btrfs: Remove EXTENT_FIRST_DELALLOC bit
[sfrench/cifs-2.6.git] / fs / btrfs / extent_io.c
1 // SPDX-License-Identifier: GPL-2.0
2
3 #include <linux/bitops.h>
4 #include <linux/slab.h>
5 #include <linux/bio.h>
6 #include <linux/mm.h>
7 #include <linux/pagemap.h>
8 #include <linux/page-flags.h>
9 #include <linux/spinlock.h>
10 #include <linux/blkdev.h>
11 #include <linux/swap.h>
12 #include <linux/writeback.h>
13 #include <linux/pagevec.h>
14 #include <linux/prefetch.h>
15 #include <linux/cleancache.h>
16 #include "extent_io.h"
17 #include "extent_map.h"
18 #include "ctree.h"
19 #include "btrfs_inode.h"
20 #include "volumes.h"
21 #include "check-integrity.h"
22 #include "locking.h"
23 #include "rcu-string.h"
24 #include "backref.h"
25 #include "disk-io.h"
26
27 static struct kmem_cache *extent_state_cache;
28 static struct kmem_cache *extent_buffer_cache;
29 static struct bio_set btrfs_bioset;
30
31 static inline bool extent_state_in_tree(const struct extent_state *state)
32 {
33         return !RB_EMPTY_NODE(&state->rb_node);
34 }
35
36 #ifdef CONFIG_BTRFS_DEBUG
37 static LIST_HEAD(buffers);
38 static LIST_HEAD(states);
39
40 static DEFINE_SPINLOCK(leak_lock);
41
42 static inline
43 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
44 {
45         unsigned long flags;
46
47         spin_lock_irqsave(&leak_lock, flags);
48         list_add(new, head);
49         spin_unlock_irqrestore(&leak_lock, flags);
50 }
51
52 static inline
53 void btrfs_leak_debug_del(struct list_head *entry)
54 {
55         unsigned long flags;
56
57         spin_lock_irqsave(&leak_lock, flags);
58         list_del(entry);
59         spin_unlock_irqrestore(&leak_lock, flags);
60 }
61
62 static inline
63 void btrfs_leak_debug_check(void)
64 {
65         struct extent_state *state;
66         struct extent_buffer *eb;
67
68         while (!list_empty(&states)) {
69                 state = list_entry(states.next, struct extent_state, leak_list);
70                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
71                        state->start, state->end, state->state,
72                        extent_state_in_tree(state),
73                        refcount_read(&state->refs));
74                 list_del(&state->leak_list);
75                 kmem_cache_free(extent_state_cache, state);
76         }
77
78         while (!list_empty(&buffers)) {
79                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
80                 pr_err("BTRFS: buffer leak start %llu len %lu refs %d bflags %lu\n",
81                        eb->start, eb->len, atomic_read(&eb->refs), eb->bflags);
82                 list_del(&eb->leak_list);
83                 kmem_cache_free(extent_buffer_cache, eb);
84         }
85 }
86
87 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
88         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
89 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
90                 struct extent_io_tree *tree, u64 start, u64 end)
91 {
92         struct inode *inode = tree->private_data;
93         u64 isize;
94
95         if (!inode || !is_data_inode(inode))
96                 return;
97
98         isize = i_size_read(inode);
99         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
100                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
101                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
102                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
103         }
104 }
105 #else
106 #define btrfs_leak_debug_add(new, head) do {} while (0)
107 #define btrfs_leak_debug_del(entry)     do {} while (0)
108 #define btrfs_leak_debug_check()        do {} while (0)
109 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
110 #endif
111
112 #define BUFFER_LRU_MAX 64
113
114 struct tree_entry {
115         u64 start;
116         u64 end;
117         struct rb_node rb_node;
118 };
119
120 struct extent_page_data {
121         struct bio *bio;
122         struct extent_io_tree *tree;
123         /* tells writepage not to lock the state bits for this range
124          * it still does the unlocking
125          */
126         unsigned int extent_locked:1;
127
128         /* tells the submit_bio code to use REQ_SYNC */
129         unsigned int sync_io:1;
130 };
131
132 static int add_extent_changeset(struct extent_state *state, unsigned bits,
133                                  struct extent_changeset *changeset,
134                                  int set)
135 {
136         int ret;
137
138         if (!changeset)
139                 return 0;
140         if (set && (state->state & bits) == bits)
141                 return 0;
142         if (!set && (state->state & bits) == 0)
143                 return 0;
144         changeset->bytes_changed += state->end - state->start + 1;
145         ret = ulist_add(&changeset->range_changed, state->start, state->end,
146                         GFP_ATOMIC);
147         return ret;
148 }
149
150 static void flush_write_bio(struct extent_page_data *epd);
151
152 int __init extent_io_init(void)
153 {
154         extent_state_cache = kmem_cache_create("btrfs_extent_state",
155                         sizeof(struct extent_state), 0,
156                         SLAB_MEM_SPREAD, NULL);
157         if (!extent_state_cache)
158                 return -ENOMEM;
159
160         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
161                         sizeof(struct extent_buffer), 0,
162                         SLAB_MEM_SPREAD, NULL);
163         if (!extent_buffer_cache)
164                 goto free_state_cache;
165
166         if (bioset_init(&btrfs_bioset, BIO_POOL_SIZE,
167                         offsetof(struct btrfs_io_bio, bio),
168                         BIOSET_NEED_BVECS))
169                 goto free_buffer_cache;
170
171         if (bioset_integrity_create(&btrfs_bioset, BIO_POOL_SIZE))
172                 goto free_bioset;
173
174         return 0;
175
176 free_bioset:
177         bioset_exit(&btrfs_bioset);
178
179 free_buffer_cache:
180         kmem_cache_destroy(extent_buffer_cache);
181         extent_buffer_cache = NULL;
182
183 free_state_cache:
184         kmem_cache_destroy(extent_state_cache);
185         extent_state_cache = NULL;
186         return -ENOMEM;
187 }
188
189 void __cold extent_io_exit(void)
190 {
191         btrfs_leak_debug_check();
192
193         /*
194          * Make sure all delayed rcu free are flushed before we
195          * destroy caches.
196          */
197         rcu_barrier();
198         kmem_cache_destroy(extent_state_cache);
199         kmem_cache_destroy(extent_buffer_cache);
200         bioset_exit(&btrfs_bioset);
201 }
202
203 void extent_io_tree_init(struct extent_io_tree *tree,
204                          void *private_data)
205 {
206         tree->state = RB_ROOT;
207         tree->ops = NULL;
208         tree->dirty_bytes = 0;
209         spin_lock_init(&tree->lock);
210         tree->private_data = private_data;
211 }
212
213 static struct extent_state *alloc_extent_state(gfp_t mask)
214 {
215         struct extent_state *state;
216
217         /*
218          * The given mask might be not appropriate for the slab allocator,
219          * drop the unsupported bits
220          */
221         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
222         state = kmem_cache_alloc(extent_state_cache, mask);
223         if (!state)
224                 return state;
225         state->state = 0;
226         state->failrec = NULL;
227         RB_CLEAR_NODE(&state->rb_node);
228         btrfs_leak_debug_add(&state->leak_list, &states);
229         refcount_set(&state->refs, 1);
230         init_waitqueue_head(&state->wq);
231         trace_alloc_extent_state(state, mask, _RET_IP_);
232         return state;
233 }
234
235 void free_extent_state(struct extent_state *state)
236 {
237         if (!state)
238                 return;
239         if (refcount_dec_and_test(&state->refs)) {
240                 WARN_ON(extent_state_in_tree(state));
241                 btrfs_leak_debug_del(&state->leak_list);
242                 trace_free_extent_state(state, _RET_IP_);
243                 kmem_cache_free(extent_state_cache, state);
244         }
245 }
246
247 static struct rb_node *tree_insert(struct rb_root *root,
248                                    struct rb_node *search_start,
249                                    u64 offset,
250                                    struct rb_node *node,
251                                    struct rb_node ***p_in,
252                                    struct rb_node **parent_in)
253 {
254         struct rb_node **p;
255         struct rb_node *parent = NULL;
256         struct tree_entry *entry;
257
258         if (p_in && parent_in) {
259                 p = *p_in;
260                 parent = *parent_in;
261                 goto do_insert;
262         }
263
264         p = search_start ? &search_start : &root->rb_node;
265         while (*p) {
266                 parent = *p;
267                 entry = rb_entry(parent, struct tree_entry, rb_node);
268
269                 if (offset < entry->start)
270                         p = &(*p)->rb_left;
271                 else if (offset > entry->end)
272                         p = &(*p)->rb_right;
273                 else
274                         return parent;
275         }
276
277 do_insert:
278         rb_link_node(node, parent, p);
279         rb_insert_color(node, root);
280         return NULL;
281 }
282
283 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
284                                       struct rb_node **prev_ret,
285                                       struct rb_node **next_ret,
286                                       struct rb_node ***p_ret,
287                                       struct rb_node **parent_ret)
288 {
289         struct rb_root *root = &tree->state;
290         struct rb_node **n = &root->rb_node;
291         struct rb_node *prev = NULL;
292         struct rb_node *orig_prev = NULL;
293         struct tree_entry *entry;
294         struct tree_entry *prev_entry = NULL;
295
296         while (*n) {
297                 prev = *n;
298                 entry = rb_entry(prev, struct tree_entry, rb_node);
299                 prev_entry = entry;
300
301                 if (offset < entry->start)
302                         n = &(*n)->rb_left;
303                 else if (offset > entry->end)
304                         n = &(*n)->rb_right;
305                 else
306                         return *n;
307         }
308
309         if (p_ret)
310                 *p_ret = n;
311         if (parent_ret)
312                 *parent_ret = prev;
313
314         if (prev_ret) {
315                 orig_prev = prev;
316                 while (prev && offset > prev_entry->end) {
317                         prev = rb_next(prev);
318                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
319                 }
320                 *prev_ret = prev;
321                 prev = orig_prev;
322         }
323
324         if (next_ret) {
325                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
326                 while (prev && offset < prev_entry->start) {
327                         prev = rb_prev(prev);
328                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
329                 }
330                 *next_ret = prev;
331         }
332         return NULL;
333 }
334
335 static inline struct rb_node *
336 tree_search_for_insert(struct extent_io_tree *tree,
337                        u64 offset,
338                        struct rb_node ***p_ret,
339                        struct rb_node **parent_ret)
340 {
341         struct rb_node *prev = NULL;
342         struct rb_node *ret;
343
344         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
345         if (!ret)
346                 return prev;
347         return ret;
348 }
349
350 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
351                                           u64 offset)
352 {
353         return tree_search_for_insert(tree, offset, NULL, NULL);
354 }
355
356 /*
357  * utility function to look for merge candidates inside a given range.
