Merge branch 'for-linus-4.3' of git://git.kernel.org/pub/scm/linux/kernel/git/mason...
[muen/linux.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 %u 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 *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 *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 *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 *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 *bits, int wake)
515 {
516         struct extent_state *next;
517         unsigned 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 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                 /*
599                  * Don't care for allocation failure here because we might end
600                  * up not needing the pre-allocated extent state at all, which
601                  * is the case if we only have in the tree extent states that
602                  * cover our input range and don't cover too any other range.
603                  * If we end up needing a new extent state we allocate it later.
604                  */
605                 prealloc = alloc_extent_state(mask);
606         }
607
608         spin_lock(&tree->lock);
609         if (cached_state) {
610                 cached = *cached_state;
611
612                 if (clear) {
613                         *cached_state = NULL;
614                         cached_state = NULL;
615                 }
616
617                 if (cached && extent_state_in_tree(cached) &&
618                     cached->start <= start && cached->end > start) {
619                         if (clear)
620                                 atomic_dec(&cached->refs);
621                         state = cached;
622                         goto hit_next;
623                 }
624                 if (clear)
625                         free_extent_state(cached);
626         }
627         /*
628          * this search will find the extents that end after
629          * our range starts
630          */
631         node = tree_search(tree, start);
632         if (!node)
633                 goto out;
634         state = rb_entry(node, struct extent_state, rb_node);
635 hit_next:
636         if (state->start > end)
637                 goto out;
638         WARN_ON(state->end < start);
639         last_end = state->end;
640
641         /* the state doesn't have the wanted bits, go ahead */
642         if (!(state->state & bits)) {
643                 state = next_state(state);
644                 goto next;
645         }
646
647         /*
648          *     | ---- desired range ---- |
649          *  | state | or
650          *  | ------------- state -------------- |
651          *
652          * We need to split the extent we found, and may flip
653          * bits on second half.
654          *
655          * If the extent we found extends past our range, we
656          * just split and search again.  It'll get split again
657          * the next time though.
658          *
659          * If the extent we found is inside our range, we clear
660          * the desired bit on it.
661          */
662
663         if (state->start < start) {
664                 prealloc = alloc_extent_state_atomic(prealloc);
665                 BUG_ON(!prealloc);
666                 err = split_state(tree, state, prealloc, start);
667                 if (err)
668                         extent_io_tree_panic(tree, err);
669
670                 prealloc = NULL;
671                 if (err)
672                         goto out;
673                 if (state->end <= end) {
674                         state = clear_state_bit(tree, state, &bits, wake);
675                         goto next;
676                 }
677                 goto search_again;
678         }
679         /*
680          * | ---- desired range ---- |
681          *                        | state |
682          * We need to split the extent, and clear the bit
683          * on the first half
684          */
685         if (state->start <= end && state->end > end) {
686                 prealloc = alloc_extent_state_atomic(prealloc);
687                 BUG_ON(!prealloc);
688                 err = split_state(tree, state, prealloc, end + 1);
689                 if (err)
690                         extent_io_tree_panic(tree, err);
691
692                 if (wake)
693                         wake_up(&state->wq);
694
695                 clear_state_bit(tree, prealloc, &bits, wake);
696
697                 prealloc = NULL;
698                 goto out;
699         }
700
701         state = clear_state_bit(tree, state, &bits, wake);
702 next:
703         if (last_end == (u64)-1)
704                 goto out;
705         start = last_end + 1;
706         if (start <= end && state && !need_resched())
707                 goto hit_next;
708         goto search_again;
709
710 out:
711         spin_unlock(&tree->lock);
712         if (prealloc)
713                 free_extent_state(prealloc);
714
715         return 0;
716
717 search_again:
718         if (start > end)
719                 goto out;
720         spin_unlock(&tree->lock);
721         if (mask & __GFP_WAIT)
722                 cond_resched();
723         goto again;
724 }
725
726 static void wait_on_state(struct extent_io_tree *tree,
727                           struct extent_state *state)
728                 __releases(tree->lock)
729                 __acquires(tree->lock)
730 {
731         DEFINE_WAIT(wait);
732         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
733         spin_unlock(&tree->lock);
734         schedule();
735         spin_lock(&tree->lock);
736         finish_wait(&state->wq, &wait);
737 }
738
739 /*
740  * waits for one or more bits to clear on a range in the state tree.
741  * The range [start, end] is inclusive.
742  * The tree lock is taken by this function
743  */
744 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
745                             unsigned long bits)
746 {
747         struct extent_state *state;
748         struct rb_node *node;
749
750         btrfs_debug_check_extent_io_range(tree, start, end);
751
752         spin_lock(&tree->lock);
753 again:
754         while (1) {
755                 /*
756                  * this search will find all the extents that end after
757                  * our range starts
758                  */
759                 node = tree_search(tree, start);
760 process_node:
761                 if (!node)
762                         break;
763
764                 state = rb_entry(node, struct extent_state, rb_node);
765
766                 if (state->start > end)
767                         goto out;
768
769                 if (state->state & bits) {
770                         start = state->start;
771                         atomic_inc(&state->refs);
772                         wait_on_state(tree, state);
773                         free_extent_state(state);
774                         goto again;
775                 }
776                 start = state->end + 1;
777
778                 if (start > end)
779                         break;
780
781                 if (!cond_resched_lock(&tree->lock)) {
782                         node = rb_next(node);
783                         goto process_node;
784                 }
785         }
786 out:
787         spin_unlock(&tree->lock);
788 }
789
790 static void set_state_bits(struct extent_io_tree *tree,
791                            struct extent_state *state,
792                            unsigned *bits)
793 {
794         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
795
796         set_state_cb(tree, state, bits);
797         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
798                 u64 range = state->end - state->start + 1;
799                 tree->dirty_bytes += range;
800         }
801         state->state |= bits_to_set;
802 }
803
804 static void cache_state_if_flags(struct extent_state *state,
805                                  struct extent_state **cached_ptr,
806                                  unsigned flags)
807 {
808         if (cached_ptr && !(*cached_ptr)) {
809                 if (!flags || (state->state & flags)) {
810                         *cached_ptr = state;
811                         atomic_inc(&state->refs);
812                 }
813         }
814 }
815
816 static void cache_state(struct extent_state *state,
817                         struct extent_state **cached_ptr)
818 {
819         return cache_state_if_flags(state, cached_ptr,
820                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
821 }
822
823 /*
824  * set some bits on a range in the tree.  This may require allocations or
825  * sleeping, so the gfp mask is used to indicate what is allowed.
826  *
827  * If any of the exclusive bits are set, this will fail with -EEXIST if some
828  * part of the range already has the desired bits set.  The start of the
829  * existing range is returned in failed_start in this case.
830  *
831  * [start, end] is inclusive This takes the tree lock.
832  */
833
834 static int __must_check
835 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
836                  unsigned bits, unsigned exclusive_bits,
837                  u64 *failed_start, struct extent_state **cached_state,
838                  gfp_t mask)
839 {
840         struct extent_state *state;
841         struct extent_state *prealloc = NULL;
842         struct rb_node *node;
843         struct rb_node **p;
844         struct rb_node *parent;
845         int err = 0;
846         u64 last_start;
847         u64 last_end;
848
849         btrfs_debug_check_extent_io_range(tree, start, end);
850
851         bits |= EXTENT_FIRST_DELALLOC;
852 again:
853         if (!prealloc && (mask & __GFP_WAIT)) {
854                 prealloc = alloc_extent_state(mask);
855                 BUG_ON(!prealloc);
856         }
857
858         spin_lock(&tree->lock);
859         if (cached_state && *cached_state) {
860                 state = *cached_state;
861                 if (state->start <= start && state->end > start &&
862                     extent_state_in_tree(state)) {
863                         node = &state->rb_node;
864                         goto hit_next;
865                 }
866         }
867         /*
868          * this search will find all the extents that end after
869          * our range starts.
870          */
871         node = tree_search_for_insert(tree, start, &p, &parent);
872         if (!node) {
873                 prealloc = alloc_extent_state_atomic(prealloc);
874                 BUG_ON(!prealloc);
875                 err = insert_state(tree, prealloc, start, end,
876                                    &p, &parent, &bits);
877                 if (err)
878                         extent_io_tree_panic(tree, err);
879
880                 cache_state(prealloc, cached_state);
881                 prealloc = NULL;
882                 goto out;
883         }
884         state = rb_entry(node, struct extent_state, rb_node);
885 hit_next:
886         last_start = state->start;
887         last_end = state->end;
888
889         /*
890          * | ---- desired range ---- |
891          * | state |
892          *
893          * Just lock what we found and keep going
894          */
895         if (state->start == start && state->end <= end) {
896                 if (state->state & exclusive_bits) {
897                         *failed_start = state->start;
898                         err = -EEXIST;
899                         goto out;
900                 }
901
902                 set_state_bits(tree, state, &bits);
903                 cache_state(state, cached_state);
904                 merge_state(tree, state);
905                 if (last_end == (u64)-1)
906                         goto out;
907                 start = last_end + 1;
908                 state = next_state(state);
909                 if (start < end && state && state->start == start &&
910                     !need_resched())
911                         goto hit_next;
912                 goto search_again;
913         }
914
915         /*
916          *     | ---- desired range ---- |
917          * | state |
918          *   or
919          * | ------------- state -------------- |
920          *
921          * We need to split the extent we found, and may flip bits on
922          * second half.
923          *
924          * If the extent we found extends past our
925          * range, we just split and search again.  It'll get split
926          * again the next time though.
927          *
928          * If the extent we found is inside our range, we set the
929          * desired bit on it.
930          */
931         if (state->start < start) {
932                 if (state->state & exclusive_bits) {
933                         *failed_start = start;
934                         err = -EEXIST;
935                         goto out;
936                 }
937
938                 prealloc = alloc_extent_state_atomic(prealloc);
939                 BUG_ON(!prealloc);
940                 err = split_state(tree, state, prealloc, start);
941                 if (err)
942                         extent_io_tree_panic(tree, err);
943
944                 prealloc = NULL;
945                 if (err)
946                         goto out;
947                 if (state->end <= end) {
948                         set_state_bits(tree, state, &bits);
949                         cache_state(state, cached_state);
950                         merge_state(tree, state);
951                         if (last_end == (u64)-1)
952                                 goto out;
953                         start = last_end + 1;
954                         state = next_state(state);
955                         if (start < end && state && state->start == start &&
956                             !need_resched())
957                                 goto hit_next;
958                 }
959                 goto search_again;
960         }
961         /*
962          * | ---- desired range ---- |
963          *     | state | or               | state |
964          *
965          * There's a hole, we need to insert something in it and
966          * ignore the extent we found.
