Merge branches 'arm/rockchip', 'arm/exynos', 'arm/smmu', 'x86/vt-d', 'x86/amd', ...
[sfrench/cifs-2.6.git] / fs / btrfs / extent_io.c
1 #include <linux/bitops.h>
2 #include <linux/slab.h>
3 #include <linux/bio.h>
4 #include <linux/mm.h>
5 #include <linux/pagemap.h>
6 #include <linux/page-flags.h>
7 #include <linux/spinlock.h>
8 #include <linux/blkdev.h>
9 #include <linux/swap.h>
10 #include <linux/writeback.h>
11 #include <linux/pagevec.h>
12 #include <linux/prefetch.h>
13 #include <linux/cleancache.h>
14 #include "extent_io.h"
15 #include "extent_map.h"
16 #include "ctree.h"
17 #include "btrfs_inode.h"
18 #include "volumes.h"
19 #include "check-integrity.h"
20 #include "locking.h"
21 #include "rcu-string.h"
22 #include "backref.h"
23
24 static struct kmem_cache *extent_state_cache;
25 static struct kmem_cache *extent_buffer_cache;
26 static struct bio_set *btrfs_bioset;
27
28 static inline bool extent_state_in_tree(const struct extent_state *state)
29 {
30         return !RB_EMPTY_NODE(&state->rb_node);
31 }
32
33 #ifdef CONFIG_BTRFS_DEBUG
34 static LIST_HEAD(buffers);
35 static LIST_HEAD(states);
36
37 static DEFINE_SPINLOCK(leak_lock);
38
39 static inline
40 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
41 {
42         unsigned long flags;
43
44         spin_lock_irqsave(&leak_lock, flags);
45         list_add(new, head);
46         spin_unlock_irqrestore(&leak_lock, flags);
47 }
48
49 static inline
50 void btrfs_leak_debug_del(struct list_head *entry)
51 {
52         unsigned long flags;
53
54         spin_lock_irqsave(&leak_lock, flags);
55         list_del(entry);
56         spin_unlock_irqrestore(&leak_lock, flags);
57 }
58
59 static inline
60 void btrfs_leak_debug_check(void)
61 {
62         struct extent_state *state;
63         struct extent_buffer *eb;
64
65         while (!list_empty(&states)) {
66                 state = list_entry(states.next, struct extent_state, leak_list);
67                 pr_err("BTRFS: state leak: start %llu end %llu state %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         return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1281 }
1282
1283 int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1284                         struct extent_state **cached_state, gfp_t mask)
1285 {
1286         return set_extent_bit(tree, start, end,
1287                               EXTENT_DELALLOC | EXTENT_UPTODATE,
1288                               NULL, cached_state, mask);
1289 }
1290
1291 int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1292                       struct extent_state **cached_state, gfp_t mask)
1293 {
1294         return set_extent_bit(tree, start, end,
1295                               EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1296                               NULL, cached_state, mask);
1297 }
1298
1299 int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1300                        gfp_t mask)
1301 {
1302         return clear_extent_bit(tree, start, end,
1303                                 EXTENT_DIRTY | EXTENT_DELALLOC |
1304                                 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1305 }
1306
1307 int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1308                      gfp_t mask)
1309 {
1310         return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1311                               NULL, mask);
1312 }
1313
1314 int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1315                         struct extent_state **cached_state, gfp_t mask)
1316 {
1317         return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1318                               cached_state, mask);
1319 }
1320
1321 int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1322                           struct extent_state **cached_state, gfp_t mask)
1323 {
1324         return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1325                                 cached_state, mask);
1326 }
1327
1328 /*
1329  * either insert or lock state struct between start and end use mask to tell
1330  * us if waiting is desired.
1331  */
1332 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1333                      unsigned bits, struct extent_state **cached_state)
1334 {
1335         int err;
1336         u64 failed_start;
1337
1338         while (1) {
1339                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1340                                        EXTENT_LOCKED, &failed_start,
1341                                        cached_state, GFP_NOFS);
1342                 if (err == -EEXIST) {
1343                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1344                         start = failed_start;
1345                 } else
1346                         break;
1347                 WARN_ON(start > end);
1348         }
1349         return err;
1350 }
1351
1352 int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1353 {
1354         return lock_extent_bits(tree, start, end, 0, NULL);
1355 }
1356
1357 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1358 {
1359         int err;
1360         u64 failed_start;
1361
1362         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1363                                &failed_start, NULL, GFP_NOFS);
1364         if (err == -EEXIST) {
1365                 if (failed_start > start)
1366                         clear_extent_bit(tree, start, failed_start - 1,
1367                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1368                 return 0;
1369         }
1370         return 1;
1371 }
1372
1373 int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1374                          struct extent_state **cached, gfp_t mask)
1375 {
1376         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1377                                 mask);
1378 }
1379
1380 int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1381 {
1382         return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1383                                 GFP_NOFS);
1384 }
1385
1386 int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1387 {
1388         unsigned long index = start >> PAGE_CACHE_SHIFT;
1389         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1390         struct page *page;
1391
1392         while (index <= end_index) {
1393                 page = find_get_page(inode->i_mapping, index);
1394                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1395                 clear_page_dirty_for_io(page);
1396                 page_cache_release(page);
1397                 index++;
1398         }
1399         return 0;
1400 }
1401
1402 int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1403 {
1404         unsigned long index = start >> PAGE_CACHE_SHIFT;
1405         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1406         struct page *page;
1407
1408         while (index <= end_index) {
1409                 page = find_get_page(inode->i_mapping, index);
1410                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1411                 __set_page_dirty_nobuffers(page);
1412                 account_page_redirty(page);
1413                 page_cache_release(page);
1414                 index++;
1415         }
1416         return 0;
1417 }
1418
1419 /*
1420  * helper function to set both pages and extents in the tree writeback
1421  */
1422 static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1423 {
1424         unsigned long index = start >> PAGE_CACHE_SHIFT;
1425         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1426         struct page *page;
1427
1428         while (index <= end_index) {
1429                 page = find_get_page(tree->mapping, index);
1430                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1431                 set_page_writeback(page);
1432                 page_cache_release(page);
1433                 index++;
1434         }
1435         return 0;
1436 }
1437
1438 /* find the first state struct with 'bits' set after 'start', and
1439  * return it.  tree->lock must be held.  NULL will returned if
1440  * nothing was found after 'start'
1441  */
1442 static struct extent_state *
1443 find_first_extent_bit_state(struct extent_io_tree *tree,
1444                             u64 start, unsigned bits)
1445 {
1446         struct rb_node *node;
1447         struct extent_state *state;
1448
1449         /*
1450          * this search will find all the extents that end after
1451          * our range starts.
1452          */
1453         node = tree_search(tree, start);
1454         if (!node)
1455                 goto out;
1456
1457         while (1) {
1458                 state = rb_entry(node, struct extent_state, rb_node);
1459                 if (state->end >= start && (state->state & bits))
1460                         return state;
1461
1462                 node = rb_next(node);
1463                 if (!node)
1464                         break;
1465         }
1466 out:
1467         return NULL;
1468 }
1469
1470 /*
1471  * find the first offset in the io tree with 'bits' set. zero is
1472  * returned if we find something, and *start_ret and *end_ret are
1473  * set to reflect the state struct that was found.
1474  *
1475  * If nothing was found, 1 is returned. If found something, return 0.
