]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - fs/dcache.c
Merge tag 'fbdev-fixes-4.6' of git://git.kernel.org/pub/scm/linux/kernel/git/tomba...
[mirror_ubuntu-artful-kernel.git] / fs / dcache.c
1 /*
2 * fs/dcache.c
3 *
4 * Complete reimplementation
5 * (C) 1997 Thomas Schoebel-Theuer,
6 * with heavy changes by Linus Torvalds
7 */
8
9 /*
10 * Notes on the allocation strategy:
11 *
12 * The dcache is a master of the icache - whenever a dcache entry
13 * exists, the inode will always exist. "iput()" is done either when
14 * the dcache entry is deleted or garbage collected.
15 */
16
17 #include <linux/syscalls.h>
18 #include <linux/string.h>
19 #include <linux/mm.h>
20 #include <linux/fs.h>
21 #include <linux/fsnotify.h>
22 #include <linux/slab.h>
23 #include <linux/init.h>
24 #include <linux/hash.h>
25 #include <linux/cache.h>
26 #include <linux/export.h>
27 #include <linux/mount.h>
28 #include <linux/file.h>
29 #include <asm/uaccess.h>
30 #include <linux/security.h>
31 #include <linux/seqlock.h>
32 #include <linux/swap.h>
33 #include <linux/bootmem.h>
34 #include <linux/fs_struct.h>
35 #include <linux/hardirq.h>
36 #include <linux/bit_spinlock.h>
37 #include <linux/rculist_bl.h>
38 #include <linux/prefetch.h>
39 #include <linux/ratelimit.h>
40 #include <linux/list_lru.h>
41 #include <linux/kasan.h>
42
43 #include "internal.h"
44 #include "mount.h"
45
46 /*
47 * Usage:
48 * dcache->d_inode->i_lock protects:
49 * - i_dentry, d_u.d_alias, d_inode of aliases
50 * dcache_hash_bucket lock protects:
51 * - the dcache hash table
52 * s_anon bl list spinlock protects:
53 * - the s_anon list (see __d_drop)
54 * dentry->d_sb->s_dentry_lru_lock protects:
55 * - the dcache lru lists and counters
56 * d_lock protects:
57 * - d_flags
58 * - d_name
59 * - d_lru
60 * - d_count
61 * - d_unhashed()
62 * - d_parent and d_subdirs
63 * - childrens' d_child and d_parent
64 * - d_u.d_alias, d_inode
65 *
66 * Ordering:
67 * dentry->d_inode->i_lock
68 * dentry->d_lock
69 * dentry->d_sb->s_dentry_lru_lock
70 * dcache_hash_bucket lock
71 * s_anon lock
72 *
73 * If there is an ancestor relationship:
74 * dentry->d_parent->...->d_parent->d_lock
75 * ...
76 * dentry->d_parent->d_lock
77 * dentry->d_lock
78 *
79 * If no ancestor relationship:
80 * if (dentry1 < dentry2)
81 * dentry1->d_lock
82 * dentry2->d_lock
83 */
84 int sysctl_vfs_cache_pressure __read_mostly = 100;
85 EXPORT_SYMBOL_GPL(sysctl_vfs_cache_pressure);
86
87 __cacheline_aligned_in_smp DEFINE_SEQLOCK(rename_lock);
88
89 EXPORT_SYMBOL(rename_lock);
90
91 static struct kmem_cache *dentry_cache __read_mostly;
92
93 /*
94 * This is the single most critical data structure when it comes
95 * to the dcache: the hashtable for lookups. Somebody should try
96 * to make this good - I've just made it work.
97 *
98 * This hash-function tries to avoid losing too many bits of hash
99 * information, yet avoid using a prime hash-size or similar.
100 */
101
102 static unsigned int d_hash_mask __read_mostly;
103 static unsigned int d_hash_shift __read_mostly;
104
105 static struct hlist_bl_head *dentry_hashtable __read_mostly;
106
107 static inline struct hlist_bl_head *d_hash(const struct dentry *parent,
108 unsigned int hash)
109 {
110 hash += (unsigned long) parent / L1_CACHE_BYTES;
111 return dentry_hashtable + hash_32(hash, d_hash_shift);
112 }
113
114 /* Statistics gathering. */
115 struct dentry_stat_t dentry_stat = {
116 .age_limit = 45,
117 };
118
119 static DEFINE_PER_CPU(long, nr_dentry);
120 static DEFINE_PER_CPU(long, nr_dentry_unused);
121
122 #if defined(CONFIG_SYSCTL) && defined(CONFIG_PROC_FS)
123
124 /*
125 * Here we resort to our own counters instead of using generic per-cpu counters
126 * for consistency with what the vfs inode code does. We are expected to harvest
127 * better code and performance by having our own specialized counters.
128 *
129 * Please note that the loop is done over all possible CPUs, not over all online
130 * CPUs. The reason for this is that we don't want to play games with CPUs going
131 * on and off. If one of them goes off, we will just keep their counters.
132 *
133 * glommer: See cffbc8a for details, and if you ever intend to change this,
134 * please update all vfs counters to match.
135 */
136 static long get_nr_dentry(void)
137 {
138 int i;
139 long sum = 0;
140 for_each_possible_cpu(i)
141 sum += per_cpu(nr_dentry, i);
142 return sum < 0 ? 0 : sum;
143 }
144
145 static long get_nr_dentry_unused(void)
146 {
147 int i;
148 long sum = 0;
149 for_each_possible_cpu(i)
150 sum += per_cpu(nr_dentry_unused, i);
151 return sum < 0 ? 0 : sum;
152 }
153
154 int proc_nr_dentry(struct ctl_table *table, int write, void __user *buffer,
155 size_t *lenp, loff_t *ppos)
156 {
157 dentry_stat.nr_dentry = get_nr_dentry();
158 dentry_stat.nr_unused = get_nr_dentry_unused();
159 return proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
160 }
161 #endif
162
163 /*
164 * Compare 2 name strings, return 0 if they match, otherwise non-zero.
165 * The strings are both count bytes long, and count is non-zero.
166 */
167 #ifdef CONFIG_DCACHE_WORD_ACCESS
168
169 #include <asm/word-at-a-time.h>
170 /*
171 * NOTE! 'cs' and 'scount' come from a dentry, so it has a
172 * aligned allocation for this particular component. We don't
173 * strictly need the load_unaligned_zeropad() safety, but it
174 * doesn't hurt either.
175 *
176 * In contrast, 'ct' and 'tcount' can be from a pathname, and do
177 * need the careful unaligned handling.
178 */
179 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
180 {
181 unsigned long a,b,mask;
182
183 for (;;) {
184 a = *(unsigned long *)cs;
185 b = load_unaligned_zeropad(ct);
186 if (tcount < sizeof(unsigned long))
187 break;
188 if (unlikely(a != b))
189 return 1;
190 cs += sizeof(unsigned long);
191 ct += sizeof(unsigned long);
192 tcount -= sizeof(unsigned long);
193 if (!tcount)
194 return 0;
195 }
196 mask = bytemask_from_count(tcount);
197 return unlikely(!!((a ^ b) & mask));
198 }
199
200 #else
201
202 static inline int dentry_string_cmp(const unsigned char *cs, const unsigned char *ct, unsigned tcount)
203 {
204 do {
205 if (*cs != *ct)
206 return 1;
207 cs++;
208 ct++;
209 tcount--;
210 } while (tcount);
211 return 0;
212 }
213
214 #endif
215
216 static inline int dentry_cmp(const struct dentry *dentry, const unsigned char *ct, unsigned tcount)
217 {
218 const unsigned char *cs;
219 /*
220 * Be careful about RCU walk racing with rename:
221 * use ACCESS_ONCE to fetch the name pointer.
222 *
223 * NOTE! Even if a rename will mean that the length
224 * was not loaded atomically, we don't care. The
225 * RCU walk will check the sequence count eventually,
226 * and catch it. And we won't overrun the buffer,
227 * because we're reading the name pointer atomically,
228 * and a dentry name is guaranteed to be properly
229 * terminated with a NUL byte.
230 *
231 * End result: even if 'len' is wrong, we'll exit
232 * early because the data cannot match (there can
233 * be no NUL in the ct/tcount data)
234 */
235 cs = ACCESS_ONCE(dentry->d_name.name);
236 smp_read_barrier_depends();
237 return dentry_string_cmp(cs, ct, tcount);
238 }
239
240 struct external_name {
241 union {
242 atomic_t count;
243 struct rcu_head head;
244 } u;
245 unsigned char name[];
246 };
247
248 static inline struct external_name *external_name(struct dentry *dentry)
249 {
250 return container_of(dentry->d_name.name, struct external_name, name[0]);
251 }
252
253 static void __d_free(struct rcu_head *head)
254 {
255 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
256
257 kmem_cache_free(dentry_cache, dentry);
258 }
259
260 static void __d_free_external(struct rcu_head *head)
261 {
262 struct dentry *dentry = container_of(head, struct dentry, d_u.d_rcu);
263 kfree(external_name(dentry));
264 kmem_cache_free(dentry_cache, dentry);
265 }
266
267 static inline int dname_external(const struct dentry *dentry)
268 {
269 return dentry->d_name.name != dentry->d_iname;
270 }
271
272 static inline void __d_set_inode_and_type(struct dentry *dentry,
273 struct inode *inode,
274 unsigned type_flags)
275 {
276 unsigned flags;
277
278 dentry->d_inode = inode;
279 flags = READ_ONCE(dentry->d_flags);
280 flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
281 flags |= type_flags;
282 WRITE_ONCE(dentry->d_flags, flags);
283 }
284
285 static inline void __d_clear_type_and_inode(struct dentry *dentry)
286 {
287 unsigned flags = READ_ONCE(dentry->d_flags);
288
289 flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
290 WRITE_ONCE(dentry->d_flags, flags);
291 dentry->d_inode = NULL;
292 }
293
294 static void dentry_free(struct dentry *dentry)
295 {
296 WARN_ON(!hlist_unhashed(&dentry->d_u.d_alias));
297 if (unlikely(dname_external(dentry))) {
298 struct external_name *p = external_name(dentry);
299 if (likely(atomic_dec_and_test(&p->u.count))) {
300 call_rcu(&dentry->d_u.d_rcu, __d_free_external);
301 return;
302 }
303 }
304 /* if dentry was never visible to RCU, immediate free is OK */
305 if (!(dentry->d_flags & DCACHE_RCUACCESS))
306 __d_free(&dentry->d_u.d_rcu);
307 else
308 call_rcu(&dentry->d_u.d_rcu, __d_free);
309 }
310
311 /**
312 * dentry_rcuwalk_invalidate - invalidate in-progress rcu-walk lookups
313 * @dentry: the target dentry
314 * After this call, in-progress rcu-walk path lookup will fail. This
315 * should be called after unhashing, and after changing d_inode (if
316 * the dentry has not already been unhashed).
317 */
318 static inline void dentry_rcuwalk_invalidate(struct dentry *dentry)
319 {
320 lockdep_assert_held(&dentry->d_lock);
321 /* Go through am invalidation barrier */
322 write_seqcount_invalidate(&dentry->d_seq);
323 }
324
325 /*
326 * Release the dentry's inode, using the filesystem
327 * d_iput() operation if defined. Dentry has no refcount
328 * and is unhashed.
329 */
330 static void dentry_iput(struct dentry * dentry)
331 __releases(dentry->d_lock)
332 __releases(dentry->d_inode->i_lock)
333 {
334 struct inode *inode = dentry->d_inode;
335 if (inode) {
336 __d_clear_type_and_inode(dentry);
337 hlist_del_init(&dentry->d_u.d_alias);
338 spin_unlock(&dentry->d_lock);
339 spin_unlock(&inode->i_lock);
340 if (!inode->i_nlink)
341 fsnotify_inoderemove(inode);
342 if (dentry->d_op && dentry->d_op->d_iput)
343 dentry->d_op->d_iput(dentry, inode);
344 else
345 iput(inode);
346 } else {
347 spin_unlock(&dentry->d_lock);
348 }
349 }
350
351 /*
352 * Release the dentry's inode, using the filesystem
353 * d_iput() operation if defined. dentry remains in-use.
354 */
355 static void dentry_unlink_inode(struct dentry * dentry)
356 __releases(dentry->d_lock)
357 __releases(dentry->d_inode->i_lock)
358 {
359 struct inode *inode = dentry->d_inode;
360
361 raw_write_seqcount_begin(&dentry->d_seq);
362 __d_clear_type_and_inode(dentry);
363 hlist_del_init(&dentry->d_u.d_alias);
364 raw_write_seqcount_end(&dentry->d_seq);
365 spin_unlock(&dentry->d_lock);
366 spin_unlock(&inode->i_lock);
367 if (!inode->i_nlink)
368 fsnotify_inoderemove(inode);
369 if (dentry->d_op && dentry->d_op->d_iput)
370 dentry->d_op->d_iput(dentry, inode);
371 else
372 iput(inode);
373 }
374
375 /*
376 * The DCACHE_LRU_LIST bit is set whenever the 'd_lru' entry
377 * is in use - which includes both the "real" per-superblock
378 * LRU list _and_ the DCACHE_SHRINK_LIST use.
379 *
380 * The DCACHE_SHRINK_LIST bit is set whenever the dentry is
381 * on the shrink list (ie not on the superblock LRU list).
382 *
383 * The per-cpu "nr_dentry_unused" counters are updated with
384 * the DCACHE_LRU_LIST bit.
385 *
386 * These helper functions make sure we always follow the
387 * rules. d_lock must be held by the caller.
388 */
389 #define D_FLAG_VERIFY(dentry,x) WARN_ON_ONCE(((dentry)->d_flags & (DCACHE_LRU_LIST | DCACHE_SHRINK_LIST)) != (x))
390 static void d_lru_add(struct dentry *dentry)
391 {
392 D_FLAG_VERIFY(dentry, 0);
393 dentry->d_flags |= DCACHE_LRU_LIST;
394 this_cpu_inc(nr_dentry_unused);
395 WARN_ON_ONCE(!list_lru_add(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
396 }
397
398 static void d_lru_del(struct dentry *dentry)
399 {
400 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
401 dentry->d_flags &= ~DCACHE_LRU_LIST;
402 this_cpu_dec(nr_dentry_unused);
403 WARN_ON_ONCE(!list_lru_del(&dentry->d_sb->s_dentry_lru, &dentry->d_lru));
404 }
405
406 static void d_shrink_del(struct dentry *dentry)
407 {
408 D_FLAG_VERIFY(dentry, DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
409 list_del_init(&dentry->d_lru);
410 dentry->d_flags &= ~(DCACHE_SHRINK_LIST | DCACHE_LRU_LIST);
411 this_cpu_dec(nr_dentry_unused);
412 }
413
414 static void d_shrink_add(struct dentry *dentry, struct list_head *list)
415 {
416 D_FLAG_VERIFY(dentry, 0);
417 list_add(&dentry->d_lru, list);
418 dentry->d_flags |= DCACHE_SHRINK_LIST | DCACHE_LRU_LIST;
419 this_cpu_inc(nr_dentry_unused);
420 }
421
422 /*
423 * These can only be called under the global LRU lock, ie during the
424 * callback for freeing the LRU list. "isolate" removes it from the
425 * LRU lists entirely, while shrink_move moves it to the indicated
426 * private list.