967          */
968         if (state->start > start) {
969                 u64 this_end;
970                 if (end < last_start)
971                         this_end = end;
972                 else
973                         this_end = last_start - 1;
974
975                 prealloc = alloc_extent_state_atomic(prealloc);
976                 BUG_ON(!prealloc);
977
978                 /*
979                  * Avoid to free 'prealloc' if it can be merged with
980                  * the later extent.
981                  */
982                 err = insert_state(tree, prealloc, start, this_end,
983                                    NULL, NULL, &bits);
984                 if (err)
985                         extent_io_tree_panic(tree, err);
986
987                 cache_state(prealloc, cached_state);
988                 prealloc = NULL;
989                 start = this_end + 1;
990                 goto search_again;
991         }
992         /*
993          * | ---- desired range ---- |
994          *                        | state |
995          * We need to split the extent, and set the bit
996          * on the first half
997          */
998         if (state->start <= end && state->end > end) {
999                 if (state->state & exclusive_bits) {
1000                         *failed_start = start;
1001                         err = -EEXIST;
1002                         goto out;
1003                 }
1004
1005                 prealloc = alloc_extent_state_atomic(prealloc);
1006                 BUG_ON(!prealloc);
1007                 err = split_state(tree, state, prealloc, end + 1);
1008                 if (err)
1009                         extent_io_tree_panic(tree, err);
1010
1011                 set_state_bits(tree, prealloc, &bits);
1012                 cache_state(prealloc, cached_state);
1013                 merge_state(tree, prealloc);
1014                 prealloc = NULL;
1015                 goto out;
1016         }
1017
1018         goto search_again;
1019
1020 out:
1021         spin_unlock(&tree->lock);
1022         if (prealloc)
1023                 free_extent_state(prealloc);
1024
1025         return err;
1026
1027 search_again:
1028         if (start > end)
1029                 goto out;
1030         spin_unlock(&tree->lock);
1031         if (mask & __GFP_WAIT)
1032                 cond_resched();
1033         goto again;
1034 }
1035
1036 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1037                    unsigned bits, u64 * failed_start,
1038                    struct extent_state **cached_state, gfp_t mask)
1039 {
1040         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1041                                 cached_state, mask);
1042 }
1043
1044
1045 /**
1046  * convert_extent_bit - convert all bits in a given range from one bit to
1047  *                      another
1048  * @tree:       the io tree to search
1049  * @start:      the start offset in bytes
1050  * @end:        the end offset in bytes (inclusive)
1051  * @bits:       the bits to set in this range
1052  * @clear_bits: the bits to clear in this range
1053  * @cached_state:       state that we're going to cache
1054  * @mask:       the allocation mask
1055  *
1056  * This will go through and set bits for the given range.  If any states exist
1057  * already in this range they are set with the given bit and cleared of the
1058  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1059  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1060  * boundary bits like LOCK.
1061  */
1062 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1063                        unsigned bits, unsigned clear_bits,
1064                        struct extent_state **cached_state, gfp_t mask)
1065 {
1066         struct extent_state *state;
1067         struct extent_state *prealloc = NULL;
1068         struct rb_node *node;
1069         struct rb_node **p;
1070         struct rb_node *parent;
1071         int err = 0;
1072         u64 last_start;
1073         u64 last_end;
1074         bool first_iteration = true;
1075
1076         btrfs_debug_check_extent_io_range(tree, start, end);
1077
1078 again:
1079         if (!prealloc && (mask & __GFP_WAIT)) {
1080                 /*
1081                  * Best effort, don't worry if extent state allocation fails
1082                  * here for the first iteration. We might have a cached state
1083                  * that matches exactly the target range, in which case no
1084                  * extent state allocations are needed. We'll only know this
1085                  * after locking the tree.
1086                  */
1087                 prealloc = alloc_extent_state(mask);
1088                 if (!prealloc && !first_iteration)
1089                         return -ENOMEM;
1090         }
1091
1092         spin_lock(&tree->lock);
1093         if (cached_state && *cached_state) {
1094                 state = *cached_state;
1095                 if (state->start <= start && state->end > start &&
1096                     extent_state_in_tree(state)) {
1097                         node = &state->rb_node;
1098                         goto hit_next;
1099                 }
1100         }
1101
1102         /*
1103          * this search will find all the extents that end after
1104          * our range starts.
1105          */
1106         node = tree_search_for_insert(tree, start, &p, &parent);
1107         if (!node) {
1108                 prealloc = alloc_extent_state_atomic(prealloc);
1109                 if (!prealloc) {
1110                         err = -ENOMEM;
1111                         goto out;
1112                 }
1113                 err = insert_state(tree, prealloc, start, end,
1114                                    &p, &parent, &bits);
1115                 if (err)
1116                         extent_io_tree_panic(tree, err);
1117                 cache_state(prealloc, cached_state);
1118                 prealloc = NULL;
1119                 goto out;
1120         }
1121         state = rb_entry(node, struct extent_state, rb_node);
1122 hit_next:
1123         last_start = state->start;
1124         last_end = state->end;
1125
1126         /*
1127          * | ---- desired range ---- |
1128          * | state |
1129          *
1130          * Just lock what we found and keep going
1131          */
1132         if (state->start == start && state->end <= end) {
1133                 set_state_bits(tree, state, &bits);
1134                 cache_state(state, cached_state);
1135                 state = clear_state_bit(tree, state, &clear_bits, 0);
1136                 if (last_end == (u64)-1)
1137                         goto out;
1138                 start = last_end + 1;
1139                 if (start < end && state && state->start == start &&
1140                     !need_resched())
1141                         goto hit_next;
1142                 goto search_again;
1143         }
1144
1145         /*
1146          *     | ---- desired range ---- |
1147          * | state |
1148          *   or
1149          * | ------------- state -------------- |
1150          *
1151          * We need to split the extent we found, and may flip bits on
1152          * second half.
1153          *
1154          * If the extent we found extends past our
1155          * range, we just split and search again.  It'll get split
1156          * again the next time though.
1157          *
1158          * If the extent we found is inside our range, we set the
1159          * desired bit on it.
1160          */
1161         if (state->start < start) {
1162                 prealloc = alloc_extent_state_atomic(prealloc);
1163                 if (!prealloc) {
1164                         err = -ENOMEM;
1165                         goto out;
1166                 }
1167                 err = split_state(tree, state, prealloc, start);
1168                 if (err)
1169                         extent_io_tree_panic(tree, err);
1170                 prealloc = NULL;
1171                 if (err)
1172                         goto out;
1173                 if (state->end <= end) {
1174                         set_state_bits(tree, state, &bits);
1175                         cache_state(state, cached_state);
1176                         state = clear_state_bit(tree, state, &clear_bits, 0);
1177                         if (last_end == (u64)-1)
1178                                 goto out;
1179                         start = last_end + 1;
1180                         if (start < end && state && state->start == start &&
1181                             !need_resched())
1182                                 goto hit_next;
1183                 }
1184                 goto search_again;
1185         }
1186         /*
1187          * | ---- desired range ---- |
1188          *     | state | or               | state |
1189          *
1190          * There's a hole, we need to insert something in it and
1191          * ignore the extent we found.
1192          */
1193         if (state->start > start) {
1194                 u64 this_end;
1195                 if (end < last_start)
1196                         this_end = end;
1197                 else
1198                         this_end = last_start - 1;
1199
1200                 prealloc = alloc_extent_state_atomic(prealloc);
1201                 if (!prealloc) {
1202                         err = -ENOMEM;
1203                         goto out;
1204                 }
1205
1206                 /*
1207                  * Avoid to free 'prealloc' if it can be merged with
1208                  * the later extent.
1209                  */
1210                 err = insert_state(tree, prealloc, start, this_end,
1211                                    NULL, NULL, &bits);
1212                 if (err)
1213                         extent_io_tree_panic(tree, err);
1214                 cache_state(prealloc, cached_state);
1215                 prealloc = NULL;
1216                 start = this_end + 1;
1217                 goto search_again;
1218         }
1219         /*
1220          * | ---- desired range ---- |
1221          *                        | state |
1222          * We need to split the extent, and set the bit
1223          * on the first half
1224          */
1225         if (state->start <= end && state->end > end) {
1226                 prealloc = alloc_extent_state_atomic(prealloc);
1227                 if (!prealloc) {
1228                         err = -ENOMEM;
1229                         goto out;
1230                 }
1231
1232                 err = split_state(tree, state, prealloc, end + 1);
1233                 if (err)
1234                         extent_io_tree_panic(tree, err);
1235
1236                 set_state_bits(tree, prealloc, &bits);
1237                 cache_state(prealloc, cached_state);
1238                 clear_state_bit(tree, prealloc, &clear_bits, 0);
1239                 prealloc = NULL;
1240                 goto out;
1241         }
1242
1243         goto search_again;
1244
1245 out:
1246         spin_unlock(&tree->lock);
1247         if (prealloc)
1248                 free_extent_state(prealloc);
1249
1250         return err;
1251
1252 search_again:
1253         if (start > end)
1254                 goto out;
1255         spin_unlock(&tree->lock);
1256         if (mask & __GFP_WAIT)
1257                 cond_resched();
1258         first_iteration = false;
1259         goto again;
1260 }
1261
1262 /* wrappers around set/clear extent bit */
1263 int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1264                      gfp_t mask)
1265 {
1266         return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1267                               NULL, mask);
1268 }
1269
1270 int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1271                     unsigned bits, gfp_t mask)
1272 {
1273         return set_extent_bit(tree, start, end, bits, NULL,
1274                               NULL, mask);
1275 }
1276
1277 int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1278                       unsigned bits, gfp_t mask)
1279 {
1280         int wake = 0;
1281
1282         if (bits & EXTENT_LOCKED)
1283                 wake = 1;
1284
1285         return clear_extent_bit(tree, start, end, bits, wake, 0, NULL, mask);
1286 }
1287
1288 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1289                         struct extent_state **cached_state, gfp_t mask)
1290 {
1291         return set_extent_bit(tree, start, end,
1292                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1293                               NULL, cached_state, mask);
1294 }
1295
1296 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1297                       struct extent_state **cached_state, gfp_t mask)
1298 {
1299         return set_extent_bit(tree, start, end,
1300                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1301                               NULL, cached_state, mask);
1302 }
1303
1304 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1305                        gfp_t mask)
1306 {
1307         return clear_extent_bit(tree, start, end,
1308                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1309                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1310 }
1311
1312 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1313                      gfp_t mask)
1314 {
1315         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1316                               NULL, mask);
1317 }
1318
1319 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1320                         struct extent_state **cached_state, gfp_t mask)
1321 {
1322         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1323                               cached_state, mask);
1324 }
1325
1326 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1327                           struct extent_state **cached_state, gfp_t mask)
1328 {
1329         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1330                                 cached_state, mask);
1331 }
1332
1333 /*
1334  * either insert or lock state struct between start and end use mask to tell
1335  * us if waiting is desired.