1476  */
1477 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1478                           u64 *start_ret, u64 *end_ret, unsigned bits,
1479                           struct extent_state **cached_state)
1480 {
1481         struct extent_state *state;
1482         struct rb_node *n;
1483         int ret = 1;
1484
1485         spin_lock(&tree->lock);
1486         if (cached_state && *cached_state) {
1487                 state = *cached_state;
1488                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1489                         n = rb_next(&state->rb_node);
1490                         while (n) {
1491                                 state = rb_entry(n, struct extent_state,
1492                                                  rb_node);
1493                                 if (state->state & bits)
1494                                         goto got_it;
1495                                 n = rb_next(n);
1496                         }
1497                         free_extent_state(*cached_state);
1498                         *cached_state = NULL;
1499                         goto out;
1500                 }
1501                 free_extent_state(*cached_state);
1502                 *cached_state = NULL;
1503         }
1504
1505         state = find_first_extent_bit_state(tree, start, bits);
1506 got_it:
1507         if (state) {
1508                 cache_state_if_flags(state, cached_state, 0);
1509                 *start_ret = state->start;
1510                 *end_ret = state->end;
1511                 ret = 0;
1512         }
1513 out:
1514         spin_unlock(&tree->lock);
1515         return ret;
1516 }
1517
1518 /*
1519  * find a contiguous range of bytes in the file marked as delalloc, not
1520  * more than 'max_bytes'.  start and end are used to return the range,
1521  *
1522  * 1 is returned if we find something, 0 if nothing was in the tree
1523  */
1524 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1525                                         u64 *start, u64 *end, u64 max_bytes,
1526                                         struct extent_state **cached_state)
1527 {
1528         struct rb_node *node;
1529         struct extent_state *state;
1530         u64 cur_start = *start;
1531         u64 found = 0;
1532         u64 total_bytes = 0;
1533
1534         spin_lock(&tree->lock);
1535
1536         /*
1537          * this search will find all the extents that end after
1538          * our range starts.
1539          */
1540         node = tree_search(tree, cur_start);
1541         if (!node) {
1542                 if (!found)
1543                         *end = (u64)-1;
1544                 goto out;
1545         }
1546
1547         while (1) {
1548                 state = rb_entry(node, struct extent_state, rb_node);
1549                 if (found && (state->start != cur_start ||
1550                               (state->state & EXTENT_BOUNDARY))) {
1551                         goto out;
1552                 }
1553                 if (!(state->state & EXTENT_DELALLOC)) {
1554                         if (!found)
1555                                 *end = state->end;
1556                         goto out;
1557                 }
1558                 if (!found) {
1559                         *start = state->start;
1560                         *cached_state = state;
1561                         atomic_inc(&state->refs);
1562                 }
1563                 found++;
1564                 *end = state->end;
1565                 cur_start = state->end + 1;
1566                 node = rb_next(node);
1567                 total_bytes += state->end - state->start + 1;
1568                 if (total_bytes >= max_bytes)
1569                         break;
1570                 if (!node)
1571                         break;
1572         }
1573 out:
1574         spin_unlock(&tree->lock);
1575         return found;
1576 }
1577
1578 static noinline void __unlock_for_delalloc(struct inode *inode,
1579                                            struct page *locked_page,
1580                                            u64 start, u64 end)
1581 {
1582         int ret;
1583         struct page *pages[16];
1584         unsigned long index = start >> PAGE_CACHE_SHIFT;
1585         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1586         unsigned long nr_pages = end_index - index + 1;
1587         int i;
1588
1589         if (index == locked_page->index && end_index == index)
1590                 return;
1591
1592         while (nr_pages > 0) {
1593                 ret = find_get_pages_contig(inode->i_mapping, index,
1594                                      min_t(unsigned long, nr_pages,
1595                                      ARRAY_SIZE(pages)), pages);
1596                 for (i = 0; i < ret; i++) {
1597                         if (pages[i] != locked_page)
1598                                 unlock_page(pages[i]);
1599                         page_cache_release(pages[i]);
1600                 }
1601                 nr_pages -= ret;
1602                 index += ret;
1603                 cond_resched();
1604         }
1605 }
1606
1607 static noinline int lock_delalloc_pages(struct inode *inode,
1608                                         struct page *locked_page,
1609                                         u64 delalloc_start,
1610                                         u64 delalloc_end)
1611 {
1612         unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1613         unsigned long start_index = index;
1614         unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1615         unsigned long pages_locked = 0;
1616         struct page *pages[16];
1617         unsigned long nrpages;
1618         int ret;
1619         int i;
1620
1621         /* the caller is responsible for locking the start index */
1622         if (index == locked_page->index && index == end_index)
1623                 return 0;
1624
1625         /* skip the page at the start index */
1626         nrpages = end_index - index + 1;
1627         while (nrpages > 0) {
1628                 ret = find_get_pages_contig(inode->i_mapping, index,
1629                                      min_t(unsigned long,
1630                                      nrpages, ARRAY_SIZE(pages)), pages);
1631                 if (ret == 0) {
1632                         ret = -EAGAIN;
1633                         goto done;
1634                 }
1635                 /* now we have an array of pages, lock them all */
1636                 for (i = 0; i < ret; i++) {
1637                         /*
1638                          * the caller is taking responsibility for
1639                          * locked_page
1640                          */
1641                         if (pages[i] != locked_page) {
1642                                 lock_page(pages[i]);
1643                                 if (!PageDirty(pages[i]) ||
1644                                     pages[i]->mapping != inode->i_mapping) {
1645                                         ret = -EAGAIN;
1646                                         unlock_page(pages[i]);
1647                                         page_cache_release(pages[i]);
1648                                         goto done;
1649                                 }
1650                         }
1651                         page_cache_release(pages[i]);
1652                         pages_locked++;
1653                 }
1654                 nrpages -= ret;
1655                 index += ret;
1656                 cond_resched();
1657         }
1658         ret = 0;
1659 done:
1660         if (ret && pages_locked) {
1661                 __unlock_for_delalloc(inode, locked_page,
1662                               delalloc_start,
1663                               ((u64)(start_index + pages_locked - 1)) <<
1664                               PAGE_CACHE_SHIFT);
1665         }
1666         return ret;
1667 }
1668
1669 /*
1670  * find a contiguous range of bytes in the file marked as delalloc, not
1671  * more than 'max_bytes'.  start and end are used to return the range,
1672  *
1673  * 1 is returned if we find something, 0 if nothing was in the tree
1674  */
1675 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1676                                     struct extent_io_tree *tree,
1677                                     struct page *locked_page, u64 *start,
1678                                     u64 *end, u64 max_bytes)
1679 {
1680         u64 delalloc_start;
1681         u64 delalloc_end;
1682         u64 found;
1683         struct extent_state *cached_state = NULL;
1684         int ret;
1685         int loops = 0;
1686
1687 again:
1688         /* step one, find a bunch of delalloc bytes starting at start */
1689         delalloc_start = *start;
1690         delalloc_end = 0;
1691         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1692                                     max_bytes, &cached_state);
1693         if (!found || delalloc_end <= *start) {
1694                 *start = delalloc_start;
1695                 *end = delalloc_end;
1696                 free_extent_state(cached_state);
