Merge branch 'for-4.9-rc3' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave...
[muen/linux.git] / fs / btrfs / extent_io.c
1 #include <linux/bitops.h>
2 #include <linux/slab.h>
3 #include <linux/bio.h>
4 #include <linux/mm.h>
5 #include <linux/pagemap.h>
6 #include <linux/page-flags.h>
7 #include <linux/spinlock.h>
8 #include <linux/blkdev.h>
9 #include <linux/swap.h>
10 #include <linux/writeback.h>
11 #include <linux/pagevec.h>
12 #include <linux/prefetch.h>
13 #include <linux/cleancache.h>
14 #include "extent_io.h"
15 #include "extent_map.h"
16 #include "ctree.h"
17 #include "btrfs_inode.h"
18 #include "volumes.h"
19 #include "check-integrity.h"
20 #include "locking.h"
21 #include "rcu-string.h"
22 #include "backref.h"
23 #include "transaction.h"
24
25 static struct kmem_cache *extent_state_cache;
26 static struct kmem_cache *extent_buffer_cache;
27 static struct bio_set *btrfs_bioset;
28
29 static inline bool extent_state_in_tree(const struct extent_state *state)
30 {
31         return !RB_EMPTY_NODE(&state->rb_node);
32 }
33
34 #ifdef CONFIG_BTRFS_DEBUG
35 static LIST_HEAD(buffers);
36 static LIST_HEAD(states);
37
38 static DEFINE_SPINLOCK(leak_lock);
39
40 static inline
41 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
42 {
43         unsigned long flags;
44
45         spin_lock_irqsave(&leak_lock, flags);
46         list_add(new, head);
47         spin_unlock_irqrestore(&leak_lock, flags);
48 }
49
50 static inline
51 void btrfs_leak_debug_del(struct list_head *entry)
52 {
53         unsigned long flags;
54
55         spin_lock_irqsave(&leak_lock, flags);
56         list_del(entry);
57         spin_unlock_irqrestore(&leak_lock, flags);
58 }
59
60 static inline
61 void btrfs_leak_debug_check(void)
62 {
63         struct extent_state *state;
64         struct extent_buffer *eb;
65
66         while (!list_empty(&states)) {
67                 state = list_entry(states.next, struct extent_state, leak_list);
68                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
69                        state->start, state->end, state->state,
70                        extent_state_in_tree(state),
71                        atomic_read(&state->refs));
72                 list_del(&state->leak_list);
73                 kmem_cache_free(extent_state_cache, state);
74         }
75
76         while (!list_empty(&buffers)) {
77                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
78                 pr_err("BTRFS: buffer leak start %llu len %lu 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                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
464                        found->start, found->end, start, end);
465                 return -EEXIST;
466         }
467         merge_state(tree, state);
468         return 0;
469 }
470
471 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
472                      u64 split)
473 {
474         if (tree->ops && tree->ops->split_extent_hook)
475                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
476 }
477
478 /*
479  * split a given extent state struct in two, inserting the preallocated
480  * struct 'prealloc' as the newly created second half.  'split' indicates an
481  * offset inside 'orig' where it should be split.
482  *
483  * Before calling,
484  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
485  * are two extent state structs in the tree:
486  * prealloc: [orig->start, split - 1]
487  * orig: [ split, orig->end ]
488  *
489  * The tree locks are not taken by this function. They need to be held
490  * by the caller.
491  */
492 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
493                        struct extent_state *prealloc, u64 split)
494 {
495         struct rb_node *node;
496
497         split_cb(tree, orig, split);
498
499         prealloc->start = orig->start;
500         prealloc->end = split - 1;
501         prealloc->state = orig->state;
502         orig->start = split;
503
504         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
505                            &prealloc->rb_node, NULL, NULL);
506         if (node) {
507                 free_extent_state(prealloc);
508                 return -EEXIST;
509         }
510         return 0;
511 }
512
513 static struct extent_state *next_state(struct extent_state *state)
514 {
515         struct rb_node *next = rb_next(&state->rb_node);
516         if (next)
517                 return rb_entry(next, struct extent_state, rb_node);
518         else
519                 return NULL;
520 }
521
522 /*
523  * utility function to clear some bits in an extent state struct.
524  * it will optionally wake up any one waiting on this state (wake == 1).
525  *
526  * If no bits are set on the state struct after clearing things, the
527  * struct is freed and removed from the tree
528  */
529 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
530                                             struct extent_state *state,
531                                             unsigned *bits, int wake,
532                                             struct extent_changeset *changeset)
533 {
534         struct extent_state *next;
535         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
536
537         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
538                 u64 range = state->end - state->start + 1;
539                 WARN_ON(range > tree->dirty_bytes);
540                 tree->dirty_bytes -= range;
541         }
542         clear_state_cb(tree, state, bits);
543         add_extent_changeset(state, bits_to_clear, changeset, 0);
544         state->state &= ~bits_to_clear;
545         if (wake)
546                 wake_up(&state->wq);
547         if (state->state == 0) {
548                 next = next_state(state);
549                 if (extent_state_in_tree(state)) {
550                         rb_erase(&state->rb_node, &tree->state);
551                         RB_CLEAR_NODE(&state->rb_node);
552                         free_extent_state(state);
553                 } else {
554                         WARN_ON(1);
555                 }
556         } else {
557                 merge_state(tree, state);
558                 next = next_state(state);
559         }
560         return next;
561 }
562
563 static struct extent_state *
564 alloc_extent_state_atomic(struct extent_state *prealloc)
565 {
566         if (!prealloc)
567                 prealloc = alloc_extent_state(GFP_ATOMIC);
568
569         return prealloc;
570 }
571
572 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
573 {
574         btrfs_panic(tree_fs_info(tree), err,
575                     "Locking error: Extent tree was modified by another thread while locked.");
576 }
577
578 /*
579  * clear some bits on a range in the tree.  This may require splitting
580  * or inserting elements in the tree, so the gfp mask is used to
581  * indicate which allocations or sleeping are allowed.
582  *
583  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
584  * the given range from the tree regardless of state (ie for truncate).
585  *
586  * the range [start, end] is inclusive.
587  *
588  * This takes the tree lock, and returns 0 on success and < 0 on error.
589  */
590 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
591                               unsigned bits, int wake, int delete,
592                               struct extent_state **cached_state,
593                               gfp_t mask, struct extent_changeset *changeset)
594 {
595         struct extent_state *state;
596         struct extent_state *cached;
597         struct extent_state *prealloc = NULL;
598         struct rb_node *node;
599         u64 last_end;
600         int err;
601         int clear = 0;
602
603         btrfs_debug_check_extent_io_range(tree, start, end);
604
605         if (bits & EXTENT_DELALLOC)
606                 bits |= EXTENT_NORESERVE;
607
608         if (delete)
609                 bits |= ~EXTENT_CTLBITS;
610         bits |= EXTENT_FIRST_DELALLOC;
611
612         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
613                 clear = 1;
614 again:
615         if (!prealloc && gfpflags_allow_blocking(mask)) {
616                 /*
617                  * Don't care for allocation failure here because we might end
618                  * up not needing the pre-allocated extent state at all, which
619                  * is the case if we only have in the tree extent states that
620                  * cover our input range and don't cover too any other range.
621                  * If we end up needing a new extent state we allocate it later.
622                  */
623                 prealloc = alloc_extent_state(mask);
624         }
625
626         spin_lock(&tree->lock);
627         if (cached_state) {
628                 cached = *cached_state;
629
630                 if (clear) {
631                         *cached_state = NULL;
632                         cached_state = NULL;
633                 }
634
635                 if (cached && extent_state_in_tree(cached) &&
636                     cached->start <= start && cached->end > start) {
637                         if (clear)
638                                 atomic_dec(&cached->refs);
639                         state = cached;
640                         goto hit_next;
641                 }
642                 if (clear)
643                         free_extent_state(cached);
644         }
645         /*
646          * this search will find the extents that end after
647          * our range starts
648          */
649         node = tree_search(tree, start);
650         if (!node)
651                 goto out;
652         state = rb_entry(node, struct extent_state, rb_node);
653 hit_next:
654         if (state->start > end)
655                 goto out;
656         WARN_ON(state->end < start);
657         last_end = state->end;
658
659         /* the state doesn't have the wanted bits, go ahead */
660         if (!(state->state & bits)) {
661                 state = next_state(state);
662                 goto next;
663         }
664
665         /*
666          *     | ---- desired range ---- |
667          *  | state | or
668          *  | ------------- state -------------- |
669          *
670          * We need to split the extent we found, and may flip
671          * bits on second half.
672          *
673          * If the extent we found extends past our range, we
674          * just split and search again.  It'll get split again
675          * the next time though.
676          *
677          * If the extent we found is inside our range, we clear
678          * the desired bit on it.
679          */
680
681         if (state->start < start) {
682                 prealloc = alloc_extent_state_atomic(prealloc);
683                 BUG_ON(!prealloc);
684                 err = split_state(tree, state, prealloc, start);
685                 if (err)
686                         extent_io_tree_panic(tree, err);
687
688                 prealloc = NULL;
689                 if (err)
690                         goto out;
691                 if (state->end <= end) {
692                         state = clear_state_bit(tree, state, &bits, wake,
693                                                 changeset);
694                         goto next;
695                 }
696                 goto search_again;
697         }
698         /*
699          * | ---- desired range ---- |
700          *                        | state |
701          * We need to split the extent, and clear the bit
702          * on the first half
703          */
704         if (state->start <= end && state->end > end) {
705                 prealloc = alloc_extent_state_atomic(prealloc);
706                 BUG_ON(!prealloc);
707                 err = split_state(tree, state, prealloc, end + 1);
708                 if (err)
709                         extent_io_tree_panic(tree, err);
710
711                 if (wake)
712                         wake_up(&state->wq);
713
714                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
715
716                 prealloc = NULL;
717                 goto out;
718         }
719
720         state = clear_state_bit(tree, state, &bits, wake, changeset);
721 next:
722         if (last_end == (u64)-1)
723                 goto out;
724         start = last_end + 1;
725         if (start <= end && state && !need_resched())
726                 goto hit_next;
727
728 search_again:
729         if (start > end)
730                 goto out;
731         spin_unlock(&tree->lock);
732         if (gfpflags_allow_blocking(mask))
733                 cond_resched();
734         goto again;
735
736 out:
737         spin_unlock(&tree->lock);
738         if (prealloc)
739                 free_extent_state(prealloc);
740
741         return 0;
742
743 }
744
745 static void wait_on_state(struct extent_io_tree *tree,
746                           struct extent_state *state)
747                 __releases(tree->lock)
748                 __acquires(tree->lock)
749 {
750         DEFINE_WAIT(wait);
751         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
752         spin_unlock(&tree->lock);
753         schedule();
754         spin_lock(&tree->lock);
755         finish_wait(&state->wq, &wait);
756 }
757
758 /*
759  * waits for one or more bits to clear on a range in the state tree.
760  * The range [start, end] is inclusive.
