]> Git Repo - J-linux.git/blob - fs/dcache.c
Merge tag 'vfs-6.13-rc7.fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs
[J-linux.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 #include <asm/runtime-const.h>
39
40 /*
41  * Usage:
42  * dcache->d_inode->i_lock protects:
43  *   - i_dentry, d_u.d_alias, d_inode of aliases
44  * dcache_hash_bucket lock protects:
45  *   - the dcache hash table
46  * s_roots bl list spinlock protects:
47  *   - the s_roots list (see __d_drop)
48  * dentry->d_sb->s_dentry_lru_lock protects:
49  *   - the dcache lru lists and counters
50  * d_lock protects:
51  *   - d_flags
52  *   - d_name
53  *   - d_lru
54  *   - d_count
55  *   - d_unhashed()
56  *   - d_parent and d_chilren
57  *   - childrens' d_sib and d_parent
58  *   - d_u.d_alias, d_inode
59  *
60  * Ordering:
61  * dentry->d_inode->i_lock
62  *   dentry->d_lock
63  *     dentry->d_sb->s_dentry_lru_lock
64  *     dcache_hash_bucket lock
65  *     s_roots lock
66  *
67  * If there is an ancestor relationship:
68  * dentry->d_parent->...->d_parent->d_lock
69  *   ...
70  *     dentry->d_parent->d_lock
71  *       dentry->d_lock
72  *
73  * If no ancestor relationship:
74  * arbitrary, since it's serialized on rename_lock
75  */
76 int sysctl_vfs_cache_pressure __read_mostly = 100;
77 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
78
79 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
80
81 EXPORT_SYMBOL(rename_lock);
82
83 static struct kmem_cache *dentry_cache __ro_after_init;
84
85 const struct qstr empty_name = QSTR_INIT("", 0);
86 EXPORT_SYMBOL(empty_name);
87 const struct qstr slash_name = QSTR_INIT("/", 1);
88 EXPORT_SYMBOL(slash_name);
89 const struct qstr dotdot_name = QSTR_INIT("..", 2);
90 EXPORT_SYMBOL(dotdot_name);
91
92 /*
93  * This is the single most critical data structure when it comes
94  * to the dcache: the hashtable for lookups. Somebody should try
95  * to make this good - I've just made it work.
96  *
97  * This hash-function tries to avoid losing too many bits of hash
98  * information, yet avoid using a prime hash-size or similar.
99  *
100  * Marking the variables "used" ensures that the compiler doesn't
101  * optimize them away completely on architectures with runtime
102  * constant infrastructure, this allows debuggers to see their
103  * values. But updating these values has no effect on those arches.
104  */
105
106 static unsigned int d_hash_shift __ro_after_init __used;
107
108 static struct hlist_bl_head *dentry_hashtable __ro_after_init __used;
109
110 static inline struct hlist_bl_head *d_hash(unsigned long hashlen)
111 {
112         return runtime_const_ptr(dentry_hashtable) +
113                 runtime_const_shift_right_32(hashlen, d_hash_shift);
114 }
115
116 #define IN_LOOKUP_SHIFT 10
117 static struct hlist_bl_head in_lookup_hashtable[1 << IN_LOOKUP_SHIFT];
118
119 static inline struct hlist_bl_head *in_lookup_hash(const struct dentry *parent,
120                                         unsigned int hash)
121 {
122         hash += (unsigned long) parent / L1_CACHE_BYTES;
123         return in_lookup_hashtable + hash_32(hash, IN_LOOKUP_SHIFT);
124 }
125
126 struct dentry_stat_t {
127         long nr_dentry;
128         long nr_unused;
129         long age_limit;         /* age in seconds */
130         long want_pages;        /* pages requested by system */
131         long nr_negative;       /* # of unused negative dentries */
132         long dummy;             /* Reserved for future use */
133 };
134
135 static DEFINE_PER_CPU(long, nr_dentry);
136 static DEFINE_PER_CPU(long, nr_dentry_unused);
137 static DEFINE_PER_CPU(long, nr_dentry_negative);
138 static int dentry_negative_policy;
139
140 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
141 /* Statistics gathering. */
142 static struct dentry_stat_t dentry_stat = {
143         .age_limit = 45,
144 };
145
146 /*
147  * Here we resort to our own counters instead of using generic per-cpu counters
148  * for consistency with what the vfs inode code does. We are expected to harvest
149  * better code and performance by having our own specialized counters.
150  *
151  * Please note that the loop is done over all possible CPUs, not over all online
152  * CPUs. The reason for this is that we don't want to play games with CPUs going
153  * on and off. If one of them goes off, we will just keep their counters.
154  *
155  * glommer: See cffbc8a for details, and if you ever intend to change this,
156  * please update all vfs counters to match.
157  */
158 static long get_nr_dentry(void)
159 {
160         int i;
161         long sum = 0;
162         for_each_possible_cpu(i)
163                 sum += per_cpu(nr_dentry, i);
164         return sum < 0 ? 0 : sum;
165 }
166
167 static long get_nr_dentry_unused(void)
168 {
169         int i;
170         long sum = 0;
171         for_each_possible_cpu(i)
172                 sum += per_cpu(nr_dentry_unused, i);
173         return sum < 0 ? 0 : sum;
174 }
175
176 static long get_nr_dentry_negative(void)
177 {
178         int i;
179         long sum = 0;
180
181         for_each_possible_cpu(i)
182                 sum += per_cpu(nr_dentry_negative, i);
183         return sum < 0 ? 0 : sum;
184 }
185
186 static int proc_nr_dentry(const struct ctl_table *table, int write, void *buffer,
187                           size_t *lenp, loff_t *ppos)
188 {
189         dentry_stat.nr_dentry = get_nr_dentry();
190         dentry_stat.nr_unused = get_nr_dentry_unused();
191         dentry_stat.nr_negative = get_nr_dentry_negative();
192         return proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
193 }
194
195 static struct ctl_table fs_dcache_sysctls[] = {
196         {
197                 .procname       = "dentry-state",
198                 .data           = &dentry_stat,
199                 .maxlen         = 6*sizeof(long),
200                 .mode           = 0444,
201                 .proc_handler   = proc_nr_dentry,
202         },
203         {
204                 .procname       = "dentry-negative",
205                 .data           = &dentry_negative_policy,
206                 .maxlen         = sizeof(dentry_negative_policy),
207                 .mode           = 0644,
208                 .proc_handler   = proc_dointvec_minmax,
209                 .extra1         = SYSCTL_ZERO,
210                 .extra2         = SYSCTL_ONE,
211         },
212 };
213
214 static int __init init_fs_dcache_sysctls(void)
215 {
216         register_sysctl_init("fs", fs_dcache_sysctls);
217         return 0;
218 }
219 fs_initcall(init_fs_dcache_sysctls);
220 #endif
221
222 /*
223  * Compare 2 name strings, return 0 if they match, otherwise non-zero.
224  * The strings are both count bytes long, and count is non-zero.
225  */
226 #ifdef CONFIG_DCACHE_WORD_ACCESS
227
228 #include <asm/word-at-a-time.h>
229 /*
230  * NOTE! 'cs' and 'scount' come from a dentry, so it has a
231  * aligned allocation for this particular component. We don't
232  * strictly need the load_unaligned_zeropad() safety, but it
233  * doesn't hurt either.
234  *
235  * In contrast, 'ct' and 'tcount' can be from a pathname, and do
236  * need the careful unaligned handling.
237  */
238 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
239 {
240         unsigned long a,b,mask;
241
242         for (;;) {
243                 a = read_word_at_a_time(cs);
244                 b = load_unaligned_zeropad(ct);
245                 if (tcount < sizeof(unsigned long))
246                         break;
247                 if (unlikely(a != b))
248                         return 1;
249                 cs += sizeof(unsigned long);
250                 ct += sizeof(unsigned long);
251                 tcount -= sizeof(unsigned long);
252                 if (!tcount)
253                         return 0;
254         }
255         mask = bytemask_from_count(tcount);
256         return unlikely(!!((a ^ b) & mask));
257 }
258
259 #else
260
261 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
262 {
263         do {
264                 if (*cs != *ct)
265                         return 1;
266                 cs++;
267                 ct++;
268                 tcount--;
269         } while (tcount);
270         return 0;
271 }
272
273 #endif
274
275 static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount)
276 {
277         /*
278          * Be careful about RCU walk racing with rename:
279          * use 'READ_ONCE' to fetch the name pointer.
280          *
281          * NOTE! Even if a rename will mean that the length
282          * was not loaded atomically, we don't care. The
283          * RCU walk will check the sequence count eventually,
284          * and catch it. And we won't overrun the buffer,
285          * because we're reading the name pointer atomically,
286          * and a dentry name is guaranteed to be properly
287          * terminated with a NUL byte.
288          *
289          * End result: even if 'len' is wrong, we'll exit
290          * early because the data cannot match (there can
291          * be no NUL in the ct/tcount data)
292          */
293         const unsigned char *cs = READ_ONCE(dentry->d_name.name);
294
295         return dentry_string_cmp(cs, ct, tcount);
296 }
297
298 struct external_name {
299         union {
300                 atomic_t count;
301                 struct rcu_head head;
302         } u;
303         unsigned char name[];
304 };
305
306 static inline struct external_name *external_name(struct dentry *dentry)
307 {
308         return container_of(dentry->d_name.name, struct external_name, name[0]);
309 }
310
311 static void __d_free(struct rcu_head *head)
312 {
313         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
314
315         kmem_cache_free(dentry_cache, dentry); 
316 }
317
318 static void __d_free_external(struct rcu_head *head)
319 {
320         struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
321         kfree(external_name(dentry));
322         kmem_cache_free(dentry_cache, dentry);
323 }
324
325 static inline int dname_external(const struct dentry *dentry)
326 {
327         return dentry->d_name.name != dentry->d_iname;
328 }
329
330 void take_dentry_name_snapshot(struct name_snapshot *name, struct dentry *dentry)
331 {
332         spin_lock(&dentry->d_lock);
333         name->name = dentry->d_name;
334         if (unlikely(dname_external(dentry))) {
335                 atomic_inc(&external_name(dentry)->u.count);
336         } else {
337                 memcpy(name->inline_name, dentry->d_iname,
338                        dentry->d_name.len + 1);
339                 name->name.name = name->inline_name;
340         }
341         spin_unlock(&dentry->d_lock);
342 }
343 EXPORT_SYMBOL(take_dentry_name_snapshot);
344
345 void release_dentry_name_snapshot(struct name_snapshot *name)
346 {
347         if (unlikely(name->name.name != name->inline_name)) {
348                 struct external_name *p;
349                 p = container_of(name->name.name, struct external_name, name[0]);
350                 if (unlikely(atomic_dec_and_test(&p->u.count)))
351                         kfree_rcu(p, u.head);
352         }
353 }
354 EXPORT_SYMBOL(release_dentry_name_snapshot);
355
356 static inline void __d_set_inode_and_type(struct dentry *dentry,
357                                           struct inode *inode,
358                                           unsigned type_flags)
359 {
360         unsigned flags;
361
362         dentry->d_inode = inode;
363         flags = READ_ONCE(dentry->d_flags);
364         flags &= ~DCACHE_ENTRY_TYPE;
365         flags |= type_flags;
366         smp_store_release(&dentry->d_flags, flags);
367 }
368
369 static inline void __d_clear_type_and_inode(struct dentry *dentry)
370 {
371         unsigned flags = READ_ONCE(dentry->d_flags);
372
373         flags &= ~DCACHE_ENTRY_TYPE;
374         WRITE_ONCE(dentry->d_flags, flags);
375         dentry->d_inode = NULL;
376         /*
377          * The negative counter only tracks dentries on the LRU. Don't inc if
378          * d_lru is on another list.
379          */
380         if ((flags & (DCACHE_LRU_LIST|DCACHE_SHRINK_LIST)) == DCACHE_LRU_LIST)
381                 this_cpu_inc(nr_dentry_negative);
382 }
383
384 static void dentry_free(struct dentry *dentry)
385 {
386         WARN_ON(!hlist_unhashed(&dentry->d_u.d_alias));
387         if (unlikely(dname_external(dentry))) {
388                 struct external_name *p = external_name(dentry);
389                 if (likely(atomic_dec_and_test(&p->u.count))) {
390                         call_rcu(&dentry->d_u.d_rcu, __d_free_external);
391                         return;
392                 }
393         }
394         /* if dentry was never visible to RCU, immediate free is OK */
395         if (dentry->d_flags & DCACHE_NORCU)
396                 __d_free(&dentry->d_u.d_rcu);
397         else
398                 call_rcu(&dentry->d_u.d_rcu, __d_free);
399 }
400
401 /*
402  * Release the dentry's inode, using the filesystem
403  * d_iput() operation if defined.
404  */
405 static void dentry_unlink_inode(struct dentry * dentry)
406         __releases(dentry->d_lock)
407         __releases(dentry->d_inode->i_lock)
408 {
409         struct inode *inode = dentry->d_inode;
410
411         raw_write_seqcount_begin(&dentry->d_seq);
412         __d_clear_type_and_inode(dentry);
413         hlist_del_init(&dentry->d_u.d_alias);
414         raw_write_seqcount_end(&dentry->d_seq);
415         spin_unlock(&dentry->d_lock);
416         spin_unlock(&inode->i_lock);
417         if (!inode->i_nlink)
418                 fsnotify_inoderemove(inode);
419         if (dentry->d_op && dentry->d_op->d_iput)
420                 dentry->d_op->d_iput(dentry, inode);
421         else
422                 iput(inode);
423 }
424
425 /*
426  * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
427  * is in use - which includes both the "real" per-superblock
428  * LRU list _and_ the DCACHE_SHRINK_LIST use.
429  *
430  * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
431  * on the shrink list (ie not on the superblock LRU list).
432  *
433  * The per-cpu "nr_dentry_unused" counters are updated with
434  * the DCACHE_LRU_LIST bit.
435  *
436  * The per-cpu "nr_dentry_negative" counters are only updated
437  * when deleted from or added to the per-superblock LRU list, not
438  * from/to the shrink list. That is to avoid an unneeded dec/inc
439  * pair when moving from LRU to shrink list in select_collect().