1697                 return 0;
1698         }
1699
1700         /*
1701          * start comes from the offset of locked_page.  We have to lock
1702          * pages in order, so we can't process delalloc bytes before
1703          * locked_page
1704          */
1705         if (delalloc_start < *start)
1706                 delalloc_start = *start;
1707
1708         /*
1709          * make sure to limit the number of pages we try to lock down
1710          */
1711         if (delalloc_end + 1 - delalloc_start > max_bytes)
1712                 delalloc_end = delalloc_start + max_bytes - 1;
1713
1714         /* step two, lock all the pages after the page that has start */
1715         ret = lock_delalloc_pages(inode, locked_page,
1716                                   delalloc_start, delalloc_end);
1717         if (ret == -EAGAIN) {
1718                 /* some of the pages are gone, lets avoid looping by
1719                  * shortening the size of the delalloc range we're searching
1720                  */
1721                 free_extent_state(cached_state);
1722                 cached_state = NULL;
1723                 if (!loops) {
1724                         max_bytes = PAGE_CACHE_SIZE;
1725                         loops = 1;
1726                         goto again;
1727                 } else {
1728                         found = 0;
1729                         goto out_failed;
1730                 }
1731         }
1732         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1733
1734         /* step three, lock the state bits for the whole range */
1735         lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1736
1737         /* then test to make sure it is all still delalloc */
1738         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1739                              EXTENT_DELALLOC, 1, cached_state);
1740         if (!ret) {
1741                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1742                                      &cached_state, GFP_NOFS);
1743                 __unlock_for_delalloc(inode, locked_page,
1744                               delalloc_start, delalloc_end);
1745                 cond_resched();
1746                 goto again;
1747         }
1748         free_extent_state(cached_state);
1749         *start = delalloc_start;
1750         *end = delalloc_end;
1751 out_failed:
1752         return found;
1753 }
1754
1755 int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1756                                  struct page *locked_page,
1757                                  unsigned clear_bits,
1758                                  unsigned long page_ops)
1759 {
1760         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1761         int ret;
1762         struct page *pages[16];
1763         unsigned long index = start >> PAGE_CACHE_SHIFT;
1764         unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1765         unsigned long nr_pages = end_index - index + 1;
1766         int i;
1767
1768         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1769         if (page_ops == 0)
1770                 return 0;
1771
1772         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1773                 mapping_set_error(inode->i_mapping, -EIO);
1774
1775         while (nr_pages > 0) {
1776                 ret = find_get_pages_contig(inode->i_mapping, index,
1777                                      min_t(unsigned long,
1778                                      nr_pages, ARRAY_SIZE(pages)), pages);
1779                 for (i = 0; i < ret; i++) {
1780
1781                         if (page_ops & PAGE_SET_PRIVATE2)
1782                                 SetPagePrivate2(pages[i]);
1783
1784                         if (pages[i] == locked_page) {
1785                                 page_cache_release(pages[i]);
1786                                 continue;
1787                         }
1788                         if (page_ops & PAGE_CLEAR_DIRTY)
1789                                 clear_page_dirty_for_io(pages[i]);
1790                         if (page_ops & PAGE_SET_WRITEBACK)
1791                                 set_page_writeback(pages[i]);
1792                         if (page_ops & PAGE_SET_ERROR)
1793                                 SetPageError(pages[i]);
1794                         if (page_ops & PAGE_END_WRITEBACK)
1795                                 end_page_writeback(pages[i]);
1796                         if (page_ops & PAGE_UNLOCK)
1797                                 unlock_page(pages[i]);
1798                         page_cache_release(pages[i]);
1799                 }
1800                 nr_pages -= ret;
1801                 index += ret;
1802                 cond_resched();
1803         }
1804         return 0;
1805 }
1806
1807 /*
1808  * count the number of bytes in the tree that have a given bit(s)
1809  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1810  * cached.  The total number found is returned.
1811  */
1812 u64 count_range_bits(struct extent_io_tree *tree,
1813                      u64 *start, u64 search_end, u64 max_bytes,
1814                      unsigned bits, int contig)
1815 {
1816         struct rb_node *node;
1817         struct extent_state *state;
1818         u64 cur_start = *start;
1819         u64 total_bytes = 0;
1820         u64 last = 0;
1821         int found = 0;
1822
1823         if (WARN_ON(search_end <= cur_start))
1824                 return 0;
1825
1826         spin_lock(&tree->lock);
1827         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1828                 total_bytes = tree->dirty_bytes;
1829                 goto out;
1830         }
1831         /*
1832          * this search will find all the extents that end after
1833          * our range starts.
1834          */
1835         node = tree_search(tree, cur_start);
1836         if (!node)
1837                 goto out;
1838
1839         while (1) {
1840                 state = rb_entry(node, struct extent_state, rb_node);
1841                 if (state->start > search_end)
1842                         break;
1843                 if (contig && found && state->start > last + 1)
1844                         break;
1845                 if (state->end >= cur_start && (state->state & bits) == bits) {
1846                         total_bytes += min(search_end, state->end) + 1 -
1847                                        max(cur_start, state->start);
1848                         if (total_bytes >= max_bytes)
1849                                 break;
1850                         if (!found) {
1851                                 *start = max(cur_start, state->start);
1852                                 found = 1;
1853                         }
1854                         last = state->end;
1855                 } else if (contig && found) {
1856                         break;
1857                 }
1858                 node = rb_next(node);
1859                 if (!node)
1860                         break;
1861         }
1862 out:
1863         spin_unlock(&tree->lock);
1864         return total_bytes;
1865 }
1866
1867 /*
1868  * set the private field for a given byte offset in the tree.  If there isn't
1869  * an extent_state there already, this does nothing.
1870  */
1871 static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1872 {
1873         struct rb_node *node;
1874         struct extent_state *state;
1875         int ret = 0;
1876
1877         spin_lock(&tree->lock);
1878         /*
1879          * this search will find all the extents that end after
1880          * our range starts.
1881          */
1882         node = tree_search(tree, start);
1883         if (!node) {
1884                 ret = -ENOENT;
1885                 goto out;
1886         }
1887         state = rb_entry(node, struct extent_state, rb_node);
1888         if (state->start != start) {
1889                 ret = -ENOENT;
1890                 goto out;
1891         }
1892         state->private = private;
1893 out:
1894         spin_unlock(&tree->lock);
1895         return ret;
1896 }
1897
1898 int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1899 {
1900         struct rb_node *node;
1901         struct extent_state *state;
1902         int ret = 0;
1903
1904         spin_lock(&tree->lock);
1905         /*
1906          * this search will find all the extents that end after
1907          * our range starts.
1908          */
1909         node = tree_search(tree, start);
1910         if (!node) {
1911                 ret = -ENOENT;
1912                 goto out;
1913         }
1914         state = rb_entry(node, struct extent_state, rb_node);
1915         if (state->start != start) {
1916                 ret = -ENOENT;
1917                 goto out;
1918         }
1919         *private = state->private;
1920 out:
1921         spin_unlock(&tree->lock);
1922         return ret;
1923 }
1924
1925 /*
1926  * searches a range in the state tree for a given mask.
1927  * If 'filled' == 1, this returns 1 only if every extent in the tree
1928  * has the bits set.  Otherwise, 1 is returned if any bit in the
1929  * range is found set.
