unsigned n;
2770
2771         WARN_ON(!dentry->d_inode);
2772         if (WARN_ON(dentry == target))
2773                 return;
2774
2775         BUG_ON(d_ancestor(target, dentry));
2776         old_parent = dentry->d_parent;
2777         p = d_ancestor(old_parent, target);
2778         if (IS_ROOT(dentry)) {
2779                 BUG_ON(p);
2780                 spin_lock(&target->d_parent->d_lock);
2781         } else if (!p) {
2782                 /* target is not a descendent of dentry->d_parent */
2783                 spin_lock(&target->d_parent->d_lock);
2784                 spin_lock_nested(&old_parent->d_lock, DENTRY_D_LOCK_NESTED);
2785         } else {
2786                 BUG_ON(p == dentry);
2787                 spin_lock(&old_parent->d_lock);
2788                 if (p != target)
2789                         spin_lock_nested(&target->d_parent->d_lock,
2790                                         DENTRY_D_LOCK_NESTED);
2791         }
2792         spin_lock_nested(&dentry->d_lock, 2);
2793         spin_lock_nested(&target->d_lock, 3);
2794
2795         if (unlikely(d_in_lookup(target))) {
2796                 dir = target->d_parent->d_inode;
2797                 n = start_dir_add(dir);
2798                 d_wait = __d_lookup_unhash(target);
2799         }
2800
2801         write_seqcount_begin(&dentry->d_seq);
2802         write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2803
2804         /* unhash both */
2805         if (!d_unhashed(dentry))
2806                 ___d_drop(dentry);
2807         if (!d_unhashed(target))
2808                 ___d_drop(target);
2809
2810         /* ... and switch them in the tree */
2811         dentry->d_parent = target->d_parent;
2812         if (!exchange) {
2813                 copy_name(dentry, target);
2814                 target->d_hash.pprev = NULL;
2815                 dentry->d_parent->d_lockref.count++;
2816                 if (dentry != old_parent) /* wasn't IS_ROOT */
2817                         WARN_ON(!--old_parent->d_lockref.count);
2818         } else {
2819                 target->d_parent = old_parent;
2820                 swap_names(dentry, target);
2821                 if (!hlist_unhashed(&target->d_sib))
2822                         __hlist_del(&target->d_sib);
2823                 hlist_add_head(&target->d_sib, &target->d_parent->d_children);
2824                 __d_rehash(target);
2825                 fsnotify_update_flags(target);
2826         }
2827         if (!hlist_unhashed(&dentry->d_sib))
2828                 __hlist_del(&dentry->d_sib);
2829         hlist_add_head(&dentry->d_sib, &dentry->d_parent->d_children);
2830         __d_rehash(dentry);
2831         fsnotify_update_flags(dentry);
2832         fscrypt_handle_d_move(dentry);
2833
2834         write_seqcount_end(&target->d_seq);
2835         write_seqcount_end(&dentry->d_seq);
2836
2837         if (dir)
2838                 end_dir_add(dir, n, d_wait);
2839
2840         if (dentry->d_parent != old_parent)
2841                 spin_unlock(&dentry->d_parent->d_lock);
2842         if (dentry != old_parent)
2843                 spin_unlock(&old_parent->d_lock);
2844         spin_unlock(&target->d_lock);
2845         spin_unlock(&dentry->d_lock);
2846 }
2847
2848 /*
2849  * d_move - move a dentry
2850  * @dentry: entry to move
2851  * @target: new dentry
2852  *
2853  * Update the dcache to reflect the move of a file name. Negative
2854  * dcache entries should not be moved in this way. See the locking
2855  * requirements for __d_move.
2856  */
2857 void d_move(struct dentry *dentry, struct dentry *target)
2858 {
2859         write_seqlock(&rename_lock);
2860         __d_move(dentry, target, false);
2861         write_sequnlock(&rename_lock);
2862 }
2863 EXPORT_SYMBOL(d_move);
2864
2865 /*
2866  * d_exchange - exchange two dentries
2867  * @dentry1: first dentry
2868  * @dentry2: second dentry
2869  */
2870 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2871 {
2872         write_seqlock(&rename_lock);
2873
2874         WARN_ON(!dentry1->d_inode);
2875         WARN_ON(!dentry2->d_inode);
2876         WARN_ON(IS_ROOT(dentry1));
2877         WARN_ON(IS_ROOT(dentry2));
2878
2879         __d_move(dentry1, dentry2, true);
2880
2881         write_sequnlock(&rename_lock);
2882 }
2883
2884 /**
2885  * d_ancestor - search for an ancestor
2886  * @p1: ancestor dentry
2887  * @p2: child dentry
2888  *
2889  * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2890  * an ancestor of p2, else NULL.
2891  */
2892 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2893 {
2894         struct dentry *p;
2895
2896         for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2897                 if (p->d_parent == p1)
2898                         return p;
2899         }
2900         return NULL;
2901 }
2902
2903 /*
2904  * This helper attempts to cope with remotely renamed directories
2905  *
2906  * It assumes that the caller is already holding
2907  * dentry->d_parent->d_inode->i_mutex, and rename_lock
2908  *
2909  * Note: If ever the locking in lock_rename() changes, then please
2910  * remember to update this too...
