btrfs: drop constant param from btrfs_release_extent_buffer_page
[sfrench/cifs-2.6.git] / fs / btrfs / extent_io.c
1 #include <linux/bitops.h>
2 #include <linux/slab.h>
3 #include <linux/bio.h>
4 #include <linux/mm.h>
5 #include <linux/pagemap.h>
6 #include <linux/page-flags.h>
7 #include <linux/spinlock.h>
8 #include <linux/blkdev.h>
9 #include <linux/swap.h>
10 #include <linux/writeback.h>
11 #include <linux/pagevec.h>
12 #include <linux/prefetch.h>
13 #include <linux/cleancache.h>
14 #include "extent_io.h"
15 #include "extent_map.h"
16 #include "ctree.h"
17 #include "btrfs_inode.h"
18 #include "volumes.h"
19 #include "check-integrity.h"
20 #include "locking.h"
21 #include "rcu-string.h"
22 #include "backref.h"
23
24 static struct kmem_cache *extent_state_cache;
25 static struct kmem_cache *extent_buffer_cache;
26 static struct bio_set *btrfs_bioset;
27
28 static inline bool extent_state_in_tree(const struct extent_state *state)
29 {
30         return !RB_EMPTY_NODE(&state->rb_node);
31 }
32
33 #ifdef CONFIG_BTRFS_DEBUG
34 static LIST_HEAD(buffers);
35 static LIST_HEAD(states);
36
37 static DEFINE_SPINLOCK(leak_lock);
38
39 static inline
40 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
41 {
42         unsigned long flags;
43
44         spin_lock_irqsave(&leak_lock, flags);
45         list_add(new, head);
46         spin_unlock_irqrestore(&leak_lock, flags);
47 }
48
49 static inline
50 void btrfs_leak_debug_del(struct list_head *entry)
51 {
52         unsigned long flags;
53
54         spin_lock_irqsave(&leak_lock, flags);
55         list_del(entry);
56         spin_unlock_irqrestore(&leak_lock, flags);
57 }
58
59 static inline
60 void btrfs_leak_debug_check(void)
61 {
62         struct extent_state *state;
63         struct extent_buffer *eb;
64
65         while (!list_empty(&states)) {
66                 state = list_entry(states.next, struct extent_state, leak_list);
67                 pr_err("BTRFS: state leak: start %llu end %llu state %lu in tree %d refs %d\n",
68                        state->start, state->end, state->state,
69                        extent_state_in_tree(state),
70                        atomic_read(&state->refs));
71                 list_del(&state->leak_list);
72                 kmem_cache_free(extent_state_cache, state);
73         }
74
75         while (!list_empty(&buffers)) {
76                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
77                 printk(KERN_ERR "BTRFS: buffer leak start %llu len %lu "
78                        "refs %d\n",
79                        eb->start, eb->len, atomic_read(&eb->refs));
80                 list_del(&eb->leak_list);
81                 kmem_cache_free(extent_buffer_cache, eb);
82         }
83 }
84
85 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
86         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
87 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
88                 struct extent_io_tree *tree, u64 start, u64 end)
89 {
90         struct inode *inode;
91         u64 isize;
92
93         if (!tree->mapping)
94                 return;
95
96         inode = tree->mapping->host;
97         isize = i_size_read(inode);
98         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
99                 printk_ratelimited(KERN_DEBUG
100                     "BTRFS: %s: ino %llu isize %llu odd range [%llu,%llu]\n",
101                                 caller, btrfs_ino(inode), isize, start, end);
102         }
103 }
104 #else
105 #define btrfs_leak_debug_add(new, head) do {} while (0)
106 #define btrfs_leak_debug_del(entry)     do {} while (0)
107 #define btrfs_leak_debug_check()        do {} while (0)
108 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
109 #endif
110
111 #define BUFFER_LRU_MAX 64
112
113 struct tree_entry {
114         u64 start;
115         u64 end;
116         struct rb_node rb_node;
117 };
118
119 struct extent_page_data {
120         struct bio *bio;
121         struct extent_io_tree *tree;
122         get_extent_t *get_extent;
123         unsigned long bio_flags;
124
125         /* tells writepage not to lock the state bits for this range
126          * it still does the unlocking
127          */
128         unsigned int extent_locked:1;
129
130         /* tells the submit_bio code to use a WRITE_SYNC */
131         unsigned int sync_io:1;
132 };
133
134 static noinline void flush_write_bio(void *data);
135 static inline struct btrfs_fs_info *
136 tree_fs_info(struct extent_io_tree *tree)
137 {
138         if (!tree->mapping)
139                 return NULL;
140         return btrfs_sb(tree->mapping->host->i_sb);
141 }
142
143 int __init extent_io_init(void)
144 {
145         extent_state_cache = kmem_cache_create("btrfs_extent_state",
146                         sizeof(struct extent_state), 0,
147                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
148         if (!extent_state_cache)
149                 return -ENOMEM;
150
151         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
152                         sizeof(struct extent_buffer), 0,
153                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
154         if (!extent_buffer_cache)
155                 goto free_state_cache;
156
157         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
158                                      offsetof(struct btrfs_io_bio, bio));
159         if (!btrfs_bioset)
160                 goto free_buffer_cache;
161
162         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
163                 goto free_bioset;
164
165         return 0;
166
167 free_bioset:
168         bioset_free(btrfs_bioset);
169         btrfs_bioset = NULL;
170
171 free_buffer_cache:
172         kmem_cache_destroy(extent_buffer_cache);
173         extent_buffer_cache = NULL;
174
175 free_state_cache:
176         kmem_cache_destroy(extent_state_cache);
177         extent_state_cache = NULL;
178         return -ENOMEM;
179 }
180
181 void extent_io_exit(void)
182 {
183         btrfs_leak_debug_check();
184
185         /*
186          * Make sure all delayed rcu free are flushed before we
187          * destroy caches.
188          */
189         rcu_barrier();
190         if (extent_state_cache)
191                 kmem_cache_destroy(extent_state_cache);
192         if (extent_buffer_cache)
193                 kmem_cache_destroy(extent_buffer_cache);
194         if (btrfs_bioset)
195                 bioset_free(btrfs_bioset);
196 }
197
198 void extent_io_tree_init(struct extent_io_tree *tree,
199                          struct address_space *mapping)
200 {
201         tree->state = RB_ROOT;
202         tree->ops = NULL;
203         tree->dirty_bytes = 0;
204         spin_lock_init(&tree->lock);
205         tree->mapping = mapping;
206 }
207
208 static struct extent_state *alloc_extent_state(gfp_t mask)
209 {
210         struct extent_state *state;
211
212         state = kmem_cache_alloc(extent_state_cache, mask);
213         if (!state)
214                 return state;
215         state->state = 0;
216         state->private = 0;
217         RB_CLEAR_NODE(&state->rb_node);
218         btrfs_leak_debug_add(&state->leak_list, &states);
219         atomic_set(&state->refs, 1);
220         init_waitqueue_head(&state->wq);
221         trace_alloc_extent_state(state, mask, _RET_IP_);
222         return state;
223 }
224
225 void free_extent_state(struct extent_state *state)
226 {
227         if (!state)
228                 return;
229         if (atomic_dec_and_test(&state->refs)) {
230                 WARN_ON(extent_state_in_tree(state));
231                 btrfs_leak_debug_del(&state->leak_list);
232                 trace_free_extent_state(state, _RET_IP_);
233                 kmem_cache_free(extent_state_cache, state);
234         }
235 }
236
237 static struct rb_node *tree_insert(struct rb_root *root,
238                                    struct rb_node *search_start,
239                                    u64 offset,
240                                    struct rb_node *node,
241                                    struct rb_node ***p_in,
242                                    struct rb_node **parent_in)
243 {
244         struct rb_node **p;
245         struct rb_node *parent = NULL;
246         struct tree_entry *entry;
247
248         if (p_in && parent_in) {
249                 p = *p_in;
250                 parent = *parent_in;
251                 goto do_insert;
252         }
253
254         p = search_start ? &search_start : &root->rb_node;
255         while (*p) {
256                 parent = *p;
257                 entry = rb_entry(parent, struct tree_entry, rb_node);
258
259                 if (offset < entry->start)
260                         p = &(*p)->rb_left;
261                 else if (offset > entry->end)
262                         p = &(*p)->rb_right;
263                 else
264                         return parent;
265         }
266
267 do_insert:
268         rb_link_node(node, parent, p);
269         rb_insert_color(node, root);
270         return NULL;
271 }
272
273 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
274                                       struct rb_node **prev_ret,
275                                       struct rb_node **next_ret,
276                                       struct rb_node ***p_ret,
277                                       struct rb_node **parent_ret)
278 {
279         struct rb_root *root = &tree->state;
280         struct rb_node **n = &root->rb_node;
281         struct rb_node *prev = NULL;
282         struct rb_node *orig_prev = NULL;
283         struct tree_entry *entry;
284         struct tree_entry *prev_entry = NULL;
285
286         while (*n) {
287                 prev = *n;
288                 entry = rb_entry(prev, struct tree_entry, rb_node);
289                 prev_entry = entry;
290
291                 if (offset < entry->start)
292                         n = &(*n)->rb_left;
293                 else if (offset > entry->end)
294                         n = &(*n)->rb_right;
295                 else
296                         return *n;
297         }
298
299         if (p_ret)
300                 *p_ret = n;
301         if (parent_ret)
302                 *parent_ret = prev;
303
304         if (prev_ret) {
305                 orig_prev = prev;
306                 while (prev && offset > prev_entry->end) {
307                         prev = rb_next(prev);
308                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
309                 }
310                 *prev_ret = prev;
311                 prev = orig_prev;
312         }
313
314         if (next_ret) {
315                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
316                 while (prev && offset < prev_entry->start) {
317                         prev = rb_prev(prev);
318                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
319                 }
320                 *next_ret = prev;
321         }
322         return NULL;
323 }
324
325 static inline struct rb_node *
326 tree_search_for_insert(struct extent_io_tree *tree,
327                        u64 offset,
328                        struct rb_node ***p_ret,
329                        struct rb_node **parent_ret)
330 {
331         struct rb_node *prev = NULL;
332         struct rb_node *ret;
333
334         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
335         if (!ret)
336                 return prev;
337         return ret;
338 }
339
340 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
341                                           u64 offset)
342 {
343         return tree_search_for_insert(tree, offset, NULL, NULL);
344 }
345
346 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
347                      struct extent_state *other)
348 {
349         if (tree->ops && tree->ops->merge_extent_hook)
350                 tree->ops->merge_extent_hook(tree->mapping->host, new,
351                                              other);
352 }
353
354 /*
355  * utility function to look for merge candidates inside a given range.
356  * Any extents with matching state are merged together into a single
357  * extent in the tree.  Extents with EXTENT_IO in their state field
358  * are not merged because the end_io handlers need to be able to do
359  * operations on them without sleeping (or doing allocations/splits).
360  *
361  * This should be called with the tree lock held.
362  */
363 static void merge_state(struct extent_io_tree *tree,
364                         struct extent_state *state)
365 {
366         struct extent_state *other;
367         struct rb_node *other_node;
368
369         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
370                 return;
371
372         other_node = rb_prev(&state->rb_node);
373         if (other_node) {
374                 other = rb_entry(other_node, struct extent_state, rb_node);
375                 if (other->end == state->start - 1 &&
376                     other->state == state->state) {
377                         merge_cb(tree, state, other);
378                         state->start = other->start;
379                         rb_erase(&other->rb_node, &tree->state);
380                         RB_CLEAR_NODE(&other->rb_node);
381                         free_extent_state(other);
382                 }
383         }
384         other_node = rb_next(&state->rb_node);
385         if (other_node) {
386                 other = rb_entry(other_node, struct extent_state, rb_node);
387                 if (other->start == state->end + 1 &&
388                     other->state == state->state) {
389                         merge_cb(tree, state, other);
390                         state->end = other->end;
391                         rb_erase(&other->rb_node, &tree->state);
392                         RB_CLEAR_NODE(&other->rb_node);
393                         free_extent_state(other);
394                 }
395         }
396 }
397
398 static void set_state_cb(struct extent_io_tree *tree,
399                          struct extent_state *state, unsigned long *bits)
400 {
401         if (tree->ops && tree->ops->set_bit_hook)
402                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
403 }
404
405 static void clear_state_cb(struct extent_io_tree *tree,
406                            struct extent_state *state, unsigned long *bits)
407 {
408         if (tree->ops && tree->ops->clear_bit_hook)
409                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
410 }
411
412 static void set_state_bits(struct extent_io_tree *tree,
413                            struct extent_state *state, unsigned long *bits);
414
415 /*
416  * insert an extent_state struct into the tree.  'bits' are set on the
417  * struct before it is inserted.