1930  */
1931 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1932                    unsigned bits, int filled, struct extent_state *cached)
1933 {
1934         struct extent_state *state = NULL;
1935         struct rb_node *node;
1936         int bitset = 0;
1937
1938         spin_lock(&tree->lock);
1939         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1940             cached->end > start)
1941                 node = &cached->rb_node;
1942         else
1943                 node = tree_search(tree, start);
1944         while (node && start <= end) {
1945                 state = rb_entry(node, struct extent_state, rb_node);
1946
1947                 if (filled && state->start > start) {
1948                         bitset = 0;
1949                         break;
1950                 }
1951
1952                 if (state->start > end)
1953                         break;
1954
1955                 if (state->state & bits) {
1956                         bitset = 1;
1957                         if (!filled)
1958                                 break;
1959                 } else if (filled) {
1960                         bitset = 0;
1961                         break;
1962                 }
1963
1964                 if (state->end == (u64)-1)
1965                         break;
1966
1967                 start = state->end + 1;
1968                 if (start > end)
1969                         break;
1970                 node = rb_next(node);
1971                 if (!node) {
1972                         if (filled)
1973                                 bitset = 0;
1974                         break;
1975                 }
1976         }
1977         spin_unlock(&tree->lock);
1978         return bitset;
1979 }
1980
1981 /*
1982  * helper function to set a given page up to date if all the
1983  * extents in the tree for that page are up to date
1984  */
1985 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1986 {
1987         u64 start = page_offset(page);
1988         u64 end = start + PAGE_CACHE_SIZE - 1;
1989         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1990                 SetPageUptodate(page);
1991 }
1992
1993 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1994 {
1995         int ret;
1996         int err = 0;
1997         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1998
1999         set_state_private(failure_tree, rec->start, 0);
2000         ret = clear_extent_bits(failure_tree, rec->start,
2001                                 rec->start + rec->len - 1,
2002                                 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2003         if (ret)
2004                 err = ret;
2005
2006         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
2007                                 rec->start + rec->len - 1,
2008                                 EXTENT_DAMAGED, GFP_NOFS);
2009         if (ret && !err)
2010                 err = ret;
2011
2012         kfree(rec);
2013         return err;
2014 }
2015
2016 /*
2017  * this bypasses the standard btrfs submit functions deliberately, as
2018  * the standard behavior is to write all copies in a raid setup. here we only
2019  * want to write the one bad copy. so we do the mapping for ourselves and issue
2020  * submit_bio directly.
2021  * to avoid any synchronization issues, wait for the data after writing, which
2022  * actually prevents the read that triggered the error from finishing.
2023  * currently, there can be no more than two copies of every data bit. thus,
2024  * exactly one rewrite is required.
2025  */
2026 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2027                       struct page *page, unsigned int pg_offset, int mirror_num)
2028 {
2029         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2030         struct bio *bio;
2031         struct btrfs_device *dev;
2032         u64 map_length = 0;
2033         u64 sector;
2034         struct btrfs_bio *bbio = NULL;
2035         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2036         int ret;
2037
2038         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2039         BUG_ON(!mirror_num);
2040
2041         /* we can't repair anything in raid56 yet */
2042         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2043                 return 0;
2044
2045         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2046         if (!bio)
2047                 return -EIO;
2048         bio->bi_iter.bi_size = 0;
2049         map_length = length;
2050
2051         ret = btrfs_map_block(fs_info, WRITE, logical,
2052                               &map_length, &bbio, mirror_num);
2053         if (ret) {
2054                 bio_put(bio);
2055                 return -EIO;
2056         }
2057         BUG_ON(mirror_num != bbio->mirror_num);
2058         sector = bbio->stripes[mirror_num-1].physical >> 9;
2059         bio->bi_iter.bi_sector = sector;
2060         dev = bbio->stripes[mirror_num-1].dev;
2061         btrfs_put_bbio(bbio);
2062         if (!dev || !dev->bdev || !dev->writeable) {
2063                 bio_put(bio);
2064                 return -EIO;
2065         }
2066         bio->bi_bdev = dev->bdev;
2067         bio_add_page(bio, page, length, pg_offset);
2068
2069         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2070                 /* try to remap that extent elsewhere? */
2071                 bio_put(bio);
2072                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2073                 return -EIO;
2074         }
2075
2076         printk_ratelimited_in_rcu(KERN_INFO
2077                                   "BTRFS: read error corrected: ino %llu off %llu (dev %s sector %llu)\n",
2078                                   btrfs_ino(inode), start,
2079                                   rcu_str_deref(dev->name), sector);
2080         bio_put(bio);
2081         return 0;
2082 }
2083
2084 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2085                          int mirror_num)
2086 {
2087         u64 start = eb->start;
2088         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2089         int ret = 0;
2090
2091         if (root->fs_info->sb->s_flags & MS_RDONLY)
2092                 return -EROFS;
2093
2094         for (i = 0; i < num_pages; i++) {
2095                 struct page *p = eb->pages[i];
2096
2097                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2098                                         PAGE_CACHE_SIZE, start, p,
2099                                         start - page_offset(p), mirror_num);
2100                 if (ret)
2101                         break;
2102                 start += PAGE_CACHE_SIZE;
2103         }
2104
2105         return ret;
2106 }
2107
2108 /*
2109  * each time an IO finishes, we do a fast check in the IO failure tree
2110  * to see if we need to process or clean up an io_failure_record
2111  */
2112 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2113                      unsigned int pg_offset)
2114 {
2115         u64 private;
2116         u64 private_failure;
2117         struct io_failure_record *failrec;
2118         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2119         struct extent_state *state;
2120         int num_copies;
2121         int ret;
2122
2123         private = 0;
2124         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2125                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2126         if (!ret)
2127                 return 0;
2128
2129         ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2130                                 &private_failure);
2131         if (ret)
2132                 return 0;
2133
2134         failrec = (struct io_failure_record *)(unsigned long) private_failure;
2135         BUG_ON(!failrec->this_mirror);
2136
2137         if (failrec->in_validation) {
2138                 /* there was no real error, just free the record */
2139                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2140                          failrec->start);
2141                 goto out;
2142         }
2143         if (fs_info->sb->s_flags & MS_RDONLY)
2144                 goto out;
2145
2146         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2147         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2148                                             failrec->start,
2149                                             EXTENT_LOCKED);
2150         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2151
2152         if (state && state->start <= failrec->start &&
2153             state->end >= failrec->start + failrec->len - 1) {
2154                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2155                                               failrec->len);
2156                 if (num_copies > 1)  {
2157                         repair_io_failure(inode, start, failrec->len,
2158                                           failrec->logical, page,
2159                                           pg_offset, failrec->failed_mirror);
2160                 }
2161         }
2162
2163 out:
2164         free_io_failure(inode, failrec);
2165
2166         return 0;
2167 }
2168
2169 /*
2170  * Can be called when
2171  * - hold extent lock
2172  * - under ordered extent
2173  * - the inode is freeing
2174  */
2175 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2176 {
2177         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2178         struct io_failure_record *failrec;
2179         struct extent_state *state, *next;
2180
2181         if (RB_EMPTY_ROOT(&failure_tree->state))
2182                 return;
2183
2184         spin_lock(&failure_tree->lock);
2185         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2186         while (state) {
2187                 if (state->start > end)
2188                         break;
2189
2190                 ASSERT(state->end <= end);
2191
2192                 next = next_state(state);
2193
2194                 failrec = (struct io_failure_record *)(unsigned long)state->private;
2195                 free_extent_state(state);
2196                 kfree(failrec);
2197
2198                 state = next;
2199         }
2200         spin_unlock(&failure_tree->lock);
2201 }
2202
2203 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2204                                 struct io_failure_record **failrec_ret)
2205 {
2206         struct io_failure_record *failrec;
2207         u64 private;
2208         struct extent_map *em;
2209         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2210         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2211         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2212         int ret;
2213         u64 logical;
2214
2215         ret = get_state_private(failure_tree, start, &private);
2216         if (ret) {
2217                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2218                 if (!failrec)
2219                         return -ENOMEM;
2220
2221                 failrec->start = start;
2222                 failrec->len = end - start + 1;
2223                 failrec->this_mirror = 0;
2224                 failrec->bio_flags = 0;
2225                 failrec->in_validation = 0;
2226
2227                 read_lock(&em_tree->lock);
2228                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2229                 if (!em) {
2230                         read_unlock(&em_tree->lock);
2231                         kfree(failrec);
2232                         return -EIO;
2233                 }
2234
2235                 if (em->start > start || em->start + em->len <= start) {
2236                         free_extent_map(em);
2237                         em = NULL;
2238                 }
2239                 read_unlock(&em_tree->lock);
2240                 if (!em) {
2241                         kfree(failrec);
2242                         return -EIO;
2243                 }
2244
2245                 logical = start - em->start;
2246                 logical = em->block_start + logical;
2247                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2248                         logical = em->block_start;
2249                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2250                         extent_set_compress_type(&failrec->bio_flags,
2251                                                  em->compress_type);
2252                 }
2253
2254                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2255                          logical, start, failrec->len);
2256
2257                 failrec->logical = logical;
2258                 free_extent_map(em);
2259
2260                 /* set the bits in the private failure tree */
2261                 ret = set_extent_bits(failure_tree, start, end,
2262                                         EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2263                 if (ret >= 0)
2264                         ret = set_state_private(failure_tree, start,
2265                                                 (u64)(unsigned long)failrec);
2266                 /* set the bits in the inode's tree */
2267                 if (ret >= 0)
2268                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2269                                                 GFP_NOFS);
2270                 if (ret < 0) {
2271                         kfree(failrec);
2272                         return ret;
2273                 }
2274         } else {
2275                 failrec = (struct io_failure_record *)(unsigned long)private;
2276                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2277                          failrec->logical, failrec->start, failrec->len,
2278                          failrec->in_validation);
2279                 /*
2280                  * when data can be on disk more than twice, add to failrec here
2281                  * (e.g. with a list for failed_mirror) to make
2282                  * clean_io_failure() clean all those errors at once.