427 */
428 static void d_lru_isolate(struct list_lru_one *lru, struct dentry *dentry)
429 {
430 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
431 dentry->d_flags &= ~DCACHE_LRU_LIST;
432 this_cpu_dec(nr_dentry_unused);
433 list_lru_isolate(lru, &dentry->d_lru);
434 }
435
436 static void d_lru_shrink_move(struct list_lru_one *lru, struct dentry *dentry,
437 struct list_head *list)
438 {
439 D_FLAG_VERIFY(dentry, DCACHE_LRU_LIST);
440 dentry->d_flags |= DCACHE_SHRINK_LIST;
441 list_lru_isolate_move(lru, &dentry->d_lru, list);
442 }
443
444 /*
445 * dentry_lru_(add|del)_list) must be called with d_lock held.
446 */
447 static void dentry_lru_add(struct dentry *dentry)
448 {
449 if (unlikely(!(dentry->d_flags & DCACHE_LRU_LIST)))
450 d_lru_add(dentry);
451 }
452
453 /**
454 * d_drop - drop a dentry
455 * @dentry: dentry to drop
456 *
457 * d_drop() unhashes the entry from the parent dentry hashes, so that it won't
458 * be found through a VFS lookup any more. Note that this is different from
459 * deleting the dentry - d_delete will try to mark the dentry negative if
460 * possible, giving a successful _negative_ lookup, while d_drop will
461 * just make the cache lookup fail.
462 *
463 * d_drop() is used mainly for stuff that wants to invalidate a dentry for some
464 * reason (NFS timeouts or autofs deletes).
465 *
466 * __d_drop requires dentry->d_lock.
467 */
468 void __d_drop(struct dentry *dentry)
469 {
470 if (!d_unhashed(dentry)) {
471 struct hlist_bl_head *b;
472 /*
473 * Hashed dentries are normally on the dentry hashtable,
474 * with the exception of those newly allocated by
475 * d_obtain_alias, which are always IS_ROOT:
476 */
477 if (unlikely(IS_ROOT(dentry)))
478 b = &dentry->d_sb->s_anon;
479 else
480 b = d_hash(dentry->d_parent, dentry->d_name.hash);
481
482 hlist_bl_lock(b);
483 __hlist_bl_del(&dentry->d_hash);
484 dentry->d_hash.pprev = NULL;
485 hlist_bl_unlock(b);
486 dentry_rcuwalk_invalidate(dentry);
487 }
488 }
489 EXPORT_SYMBOL(__d_drop);
490
491 void d_drop(struct dentry *dentry)
492 {
493 spin_lock(&dentry->d_lock);
494 __d_drop(dentry);
495 spin_unlock(&dentry->d_lock);
496 }
497 EXPORT_SYMBOL(d_drop);
498
499 static void __dentry_kill(struct dentry *dentry)
500 {
501 struct dentry *parent = NULL;
502 bool can_free = true;
503 if (!IS_ROOT(dentry))
504 parent = dentry->d_parent;
505
506 /*
507 * The dentry is now unrecoverably dead to the world.
508 */
509 lockref_mark_dead(&dentry->d_lockref);
510
511 /*
512 * inform the fs via d_prune that this dentry is about to be
513 * unhashed and destroyed.
514 */
515 if (dentry->d_flags & DCACHE_OP_PRUNE)
516 dentry->d_op->d_prune(dentry);
517
518 if (dentry->d_flags & DCACHE_LRU_LIST) {
519 if (!(dentry->d_flags & DCACHE_SHRINK_LIST))
520 d_lru_del(dentry);
521 }
522 /* if it was on the hash then remove it */
523 __d_drop(dentry);
524 __list_del_entry(&dentry->d_child);
525 /*
526 * Inform d_walk() that we are no longer attached to the
527 * dentry tree
528 */
529 dentry->d_flags |= DCACHE_DENTRY_KILLED;
530 if (parent)
531 spin_unlock(&parent->d_lock);
532 dentry_iput(dentry);
533 /*
534 * dentry_iput drops the locks, at which point nobody (except
535 * transient RCU lookups) can reach this dentry.
536 */
537 BUG_ON(dentry->d_lockref.count > 0);
538 this_cpu_dec(nr_dentry);
539 if (dentry->d_op && dentry->d_op->d_release)
540 dentry->d_op->d_release(dentry);
541
542 spin_lock(&dentry->d_lock);
543 if (dentry->d_flags & DCACHE_SHRINK_LIST) {
544 dentry->d_flags |= DCACHE_MAY_FREE;
545 can_free = false;
546 }
547 spin_unlock(&dentry->d_lock);
548 if (likely(can_free))
549 dentry_free(dentry);
550 }
551
552 /*
553 * Finish off a dentry we've decided to kill.
554 * dentry->d_lock must be held, returns with it unlocked.
555 * If ref is non-zero, then decrement the refcount too.
556 * Returns dentry requiring refcount drop, or NULL if we're done.
557 */
558 static struct dentry *dentry_kill(struct dentry *dentry)
559 __releases(dentry->d_lock)
560 {
561 struct inode *inode = dentry->d_inode;
562 struct dentry *parent = NULL;
563
564 if (inode && unlikely(!spin_trylock(&inode->i_lock)))
565 goto failed;
566
567 if (!IS_ROOT(dentry)) {
568 parent = dentry->d_parent;
569 if (unlikely(!spin_trylock(&parent->d_lock))) {
570 if (inode)
571 spin_unlock(&inode->i_lock);
572 goto failed;
573 }
574 }
575
576 __dentry_kill(dentry);
577 return parent;
578
579 failed:
580 spin_unlock(&dentry->d_lock);
581 cpu_relax();
582 return dentry; /* try again with same dentry */
583 }
584
585 static inline struct dentry *lock_parent(struct dentry *dentry)
586 {
587 struct dentry *parent = dentry->d_parent;
588 if (IS_ROOT(dentry))
589 return NULL;
590 if (unlikely(dentry->d_lockref.count < 0))
591 return NULL;
592 if (likely(spin_trylock(&parent->d_lock)))
593 return parent;
594 rcu_read_lock();
595 spin_unlock(&dentry->d_lock);
596 again:
597 parent = ACCESS_ONCE(dentry->d_parent);
598 spin_lock(&parent->d_lock);
599 /*
600 * We can't blindly lock dentry until we are sure
601 * that we won't violate the locking order.
602 * Any changes of dentry->d_parent must have
603 * been done with parent->d_lock held, so
604 * spin_lock() above is enough of a barrier
605 * for checking if it's still our child.
606 */
607 if (unlikely(parent != dentry->d_parent)) {
608 spin_unlock(&parent->d_lock);
609 goto again;
610 }
611 rcu_read_unlock();
612 if (parent != dentry)
613 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
614 else
615 parent = NULL;
616 return parent;
617 }
618
619 /*
620 * Try to do a lockless dput(), and return whether that was successful.
621 *
622 * If unsuccessful, we return false, having already taken the dentry lock.
623 *
624 * The caller needs to hold the RCU read lock, so that the dentry is
625 * guaranteed to stay around even if the refcount goes down to zero!
626 */
627 static inline bool fast_dput(struct dentry *dentry)
628 {
629 int ret;
630 unsigned int d_flags;
631
632 /*
633 * If we have a d_op->d_delete() operation, we sould not
634 * let the dentry count go to zero, so use "put_or_lock".
635 */
636 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE))
637 return lockref_put_or_lock(&dentry->d_lockref);
638
639 /*
640 * .. otherwise, we can try to just decrement the
641 * lockref optimistically.
642 */
643 ret = lockref_put_return(&dentry->d_lockref);
644
645 /*
646 * If the lockref_put_return() failed due to the lock being held
647 * by somebody else, the fast path has failed. We will need to
648 * get the lock, and then check the count again.
649 */
650 if (unlikely(ret < 0)) {
651 spin_lock(&dentry->d_lock);
652 if (dentry->d_lockref.count > 1) {
653 dentry->d_lockref.count--;
654 spin_unlock(&dentry->d_lock);
655 return 1;
656 }
657 return 0;
658 }
659
660 /*
661 * If we weren't the last ref, we're done.
662 */
663 if (ret)
664 return 1;
665
666 /*
667 * Careful, careful. The reference count went down
668 * to zero, but we don't hold the dentry lock, so
669 * somebody else could get it again, and do another
670 * dput(), and we need to not race with that.
671 *
672 * However, there is a very special and common case
673 * where we don't care, because there is nothing to
674 * do: the dentry is still hashed, it does not have
675 * a 'delete' op, and it's referenced and already on
676 * the LRU list.
677 *
678 * NOTE! Since we aren't locked, these values are
679 * not "stable". However, it is sufficient that at
680 * some point after we dropped the reference the
681 * dentry was hashed and the flags had the proper
682 * value. Other dentry users may have re-gotten
683 * a reference to the dentry and change that, but
684 * our work is done - we can leave the dentry
685 * around with a zero refcount.
686 */
687 smp_rmb();
688 d_flags = ACCESS_ONCE(dentry->d_flags);
689 d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST | DCACHE_DISCONNECTED;
690
691 /* Nothing to do? Dropping the reference was all we needed? */
692 if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry))
693 return 1;
694
695 /*
696 * Not the fast normal case? Get the lock. We've already decremented
697 * the refcount, but we'll need to re-check the situation after
698 * getting the lock.
699 */
700 spin_lock(&dentry->d_lock);
701
702 /*
703 * Did somebody else grab a reference to it in the meantime, and
704 * we're no longer the last user after all? Alternatively, somebody
705 * else could have killed it and marked it dead. Either way, we
706 * don't need to do anything else.
707 */
708 if (dentry->d_lockref.count) {
709 spin_unlock(&dentry->d_lock);
710 return 1;
711 }
712
713 /*
714 * Re-get the reference we optimistically dropped. We hold the
715 * lock, and we just tested that it was zero, so we can just
716 * set it to 1.
717 */
718 dentry->d_lockref.count = 1;
719 return 0;
720 }
721
722
723 /*
724 * This is dput
725 *
726 * This is complicated by the fact that we do not want to put
727 * dentries that are no longer on any hash chain on the unused
728 * list: we'd much rather just get rid of them immediately.
729 *
730 * However, that implies that we have to traverse the dentry
731 * tree upwards to the parents which might _also_ now be
732 * scheduled for deletion (it may have been only waiting for
733 * its last child to go away).
734 *
735 * This tail recursion is done by hand as we don't want to depend
736 * on the compiler to always get this right (gcc generally doesn't).
737 * Real recursion would eat up our stack space.
738 */
739
740 /*
741 * dput - release a dentry
742 * @dentry: dentry to release
743 *
744 * Release a dentry. This will drop the usage count and if appropriate
745 * call the dentry unlink method as well as removing it from the queues and
746 * releasing its resources. If the parent dentries were scheduled for release
747 * they too may now get deleted.
748 */
749 void dput(struct dentry *dentry)
750 {
751 if (unlikely(!dentry))
752 return;
753
754 repeat:
755 rcu_read_lock();
756 if (likely(fast_dput(dentry))) {
757 rcu_read_unlock();
758 return;
759 }
760
761 /* Slow case: now with the dentry lock held */
762 rcu_read_unlock();
763
764 /* Unreachable? Get rid of it */
765 if (unlikely(d_unhashed(dentry)))
766 goto kill_it;
767
768 if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
769 goto kill_it;
770
771 if (unlikely(dentry->d_flags & DCACHE_OP_DELETE)) {
772 if (dentry->d_op->d_delete(dentry))
773 goto kill_it;
774 }
775
776 if (!(dentry->d_flags & DCACHE_REFERENCED))
777 dentry->d_flags |= DCACHE_REFERENCED;
778 dentry_lru_add(dentry);
779
780 dentry->d_lockref.count--;
781 spin_unlock(&dentry->d_lock);
782 return;
783
784 kill_it:
785 dentry = dentry_kill(dentry);
786 if (dentry)
787 goto repeat;
788 }
789 EXPORT_SYMBOL(dput);
790
791
792 /* This must be called with d_lock held */
793 static inline void __dget_dlock(struct dentry *dentry)
794 {
795 dentry->d_lockref.count++;
796 }
797
798 static inline void __dget(struct dentry *dentry)
799 {
800 lockref_get(&dentry->d_lockref);
801 }
802
803 struct dentry *dget_parent(struct dentry *dentry)
804 {
805 int gotref;
806 struct dentry *ret;
807
808 /*
809 * Do optimistic parent lookup without any
810 * locking.
811 */
812 rcu_read_lock();
813 ret = ACCESS_ONCE(dentry->d_parent);
814 gotref = lockref_get_not_zero(&ret->d_lockref);
815 rcu_read_unlock();
816 if (likely(gotref)) {
817 if (likely(ret == ACCESS_ONCE(dentry->d_parent)))
818 return ret;
819 dput(ret);
820 }
821
822 repeat:
823 /*
824 * Don't need rcu_dereference because we re-check it was correct under
825 * the lock.
826 */
827 rcu_read_lock();
828 ret = dentry->d_parent;
829 spin_lock(&ret->d_lock);
830 if (unlikely(ret != dentry->d_parent)) {
831 spin_unlock(&ret->d_lock);
832 rcu_read_unlock();
833 goto repeat;
834 }
835 rcu_read_unlock();
836 BUG_ON(!ret->d_lockref.count);
837 ret->d_lockref.count++;
838 spin_unlock(&ret->d_lock);
839 return ret;
840 }
841 EXPORT_SYMBOL(dget_parent);
842
843 /**
844 * d_find_alias - grab a hashed alias of inode
845 * @inode: inode in question
846 *
847 * If inode has a hashed alias, or is a directory and has any alias,
848 * acquire the reference to alias and return it. Otherwise return NULL.
849 * Notice that if inode is a directory there can be only one alias and
850 * it can be unhashed only if it has no children, or if it is the root
851 * of a filesystem, or if the directory was renamed and d_revalidate
852 * was the first vfs operation to notice.
853 *
854 * If the inode has an IS_ROOT, DCACHE_DISCONNECTED alias, then prefer
855 * any other hashed alias over that one.
856 */
857 static struct dentry *__d_find_alias(struct inode *inode)
858 {
859 struct dentry *alias, *discon_alias;
860
861 again:
862 discon_alias = NULL;
863 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
864 spin_lock(&alias->d_lock);
865 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
866 if (IS_ROOT(alias) &&
867 (alias->d_flags & DCACHE_DISCONNECTED)) {
868 discon_alias = alias;
869 } else {
870 __dget_dlock(alias);
871 spin_unlock(&alias->d_lock);
872 return alias;
873 }
874 }
875 spin_unlock(&alias->d_lock);
876 }
877 if (discon_alias) {
878 alias = discon_alias;
879 spin_lock(&alias->d_lock);
880 if (S_ISDIR(inode->i_mode) || !d_unhashed(alias)) {
881 __dget_dlock(alias);
882 spin_unlock(&alias->d_lock);
883 return alias;
884 }
885 spin_unlock(&alias->d_lock);
886 goto again;
887 }
888 return NULL;
889 }
890
891 struct dentry *d_find_alias(struct inode *inode)
892 {
893 struct dentry *de = NULL;
894
895 if (!hlist_empty(&inode->i_dentry)) {
896 spin_lock(&inode->i_lock);
897 de = __d_find_alias(inode);
898 spin_unlock(&inode->i_lock);
899 }
900 return de;
901 }
902 EXPORT_SYMBOL(d_find_alias);
903
904 /*
905 * Try to kill dentries associated with this inode.