761  * The tree lock is taken by this function
762  */
763 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
764                             unsigned long bits)
765 {
766         struct extent_state *state;
767         struct rb_node *node;
768
769         btrfs_debug_check_extent_io_range(tree, start, end);
770
771         spin_lock(&tree->lock);
772 again:
773         while (1) {
774                 /*
775                  * this search will find all the extents that end after
776                  * our range starts
777                  */
778                 node = tree_search(tree, start);
779 process_node:
780                 if (!node)
781                         break;
782
783                 state = rb_entry(node, struct extent_state, rb_node);
784
785                 if (state->start > end)
786                         goto out;
787
788                 if (state->state & bits) {
789                         start = state->start;
790                         atomic_inc(&state->refs);
791                         wait_on_state(tree, state);
792                         free_extent_state(state);
793                         goto again;
794                 }
795                 start = state->end + 1;
796
797                 if (start > end)
798                         break;
799
800                 if (!cond_resched_lock(&tree->lock)) {
801                         node = rb_next(node);
802                         goto process_node;
803                 }
804         }
805 out:
806         spin_unlock(&tree->lock);
807 }
808
809 static void set_state_bits(struct extent_io_tree *tree,
810                            struct extent_state *state,
811                            unsigned *bits, struct extent_changeset *changeset)
812 {
813         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
814
815         set_state_cb(tree, state, bits);
816         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
817                 u64 range = state->end - state->start + 1;
818                 tree->dirty_bytes += range;
819         }
820         add_extent_changeset(state, bits_to_set, changeset, 1);
821         state->state |= bits_to_set;
822 }
823
824 static void cache_state_if_flags(struct extent_state *state,
825                                  struct extent_state **cached_ptr,
826                                  unsigned flags)
827 {
828         if (cached_ptr && !(*cached_ptr)) {
829                 if (!flags || (state->state & flags)) {
830                         *cached_ptr = state;
831                         atomic_inc(&state->refs);
832                 }
833         }
834 }
835
836 static void cache_state(struct extent_state *state,
837                         struct extent_state **cached_ptr)
838 {
839         return cache_state_if_flags(state, cached_ptr,
840                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
841 }
842
843 /*
844  * set some bits on a range in the tree.  This may require allocations or
845  * sleeping, so the gfp mask is used to indicate what is allowed.
846  *
847  * If any of the exclusive bits are set, this will fail with -EEXIST if some
848  * part of the range already has the desired bits set.  The start of the
849  * existing range is returned in failed_start in this case.
850  *
851  * [start, end] is inclusive This takes the tree lock.
852  */
853
854 static int __must_check
855 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
856                  unsigned bits, unsigned exclusive_bits,
857                  u64 *failed_start, struct extent_state **cached_state,
858                  gfp_t mask, struct extent_changeset *changeset)
859 {
860         struct extent_state *state;
861         struct extent_state *prealloc = NULL;
862         struct rb_node *node;
863         struct rb_node **p;
864         struct rb_node *parent;
865         int err = 0;
866         u64 last_start;
867         u64 last_end;
868
869         btrfs_debug_check_extent_io_range(tree, start, end);
870
871         bits |= EXTENT_FIRST_DELALLOC;
872 again:
873         if (!prealloc && gfpflags_allow_blocking(mask)) {
874                 /*
875                  * Don't care for allocation failure here because we might end
876                  * up not needing the pre-allocated extent state at all, which
877                  * is the case if we only have in the tree extent states that
878                  * cover our input range and don't cover too any other range.
879                  * If we end up needing a new extent state we allocate it later.
880                  */
881                 prealloc = alloc_extent_state(mask);
882         }
883
884         spin_lock(&tree->lock);
885         if (cached_state && *cached_state) {
886                 state = *cached_state;
887                 if (state->start <= start && state->end > start &&
888                     extent_state_in_tree(state)) {
889                         node = &state->rb_node;
890                         goto hit_next;
891                 }
892         }
893         /*
894          * this search will find all the extents that end after
895          * our range starts.
896          */
897         node = tree_search_for_insert(tree, start, &p, &parent);
898         if (!node) {
899                 prealloc = alloc_extent_state_atomic(prealloc);
900                 BUG_ON(!prealloc);
901                 err = insert_state(tree, prealloc, start, end,
902                                    &p, &parent, &bits, changeset);
903                 if (err)
904                         extent_io_tree_panic(tree, err);
905
906                 cache_state(prealloc, cached_state);
907                 prealloc = NULL;
908                 goto out;
909         }
910         state = rb_entry(node, struct extent_state, rb_node);
911 hit_next:
912         last_start = state->start;
913         last_end = state->end;
914
915         /*
916          * | ---- desired range ---- |
917          * | state |
918          *
919          * Just lock what we found and keep going
920          */
921         if (state->start == start && state->end <= end) {
922                 if (state->state & exclusive_bits) {
923                         *failed_start = state->start;
924                         err = -EEXIST;
925                         goto out;
926                 }
927
928                 set_state_bits(tree, state, &bits, changeset);
929                 cache_state(state, cached_state);
930                 merge_state(tree, state);
931                 if (last_end == (u64)-1)
932                         goto out;
933                 start = last_end + 1;
934                 state = next_state(state);
935                 if (start < end && state && state->start == start &&
936                     !need_resched())
937                         goto hit_next;
938                 goto search_again;
939         }
940
941         /*
942          *     | ---- desired range ---- |
943          * | state |
944          *   or
945          * | ------------- state -------------- |
946          *
947          * We need to split the extent we found, and may flip bits on
948          * second half.
949          *
950          * If the extent we found extends past our
951          * range, we just split and search again.  It'll get split
952          * again the next time though.
953          *
954          * If the extent we found is inside our range, we set the
955          * desired bit on it.
956          */
957         if (state->start < start) {
958                 if (state->state & exclusive_bits) {
959                         *failed_start = start;
960                         err = -EEXIST;
961                         goto out;
962                 }
963
964                 prealloc = alloc_extent_state_atomic(prealloc);
965                 BUG_ON(!prealloc);
966                 err = split_state(tree, state, prealloc, start);
967                 if (err)
968                         extent_io_tree_panic(tree, err);
969
970                 prealloc = NULL;
971                 if (err)
972                         goto out;
973                 if (state->end <= end) {
974                         set_state_bits(tree, state, &bits, changeset);
975                         cache_state(state, cached_state);
976                         merge_state(tree, state);
977                         if (last_end == (u64)-1)
978                                 goto out;
979                         start = last_end + 1;
980                         state = next_state(state);
981                         if (start < end && state && state->start == start &&
982                             !need_resched())
983                                 goto hit_next;
984                 }
985                 goto search_again;
986         }
987         /*
988          * | ---- desired range ---- |
989          *     | state | or               | state |
990          *
991          * There's a hole, we need to insert something in it and
992          * ignore the extent we found.
993          */
994         if (state->start > start) {
995                 u64 this_end;
996                 if (end < last_start)
997                         this_end = end;
998                 else
999                         this_end = last_start - 1;
1000
1001                 prealloc = alloc_extent_state_atomic(prealloc);
1002                 BUG_ON(!prealloc);
1003
1004                 /*
1005                  * Avoid to free 'prealloc' if it can be merged with
1006                  * the later extent.
1007                  */
1008                 err = insert_state(tree, prealloc, start, this_end,
1009                                    NULL, NULL, &bits, changeset);
1010                 if (err)
1011                         extent_io_tree_panic(tree, err);
1012
1013                 cache_state(prealloc, cached_state);
1014                 prealloc = NULL;
1015                 start = this_end + 1;
1016                 goto search_again;
1017         }
1018         /*
1019          * | ---- desired range ---- |
1020          *                        | state |
1021          * We need to split the extent, and set the bit
1022          * on the first half
1023          */
1024         if (state->start <= end && state->end > end) {
1025                 if (state->state & exclusive_bits) {
1026                         *failed_start = start;
1027                         err = -EEXIST;
1028                         goto out;
1029                 }
1030
1031                 prealloc = alloc_extent_state_atomic(prealloc);
1032                 BUG_ON(!prealloc);
1033                 err = split_state(tree, state, prealloc, end + 1);
1034                 if (err)
1035                         extent_io_tree_panic(tree, err);
1036
1037                 set_state_bits(tree, prealloc, &bits, changeset);
1038                 cache_state(prealloc, cached_state);
1039                 merge_state(tree, prealloc);
1040                 prealloc = NULL;
1041                 goto out;
1042         }
1043
1044 search_again:
1045         if (start > end)
1046                 goto out;
1047         spin_unlock(&tree->lock);
1048         if (gfpflags_allow_blocking(mask))
1049                 cond_resched();
1050         goto again;
1051
1052 out:
1053         spin_unlock(&tree->lock);
1054         if (prealloc)
1055                 free_extent_state(prealloc);
1056
1057         return err;
1058
1059 }
1060
1061 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1062                    unsigned bits, u64 * failed_start,
1063                    struct extent_state **cached_state, gfp_t mask)
1064 {
1065         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1066                                 cached_state, mask, NULL);
1067 }
1068
1069
1070 /**
1071  * convert_extent_bit - convert all bits in a given range from one bit to
1072  *                      another
1073  * @tree:       the io tree to search
1074  * @start:      the start offset in bytes
1075  * @end:        the end offset in bytes (inclusive)
1076  * @bits:       the bits to set in this range
1077  * @clear_bits: the bits to clear in this range
1078  * @cached_state:       state that we're going to cache
1079  *
1080  * This will go through and set bits for the given range.  If any states exist
1081  * already in this range they are set with the given bit and cleared of the
1082  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1083  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1084  * boundary bits like LOCK.
1085  *
1086  * All allocations are done with GFP_NOFS.
1087  */
1088 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1089                        unsigned bits, unsigned clear_bits,
1090                        struct extent_state **cached_state)
1091 {
1092         struct extent_state *state;
1093         struct extent_state *prealloc = NULL;
1094         struct rb_node *node;
1095         struct rb_node **p;
1096         struct rb_node *parent;
1097         int err = 0;
1098         u64 last_start;
1099         u64 last_end;
1100         bool first_iteration = true;
1101
1102         btrfs_debug_check_extent_io_range(tree, start, end);
1103
1104 again:
1105         if (!prealloc) {
1106                 /*
1107                  * Best effort, don't worry if extent state allocation fails
1108                  * here for the first iteration. We might have a cached state
1109                  * that matches exactly the target range, in which case no
1110                  * extent state allocations are needed. We'll only know this
1111                  * after locking the tree.
1112                  */
1113                 prealloc = alloc_extent_state(GFP_NOFS);
1114                 if (!prealloc && !first_iteration)
1115                         return -ENOMEM;
1116         }
1117
1118         spin_lock(&tree->lock);
1119         if (cached_state && *cached_state) {
1120                 state = *cached_state;
1121                 if (state->start <= start && state->end > start &&
1122                     extent_state_in_tree(state)) {
1123                         node = &state->rb_node;
1124                         goto hit_next;
1125                 }
1126         }
1127
1128         /*
1129          * this search will find all the extents that end after
1130          * our range starts.