2283                  */
2284         }
2285
2286         *failrec_ret = failrec;
2287
2288         return 0;
2289 }
2290
2291 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2292                            struct io_failure_record *failrec, int failed_mirror)
2293 {
2294         int num_copies;
2295
2296         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2297                                       failrec->logical, failrec->len);
2298         if (num_copies == 1) {
2299                 /*
2300                  * we only have a single copy of the data, so don't bother with
2301                  * all the retry and error correction code that follows. no
2302                  * matter what the error is, it is very likely to persist.
2303                  */
2304                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2305                          num_copies, failrec->this_mirror, failed_mirror);
2306                 return 0;
2307         }
2308
2309         /*
2310          * there are two premises:
2311          *      a) deliver good data to the caller
2312          *      b) correct the bad sectors on disk
2313          */
2314         if (failed_bio->bi_vcnt > 1) {
2315                 /*
2316                  * to fulfill b), we need to know the exact failing sectors, as
2317                  * we don't want to rewrite any more than the failed ones. thus,
2318                  * we need separate read requests for the failed bio
2319                  *
2320                  * if the following BUG_ON triggers, our validation request got
2321                  * merged. we need separate requests for our algorithm to work.
2322                  */
2323                 BUG_ON(failrec->in_validation);
2324                 failrec->in_validation = 1;
2325                 failrec->this_mirror = failed_mirror;
2326         } else {
2327                 /*
2328                  * we're ready to fulfill a) and b) alongside. get a good copy
2329                  * of the failed sector and if we succeed, we have setup
2330                  * everything for repair_io_failure to do the rest for us.
2331                  */
2332                 if (failrec->in_validation) {
2333                         BUG_ON(failrec->this_mirror != failed_mirror);
2334                         failrec->in_validation = 0;
2335                         failrec->this_mirror = 0;
2336                 }
2337                 failrec->failed_mirror = failed_mirror;
2338                 failrec->this_mirror++;
2339                 if (failrec->this_mirror == failed_mirror)
2340                         failrec->this_mirror++;
2341         }
2342
2343         if (failrec->this_mirror > num_copies) {
2344                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2345                          num_copies, failrec->this_mirror, failed_mirror);
2346                 return 0;
2347         }
2348
2349         return 1;
2350 }
2351
2352
2353 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2354                                     struct io_failure_record *failrec,
2355                                     struct page *page, int pg_offset, int icsum,
2356                                     bio_end_io_t *endio_func, void *data)
2357 {
2358         struct bio *bio;
2359         struct btrfs_io_bio *btrfs_failed_bio;
2360         struct btrfs_io_bio *btrfs_bio;
2361
2362         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2363         if (!bio)
2364                 return NULL;
2365
2366         bio->bi_end_io = endio_func;
2367         bio->bi_iter.bi_sector = failrec->logical >> 9;
2368         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2369         bio->bi_iter.bi_size = 0;
2370         bio->bi_private = data;
2371
2372         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2373         if (btrfs_failed_bio->csum) {
2374                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2375                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2376
2377                 btrfs_bio = btrfs_io_bio(bio);
2378                 btrfs_bio->csum = btrfs_bio->csum_inline;
2379                 icsum *= csum_size;
2380                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2381                        csum_size);
2382         }
2383
2384         bio_add_page(bio, page, failrec->len, pg_offset);
2385
2386         return bio;
2387 }
2388
2389 /*
2390  * this is a generic handler for readpage errors (default
2391  * readpage_io_failed_hook). if other copies exist, read those and write back
2392  * good data to the failed position. does not investigate in remapping the
2393  * failed extent elsewhere, hoping the device will be smart enough to do this as
2394  * needed
2395  */
2396
2397 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2398                               struct page *page, u64 start, u64 end,
2399                               int failed_mirror)
2400 {
2401         struct io_failure_record *failrec;
2402         struct inode *inode = page->mapping->host;
2403         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2404         struct bio *bio;
2405         int read_mode;
2406         int ret;
2407
2408         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2409
2410         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2411         if (ret)
2412                 return ret;
2413
2414         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2415         if (!ret) {
2416                 free_io_failure(inode, failrec);
2417                 return -EIO;
2418         }
2419
2420         if (failed_bio->bi_vcnt > 1)
2421                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2422         else
2423                 read_mode = READ_SYNC;
2424
2425         phy_offset >>= inode->i_sb->s_blocksize_bits;
2426         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2427                                       start - page_offset(page),
2428                                       (int)phy_offset, failed_bio->bi_end_io,
2429                                       NULL);
2430         if (!bio) {
2431                 free_io_failure(inode, failrec);
2432                 return -EIO;
2433         }
2434
2435         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2436                  read_mode, failrec->this_mirror, failrec->in_validation);
2437
2438         ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2439                                          failrec->this_mirror,
2440                                          failrec->bio_flags, 0);
2441         if (ret) {
2442                 free_io_failure(inode, failrec);
2443                 bio_put(bio);
2444         }
2445
2446         return ret;
2447 }
2448
2449 /* lots and lots of room for performance fixes in the end_bio funcs */
2450
2451 int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2452 {
2453         int uptodate = (err == 0);
2454         struct extent_io_tree *tree;
2455         int ret = 0;
2456
2457         tree = &BTRFS_I(page->mapping->host)->io_tree;
2458
2459         if (tree->ops && tree->ops->writepage_end_io_hook) {
2460                 ret = tree->ops->writepage_end_io_hook(page, start,
2461                                                end, NULL, uptodate);
2462                 if (ret)
2463                         uptodate = 0;
2464         }
2465
2466         if (!uptodate) {
2467                 ClearPageUptodate(page);
2468                 SetPageError(page);
2469                 ret = ret < 0 ? ret : -EIO;
2470                 mapping_set_error(page->mapping, ret);
2471         }
2472         return 0;
2473 }
2474
2475 /*
2476  * after a writepage IO is done, we need to:
2477  * clear the uptodate bits on error
2478  * clear the writeback bits in the extent tree for this IO
2479  * end_page_writeback if the page has no more pending IO
2480  *
2481  * Scheduling is not allowed, so the extent state tree is expected
2482  * to have one and only one object corresponding to this IO.