1336  */
1337 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1338                      unsigned bits, struct extent_state **cached_state)
1339 {
1340         int err;
1341         u64 failed_start;
1342
1343         while (1) {
1344                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1345                                        EXTENT_LOCKED, &failed_start,
1346                                        cached_state, GFP_NOFS);
1347                 if (err == -EEXIST) {
1348                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1349                         start = failed_start;
1350                 } else
1351                         break;
1352                 WARN_ON(start > end);
1353         }
1354         return err;
1355 }
1356
1357 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1358 {
1359         return lock_extent_bits(tree, start, end, 0, NULL);
1360 }
1361
1362 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1363 {
1364         int err;
1365         u64 failed_start;
1366
1367         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1368                                &failed_start, NULL, GFP_NOFS);
1369         if (err == -EEXIST) {
1370                 if (failed_start > start)
1371                         clear_extent_bit(tree, start, failed_start - 1,
1372                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1373                 return 0;
1374         }
1375         return 1;
1376 }
1377
1378 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1379                          struct extent_state **cached, gfp_t mask)
1380 {
1381         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1382                                 mask);
1383 }
1384
1385 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1386 {
1387         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1388                                 GFP_NOFS);
1389 }
1390
1391 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1392 {
1393         unsigned long index = start >> PAGE_CACHE_SHIFT;
1394         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1395         struct page *page;
1396
1397         while (index <= end_index) {
1398                 page = find_get_page(inode->i_mapping, index);
1399                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1400                 clear_page_dirty_for_io(page);
1401                 page_cache_release(page);
1402                 index++;
1403         }
1404         return 0;
1405 }
1406
1407 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1408 {
1409         unsigned long index = start >> PAGE_CACHE_SHIFT;
1410         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1411         struct page *page;
1412
1413         while (index <= end_index) {
1414                 page = find_get_page(inode->i_mapping, index);
1415                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1416                 __set_page_dirty_nobuffers(page);
1417                 account_page_redirty(page);
1418                 page_cache_release(page);
1419                 index++;
1420         }
1421         return 0;
1422 }
1423
1424 /*
1425  * helper function to set both pages and extents in the tree writeback
1426  */
1427 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1428 {
1429         unsigned long index = start >> PAGE_CACHE_SHIFT;
1430         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1431         struct page *page;
1432
1433         while (index <= end_index) {
1434                 page = find_get_page(tree->mapping, index);
1435                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1436                 set_page_writeback(page);
1437                 page_cache_release(page);
1438                 index++;
1439         }
1440         return 0;
1441 }
1442
1443 /* find the first state struct with 'bits' set after 'start', and
1444  * return it.  tree->lock must be held.  NULL will returned if
1445  * nothing was found after 'start'
1446  */
1447 static struct extent_state *
1448 find_first_extent_bit_state(struct extent_io_tree *tree,
1449                             u64 start, unsigned bits)
1450 {
1451         struct rb_node *node;
1452         struct extent_state *state;
1453
1454         /*
1455          * this search will find all the extents that end after
1456          * our range starts.
1457          */
1458         node = tree_search(tree, start);
1459         if (!node)
1460                 goto out;
1461
1462         while (1) {
1463                 state = rb_entry(node, struct extent_state, rb_node);
1464                 if (state->end >= start && (state->state & bits))
1465                         return state;
1466
1467                 node = rb_next(node);
1468                 if (!node)
1469                         break;
1470         }
1471 out:
1472         return NULL;
1473 }
1474
1475 /*
1476  * find the first offset in the io tree with 'bits' set. zero is
1477  * returned if we find something, and *start_ret and *end_ret are
1478  * set to reflect the state struct that was found.
1479  *
1480  * If nothing was found, 1 is returned. If found something, return 0.
1481  */
1482 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1483                           u64 *start_ret, u64 *end_ret, unsigned bits,
1484                           struct extent_state **cached_state)
1485 {
1486         struct extent_state *state;
1487         struct rb_node *n;
1488         int ret = 1;
1489
1490         spin_lock(&tree->lock);
1491         if (cached_state && *cached_state) {
1492                 state = *cached_state;
1493                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1494                         n = rb_next(&state->rb_node);
1495                         while (n) {
1496                                 state = rb_entry(n, struct extent_state,
1497                                                  rb_node);
1498                                 if (state->state & bits)
1499                                         goto got_it;
1500                                 n = rb_next(n);
1501                         }
1502                         free_extent_state(*cached_state);
1503                         *cached_state = NULL;
1504                         goto out;
1505                 }
1506                 free_extent_state(*cached_state);
1507                 *cached_state = NULL;
1508         }
1509
1510         state = find_first_extent_bit_state(tree, start, bits);
1511 got_it:
1512         if (state) {
1513                 cache_state_if_flags(state, cached_state, 0);
1514                 *start_ret = state->start;
1515                 *end_ret = state->end;
1516                 ret = 0;
1517         }
1518 out:
1519         spin_unlock(&tree->lock);
1520         return ret;
1521 }
1522
1523 /*
1524  * find a contiguous range of bytes in the file marked as delalloc, not
1525  * more than 'max_bytes'.  start and end are used to return the range,
1526  *
1527  * 1 is returned if we find something, 0 if nothing was in the tree
1528  */
1529 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1530                                         u64 *start, u64 *end, u64 max_bytes,
1531                                         struct extent_state **cached_state)
1532 {
1533         struct rb_node *node;
1534         struct extent_state *state;
1535         u64 cur_start = *start;
1536         u64 found = 0;
1537         u64 total_bytes = 0;
1538
1539         spin_lock(&tree->lock);
1540
1541         /*
1542          * this search will find all the extents that end after
1543          * our range starts.
1544          */
1545         node = tree_search(tree, cur_start);
1546         if (!node) {
1547                 if (!found)
1548                         *end = (u64)-1;
1549                 goto out;
1550         }
1551
1552         while (1) {
1553                 state = rb_entry(node, struct extent_state, rb_node);
1554                 if (found && (state->start != cur_start ||
1555                               (state->state & EXTENT_BOUNDARY))) {
1556                         goto out;
1557                 }
1558                 if (!(state->state & EXTENT_DELALLOC)) {
1559                         if (!found)
1560                                 *end = state->end;
1561                         goto out;
1562                 }
1563                 if (!found) {
1564                         *start = state->start;
1565                         *cached_state = state;
1566                         atomic_inc(&state->refs);
1567                 }
1568                 found++;
1569                 *end = state->end;
1570                 cur_start = state->end + 1;
1571                 node = rb_next(node);
1572                 total_bytes += state->end - state->start + 1;
1573                 if (total_bytes >= max_bytes)
1574                         break;
1575                 if (!node)
1576                         break;
1577         }
1578 out:
1579         spin_unlock(&tree->lock);
1580         return found;
1581 }
1582
1583 static noinline void __unlock_for_delalloc(struct inode *inode,
1584                                            struct page *locked_page,
1585                                            u64 start, u64 end)
1586 {
1587         int ret;
1588         struct page *pages[16];
1589         unsigned long index = start >> PAGE_CACHE_SHIFT;
1590         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1591         unsigned long nr_pages = end_index - index + 1;
1592         int i;
1593
1594         if (index == locked_page->index && end_index == index)
1595                 return;
1596
1597         while (nr_pages > 0) {
1598                 ret = find_get_pages_contig(inode->i_mapping, index,
1599                                      min_t(unsigned long, nr_pages,
1600                                      ARRAY_SIZE(pages)), pages);
1601                 for (i = 0; i < ret; i++) {
1602                         if (pages[i] != locked_page)
1603                                 unlock_page(pages[i]);
1604                         page_cache_release(pages[i]);
1605                 }
1606                 nr_pages -= ret;
1607                 index += ret;
1608                 cond_resched();
1609         }
1610 }
1611
1612 static noinline int lock_delalloc_pages(struct inode *inode,
1613                                         struct page *locked_page,
1614                                         u64 delalloc_start,
1615                                         u64 delalloc_end)
1616 {
1617         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1618         unsigned long start_index = index;
1619         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1620         unsigned long pages_locked = 0;
1621         struct page *pages[16];
1622         unsigned long nrpages;
1623         int ret;
1624         int i;
1625
1626         /* the caller is responsible for locking the start index */
1627         if (index == locked_page->index && index == end_index)
1628                 return 0;
1629
1630         /* skip the page at the start index */
1631         nrpages = end_index - index + 1;
1632         while (nrpages > 0) {
1633                 ret = find_get_pages_contig(inode->i_mapping, index,
1634                                      min_t(unsigned long,
1635                                      nrpages, ARRAY_SIZE(pages)), pages);
1636                 if (ret == 0) {
1637                         ret = -EAGAIN;
1638                         goto done;
1639                 }
1640                 /* now we have an array of pages, lock them all */
1641                 for (i = 0; i < ret; i++) {
1642                         /*
1643                          * the caller is taking responsibility for
1644                          * locked_page
1645                          */
1646                         if (pages[i] != locked_page) {
1647                                 lock_page(pages[i]);
1648                                 if (!PageDirty(pages[i]) ||
1649                                     pages[i]->mapping != inode->i_mapping) {
1650                                         ret = -EAGAIN;
1651                                         unlock_page(pages[i]);
1652                                         page_cache_release(pages[i]);
1653                                         goto done;
1654                                 }
1655                         }
1656                         page_cache_release(pages[i]);
1657                         pages_locked++;
1658                 }
1659                 nrpages -= ret;
1660                 index += ret;
1661                 cond_resched();
1662         }
1663         ret = 0;
1664 done:
1665         if (ret && pages_locked) {
1666                 __unlock_for_delalloc(inode, locked_page,
1667                               delalloc_start,
1668                               ((u64)(start_index + pages_locked - 1)) <<
1669                               PAGE_CACHE_SHIFT);
1670         }
1671         return ret;
1672 }
1673
1674 /*
1675  * find a contiguous range of bytes in the file marked as delalloc, not
1676  * more than 'max_bytes'.  start and end are used to return the range,
1677  *
1678  * 1 is returned if we find something, 0 if nothing was in the tree