440  *
441  * These helper functions make sure we always follow the
442  * rules. d_lock must be held by the caller.
443  */
444 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
445 static void d_lru_add(struct dentry *dentry)
446 {
447         D_FLAG_VERIFY(dentry, 0);
448         dentry->d_flags |= DCACHE_LRU_LIST;
449         this_cpu_inc(nr_dentry_unused);
450         if (d_is_negative(dentry))
451                 this_cpu_inc(nr_dentry_negative);
452         WARN_ON_ONCE(!list_lru_add_obj(
453                         &dentry->d_sb->s_dentry_lru, &dentry->d_lru));
454 }
455
456 static void d_lru_del(struct dentry *dentry)
457 {
458         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
459         dentry->d_flags &= ~DCACHE_LRU_LIST;
460         this_cpu_dec(nr_dentry_unused);
461         if (d_is_negative(dentry))
462                 this_cpu_dec(nr_dentry_negative);
463         WARN_ON_ONCE(!list_lru_del_obj(
464                         &dentry->d_sb->s_dentry_lru, &dentry->d_lru));
465 }
466
467 static void d_shrink_del(struct dentry *dentry)
468 {
469         D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
470         list_del_init(&dentry->d_lru);
471         dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
472         this_cpu_dec(nr_dentry_unused);
473 }
474
475 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
476 {
477         D_FLAG_VERIFY(dentry, 0);
478         list_add(&dentry->d_lru, list);
479         dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
480         this_cpu_inc(nr_dentry_unused);
481 }
482
483 /*
484  * These can only be called under the global LRU lock, ie during the
485  * callback for freeing the LRU list. "isolate" removes it from the
486  * LRU lists entirely, while shrink_move moves it to the indicated
487  * private list.
488  */
489 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
490 {
491         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
492         dentry->d_flags &= ~DCACHE_LRU_LIST;
493         this_cpu_dec(nr_dentry_unused);
494         if (d_is_negative(dentry))
495                 this_cpu_dec(nr_dentry_negative);
496         list_lru_isolate(lru, &dentry->d_lru);
497 }
498
499 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
500                               struct list_head *list)
501 {
502         D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
503         dentry->d_flags |= DCACHE_SHRINK_LIST;
504         if (d_is_negative(dentry))
505                 this_cpu_dec(nr_dentry_negative);
506         list_lru_isolate_move(lru, &dentry->d_lru, list);
507 }
508
509 static void ___d_drop(struct dentry *dentry)
510 {
511         struct hlist_bl_head *b;
512         /*
513          * Hashed dentries are normally on the dentry hashtable,
514          * with the exception of those newly allocated by
515          * d_obtain_root, which are always IS_ROOT:
516          */
517         if (unlikely(IS_ROOT(dentry)))
518                 b = &dentry->d_sb->s_roots;
519         else
520                 b = d_hash(dentry->d_name.hash);
521
522         hlist_bl_lock(b);
523         __hlist_bl_del(&dentry->d_hash);
524         hlist_bl_unlock(b);
525 }
526
527 void __d_drop(struct dentry *dentry)
528 {
529         if (!d_unhashed(dentry)) {
530                 ___d_drop(dentry);
531                 dentry->d_hash.pprev = NULL;
532                 write_seqcount_invalidate(&dentry->d_seq);
533         }
534 }
535 EXPORT_SYMBOL(__d_drop);
536
537 /**
538  * d_drop - drop a dentry
539  * @dentry: dentry to drop
540  *
541  * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
542  * be found through a VFS lookup any more. Note that this is different from
543  * deleting the dentry - d_delete will try to mark the dentry negative if
544  * possible, giving a successful _negative_ lookup, while d_drop will
545  * just make the cache lookup fail.
546  *
547  * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
548  * reason (NFS timeouts or autofs deletes).
549  *
550  * __d_drop requires dentry->d_lock
551  *
552  * ___d_drop doesn't mark dentry as "unhashed"
553  * (dentry->d_hash.pprev will be LIST_POISON2, not NULL).
554  */
555 void d_drop(struct dentry *dentry)
556 {
557         spin_lock(&dentry->d_lock);
558         __d_drop(dentry);
559         spin_unlock(&dentry->d_lock);
560 }
561 EXPORT_SYMBOL(d_drop);
562
563 static inline void dentry_unlist(struct dentry *dentry)
564 {
565         struct dentry *next;
566         /*
567          * Inform d_walk() and shrink_dentry_list() that we are no longer
568          * attached to the dentry tree
569          */
570         dentry->d_flags |= DCACHE_DENTRY_KILLED;
571         if (unlikely(hlist_unhashed(&dentry->d_sib)))
572                 return;
573         __hlist_del(&dentry->d_sib);
574         /*
575          * Cursors can move around the list of children.  While we'd been
576          * a normal list member, it didn't matter - ->d_sib.next would've
577          * been updated.  However, from now on it won't be and for the
578          * things like d_walk() it might end up with a nasty surprise.
579          * Normally d_walk() doesn't care about cursors moving around -
580          * ->d_lock on parent prevents that and since a cursor has no children
581          * of its own, we get through it without ever unlocking the parent.
582          * There is one exception, though - if we ascend from a child that
583          * gets killed as soon as we unlock it, the next sibling is found
584          * using the value left in its ->d_sib.next.  And if _that_
585          * pointed to a cursor, and cursor got moved (e.g. by lseek())
586          * before d_walk() regains parent->d_lock, we'll end up skipping
587          * everything the cursor had been moved past.
588          *
589          * Solution: make sure that the pointer left behind in ->d_sib.next
590          * points to something that won't be moving around.  I.e. skip the
591          * cursors.
592          */
593         while (dentry->d_sib.next) {
594                 next = hlist_entry(dentry->d_sib.next, struct dentry, d_sib);
595                 if (likely(!(next->d_flags & DCACHE_DENTRY_CURSOR)))
596                         break;
597                 dentry->d_sib.next = next->d_sib.next;
598         }
599 }
600
601 static struct dentry *__dentry_kill(struct dentry *dentry)
602 {
603         struct dentry *parent = NULL;
604         bool can_free = true;
605
606         /*
607          * The dentry is now unrecoverably dead to the world.
608          */
609         lockref_mark_dead(&dentry->d_lockref);
610
611         /*
612          * inform the fs via d_prune that this dentry is about to be
613          * unhashed and destroyed.
614          */
615         if (dentry->d_flags & DCACHE_OP_PRUNE)
616                 dentry->d_op->d_prune(dentry);
617
618         if (dentry->d_flags & DCACHE_LRU_LIST) {
619                 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
620                         d_lru_del(dentry);
621         }
622         /* if it was on the hash then remove it */
623         __d_drop(dentry);
624         if (dentry->d_inode)
625                 dentry_unlink_inode(dentry);
626         else
627                 spin_unlock(&dentry->d_lock);
628         this_cpu_dec(nr_dentry);
629         if (dentry->d_op && dentry->d_op->d_release)
630                 dentry->d_op->d_release(dentry);
631
632         cond_resched();
633         /* now that it's negative, ->d_parent is stable */
634         if (!IS_ROOT(dentry)) {
635                 parent = dentry->d_parent;
636                 spin_lock(&parent->d_lock);
637         }
638         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
639         dentry_unlist(dentry);
640         if (dentry->d_flags & DCACHE_SHRINK_LIST)
641                 can_free = false;
642         spin_unlock(&dentry->d_lock);
643         if (likely(can_free))
644                 dentry_free(dentry);
645         if (parent && --parent->d_lockref.count) {
646                 spin_unlock(&parent->d_lock);
647                 return NULL;
648         }
649         return parent;
650 }
651
652 /*
653  * Lock a dentry for feeding it to __dentry_kill().
654  * Called under rcu_read_lock() and dentry->d_lock; the former
655  * guarantees that nothing we access will be freed under us.
656  * Note that dentry is *not* protected from concurrent dentry_kill(),
657  * d_delete(), etc.
658  *
659  * Return false if dentry is busy.  Otherwise, return true and have
660  * that dentry's inode locked.
661  */
662
663 static bool lock_for_kill(struct dentry *dentry)
664 {
665         struct inode *inode = dentry->d_inode;
666
667         if (unlikely(dentry->d_lockref.count))
668                 return false;
669
670         if (!inode || likely(spin_trylock(&inode->i_lock)))
671                 return true;
672
673         do {
674                 spin_unlock(&dentry->d_lock);
675                 spin_lock(&inode->i_lock);
676                 spin_lock(&dentry->d_lock);
677                 if (likely(inode == dentry->d_inode))
678                         break;
679                 spin_unlock(&inode->i_lock);
680                 inode = dentry->d_inode;
681         } while (inode);
682         if (likely(!dentry->d_lockref.count))
683                 return true;
684         if (inode)
685                 spin_unlock(&inode->i_lock);
686         return false;
687 }
688
689 /*
690  * Decide if dentry is worth retaining.  Usually this is called with dentry
691  * locked; if not locked, we are more limited and might not be able to tell
692  * without a lock.  False in this case means "punt to locked path and recheck".
693  *
694  * In case we aren't locked, these predicates are not "stable". However, it is
695  * sufficient that at some point after we dropped the reference the dentry was
696  * hashed and the flags had the proper value. Other dentry users may have
697  * re-gotten a reference to the dentry and change that, but our work is done -
698  * we can leave the dentry around with a zero refcount.
699  */
700 static inline bool retain_dentry(struct dentry *dentry, bool locked)
701 {
702         unsigned int d_flags;
703
704         smp_rmb();
705         d_flags = READ_ONCE(dentry->d_flags);
706
707         // Unreachable? Nobody would be able to look it up, no point retaining
708         if (unlikely(d_unhashed(dentry)))
709                 return false;
710
711         // Same if it's disconnected
712         if (unlikely(d_flags & DCACHE_DISCONNECTED))
713                 return false;
714
715         // ->d_delete() might tell us not to bother, but that requires
716         // ->d_lock; can't decide without it
717         if (unlikely(d_flags & DCACHE_OP_DELETE)) {
718                 if (!locked || dentry->d_op->d_delete(dentry))
719                         return false;
720         }
721
722         // Explicitly told not to bother
723         if (unlikely(d_flags & DCACHE_DONTCACHE))
724                 return false;
725
726         // At this point it looks like we ought to keep it.  We also might
727         // need to do something - put it on LRU if it wasn't there already
728         // and mark it referenced if it was on LRU, but not marked yet.
729         // Unfortunately, both actions require ->d_lock, so in lockless
730         // case we'd have to punt rather than doing those.
731         if (unlikely(!(d_flags & DCACHE_LRU_LIST))) {
732                 if (!locked)
733                         return false;
734                 d_lru_add(dentry);
735         } else if (unlikely(!(d_flags & DCACHE_REFERENCED))) {
736                 if (!locked)
737                         return false;
738                 dentry->d_flags |= DCACHE_REFERENCED;
739         }
740         return true;
741 }
742
743 void d_mark_dontcache(struct inode *inode)
744 {
745         struct dentry *de;
746
747         spin_lock(&inode->i_lock);
748         hlist_for_each_entry(de, &inode->i_dentry, d_u.d_alias) {
749                 spin_lock(&de->d_lock);
750                 de->d_flags |= DCACHE_DONTCACHE;
751                 spin_unlock(&de->d_lock);
752         }
753         inode->i_state |= I_DONTCACHE;
754         spin_unlock(&inode->i_lock);
755 }
756 EXPORT_SYMBOL(d_mark_dontcache);
757
758 /*
759  * Try to do a lockless dput(), and return whether that was successful.
760  *
761  * If unsuccessful, we return false, having already taken the dentry lock.
762  * In that case refcount is guaranteed to be zero and we have already
763  * decided that it's not worth keeping around.
764  *
765  * The caller needs to hold the RCU read lock, so that the dentry is
766  * guaranteed to stay around even if the refcount goes down to zero!
767  */
768 static inline bool fast_dput(struct dentry *dentry)
769 {
770         int ret;
771
772         /*
773          * try to decrement the lockref optimistically.
774          */
775         ret = lockref_put_return(&dentry->d_lockref);
776
777         /*
778          * If the lockref_put_return() failed due to the lock being held
779          * by somebody else, the fast path has failed. We will need to
780          * get the lock, and then check the count again.
781          */
782         if (unlikely(ret < 0)) {
783                 spin_lock(&dentry->d_lock);
784                 if (WARN_ON_ONCE(dentry->d_lockref.count <= 0)) {
785                         spin_unlock(&dentry->d_lock);
786                         return true;
787                 }
788                 dentry->d_lockref.count--;
789                 goto locked;
790         }
791
792         /*
793          * If we weren't the last ref, we're done.
794          */
795         if (ret)
796                 return true;
797
798         /*
799          * Can we decide that decrement of refcount is all we needed without
800          * taking the lock?  There's a very common case when it's all we need -
801          * dentry looks like it ought to be retained and there's nothing else
802          * to do.
803          */
804         if (retain_dentry(dentry, false))
805                 return true;
806
807         /*
808          * Either not worth retaining or we can't tell without the lock.
809          * Get the lock, then.  We've already decremented the refcount to 0,
810          * but we'll need to re-check the situation after getting the lock.
811          */
812         spin_lock(&dentry->d_lock);
813
814         /*
815          * Did somebody else grab a reference to it in the meantime, and
816          * we're no longer the last user after all? Alternatively, somebody
817          * else could have killed it and marked it dead. Either way, we
818          * don't need to do anything else.
819          */
820 locked:
821         if (dentry->d_lockref.count || retain_dentry(dentry, true)) {
822                 spin_unlock(&dentry->d_lock);
823                 return true;
824         }
825         return false;
826 }
827
828
829 /* 
830  * This is dput
831  *
832  * This is complicated by the fact that we do not want to put
833  * dentries that are no longer on any hash chain on the unused
834  * list: we'd much rather just get rid of them immediately.
835  *
836  * However, that implies that we have to traverse the dentry
837  * tree upwards to the parents which might _also_ now be
838  * scheduled for deletion (it may have been only waiting for
839  * its last child to go away).
840  *
841  * This tail recursion is done by hand as we don't want to depend
842  * on the compiler to always get this right (gcc generally doesn't).
843  * Real recursion would eat up our stack space.
