Merge branch 'for-linus-4.8' of git://git.kernel.org/pub/scm/linux/kernel/git/mason...
[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                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
100                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
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 void add_extent_changeset(struct extent_state *state, unsigned bits,
135                                  struct extent_changeset *changeset,
136                                  int set)
137 {
138         int ret;
139
140         if (!changeset)
141                 return;
142         if (set && (state->state & bits) == bits)
143                 return;
144         if (!set && (state->state & bits) == 0)
145                 return;
146         changeset->bytes_changed += state->end - state->start + 1;
147         ret = ulist_add(changeset->range_changed, state->start, state->end,
148                         GFP_ATOMIC);
149         /* ENOMEM */
150         BUG_ON(ret < 0);
151 }
152
153 static noinline void flush_write_bio(void *data);
154 static inline struct btrfs_fs_info *
155 tree_fs_info(struct extent_io_tree *tree)
156 {
157         if (!tree->mapping)
158                 return NULL;
159         return btrfs_sb(tree->mapping->host->i_sb);
160 }
161
162 int __init extent_io_init(void)
163 {
164         extent_state_cache = kmem_cache_create("btrfs_extent_state",
165                         sizeof(struct extent_state), 0,
166                         SLAB_MEM_SPREAD, NULL);
167         if (!extent_state_cache)
168                 return -ENOMEM;
169
170         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
171                         sizeof(struct extent_buffer), 0,
172                         SLAB_MEM_SPREAD, NULL);
173         if (!extent_buffer_cache)
174                 goto free_state_cache;
175
176         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
177                                      offsetof(struct btrfs_io_bio, bio));
178         if (!btrfs_bioset)
179                 goto free_buffer_cache;
180
181         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
182                 goto free_bioset;
183
184         return 0;
185
186 free_bioset:
187         bioset_free(btrfs_bioset);
188         btrfs_bioset = NULL;
189
190 free_buffer_cache:
191         kmem_cache_destroy(extent_buffer_cache);
192         extent_buffer_cache = NULL;
193
194 free_state_cache:
195         kmem_cache_destroy(extent_state_cache);
196         extent_state_cache = NULL;
197         return -ENOMEM;
198 }
199
200 void extent_io_exit(void)
201 {
202         btrfs_leak_debug_check();
203
204         /*
205          * Make sure all delayed rcu free are flushed before we
206          * destroy caches.
207          */
208         rcu_barrier();
209         kmem_cache_destroy(extent_state_cache);
210         kmem_cache_destroy(extent_buffer_cache);
211         if (btrfs_bioset)
212                 bioset_free(btrfs_bioset);
213 }
214
215 void extent_io_tree_init(struct extent_io_tree *tree,
216                          struct address_space *mapping)
217 {
218         tree->state = RB_ROOT;
219         tree->ops = NULL;
220         tree->dirty_bytes = 0;
221         spin_lock_init(&tree->lock);
222         tree->mapping = mapping;
223 }
224
225 static struct extent_state *alloc_extent_state(gfp_t mask)
226 {
227         struct extent_state *state;
228
229         state = kmem_cache_alloc(extent_state_cache, mask);
230         if (!state)
231                 return state;
232         state->state = 0;
233         state->failrec = NULL;
234         RB_CLEAR_NODE(&state->rb_node);
235         btrfs_leak_debug_add(&state->leak_list, &states);
236         atomic_set(&state->refs, 1);
237         init_waitqueue_head(&state->wq);
238         trace_alloc_extent_state(state, mask, _RET_IP_);
239         return state;
240 }
241
242 void free_extent_state(struct extent_state *state)
243 {
244         if (!state)
245                 return;
246         if (atomic_dec_and_test(&state->refs)) {
247                 WARN_ON(extent_state_in_tree(state));
248                 btrfs_leak_debug_del(&state->leak_list);
249                 trace_free_extent_state(state, _RET_IP_);
250                 kmem_cache_free(extent_state_cache, state);
251         }
252 }
253
254 static struct rb_node *tree_insert(struct rb_root *root,
255                                    struct rb_node *search_start,
256                                    u64 offset,
257                                    struct rb_node *node,
258                                    struct rb_node ***p_in,
259                                    struct rb_node **parent_in)
260 {
261         struct rb_node **p;
262         struct rb_node *parent = NULL;
263         struct tree_entry *entry;
264
265         if (p_in && parent_in) {
266                 p = *p_in;
267                 parent = *parent_in;
268                 goto do_insert;
269         }
270
271         p = search_start ? &search_start : &root->rb_node;
272         while (*p) {
273                 parent = *p;
274                 entry = rb_entry(parent, struct tree_entry, rb_node);
275
276                 if (offset < entry->start)
277                         p = &(*p)->rb_left;
278                 else if (offset > entry->end)
279                         p = &(*p)->rb_right;
280                 else
281                         return parent;
282         }
283
284 do_insert:
285         rb_link_node(node, parent, p);
286         rb_insert_color(node, root);
287         return NULL;
288 }
289
290 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
291                                       struct rb_node **prev_ret,
292                                       struct rb_node **next_ret,
293                                       struct rb_node ***p_ret,
294                                       struct rb_node **parent_ret)
295 {
296         struct rb_root *root = &tree->state;
297         struct rb_node **n = &root->rb_node;
298         struct rb_node *prev = NULL;
299         struct rb_node *orig_prev = NULL;
300         struct tree_entry *entry;
301         struct tree_entry *prev_entry = NULL;
302
303         while (*n) {
304                 prev = *n;
305                 entry = rb_entry(prev, struct tree_entry, rb_node);
306                 prev_entry = entry;
307
308                 if (offset < entry->start)
309                         n = &(*n)->rb_left;
310                 else if (offset > entry->end)
311                         n = &(*n)->rb_right;
312                 else
313                         return *n;
314         }
315
316         if (p_ret)
317                 *p_ret = n;
318         if (parent_ret)
319                 *parent_ret = prev;
320
321         if (prev_ret) {
322                 orig_prev = prev;
323                 while (prev && offset > prev_entry->end) {
324                         prev = rb_next(prev);
325                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
326                 }
327                 *prev_ret = prev;
328                 prev = orig_prev;
329         }
330
331         if (next_ret) {
332                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
333                 while (prev && offset < prev_entry->start) {
334                         prev = rb_prev(prev);
335                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
336                 }
337                 *next_ret = prev;
338         }
339         return NULL;
340 }
341
342 static inline struct rb_node *
343 tree_search_for_insert(struct extent_io_tree *tree,
344                        u64 offset,
345                        struct rb_node ***p_ret,
346                        struct rb_node **parent_ret)
347 {
348         struct rb_node *prev = NULL;
349         struct rb_node *ret;
350
351         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
352         if (!ret)
353                 return prev;
354         return ret;
355 }
356
357 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
358                                           u64 offset)
359 {
360         return tree_search_for_insert(tree, offset, NULL, NULL);
361 }
362
363 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
364                      struct extent_state *other)
365 {
366         if (tree->ops && tree->ops->merge_extent_hook)
367                 tree->ops->merge_extent_hook(tree->mapping->host, new,
368                                              other);
369 }
370
371 /*
372  * utility function to look for merge candidates inside a given range.
373  * Any extents with matching state are merged together into a single
374  * extent in the tree.  Extents with EXTENT_IO in their state field
375  * are not merged because the end_io handlers need to be able to do
376  * operations on them without sleeping (or doing allocations/splits).
377  *
378  * This should be called with the tree lock held.
379  */
380 static void merge_state(struct extent_io_tree *tree,
381                         struct extent_state *state)
382 {
383         struct extent_state *other;
384         struct rb_node *other_node;
385
386         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
387                 return;
388
389         other_node = rb_prev(&state->rb_node);
390         if (other_node) {
391                 other = rb_entry(other_node, struct extent_state, rb_node);
392                 if (other->end == state->start - 1 &&
393                     other->state == state->state) {
394                         merge_cb(tree, state, other);
395                         state->start = other->start;
396                         rb_erase(&other->rb_node, &tree->state);
397                         RB_CLEAR_NODE(&other->rb_node);
398                         free_extent_state(other);
399                 }
400         }
401         other_node = rb_next(&state->rb_node);
402         if (other_node) {
403                 other = rb_entry(other_node, struct extent_state, rb_node);
404                 if (other->start == state->end + 1 &&
405                     other->state == state->state) {
406                         merge_cb(tree, state, other);
407                         state->end = other->end;
408                         rb_erase(&other->rb_node, &tree->state);
409                         RB_CLEAR_NODE(&other->rb_node);
410                         free_extent_state(other);
411                 }
412         }
413 }
414
415 static void set_state_cb(struct extent_io_tree *tree,
416                          struct extent_state *state, unsigned *bits)
417 {
418         if (tree->ops && tree->ops->set_bit_hook)
419                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
420 }
421
422 static void clear_state_cb(struct extent_io_tree *tree,
423                            struct extent_state *state, unsigned *bits)
424 {
425         if (tree->ops && tree->ops->clear_bit_hook)
426                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
427 }
428
429 static void set_state_bits(struct extent_io_tree *tree,
430                            struct extent_state *state, unsigned *bits,
431                            struct extent_changeset *changeset);
432
433 /*
434  * insert an extent_state struct into the tree.  'bits' are set on the
435  * struct before it is inserted.
436  *
437  * This may return -EEXIST if the extent is already there, in which case the
438  * state struct is freed.
439  *
440  * The tree lock is not taken internally.  This is a utility function and
441  * probably isn't what you want to call (see set/clear_extent_bit).
442  */
443 static int insert_state(struct extent_io_tree *tree,
444                         struct extent_state *state, u64 start, u64 end,
445                         struct rb_node ***p,
446                         struct rb_node **parent,
447                         unsigned *bits, struct extent_changeset *changeset)
448 {
449         struct rb_node *node;
450
451         if (end < start)
452                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
453                        end, start);
454         state->start = start;
455         state->end = end;
456
457         set_state_bits(tree, state, bits, changeset);
458
459         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
460         if (node) {
461                 struct extent_state *found;
462                 found = rb_entry(node, struct extent_state, rb_node);
463                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
464                        "%llu %llu\n",
465                        found->start, found->end, start, end);
466                 return -EEXIST;
467         }
468         merge_state(tree, state);
469         return 0;
470 }
471
472 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
473                      u64 split)
474 {
475         if (tree->ops && tree->ops->split_extent_hook)
476                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
477 }
478
479 /*
480  * split a given extent state struct in two, inserting the preallocated
481  * struct 'prealloc' as the newly created second half.  'split' indicates an
482  * offset inside 'orig' where it should be split.
483  *
484  * Before calling,
485  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
486  * are two extent state structs in the tree:
487  * prealloc: [orig->start, split - 1]
488  * orig: [ split, orig->end ]
489  *
490  * The tree locks are not taken by this function. They need to be held
491  * by the caller.
492  */
493 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
494                        struct extent_state *prealloc, u64 split)
495 {
496         struct rb_node *node;
497
498         split_cb(tree, orig, split);
499
500         prealloc->start = orig->start;
501         prealloc->end = split - 1;
502         prealloc->state = orig->state;
503         orig->start = split;
504
505         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
506                            &prealloc->rb_node, NULL, NULL);
507         if (node) {
508                 free_extent_state(prealloc);
509                 return -EEXIST;
510         }
511         return 0;
512 }
513
514 static struct extent_state *next_state(struct extent_state *state)
515 {
516         struct rb_node *next = rb_next(&state->rb_node);
517         if (next)
518                 return rb_entry(next, struct extent_state, rb_node);
519         else
520                 return NULL;
521 }
522
523 /*
524  * utility function to clear some bits in an extent state struct.