418  *
419  * This may return -EEXIST if the extent is already there, in which case the
420  * state struct is freed.
421  *
422  * The tree lock is not taken internally.  This is a utility function and
423  * probably isn't what you want to call (see set/clear_extent_bit).
424  */
425 static int insert_state(struct extent_io_tree *tree,
426                         struct extent_state *state, u64 start, u64 end,
427                         struct rb_node ***p,
428                         struct rb_node **parent,
429                         unsigned long *bits)
430 {
431         struct rb_node *node;
432
433         if (end < start)
434                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
435                        end, start);
436         state->start = start;
437         state->end = end;
438
439         set_state_bits(tree, state, bits);
440
441         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
442         if (node) {
443                 struct extent_state *found;
444                 found = rb_entry(node, struct extent_state, rb_node);
445                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
446                        "%llu %llu\n",
447                        found->start, found->end, start, end);
448                 return -EEXIST;
449         }
450         merge_state(tree, state);
451         return 0;
452 }
453
454 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
455                      u64 split)
456 {
457         if (tree->ops && tree->ops->split_extent_hook)
458                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
459 }
460
461 /*
462  * split a given extent state struct in two, inserting the preallocated
463  * struct 'prealloc' as the newly created second half.  'split' indicates an
464  * offset inside 'orig' where it should be split.
465  *
466  * Before calling,
467  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
468  * are two extent state structs in the tree:
469  * prealloc: [orig->start, split - 1]
470  * orig: [ split, orig->end ]
471  *
472  * The tree locks are not taken by this function. They need to be held
473  * by the caller.
474  */
475 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
476                        struct extent_state *prealloc, u64 split)
477 {
478         struct rb_node *node;
479
480         split_cb(tree, orig, split);
481
482         prealloc->start = orig->start;
483         prealloc->end = split - 1;
484         prealloc->state = orig->state;
485         orig->start = split;
486
487         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
488                            &prealloc->rb_node, NULL, NULL);
489         if (node) {
490                 free_extent_state(prealloc);
491                 return -EEXIST;
492         }
493         return 0;
494 }
495
496 static struct extent_state *next_state(struct extent_state *state)
497 {
498         struct rb_node *next = rb_next(&state->rb_node);
499         if (next)
500                 return rb_entry(next, struct extent_state, rb_node);
501         else
502                 return NULL;
503 }
504
505 /*
506  * utility function to clear some bits in an extent state struct.
507  * it will optionally wake up any one waiting on this state (wake == 1).
508  *
509  * If no bits are set on the state struct after clearing things, the
510  * struct is freed and removed from the tree
511  */
512 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
513                                             struct extent_state *state,
514                                             unsigned long *bits, int wake)
515 {
516         struct extent_state *next;
517         unsigned long bits_to_clear = *bits & ~EXTENT_CTLBITS;
518
519         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
520                 u64 range = state->end - state->start + 1;
521                 WARN_ON(range > tree->dirty_bytes);
522                 tree->dirty_bytes -= range;
523         }
524         clear_state_cb(tree, state, bits);
525         state->state &= ~bits_to_clear;
526         if (wake)
527                 wake_up(&state->wq);
528         if (state->state == 0) {
529                 next = next_state(state);
530                 if (extent_state_in_tree(state)) {
531                         rb_erase(&state->rb_node, &tree->state);
532                         RB_CLEAR_NODE(&state->rb_node);
533                         free_extent_state(state);
534                 } else {
535                         WARN_ON(1);
536                 }
537         } else {
538                 merge_state(tree, state);
539                 next = next_state(state);
540         }
541         return next;
542 }
543
544 static struct extent_state *
545 alloc_extent_state_atomic(struct extent_state *prealloc)
546 {
547         if (!prealloc)
548                 prealloc = alloc_extent_state(GFP_ATOMIC);
549
550         return prealloc;
551 }
552
553 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
554 {
555         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
556                     "Extent tree was modified by another "
557                     "thread while locked.");
558 }
559
560 /*
561  * clear some bits on a range in the tree.  This may require splitting
562  * or inserting elements in the tree, so the gfp mask is used to
563  * indicate which allocations or sleeping are allowed.
564  *
565  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
566  * the given range from the tree regardless of state (ie for truncate).
567  *
568  * the range [start, end] is inclusive.
569  *
570  * This takes the tree lock, and returns 0 on success and < 0 on error.
571  */
572 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
573                      unsigned long bits, int wake, int delete,
574                      struct extent_state **cached_state,
575                      gfp_t mask)
576 {
577         struct extent_state *state;
578         struct extent_state *cached;
579         struct extent_state *prealloc = NULL;
580         struct rb_node *node;
581         u64 last_end;
582         int err;
583         int clear = 0;
584
585         btrfs_debug_check_extent_io_range(tree, start, end);
586
587         if (bits & EXTENT_DELALLOC)
588                 bits |= EXTENT_NORESERVE;
589
590         if (delete)
591                 bits |= ~EXTENT_CTLBITS;
592         bits |= EXTENT_FIRST_DELALLOC;
593
594         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
595                 clear = 1;
596 again:
597         if (!prealloc && (mask & __GFP_WAIT)) {
598                 prealloc = alloc_extent_state(mask);
599                 if (!prealloc)
600                         return -ENOMEM;
601         }
602
603         spin_lock(&tree->lock);
604         if (cached_state) {
605                 cached = *cached_state;
606
607                 if (clear) {
608                         *cached_state = NULL;
609                         cached_state = NULL;
610                 }
611
612                 if (cached && extent_state_in_tree(cached) &&
613                     cached->start <= start && cached->end > start) {
614                         if (clear)
615                                 atomic_dec(&cached->refs);
616                         state = cached;
617                         goto hit_next;
618                 }
619                 if (clear)
620                         free_extent_state(cached);
621         }
622         /*
623          * this search will find the extents that end after
624          * our range starts
625          */
626         node = tree_search(tree, start);
627         if (!node)
628                 goto out;
629         state = rb_entry(node, struct extent_state, rb_node);
630 hit_next:
631         if (state->start > end)
632                 goto out;
633         WARN_ON(state->end < start);
634         last_end = state->end;
635
636         /* the state doesn't have the wanted bits, go ahead */
637         if (!(state->state & bits)) {
638                 state = next_state(state);
639                 goto next;
640         }
641
642         /*
643          *     | ---- desired range ---- |
644          *  | state | or
645          *  | ------------- state -------------- |
646          *
647          * We need to split the extent we found, and may flip
648          * bits on second half.
649          *
650          * If the extent we found extends past our range, we
651          * just split and search again.  It'll get split again
652          * the next time though.
653          *
654          * If the extent we found is inside our range, we clear
655          * the desired bit on it.
656          */
657
658         if (state->start < start) {
659                 prealloc = alloc_extent_state_atomic(prealloc);
660                 BUG_ON(!prealloc);
661                 err = split_state(tree, state, prealloc, start);
662                 if (err)
663                         extent_io_tree_panic(tree, err);
664
665                 prealloc = NULL;
666                 if (err)
667                         goto out;
668                 if (state->end <= end) {
669                         state = clear_state_bit(tree, state, &bits, wake);
670                         goto next;
671                 }
672                 goto search_again;
673         }
674         /*
675          * | ---- desired range ---- |
676          *                        | state |
677          * We need to split the extent, and clear the bit
678          * on the first half
679          */
680         if (state->start <= end && state->end > end) {
681                 prealloc = alloc_extent_state_atomic(prealloc);
682                 BUG_ON(!prealloc);
683                 err = split_state(tree, state, prealloc, end + 1);
684                 if (err)
685                         extent_io_tree_panic(tree, err);
686
687                 if (wake)
688                         wake_up(&state->wq);
689
690                 clear_state_bit(tree, prealloc, &bits, wake);
691
692                 prealloc = NULL;
693                 goto out;
694         }
695
696         state = clear_state_bit(tree, state, &bits, wake);
697 next:
698         if (last_end == (u64)-1)
699                 goto out;
700         start = last_end + 1;
701         if (start <= end && state && !need_resched())
702                 goto hit_next;
703         goto search_again;
704
705 out:
706         spin_unlock(&tree->lock);
707         if (prealloc)
708                 free_extent_state(prealloc);
709
710         return 0;
711
712 search_again:
713         if (start > end)
714                 goto out;
715         spin_unlock(&tree->lock);
716         if (mask & __GFP_WAIT)
717                 cond_resched();
718         goto again;
719 }
720
721 static void wait_on_state(struct extent_io_tree *tree,
722                           struct extent_state *state)
723                 __releases(tree->lock)
724                 __acquires(tree->lock)
725 {
726         DEFINE_WAIT(wait);
727         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
728         spin_unlock(&tree->lock);
729         schedule();
730         spin_lock(&tree->lock);
731         finish_wait(&state->wq, &wait);
732 }
733
734 /*
735  * waits for one or more bits to clear on a range in the state tree.
736  * The range [start, end] is inclusive.
737  * The tree lock is taken by this function
738  */
739 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
740                             unsigned long bits)
741 {
742         struct extent_state *state;
743         struct rb_node *node;
744
745         btrfs_debug_check_extent_io_range(tree, start, end);
746
747         spin_lock(&tree->lock);
748 again:
749         while (1) {
750                 /*
751                  * this search will find all the extents that end after
752                  * our range starts
753                  */
754                 node = tree_search(tree, start);
755 process_node:
756                 if (!node)
757                         break;
758
759                 state = rb_entry(node, struct extent_state, rb_node);
760
761                 if (state->start > end)
762                         goto out;
763
764                 if (state->state & bits) {
765                         start = state->start;
766                         atomic_inc(&state->refs);
767                         wait_on_state(tree, state);
768                         free_extent_state(state);
769                         goto again;
770                 }
771                 start = state->end + 1;
772
773                 if (start > end)
774                         break;
775
776                 if (!cond_resched_lock(&tree->lock)) {
777                         node = rb_next(node);
778                         goto process_node;
779                 }
780         }
781 out:
782         spin_unlock(&tree->lock);
783 }
784
785 static void set_state_bits(struct extent_io_tree *tree,
786                            struct extent_state *state,
787                            unsigned long *bits)
788 {
789         unsigned long bits_to_set = *bits & ~EXTENT_CTLBITS;
790
791         set_state_cb(tree, state, bits);
792         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
793                 u64 range = state->end - state->start + 1;
794                 tree->dirty_bytes += range;
795         }
796         state->state |= bits_to_set;
797 }
798
799 static void cache_state(struct extent_state *state,
800                         struct extent_state **cached_ptr)
801 {
802         if (cached_ptr && !(*cached_ptr)) {
803                 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
804                         *cached_ptr = state;
805                         atomic_inc(&state->refs);
806                 }
807         }
808 }
809
810 /*
811  * set some bits on a range in the tree.  This may require allocations or
812  * sleeping, so the gfp mask is used to indicate what is allowed.
813  *
814  * If any of the exclusive bits are set, this will fail with -EEXIST if some
815  * part of the range already has the desired bits set.  The start of the
816  * existing range is returned in failed_start in this case.
817  *
818  * [start, end] is inclusive This takes the tree lock.
819  */
820
821 static int __must_check
822 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
823                  unsigned long bits, unsigned long exclusive_bits,
824                  u64 *failed_start, struct extent_state **cached_state,
825                  gfp_t mask)
826 {
827         struct extent_state *state;
828         struct extent_state *prealloc = NULL;
829         struct rb_node *node;
830         struct rb_node **p;
831         struct rb_node *parent;
832         int err = 0;
833         u64 last_start;
834         u64 last_end;
835
836         btrfs_debug_check_extent_io_range(tree, start, end);
837
838         bits |= EXTENT_FIRST_DELALLOC;
839 again:
840         if (!prealloc && (mask & __GFP_WAIT)) {
841                 prealloc = alloc_extent_state(mask);
842                 BUG_ON(!prealloc);
843         }
844
845         spin_lock(&tree->lock);
846         if (cached_state && *cached_state) {
847                 state = *cached_state;
848                 if (state->start <= start && state->end > start &&
849                     extent_state_in_tree(state)) {
850                         node = &state->rb_node;
851                         goto hit_next;
852                 }
853         }
854         /*
855          * this search will find all the extents that end after
856          * our range starts.