1679  */
1680 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1681                                     struct extent_io_tree *tree,
1682                                     struct page *locked_page, u64 *start,
1683                                     u64 *end, u64 max_bytes)
1684 {
1685         u64 delalloc_start;
1686         u64 delalloc_end;
1687         u64 found;
1688         struct extent_state *cached_state = NULL;
1689         int ret;
1690         int loops = 0;
1691
1692 again:
1693         /* step one, find a bunch of delalloc bytes starting at start */
1694         delalloc_start = *start;
1695         delalloc_end = 0;
1696         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1697                                     max_bytes, &cached_state);
1698         if (!found || delalloc_end <= *start) {
1699                 *start = delalloc_start;
1700                 *end = delalloc_end;
1701                 free_extent_state(cached_state);
1702                 return 0;
1703         }
1704
1705         /*
1706          * start comes from the offset of locked_page.  We have to lock
1707          * pages in order, so we can't process delalloc bytes before
1708          * locked_page
1709          */
1710         if (delalloc_start < *start)
1711                 delalloc_start = *start;
1712
1713         /*
1714          * make sure to limit the number of pages we try to lock down
1715          */
1716         if (delalloc_end + 1 - delalloc_start > max_bytes)
1717                 delalloc_end = delalloc_start + max_bytes - 1;
1718
1719         /* step two, lock all the pages after the page that has start */
1720         ret = lock_delalloc_pages(inode, locked_page,
1721                                   delalloc_start, delalloc_end);
1722         if (ret == -EAGAIN) {
1723                 /* some of the pages are gone, lets avoid looping by
1724                  * shortening the size of the delalloc range we're searching
1725                  */
1726                 free_extent_state(cached_state);
1727                 cached_state = NULL;
1728                 if (!loops) {
1729                         max_bytes = PAGE_CACHE_SIZE;
1730                         loops = 1;
1731                         goto again;
1732                 } else {
1733                         found = 0;
1734                         goto out_failed;
1735                 }
1736         }
1737         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1738
1739         /* step three, lock the state bits for the whole range */
1740         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1741
1742         /* then test to make sure it is all still delalloc */
1743         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1744                              EXTENT_DELALLOC, 1, cached_state);
1745         if (!ret) {
1746                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1747                                      &cached_state, GFP_NOFS);
1748                 __unlock_for_delalloc(inode, locked_page,
1749                               delalloc_start, delalloc_end);
1750                 cond_resched();
1751                 goto again;
1752         }
1753         free_extent_state(cached_state);
1754         *start = delalloc_start;
1755         *end = delalloc_end;
1756 out_failed:
1757         return found;
1758 }
1759
1760 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1761                                  struct page *locked_page,
1762                                  unsigned clear_bits,
1763                                  unsigned long page_ops)
1764 {
1765         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1766         int ret;
1767         struct page *pages[16];
1768         unsigned long index = start >> PAGE_CACHE_SHIFT;
1769         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1770         unsigned long nr_pages = end_index - index + 1;
1771         int i;
1772
1773         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1774         if (page_ops == 0)
1775                 return 0;
1776
1777         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1778                 mapping_set_error(inode->i_mapping, -EIO);
1779
1780         while (nr_pages > 0) {
1781                 ret = find_get_pages_contig(inode->i_mapping, index,
1782                                      min_t(unsigned long,
1783                                      nr_pages, ARRAY_SIZE(pages)), pages);
1784                 for (i = 0; i < ret; i++) {
1785
1786                         if (page_ops & PAGE_SET_PRIVATE2)
1787                                 SetPagePrivate2(pages[i]);
1788
1789                         if (pages[i] == locked_page) {
1790                                 page_cache_release(pages[i]);
1791                                 continue;
1792                         }
1793                         if (page_ops & PAGE_CLEAR_DIRTY)
1794                                 clear_page_dirty_for_io(pages[i]);
1795                         if (page_ops & PAGE_SET_WRITEBACK)
1796                                 set_page_writeback(pages[i]);
1797                         if (page_ops & PAGE_SET_ERROR)
1798                                 SetPageError(pages[i]);
1799                         if (page_ops & PAGE_END_WRITEBACK)
1800                                 end_page_writeback(pages[i]);
1801                         if (page_ops & PAGE_UNLOCK)
1802                                 unlock_page(pages[i]);
1803                         page_cache_release(pages[i]);
1804                 }
1805                 nr_pages -= ret;
1806                 index += ret;
1807                 cond_resched();
1808         }
1809         return 0;
1810 }
1811
1812 /*
1813  * count the number of bytes in the tree that have a given bit(s)
1814  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1815  * cached.  The total number found is returned.
1816  */
1817 u64 count_range_bits(struct extent_io_tree *tree,
1818                      u64 *start, u64 search_end, u64 max_bytes,
1819                      unsigned bits, int contig)
1820 {
1821         struct rb_node *node;
1822         struct extent_state *state;
1823         u64 cur_start = *start;
1824         u64 total_bytes = 0;
1825         u64 last = 0;
1826         int found = 0;
1827
1828         if (WARN_ON(search_end <= cur_start))
1829                 return 0;
1830
1831         spin_lock(&tree->lock);
1832         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1833                 total_bytes = tree->dirty_bytes;
1834                 goto out;
1835         }
1836         /*
1837          * this search will find all the extents that end after
1838          * our range starts.
1839          */
1840         node = tree_search(tree, cur_start);
1841         if (!node)
1842                 goto out;
1843
1844         while (1) {
1845                 state = rb_entry(node, struct extent_state, rb_node);
1846                 if (state->start > search_end)
1847                         break;
1848                 if (contig && found && state->start > last + 1)
1849                         break;
1850                 if (state->end >= cur_start && (state->state & bits) == bits) {
1851                         total_bytes += min(search_end, state->end) + 1 -
1852                                        max(cur_start, state->start);
1853                         if (total_bytes >= max_bytes)
1854                                 break;
1855                         if (!found) {
1856                                 *start = max(cur_start, state->start);
1857                                 found = 1;
1858                         }
1859                         last = state->end;
1860                 } else if (contig && found) {
1861                         break;
1862                 }
1863                 node = rb_next(node);
1864                 if (!node)
1865                         break;
1866         }
1867 out:
1868         spin_unlock(&tree->lock);
1869         return total_bytes;
1870 }
1871
1872 /*
1873  * set the private field for a given byte offset in the tree.  If there isn't
1874  * an extent_state there already, this does nothing.
1875  */
1876 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1877 {
1878         struct rb_node *node;
1879         struct extent_state *state;
1880         int ret = 0;
1881
1882         spin_lock(&tree->lock);
1883         /*
1884          * this search will find all the extents that end after
1885          * our range starts.
1886          */
1887         node = tree_search(tree, start);
1888         if (!node) {
1889                 ret = -ENOENT;
1890                 goto out;
1891         }
1892         state = rb_entry(node, struct extent_state, rb_node);
1893         if (state->start != start) {
1894                 ret = -ENOENT;
1895                 goto out;
1896         }
1897         state->private = private;
1898 out:
1899         spin_unlock(&tree->lock);
1900         return ret;
1901 }
1902
1903 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1904 {
1905         struct rb_node *node;
1906         struct extent_state *state;
1907         int ret = 0;
1908
1909         spin_lock(&tree->lock);
1910         /*
1911          * this search will find all the extents that end after
1912          * our range starts.
1913          */
1914         node = tree_search(tree, start);
1915         if (!node) {
1916                 ret = -ENOENT;
1917                 goto out;
1918         }
1919         state = rb_entry(node, struct extent_state, rb_node);
1920         if (state->start != start) {
1921                 ret = -ENOENT;
1922                 goto out;
1923         }
1924         *private = state->private;
1925 out:
1926         spin_unlock(&tree->lock);
1927         return ret;
1928 }
1929
1930 /*
1931  * searches a range in the state tree for a given mask.
1932  * If 'filled' == 1, this returns 1 only if every extent in the tree
1933  * has the bits set.  Otherwise, 1 is returned if any bit in the
1934  * range is found set.
1935  */
1936 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1937                    unsigned bits, int filled, struct extent_state *cached)
1938 {
1939         struct extent_state *state = NULL;
1940         struct rb_node *node;
1941         int bitset = 0;
1942
1943         spin_lock(&tree->lock);
1944         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1945             cached->end > start)
1946                 node = &cached->rb_node;
1947         else
1948                 node = tree_search(tree, start);
1949         while (node && start <= end) {
1950                 state = rb_entry(node, struct extent_state, rb_node);
1951
1952                 if (filled && state->start > start) {
1953                         bitset = 0;
1954                         break;
1955                 }
1956
1957                 if (state->start > end)
1958                         break;
1959
1960                 if (state->state & bits) {
1961                         bitset = 1;
1962                         if (!filled)
1963                                 break;
1964                 } else if (filled) {
1965                         bitset = 0;
1966                         break;
1967                 }
1968
1969                 if (state->end == (u64)-1)
1970                         break;
1971
1972                 start = state->end + 1;
1973                 if (start > end)
1974                         break;
1975                 node = rb_next(node);
1976                 if (!node) {
1977                         if (filled)
1978                                 bitset = 0;
1979                         break;
1980                 }
1981         }
1982         spin_unlock(&tree->lock);
1983         return bitset;
1984 }
1985
1986 /*
1987  * helper function to set a given page up to date if all the
1988  * extents in the tree for that page are up to date
1989  */
1990 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1991 {
1992         u64 start = page_offset(page);
1993         u64 end = start + PAGE_CACHE_SIZE - 1;
1994         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1995                 SetPageUptodate(page);
1996 }
1997
1998 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1999 {
2000         int ret;
2001         int err = 0;
2002         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2003
2004         set_state_private(failure_tree, rec->start, 0);
2005         ret = clear_extent_bits(failure_tree, rec->start,
2006                                 rec->start + rec->len - 1,
2007                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2008         if (ret)
2009                 err = ret;
2010
2011         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
2012                                 rec->start + rec->len - 1,
2013                                 EXTENT_DAMAGED, GFP_NOFS);
2014         if (ret && !err)
2015                 err = ret;
2016
2017         kfree(rec);
2018         return err;
2019 }
2020
2021 /*
2022  * this bypasses the standard btrfs submit functions deliberately, as
2023  * the standard behavior is to write all copies in a raid setup. here we only
2024  * want to write the one bad copy. so we do the mapping for ourselves and issue
2025  * submit_bio directly.