358  * Any extents with matching state are merged together into a single
359  * extent in the tree.  Extents with EXTENT_IO in their state field
360  * are not merged because the end_io handlers need to be able to do
361  * operations on them without sleeping (or doing allocations/splits).
362  *
363  * This should be called with the tree lock held.
364  */
365 static void merge_state(struct extent_io_tree *tree,
366                         struct extent_state *state)
367 {
368         struct extent_state *other;
369         struct rb_node *other_node;
370
371         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
372                 return;
373
374         other_node = rb_prev(&state->rb_node);
375         if (other_node) {
376                 other = rb_entry(other_node, struct extent_state, rb_node);
377                 if (other->end == state->start - 1 &&
378                     other->state == state->state) {
379                         if (tree->private_data &&
380                             is_data_inode(tree->private_data))
381                                 btrfs_merge_delalloc_extent(tree->private_data,
382                                                             state, other);
383                         state->start = other->start;
384                         rb_erase(&other->rb_node, &tree->state);
385                         RB_CLEAR_NODE(&other->rb_node);
386                         free_extent_state(other);
387                 }
388         }
389         other_node = rb_next(&state->rb_node);
390         if (other_node) {
391                 other = rb_entry(other_node, struct extent_state, rb_node);
392                 if (other->start == state->end + 1 &&
393                     other->state == state->state) {
394                         if (tree->private_data &&
395                             is_data_inode(tree->private_data))
396                                 btrfs_merge_delalloc_extent(tree->private_data,
397                                                             state, other);
398                         state->end = other->end;
399                         rb_erase(&other->rb_node, &tree->state);
400                         RB_CLEAR_NODE(&other->rb_node);
401                         free_extent_state(other);
402                 }
403         }
404 }
405
406 static void set_state_bits(struct extent_io_tree *tree,
407                            struct extent_state *state, unsigned *bits,
408                            struct extent_changeset *changeset);
409
410 /*
411  * insert an extent_state struct into the tree.  'bits' are set on the
412  * struct before it is inserted.
413  *
414  * This may return -EEXIST if the extent is already there, in which case the
415  * state struct is freed.
416  *
417  * The tree lock is not taken internally.  This is a utility function and
418  * probably isn't what you want to call (see set/clear_extent_bit).
419  */
420 static int insert_state(struct extent_io_tree *tree,
421                         struct extent_state *state, u64 start, u64 end,
422                         struct rb_node ***p,
423                         struct rb_node **parent,
424                         unsigned *bits, struct extent_changeset *changeset)
425 {
426         struct rb_node *node;
427
428         if (end < start)
429                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
430                        end, start);
431         state->start = start;
432         state->end = end;
433
434         set_state_bits(tree, state, bits, changeset);
435
436         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
437         if (node) {
438                 struct extent_state *found;
439                 found = rb_entry(node, struct extent_state, rb_node);
440                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
441                        found->start, found->end, start, end);
442                 return -EEXIST;
443         }
444         merge_state(tree, state);
445         return 0;
446 }
447
448 /*
449  * split a given extent state struct in two, inserting the preallocated
450  * struct 'prealloc' as the newly created second half.  'split' indicates an
451  * offset inside 'orig' where it should be split.
452  *
453  * Before calling,
454  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
455  * are two extent state structs in the tree:
456  * prealloc: [orig->start, split - 1]
457  * orig: [ split, orig->end ]
458  *
459  * The tree locks are not taken by this function. They need to be held
460  * by the caller.
461  */
462 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
463                        struct extent_state *prealloc, u64 split)
464 {
465         struct rb_node *node;
466
467         if (tree->private_data && is_data_inode(tree->private_data))
468                 btrfs_split_delalloc_extent(tree->private_data, orig, split);
469
470         prealloc->start = orig->start;
471         prealloc->end = split - 1;
472         prealloc->state = orig->state;
473         orig->start = split;
474
475         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
476                            &prealloc->rb_node, NULL, NULL);
477         if (node) {
478                 free_extent_state(prealloc);
479                 return -EEXIST;
480         }
481         return 0;
482 }
483
484 static struct extent_state *next_state(struct extent_state *state)
485 {
486         struct rb_node *next = rb_next(&state->rb_node);
487         if (next)
488                 return rb_entry(next, struct extent_state, rb_node);
489         else
490                 return NULL;
491 }
492
493 /*
494  * utility function to clear some bits in an extent state struct.
495  * it will optionally wake up anyone waiting on this state (wake == 1).
496  *
497  * If no bits are set on the state struct after clearing things, the
498  * struct is freed and removed from the tree
499  */
500 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
501                                             struct extent_state *state,
502                                             unsigned *bits, int wake,
503                                             struct extent_changeset *changeset)
504 {
505         struct extent_state *next;
506         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
507         int ret;
508
509         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
510                 u64 range = state->end - state->start + 1;
511                 WARN_ON(range > tree->dirty_bytes);
512                 tree->dirty_bytes -= range;
513         }
514
515         if (tree->private_data && is_data_inode(tree->private_data))
516                 btrfs_clear_delalloc_extent(tree->private_data, state, bits);
517
518         ret = add_extent_changeset(state, bits_to_clear, changeset, 0);
519         BUG_ON(ret < 0);
520         state->state &= ~bits_to_clear;
521         if (wake)
522                 wake_up(&state->wq);
523         if (state->state == 0) {
524                 next = next_state(state);
525                 if (extent_state_in_tree(state)) {
526                         rb_erase(&state->rb_node, &tree->state);
527                         RB_CLEAR_NODE(&state->rb_node);
528                         free_extent_state(state);
529                 } else {
530                         WARN_ON(1);
531                 }
532         } else {
533                 merge_state(tree, state);
534                 next = next_state(state);
535         }
536         return next;
537 }
538
539 static struct extent_state *
540 alloc_extent_state_atomic(struct extent_state *prealloc)
541 {
542         if (!prealloc)
543                 prealloc = alloc_extent_state(GFP_ATOMIC);
544
545         return prealloc;
546 }
547
548 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
549 {
550         struct inode *inode = tree->private_data;
551
552         btrfs_panic(btrfs_sb(inode->i_sb), err,
553         "locking error: extent tree was modified by another thread while locked");
554 }
555
556 /*
557  * clear some bits on a range in the tree.  This may require splitting
558  * or inserting elements in the tree, so the gfp mask is used to
559  * indicate which allocations or sleeping are allowed.
560  *
561  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
562  * the given range from the tree regardless of state (ie for truncate).
563  *
564  * the range [start, end] is inclusive.
565  *
566  * This takes the tree lock, and returns 0 on success and < 0 on error.
567  */
568 int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
569                               unsigned bits, int wake, int delete,
570                               struct extent_state **cached_state,
571                               gfp_t mask, struct extent_changeset *changeset)
572 {
573         struct extent_state *state;
574         struct extent_state *cached;
575         struct extent_state *prealloc = NULL;
576         struct rb_node *node;
577         u64 last_end;
578         int err;
579         int clear = 0;
580
581         btrfs_debug_check_extent_io_range(tree, start, end);
582
583         if (bits & EXTENT_DELALLOC)
584                 bits |= EXTENT_NORESERVE;
585
586         if (delete)
587                 bits |= ~EXTENT_CTLBITS;
588
589         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
590                 clear = 1;
591 again:
592         if (!prealloc && gfpflags_allow_blocking(mask)) {
593                 /*
594                  * Don't care for allocation failure here because we might end
595                  * up not needing the pre-allocated extent state at all, which
596                  * is the case if we only have in the tree extent states that
597                  * cover our input range and don't cover too any other range.
598                  * If we end up needing a new extent state we allocate it later.
599                  */
600                 prealloc = alloc_extent_state(mask);
601         }
602
603         spin_lock(&tree->lock);
604         if (cached_state) {
605                 cached = *cached_state;
606
607                 if (clear) {
608                         *cached_state = NULL;
609                         cached_state = NULL;
610                 }
611
612                 if (cached && extent_state_in_tree(cached) &&
613                     cached->start <= start && cached->end > start) {
614                         if (clear)
615                                 refcount_dec(&cached->refs);
616                         state = cached;
617                         goto hit_next;
618                 }
619                 if (clear)
620                         free_extent_state(cached);
621         }
622         /*
623          * this search will find the extents that end after
624          * our range starts
625          */
626         node = tree_search(tree, start);
627         if (!node)
628                 goto out;
629         state = rb_entry(node, struct extent_state, rb_node);
630 hit_next:
631         if (state->start > end)
632                 goto out;
633         WARN_ON(state->end < start);
634         last_end = state->end;
635
636         /* the state doesn't have the wanted bits, go ahead */
637         if (!(state->state & bits)) {
638                 state = next_state(state);
639                 goto next;
640         }
641
642         /*
643          *     | ---- desired range ---- |
644          *  | state | or
645          *  | ------------- state -------------- |
646          *
647          * We need to split the extent we found, and may flip
648          * bits on second half.
649          *
650          * If the extent we found extends past our range, we
651          * just split and search again.  It'll get split again
652          * the next time though.
653          *
654          * If the extent we found is inside our range, we clear
655          * the desired bit on it.