857          */
858         node = tree_search_for_insert(tree, start, &p, &parent);
859         if (!node) {
860                 prealloc = alloc_extent_state_atomic(prealloc);
861                 BUG_ON(!prealloc);
862                 err = insert_state(tree, prealloc, start, end,
863                                    &p, &parent, &bits);
864                 if (err)
865                         extent_io_tree_panic(tree, err);
866
867                 cache_state(prealloc, cached_state);
868                 prealloc = NULL;
869                 goto out;
870         }
871         state = rb_entry(node, struct extent_state, rb_node);
872 hit_next:
873         last_start = state->start;
874         last_end = state->end;
875
876         /*
877          * | ---- desired range ---- |
878          * | state |
879          *
880          * Just lock what we found and keep going
881          */
882         if (state->start == start && state->end <= end) {
883                 if (state->state & exclusive_bits) {
884                         *failed_start = state->start;
885                         err = -EEXIST;
886                         goto out;
887                 }
888
889                 set_state_bits(tree, state, &bits);
890                 cache_state(state, cached_state);
891                 merge_state(tree, state);
892                 if (last_end == (u64)-1)
893                         goto out;
894                 start = last_end + 1;
895                 state = next_state(state);
896                 if (start < end && state && state->start == start &&
897                     !need_resched())
898                         goto hit_next;
899                 goto search_again;
900         }
901
902         /*
903          *     | ---- desired range ---- |
904          * | state |
905          *   or
906          * | ------------- state -------------- |
907          *
908          * We need to split the extent we found, and may flip bits on
909          * second half.
910          *
911          * If the extent we found extends past our
912          * range, we just split and search again.  It'll get split
913          * again the next time though.
914          *
915          * If the extent we found is inside our range, we set the
916          * desired bit on it.
917          */
918         if (state->start < start) {
919                 if (state->state & exclusive_bits) {
920                         *failed_start = start;
921                         err = -EEXIST;
922                         goto out;
923                 }
924
925                 prealloc = alloc_extent_state_atomic(prealloc);
926                 BUG_ON(!prealloc);
927                 err = split_state(tree, state, prealloc, start);
928                 if (err)
929                         extent_io_tree_panic(tree, err);
930
931                 prealloc = NULL;
932                 if (err)
933                         goto out;
934                 if (state->end <= end) {
935                         set_state_bits(tree, state, &bits);
936                         cache_state(state, cached_state);
937                         merge_state(tree, state);
938                         if (last_end == (u64)-1)
939                                 goto out;
940                         start = last_end + 1;
941                         state = next_state(state);
942                         if (start < end && state && state->start == start &&
943                             !need_resched())
944                                 goto hit_next;
945                 }
946                 goto search_again;
947         }
948         /*
949          * | ---- desired range ---- |
950          *     | state | or               | state |
951          *
952          * There's a hole, we need to insert something in it and
953          * ignore the extent we found.
954          */
955         if (state->start > start) {
956                 u64 this_end;
957                 if (end < last_start)
958                         this_end = end;
959                 else
960                         this_end = last_start - 1;
961
962                 prealloc = alloc_extent_state_atomic(prealloc);
963                 BUG_ON(!prealloc);
964
965                 /*
966                  * Avoid to free 'prealloc' if it can be merged with
967                  * the later extent.
968                  */
969                 err = insert_state(tree, prealloc, start, this_end,
970                                    NULL, NULL, &bits);
971                 if (err)
972                         extent_io_tree_panic(tree, err);
973
974                 cache_state(prealloc, cached_state);
975                 prealloc = NULL;
976                 start = this_end + 1;
977                 goto search_again;
978         }
979         /*
980          * | ---- desired range ---- |
981          *                        | state |
982          * We need to split the extent, and set the bit
983          * on the first half
984          */
985         if (state->start <= end && state->end > end) {
986                 if (state->state & exclusive_bits) {
987                         *failed_start = start;
988                         err = -EEXIST;
989                         goto out;
990                 }
991
992                 prealloc = alloc_extent_state_atomic(prealloc);
993                 BUG_ON(!prealloc);
994                 err = split_state(tree, state, prealloc, end + 1);
995                 if (err)
996                         extent_io_tree_panic(tree, err);
997
998                 set_state_bits(tree, prealloc, &bits);
999                 cache_state(prealloc, cached_state);
1000                 merge_state(tree, prealloc);
1001                 prealloc = NULL;
1002                 goto out;
1003         }
1004
1005         goto search_again;
1006
1007 out:
1008         spin_unlock(&tree->lock);
1009         if (prealloc)
1010                 free_extent_state(prealloc);
1011
1012         return err;
1013
1014 search_again:
1015         if (start > end)
1016                 goto out;
1017         spin_unlock(&tree->lock);
1018         if (mask & __GFP_WAIT)
1019                 cond_resched();
1020         goto again;
1021 }
1022
1023 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1024                    unsigned long bits, u64 * failed_start,
1025                    struct extent_state **cached_state, gfp_t mask)
1026 {
1027         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1028                                 cached_state, mask);
1029 }
1030
1031
1032 /**
1033  * convert_extent_bit - convert all bits in a given range from one bit to
1034  *                      another
1035  * @tree:       the io tree to search
1036  * @start:      the start offset in bytes
1037  * @end:        the end offset in bytes (inclusive)
1038  * @bits:       the bits to set in this range
1039  * @clear_bits: the bits to clear in this range
1040  * @cached_state:       state that we're going to cache
1041  * @mask:       the allocation mask
1042  *
1043  * This will go through and set bits for the given range.  If any states exist
1044  * already in this range they are set with the given bit and cleared of the
1045  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1046  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1047  * boundary bits like LOCK.
1048  */
1049 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1050                        unsigned long bits, unsigned long clear_bits,
1051                        struct extent_state **cached_state, gfp_t mask)
1052 {
1053         struct extent_state *state;
1054         struct extent_state *prealloc = NULL;
1055         struct rb_node *node;
1056         struct rb_node **p;
1057         struct rb_node *parent;
1058         int err = 0;
1059         u64 last_start;
1060         u64 last_end;
1061
1062         btrfs_debug_check_extent_io_range(tree, start, end);
1063
1064 again:
1065         if (!prealloc && (mask & __GFP_WAIT)) {
1066                 prealloc = alloc_extent_state(mask);
1067                 if (!prealloc)
1068                         return -ENOMEM;
1069         }
1070
1071         spin_lock(&tree->lock);
1072         if (cached_state && *cached_state) {
1073                 state = *cached_state;
1074                 if (state->start <= start && state->end > start &&
1075                     extent_state_in_tree(state)) {
1076                         node = &state->rb_node;
1077                         goto hit_next;
1078                 }
1079         }
1080
1081         /*
1082          * this search will find all the extents that end after
1083          * our range starts.
1084          */
1085         node = tree_search_for_insert(tree, start, &p, &parent);
1086         if (!node) {
1087                 prealloc = alloc_extent_state_atomic(prealloc);
1088                 if (!prealloc) {
1089                         err = -ENOMEM;
1090                         goto out;
1091                 }
1092                 err = insert_state(tree, prealloc, start, end,
1093                                    &p, &parent, &bits);
1094                 if (err)
1095                         extent_io_tree_panic(tree, err);
1096                 cache_state(prealloc, cached_state);
1097                 prealloc = NULL;
1098                 goto out;
1099         }
1100         state = rb_entry(node, struct extent_state, rb_node);
1101 hit_next:
1102         last_start = state->start;
1103         last_end = state->end;
1104
1105         /*
1106          * | ---- desired range ---- |
1107          * | state |
1108          *
1109          * Just lock what we found and keep going
1110          */
1111         if (state->start == start && state->end <= end) {
1112                 set_state_bits(tree, state, &bits);
1113                 cache_state(state, cached_state);
1114                 state = clear_state_bit(tree, state, &clear_bits, 0);
1115                 if (last_end == (u64)-1)
1116                         goto out;
1117                 start = last_end + 1;
1118                 if (start < end && state && state->start == start &&
1119                     !need_resched())
1120                         goto hit_next;
1121                 goto search_again;
1122         }
1123
1124         /*
1125          *     | ---- desired range ---- |
1126          * | state |
1127          *   or
1128          * | ------------- state -------------- |
1129          *
1130          * We need to split the extent we found, and may flip bits on
1131          * second half.
1132          *
1133          * If the extent we found extends past our
1134          * range, we just split and search again.  It'll get split
1135          * again the next time though.
1136          *
1137          * If the extent we found is inside our range, we set the
1138          * desired bit on it.
1139          */
1140         if (state->start < start) {
1141                 prealloc = alloc_extent_state_atomic(prealloc);
1142                 if (!prealloc) {
1143                         err = -ENOMEM;
1144                         goto out;
1145                 }
1146                 err = split_state(tree, state, prealloc, start);
1147                 if (err)
1148                         extent_io_tree_panic(tree, err);
1149                 prealloc = NULL;
1150                 if (err)
1151                         goto out;
1152                 if (state->end <= end) {
1153                         set_state_bits(tree, state, &bits);
1154                         cache_state(state, cached_state);
1155                         state = clear_state_bit(tree, state, &clear_bits, 0);
1156                         if (last_end == (u64)-1)
1157                                 goto out;
1158                         start = last_end + 1;
1159                         if (start < end && state && state->start == start &&
1160                             !need_resched())
1161                                 goto hit_next;
1162                 }
1163                 goto search_again;
1164         }
1165         /*
1166          * | ---- desired range ---- |
1167          *     | state | or               | state |
1168          *
1169          * There's a hole, we need to insert something in it and
1170          * ignore the extent we found.
1171          */
1172         if (state->start > start) {
1173                 u64 this_end;
1174                 if (end < last_start)
1175                         this_end = end;
1176                 else
1177                         this_end = last_start - 1;
1178
1179                 prealloc = alloc_extent_state_atomic(prealloc);
1180                 if (!prealloc) {
1181                         err = -ENOMEM;
1182                         goto out;
1183                 }
1184
1185                 /*
1186                  * Avoid to free 'prealloc' if it can be merged with
1187                  * the later extent.
1188                  */
1189                 err = insert_state(tree, prealloc, start, this_end,
1190                                    NULL, NULL, &bits);
1191                 if (err)
1192                         extent_io_tree_panic(tree, err);
1193                 cache_state(prealloc, cached_state);
1194                 prealloc = NULL;
1195                 start = this_end + 1;
1196                 goto search_again;
1197         }
1198         /*
1199          * | ---- desired range ---- |
1200          *                        | state |
1201          * We need to split the extent, and set the bit
1202          * on the first half
1203          */
1204         if (state->start <= end && state->end > end) {
1205                 prealloc = alloc_extent_state_atomic(prealloc);
1206                 if (!prealloc) {
1207                         err = -ENOMEM;
1208                         goto out;
1209                 }
1210
1211                 err = split_state(tree, state, prealloc, end + 1);
1212                 if (err)
1213                         extent_io_tree_panic(tree, err);
1214
1215                 set_state_bits(tree, prealloc, &bits);
1216                 cache_state(prealloc, cached_state);
1217                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1218                 prealloc = NULL;
1219                 goto out;
1220         }
1221
1222         goto search_again;
1223
1224 out:
1225         spin_unlock(&tree->lock);
1226         if (prealloc)
1227                 free_extent_state(prealloc);
1228
1229         return err;
1230
1231 search_again:
1232         if (start > end)
1233                 goto out;
1234         spin_unlock(&tree->lock);
1235         if (mask & __GFP_WAIT)
1236                 cond_resched();
1237         goto again;
1238 }
1239
1240 /* wrappers around set/clear extent bit */
1241 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1242                      gfp_t mask)
1243 {
1244         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1245                               NULL, mask);
1246 }
1247
1248 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1249                     unsigned long bits, gfp_t mask)
1250 {
1251         return set_extent_bit(tree, start, end, bits, NULL,
1252                               NULL, mask);
1253 }
1254
1255 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1256                       unsigned long bits, gfp_t mask)
1257 {
1258         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1259 }
1260
1261 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1262                         struct extent_state **cached_state, gfp_t mask)
1263 {
1264         return set_extent_bit(tree, start, end,
1265                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1266                               NULL, cached_state, mask);
1267 }
1268
1269 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1270                       struct extent_state **cached_state, gfp_t mask)
1271 {
1272         return set_extent_bit(tree, start, end,
1273                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1274                               NULL, cached_state, mask);
1275 }
1276
1277 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1278                        gfp_t mask)
1279 {
1280         return clear_extent_bit(tree, start, end,
1281                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1282                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1283 }
1284
1285 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1286                      gfp_t mask)
1287 {
1288         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1289                               NULL, mask);
1290 }
1291
1292 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1293                         struct extent_state **cached_state, gfp_t mask)
1294 {
1295         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1296                               cached_state, mask);
1297 }
1298
1299 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1300                           struct extent_state **cached_state, gfp_t mask)
1301 {
1302         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1303                                 cached_state, mask);
1304 }
1305
1306 /*
1307  * either insert or lock state struct between start and end use mask to tell
1308  * us if waiting is desired.