2911  */
2912 static int __d_unalias(struct dentry *dentry, struct dentry *alias)
2913 {
2914         struct mutex *m1 = NULL;
2915         struct rw_semaphore *m2 = NULL;
2916         int ret = -ESTALE;
2917
2918         /* If alias and dentry share a parent, then no extra locks required */
2919         if (alias->d_parent == dentry->d_parent)
2920                 goto out_unalias;
2921
2922         /* See lock_rename() */
2923         if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2924                 goto out_err;
2925         m1 = &dentry->d_sb->s_vfs_rename_mutex;
2926         if (!inode_trylock_shared(alias->d_parent->d_inode))
2927                 goto out_err;
2928         m2 = &alias->d_parent->d_inode->i_rwsem;
2929 out_unalias:
2930         __d_move(alias, dentry, false);
2931         ret = 0;
2932 out_err:
2933         if (m2)
2934                 up_read(m2);
2935         if (m1)
2936                 mutex_unlock(m1);
2937         return ret;
2938 }
2939
2940 /**
2941  * d_splice_alias - splice a disconnected dentry into the tree if one exists
2942  * @inode:  the inode which may have a disconnected dentry
2943  * @dentry: a negative dentry which we want to point to the inode.
2944  *
2945  * If inode is a directory and has an IS_ROOT alias, then d_move that in
2946  * place of the given dentry and return it, else simply d_add the inode
2947  * to the dentry and return NULL.
2948  *
2949  * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2950  * we should error out: directories can't have multiple aliases.
2951  *
2952  * This is needed in the lookup routine of any filesystem that is exportable
2953  * (via knfsd) so that we can build dcache paths to directories effectively.
2954  *
2955  * If a dentry was found and moved, then it is returned.  Otherwise NULL
2956  * is returned.  This matches the expected return value of ->lookup.
2957  *
2958  * Cluster filesystems may call this function with a negative, hashed dentry.
2959  * In that case, we know that the inode will be a regular file, and also this
2960  * will only occur during atomic_open. So we need to check for the dentry
2961  * being already hashed only in the final case.
2962  */
2963 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
2964 {
2965         if (IS_ERR(inode))
2966                 return ERR_CAST(inode);
2967
2968         BUG_ON(!d_unhashed(dentry));
2969
2970         if (!inode)
2971                 goto out;
2972
2973         security_d_instantiate(dentry, inode);
2974         spin_lock(&inode->i_lock);
2975         if (S_ISDIR(inode->i_mode)) {
2976                 struct dentry *new = __d_find_any_alias(inode);
2977                 if (unlikely(new)) {
2978                         /* The reference to new ensures it remains an alias */
2979                         spin_unlock(&inode->i_lock);
2980                         write_seqlock(&rename_lock);
2981                         if (unlikely(d_ancestor(new, dentry))) {
2982                                 write_sequnlock(&rename_lock);
2983                                 dput(new);
2984                                 new = ERR_PTR(-ELOOP);
2985                                 pr_warn_ratelimited(
2986                                         "VFS: Lookup of '%s' in %s %s"
2987                                         " would have caused loop\n",
2988                                         dentry->d_name.name,
2989                                         inode->i_sb->s_type->name,
2990                                         inode->i_sb->s_id);
2991                         } else if (!IS_ROOT(new)) {
2992                                 struct dentry *old_parent = dget(new->d_parent);
2993                                 int err = __d_unalias(dentry, new);
2994                                 write_sequnlock(&rename_lock);
2995                                 if (err) {
2996                                         dput(new);
2997                                         new = ERR_PTR(err);
2998                                 }
2999                                 dput(old_parent);
3000                         } else {
3001                                 __d_move(new, dentry, false);
3002                                 write_sequnlock(&rename_lock);
3003                         }
3004                         iput(inode);
3005                         return new;
3006                 }
3007         }
3008 out:
3009         __d_add(dentry, inode);
3010         return NULL;
3011 }
3012 EXPORT_SYMBOL(d_splice_alias);
3013
3014 /*
3015  * Test whether new_dentry is a subdirectory of old_dentry.
3016  *
3017  * Trivially implemented using the dcache structure
3018  */
3019
3020 /**
3021  * is_subdir - is new dentry a subdirectory of old_dentry
3022  * @new_dentry: new dentry
3023  * @old_dentry: old dentry
3024  *
3025  * Returns true if new_dentry is a subdirectory of the parent (at any depth).
3026  * Returns false otherwise.