2026  * to avoid any synchronization issues, wait for the data after writing, which
2027  * actually prevents the read that triggered the error from finishing.
2028  * currently, there can be no more than two copies of every data bit. thus,
2029  * exactly one rewrite is required.
2030  */
2031 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2032                       struct page *page, unsigned int pg_offset, int mirror_num)
2033 {
2034         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2035         struct bio *bio;
2036         struct btrfs_device *dev;
2037         u64 map_length = 0;
2038         u64 sector;
2039         struct btrfs_bio *bbio = NULL;
2040         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2041         int ret;
2042
2043         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2044         BUG_ON(!mirror_num);
2045
2046         /* we can't repair anything in raid56 yet */
2047         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2048                 return 0;
2049
2050         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2051         if (!bio)
2052                 return -EIO;
2053         bio->bi_iter.bi_size = 0;
2054         map_length = length;
2055
2056         ret = btrfs_map_block(fs_info, WRITE, logical,
2057                               &map_length, &bbio, mirror_num);
2058         if (ret) {
2059                 bio_put(bio);
2060                 return -EIO;
2061         }
2062         BUG_ON(mirror_num != bbio->mirror_num);
2063         sector = bbio->stripes[mirror_num-1].physical >> 9;
2064         bio->bi_iter.bi_sector = sector;
2065         dev = bbio->stripes[mirror_num-1].dev;
2066         btrfs_put_bbio(bbio);
2067         if (!dev || !dev->bdev || !dev->writeable) {
2068                 bio_put(bio);
2069                 return -EIO;
2070         }
2071         bio->bi_bdev = dev->bdev;
2072         bio_add_page(bio, page, length, pg_offset);
2073
2074         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2075                 /* try to remap that extent elsewhere? */
2076                 bio_put(bio);
2077                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2078                 return -EIO;
2079         }
2080
2081         printk_ratelimited_in_rcu(KERN_INFO
2082                                   "BTRFS: read error corrected: ino %llu off %llu (dev %s sector %llu)\n",
2083                                   btrfs_ino(inode), start,
2084                                   rcu_str_deref(dev->name), sector);
2085         bio_put(bio);
2086         return 0;
2087 }
2088
2089 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2090                          int mirror_num)
2091 {
2092         u64 start = eb->start;
2093         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2094         int ret = 0;
2095
2096         if (root->fs_info->sb->s_flags & MS_RDONLY)
2097                 return -EROFS;
2098
2099         for (i = 0; i < num_pages; i++) {
2100                 struct page *p = eb->pages[i];
2101
2102                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2103                                         PAGE_CACHE_SIZE, start, p,
2104                                         start - page_offset(p), mirror_num);
2105                 if (ret)
2106                         break;
2107                 start += PAGE_CACHE_SIZE;
2108         }
2109
2110         return ret;
2111 }
2112
2113 /*
2114  * each time an IO finishes, we do a fast check in the IO failure tree
2115  * to see if we need to process or clean up an io_failure_record
2116  */
2117 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2118                      unsigned int pg_offset)
2119 {
2120         u64 private;
2121         u64 private_failure;
2122         struct io_failure_record *failrec;
2123         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2124         struct extent_state *state;
2125         int num_copies;
2126         int ret;
2127
2128         private = 0;
2129         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2130                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2131         if (!ret)
2132                 return 0;
2133
2134         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2135                                 &private_failure);
2136         if (ret)
2137                 return 0;
2138
2139         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2140         BUG_ON(!failrec->this_mirror);
2141
2142         if (failrec->in_validation) {
2143                 /* there was no real error, just free the record */
2144                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2145                          failrec->start);
2146                 goto out;
2147         }
2148         if (fs_info->sb->s_flags & MS_RDONLY)
2149                 goto out;
2150
2151         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2152         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2153                                             failrec->start,
2154                                             EXTENT_LOCKED);
2155         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2156
2157         if (state && state->start <= failrec->start &&
2158             state->end >= failrec->start + failrec->len - 1) {
2159                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2160                                               failrec->len);
2161                 if (num_copies > 1)  {
2162                         repair_io_failure(inode, start, failrec->len,
2163                                           failrec->logical, page,
2164                                           pg_offset, failrec->failed_mirror);
2165                 }
2166         }
2167
2168 out:
2169         free_io_failure(inode, failrec);
2170
2171         return 0;
2172 }
2173
2174 /*
2175  * Can be called when
2176  * - hold extent lock
2177  * - under ordered extent
2178  * - the inode is freeing
2179  */
2180 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2181 {
2182         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2183         struct io_failure_record *failrec;
2184         struct extent_state *state, *next;
2185
2186         if (RB_EMPTY_ROOT(&failure_tree->state))
2187                 return;
2188
2189         spin_lock(&failure_tree->lock);
2190         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2191         while (state) {
2192                 if (state->start > end)
2193                         break;
2194
2195                 ASSERT(state->end <= end);
2196
2197                 next = next_state(state);
2198
2199                 failrec = (struct io_failure_record *)(unsigned long)state->private;
2200                 free_extent_state(state);
2201                 kfree(failrec);
2202
2203                 state = next;
2204         }
2205         spin_unlock(&failure_tree->lock);
2206 }
2207
2208 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2209                                 struct io_failure_record **failrec_ret)
2210 {
2211         struct io_failure_record *failrec;
2212         u64 private;
2213         struct extent_map *em;
2214         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2215         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2216         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2217         int ret;
2218         u64 logical;
2219
2220         ret = get_state_private(failure_tree, start, &private);
2221         if (ret) {
2222                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2223                 if (!failrec)
2224                         return -ENOMEM;
2225
2226                 failrec->start = start;
2227                 failrec->len = end - start + 1;
2228                 failrec->this_mirror = 0;
2229                 failrec->bio_flags = 0;
2230                 failrec->in_validation = 0;
2231
2232                 read_lock(&em_tree->lock);
2233                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2234                 if (!em) {
2235                         read_unlock(&em_tree->lock);
2236                         kfree(failrec);
2237                         return -EIO;
2238                 }
2239
2240                 if (em->start > start || em->start + em->len <= start) {
2241                         free_extent_map(em);
2242                         em = NULL;
2243                 }
2244                 read_unlock(&em_tree->lock);
2245                 if (!em) {
2246                         kfree(failrec);
2247                         return -EIO;
2248                 }
2249
2250                 logical = start - em->start;
2251                 logical = em->block_start + logical;
2252                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2253                         logical = em->block_start;
2254                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2255                         extent_set_compress_type(&failrec->bio_flags,
2256                                                  em->compress_type);
2257                 }
2258
2259                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2260                          logical, start, failrec->len);
2261
2262                 failrec->logical = logical;
2263                 free_extent_map(em);
2264
2265                 /* set the bits in the private failure tree */
2266                 ret = set_extent_bits(failure_tree, start, end,
2267                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2268                 if (ret >= 0)
2269                         ret = set_state_private(failure_tree, start,
2270                                                 (u64)(unsigned long)failrec);
2271                 /* set the bits in the inode's tree */
2272                 if (ret >= 0)
2273                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2274                                                 GFP_NOFS);
2275                 if (ret < 0) {
2276                         kfree(failrec);
2277                         return ret;
2278                 }
2279         } else {
2280                 failrec = (struct io_failure_record *)(unsigned long)private;
2281                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2282                          failrec->logical, failrec->start, failrec->len,
2283                          failrec->in_validation);
2284                 /*
2285                  * when data can be on disk more than twice, add to failrec here
2286                  * (e.g. with a list for failed_mirror) to make
2287                  * clean_io_failure() clean all those errors at once.
2288                  */
2289         }
2290
2291         *failrec_ret = failrec;
2292
2293         return 0;
2294 }
2295
2296 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2297                            struct io_failure_record *failrec, int failed_mirror)
2298 {
2299         int num_copies;
2300
2301         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2302                                       failrec->logical, failrec->len);
2303         if (num_copies == 1) {
2304                 /*
2305                  * we only have a single copy of the data, so don't bother with
2306                  * all the retry and error correction code that follows. no
2307                  * matter what the error is, it is very likely to persist.
2308                  */
2309                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2310                          num_copies, failrec->this_mirror, failed_mirror);
2311                 return 0;
2312         }
2313
2314         /*
2315          * there are two premises:
2316          *      a) deliver good data to the caller
2317          *      b) correct the bad sectors on disk
2318          */
2319         if (failed_bio->bi_vcnt > 1) {
2320                 /*
2321                  * to fulfill b), we need to know the exact failing sectors, as
2322                  * we don't want to rewrite any more than the failed ones. thus,
2323                  * we need separate read requests for the failed bio
2324                  *
2325                  * if the following BUG_ON triggers, our validation request got
2326                  * merged. we need separate requests for our algorithm to work.
2327                  */
2328                 BUG_ON(failrec->in_validation);
2329                 failrec->in_validation = 1;
2330                 failrec->this_mirror = failed_mirror;
2331         } else {
2332                 /*
2333                  * we're ready to fulfill a) and b) alongside. get a good copy
2334                  * of the failed sector and if we succeed, we have setup
2335                  * everything for repair_io_failure to do the rest for us.