1131          */
1132         node = tree_search_for_insert(tree, start, &p, &parent);
1133         if (!node) {
1134                 prealloc = alloc_extent_state_atomic(prealloc);
1135                 if (!prealloc) {
1136                         err = -ENOMEM;
1137                         goto out;
1138                 }
1139                 err = insert_state(tree, prealloc, start, end,
1140                                    &p, &parent, &bits, NULL);
1141                 if (err)
1142                         extent_io_tree_panic(tree, err);
1143                 cache_state(prealloc, cached_state);
1144                 prealloc = NULL;
1145                 goto out;
1146         }
1147         state = rb_entry(node, struct extent_state, rb_node);
1148 hit_next:
1149         last_start = state->start;
1150         last_end = state->end;
1151
1152         /*
1153          * | ---- desired range ---- |
1154          * | state |
1155          *
1156          * Just lock what we found and keep going
1157          */
1158         if (state->start == start && state->end <= end) {
1159                 set_state_bits(tree, state, &bits, NULL);
1160                 cache_state(state, cached_state);
1161                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1162                 if (last_end == (u64)-1)
1163                         goto out;
1164                 start = last_end + 1;
1165                 if (start < end && state && state->start == start &&
1166                     !need_resched())
1167                         goto hit_next;
1168                 goto search_again;
1169         }
1170
1171         /*
1172          *     | ---- desired range ---- |
1173          * | state |
1174          *   or
1175          * | ------------- state -------------- |
1176          *
1177          * We need to split the extent we found, and may flip bits on
1178          * second half.
1179          *
1180          * If the extent we found extends past our
1181          * range, we just split and search again.  It'll get split
1182          * again the next time though.
1183          *
1184          * If the extent we found is inside our range, we set the
1185          * desired bit on it.
1186          */
1187         if (state->start < start) {
1188                 prealloc = alloc_extent_state_atomic(prealloc);
1189                 if (!prealloc) {
1190                         err = -ENOMEM;
1191                         goto out;
1192                 }
1193                 err = split_state(tree, state, prealloc, start);
1194                 if (err)
1195                         extent_io_tree_panic(tree, err);
1196                 prealloc = NULL;
1197                 if (err)
1198                         goto out;
1199                 if (state->end <= end) {
1200                         set_state_bits(tree, state, &bits, NULL);
1201                         cache_state(state, cached_state);
1202                         state = clear_state_bit(tree, state, &clear_bits, 0,
1203                                                 NULL);
1204                         if (last_end == (u64)-1)
1205                                 goto out;
1206                         start = last_end + 1;
1207                         if (start < end && state && state->start == start &&
1208                             !need_resched())
1209                                 goto hit_next;
1210                 }
1211                 goto search_again;
1212         }
1213         /*
1214          * | ---- desired range ---- |
1215          *     | state | or               | state |
1216          *
1217          * There's a hole, we need to insert something in it and
1218          * ignore the extent we found.
1219          */
1220         if (state->start > start) {
1221                 u64 this_end;
1222                 if (end < last_start)
1223                         this_end = end;
1224                 else
1225                         this_end = last_start - 1;
1226
1227                 prealloc = alloc_extent_state_atomic(prealloc);
1228                 if (!prealloc) {
1229                         err = -ENOMEM;
1230                         goto out;
1231                 }
1232
1233                 /*
1234                  * Avoid to free 'prealloc' if it can be merged with
1235                  * the later extent.
1236                  */
1237                 err = insert_state(tree, prealloc, start, this_end,
1238                                    NULL, NULL, &bits, NULL);
1239                 if (err)
1240                         extent_io_tree_panic(tree, err);
1241                 cache_state(prealloc, cached_state);
1242                 prealloc = NULL;
1243                 start = this_end + 1;
1244                 goto search_again;
1245         }
1246         /*
1247          * | ---- desired range ---- |
1248          *                        | state |
1249          * We need to split the extent, and set the bit
1250          * on the first half
1251          */
1252         if (state->start <= end && state->end > end) {
1253                 prealloc = alloc_extent_state_atomic(prealloc);
1254                 if (!prealloc) {
1255                         err = -ENOMEM;
1256                         goto out;
1257                 }
1258
1259                 err = split_state(tree, state, prealloc, end + 1);
1260                 if (err)
1261                         extent_io_tree_panic(tree, err);
1262
1263                 set_state_bits(tree, prealloc, &bits, NULL);
1264                 cache_state(prealloc, cached_state);
1265                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1266                 prealloc = NULL;
1267                 goto out;
1268         }
1269
1270 search_again:
1271         if (start > end)
1272                 goto out;
1273         spin_unlock(&tree->lock);
1274         cond_resched();
1275         first_iteration = false;
1276         goto again;
1277
1278 out:
1279         spin_unlock(&tree->lock);
1280         if (prealloc)
1281                 free_extent_state(prealloc);
1282
1283         return err;
1284 }
1285
1286 /* wrappers around set/clear extent bit */
1287 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1288                            unsigned bits, struct extent_changeset *changeset)
1289 {
1290         /*
1291          * We don't support EXTENT_LOCKED yet, as current changeset will
1292          * record any bits changed, so for EXTENT_LOCKED case, it will
1293          * either fail with -EEXIST or changeset will record the whole
1294          * range.
1295          */
1296         BUG_ON(bits & EXTENT_LOCKED);
1297
1298         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1299                                 changeset);
1300 }
1301
1302 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1303                      unsigned bits, int wake, int delete,
1304                      struct extent_state **cached, gfp_t mask)
1305 {
1306         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1307                                   cached, mask, NULL);
1308 }
1309
1310 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1311                 unsigned bits, struct extent_changeset *changeset)
1312 {
1313         /*
1314          * Don't support EXTENT_LOCKED case, same reason as
1315          * set_record_extent_bits().
1316          */
1317         BUG_ON(bits & EXTENT_LOCKED);
1318
1319         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1320                                   changeset);
1321 }
1322
1323 /*
1324  * either insert or lock state struct between start and end use mask to tell
1325  * us if waiting is desired.
1326  */
1327 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1328                      struct extent_state **cached_state)
1329 {
1330         int err;
1331         u64 failed_start;
1332
1333         while (1) {
1334                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1335                                        EXTENT_LOCKED, &failed_start,
1336                                        cached_state, GFP_NOFS, NULL);
1337                 if (err == -EEXIST) {
1338                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1339                         start = failed_start;
1340                 } else
1341                         break;
1342                 WARN_ON(start > end);
1343         }
1344         return err;
1345 }
1346
1347 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1348 {
1349         int err;
1350         u64 failed_start;
1351
1352         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1353                                &failed_start, NULL, GFP_NOFS, NULL);
1354         if (err == -EEXIST) {
1355                 if (failed_start > start)
1356                         clear_extent_bit(tree, start, failed_start - 1,
1357                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1358                 return 0;
1359         }
1360         return 1;
1361 }
1362
1363 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1364 {
1365         unsigned long index = start >> PAGE_SHIFT;
1366         unsigned long end_index = end >> PAGE_SHIFT;
1367         struct page *page;
1368
1369         while (index <= end_index) {
1370                 page = find_get_page(inode->i_mapping, index);
1371                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1372                 clear_page_dirty_for_io(page);
1373                 put_page(page);
1374                 index++;
1375         }
1376 }
1377
1378 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1379 {
1380         unsigned long index = start >> PAGE_SHIFT;
1381         unsigned long end_index = end >> PAGE_SHIFT;
1382         struct page *page;
1383
1384         while (index <= end_index) {
1385                 page = find_get_page(inode->i_mapping, index);
1386                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1387                 __set_page_dirty_nobuffers(page);
1388                 account_page_redirty(page);
1389                 put_page(page);
1390                 index++;
1391         }
1392 }
1393
1394 /*
1395  * helper function to set both pages and extents in the tree writeback
1396  */
1397 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1398 {
1399         unsigned long index = start >> PAGE_SHIFT;
1400         unsigned long end_index = end >> PAGE_SHIFT;
1401         struct page *page;
1402
1403         while (index <= end_index) {
1404                 page = find_get_page(tree->mapping, index);
1405                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1406                 set_page_writeback(page);
1407                 put_page(page);
1408                 index++;
1409         }
1410 }
1411
1412 /* find the first state struct with 'bits' set after 'start', and
1413  * return it.  tree->lock must be held.  NULL will returned if
1414  * nothing was found after 'start'
1415  */
1416 static struct extent_state *
1417 find_first_extent_bit_state(struct extent_io_tree *tree,
1418                             u64 start, unsigned bits)
1419 {
1420         struct rb_node *node;
1421         struct extent_state *state;
1422
1423         /*
1424          * this search will find all the extents that end after
1425          * our range starts.
1426          */
1427         node = tree_search(tree, start);
1428         if (!node)
1429                 goto out;
1430
1431         while (1) {
1432                 state = rb_entry(node, struct extent_state, rb_node);
1433                 if (state->end >= start && (state->state & bits))
1434                         return state;
1435
1436                 node = rb_next(node);
1437                 if (!node)
1438                         break;
1439         }
1440 out:
1441         return NULL;
1442 }
1443
1444 /*
1445  * find the first offset in the io tree with 'bits' set. zero is
1446  * returned if we find something, and *start_ret and *end_ret are
1447  * set to reflect the state struct that was found.
1448  *
1449  * If nothing was found, 1 is returned. If found something, return 0.
1450  */
1451 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1452                           u64 *start_ret, u64 *end_ret, unsigned bits,
1453                           struct extent_state **cached_state)
1454 {
1455         struct extent_state *state;
1456         struct rb_node *n;
1457         int ret = 1;
1458
1459         spin_lock(&tree->lock);
1460         if (cached_state && *cached_state) {
1461                 state = *cached_state;
1462                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1463                         n = rb_next(&state->rb_node);
1464                         while (n) {
1465                                 state = rb_entry(n, struct extent_state,
1466                                                  rb_node);
1467                                 if (state->state & bits)
1468                                         goto got_it;
1469                                 n = rb_next(n);
1470                         }
1471                         free_extent_state(*cached_state);
1472                         *cached_state = NULL;
1473                         goto out;
1474                 }
1475                 free_extent_state(*cached_state);
1476                 *cached_state = NULL;
1477         }
1478
1479         state = find_first_extent_bit_state(tree, start, bits);
1480 got_it:
1481         if (state) {
1482                 cache_state_if_flags(state, cached_state, 0);
1483                 *start_ret = state->start;
1484                 *end_ret = state->end;
1485                 ret = 0;
1486         }
1487 out:
1488         spin_unlock(&tree->lock);
1489         return ret;
1490 }
1491
1492 /*
1493  * find a contiguous range of bytes in the file marked as delalloc, not
1494  * more than 'max_bytes'.  start and end are used to return the range,
1495  *
1496  * 1 is returned if we find something, 0 if nothing was in the tree
1497  */
1498 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1499                                         u64 *start, u64 *end, u64 max_bytes,
1500                                         struct extent_state **cached_state)
1501 {
1502         struct rb_node *node;
1503         struct extent_state *state;
1504         u64 cur_start = *start;
1505         u64 found = 0;
1506         u64 total_bytes = 0;
1507
1508         spin_lock(&tree->lock);
1509
1510         /*
1511          * this search will find all the extents that end after
1512          * our range starts.