525  * it will optionally wake up any one waiting on this state (wake == 1).
526  *
527  * If no bits are set on the state struct after clearing things, the
528  * struct is freed and removed from the tree
529  */
530 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
531                                             struct extent_state *state,
532                                             unsigned *bits, int wake,
533                                             struct extent_changeset *changeset)
534 {
535         struct extent_state *next;
536         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
537
538         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
539                 u64 range = state->end - state->start + 1;
540                 WARN_ON(range > tree->dirty_bytes);
541                 tree->dirty_bytes -= range;
542         }
543         clear_state_cb(tree, state, bits);
544         add_extent_changeset(state, bits_to_clear, changeset, 0);
545         state->state &= ~bits_to_clear;
546         if (wake)
547                 wake_up(&state->wq);
548         if (state->state == 0) {
549                 next = next_state(state);
550                 if (extent_state_in_tree(state)) {
551                         rb_erase(&state->rb_node, &tree->state);
552                         RB_CLEAR_NODE(&state->rb_node);
553                         free_extent_state(state);
554                 } else {
555                         WARN_ON(1);
556                 }
557         } else {
558                 merge_state(tree, state);
559                 next = next_state(state);
560         }
561         return next;
562 }
563
564 static struct extent_state *
565 alloc_extent_state_atomic(struct extent_state *prealloc)
566 {
567         if (!prealloc)
568                 prealloc = alloc_extent_state(GFP_ATOMIC);
569
570         return prealloc;
571 }
572
573 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
574 {
575         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
576                     "Extent tree was modified by another "
577                     "thread while locked.");
578 }
579
580 /*
581  * clear some bits on a range in the tree.  This may require splitting
582  * or inserting elements in the tree, so the gfp mask is used to
583  * indicate which allocations or sleeping are allowed.
584  *
585  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
586  * the given range from the tree regardless of state (ie for truncate).
587  *
588  * the range [start, end] is inclusive.
589  *
590  * This takes the tree lock, and returns 0 on success and < 0 on error.
591  */
592 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
593                               unsigned bits, int wake, int delete,
594                               struct extent_state **cached_state,
595                               gfp_t mask, struct extent_changeset *changeset)
596 {
597         struct extent_state *state;
598         struct extent_state *cached;
599         struct extent_state *prealloc = NULL;
600         struct rb_node *node;
601         u64 last_end;
602         int err;
603         int clear = 0;
604
605         btrfs_debug_check_extent_io_range(tree, start, end);
606
607         if (bits & EXTENT_DELALLOC)
608                 bits |= EXTENT_NORESERVE;
609
610         if (delete)
611                 bits |= ~EXTENT_CTLBITS;
612         bits |= EXTENT_FIRST_DELALLOC;
613
614         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
615                 clear = 1;
616 again:
617         if (!prealloc && gfpflags_allow_blocking(mask)) {
618                 /*
619                  * Don't care for allocation failure here because we might end
620                  * up not needing the pre-allocated extent state at all, which
621                  * is the case if we only have in the tree extent states that
622                  * cover our input range and don't cover too any other range.
623                  * If we end up needing a new extent state we allocate it later.
624                  */
625                 prealloc = alloc_extent_state(mask);
626         }
627
628         spin_lock(&tree->lock);
629         if (cached_state) {
630                 cached = *cached_state;
631
632                 if (clear) {
633                         *cached_state = NULL;
634                         cached_state = NULL;
635                 }
636
637                 if (cached && extent_state_in_tree(cached) &&
638                     cached->start <= start && cached->end > start) {
639                         if (clear)
640                                 atomic_dec(&cached->refs);
641                         state = cached;
642                         goto hit_next;
643                 }
644                 if (clear)
645                         free_extent_state(cached);
646         }
647         /*
648          * this search will find the extents that end after
649          * our range starts
650          */
651         node = tree_search(tree, start);
652         if (!node)
653                 goto out;
654         state = rb_entry(node, struct extent_state, rb_node);
655 hit_next:
656         if (state->start > end)
657                 goto out;
658         WARN_ON(state->end < start);
659         last_end = state->end;
660
661         /* the state doesn't have the wanted bits, go ahead */
662         if (!(state->state & bits)) {
663                 state = next_state(state);
664                 goto next;
665         }
666
667         /*
668          *     | ---- desired range ---- |
669          *  | state | or
670          *  | ------------- state -------------- |
671          *
672          * We need to split the extent we found, and may flip
673          * bits on second half.
674          *
675          * If the extent we found extends past our range, we
676          * just split and search again.  It'll get split again
677          * the next time though.
678          *
679          * If the extent we found is inside our range, we clear
680          * the desired bit on it.
681          */
682
683         if (state->start < start) {
684                 prealloc = alloc_extent_state_atomic(prealloc);
685                 BUG_ON(!prealloc);
686                 err = split_state(tree, state, prealloc, start);
687                 if (err)
688                         extent_io_tree_panic(tree, err);
689
690                 prealloc = NULL;
691                 if (err)
692                         goto out;
693                 if (state->end <= end) {
694                         state = clear_state_bit(tree, state, &bits, wake,
695                                                 changeset);
696                         goto next;
697                 }
698                 goto search_again;
699         }
700         /*
701          * | ---- desired range ---- |
702          *                        | state |
703          * We need to split the extent, and clear the bit
704          * on the first half
705          */
706         if (state->start <= end && state->end > end) {
707                 prealloc = alloc_extent_state_atomic(prealloc);
708                 BUG_ON(!prealloc);
709                 err = split_state(tree, state, prealloc, end + 1);
710                 if (err)
711                         extent_io_tree_panic(tree, err);
712
713                 if (wake)
714                         wake_up(&state->wq);
715
716                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
717
718                 prealloc = NULL;
719                 goto out;
720         }
721
722         state = clear_state_bit(tree, state, &bits, wake, changeset);
723 next:
724         if (last_end == (u64)-1)
725                 goto out;
726         start = last_end + 1;
727         if (start <= end && state && !need_resched())
728                 goto hit_next;
729
730 search_again:
731         if (start > end)
732                 goto out;
733         spin_unlock(&tree->lock);
734         if (gfpflags_allow_blocking(mask))
735                 cond_resched();
736         goto again;
737
738 out:
739         spin_unlock(&tree->lock);
740         if (prealloc)
741                 free_extent_state(prealloc);
742
743         return 0;
744
745 }
746
747 static void wait_on_state(struct extent_io_tree *tree,
748                           struct extent_state *state)
749                 __releases(tree->lock)
750                 __acquires(tree->lock)
751 {
752         DEFINE_WAIT(wait);
753         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
754         spin_unlock(&tree->lock);
755         schedule();
756         spin_lock(&tree->lock);
757         finish_wait(&state->wq, &wait);
758 }
759
760 /*
761  * waits for one or more bits to clear on a range in the state tree.
762  * The range [start, end] is inclusive.
763  * The tree lock is taken by this function
764  */
765 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
766                             unsigned long bits)
767 {
768         struct extent_state *state;
769         struct rb_node *node;
770
771         btrfs_debug_check_extent_io_range(tree, start, end);
772
773         spin_lock(&tree->lock);
774 again:
775         while (1) {
776                 /*
777                  * this search will find all the extents that end after
778                  * our range starts
779                  */
780                 node = tree_search(tree, start);
781 process_node:
782                 if (!node)
783                         break;
784
785                 state = rb_entry(node, struct extent_state, rb_node);
786
787                 if (state->start > end)
788                         goto out;
789
790                 if (state->state & bits) {
791                         start = state->start;
792                         atomic_inc(&state->refs);
793                         wait_on_state(tree, state);
794                         free_extent_state(state);
795                         goto again;
796                 }
797                 start = state->end + 1;
798
799                 if (start > end)
800                         break;
801
802                 if (!cond_resched_lock(&tree->lock)) {
803                         node = rb_next(node);
804                         goto process_node;
805                 }
806         }
807 out:
808         spin_unlock(&tree->lock);
809 }
810
811 static void set_state_bits(struct extent_io_tree *tree,
812                            struct extent_state *state,
813                            unsigned *bits, struct extent_changeset *changeset)
814 {
815         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
816
817         set_state_cb(tree, state, bits);
818         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
819                 u64 range = state->end - state->start + 1;
820                 tree->dirty_bytes += range;
821         }
822         add_extent_changeset(state, bits_to_set, changeset, 1);
823         state->state |= bits_to_set;
824 }
825
826 static void cache_state_if_flags(struct extent_state *state,
827                                  struct extent_state **cached_ptr,
828                                  unsigned flags)
829 {
830         if (cached_ptr && !(*cached_ptr)) {
831                 if (!flags || (state->state & flags)) {
832                         *cached_ptr = state;
833                         atomic_inc(&state->refs);
834                 }
835         }
836 }
837
838 static void cache_state(struct extent_state *state,
839                         struct extent_state **cached_ptr)
840 {
841         return cache_state_if_flags(state, cached_ptr,
842                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
843 }
844
845 /*
846  * set some bits on a range in the tree.  This may require allocations or
847  * sleeping, so the gfp mask is used to indicate what is allowed.
848  *
849  * If any of the exclusive bits are set, this will fail with -EEXIST if some
850  * part of the range already has the desired bits set.  The start of the
851  * existing range is returned in failed_start in this case.
852  *
853  * [start, end] is inclusive This takes the tree lock.
854  */
855
856 static int __must_check
857 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
858                  unsigned bits, unsigned exclusive_bits,
859                  u64 *failed_start, struct extent_state **cached_state,
860                  gfp_t mask, struct extent_changeset *changeset)
861 {
862         struct extent_state *state;
863         struct extent_state *prealloc = NULL;
864         struct rb_node *node;
865         struct rb_node **p;
866         struct rb_node *parent;
867         int err = 0;
868         u64 last_start;
869         u64 last_end;
870
871         btrfs_debug_check_extent_io_range(tree, start, end);
872
873         bits |= EXTENT_FIRST_DELALLOC;
874 again:
875         if (!prealloc && gfpflags_allow_blocking(mask)) {
876                 /*
877                  * Don't care for allocation failure here because we might end
878                  * up not needing the pre-allocated extent state at all, which
879                  * is the case if we only have in the tree extent states that
880                  * cover our input range and don't cover too any other range.
881                  * If we end up needing a new extent state we allocate it later.
882                  */
883                 prealloc = alloc_extent_state(mask);
884         }
885
886         spin_lock(&tree->lock);
887         if (cached_state && *cached_state) {
888                 state = *cached_state;
889                 if (state->start <= start && state->end > start &&
890                     extent_state_in_tree(state)) {
891                         node = &state->rb_node;
892                         goto hit_next;
893                 }
894         }
895         /*
896          * this search will find all the extents that end after
897          * our range starts.