2483  */
2484 static void end_bio_extent_writepage(struct bio *bio, int err)
2485 {
2486         struct bio_vec *bvec;
2487         u64 start;
2488         u64 end;
2489         int i;
2490
2491         bio_for_each_segment_all(bvec, bio, i) {
2492                 struct page *page = bvec->bv_page;
2493
2494                 /* We always issue full-page reads, but if some block
2495                  * in a page fails to read, blk_update_request() will
2496                  * advance bv_offset and adjust bv_len to compensate.
2497                  * Print a warning for nonzero offsets, and an error
2498                  * if they don't add up to a full page.  */
2499                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2500                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2501                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2502                                    "partial page write in btrfs with offset %u and length %u",
2503                                         bvec->bv_offset, bvec->bv_len);
2504                         else
2505                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2506                                    "incomplete page write in btrfs with offset %u and "
2507                                    "length %u",
2508                                         bvec->bv_offset, bvec->bv_len);
2509                 }
2510
2511                 start = page_offset(page);
2512                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2513
2514                 if (end_extent_writepage(page, err, start, end))
2515                         continue;
2516
2517                 end_page_writeback(page);
2518         }
2519
2520         bio_put(bio);
2521 }
2522
2523 static void
2524 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2525                               int uptodate)
2526 {
2527         struct extent_state *cached = NULL;
2528         u64 end = start + len - 1;
2529
2530         if (uptodate && tree->track_uptodate)
2531                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2532         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2533 }
2534
2535 /*
2536  * after a readpage IO is done, we need to:
2537  * clear the uptodate bits on error
2538  * set the uptodate bits if things worked
2539  * set the page up to date if all extents in the tree are uptodate
2540  * clear the lock bit in the extent tree
2541  * unlock the page if there are no other extents locked for it
2542  *
2543  * Scheduling is not allowed, so the extent state tree is expected
2544  * to have one and only one object corresponding to this IO.
2545  */
2546 static void end_bio_extent_readpage(struct bio *bio, int err)
2547 {
2548         struct bio_vec *bvec;
2549         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2550         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2551         struct extent_io_tree *tree;
2552         u64 offset = 0;
2553         u64 start;
2554         u64 end;
2555         u64 len;
2556         u64 extent_start = 0;
2557         u64 extent_len = 0;
2558         int mirror;
2559         int ret;
2560         int i;
2561
2562         if (err)
2563                 uptodate = 0;
2564
2565         bio_for_each_segment_all(bvec, bio, i) {
2566                 struct page *page = bvec->bv_page;
2567                 struct inode *inode = page->mapping->host;
2568
2569                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2570                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector, err,
2571                          io_bio->mirror_num);
2572                 tree = &BTRFS_I(inode)->io_tree;
2573
2574                 /* We always issue full-page reads, but if some block
2575                  * in a page fails to read, blk_update_request() will
2576                  * advance bv_offset and adjust bv_len to compensate.
2577                  * Print a warning for nonzero offsets, and an error
2578                  * if they don't add up to a full page.  */
2579                 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2580                         if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2581                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2582                                    "partial page read in btrfs with offset %u and length %u",
2583                                         bvec->bv_offset, bvec->bv_len);
2584                         else
2585                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2586                                    "incomplete page read in btrfs with offset %u and "
2587                                    "length %u",
2588                                         bvec->bv_offset, bvec->bv_len);
2589                 }
2590
2591                 start = page_offset(page);
2592                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2593                 len = bvec->bv_len;
2594
2595                 mirror = io_bio->mirror_num;
2596                 if (likely(uptodate && tree->ops &&
2597                            tree->ops->readpage_end_io_hook)) {
2598                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2599                                                               page, start, end,
2600                                                               mirror);
2601                         if (ret)
2602                                 uptodate = 0;
2603                         else
2604                                 clean_io_failure(inode, start, page, 0);
2605                 }
2606
2607                 if (likely(uptodate))
2608                         goto readpage_ok;
2609
2610                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2611                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2612                         if (!ret && !err &&
2613                             test_bit(BIO_UPTODATE, &bio->bi_flags))
2614                                 uptodate = 1;
2615                 } else {
2616                         /*
2617                          * The generic bio_readpage_error handles errors the
2618                          * following way: If possible, new read requests are
2619                          * created and submitted and will end up in
2620                          * end_bio_extent_readpage as well (if we're lucky, not
2621                          * in the !uptodate case). In that case it returns 0 and
2622                          * we just go on with the next page in our bio. If it
2623                          * can't handle the error it will return -EIO and we
2624                          * remain responsible for that page.
2625                          */
2626                         ret = bio_readpage_error(bio, offset, page, start, end,
2627                                                  mirror);
2628                         if (ret == 0) {
2629                                 uptodate =
2630                                         test_bit(BIO_UPTODATE, &bio->bi_flags);
2631                                 if (err)
2632                                         uptodate = 0;
2633                                 offset += len;
2634                                 continue;
2635                         }
2636                 }
2637 readpage_ok:
2638                 if (likely(uptodate)) {
2639                         loff_t i_size = i_size_read(inode);
2640                         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2641                         unsigned off;
2642
2643                         /* Zero out the end if this page straddles i_size */
2644                         off = i_size & (PAGE_CACHE_SIZE-1);
2645                         if (page->index == end_index && off)
2646                                 zero_user_segment(page, off, PAGE_CACHE_SIZE);
2647                         SetPageUptodate(page);
2648                 } else {
2649                         ClearPageUptodate(page);
2650                         SetPageError(page);
2651                 }
2652                 unlock_page(page);
2653                 offset += len;
2654
2655                 if (unlikely(!uptodate)) {
2656                         if (extent_len) {
2657                                 endio_readpage_release_extent(tree,
2658                                                               extent_start,
2659                                                               extent_len, 1);
2660                                 extent_start = 0;
2661                                 extent_len = 0;
2662                         }
2663                         endio_readpage_release_extent(tree, start,
2664                                                       end - start + 1, 0);
2665                 } else if (!extent_len) {
2666                         extent_start = start;
2667                         extent_len = end + 1 - start;
2668                 } else if (extent_start + extent_len == start) {
2669                         extent_len += end + 1 - start;
2670                 } else {
2671                         endio_readpage_release_extent(tree, extent_start,
2672                                                       extent_len, uptodate);
2673                         extent_start = start;
2674                         extent_len = end + 1 - start;
2675                 }
2676         }
2677
2678         if (extent_len)
2679                 endio_readpage_release_extent(tree, extent_start, extent_len,
2680                                               uptodate);
2681         if (io_bio->end_io)