844  */
845
846 /*
847  * dput - release a dentry
848  * @dentry: dentry to release 
849  *
850  * Release a dentry. This will drop the usage count and if appropriate
851  * call the dentry unlink method as well as removing it from the queues and
852  * releasing its resources. If the parent dentries were scheduled for release
853  * they too may now get deleted.
854  */
855 void dput(struct dentry *dentry)
856 {
857         if (!dentry)
858                 return;
859         might_sleep();
860         rcu_read_lock();
861         if (likely(fast_dput(dentry))) {
862                 rcu_read_unlock();
863                 return;
864         }
865         while (lock_for_kill(dentry)) {
866                 rcu_read_unlock();
867                 dentry = __dentry_kill(dentry);
868                 if (!dentry)
869                         return;
870                 if (retain_dentry(dentry, true)) {
871                         spin_unlock(&dentry->d_lock);
872                         return;
873                 }
874                 rcu_read_lock();
875         }
876         rcu_read_unlock();
877         spin_unlock(&dentry->d_lock);
878 }
879 EXPORT_SYMBOL(dput);
880
881 static void to_shrink_list(struct dentry *dentry, struct list_head *list)
882 __must_hold(&dentry->d_lock)
883 {
884         if (!(dentry->d_flags & DCACHE_SHRINK_LIST)) {
885                 if (dentry->d_flags & DCACHE_LRU_LIST)
886                         d_lru_del(dentry);
887                 d_shrink_add(dentry, list);
888         }
889 }
890
891 void dput_to_list(struct dentry *dentry, struct list_head *list)
892 {
893         rcu_read_lock();
894         if (likely(fast_dput(dentry))) {
895                 rcu_read_unlock();
896                 return;
897         }
898         rcu_read_unlock();
899         to_shrink_list(dentry, list);
900         spin_unlock(&dentry->d_lock);
901 }
902
903 struct dentry *dget_parent(struct dentry *dentry)
904 {
905         int gotref;
906         struct dentry *ret;
907         unsigned seq;
908
909         /*
910          * Do optimistic parent lookup without any
911          * locking.
912          */
913         rcu_read_lock();
914         seq = raw_seqcount_begin(&dentry->d_seq);
915         ret = READ_ONCE(dentry->d_parent);
916         gotref = lockref_get_not_zero(&ret->d_lockref);
917         rcu_read_unlock();
918         if (likely(gotref)) {
919                 if (!read_seqcount_retry(&dentry->d_seq, seq))
920                         return ret;
921                 dput(ret);
922         }
923
924 repeat:
925         /*
926          * Don't need rcu_dereference because we re-check it was correct under
927          * the lock.
928          */
929         rcu_read_lock();
930         ret = dentry->d_parent;
931         spin_lock(&ret->d_lock);
932         if (unlikely(ret != dentry->d_parent)) {
933                 spin_unlock(&ret->d_lock);
934                 rcu_read_unlock();
935                 goto repeat;
936         }
937         rcu_read_unlock();
938         BUG_ON(!ret->d_lockref.count);
939         ret->d_lockref.count++;
940         spin_unlock(&ret->d_lock);
941         return ret;
942 }
943 EXPORT_SYMBOL(dget_parent);
944
945 static struct dentry * __d_find_any_alias(struct inode *inode)
946 {
947         struct dentry *alias;
948
949         if (hlist_empty(&inode->i_dentry))
950                 return NULL;
951         alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
952         lockref_get(&alias->d_lockref);
953         return alias;
954 }
955
956 /**
957  * d_find_any_alias - find any alias for a given inode
958  * @inode: inode to find an alias for
959  *
960  * If any aliases exist for the given inode, take and return a
961  * reference for one of them.  If no aliases exist, return %NULL.
962  */
963 struct dentry *d_find_any_alias(struct inode *inode)
964 {
965         struct dentry *de;
966
967         spin_lock(&inode->i_lock);
968         de = __d_find_any_alias(inode);
969         spin_unlock(&inode->i_lock);
970         return de;
971 }
972 EXPORT_SYMBOL(d_find_any_alias);
973
974 static struct dentry *__d_find_alias(struct inode *inode)
975 {
976         struct dentry *alias;
977
978         if (S_ISDIR(inode->i_mode))
979                 return __d_find_any_alias(inode);
980
981         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
982                 spin_lock(&alias->d_lock);
983                 if (!d_unhashed(alias)) {
984                         dget_dlock(alias);
985                         spin_unlock(&alias->d_lock);
986                         return alias;
987                 }
988                 spin_unlock(&alias->d_lock);
989         }
990         return NULL;
991 }
992
993 /**
994  * d_find_alias - grab a hashed alias of inode
995  * @inode: inode in question
996  *
997  * If inode has a hashed alias, or is a directory and has any alias,
998  * acquire the reference to alias and return it. Otherwise return NULL.
999  * Notice that if inode is a directory there can be only one alias and
1000  * it can be unhashed only if it has no children, or if it is the root
1001  * of a filesystem, or if the directory was renamed and d_revalidate
1002  * was the first vfs operation to notice.
1003  *
1004  * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
1005  * any other hashed alias over that one.
1006  */
1007 struct dentry *d_find_alias(struct inode *inode)
1008 {
1009         struct dentry *de = NULL;
1010
1011         if (!hlist_empty(&inode->i_dentry)) {
1012                 spin_lock(&inode->i_lock);
1013                 de = __d_find_alias(inode);
1014                 spin_unlock(&inode->i_lock);
1015         }
1016         return de;
1017 }
1018 EXPORT_SYMBOL(d_find_alias);
1019
1020 /*
1021  *  Caller MUST be holding rcu_read_lock() and be guaranteed
1022  *  that inode won't get freed until rcu_read_unlock().
1023  */
1024 struct dentry *d_find_alias_rcu(struct inode *inode)
1025 {
1026         struct hlist_head *l = &inode->i_dentry;
1027         struct dentry *de = NULL;
1028
1029         spin_lock(&inode->i_lock);
1030         // ->i_dentry and ->i_rcu are colocated, but the latter won't be
1031         // used without having I_FREEING set, which means no aliases left
1032         if (likely(!(inode->i_state & I_FREEING) && !hlist_empty(l))) {
1033                 if (S_ISDIR(inode->i_mode)) {
1034                         de = hlist_entry(l->first, struct dentry, d_u.d_alias);
1035                 } else {
1036                         hlist_for_each_entry(de, l, d_u.d_alias)
1037                                 if (!d_unhashed(de))
1038                                         break;
1039                 }
1040         }
1041         spin_unlock(&inode->i_lock);
1042         return de;
1043 }
1044
1045 /*
1046  *      Try to kill dentries associated with this inode.
1047  * WARNING: you must own a reference to inode.
1048  */
1049 void d_prune_aliases(struct inode *inode)
1050 {
1051         LIST_HEAD(dispose);
1052         struct dentry *dentry;
1053
1054         spin_lock(&inode->i_lock);
1055         hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
1056                 spin_lock(&dentry->d_lock);
1057                 if (!dentry->d_lockref.count)
1058                         to_shrink_list(dentry, &dispose);
1059                 spin_unlock(&dentry->d_lock);
1060         }
1061         spin_unlock(&inode->i_lock);
1062         shrink_dentry_list(&dispose);
1063 }
1064 EXPORT_SYMBOL(d_prune_aliases);
1065
1066 static inline void shrink_kill(struct dentry *victim)
1067 {
1068         do {
1069                 rcu_read_unlock();
1070                 victim = __dentry_kill(victim);
1071                 rcu_read_lock();
1072         } while (victim && lock_for_kill(victim));
1073         rcu_read_unlock();
1074         if (victim)
1075                 spin_unlock(&victim->d_lock);
1076 }
1077
1078 void shrink_dentry_list(struct list_head *list)
1079 {
1080         while (!list_empty(list)) {
1081                 struct dentry *dentry;
1082
1083                 dentry = list_entry(list->prev, struct dentry, d_lru);
1084                 spin_lock(&dentry->d_lock);
1085                 rcu_read_lock();
1086                 if (!lock_for_kill(dentry)) {
1087                         bool can_free;
1088                         rcu_read_unlock();
1089                         d_shrink_del(dentry);
1090                         can_free = dentry->d_flags & DCACHE_DENTRY_KILLED;
1091                         spin_unlock(&dentry->d_lock);
1092                         if (can_free)
1093                                 dentry_free(dentry);
1094                         continue;
1095                 }
1096                 d_shrink_del(dentry);
1097                 shrink_kill(dentry);
1098         }
1099 }
1100
1101 static enum lru_status dentry_lru_isolate(struct list_head *item,
1102                 struct list_lru_one *lru, void *arg)
1103 {
1104         struct list_head *freeable = arg;
1105         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1106
1107
1108         /*
1109          * we are inverting the lru lock/dentry->d_lock here,
1110          * so use a trylock. If we fail to get the lock, just skip
1111          * it
1112          */
1113         if (!spin_trylock(&dentry->d_lock))
1114                 return LRU_SKIP;
1115
1116         /*
1117          * Referenced dentries are still in use. If they have active
1118          * counts, just remove them from the LRU. Otherwise give them
1119          * another pass through the LRU.
1120          */
1121         if (dentry->d_lockref.count) {
1122                 d_lru_isolate(lru, dentry);
1123                 spin_unlock(&dentry->d_lock);
1124                 return LRU_REMOVED;
1125         }
1126
1127         if (dentry->d_flags & DCACHE_REFERENCED) {
1128                 dentry->d_flags &= ~DCACHE_REFERENCED;
1129                 spin_unlock(&dentry->d_lock);
1130
1131                 /*
1132                  * The list move itself will be made by the common LRU code. At
1133                  * this point, we've dropped the dentry->d_lock but keep the
1134                  * lru lock. This is safe to do, since every list movement is
1135                  * protected by the lru lock even if both locks are held.
1136                  *
1137                  * This is guaranteed by the fact that all LRU management
1138                  * functions are intermediated by the LRU API calls like
1139                  * list_lru_add_obj and list_lru_del_obj. List movement in this file
1140                  * only ever occur through this functions or through callbacks
1141                  * like this one, that are called from the LRU API.
1142                  *
1143                  * The only exceptions to this are functions like
1144                  * shrink_dentry_list, and code that first checks for the
1145                  * DCACHE_SHRINK_LIST flag.  Those are guaranteed to be
1146                  * operating only with stack provided lists after they are
1147                  * properly isolated from the main list.  It is thus, always a
1148                  * local access.
1149                  */
1150                 return LRU_ROTATE;
1151         }
1152
1153         d_lru_shrink_move(lru, dentry, freeable);
1154         spin_unlock(&dentry->d_lock);
1155
1156         return LRU_REMOVED;
1157 }
1158
1159 /**
1160  * prune_dcache_sb - shrink the dcache
1161  * @sb: superblock
1162  * @sc: shrink control, passed to list_lru_shrink_walk()
1163  *
1164  * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1165  * is done when we need more memory and called from the superblock shrinker
1166  * function.
1167  *
1168  * This function may fail to free any resources if all the dentries are in
1169  * use.
1170  */
1171 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1172 {
1173         LIST_HEAD(dispose);
1174         long freed;
1175
1176         freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1177                                      dentry_lru_isolate, &dispose);
1178         shrink_dentry_list(&dispose);
1179         return freed;
1180 }
1181
1182 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1183                 struct list_lru_one *lru, void *arg)
1184 {
1185         struct list_head *freeable = arg;
1186         struct dentry   *dentry = container_of(item, struct dentry, d_lru);
1187
1188         /*
1189          * we are inverting the lru lock/dentry->d_lock here,
1190          * so use a trylock. If we fail to get the lock, just skip
1191          * it
1192          */
1193         if (!spin_trylock(&dentry->d_lock))
1194                 return LRU_SKIP;
1195
1196         d_lru_shrink_move(lru, dentry, freeable);
1197         spin_unlock(&dentry->d_lock);
1198
1199         return LRU_REMOVED;
1200 }
1201
1202
1203 /**
1204  * shrink_dcache_sb - shrink dcache for a superblock
1205  * @sb: superblock
1206  *
1207  * Shrink the dcache for the specified super block. This is used to free
1208  * the dcache before unmounting a file system.
1209  */
1210 void shrink_dcache_sb(struct super_block *sb)
1211 {
1212         do {
1213                 LIST_HEAD(dispose);
1214
1215                 list_lru_walk(&sb->s_dentry_lru,
1216                         dentry_lru_isolate_shrink, &dispose, 1024);
1217                 shrink_dentry_list(&dispose);
1218         } while (list_lru_count(&sb->s_dentry_lru) > 0);
1219 }
1220 EXPORT_SYMBOL(shrink_dcache_sb);
1221
1222 /**
1223  * enum d_walk_ret - action to talke during tree walk
1224  * @D_WALK_CONTINUE:    contrinue walk
1225  * @D_WALK_QUIT:        quit walk
1226  * @D_WALK_NORETRY:     quit when retry is needed
1227  * @D_WALK_SKIP:        skip this dentry and its children
1228  */
1229 enum d_walk_ret {
1230         D_WALK_CONTINUE,
1231         D_WALK_QUIT,
1232         D_WALK_NORETRY,
1233         D_WALK_SKIP,
1234 };
1235
1236 /**
1237  * d_walk - walk the dentry tree
1238  * @parent:     start of walk
1239  * @data:       data passed to @enter() and @finish()
1240  * @enter:      callback when first entering the dentry
1241  *
1242  * The @enter() callbacks are called with d_lock held.