898          */
899         node = tree_search_for_insert(tree, start, &p, &parent);
900         if (!node) {
901                 prealloc = alloc_extent_state_atomic(prealloc);
902                 BUG_ON(!prealloc);
903                 err = insert_state(tree, prealloc, start, end,
904                                    &p, &parent, &bits, changeset);
905                 if (err)
906                         extent_io_tree_panic(tree, err);
907
908                 cache_state(prealloc, cached_state);
909                 prealloc = NULL;
910                 goto out;
911         }
912         state = rb_entry(node, struct extent_state, rb_node);
913 hit_next:
914         last_start = state->start;
915         last_end = state->end;
916
917         /*
918          * | ---- desired range ---- |
919          * | state |
920          *
921          * Just lock what we found and keep going
922          */
923         if (state->start == start && state->end <= end) {
924                 if (state->state & exclusive_bits) {
925                         *failed_start = state->start;
926                         err = -EEXIST;
927                         goto out;
928                 }
929
930                 set_state_bits(tree, state, &bits, changeset);
931                 cache_state(state, cached_state);
932                 merge_state(tree, state);
933                 if (last_end == (u64)-1)
934                         goto out;
935                 start = last_end + 1;
936                 state = next_state(state);
937                 if (start < end && state && state->start == start &&
938                     !need_resched())
939                         goto hit_next;
940                 goto search_again;
941         }
942
943         /*
944          *     | ---- desired range ---- |
945          * | state |
946          *   or
947          * | ------------- state -------------- |
948          *
949          * We need to split the extent we found, and may flip bits on
950          * second half.
951          *
952          * If the extent we found extends past our
953          * range, we just split and search again.  It'll get split
954          * again the next time though.
955          *
956          * If the extent we found is inside our range, we set the
957          * desired bit on it.
958          */
959         if (state->start < start) {
960                 if (state->state & exclusive_bits) {
961                         *failed_start = start;
962                         err = -EEXIST;
963                         goto out;
964                 }
965
966                 prealloc = alloc_extent_state_atomic(prealloc);
967                 BUG_ON(!prealloc);
968                 err = split_state(tree, state, prealloc, start);
969                 if (err)
970                         extent_io_tree_panic(tree, err);
971
972                 prealloc = NULL;
973                 if (err)
974                         goto out;
975                 if (state->end <= end) {
976                         set_state_bits(tree, state, &bits, changeset);
977                         cache_state(state, cached_state);
978                         merge_state(tree, state);
979                         if (last_end == (u64)-1)
980                                 goto out;
981                         start = last_end + 1;
982                         state = next_state(state);
983                         if (start < end && state && state->start == start &&
984                             !need_resched())
985                                 goto hit_next;
986                 }
987                 goto search_again;
988         }
989         /*
990          * | ---- desired range ---- |
991          *     | state | or               | state |
992          *
993          * There's a hole, we need to insert something in it and
994          * ignore the extent we found.
995          */
996         if (state->start > start) {
997                 u64 this_end;
998                 if (end < last_start)
999                         this_end = end;
1000                 else
1001                         this_end = last_start - 1;
1002
1003                 prealloc = alloc_extent_state_atomic(prealloc);
1004                 BUG_ON(!prealloc);
1005
1006                 /*
1007                  * Avoid to free 'prealloc' if it can be merged with
1008                  * the later extent.
1009                  */
1010                 err = insert_state(tree, prealloc, start, this_end,
1011                                    NULL, NULL, &bits, changeset);
1012                 if (err)
1013                         extent_io_tree_panic(tree, err);
1014
1015                 cache_state(prealloc, cached_state);
1016                 prealloc = NULL;
1017                 start = this_end + 1;
1018                 goto search_again;
1019         }
1020         /*
1021          * | ---- desired range ---- |
1022          *                        | state |
1023          * We need to split the extent, and set the bit
1024          * on the first half
1025          */
1026         if (state->start <= end && state->end > end) {
1027                 if (state->state & exclusive_bits) {
1028                         *failed_start = start;
1029                         err = -EEXIST;
1030                         goto out;
1031                 }
1032
1033                 prealloc = alloc_extent_state_atomic(prealloc);
1034                 BUG_ON(!prealloc);
1035                 err = split_state(tree, state, prealloc, end + 1);
1036                 if (err)
1037                         extent_io_tree_panic(tree, err);
1038
1039                 set_state_bits(tree, prealloc, &bits, changeset);
1040                 cache_state(prealloc, cached_state);
1041                 merge_state(tree, prealloc);
1042                 prealloc = NULL;
1043                 goto out;
1044         }
1045
1046 search_again:
1047         if (start > end)
1048                 goto out;
1049         spin_unlock(&tree->lock);
1050         if (gfpflags_allow_blocking(mask))
1051                 cond_resched();
1052         goto again;
1053
1054 out:
1055         spin_unlock(&tree->lock);
1056         if (prealloc)
1057                 free_extent_state(prealloc);
1058
1059         return err;
1060
1061 }
1062
1063 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1064                    unsigned bits, u64 * failed_start,
1065                    struct extent_state **cached_state, gfp_t mask)
1066 {
1067         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1068                                 cached_state, mask, NULL);
1069 }
1070
1071
1072 /**
1073  * convert_extent_bit - convert all bits in a given range from one bit to
1074  *                      another
1075  * @tree:       the io tree to search
1076  * @start:      the start offset in bytes
1077  * @end:        the end offset in bytes (inclusive)
1078  * @bits:       the bits to set in this range
1079  * @clear_bits: the bits to clear in this range
1080  * @cached_state:       state that we're going to cache
1081  *
1082  * This will go through and set bits for the given range.  If any states exist
1083  * already in this range they are set with the given bit and cleared of the
1084  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1085  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1086  * boundary bits like LOCK.
1087  *
1088  * All allocations are done with GFP_NOFS.
1089  */
1090 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1091                        unsigned bits, unsigned clear_bits,
1092                        struct extent_state **cached_state)
1093 {
1094         struct extent_state *state;
1095         struct extent_state *prealloc = NULL;
1096         struct rb_node *node;
1097         struct rb_node **p;
1098         struct rb_node *parent;
1099         int err = 0;
1100         u64 last_start;
1101         u64 last_end;
1102         bool first_iteration = true;
1103
1104         btrfs_debug_check_extent_io_range(tree, start, end);
1105
1106 again:
1107         if (!prealloc) {
1108                 /*
1109                  * Best effort, don't worry if extent state allocation fails
1110                  * here for the first iteration. We might have a cached state
1111                  * that matches exactly the target range, in which case no
1112                  * extent state allocations are needed. We'll only know this
1113                  * after locking the tree.
1114                  */
1115                 prealloc = alloc_extent_state(GFP_NOFS);
1116                 if (!prealloc && !first_iteration)
1117                         return -ENOMEM;
1118         }
1119
1120         spin_lock(&tree->lock);
1121         if (cached_state && *cached_state) {
1122                 state = *cached_state;
1123                 if (state->start <= start && state->end > start &&
1124                     extent_state_in_tree(state)) {
1125                         node = &state->rb_node;
1126                         goto hit_next;
1127                 }
1128         }
1129
1130         /*
1131          * this search will find all the extents that end after
1132          * our range starts.
1133          */
1134         node = tree_search_for_insert(tree, start, &p, &parent);
1135         if (!node) {
1136                 prealloc = alloc_extent_state_atomic(prealloc);
1137                 if (!prealloc) {
1138                         err = -ENOMEM;
1139                         goto out;
1140                 }
1141                 err = insert_state(tree, prealloc, start, end,
1142                                    &p, &parent, &bits, NULL);
1143                 if (err)
1144                         extent_io_tree_panic(tree, err);
1145                 cache_state(prealloc, cached_state);
1146                 prealloc = NULL;
1147                 goto out;
1148         }
1149         state = rb_entry(node, struct extent_state, rb_node);
1150 hit_next:
1151         last_start = state->start;
1152         last_end = state->end;
1153
1154         /*
1155          * | ---- desired range ---- |
1156          * | state |
1157          *
1158          * Just lock what we found and keep going
1159          */
1160         if (state->start == start && state->end <= end) {
1161                 set_state_bits(tree, state, &bits, NULL);
1162                 cache_state(state, cached_state);
1163                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1164                 if (last_end == (u64)-1)
1165                         goto out;
1166                 start = last_end + 1;
1167                 if (start < end && state && state->start == start &&
1168                     !need_resched())
1169                         goto hit_next;
1170                 goto search_again;
1171         }
1172
1173         /*
1174          *     | ---- desired range ---- |
1175          * | state |
1176          *   or
1177          * | ------------- state -------------- |
1178          *
1179          * We need to split the extent we found, and may flip bits on
1180          * second half.
1181          *
1182          * If the extent we found extends past our
1183          * range, we just split and search again.  It'll get split
1184          * again the next time though.
1185          *
1186          * If the extent we found is inside our range, we set the
1187          * desired bit on it.
1188          */
1189         if (state->start < start) {
1190                 prealloc = alloc_extent_state_atomic(prealloc);
1191                 if (!prealloc) {
1192                         err = -ENOMEM;
1193                         goto out;
1194                 }
1195                 err = split_state(tree, state, prealloc, start);
1196                 if (err)
1197                         extent_io_tree_panic(tree, err);
1198                 prealloc = NULL;
1199                 if (err)
1200                         goto out;
1201                 if (state->end <= end) {
1202                         set_state_bits(tree, state, &bits, NULL);
1203                         cache_state(state, cached_state);
1204                         state = clear_state_bit(tree, state, &clear_bits, 0,
1205                                                 NULL);
1206                         if (last_end == (u64)-1)
1207                                 goto out;
1208                         start = last_end + 1;
1209                         if (start < end && state && state->start == start &&
1210                             !need_resched())
1211                                 goto hit_next;
1212                 }
1213                 goto search_again;
1214         }
1215         /*
1216          * | ---- desired range ---- |
1217          *     | state | or               | state |
1218          *
1219          * There's a hole, we need to insert something in it and
1220          * ignore the extent we found.
1221          */
1222         if (state->start > start) {
1223                 u64 this_end;
1224                 if (end < last_start)
1225                         this_end = end;
1226                 else
1227                         this_end = last_start - 1;
1228
1229                 prealloc = alloc_extent_state_atomic(prealloc);
1230                 if (!prealloc) {
1231                         err = -ENOMEM;
1232                         goto out;
1233                 }
1234
1235                 /*
1236                  * Avoid to free 'prealloc' if it can be merged with
1237                  * the later extent.
1238                  */
1239                 err = insert_state(tree, prealloc, start, this_end,
1240                                    NULL, NULL, &bits, NULL);
1241                 if (err)
1242                         extent_io_tree_panic(tree, err);
1243                 cache_state(prealloc, cached_state);
1244                 prealloc = NULL;
1245                 start = this_end + 1;
1246                 goto search_again;
1247         }
1248         /*
1249          * | ---- desired range ---- |
1250          *                        | state |
1251          * We need to split the extent, and set the bit
1252          * on the first half
1253          */
1254         if (state->start <= end && state->end > end) {
1255                 prealloc = alloc_extent_state_atomic(prealloc);
1256                 if (!prealloc) {
1257                         err = -ENOMEM;
1258                         goto out;
1259                 }
1260
1261                 err = split_state(tree, state, prealloc, end + 1);
1262                 if (err)
1263                         extent_io_tree_panic(tree, err);
1264
1265                 set_state_bits(tree, prealloc, &bits, NULL);
1266                 cache_state(prealloc, cached_state);
1267                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1268                 prealloc = NULL;
1269                 goto out;
1270         }
1271
1272 search_again:
1273         if (start > end)
1274                 goto out;
1275         spin_unlock(&tree->lock);
1276         cond_resched();
1277         first_iteration = false;
1278         goto again;
1279
1280 out:
1281         spin_unlock(&tree->lock);
1282         if (prealloc)
1283                 free_extent_state(prealloc);
1284
1285         return err;
1286 }
1287
1288 /* wrappers around set/clear extent bit */
1289 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1290                            unsigned bits, struct extent_changeset *changeset)
1291 {
1292         /*
1293          * We don't support EXTENT_LOCKED yet, as current changeset will
1294          * record any bits changed, so for EXTENT_LOCKED case, it will
1295          * either fail with -EEXIST or changeset will record the whole
1296          * range.