1513          */
1514         node = tree_search(tree, cur_start);
1515         if (!node) {
1516                 if (!found)
1517                         *end = (u64)-1;
1518                 goto out;
1519         }
1520
1521         while (1) {
1522                 state = rb_entry(node, struct extent_state, rb_node);
1523                 if (found && (state->start != cur_start ||
1524                               (state->state & EXTENT_BOUNDARY))) {
1525                         goto out;
1526                 }
1527                 if (!(state->state & EXTENT_DELALLOC)) {
1528                         if (!found)
1529                                 *end = state->end;
1530                         goto out;
1531                 }
1532                 if (!found) {
1533                         *start = state->start;
1534                         *cached_state = state;
1535                         atomic_inc(&state->refs);
1536                 }
1537                 found++;
1538                 *end = state->end;
1539                 cur_start = state->end + 1;
1540                 node = rb_next(node);
1541                 total_bytes += state->end - state->start + 1;
1542                 if (total_bytes >= max_bytes)
1543                         break;
1544                 if (!node)
1545                         break;
1546         }
1547 out:
1548         spin_unlock(&tree->lock);
1549         return found;
1550 }
1551
1552 static noinline void __unlock_for_delalloc(struct inode *inode,
1553                                            struct page *locked_page,
1554                                            u64 start, u64 end)
1555 {
1556         int ret;
1557         struct page *pages[16];
1558         unsigned long index = start >> PAGE_SHIFT;
1559         unsigned long end_index = end >> PAGE_SHIFT;
1560         unsigned long nr_pages = end_index - index + 1;
1561         int i;
1562
1563         if (index == locked_page->index && end_index == index)
1564                 return;
1565
1566         while (nr_pages > 0) {
1567                 ret = find_get_pages_contig(inode->i_mapping, index,
1568                                      min_t(unsigned long, nr_pages,
1569                                      ARRAY_SIZE(pages)), pages);
1570                 for (i = 0; i < ret; i++) {
1571                         if (pages[i] != locked_page)
1572                                 unlock_page(pages[i]);
1573                         put_page(pages[i]);
1574                 }
1575                 nr_pages -= ret;
1576                 index += ret;
1577                 cond_resched();
1578         }
1579 }
1580
1581 static noinline int lock_delalloc_pages(struct inode *inode,
1582                                         struct page *locked_page,
1583                                         u64 delalloc_start,
1584                                         u64 delalloc_end)
1585 {
1586         unsigned long index = delalloc_start >> PAGE_SHIFT;
1587         unsigned long start_index = index;
1588         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1589         unsigned long pages_locked = 0;
1590         struct page *pages[16];
1591         unsigned long nrpages;
1592         int ret;
1593         int i;
1594
1595         /* the caller is responsible for locking the start index */
1596         if (index == locked_page->index && index == end_index)
1597                 return 0;
1598
1599         /* skip the page at the start index */
1600         nrpages = end_index - index + 1;
1601         while (nrpages > 0) {
1602                 ret = find_get_pages_contig(inode->i_mapping, index,
1603                                      min_t(unsigned long,
1604                                      nrpages, ARRAY_SIZE(pages)), pages);
1605                 if (ret == 0) {
1606                         ret = -EAGAIN;
1607                         goto done;
1608                 }
1609                 /* now we have an array of pages, lock them all */
1610                 for (i = 0; i < ret; i++) {
1611                         /*
1612                          * the caller is taking responsibility for
1613                          * locked_page
1614                          */
1615                         if (pages[i] != locked_page) {
1616                                 lock_page(pages[i]);
1617                                 if (!PageDirty(pages[i]) ||
1618                                     pages[i]->mapping != inode->i_mapping) {
1619                                         ret = -EAGAIN;
1620                                         unlock_page(pages[i]);
1621                                         put_page(pages[i]);
1622                                         goto done;
1623                                 }
1624                         }
1625                         put_page(pages[i]);
1626                         pages_locked++;
1627                 }
1628                 nrpages -= ret;
1629                 index += ret;
1630                 cond_resched();
1631         }
1632         ret = 0;
1633 done:
1634         if (ret && pages_locked) {
1635                 __unlock_for_delalloc(inode, locked_page,
1636                               delalloc_start,
1637                               ((u64)(start_index + pages_locked - 1)) <<
1638                               PAGE_SHIFT);
1639         }
1640         return ret;
1641 }
1642
1643 /*
1644  * find a contiguous range of bytes in the file marked as delalloc, not
1645  * more than 'max_bytes'.  start and end are used to return the range,
1646  *
1647  * 1 is returned if we find something, 0 if nothing was in the tree
1648  */
1649 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1650                                     struct extent_io_tree *tree,
1651                                     struct page *locked_page, u64 *start,
1652                                     u64 *end, u64 max_bytes)
1653 {
1654         u64 delalloc_start;
1655         u64 delalloc_end;
1656         u64 found;
1657         struct extent_state *cached_state = NULL;
1658         int ret;
1659         int loops = 0;
1660
1661 again:
1662         /* step one, find a bunch of delalloc bytes starting at start */
1663         delalloc_start = *start;
1664         delalloc_end = 0;
1665         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1666                                     max_bytes, &cached_state);
1667         if (!found || delalloc_end <= *start) {
1668                 *start = delalloc_start;
1669                 *end = delalloc_end;
1670                 free_extent_state(cached_state);
1671                 return 0;
1672         }
1673
1674         /*
1675          * start comes from the offset of locked_page.  We have to lock
1676          * pages in order, so we can't process delalloc bytes before
1677          * locked_page
1678          */
1679         if (delalloc_start < *start)
1680                 delalloc_start = *start;
1681
1682         /*
1683          * make sure to limit the number of pages we try to lock down
1684          */
1685         if (delalloc_end + 1 - delalloc_start > max_bytes)
1686                 delalloc_end = delalloc_start + max_bytes - 1;
1687
1688         /* step two, lock all the pages after the page that has start */
1689         ret = lock_delalloc_pages(inode, locked_page,
1690                                   delalloc_start, delalloc_end);
1691         if (ret == -EAGAIN) {
1692                 /* some of the pages are gone, lets avoid looping by
1693                  * shortening the size of the delalloc range we're searching
1694                  */
1695                 free_extent_state(cached_state);
1696                 cached_state = NULL;
1697                 if (!loops) {
1698                         max_bytes = PAGE_SIZE;
1699                         loops = 1;
1700                         goto again;
1701                 } else {
1702                         found = 0;
1703                         goto out_failed;
1704                 }
1705         }
1706         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1707
1708         /* step three, lock the state bits for the whole range */
1709         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1710
1711         /* then test to make sure it is all still delalloc */
1712         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1713                              EXTENT_DELALLOC, 1, cached_state);
1714         if (!ret) {
1715                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1716                                      &cached_state, GFP_NOFS);
1717                 __unlock_for_delalloc(inode, locked_page,
1718                               delalloc_start, delalloc_end);
1719                 cond_resched();
1720                 goto again;
1721         }
1722         free_extent_state(cached_state);
1723         *start = delalloc_start;
1724         *end = delalloc_end;
1725 out_failed:
1726         return found;
1727 }
1728
1729 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1730                                  u64 delalloc_end, struct page *locked_page,
1731                                  unsigned clear_bits,
1732                                  unsigned long page_ops)
1733 {
1734         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1735         int ret;
1736         struct page *pages[16];
1737         unsigned long index = start >> PAGE_SHIFT;
1738         unsigned long end_index = end >> PAGE_SHIFT;
1739         unsigned long nr_pages = end_index - index + 1;
1740         int i;
1741
1742         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1743         if (page_ops == 0)
1744                 return;
1745
1746         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1747                 mapping_set_error(inode->i_mapping, -EIO);
1748
1749         while (nr_pages > 0) {
1750                 ret = find_get_pages_contig(inode->i_mapping, index,
1751                                      min_t(unsigned long,
1752                                      nr_pages, ARRAY_SIZE(pages)), pages);
1753                 for (i = 0; i < ret; i++) {
1754
1755                         if (page_ops & PAGE_SET_PRIVATE2)
1756                                 SetPagePrivate2(pages[i]);
1757
1758                         if (pages[i] == locked_page) {
1759                                 put_page(pages[i]);
1760                                 continue;
1761                         }
1762                         if (page_ops & PAGE_CLEAR_DIRTY)
1763                                 clear_page_dirty_for_io(pages[i]);
1764                         if (page_ops & PAGE_SET_WRITEBACK)
1765                                 set_page_writeback(pages[i]);
1766                         if (page_ops & PAGE_SET_ERROR)
1767                                 SetPageError(pages[i]);
1768                         if (page_ops & PAGE_END_WRITEBACK)
1769                                 end_page_writeback(pages[i]);
1770                         if (page_ops & PAGE_UNLOCK)
1771                                 unlock_page(pages[i]);
1772                         put_page(pages[i]);
1773                 }
1774                 nr_pages -= ret;
1775                 index += ret;
1776                 cond_resched();
1777         }
1778 }
1779
1780 /*
1781  * count the number of bytes in the tree that have a given bit(s)
1782  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1783  * cached.  The total number found is returned.
1784  */
1785 u64 count_range_bits(struct extent_io_tree *tree,
1786                      u64 *start, u64 search_end, u64 max_bytes,
1787                      unsigned bits, int contig)
1788 {
1789         struct rb_node *node;
1790         struct extent_state *state;
1791         u64 cur_start = *start;
1792         u64 total_bytes = 0;
1793         u64 last = 0;
1794         int found = 0;
1795
1796         if (WARN_ON(search_end <= cur_start))
1797                 return 0;
1798
1799         spin_lock(&tree->lock);
1800         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1801                 total_bytes = tree->dirty_bytes;
1802                 goto out;
1803         }
1804         /*
1805          * this search will find all the extents that end after
1806          * our range starts.
1807          */
1808         node = tree_search(tree, cur_start);
1809         if (!node)
1810                 goto out;
1811
1812         while (1) {
1813                 state = rb_entry(node, struct extent_state, rb_node);
1814                 if (state->start > search_end)
1815                         break;
1816                 if (contig && found && state->start > last + 1)
1817                         break;
1818                 if (state->end >= cur_start && (state->state & bits) == bits) {
1819                         total_bytes += min(search_end, state->end) + 1 -
1820                                        max(cur_start, state->start);
1821                         if (total_bytes >= max_bytes)
1822                                 break;
1823                         if (!found) {
1824                                 *start = max(cur_start, state->start);
1825                                 found = 1;
1826                         }
1827                         last = state->end;
1828                 } else if (contig && found) {
1829                         break;
1830                 }
1831                 node = rb_next(node);
1832                 if (!node)
1833                         break;
1834         }
1835 out:
1836         spin_unlock(&tree->lock);
1837         return total_bytes;
1838 }
1839
1840 /*
1841  * set the private field for a given byte offset in the tree.  If there isn't
1842  * an extent_state there already, this does nothing.