906 * WARNING: you must own a reference to inode.
907 */
908 void d_prune_aliases(struct inode *inode)
909 {
910 struct dentry *dentry;
911 restart:
912 spin_lock(&inode->i_lock);
913 hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
914 spin_lock(&dentry->d_lock);
915 if (!dentry->d_lockref.count) {
916 struct dentry *parent = lock_parent(dentry);
917 if (likely(!dentry->d_lockref.count)) {
918 __dentry_kill(dentry);
919 dput(parent);
920 goto restart;
921 }
922 if (parent)
923 spin_unlock(&parent->d_lock);
924 }
925 spin_unlock(&dentry->d_lock);
926 }
927 spin_unlock(&inode->i_lock);
928 }
929 EXPORT_SYMBOL(d_prune_aliases);
930
931 static void shrink_dentry_list(struct list_head *list)
932 {
933 struct dentry *dentry, *parent;
934
935 while (!list_empty(list)) {
936 struct inode *inode;
937 dentry = list_entry(list->prev, struct dentry, d_lru);
938 spin_lock(&dentry->d_lock);
939 parent = lock_parent(dentry);
940
941 /*
942 * The dispose list is isolated and dentries are not accounted
943 * to the LRU here, so we can simply remove it from the list
944 * here regardless of whether it is referenced or not.
945 */
946 d_shrink_del(dentry);
947
948 /*
949 * We found an inuse dentry which was not removed from
950 * the LRU because of laziness during lookup. Do not free it.
951 */
952 if (dentry->d_lockref.count > 0) {
953 spin_unlock(&dentry->d_lock);
954 if (parent)
955 spin_unlock(&parent->d_lock);
956 continue;
957 }
958
959
960 if (unlikely(dentry->d_flags & DCACHE_DENTRY_KILLED)) {
961 bool can_free = dentry->d_flags & DCACHE_MAY_FREE;
962 spin_unlock(&dentry->d_lock);
963 if (parent)
964 spin_unlock(&parent->d_lock);
965 if (can_free)
966 dentry_free(dentry);
967 continue;
968 }
969
970 inode = dentry->d_inode;
971 if (inode && unlikely(!spin_trylock(&inode->i_lock))) {
972 d_shrink_add(dentry, list);
973 spin_unlock(&dentry->d_lock);
974 if (parent)
975 spin_unlock(&parent->d_lock);
976 continue;
977 }
978
979 __dentry_kill(dentry);
980
981 /*
982 * We need to prune ancestors too. This is necessary to prevent
983 * quadratic behavior of shrink_dcache_parent(), but is also
984 * expected to be beneficial in reducing dentry cache
985 * fragmentation.
986 */
987 dentry = parent;
988 while (dentry && !lockref_put_or_lock(&dentry->d_lockref)) {
989 parent = lock_parent(dentry);
990 if (dentry->d_lockref.count != 1) {
991 dentry->d_lockref.count--;
992 spin_unlock(&dentry->d_lock);
993 if (parent)
994 spin_unlock(&parent->d_lock);
995 break;
996 }
997 inode = dentry->d_inode; /* can't be NULL */
998 if (unlikely(!spin_trylock(&inode->i_lock))) {
999 spin_unlock(&dentry->d_lock);
1000 if (parent)
1001 spin_unlock(&parent->d_lock);
1002 cpu_relax();
1003 continue;
1004 }
1005 __dentry_kill(dentry);
1006 dentry = parent;
1007 }
1008 }
1009 }
1010
1011 static enum lru_status dentry_lru_isolate(struct list_head *item,
1012 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1013 {
1014 struct list_head *freeable = arg;
1015 struct dentry *dentry = container_of(item, struct dentry, d_lru);
1016
1017
1018 /*
1019 * we are inverting the lru lock/dentry->d_lock here,
1020 * so use a trylock. If we fail to get the lock, just skip
1021 * it
1022 */
1023 if (!spin_trylock(&dentry->d_lock))
1024 return LRU_SKIP;
1025
1026 /*
1027 * Referenced dentries are still in use. If they have active
1028 * counts, just remove them from the LRU. Otherwise give them
1029 * another pass through the LRU.
1030 */
1031 if (dentry->d_lockref.count) {
1032 d_lru_isolate(lru, dentry);
1033 spin_unlock(&dentry->d_lock);
1034 return LRU_REMOVED;
1035 }
1036
1037 if (dentry->d_flags & DCACHE_REFERENCED) {
1038 dentry->d_flags &= ~DCACHE_REFERENCED;
1039 spin_unlock(&dentry->d_lock);
1040
1041 /*
1042 * The list move itself will be made by the common LRU code. At
1043 * this point, we've dropped the dentry->d_lock but keep the
1044 * lru lock. This is safe to do, since every list movement is
1045 * protected by the lru lock even if both locks are held.
1046 *
1047 * This is guaranteed by the fact that all LRU management
1048 * functions are intermediated by the LRU API calls like
1049 * list_lru_add and list_lru_del. List movement in this file
1050 * only ever occur through this functions or through callbacks
1051 * like this one, that are called from the LRU API.
1052 *
1053 * The only exceptions to this are functions like
1054 * shrink_dentry_list, and code that first checks for the
1055 * DCACHE_SHRINK_LIST flag. Those are guaranteed to be
1056 * operating only with stack provided lists after they are
1057 * properly isolated from the main list. It is thus, always a
1058 * local access.
1059 */
1060 return LRU_ROTATE;
1061 }
1062
1063 d_lru_shrink_move(lru, dentry, freeable);
1064 spin_unlock(&dentry->d_lock);
1065
1066 return LRU_REMOVED;
1067 }
1068
1069 /**
1070 * prune_dcache_sb - shrink the dcache
1071 * @sb: superblock
1072 * @sc: shrink control, passed to list_lru_shrink_walk()
1073 *
1074 * Attempt to shrink the superblock dcache LRU by @sc->nr_to_scan entries. This
1075 * is done when we need more memory and called from the superblock shrinker
1076 * function.
1077 *
1078 * This function may fail to free any resources if all the dentries are in
1079 * use.
1080 */
1081 long prune_dcache_sb(struct super_block *sb, struct shrink_control *sc)
1082 {
1083 LIST_HEAD(dispose);
1084 long freed;
1085
1086 freed = list_lru_shrink_walk(&sb->s_dentry_lru, sc,
1087 dentry_lru_isolate, &dispose);
1088 shrink_dentry_list(&dispose);
1089 return freed;
1090 }
1091
1092 static enum lru_status dentry_lru_isolate_shrink(struct list_head *item,
1093 struct list_lru_one *lru, spinlock_t *lru_lock, void *arg)
1094 {
1095 struct list_head *freeable = arg;
1096 struct dentry *dentry = container_of(item, struct dentry, d_lru);
1097
1098 /*
1099 * we are inverting the lru lock/dentry->d_lock here,
1100 * so use a trylock. If we fail to get the lock, just skip
1101 * it
1102 */
1103 if (!spin_trylock(&dentry->d_lock))
1104 return LRU_SKIP;
1105
1106 d_lru_shrink_move(lru, dentry, freeable);
1107 spin_unlock(&dentry->d_lock);
1108
1109 return LRU_REMOVED;
1110 }
1111
1112
1113 /**
1114 * shrink_dcache_sb - shrink dcache for a superblock
1115 * @sb: superblock
1116 *
1117 * Shrink the dcache for the specified super block. This is used to free
1118 * the dcache before unmounting a file system.
1119 */
1120 void shrink_dcache_sb(struct super_block *sb)
1121 {
1122 long freed;
1123
1124 do {
1125 LIST_HEAD(dispose);
1126
1127 freed = list_lru_walk(&sb->s_dentry_lru,
1128 dentry_lru_isolate_shrink, &dispose, UINT_MAX);
1129
1130 this_cpu_sub(nr_dentry_unused, freed);
1131 shrink_dentry_list(&dispose);
1132 } while (freed > 0);
1133 }
1134 EXPORT_SYMBOL(shrink_dcache_sb);
1135
1136 /**
1137 * enum d_walk_ret - action to talke during tree walk
1138 * @D_WALK_CONTINUE: contrinue walk
1139 * @D_WALK_QUIT: quit walk
1140 * @D_WALK_NORETRY: quit when retry is needed
1141 * @D_WALK_SKIP: skip this dentry and its children
1142 */
1143 enum d_walk_ret {
1144 D_WALK_CONTINUE,
1145 D_WALK_QUIT,
1146 D_WALK_NORETRY,
1147 D_WALK_SKIP,
1148 };
1149
1150 /**
1151 * d_walk - walk the dentry tree
1152 * @parent: start of walk
1153 * @data: data passed to @enter() and @finish()
1154 * @enter: callback when first entering the dentry
1155 * @finish: callback when successfully finished the walk
1156 *
1157 * The @enter() and @finish() callbacks are called with d_lock held.
1158 */
1159 static void d_walk(struct dentry *parent, void *data,
1160 enum d_walk_ret (*enter)(void *, struct dentry *),
1161 void (*finish)(void *))
1162 {
1163 struct dentry *this_parent;
1164 struct list_head *next;
1165 unsigned seq = 0;
1166 enum d_walk_ret ret;
1167 bool retry = true;
1168
1169 again:
1170 read_seqbegin_or_lock(&rename_lock, &seq);
1171 this_parent = parent;
1172 spin_lock(&this_parent->d_lock);
1173
1174 ret = enter(data, this_parent);
1175 switch (ret) {
1176 case D_WALK_CONTINUE:
1177 break;
1178 case D_WALK_QUIT:
1179 case D_WALK_SKIP:
1180 goto out_unlock;
1181 case D_WALK_NORETRY:
1182 retry = false;
1183 break;
1184 }
1185 repeat:
1186 next = this_parent->d_subdirs.next;
1187 resume:
1188 while (next != &this_parent->d_subdirs) {
1189 struct list_head *tmp = next;
1190 struct dentry *dentry = list_entry(tmp, struct dentry, d_child);
1191 next = tmp->next;
1192
1193 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
1194
1195 ret = enter(data, dentry);
1196 switch (ret) {
1197 case D_WALK_CONTINUE:
1198 break;
1199 case D_WALK_QUIT:
1200 spin_unlock(&dentry->d_lock);
1201 goto out_unlock;
1202 case D_WALK_NORETRY:
1203 retry = false;
1204 break;
1205 case D_WALK_SKIP:
1206 spin_unlock(&dentry->d_lock);
1207 continue;
1208 }
1209
1210 if (!list_empty(&dentry->d_subdirs)) {
1211 spin_unlock(&this_parent->d_lock);
1212 spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
1213 this_parent = dentry;
1214 spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
1215 goto repeat;
1216 }
1217 spin_unlock(&dentry->d_lock);
1218 }
1219 /*
1220 * All done at this level ... ascend and resume the search.
1221 */
1222 rcu_read_lock();
1223 ascend:
1224 if (this_parent != parent) {
1225 struct dentry *child = this_parent;
1226 this_parent = child->d_parent;
1227
1228 spin_unlock(&child->d_lock);
1229 spin_lock(&this_parent->d_lock);
1230
1231 /* might go back up the wrong parent if we have had a rename. */
1232 if (need_seqretry(&rename_lock, seq))
1233 goto rename_retry;
1234 /* go into the first sibling still alive */
1235 do {
1236 next = child->d_child.next;
1237 if (next == &this_parent->d_subdirs)
1238 goto ascend;
1239 child = list_entry(next, struct dentry, d_child);
1240 } while (unlikely(child->d_flags & DCACHE_DENTRY_KILLED));
1241 rcu_read_unlock();
1242 goto resume;
1243 }
1244 if (need_seqretry(&rename_lock, seq))
1245 goto rename_retry;
1246 rcu_read_unlock();
1247 if (finish)
1248 finish(data);
1249
1250 out_unlock:
1251 spin_unlock(&this_parent->d_lock);
1252 done_seqretry(&rename_lock, seq);
1253 return;
1254
1255 rename_retry:
1256 spin_unlock(&this_parent->d_lock);
1257 rcu_read_unlock();
1258 BUG_ON(seq & 1);
1259 if (!retry)
1260 return;
1261 seq = 1;
1262 goto again;
1263 }
1264
1265 /*
1266 * Search for at least 1 mount point in the dentry's subdirs.
1267 * We descend to the next level whenever the d_subdirs
1268 * list is non-empty and continue searching.
1269 */
1270
1271 static enum d_walk_ret check_mount(void *data, struct dentry *dentry)
1272 {
1273 int *ret = data;
1274 if (d_mountpoint(dentry)) {
1275 *ret = 1;
1276 return D_WALK_QUIT;
1277 }
1278 return D_WALK_CONTINUE;
1279 }
1280
1281 /**
1282 * have_submounts - check for mounts over a dentry
1283 * @parent: dentry to check.
1284 *
1285 * Return true if the parent or its subdirectories contain
1286 * a mount point
1287 */
1288 int have_submounts(struct dentry *parent)
1289 {
1290 int ret = 0;
1291
1292 d_walk(parent, &ret, check_mount, NULL);
1293
1294 return ret;
1295 }
1296 EXPORT_SYMBOL(have_submounts);
1297
1298 /*
1299 * Called by mount code to set a mountpoint and check if the mountpoint is
1300 * reachable (e.g. NFS can unhash a directory dentry and then the complete
1301 * subtree can become unreachable).
1302 *
1303 * Only one of d_invalidate() and d_set_mounted() must succeed. For
1304 * this reason take rename_lock and d_lock on dentry and ancestors.
1305 */
1306 int d_set_mounted(struct dentry *dentry)
1307 {
1308 struct dentry *p;
1309 int ret = -ENOENT;
1310 write_seqlock(&rename_lock);
1311 for (p = dentry->d_parent; !IS_ROOT(p); p = p->d_parent) {
1312 /* Need exclusion wrt. d_invalidate() */
1313 spin_lock(&p->d_lock);
1314 if (unlikely(d_unhashed(p))) {
1315 spin_unlock(&p->d_lock);
1316 goto out;
1317 }
1318 spin_unlock(&p->d_lock);
1319 }
1320 spin_lock(&dentry->d_lock);
1321 if (!d_unlinked(dentry)) {
1322 dentry->d_flags |= DCACHE_MOUNTED;
1323 ret = 0;
1324 }
1325 spin_unlock(&dentry->d_lock);
1326 out:
1327 write_sequnlock(&rename_lock);
1328 return ret;
1329 }
1330
1331 /*
1332 * Search the dentry child list of the specified parent,
1333 * and move any unused dentries to the end of the unused
1334 * list for prune_dcache(). We descend to the next level
1335 * whenever the d_subdirs list is non-empty and continue
1336 * searching.
1337 *
1338 * It returns zero iff there are no unused children,
1339 * otherwise it returns the number of children moved to
1340 * the end of the unused list. This may not be the total
1341 * number of unused children, because select_parent can
1342 * drop the lock and return early due to latency
1343 * constraints.