3027  * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3028  */
3029   
3030 bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3031 {
3032         bool result;
3033         unsigned seq;
3034
3035         if (new_dentry == old_dentry)
3036                 return true;
3037
3038         do {
3039                 /* for restarting inner loop in case of seq retry */
3040                 seq = read_seqbegin(&rename_lock);
3041                 /*
3042                  * Need rcu_readlock to protect against the d_parent trashing
3043                  * due to d_move
3044                  */
3045                 rcu_read_lock();
3046                 if (d_ancestor(old_dentry, new_dentry))
3047                         result = true;
3048                 else
3049                         result = false;
3050                 rcu_read_unlock();
3051         } while (read_seqretry(&rename_lock, seq));
3052
3053         return result;
3054 }
3055 EXPORT_SYMBOL(is_subdir);
3056
3057 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3058 {
3059         struct dentry *root = data;
3060         if (dentry != root) {
3061                 if (d_unhashed(dentry) || !dentry->d_inode)
3062                         return D_WALK_SKIP;
3063
3064                 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3065                         dentry->d_flags |= DCACHE_GENOCIDE;
3066                         dentry->d_lockref.count--;
3067                 }
3068         }
3069         return D_WALK_CONTINUE;
3070 }
3071
3072 void d_genocide(struct dentry *parent)
3073 {
3074         d_walk(parent, parent, d_genocide_kill);
3075 }
3076
3077 void d_mark_tmpfile(struct file *file, struct inode *inode)
3078 {
3079         struct dentry *dentry = file->f_path.dentry;
3080
3081         BUG_ON(dentry->d_name.name != dentry->d_iname ||
3082                 !hlist_unhashed(&dentry->d_u.d_alias) ||
3083                 !d_unlinked(dentry));
3084         spin_lock(&dentry->d_parent->d_lock);
3085         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3086         dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3087                                 (unsigned long long)inode->i_ino);
3088         spin_unlock(&dentry->d_lock);
3089         spin_unlock(&dentry->d_parent->d_lock);
3090 }
3091 EXPORT_SYMBOL(d_mark_tmpfile);
3092
3093 void d_tmpfile(struct file *file, struct inode *inode)
3094 {
3095         struct dentry *dentry = file->f_path.dentry;
3096
3097         inode_dec_link_count(inode);
3098         d_mark_tmpfile(file, inode);
3099         d_instantiate(dentry, inode);
3100 }
3101 EXPORT_SYMBOL(d_tmpfile);
3102
3103 static __initdata unsigned long dhash_entries;
3104 static int __init set_dhash_entries(char *str)
3105 {
3106         if (!str)
3107                 return 0;
3108         dhash_entries = simple_strtoul(str, &str, 0);
3109         return 1;
3110 }
3111 __setup("dhash_entries=", set_dhash_entries);
3112
3113 static void __init dcache_init_early(void)
3114 {
3115         /* If hashes are distributed across NUMA nodes, defer
3116          * hash allocation until vmalloc space is available.
3117          */
3118         if (hashdist)
3119                 return;
3120
3121         dentry_hashtable =
3122                 alloc_large_system_hash("Dentry cache",
3123                                         sizeof(struct hlist_bl_head),
3124                                         dhash_entries,
3125                                         13,
3126                                         HASH_EARLY | HASH_ZERO,
3127                                         &d_hash_shift,
3128                                         NULL,
3129                                         0,
3130                                         0);
3131         d_hash_shift = 32 - d_hash_shift;
3132 }
3133
3134 static void __init dcache_init(void)
3135 {
3136         /*
3137          * A constructor could be added for stable state like the lists,
3138          * but it is probably not worth it because of the cache nature
3139          * of the dcache.
3140          */
3141         dentry_cache = KMEM_CACHE_USERCOPY(dentry,
3142                 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_ACCOUNT,
3143                 d_iname);
3144
3145         /* Hash may have been set up in dcache_init_early */
3146         if (!hashdist)
3147                 return;
3148
3149         dentry_hashtable =
3150                 alloc_large_system_hash("Dentry cache",
3151                                         sizeof(struct hlist_bl_head),
3152                                         dhash_entries,
3153                                         13,
3154                                         HASH_ZERO,
3155                                         &d_hash_shift,
3156                                         NULL,
3157                                         0,
3158                                         0);
3159         d_hash_shift = 32 - d_hash_shift;
3160 }
3161
3162 /* SLAB cache for __getname() consumers */
3163 struct kmem_cache *names_cachep __ro_after_init;
3164 EXPORT_SYMBOL(names_cachep);
3165
3166 void __init vfs_caches_init_early(void)
3167 {
3168         int i;
3169
3170         for (i = 0; i < ARRAY_SIZE(in_lookup_hashtable); i++)
3171                 INIT_HLIST_BL_HEAD(&in_lookup_hashtable[i]);
3172
3173         dcache_init_early();
3174         inode_init_early();
3175 }
3176
3177 void __init vfs_caches_init(void)
3178 {
3179         names_cachep = kmem_cache_create_usercopy("names_cache", PATH_MAX, 0,
3180                         SLAB_HWCACHE_ALIGN|SLAB_PANIC, 0, PATH_MAX, NULL);
3181
3182         dcache_init();
3183         inode_init();
3184         files_init();
3185         files_maxfiles_init();
3186         mnt_init();
3187         bdev_cache_init();
3188         chrdev_init();
3189 }