2682                 io_bio->end_io(io_bio, err);
2683         bio_put(bio);
2684 }
2685
2686 /*
2687  * this allocates from the btrfs_bioset.  We're returning a bio right now
2688  * but you can call btrfs_io_bio for the appropriate container_of magic
2689  */
2690 struct bio *
2691 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2692                 gfp_t gfp_flags)
2693 {
2694         struct btrfs_io_bio *btrfs_bio;
2695         struct bio *bio;
2696
2697         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2698
2699         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2700                 while (!bio && (nr_vecs /= 2)) {
2701                         bio = bio_alloc_bioset(gfp_flags,
2702                                                nr_vecs, btrfs_bioset);
2703                 }
2704         }
2705
2706         if (bio) {
2707                 bio->bi_bdev = bdev;
2708                 bio->bi_iter.bi_sector = first_sector;
2709                 btrfs_bio = btrfs_io_bio(bio);
2710                 btrfs_bio->csum = NULL;
2711                 btrfs_bio->csum_allocated = NULL;
2712                 btrfs_bio->end_io = NULL;
2713         }
2714         return bio;
2715 }
2716
2717 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2718 {
2719         struct btrfs_io_bio *btrfs_bio;
2720         struct bio *new;
2721
2722         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2723         if (new) {
2724                 btrfs_bio = btrfs_io_bio(new);
2725                 btrfs_bio->csum = NULL;
2726                 btrfs_bio->csum_allocated = NULL;
2727                 btrfs_bio->end_io = NULL;
2728         }
2729         return new;
2730 }
2731
2732 /* this also allocates from the btrfs_bioset */
2733 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2734 {
2735         struct btrfs_io_bio *btrfs_bio;
2736         struct bio *bio;
2737
2738         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2739         if (bio) {
2740                 btrfs_bio = btrfs_io_bio(bio);
2741                 btrfs_bio->csum = NULL;
2742                 btrfs_bio->csum_allocated = NULL;
2743                 btrfs_bio->end_io = NULL;
2744         }
2745         return bio;
2746 }
2747
2748
2749 static int __must_check submit_one_bio(int rw, struct bio *bio,
2750                                        int mirror_num, unsigned long bio_flags)
2751 {
2752         int ret = 0;
2753         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2754         struct page *page = bvec->bv_page;
2755         struct extent_io_tree *tree = bio->bi_private;
2756         u64 start;
2757
2758         start = page_offset(page) + bvec->bv_offset;
2759
2760         bio->bi_private = NULL;
2761
2762         bio_get(bio);
2763
2764         if (tree->ops && tree->ops->submit_bio_hook)
2765                 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2766                                            mirror_num, bio_flags, start);
2767         else
2768                 btrfsic_submit_bio(rw, bio);
2769
2770         if (bio_flagged(bio, BIO_EOPNOTSUPP))
2771                 ret = -EOPNOTSUPP;
2772         bio_put(bio);
2773         return ret;
2774 }
2775
2776 static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2777                      unsigned long offset, size_t size, struct bio *bio,
2778                      unsigned long bio_flags)
2779 {
2780         int ret = 0;
2781         if (tree->ops && tree->ops->merge_bio_hook)
2782                 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2783                                                 bio_flags);
2784         BUG_ON(ret < 0);
2785         return ret;
2786
2787 }
2788
2789 static int submit_extent_page(int rw, struct extent_io_tree *tree,
2790                               struct page *page, sector_t sector,
2791                               size_t size, unsigned long offset,
2792                               struct block_device *bdev,
2793                               struct bio **bio_ret,
2794                               unsigned long max_pages,
2795                               bio_end_io_t end_io_func,
2796                               int mirror_num,
2797                               unsigned long prev_bio_flags,
2798                               unsigned long bio_flags)
2799 {
2800         int ret = 0;
2801         struct bio *bio;
2802         int nr;
2803         int contig = 0;
2804         int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2805         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2806         size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2807
2808         if (bio_ret && *bio_ret) {
2809                 bio = *bio_ret;
2810                 if (old_compressed)
2811                         contig = bio->bi_iter.bi_sector == sector;
2812                 else
2813                         contig = bio_end_sector(bio) == sector;
2814
2815                 if (prev_bio_flags != bio_flags || !contig ||
2816                     merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2817                     bio_add_page(bio, page, page_size, offset) < page_size) {
2818                         ret = submit_one_bio(rw, bio, mirror_num,
2819                                              prev_bio_flags);
2820                         if (ret < 0) {
2821                                 *bio_ret = NULL;
2822                                 return ret;
2823                         }
2824                         bio = NULL;
2825                 } else {
2826                         return 0;
2827                 }
2828         }
2829         if (this_compressed)
2830                 nr = BIO_MAX_PAGES;
2831         else
2832                 nr = bio_get_nr_vecs(bdev);
2833
2834         bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2835         if (!bio)
2836                 return -ENOMEM;
2837
2838         bio_add_page(bio, page, page_size, offset);
2839         bio->bi_end_io = end_io_func;
2840         bio->bi_private = tree;
2841
2842         if (bio_ret)
2843                 *bio_ret = bio;
2844         else
2845                 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2846
2847         return ret;
2848 }
2849
2850 static void attach_extent_buffer_page(struct extent_buffer *eb,
2851                                       struct page *page)
2852 {
2853         if (!PagePrivate(page)) {
2854                 SetPagePrivate(page);
2855                 page_cache_get(page);
2856                 set_page_private(page, (unsigned long)eb);
2857         } else {
2858                 WARN_ON(page->private != (unsigned long)eb);
2859         }
2860 }
2861
2862 void set_page_extent_mapped(struct page *page)
2863 {
2864         if (!PagePrivate(page)) {
2865                 SetPagePrivate(page);
2866                 page_cache_get(page);
2867                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2868         }
2869 }
2870
2871 static struct extent_map *
2872 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2873                  u64 start, u64 len, get_extent_t *get_extent,
2874                  struct extent_map **em_cached)
2875 {
2876         struct extent_map *em;
2877
2878         if (em_cached && *em_cached) {
2879                 em = *em_cached;
2880                 if (extent_map_in_tree(em) && start >= em->start &&
2881                     start < extent_map_end(em)) {
2882                         atomic_inc(&em->refs);
2883                         return em;
2884                 }
2885
2886                 free_extent_map(em);
2887                 *em_cached = NULL;
2888         }
2889
2890         em = get_extent(inode, page, pg_offset, start, len, 0);
2891         if (em_cached && !IS_ERR_OR_NULL(em)) {
2892                 BUG_ON(*em_cached);
2893                 atomic_inc(&em->refs);
2894                 *em_cached = em;
2895         }
2896         return em;
2897 }
2898 /*
2899  * basic readpage implementation.  Locked extent state structs are inserted
2900  * into the tree that are removed when the IO is done (by the end_io
2901  * handlers)
2902  * XXX JDM: This needs looking at to ensure proper page locking
2903  */
2904 static int __do_readpage(struct extent_io_tree *tree,
2905                          struct page *page,
2906                          get_extent_t *get_extent,
2907                          struct extent_map **em_cached,
2908                          struct bio **bio, int mirror_num,
2909                          unsigned long *bio_flags, int rw)
2910 {
2911         struct inode *inode = page->mapping->host;
2912         u64 start = page_offset(page);
2913         u64 page_end = start + PAGE_CACHE_SIZE - 1;
2914         u64 end;
2915         u64 cur = start;
2916         u64 extent_offset;
2917         u64 last_byte = i_size_read(inode);
2918         u64 block_start;
2919         u64 cur_end;
2920         sector_t sector;
2921         struct extent_map *em;
2922         struct block_device *bdev;
2923         int ret;
2924         int nr = 0;
2925         int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2926         size_t pg_offset = 0;
2927         size_t iosize;
2928         size_t disk_io_size;
2929         size_t blocksize = inode->i_sb->s_blocksize;
2930         unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2931
2932         set_page_extent_mapped(page);
2933
2934         end = page_end;
2935         if (!PageUptodate(page)) {
2936                 if (cleancache_get_page(page) == 0) {
2937                         BUG_ON(blocksize != PAGE_SIZE);
2938                         unlock_extent(tree, start, end);