2336                  */
2337                 if (failrec->in_validation) {
2338                         BUG_ON(failrec->this_mirror != failed_mirror);
2339                         failrec->in_validation = 0;
2340                         failrec->this_mirror = 0;
2341                 }
2342                 failrec->failed_mirror = failed_mirror;
2343                 failrec->this_mirror++;
2344                 if (failrec->this_mirror == failed_mirror)
2345                         failrec->this_mirror++;
2346         }
2347
2348         if (failrec->this_mirror > num_copies) {
2349                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2350                          num_copies, failrec->this_mirror, failed_mirror);
2351                 return 0;
2352         }
2353
2354         return 1;
2355 }
2356
2357
2358 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2359                                     struct io_failure_record *failrec,
2360                                     struct page *page, int pg_offset, int icsum,
2361                                     bio_end_io_t *endio_func, void *data)
2362 {
2363         struct bio *bio;
2364         struct btrfs_io_bio *btrfs_failed_bio;
2365         struct btrfs_io_bio *btrfs_bio;
2366
2367         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2368         if (!bio)
2369                 return NULL;
2370
2371         bio->bi_end_io = endio_func;
2372         bio->bi_iter.bi_sector = failrec->logical >> 9;
2373         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2374         bio->bi_iter.bi_size = 0;
2375         bio->bi_private = data;
2376
2377         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2378         if (btrfs_failed_bio->csum) {
2379                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2380                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2381
2382                 btrfs_bio = btrfs_io_bio(bio);
2383                 btrfs_bio->csum = btrfs_bio->csum_inline;
2384                 icsum *= csum_size;
2385                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2386                        csum_size);
2387         }
2388
2389         bio_add_page(bio, page, failrec->len, pg_offset);
2390
2391         return bio;
2392 }
2393
2394 /*
2395  * this is a generic handler for readpage errors (default
2396  * readpage_io_failed_hook). if other copies exist, read those and write back
2397  * good data to the failed position. does not investigate in remapping the
2398  * failed extent elsewhere, hoping the device will be smart enough to do this as
2399  * needed
2400  */
2401
2402 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2403                               struct page *page, u64 start, u64 end,
2404                               int failed_mirror)
2405 {
2406         struct io_failure_record *failrec;
2407         struct inode *inode = page->mapping->host;
2408         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2409         struct bio *bio;
2410         int read_mode;
2411         int ret;
2412
2413         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2414
2415         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2416         if (ret)
2417                 return ret;
2418
2419         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2420         if (!ret) {
2421                 free_io_failure(inode, failrec);
2422                 return -EIO;
2423         }
2424
2425         if (failed_bio->bi_vcnt > 1)
2426                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2427         else
2428                 read_mode = READ_SYNC;
2429
2430         phy_offset >>= inode->i_sb->s_blocksize_bits;
2431         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2432                                       start - page_offset(page),
2433                                       (int)phy_offset, failed_bio->bi_end_io,
2434                                       NULL);
2435         if (!bio) {
2436                 free_io_failure(inode, failrec);
2437                 return -EIO;
2438         }
2439
2440         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2441                  read_mode, failrec->this_mirror, failrec->in_validation);
2442
2443         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2444                                          failrec->this_mirror,
2445                                          failrec->bio_flags, 0);
2446         if (ret) {
2447                 free_io_failure(inode, failrec);
2448                 bio_put(bio);
2449         }
2450
2451         return ret;
2452 }
2453
2454 /* lots and lots of room for performance fixes in the end_bio funcs */
2455
2456 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2457 {
2458         int uptodate = (err == 0);
2459         struct extent_io_tree *tree;
2460         int ret = 0;
2461
2462         tree = &BTRFS_I(page->mapping->host)->io_tree;
2463
2464         if (tree->ops && tree->ops->writepage_end_io_hook) {
2465                 ret = tree->ops->writepage_end_io_hook(page, start,
2466                                                end, NULL, uptodate);
2467                 if (ret)
2468                         uptodate = 0;
2469         }
2470
2471         if (!uptodate) {
2472                 ClearPageUptodate(page);
2473                 SetPageError(page);
2474                 ret = ret < 0 ? ret : -EIO;
2475                 mapping_set_error(page->mapping, ret);
2476         }
2477         return 0;
2478 }
2479
2480 /*
2481  * after a writepage IO is done, we need to:
2482  * clear the uptodate bits on error
2483  * clear the writeback bits in the extent tree for this IO
2484  * end_page_writeback if the page has no more pending IO
2485  *
2486  * Scheduling is not allowed, so the extent state tree is expected
2487  * to have one and only one object corresponding to this IO.
2488  */
2489 static void end_bio_extent_writepage(struct bio *bio)
2490 {
2491         struct bio_vec *bvec;
2492         u64 start;
2493         u64 end;
2494         int i;
2495
2496         bio_for_each_segment_all(bvec, bio, i) {
2497                 struct page *page = bvec->bv_page;
2498
2499                 /* We always issue full-page reads, but if some block
2500                  * in a page fails to read, blk_update_request() will
2501                  * advance bv_offset and adjust bv_len to compensate.
2502                  * Print a warning for nonzero offsets, and an error
2503                  * if they don't add up to a full page.  */
2504                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2505                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2506                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2507                                    "partial page write in btrfs with offset %u and length %u",
2508                                         bvec->bv_offset, bvec->bv_len);
2509                         else
2510                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2511                                    "incomplete page write in btrfs with offset %u and "
2512                                    "length %u",
2513                                         bvec->bv_offset, bvec->bv_len);
2514                 }
2515
2516                 start = page_offset(page);
2517                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2518
2519                 if (end_extent_writepage(page, bio->bi_error, start, end))
2520                         continue;
2521
2522                 end_page_writeback(page);
2523         }
2524
2525         bio_put(bio);
2526 }
2527
2528 static void
2529 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2530                               int uptodate)
2531 {
2532         struct extent_state *cached = NULL;
2533         u64 end = start + len - 1;
2534
2535         if (uptodate && tree->track_uptodate)
2536                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2537         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2538 }
2539
2540 /*
2541  * after a readpage IO is done, we need to:
2542  * clear the uptodate bits on error
2543  * set the uptodate bits if things worked
2544  * set the page up to date if all extents in the tree are uptodate
2545  * clear the lock bit in the extent tree
2546  * unlock the page if there are no other extents locked for it
2547  *
2548  * Scheduling is not allowed, so the extent state tree is expected
2549  * to have one and only one object corresponding to this IO.
2550  */
2551 static void end_bio_extent_readpage(struct bio *bio)
2552 {
2553         struct bio_vec *bvec;
2554         int uptodate = !bio->bi_error;
2555         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2556         struct extent_io_tree *tree;
2557         u64 offset = 0;
2558         u64 start;
2559         u64 end;
2560         u64 len;
2561         u64 extent_start = 0;
2562         u64 extent_len = 0;
2563         int mirror;
2564         int ret;
2565         int i;
2566
2567         bio_for_each_segment_all(bvec, bio, i) {
2568                 struct page *page = bvec->bv_page;
2569                 struct inode *inode = page->mapping->host;
2570
2571                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2572                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector,
2573                          bio->bi_error, io_bio->mirror_num);
2574                 tree = &BTRFS_I(inode)->io_tree;
2575
2576                 /* We always issue full-page reads, but if some block
2577                  * in a page fails to read, blk_update_request() will
2578                  * advance bv_offset and adjust bv_len to compensate.
2579                  * Print a warning for nonzero offsets, and an error
2580                  * if they don't add up to a full page.  */
2581                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2582                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2583                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2584                                    "partial page read in btrfs with offset %u and length %u",
2585                                         bvec->bv_offset, bvec->bv_len);
2586                         else
2587                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2588                                    "incomplete page read in btrfs with offset %u and "
2589                                    "length %u",
2590                                         bvec->bv_offset, bvec->bv_len);
2591                 }
2592
2593                 start = page_offset(page);
2594                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2595                 len = bvec->bv_len;
2596
2597                 mirror = io_bio->mirror_num;
2598                 if (likely(uptodate && tree->ops &&
2599                            tree->ops->readpage_end_io_hook)) {
2600                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2601                                                               page, start, end,
2602                                                               mirror);
2603                         if (ret)
2604                                 uptodate = 0;
2605                         else
2606                                 clean_io_failure(inode, start, page, 0);
2607                 }
2608
2609                 if (likely(uptodate))
2610                         goto readpage_ok;
2611
2612                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2613                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2614                         if (!ret && !bio->bi_error)
2615                                 uptodate = 1;
2616                 } else {
2617                         /*
2618                          * The generic bio_readpage_error handles errors the
2619                          * following way: If possible, new read requests are
2620                          * created and submitted and will end up in
2621                          * end_bio_extent_readpage as well (if we're lucky, not
2622                          * in the !uptodate case). In that case it returns 0 and
2623                          * we just go on with the next page in our bio. If it
2624                          * can't handle the error it will return -EIO and we
2625                          * remain responsible for that page.