1309  */
1310 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1311                      unsigned long bits, struct extent_state **cached_state)
1312 {
1313         int err;
1314         u64 failed_start;
1315         while (1) {
1316                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1317                                        EXTENT_LOCKED, &failed_start,
1318                                        cached_state, GFP_NOFS);
1319                 if (err == -EEXIST) {
1320                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1321                         start = failed_start;
1322                 } else
1323                         break;
1324                 WARN_ON(start > end);
1325         }
1326         return err;
1327 }
1328
1329 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1330 {
1331         return lock_extent_bits(tree, start, end, 0, NULL);
1332 }
1333
1334 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1335 {
1336         int err;
1337         u64 failed_start;
1338
1339         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1340                                &failed_start, NULL, GFP_NOFS);
1341         if (err == -EEXIST) {
1342                 if (failed_start > start)
1343                         clear_extent_bit(tree, start, failed_start - 1,
1344                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1345                 return 0;
1346         }
1347         return 1;
1348 }
1349
1350 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1351                          struct extent_state **cached, gfp_t mask)
1352 {
1353         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1354                                 mask);
1355 }
1356
1357 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1358 {
1359         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1360                                 GFP_NOFS);
1361 }
1362
1363 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1364 {
1365         unsigned long index = start >> PAGE_CACHE_SHIFT;
1366         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1367         struct page *page;
1368
1369         while (index <= end_index) {
1370                 page = find_get_page(inode->i_mapping, index);
1371                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1372                 clear_page_dirty_for_io(page);
1373                 page_cache_release(page);
1374                 index++;
1375         }
1376         return 0;
1377 }
1378
1379 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1380 {
1381         unsigned long index = start >> PAGE_CACHE_SHIFT;
1382         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1383         struct page *page;
1384
1385         while (index <= end_index) {
1386                 page = find_get_page(inode->i_mapping, index);
1387                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1388                 account_page_redirty(page);
1389                 __set_page_dirty_nobuffers(page);
1390                 page_cache_release(page);
1391                 index++;
1392         }
1393         return 0;
1394 }
1395
1396 /*
1397  * helper function to set both pages and extents in the tree writeback
1398  */
1399 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1400 {
1401         unsigned long index = start >> PAGE_CACHE_SHIFT;
1402         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1403         struct page *page;
1404
1405         while (index <= end_index) {
1406                 page = find_get_page(tree->mapping, index);
1407                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1408                 set_page_writeback(page);
1409                 page_cache_release(page);
1410                 index++;
1411         }
1412         return 0;
1413 }
1414
1415 /* find the first state struct with 'bits' set after 'start', and
1416  * return it.  tree->lock must be held.  NULL will returned if
1417  * nothing was found after 'start'
1418  */
1419 static struct extent_state *
1420 find_first_extent_bit_state(struct extent_io_tree *tree,
1421                             u64 start, unsigned long bits)
1422 {
1423         struct rb_node *node;
1424         struct extent_state *state;
1425
1426         /*
1427          * this search will find all the extents that end after
1428          * our range starts.
1429          */
1430         node = tree_search(tree, start);
1431         if (!node)
1432                 goto out;
1433
1434         while (1) {
1435                 state = rb_entry(node, struct extent_state, rb_node);
1436                 if (state->end >= start && (state->state & bits))
1437                         return state;
1438
1439                 node = rb_next(node);
1440                 if (!node)
1441                         break;
1442         }
1443 out:
1444         return NULL;
1445 }
1446
1447 /*
1448  * find the first offset in the io tree with 'bits' set. zero is
1449  * returned if we find something, and *start_ret and *end_ret are
1450  * set to reflect the state struct that was found.
1451  *
1452  * If nothing was found, 1 is returned. If found something, return 0.
1453  */
1454 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1455                           u64 *start_ret, u64 *end_ret, unsigned long bits,
1456                           struct extent_state **cached_state)
1457 {
1458         struct extent_state *state;
1459         struct rb_node *n;
1460         int ret = 1;
1461
1462         spin_lock(&tree->lock);
1463         if (cached_state && *cached_state) {
1464                 state = *cached_state;
1465                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1466                         n = rb_next(&state->rb_node);
1467                         while (n) {
1468                                 state = rb_entry(n, struct extent_state,
1469                                                  rb_node);
1470                                 if (state->state & bits)
1471                                         goto got_it;
1472                                 n = rb_next(n);
1473                         }
1474                         free_extent_state(*cached_state);
1475                         *cached_state = NULL;
1476                         goto out;
1477                 }
1478                 free_extent_state(*cached_state);
1479                 *cached_state = NULL;
1480         }
1481
1482         state = find_first_extent_bit_state(tree, start, bits);
1483 got_it:
1484         if (state) {
1485                 cache_state(state, cached_state);
1486                 *start_ret = state->start;
1487                 *end_ret = state->end;
1488                 ret = 0;
1489         }
1490 out:
1491         spin_unlock(&tree->lock);
1492         return ret;
1493 }
1494
1495 /*
1496  * find a contiguous range of bytes in the file marked as delalloc, not
1497  * more than 'max_bytes'.  start and end are used to return the range,
1498  *
1499  * 1 is returned if we find something, 0 if nothing was in the tree
1500  */
1501 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1502                                         u64 *start, u64 *end, u64 max_bytes,
1503                                         struct extent_state **cached_state)
1504 {
1505         struct rb_node *node;
1506         struct extent_state *state;
1507         u64 cur_start = *start;
1508         u64 found = 0;
1509         u64 total_bytes = 0;
1510
1511         spin_lock(&tree->lock);
1512
1513         /*
1514          * this search will find all the extents that end after
1515          * our range starts.
1516          */
1517         node = tree_search(tree, cur_start);
1518         if (!node) {
1519                 if (!found)
1520                         *end = (u64)-1;
1521                 goto out;
1522         }
1523
1524         while (1) {
1525                 state = rb_entry(node, struct extent_state, rb_node);
1526                 if (found && (state->start != cur_start ||
1527                               (state->state & EXTENT_BOUNDARY))) {
1528                         goto out;
1529                 }
1530                 if (!(state->state & EXTENT_DELALLOC)) {
1531                         if (!found)
1532                                 *end = state->end;
1533                         goto out;
1534                 }
1535                 if (!found) {
1536                         *start = state->start;
1537                         *cached_state = state;
1538                         atomic_inc(&state->refs);
1539                 }
1540                 found++;
1541                 *end = state->end;
1542                 cur_start = state->end + 1;
1543                 node = rb_next(node);
1544                 total_bytes += state->end - state->start + 1;
1545                 if (total_bytes >= max_bytes)
1546                         break;
1547                 if (!node)
1548                         break;
1549         }
1550 out:
1551         spin_unlock(&tree->lock);
1552         return found;
1553 }
1554
1555 static noinline void __unlock_for_delalloc(struct inode *inode,
1556                                            struct page *locked_page,
1557                                            u64 start, u64 end)
1558 {
1559         int ret;
1560         struct page *pages[16];
1561         unsigned long index = start >> PAGE_CACHE_SHIFT;
1562         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1563         unsigned long nr_pages = end_index - index + 1;
1564         int i;
1565
1566         if (index == locked_page->index && end_index == index)
1567                 return;
1568
1569         while (nr_pages > 0) {
1570                 ret = find_get_pages_contig(inode->i_mapping, index,
1571                                      min_t(unsigned long, nr_pages,
1572                                      ARRAY_SIZE(pages)), pages);
1573                 for (i = 0; i < ret; i++) {
1574                         if (pages[i] != locked_page)
1575                                 unlock_page(pages[i]);
1576                         page_cache_release(pages[i]);
1577                 }
1578                 nr_pages -= ret;
1579                 index += ret;
1580                 cond_resched();
1581         }
1582 }
1583
1584 static noinline int lock_delalloc_pages(struct inode *inode,
1585                                         struct page *locked_page,
1586                                         u64 delalloc_start,
1587                                         u64 delalloc_end)
1588 {
1589         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1590         unsigned long start_index = index;
1591         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1592         unsigned long pages_locked = 0;
1593         struct page *pages[16];
1594         unsigned long nrpages;
1595         int ret;
1596         int i;
1597
1598         /* the caller is responsible for locking the start index */
1599         if (index == locked_page->index && index == end_index)
1600                 return 0;
1601
1602         /* skip the page at the start index */
1603         nrpages = end_index - index + 1;
1604         while (nrpages > 0) {
1605                 ret = find_get_pages_contig(inode->i_mapping, index,
1606                                      min_t(unsigned long,
1607                                      nrpages, ARRAY_SIZE(pages)), pages);
1608                 if (ret == 0) {
1609                         ret = -EAGAIN;
1610                         goto done;
1611                 }
1612                 /* now we have an array of pages, lock them all */
1613                 for (i = 0; i < ret; i++) {
1614                         /*
1615                          * the caller is taking responsibility for
1616                          * locked_page
1617                          */
1618                         if (pages[i] != locked_page) {
1619                                 lock_page(pages[i]);
1620                                 if (!PageDirty(pages[i]) ||
1621                                     pages[i]->mapping != inode->i_mapping) {
1622                                         ret = -EAGAIN;
1623                                         unlock_page(pages[i]);
1624                                         page_cache_release(pages[i]);
1625                                         goto done;
1626                                 }
1627                         }
1628                         page_cache_release(pages[i]);
1629                         pages_locked++;
1630                 }
1631                 nrpages -= ret;
1632                 index += ret;
1633                 cond_resched();
1634         }
1635         ret = 0;
1636 done:
1637         if (ret && pages_locked) {
1638                 __unlock_for_delalloc(inode, locked_page,
1639                               delalloc_start,
1640                               ((u64)(start_index + pages_locked - 1)) <<
1641                               PAGE_CACHE_SHIFT);
1642         }
1643         return ret;
1644 }
1645
1646 /*
1647  * find a contiguous range of bytes in the file marked as delalloc, not
1648  * more than 'max_bytes'.  start and end are used to return the range,
1649  *
1650  * 1 is returned if we find something, 0 if nothing was in the tree
1651  */
1652 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1653                                     struct extent_io_tree *tree,
1654                                     struct page *locked_page, u64 *start,
1655                                     u64 *end, u64 max_bytes)
1656 {
1657         u64 delalloc_start;
1658         u64 delalloc_end;
1659         u64 found;
1660         struct extent_state *cached_state = NULL;
1661         int ret;
1662         int loops = 0;
1663
1664 again:
1665         /* step one, find a bunch of delalloc bytes starting at start */
1666         delalloc_start = *start;
1667         delalloc_end = 0;
1668         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1669                                     max_bytes, &cached_state);
1670         if (!found || delalloc_end <= *start) {
1671                 *start = delalloc_start;
1672                 *end = delalloc_end;
1673                 free_extent_state(cached_state);
1674                 return 0;
1675         }
1676
1677         /*
1678          * start comes from the offset of locked_page.  We have to lock
1679          * pages in order, so we can't process delalloc bytes before
1680          * locked_page
1681          */
1682         if (delalloc_start < *start)
1683                 delalloc_start = *start;
1684
1685         /*
1686          * make sure to limit the number of pages we try to lock down
1687          */
1688         if (delalloc_end + 1 - delalloc_start > max_bytes)
1689                 delalloc_end = delalloc_start + max_bytes - 1;
1690
1691         /* step two, lock all the pages after the page that has start */
1692         ret = lock_delalloc_pages(inode, locked_page,
1693                                   delalloc_start, delalloc_end);
1694         if (ret == -EAGAIN) {
1695                 /* some of the pages are gone, lets avoid looping by
1696                  * shortening the size of the delalloc range we're searching
1697                  */
1698                 free_extent_state(cached_state);
1699                 cached_state = NULL;
1700                 if (!loops) {
1701                         max_bytes = PAGE_CACHE_SIZE;
1702                         loops = 1;
1703                         goto again;
1704                 } else {
1705                         found = 0;
1706                         goto out_failed;
1707                 }
1708         }
1709         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1710
1711         /* step three, lock the state bits for the whole range */
1712         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1713
1714         /* then test to make sure it is all still delalloc */
1715         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1716                              EXTENT_DELALLOC, 1, cached_state);
1717         if (!ret) {
1718                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1719                                      &cached_state, GFP_NOFS);
1720                 __unlock_for_delalloc(inode, locked_page,
1721                               delalloc_start, delalloc_end);
1722                 cond_resched();
1723                 goto again;
1724         }
1725         free_extent_state(cached_state);
1726         *start = delalloc_start;
1727         *end = delalloc_end;
1728 out_failed:
1729         return found;
1730 }
1731
1732 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1733                                  struct page *locked_page,
1734                                  unsigned long clear_bits,
1735                                  unsigned long page_ops)
1736 {
1737         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1738         int ret;
1739         struct page *pages[16];
1740         unsigned long index = start >> PAGE_CACHE_SHIFT;
1741         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1742         unsigned long nr_pages = end_index - index + 1;
1743         int i;
1744
1745         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1746         if (page_ops == 0)
1747                 return 0;
1748
1749         while (nr_pages > 0) {
1750                 ret = find_get_pages_contig(inode->i_mapping, index,
1751                                      min_t(unsigned long,
1752                                      nr_pages, ARRAY_SIZE(pages)), pages);
1753                 for (i = 0; i < ret; i++) {
1754
1755                         if (page_ops & PAGE_SET_PRIVATE2)
1756                                 SetPagePrivate2(pages[i]);
1757
1758                         if (pages[i] == locked_page) {
1759                                 page_cache_release(pages[i]);
1760                                 continue;
1761                         }
1762                         if (page_ops & PAGE_CLEAR_DIRTY)
1763                                 clear_page_dirty_for_io(pages[i]);
1764                         if (page_ops & PAGE_SET_WRITEBACK)
1765                                 set_page_writeback(pages[i]);
1766                         if (page_ops & PAGE_END_WRITEBACK)
1767                                 end_page_writeback(pages[i]);
1768                         if (page_ops & PAGE_UNLOCK)
1769                                 unlock_page(pages[i]);
1770                         page_cache_release(pages[i]);
1771                 }
1772                 nr_pages -= ret;
1773                 index += ret;
1774                 cond_resched();
1775         }
1776         return 0;
1777 }
1778
1779 /*
1780  * count the number of bytes in the tree that have a given bit(s)
1781  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1782  * cached.  The total number found is returned.