1243  */
1244 static void d_walk(struct dentry *parent, void *data,
1245                    enum d_walk_ret (*enter)(void *, struct dentry *))
1246 {
1247         struct dentry *this_parent, *dentry;
1248         unsigned seq = 0;
1249         enum d_walk_ret ret;
1250         bool retry = true;
1251
1252 again:
1253         read_seqbegin_or_lock(&rename_lock, &seq);
1254         this_parent = parent;
1255         spin_lock(&this_parent->d_lock);
1256
1257         ret = enter(data, this_parent);
1258         switch (ret) {
1259         case D_WALK_CONTINUE:
1260                 break;
1261         case D_WALK_QUIT:
1262         case D_WALK_SKIP:
1263                 goto out_unlock;
1264         case D_WALK_NORETRY:
1265                 retry = false;
1266                 break;
1267         }
1268 repeat:
1269         dentry = d_first_child(this_parent);
1270 resume:
1271         hlist_for_each_entry_from(dentry, d_sib) {
1272                 if (unlikely(dentry->d_flags & DCACHE_DENTRY_CURSOR))
1273                         continue;
1274
1275                 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1276
1277                 ret = enter(data, dentry);
1278                 switch (ret) {
1279                 case D_WALK_CONTINUE:
1280                         break;
1281                 case D_WALK_QUIT:
1282                         spin_unlock(&dentry->d_lock);
1283                         goto out_unlock;
1284                 case D_WALK_NORETRY:
1285                         retry = false;
1286                         break;
1287                 case D_WALK_SKIP:
1288                         spin_unlock(&dentry->d_lock);
1289                         continue;
1290                 }
1291
1292                 if (!hlist_empty(&dentry->d_children)) {
1293                         spin_unlock(&this_parent->d_lock);
1294                         spin_release(&dentry->d_lock.dep_map, _RET_IP_);
1295                         this_parent = dentry;
1296                         spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1297                         goto repeat;
1298                 }
1299                 spin_unlock(&dentry->d_lock);
1300         }
1301         /*
1302          * All done at this level ... ascend and resume the search.
1303          */
1304         rcu_read_lock();
1305 ascend:
1306         if (this_parent != parent) {
1307                 dentry = this_parent;
1308                 this_parent = dentry->d_parent;
1309
1310                 spin_unlock(&dentry->d_lock);
1311                 spin_lock(&this_parent->d_lock);
1312
1313                 /* might go back up the wrong parent if we have had a rename. */
1314                 if (need_seqretry(&rename_lock, seq))
1315                         goto rename_retry;
1316                 /* go into the first sibling still alive */
1317                 hlist_for_each_entry_continue(dentry, d_sib) {
1318                         if (likely(!(dentry->d_flags & DCACHE_DENTRY_KILLED))) {
1319                                 rcu_read_unlock();
1320                                 goto resume;
1321                         }
1322                 }
1323                 goto ascend;
1324         }
1325         if (need_seqretry(&rename_lock, seq))
1326                 goto rename_retry;
1327         rcu_read_unlock();
1328
1329 out_unlock:
1330         spin_unlock(&this_parent->d_lock);
1331         done_seqretry(&rename_lock, seq);
1332         return;
1333
1334 rename_retry:
1335         spin_unlock(&this_parent->d_lock);
1336         rcu_read_unlock();
1337         BUG_ON(seq & 1);
1338         if (!retry)
1339                 return;
1340         seq = 1;
1341         goto again;
1342 }
1343
1344 struct check_mount {
1345         struct vfsmount *mnt;
1346         unsigned int mounted;
1347 };
1348
1349 static enum d_walk_ret path_check_mount(void *data, struct dentry *dentry)
1350 {
1351         struct check_mount *info = data;
1352         struct path path = { .mnt = info->mnt, .dentry = dentry };
1353
1354         if (likely(!d_mountpoint(dentry)))
1355                 return D_WALK_CONTINUE;
1356         if (__path_is_mountpoint(&path)) {
1357                 info->mounted = 1;
1358                 return D_WALK_QUIT;
1359         }
1360         return D_WALK_CONTINUE;
1361 }
1362
1363 /**
1364  * path_has_submounts - check for mounts over a dentry in the
1365  *                      current namespace.
1366  * @parent: path to check.
1367  *
1368  * Return true if the parent or its subdirectories contain
1369  * a mount point in the current namespace.
1370  */
1371 int path_has_submounts(const struct path *parent)
1372 {
1373         struct check_mount data = { .mnt = parent->mnt, .mounted = 0 };
1374
1375         read_seqlock_excl(&mount_lock);
1376         d_walk(parent->dentry, &data, path_check_mount);
1377         read_sequnlock_excl(&mount_lock);
1378
1379         return data.mounted;
1380 }
1381 EXPORT_SYMBOL(path_has_submounts);
1382
1383 /*
1384  * Called by mount code to set a mountpoint and check if the mountpoint is
1385  * reachable (e.g. NFS can unhash a directory dentry and then the complete
1386  * subtree can become unreachable).
1387  *
1388  * Only one of d_invalidate() and d_set_mounted() must succeed.  For
1389  * this reason take rename_lock and d_lock on dentry and ancestors.
1390  */
1391 int d_set_mounted(struct dentry *dentry)
1392 {
1393         struct dentry *p;
1394         int ret = -ENOENT;
1395         write_seqlock(&rename_lock);
1396         for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1397                 /* Need exclusion wrt. d_invalidate() */
1398                 spin_lock(&p->d_lock);
1399                 if (unlikely(d_unhashed(p))) {
1400                         spin_unlock(&p->d_lock);
1401                         goto out;
1402                 }
1403                 spin_unlock(&p->d_lock);
1404         }
1405         spin_lock(&dentry->d_lock);
1406         if (!d_unlinked(dentry)) {
1407                 ret = -EBUSY;
1408                 if (!d_mountpoint(dentry)) {
1409                         dentry->d_flags |= DCACHE_MOUNTED;
1410                         ret = 0;
1411                 }
1412         }
1413         spin_unlock(&dentry->d_lock);
1414 out:
1415         write_sequnlock(&rename_lock);
1416         return ret;
1417 }
1418
1419 /*
1420  * Search the dentry child list of the specified parent,
1421  * and move any unused dentries to the end of the unused
1422  * list for prune_dcache(). We descend to the next level
1423  * whenever the d_children list is non-empty and continue
1424  * searching.
1425  *
1426  * It returns zero iff there are no unused children,
1427  * otherwise  it returns the number of children moved to
1428  * the end of the unused list. This may not be the total
1429  * number of unused children, because select_parent can
1430  * drop the lock and return early due to latency
1431  * constraints.
1432  */
1433
1434 struct select_data {
1435         struct dentry *start;
1436         union {
1437                 long found;
1438                 struct dentry *victim;
1439         };
1440         struct list_head dispose;
1441 };
1442
1443 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1444 {
1445         struct select_data *data = _data;
1446         enum d_walk_ret ret = D_WALK_CONTINUE;
1447
1448         if (data->start == dentry)
1449                 goto out;
1450
1451         if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1452                 data->found++;
1453         } else if (!dentry->d_lockref.count) {
1454                 to_shrink_list(dentry, &data->dispose);
1455                 data->found++;
1456         } else if (dentry->d_lockref.count < 0) {
1457                 data->found++;
1458         }
1459         /*
1460          * We can return to the caller if we have found some (this
1461          * ensures forward progress). We'll be coming back to find
1462          * the rest.
1463          */
1464         if (!list_empty(&data->dispose))
1465                 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1466 out:
1467         return ret;
1468 }
1469
1470 static enum d_walk_ret select_collect2(void *_data, struct dentry *dentry)
1471 {
1472         struct select_data *data = _data;
1473         enum d_walk_ret ret = D_WALK_CONTINUE;
1474
1475         if (data->start == dentry)
1476                 goto out;
1477
1478         if (!dentry->d_lockref.count) {
1479                 if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1480                         rcu_read_lock();
1481                         data->victim = dentry;
1482                         return D_WALK_QUIT;
1483                 }
1484                 to_shrink_list(dentry, &data->dispose);
1485         }
1486         /*
1487          * We can return to the caller if we have found some (this
1488          * ensures forward progress). We'll be coming back to find
1489          * the rest.
1490          */
1491         if (!list_empty(&data->dispose))
1492                 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1493 out:
1494         return ret;
1495 }
1496
1497 /**
1498  * shrink_dcache_parent - prune dcache
1499  * @parent: parent of entries to prune
1500  *
1501  * Prune the dcache to remove unused children of the parent dentry.
1502  */
1503 void shrink_dcache_parent(struct dentry *parent)
1504 {
1505         for (;;) {
1506                 struct select_data data = {.start = parent};
1507
1508                 INIT_LIST_HEAD(&data.dispose);
1509                 d_walk(parent, &data, select_collect);
1510
1511                 if (!list_empty(&data.dispose)) {
1512                         shrink_dentry_list(&data.dispose);
1513                         continue;
1514                 }
1515
1516                 cond_resched();
1517                 if (!data.found)
1518                         break;
1519                 data.victim = NULL;
1520                 d_walk(parent, &data, select_collect2);
1521                 if (data.victim) {
1522                         spin_lock(&data.victim->d_lock);
1523                         if (!lock_for_kill(data.victim)) {
1524                                 spin_unlock(&data.victim->d_lock);
1525                                 rcu_read_unlock();
1526                         } else {
1527                                 shrink_kill(data.victim);
1528                         }
1529                 }
1530                 if (!list_empty(&data.dispose))
1531                         shrink_dentry_list(&data.dispose);
1532         }
1533 }
1534 EXPORT_SYMBOL(shrink_dcache_parent);
1535
1536 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1537 {
1538         /* it has busy descendents; complain about those instead */
1539         if (!hlist_empty(&dentry->d_children))
1540                 return D_WALK_CONTINUE;
1541
1542         /* root with refcount 1 is fine */
1543         if (dentry == _data && dentry->d_lockref.count == 1)
1544                 return D_WALK_CONTINUE;
1545
1546         WARN(1, "BUG: Dentry %p{i=%lx,n=%pd} "
1547                         " still in use (%d) [unmount of %s %s]\n",
1548                        dentry,
1549                        dentry->d_inode ?
1550                        dentry->d_inode->i_ino : 0UL,
1551                        dentry,
1552                        dentry->d_lockref.count,
1553                        dentry->d_sb->s_type->name,
1554                        dentry->d_sb->s_id);
1555         return D_WALK_CONTINUE;
1556 }
1557
1558 static void do_one_tree(struct dentry *dentry)
1559 {
1560         shrink_dcache_parent(dentry);
1561         d_walk(dentry, dentry, umount_check);
1562         d_drop(dentry);
1563         dput(dentry);
1564 }
1565
1566 /*
1567  * destroy the dentries attached to a superblock on unmounting
1568  */
1569 void shrink_dcache_for_umount(struct super_block *sb)
1570 {
1571         struct dentry *dentry;
1572
1573         rwsem_assert_held_write(&sb->s_umount);
1574
1575         dentry = sb->s_root;
1576         sb->s_root = NULL;
1577         do_one_tree(dentry);
1578
1579         while (!hlist_bl_empty(&sb->s_roots)) {
1580                 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_roots), struct dentry, d_hash));
1581                 do_one_tree(dentry);
1582         }
1583 }
1584
1585 static enum d_walk_ret find_submount(void *_data, struct dentry *dentry)
1586 {
1587         struct dentry **victim = _data;
1588         if (d_mountpoint(dentry)) {
1589                 *victim = dget_dlock(dentry);
1590                 return D_WALK_QUIT;
1591         }
1592         return D_WALK_CONTINUE;
1593 }
1594
1595 /**
1596  * d_invalidate - detach submounts, prune dcache, and drop
1597  * @dentry: dentry to invalidate (aka detach, prune and drop)
1598  */
1599 void d_invalidate(struct dentry *dentry)
1600 {
1601         bool had_submounts = false;
1602         spin_lock(&dentry->d_lock);
1603         if (d_unhashed(dentry)) {
1604                 spin_unlock(&dentry->d_lock);
1605                 return;
1606         }
1607         __d_drop(dentry);
1608         spin_unlock(&dentry->d_lock);
1609
1610         /* Negative dentries can be dropped without further checks */
1611         if (!dentry->d_inode)
1612                 return;
1613
1614         shrink_dcache_parent(dentry);
1615         for (;;) {
1616                 struct dentry *victim = NULL;
1617                 d_walk(dentry, &victim, find_submount);
1618                 if (!victim) {
1619                         if (had_submounts)
1620                                 shrink_dcache_parent(dentry);
1621                         return;
1622                 }
1623                 had_submounts = true;
1624                 detach_mounts(victim);
1625                 dput(victim);
1626         }
1627 }
1628 EXPORT_SYMBOL(d_invalidate);
1629
1630 /**
1631  * __d_alloc    -       allocate a dcache entry
1632  * @sb: filesystem it will belong to
1633  * @name: qstr of the name
1634  *
1635  * Allocates a dentry. It returns %NULL if there is insufficient memory
1636  * available. On a success the dentry is returned. The name passed in is
1637  * copied and the copy passed in may be reused after this call.
1638  */
1639  
1640 static struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1641 {
1642         struct dentry *dentry;
1643         char *dname;
1644         int err;
1645
1646         dentry = kmem_cache_alloc_lru(dentry_cache, &sb->s_dentry_lru,
1647                                       GFP_KERNEL);
1648         if (!dentry)
1649                 return NULL;
1650
1651         /*
1652          * We guarantee that the inline name is always NUL-terminated.