1297          */
1298         BUG_ON(bits & EXTENT_LOCKED);
1299
1300         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1301                                 changeset);
1302 }
1303
1304 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1305                      unsigned bits, int wake, int delete,
1306                      struct extent_state **cached, gfp_t mask)
1307 {
1308         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1309                                   cached, mask, NULL);
1310 }
1311
1312 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1313                 unsigned bits, struct extent_changeset *changeset)
1314 {
1315         /*
1316          * Don't support EXTENT_LOCKED case, same reason as
1317          * set_record_extent_bits().
1318          */
1319         BUG_ON(bits & EXTENT_LOCKED);
1320
1321         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1322                                   changeset);
1323 }
1324
1325 /*
1326  * either insert or lock state struct between start and end use mask to tell
1327  * us if waiting is desired.
1328  */
1329 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1330                      struct extent_state **cached_state)
1331 {
1332         int err;
1333         u64 failed_start;
1334
1335         while (1) {
1336                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1337                                        EXTENT_LOCKED, &failed_start,
1338                                        cached_state, GFP_NOFS, NULL);
1339                 if (err == -EEXIST) {
1340                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1341                         start = failed_start;
1342                 } else
1343                         break;
1344                 WARN_ON(start > end);
1345         }
1346         return err;
1347 }
1348
1349 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1350 {
1351         int err;
1352         u64 failed_start;
1353
1354         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1355                                &failed_start, NULL, GFP_NOFS, NULL);
1356         if (err == -EEXIST) {
1357                 if (failed_start > start)
1358                         clear_extent_bit(tree, start, failed_start - 1,
1359                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1360                 return 0;
1361         }
1362         return 1;
1363 }
1364
1365 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1366 {
1367         unsigned long index = start >> PAGE_SHIFT;
1368         unsigned long end_index = end >> PAGE_SHIFT;
1369         struct page *page;
1370
1371         while (index <= end_index) {
1372                 page = find_get_page(inode->i_mapping, index);
1373                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1374                 clear_page_dirty_for_io(page);
1375                 put_page(page);
1376                 index++;
1377         }
1378 }
1379
1380 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1381 {
1382         unsigned long index = start >> PAGE_SHIFT;
1383         unsigned long end_index = end >> PAGE_SHIFT;
1384         struct page *page;
1385
1386         while (index <= end_index) {
1387                 page = find_get_page(inode->i_mapping, index);
1388                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1389                 __set_page_dirty_nobuffers(page);
1390                 account_page_redirty(page);
1391                 put_page(page);
1392                 index++;
1393         }
1394 }
1395
1396 /*
1397  * helper function to set both pages and extents in the tree writeback
1398  */
1399 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1400 {
1401         unsigned long index = start >> PAGE_SHIFT;
1402         unsigned long end_index = end >> PAGE_SHIFT;
1403         struct page *page;
1404
1405         while (index <= end_index) {
1406                 page = find_get_page(tree->mapping, index);
1407                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1408                 set_page_writeback(page);
1409                 put_page(page);
1410                 index++;
1411         }
1412 }
1413
1414 /* find the first state struct with 'bits' set after 'start', and
1415  * return it.  tree->lock must be held.  NULL will returned if
1416  * nothing was found after 'start'
1417  */
1418 static struct extent_state *
1419 find_first_extent_bit_state(struct extent_io_tree *tree,
1420                             u64 start, unsigned bits)
1421 {
1422         struct rb_node *node;
1423         struct extent_state *state;
1424
1425         /*
1426          * this search will find all the extents that end after
1427          * our range starts.
1428          */
1429         node = tree_search(tree, start);
1430         if (!node)
1431                 goto out;
1432
1433         while (1) {
1434                 state = rb_entry(node, struct extent_state, rb_node);
1435                 if (state->end >= start && (state->state & bits))
1436                         return state;
1437
1438                 node = rb_next(node);
1439                 if (!node)
1440                         break;
1441         }
1442 out:
1443         return NULL;
1444 }
1445
1446 /*
1447  * find the first offset in the io tree with 'bits' set. zero is
1448  * returned if we find something, and *start_ret and *end_ret are
1449  * set to reflect the state struct that was found.
1450  *
1451  * If nothing was found, 1 is returned. If found something, return 0.
1452  */
1453 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1454                           u64 *start_ret, u64 *end_ret, unsigned bits,
1455                           struct extent_state **cached_state)
1456 {
1457         struct extent_state *state;
1458         struct rb_node *n;
1459         int ret = 1;
1460
1461         spin_lock(&tree->lock);
1462         if (cached_state && *cached_state) {
1463                 state = *cached_state;
1464                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1465                         n = rb_next(&state->rb_node);
1466                         while (n) {
1467                                 state = rb_entry(n, struct extent_state,
1468                                                  rb_node);
1469                                 if (state->state & bits)
1470                                         goto got_it;
1471                                 n = rb_next(n);
1472                         }
1473                         free_extent_state(*cached_state);
1474                         *cached_state = NULL;
1475                         goto out;
1476                 }
1477                 free_extent_state(*cached_state);
1478                 *cached_state = NULL;
1479         }
1480
1481         state = find_first_extent_bit_state(tree, start, bits);
1482 got_it:
1483         if (state) {
1484                 cache_state_if_flags(state, cached_state, 0);
1485                 *start_ret = state->start;
1486                 *end_ret = state->end;
1487                 ret = 0;
1488         }
1489 out:
1490         spin_unlock(&tree->lock);
1491         return ret;
1492 }
1493
1494 /*
1495  * find a contiguous range of bytes in the file marked as delalloc, not
1496  * more than 'max_bytes'.  start and end are used to return the range,
1497  *
1498  * 1 is returned if we find something, 0 if nothing was in the tree
1499  */
1500 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1501                                         u64 *start, u64 *end, u64 max_bytes,
1502                                         struct extent_state **cached_state)
1503 {
1504         struct rb_node *node;
1505         struct extent_state *state;
1506         u64 cur_start = *start;
1507         u64 found = 0;
1508         u64 total_bytes = 0;
1509
1510         spin_lock(&tree->lock);
1511
1512         /*
1513          * this search will find all the extents that end after
1514          * our range starts.
1515          */
1516         node = tree_search(tree, cur_start);
1517         if (!node) {
1518                 if (!found)
1519                         *end = (u64)-1;
1520                 goto out;
1521         }
1522
1523         while (1) {
1524                 state = rb_entry(node, struct extent_state, rb_node);
1525                 if (found && (state->start != cur_start ||
1526                               (state->state & EXTENT_BOUNDARY))) {
1527                         goto out;
1528                 }
1529                 if (!(state->state & EXTENT_DELALLOC)) {
1530                         if (!found)
1531                                 *end = state->end;
1532                         goto out;
1533                 }
1534                 if (!found) {
1535                         *start = state->start;
1536                         *cached_state = state;
1537                         atomic_inc(&state->refs);
1538                 }
1539                 found++;
1540                 *end = state->end;
1541                 cur_start = state->end + 1;
1542                 node = rb_next(node);
1543                 total_bytes += state->end - state->start + 1;
1544                 if (total_bytes >= max_bytes)
1545                         break;
1546                 if (!node)
1547                         break;
1548         }
1549 out:
1550         spin_unlock(&tree->lock);
1551         return found;
1552 }
1553
1554 static noinline void __unlock_for_delalloc(struct inode *inode,
1555                                            struct page *locked_page,
1556                                            u64 start, u64 end)
1557 {
1558         int ret;
1559         struct page *pages[16];
1560         unsigned long index = start >> PAGE_SHIFT;
1561         unsigned long end_index = end >> PAGE_SHIFT;
1562         unsigned long nr_pages = end_index - index + 1;
1563         int i;
1564
1565         if (index == locked_page->index && end_index == index)
1566                 return;
1567
1568         while (nr_pages > 0) {
1569                 ret = find_get_pages_contig(inode->i_mapping, index,
1570                                      min_t(unsigned long, nr_pages,
1571                                      ARRAY_SIZE(pages)), pages);
1572                 for (i = 0; i < ret; i++) {
1573                         if (pages[i] != locked_page)
1574                                 unlock_page(pages[i]);
1575                         put_page(pages[i]);
1576                 }
1577                 nr_pages -= ret;
1578                 index += ret;
1579                 cond_resched();
1580         }
1581 }
1582
1583 static noinline int lock_delalloc_pages(struct inode *inode,
1584                                         struct page *locked_page,
1585                                         u64 delalloc_start,
1586                                         u64 delalloc_end)
1587 {
1588         unsigned long index = delalloc_start >> PAGE_SHIFT;
1589         unsigned long start_index = index;
1590         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1591         unsigned long pages_locked = 0;
1592         struct page *pages[16];
1593         unsigned long nrpages;
1594         int ret;
1595         int i;
1596
1597         /* the caller is responsible for locking the start index */
1598         if (index == locked_page->index && index == end_index)
1599                 return 0;
1600
1601         /* skip the page at the start index */
1602         nrpages = end_index - index + 1;
1603         while (nrpages > 0) {
1604                 ret = find_get_pages_contig(inode->i_mapping, index,
1605                                      min_t(unsigned long,
1606                                      nrpages, ARRAY_SIZE(pages)), pages);
1607                 if (ret == 0) {
1608                         ret = -EAGAIN;
1609                         goto done;
1610                 }
1611                 /* now we have an array of pages, lock them all */
1612                 for (i = 0; i < ret; i++) {
1613                         /*
1614                          * the caller is taking responsibility for
1615                          * locked_page
1616                          */
1617                         if (pages[i] != locked_page) {
1618                                 lock_page(pages[i]);
1619                                 if (!PageDirty(pages[i]) ||
1620                                     pages[i]->mapping != inode->i_mapping) {
1621                                         ret = -EAGAIN;
1622                                         unlock_page(pages[i]);
1623                                         put_page(pages[i]);
1624                                         goto done;
1625                                 }
1626                         }
1627                         put_page(pages[i]);
1628                         pages_locked++;
1629                 }
1630                 nrpages -= ret;
1631                 index += ret;
1632                 cond_resched();
1633         }
1634         ret = 0;
1635 done:
1636         if (ret && pages_locked) {
1637                 __unlock_for_delalloc(inode, locked_page,
1638                               delalloc_start,
1639                               ((u64)(start_index + pages_locked - 1)) <<
1640                               PAGE_SHIFT);
1641         }
1642         return ret;
1643 }
1644
1645 /*
1646  * find a contiguous range of bytes in the file marked as delalloc, not
1647  * more than 'max_bytes'.  start and end are used to return the range,
1648  *
1649  * 1 is returned if we find something, 0 if nothing was in the tree
1650  */
1651 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1652                                     struct extent_io_tree *tree,
1653                                     struct page *locked_page, u64 *start,
1654                                     u64 *end, u64 max_bytes)
1655 {
1656         u64 delalloc_start;
1657         u64 delalloc_end;
1658         u64 found;
1659         struct extent_state *cached_state = NULL;
1660         int ret;
1661         int loops = 0;
1662
1663 again:
1664         /* step one, find a bunch of delalloc bytes starting at start */