1783  */
1784 u64 count_range_bits(struct extent_io_tree *tree,
1785                      u64 *start, u64 search_end, u64 max_bytes,
1786                      unsigned long bits, int contig)
1787 {
1788         struct rb_node *node;
1789         struct extent_state *state;
1790         u64 cur_start = *start;
1791         u64 total_bytes = 0;
1792         u64 last = 0;
1793         int found = 0;
1794
1795         if (WARN_ON(search_end <= cur_start))
1796                 return 0;
1797
1798         spin_lock(&tree->lock);
1799         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1800                 total_bytes = tree->dirty_bytes;
1801                 goto out;
1802         }
1803         /*
1804          * this search will find all the extents that end after
1805          * our range starts.
1806          */
1807         node = tree_search(tree, cur_start);
1808         if (!node)
1809                 goto out;
1810
1811         while (1) {
1812                 state = rb_entry(node, struct extent_state, rb_node);
1813                 if (state->start > search_end)
1814                         break;
1815                 if (contig && found && state->start > last + 1)
1816                         break;
1817                 if (state->end >= cur_start && (state->state & bits) == bits) {
1818                         total_bytes += min(search_end, state->end) + 1 -
1819                                        max(cur_start, state->start);
1820                         if (total_bytes >= max_bytes)
1821                                 break;
1822                         if (!found) {
1823                                 *start = max(cur_start, state->start);
1824                                 found = 1;
1825                         }
1826                         last = state->end;
1827                 } else if (contig && found) {
1828                         break;
1829                 }
1830                 node = rb_next(node);
1831                 if (!node)
1832                         break;
1833         }
1834 out:
1835         spin_unlock(&tree->lock);
1836         return total_bytes;
1837 }
1838
1839 /*
1840  * set the private field for a given byte offset in the tree.  If there isn't
1841  * an extent_state there already, this does nothing.
1842  */
1843 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1844 {
1845         struct rb_node *node;
1846         struct extent_state *state;
1847         int ret = 0;
1848
1849         spin_lock(&tree->lock);
1850         /*
1851          * this search will find all the extents that end after
1852          * our range starts.
1853          */
1854         node = tree_search(tree, start);
1855         if (!node) {
1856                 ret = -ENOENT;
1857                 goto out;
1858         }
1859         state = rb_entry(node, struct extent_state, rb_node);
1860         if (state->start != start) {
1861                 ret = -ENOENT;
1862                 goto out;
1863         }
1864         state->private = private;
1865 out:
1866         spin_unlock(&tree->lock);
1867         return ret;
1868 }
1869
1870 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1871 {
1872         struct rb_node *node;
1873         struct extent_state *state;
1874         int ret = 0;
1875
1876         spin_lock(&tree->lock);
1877         /*
1878          * this search will find all the extents that end after
1879          * our range starts.
1880          */
1881         node = tree_search(tree, start);
1882         if (!node) {
1883                 ret = -ENOENT;
1884                 goto out;
1885         }
1886         state = rb_entry(node, struct extent_state, rb_node);
1887         if (state->start != start) {
1888                 ret = -ENOENT;
1889                 goto out;
1890         }
1891         *private = state->private;
1892 out:
1893         spin_unlock(&tree->lock);
1894         return ret;
1895 }
1896
1897 /*
1898  * searches a range in the state tree for a given mask.
1899  * If 'filled' == 1, this returns 1 only if every extent in the tree
1900  * has the bits set.  Otherwise, 1 is returned if any bit in the
1901  * range is found set.
1902  */
1903 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1904                    unsigned long bits, int filled, struct extent_state *cached)
1905 {
1906         struct extent_state *state = NULL;
1907         struct rb_node *node;
1908         int bitset = 0;
1909
1910         spin_lock(&tree->lock);
1911         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1912             cached->end > start)
1913                 node = &cached->rb_node;
1914         else
1915                 node = tree_search(tree, start);
1916         while (node && start <= end) {
1917                 state = rb_entry(node, struct extent_state, rb_node);
1918
1919                 if (filled && state->start > start) {
1920                         bitset = 0;
1921                         break;
1922                 }
1923
1924                 if (state->start > end)
1925                         break;
1926
1927                 if (state->state & bits) {
1928                         bitset = 1;
1929                         if (!filled)
1930                                 break;
1931                 } else if (filled) {
1932                         bitset = 0;
1933                         break;
1934                 }
1935
1936                 if (state->end == (u64)-1)
1937                         break;
1938
1939                 start = state->end + 1;
1940                 if (start > end)
1941                         break;
1942                 node = rb_next(node);
1943                 if (!node) {
1944                         if (filled)
1945                                 bitset = 0;
1946                         break;
1947                 }
1948         }
1949         spin_unlock(&tree->lock);
1950         return bitset;
1951 }
1952
1953 /*
1954  * helper function to set a given page up to date if all the
1955  * extents in the tree for that page are up to date
1956  */
1957 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1958 {
1959         u64 start = page_offset(page);
1960         u64 end = start + PAGE_CACHE_SIZE - 1;
1961         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1962                 SetPageUptodate(page);
1963 }
1964
1965 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1966 {
1967         int ret;
1968         int err = 0;
1969         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1970
1971         set_state_private(failure_tree, rec->start, 0);
1972         ret = clear_extent_bits(failure_tree, rec->start,
1973                                 rec->start + rec->len - 1,
1974                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1975         if (ret)
1976                 err = ret;
1977
1978         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1979                                 rec->start + rec->len - 1,
1980                                 EXTENT_DAMAGED, GFP_NOFS);
1981         if (ret && !err)
1982                 err = ret;
1983
1984         kfree(rec);
1985         return err;
1986 }
1987
1988 /*
1989  * this bypasses the standard btrfs submit functions deliberately, as
1990  * the standard behavior is to write all copies in a raid setup. here we only
1991  * want to write the one bad copy. so we do the mapping for ourselves and issue
1992  * submit_bio directly.
1993  * to avoid any synchronization issues, wait for the data after writing, which
1994  * actually prevents the read that triggered the error from finishing.
1995  * currently, there can be no more than two copies of every data bit. thus,
1996  * exactly one rewrite is required.