1344 */
1345
1346 struct select_data {
1347 struct dentry *start;
1348 struct list_head dispose;
1349 int found;
1350 };
1351
1352 static enum d_walk_ret select_collect(void *_data, struct dentry *dentry)
1353 {
1354 struct select_data *data = _data;
1355 enum d_walk_ret ret = D_WALK_CONTINUE;
1356
1357 if (data->start == dentry)
1358 goto out;
1359
1360 if (dentry->d_flags & DCACHE_SHRINK_LIST) {
1361 data->found++;
1362 } else {
1363 if (dentry->d_flags & DCACHE_LRU_LIST)
1364 d_lru_del(dentry);
1365 if (!dentry->d_lockref.count) {
1366 d_shrink_add(dentry, &data->dispose);
1367 data->found++;
1368 }
1369 }
1370 /*
1371 * We can return to the caller if we have found some (this
1372 * ensures forward progress). We'll be coming back to find
1373 * the rest.
1374 */
1375 if (!list_empty(&data->dispose))
1376 ret = need_resched() ? D_WALK_QUIT : D_WALK_NORETRY;
1377 out:
1378 return ret;
1379 }
1380
1381 /**
1382 * shrink_dcache_parent - prune dcache
1383 * @parent: parent of entries to prune
1384 *
1385 * Prune the dcache to remove unused children of the parent dentry.
1386 */
1387 void shrink_dcache_parent(struct dentry *parent)
1388 {
1389 for (;;) {
1390 struct select_data data;
1391
1392 INIT_LIST_HEAD(&data.dispose);
1393 data.start = parent;
1394 data.found = 0;
1395
1396 d_walk(parent, &data, select_collect, NULL);
1397 if (!data.found)
1398 break;
1399
1400 shrink_dentry_list(&data.dispose);
1401 cond_resched();
1402 }
1403 }
1404 EXPORT_SYMBOL(shrink_dcache_parent);
1405
1406 static enum d_walk_ret umount_check(void *_data, struct dentry *dentry)
1407 {
1408 /* it has busy descendents; complain about those instead */
1409 if (!list_empty(&dentry->d_subdirs))
1410 return D_WALK_CONTINUE;
1411
1412 /* root with refcount 1 is fine */
1413 if (dentry == _data && dentry->d_lockref.count == 1)
1414 return D_WALK_CONTINUE;
1415
1416 printk(KERN_ERR "BUG: Dentry %p{i=%lx,n=%pd} "
1417 " still in use (%d) [unmount of %s %s]\n",
1418 dentry,
1419 dentry->d_inode ?
1420 dentry->d_inode->i_ino : 0UL,
1421 dentry,
1422 dentry->d_lockref.count,
1423 dentry->d_sb->s_type->name,
1424 dentry->d_sb->s_id);
1425 WARN_ON(1);
1426 return D_WALK_CONTINUE;
1427 }
1428
1429 static void do_one_tree(struct dentry *dentry)
1430 {
1431 shrink_dcache_parent(dentry);
1432 d_walk(dentry, dentry, umount_check, NULL);
1433 d_drop(dentry);
1434 dput(dentry);
1435 }
1436
1437 /*
1438 * destroy the dentries attached to a superblock on unmounting
1439 */
1440 void shrink_dcache_for_umount(struct super_block *sb)
1441 {
1442 struct dentry *dentry;
1443
1444 WARN(down_read_trylock(&sb->s_umount), "s_umount should've been locked");
1445
1446 dentry = sb->s_root;
1447 sb->s_root = NULL;
1448 do_one_tree(dentry);
1449
1450 while (!hlist_bl_empty(&sb->s_anon)) {
1451 dentry = dget(hlist_bl_entry(hlist_bl_first(&sb->s_anon), struct dentry, d_hash));
1452 do_one_tree(dentry);
1453 }
1454 }
1455
1456 struct detach_data {
1457 struct select_data select;
1458 struct dentry *mountpoint;
1459 };
1460 static enum d_walk_ret detach_and_collect(void *_data, struct dentry *dentry)
1461 {
1462 struct detach_data *data = _data;
1463
1464 if (d_mountpoint(dentry)) {
1465 __dget_dlock(dentry);
1466 data->mountpoint = dentry;
1467 return D_WALK_QUIT;
1468 }
1469
1470 return select_collect(&data->select, dentry);
1471 }
1472
1473 static void check_and_drop(void *_data)
1474 {
1475 struct detach_data *data = _data;
1476
1477 if (!data->mountpoint && !data->select.found)
1478 __d_drop(data->select.start);
1479 }
1480
1481 /**
1482 * d_invalidate - detach submounts, prune dcache, and drop
1483 * @dentry: dentry to invalidate (aka detach, prune and drop)
1484 *
1485 * no dcache lock.
1486 *
1487 * The final d_drop is done as an atomic operation relative to
1488 * rename_lock ensuring there are no races with d_set_mounted. This
1489 * ensures there are no unhashed dentries on the path to a mountpoint.
1490 */
1491 void d_invalidate(struct dentry *dentry)
1492 {
1493 /*
1494 * If it's already been dropped, return OK.
1495 */
1496 spin_lock(&dentry->d_lock);
1497 if (d_unhashed(dentry)) {
1498 spin_unlock(&dentry->d_lock);
1499 return;
1500 }
1501 spin_unlock(&dentry->d_lock);
1502
1503 /* Negative dentries can be dropped without further checks */
1504 if (!dentry->d_inode) {
1505 d_drop(dentry);
1506 return;
1507 }
1508
1509 for (;;) {
1510 struct detach_data data;
1511
1512 data.mountpoint = NULL;
1513 INIT_LIST_HEAD(&data.select.dispose);
1514 data.select.start = dentry;
1515 data.select.found = 0;
1516
1517 d_walk(dentry, &data, detach_and_collect, check_and_drop);
1518
1519 if (data.select.found)
1520 shrink_dentry_list(&data.select.dispose);
1521
1522 if (data.mountpoint) {
1523 detach_mounts(data.mountpoint);
1524 dput(data.mountpoint);
1525 }
1526
1527 if (!data.mountpoint && !data.select.found)
1528 break;
1529
1530 cond_resched();
1531 }
1532 }
1533 EXPORT_SYMBOL(d_invalidate);
1534
1535 /**
1536 * __d_alloc - allocate a dcache entry
1537 * @sb: filesystem it will belong to
1538 * @name: qstr of the name
1539 *
1540 * Allocates a dentry. It returns %NULL if there is insufficient memory
1541 * available. On a success the dentry is returned. The name passed in is
1542 * copied and the copy passed in may be reused after this call.
1543 */
1544
1545 struct dentry *__d_alloc(struct super_block *sb, const struct qstr *name)
1546 {
1547 struct dentry *dentry;
1548 char *dname;
1549
1550 dentry = kmem_cache_alloc(dentry_cache, GFP_KERNEL);
1551 if (!dentry)
1552 return NULL;
1553
1554 /*
1555 * We guarantee that the inline name is always NUL-terminated.
1556 * This way the memcpy() done by the name switching in rename
1557 * will still always have a NUL at the end, even if we might
1558 * be overwriting an internal NUL character
1559 */
1560 dentry->d_iname[DNAME_INLINE_LEN-1] = 0;
1561 if (name->len > DNAME_INLINE_LEN-1) {
1562 size_t size = offsetof(struct external_name, name[1]);
1563 struct external_name *p = kmalloc(size + name->len,
1564 GFP_KERNEL_ACCOUNT);
1565 if (!p) {
1566 kmem_cache_free(dentry_cache, dentry);
1567 return NULL;
1568 }
1569 atomic_set(&p->u.count, 1);
1570 dname = p->name;
1571 if (IS_ENABLED(CONFIG_DCACHE_WORD_ACCESS))
1572 kasan_unpoison_shadow(dname,
1573 round_up(name->len + 1, sizeof(unsigned long)));
1574 } else {
1575 dname = dentry->d_iname;
1576 }
1577
1578 dentry->d_name.len = name->len;
1579 dentry->d_name.hash = name->hash;
1580 memcpy(dname, name->name, name->len);
1581 dname[name->len] = 0;
1582
1583 /* Make sure we always see the terminating NUL character */
1584 smp_wmb();
1585 dentry->d_name.name = dname;
1586
1587 dentry->d_lockref.count = 1;
1588 dentry->d_flags = 0;
1589 spin_lock_init(&dentry->d_lock);
1590 seqcount_init(&dentry->d_seq);
1591 dentry->d_inode = NULL;
1592 dentry->d_parent = dentry;
1593 dentry->d_sb = sb;
1594 dentry->d_op = NULL;
1595 dentry->d_fsdata = NULL;
1596 INIT_HLIST_BL_NODE(&dentry->d_hash);
1597 INIT_LIST_HEAD(&dentry->d_lru);
1598 INIT_LIST_HEAD(&dentry->d_subdirs);
1599 INIT_HLIST_NODE(&dentry->d_u.d_alias);
1600 INIT_LIST_HEAD(&dentry->d_child);
1601 d_set_d_op(dentry, dentry->d_sb->s_d_op);
1602
1603 this_cpu_inc(nr_dentry);
1604
1605 return dentry;
1606 }
1607
1608 /**
1609 * d_alloc - allocate a dcache entry
1610 * @parent: parent of entry to allocate
1611 * @name: qstr of the name
1612 *
1613 * Allocates a dentry. It returns %NULL if there is insufficient memory
1614 * available. On a success the dentry is returned. The name passed in is
1615 * copied and the copy passed in may be reused after this call.
1616 */
1617 struct dentry *d_alloc(struct dentry * parent, const struct qstr *name)
1618 {
1619 struct dentry *dentry = __d_alloc(parent->d_sb, name);
1620 if (!dentry)
1621 return NULL;
1622
1623 spin_lock(&parent->d_lock);
1624 /*
1625 * don't need child lock because it is not subject
1626 * to concurrency here
1627 */
1628 __dget_dlock(parent);
1629 dentry->d_parent = parent;
1630 list_add(&dentry->d_child, &parent->d_subdirs);
1631 spin_unlock(&parent->d_lock);
1632
1633 return dentry;
1634 }
1635 EXPORT_SYMBOL(d_alloc);
1636
1637 /**
1638 * d_alloc_pseudo - allocate a dentry (for lookup-less filesystems)
1639 * @sb: the superblock
1640 * @name: qstr of the name
1641 *
1642 * For a filesystem that just pins its dentries in memory and never
1643 * performs lookups at all, return an unhashed IS_ROOT dentry.
1644 */
1645 struct dentry *d_alloc_pseudo(struct super_block *sb, const struct qstr *name)
1646 {
1647 return __d_alloc(sb, name);
1648 }
1649 EXPORT_SYMBOL(d_alloc_pseudo);
1650
1651 struct dentry *d_alloc_name(struct dentry *parent, const char *name)
1652 {
1653 struct qstr q;
1654
1655 q.name = name;
1656 q.len = strlen(name);
1657 q.hash = full_name_hash(q.name, q.len);
1658 return d_alloc(parent, &q);
1659 }
1660 EXPORT_SYMBOL(d_alloc_name);
1661
1662 void d_set_d_op(struct dentry *dentry, const struct dentry_operations *op)
1663 {
1664 WARN_ON_ONCE(dentry->d_op);
1665 WARN_ON_ONCE(dentry->d_flags & (DCACHE_OP_HASH |
1666 DCACHE_OP_COMPARE |
1667 DCACHE_OP_REVALIDATE |
1668 DCACHE_OP_WEAK_REVALIDATE |
1669 DCACHE_OP_DELETE |
1670 DCACHE_OP_SELECT_INODE |
1671 DCACHE_OP_REAL));
1672 dentry->d_op = op;
1673 if (!op)
1674 return;
1675 if (op->d_hash)
1676 dentry->d_flags |= DCACHE_OP_HASH;
1677 if (op->d_compare)
1678 dentry->d_flags |= DCACHE_OP_COMPARE;
1679 if (op->d_revalidate)
1680 dentry->d_flags |= DCACHE_OP_REVALIDATE;
1681 if (op->d_weak_revalidate)
1682 dentry->d_flags |= DCACHE_OP_WEAK_REVALIDATE;
1683 if (op->d_delete)
1684 dentry->d_flags |= DCACHE_OP_DELETE;
1685 if (op->d_prune)
1686 dentry->d_flags |= DCACHE_OP_PRUNE;
1687 if (op->d_select_inode)
1688 dentry->d_flags |= DCACHE_OP_SELECT_INODE;
1689 if (op->d_real)
1690 dentry->d_flags |= DCACHE_OP_REAL;
1691
1692 }
1693 EXPORT_SYMBOL(d_set_d_op);
1694
1695
1696 /*
1697 * d_set_fallthru - Mark a dentry as falling through to a lower layer
1698 * @dentry - The dentry to mark
1699 *
1700 * Mark a dentry as falling through to the lower layer (as set with
1701 * d_pin_lower()). This flag may be recorded on the medium.
1702 */
1703 void d_set_fallthru(struct dentry *dentry)
1704 {
1705 spin_lock(&dentry->d_lock);
1706 dentry->d_flags |= DCACHE_FALLTHRU;
1707 spin_unlock(&dentry->d_lock);
1708 }
1709 EXPORT_SYMBOL(d_set_fallthru);
1710
1711 static unsigned d_flags_for_inode(struct inode *inode)
1712 {
1713 unsigned add_flags = DCACHE_REGULAR_TYPE;
1714
1715 if (!inode)
1716 return DCACHE_MISS_TYPE;
1717
1718 if (S_ISDIR(inode->i_mode)) {
1719 add_flags = DCACHE_DIRECTORY_TYPE;
1720 if (unlikely(!(inode->i_opflags & IOP_LOOKUP))) {
1721 if (unlikely(!inode->i_op->lookup))
1722 add_flags = DCACHE_AUTODIR_TYPE;
1723 else
1724 inode->i_opflags |= IOP_LOOKUP;
1725 }
1726 goto type_determined;
1727 }
1728
1729 if (unlikely(!(inode->i_opflags & IOP_NOFOLLOW))) {
1730 if (unlikely(inode->i_op->get_link)) {
1731 add_flags = DCACHE_SYMLINK_TYPE;
1732 goto type_determined;
1733 }
1734 inode->i_opflags |= IOP_NOFOLLOW;
1735 }
1736
1737 if (unlikely(!S_ISREG(inode->i_mode)))
1738 add_flags = DCACHE_SPECIAL_TYPE;
1739
1740 type_determined:
1741 if (unlikely(IS_AUTOMOUNT(inode)))
1742 add_flags |= DCACHE_NEED_AUTOMOUNT;
1743 return add_flags;
1744 }
1745
1746 static void __d_instantiate(struct dentry *dentry, struct inode *inode)
1747 {
1748 unsigned add_flags = d_flags_for_inode(inode);
1749
1750 spin_lock(&dentry->d_lock);
1751 hlist_add_head(&dentry->d_u.d_alias, &inode->i_dentry);
1752 raw_write_seqcount_begin(&dentry->d_seq);
1753 __d_set_inode_and_type(dentry, inode, add_flags);
1754 raw_write_seqcount_end(&dentry->d_seq);
1755 __fsnotify_d_instantiate(dentry);
1756 spin_unlock(&dentry->d_lock);
1757 }
1758
1759 /**
1760 * d_instantiate - fill in inode information for a dentry
1761 * @entry: dentry to complete
1762 * @inode: inode to attach to this dentry
1763 *
1764 * Fill in inode information in the entry.