1665         delalloc_start = *start;
1666         delalloc_end = 0;
1667         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1668                                     max_bytes, &cached_state);
1669         if (!found || delalloc_end <= *start) {
1670                 *start = delalloc_start;
1671                 *end = delalloc_end;
1672                 free_extent_state(cached_state);
1673                 return 0;
1674         }
1675
1676         /*
1677          * start comes from the offset of locked_page.  We have to lock
1678          * pages in order, so we can't process delalloc bytes before
1679          * locked_page
1680          */
1681         if (delalloc_start < *start)
1682                 delalloc_start = *start;
1683
1684         /*
1685          * make sure to limit the number of pages we try to lock down
1686          */
1687         if (delalloc_end + 1 - delalloc_start > max_bytes)
1688                 delalloc_end = delalloc_start + max_bytes - 1;
1689
1690         /* step two, lock all the pages after the page that has start */
1691         ret = lock_delalloc_pages(inode, locked_page,
1692                                   delalloc_start, delalloc_end);
1693         if (ret == -EAGAIN) {
1694                 /* some of the pages are gone, lets avoid looping by
1695                  * shortening the size of the delalloc range we're searching
1696                  */
1697                 free_extent_state(cached_state);
1698                 cached_state = NULL;
1699                 if (!loops) {
1700                         max_bytes = PAGE_SIZE;
1701                         loops = 1;
1702                         goto again;
1703                 } else {
1704                         found = 0;
1705                         goto out_failed;
1706                 }
1707         }
1708         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1709
1710         /* step three, lock the state bits for the whole range */
1711         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1712
1713         /* then test to make sure it is all still delalloc */
1714         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1715                              EXTENT_DELALLOC, 1, cached_state);
1716         if (!ret) {
1717                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1718                                      &cached_state, GFP_NOFS);
1719                 __unlock_for_delalloc(inode, locked_page,
1720                               delalloc_start, delalloc_end);
1721                 cond_resched();
1722                 goto again;
1723         }
1724         free_extent_state(cached_state);
1725         *start = delalloc_start;
1726         *end = delalloc_end;
1727 out_failed:
1728         return found;
1729 }
1730
1731 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1732                                  struct page *locked_page,
1733                                  unsigned clear_bits,
1734                                  unsigned long page_ops)
1735 {
1736         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1737         int ret;
1738         struct page *pages[16];
1739         unsigned long index = start >> PAGE_SHIFT;
1740         unsigned long end_index = end >> PAGE_SHIFT;
1741         unsigned long nr_pages = end_index - index + 1;
1742         int i;
1743
1744         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1745         if (page_ops == 0)
1746                 return;
1747
1748         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1749                 mapping_set_error(inode->i_mapping, -EIO);
1750
1751         while (nr_pages > 0) {
1752                 ret = find_get_pages_contig(inode->i_mapping, index,
1753                                      min_t(unsigned long,
1754                                      nr_pages, ARRAY_SIZE(pages)), pages);
1755                 for (i = 0; i < ret; i++) {
1756
1757                         if (page_ops & PAGE_SET_PRIVATE2)
1758                                 SetPagePrivate2(pages[i]);
1759
1760                         if (pages[i] == locked_page) {
1761                                 put_page(pages[i]);
1762                                 continue;
1763                         }
1764                         if (page_ops & PAGE_CLEAR_DIRTY)
1765                                 clear_page_dirty_for_io(pages[i]);
1766                         if (page_ops & PAGE_SET_WRITEBACK)
1767                                 set_page_writeback(pages[i]);
1768                         if (page_ops & PAGE_SET_ERROR)
1769                                 SetPageError(pages[i]);
1770                         if (page_ops & PAGE_END_WRITEBACK)
1771                                 end_page_writeback(pages[i]);
1772                         if (page_ops & PAGE_UNLOCK)
1773                                 unlock_page(pages[i]);
1774                         put_page(pages[i]);
1775                 }
1776                 nr_pages -= ret;
1777                 index += ret;
1778                 cond_resched();
1779         }
1780 }
1781
1782 /*
1783  * count the number of bytes in the tree that have a given bit(s)
1784  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1785  * cached.  The total number found is returned.
1786  */
1787 u64 count_range_bits(struct extent_io_tree *tree,
1788                      u64 *start, u64 search_end, u64 max_bytes,
1789                      unsigned bits, int contig)
1790 {
1791         struct rb_node *node;
1792         struct extent_state *state;
1793         u64 cur_start = *start;
1794         u64 total_bytes = 0;
1795         u64 last = 0;
1796         int found = 0;
1797
1798         if (WARN_ON(search_end <= cur_start))
1799                 return 0;
1800
1801         spin_lock(&tree->lock);
1802         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1803                 total_bytes = tree->dirty_bytes;
1804                 goto out;
1805         }
1806         /*
1807          * this search will find all the extents that end after
1808          * our range starts.
1809          */
1810         node = tree_search(tree, cur_start);
1811         if (!node)
1812                 goto out;
1813
1814         while (1) {
1815                 state = rb_entry(node, struct extent_state, rb_node);
1816                 if (state->start > search_end)
1817                         break;
1818                 if (contig && found && state->start > last + 1)
1819                         break;
1820                 if (state->end >= cur_start && (state->state & bits) == bits) {
1821                         total_bytes += min(search_end, state->end) + 1 -
1822                                        max(cur_start, state->start);
1823                         if (total_bytes >= max_bytes)
1824                                 break;
1825                         if (!found) {
1826                                 *start = max(cur_start, state->start);
1827                                 found = 1;
1828                         }
1829                         last = state->end;
1830                 } else if (contig && found) {
1831                         break;
1832                 }
1833                 node = rb_next(node);
1834                 if (!node)
1835                         break;
1836         }
1837 out:
1838         spin_unlock(&tree->lock);
1839         return total_bytes;
1840 }
1841
1842 /*
1843  * set the private field for a given byte offset in the tree.  If there isn't
1844  * an extent_state there already, this does nothing.
1845  */
1846 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1847                 struct io_failure_record *failrec)
1848 {
1849         struct rb_node *node;
1850         struct extent_state *state;
1851         int ret = 0;
1852
1853         spin_lock(&tree->lock);
1854         /*
1855          * this search will find all the extents that end after
1856          * our range starts.
1857          */
1858         node = tree_search(tree, start);
1859         if (!node) {
1860                 ret = -ENOENT;
1861                 goto out;
1862         }
1863         state = rb_entry(node, struct extent_state, rb_node);
1864         if (state->start != start) {
1865                 ret = -ENOENT;
1866                 goto out;
1867         }
1868         state->failrec = failrec;
1869 out:
1870         spin_unlock(&tree->lock);
1871         return ret;
1872 }
1873
1874 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1875                 struct io_failure_record **failrec)
1876 {
1877         struct rb_node *node;
1878         struct extent_state *state;
1879         int ret = 0;
1880
1881         spin_lock(&tree->lock);
1882         /*
1883          * this search will find all the extents that end after
1884          * our range starts.
1885          */
1886         node = tree_search(tree, start);
1887         if (!node) {
1888                 ret = -ENOENT;
1889                 goto out;
1890         }
1891         state = rb_entry(node, struct extent_state, rb_node);
1892         if (state->start != start) {
1893                 ret = -ENOENT;
1894                 goto out;
1895         }
1896         *failrec = state->failrec;
1897 out:
1898         spin_unlock(&tree->lock);
1899         return ret;
1900 }
1901
1902 /*
1903  * searches a range in the state tree for a given mask.
1904  * If 'filled' == 1, this returns 1 only if every extent in the tree
1905  * has the bits set.  Otherwise, 1 is returned if any bit in the
1906  * range is found set.
1907  */
1908 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1909                    unsigned bits, int filled, struct extent_state *cached)
1910 {
1911         struct extent_state *state = NULL;
1912         struct rb_node *node;
1913         int bitset = 0;
1914
1915         spin_lock(&tree->lock);
1916         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1917             cached->end > start)
1918                 node = &cached->rb_node;
1919         else
1920                 node = tree_search(tree, start);
1921         while (node && start <= end) {
1922                 state = rb_entry(node, struct extent_state, rb_node);
1923
1924                 if (filled && state->start > start) {
1925                         bitset = 0;
1926                         break;
1927                 }
1928
1929                 if (state->start > end)
1930                         break;
1931
1932                 if (state->state & bits) {
1933                         bitset = 1;
1934                         if (!filled)
1935                                 break;
1936                 } else if (filled) {
1937                         bitset = 0;
1938                         break;
1939                 }
1940
1941                 if (state->end == (u64)-1)
1942                         break;
1943
1944                 start = state->end + 1;
1945                 if (start > end)
1946                         break;
1947                 node = rb_next(node);
1948                 if (!node) {
1949                         if (filled)
1950                                 bitset = 0;
1951                         break;
1952                 }
1953         }
1954         spin_unlock(&tree->lock);
1955         return bitset;
1956 }
1957
1958 /*
1959  * helper function to set a given page up to date if all the
1960  * extents in the tree for that page are up to date
1961  */
1962 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1963 {
1964         u64 start = page_offset(page);
1965         u64 end = start + PAGE_SIZE - 1;
1966         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1967                 SetPageUptodate(page);
1968 }
1969
1970 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1971 {
1972         int ret;
1973         int err = 0;
1974         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1975
1976         set_state_failrec(failure_tree, rec->start, NULL);
1977         ret = clear_extent_bits(failure_tree, rec->start,
1978                                 rec->start + rec->len - 1,
1979                                 EXTENT_LOCKED | EXTENT_DIRTY);
1980         if (ret)
1981                 err = ret;
1982
1983         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1984                                 rec->start + rec->len - 1,
1985                                 EXTENT_DAMAGED);
1986         if (ret && !err)
1987                 err = ret;
1988
1989         kfree(rec);
1990         return err;
1991 }
1992
1993 /*
1994  * this bypasses the standard btrfs submit functions deliberately, as
1995  * the standard behavior is to write all copies in a raid setup. here we only
1996  * want to write the one bad copy. so we do the mapping for ourselves and issue
1997  * submit_bio directly.
1998  * to avoid any synchronization issues, wait for the data after writing, which
1999  * actually prevents the read that triggered the error from finishing.
2000  * currently, there can be no more than two copies of every data bit. thus,
2001  * exactly one rewrite is required.
2002  */
2003 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2004                       struct page *page, unsigned int pg_offset, int mirror_num)
2005 {
2006         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2007         struct bio *bio;
2008         struct btrfs_device *dev;
2009         u64 map_length = 0;
2010         u64 sector;
2011         struct btrfs_bio *bbio = NULL;
2012         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2013         int ret;
2014
2015         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2016         BUG_ON(!mirror_num);
2017
2018         /* we can't repair anything in raid56 yet */
2019         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2020                 return 0;
2021
2022         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2023         if (!bio)
2024                 return -EIO;
2025         bio->bi_iter.bi_size = 0;
2026         map_length = length;
2027
2028         /*
2029          * Avoid races with device replace and make sure our bbio has devices
2030          * associated to its stripes that don't go away while we are doing the
2031          * read repair operation.