1843  */
1844 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1845                 struct io_failure_record *failrec)
1846 {
1847         struct rb_node *node;
1848         struct extent_state *state;
1849         int ret = 0;
1850
1851         spin_lock(&tree->lock);
1852         /*
1853          * this search will find all the extents that end after
1854          * our range starts.
1855          */
1856         node = tree_search(tree, start);
1857         if (!node) {
1858                 ret = -ENOENT;
1859                 goto out;
1860         }
1861         state = rb_entry(node, struct extent_state, rb_node);
1862         if (state->start != start) {
1863                 ret = -ENOENT;
1864                 goto out;
1865         }
1866         state->failrec = failrec;
1867 out:
1868         spin_unlock(&tree->lock);
1869         return ret;
1870 }
1871
1872 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1873                 struct io_failure_record **failrec)
1874 {
1875         struct rb_node *node;
1876         struct extent_state *state;
1877         int ret = 0;
1878
1879         spin_lock(&tree->lock);
1880         /*
1881          * this search will find all the extents that end after
1882          * our range starts.
1883          */
1884         node = tree_search(tree, start);
1885         if (!node) {
1886                 ret = -ENOENT;
1887                 goto out;
1888         }
1889         state = rb_entry(node, struct extent_state, rb_node);
1890         if (state->start != start) {
1891                 ret = -ENOENT;
1892                 goto out;
1893         }
1894         *failrec = state->failrec;
1895 out:
1896         spin_unlock(&tree->lock);
1897         return ret;
1898 }
1899
1900 /*
1901  * searches a range in the state tree for a given mask.
1902  * If 'filled' == 1, this returns 1 only if every extent in the tree
1903  * has the bits set.  Otherwise, 1 is returned if any bit in the
1904  * range is found set.
1905  */
1906 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1907                    unsigned bits, int filled, struct extent_state *cached)
1908 {
1909         struct extent_state *state = NULL;
1910         struct rb_node *node;
1911         int bitset = 0;
1912
1913         spin_lock(&tree->lock);
1914         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1915             cached->end > start)
1916                 node = &cached->rb_node;
1917         else
1918                 node = tree_search(tree, start);
1919         while (node && start <= end) {
1920                 state = rb_entry(node, struct extent_state, rb_node);
1921
1922                 if (filled && state->start > start) {
1923                         bitset = 0;
1924                         break;
1925                 }
1926
1927                 if (state->start > end)
1928                         break;
1929
1930                 if (state->state & bits) {
1931                         bitset = 1;
1932                         if (!filled)
1933                                 break;
1934                 } else if (filled) {
1935                         bitset = 0;
1936                         break;
1937                 }
1938
1939                 if (state->end == (u64)-1)
1940                         break;
1941
1942                 start = state->end + 1;
1943                 if (start > end)
1944                         break;
1945                 node = rb_next(node);
1946                 if (!node) {
1947                         if (filled)
1948                                 bitset = 0;
1949                         break;
1950                 }
1951         }
1952         spin_unlock(&tree->lock);
1953         return bitset;
1954 }
1955
1956 /*
1957  * helper function to set a given page up to date if all the
1958  * extents in the tree for that page are up to date
1959  */
1960 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1961 {
1962         u64 start = page_offset(page);
1963         u64 end = start + PAGE_SIZE - 1;
1964         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1965                 SetPageUptodate(page);
1966 }
1967
1968 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1969 {
1970         int ret;
1971         int err = 0;
1972         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1973
1974         set_state_failrec(failure_tree, rec->start, NULL);
1975         ret = clear_extent_bits(failure_tree, rec->start,
1976                                 rec->start + rec->len - 1,
1977                                 EXTENT_LOCKED | EXTENT_DIRTY);
1978         if (ret)
1979                 err = ret;
1980
1981         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1982                                 rec->start + rec->len - 1,
1983                                 EXTENT_DAMAGED);
1984         if (ret && !err)
1985                 err = ret;
1986
1987         kfree(rec);
1988         return err;
1989 }
1990
1991 /*
1992  * this bypasses the standard btrfs submit functions deliberately, as
1993  * the standard behavior is to write all copies in a raid setup. here we only
1994  * want to write the one bad copy. so we do the mapping for ourselves and issue
1995  * submit_bio directly.
1996  * to avoid any synchronization issues, wait for the data after writing, which
1997  * actually prevents the read that triggered the error from finishing.
1998  * currently, there can be no more than two copies of every data bit. thus,
1999  * exactly one rewrite is required.
2000  */
2001 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2002                       struct page *page, unsigned int pg_offset, int mirror_num)
2003 {
2004         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2005         struct bio *bio;
2006         struct btrfs_device *dev;
2007         u64 map_length = 0;
2008         u64 sector;
2009         struct btrfs_bio *bbio = NULL;
2010         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2011         int ret;
2012
2013         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2014         BUG_ON(!mirror_num);
2015
2016         /* we can't repair anything in raid56 yet */
2017         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2018                 return 0;
2019
2020         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2021         if (!bio)
2022                 return -EIO;
2023         bio->bi_iter.bi_size = 0;
2024         map_length = length;
2025
2026         /*
2027          * Avoid races with device replace and make sure our bbio has devices
2028          * associated to its stripes that don't go away while we are doing the
2029          * read repair operation.
2030          */
2031         btrfs_bio_counter_inc_blocked(fs_info);
2032         ret = btrfs_map_block(fs_info, WRITE, logical,
2033                               &map_length, &bbio, mirror_num);
2034         if (ret) {
2035                 btrfs_bio_counter_dec(fs_info);
2036                 bio_put(bio);
2037                 return -EIO;
2038         }
2039         BUG_ON(mirror_num != bbio->mirror_num);
2040         sector = bbio->stripes[mirror_num-1].physical >> 9;
2041         bio->bi_iter.bi_sector = sector;
2042         dev = bbio->stripes[mirror_num-1].dev;
2043         btrfs_put_bbio(bbio);
2044         if (!dev || !dev->bdev || !dev->writeable) {
2045                 btrfs_bio_counter_dec(fs_info);
2046                 bio_put(bio);
2047                 return -EIO;
2048         }
2049         bio->bi_bdev = dev->bdev;
2050         bio_set_op_attrs(bio, REQ_OP_WRITE, WRITE_SYNC);
2051         bio_add_page(bio, page, length, pg_offset);
2052
2053         if (btrfsic_submit_bio_wait(bio)) {
2054                 /* try to remap that extent elsewhere? */
2055                 btrfs_bio_counter_dec(fs_info);
2056                 bio_put(bio);
2057                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2058                 return -EIO;
2059         }
2060
2061         btrfs_info_rl_in_rcu(fs_info,
2062                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2063                                   btrfs_ino(inode), start,
2064                                   rcu_str_deref(dev->name), sector);
2065         btrfs_bio_counter_dec(fs_info);
2066         bio_put(bio);
2067         return 0;
2068 }
2069
2070 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2071                          int mirror_num)
2072 {
2073         u64 start = eb->start;
2074         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2075         int ret = 0;
2076
2077         if (root->fs_info->sb->s_flags & MS_RDONLY)
2078                 return -EROFS;
2079
2080         for (i = 0; i < num_pages; i++) {
2081                 struct page *p = eb->pages[i];
2082
2083                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2084                                         PAGE_SIZE, start, p,
2085                                         start - page_offset(p), mirror_num);
2086                 if (ret)
2087                         break;
2088                 start += PAGE_SIZE;
2089         }
2090
2091         return ret;
2092 }
2093
2094 /*
2095  * each time an IO finishes, we do a fast check in the IO failure tree
2096  * to see if we need to process or clean up an io_failure_record
2097  */
2098 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2099                      unsigned int pg_offset)
2100 {
2101         u64 private;
2102         struct io_failure_record *failrec;
2103         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2104         struct extent_state *state;
2105         int num_copies;
2106         int ret;
2107
2108         private = 0;
2109         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2110                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2111         if (!ret)
2112                 return 0;
2113
2114         ret = get_state_failrec(&BTRFS_I(inode)->io_failure_tree, start,
2115                         &failrec);
2116         if (ret)
2117                 return 0;
2118
2119         BUG_ON(!failrec->this_mirror);
2120
2121         if (failrec->in_validation) {
2122                 /* there was no real error, just free the record */
2123                 btrfs_debug(fs_info,
2124                         "clean_io_failure: freeing dummy error at %llu",
2125                         failrec->start);
2126                 goto out;
2127         }
2128         if (fs_info->sb->s_flags & MS_RDONLY)
2129                 goto out;
2130
2131         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2132         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2133                                             failrec->start,
2134                                             EXTENT_LOCKED);
2135         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2136
2137         if (state && state->start <= failrec->start &&
2138             state->end >= failrec->start + failrec->len - 1) {
2139                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2140                                               failrec->len);
2141                 if (num_copies > 1)  {
2142                         repair_io_failure(inode, start, failrec->len,
2143                                           failrec->logical, page,
2144                                           pg_offset, failrec->failed_mirror);
2145                 }
2146         }
2147
2148 out:
2149         free_io_failure(inode, failrec);
2150
2151         return 0;
2152 }
2153
2154 /*
2155  * Can be called when
2156  * - hold extent lock
2157  * - under ordered extent
2158  * - the inode is freeing
2159  */
2160 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2161 {
2162         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2163         struct io_failure_record *failrec;
2164         struct extent_state *state, *next;
2165
2166         if (RB_EMPTY_ROOT(&failure_tree->state))
2167                 return;
2168
2169         spin_lock(&failure_tree->lock);
2170         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2171         while (state) {
2172                 if (state->start > end)
2173                         break;
2174
2175                 ASSERT(state->end <= end);
2176
2177                 next = next_state(state);
2178
2179                 failrec = state->failrec;
2180                 free_extent_state(state);
2181                 kfree(failrec);
2182
2183                 state = next;
2184         }
2185         spin_unlock(&failure_tree->lock);
2186 }
2187
2188 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2189                 struct io_failure_record **failrec_ret)
2190 {
2191         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
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                 btrfs_debug(fs_info,
2240                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2241                         logical, start, failrec->len);
2242
2243                 failrec->logical = logical;
2244                 free_extent_map(em);
2245
2246                 /* set the bits in the private failure tree */
2247                 ret = set_extent_bits(failure_tree, start, end,
2248                                         EXTENT_LOCKED | EXTENT_DIRTY);
2249                 if (ret >= 0)
2250                         ret = set_state_failrec(failure_tree, start, failrec);
2251                 /* set the bits in the inode's tree */
2252                 if (ret >= 0)
2253                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2254                 if (ret < 0) {
2255                         kfree(failrec);
2256                         return ret;
2257                 }
2258         } else {
2259                 btrfs_debug(fs_info,
2260                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2261                         failrec->logical, failrec->start, failrec->len,
2262                         failrec->in_validation);
2263                 /*
2264                  * when data can be on disk more than twice, add to failrec here
2265                  * (e.g. with a list for failed_mirror) to make
2266                  * clean_io_failure() clean all those errors at once.