656          */
657
658         if (state->start < start) {
659                 prealloc = alloc_extent_state_atomic(prealloc);
660                 BUG_ON(!prealloc);
661                 err = split_state(tree, state, prealloc, start);
662                 if (err)
663                         extent_io_tree_panic(tree, err);
664
665                 prealloc = NULL;
666                 if (err)
667                         goto out;
668                 if (state->end <= end) {
669                         state = clear_state_bit(tree, state, &bits, wake,
670                                                 changeset);
671                         goto next;
672                 }
673                 goto search_again;
674         }
675         /*
676          * | ---- desired range ---- |
677          *                        | state |
678          * We need to split the extent, and clear the bit
679          * on the first half
680          */
681         if (state->start <= end && state->end > end) {
682                 prealloc = alloc_extent_state_atomic(prealloc);
683                 BUG_ON(!prealloc);
684                 err = split_state(tree, state, prealloc, end + 1);
685                 if (err)
686                         extent_io_tree_panic(tree, err);
687
688                 if (wake)
689                         wake_up(&state->wq);
690
691                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
692
693                 prealloc = NULL;
694                 goto out;
695         }
696
697         state = clear_state_bit(tree, state, &bits, wake, changeset);
698 next:
699         if (last_end == (u64)-1)
700                 goto out;
701         start = last_end + 1;
702         if (start <= end && state && !need_resched())
703                 goto hit_next;
704
705 search_again:
706         if (start > end)
707                 goto out;
708         spin_unlock(&tree->lock);
709         if (gfpflags_allow_blocking(mask))
710                 cond_resched();
711         goto again;
712
713 out:
714         spin_unlock(&tree->lock);
715         if (prealloc)
716                 free_extent_state(prealloc);
717
718         return 0;
719
720 }
721
722 static void wait_on_state(struct extent_io_tree *tree,
723                           struct extent_state *state)
724                 __releases(tree->lock)
725                 __acquires(tree->lock)
726 {
727         DEFINE_WAIT(wait);
728         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
729         spin_unlock(&tree->lock);
730         schedule();
731         spin_lock(&tree->lock);
732         finish_wait(&state->wq, &wait);
733 }
734
735 /*
736  * waits for one or more bits to clear on a range in the state tree.
737  * The range [start, end] is inclusive.
738  * The tree lock is taken by this function
739  */
740 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
741                             unsigned long bits)
742 {
743         struct extent_state *state;
744         struct rb_node *node;
745
746         btrfs_debug_check_extent_io_range(tree, start, end);
747
748         spin_lock(&tree->lock);
749 again:
750         while (1) {
751                 /*
752                  * this search will find all the extents that end after
753                  * our range starts
754                  */
755                 node = tree_search(tree, start);
756 process_node:
757                 if (!node)
758                         break;
759
760                 state = rb_entry(node, struct extent_state, rb_node);
761
762                 if (state->start > end)
763                         goto out;
764
765                 if (state->state & bits) {
766                         start = state->start;
767                         refcount_inc(&state->refs);
768                         wait_on_state(tree, state);
769                         free_extent_state(state);
770                         goto again;
771                 }
772                 start = state->end + 1;
773
774                 if (start > end)
775                         break;
776
777                 if (!cond_resched_lock(&tree->lock)) {
778                         node = rb_next(node);
779                         goto process_node;
780                 }
781         }
782 out:
783         spin_unlock(&tree->lock);
784 }
785
786 static void set_state_bits(struct extent_io_tree *tree,
787                            struct extent_state *state,
788                            unsigned *bits, struct extent_changeset *changeset)
789 {
790         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
791         int ret;
792
793         if (tree->private_data && is_data_inode(tree->private_data))
794                 btrfs_set_delalloc_extent(tree->private_data, state, bits);
795
796         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
797                 u64 range = state->end - state->start + 1;
798                 tree->dirty_bytes += range;
799         }
800         ret = add_extent_changeset(state, bits_to_set, changeset, 1);
801         BUG_ON(ret < 0);
802         state->state |= bits_to_set;
803 }
804
805 static void cache_state_if_flags(struct extent_state *state,
806                                  struct extent_state **cached_ptr,
807                                  unsigned flags)
808 {
809         if (cached_ptr && !(*cached_ptr)) {
810                 if (!flags || (state->state & flags)) {
811                         *cached_ptr = state;
812                         refcount_inc(&state->refs);
813                 }
814         }
815 }
816
817 static void cache_state(struct extent_state *state,
818                         struct extent_state **cached_ptr)
819 {
820         return cache_state_if_flags(state, cached_ptr,
821                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
822 }
823
824 /*
825  * set some bits on a range in the tree.  This may require allocations or
826  * sleeping, so the gfp mask is used to indicate what is allowed.
827  *
828  * If any of the exclusive bits are set, this will fail with -EEXIST if some
829  * part of the range already has the desired bits set.  The start of the
830  * existing range is returned in failed_start in this case.
831  *
832  * [start, end] is inclusive This takes the tree lock.
833  */
834
835 static int __must_check
836 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
837                  unsigned bits, unsigned exclusive_bits,
838                  u64 *failed_start, struct extent_state **cached_state,
839                  gfp_t mask, struct extent_changeset *changeset)
840 {
841         struct extent_state *state;
842         struct extent_state *prealloc = NULL;
843         struct rb_node *node;
844         struct rb_node **p;
845         struct rb_node *parent;
846         int err = 0;
847         u64 last_start;
848         u64 last_end;
849
850         btrfs_debug_check_extent_io_range(tree, start, end);
851
852 again:
853         if (!prealloc && gfpflags_allow_blocking(mask)) {
854                 /*
855                  * Don't care for allocation failure here because we might end
856                  * up not needing the pre-allocated extent state at all, which
857                  * is the case if we only have in the tree extent states that
858                  * cover our input range and don't cover too any other range.
859                  * If we end up needing a new extent state we allocate it later.
860                  */
861                 prealloc = alloc_extent_state(mask);
862         }
863
864         spin_lock(&tree->lock);
865         if (cached_state && *cached_state) {
866                 state = *cached_state;
867                 if (state->start <= start && state->end > start &&
868                     extent_state_in_tree(state)) {
869                         node = &state->rb_node;
870                         goto hit_next;
871                 }
872         }
873         /*
874          * this search will find all the extents that end after
875          * our range starts.
876          */
877         node = tree_search_for_insert(tree, start, &p, &parent);
878         if (!node) {
879                 prealloc = alloc_extent_state_atomic(prealloc);
880                 BUG_ON(!prealloc);
881                 err = insert_state(tree, prealloc, start, end,
882                                    &p, &parent, &bits, changeset);
883                 if (err)
884                         extent_io_tree_panic(tree, err);
885
886                 cache_state(prealloc, cached_state);
887                 prealloc = NULL;
888                 goto out;
889         }
890         state = rb_entry(node, struct extent_state, rb_node);
891 hit_next:
892         last_start = state->start;
893         last_end = state->end;
894
895         /*
896          * | ---- desired range ---- |
897          * | state |
898          *
899          * Just lock what we found and keep going
900          */
901         if (state->start == start && state->end <= end) {
902                 if (state->state & exclusive_bits) {
903                         *failed_start = state->start;
904                         err = -EEXIST;
905                         goto out;
906                 }
907
908                 set_state_bits(tree, state, &bits, changeset);
909                 cache_state(state, cached_state);
910                 merge_state(tree, state);
911                 if (last_end == (u64)-1)
912                         goto out;
913                 start = last_end + 1;
914                 state = next_state(state);
915                 if (start < end && state && state->start == start &&
916                     !need_resched())
917                         goto hit_next;
918                 goto search_again;
919         }
920
921         /*
922          *     | ---- desired range ---- |
923          * | state |
924          *   or
925          * | ------------- state -------------- |
926          *
927          * We need to split the extent we found, and may flip bits on
928          * second half.
929          *
930          * If the extent we found extends past our
931          * range, we just split and search again.  It'll get split
932          * again the next time though.
933          *
934          * If the extent we found is inside our range, we set the
935          * desired bit on it.
936          */
937         if (state->start < start) {
938                 if (state->state & exclusive_bits) {
939                         *failed_start = start;
940                         err = -EEXIST;
941                         goto out;
942                 }
943
944                 prealloc = alloc_extent_state_atomic(prealloc);
945                 BUG_ON(!prealloc);
946                 err = split_state(tree, state, prealloc, start);
947                 if (err)
948                         extent_io_tree_panic(tree, err);
949
950                 prealloc = NULL;
951                 if (err)
952                         goto out;
953                 if (state->end <= end) {
954                         set_state_bits(tree, state, &bits, changeset);
955                         cache_state(state, cached_state);
956                         merge_state(tree, state);
957                         if (last_end == (u64)-1)
958                                 goto out;
959                         start = last_end + 1;
960                         state = next_state(state);
961                         if (start < end && state && state->start == start &&
962                             !need_resched())
963                                 goto hit_next;
964                 }
965                 goto search_again;
966         }
967         /*
968          * | ---- desired range ---- |
969          *     | state | or               | state |
970          *
971          * There's a hole, we need to insert something in it and
972          * ignore the extent we found.
973          */
974         if (state->start > start) {
975                 u64 this_end;
976                 if (end < last_start)
977                         this_end = end;
978                 else
979                         this_end = last_start - 1;
980
981                 prealloc = alloc_extent_state_atomic(prealloc);
982                 BUG_ON(!prealloc);
983
984                 /*
985                  * Avoid to free 'prealloc' if it can be merged with
986                  * the later extent.
987                  */
988                 err = insert_state(tree, prealloc, start, this_end,
989                                    NULL, NULL, &bits, changeset);
990                 if (err)
991                         extent_io_tree_panic(tree, err);
992
993                 cache_state(prealloc, cached_state);
994                 prealloc = NULL;
995                 start = this_end + 1;
996                 goto search_again;
997         }
998         /*
999          * | ---- desired range ---- |
1000          *                        | state |
1001          * We need to split the extent, and set the bit
1002          * on the first half
1003          */
1004         if (state->start <= end && state->end > end) {
1005                 if (state->state & exclusive_bits) {
1006                         *failed_start = start;
1007                         err = -EEXIST;
1008                         goto out;
1009                 }
1010
1011                 prealloc = alloc_extent_state_atomic(prealloc);
1012                 BUG_ON(!prealloc);
1013                 err = split_state(tree, state, prealloc, end + 1);
1014                 if (err)
1015                         extent_io_tree_panic(tree, err);
1016
1017                 set_state_bits(tree, prealloc, &bits, changeset);
1018                 cache_state(prealloc, cached_state);
1019                 merge_state(tree, prealloc);
1020                 prealloc = NULL;
1021                 goto out;
1022         }
1023
1024 search_again:
1025         if (start > end)
1026                 goto out;
1027         spin_unlock(&tree->lock);
1028         if (gfpflags_allow_blocking(mask))
1029                 cond_resched();
1030         goto again;
1031
1032 out:
1033         spin_unlock(&tree->lock);
1034         if (prealloc)
1035                 free_extent_state(prealloc);
1036
1037         return err;
1038
1039 }
1040
1041 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1042                    unsigned bits, u64 * failed_start,
1043                    struct extent_state **cached_state, gfp_t mask)
1044 {
1045         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1046                                 cached_state, mask, NULL);
1047 }
1048
1049
1050 /**
1051  * convert_extent_bit - convert all bits in a given range from one bit to
1052  *                      another
1053  * @tree:       the io tree to search
1054  * @start:      the start offset in bytes
1055  * @end:        the end offset in bytes (inclusive)
1056  * @bits:       the bits to set in this range
1057  * @clear_bits: the bits to clear in this range
1058  * @cached_state:       state that we're going to cache
1059  *
1060  * This will go through and set bits for the given range.  If any states exist
1061  * already in this range they are set with the given bit and cleared of the
1062  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1063  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1064  * boundary bits like LOCK.