2032          */
2033         btrfs_bio_counter_inc_blocked(fs_info);
2034         ret = btrfs_map_block(fs_info, WRITE, logical,
2035                               &map_length, &bbio, mirror_num);
2036         if (ret) {
2037                 btrfs_bio_counter_dec(fs_info);
2038                 bio_put(bio);
2039                 return -EIO;
2040         }
2041         BUG_ON(mirror_num != bbio->mirror_num);
2042         sector = bbio->stripes[mirror_num-1].physical >> 9;
2043         bio->bi_iter.bi_sector = sector;
2044         dev = bbio->stripes[mirror_num-1].dev;
2045         btrfs_put_bbio(bbio);
2046         if (!dev || !dev->bdev || !dev->writeable) {
2047                 btrfs_bio_counter_dec(fs_info);
2048                 bio_put(bio);
2049                 return -EIO;
2050         }
2051         bio->bi_bdev = dev->bdev;
2052         bio->bi_rw = WRITE_SYNC;
2053         bio_add_page(bio, page, length, pg_offset);
2054
2055         if (btrfsic_submit_bio_wait(bio)) {
2056                 /* try to remap that extent elsewhere? */
2057                 btrfs_bio_counter_dec(fs_info);
2058                 bio_put(bio);
2059                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2060                 return -EIO;
2061         }
2062
2063         btrfs_info_rl_in_rcu(fs_info,
2064                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2065                                   btrfs_ino(inode), start,
2066                                   rcu_str_deref(dev->name), sector);
2067         btrfs_bio_counter_dec(fs_info);
2068         bio_put(bio);
2069         return 0;
2070 }
2071
2072 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2073                          int mirror_num)
2074 {
2075         u64 start = eb->start;
2076         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2077         int ret = 0;
2078
2079         if (root->fs_info->sb->s_flags & MS_RDONLY)
2080                 return -EROFS;
2081
2082         for (i = 0; i < num_pages; i++) {
2083                 struct page *p = eb->pages[i];
2084
2085                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2086                                         PAGE_SIZE, start, p,
2087                                         start - page_offset(p), mirror_num);
2088                 if (ret)
2089                         break;
2090                 start += PAGE_SIZE;
2091         }
2092
2093         return ret;
2094 }
2095
2096 /*
2097  * each time an IO finishes, we do a fast check in the IO failure tree
2098  * to see if we need to process or clean up an io_failure_record
2099  */
2100 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2101                      unsigned int pg_offset)
2102 {
2103         u64 private;
2104         struct io_failure_record *failrec;
2105         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2106         struct extent_state *state;
2107         int num_copies;
2108         int ret;
2109
2110         private = 0;
2111         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2112                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2113         if (!ret)
2114                 return 0;
2115
2116         ret = get_state_failrec(&BTRFS_I(inode)->io_failure_tree, start,
2117                         &failrec);
2118         if (ret)
2119                 return 0;
2120
2121         BUG_ON(!failrec->this_mirror);
2122
2123         if (failrec->in_validation) {
2124                 /* there was no real error, just free the record */
2125                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2126                          failrec->start);
2127                 goto out;
2128         }
2129         if (fs_info->sb->s_flags & MS_RDONLY)
2130                 goto out;
2131
2132         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2133         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2134                                             failrec->start,
2135                                             EXTENT_LOCKED);
2136         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2137
2138         if (state && state->start <= failrec->start &&
2139             state->end >= failrec->start + failrec->len - 1) {
2140                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2141                                               failrec->len);
2142                 if (num_copies > 1)  {
2143                         repair_io_failure(inode, start, failrec->len,
2144                                           failrec->logical, page,
2145                                           pg_offset, failrec->failed_mirror);
2146                 }
2147         }
2148
2149 out:
2150         free_io_failure(inode, failrec);
2151
2152         return 0;
2153 }
2154
2155 /*
2156  * Can be called when
2157  * - hold extent lock
2158  * - under ordered extent
2159  * - the inode is freeing
2160  */
2161 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2162 {
2163         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2164         struct io_failure_record *failrec;
2165         struct extent_state *state, *next;
2166
2167         if (RB_EMPTY_ROOT(&failure_tree->state))
2168                 return;
2169
2170         spin_lock(&failure_tree->lock);
2171         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2172         while (state) {
2173                 if (state->start > end)
2174                         break;
2175
2176                 ASSERT(state->end <= end);
2177
2178                 next = next_state(state);
2179
2180                 failrec = state->failrec;
2181                 free_extent_state(state);
2182                 kfree(failrec);
2183
2184                 state = next;
2185         }
2186         spin_unlock(&failure_tree->lock);
2187 }
2188
2189 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2190                 struct io_failure_record **failrec_ret)
2191 {
2192         struct io_failure_record *failrec;
2193         struct extent_map *em;
2194         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2195         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2196         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2197         int ret;
2198         u64 logical;
2199
2200         ret = get_state_failrec(failure_tree, start, &failrec);
2201         if (ret) {
2202                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2203                 if (!failrec)
2204                         return -ENOMEM;
2205
2206                 failrec->start = start;
2207                 failrec->len = end - start + 1;
2208                 failrec->this_mirror = 0;
2209                 failrec->bio_flags = 0;
2210                 failrec->in_validation = 0;
2211
2212                 read_lock(&em_tree->lock);
2213                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2214                 if (!em) {
2215                         read_unlock(&em_tree->lock);
2216                         kfree(failrec);
2217                         return -EIO;
2218                 }
2219
2220                 if (em->start > start || em->start + em->len <= start) {
2221                         free_extent_map(em);
2222                         em = NULL;
2223                 }
2224                 read_unlock(&em_tree->lock);
2225                 if (!em) {
2226                         kfree(failrec);
2227                         return -EIO;
2228                 }
2229
2230                 logical = start - em->start;
2231                 logical = em->block_start + logical;
2232                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2233                         logical = em->block_start;
2234                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2235                         extent_set_compress_type(&failrec->bio_flags,
2236                                                  em->compress_type);
2237                 }
2238
2239                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2240                          logical, start, failrec->len);
2241
2242                 failrec->logical = logical;
2243                 free_extent_map(em);
2244
2245                 /* set the bits in the private failure tree */
2246                 ret = set_extent_bits(failure_tree, start, end,
2247                                         EXTENT_LOCKED | EXTENT_DIRTY);
2248                 if (ret >= 0)
2249                         ret = set_state_failrec(failure_tree, start, failrec);
2250                 /* set the bits in the inode's tree */
2251                 if (ret >= 0)
2252                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2253                 if (ret < 0) {
2254                         kfree(failrec);
2255                         return ret;
2256                 }
2257         } else {
2258                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2259                          failrec->logical, failrec->start, failrec->len,
2260                          failrec->in_validation);
2261                 /*
2262                  * when data can be on disk more than twice, add to failrec here
2263                  * (e.g. with a list for failed_mirror) to make
2264                  * clean_io_failure() clean all those errors at once.
2265                  */
2266         }
2267
2268         *failrec_ret = failrec;
2269
2270         return 0;
2271 }
2272
2273 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2274                            struct io_failure_record *failrec, int failed_mirror)
2275 {
2276         int num_copies;
2277
2278         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2279                                       failrec->logical, failrec->len);
2280         if (num_copies == 1) {
2281                 /*
2282                  * we only have a single copy of the data, so don't bother with
2283                  * all the retry and error correction code that follows. no
2284                  * matter what the error is, it is very likely to persist.
2285                  */
2286                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2287                          num_copies, failrec->this_mirror, failed_mirror);
2288                 return 0;
2289         }
2290
2291         /*
2292          * there are two premises:
2293          *      a) deliver good data to the caller
2294          *      b) correct the bad sectors on disk
2295          */
2296         if (failed_bio->bi_vcnt > 1) {
2297                 /*
2298                  * to fulfill b), we need to know the exact failing sectors, as
2299                  * we don't want to rewrite any more than the failed ones. thus,
2300                  * we need separate read requests for the failed bio
2301                  *
2302                  * if the following BUG_ON triggers, our validation request got
2303                  * merged. we need separate requests for our algorithm to work.
2304                  */
2305                 BUG_ON(failrec->in_validation);
2306                 failrec->in_validation = 1;
2307                 failrec->this_mirror = failed_mirror;
2308         } else {
2309                 /*
2310                  * we're ready to fulfill a) and b) alongside. get a good copy
2311                  * of the failed sector and if we succeed, we have setup
2312                  * everything for repair_io_failure to do the rest for us.
2313                  */
2314                 if (failrec->in_validation) {
2315                         BUG_ON(failrec->this_mirror != failed_mirror);
2316                         failrec->in_validation = 0;
2317                         failrec->this_mirror = 0;
2318                 }
2319                 failrec->failed_mirror = failed_mirror;
2320                 failrec->this_mirror++;
2321                 if (failrec->this_mirror == failed_mirror)
2322                         failrec->this_mirror++;
2323         }
2324
2325         if (failrec->this_mirror > num_copies) {
2326                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2327                          num_copies, failrec->this_mirror, failed_mirror);
2328                 return 0;
2329         }
2330
2331         return 1;
2332 }
2333
2334
2335 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2336                                     struct io_failure_record *failrec,
2337                                     struct page *page, int pg_offset, int icsum,
2338                                     bio_end_io_t *endio_func, void *data)
2339 {
2340         struct bio *bio;
2341         struct btrfs_io_bio *btrfs_failed_bio;
2342         struct btrfs_io_bio *btrfs_bio;
2343
2344         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2345         if (!bio)
2346                 return NULL;
2347
2348         bio->bi_end_io = endio_func;
2349         bio->bi_iter.bi_sector = failrec->logical >> 9;
2350         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2351         bio->bi_iter.bi_size = 0;
2352         bio->bi_private = data;
2353
2354         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2355         if (btrfs_failed_bio->csum) {
2356                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2357                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2358
2359                 btrfs_bio = btrfs_io_bio(bio);
2360                 btrfs_bio->csum = btrfs_bio->csum_inline;
2361                 icsum *= csum_size;
2362                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2363                        csum_size);
2364         }
2365
2366         bio_add_page(bio, page, failrec->len, pg_offset);
2367
2368         return bio;
2369 }
2370
2371 /*
2372  * this is a generic handler for readpage errors (default
2373  * readpage_io_failed_hook). if other copies exist, read those and write back
2374  * good data to the failed position. does not investigate in remapping the
2375  * failed extent elsewhere, hoping the device will be smart enough to do this as
2376  * needed
2377  */
2378
2379 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2380                               struct page *page, u64 start, u64 end,
2381                               int failed_mirror)
2382 {
2383         struct io_failure_record *failrec;
2384         struct inode *inode = page->mapping->host;
2385         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2386         struct bio *bio;
2387         int read_mode;
2388         int ret;
2389
2390         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2391
2392         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2393         if (ret)
2394                 return ret;
2395
2396         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2397         if (!ret) {
2398                 free_io_failure(inode, failrec);
2399                 return -EIO;
2400         }
2401
2402         if (failed_bio->bi_vcnt > 1)
2403                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2404         else
2405                 read_mode = READ_SYNC;
2406
2407         phy_offset >>= inode->i_sb->s_blocksize_bits;
2408         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2409                                       start - page_offset(page),
2410                                       (int)phy_offset, failed_bio->bi_end_io,
2411                                       NULL);
2412         if (!bio) {
2413                 free_io_failure(inode, failrec);
2414                 return -EIO;
2415         }
2416         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2417
2418         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2419                  read_mode, failrec->this_mirror, failrec->in_validation);
2420
2421         ret = tree->ops->submit_bio_hook(inode, bio, failrec->this_mirror,
2422                                          failrec->bio_flags, 0);
2423         if (ret) {
2424                 free_io_failure(inode, failrec);
2425                 bio_put(bio);
2426         }
2427
2428         return ret;
2429 }
2430
2431 /* lots and lots of room for performance fixes in the end_bio funcs */
2432
2433 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2434 {
2435         int uptodate = (err == 0);
2436         struct extent_io_tree *tree;
2437         int ret = 0;
2438
2439         tree = &BTRFS_I(page->mapping->host)->io_tree;
2440
2441         if (tree->ops && tree->ops->writepage_end_io_hook) {
2442                 ret = tree->ops->writepage_end_io_hook(page, start,
2443                                                end, NULL, uptodate);
2444                 if (ret)
2445                         uptodate = 0;
2446         }
2447
2448         if (!uptodate) {
2449                 ClearPageUptodate(page);
2450                 SetPageError(page);
2451                 ret = ret < 0 ? ret : -EIO;
2452                 mapping_set_error(page->mapping, ret);
2453         }
2454 }
2455
2456 /*
2457  * after a writepage IO is done, we need to:
2458  * clear the uptodate bits on error
2459  * clear the writeback bits in the extent tree for this IO
2460  * end_page_writeback if the page has no more pending IO
2461  *
2462  * Scheduling is not allowed, so the extent state tree is expected
2463  * to have one and only one object corresponding to this IO.