2626                          */
2627                         ret = bio_readpage_error(bio, offset, page, start, end,
2628                                                  mirror);
2629                         if (ret == 0) {
2630                                 uptodate = !bio->bi_error;
2631                                 offset += len;
2632                                 continue;
2633                         }
2634                 }
2635 readpage_ok:
2636                 if (likely(uptodate)) {
2637                         loff_t i_size = i_size_read(inode);
2638                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2639                         unsigned off;
2640
2641                         /* Zero out the end if this page straddles i_size */
2642                         off = i_size & (PAGE_CACHE_SIZE-1);
2643                         if (page->index == end_index && off)
2644                                 zero_user_segment(page, off, PAGE_CACHE_SIZE);
2645                         SetPageUptodate(page);
2646                 } else {
2647                         ClearPageUptodate(page);
2648                         SetPageError(page);
2649                 }
2650                 unlock_page(page);
2651                 offset += len;
2652
2653                 if (unlikely(!uptodate)) {
2654                         if (extent_len) {
2655                                 endio_readpage_release_extent(tree,
2656                                                               extent_start,
2657                                                               extent_len, 1);
2658                                 extent_start = 0;
2659                                 extent_len = 0;
2660                         }
2661                         endio_readpage_release_extent(tree, start,
2662                                                       end - start + 1, 0);
2663                 } else if (!extent_len) {
2664                         extent_start = start;
2665                         extent_len = end + 1 - start;
2666                 } else if (extent_start + extent_len == start) {
2667                         extent_len += end + 1 - start;
2668                 } else {
2669                         endio_readpage_release_extent(tree, extent_start,
2670                                                       extent_len, uptodate);
2671                         extent_start = start;
2672                         extent_len = end + 1 - start;
2673                 }
2674         }
2675
2676         if (extent_len)
2677                 endio_readpage_release_extent(tree, extent_start, extent_len,
2678                                               uptodate);
2679         if (io_bio->end_io)
2680                 io_bio->end_io(io_bio, bio->bi_error);
2681         bio_put(bio);
2682 }
2683
2684 /*
2685  * this allocates from the btrfs_bioset.  We're returning a bio right now
2686  * but you can call btrfs_io_bio for the appropriate container_of magic
2687  */
2688 struct bio *
2689 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2690                 gfp_t gfp_flags)
2691 {
2692         struct btrfs_io_bio *btrfs_bio;
2693         struct bio *bio;
2694
2695         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2696
2697         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2698                 while (!bio && (nr_vecs /= 2)) {
2699                         bio = bio_alloc_bioset(gfp_flags,
2700                                                nr_vecs, btrfs_bioset);
2701                 }
2702         }
2703
2704         if (bio) {
2705                 bio->bi_bdev = bdev;
2706                 bio->bi_iter.bi_sector = first_sector;
2707                 btrfs_bio = btrfs_io_bio(bio);
2708                 btrfs_bio->csum = NULL;
2709                 btrfs_bio->csum_allocated = NULL;
2710                 btrfs_bio->end_io = NULL;
2711         }
2712         return bio;
2713 }
2714
2715 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2716 {
2717         struct btrfs_io_bio *btrfs_bio;
2718         struct bio *new;
2719
2720         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2721         if (new) {
2722                 btrfs_bio = btrfs_io_bio(new);
2723                 btrfs_bio->csum = NULL;
2724                 btrfs_bio->csum_allocated = NULL;
2725                 btrfs_bio->end_io = NULL;
2726
2727 #ifdef CONFIG_BLK_CGROUP
2728                 /* FIXME, put this into bio_clone_bioset */
2729                 if (bio->bi_css)
2730                         bio_associate_blkcg(new, bio->bi_css);
2731 #endif
2732         }
2733         return new;
2734 }
2735
2736 /* this also allocates from the btrfs_bioset */
2737 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2738 {
2739         struct btrfs_io_bio *btrfs_bio;
2740         struct bio *bio;
2741
2742         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2743         if (bio) {
2744                 btrfs_bio = btrfs_io_bio(bio);
2745                 btrfs_bio->csum = NULL;
2746                 btrfs_bio->csum_allocated = NULL;
2747                 btrfs_bio->end_io = NULL;
2748         }
2749         return bio;
2750 }
2751
2752
2753 static int __must_check submit_one_bio(int rw, struct bio *bio,
2754                                        int mirror_num, unsigned long bio_flags)
2755 {
2756         int ret = 0;
2757         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2758         struct page *page = bvec->bv_page;
2759         struct extent_io_tree *tree = bio->bi_private;
2760         u64 start;
2761
2762         start = page_offset(page) + bvec->bv_offset;
2763
2764         bio->bi_private = NULL;
2765
2766         bio_get(bio);
2767
2768         if (tree->ops && tree->ops->submit_bio_hook)
2769                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2770                                            mirror_num, bio_flags, start);
2771         else
2772                 btrfsic_submit_bio(rw, bio);
2773
2774         bio_put(bio);
2775         return ret;
2776 }
2777
2778 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2779                      unsigned long offset, size_t size, struct bio *bio,
2780                      unsigned long bio_flags)
2781 {
2782         int ret = 0;
2783         if (tree->ops && tree->ops->merge_bio_hook)
2784                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2785                                                 bio_flags);
2786         BUG_ON(ret < 0);
2787         return ret;
2788
2789 }
2790
2791 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2792                               struct writeback_control *wbc,
2793                               struct page *page, sector_t sector,
2794                               size_t size, unsigned long offset,
2795                               struct block_device *bdev,
2796                               struct bio **bio_ret,
2797                               unsigned long max_pages,
2798                               bio_end_io_t end_io_func,
2799                               int mirror_num,
2800                               unsigned long prev_bio_flags,
2801                               unsigned long bio_flags,
2802                               bool force_bio_submit)
2803 {
2804         int ret = 0;
2805         struct bio *bio;
2806         int contig = 0;
2807         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2808         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2809
2810         if (bio_ret && *bio_ret) {
2811                 bio = *bio_ret;
2812                 if (old_compressed)
2813                         contig = bio->bi_iter.bi_sector == sector;
2814                 else
2815                         contig = bio_end_sector(bio) == sector;
2816
2817                 if (prev_bio_flags != bio_flags || !contig ||
2818                     force_bio_submit ||
2819                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2820                     bio_add_page(bio, page, page_size, offset) < page_size) {
2821                         ret = submit_one_bio(rw, bio, mirror_num,
2822                                              prev_bio_flags);
2823                         if (ret < 0) {
2824                                 *bio_ret = NULL;
2825                                 return ret;
2826                         }
2827                         bio = NULL;
2828                 } else {
2829                         if (wbc)
2830                                 wbc_account_io(wbc, page, page_size);
2831                         return 0;
2832                 }
2833         }
2834
2835         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2836                         GFP_NOFS | __GFP_HIGH);
2837         if (!bio)
2838                 return -ENOMEM;
2839
2840         bio_add_page(bio, page, page_size, offset);
2841         bio->bi_end_io = end_io_func;
2842         bio->bi_private = tree;
2843         if (wbc) {
2844                 wbc_init_bio(wbc, bio);
2845                 wbc_account_io(wbc, page, page_size);
2846         }
2847
2848         if (bio_ret)
2849                 *bio_ret = bio;
2850         else
2851                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2852
2853         return ret;
2854 }
2855
2856 static void attach_extent_buffer_page(struct extent_buffer *eb,
2857                                       struct page *page)
2858 {
2859         if (!PagePrivate(page)) {
2860                 SetPagePrivate(page);
2861                 page_cache_get(page);
2862                 set_page_private(page, (unsigned long)eb);
2863         } else {
2864                 WARN_ON(page->private != (unsigned long)eb);
2865         }
2866 }
2867
2868 void set_page_extent_mapped(struct page *page)
2869 {
2870         if (!PagePrivate(page)) {
2871                 SetPagePrivate(page);
2872                 page_cache_get(page);
2873                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2874         }
2875 }
2876
2877 static struct extent_map *
2878 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2879                  u64 start, u64 len, get_extent_t *get_extent,
2880                  struct extent_map **em_cached)
2881 {
2882         struct extent_map *em;
2883
2884         if (em_cached && *em_cached) {
2885                 em = *em_cached;
2886                 if (extent_map_in_tree(em) && start >= em->start &&
2887                     start < extent_map_end(em)) {
2888                         atomic_inc(&em->refs);
2889                         return em;
2890                 }
2891
2892                 free_extent_map(em);
2893                 *em_cached = NULL;
2894         }
2895
2896         em = get_extent(inode, page, pg_offset, start, len, 0);
2897         if (em_cached && !IS_ERR_OR_NULL(em)) {
2898                 BUG_ON(*em_cached);
2899                 atomic_inc(&em->refs);
2900                 *em_cached = em;
2901         }
2902         return em;
2903 }
2904 /*
2905  * basic readpage implementation.  Locked extent state structs are inserted
2906  * into the tree that are removed when the IO is done (by the end_io
2907  * handlers)
2908  * XXX JDM: This needs looking at to ensure proper page locking
2909  */
2910 static int __do_readpage(struct extent_io_tree *tree,
2911                          struct page *page,
2912                          get_extent_t *get_extent,
2913                          struct extent_map **em_cached,
2914                          struct bio **bio, int mirror_num,
2915                          unsigned long *bio_flags, int rw,
2916                          u64 *prev_em_start)
2917 {
2918         struct inode *inode = page->mapping->host;
2919         u64 start = page_offset(page);
2920         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2921         u64 end;
2922         u64 cur = start;
2923         u64 extent_offset;
2924         u64 last_byte = i_size_read(inode);
2925         u64 block_start;
2926         u64 cur_end;
2927         sector_t sector;
2928         struct extent_map *em;
2929         struct block_device *bdev;
2930         int ret;
2931         int nr = 0;
2932         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2933         size_t pg_offset = 0;
2934         size_t iosize;
2935         size_t disk_io_size;
2936         size_t blocksize = inode->i_sb->s_blocksize;
2937         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2938
2939         set_page_extent_mapped(page);
2940
2941         end = page_end;
2942         if (!PageUptodate(page)) {
2943                 if (cleancache_get_page(page) == 0) {
2944                         BUG_ON(blocksize != PAGE_SIZE);
2945                         unlock_extent(tree, start, end);
2946                         goto out;
2947                 }
2948         }
2949
2950         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2951                 char *userpage;
2952                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2953
2954                 if (zero_offset) {
2955                         iosize = PAGE_CACHE_SIZE - zero_offset;
2956                         userpage = kmap_atomic(page);
2957                         memset(userpage + zero_offset, 0, iosize);
2958                         flush_dcache_page(page);
2959                         kunmap_atomic(userpage);
2960                 }
2961         }
2962         while (cur <= end) {
2963                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2964                 bool force_bio_submit = false;
2965
2966                 if (cur >= last_byte) {
2967                         char *userpage;
2968                         struct extent_state *cached = NULL;
2969
2970                         iosize = PAGE_CACHE_SIZE - pg_offset;
2971                         userpage = kmap_atomic(page);
2972                         memset(userpage + pg_offset, 0, iosize);
2973                         flush_dcache_page(page);
2974                         kunmap_atomic(userpage);
2975                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2976                                             &cached, GFP_NOFS);
2977                         if (!parent_locked)
2978                                 unlock_extent_cached(tree, cur,
2979                                                      cur + iosize - 1,
2980                                                      &cached, GFP_NOFS);
2981                         break;
2982                 }
2983                 em = __get_extent_map(inode, page, pg_offset, cur,
2984                                       end - cur + 1, get_extent, em_cached);
2985                 if (IS_ERR_OR_NULL(em)) {
2986                         SetPageError(page);
2987                         if (!parent_locked)
2988                                 unlock_extent(tree, cur, end);
2989                         break;
2990                 }
2991                 extent_offset = cur - em->start;
2992                 BUG_ON(extent_map_end(em) <= cur);
2993                 BUG_ON(end < cur);
2994
2995                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2996                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2997                         extent_set_compress_type(&this_bio_flag,
2998                                                  em->compress_type);
2999                 }
3000
3001                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
3002                 cur_end = min(extent_map_end(em) - 1, end);
3003                 iosize = ALIGN(iosize, blocksize);
3004                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
3005                         disk_io_size = em->block_len;
3006                         sector = em->block_start >> 9;
3007                 } else {
3008                         sector = (em->block_start + extent_offset) >> 9;
3009                         disk_io_size = iosize;
3010                 }
3011                 bdev = em->bdev;
3012                 block_start = em->block_start;
3013                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3014                         block_start = EXTENT_MAP_HOLE;
3015
3016                 /*
3017                  * If we have a file range that points to a compressed extent
3018                  * and it's followed by a consecutive file range that points to
3019                  * to the same compressed extent (possibly with a different
3020                  * offset and/or length, so it either points to the whole extent
3021                  * or only part of it), we must make sure we do not submit a
3022                  * single bio to populate the pages for the 2 ranges because
3023                  * this makes the compressed extent read zero out the pages
3024                  * belonging to the 2nd range. Imagine the following scenario:
3025                  *
3026                  *  File layout
3027                  *  [0 - 8K]                     [8K - 24K]
3028                  *    |                               |
3029                  *    |                               |
3030                  * points to extent X,         points to extent X,
3031                  * offset 4K, length of 8K     offset 0, length 16K
3032                  *
3033                  * [extent X, compressed length = 4K uncompressed length = 16K]
3034                  *
3035                  * If the bio to read the compressed extent covers both ranges,
3036                  * it will decompress extent X into the pages belonging to the
3037                  * first range and then it will stop, zeroing out the remaining
3038                  * pages that belong to the other range that points to extent X.