1653          * This way the memcpy() done by the name switching in rename
1654          * will still always have a NUL at the end, even if we might
1655          * be overwriting an internal NUL character
1656          */
1657         dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1658         if (unlikely(!name)) {
1659                 name = &slash_name;
1660                 dname = dentry->d_iname;
1661         } else if (name->len > DNAME_INLINE_LEN-1) {
1662                 size_t size = offsetof(struct external_name, name[1]);
1663                 struct external_name *p = kmalloc(size + name->len,
1664                                                   GFP_KERNEL_ACCOUNT |
1665                                                   __GFP_RECLAIMABLE);
1666                 if (!p) {
1667                         kmem_cache_free(dentry_cache, dentry); 
1668                         return NULL;
1669                 }
1670                 atomic_set(&p->u.count, 1);
1671                 dname = p->name;
1672         } else  {
1673                 dname = dentry->d_iname;
1674         }       
1675
1676         dentry->d_name.len = name->len;
1677         dentry->d_name.hash = name->hash;
1678         memcpy(dname, name->name, name->len);
1679         dname[name->len] = 0;
1680
1681         /* Make sure we always see the terminating NUL character */
1682         smp_store_release(&dentry->d_name.name, dname); /* ^^^ */
1683
1684         dentry->d_lockref.count = 1;
1685         dentry->d_flags = 0;
1686         spin_lock_init(&dentry->d_lock);
1687         seqcount_spinlock_init(&dentry->d_seq, &dentry->d_lock);
1688         dentry->d_inode = NULL;
1689         dentry->d_parent = dentry;
1690         dentry->d_sb = sb;
1691         dentry->d_op = NULL;
1692         dentry->d_fsdata = NULL;
1693         INIT_HLIST_BL_NODE(&dentry->d_hash);
1694         INIT_LIST_HEAD(&dentry->d_lru);
1695         INIT_HLIST_HEAD(&dentry->d_children);
1696         INIT_HLIST_NODE(&dentry->d_u.d_alias);
1697         INIT_HLIST_NODE(&dentry->d_sib);
1698         d_set_d_op(dentry, dentry->d_sb->s_d_op);
1699
1700         if (dentry->d_op && dentry->d_op->d_init) {
1701                 err = dentry->d_op->d_init(dentry);
1702                 if (err) {
1703                         if (dname_external(dentry))
1704                                 kfree(external_name(dentry));
1705                         kmem_cache_free(dentry_cache, dentry);
1706                         return NULL;
1707                 }
1708         }
1709
1710         this_cpu_inc(nr_dentry);
1711
1712         return dentry;
1713 }
1714
1715 /**
1716  * d_alloc      -       allocate a dcache entry
1717  * @parent: parent of entry to allocate
1718  * @name: qstr of the name
1719  *
1720  * Allocates a dentry. It returns %NULL if there is insufficient memory
1721  * available. On a success the dentry is returned. The name passed in is
1722  * copied and the copy passed in may be reused after this call.
1723  */
1724 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1725 {
1726         struct dentry *dentry = __d_alloc(parent->d_sb, name);
1727         if (!dentry)
1728                 return NULL;
1729         spin_lock(&parent->d_lock);
1730         /*
1731          * don't need child lock because it is not subject
1732          * to concurrency here
1733          */
1734         dentry->d_parent = dget_dlock(parent);
1735         hlist_add_head(&dentry->d_sib, &parent->d_children);
1736         spin_unlock(&parent->d_lock);
1737
1738         return dentry;
1739 }
1740 EXPORT_SYMBOL(d_alloc);
1741
1742 struct dentry *d_alloc_anon(struct super_block *sb)
1743 {
1744         return __d_alloc(sb, NULL);
1745 }
1746 EXPORT_SYMBOL(d_alloc_anon);
1747
1748 struct dentry *d_alloc_cursor(struct dentry * parent)
1749 {
1750         struct dentry *dentry = d_alloc_anon(parent->d_sb);
1751         if (dentry) {
1752                 dentry->d_flags |= DCACHE_DENTRY_CURSOR;
1753                 dentry->d_parent = dget(parent);
1754         }
1755         return dentry;
1756 }
1757
1758 /**
1759  * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1760  * @sb: the superblock
1761  * @name: qstr of the name
1762  *
1763  * For a filesystem that just pins its dentries in memory and never
1764  * performs lookups at all, return an unhashed IS_ROOT dentry.
1765  * This is used for pipes, sockets et.al. - the stuff that should
1766  * never be anyone's children or parents.  Unlike all other
1767  * dentries, these will not have RCU delay between dropping the
1768  * last reference and freeing them.
1769  *
1770  * The only user is alloc_file_pseudo() and that's what should
1771  * be considered a public interface.  Don't use directly.
1772  */
1773 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1774 {
1775         static const struct dentry_operations anon_ops = {
1776                 .d_dname = simple_dname
1777         };
1778         struct dentry *dentry = __d_alloc(sb, name);
1779         if (likely(dentry)) {
1780                 dentry->d_flags |= DCACHE_NORCU;
1781                 if (!sb->s_d_op)
1782                         d_set_d_op(dentry, &anon_ops);
1783         }
1784         return dentry;
1785 }
1786
1787 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1788 {
1789         struct qstr q;
1790
1791         q.name = name;
1792         q.hash_len = hashlen_string(parent, name);
1793         return d_alloc(parent, &q);
1794 }
1795 EXPORT_SYMBOL(d_alloc_name);
1796
1797 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1798 {
1799         WARN_ON_ONCE(dentry->d_op);
1800         WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH  |
1801                                 DCACHE_OP_COMPARE       |
1802                                 DCACHE_OP_REVALIDATE    |
1803                                 DCACHE_OP_WEAK_REVALIDATE       |
1804                                 DCACHE_OP_DELETE        |
1805                                 DCACHE_OP_REAL));
1806         dentry->d_op = op;
1807         if (!op)
1808                 return;
1809         if (op->d_hash)
1810                 dentry->d_flags |= DCACHE_OP_HASH;
1811         if (op->d_compare)
1812                 dentry->d_flags |= DCACHE_OP_COMPARE;
1813         if (op->d_revalidate)
1814                 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1815         if (op->d_weak_revalidate)
1816                 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1817         if (op->d_delete)
1818                 dentry->d_flags |= DCACHE_OP_DELETE;
1819         if (op->d_prune)
1820                 dentry->d_flags |= DCACHE_OP_PRUNE;
1821         if (op->d_real)
1822                 dentry->d_flags |= DCACHE_OP_REAL;
1823
1824 }
1825 EXPORT_SYMBOL(d_set_d_op);
1826
1827 static unsigned d_flags_for_inode(struct inode *inode)
1828 {
1829         unsigned add_flags = DCACHE_REGULAR_TYPE;
1830
1831         if (!inode)
1832                 return DCACHE_MISS_TYPE;
1833
1834         if (S_ISDIR(inode->i_mode)) {
1835                 add_flags = DCACHE_DIRECTORY_TYPE;
1836                 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1837                         if (unlikely(!inode->i_op->lookup))
1838                                 add_flags = DCACHE_AUTODIR_TYPE;
1839                         else
1840                                 inode->i_opflags |= IOP_LOOKUP;
1841                 }
1842                 goto type_determined;
1843         }
1844
1845         if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1846                 if (unlikely(inode->i_op->get_link)) {
1847                         add_flags = DCACHE_SYMLINK_TYPE;
1848                         goto type_determined;
1849                 }
1850                 inode->i_opflags |= IOP_NOFOLLOW;
1851         }
1852
1853         if (unlikely(!S_ISREG(inode->i_mode)))
1854                 add_flags = DCACHE_SPECIAL_TYPE;
1855
1856 type_determined:
1857         if (unlikely(IS_AUTOMOUNT(inode)))
1858                 add_flags |= DCACHE_NEED_AUTOMOUNT;
1859         return add_flags;
1860 }
1861
1862 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1863 {
1864         unsigned add_flags = d_flags_for_inode(inode);
1865         WARN_ON(d_in_lookup(dentry));
1866
1867         spin_lock(&dentry->d_lock);
1868         /*
1869          * The negative counter only tracks dentries on the LRU. Don't dec if
1870          * d_lru is on another list.
1871          */
1872         if ((dentry->d_flags &
1873              (DCACHE_LRU_LIST|DCACHE_SHRINK_LIST)) == DCACHE_LRU_LIST)
1874                 this_cpu_dec(nr_dentry_negative);
1875         hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1876         raw_write_seqcount_begin(&dentry->d_seq);
1877         __d_set_inode_and_type(dentry, inode, add_flags);
1878         raw_write_seqcount_end(&dentry->d_seq);
1879         fsnotify_update_flags(dentry);
1880         spin_unlock(&dentry->d_lock);
1881 }
1882
1883 /**
1884  * d_instantiate - fill in inode information for a dentry
1885  * @entry: dentry to complete
1886  * @inode: inode to attach to this dentry
1887  *
1888  * Fill in inode information in the entry.
1889  *
1890  * This turns negative dentries into productive full members
1891  * of society.
1892  *
1893  * NOTE! This assumes that the inode count has been incremented
1894  * (or otherwise set) by the caller to indicate that it is now
1895  * in use by the dcache.
1896  */
1897  
1898 void d_instantiate(struct dentry *entry, struct inode * inode)
1899 {
1900         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1901         if (inode) {
1902                 security_d_instantiate(entry, inode);
1903                 spin_lock(&inode->i_lock);
1904                 __d_instantiate(entry, inode);
1905                 spin_unlock(&inode->i_lock);
1906         }
1907 }
1908 EXPORT_SYMBOL(d_instantiate);
1909
1910 /*
1911  * This should be equivalent to d_instantiate() + unlock_new_inode(),
1912  * with lockdep-related part of unlock_new_inode() done before
1913  * anything else.  Use that instead of open-coding d_instantiate()/
1914  * unlock_new_inode() combinations.
1915  */
1916 void d_instantiate_new(struct dentry *entry, struct inode *inode)
1917 {
1918         BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1919         BUG_ON(!inode);
1920         lockdep_annotate_inode_mutex_key(inode);
1921         security_d_instantiate(entry, inode);
1922         spin_lock(&inode->i_lock);
1923         __d_instantiate(entry, inode);
1924         WARN_ON(!(inode->i_state & I_NEW));
1925         inode->i_state &= ~I_NEW & ~I_CREATING;
1926         /*
1927          * Pairs with the barrier in prepare_to_wait_event() to make sure
1928          * ___wait_var_event() either sees the bit cleared or
1929          * waitqueue_active() check in wake_up_var() sees the waiter.
1930          */
1931         smp_mb();
1932         inode_wake_up_bit(inode, __I_NEW);
1933         spin_unlock(&inode->i_lock);
1934 }
1935 EXPORT_SYMBOL(d_instantiate_new);
1936
1937 struct dentry *d_make_root(struct inode *root_inode)
1938 {
1939         struct dentry *res = NULL;
1940
1941         if (root_inode) {
1942                 res = d_alloc_anon(root_inode->i_sb);
1943                 if (res)
1944                         d_instantiate(res, root_inode);
1945                 else
1946                         iput(root_inode);
1947         }
1948         return res;
1949 }
1950 EXPORT_SYMBOL(d_make_root);
1951
1952 static struct dentry *__d_obtain_alias(struct inode *inode, bool disconnected)
1953 {
1954         struct super_block *sb;
1955         struct dentry *new, *res;
1956
1957         if (!inode)
1958                 return ERR_PTR(-ESTALE);
1959         if (IS_ERR(inode))
1960                 return ERR_CAST(inode);
1961
1962         sb = inode->i_sb;
1963
1964         res = d_find_any_alias(inode); /* existing alias? */
1965         if (res)
1966                 goto out;
1967
1968         new = d_alloc_anon(sb);
1969         if (!new) {
1970                 res = ERR_PTR(-ENOMEM);
1971                 goto out;
1972         }
1973
1974         security_d_instantiate(new, inode);
1975         spin_lock(&inode->i_lock);
1976         res = __d_find_any_alias(inode); /* recheck under lock */
1977         if (likely(!res)) { /* still no alias, attach a disconnected dentry */
1978                 unsigned add_flags = d_flags_for_inode(inode);
1979
1980                 if (disconnected)
1981                         add_flags |= DCACHE_DISCONNECTED;
1982
1983                 spin_lock(&new->d_lock);
1984                 __d_set_inode_and_type(new, inode, add_flags);
1985                 hlist_add_head(&new->d_u.d_alias, &inode->i_dentry);
1986                 if (!disconnected) {
1987                         hlist_bl_lock(&sb->s_roots);
1988                         hlist_bl_add_head(&new->d_hash, &sb->s_roots);
1989                         hlist_bl_unlock(&sb->s_roots);
1990                 }
1991                 spin_unlock(&new->d_lock);
1992                 spin_unlock(&inode->i_lock);
1993                 inode = NULL; /* consumed by new->d_inode */
1994                 res = new;
1995         } else {
1996                 spin_unlock(&inode->i_lock);
1997                 dput(new);
1998         }
1999
2000  out:
2001         iput(inode);
2002         return res;
2003 }
2004
2005 /**
2006  * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
2007  * @inode: inode to allocate the dentry for
2008  *
2009  * Obtain a dentry for an inode resulting from NFS filehandle conversion or
2010  * similar open by handle operations.  The returned dentry may be anonymous,
2011  * or may have a full name (if the inode was already in the cache).
2012  *
2013  * When called on a directory inode, we must ensure that the inode only ever
2014  * has one dentry.  If a dentry is found, that is returned instead of
2015  * allocating a new one.
2016  *
2017  * On successful return, the reference to the inode has been transferred
2018  * to the dentry.  In case of an error the reference on the inode is released.
2019  * To make it easier to use in export operations a %NULL or IS_ERR inode may
2020  * be passed in and the error will be propagated to the return value,
2021  * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
2022  */
2023 struct dentry *d_obtain_alias(struct inode *inode)
2024 {
2025         return __d_obtain_alias(inode, true);
2026 }
2027 EXPORT_SYMBOL(d_obtain_alias);
2028
2029 /**
2030  * d_obtain_root - find or allocate a dentry for a given inode
2031  * @inode: inode to allocate the dentry for
2032  *
2033  * Obtain an IS_ROOT dentry for the root of a filesystem.
2034  *
2035  * We must ensure that directory inodes only ever have one dentry.  If a
2036  * dentry is found, that is returned instead of allocating a new one.
2037  *
2038  * On successful return, the reference to the inode has been transferred
2039  * to the dentry.  In case of an error the reference on the inode is
2040  * released.  A %NULL or IS_ERR inode may be passed in and will be the
2041  * error will be propagate to the return value, with a %NULL @inode
2042  * replaced by ERR_PTR(-ESTALE).
2043  */
2044 struct dentry *d_obtain_root(struct inode *inode)
2045 {
2046         return __d_obtain_alias(inode, false);
2047 }
2048 EXPORT_SYMBOL(d_obtain_root);
2049
2050 /**
2051  * d_add_ci - lookup or allocate new dentry with case-exact name
2052  * @dentry: the negative dentry that was passed to the parent's lookup func
2053  * @inode:  the inode case-insensitive lookup has found
2054  * @name:   the case-exact name to be associated with the returned dentry
2055  *
2056  * This is to avoid filling the dcache with case-insensitive names to the
2057  * same inode, only the actual correct case is stored in the dcache for
2058  * case-insensitive filesystems.
2059  *
2060  * For a case-insensitive lookup match and if the case-exact dentry
2061  * already exists in the dcache, use it and return it.
2062  *
2063  * If no entry exists with the exact case name, allocate new dentry with
2064  * the exact case, and return the spliced entry.