2464  */
2465 static void end_bio_extent_writepage(struct bio *bio)
2466 {
2467         struct bio_vec *bvec;
2468         u64 start;
2469         u64 end;
2470         int i;
2471
2472         bio_for_each_segment_all(bvec, bio, i) {
2473                 struct page *page = bvec->bv_page;
2474
2475                 /* We always issue full-page reads, but if some block
2476                  * in a page fails to read, blk_update_request() will
2477                  * advance bv_offset and adjust bv_len to compensate.
2478                  * Print a warning for nonzero offsets, and an error
2479                  * if they don't add up to a full page.  */
2480                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2481                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2482                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2483                                    "partial page write in btrfs with offset %u and length %u",
2484                                         bvec->bv_offset, bvec->bv_len);
2485                         else
2486                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2487                                    "incomplete page write in btrfs with offset %u and "
2488                                    "length %u",
2489                                         bvec->bv_offset, bvec->bv_len);
2490                 }
2491
2492                 start = page_offset(page);
2493                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2494
2495                 end_extent_writepage(page, bio->bi_error, start, end);
2496                 end_page_writeback(page);
2497         }
2498
2499         bio_put(bio);
2500 }
2501
2502 static void
2503 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2504                               int uptodate)
2505 {
2506         struct extent_state *cached = NULL;
2507         u64 end = start + len - 1;
2508
2509         if (uptodate && tree->track_uptodate)
2510                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2511         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2512 }
2513
2514 /*
2515  * after a readpage IO is done, we need to:
2516  * clear the uptodate bits on error
2517  * set the uptodate bits if things worked
2518  * set the page up to date if all extents in the tree are uptodate
2519  * clear the lock bit in the extent tree
2520  * unlock the page if there are no other extents locked for it
2521  *
2522  * Scheduling is not allowed, so the extent state tree is expected
2523  * to have one and only one object corresponding to this IO.
2524  */
2525 static void end_bio_extent_readpage(struct bio *bio)
2526 {
2527         struct bio_vec *bvec;
2528         int uptodate = !bio->bi_error;
2529         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2530         struct extent_io_tree *tree;
2531         u64 offset = 0;
2532         u64 start;
2533         u64 end;
2534         u64 len;
2535         u64 extent_start = 0;
2536         u64 extent_len = 0;
2537         int mirror;
2538         int ret;
2539         int i;
2540
2541         bio_for_each_segment_all(bvec, bio, i) {
2542                 struct page *page = bvec->bv_page;
2543                 struct inode *inode = page->mapping->host;
2544
2545                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2546                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector,
2547                          bio->bi_error, io_bio->mirror_num);
2548                 tree = &BTRFS_I(inode)->io_tree;
2549
2550                 /* We always issue full-page reads, but if some block
2551                  * in a page fails to read, blk_update_request() will
2552                  * advance bv_offset and adjust bv_len to compensate.
2553                  * Print a warning for nonzero offsets, and an error
2554                  * if they don't add up to a full page.  */
2555                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2556                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2557                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2558                                    "partial page read in btrfs with offset %u and length %u",
2559                                         bvec->bv_offset, bvec->bv_len);
2560                         else
2561                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2562                                    "incomplete page read in btrfs with offset %u and "
2563                                    "length %u",
2564                                         bvec->bv_offset, bvec->bv_len);
2565                 }
2566
2567                 start = page_offset(page);
2568                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2569                 len = bvec->bv_len;
2570
2571                 mirror = io_bio->mirror_num;
2572                 if (likely(uptodate && tree->ops &&
2573                            tree->ops->readpage_end_io_hook)) {
2574                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2575                                                               page, start, end,
2576                                                               mirror);
2577                         if (ret)
2578                                 uptodate = 0;
2579                         else
2580                                 clean_io_failure(inode, start, page, 0);
2581                 }
2582
2583                 if (likely(uptodate))
2584                         goto readpage_ok;
2585
2586                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2587                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2588                         if (!ret && !bio->bi_error)
2589                                 uptodate = 1;
2590                 } else {
2591                         /*
2592                          * The generic bio_readpage_error handles errors the
2593                          * following way: If possible, new read requests are
2594                          * created and submitted and will end up in
2595                          * end_bio_extent_readpage as well (if we're lucky, not
2596                          * in the !uptodate case). In that case it returns 0 and
2597                          * we just go on with the next page in our bio. If it
2598                          * can't handle the error it will return -EIO and we
2599                          * remain responsible for that page.
2600                          */
2601                         ret = bio_readpage_error(bio, offset, page, start, end,
2602                                                  mirror);
2603                         if (ret == 0) {
2604                                 uptodate = !bio->bi_error;
2605                                 offset += len;
2606                                 continue;
2607                         }
2608                 }
2609 readpage_ok:
2610                 if (likely(uptodate)) {
2611                         loff_t i_size = i_size_read(inode);
2612                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2613                         unsigned off;
2614
2615                         /* Zero out the end if this page straddles i_size */
2616                         off = i_size & (PAGE_SIZE-1);
2617                         if (page->index == end_index && off)
2618                                 zero_user_segment(page, off, PAGE_SIZE);
2619                         SetPageUptodate(page);
2620                 } else {
2621                         ClearPageUptodate(page);
2622                         SetPageError(page);
2623                 }
2624                 unlock_page(page);
2625                 offset += len;
2626
2627                 if (unlikely(!uptodate)) {
2628                         if (extent_len) {
2629                                 endio_readpage_release_extent(tree,
2630                                                               extent_start,
2631                                                               extent_len, 1);
2632                                 extent_start = 0;
2633                                 extent_len = 0;
2634                         }
2635                         endio_readpage_release_extent(tree, start,
2636                                                       end - start + 1, 0);
2637                 } else if (!extent_len) {
2638                         extent_start = start;
2639                         extent_len = end + 1 - start;
2640                 } else if (extent_start + extent_len == start) {
2641                         extent_len += end + 1 - start;
2642                 } else {
2643                         endio_readpage_release_extent(tree, extent_start,
2644                                                       extent_len, uptodate);
2645                         extent_start = start;
2646                         extent_len = end + 1 - start;
2647                 }
2648         }
2649
2650         if (extent_len)
2651                 endio_readpage_release_extent(tree, extent_start, extent_len,
2652                                               uptodate);
2653         if (io_bio->end_io)
2654                 io_bio->end_io(io_bio, bio->bi_error);
2655         bio_put(bio);
2656 }
2657
2658 /*
2659  * this allocates from the btrfs_bioset.  We're returning a bio right now
2660  * but you can call btrfs_io_bio for the appropriate container_of magic
2661  */
2662 struct bio *
2663 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2664                 gfp_t gfp_flags)
2665 {
2666         struct btrfs_io_bio *btrfs_bio;
2667         struct bio *bio;
2668
2669         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2670
2671         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2672                 while (!bio && (nr_vecs /= 2)) {
2673                         bio = bio_alloc_bioset(gfp_flags,
2674                                                nr_vecs, btrfs_bioset);
2675                 }
2676         }
2677
2678         if (bio) {
2679                 bio->bi_bdev = bdev;
2680                 bio->bi_iter.bi_sector = first_sector;
2681                 btrfs_bio = btrfs_io_bio(bio);
2682                 btrfs_bio->csum = NULL;
2683                 btrfs_bio->csum_allocated = NULL;
2684                 btrfs_bio->end_io = NULL;
2685         }
2686         return bio;
2687 }
2688
2689 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2690 {
2691         struct btrfs_io_bio *btrfs_bio;
2692         struct bio *new;
2693
2694         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2695         if (new) {
2696                 btrfs_bio = btrfs_io_bio(new);
2697                 btrfs_bio->csum = NULL;
2698                 btrfs_bio->csum_allocated = NULL;
2699                 btrfs_bio->end_io = NULL;
2700
2701 #ifdef CONFIG_BLK_CGROUP
2702                 /* FIXME, put this into bio_clone_bioset */
2703                 if (bio->bi_css)
2704                         bio_associate_blkcg(new, bio->bi_css);
2705 #endif
2706         }
2707         return new;
2708 }
2709
2710 /* this also allocates from the btrfs_bioset */
2711 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2712 {
2713         struct btrfs_io_bio *btrfs_bio;
2714         struct bio *bio;
2715
2716         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2717         if (bio) {
2718                 btrfs_bio = btrfs_io_bio(bio);
2719                 btrfs_bio->csum = NULL;
2720                 btrfs_bio->csum_allocated = NULL;
2721                 btrfs_bio->end_io = NULL;
2722         }
2723         return bio;
2724 }
2725
2726
2727 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2728                                        unsigned long bio_flags)
2729 {
2730         int ret = 0;
2731         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2732         struct page *page = bvec->bv_page;
2733         struct extent_io_tree *tree = bio->bi_private;
2734         u64 start;
2735
2736         start = page_offset(page) + bvec->bv_offset;
2737
2738         bio->bi_private = NULL;
2739         bio_get(bio);
2740
2741         if (tree->ops && tree->ops->submit_bio_hook)
2742                 ret = tree->ops->submit_bio_hook(page->mapping->host, bio,
2743                                            mirror_num, bio_flags, start);
2744         else
2745                 btrfsic_submit_bio(bio);
2746
2747         bio_put(bio);
2748         return ret;
2749 }
2750
2751 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2752                      unsigned long offset, size_t size, struct bio *bio,
2753                      unsigned long bio_flags)
2754 {
2755         int ret = 0;
2756         if (tree->ops && tree->ops->merge_bio_hook)
2757                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2758                                                 bio_flags);
2759         return ret;
2760
2761 }
2762
2763 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2764                               struct writeback_control *wbc,
2765                               struct page *page, sector_t sector,
2766                               size_t size, unsigned long offset,
2767                               struct block_device *bdev,
2768                               struct bio **bio_ret,
2769                               unsigned long max_pages,
2770                               bio_end_io_t end_io_func,
2771                               int mirror_num,
2772                               unsigned long prev_bio_flags,
2773                               unsigned long bio_flags,
2774                               bool force_bio_submit)
2775 {
2776         int ret = 0;
2777         struct bio *bio;
2778         int contig = 0;
2779         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2780         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2781
2782         if (bio_ret && *bio_ret) {
2783                 bio = *bio_ret;
2784                 if (old_compressed)
2785                         contig = bio->bi_iter.bi_sector == sector;
2786                 else
2787                         contig = bio_end_sector(bio) == sector;
2788
2789                 if (prev_bio_flags != bio_flags || !contig ||
2790                     force_bio_submit ||
2791                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2792                     bio_add_page(bio, page, page_size, offset) < page_size) {
2793                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2794                         if (ret < 0) {
2795                                 *bio_ret = NULL;
2796                                 return ret;
2797                         }
2798                         bio = NULL;
2799                 } else {
2800                         if (wbc)
2801                                 wbc_account_io(wbc, page, page_size);
2802                         return 0;