1065  *
1066  * All allocations are done with GFP_NOFS.
1067  */
1068 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1069                        unsigned bits, unsigned clear_bits,
1070                        struct extent_state **cached_state)
1071 {
1072         struct extent_state *state;
1073         struct extent_state *prealloc = NULL;
1074         struct rb_node *node;
1075         struct rb_node **p;
1076         struct rb_node *parent;
1077         int err = 0;
1078         u64 last_start;
1079         u64 last_end;
1080         bool first_iteration = true;
1081
1082         btrfs_debug_check_extent_io_range(tree, start, end);
1083
1084 again:
1085         if (!prealloc) {
1086                 /*
1087                  * Best effort, don't worry if extent state allocation fails
1088                  * here for the first iteration. We might have a cached state
1089                  * that matches exactly the target range, in which case no
1090                  * extent state allocations are needed. We'll only know this
1091                  * after locking the tree.
1092                  */
1093                 prealloc = alloc_extent_state(GFP_NOFS);
1094                 if (!prealloc && !first_iteration)
1095                         return -ENOMEM;
1096         }
1097
1098         spin_lock(&tree->lock);
1099         if (cached_state && *cached_state) {
1100                 state = *cached_state;
1101                 if (state->start <= start && state->end > start &&
1102                     extent_state_in_tree(state)) {
1103                         node = &state->rb_node;
1104                         goto hit_next;
1105                 }
1106         }
1107
1108         /*
1109          * this search will find all the extents that end after
1110          * our range starts.
1111          */
1112         node = tree_search_for_insert(tree, start, &p, &parent);
1113         if (!node) {
1114                 prealloc = alloc_extent_state_atomic(prealloc);
1115                 if (!prealloc) {
1116                         err = -ENOMEM;
1117                         goto out;
1118                 }
1119                 err = insert_state(tree, prealloc, start, end,
1120                                    &p, &parent, &bits, NULL);
1121                 if (err)
1122                         extent_io_tree_panic(tree, err);
1123                 cache_state(prealloc, cached_state);
1124                 prealloc = NULL;
1125                 goto out;
1126         }
1127         state = rb_entry(node, struct extent_state, rb_node);
1128 hit_next:
1129         last_start = state->start;
1130         last_end = state->end;
1131
1132         /*
1133          * | ---- desired range ---- |
1134          * | state |
1135          *
1136          * Just lock what we found and keep going
1137          */
1138         if (state->start == start && state->end <= end) {
1139                 set_state_bits(tree, state, &bits, NULL);
1140                 cache_state(state, cached_state);
1141                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1142                 if (last_end == (u64)-1)
1143                         goto out;
1144                 start = last_end + 1;
1145                 if (start < end && state && state->start == start &&
1146                     !need_resched())
1147                         goto hit_next;
1148                 goto search_again;
1149         }
1150
1151         /*
1152          *     | ---- desired range ---- |
1153          * | state |
1154          *   or
1155          * | ------------- state -------------- |
1156          *
1157          * We need to split the extent we found, and may flip bits on
1158          * second half.
1159          *
1160          * If the extent we found extends past our
1161          * range, we just split and search again.  It'll get split
1162          * again the next time though.
1163          *
1164          * If the extent we found is inside our range, we set the
1165          * desired bit on it.
1166          */
1167         if (state->start < start) {
1168                 prealloc = alloc_extent_state_atomic(prealloc);
1169                 if (!prealloc) {
1170                         err = -ENOMEM;
1171                         goto out;
1172                 }
1173                 err = split_state(tree, state, prealloc, start);
1174                 if (err)
1175                         extent_io_tree_panic(tree, err);
1176                 prealloc = NULL;
1177                 if (err)
1178                         goto out;
1179                 if (state->end <= end) {
1180                         set_state_bits(tree, state, &bits, NULL);
1181                         cache_state(state, cached_state);
1182                         state = clear_state_bit(tree, state, &clear_bits, 0,
1183                                                 NULL);
1184                         if (last_end == (u64)-1)
1185                                 goto out;
1186                         start = last_end + 1;
1187                         if (start < end && state && state->start == start &&
1188                             !need_resched())
1189                                 goto hit_next;
1190                 }
1191                 goto search_again;
1192         }
1193         /*
1194          * | ---- desired range ---- |
1195          *     | state | or               | state |
1196          *
1197          * There's a hole, we need to insert something in it and
1198          * ignore the extent we found.
1199          */
1200         if (state->start > start) {
1201                 u64 this_end;
1202                 if (end < last_start)
1203                         this_end = end;
1204                 else
1205                         this_end = last_start - 1;
1206
1207                 prealloc = alloc_extent_state_atomic(prealloc);
1208                 if (!prealloc) {
1209                         err = -ENOMEM;
1210                         goto out;
1211                 }
1212
1213                 /*
1214                  * Avoid to free 'prealloc' if it can be merged with
1215                  * the later extent.
1216                  */
1217                 err = insert_state(tree, prealloc, start, this_end,
1218                                    NULL, NULL, &bits, NULL);
1219                 if (err)
1220                         extent_io_tree_panic(tree, err);
1221                 cache_state(prealloc, cached_state);
1222                 prealloc = NULL;
1223                 start = this_end + 1;
1224                 goto search_again;
1225         }
1226         /*
1227          * | ---- desired range ---- |
1228          *                        | state |
1229          * We need to split the extent, and set the bit
1230          * on the first half
1231          */
1232         if (state->start <= end && state->end > end) {
1233                 prealloc = alloc_extent_state_atomic(prealloc);
1234                 if (!prealloc) {
1235                         err = -ENOMEM;
1236                         goto out;
1237                 }
1238
1239                 err = split_state(tree, state, prealloc, end + 1);
1240                 if (err)
1241                         extent_io_tree_panic(tree, err);
1242
1243                 set_state_bits(tree, prealloc, &bits, NULL);
1244                 cache_state(prealloc, cached_state);
1245                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1246                 prealloc = NULL;
1247                 goto out;
1248         }
1249
1250 search_again:
1251         if (start > end)
1252                 goto out;
1253         spin_unlock(&tree->lock);
1254         cond_resched();
1255         first_iteration = false;
1256         goto again;
1257
1258 out:
1259         spin_unlock(&tree->lock);
1260         if (prealloc)
1261                 free_extent_state(prealloc);
1262
1263         return err;
1264 }
1265
1266 /* wrappers around set/clear extent bit */
1267 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1268                            unsigned bits, struct extent_changeset *changeset)
1269 {
1270         /*
1271          * We don't support EXTENT_LOCKED yet, as current changeset will
1272          * record any bits changed, so for EXTENT_LOCKED case, it will
1273          * either fail with -EEXIST or changeset will record the whole
1274          * range.
1275          */
1276         BUG_ON(bits & EXTENT_LOCKED);
1277
1278         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1279                                 changeset);
1280 }
1281
1282 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1283                      unsigned bits, int wake, int delete,
1284                      struct extent_state **cached)
1285 {
1286         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1287                                   cached, GFP_NOFS, NULL);
1288 }
1289
1290 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1291                 unsigned bits, struct extent_changeset *changeset)
1292 {
1293         /*
1294          * Don't support EXTENT_LOCKED case, same reason as
1295          * set_record_extent_bits().
1296          */
1297         BUG_ON(bits & EXTENT_LOCKED);
1298
1299         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1300                                   changeset);
1301 }
1302
1303 /*
1304  * either insert or lock state struct between start and end use mask to tell
1305  * us if waiting is desired.
1306  */
1307 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1308                      struct extent_state **cached_state)
1309 {
1310         int err;
1311         u64 failed_start;
1312
1313         while (1) {
1314                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1315                                        EXTENT_LOCKED, &failed_start,
1316                                        cached_state, GFP_NOFS, NULL);
1317                 if (err == -EEXIST) {
1318                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1319                         start = failed_start;
1320                 } else
1321                         break;
1322                 WARN_ON(start > end);
1323         }
1324         return err;
1325 }
1326
1327 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1328 {
1329         int err;
1330         u64 failed_start;
1331
1332         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1333                                &failed_start, NULL, GFP_NOFS, NULL);
1334         if (err == -EEXIST) {
1335                 if (failed_start > start)
1336                         clear_extent_bit(tree, start, failed_start - 1,
1337                                          EXTENT_LOCKED, 1, 0, NULL);
1338                 return 0;
1339         }
1340         return 1;
1341 }
1342
1343 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1344 {
1345         unsigned long index = start >> PAGE_SHIFT;
1346         unsigned long end_index = end >> PAGE_SHIFT;
1347         struct page *page;
1348
1349         while (index <= end_index) {
1350                 page = find_get_page(inode->i_mapping, index);
1351                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1352                 clear_page_dirty_for_io(page);
1353                 put_page(page);
1354                 index++;
1355         }
1356 }
1357
1358 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1359 {
1360         unsigned long index = start >> PAGE_SHIFT;
1361         unsigned long end_index = end >> PAGE_SHIFT;
1362         struct page *page;
1363
1364         while (index <= end_index) {
1365                 page = find_get_page(inode->i_mapping, index);
1366                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1367                 __set_page_dirty_nobuffers(page);
1368                 account_page_redirty(page);
1369                 put_page(page);
1370                 index++;
1371         }
1372 }
1373
1374 /* find the first state struct with 'bits' set after 'start', and
1375  * return it.  tree->lock must be held.  NULL will returned if
1376  * nothing was found after 'start'
1377  */
1378 static struct extent_state *
1379 find_first_extent_bit_state(struct extent_io_tree *tree,
1380                             u64 start, unsigned bits)
1381 {
1382         struct rb_node *node;
1383         struct extent_state *state;
1384
1385         /*
1386          * this search will find all the extents that end after
1387          * our range starts.