1765 *
1766 * This turns negative dentries into productive full members
1767 * of society.
1768 *
1769 * NOTE! This assumes that the inode count has been incremented
1770 * (or otherwise set) by the caller to indicate that it is now
1771 * in use by the dcache.
1772 */
1773
1774 void d_instantiate(struct dentry *entry, struct inode * inode)
1775 {
1776 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1777 if (inode) {
1778 spin_lock(&inode->i_lock);
1779 __d_instantiate(entry, inode);
1780 spin_unlock(&inode->i_lock);
1781 }
1782 security_d_instantiate(entry, inode);
1783 }
1784 EXPORT_SYMBOL(d_instantiate);
1785
1786 /**
1787 * d_instantiate_no_diralias - instantiate a non-aliased dentry
1788 * @entry: dentry to complete
1789 * @inode: inode to attach to this dentry
1790 *
1791 * Fill in inode information in the entry. If a directory alias is found, then
1792 * return an error (and drop inode). Together with d_materialise_unique() this
1793 * guarantees that a directory inode may never have more than one alias.
1794 */
1795 int d_instantiate_no_diralias(struct dentry *entry, struct inode *inode)
1796 {
1797 BUG_ON(!hlist_unhashed(&entry->d_u.d_alias));
1798
1799 spin_lock(&inode->i_lock);
1800 if (S_ISDIR(inode->i_mode) && !hlist_empty(&inode->i_dentry)) {
1801 spin_unlock(&inode->i_lock);
1802 iput(inode);
1803 return -EBUSY;
1804 }
1805 __d_instantiate(entry, inode);
1806 spin_unlock(&inode->i_lock);
1807 security_d_instantiate(entry, inode);
1808
1809 return 0;
1810 }
1811 EXPORT_SYMBOL(d_instantiate_no_diralias);
1812
1813 struct dentry *d_make_root(struct inode *root_inode)
1814 {
1815 struct dentry *res = NULL;
1816
1817 if (root_inode) {
1818 static const struct qstr name = QSTR_INIT("/", 1);
1819
1820 res = __d_alloc(root_inode->i_sb, &name);
1821 if (res)
1822 d_instantiate(res, root_inode);
1823 else
1824 iput(root_inode);
1825 }
1826 return res;
1827 }
1828 EXPORT_SYMBOL(d_make_root);
1829
1830 static struct dentry * __d_find_any_alias(struct inode *inode)
1831 {
1832 struct dentry *alias;
1833
1834 if (hlist_empty(&inode->i_dentry))
1835 return NULL;
1836 alias = hlist_entry(inode->i_dentry.first, struct dentry, d_u.d_alias);
1837 __dget(alias);
1838 return alias;
1839 }
1840
1841 /**
1842 * d_find_any_alias - find any alias for a given inode
1843 * @inode: inode to find an alias for
1844 *
1845 * If any aliases exist for the given inode, take and return a
1846 * reference for one of them. If no aliases exist, return %NULL.
1847 */
1848 struct dentry *d_find_any_alias(struct inode *inode)
1849 {
1850 struct dentry *de;
1851
1852 spin_lock(&inode->i_lock);
1853 de = __d_find_any_alias(inode);
1854 spin_unlock(&inode->i_lock);
1855 return de;
1856 }
1857 EXPORT_SYMBOL(d_find_any_alias);
1858
1859 static struct dentry *__d_obtain_alias(struct inode *inode, int disconnected)
1860 {
1861 static const struct qstr anonstring = QSTR_INIT("/", 1);
1862 struct dentry *tmp;
1863 struct dentry *res;
1864 unsigned add_flags;
1865
1866 if (!inode)
1867 return ERR_PTR(-ESTALE);
1868 if (IS_ERR(inode))
1869 return ERR_CAST(inode);
1870
1871 res = d_find_any_alias(inode);
1872 if (res)
1873 goto out_iput;
1874
1875 tmp = __d_alloc(inode->i_sb, &anonstring);
1876 if (!tmp) {
1877 res = ERR_PTR(-ENOMEM);
1878 goto out_iput;
1879 }
1880
1881 spin_lock(&inode->i_lock);
1882 res = __d_find_any_alias(inode);
1883 if (res) {
1884 spin_unlock(&inode->i_lock);
1885 dput(tmp);
1886 goto out_iput;
1887 }
1888
1889 /* attach a disconnected dentry */
1890 add_flags = d_flags_for_inode(inode);
1891
1892 if (disconnected)
1893 add_flags |= DCACHE_DISCONNECTED;
1894
1895 spin_lock(&tmp->d_lock);
1896 __d_set_inode_and_type(tmp, inode, add_flags);
1897 hlist_add_head(&tmp->d_u.d_alias, &inode->i_dentry);
1898 hlist_bl_lock(&tmp->d_sb->s_anon);
1899 hlist_bl_add_head(&tmp->d_hash, &tmp->d_sb->s_anon);
1900 hlist_bl_unlock(&tmp->d_sb->s_anon);
1901 spin_unlock(&tmp->d_lock);
1902 spin_unlock(&inode->i_lock);
1903 security_d_instantiate(tmp, inode);
1904
1905 return tmp;
1906
1907 out_iput:
1908 if (res && !IS_ERR(res))
1909 security_d_instantiate(res, inode);
1910 iput(inode);
1911 return res;
1912 }
1913
1914 /**
1915 * d_obtain_alias - find or allocate a DISCONNECTED dentry for a given inode
1916 * @inode: inode to allocate the dentry for
1917 *
1918 * Obtain a dentry for an inode resulting from NFS filehandle conversion or
1919 * similar open by handle operations. The returned dentry may be anonymous,
1920 * or may have a full name (if the inode was already in the cache).
1921 *
1922 * When called on a directory inode, we must ensure that the inode only ever
1923 * has one dentry. If a dentry is found, that is returned instead of
1924 * allocating a new one.
1925 *
1926 * On successful return, the reference to the inode has been transferred
1927 * to the dentry. In case of an error the reference on the inode is released.
1928 * To make it easier to use in export operations a %NULL or IS_ERR inode may
1929 * be passed in and the error will be propagated to the return value,
1930 * with a %NULL @inode replaced by ERR_PTR(-ESTALE).
1931 */
1932 struct dentry *d_obtain_alias(struct inode *inode)
1933 {
1934 return __d_obtain_alias(inode, 1);
1935 }
1936 EXPORT_SYMBOL(d_obtain_alias);
1937
1938 /**
1939 * d_obtain_root - find or allocate a dentry for a given inode
1940 * @inode: inode to allocate the dentry for
1941 *
1942 * Obtain an IS_ROOT dentry for the root of a filesystem.
1943 *
1944 * We must ensure that directory inodes only ever have one dentry. If a
1945 * dentry is found, that is returned instead of allocating a new one.
1946 *
1947 * On successful return, the reference to the inode has been transferred
1948 * to the dentry. In case of an error the reference on the inode is
1949 * released. A %NULL or IS_ERR inode may be passed in and will be the
1950 * error will be propagate to the return value, with a %NULL @inode
1951 * replaced by ERR_PTR(-ESTALE).
1952 */
1953 struct dentry *d_obtain_root(struct inode *inode)
1954 {
1955 return __d_obtain_alias(inode, 0);
1956 }
1957 EXPORT_SYMBOL(d_obtain_root);
1958
1959 /**
1960 * d_add_ci - lookup or allocate new dentry with case-exact name
1961 * @inode: the inode case-insensitive lookup has found
1962 * @dentry: the negative dentry that was passed to the parent's lookup func
1963 * @name: the case-exact name to be associated with the returned dentry
1964 *
1965 * This is to avoid filling the dcache with case-insensitive names to the
1966 * same inode, only the actual correct case is stored in the dcache for
1967 * case-insensitive filesystems.
1968 *
1969 * For a case-insensitive lookup match and if the the case-exact dentry
1970 * already exists in in the dcache, use it and return it.
1971 *
1972 * If no entry exists with the exact case name, allocate new dentry with
1973 * the exact case, and return the spliced entry.
1974 */
1975 struct dentry *d_add_ci(struct dentry *dentry, struct inode *inode,
1976 struct qstr *name)
1977 {
1978 struct dentry *found;
1979 struct dentry *new;
1980
1981 /*
1982 * First check if a dentry matching the name already exists,
1983 * if not go ahead and create it now.
1984 */
1985 found = d_hash_and_lookup(dentry->d_parent, name);
1986 if (!found) {
1987 new = d_alloc(dentry->d_parent, name);
1988 if (!new) {
1989 found = ERR_PTR(-ENOMEM);
1990 } else {
1991 found = d_splice_alias(inode, new);
1992 if (found) {
1993 dput(new);
1994 return found;
1995 }
1996 return new;
1997 }
1998 }
1999 iput(inode);
2000 return found;
2001 }
2002 EXPORT_SYMBOL(d_add_ci);
2003
2004 /*
2005 * Do the slow-case of the dentry name compare.
2006 *
2007 * Unlike the dentry_cmp() function, we need to atomically
2008 * load the name and length information, so that the
2009 * filesystem can rely on them, and can use the 'name' and
2010 * 'len' information without worrying about walking off the
2011 * end of memory etc.
2012 *
2013 * Thus the read_seqcount_retry() and the "duplicate" info
2014 * in arguments (the low-level filesystem should not look
2015 * at the dentry inode or name contents directly, since
2016 * rename can change them while we're in RCU mode).
2017 */
2018 enum slow_d_compare {
2019 D_COMP_OK,
2020 D_COMP_NOMATCH,
2021 D_COMP_SEQRETRY,
2022 };
2023
2024 static noinline enum slow_d_compare slow_dentry_cmp(
2025 const struct dentry *parent,
2026 struct dentry *dentry,
2027 unsigned int seq,
2028 const struct qstr *name)
2029 {
2030 int tlen = dentry->d_name.len;
2031 const char *tname = dentry->d_name.name;
2032
2033 if (read_seqcount_retry(&dentry->d_seq, seq)) {
2034 cpu_relax();
2035 return D_COMP_SEQRETRY;
2036 }
2037 if (parent->d_op->d_compare(parent, dentry, tlen, tname, name))
2038 return D_COMP_NOMATCH;
2039 return D_COMP_OK;
2040 }
2041
2042 /**
2043 * __d_lookup_rcu - search for a dentry (racy, store-free)
2044 * @parent: parent dentry
2045 * @name: qstr of name we wish to find
2046 * @seqp: returns d_seq value at the point where the dentry was found
2047 * Returns: dentry, or NULL
2048 *
2049 * __d_lookup_rcu is the dcache lookup function for rcu-walk name
2050 * resolution (store-free path walking) design described in
2051 * Documentation/filesystems/path-lookup.txt.
2052 *
2053 * This is not to be used outside core vfs.
2054 *
2055 * __d_lookup_rcu must only be used in rcu-walk mode, ie. with vfsmount lock
2056 * held, and rcu_read_lock held. The returned dentry must not be stored into
2057 * without taking d_lock and checking d_seq sequence count against @seq
2058 * returned here.
2059 *
2060 * A refcount may be taken on the found dentry with the d_rcu_to_refcount
2061 * function.
2062 *
2063 * Alternatively, __d_lookup_rcu may be called again to look up the child of
2064 * the returned dentry, so long as its parent's seqlock is checked after the
2065 * child is looked up. Thus, an interlocking stepping of sequence lock checks
2066 * is formed, giving integrity down the path walk.
2067 *
2068 * NOTE! The caller *has* to check the resulting dentry against the sequence
2069 * number we've returned before using any of the resulting dentry state!
2070 */
2071 struct dentry *__d_lookup_rcu(const struct dentry *parent,
2072 const struct qstr *name,
2073 unsigned *seqp)
2074 {
2075 u64 hashlen = name->hash_len;
2076 const unsigned char *str = name->name;
2077 struct hlist_bl_head *b = d_hash(parent, hashlen_hash(hashlen));
2078 struct hlist_bl_node *node;
2079 struct dentry *dentry;
2080
2081 /*
2082 * Note: There is significant duplication with __d_lookup_rcu which is
2083 * required to prevent single threaded performance regressions
2084 * especially on architectures where smp_rmb (in seqcounts) are costly.
2085 * Keep the two functions in sync.
2086 */
2087
2088 /*
2089 * The hash list is protected using RCU.
2090 *
2091 * Carefully use d_seq when comparing a candidate dentry, to avoid
2092 * races with d_move().
2093 *
2094 * It is possible that concurrent renames can mess up our list
2095 * walk here and result in missing our dentry, resulting in the
2096 * false-negative result. d_lookup() protects against concurrent
2097 * renames using rename_lock seqlock.
2098 *
2099 * See Documentation/filesystems/path-lookup.txt for more details.
2100 */
2101 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2102 unsigned seq;
2103
2104 seqretry:
2105 /*
2106 * The dentry sequence count protects us from concurrent
2107 * renames, and thus protects parent and name fields.
2108 *
2109 * The caller must perform a seqcount check in order
2110 * to do anything useful with the returned dentry.
2111 *
2112 * NOTE! We do a "raw" seqcount_begin here. That means that
2113 * we don't wait for the sequence count to stabilize if it
2114 * is in the middle of a sequence change. If we do the slow
2115 * dentry compare, we will do seqretries until it is stable,
2116 * and if we end up with a successful lookup, we actually
2117 * want to exit RCU lookup anyway.
2118 */
2119 seq = raw_seqcount_begin(&dentry->d_seq);
2120 if (dentry->d_parent != parent)
2121 continue;
2122 if (d_unhashed(dentry))
2123 continue;
2124
2125 if (unlikely(parent->d_flags & DCACHE_OP_COMPARE)) {
2126 if (dentry->d_name.hash != hashlen_hash(hashlen))
2127 continue;
2128 *seqp = seq;
2129 switch (slow_dentry_cmp(parent, dentry, seq, name)) {
2130 case D_COMP_OK:
2131 return dentry;
2132 case D_COMP_NOMATCH:
2133 continue;
2134 default:
2135 goto seqretry;
2136 }
2137 }
2138
2139 if (dentry->d_name.hash_len != hashlen)
2140 continue;
2141 *seqp = seq;
2142 if (!dentry_cmp(dentry, str, hashlen_len(hashlen)))
2143 return dentry;
2144 }
2145 return NULL;
2146 }
2147
2148 /**
2149 * d_lookup - search for a dentry
2150 * @parent: parent dentry
2151 * @name: qstr of name we wish to find
2152 * Returns: dentry, or NULL
2153 *
2154 * d_lookup searches the children of the parent dentry for the name in
2155 * question. If the dentry is found its reference count is incremented and the
2156 * dentry is returned. The caller must use dput to free the entry when it has
2157 * finished using it. %NULL is returned if the dentry does not exist.