2803                 }
2804         }
2805
2806         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2807                         GFP_NOFS | __GFP_HIGH);
2808         if (!bio)
2809                 return -ENOMEM;
2810
2811         bio_add_page(bio, page, page_size, offset);
2812         bio->bi_end_io = end_io_func;
2813         bio->bi_private = tree;
2814         bio_set_op_attrs(bio, op, op_flags);
2815         if (wbc) {
2816                 wbc_init_bio(wbc, bio);
2817                 wbc_account_io(wbc, page, page_size);
2818         }
2819
2820         if (bio_ret)
2821                 *bio_ret = bio;
2822         else
2823                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2824
2825         return ret;
2826 }
2827
2828 static void attach_extent_buffer_page(struct extent_buffer *eb,
2829                                       struct page *page)
2830 {
2831         if (!PagePrivate(page)) {
2832                 SetPagePrivate(page);
2833                 get_page(page);
2834                 set_page_private(page, (unsigned long)eb);
2835         } else {
2836                 WARN_ON(page->private != (unsigned long)eb);
2837         }
2838 }
2839
2840 void set_page_extent_mapped(struct page *page)
2841 {
2842         if (!PagePrivate(page)) {
2843                 SetPagePrivate(page);
2844                 get_page(page);
2845                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2846         }
2847 }
2848
2849 static struct extent_map *
2850 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2851                  u64 start, u64 len, get_extent_t *get_extent,
2852                  struct extent_map **em_cached)
2853 {
2854         struct extent_map *em;
2855
2856         if (em_cached && *em_cached) {
2857                 em = *em_cached;
2858                 if (extent_map_in_tree(em) && start >= em->start &&
2859                     start < extent_map_end(em)) {
2860                         atomic_inc(&em->refs);
2861                         return em;
2862                 }
2863
2864                 free_extent_map(em);
2865                 *em_cached = NULL;
2866         }
2867
2868         em = get_extent(inode, page, pg_offset, start, len, 0);
2869         if (em_cached && !IS_ERR_OR_NULL(em)) {
2870                 BUG_ON(*em_cached);
2871                 atomic_inc(&em->refs);
2872                 *em_cached = em;
2873         }
2874         return em;
2875 }
2876 /*
2877  * basic readpage implementation.  Locked extent state structs are inserted
2878  * into the tree that are removed when the IO is done (by the end_io
2879  * handlers)
2880  * XXX JDM: This needs looking at to ensure proper page locking
2881  * return 0 on success, otherwise return error
2882  */
2883 static int __do_readpage(struct extent_io_tree *tree,
2884                          struct page *page,
2885                          get_extent_t *get_extent,
2886                          struct extent_map **em_cached,
2887                          struct bio **bio, int mirror_num,
2888                          unsigned long *bio_flags, int read_flags,
2889                          u64 *prev_em_start)
2890 {
2891         struct inode *inode = page->mapping->host;
2892         u64 start = page_offset(page);
2893         u64 page_end = start + PAGE_SIZE - 1;
2894         u64 end;
2895         u64 cur = start;
2896         u64 extent_offset;
2897         u64 last_byte = i_size_read(inode);
2898         u64 block_start;
2899         u64 cur_end;
2900         sector_t sector;
2901         struct extent_map *em;
2902         struct block_device *bdev;
2903         int ret = 0;
2904         int nr = 0;
2905         size_t pg_offset = 0;
2906         size_t iosize;
2907         size_t disk_io_size;
2908         size_t blocksize = inode->i_sb->s_blocksize;
2909         unsigned long this_bio_flag = 0;
2910
2911         set_page_extent_mapped(page);
2912
2913         end = page_end;
2914         if (!PageUptodate(page)) {
2915                 if (cleancache_get_page(page) == 0) {
2916                         BUG_ON(blocksize != PAGE_SIZE);
2917                         unlock_extent(tree, start, end);
2918                         goto out;
2919                 }
2920         }
2921
2922         if (page->index == last_byte >> PAGE_SHIFT) {
2923                 char *userpage;
2924                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2925
2926                 if (zero_offset) {
2927                         iosize = PAGE_SIZE - zero_offset;
2928                         userpage = kmap_atomic(page);
2929                         memset(userpage + zero_offset, 0, iosize);
2930                         flush_dcache_page(page);
2931                         kunmap_atomic(userpage);
2932                 }
2933         }
2934         while (cur <= end) {
2935                 unsigned long pnr = (last_byte >> PAGE_SHIFT) + 1;
2936                 bool force_bio_submit = false;
2937
2938                 if (cur >= last_byte) {
2939                         char *userpage;
2940                         struct extent_state *cached = NULL;
2941
2942                         iosize = PAGE_SIZE - pg_offset;
2943                         userpage = kmap_atomic(page);
2944                         memset(userpage + pg_offset, 0, iosize);
2945                         flush_dcache_page(page);
2946                         kunmap_atomic(userpage);
2947                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2948                                             &cached, GFP_NOFS);
2949                         unlock_extent_cached(tree, cur,
2950                                              cur + iosize - 1,
2951                                              &cached, GFP_NOFS);
2952                         break;
2953                 }
2954                 em = __get_extent_map(inode, page, pg_offset, cur,
2955                                       end - cur + 1, get_extent, em_cached);
2956                 if (IS_ERR_OR_NULL(em)) {
2957                         SetPageError(page);
2958                         unlock_extent(tree, cur, end);
2959                         break;
2960                 }
2961                 extent_offset = cur - em->start;
2962                 BUG_ON(extent_map_end(em) <= cur);
2963                 BUG_ON(end < cur);
2964
2965                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2966                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2967                         extent_set_compress_type(&this_bio_flag,
2968                                                  em->compress_type);
2969                 }
2970
2971                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2972                 cur_end = min(extent_map_end(em) - 1, end);
2973                 iosize = ALIGN(iosize, blocksize);
2974                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2975                         disk_io_size = em->block_len;
2976                         sector = em->block_start >> 9;
2977                 } else {
2978                         sector = (em->block_start + extent_offset) >> 9;
2979                         disk_io_size = iosize;
2980                 }
2981                 bdev = em->bdev;
2982                 block_start = em->block_start;
2983                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2984                         block_start = EXTENT_MAP_HOLE;
2985
2986                 /*
2987                  * If we have a file range that points to a compressed extent
2988                  * and it's followed by a consecutive file range that points to
2989                  * to the same compressed extent (possibly with a different
2990                  * offset and/or length, so it either points to the whole extent
2991                  * or only part of it), we must make sure we do not submit a
2992                  * single bio to populate the pages for the 2 ranges because
2993                  * this makes the compressed extent read zero out the pages
2994                  * belonging to the 2nd range. Imagine the following scenario:
2995                  *
2996                  *  File layout
2997                  *  [0 - 8K]                     [8K - 24K]
2998                  *    |                               |
2999                  *    |                               |
3000                  * points to extent X,         points to extent X,
3001                  * offset 4K, length of 8K     offset 0, length 16K
3002                  *
3003                  * [extent X, compressed length = 4K uncompressed length = 16K]
3004                  *
3005                  * If the bio to read the compressed extent covers both ranges,
3006                  * it will decompress extent X into the pages belonging to the
3007                  * first range and then it will stop, zeroing out the remaining
3008                  * pages that belong to the other range that points to extent X.
3009                  * So here we make sure we submit 2 bios, one for the first
3010                  * range and another one for the third range. Both will target
3011                  * the same physical extent from disk, but we can't currently
3012                  * make the compressed bio endio callback populate the pages
3013                  * for both ranges because each compressed bio is tightly
3014                  * coupled with a single extent map, and each range can have
3015                  * an extent map with a different offset value relative to the
3016                  * uncompressed data of our extent and different lengths. This
3017                  * is a corner case so we prioritize correctness over
3018                  * non-optimal behavior (submitting 2 bios for the same extent).
3019                  */
3020                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3021                     prev_em_start && *prev_em_start != (u64)-1 &&
3022                     *prev_em_start != em->orig_start)
3023                         force_bio_submit = true;
3024
3025                 if (prev_em_start)
3026                         *prev_em_start = em->orig_start;
3027
3028                 free_extent_map(em);
3029                 em = NULL;
3030
3031                 /* we've found a hole, just zero and go on */
3032                 if (block_start == EXTENT_MAP_HOLE) {
3033                         char *userpage;
3034                         struct extent_state *cached = NULL;
3035
3036                         userpage = kmap_atomic(page);
3037                         memset(userpage + pg_offset, 0, iosize);
3038                         flush_dcache_page(page);
3039                         kunmap_atomic(userpage);
3040
3041                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3042                                             &cached, GFP_NOFS);
3043                         unlock_extent_cached(tree, cur,
3044                                              cur + iosize - 1,
3045                                              &cached, GFP_NOFS);
3046                         cur = cur + iosize;
3047                         pg_offset += iosize;
3048                         continue;
3049                 }
3050                 /* the get_extent function already copied into the page */
3051                 if (test_range_bit(tree, cur, cur_end,
3052                                    EXTENT_UPTODATE, 1, NULL)) {
3053                         check_page_uptodate(tree, page);
3054                         unlock_extent(tree, cur, cur + iosize - 1);
3055                         cur = cur + iosize;
3056                         pg_offset += iosize;
3057                         continue;
3058                 }
3059                 /* we have an inline extent but it didn't get marked up
3060                  * to date.  Error out
3061                  */
3062                 if (block_start == EXTENT_MAP_INLINE) {
3063                         SetPageError(page);
3064                         unlock_extent(tree, cur, cur + iosize - 1);
3065                         cur = cur + iosize;
3066                         pg_offset += iosize;
3067                         continue;
3068                 }
3069
3070                 pnr -= page->index;
3071                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3072                                          page, sector, disk_io_size, pg_offset,
3073                                          bdev, bio, pnr,
3074                                          end_bio_extent_readpage, mirror_num,
3075                                          *bio_flags,
3076                                          this_bio_flag,
3077                                          force_bio_submit);
3078                 if (!ret) {
3079                         nr++;
3080                         *bio_flags = this_bio_flag;
3081                 } else {
3082                         SetPageError(page);
3083                         unlock_extent(tree, cur, cur + iosize - 1);
3084                         goto out;
3085                 }
3086                 cur = cur + iosize;
3087                 pg_offset += iosize;
3088         }
3089 out:
3090         if (!nr) {
3091                 if (!PageError(page))
3092                         SetPageUptodate(page);
3093                 unlock_page(page);
3094         }
3095         return ret;
3096 }
3097
3098 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3099                                              struct page *pages[], int nr_pages,
3100                                              u64 start, u64 end,
3101                                              get_extent_t *get_extent,
3102                                              struct extent_map **em_cached,
3103                                              struct bio **bio, int mirror_num,
3104                                              unsigned long *bio_flags,
3105                                              u64 *prev_em_start)
3106 {
3107         struct inode *inode;
3108         struct btrfs_ordered_extent *ordered;
3109         int index;
3110
3111         inode = pages[0]->mapping->host;
3112         while (1) {
3113                 lock_extent(tree, start, end);
3114                 ordered = btrfs_lookup_ordered_range(inode, start,
3115                                                      end - start + 1);
3116                 if (!ordered)
3117                         break;
3118                 unlock_extent(tree, start, end);
3119                 btrfs_start_ordered_extent(inode, ordered, 1);
3120                 btrfs_put_ordered_extent(ordered);
3121         }
3122
3123         for (index = 0; index < nr_pages; index++) {
3124                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3125                               mirror_num, bio_flags, 0, prev_em_start);
3126                 put_page(pages[index]);
3127         }