1388          */
1389         node = tree_search(tree, start);
1390         if (!node)
1391                 goto out;
1392
1393         while (1) {
1394                 state = rb_entry(node, struct extent_state, rb_node);
1395                 if (state->end >= start && (state->state & bits))
1396                         return state;
1397
1398                 node = rb_next(node);
1399                 if (!node)
1400                         break;
1401         }
1402 out:
1403         return NULL;
1404 }
1405
1406 /*
1407  * find the first offset in the io tree with 'bits' set. zero is
1408  * returned if we find something, and *start_ret and *end_ret are
1409  * set to reflect the state struct that was found.
1410  *
1411  * If nothing was found, 1 is returned. If found something, return 0.
1412  */
1413 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1414                           u64 *start_ret, u64 *end_ret, unsigned bits,
1415                           struct extent_state **cached_state)
1416 {
1417         struct extent_state *state;
1418         int ret = 1;
1419
1420         spin_lock(&tree->lock);
1421         if (cached_state && *cached_state) {
1422                 state = *cached_state;
1423                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1424                         while ((state = next_state(state)) != NULL) {
1425                                 if (state->state & bits)
1426                                         goto got_it;
1427                         }
1428                         free_extent_state(*cached_state);
1429                         *cached_state = NULL;
1430                         goto out;
1431                 }
1432                 free_extent_state(*cached_state);
1433                 *cached_state = NULL;
1434         }
1435
1436         state = find_first_extent_bit_state(tree, start, bits);
1437 got_it:
1438         if (state) {
1439                 cache_state_if_flags(state, cached_state, 0);
1440                 *start_ret = state->start;
1441                 *end_ret = state->end;
1442                 ret = 0;
1443         }
1444 out:
1445         spin_unlock(&tree->lock);
1446         return ret;
1447 }
1448
1449 /*
1450  * find a contiguous range of bytes in the file marked as delalloc, not
1451  * more than 'max_bytes'.  start and end are used to return the range,
1452  *
1453  * true is returned if we find something, false if nothing was in the tree
1454  */
1455 static noinline bool find_delalloc_range(struct extent_io_tree *tree,
1456                                         u64 *start, u64 *end, u64 max_bytes,
1457                                         struct extent_state **cached_state)
1458 {
1459         struct rb_node *node;
1460         struct extent_state *state;
1461         u64 cur_start = *start;
1462         bool found = false;
1463         u64 total_bytes = 0;
1464
1465         spin_lock(&tree->lock);
1466
1467         /*
1468          * this search will find all the extents that end after
1469          * our range starts.
1470          */
1471         node = tree_search(tree, cur_start);
1472         if (!node) {
1473                 *end = (u64)-1;
1474                 goto out;
1475         }
1476
1477         while (1) {
1478                 state = rb_entry(node, struct extent_state, rb_node);
1479                 if (found && (state->start != cur_start ||
1480                               (state->state & EXTENT_BOUNDARY))) {
1481                         goto out;
1482                 }
1483                 if (!(state->state & EXTENT_DELALLOC)) {
1484                         if (!found)
1485                                 *end = state->end;
1486                         goto out;
1487                 }
1488                 if (!found) {
1489                         *start = state->start;
1490                         *cached_state = state;
1491                         refcount_inc(&state->refs);
1492                 }
1493                 found = true;
1494                 *end = state->end;
1495                 cur_start = state->end + 1;
1496                 node = rb_next(node);
1497                 total_bytes += state->end - state->start + 1;
1498                 if (total_bytes >= max_bytes)
1499                         break;
1500                 if (!node)
1501                         break;
1502         }
1503 out:
1504         spin_unlock(&tree->lock);
1505         return found;
1506 }
1507
1508 static int __process_pages_contig(struct address_space *mapping,
1509                                   struct page *locked_page,
1510                                   pgoff_t start_index, pgoff_t end_index,
1511                                   unsigned long page_ops, pgoff_t *index_ret);
1512
1513 static noinline void __unlock_for_delalloc(struct inode *inode,
1514                                            struct page *locked_page,
1515                                            u64 start, u64 end)
1516 {
1517         unsigned long index = start >> PAGE_SHIFT;
1518         unsigned long end_index = end >> PAGE_SHIFT;
1519
1520         ASSERT(locked_page);
1521         if (index == locked_page->index && end_index == index)
1522                 return;
1523
1524         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1525                                PAGE_UNLOCK, NULL);
1526 }
1527
1528 static noinline int lock_delalloc_pages(struct inode *inode,
1529                                         struct page *locked_page,
1530                                         u64 delalloc_start,
1531                                         u64 delalloc_end)
1532 {
1533         unsigned long index = delalloc_start >> PAGE_SHIFT;
1534         unsigned long index_ret = index;
1535         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1536         int ret;
1537
1538         ASSERT(locked_page);
1539         if (index == locked_page->index && index == end_index)
1540                 return 0;
1541
1542         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1543                                      end_index, PAGE_LOCK, &index_ret);
1544         if (ret == -EAGAIN)
1545                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1546                                       (u64)index_ret << PAGE_SHIFT);
1547         return ret;
1548 }
1549
1550 /*
1551  * Find and lock a contiguous range of bytes in the file marked as delalloc, no
1552  * more than @max_bytes.  @Start and @end are used to return the range,
1553  *
1554  * Return: true if we find something
1555  *         false if nothing was in the tree
1556  */
1557 EXPORT_FOR_TESTS
1558 noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
1559                                     struct extent_io_tree *tree,
1560                                     struct page *locked_page, u64 *start,
1561                                     u64 *end)
1562 {
1563         u64 max_bytes = BTRFS_MAX_EXTENT_SIZE;
1564         u64 delalloc_start;
1565         u64 delalloc_end;
1566         bool found;
1567         struct extent_state *cached_state = NULL;
1568         int ret;
1569         int loops = 0;
1570
1571 again:
1572         /* step one, find a bunch of delalloc bytes starting at start */
1573         delalloc_start = *start;
1574         delalloc_end = 0;
1575         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1576                                     max_bytes, &cached_state);
1577         if (!found || delalloc_end <= *start) {
1578                 *start = delalloc_start;
1579                 *end = delalloc_end;
1580                 free_extent_state(cached_state);
1581                 return false;
1582         }
1583
1584         /*
1585          * start comes from the offset of locked_page.  We have to lock
1586          * pages in order, so we can't process delalloc bytes before
1587          * locked_page
1588          */
1589         if (delalloc_start < *start)
1590                 delalloc_start = *start;
1591
1592         /*
1593          * make sure to limit the number of pages we try to lock down
1594          */
1595         if (delalloc_end + 1 - delalloc_start > max_bytes)
1596                 delalloc_end = delalloc_start + max_bytes - 1;
1597
1598         /* step two, lock all the pages after the page that has start */
1599         ret = lock_delalloc_pages(inode, locked_page,
1600                                   delalloc_start, delalloc_end);
1601         ASSERT(!ret || ret == -EAGAIN);
1602         if (ret == -EAGAIN) {
1603                 /* some of the pages are gone, lets avoid looping by
1604                  * shortening the size of the delalloc range we're searching
1605                  */
1606                 free_extent_state(cached_state);
1607                 cached_state = NULL;
1608                 if (!loops) {
1609                         max_bytes = PAGE_SIZE;
1610                         loops = 1;
1611                         goto again;
1612                 } else {
1613                         found = false;
1614                         goto out_failed;
1615                 }
1616         }
1617
1618         /* step three, lock the state bits for the whole range */
1619         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1620
1621         /* then test to make sure it is all still delalloc */
1622         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1623                              EXTENT_DELALLOC, 1, cached_state);
1624         if (!ret) {
1625                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1626                                      &cached_state);
1627                 __unlock_for_delalloc(inode, locked_page,
1628                               delalloc_start, delalloc_end);
1629                 cond_resched();
1630                 goto again;
1631         }
1632         free_extent_state(cached_state);
1633         *start = delalloc_start;
1634         *end = delalloc_end;
1635 out_failed:
1636         return found;
1637 }
1638
1639 static int __process_pages_contig(struct address_space *mapping,
1640                                   struct page *locked_page,
1641                                   pgoff_t start_index, pgoff_t end_index,
1642                                   unsigned long page_ops, pgoff_t *index_ret)
1643 {
1644         unsigned long nr_pages = end_index - start_index + 1;
1645         unsigned long pages_locked = 0;
1646         pgoff_t index = start_index;
1647         struct page *pages[16];
1648         unsigned ret;
1649         int err = 0;
1650         int i;
1651
1652         if (page_ops & PAGE_LOCK) {
1653                 ASSERT(page_ops == PAGE_LOCK);
1654                 ASSERT(index_ret && *index_ret == start_index);
1655         }
1656
1657         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1658                 mapping_set_error(mapping, -EIO);
1659
1660         while (nr_pages > 0) {
1661                 ret = find_get_pages_contig(mapping, index,
1662                                      min_t(unsigned long,
1663                                      nr_pages, ARRAY_SIZE(pages)), pages);
1664                 if (ret == 0) {
1665                         /*
1666                          * Only if we're going to lock these pages,
1667                          * can we find nothing at @index.