2065  */
2066 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
2067                         struct qstr *name)
2068 {
2069         struct dentry *found, *res;
2070
2071         /*
2072          * First check if a dentry matching the name already exists,
2073          * if not go ahead and create it now.
2074          */
2075         found = d_hash_and_lookup(dentry->d_parent, name);
2076         if (found) {
2077                 iput(inode);
2078                 return found;
2079         }
2080         if (d_in_lookup(dentry)) {
2081                 found = d_alloc_parallel(dentry->d_parent, name,
2082                                         dentry->d_wait);
2083                 if (IS_ERR(found) || !d_in_lookup(found)) {
2084                         iput(inode);
2085                         return found;
2086                 }
2087         } else {
2088                 found = d_alloc(dentry->d_parent, name);
2089                 if (!found) {
2090                         iput(inode);
2091                         return ERR_PTR(-ENOMEM);
2092                 } 
2093         }
2094         res = d_splice_alias(inode, found);
2095         if (res) {
2096                 d_lookup_done(found);
2097                 dput(found);
2098                 return res;
2099         }
2100         return found;
2101 }
2102 EXPORT_SYMBOL(d_add_ci);
2103
2104 /**
2105  * d_same_name - compare dentry name with case-exact name
2106  * @dentry: the negative dentry that was passed to the parent's lookup func
2107  * @parent: parent dentry
2108  * @name:   the case-exact name to be associated with the returned dentry
2109  *
2110  * Return: true if names are same, or false
2111  */
2112 bool d_same_name(const struct dentry *dentry, const struct dentry *parent,
2113                  const struct qstr *name)
2114 {
2115         if (likely(!(parent->d_flags & DCACHE_OP_COMPARE))) {
2116                 if (dentry->d_name.len != name->len)
2117                         return false;
2118                 return dentry_cmp(dentry, name->name, name->len) == 0;
2119         }
2120         return parent->d_op->d_compare(dentry,
2121                                        dentry->d_name.len, dentry->d_name.name,
2122                                        name) == 0;
2123 }
2124 EXPORT_SYMBOL_GPL(d_same_name);
2125
2126 /*
2127  * This is __d_lookup_rcu() when the parent dentry has
2128  * DCACHE_OP_COMPARE, which makes things much nastier.
2129  */
2130 static noinline struct dentry *__d_lookup_rcu_op_compare(
2131         const struct dentry *parent,
2132         const struct qstr *name,
2133         unsigned *seqp)
2134 {
2135         u64 hashlen = name->hash_len;
2136         struct hlist_bl_head *b = d_hash(hashlen);
2137         struct hlist_bl_node *node;
2138         struct dentry *dentry;
2139
2140         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2141                 int tlen;
2142                 const char *tname;
2143                 unsigned seq;
2144
2145 seqretry:
2146                 seq = raw_seqcount_begin(&dentry->d_seq);
2147                 if (dentry->d_parent != parent)
2148                         continue;
2149                 if (d_unhashed(dentry))
2150                         continue;
2151                 if (dentry->d_name.hash != hashlen_hash(hashlen))
2152                         continue;
2153                 tlen = dentry->d_name.len;
2154                 tname = dentry->d_name.name;
2155                 /* we want a consistent (name,len) pair */
2156                 if (read_seqcount_retry(&dentry->d_seq, seq)) {
2157                         cpu_relax();
2158                         goto seqretry;
2159                 }
2160                 if (parent->d_op->d_compare(dentry, tlen, tname, name) != 0)
2161                         continue;
2162                 *seqp = seq;
2163                 return dentry;
2164         }
2165         return NULL;
2166 }
2167
2168 /**
2169  * __d_lookup_rcu - search for a dentry (racy, store-free)
2170  * @parent: parent dentry
2171  * @name: qstr of name we wish to find
2172  * @seqp: returns d_seq value at the point where the dentry was found
2173  * Returns: dentry, or NULL
2174  *
2175  * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2176  * resolution (store-free path walking) design described in
2177  * Documentation/filesystems/path-lookup.txt.
2178  *
2179  * This is not to be used outside core vfs.
2180  *
2181  * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2182  * held, and rcu_read_lock held. The returned dentry must not be stored into
2183  * without taking d_lock and checking d_seq sequence count against @seq
2184  * returned here.
2185  *
2186  * Alternatively, __d_lookup_rcu may be called again to look up the child of
2187  * the returned dentry, so long as its parent's seqlock is checked after the
2188  * child is looked up. Thus, an interlocking stepping of sequence lock checks
2189  * is formed, giving integrity down the path walk.
2190  *
2191  * NOTE! The caller *has* to check the resulting dentry against the sequence
2192  * number we've returned before using any of the resulting dentry state!
2193  */
2194 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2195                                 const struct qstr *name,
2196                                 unsigned *seqp)
2197 {
2198         u64 hashlen = name->hash_len;
2199         const unsigned char *str = name->name;
2200         struct hlist_bl_head *b = d_hash(hashlen);
2201         struct hlist_bl_node *node;
2202         struct dentry *dentry;
2203
2204         /*
2205          * Note: There is significant duplication with __d_lookup_rcu which is
2206          * required to prevent single threaded performance regressions
2207          * especially on architectures where smp_rmb (in seqcounts) are costly.
2208          * Keep the two functions in sync.
2209          */
2210
2211         if (unlikely(parent->d_flags & DCACHE_OP_COMPARE))
2212                 return __d_lookup_rcu_op_compare(parent, name, seqp);
2213
2214         /*
2215          * The hash list is protected using RCU.
2216          *
2217          * Carefully use d_seq when comparing a candidate dentry, to avoid
2218          * races with d_move().
2219          *
2220          * It is possible that concurrent renames can mess up our list
2221          * walk here and result in missing our dentry, resulting in the
2222          * false-negative result. d_lookup() protects against concurrent
2223          * renames using rename_lock seqlock.
2224          *
2225          * See Documentation/filesystems/path-lookup.txt for more details.
2226          */
2227         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2228                 unsigned seq;
2229
2230                 /*
2231                  * The dentry sequence count protects us from concurrent
2232                  * renames, and thus protects parent and name fields.
2233                  *
2234                  * The caller must perform a seqcount check in order
2235                  * to do anything useful with the returned dentry.
2236                  *
2237                  * NOTE! We do a "raw" seqcount_begin here. That means that
2238                  * we don't wait for the sequence count to stabilize if it
2239                  * is in the middle of a sequence change. If we do the slow
2240                  * dentry compare, we will do seqretries until it is stable,
2241                  * and if we end up with a successful lookup, we actually
2242                  * want to exit RCU lookup anyway.
2243                  *
2244                  * Note that raw_seqcount_begin still *does* smp_rmb(), so
2245                  * we are still guaranteed NUL-termination of ->d_name.name.
2246                  */
2247                 seq = raw_seqcount_begin(&dentry->d_seq);
2248                 if (dentry->d_parent != parent)
2249                         continue;
2250                 if (d_unhashed(dentry))
2251                         continue;
2252                 if (dentry->d_name.hash_len != hashlen)
2253                         continue;
2254                 if (dentry_cmp(dentry, str, hashlen_len(hashlen)) != 0)
2255                         continue;
2256                 *seqp = seq;
2257                 return dentry;
2258         }
2259         return NULL;
2260 }
2261
2262 /**
2263  * d_lookup - search for a dentry
2264  * @parent: parent dentry
2265  * @name: qstr of name we wish to find
2266  * Returns: dentry, or NULL
2267  *
2268  * d_lookup searches the children of the parent dentry for the name in
2269  * question. If the dentry is found its reference count is incremented and the
2270  * dentry is returned. The caller must use dput to free the entry when it has
2271  * finished using it. %NULL is returned if the dentry does not exist.
2272  */
2273 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2274 {
2275         struct dentry *dentry;
2276         unsigned seq;
2277
2278         do {
2279                 seq = read_seqbegin(&rename_lock);
2280                 dentry = __d_lookup(parent, name);
2281                 if (dentry)
2282                         break;
2283         } while (read_seqretry(&rename_lock, seq));
2284         return dentry;
2285 }
2286 EXPORT_SYMBOL(d_lookup);
2287
2288 /**
2289  * __d_lookup - search for a dentry (racy)
2290  * @parent: parent dentry
2291  * @name: qstr of name we wish to find
2292  * Returns: dentry, or NULL
2293  *
2294  * __d_lookup is like d_lookup, however it may (rarely) return a
2295  * false-negative result due to unrelated rename activity.
2296  *
2297  * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2298  * however it must be used carefully, eg. with a following d_lookup in
2299  * the case of failure.
2300  *
2301  * __d_lookup callers must be commented.
2302  */
2303 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2304 {
2305         unsigned int hash = name->hash;
2306         struct hlist_bl_head *b = d_hash(hash);
2307         struct hlist_bl_node *node;
2308         struct dentry *found = NULL;
2309         struct dentry *dentry;
2310
2311         /*
2312          * Note: There is significant duplication with __d_lookup_rcu which is
2313          * required to prevent single threaded performance regressions
2314          * especially on architectures where smp_rmb (in seqcounts) are costly.
2315          * Keep the two functions in sync.
2316          */
2317
2318         /*
2319          * The hash list is protected using RCU.
2320          *
2321          * Take d_lock when comparing a candidate dentry, to avoid races
2322          * with d_move().
2323          *
2324          * It is possible that concurrent renames can mess up our list
2325          * walk here and result in missing our dentry, resulting in the
2326          * false-negative result. d_lookup() protects against concurrent
2327          * renames using rename_lock seqlock.
2328          *
2329          * See Documentation/filesystems/path-lookup.txt for more details.
2330          */
2331         rcu_read_lock();
2332         
2333         hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2334
2335                 if (dentry->d_name.hash != hash)
2336                         continue;
2337
2338                 spin_lock(&dentry->d_lock);
2339                 if (dentry->d_parent != parent)
2340                         goto next;
2341                 if (d_unhashed(dentry))
2342                         goto next;
2343
2344                 if (!d_same_name(dentry, parent, name))
2345                         goto next;
2346
2347                 dentry->d_lockref.count++;
2348                 found = dentry;
2349                 spin_unlock(&dentry->d_lock);
2350                 break;
2351 next:
2352                 spin_unlock(&dentry->d_lock);
2353         }
2354         rcu_read_unlock();
2355
2356         return found;
2357 }
2358
2359 /**
2360  * d_hash_and_lookup - hash the qstr then search for a dentry
2361  * @dir: Directory to search in
2362  * @name: qstr of name we wish to find
2363  *
2364  * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2365  */
2366 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2367 {
2368         /*
2369          * Check for a fs-specific hash function. Note that we must
2370          * calculate the standard hash first, as the d_op->d_hash()
2371          * routine may choose to leave the hash value unchanged.
2372          */
2373         name->hash = full_name_hash(dir, name->name, name->len);
2374         if (dir->d_flags & DCACHE_OP_HASH) {
2375                 int err = dir->d_op->d_hash(dir, name);
2376                 if (unlikely(err < 0))
2377                         return ERR_PTR(err);
2378         }
2379         return d_lookup(dir, name);
2380 }
2381 EXPORT_SYMBOL(d_hash_and_lookup);
2382
2383 /*
2384  * When a file is deleted, we have two options:
2385  * - turn this dentry into a negative dentry
2386  * - unhash this dentry and free it.
2387  *
2388  * Usually, we want to just turn this into
2389  * a negative dentry, but if anybody else is
2390  * currently using the dentry or the inode
2391  * we can't do that and we fall back on removing
2392  * it from the hash queues and waiting for
2393  * it to be deleted later when it has no users
2394  */
2395  
2396 /**
2397  * d_delete - delete a dentry
2398  * @dentry: The dentry to delete
2399  *
2400  * Turn the dentry into a negative dentry if possible, otherwise
2401  * remove it from the hash queues so it can be deleted later
2402  */
2403  
2404 void d_delete(struct dentry * dentry)
2405 {
2406         struct inode *inode = dentry->d_inode;
2407
2408         spin_lock(&inode->i_lock);
2409         spin_lock(&dentry->d_lock);
2410         /*
2411          * Are we the only user?
2412          */
2413         if (dentry->d_lockref.count == 1) {
2414                 if (dentry_negative_policy)
2415                         __d_drop(dentry);
2416                 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2417                 dentry_unlink_inode(dentry);
2418         } else {
2419                 __d_drop(dentry);
2420                 spin_unlock(&dentry->d_lock);
2421                 spin_unlock(&inode->i_lock);
2422         }
2423 }
2424 EXPORT_SYMBOL(d_delete);
2425
2426 static void __d_rehash(struct dentry *entry)
2427 {
2428         struct hlist_bl_head *b = d_hash(entry->d_name.hash);
2429
2430         hlist_bl_lock(b);
2431         hlist_bl_add_head_rcu(&entry->d_hash, b);
2432         hlist_bl_unlock(b);
2433 }
2434
2435 /**
2436  * d_rehash     - add an entry back to the hash
2437  * @entry: dentry to add to the hash
2438  *
2439  * Adds a dentry to the hash according to its name.