2267                  */
2268         }
2269
2270         *failrec_ret = failrec;
2271
2272         return 0;
2273 }
2274
2275 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2276                            struct io_failure_record *failrec, int failed_mirror)
2277 {
2278         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2279         int num_copies;
2280
2281         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2282         if (num_copies == 1) {
2283                 /*
2284                  * we only have a single copy of the data, so don't bother with
2285                  * all the retry and error correction code that follows. no
2286                  * matter what the error is, it is very likely to persist.
2287                  */
2288                 btrfs_debug(fs_info,
2289                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2290                         num_copies, failrec->this_mirror, failed_mirror);
2291                 return 0;
2292         }
2293
2294         /*
2295          * there are two premises:
2296          *      a) deliver good data to the caller
2297          *      b) correct the bad sectors on disk
2298          */
2299         if (failed_bio->bi_vcnt > 1) {
2300                 /*
2301                  * to fulfill b), we need to know the exact failing sectors, as
2302                  * we don't want to rewrite any more than the failed ones. thus,
2303                  * we need separate read requests for the failed bio
2304                  *
2305                  * if the following BUG_ON triggers, our validation request got
2306                  * merged. we need separate requests for our algorithm to work.
2307                  */
2308                 BUG_ON(failrec->in_validation);
2309                 failrec->in_validation = 1;
2310                 failrec->this_mirror = failed_mirror;
2311         } else {
2312                 /*
2313                  * we're ready to fulfill a) and b) alongside. get a good copy
2314                  * of the failed sector and if we succeed, we have setup
2315                  * everything for repair_io_failure to do the rest for us.
2316                  */
2317                 if (failrec->in_validation) {
2318                         BUG_ON(failrec->this_mirror != failed_mirror);
2319                         failrec->in_validation = 0;
2320                         failrec->this_mirror = 0;
2321                 }
2322                 failrec->failed_mirror = failed_mirror;
2323                 failrec->this_mirror++;
2324                 if (failrec->this_mirror == failed_mirror)
2325                         failrec->this_mirror++;
2326         }
2327
2328         if (failrec->this_mirror > num_copies) {
2329                 btrfs_debug(fs_info,
2330                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2331                         num_copies, failrec->this_mirror, failed_mirror);
2332                 return 0;
2333         }
2334
2335         return 1;
2336 }
2337
2338
2339 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2340                                     struct io_failure_record *failrec,
2341                                     struct page *page, int pg_offset, int icsum,
2342                                     bio_end_io_t *endio_func, void *data)
2343 {
2344         struct bio *bio;
2345         struct btrfs_io_bio *btrfs_failed_bio;
2346         struct btrfs_io_bio *btrfs_bio;
2347
2348         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2349         if (!bio)
2350                 return NULL;
2351
2352         bio->bi_end_io = endio_func;
2353         bio->bi_iter.bi_sector = failrec->logical >> 9;
2354         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2355         bio->bi_iter.bi_size = 0;
2356         bio->bi_private = data;
2357
2358         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2359         if (btrfs_failed_bio->csum) {
2360                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2361                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2362
2363                 btrfs_bio = btrfs_io_bio(bio);
2364                 btrfs_bio->csum = btrfs_bio->csum_inline;
2365                 icsum *= csum_size;
2366                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2367                        csum_size);
2368         }
2369
2370         bio_add_page(bio, page, failrec->len, pg_offset);
2371
2372         return bio;
2373 }
2374
2375 /*
2376  * this is a generic handler for readpage errors (default
2377  * readpage_io_failed_hook). if other copies exist, read those and write back
2378  * good data to the failed position. does not investigate in remapping the
2379  * failed extent elsewhere, hoping the device will be smart enough to do this as
2380  * needed
2381  */
2382
2383 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2384                               struct page *page, u64 start, u64 end,
2385                               int failed_mirror)
2386 {
2387         struct io_failure_record *failrec;
2388         struct inode *inode = page->mapping->host;
2389         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2390         struct bio *bio;
2391         int read_mode;
2392         int ret;
2393
2394         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2395
2396         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2397         if (ret)
2398                 return ret;
2399
2400         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2401         if (!ret) {
2402                 free_io_failure(inode, failrec);
2403                 return -EIO;
2404         }
2405
2406         if (failed_bio->bi_vcnt > 1)
2407                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2408         else
2409                 read_mode = READ_SYNC;
2410
2411         phy_offset >>= inode->i_sb->s_blocksize_bits;
2412         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2413                                       start - page_offset(page),
2414                                       (int)phy_offset, failed_bio->bi_end_io,
2415                                       NULL);
2416         if (!bio) {
2417                 free_io_failure(inode, failrec);
2418                 return -EIO;
2419         }
2420         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2421
2422         btrfs_debug(btrfs_sb(inode->i_sb),
2423                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2424                 read_mode, failrec->this_mirror, failrec->in_validation);
2425
2426         ret = tree->ops->submit_bio_hook(inode, bio, failrec->this_mirror,
2427                                          failrec->bio_flags, 0);
2428         if (ret) {
2429                 free_io_failure(inode, failrec);
2430                 bio_put(bio);
2431         }
2432
2433         return ret;
2434 }
2435
2436 /* lots and lots of room for performance fixes in the end_bio funcs */
2437
2438 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2439 {
2440         int uptodate = (err == 0);
2441         struct extent_io_tree *tree;
2442         int ret = 0;
2443
2444         tree = &BTRFS_I(page->mapping->host)->io_tree;
2445
2446         if (tree->ops && tree->ops->writepage_end_io_hook) {
2447                 ret = tree->ops->writepage_end_io_hook(page, start,
2448                                                end, NULL, uptodate);
2449                 if (ret)
2450                         uptodate = 0;
2451         }
2452
2453         if (!uptodate) {
2454                 ClearPageUptodate(page);
2455                 SetPageError(page);
2456                 ret = ret < 0 ? ret : -EIO;
2457                 mapping_set_error(page->mapping, ret);
2458         }
2459 }
2460
2461 /*
2462  * after a writepage IO is done, we need to:
2463  * clear the uptodate bits on error
2464  * clear the writeback bits in the extent tree for this IO
2465  * end_page_writeback if the page has no more pending IO
2466  *
2467  * Scheduling is not allowed, so the extent state tree is expected
2468  * to have one and only one object corresponding to this IO.
2469  */
2470 static void end_bio_extent_writepage(struct bio *bio)
2471 {
2472         struct bio_vec *bvec;
2473         u64 start;
2474         u64 end;
2475         int i;
2476
2477         bio_for_each_segment_all(bvec, bio, i) {
2478                 struct page *page = bvec->bv_page;
2479
2480                 /* We always issue full-page reads, but if some block
2481                  * in a page fails to read, blk_update_request() will
2482                  * advance bv_offset and adjust bv_len to compensate.
2483                  * Print a warning for nonzero offsets, and an error
2484                  * if they don't add up to a full page.  */
2485                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2486                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2487                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2488                                    "partial page write in btrfs with offset %u and length %u",
2489                                         bvec->bv_offset, bvec->bv_len);
2490                         else
2491                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2492                                    "incomplete page write in btrfs with offset %u and length %u",
2493                                         bvec->bv_offset, bvec->bv_len);
2494                 }
2495
2496                 start = page_offset(page);
2497                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2498
2499                 end_extent_writepage(page, bio->bi_error, start, end);
2500                 end_page_writeback(page);
2501         }
2502
2503         bio_put(bio);
2504 }
2505
2506 static void
2507 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2508                               int uptodate)
2509 {
2510         struct extent_state *cached = NULL;
2511         u64 end = start + len - 1;
2512
2513         if (uptodate && tree->track_uptodate)
2514                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2515         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2516 }
2517
2518 /*
2519  * after a readpage IO is done, we need to:
2520  * clear the uptodate bits on error
2521  * set the uptodate bits if things worked
2522  * set the page up to date if all extents in the tree are uptodate
2523  * clear the lock bit in the extent tree
2524  * unlock the page if there are no other extents locked for it
2525  *
2526  * Scheduling is not allowed, so the extent state tree is expected
2527  * to have one and only one object corresponding to this IO.
2528  */
2529 static void end_bio_extent_readpage(struct bio *bio)
2530 {
2531         struct bio_vec *bvec;
2532         int uptodate = !bio->bi_error;
2533         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2534         struct extent_io_tree *tree;
2535         u64 offset = 0;
2536         u64 start;
2537         u64 end;
2538         u64 len;
2539         u64 extent_start = 0;
2540         u64 extent_len = 0;
2541         int mirror;
2542         int ret;
2543         int i;
2544
2545         bio_for_each_segment_all(bvec, bio, i) {
2546                 struct page *page = bvec->bv_page;
2547                 struct inode *inode = page->mapping->host;
2548                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2549
2550                 btrfs_debug(fs_info,
2551                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2552                         (u64)bio->bi_iter.bi_sector, bio->bi_error,
2553                         io_bio->mirror_num);
2554                 tree = &BTRFS_I(inode)->io_tree;
2555
2556                 /* We always issue full-page reads, but if some block
2557                  * in a page fails to read, blk_update_request() will
2558                  * advance bv_offset and adjust bv_len to compensate.
2559                  * Print a warning for nonzero offsets, and an error
2560                  * if they don't add up to a full page.  */
2561                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2562                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2563                                 btrfs_err(fs_info,
2564                                         "partial page read in btrfs with offset %u and length %u",
2565                                         bvec->bv_offset, bvec->bv_len);
2566                         else
2567                                 btrfs_info(fs_info,
2568                                         "incomplete page read in btrfs with offset %u and length %u",
2569                                         bvec->bv_offset, bvec->bv_len);
2570                 }
2571
2572                 start = page_offset(page);
2573                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2574                 len = bvec->bv_len;
2575
2576                 mirror = io_bio->mirror_num;
2577                 if (likely(uptodate && tree->ops &&
2578                            tree->ops->readpage_end_io_hook)) {
2579                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2580                                                               page, start, end,
2581                                                               mirror);
2582                         if (ret)
2583                                 uptodate = 0;
2584                         else
2585                                 clean_io_failure(inode, start, page, 0);
2586                 }
2587
2588                 if (likely(uptodate))
2589                         goto readpage_ok;
2590
2591                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2592                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2593                         if (!ret && !bio->bi_error)
2594                                 uptodate = 1;
2595                 } else {
2596                         /*
2597                          * The generic bio_readpage_error handles errors the
2598                          * following way: If possible, new read requests are
2599                          * created and submitted and will end up in
2600                          * end_bio_extent_readpage as well (if we're lucky, not
2601                          * in the !uptodate case). In that case it returns 0 and
2602                          * we just go on with the next page in our bio. If it
2603                          * can't handle the error it will return -EIO and we
2604                          * remain responsible for that page.