1668                          */
1669                         ASSERT(page_ops & PAGE_LOCK);
1670                         err = -EAGAIN;
1671                         goto out;
1672                 }
1673
1674                 for (i = 0; i < ret; i++) {
1675                         if (page_ops & PAGE_SET_PRIVATE2)
1676                                 SetPagePrivate2(pages[i]);
1677
1678                         if (pages[i] == locked_page) {
1679                                 put_page(pages[i]);
1680                                 pages_locked++;
1681                                 continue;
1682                         }
1683                         if (page_ops & PAGE_CLEAR_DIRTY)
1684                                 clear_page_dirty_for_io(pages[i]);
1685                         if (page_ops & PAGE_SET_WRITEBACK)
1686                                 set_page_writeback(pages[i]);
1687                         if (page_ops & PAGE_SET_ERROR)
1688                                 SetPageError(pages[i]);
1689                         if (page_ops & PAGE_END_WRITEBACK)
1690                                 end_page_writeback(pages[i]);
1691                         if (page_ops & PAGE_UNLOCK)
1692                                 unlock_page(pages[i]);
1693                         if (page_ops & PAGE_LOCK) {
1694                                 lock_page(pages[i]);
1695                                 if (!PageDirty(pages[i]) ||
1696                                     pages[i]->mapping != mapping) {
1697                                         unlock_page(pages[i]);
1698                                         put_page(pages[i]);
1699                                         err = -EAGAIN;
1700                                         goto out;
1701                                 }
1702                         }
1703                         put_page(pages[i]);
1704                         pages_locked++;
1705                 }
1706                 nr_pages -= ret;
1707                 index += ret;
1708                 cond_resched();
1709         }
1710 out:
1711         if (err && index_ret)
1712                 *index_ret = start_index + pages_locked - 1;
1713         return err;
1714 }
1715
1716 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1717                                  u64 delalloc_end, struct page *locked_page,
1718                                  unsigned clear_bits,
1719                                  unsigned long page_ops)
1720 {
1721         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1722                          NULL);
1723
1724         __process_pages_contig(inode->i_mapping, locked_page,
1725                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1726                                page_ops, NULL);
1727 }
1728
1729 /*
1730  * count the number of bytes in the tree that have a given bit(s)
1731  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1732  * cached.  The total number found is returned.
1733  */
1734 u64 count_range_bits(struct extent_io_tree *tree,
1735                      u64 *start, u64 search_end, u64 max_bytes,
1736                      unsigned bits, int contig)
1737 {
1738         struct rb_node *node;
1739         struct extent_state *state;
1740         u64 cur_start = *start;
1741         u64 total_bytes = 0;
1742         u64 last = 0;
1743         int found = 0;
1744
1745         if (WARN_ON(search_end <= cur_start))
1746                 return 0;
1747
1748         spin_lock(&tree->lock);
1749         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1750                 total_bytes = tree->dirty_bytes;
1751                 goto out;
1752         }
1753         /*
1754          * this search will find all the extents that end after
1755          * our range starts.
1756          */
1757         node = tree_search(tree, cur_start);
1758         if (!node)
1759                 goto out;
1760
1761         while (1) {
1762                 state = rb_entry(node, struct extent_state, rb_node);
1763                 if (state->start > search_end)
1764                         break;
1765                 if (contig && found && state->start > last + 1)
1766                         break;
1767                 if (state->end >= cur_start && (state->state & bits) == bits) {
1768                         total_bytes += min(search_end, state->end) + 1 -
1769                                        max(cur_start, state->start);
1770                         if (total_bytes >= max_bytes)
1771                                 break;
1772                         if (!found) {
1773                                 *start = max(cur_start, state->start);
1774                                 found = 1;
1775                         }
1776                         last = state->end;
1777                 } else if (contig && found) {
1778                         break;
1779                 }
1780                 node = rb_next(node);
1781                 if (!node)
1782                         break;
1783         }
1784 out:
1785         spin_unlock(&tree->lock);
1786         return total_bytes;
1787 }
1788
1789 /*
1790  * set the private field for a given byte offset in the tree.  If there isn't
1791  * an extent_state there already, this does nothing.
1792  */
1793 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1794                 struct io_failure_record *failrec)
1795 {
1796         struct rb_node *node;
1797         struct extent_state *state;
1798         int ret = 0;
1799
1800         spin_lock(&tree->lock);
1801         /*
1802          * this search will find all the extents that end after
1803          * our range starts.
1804          */
1805         node = tree_search(tree, start);
1806         if (!node) {
1807                 ret = -ENOENT;
1808                 goto out;
1809         }
1810         state = rb_entry(node, struct extent_state, rb_node);
1811         if (state->start != start) {
1812                 ret = -ENOENT;
1813                 goto out;
1814         }
1815         state->failrec = failrec;
1816 out:
1817         spin_unlock(&tree->lock);
1818         return ret;
1819 }
1820
1821 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1822                 struct io_failure_record **failrec)
1823 {
1824         struct rb_node *node;
1825         struct extent_state *state;
1826         int ret = 0;
1827
1828         spin_lock(&tree->lock);
1829         /*
1830          * this search will find all the extents that end after
1831          * our range starts.
1832          */
1833         node = tree_search(tree, start);
1834         if (!node) {
1835                 ret = -ENOENT;
1836                 goto out;
1837         }
1838         state = rb_entry(node, struct extent_state, rb_node);
1839         if (state->start != start) {
1840                 ret = -ENOENT;
1841                 goto out;
1842         }
1843         *failrec = state->failrec;
1844 out:
1845         spin_unlock(&tree->lock);
1846         return ret;
1847 }
1848
1849 /*
1850  * searches a range in the state tree for a given mask.
1851  * If 'filled' == 1, this returns 1 only if every extent in the tree
1852  * has the bits set.  Otherwise, 1 is returned if any bit in the
1853  * range is found set.
1854  */
1855 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1856                    unsigned bits, int filled, struct extent_state *cached)
1857 {
1858         struct extent_state *state = NULL;
1859         struct rb_node *node;
1860         int bitset = 0;
1861
1862         spin_lock(&tree->lock);
1863         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1864             cached->end > start)
1865                 node = &cached->rb_node;
1866         else
1867                 node = tree_search(tree, start);
1868         while (node && start <= end) {
1869                 state = rb_entry(node, struct extent_state, rb_node);
1870
1871                 if (filled && state->start > start) {
1872                         bitset = 0;
1873                         break;
1874                 }
1875
1876                 if (state->start > end)
1877                         break;
1878
1879                 if (state->state & bits) {
1880                         bitset = 1;
1881                         if (!filled)
1882                                 break;
1883                 } else if (filled) {
1884                         bitset = 0;
1885                         break;
1886                 }
1887
1888                 if (state->end == (u64)-1)
1889                         break;
1890
1891                 start = state->end + 1;
1892                 if (start > end)
1893                         break;
1894                 node = rb_next(node);
1895                 if (!node) {
1896                         if (filled)
1897                                 bitset = 0;
1898                         break;
1899                 }
1900         }
1901         spin_unlock(&tree->lock);
1902         return bitset;
1903 }
1904
1905 /*
1906  * helper function to set a given page up to date if all the
1907  * extents in the tree for that page are up to date
1908  */
1909 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1910 {
1911         u64 start = page_offset(page);
1912         u64 end = start + PAGE_SIZE - 1;
1913         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1914                 SetPageUptodate(page);
1915 }
1916
1917 int free_io_failure(struct extent_io_tree *failure_tree,
1918                     struct extent_io_tree *io_tree,
1919                     struct io_failure_record *rec)
1920 {
1921         int ret;
1922         int err = 0;
1923
1924         set_state_failrec(failure_tree, rec->start, NULL);
1925         ret = clear_extent_bits(failure_tree, rec->start,
1926                                 rec->start + rec->len - 1,
1927                                 EXTENT_LOCKED | EXTENT_DIRTY);
1928         if (ret)
1929                 err = ret;
1930
1931         ret = clear_extent_bits(io_tree, rec->start,
1932                                 rec->start + rec->len - 1,
1933                                 EXTENT_DAMAGED);
1934         if (ret && !err)
1935                 err = ret;
1936
1937         kfree(rec);
1938         return err;
1939 }
1940
1941 /*
1942  * this bypasses the standard btrfs submit functions deliberately, as
1943  * the standard behavior is to write all copies in a raid setup. here we only
1944  * want to write the one bad copy. so we do the mapping for ourselves and issue
1945  * submit_bio directly.
1946  * to avoid any synchronization issues, wait for the data after writing, which
1947  * actually prevents the read that triggered the error from finishing.
1948  * currently, there can be no more than two copies of every data bit. thus,
1949  * exactly one rewrite is required.
1950  */
1951 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
1952                       u64 length, u64 logical, struct page *page,
1953                       unsigned int pg_offset, int mirror_num)
1954 {
1955         struct bio *bio;
1956         struct btrfs_device *dev;
1957         u64 map_length = 0;
1958         u64 sector;
1959         struct btrfs_bio *bbio = NULL;
1960         int ret;
1961
1962         ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
1963         BUG_ON(!mirror_num);
1964
1965         bio = btrfs_io_bio_alloc(1);
1966         bio->bi_iter.bi_size = 0;
1967         map_length = length;
1968
1969         /*
1970          * Avoid races with device replace and make sure our bbio has devices
1971          * associated to its stripes that don't go away while we are doing the
1972          * read repair operation.
1973          */
1974         btrfs_bio_counter_inc_blocked(fs_info);
1975         if (btrfs_is_parity_mirror(fs_info, logical, length)) {
1976                 /*
1977                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
1978                  * to update all raid stripes, but here we just want to correct
1979                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
1980                  * stripe's dev and sector.