2158 */
2159 struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
2160 {
2161 struct dentry *dentry;
2162 unsigned seq;
2163
2164 do {
2165 seq = read_seqbegin(&rename_lock);
2166 dentry = __d_lookup(parent, name);
2167 if (dentry)
2168 break;
2169 } while (read_seqretry(&rename_lock, seq));
2170 return dentry;
2171 }
2172 EXPORT_SYMBOL(d_lookup);
2173
2174 /**
2175 * __d_lookup - search for a dentry (racy)
2176 * @parent: parent dentry
2177 * @name: qstr of name we wish to find
2178 * Returns: dentry, or NULL
2179 *
2180 * __d_lookup is like d_lookup, however it may (rarely) return a
2181 * false-negative result due to unrelated rename activity.
2182 *
2183 * __d_lookup is slightly faster by avoiding rename_lock read seqlock,
2184 * however it must be used carefully, eg. with a following d_lookup in
2185 * the case of failure.
2186 *
2187 * __d_lookup callers must be commented.
2188 */
2189 struct dentry *__d_lookup(const struct dentry *parent, const struct qstr *name)
2190 {
2191 unsigned int len = name->len;
2192 unsigned int hash = name->hash;
2193 const unsigned char *str = name->name;
2194 struct hlist_bl_head *b = d_hash(parent, hash);
2195 struct hlist_bl_node *node;
2196 struct dentry *found = NULL;
2197 struct dentry *dentry;
2198
2199 /*
2200 * Note: There is significant duplication with __d_lookup_rcu which is
2201 * required to prevent single threaded performance regressions
2202 * especially on architectures where smp_rmb (in seqcounts) are costly.
2203 * Keep the two functions in sync.
2204 */
2205
2206 /*
2207 * The hash list is protected using RCU.
2208 *
2209 * Take d_lock when comparing a candidate dentry, to avoid races
2210 * with d_move().
2211 *
2212 * It is possible that concurrent renames can mess up our list
2213 * walk here and result in missing our dentry, resulting in the
2214 * false-negative result. d_lookup() protects against concurrent
2215 * renames using rename_lock seqlock.
2216 *
2217 * See Documentation/filesystems/path-lookup.txt for more details.
2218 */
2219 rcu_read_lock();
2220
2221 hlist_bl_for_each_entry_rcu(dentry, node, b, d_hash) {
2222
2223 if (dentry->d_name.hash != hash)
2224 continue;
2225
2226 spin_lock(&dentry->d_lock);
2227 if (dentry->d_parent != parent)
2228 goto next;
2229 if (d_unhashed(dentry))
2230 goto next;
2231
2232 /*
2233 * It is safe to compare names since d_move() cannot
2234 * change the qstr (protected by d_lock).
2235 */
2236 if (parent->d_flags & DCACHE_OP_COMPARE) {
2237 int tlen = dentry->d_name.len;
2238 const char *tname = dentry->d_name.name;
2239 if (parent->d_op->d_compare(parent, dentry, tlen, tname, name))
2240 goto next;
2241 } else {
2242 if (dentry->d_name.len != len)
2243 goto next;
2244 if (dentry_cmp(dentry, str, len))
2245 goto next;
2246 }
2247
2248 dentry->d_lockref.count++;
2249 found = dentry;
2250 spin_unlock(&dentry->d_lock);
2251 break;
2252 next:
2253 spin_unlock(&dentry->d_lock);
2254 }
2255 rcu_read_unlock();
2256
2257 return found;
2258 }
2259
2260 /**
2261 * d_hash_and_lookup - hash the qstr then search for a dentry
2262 * @dir: Directory to search in
2263 * @name: qstr of name we wish to find
2264 *
2265 * On lookup failure NULL is returned; on bad name - ERR_PTR(-error)
2266 */
2267 struct dentry *d_hash_and_lookup(struct dentry *dir, struct qstr *name)
2268 {
2269 /*
2270 * Check for a fs-specific hash function. Note that we must
2271 * calculate the standard hash first, as the d_op->d_hash()
2272 * routine may choose to leave the hash value unchanged.
2273 */
2274 name->hash = full_name_hash(name->name, name->len);
2275 if (dir->d_flags & DCACHE_OP_HASH) {
2276 int err = dir->d_op->d_hash(dir, name);
2277 if (unlikely(err < 0))
2278 return ERR_PTR(err);
2279 }
2280 return d_lookup(dir, name);
2281 }
2282 EXPORT_SYMBOL(d_hash_and_lookup);
2283
2284 /*
2285 * When a file is deleted, we have two options:
2286 * - turn this dentry into a negative dentry
2287 * - unhash this dentry and free it.
2288 *
2289 * Usually, we want to just turn this into
2290 * a negative dentry, but if anybody else is
2291 * currently using the dentry or the inode
2292 * we can't do that and we fall back on removing
2293 * it from the hash queues and waiting for
2294 * it to be deleted later when it has no users
2295 */
2296
2297 /**
2298 * d_delete - delete a dentry
2299 * @dentry: The dentry to delete
2300 *
2301 * Turn the dentry into a negative dentry if possible, otherwise
2302 * remove it from the hash queues so it can be deleted later
2303 */
2304
2305 void d_delete(struct dentry * dentry)
2306 {
2307 struct inode *inode;
2308 int isdir = 0;
2309 /*
2310 * Are we the only user?
2311 */
2312 again:
2313 spin_lock(&dentry->d_lock);
2314 inode = dentry->d_inode;
2315 isdir = S_ISDIR(inode->i_mode);
2316 if (dentry->d_lockref.count == 1) {
2317 if (!spin_trylock(&inode->i_lock)) {
2318 spin_unlock(&dentry->d_lock);
2319 cpu_relax();
2320 goto again;
2321 }
2322 dentry->d_flags &= ~DCACHE_CANT_MOUNT;
2323 dentry_unlink_inode(dentry);
2324 fsnotify_nameremove(dentry, isdir);
2325 return;
2326 }
2327
2328 if (!d_unhashed(dentry))
2329 __d_drop(dentry);
2330
2331 spin_unlock(&dentry->d_lock);
2332
2333 fsnotify_nameremove(dentry, isdir);
2334 }
2335 EXPORT_SYMBOL(d_delete);
2336
2337 static void __d_rehash(struct dentry * entry, struct hlist_bl_head *b)
2338 {
2339 BUG_ON(!d_unhashed(entry));
2340 hlist_bl_lock(b);
2341 entry->d_flags |= DCACHE_RCUACCESS;
2342 hlist_bl_add_head_rcu(&entry->d_hash, b);
2343 hlist_bl_unlock(b);
2344 }
2345
2346 static void _d_rehash(struct dentry * entry)
2347 {
2348 __d_rehash(entry, d_hash(entry->d_parent, entry->d_name.hash));
2349 }
2350
2351 /**
2352 * d_rehash - add an entry back to the hash
2353 * @entry: dentry to add to the hash
2354 *
2355 * Adds a dentry to the hash according to its name.
2356 */
2357
2358 void d_rehash(struct dentry * entry)
2359 {
2360 spin_lock(&entry->d_lock);
2361 _d_rehash(entry);
2362 spin_unlock(&entry->d_lock);
2363 }
2364 EXPORT_SYMBOL(d_rehash);
2365
2366
2367 /* inode->i_lock held if inode is non-NULL */
2368
2369 static inline void __d_add(struct dentry *dentry, struct inode *inode)
2370 {
2371 if (inode) {
2372 __d_instantiate(dentry, inode);
2373 spin_unlock(&inode->i_lock);
2374 }
2375 security_d_instantiate(dentry, inode);
2376 d_rehash(dentry);
2377 }
2378
2379 /**
2380 * d_add - add dentry to hash queues
2381 * @entry: dentry to add
2382 * @inode: The inode to attach to this dentry
2383 *
2384 * This adds the entry to the hash queues and initializes @inode.
2385 * The entry was actually filled in earlier during d_alloc().
2386 */
2387
2388 void d_add(struct dentry *entry, struct inode *inode)
2389 {
2390 if (inode)
2391 spin_lock(&inode->i_lock);
2392 __d_add(entry, inode);
2393 }
2394 EXPORT_SYMBOL(d_add);
2395
2396 /**
2397 * d_exact_alias - find and hash an exact unhashed alias
2398 * @entry: dentry to add
2399 * @inode: The inode to go with this dentry
2400 *
2401 * If an unhashed dentry with the same name/parent and desired
2402 * inode already exists, hash and return it. Otherwise, return
2403 * NULL.
2404 *
2405 * Parent directory should be locked.
2406 */
2407 struct dentry *d_exact_alias(struct dentry *entry, struct inode *inode)
2408 {
2409 struct dentry *alias;
2410 int len = entry->d_name.len;
2411 const char *name = entry->d_name.name;
2412 unsigned int hash = entry->d_name.hash;
2413
2414 spin_lock(&inode->i_lock);
2415 hlist_for_each_entry(alias, &inode->i_dentry, d_u.d_alias) {
2416 /*
2417 * Don't need alias->d_lock here, because aliases with
2418 * d_parent == entry->d_parent are not subject to name or
2419 * parent changes, because the parent inode i_mutex is held.
2420 */
2421 if (alias->d_name.hash != hash)
2422 continue;
2423 if (alias->d_parent != entry->d_parent)
2424 continue;
2425 if (alias->d_name.len != len)
2426 continue;
2427 if (dentry_cmp(alias, name, len))
2428 continue;
2429 spin_lock(&alias->d_lock);
2430 if (!d_unhashed(alias)) {
2431 spin_unlock(&alias->d_lock);
2432 alias = NULL;
2433 } else {
2434 __dget_dlock(alias);
2435 _d_rehash(alias);
2436 spin_unlock(&alias->d_lock);
2437 }
2438 spin_unlock(&inode->i_lock);
2439 return alias;
2440 }
2441 spin_unlock(&inode->i_lock);
2442 return NULL;
2443 }
2444 EXPORT_SYMBOL(d_exact_alias);
2445
2446 /**
2447 * dentry_update_name_case - update case insensitive dentry with a new name
2448 * @dentry: dentry to be updated
2449 * @name: new name
2450 *
2451 * Update a case insensitive dentry with new case of name.
2452 *
2453 * dentry must have been returned by d_lookup with name @name. Old and new
2454 * name lengths must match (ie. no d_compare which allows mismatched name
2455 * lengths).
2456 *
2457 * Parent inode i_mutex must be held over d_lookup and into this call (to
2458 * keep renames and concurrent inserts, and readdir(2) away).
2459 */
2460 void dentry_update_name_case(struct dentry *dentry, struct qstr *name)
2461 {
2462 BUG_ON(!inode_is_locked(dentry->d_parent->d_inode));
2463 BUG_ON(dentry->d_name.len != name->len); /* d_lookup gives this */
2464
2465 spin_lock(&dentry->d_lock);
2466 write_seqcount_begin(&dentry->d_seq);
2467 memcpy((unsigned char *)dentry->d_name.name, name->name, name->len);
2468 write_seqcount_end(&dentry->d_seq);
2469 spin_unlock(&dentry->d_lock);
2470 }
2471 EXPORT_SYMBOL(dentry_update_name_case);
2472
2473 static void swap_names(struct dentry *dentry, struct dentry *target)
2474 {
2475 if (unlikely(dname_external(target))) {
2476 if (unlikely(dname_external(dentry))) {
2477 /*
2478 * Both external: swap the pointers
2479 */
2480 swap(target->d_name.name, dentry->d_name.name);
2481 } else {
2482 /*
2483 * dentry:internal, target:external. Steal target's
2484 * storage and make target internal.
2485 */
2486 memcpy(target->d_iname, dentry->d_name.name,
2487 dentry->d_name.len + 1);
2488 dentry->d_name.name = target->d_name.name;
2489 target->d_name.name = target->d_iname;
2490 }
2491 } else {
2492 if (unlikely(dname_external(dentry))) {
2493 /*
2494 * dentry:external, target:internal. Give dentry's
2495 * storage to target and make dentry internal
2496 */
2497 memcpy(dentry->d_iname, target->d_name.name,
2498 target->d_name.len + 1);
2499 target->d_name.name = dentry->d_name.name;
2500 dentry->d_name.name = dentry->d_iname;
2501 } else {
2502 /*
2503 * Both are internal.
2504 */
2505 unsigned int i;
2506 BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
2507 kmemcheck_mark_initialized(dentry->d_iname, DNAME_INLINE_LEN);
2508 kmemcheck_mark_initialized(target->d_iname, DNAME_INLINE_LEN);
2509 for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
2510 swap(((long *) &dentry->d_iname)[i],
2511 ((long *) &target->d_iname)[i]);
2512 }
2513 }
2514 }
2515 swap(dentry->d_name.hash_len, target->d_name.hash_len);
2516 }
2517
2518 static void copy_name(struct dentry *dentry, struct dentry *target)
2519 {
2520 struct external_name *old_name = NULL;
2521 if (unlikely(dname_external(dentry)))
2522 old_name = external_name(dentry);
2523 if (unlikely(dname_external(target))) {
2524 atomic_inc(&external_name(target)->u.count);
2525 dentry->d_name = target->d_name;
2526 } else {
2527 memcpy(dentry->d_iname, target->d_name.name,
2528 target->d_name.len + 1);
2529 dentry->d_name.name = dentry->d_iname;
2530 dentry->d_name.hash_len = target->d_name.hash_len;
2531 }
2532 if (old_name && likely(atomic_dec_and_test(&old_name->u.count)))
2533 kfree_rcu(old_name, u.head);
2534 }
2535
2536 static void dentry_lock_for_move(struct dentry *dentry, struct dentry *target)
2537 {
2538 /*
2539 * XXXX: do we really need to take target->d_lock?
2540 */
2541 if (IS_ROOT(dentry) || dentry->d_parent == target->d_parent)
2542 spin_lock(&target->d_parent->d_lock);
2543 else {
2544 if (d_ancestor(dentry->d_parent, target->d_parent)) {
2545 spin_lock(&dentry->d_parent->d_lock);
2546 spin_lock_nested(&target->d_parent->d_lock,
2547 DENTRY_D_LOCK_NESTED);
2548 } else {
2549 spin_lock(&target->d_parent->d_lock);
2550 spin_lock_nested(&dentry->d_parent->d_lock,
2551 DENTRY_D_LOCK_NESTED);
2552 }
2553 }
2554 if (target < dentry) {
2555 spin_lock_nested(&target->d_lock, 2);
2556 spin_lock_nested(&dentry->d_lock, 3);
2557 } else {
2558 spin_lock_nested(&dentry->d_lock, 2);
2559 spin_lock_nested(&target->d_lock, 3);
2560 }
2561 }
2562
2563 static void dentry_unlock_for_move(struct dentry *dentry, struct dentry *target)
2564 {
2565 if (target->d_parent != dentry->d_parent)
2566 spin_unlock(&dentry->d_parent->d_lock);
2567 if (target->d_parent != target)
2568 spin_unlock(&target->d_parent->d_lock);
2569 spin_unlock(&target->d_lock);
2570 spin_unlock(&dentry->d_lock);
2571 }
2572
2573 /*
2574 * When switching names, the actual string doesn't strictly have to
2575 * be preserved in the target - because we're dropping the target
2576 * anyway. As such, we can just do a simple memcpy() to copy over
2577 * the new name before we switch, unless we are going to rehash
2578 * it. Note that if we *do* unhash the target, we are not allowed
2579 * to rehash it without giving it a new name/hash key - whether
2580 * we swap or overwrite the names here, resulting name won't match
2581 * the reality in filesystem; it's only there for d_path() purposes.