2605                          */
2606                         ret = bio_readpage_error(bio, offset, page, start, end,
2607                                                  mirror);
2608                         if (ret == 0) {
2609                                 uptodate = !bio->bi_error;
2610                                 offset += len;
2611                                 continue;
2612                         }
2613                 }
2614 readpage_ok:
2615                 if (likely(uptodate)) {
2616                         loff_t i_size = i_size_read(inode);
2617                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2618                         unsigned off;
2619
2620                         /* Zero out the end if this page straddles i_size */
2621                         off = i_size & (PAGE_SIZE-1);
2622                         if (page->index == end_index && off)
2623                                 zero_user_segment(page, off, PAGE_SIZE);
2624                         SetPageUptodate(page);
2625                 } else {
2626                         ClearPageUptodate(page);
2627                         SetPageError(page);
2628                 }
2629                 unlock_page(page);
2630                 offset += len;
2631
2632                 if (unlikely(!uptodate)) {
2633                         if (extent_len) {
2634                                 endio_readpage_release_extent(tree,
2635                                                               extent_start,
2636                                                               extent_len, 1);
2637                                 extent_start = 0;
2638                                 extent_len = 0;
2639                         }
2640                         endio_readpage_release_extent(tree, start,
2641                                                       end - start + 1, 0);
2642                 } else if (!extent_len) {
2643                         extent_start = start;
2644                         extent_len = end + 1 - start;
2645                 } else if (extent_start + extent_len == start) {
2646                         extent_len += end + 1 - start;
2647                 } else {
2648                         endio_readpage_release_extent(tree, extent_start,
2649                                                       extent_len, uptodate);
2650                         extent_start = start;
2651                         extent_len = end + 1 - start;
2652                 }
2653         }
2654
2655         if (extent_len)
2656                 endio_readpage_release_extent(tree, extent_start, extent_len,
2657                                               uptodate);
2658         if (io_bio->end_io)
2659                 io_bio->end_io(io_bio, bio->bi_error);
2660         bio_put(bio);
2661 }
2662
2663 /*
2664  * this allocates from the btrfs_bioset.  We're returning a bio right now
2665  * but you can call btrfs_io_bio for the appropriate container_of magic
2666  */
2667 struct bio *
2668 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2669                 gfp_t gfp_flags)
2670 {
2671         struct btrfs_io_bio *btrfs_bio;
2672         struct bio *bio;
2673
2674         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2675
2676         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2677                 while (!bio && (nr_vecs /= 2)) {
2678                         bio = bio_alloc_bioset(gfp_flags,
2679                                                nr_vecs, btrfs_bioset);
2680                 }
2681         }
2682
2683         if (bio) {
2684                 bio->bi_bdev = bdev;
2685                 bio->bi_iter.bi_sector = first_sector;
2686                 btrfs_bio = btrfs_io_bio(bio);
2687                 btrfs_bio->csum = NULL;
2688                 btrfs_bio->csum_allocated = NULL;
2689                 btrfs_bio->end_io = NULL;
2690         }
2691         return bio;
2692 }
2693
2694 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2695 {
2696         struct btrfs_io_bio *btrfs_bio;
2697         struct bio *new;
2698
2699         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2700         if (new) {
2701                 btrfs_bio = btrfs_io_bio(new);
2702                 btrfs_bio->csum = NULL;
2703                 btrfs_bio->csum_allocated = NULL;
2704                 btrfs_bio->end_io = NULL;
2705         }
2706         return new;
2707 }
2708
2709 /* this also allocates from the btrfs_bioset */
2710 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2711 {
2712         struct btrfs_io_bio *btrfs_bio;
2713         struct bio *bio;
2714
2715         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2716         if (bio) {
2717                 btrfs_bio = btrfs_io_bio(bio);
2718                 btrfs_bio->csum = NULL;
2719                 btrfs_bio->csum_allocated = NULL;
2720                 btrfs_bio->end_io = NULL;
2721         }
2722         return bio;
2723 }
2724
2725
2726 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2727                                        unsigned long bio_flags)
2728 {
2729         int ret = 0;
2730         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2731         struct page *page = bvec->bv_page;
2732         struct extent_io_tree *tree = bio->bi_private;
2733         u64 start;
2734
2735         start = page_offset(page) + bvec->bv_offset;
2736
2737         bio->bi_private = NULL;
2738         bio_get(bio);
2739
2740         if (tree->ops && tree->ops->submit_bio_hook)
2741                 ret = tree->ops->submit_bio_hook(page->mapping->host, bio,
2742                                            mirror_num, bio_flags, start);
2743         else
2744                 btrfsic_submit_bio(bio);
2745
2746         bio_put(bio);
2747         return ret;
2748 }
2749
2750 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2751                      unsigned long offset, size_t size, struct bio *bio,
2752                      unsigned long bio_flags)
2753 {
2754         int ret = 0;
2755         if (tree->ops && tree->ops->merge_bio_hook)
2756                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2757                                                 bio_flags);
2758         return ret;
2759
2760 }
2761
2762 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2763                               struct writeback_control *wbc,
2764                               struct page *page, sector_t sector,
2765                               size_t size, unsigned long offset,
2766                               struct block_device *bdev,
2767                               struct bio **bio_ret,
2768                               unsigned long max_pages,
2769                               bio_end_io_t end_io_func,
2770                               int mirror_num,
2771                               unsigned long prev_bio_flags,
2772                               unsigned long bio_flags,
2773                               bool force_bio_submit)
2774 {
2775         int ret = 0;
2776         struct bio *bio;
2777         int contig = 0;
2778         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2779         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2780
2781         if (bio_ret && *bio_ret) {
2782                 bio = *bio_ret;
2783                 if (old_compressed)
2784                         contig = bio->bi_iter.bi_sector == sector;
2785                 else
2786                         contig = bio_end_sector(bio) == sector;
2787
2788                 if (prev_bio_flags != bio_flags || !contig ||
2789                     force_bio_submit ||
2790                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2791                     bio_add_page(bio, page, page_size, offset) < page_size) {
2792                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2793                         if (ret < 0) {
2794                                 *bio_ret = NULL;
2795                                 return ret;
2796                         }
2797                         bio = NULL;
2798                 } else {
2799                         if (wbc)
2800                                 wbc_account_io(wbc, page, page_size);
2801                         return 0;
2802                 }
2803         }
2804
2805         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2806                         GFP_NOFS | __GFP_HIGH);
2807         if (!bio)
2808                 return -ENOMEM;
2809
2810         bio_add_page(bio, page, page_size, offset);
2811         bio->bi_end_io = end_io_func;
2812         bio->bi_private = tree;
2813         bio_set_op_attrs(bio, op, op_flags);
2814         if (wbc) {
2815                 wbc_init_bio(wbc, bio);
2816                 wbc_account_io(wbc, page, page_size);
2817         }
2818
2819         if (bio_ret)
2820                 *bio_ret = bio;
2821         else
2822                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2823
2824         return ret;
2825 }
2826
2827 static void attach_extent_buffer_page(struct extent_buffer *eb,
2828                                       struct page *page)
2829 {
2830         if (!PagePrivate(page)) {
2831                 SetPagePrivate(page);
2832                 get_page(page);
2833                 set_page_private(page, (unsigned long)eb);
2834         } else {
2835                 WARN_ON(page->private != (unsigned long)eb);
2836         }
2837 }
2838
2839 void set_page_extent_mapped(struct page *page)
2840 {
2841         if (!PagePrivate(page)) {
2842                 SetPagePrivate(page);
2843                 get_page(page);
2844                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2845         }
2846 }
2847
2848 static struct extent_map *
2849 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2850                  u64 start, u64 len, get_extent_t *get_extent,
2851                  struct extent_map **em_cached)
2852 {
2853         struct extent_map *em;
2854
2855         if (em_cached && *em_cached) {
2856                 em = *em_cached;
2857                 if (extent_map_in_tree(em) && start >= em->start &&
2858                     start < extent_map_end(em)) {
2859                         atomic_inc(&em->refs);
2860                         return em;
2861                 }
2862
2863                 free_extent_map(em);
2864                 *em_cached = NULL;
2865         }
2866
2867         em = get_extent(inode, page, pg_offset, start, len, 0);
2868         if (em_cached && !IS_ERR_OR_NULL(em)) {
2869                 BUG_ON(*em_cached);
2870                 atomic_inc(&em->refs);
2871                 *em_cached = em;
2872         }
2873         return em;
2874 }
2875 /*
2876  * basic readpage implementation.  Locked extent state structs are inserted
2877  * into the tree that are removed when the IO is done (by the end_io
2878  * handlers)
2879  * XXX JDM: This needs looking at to ensure proper page locking
2880  * return 0 on success, otherwise return error
2881  */
2882 static int __do_readpage(struct extent_io_tree *tree,
2883                          struct page *page,
2884                          get_extent_t *get_extent,
2885                          struct extent_map **em_cached,
2886                          struct bio **bio, int mirror_num,
2887                          unsigned long *bio_flags, int read_flags,
2888                          u64 *prev_em_start)
2889 {
2890         struct inode *inode = page->mapping->host;
2891         u64 start = page_offset(page);
2892         u64 page_end = start + PAGE_SIZE - 1;
2893         u64 end;
2894         u64 cur = start;
2895         u64 extent_offset;
2896         u64 last_byte = i_size_read(inode);
2897         u64 block_start;
2898         u64 cur_end;
2899         sector_t sector;
2900         struct extent_map *em;
2901         struct block_device *bdev;
2902         int ret = 0;
2903         int nr = 0;
2904         size_t pg_offset = 0;
2905         size_t iosize;
2906         size_t disk_io_size;
2907         size_t blocksize = inode->i_sb->s_blocksize;
2908         unsigned long this_bio_flag = 0;
2909
2910         set_page_extent_mapped(page);
2911
2912         end = page_end;
2913         if (!PageUptodate(page)) {
2914                 if (cleancache_get_page(page) == 0) {
2915                         BUG_ON(blocksize != PAGE_SIZE);
2916                         unlock_extent(tree, start, end);
2917                         goto out;
2918                 }
2919         }
2920
2921         if (page->index == last_byte >> PAGE_SHIFT) {
2922                 char *userpage;
2923                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2924
2925                 if (zero_offset) {
2926                         iosize = PAGE_SIZE - zero_offset;
2927                         userpage = kmap_atomic(page);
2928                         memset(userpage + zero_offset, 0, iosize);
2929                         flush_dcache_page(page);
2930                         kunmap_atomic(userpage);
2931                 }
2932         }
2933         while (cur <= end) {
2934                 unsigned long pnr = (last_byte >> PAGE_SHIFT) + 1;
2935                 bool force_bio_submit = false;
2936
2937                 if (cur >= last_byte) {
2938                         char *userpage;
2939                         struct extent_state *cached = NULL;
2940
2941                         iosize = PAGE_SIZE - pg_offset;
2942                         userpage = kmap_atomic(page);
2943                         memset(userpage + pg_offset, 0, iosize);
2944                         flush_dcache_page(page);
2945                         kunmap_atomic(userpage);
2946                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2947                                             &cached, GFP_NOFS);
2948                         unlock_extent_cached(tree, cur,
2949                                              cur + iosize - 1,
2950                                              &cached, GFP_NOFS);
2951                         break;
2952                 }
2953                 em = __get_extent_map(inode, page, pg_offset, cur,
2954                                       end - cur + 1, get_extent, em_cached);
2955                 if (IS_ERR_OR_NULL(em)) {
2956                         SetPageError(page);
2957                         unlock_extent(tree, cur, end);
2958                         break;
2959                 }
2960                 extent_offset = cur - em->start;
2961                 BUG_ON(extent_map_end(em) <= cur);
2962                 BUG_ON(end < cur);
2963
2964                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2965                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2966                         extent_set_compress_type(&this_bio_flag,
2967                                                  em->compress_type);
2968                 }
2969
2970                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2971                 cur_end = min(extent_map_end(em) - 1, end);
2972                 iosize = ALIGN(iosize, blocksize);
2973                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2974                         disk_io_size = em->block_len;
2975                         sector = em->block_start >> 9;
2976                 } else {
2977                         sector = (em->block_start + extent_offset) >> 9;
2978                         disk_io_size = iosize;
2979                 }
2980                 bdev = em->bdev;
2981                 block_start = em->block_start;
2982                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2983                         block_start = EXTENT_MAP_HOLE;
2984
2985                 /*
2986                  * If we have a file range that points to a compressed extent
2987                  * and it's followed by a consecutive file range that points to
2988                  * to the same compressed extent (possibly with a different
2989                  * offset and/or length, so it either points to the whole extent
2990                  * or only part of it), we must make sure we do not submit a
2991                  * single bio to populate the pages for the 2 ranges because
2992                  * this makes the compressed extent read zero out the pages
2993                  * belonging to the 2nd range. Imagine the following scenario:
2994                  *
2995                  *  File layout
2996                  *  [0 - 8K]                     [8K - 24K]
2997                  *    |                               |
2998                  *    |                               |
2999                  * points to extent X,         points to extent X,
3000                  * offset 4K, length of 8K     offset 0, length 16K
3001                  *
3002                  * [extent X, compressed length = 4K uncompressed length = 16K]
3003                  *
3004                  * If the bio to read the compressed extent covers both ranges,
3005                  * it will decompress extent X into the pages belonging to the
3006                  * first range and then it will stop, zeroing out the remaining
3007                  * pages that belong to the other range that points to extent X.