3039                  * So here we make sure we submit 2 bios, one for the first
3040                  * range and another one for the third range. Both will target
3041                  * the same physical extent from disk, but we can't currently
3042                  * make the compressed bio endio callback populate the pages
3043                  * for both ranges because each compressed bio is tightly
3044                  * coupled with a single extent map, and each range can have
3045                  * an extent map with a different offset value relative to the
3046                  * uncompressed data of our extent and different lengths. This
3047                  * is a corner case so we prioritize correctness over
3048                  * non-optimal behavior (submitting 2 bios for the same extent).
3049                  */
3050                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3051                     prev_em_start && *prev_em_start != (u64)-1 &&
3052                     *prev_em_start != em->orig_start)
3053                         force_bio_submit = true;
3054
3055                 if (prev_em_start)
3056                         *prev_em_start = em->orig_start;
3057
3058                 free_extent_map(em);
3059                 em = NULL;
3060
3061                 /* we've found a hole, just zero and go on */
3062                 if (block_start == EXTENT_MAP_HOLE) {
3063                         char *userpage;
3064                         struct extent_state *cached = NULL;
3065
3066                         userpage = kmap_atomic(page);
3067                         memset(userpage + pg_offset, 0, iosize);
3068                         flush_dcache_page(page);
3069                         kunmap_atomic(userpage);
3070
3071                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3072                                             &cached, GFP_NOFS);
3073                         unlock_extent_cached(tree, cur, cur + iosize - 1,
3074                                              &cached, GFP_NOFS);
3075                         cur = cur + iosize;
3076                         pg_offset += iosize;
3077                         continue;
3078                 }
3079                 /* the get_extent function already copied into the page */
3080                 if (test_range_bit(tree, cur, cur_end,
3081                                    EXTENT_UPTODATE, 1, NULL)) {
3082                         check_page_uptodate(tree, page);
3083                         if (!parent_locked)
3084                                 unlock_extent(tree, cur, cur + iosize - 1);
3085                         cur = cur + iosize;
3086                         pg_offset += iosize;
3087                         continue;
3088                 }
3089                 /* we have an inline extent but it didn't get marked up
3090                  * to date.  Error out
3091                  */
3092                 if (block_start == EXTENT_MAP_INLINE) {
3093                         SetPageError(page);
3094                         if (!parent_locked)
3095                                 unlock_extent(tree, cur, cur + iosize - 1);
3096                         cur = cur + iosize;
3097                         pg_offset += iosize;
3098                         continue;
3099                 }
3100
3101                 pnr -= page->index;
3102                 ret = submit_extent_page(rw, tree, NULL, page,
3103                                          sector, disk_io_size, pg_offset,
3104                                          bdev, bio, pnr,
3105                                          end_bio_extent_readpage, mirror_num,
3106                                          *bio_flags,
3107                                          this_bio_flag,
3108                                          force_bio_submit);
3109                 if (!ret) {
3110                         nr++;
3111                         *bio_flags = this_bio_flag;
3112                 } else {
3113                         SetPageError(page);
3114                         if (!parent_locked)
3115                                 unlock_extent(tree, cur, cur + iosize - 1);
3116                 }
3117                 cur = cur + iosize;
3118                 pg_offset += iosize;
3119         }
3120 out:
3121         if (!nr) {
3122                 if (!PageError(page))
3123                         SetPageUptodate(page);
3124                 unlock_page(page);
3125         }
3126         return 0;
3127 }
3128
3129 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3130                                              struct page *pages[], int nr_pages,
3131                                              u64 start, u64 end,
3132                                              get_extent_t *get_extent,
3133                                              struct extent_map **em_cached,
3134                                              struct bio **bio, int mirror_num,
3135                                              unsigned long *bio_flags, int rw)
3136 {
3137         struct inode *inode;
3138         struct btrfs_ordered_extent *ordered;
3139         int index;
3140         u64 prev_em_start = (u64)-1;
3141
3142         inode = pages[0]->mapping->host;
3143         while (1) {
3144                 lock_extent(tree, start, end);
3145                 ordered = btrfs_lookup_ordered_range(inode, start,
3146                                                      end - start + 1);
3147                 if (!ordered)
3148                         break;
3149                 unlock_extent(tree, start, end);
3150                 btrfs_start_ordered_extent(inode, ordered, 1);
3151                 btrfs_put_ordered_extent(ordered);
3152         }
3153
3154         for (index = 0; index < nr_pages; index++) {
3155                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3156                               mirror_num, bio_flags, rw, &prev_em_start);
3157                 page_cache_release(pages[index]);
3158         }
3159 }
3160
3161 static void __extent_readpages(struct extent_io_tree *tree,
3162                                struct page *pages[],
3163                                int nr_pages, get_extent_t *get_extent,
3164                                struct extent_map **em_cached,
3165                                struct bio **bio, int mirror_num,
3166                                unsigned long *bio_flags, int rw)
3167 {
3168         u64 start = 0;
3169         u64 end = 0;
3170         u64 page_start;
3171         int index;
3172         int first_index = 0;
3173
3174         for (index = 0; index < nr_pages; index++) {
3175                 page_start = page_offset(pages[index]);
3176                 if (!end) {
3177                         start = page_start;
3178                         end = start + PAGE_CACHE_SIZE - 1;
3179                         first_index = index;
3180                 } else if (end + 1 == page_start) {
3181                         end += PAGE_CACHE_SIZE;
3182                 } else {
3183                         __do_contiguous_readpages(tree, &pages[first_index],
3184                                                   index - first_index, start,
3185                                                   end, get_extent, em_cached,
3186                                                   bio, mirror_num, bio_flags,
3187                                                   rw);
3188                         start = page_start;
3189                         end = start + PAGE_CACHE_SIZE - 1;
3190                         first_index = index;
3191                 }
3192         }
3193
3194         if (end)
3195                 __do_contiguous_readpages(tree, &pages[first_index],
3196                                           index - first_index, start,
3197                                           end, get_extent, em_cached, bio,
3198                                           mirror_num, bio_flags, rw);
3199 }
3200
3201 static int __extent_read_full_page(struct extent_io_tree *tree,
3202                                    struct page *page,
3203                                    get_extent_t *get_extent,
3204                                    struct bio **bio, int mirror_num,
3205                                    unsigned long *bio_flags, int rw)
3206 {
3207         struct inode *inode = page->mapping->host;
3208         struct btrfs_ordered_extent *ordered;
3209         u64 start = page_offset(page);
3210         u64 end = start + PAGE_CACHE_SIZE - 1;
3211         int ret;
3212
3213         while (1) {
3214                 lock_extent(tree, start, end);
3215                 ordered = btrfs_lookup_ordered_extent(inode, start);
3216                 if (!ordered)
3217                         break;
3218                 unlock_extent(tree, start, end);
3219                 btrfs_start_ordered_extent(inode, ordered, 1);
3220                 btrfs_put_ordered_extent(ordered);
3221         }
3222
3223         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3224                             bio_flags, rw, NULL);
3225         return ret;
3226 }
3227
3228 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3229                             get_extent_t *get_extent, int mirror_num)
3230 {
3231         struct bio *bio = NULL;
3232         unsigned long bio_flags = 0;
3233         int ret;
3234
3235         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3236                                       &bio_flags, READ);
3237         if (bio)
3238                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3239         return ret;
3240 }
3241
3242 int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3243                                  get_extent_t *get_extent, int mirror_num)
3244 {
3245         struct bio *bio = NULL;
3246         unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3247         int ret;
3248
3249         ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3250                             &bio_flags, READ, NULL);
3251         if (bio)
3252                 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3253         return ret;
3254 }
3255
3256 static noinline void update_nr_written(struct page *page,
3257                                       struct writeback_control *wbc,
3258                                       unsigned long nr_written)
3259 {
3260         wbc->nr_to_write -= nr_written;
3261         if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3262             wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3263                 page->mapping->writeback_index = page->index + nr_written;
3264 }
3265
3266 /*
3267  * helper for __extent_writepage, doing all of the delayed allocation setup.
3268  *
3269  * This returns 1 if our fill_delalloc function did all the work required
3270  * to write the page (copy into inline extent).  In this case the IO has
3271  * been started and the page is already unlocked.
3272  *
3273  * This returns 0 if all went well (page still locked)
3274  * This returns < 0 if there were errors (page still locked)
3275  */
3276 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3277                               struct page *page, struct writeback_control *wbc,
3278                               struct extent_page_data *epd,
3279                               u64 delalloc_start,
3280                               unsigned long *nr_written)
3281 {
3282         struct extent_io_tree *tree = epd->tree;
3283         u64 page_end = delalloc_start + PAGE_CACHE_SIZE - 1;
3284         u64 nr_delalloc;
3285         u64