2582 * Note that all of this is happening under rename_lock, so the
2583 * any hash lookup seeing it in the middle of manipulations will
2584 * be discarded anyway. So we do not care what happens to the hash
2585 * key in that case.
2586 */
2587 /*
2588 * __d_move - move a dentry
2589 * @dentry: entry to move
2590 * @target: new dentry
2591 * @exchange: exchange the two dentries
2592 *
2593 * Update the dcache to reflect the move of a file name. Negative
2594 * dcache entries should not be moved in this way. Caller must hold
2595 * rename_lock, the i_mutex of the source and target directories,
2596 * and the sb->s_vfs_rename_mutex if they differ. See lock_rename().
2597 */
2598 static void __d_move(struct dentry *dentry, struct dentry *target,
2599 bool exchange)
2600 {
2601 if (!dentry->d_inode)
2602 printk(KERN_WARNING "VFS: moving negative dcache entry\n");
2603
2604 BUG_ON(d_ancestor(dentry, target));
2605 BUG_ON(d_ancestor(target, dentry));
2606
2607 dentry_lock_for_move(dentry, target);
2608
2609 write_seqcount_begin(&dentry->d_seq);
2610 write_seqcount_begin_nested(&target->d_seq, DENTRY_D_LOCK_NESTED);
2611
2612 /* __d_drop does write_seqcount_barrier, but they're OK to nest. */
2613
2614 /*
2615 * Move the dentry to the target hash queue. Don't bother checking
2616 * for the same hash queue because of how unlikely it is.
2617 */
2618 __d_drop(dentry);
2619 __d_rehash(dentry, d_hash(target->d_parent, target->d_name.hash));
2620
2621 /*
2622 * Unhash the target (d_delete() is not usable here). If exchanging
2623 * the two dentries, then rehash onto the other's hash queue.
2624 */
2625 __d_drop(target);
2626 if (exchange) {
2627 __d_rehash(target,
2628 d_hash(dentry->d_parent, dentry->d_name.hash));
2629 }
2630
2631 /* Switch the names.. */
2632 if (exchange)
2633 swap_names(dentry, target);
2634 else
2635 copy_name(dentry, target);
2636
2637 /* ... and switch them in the tree */
2638 if (IS_ROOT(dentry)) {
2639 /* splicing a tree */
2640 dentry->d_parent = target->d_parent;
2641 target->d_parent = target;
2642 list_del_init(&target->d_child);
2643 list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2644 } else {
2645 /* swapping two dentries */
2646 swap(dentry->d_parent, target->d_parent);
2647 list_move(&target->d_child, &target->d_parent->d_subdirs);
2648 list_move(&dentry->d_child, &dentry->d_parent->d_subdirs);
2649 if (exchange)
2650 fsnotify_d_move(target);
2651 fsnotify_d_move(dentry);
2652 }
2653
2654 write_seqcount_end(&target->d_seq);
2655 write_seqcount_end(&dentry->d_seq);
2656
2657 dentry_unlock_for_move(dentry, target);
2658 }
2659
2660 /*
2661 * d_move - move a dentry
2662 * @dentry: entry to move
2663 * @target: new dentry
2664 *
2665 * Update the dcache to reflect the move of a file name. Negative
2666 * dcache entries should not be moved in this way. See the locking
2667 * requirements for __d_move.
2668 */
2669 void d_move(struct dentry *dentry, struct dentry *target)
2670 {
2671 write_seqlock(&rename_lock);
2672 __d_move(dentry, target, false);
2673 write_sequnlock(&rename_lock);
2674 }
2675 EXPORT_SYMBOL(d_move);
2676
2677 /*
2678 * d_exchange - exchange two dentries
2679 * @dentry1: first dentry
2680 * @dentry2: second dentry
2681 */
2682 void d_exchange(struct dentry *dentry1, struct dentry *dentry2)
2683 {
2684 write_seqlock(&rename_lock);
2685
2686 WARN_ON(!dentry1->d_inode);
2687 WARN_ON(!dentry2->d_inode);
2688 WARN_ON(IS_ROOT(dentry1));
2689 WARN_ON(IS_ROOT(dentry2));
2690
2691 __d_move(dentry1, dentry2, true);
2692
2693 write_sequnlock(&rename_lock);
2694 }
2695
2696 /**
2697 * d_ancestor - search for an ancestor
2698 * @p1: ancestor dentry
2699 * @p2: child dentry
2700 *
2701 * Returns the ancestor dentry of p2 which is a child of p1, if p1 is
2702 * an ancestor of p2, else NULL.
2703 */
2704 struct dentry *d_ancestor(struct dentry *p1, struct dentry *p2)
2705 {
2706 struct dentry *p;
2707
2708 for (p = p2; !IS_ROOT(p); p = p->d_parent) {
2709 if (p->d_parent == p1)
2710 return p;
2711 }
2712 return NULL;
2713 }
2714
2715 /*
2716 * This helper attempts to cope with remotely renamed directories
2717 *
2718 * It assumes that the caller is already holding
2719 * dentry->d_parent->d_inode->i_mutex, and rename_lock
2720 *
2721 * Note: If ever the locking in lock_rename() changes, then please
2722 * remember to update this too...
2723 */
2724 static int __d_unalias(struct inode *inode,
2725 struct dentry *dentry, struct dentry *alias)
2726 {
2727 struct mutex *m1 = NULL, *m2 = NULL;
2728 int ret = -ESTALE;
2729
2730 /* If alias and dentry share a parent, then no extra locks required */
2731 if (alias->d_parent == dentry->d_parent)
2732 goto out_unalias;
2733
2734 /* See lock_rename() */
2735 if (!mutex_trylock(&dentry->d_sb->s_vfs_rename_mutex))
2736 goto out_err;
2737 m1 = &dentry->d_sb->s_vfs_rename_mutex;
2738 if (!inode_trylock(alias->d_parent->d_inode))
2739 goto out_err;
2740 m2 = &alias->d_parent->d_inode->i_mutex;
2741 out_unalias:
2742 __d_move(alias, dentry, false);
2743 ret = 0;
2744 out_err:
2745 if (m2)
2746 mutex_unlock(m2);
2747 if (m1)
2748 mutex_unlock(m1);
2749 return ret;
2750 }
2751
2752 /**
2753 * d_splice_alias - splice a disconnected dentry into the tree if one exists
2754 * @inode: the inode which may have a disconnected dentry
2755 * @dentry: a negative dentry which we want to point to the inode.
2756 *
2757 * If inode is a directory and has an IS_ROOT alias, then d_move that in
2758 * place of the given dentry and return it, else simply d_add the inode
2759 * to the dentry and return NULL.
2760 *
2761 * If a non-IS_ROOT directory is found, the filesystem is corrupt, and
2762 * we should error out: directories can't have multiple aliases.
2763 *
2764 * This is needed in the lookup routine of any filesystem that is exportable
2765 * (via knfsd) so that we can build dcache paths to directories effectively.
2766 *
2767 * If a dentry was found and moved, then it is returned. Otherwise NULL
2768 * is returned. This matches the expected return value of ->lookup.
2769 *
2770 * Cluster filesystems may call this function with a negative, hashed dentry.
2771 * In that case, we know that the inode will be a regular file, and also this
2772 * will only occur during atomic_open. So we need to check for the dentry
2773 * being already hashed only in the final case.
2774 */
2775 struct dentry *d_splice_alias(struct inode *inode, struct dentry *dentry)
2776 {
2777 if (IS_ERR(inode))
2778 return ERR_CAST(inode);
2779
2780 BUG_ON(!d_unhashed(dentry));
2781
2782 if (!inode)
2783 goto out;
2784
2785 spin_lock(&inode->i_lock);
2786 if (S_ISDIR(inode->i_mode)) {
2787 struct dentry *new = __d_find_any_alias(inode);
2788 if (unlikely(new)) {
2789 /* The reference to new ensures it remains an alias */
2790 spin_unlock(&inode->i_lock);
2791 write_seqlock(&rename_lock);
2792 if (unlikely(d_ancestor(new, dentry))) {
2793 write_sequnlock(&rename_lock);
2794 dput(new);
2795 new = ERR_PTR(-ELOOP);
2796 pr_warn_ratelimited(
2797 "VFS: Lookup of '%s' in %s %s"
2798 " would have caused loop\n",
2799 dentry->d_name.name,
2800 inode->i_sb->s_type->name,
2801 inode->i_sb->s_id);
2802 } else if (!IS_ROOT(new)) {
2803 int err = __d_unalias(inode, dentry, new);
2804 write_sequnlock(&rename_lock);
2805 if (err) {
2806 dput(new);
2807 new = ERR_PTR(err);
2808 }
2809 } else {
2810 __d_move(new, dentry, false);
2811 write_sequnlock(&rename_lock);
2812 security_d_instantiate(new, inode);
2813 }
2814 iput(inode);
2815 return new;
2816 }
2817 }
2818 out:
2819 __d_add(dentry, inode);
2820 return NULL;
2821 }
2822 EXPORT_SYMBOL(d_splice_alias);
2823
2824 static int prepend(char **buffer, int *buflen, const char *str, int namelen)
2825 {
2826 *buflen -= namelen;
2827 if (*buflen < 0)
2828 return -ENAMETOOLONG;
2829 *buffer -= namelen;
2830 memcpy(*buffer, str, namelen);
2831 return 0;
2832 }
2833
2834 /**
2835 * prepend_name - prepend a pathname in front of current buffer pointer
2836 * @buffer: buffer pointer
2837 * @buflen: allocated length of the buffer
2838 * @name: name string and length qstr structure
2839 *
2840 * With RCU path tracing, it may race with d_move(). Use ACCESS_ONCE() to
2841 * make sure that either the old or the new name pointer and length are
2842 * fetched. However, there may be mismatch between length and pointer.
2843 * The length cannot be trusted, we need to copy it byte-by-byte until
2844 * the length is reached or a null byte is found. It also prepends "/" at
2845 * the beginning of the name. The sequence number check at the caller will
2846 * retry it again when a d_move() does happen. So any garbage in the buffer
2847 * due to mismatched pointer and length will be discarded.
2848 *
2849 * Data dependency barrier is needed to make sure that we see that terminating
2850 * NUL. Alpha strikes again, film at 11...
2851 */
2852 static int prepend_name(char **buffer, int *buflen, struct qstr *name)
2853 {
2854 const char *dname = ACCESS_ONCE(name->name);
2855 u32 dlen = ACCESS_ONCE(name->len);
2856 char *p;
2857
2858 smp_read_barrier_depends();
2859
2860 *buflen -= dlen + 1;
2861 if (*buflen < 0)
2862 return -ENAMETOOLONG;
2863 p = *buffer -= dlen + 1;
2864 *p++ = '/';
2865 while (dlen--) {
2866 char c = *dname++;
2867 if (!c)
2868 break;
2869 *p++ = c;
2870 }
2871 return 0;
2872 }
2873
2874 /**
2875 * prepend_path - Prepend path string to a buffer
2876 * @path: the dentry/vfsmount to report
2877 * @root: root vfsmnt/dentry
2878 * @buffer: pointer to the end of the buffer
2879 * @buflen: pointer to buffer length
2880 *
2881 * The function will first try to write out the pathname without taking any
2882 * lock other than the RCU read lock to make sure that dentries won't go away.
2883 * It only checks the sequence number of the global rename_lock as any change
2884 * in the dentry's d_seq will be preceded by changes in the rename_lock
2885 * sequence number. If the sequence number had been changed, it will restart
2886 * the whole pathname back-tracing sequence again by taking the rename_lock.
2887 * In this case, there is no need to take the RCU read lock as the recursive
2888 * parent pointer references will keep the dentry chain alive as long as no
2889 * rename operation is performed.
2890 */
2891 static int prepend_path(const struct path *path,
2892 const struct path *root,
2893 char **buffer, int *buflen)
2894 {
2895 struct dentry *dentry;
2896 struct vfsmount *vfsmnt;
2897 struct mount *mnt;
2898 int error = 0;
2899 unsigned seq, m_seq = 0;
2900 char *bptr;
2901 int blen;
2902
2903 rcu_read_lock();
2904 restart_mnt:
2905 read_seqbegin_or_lock(&mount_lock, &m_seq);
2906 seq = 0;
2907 rcu_read_lock();
2908 restart:
2909 bptr = *buffer;
2910 blen = *buflen;
2911 error = 0;
2912 dentry = path->dentry;
2913 vfsmnt = path->mnt;
2914 mnt = real_mount(vfsmnt);
2915 read_seqbegin_or_lock(&rename_lock, &seq);
2916 while (dentry != root->dentry || vfsmnt != root->mnt) {
2917 struct dentry * parent;
2918
2919 if (dentry == vfsmnt->mnt_root || IS_ROOT(dentry)) {
2920 struct mount *parent = ACCESS_ONCE(mnt->mnt_parent);
2921 /* Escaped? */
2922 if (dentry != vfsmnt->mnt_root) {
2923 bptr = *buffer;
2924 blen = *buflen;
2925 error = 3;
2926 break;
2927 }
2928 /* Global root? */
2929 if (mnt != parent) {
2930 dentry = ACCESS_ONCE(mnt->mnt_mountpoint);
2931 mnt = parent;
2932 vfsmnt = &mnt->mnt;
2933 continue;
2934 }
2935 if (!error)
2936 error = is_mounted(vfsmnt) ? 1 : 2;
2937 break;
2938 }
2939 parent = dentry->d_parent;
2940 prefetch(parent);
2941 error = prepend_name(&bptr, &blen, &dentry->d_name);
2942 if (error)
2943 break;
2944
2945 dentry = parent;
2946 }
2947 if (!(seq & 1))
2948 rcu_read_unlock();
2949 if (need_seqretry(&rename_lock, seq)) {
2950 seq = 1;
2951 goto restart;
2952 }
2953 done_seqretry(&rename_lock, seq);
2954
2955 if (!(m_seq & 1))
2956 rcu_read_unlock();
2957 if (need_seqretry(&mount_lock, m_seq)) {
2958 m_seq = 1;
2959 goto restart_mnt;
2960 }
2961 done_seqretry(&mount_lock, m_seq);
2962
2963 if (error >= 0 && bptr == *buffer) {
2964 if (--blen < 0)
2965 error = -ENAMETOOLONG;
2966 else
2967 *--bptr = '/';
2968 }
2969 *buffer = bptr;
2970 *buflen = blen;
2971 return error;
2972 }
2973
2974 /**
2975 * __d_path - return the path of a dentry
2976 * @path: the dentry/vfsmount to report
2977 * @root: root vfsmnt/dentry
2978 * @buf: buffer to return value in
2979 * @buflen: buffer length
2980 *
2981 * Convert a dentry into an ASCII path name.
2982 *
2983 * Returns a pointer into the buffer or an error code if the
2984 * path was too long.
2985 *
2986 * "buflen" should be positive.
2987 *
2988 * If the path is not reachable from the supplied root, return %NULL.