2440  */
2441  
2442 void d_rehash(struct dentry * entry)
2443 {
2444         spin_lock(&entry->d_lock);
2445         __d_rehash(entry);
2446         spin_unlock(&entry->d_lock);
2447 }
2448 EXPORT_SYMBOL(d_rehash);
2449
2450 static inline unsigned start_dir_add(struct inode *dir)
2451 {
2452         preempt_disable_nested();
2453         for (;;) {
2454                 unsigned n = dir->i_dir_seq;
2455                 if (!(n & 1) && cmpxchg(&dir->i_dir_seq, n, n + 1) == n)
2456                         return n;
2457                 cpu_relax();
2458         }
2459 }
2460
2461 static inline void end_dir_add(struct inode *dir, unsigned int n,
2462                                wait_queue_head_t *d_wait)
2463 {
2464         smp_store_release(&dir->i_dir_seq, n + 2);
2465         preempt_enable_nested();
2466         wake_up_all(d_wait);
2467 }
2468
2469 static void d_wait_lookup(struct dentry *dentry)
2470 {
2471         if (d_in_lookup(dentry)) {
2472                 DECLARE_WAITQUEUE(wait, current);
2473                 add_wait_queue(dentry->d_wait, &wait);
2474                 do {
2475                         set_current_state(TASK_UNINTERRUPTIBLE);
2476                         spin_unlock(&dentry->d_lock);
2477                         schedule();
2478                         spin_lock(&dentry->d_lock);
2479                 } while (d_in_lookup(dentry));
2480         }
2481 }
2482
2483 struct dentry *d_alloc_parallel(struct dentry *parent,
2484                                 const struct qstr *name,
2485                                 wait_queue_head_t *wq)
2486 {
2487         unsigned int hash = name->hash;
2488         struct hlist_bl_head *b = in_lookup_hash(parent, hash);
2489         struct hlist_bl_node *node;
2490         struct dentry *new = d_alloc(parent, name);
2491         struct dentry *dentry;
2492         unsigned seq, r_seq, d_seq;
2493
2494         if (unlikely(!new))
2495                 return ERR_PTR(-ENOMEM);
2496
2497 retry:
2498         rcu_read_lock();
2499         seq = smp_load_acquire(&parent->d_inode->i_dir_seq);
2500         r_seq = read_seqbegin(&rename_lock);
2501         dentry = __d_lookup_rcu(parent, name, &d_seq);
2502         if (unlikely(dentry)) {
2503                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2504                         rcu_read_unlock();
2505                         goto retry;
2506                 }
2507                 if (read_seqcount_retry(&dentry->d_seq, d_seq)) {
2508                         rcu_read_unlock();
2509                         dput(dentry);
2510                         goto retry;
2511                 }
2512                 rcu_read_unlock();
2513                 dput(new);
2514                 return dentry;
2515         }
2516         if (unlikely(read_seqretry(&rename_lock, r_seq))) {
2517                 rcu_read_unlock();
2518                 goto retry;
2519         }
2520
2521         if (unlikely(seq & 1)) {
2522                 rcu_read_unlock();
2523                 goto retry;
2524         }
2525
2526         hlist_bl_lock(b);
2527         if (unlikely(READ_ONCE(parent->d_inode->i_dir_seq) != seq)) {
2528                 hlist_bl_unlock(b);
2529                 rcu_read_unlock();
2530                 goto retry;
2531         }
2532         /*
2533          * No changes for the parent since the beginning of d_lookup().
2534          * Since all removals from the chain happen with hlist_bl_lock(),
2535          * any potential in-lookup matches are going to stay here until
2536          * we unlock the chain.  All fields are stable in everything
2537          * we encounter.
2538          */
2539         hlist_bl_for_each_entry(dentry, node, b, d_u.d_in_lookup_hash) {
2540                 if (dentry->d_name.hash != hash)
2541                         continue;
2542                 if (dentry->d_parent != parent)
2543                         continue;
2544                 if (!d_same_name(dentry, parent, name))
2545                         continue;
2546                 hlist_bl_unlock(b);
2547                 /* now we can try to grab a reference */
2548                 if (!lockref_get_not_dead(&dentry->d_lockref)) {
2549                         rcu_read_unlock();
2550                         goto retry;
2551                 }
2552
2553                 rcu_read_unlock();
2554                 /*
2555                  * somebody is likely to be still doing lookup for it;
2556                  * wait for them to finish
2557                  */
2558                 spin_lock(&dentry->d_lock);
2559                 d_wait_lookup(dentry);
2560                 /*
2561                  * it's not in-lookup anymore; in principle we should repeat
2562                  * everything from dcache lookup, but it's likely to be what
2563                  * d_lookup() would've found anyway.  If it is, just return it;
2564                  * otherwise we really have to repeat the whole thing.
2565                  */
2566                 if (unlikely(dentry->d_name.hash != hash))
2567                         goto mismatch;
2568                 if (unlikely(dentry->d_parent != parent))
2569                         goto mismatch;
2570                 if (unlikely(d_unhashed(dentry)))
2571                         goto mismatch;
2572                 if (unlikely(!d_same_name(dentry, parent, name)))
2573                         goto mismatch;
2574                 /* OK, it *is* a hashed match; return it */
2575                 spin_unlock(&dentry->d_lock);
2576                 dput(new);
2577                 return dentry;
2578         }
2579         rcu_read_unlock();
2580         /* we can't take ->d_lock here; it's OK, though. */
2581         new->d_flags |= DCACHE_PAR_LOOKUP;
2582         new->d_wait = wq;
2583         hlist_bl_add_head(&new->d_u.d_in_lookup_hash, b);
2584         hlist_bl_unlock(b);
2585         return new;
2586 mismatch:
2587         spin_unlock(&dentry->d_lock);
2588         dput(dentry);
2589         goto retry;
2590 }
2591 EXPORT_SYMBOL(d_alloc_parallel);
2592
2593 /*
2594  * - Unhash the dentry
2595  * - Retrieve and clear the waitqueue head in dentry
2596  * - Return the waitqueue head
2597  */
2598 static wait_queue_head_t *__d_lookup_unhash(struct dentry *dentry)
2599 {
2600         wait_queue_head_t *d_wait;
2601         struct hlist_bl_head *b;
2602
2603         lockdep_assert_held(&dentry->d_lock);
2604
2605         b = in_lookup_hash(dentry->d_parent, dentry->d_name.hash);
2606         hlist_bl_lock(b);
2607         dentry->d_flags &= ~DCACHE_PAR_LOOKUP;
2608         __hlist_bl_del(&dentry->d_u.d_in_lookup_hash);
2609         d_wait = dentry->d_wait;
2610         dentry->d_wait = NULL;
2611         hlist_bl_unlock(b);
2612         INIT_HLIST_NODE(&dentry->d_u.d_alias);
2613         INIT_LIST_HEAD(&dentry->d_lru);
2614         return d_wait;
2615 }
2616
2617 void __d_lookup_unhash_wake(struct dentry *dentry)
2618 {
2619         spin_lock(&dentry->d_lock);
2620         wake_up_all(__d_lookup_unhash(dentry));
2621         spin_unlock(&dentry->d_lock);
2622 }
2623 EXPORT_SYMBOL(__d_lookup_unhash_wake);
2624
2625 /* inode->i_lock held if inode is non-NULL */
2626
2627 static inline void __d_add(struct dentry *dentry, struct inode *inode)
2628 {
2629         wait_queue_head_t *d_wait;
2630         struct inode *dir = NULL;
2631         unsigned n;
2632         spin_lock(&dentry->d_lock);
2633         if (unlikely(d_in_lookup(dentry))) {
2634                 dir = dentry->d_parent->d_inode;
2635                 n = start_dir_add(dir);
2636                 d_wait = __d_lookup_unhash(dentry);
2637         }
2638         if (inode) {
2639                 unsigned add_flags = d_flags_for_inode(inode);
2640                 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
2641                 raw_write_seqcount_begin(&dentry->d_seq);
2642                 __d_set_inode_and_type(dentry, inode, add_flags);
2643                 raw_write_seqcount_end(&dentry->d_seq);
2644                 fsnotify_update_flags(dentry);
2645         }
2646         __d_rehash(dentry);
2647         if (dir)
2648                 end_dir_add(dir, n, d_wait);
2649         spin_unlock(&dentry->d_lock);
2650         if (inode)
2651                 spin_unlock(&inode->i_lock);
2652 }
2653
2654 /**
2655  * d_add - add dentry to hash queues
2656  * @entry: dentry to add
2657  * @inode: The inode to attach to this dentry
2658  *
2659  * This adds the entry to the hash queues and initializes @inode.
2660  * The entry was actually filled in earlier during d_alloc().
2661  */
2662
2663 void d_add(struct dentry *entry, struct inode *inode)
2664 {
2665         if (inode) {
2666                 security_d_instantiate(entry, inode);
2667                 spin_lock(&inode->i_lock);
2668         }
2669         __d_add(entry, inode);
2670 }
2671 EXPORT_SYMBOL(d_add);
2672
2673 /**
2674  * d_exact_alias - find and hash an exact unhashed alias
2675  * @entry: dentry to add
2676  * @inode: The inode to go with this dentry
2677  *
2678  * If an unhashed dentry with the same name/parent and desired
2679  * inode already exists, hash and return it.  Otherwise, return
2680  * NULL.
2681  *
2682  * Parent directory should be locked.
2683  */
2684 struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2685 {
2686         struct dentry *alias;
2687         unsigned int hash = entry->d_name.hash;
2688
2689         spin_lock(&inode->i_lock);
2690         hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2691                 /*
2692                  * Don't need alias->d_lock here, because aliases with
2693                  * d_parent == entry->d_parent are not subject to name or
2694                  * parent changes, because the parent inode i_mutex is held.
2695                  */
2696                 if (alias->d_name.hash != hash)
2697                         continue;
2698                 if (alias->d_parent != entry->d_parent)
2699                         continue;
2700                 if (!d_same_name(alias, entry->d_parent, &entry->d_name))
2701                         continue;
2702                 spin_lock(&alias->d_lock);
2703                 if (!d_unhashed(alias)) {
2704                         spin_unlock(&alias->d_lock);
2705                         alias = NULL;
2706                 } else {
2707                         dget_dlock(alias);
2708                         __d_rehash(alias);
2709                         spin_unlock(&alias->d_lock);
2710                 }
2711                 spin_unlock(&inode->i_lock);
2712                 return alias;
2713         }
2714         spin_unlock(&inode->i_lock);
2715         return NULL;
2716 }
2717 EXPORT_SYMBOL(d_exact_alias);
2718
2719 static void swap_names(struct dentry *dentry, struct dentry *target)
2720 {
2721         if (unlikely(dname_external(target))) {
2722                 if (unlikely(dname_external(dentry))) {
2723                         /*
2724                          * Both external: swap the pointers
2725                          */
2726                         swap(target->d_name.name, dentry->d_name.name);
2727                 } else {
2728                         /*
2729                          * dentry:internal, target:external.  Steal target's
2730                          * storage and make target internal.
2731                          */
2732                         memcpy(target->d_iname, dentry->d_name.name,
2733                                         dentry->d_name.len + 1);
2734                         dentry->d_name.name = target->d_name.name;
2735                         target->d_name.name = target->d_iname;
2736                 }
2737         } else {
2738                 if (unlikely(dname_external(dentry))) {
2739                         /*
2740                          * dentry:external, target:internal.  Give dentry's
2741                          * storage to target and make dentry internal
2742                          */
2743                         memcpy(dentry->d_iname, target->d_name.name,
2744                                         target->d_name.len + 1);
2745                         target->d_name.name = dentry->d_name.name;
2746                         dentry->d_name.name = dentry->d_iname;
2747                 } else {
2748                         /*
2749                          * Both are internal.
2750                          */
2751                         unsigned int i;
2752                         BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2753                         for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2754                                 swap(((long *) &dentry->d_iname)[i],
2755                                      ((long *) &target->d_iname)[i]);
2756                         }
2757                 }
2758         }
2759         swap(dentry->d_name.hash_len, target->d_name.hash_len);
2760 }
2761
2762 static void copy_name(struct dentry *dentry, struct dentry *target)
2763 {
2764         struct external_name *old_name = NULL;
2765         if (unlikely(dname_external(dentry)))
2766                 old_name = external_name(dentry);
2767         if (unlikely(dname_external(target))) {
2768                 atomic_inc(&external_name(target)->u.count);
2769                 dentry->d_name = target->d_name;
2770         } else {
2771                 memcpy(dentry->d_iname, target->d_name.name,
2772                                 target->d_name.len + 1);
2773                 dentry->d_name.name = dentry->d_iname;
2774                 dentry->d_name.hash_len = target->d_name.hash_len;
2775         }
2776         if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2777                 kfree_rcu(old_name, u.head);
2778 }
2779
2780 /*
2781  * __d_move - move a dentry
2782  * @dentry: entry to move
2783  * @target: new dentry
2784  * @exchange: exchange the two dentries
2785  *
2786  * Update the dcache to reflect the move of a file name. Negative
2787  * dcache entries should not be moved in this way. Caller must hold
2788  * rename_lock, the i_mutex of the source and target directories,
2789  * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2790  */
2791 static void __d_move(struct dentry *dentry, struct dentry *target,
2792                      bool exchange)
2793 {
2794         struct dentry *old_parent, *p;
2795         wait_queue_head_t *d_wait;
2796         struct inode *dir = NULL;
2797         unsigned n;
2798
2799         WARN_ON(!dentry->d_inode);
2800         if (WARN_ON(dentry == target))
2801                 return;
2802
2803         BUG_ON(d_ancestor(target, dentry));
2804         old_parent = dentry->d_parent;
2805         p = d_ancestor(old_parent, target);
2806         if (IS_ROOT(dentry)) {
2807                 BUG_ON(p);
2808                 spin_lock(&target->d_parent->d_lock);
2809         } else if (!p) {
2810                 /* target is not a descendent of dentry->d_parent */
2811                 spin_lock(&target->d_parent->d_lock);
2812                 spin_lock_nested(&old_parent->d_lock, DENTRY_D_LOCK_NESTED);
2813         } else {
2814                 BUG_ON(p == dentry);
2815                 spin_lock(&old_parent->d_lock);
2816                 if (p != target)
2817                         spin_lock_nested(&target->d_parent->d_lock,
2818                                         DENTRY_D_LOCK_NESTED);
2819         }
2820         spin_lock_nested(&dentry->d_lock, 2);
2821         spin_lock_nested(&target->d_lock, 3);
2822
2823         if (unlikely(d_in_lookup(target))) {
2824                 dir = target->d_parent->d_inode;
2825                 n = start_dir_add(dir);
2826                 d_wait = __d_lookup_unhash(target);
2827         }
2828
2829         write_seqcount_begin(&dentry->d_seq);
2830         write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2831
2832         /* unhash both */
2833         if (!d_unhashed(dentry))
2834                 ___d_drop(dentry);
2835         if (!d_unhashed(target))
2836                 ___d_drop(target);
2837
2838         /* ... and switch them in the tree */
2839         dentry->d_parent = target->d_parent;
2840         if (!exchange) {
2841                 copy_name(dentry, target);
2842                 target->d_hash.pprev = NULL;
2843                 dentry->d_parent->d_lockref.count++;
2844                 if (dentry != old_parent) /* wasn't IS_ROOT */
2845                         WARN_ON(!--old_parent->d_lockref.count);
2846         } else {
2847                 target->d_parent = old_parent;
2848                 swap_names(dentry, target);
2849                 if (!hlist_unhashed(&target->d_sib))
2850                         __hlist_del(&target->d_sib);
2851                 hlist_add_head(&target->d_sib, &target->d_parent->d_children);
2852                 __d_rehash(target);
2853                 fsnotify_update_flags(target);
2854         }
2855         if (!hlist_unhashed(&dentry->d_sib))
2856                 __hlist_del(&dentry->d_sib);
2857         hlist_add_head(&dentry->d_sib, &dentry->d_parent->d_children);
2858         __d_rehash(dentry);
2859         fsnotify_update_flags(dentry);
2860         fscrypt_handle_d_move(dentry);
2861
2862         write_seqcount_end(&target->d_seq);
2863         write_seqcount_end(&dentry->d_seq);
2864
2865         if (dir)
2866                 end_dir_add(dir, n, d_wait);
2867
2868         if (dentry->d_parent != old_parent)
2869                 spin_unlock(&dentry->d_parent->d_lock);
2870         if (dentry != old_parent)
2871                 spin_unlock(&old_parent->d_lock);
2872         spin_unlock(&target->d_lock);
2873         spin_unlock(&dentry->d_lock);
2874 }
2875
2876 /*
2877  * d_move - move a dentry
2878  * @dentry: entry to move
2879  * @target: new dentry
2880  *
2881  * Update the dcache to reflect the move of a file name. Negative
2882  * dcache entries should not be moved in this way. See the locking
2883  * requirements for __d_move.