2939                         goto out;
2940                 }
2941         }
2942
2943         if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2944                 char *userpage;
2945                 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2946
2947                 if (zero_offset) {
2948                         iosize = PAGE_CACHE_SIZE - zero_offset;
2949                         userpage = kmap_atomic(page);
2950                         memset(userpage + zero_offset, 0, iosize);
2951                         flush_dcache_page(page);
2952                         kunmap_atomic(userpage);
2953                 }
2954         }
2955         while (cur <= end) {
2956                 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2957
2958                 if (cur >= last_byte) {
2959                         char *userpage;
2960                         struct extent_state *cached = NULL;
2961
2962                         iosize = PAGE_CACHE_SIZE - pg_offset;
2963                         userpage = kmap_atomic(page);
2964                         memset(userpage + pg_offset, 0, iosize);
2965                         flush_dcache_page(page);
2966                         kunmap_atomic(userpage);
2967                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2968                                             &cached, GFP_NOFS);
2969                         if (!parent_locked)
2970                                 unlock_extent_cached(tree, cur,
2971                                                      cur + iosize - 1,
2972                                                      &cached, GFP_NOFS);
2973                         break;
2974                 }
2975                 em = __get_extent_map(inode, page, pg_offset, cur,
2976                                       end - cur + 1, get_extent, em_cached);
2977                 if (IS_ERR_OR_NULL(em)) {
2978                         SetPageError(page);
2979                         if (!parent_locked)
2980                                 unlock_extent(tree, cur, end);
2981                         break;
2982                 }
2983                 extent_offset = cur - em->start;
2984                 BUG_ON(extent_map_end(em) <= cur);
2985                 BUG_ON(end < cur);
2986
2987                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2988                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2989                         extent_set_compress_type(&this_bio_flag,
2990                                                  em->compress_type);
2991                 }
2992
2993                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2994                 cur_end = min(extent_map_end(em) - 1, end);
2995                 iosize = ALIGN(iosize, blocksize);
2996                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2997                         disk_io_size = em->block_len;
2998                         sector = em->block_start >> 9;
2999                 } else {
3000                         sector = (em->block_start + extent_offset) >> 9;
3001                         disk_io_size = iosize;
3002                 }
3003                 bdev = em->bdev;
3004                 block_start = em->block_start;
3005                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3006                         block_start = EXTENT_MAP_HOLE;
3007                 free_extent_map(em);
3008                 em = NULL;
3009
3010                 /* we've found a hole, just zero and go on */
3011                 if (block_start == EXTENT_MAP_HOLE) {
3012                         char *userpage;
3013                         struct extent_state *cached = NULL;
3014
3015                         userpage = kmap_atomic(page);
3016                         memset(userpage + pg_offset, 0, iosize);
3017                         flush_dcache_page(page);
3018                         kunmap_atomic(userpage);
3019
3020                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3021                                             &cached, GFP_NOFS);
3022                         unlock_extent_cached(tree, cur, cur + iosize - 1,
3023                                              &cached, GFP_NOFS);
3024                         cur = cur + iosize;
3025                         pg_offset += iosize;
3026                         continue;
3027                 }
3028                 /* the get_extent function already copied into the page */
3029                 if (test_range_bit(tree, cur, cur_end,
3030                                    EXTENT_UPTODATE, 1, NULL)) {
3031                         check_page_uptodate(tree, page);
3032                         if (!parent_locked)
3033                                 unlock_extent(tree, cur, cur + iosize - 1);
3034                         cur = cur + iosize;
3035                         pg_offset += iosize;
3036                         continue;
3037                 }
3038                 /* we have an inline extent but it didn't get marked up
3039                  * to date.  Error out
3040                  */
3041                 if (block_start == EXTENT_MAP_INLINE) {
3042                         SetPageError(page);
3043                         if (!parent_locked)
3044                                 unlock_extent(tree, cur, cur + iosize - 1);
3045                         cur = cur + iosize;
3046                         pg_offset += iosize;
3047                         continue;
3048                 }
3049
3050                 pnr -= page->index;
3051                 ret = submit_extent_page(rw, tree, page,
3052                                          sector, disk_io_size, pg_offset,
3053                                          bdev, bio, pnr,
3054                                          end_bio_extent_readpage, mirror_num,
3055                                          *bio_flags,
3056                                          this_bio_flag);
3057                 if (!ret) {
3058                         nr++;
3059                         *bio_flags = this_bio_flag;
3060                 } else {
3061                         SetPageError(page);
3062                         if (!parent_locked)
3063                                 unlock_extent(tree, cur, cur + iosize - 1);
3064                 }
3065                 cur = cur + iosize;
3066                 pg_offset += iosize;
3067         }
3068 out:
3069         if (!nr) {
3070                 if (!PageError(page))
3071                         SetPageUptodate(page);
3072                 unlock_page(page);
3073         }
3074         return 0;
3075 }
3076
3077 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3078                                              struct page *pages[], int nr_pages,
3079                                              u64 start, u64 end,
3080                                              get_extent_t *get_extent,
3081                                              struct extent_map **em_cached,
3082                                              struct bio **bio, int mirror_num,
3083                                              unsigned long *bio_flags, int rw)
3084 {
3085         struct inode *inode;
3086         struct btrfs_ordered_extent *ordered;
3087         int index;
3088
3089         inode = pages[0]->mapping->host;
3090         while (1) {
3091                 lock_extent(tree, start, end);
3092                 ordered = btrfs_lookup_ordered_range(inode, start,
3093                                                      end - start + 1);
3094                 if (!ordered)
3095                         break;
3096                 unlock_extent(tree, start, end);
3097                 btrfs_start_ordered_extent(inode, ordered, 1);
3098                 btrfs_put_ordered_extent(ordered);
3099         }
3100
3101         for (index = 0; index < nr_pages; index++) {
3102                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3103                               mirror_num, bio_flags, rw);
3104                 page_cache_release(pages[index]);
3105         }
3106 }
3107
3108 static void __extent_readpages(struct extent_io_tree *tree,
3109                                struct page *pages[],
3110                                int nr_pages, get_extent_t *get_extent,
3111                                struct extent_map **em_cached,
3112                                struct bio **bio, int mirror_num,
3113                                unsigned long *bio_flags, int rw)
3114 {
3115         u64 start = 0;
3116         u64 end = 0;
3117         u64 page_start;
3118         int index;
3119         int first_index = 0;
3120
3121         for (index = 0; index < nr_pages; index++) {
3122                 page_start = page_offset(pages[index]);
3123                 if (!end) {
3124                         start = page_start;
3125                         end = start + PAGE_CACHE_SIZE - 1;
3126                         first_index = index;
3127                 } else if (end + 1 == page_start) {
3128                         end += PAGE_CACHE_SIZE;
3129                 } else {
3130                         __do_contiguous_readpages(tree, &pages[first_index],
3131                                                   index - first_index, start,
3132                                                   end, get_extent, em_cached,
3133                                                   bio, mirror_num, bio_flags,
3134                                                   rw);
3135                         start = page_start;
3136                         end = start + PAGE_CACHE_SIZE - 1;