3008                  * So here we make sure we submit 2 bios, one for the first
3009                  * range and another one for the third range. Both will target
3010                  * the same physical extent from disk, but we can't currently
3011                  * make the compressed bio endio callback populate the pages
3012                  * for both ranges because each compressed bio is tightly
3013                  * coupled with a single extent map, and each range can have
3014                  * an extent map with a different offset value relative to the
3015                  * uncompressed data of our extent and different lengths. This
3016                  * is a corner case so we prioritize correctness over
3017                  * non-optimal behavior (submitting 2 bios for the same extent).
3018                  */
3019                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3020                     prev_em_start && *prev_em_start != (u64)-1 &&
3021                     *prev_em_start != em->orig_start)
3022                         force_bio_submit = true;
3023
3024                 if (prev_em_start)
3025                         *prev_em_start = em->orig_start;
3026
3027                 free_extent_map(em);
3028                 em = NULL;
3029
3030                 /* we've found a hole, just zero and go on */
3031                 if (block_start == EXTENT_MAP_HOLE) {
3032                         char *userpage;
3033                         struct extent_state *cached = NULL;
3034
3035                         userpage = kmap_atomic(page);
3036                         memset(userpage + pg_offset, 0, iosize);
3037                         flush_dcache_page(page);
3038                         kunmap_atomic(userpage);
3039
3040                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3041                                             &cached, GFP_NOFS);
3042                         unlock_extent_cached(tree, cur,
3043                                              cur + iosize - 1,
3044                                              &cached, GFP_NOFS);
3045                         cur = cur + iosize;
3046                         pg_offset += iosize;
3047                         continue;
3048                 }
3049                 /* the get_extent function already copied into the page */
3050                 if (test_range_bit(tree, cur, cur_end,
3051                                    EXTENT_UPTODATE, 1, NULL)) {
3052                         check_page_uptodate(tree, page);
3053                         unlock_extent(tree, cur, cur + iosize - 1);
3054                         cur = cur + iosize;
3055                         pg_offset += iosize;
3056                         continue;
3057                 }
3058                 /* we have an inline extent but it didn't get marked up
3059                  * to date.  Error out
3060                  */
3061                 if (block_start == EXTENT_MAP_INLINE) {
3062                         SetPageError(page);
3063                         unlock_extent(tree, cur, cur + iosize - 1);
3064                         cur = cur + iosize;
3065                         pg_offset += iosize;
3066                         continue;
3067                 }
3068
3069                 pnr -= page->index;
3070                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3071                                          page, sector, disk_io_size, pg_offset,
3072                                          bdev, bio, pnr,
3073                                          end_bio_extent_readpage, mirror_num,
3074                                          *bio_flags,
3075                                          this_bio_flag,
3076                                          force_bio_submit);
3077                 if (!ret) {
3078                         nr++;
3079                         *bio_flags = this_bio_flag;
3080                 } else {
3081                         SetPageError(page);
3082                         unlock_extent(tree, cur, cur + iosize - 1);
3083                         goto out;
3084                 }
3085                 cur = cur + iosize;
3086                 pg_offset += iosize;
3087         }
3088 out:
3089         if (!nr) {
3090                 if (!PageError(page))
3091                         SetPageUptodate(page);
3092                 unlock_page(page);
3093         }
3094         return ret;
3095 }
3096
3097 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3098                                              struct page *pages[], int nr_pages,
3099                                              u64 start, u64 end,
3100                                              get_extent_t *get_extent,
3101                                              struct extent_map **em_cached,
3102                                              struct bio **bio, int mirror_num,
3103                                              unsigned long *bio_flags,
3104                                              u64 *prev_em_start)
3105 {
3106         struct inode *inode;
3107         struct btrfs_ordered_extent *ordered;
3108         int index;
3109
3110         inode = pages[0]->mapping->host;
3111         while (1) {
3112                 lock_extent(tree, start, end);
3113                 ordered = btrfs_lookup_ordered_range(inode, start,
3114                                                      end - start + 1);
3115                 if (!ordered)
3116                         break;
3117                 unlock_extent(tree, start, end);
3118                 btrfs_start_ordered_extent(inode, ordered, 1);
3119                 btrfs_put_ordered_extent(ordered);
3120         }
3121
3122         for (index = 0; index < nr_pages; index++) {
3123                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3124                               mirror_num, bio_flags, 0, prev_em_start);
3125                 put_page(pages[index]);
3126         }
3127 }
3128
3129 static void __extent_readpages(struct extent_io_tree *tree,
3130                                struct page *pages[],
3131                                int nr_pages, get_extent_t *get_extent,
3132                                struct extent_map **em_cached,
3133                                struct bio **bio, int mirror_num,
3134                                unsigned long *bio_flags,
3135                                u64 *prev_em_start)
3136 {
3137         u64 start = 0;
3138         u64 end = 0;
3139         u64 page_start;
3140         int index;
3141         int first_index = 0;
3142
3143         for (index = 0; index < nr_pages; index++) {
3144                 page_start = page_offset(pages[index]);
3145                 if (!end) {
3146                         start = page_start;
3147                         end = start + PAGE_SIZE - 1;
3148                         first_index = index;
3149                 } else if (end + 1 == page_start) {
3150                         end += PAGE_SIZE;
3151                 } else {
3152                         __do_contiguous_readpages(tree, &pages[first_index],
3153                                                   index - first_index, start,
3154                                                   end, get_extent, em_cached,
3155                                                   bio, mirror_num, bio_flags,
3156                                                   prev_em_start);
3157                         start = page_start;
3158                         end = start + PAGE_SIZE - 1;
3159                         first_index = index;
3160                 }
3161         }
3162
3163         if (end)
3164                 __do_contiguous_readpages(tree, &pages[first_index],
3165                                           index - first_index, start,
3166                                           end, get_extent, em_cached, bio,
3167                                           mirror_num, bio_flags,
3168                                           prev_em_start);
3169 }
3170
3171 static int __extent_read_full_page(struct extent_io_tree *tree,
3172                                    struct page *page,
3173                                    get_extent_t *get_extent,
3174                                    struct bio **bio, int mirror_num,
3175                                    unsigned long *bio_flags, int read_flags)
3176 {
3177         struct inode *inode = page->mapping->host;
3178         struct btrfs_ordered_extent *ordered;
3179         u64 start = page_offset(page);
3180         u64 end = start + PAGE_SIZE - 1;
3181         int ret;
3182
3183         while (1) {
3184                 lock_extent(tree, start, end);
3185                 ordered = btrfs_lookup_ordered_range(inode, start,
3186                                                 PAGE_SIZE);
3187                 if (!ordered)
3188                         break;
3189                 unlock_extent(tree, start, end);
3190                 btrfs_start_ordered_extent(inode, ordered, 1);
3191                 btrfs_put_ordered_extent(ordered);
3192         }
3193
3194         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3195                             bio_flags, read_flags, NULL);
3196         return ret;
3197 }
3198
3199 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3200                             get_extent_t *get_extent, int mirror_num)
3201 {
3202         struct bio *bio = NULL;
3203         unsigned long bio_flags = 0;
3204         int ret;
3205
3206         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3207                                       &bio_flags, 0);
3208         if (bio)
3209                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3210         return ret;
3211 }
3212
3213 static void update_nr_written(struct page *page, struct writeback_control *wbc,
3214                               unsigned long nr_written)
3215 {
3216         wbc->nr_to_write -= nr_written;
3217 }
3218
3219 /*
3220  * helper for __extent_writepage, doing all of the delayed allocation setup.
3221  *
3222  * This returns 1 if our fill_delalloc function did all the work required
3223  * to write the page (copy into inline extent).  In this case the IO has
3224  * been started and the page is already unlocked.
3225  *
3226  * This returns 0 if all went well (page still locked)
3227  * This returns < 0 if there were errors (page still locked)
3228  */
3229 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3230                               struct page *page, struct writeback_control *wbc,
3231                               struct extent_page_data *epd,
3232                               u64 delalloc_start,
3233                               unsigned long *nr_written)
3234 {
3235         struct extent_io_tree *tree = epd->tree;
3236         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3237         u64 nr_delalloc;
3238         u64 delalloc_to_write = 0;
3239         u64 delalloc_end = 0;
3240         int ret;
3241         int page_started = 0;
3242
3243         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3244                 return 0;
3245
3246         while (delalloc_end < page_end) {
3247                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3248                                                page,
3249                                                &delalloc_start,
3250                                                &delalloc_end,
3251                                                BTRFS_MAX_EXTENT_SIZE);
3252                 if (nr_delalloc == 0) {
3253                         delalloc_start = delalloc_end + 1;
3254                         continue;
3255                 }
3256                 ret = tree->ops->fill_delalloc(inode, page,
3257                                                delalloc_start,
3258                                                delalloc_end,
3259                                                &page_started,
3260                                                nr_written);
3261                 /* File system has been set read-only */
3262                 if (ret) {
3263                         SetPageError(page);
3264                         /* fill_delalloc should be return < 0 for error
3265                          * but just in case, we use > 0 here meaning the
3266                          * IO is started, so we don't want to return > 0
3267                          * unless things are going well.
3268                          */
3269                         ret = ret < 0 ? ret : -EIO;
3270                         goto done;
3271                 }
3272                 /*
3273