2884  */
2885 void d_move(struct dentry *dentry, struct dentry *target)
2886 {
2887         write_seqlock(&rename_lock);
2888         __d_move(dentry, target, false);
2889         write_sequnlock(&rename_lock);
2890 }
2891 EXPORT_SYMBOL(d_move);
2892
2893 /*
2894  * d_exchange - exchange two dentries
2895  * @dentry1: first dentry
2896  * @dentry2: second dentry
2897  */
2898 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2899 {
2900         write_seqlock(&rename_lock);
2901
2902         WARN_ON(!dentry1->d_inode);
2903         WARN_ON(!dentry2->d_inode);
2904         WARN_ON(IS_ROOT(dentry1));
2905         WARN_ON(IS_ROOT(dentry2));
2906
2907         __d_move(dentry1, dentry2, true);
2908
2909         write_sequnlock(&rename_lock);
2910 }
2911
2912 /**
2913  * d_ancestor - search for an ancestor
2914  * @p1: ancestor dentry
2915  * @p2: child dentry
2916  *
2917  * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2918  * an ancestor of p2, else NULL.
2919  */
2920 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2921 {
2922         struct dentry *p;
2923
2924         for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2925                 if (p->d_parent == p1)
2926                         return p;
2927         }
2928         return NULL;
2929 }
2930
2931 /*
2932  * This helper attempts to cope with remotely renamed directories
2933  *
2934  * It assumes that the caller is already holding
2935  * dentry->d_parent->d_inode->i_mutex, and rename_lock
2936  *
2937  * Note: If ever the locking in lock_rename() changes, then please
2938  * remember to update this too...
2939  */
2940 static int __d_unalias(struct dentry *dentry, struct dentry *alias)
2941 {
2942         struct mutex *m1 = NULL;
2943         struct rw_semaphore *m2 = NULL;
2944         int ret = -ESTALE;
2945
2946         /* If alias and dentry share a parent, then no extra locks required */
2947         if (alias->d_parent == dentry->d_parent)
2948                 goto out_unalias;
2949
2950         /* See lock_rename() */
2951         if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2952                 goto out_err;
2953         m1 = &dentry->d_sb->s_vfs_rename_mutex;
2954         if (!inode_trylock_shared(alias->d_parent->d_inode))
2955                 goto out_err;
2956         m2 = &alias->d_parent->d_inode->i_rwsem;
2957 out_unalias:
2958         __d_move(alias, dentry, false);
2959         ret = 0;
2960 out_err:
2961         if (m2)
2962                 up_read(m2);
2963         if (m1)
2964                 mutex_unlock(m1);
2965         return ret;
2966 }
2967
2968 /**
2969  * d_splice_alias - splice a disconnected dentry into the tree if one exists
2970  * @inode:  the inode which may have a disconnected dentry
2971  * @dentry: a negative dentry which we want to point to the inode.
2972  *
2973  * If inode is a directory and has an IS_ROOT alias, then d_move that in
2974  * place of the given dentry and return it, else simply d_add the inode
2975  * to the dentry and return NULL.
2976  *
2977  * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2978  * we should error out: directories can't have multiple aliases.
2979  *
2980  * This is needed in the lookup routine of any filesystem that is exportable
2981  * (via knfsd) so that we can build dcache paths to directories effectively.
2982  *
2983  * If a dentry was found and moved, then it is returned.  Otherwise NULL
2984  * is returned.  This matches the expected return value of ->lookup.
2985  *
2986  * Cluster filesystems may call this function with a negative, hashed dentry.
2987  * In that case, we know that the inode will be a regular file, and also this
2988  * will only occur during atomic_open. So we need to check for the dentry
2989  * being already hashed only in the final case.
2990  */
2991 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
2992 {
2993         if (IS_ERR(inode))
2994                 return ERR_CAST(inode);
2995
2996         BUG_ON(!d_unhashed(dentry));
2997
2998         if (!inode)
2999                 goto out;
3000
3001         security_d_instantiate(dentry, inode);
3002         spin_lock(&inode->i_lock);
3003         if (S_ISDIR(inode->i_mode)) {
3004                 struct dentry *new = __d_find_any_alias(inode);
3005                 if (unlikely(new)) {
3006                         /* The reference to new ensures it remains an alias */
3007                         spin_unlock(&inode->i_lock);
3008                         write_seqlock(&rename_lock);
3009                         if (unlikely(d_ancestor(new, dentry))) {
3010                                 write_sequnlock(&rename_lock);
3011                                 dput(new);
3012                                 new = ERR_PTR(-ELOOP);
3013                                 pr_warn_ratelimited(
3014                                         "VFS: Lookup of '%s' in %s %s"
3015                                         " would have caused loop\n",
3016                                         dentry->d_name.name,
3017                                         inode->i_sb->s_type->name,
3018                                         inode->i_sb->s_id);
3019                         } else if (!IS_ROOT(new)) {
3020                                 struct dentry *old_parent = dget(new->d_parent);
3021                                 int err = __d_unalias(dentry, new);
3022                                 write_sequnlock(&rename_lock);
3023                                 if (err) {
3024                                         dput(new);
3025                                         new = ERR_PTR(err);
3026                                 }
3027                                 dput(old_parent);
3028                         } else {
3029                                 __d_move(new, dentry, false);
3030                                 write_sequnlock(&rename_lock);
3031                         }
3032                         iput(inode);
3033                         return new;
3034                 }
3035         }
3036 out:
3037         __d_add(dentry, inode);
3038         return NULL;
3039 }
3040 EXPORT_SYMBOL(d_splice_alias);
3041
3042 /*
3043  * Test whether new_dentry is a subdirectory of old_dentry.
3044  *
3045  * Trivially implemented using the dcache structure
3046  */
3047
3048 /**
3049  * is_subdir - is new dentry a subdirectory of old_dentry
3050  * @new_dentry: new dentry
3051  * @old_dentry: old dentry
3052  *
3053  * Returns true if new_dentry is a subdirectory of the parent (at any depth).
3054  * Returns false otherwise.
3055  * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3056  */
3057   
3058 bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3059 {
3060         bool subdir;
3061         unsigned seq;
3062
3063         if (new_dentry == old_dentry)
3064                 return true;
3065
3066         /* Access d_parent under rcu as d_move() may change it. */
3067         rcu_read_lock();
3068         seq = read_seqbegin(&rename_lock);
3069         subdir = d_ancestor(old_dentry, new_dentry);
3070          /* Try lockless once... */
3071         if (read_seqretry(&rename_lock, seq)) {
3072                 /* ...else acquire lock for progress even on deep chains. */
3073                 read_seqlock_excl(&rename_lock);
3074                 subdir = d_ancestor(old_dentry, new_dentry);
3075                 read_sequnlock_excl(&rename_lock);
3076         }
3077         rcu_read_unlock();
3078         return subdir;
3079 }
3080 EXPORT_SYMBOL(is_subdir);
3081
3082 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3083 {
3084         struct dentry *root = data;
3085         if (dentry != root) {
3086                 if (d_unhashed(dentry) || !dentry->d_inode)
3087                         return D_WALK_SKIP;
3088
3089                 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3090                         dentry->d_flags |= DCACHE_GENOCIDE;
3091                         dentry->d_lockref.count--;
3092                 }
3093         }
3094         return D_WALK_CONTINUE;
3095 }
3096
3097 void d_genocide(struct dentry *parent)
3098 {
3099         d_walk(parent, parent, d_genocide_kill);
3100 }
3101
3102 void d_mark_tmpfile(struct file *file, struct inode *inode)
3103 {
3104         struct dentry *dentry = file->f_path.dentry;
3105
3106         BUG_ON(dentry->d_name.name != dentry->d_iname ||
3107                 !hlist_unhashed(&dentry->d_u.d_alias) ||
3108                 !d_unlinked(dentry));
3109         spin_lock(&dentry->d_parent->d_lock);
3110         spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3111         dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3112                                 (unsigned long long)inode->i_ino);
3113         spin_unlock(&dentry->d_lock);
3114         spin_unlock(&dentry->d_parent->d_lock);
3115 }
3116 EXPORT_SYMBOL(d_mark_tmpfile);
3117
3118 void d_tmpfile(struct file *file, struct inode *inode)
3119 {
3120         struct dentry *dentry = file->f_path.dentry;
3121
3122         inode_dec_link_count(inode);
3123         d_mark_tmpfile(file, inode);
3124         d_instantiate(dentry, inode);
3125 }
3126 EXPORT_SYMBOL(d_tmpfile);
3127
3128 /*
3129  * Obtain inode number of the parent dentry.
3130  */
3131 ino_t d_parent_ino(struct dentry *dentry)
3132 {
3133         struct dentry *parent;
3134         struct inode *iparent;
3135         unsigned seq;
3136         ino_t ret;
3137
3138         scoped_guard(rcu) {
3139                 seq = raw_seqcount_begin(&dentry->d_seq);
3140                 parent = READ_ONCE(dentry->d_parent);
3141                 iparent = d_inode_rcu(parent);
3142                 if (likely(iparent)) {
3143                         ret = iparent->i_ino;
3144                         if (!read_seqcount_retry(&dentry->d_seq, seq))
3145                                 return ret;
3146                 }
3147         }
3148
3149         spin_lock(&dentry->d_lock);
3150         ret = dentry->d_parent->d_inode->i_ino;
3151         spin_unlock(&dentry->d_lock);
3152         return ret;
3153 }
3154 EXPORT_SYMBOL(d_parent_ino);
3155
3156 static __initdata unsigned long dhash_entries;
3157 static int __init set_dhash_entries(char *str)
3158 {
3159         if (!str)
3160                 return 0;
3161         dhash_entries = simple_strtoul(str, &str, 0);
3162         return 1;
3163 }
3164 __setup("dhash_entries=", set_dhash_entries);
3165
3166 static void __init dcache_init_early(void)
3167 {
3168         /* If hashes are distributed across NUMA nodes, defer
3169          * hash allocation until vmalloc space is available.
3170          */
3171         if (hashdist)
3172                 return;
3173
3174         dentry_hashtable =
3175                 alloc_large_system_hash("Dentry cache",
3176                                         sizeof(struct hlist_bl_head),
3177                                         dhash_entries,
3178                                         13,
3179                                         HASH_EARLY | HASH_ZERO,
3180                                         &d_hash_shift,
3181                                         NULL,
3182                                         0,
3183                                         0);
3184         d_hash_shift = 32 - d_hash_shift;
3185
3186         runtime_const_init(shift, d_hash_shift);
3187         runtime_const_init(ptr, dentry_hashtable);
3188 }
3189
3190 static void __init dcache_init(void)
3191 {
3192         /*
3193          * A constructor could be added for stable state like the lists,
3194          * but it is probably not worth it because of the cache nature
3195          * of the dcache.
3196          */
3197         dentry_cache = KMEM_CACHE_USERCOPY(dentry,
3198                 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_ACCOUNT,
3199                 d_iname);
3200
3201         /* Hash may have been set up in dcache_init_early */
3202         if (!hashdist)
3203                 return;
3204
3205         dentry_hashtable =
3206                 alloc_large_system_hash("Dentry cache",
3207                                         sizeof(struct hlist_bl_head),
3208                                         dhash_entries,
3209                                         13,
3210                                         HASH_ZERO,
3211                                         &d_hash_shift,
3212                                         NULL,
3213                                         0,
3214                                         0);
3215         d_hash_shift = 32 - d_hash_shift;
3216
3217         runtime_const_init(shift, d_hash_shift);
3218         runtime_const_init(ptr, dentry_hashtable);
3219 }
3220
3221 /* SLAB cache for __getname() consumers */
3222 struct kmem_cache *names_cachep __ro_after_init;
3223 EXPORT_SYMBOL(names_cachep);
3224
3225 void __init vfs_caches_init_early(void)
3226 {
3227         int i;
3228
3229         for (i = 0; i < ARRAY_SIZE(in_lookup_hashtable); i++)
3230                 INIT_HLIST_BL_HEAD(&in_lookup_hashtable[i]);
3231
3232         dcache_init_early();
3233         inode_init_early();
3234 }
3235
3236 void __init vfs_caches_init(void)
3237 {
3238         names_cachep = kmem_cache_create_usercopy("names_cache", PATH_MAX, 0,
3239                         SLAB_HWCACHE_ALIGN|SLAB_PANIC, 0, PATH_MAX, NULL);
3240
3241         dcache_init();
3242         inode_init();
3243         files_init();
3244         files_maxfiles_init();
3245         mnt_init();
3246         bdev_cache_init();
3247         chrdev_init();
3248 }
This page took 0.200451 seconds and 4 git commands to generate.