1981                  */
1982                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
1983                                       &map_length, &bbio, 0);
1984                 if (ret) {
1985                         btrfs_bio_counter_dec(fs_info);
1986                         bio_put(bio);
1987                         return -EIO;
1988                 }
1989                 ASSERT(bbio->mirror_num == 1);
1990         } else {
1991                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
1992                                       &map_length, &bbio, mirror_num);
1993                 if (ret) {
1994                         btrfs_bio_counter_dec(fs_info);
1995                         bio_put(bio);
1996                         return -EIO;
1997                 }
1998                 BUG_ON(mirror_num != bbio->mirror_num);
1999         }
2000
2001         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2002         bio->bi_iter.bi_sector = sector;
2003         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2004         btrfs_put_bbio(bbio);
2005         if (!dev || !dev->bdev ||
2006             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2007                 btrfs_bio_counter_dec(fs_info);
2008                 bio_put(bio);
2009                 return -EIO;
2010         }
2011         bio_set_dev(bio, dev->bdev);
2012         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2013         bio_add_page(bio, page, length, pg_offset);
2014
2015         if (btrfsic_submit_bio_wait(bio)) {
2016                 /* try to remap that extent elsewhere? */
2017                 btrfs_bio_counter_dec(fs_info);
2018                 bio_put(bio);
2019                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2020                 return -EIO;
2021         }
2022
2023         btrfs_info_rl_in_rcu(fs_info,
2024                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2025                                   ino, start,
2026                                   rcu_str_deref(dev->name), sector);
2027         btrfs_bio_counter_dec(fs_info);
2028         bio_put(bio);
2029         return 0;
2030 }
2031
2032 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2033                          struct extent_buffer *eb, int mirror_num)
2034 {
2035         u64 start = eb->start;
2036         int i, num_pages = num_extent_pages(eb);
2037         int ret = 0;
2038
2039         if (sb_rdonly(fs_info->sb))
2040                 return -EROFS;
2041
2042         for (i = 0; i < num_pages; i++) {
2043                 struct page *p = eb->pages[i];
2044
2045                 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2046                                         start - page_offset(p), mirror_num);
2047                 if (ret)
2048                         break;
2049                 start += PAGE_SIZE;
2050         }
2051
2052         return ret;
2053 }
2054
2055 /*
2056  * each time an IO finishes, we do a fast check in the IO failure tree
2057  * to see if we need to process or clean up an io_failure_record
2058  */
2059 int clean_io_failure(struct btrfs_fs_info *fs_info,
2060                      struct extent_io_tree *failure_tree,
2061                      struct extent_io_tree *io_tree, u64 start,
2062                      struct page *page, u64 ino, unsigned int pg_offset)
2063 {
2064         u64 private;
2065         struct io_failure_record *failrec;
2066         struct extent_state *state;
2067         int num_copies;
2068         int ret;
2069
2070         private = 0;
2071         ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2072                                EXTENT_DIRTY, 0);
2073         if (!ret)
2074                 return 0;
2075
2076         ret = get_state_failrec(failure_tree, start, &failrec);
2077         if (ret)
2078                 return 0;
2079
2080         BUG_ON(!failrec->this_mirror);
2081
2082         if (failrec->in_validation) {
2083                 /* there was no real error, just free the record */
2084                 btrfs_debug(fs_info,
2085                         "clean_io_failure: freeing dummy error at %llu",
2086                         failrec->start);
2087                 goto out;
2088         }
2089         if (sb_rdonly(fs_info->sb))
2090                 goto out;
2091
2092         spin_lock(&io_tree->lock);
2093         state = find_first_extent_bit_state(io_tree,
2094                                             failrec->start,
2095                                             EXTENT_LOCKED);
2096         spin_unlock(&io_tree->lock);
2097
2098         if (state && state->start <= failrec->start &&
2099             state->end >= failrec->start + failrec->len - 1) {
2100                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2101                                               failrec->len);
2102                 if (num_copies > 1)  {
2103                         repair_io_failure(fs_info, ino, start, failrec->len,
2104                                           failrec->logical, page, pg_offset,
2105                                           failrec->failed_mirror);
2106                 }
2107         }
2108
2109 out:
2110         free_io_failure(failure_tree, io_tree, failrec);
2111
2112         return 0;
2113 }
2114
2115 /*
2116  * Can be called when
2117  * - hold extent lock
2118  * - under ordered extent
2119  * - the inode is freeing
2120  */
2121 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2122 {
2123         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2124         struct io_failure_record *failrec;
2125         struct extent_state *state, *next;
2126
2127         if (RB_EMPTY_ROOT(&failure_tree->state))
2128                 return;
2129
2130         spin_lock(&failure_tree->lock);
2131         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2132         while (state) {
2133                 if (state->start > end)
2134                         break;
2135
2136                 ASSERT(state->end <= end);
2137
2138                 next = next_state(state);
2139
2140                 failrec = state->failrec;
2141                 free_extent_state(state);
2142                 kfree(failrec);
2143
2144                 state = next;
2145         }
2146         spin_unlock(&failure_tree->lock);
2147 }
2148
2149 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2150                 struct io_failure_record **failrec_ret)
2151 {
2152         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2153         struct io_failure_record *failrec;
2154         struct extent_map *em;
2155         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2156         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2157         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2158         int ret;
2159         u64 logical;
2160
2161         ret = get_state_failrec(failure_tree, start, &failrec);
2162         if (ret) {
2163                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2164                 if (!failrec)
2165                         return -ENOMEM;
2166
2167                 failrec->start = start;
2168                 failrec->len = end - start + 1;
2169                 failrec->this_mirror = 0;
2170                 failrec->bio_flags = 0;
2171                 failrec->in_validation = 0;
2172
2173                 read_lock(&em_tree->lock);
2174                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2175                 if (!em) {
2176                         read_unlock(&em_tree->lock);
2177                         kfree(failrec);
2178                         return -EIO;
2179                 }
2180
2181                 if (em->start > start || em->start + em->len <= start) {
2182                         free_extent_map(em);
2183                         em = NULL;
2184                 }
2185                 read_unlock(&em_tree->lock);
2186                 if (!em) {
2187                         kfree(failrec);
2188                         return -EIO;
2189                 }
2190
2191                 logical = start - em->start;
2192                 logical = em->block_start + logical;
2193                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2194                         logical = em->block_start;
2195                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2196                         extent_set_compress_type(&failrec->bio_flags,
2197                                                  em->compress_type);
2198                 }
2199
2200                 btrfs_debug(fs_info,
2201                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2202                         logical, start, failrec->len);
2203
2204                 failrec->logical = logical;
2205                 free_extent_map(em);
2206
2207                 /* set the bits in the private failure tree */
2208                 ret = set_extent_bits(failure_tree, start, end,
2209                                         EXTENT_LOCKED | EXTENT_DIRTY);
2210                 if (ret >= 0)
2211                         ret = set_state_failrec(failure_tree, start, failrec);
2212                 /* set the bits in the inode's tree */
2213                 if (ret >= 0)
2214                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2215                 if (ret < 0) {
2216                         kfree(failrec);
2217                         return ret;
2218                 }
2219         } else {
2220                 btrfs_debug(fs_info,
2221                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2222                         failrec->logical, failrec->start, failrec->len,
2223                         failrec->in_validation);
2224                 /*
2225                  * when data can be on disk more than twice, add to failrec here
2226                  * (e.g. with a list for failed_mirror) to make
2227                  * clean_io_failure() clean all those errors at once.
2228                  */
2229         }
2230
2231         *failrec_ret = failrec;
2232
2233         return 0;
2234 }
2235
2236 bool btrfs_check_repairable(struct inode *inode, unsigned failed_bio_pages,
2237                            struct io_failure_record *failrec, int failed_mirror)
2238 {
2239         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2240         int num_copies;
2241
2242         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2243         if (num_copies == 1) {
2244                 /*
2245                  * we only have a single copy of the data, so don't bother with
2246                  * all the retry and error correction code that follows. no
2247                  * matter what the error is, it is very likely to persist.
2248                  */
2249                 btrfs_debug(fs_info,
2250                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2251                         num_copies, failrec->this_mirror, failed_mirror);
2252                 return false;
2253         }
2254
2255         /*
2256          * there are two premises:
2257          *      a) deliver good data to the caller
2258          *      b) correct the bad sectors on disk
2259          */
2260         if (failed_bio_pages > 1) {
2261                 /*
2262                  * to fulfill b), we need to know the exact failing sectors, as
2263                  * we don't want to rewrite any more than the failed ones. thus,
2264                  * we need separate read requests for the failed bio
2265                  *
2266                  * if the following BUG_ON triggers, our validation request got
2267                  * merged. we need separate requests for our algorithm to work.
2268                  */
2269                 BUG_ON(failrec->in_validation);
2270                 failrec->in_validation = 1;
2271                 failrec->this_mirror = failed_mirror;
2272         } else {
2273                 /*
2274                  * we're ready to fulfill a) and b) alongside. get a good copy
2275                  * of the failed sector and if we succeed, we have setup
2276                  * everything for repair_io_failure to do the rest for us.
2277                  */
2278                 if (failrec->in_validation) {
2279                         BUG_ON(failrec->this_mirror != failed_mirror);
2280                         failrec->in_validation = 0;
2281                         failrec->this_mirror = 0;
2282                 }
2283                 failrec->failed_mirror = failed_mirror;
2284                 failrec->this_mirror++;
2285                 if (failrec->this_mirror == failed_mirror)
2286                         failrec->this_mirror++;
2287         }
2288
2289         if (failrec->this_mirror > num_copies) {
2290                 btrfs_debug(fs_info,
2291                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2292                         num_copies, failrec->this_mirror, failed_mirror);
2293                 return false;
2294         }
2295
2296         return true;
2297 }
2298
2299
2300 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2301                                     struct io_failure_record *failrec,
2302                                     struct page *page, int pg_offset, int icsum,
2303                                     bio_end_io_t *endio_func, void *data)
2304 {
2305         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2306         struct bio *bio;
2307         struct btrfs_io_bio *btrfs_failed_bio;
2308         struct btrfs_io_bio *btrfs_bio;
2309
2310         bio = btrfs_io_bio_alloc(1);
2311         bio->bi_end_io = endio_func;
2312         bio->bi_iter.bi_sector = failrec->logical >> 9;
2313         bio_set_dev(bio, fs_info->fs_devices->latest_bdev);
2314         bio->bi_iter.bi_size = 0;
2315         bio->bi_private = data;
2316
2317         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2318         if (btrfs_failed_bio->csum) {
2319                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2320
2321                 btrfs_bio = btrfs_io_bio(bio);
2322                 btrfs_bio->csum = btrfs_bio->csum_inline;
2323                 icsum *= csum_size;
2324                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2325                        csum_size);
2326         }
2327
2328         bio_add_page(bio, page, failrec->len, pg_offset);
2329
2330         return bio;
2331 }
2332
2333 /*
2334  * This is a generic handler for readpage errors. If other copies exist, read
2335  * those and write back good data to the failed position. Does not investigate
2336  * in remapping the failed extent elsewhere, hoping the device will be smart
2337  * enough to do this as needed
2338  */
2339 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2340                               struct page *page, u64 start, u64 end,
2341                               int failed_mirror)
2342 {
2343         struct io_failure_record *failrec;
2344         struct inode *inode = page->mapping->host;
2345         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2346         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2347         struct bio *bio;
2348         int read_mode = 0;
2349         blk_status_t status;
2350         int ret;
2351         unsigned failed_bio_pages = bio_pages_all(failed_bio);
2352
2353         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2354
2355         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2356         if (ret)
2357                 return ret;
2358
2359         if (!btrfs_check_repairable(inode, failed_bio_pages, failrec,
2360                                     failed_mirror)) {
2361                 free_io_failure(failure_tree, tree, failrec);
2362                 return -EIO;
2363         }
2364
2365         if (failed_bio_pages > 1)
2366                 read_mode |= REQ_FAILFAST_DEV;
2367
2368         phy_offset >>= inode->i_sb->s_blocksize_bits;
2369         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2370                                       start - page_offset(page),
2371                                       (int)phy_offset, failed_bio->bi_end_io,
2372                                       NULL);
2373         bio->bi_opf = REQ_OP_READ | read_mode;
2374
2375         btrfs_debug(btrfs_sb(inode->i_sb),
2376                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2377                 read_mode, failrec->this_mirror, failrec->in_validation);
2378
2379         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2380                                          failrec->bio_flags, 0);
2381         if (status) {
2382                 free_io_failure(failure_tree, tree, failrec);
2383                 bio_put(bio);
2384                 ret = blk_status_to_errno(status);
2385         }
2386
2387         return ret;
2388 }
2389
2390 /* lots and lots of room for performance fixes in the end_bio funcs */
2391
2392 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2393 {
2394         int uptodate = (err == 0);
2395         int ret = 0;
2396
2397         btrfs_writepage_endio_finish_ordered(page, start, end, uptodate);
2398
2399         if (!uptodate) {
2400                 ClearPageUptodate(page);
2401                 SetPageError(page);
2402                 ret = err < 0 ? err : -EIO;
2403                 mapping_set_error(page->mapping, ret);
2404         }
2405 }
2406
2407 /*
2408  * after a writepage IO is done, we need to:
2409  * clear the uptodate bits on error
2410  * clear the writeback bits in the extent tree for this IO
2411  * end_page_writeback if the page has no more pending IO
2412  *
2413  * Scheduling is not allowed, so the extent state tree is expected
2414  * to have one and only one object corresponding to this IO.