1997  */
1998 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
1999                       struct page *page, unsigned int pg_offset, int mirror_num)
2000 {
2001         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2002         struct bio *bio;
2003         struct btrfs_device *dev;
2004         u64 map_length = 0;
2005         u64 sector;
2006         struct btrfs_bio *bbio = NULL;
2007         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2008         int ret;
2009
2010         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2011         BUG_ON(!mirror_num);
2012
2013         /* we can't repair anything in raid56 yet */
2014         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2015                 return 0;
2016
2017         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2018         if (!bio)
2019                 return -EIO;
2020         bio->bi_iter.bi_size = 0;
2021         map_length = length;
2022
2023         ret = btrfs_map_block(fs_info, WRITE, logical,
2024                               &map_length, &bbio, mirror_num);
2025         if (ret) {
2026                 bio_put(bio);
2027                 return -EIO;
2028         }
2029         BUG_ON(mirror_num != bbio->mirror_num);
2030         sector = bbio->stripes[mirror_num-1].physical >> 9;
2031         bio->bi_iter.bi_sector = sector;
2032         dev = bbio->stripes[mirror_num-1].dev;
2033         kfree(bbio);
2034         if (!dev || !dev->bdev || !dev->writeable) {
2035                 bio_put(bio);
2036                 return -EIO;
2037         }
2038         bio->bi_bdev = dev->bdev;
2039         bio_add_page(bio, page, length, pg_offset);
2040
2041         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2042                 /* try to remap that extent elsewhere? */
2043                 bio_put(bio);
2044                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2045                 return -EIO;
2046         }
2047
2048         printk_ratelimited_in_rcu(KERN_INFO
2049                                   "BTRFS: read error corrected: ino %llu off %llu (dev %s sector %llu)\n",
2050                                   btrfs_ino(inode), start,
2051                                   rcu_str_deref(dev->name), sector);
2052         bio_put(bio);
2053         return 0;
2054 }
2055
2056 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2057                          int mirror_num)
2058 {
2059         u64 start = eb->start;
2060         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2061         int ret = 0;
2062
2063         if (root->fs_info->sb->s_flags & MS_RDONLY)
2064                 return -EROFS;
2065
2066         for (i = 0; i < num_pages; i++) {
2067                 struct page *p = extent_buffer_page(eb, i);
2068
2069                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2070                                         PAGE_CACHE_SIZE, start, p,
2071                                         start - page_offset(p), mirror_num);
2072                 if (ret)
2073                         break;
2074                 start += PAGE_CACHE_SIZE;
2075         }
2076
2077         return ret;
2078 }
2079
2080 /*
2081  * each time an IO finishes, we do a fast check in the IO failure tree
2082  * to see if we need to process or clean up an io_failure_record
2083  */
2084 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2085                      unsigned int pg_offset)
2086 {
2087         u64 private;
2088         u64 private_failure;
2089         struct io_failure_record *failrec;
2090         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2091         struct extent_state *state;
2092         int num_copies;
2093         int ret;
2094
2095         private = 0;
2096         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2097                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2098         if (!ret)
2099                 return 0;
2100
2101         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2102                                 &private_failure);
2103         if (ret)
2104                 return 0;
2105
2106         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2107         BUG_ON(!failrec->this_mirror);
2108
2109         if (failrec->in_validation) {
2110                 /* there was no real error, just free the record */
2111                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2112                          failrec->start);
2113                 goto out;
2114         }
2115         if (fs_info->sb->s_flags & MS_RDONLY)
2116                 goto out;
2117
2118         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2119         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2120                                             failrec->start,
2121                                             EXTENT_LOCKED);
2122         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2123
2124         if (state && state->start <= failrec->start &&
2125             state->end >= failrec->start + failrec->len - 1) {
2126                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2127                                               failrec->len);
2128                 if (num_copies > 1)  {
2129                         repair_io_failure(inode, start, failrec->len,
2130                                           failrec->logical, page,
2131                                           pg_offset, failrec->failed_mirror);
2132                 }
2133         }
2134
2135 out:
2136         free_io_failure(inode, failrec);
2137
2138         return 0;
2139 }
2140
2141 /*
2142  * Can be called when
2143  * - hold extent lock
2144  * - under ordered extent
2145  * - the inode is freeing
2146  */
2147 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2148 {
2149         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2150         struct io_failure_record *failrec;
2151         struct extent_state *state, *next;
2152
2153         if (RB_EMPTY_ROOT(&failure_tree->state))
2154                 return;
2155
2156         spin_lock(&failure_tree->lock);
2157         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2158         while (state) {
2159                 if (state->start > end)
2160                         break;
2161
2162                 ASSERT(state->end <= end);
2163
2164                 next = next_state(state);
2165
2166                 failrec = (struct io_failure_record *)state->private;
2167                 free_extent_state(state);
2168                 kfree(failrec);
2169
2170                 state = next;
2171         }
2172         spin_unlock(&failure_tree->lock);
2173 }
2174
2175 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2176                                 struct io_failure_record **failrec_ret)
2177 {
2178         struct io_failure_record *failrec;
2179         u64 private;
2180         struct extent_map *em;
2181         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2182         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2183         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2184         int ret;
2185         u64 logical;
2186
2187         ret = get_state_private(failure_tree, start, &private);
2188         if (ret) {
2189                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2190                 if (!failrec)
2191                         return -ENOMEM;
2192
2193                 failrec->start = start;
2194                 failrec->len = end - start + 1;
2195                 failrec->this_mirror = 0;
2196                 failrec->bio_flags = 0;
2197                 failrec->in_validation = 0;
2198
2199                 read_lock(&em_tree->lock);
2200                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2201                 if (!em) {
2202                         read_unlock(&em_tree->lock);
2203                         kfree(failrec);
2204                         return -EIO;
2205                 }
2206
2207                 if (em->start > start || em->start + em->len <= start) {
2208                         free_extent_map(em);
2209                         em = NULL;
2210                 }
2211                 read_unlock(&em_tree->lock);
2212                 if (!em) {
2213                         kfree(failrec);
2214                         return -EIO;
2215                 }
2216
2217                 logical = start - em->start;
2218                 logical = em->block_start + logical;
2219                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2220                         logical = em->block_start;
2221                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2222                         extent_set_compress_type(&failrec->bio_flags,
2223                                                  em->compress_type);
2224                 }
2225
2226                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2227                          logical, start, failrec->len);
2228
2229                 failrec->logical = logical;
2230                 free_extent_map(em);
2231
2232                 /* set the bits in the private failure tree */
2233                 ret = set_extent_bits(failure_tree, start, end,
2234                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2235                 if (ret >= 0)
2236                         ret = set_state_private(failure_tree, start,
2237                                                 (u64)(unsigned long)failrec);
2238                 /* set the bits in the inode's tree */
2239                 if (ret >= 0)
2240                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2241                                                 GFP_NOFS);
2242                 if (ret < 0) {
2243                         kfree(failrec);
2244                         return ret;
2245                 }
2246         } else {
2247                 failrec = (struct io_failure_record *)(unsigned long)private;
2248                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2249                          failrec->logical, failrec->start, failrec->len,
2250                          failrec->in_validation);
2251                 /*
2252                  * when data can be on disk more than twice, add to failrec here
2253                  * (e.g. with a list for failed_mirror) to make
2254                  * clean_io_failure() clean all those errors at once.
2255                  */
2256         }
2257
2258         *failrec_ret = failrec;
2259
2260         return 0;
2261 }
2262
2263 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2264                            struct io_failure_record *failrec, int failed_mirror)
2265 {
2266         int num_copies;
2267
2268         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2269                                       failrec->logical, failrec->len);
2270         if (num_copies == 1) {
2271                 /*
2272                  * we only have a single copy of the data, so don't bother with
2273                  * all the retry and error correction code that follows. no
2274                  * matter what the error is, it is very likely to persist.
2275                  */
2276                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2277                          num_copies, failrec->this_mirror, failed_mirror);
2278                 return 0;
2279         }
2280
2281         /*
2282          * there are two premises:
2283          *      a) deliver good data to the caller
2284          *      b) correct the bad sectors on disk
2285          */
2286         if (failed_bio->bi_vcnt > 1) {
2287                 /*
2288                  * to fulfill b), we need to know the exact failing sectors, as
2289                  * we don't want to rewrite any more than the failed ones. thus,
2290                  * we need separate read requests for the failed bio
2291                  *
2292                  * if the following BUG_ON triggers, our validation request got
2293                  * merged. we need separate requests for our algorithm to work.
2294                  */
2295                 BUG_ON(failrec->in_validation);
2296                 failrec->in_validation = 1;
2297                 failrec->this_mirror = failed_mirror;
2298         } else {
2299                 /*
2300                  * we're ready to fulfill a) and b) alongside. get a good copy
2301                  * of the failed sector and if we succeed, we have setup
2302                  * everything for repair_io_failure to do the rest for us.
2303                  */
2304                 if (failrec->in_validation) {
2305                         BUG_ON(failrec->this_mirror != failed_mirror);
2306                         failrec->in_validation = 0;
2307                         failrec->this_mirror = 0;
2308                 }
2309                 failrec->failed_mirror = failed_mirror;
2310                 failrec->this_mirror++;
2311                 if (failrec->this_mirror == failed_mirror)
2312                         failrec->this_mirror++;
2313         }
2314
2315         if (failrec->this_mirror > num_copies) {
2316                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2317                          num_copies, failrec->this_mirror, failed_mirror);
2318                 return 0;
2319         }
2320
2321         return 1;
2322 }
2323
2324
2325 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2326                                     struct io_failure_record *failrec,
2327                                     struct page *page, int pg_offset, int icsum,
2328                                     bio_end_io_t *endio_func, void *data)
2329 {
2330         struct bio *bio;
2331         struct btrfs_io_bio *btrfs_failed_bio;
2332         struct btrfs_io_bio *btrfs_bio;
2333
2334         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2335         if (!bio)
2336                 return NULL;
2337
2338         bio->bi_end_io = endio_func;
2339         bio->bi_iter.bi_sector = failrec->logical >> 9;
2340         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2341         bio->bi_iter.bi_size = 0;
2342         bio->bi_private = data;
2343
2344         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2345         if (btrfs_failed_bio->csum) {
2346                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2347                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2348
2349                 btrfs_bio = btrfs_io_bio(bio);
2350                 btrfs_bio->csum = btrfs_bio->csum_inline;
2351                 icsum *= csum_size;
2352                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2353                        csum_size);
2354         }
2355
2356         bio_add_page(bio, page, failrec->len, pg_offset);
2357
2358         return bio;
2359 }
2360
2361 /*
2362  * this is a generic handler for readpage errors (default
2363  * readpage_io_failed_hook). if other copies exist, read those and write back
2364  * good data to the failed position. does not investigate in remapping the
2365  * failed extent elsewhere, hoping the device will be smart enough to do this as
2366  * needed
2367  */
2368
2369 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2370                               struct page *page, u64 start, u64 end,
2371                               int failed_mirror)
2372 {
2373         struct io_failure_record *failrec;
2374         struct inode *inode = page->mapping->host;
2375         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2376         struct bio *bio;
2377         int read_mode;
2378         int ret;
2379
2380         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2381
2382         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2383         if (ret)
2384                 return ret;
2385
2386         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2387         if (!ret) {
2388                 free_io_failure(inode, failrec);
2389                 return -EIO;
2390         }
2391
2392         if (failed_bio->bi_vcnt > 1)
2393                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2394         else
2395                 read_mode = READ_SYNC;
2396
2397         phy_offset >>= inode->i_sb->s_blocksize_bits;
2398         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2399                                       start - page_offset(page),
2400                                       (int)phy_offset, failed_bio->bi_end_io,
2401                                       NULL);
2402         if (!bio) {
2403                 free_io_failure(inode, failrec);
2404                 return -EIO;
2405         }
2406
2407         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2408                  read_mode, failrec->this_mirror, failrec->in_validation);
2409
2410         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2411                                          failrec->this_mirror,
2412                                          failrec->bio_flags, 0);
2413         if (ret) {
2414                 free_io_failure(inode, failrec);
2415                 bio_put(bio);
2416         }
2417
2418         return ret;
2419 }
2420
2421 /* lots and lots of room for performance fixes in the end_bio funcs */
2422
2423 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2424 {
2425         int uptodate = (err == 0);
2426         struct extent_io_tree *tree;
2427         int ret = 0;
2428
2429         tree = &BTRFS_I(page->mapping->host)->io_tree;
2430
2431         if (tree->ops && tree->ops->writepage_end_io_hook) {
2432                 ret = tree->ops->writepage_end_io_hook(page, start,
2433                                                end, NULL, uptodate);
2434                 if (ret)
2435                         uptodate = 0;
2436         }
2437
2438         if (!uptodate) {
2439                 ClearPageUptodate(page);
2440                 SetPageError(page);
2441                 ret = ret < 0 ? ret : -EIO;
2442                 mapping_set_error(page->mapping, ret);
2443         }
2444         return 0;
2445 }
2446
2447 /*
2448  * after a writepage IO is done, we need to:
2449  * clear the uptodate bits on error
2450  * clear the writeback bits in the extent tree for this IO
2451  * end_page_writeback if the page has no more pending IO
2452  *
2453  * Scheduling is not allowed, so the extent state tree is expected
2454  * to have one and only one object corresponding to this IO.
2455  */
2456 static void end_bio_extent_writepage(struct bio *bio, int err)
2457 {
2458         struct bio_vec *bvec;
2459         u64 start;
2460         u64 end;
2461         int i;
2462
2463         bio_for_each_segment_all(bvec, bio, i) {
2464                 struct page *page = bvec->bv_page;
2465
2466                 /* We always issue full-page reads, but if some block
2467                  * in a page fails to read, blk_update_request() will
2468                  * advance bv_offset and adjust bv_len to compensate.
2469                  * Print a warning for nonzero offsets, and an error
2470                  * if they don't add up to a full page.  */
2471                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2472                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2473                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2474                                    "partial page write in btrfs with offset %u and length %u",
2475                                         bvec->bv_offset, bvec->bv_len);
2476                         else
2477                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2478                                    "incomplete page write in btrfs with offset %u and "
2479                                    "length %u",
2480                                         bvec->bv_offset, bvec->bv_len);
2481                 }
2482
2483                 start = page_offset(page);
2484                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2485
2486                 if (end_extent_writepage(page, err, start, end))
2487                         continue;
2488
2489                 end_page_writeback(page);
2490         }
2491
2492         bio_put(bio);
2493 }
2494
2495 static void
2496 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2497                               int uptodate)
2498 {
2499         struct extent_state *cached = NULL;
2500         u64 end = start + len - 1;
2501
2502         if (uptodate && tree->track_uptodate)
2503                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2504         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2505 }
2506
2507 /*
2508  * after a readpage IO is done, we need to:
2509  * clear the uptodate bits on error
2510  * set the uptodate bits if things worked
2511  * set the page up to date if all extents in the tree are uptodate
2512  * clear the lock bit in the extent tree
2513  * unlock the page if there are no other extents locked for it
2514  *
2515  * Scheduling is not allowed, so the extent state tree is expected
2516  * to have one and only one object corresponding to this IO.
2517  */
2518 static void end_bio_extent_readpage(struct bio *bio, int err)
2519 {
2520         struct bio_vec *bvec;
2521         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2522         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2523         struct extent_io_tree *tree;
2524         u64 offset = 0;
2525         u64 start;
2526         u64 end;
2527         u64 len;
2528         u64 extent_start = 0;
2529         u64 extent_len = 0;
2530         int mirror;
2531         int ret;
2532         int i;
2533
2534         if (err)
2535                 uptodate = 0;
2536
2537         bio_for_each_segment_all(bvec, bio, i) {
2538                 struct page *page = bvec->bv_page;
2539                 struct inode *inode = page->mapping->host;
2540
2541                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2542                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector, err,
2543                          io_bio->mirror_num);
2544                 tree = &BTRFS_I(inode)->io_tree;
2545
2546                 /* We always issue full-page reads, but if some block
2547                  * in a page fails to read, blk_update_request() will
2548                  * advance bv_offset and adjust bv_len to compensate.