2989 */
2990 char *__d_path(const struct path *path,
2991 const struct path *root,
2992 char *buf, int buflen)
2993 {
2994 char *res = buf + buflen;
2995 int error;
2996
2997 prepend(&res, &buflen, "\0", 1);
2998 error = prepend_path(path, root, &res, &buflen);
2999
3000 if (error < 0)
3001 return ERR_PTR(error);
3002 if (error > 0)
3003 return NULL;
3004 return res;
3005 }
3006
3007 char *d_absolute_path(const struct path *path,
3008 char *buf, int buflen)
3009 {
3010 struct path root = {};
3011 char *res = buf + buflen;
3012 int error;
3013
3014 prepend(&res, &buflen, "\0", 1);
3015 error = prepend_path(path, &root, &res, &buflen);
3016
3017 if (error > 1)
3018 error = -EINVAL;
3019 if (error < 0)
3020 return ERR_PTR(error);
3021 return res;
3022 }
3023
3024 /*
3025 * same as __d_path but appends "(deleted)" for unlinked files.
3026 */
3027 static int path_with_deleted(const struct path *path,
3028 const struct path *root,
3029 char **buf, int *buflen)
3030 {
3031 prepend(buf, buflen, "\0", 1);
3032 if (d_unlinked(path->dentry)) {
3033 int error = prepend(buf, buflen, " (deleted)", 10);
3034 if (error)
3035 return error;
3036 }
3037
3038 return prepend_path(path, root, buf, buflen);
3039 }
3040
3041 static int prepend_unreachable(char **buffer, int *buflen)
3042 {
3043 return prepend(buffer, buflen, "(unreachable)", 13);
3044 }
3045
3046 static void get_fs_root_rcu(struct fs_struct *fs, struct path *root)
3047 {
3048 unsigned seq;
3049
3050 do {
3051 seq = read_seqcount_begin(&fs->seq);
3052 *root = fs->root;
3053 } while (read_seqcount_retry(&fs->seq, seq));
3054 }
3055
3056 /**
3057 * d_path - return the path of a dentry
3058 * @path: path to report
3059 * @buf: buffer to return value in
3060 * @buflen: buffer length
3061 *
3062 * Convert a dentry into an ASCII path name. If the entry has been deleted
3063 * the string " (deleted)" is appended. Note that this is ambiguous.
3064 *
3065 * Returns a pointer into the buffer or an error code if the path was
3066 * too long. Note: Callers should use the returned pointer, not the passed
3067 * in buffer, to use the name! The implementation often starts at an offset
3068 * into the buffer, and may leave 0 bytes at the start.
3069 *
3070 * "buflen" should be positive.
3071 */
3072 char *d_path(const struct path *path, char *buf, int buflen)
3073 {
3074 char *res = buf + buflen;
3075 struct path root;
3076 int error;
3077
3078 /*
3079 * We have various synthetic filesystems that never get mounted. On
3080 * these filesystems dentries are never used for lookup purposes, and
3081 * thus don't need to be hashed. They also don't need a name until a
3082 * user wants to identify the object in /proc/pid/fd/. The little hack
3083 * below allows us to generate a name for these objects on demand:
3084 *
3085 * Some pseudo inodes are mountable. When they are mounted
3086 * path->dentry == path->mnt->mnt_root. In that case don't call d_dname
3087 * and instead have d_path return the mounted path.
3088 */
3089 if (path->dentry->d_op && path->dentry->d_op->d_dname &&
3090 (!IS_ROOT(path->dentry) || path->dentry != path->mnt->mnt_root))
3091 return path->dentry->d_op->d_dname(path->dentry, buf, buflen);
3092
3093 rcu_read_lock();
3094 get_fs_root_rcu(current->fs, &root);
3095 error = path_with_deleted(path, &root, &res, &buflen);
3096 rcu_read_unlock();
3097
3098 if (error < 0)
3099 res = ERR_PTR(error);
3100 return res;
3101 }
3102 EXPORT_SYMBOL(d_path);
3103
3104 /*
3105 * Helper function for dentry_operations.d_dname() members
3106 */
3107 char *dynamic_dname(struct dentry *dentry, char *buffer, int buflen,
3108 const char *fmt, ...)
3109 {
3110 va_list args;
3111 char temp[64];
3112 int sz;
3113
3114 va_start(args, fmt);
3115 sz = vsnprintf(temp, sizeof(temp), fmt, args) + 1;
3116 va_end(args);
3117
3118 if (sz > sizeof(temp) || sz > buflen)
3119 return ERR_PTR(-ENAMETOOLONG);
3120
3121 buffer += buflen - sz;
3122 return memcpy(buffer, temp, sz);
3123 }
3124
3125 char *simple_dname(struct dentry *dentry, char *buffer, int buflen)
3126 {
3127 char *end = buffer + buflen;
3128 /* these dentries are never renamed, so d_lock is not needed */
3129 if (prepend(&end, &buflen, " (deleted)", 11) ||
3130 prepend(&end, &buflen, dentry->d_name.name, dentry->d_name.len) ||
3131 prepend(&end, &buflen, "/", 1))
3132 end = ERR_PTR(-ENAMETOOLONG);
3133 return end;
3134 }
3135 EXPORT_SYMBOL(simple_dname);
3136
3137 /*
3138 * Write full pathname from the root of the filesystem into the buffer.
3139 */
3140 static char *__dentry_path(struct dentry *d, char *buf, int buflen)
3141 {
3142 struct dentry *dentry;
3143 char *end, *retval;
3144 int len, seq = 0;
3145 int error = 0;
3146
3147 if (buflen < 2)
3148 goto Elong;
3149
3150 rcu_read_lock();
3151 restart:
3152 dentry = d;
3153 end = buf + buflen;
3154 len = buflen;
3155 prepend(&end, &len, "\0", 1);
3156 /* Get '/' right */
3157 retval = end-1;
3158 *retval = '/';
3159 read_seqbegin_or_lock(&rename_lock, &seq);
3160 while (!IS_ROOT(dentry)) {
3161 struct dentry *parent = dentry->d_parent;
3162
3163 prefetch(parent);
3164 error = prepend_name(&end, &len, &dentry->d_name);
3165 if (error)
3166 break;
3167
3168 retval = end;
3169 dentry = parent;
3170 }
3171 if (!(seq & 1))
3172 rcu_read_unlock();
3173 if (need_seqretry(&rename_lock, seq)) {
3174 seq = 1;
3175 goto restart;
3176 }
3177 done_seqretry(&rename_lock, seq);
3178 if (error)
3179 goto Elong;
3180 return retval;
3181 Elong:
3182 return ERR_PTR(-ENAMETOOLONG);
3183 }
3184
3185 char *dentry_path_raw(struct dentry *dentry, char *buf, int buflen)
3186 {
3187 return __dentry_path(dentry, buf, buflen);
3188 }
3189 EXPORT_SYMBOL(dentry_path_raw);
3190
3191 char *dentry_path(struct dentry *dentry, char *buf, int buflen)
3192 {
3193 char *p = NULL;
3194 char *retval;
3195
3196 if (d_unlinked(dentry)) {
3197 p = buf + buflen;
3198 if (prepend(&p, &buflen, "//deleted", 10) != 0)
3199 goto Elong;
3200 buflen++;
3201 }
3202 retval = __dentry_path(dentry, buf, buflen);
3203 if (!IS_ERR(retval) && p)
3204 *p = '/'; /* restore '/' overriden with '\0' */
3205 return retval;
3206 Elong:
3207 return ERR_PTR(-ENAMETOOLONG);
3208 }
3209
3210 static void get_fs_root_and_pwd_rcu(struct fs_struct *fs, struct path *root,
3211 struct path *pwd)
3212 {
3213 unsigned seq;
3214
3215 do {
3216 seq = read_seqcount_begin(&fs->seq);
3217 *root = fs->root;
3218 *pwd = fs->pwd;
3219 } while (read_seqcount_retry(&fs->seq, seq));
3220 }
3221
3222 /*
3223 * NOTE! The user-level library version returns a
3224 * character pointer. The kernel system call just
3225 * returns the length of the buffer filled (which
3226 * includes the ending '\0' character), or a negative
3227 * error value. So libc would do something like
3228 *
3229 * char *getcwd(char * buf, size_t size)
3230 * {
3231 * int retval;
3232 *
3233 * retval = sys_getcwd(buf, size);
3234 * if (retval >= 0)
3235 * return buf;
3236 * errno = -retval;
3237 * return NULL;
3238 * }
3239 */
3240 SYSCALL_DEFINE2(getcwd, char __user *, buf, unsigned long, size)
3241 {
3242 int error;
3243 struct path pwd, root;
3244 char *page = __getname();
3245
3246 if (!page)
3247 return -ENOMEM;
3248
3249 rcu_read_lock();
3250 get_fs_root_and_pwd_rcu(current->fs, &root, &pwd);
3251
3252 error = -ENOENT;
3253 if (!d_unlinked(pwd.dentry)) {
3254 unsigned long len;
3255 char *cwd = page + PATH_MAX;
3256 int buflen = PATH_MAX;
3257
3258 prepend(&cwd, &buflen, "\0", 1);
3259 error = prepend_path(&pwd, &root, &cwd, &buflen);
3260 rcu_read_unlock();
3261
3262 if (error < 0)
3263 goto out;
3264
3265 /* Unreachable from current root */
3266 if (error > 0) {
3267 error = prepend_unreachable(&cwd, &buflen);
3268 if (error)
3269 goto out;
3270 }
3271
3272 error = -ERANGE;
3273 len = PATH_MAX + page - cwd;
3274 if (len <= size) {
3275 error = len;
3276 if (copy_to_user(buf, cwd, len))
3277 error = -EFAULT;
3278 }
3279 } else {
3280 rcu_read_unlock();
3281 }
3282
3283 out:
3284 __putname(page);
3285 return error;
3286 }
3287
3288 /*
3289 * Test whether new_dentry is a subdirectory of old_dentry.
3290 *
3291 * Trivially implemented using the dcache structure
3292 */
3293
3294 /**
3295 * is_subdir - is new dentry a subdirectory of old_dentry
3296 * @new_dentry: new dentry
3297 * @old_dentry: old dentry
3298 *
3299 * Returns true if new_dentry is a subdirectory of the parent (at any depth).
3300 * Returns false otherwise.
3301 * Caller must ensure that "new_dentry" is pinned before calling is_subdir()
3302 */
3303
3304 bool is_subdir(struct dentry *new_dentry, struct dentry *old_dentry)
3305 {
3306 bool result;
3307 unsigned seq;
3308
3309 if (new_dentry == old_dentry)
3310 return true;
3311
3312 do {
3313 /* for restarting inner loop in case of seq retry */
3314 seq = read_seqbegin(&rename_lock);
3315 /*
3316 * Need rcu_readlock to protect against the d_parent trashing
3317 * due to d_move
3318 */
3319 rcu_read_lock();
3320 if (d_ancestor(old_dentry, new_dentry))
3321 result = true;
3322 else
3323 result = false;
3324 rcu_read_unlock();
3325 } while (read_seqretry(&rename_lock, seq));
3326
3327 return result;
3328 }
3329
3330 static enum d_walk_ret d_genocide_kill(void *data, struct dentry *dentry)
3331 {
3332 struct dentry *root = data;
3333 if (dentry != root) {
3334 if (d_unhashed(dentry) || !dentry->d_inode)
3335 return D_WALK_SKIP;
3336
3337 if (!(dentry->d_flags & DCACHE_GENOCIDE)) {
3338 dentry->d_flags |= DCACHE_GENOCIDE;
3339 dentry->d_lockref.count--;
3340 }
3341 }
3342 return D_WALK_CONTINUE;
3343 }
3344
3345 void d_genocide(struct dentry *parent)
3346 {
3347 d_walk(parent, parent, d_genocide_kill, NULL);
3348 }
3349
3350 void d_tmpfile(struct dentry *dentry, struct inode *inode)
3351 {
3352 inode_dec_link_count(inode);
3353 BUG_ON(dentry->d_name.name != dentry->d_iname ||
3354 !hlist_unhashed(&dentry->d_u.d_alias) ||
3355 !d_unlinked(dentry));
3356 spin_lock(&dentry->d_parent->d_lock);
3357 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED);
3358 dentry->d_name.len = sprintf(dentry->d_iname, "#%llu",
3359 (unsigned long long)inode->i_ino);
3360 spin_unlock(&dentry->d_lock);
3361 spin_unlock(&dentry->d_parent->d_lock);
3362 d_instantiate(dentry, inode);
3363 }
3364 EXPORT_SYMBOL(d_tmpfile);
3365
3366 static __initdata unsigned long dhash_entries;
3367 static int __init set_dhash_entries(char *str)
3368 {
3369 if (!str)
3370 return 0;
3371 dhash_entries = simple_strtoul(str, &str, 0);
3372 return 1;
3373 }
3374 __setup("dhash_entries=", set_dhash_entries);
3375
3376 static void __init dcache_init_early(void)
3377 {
3378 unsigned int loop;
3379
3380 /* If hashes are distributed across NUMA nodes, defer
3381 * hash allocation until vmalloc space is available.
3382 */
3383 if (hashdist)
3384 return;
3385
3386 dentry_hashtable =
3387 alloc_large_system_hash("Dentry cache",
3388 sizeof(struct hlist_bl_head),
3389 dhash_entries,
3390 13,
3391 HASH_EARLY,
3392 &d_hash_shift,
3393 &d_hash_mask,
3394 0,
3395 0);
3396
3397 for (loop = 0; loop < (1U << d_hash_shift); loop++)
3398 INIT_HLIST_BL_HEAD(dentry_hashtable + loop);
3399 }
3400
3401 static void __init dcache_init(void)
3402 {
3403 unsigned int loop;
3404
3405 /*
3406 * A constructor could be added for stable state like the lists,
3407 * but it is probably not worth it because of the cache nature
3408 * of the dcache.
3409 */
3410 dentry_cache = KMEM_CACHE(dentry,
3411 SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|SLAB_MEM_SPREAD|SLAB_ACCOUNT);
3412
3413 /* Hash may have been set up in dcache_init_early */
3414 if (!hashdist)
3415 return;
3416
3417 dentry_hashtable =
3418 alloc_large_system_hash("Dentry cache",
3419 sizeof(struct hlist_bl_head),
3420 dhash_entries,
3421 13,
3422 0,
3423 &d_hash_shift,
3424 &d_hash_mask,
3425 0,
3426 0);
3427
3428 for (loop = 0; loop < (1U << d_hash_shift); loop++)
3429 INIT_HLIST_BL_HEAD(dentry_hashtable + loop);
3430 }
3431
3432 /* SLAB cache for __getname() consumers */
3433 struct kmem_cache *names_cachep __read_mostly;
3434 EXPORT_SYMBOL(names_cachep);
3435
3436 EXPORT_SYMBOL(d_genocide);
3437
3438 void __init vfs_caches_init_early(void)
3439 {
3440 dcache_init_early();
3441 inode_init_early();
3442 }
3443
3444 void __init vfs_caches_init(void)
3445 {
3446 names_cachep = kmem_cache_create("names_cache", PATH_MAX, 0,
3447 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3448
3449 dcache_init();
3450 inode_init();
3451 files_init();
3452 files_maxfiles_init();
3453 mnt_init();
3454 bdev_cache_init();
3455 chrdev_init();
3456 }