2549                  * Print a warning for nonzero offsets, and an error
2550                  * if they don't add up to a full page.  */
2551                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2552                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2553                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2554                                    "partial page read in btrfs with offset %u and length %u",
2555                                         bvec->bv_offset, bvec->bv_len);
2556                         else
2557                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2558                                    "incomplete page read in btrfs with offset %u and "
2559                                    "length %u",
2560                                         bvec->bv_offset, bvec->bv_len);
2561                 }
2562
2563                 start = page_offset(page);
2564                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2565                 len = bvec->bv_len;
2566
2567                 mirror = io_bio->mirror_num;
2568                 if (likely(uptodate && tree->ops &&
2569                            tree->ops->readpage_end_io_hook)) {
2570                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2571                                                               page, start, end,
2572                                                               mirror);
2573                         if (ret)
2574                                 uptodate = 0;
2575                         else
2576                                 clean_io_failure(inode, start, page, 0);
2577                 }
2578
2579                 if (likely(uptodate))
2580                         goto readpage_ok;
2581
2582                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2583                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2584                         if (!ret && !err &&
2585                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2586                                 uptodate = 1;
2587                 } else {
2588                         /*
2589                          * The generic bio_readpage_error handles errors the
2590                          * following way: If possible, new read requests are
2591                          * created and submitted and will end up in
2592                          * end_bio_extent_readpage as well (if we're lucky, not
2593                          * in the !uptodate case). In that case it returns 0 and
2594                          * we just go on with the next page in our bio. If it
2595                          * can't handle the error it will return -EIO and we
2596                          * remain responsible for that page.
2597                          */
2598                         ret = bio_readpage_error(bio, offset, page, start, end,
2599                                                  mirror);
2600                         if (ret == 0) {
2601                                 uptodate =
2602                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2603                                 if (err)
2604                                         uptodate = 0;
2605                                 offset += len;
2606                                 continue;
2607                         }
2608                 }
2609 readpage_ok:
2610                 if (likely(uptodate)) {
2611                         loff_t i_size = i_size_read(inode);
2612                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2613                         unsigned off;
2614
2615                         /* Zero out the end if this page straddles i_size */
2616                         off = i_size & (PAGE_CACHE_SIZE-1);
2617                         if (page->index == end_index && off)
2618                                 zero_user_segment(page, off, PAGE_CACHE_SIZE);
2619                         SetPageUptodate(page);
2620                 } else {
2621                         ClearPageUptodate(page);
2622                         SetPageError(page);
2623                 }
2624                 unlock_page(page);
2625                 offset += len;
2626
2627                 if (unlikely(!uptodate)) {
2628                         if (extent_len) {
2629                                 endio_readpage_release_extent(tree,
2630                                                               extent_start,
2631                                                               extent_len, 1);
2632                                 extent_start = 0;
2633                                 extent_len = 0;
2634                         }
2635                         endio_readpage_release_extent(tree, start,
2636                                                       end - start + 1, 0);
2637                 } else if (!extent_len) {
2638                         extent_start = start;
2639                         extent_len = end + 1 - start;
2640                 } else if (extent_start + extent_len == start) {
2641                         extent_len += end + 1 - start;
2642                 } else {
2643                         endio_readpage_release_extent(tree, extent_start,
2644                                                       extent_len, uptodate);
2645                         extent_start = start;
2646                         extent_len = end + 1 - start;
2647                 }
2648         }
2649
2650         if (extent_len)
2651                 endio_readpage_release_extent(tree, extent_start, extent_len,
2652                                               uptodate);
2653         if (io_bio->end_io)
2654                 io_bio->end_io(io_bio, err);
2655         bio_put(bio);
2656 }
2657
2658 /*
2659  * this allocates from the btrfs_bioset.  We're returning a bio right now
2660  * but you can call btrfs_io_bio for the appropriate container_of magic
2661  */
2662 struct bio *
2663 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2664                 gfp_t gfp_flags)
2665 {
2666         struct btrfs_io_bio *btrfs_bio;
2667         struct bio *bio;
2668
2669         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2670
2671         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2672                 while (!bio && (nr_vecs /= 2)) {
2673                         bio = bio_alloc_bioset(gfp_flags,
2674                                                nr_vecs, btrfs_bioset);
2675                 }
2676         }
2677
2678         if (bio) {
2679                 bio->bi_bdev = bdev;
2680                 bio->bi_iter.bi_sector = first_sector;
2681                 btrfs_bio = btrfs_io_bio(bio);
2682                 btrfs_bio->csum = NULL;
2683                 btrfs_bio->csum_allocated = NULL;
2684                 btrfs_bio->end_io = NULL;
2685         }
2686         return bio;
2687 }
2688
2689 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2690 {
2691         struct btrfs_io_bio *btrfs_bio;
2692         struct bio *new;
2693
2694         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2695         if (new) {
2696                 btrfs_bio = btrfs_io_bio(new);
2697                 btrfs_bio->csum = NULL;
2698                 btrfs_bio->csum_allocated = NULL;
2699                 btrfs_bio->end_io = NULL;
2700         }
2701         return new;
2702 }
2703
2704 /* this also allocates from the btrfs_bioset */
2705 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2706 {
2707         struct btrfs_io_bio *btrfs_bio;
2708         struct bio *bio;
2709
2710         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2711         if (bio) {
2712                 btrfs_bio = btrfs_io_bio(bio);
2713                 btrfs_bio->csum = NULL;
2714                 btrfs_bio->csum_allocated = NULL;
2715                 btrfs_bio->end_io = NULL;
2716         }
2717         return bio;
2718 }
2719
2720
2721 static int __must_check submit_one_bio(int rw, struct bio *bio,
2722                                        int mirror_num, unsigned long bio_flags)
2723 {
2724         int ret = 0;
2725         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2726         struct page *page = bvec->bv_page;
2727         struct extent_io_tree *tree = bio->bi_private;
2728         u64 start;
2729
2730         start = page_offset(page) + bvec->bv_offset;
2731
2732         bio->bi_private = NULL;
2733
2734         bio_get(bio);
2735
2736         if (tree->ops && tree->ops->submit_bio_hook)
2737                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2738                                            mirror_num, bio_flags, start);
2739         else
2740                 btrfsic_submit_bio(rw, bio);
2741
2742         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2743                 ret = -EOPNOTSUPP;
2744         bio_put(bio);
2745         return ret;
2746 }
2747
2748 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2749                      unsigned long offset, size_t size, struct bio *bio,
2750                      unsigned long bio_flags)
2751 {
2752         int ret = 0;
2753         if (tree->ops && tree->ops->merge_bio_hook)
2754                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2755                                                 bio_flags);
2756         BUG_ON(ret < 0);
2757         return ret;
2758
2759 }
2760
2761 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2762                               struct page *page, sector_t sector,
2763                               size_t size, unsigned long offset,
2764                               struct block_device *bdev,
2765                               struct bio **bio_ret,
2766                               unsigned long max_pages,
2767                               bio_end_io_t end_io_func,
2768                               int mirror_num,
2769                               unsigned long prev_bio_flags,
2770                               unsigned long bio_flags)
2771 {
2772         int ret = 0;
2773         struct bio *bio;
2774         int nr;
2775         int contig = 0;
2776         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2777         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2778         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2779
2780         if (bio_ret && *bio_ret) {
2781                 bio = *bio_ret;
2782                 if (old_compressed)
2783                         contig = bio->bi_iter.bi_sector == sector;
2784                 else
2785                         contig = bio_end_sector(bio) == sector;
2786
2787                 if (prev_bio_flags != bio_flags || !contig ||
2788                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2789                     bio_add_page(bio, page, page_size, offset) < page_size) {
2790                         ret = submit_one_bio(rw, bio, mirror_num,
2791                                              prev_bio_flags);
2792                         if (ret < 0)
2793                                 return ret;
2794                         bio = NULL;
2795                 } else {
2796                         return 0;
2797                 }
2798         }
2799         if (this_compressed)
2800                 nr = BIO_MAX_PAGES;
2801         else
2802                 nr = bio_get_nr_vecs(bdev);
2803
2804         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2805         if (!bio)
2806                 return -ENOMEM;
2807
2808         bio_add_page(bio, page, page_size, offset);
2809         bio->bi_end_io = end_io_func;
2810         bio->bi_private = tree;
2811
2812         if (bio_ret)
2813                 *bio_ret = bio;
2814         else
2815                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2816
2817         return ret;
2818 }
2819
2820 static void attach_extent_buffer_page(struct extent_buffer *eb,
2821                                       struct page *page)
2822 {
2823         if (!PagePrivate(page)) {
2824                 SetPagePrivate(page);
2825                 page_cache_get(page);
2826                 set_page_private(page, (unsigned long)eb);
2827         } else {
2828                 WARN_ON(page->private != (unsigned long)eb);
2829         }
2830 }
2831
2832 void set_page_extent_mapped(struct page *page)
2833 {
2834         if (!PagePrivate(page)) {
2835                 SetPagePrivate(page);
2836                 page_cache_get(page);
2837                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2838         }
2839 }
2840
2841 static struct extent_map *
2842 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2843                  u64 start, u64 len, get_extent_t *get_extent,
2844                  struct extent_map **em_cached)
2845 {
2846         struct extent_map *em;
2847
2848         if (em_cached && *em_cached) {
2849                 em = *em_cached;
2850                 if (extent_map_in_tree(em) && start >= em->start &&
2851                     start < extent_map_end(em)) {
2852                         atomic_inc(&em->refs);
2853                         return em;
2854                 }
2855
2856                 free_extent_map(em);
2857                 *em_cached = NULL;
2858         }
2859
2860         em = get_extent(inode, page, pg_offset, start, len, 0);
2861         if (em_cached && !IS_ERR_OR_NULL(em)) {
2862                 BUG_ON(*em_cached);
2863                 atomic_inc(&em->refs);
2864                 *em_cached = em;
2865         }
2866         return em;
2867 }
2868 /*
2869  * basic readpage implementation.  Locked extent state structs are inserted
2870  * into the tree that are removed when the IO is done (by the end_io
2871  * handlers)
2872  * XXX JDM: This needs looking at to ensure proper page locking
2873  */
2874 static int __do_readpage(struct extent_io_tree *tree,
2875                          struct page *page,
2876                          get_extent_t *get_extent,
2877                          struct extent_map **em_cached,
2878                          struct bio **bio, int mirror_num,
2879                          unsigned long *bio_flags, int rw)
2880 {
2881         struct inode *inode = page->mapping->host;
2882         u64 start = page_offset(page);
2883         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2884         u64 end;
2885         u64 cur = start;
2886         u64 extent_offset;
2887         u64 last_byte = i_size_read(inode);
2888         u64 block_start;
2889         u64 cur_end;
2890         sector_t sector;
2891         struct extent_map *em;
2892         struct block_device *bdev;
2893         int ret;
2894         int nr = 0;
2895         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2896         size_t pg_offset = 0;
2897         size_t iosize;
2898         size_t disk_io_size;
2899         size_t blocksize = inode->i_sb->s_blocksize;
2900         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2901
2902         set_page_extent_mapped(page);
2903
2904         end = page_end;
2905         if (!PageUptodate(page)) {
2906                 if (cleancache_get_page(page) == 0) {
2907                         BUG_ON(blocksize != PAGE_SIZE);
2908                         unlock_extent(tree, start, end);
2909                         goto out;
2910                 }
2911         }
2912
2913         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2914                 char *userpage;
2915                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2916
2917                 if (zero_offset) {
2918                         iosize = PAGE_CACHE_SIZE - zero_offset;
2919                         userpage = kmap_atomic(page);
2920                         memset(userpage + zero_offset, 0, iosize);
2921                         flush_dcache_page(page);
2922                         kunmap_atomic(userpage);
2923                 }
2924         }
2925         while (cur <= end) {
2926                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2927
2928                 if (cur >= last_byte) {
2929                         char *userpage;
2930                         struct extent_state *cached = NULL;
2931
2932                         iosize = PAGE_CACHE_SIZE - pg_offset;
2933                         userpage = kmap_atomic(page);
2934                         memset(userpage + pg_offset, 0, iosize);
2935                         flush_dcache_page(page);
2936                         kunmap_atomic(userpage);
2937                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2938                                             &cached, GFP_NOFS);
2939                         if (!parent_locked)
2940                                 unlock_extent_cached(tree, cur,
2941                                                      cur + iosize - 1,
2942                                                      &cached, GFP_NOFS);
2943                         break;
2944                 }
2945                 em = __get_extent_map(inode, page, pg_offset, cur,
2946                                       end - cur + 1, get_extent, em_cached);
2947                 if (IS_ERR_OR_NULL(em)) {
2948                         SetPageError(page);
2949                         if (!parent_locked)
2950                                 unlock_extent(tree, cur, end);
2951                         break;
2952                 }
2953                 extent_offset = cur - em->start;
2954                 BUG_ON(extent_map_end(em) <= cur);
2955                 BUG_ON(end < cur);
2956
2957                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2958                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2959                         extent_set_compress_type(&this_bio_flag,
2960                                                  em->compress_type);
2961                 }
2962
2963                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2964                 cur_end = min(extent_map_end(em) - 1, end);
2965                 iosize = ALIGN(iosize, blocksize);
2966                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2967                         disk_io_size = em->block_len;
2968                         sector = em->block_start >> 9;
2969                 } else {
2970                         sector = (em->block_start + extent_offset) >> 9;
2971                         disk_io_size = iosize;
2972                 }
2973                 bdev = em->bdev;
2974                 block_start = em->block_start;
2975                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2976                         block_start = EXTENT_MAP_HOLE;
2977                 free_extent_map(em);
2978                 em = NULL;
2979
2980                 /* we've found a hole, just zero and go on */
2981                 if (block_start == EXTENT_MAP_HOLE) {
2982                         char *userpage;
2983                         struct extent_state *cached = NULL;
2984
2985                         userpage = kmap_atomic(page);
2986                         memset(userpage + pg_offset, 0, iosize);
2987                         flush_dcache_page(page);
2988                         kunmap_atomic(userpage);
2989
2990                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2991                                             &cached, GFP_NOFS);
2992                         unlock_extent_cached(tree, cur, cur + iosize - 1,
2993                                              &cached, GFP_NOFS);
2994                         cur = cur + iosize;
2995                         pg_offset += iosize;
2996                         continue;
2997                 }
2998                 /* the get_extent function already copied into the page */
2999                 if (test_range_bit(tree, cur, cur_end,
3000                                    EXTENT_UPTODATE, 1, NULL)) {
3001                         check_page_uptodate(tree, page);
3002                         if (!parent_locked)
3003                                 unlock_extent(tree, cur, cur + iosize - 1);
3004                         cur = cur + iosize;
3005                         pg_offset += iosize;
3006                         continue;
3007                 }
3008                 /* we have an inline extent but it didn't get marked up
3009                  * to date.  Error out
3010                  */
3011                 if (block_start == EXTENT_MAP_INLINE) {
3012                         SetPageError(page);
3013                         if (!parent_locked)
3014                                 unlock_extent(tree, cur, cur + iosize - 1);
3015                         cur = cur + iosize;
3016                         pg_offset += iosize;
3017                         continue;
3018                 }
3019
3020                 pnr -= page->index;
3021                 ret = submit_extent_page(rw, tree, page,
3022                                          sector, disk_io_size, pg_offset,
3023                                          bdev, bio, pnr,
3024                                          end_bio_extent_readpage, mirror_num,
3025                                          *bio_flags,
3026                                          this_bio_flag);
3027                 if (!ret) {
3028                         nr++;
3029                         *bio_flags = this_bio_flag;
3030                 } else {
3031                         SetPageError(page);
3032                         if (!parent_locked)
3033                                 unlock_extent(tree, cur, cur + iosize - 1);
3034                 }
3035                 cur = cur + iosize;
3036                 pg_offset += iosize;
3037         }
3038 out:
3039         if (!nr) {
3040                 if (!PageError(page))
3041                         SetPageUptodate(page);
3042                 unlock_page(page);
3043         }
3044         return 0;
3045 }
3046
3047 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3048                                              struct page *pages[], int nr_pages,
3049                                              u64 start, u64 end,
3050                                              get_extent_t *get_extent,
3051                                              struct extent_map **em_cached,
3052                                              struct bio **bio, int mirror_num,
3053                                              unsigned long *bio_flags, int rw)
3054 {
3055         struct inode *inode;
3056         struct btrfs_ordered_extent *ordered;
3057         int index;
3058
3059         inode = pages[0]->mapping->host;
3060         while (1) {
3061                 lock_extent(tree, start, end);
3062                 ordered = btrfs_lookup_ordered_range(inode, start,
3063                                                      end - start + 1);
3064                 if (!ordered)
3065                         break;
3066                 unlock_extent(tree, start, end);
3067                 btrfs_start_ordered_extent(inode, ordered, 1);
3068                 btrfs_put_ordered_extent(ordered);
3069         }
3070
3071         for (index = 0; index < nr_pages; index++) {
3072                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3073                               mirror_num, bio_flags, rw);
3074                 page_cache_release(pages[index]);
3075         }
3076 }
3077
3078 static void __extent_readpages(struct extent_io_tree *tree,
3079                                struct page *pages[],
3080                                int nr_pages, get_extent_t *get_extent,
3081                                struct extent_map **em_cached,
3082                                struct bio **bio, int mirror_num,
3083                                unsigned long *bio_flags, int rw)
3084 {
3085         u64 start = 0;
3086         u64 end = 0;
3087         u64 page_start;
3088         int index;
3089         int first_index = 0;
3090
3091         for (index = 0; index < nr_pages; index++) {
3092                 page_start = page_offset(pages[index]);
3093                 if (!end) {
3094                         start = page_start;
3095                         end = start + PAGE_CACHE_SIZE - 1;
3096                         first_index = index;
3097                 } else if (end + 1 == page_start) {
3098                         end += PAGE_CACHE_SIZE;
3099                 } else {
3100                         __do_contiguous_readpages(tree, &pages[first_index],
3101                                                   index - first_index, start,
3102                                                   end, get_extent, em_cached,
3103                                                   bio, mirror_num, bio_flags,
3104                                                   rw);
3105                         start = page_start;
3106                         end = start + PAGE_CACHE_SIZE - 1;
3107                         first_index = index;
3108                 }
3109         }
3110
3111         if (end)
3112                 __do_contiguous_readpages(tree, &pages[first_index],
3113                                           index - first_index, start,
3114                                           end, get_extent, em_cached, bio,
3115                                           mirror_num, bio_flags, rw);
3116 }
3117
3118 static int __extent_read_full_page(struct extent_io_tree *tree,
3119                                    struct page *page,
3120                                    get_extent_t *get_extent,
3121                                    struct bio **bio, int mirror_num,
3122                                    unsigned long *bio_flags, int rw)
3123 {
3124         struct inode *inode = page->mapping->host;
3125         struct btrfs_ordered_extent *ordered;
3126         u64 start = page_offset(page);
3127         u64 end = start + PAGE_CACHE_SIZE - 1;
3128         int ret;
3129
3130         while (1) {
3131                 lock_extent(tree, start, end);
3132                 ordered = btrfs_lookup_ordered_extent(inode, start);
3133                 if (!ordered)
3134                         break;
3135                 unlock_extent(tree, start, end);
3136                 btrfs_start_ordered_extent(inode, ordered, 1);
3137                 btrfs_put_ordered_extent(ordered);
3138         }
3139
3140         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3141                             bio_flags, rw);
3142         return ret;
3143 }
3144
3145 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3146                             get_extent_t *get_extent, int mirror_num)
3147 {
3148         struct bio *bio = NULL;
3149         unsigned long bio_flags = 0;
3150         int ret;
3151
3152         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3153                                       &bio_flags, READ);
3154         if (bio)
3155                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3156         return ret;
3157 }
3158
3159 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3160                                  get_extent_t *get_extent, int mirror_num)
3161 {
3162         struct bio *bio = NULL;
3163         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3164         int ret;
3165
3166         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3167                                       &bio_flags, READ);
3168         if (bio)
3169                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3170         return ret;
3171 }
3172
3173 static noinline void update_nr_written(struct page *page,
3174                                       struct writeback_control *wbc,
3175                                       unsigned long nr_written)
3176 {
3177         wbc->nr_to_write -= nr_written;
3178         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3179             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3180                 page->mapping->writeback_index = page->index + nr_written;
3181 }
3182
3183 /*
3184  * helper for __extent_writepage, doing all of the delayed allocation setup.
3185  *
3186  * This returns 1 if our fill_delalloc function did all the work required
3187  * to write the page (copy into inline extent).  In this case the IO has
3188  * been started and the page is already unlocked.
3189  *
3190  * This returns 0 if all went well (page still locked)
3191  * This returns < 0 if there were errors (page still locked)
3192  */
3193 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3194                               struct page *page, struct writeback_control *wbc,
3195                               struct extent_page_data *epd,
3196                               u64 delalloc_start,
3197                               unsigned long *nr_written)
3198 {
3199         struct extent_io_tree *tree = epd->tree;
3200         u64 page_end = delalloc_start + PAGE_CACHE_SIZE - 1;
3201         u64 nr_delalloc;
3202         u64 delalloc_to_write = 0;
3203         u64 delalloc_end = 0;
3204         int ret;
3205         int page_started = 0;
3206
3207         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3208                 return 0;
3209
3210         while (delalloc_end < page_end) {
3211                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3212                                                page,
3213                                                &delalloc_start,
3214                                                &delalloc_end,
3215                                                128 * 1024 * 1024);
3216                 if (nr_delalloc == 0) {
3217                         delalloc_start = delalloc_end + 1;
3218                         continue;
3219                 }
3220                 ret = tree->ops->fill_delalloc(inode, page,
3221                                                delalloc_start,
3222                                                delalloc_end,
3223                                                &page_started,
3224                                                nr_written);
3225                 /* File system has been set read-only */
3226                 if (ret) {
3227                         SetPageError(page);
3228                         /* fill_delalloc should be return < 0 for error
3229                          * but just in case, we use > 0 here meaning the
3230                          * IO is started, so we don't want to return > 0
3231                          * unless things are going well.
3232                          */
3233                         ret = ret < 0 ? ret : -EIO;
3234                         goto done;
3235                 }
3236                 /*
3237                  * delalloc_end is already one less than the total
3238                  * length, so we don't subtract one from
3239                  * PAGE_CACHE_SIZE
3240                  */
3241                 delalloc_to_write += (delalloc_end - delalloc_start +
3242                                       PAGE_CACHE_SIZE) >>
3243                                       PAGE_CACHE_SHIFT;
3244                 delalloc_start = delalloc_end + 1;
3245         }
3246         if (wbc->nr_to_write < delalloc_to_write) {
3247                 int thresh = 8192;
3248
3249                 if (delalloc_to_write < thresh * 2)
3250                         thresh = delalloc_to_write;
3251                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3252                                          thresh);
3253         }
3254
3255         /* did the fill delalloc function already unlock and start
3256          * the IO?
3257          */
3258         if (page_started) {
3259                 /*
3260                  * we've unlocked the page, so we can't update
3261                  * the mapping's writeback index, just update
3262                  * nr_to_write.
3263                  */
3264                 wbc->nr_to_write -= *nr_written;
3265                 return 1;
3266         }
3267
3268         ret = 0;
3269
3270 done:
3271         return ret;
3272 }
3273
3274 /*
3275  * helper for __extent_writepage.  This calls the writepage start hooks,
3276  * and does the loop to map the page into extents and bios.
3277  *
3278  * We return 1 if the IO is started and the page is unlocked,
3279  * 0 if all went well (page still locked)
3280  * < 0 if there were errors (page still locked)
3281  */
3282 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3283                                  struct page *page,
3284                                  struct writeback_control *wbc,
3285                                  struct extent_page_data *epd,
3286                                  loff_t i_size,
3287                                  unsigned long nr_written,
3288                                  int write_flags, int *nr_ret)
3289 {
3290         struct extent_io_tree *tree = epd->tree;
3291         u64 start = page_offset(page);
3292         u64 page_end = start + PAGE_CACHE_SIZE - 1;
3293         u64 end;
3294         u64 cur = start;
3295         u64 extent_offset;
3296         u64 block_start;
3297         u64 iosize;
3298         sector_t sector;
3299         struct extent_state *cached_state = NULL;
3300         struct