2415  */
2416 static void end_bio_extent_writepage(struct bio *bio)
2417 {
2418         int error = blk_status_to_errno(bio->bi_status);
2419         struct bio_vec *bvec;
2420         u64 start;
2421         u64 end;
2422         int i;
2423
2424         ASSERT(!bio_flagged(bio, BIO_CLONED));
2425         bio_for_each_segment_all(bvec, bio, i) {
2426                 struct page *page = bvec->bv_page;
2427                 struct inode *inode = page->mapping->host;
2428                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2429
2430                 /* We always issue full-page reads, but if some block
2431                  * in a page fails to read, blk_update_request() will
2432                  * advance bv_offset and adjust bv_len to compensate.
2433                  * Print a warning for nonzero offsets, and an error
2434                  * if they don't add up to a full page.  */
2435                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2436                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2437                                 btrfs_err(fs_info,
2438                                    "partial page write in btrfs with offset %u and length %u",
2439                                         bvec->bv_offset, bvec->bv_len);
2440                         else
2441                                 btrfs_info(fs_info,
2442                                    "incomplete page write in btrfs with offset %u and length %u",
2443                                         bvec->bv_offset, bvec->bv_len);
2444                 }
2445
2446                 start = page_offset(page);
2447                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2448
2449                 end_extent_writepage(page, error, start, end);
2450                 end_page_writeback(page);
2451         }
2452
2453         bio_put(bio);
2454 }
2455
2456 static void
2457 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2458                               int uptodate)
2459 {
2460         struct extent_state *cached = NULL;
2461         u64 end = start + len - 1;
2462
2463         if (uptodate && tree->track_uptodate)
2464                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2465         unlock_extent_cached_atomic(tree, start, end, &cached);
2466 }
2467
2468 /*
2469  * after a readpage IO is done, we need to:
2470  * clear the uptodate bits on error
2471  * set the uptodate bits if things worked
2472  * set the page up to date if all extents in the tree are uptodate
2473  * clear the lock bit in the extent tree
2474  * unlock the page if there are no other extents locked for it
2475  *
2476  * Scheduling is not allowed, so the extent state tree is expected
2477  * to have one and only one object corresponding to this IO.
2478  */
2479 static void end_bio_extent_readpage(struct bio *bio)
2480 {
2481         struct bio_vec *bvec;
2482         int uptodate = !bio->bi_status;
2483         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2484         struct extent_io_tree *tree, *failure_tree;
2485         u64 offset = 0;
2486         u64 start;
2487         u64 end;
2488         u64 len;
2489         u64 extent_start = 0;
2490         u64 extent_len = 0;
2491         int mirror;
2492         int ret;
2493         int i;
2494
2495         ASSERT(!bio_flagged(bio, BIO_CLONED));
2496         bio_for_each_segment_all(bvec, bio, i) {
2497                 struct page *page = bvec->bv_page;
2498                 struct inode *inode = page->mapping->host;
2499                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2500                 bool data_inode = btrfs_ino(BTRFS_I(inode))
2501                         != BTRFS_BTREE_INODE_OBJECTID;
2502
2503                 btrfs_debug(fs_info,
2504                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2505                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2506                         io_bio->mirror_num);
2507                 tree = &BTRFS_I(inode)->io_tree;
2508                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2509
2510                 /* We always issue full-page reads, but if some block
2511                  * in a page fails to read, blk_update_request() will
2512                  * advance bv_offset and adjust bv_len to compensate.
2513                  * Print a warning for nonzero offsets, and an error
2514                  * if they don't add up to a full page.  */
2515                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2516                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2517                                 btrfs_err(fs_info,
2518                                         "partial page read in btrfs with offset %u and length %u",
2519                                         bvec->bv_offset, bvec->bv_len);
2520                         else
2521                                 btrfs_info(fs_info,
2522                                         "incomplete page read in btrfs with offset %u and length %u",
2523                                         bvec->bv_offset, bvec->bv_len);
2524                 }
2525
2526                 start = page_offset(page);
2527                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2528                 len = bvec->bv_len;
2529
2530                 mirror = io_bio->mirror_num;
2531                 if (likely(uptodate)) {
2532                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2533                                                               page, start, end,
2534                                                               mirror);
2535                         if (ret)
2536                                 uptodate = 0;
2537                         else
2538                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2539                                                  failure_tree, tree, start,
2540                                                  page,
2541                                                  btrfs_ino(BTRFS_I(inode)), 0);
2542                 }
2543
2544                 if (likely(uptodate))
2545                         goto readpage_ok;
2546
2547                 if (data_inode) {
2548
2549                         /*
2550                          * The generic bio_readpage_error handles errors the
2551                          * following way: If possible, new read requests are
2552                          * created and submitted and will end up in
2553                          * end_bio_extent_readpage as well (if we're lucky,
2554                          * not in the !uptodate case). In that case it returns
2555                          * 0 and we just go on with the next page in our bio.
2556                          * If it can't handle the error it will return -EIO and
2557                          * we remain responsible for that page.
2558                          */
2559                         ret = bio_readpage_error(bio, offset, page, start, end,
2560                                                  mirror);
2561                         if (ret == 0) {
2562                                 uptodate = !bio->bi_status;
2563                                 offset += len;
2564                                 continue;
2565                         }
2566                 } else {
2567                         struct extent_buffer *eb;
2568
2569                         eb = (struct extent_buffer *)page->private;
2570                         set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
2571                         eb->read_mirror = mirror;
2572                         atomic_dec(&eb->io_pages);
2573                         if (test_and_clear_bit(EXTENT_BUFFER_READAHEAD,
2574                                                &eb->bflags))
2575                                 btree_readahead_hook(eb, -EIO);
2576
2577                         ret = -EIO;
2578                 }
2579 readpage_ok:
2580                 if (likely(uptodate)) {
2581                         loff_t i_size = i_size_read(inode);
2582                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2583                         unsigned off;
2584
2585                         /* Zero out the end if this page straddles i_size */
2586                         off = offset_in_page(i_size);
2587                         if (page->index == end_index && off)
2588                                 zero_user_segment(page, off, PAGE_SIZE);
2589                         SetPageUptodate(page);
2590                 } else {
2591                         ClearPageUptodate(page);
2592                         SetPageError(page);
2593                 }
2594                 unlock_page(page);
2595                 offset += len;
2596
2597                 if (unlikely(!uptodate)) {
2598                         if (extent_len) {
2599                                 endio_readpage_release_extent(tree,
2600                                                               extent_start,
2601                                                               extent_len, 1);
2602                                 extent_start = 0;
2603                                 extent_len = 0;
2604                         }
2605                         endio_readpage_release_extent(tree, start,
2606                                                       end - start + 1, 0);
2607                 } else if (!extent_len) {
2608                         extent_start = start;
2609                         extent_len = end + 1 - start;
2610                 } else if (extent_start + extent_len == start) {
2611                         extent_len += end + 1 - start;
2612                 } else {
2613                         endio_readpage_release_extent(tree, extent_start,
2614                                                       extent_len, uptodate);
2615                         extent_start = start;
2616                         extent_len = end + 1 - start;
2617                 }
2618         }
2619
2620         if (extent_len)
2621                 endio_readpage_release_extent(tree, extent_start, extent_len,
2622                                               uptodate);
2623         btrfs_io_bio_free_csum(io_bio);
2624         bio_put(bio);
2625 }
2626
2627 /*
2628  * Initialize the members up to but not including 'bio'. Use after allocating a
2629  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2630  * 'bio' because use of __GFP_ZERO is not supported.
2631  */
2632 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2633 {
2634         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2635 }
2636
2637 /*
2638  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2639  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2640  * for the appropriate container_of magic
2641  */
2642 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte)
2643 {
2644         struct bio *bio;
2645
2646         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, &btrfs_bioset);
2647         bio_set_dev(bio, bdev);
2648         bio->bi_iter.bi_sector = first_byte >> 9;
2649         btrfs_io_bio_init(btrfs_io_bio(bio));
2650         return bio;
2651 }
2652
2653 struct bio *btrfs_bio_clone(struct bio *bio)
2654 {
2655         struct btrfs_io_bio *btrfs_bio;
2656         struct bio *new;
2657
2658         /* Bio allocation backed by a bioset does not fail */
2659         new = bio_clone_fast(bio, GFP_NOFS, &btrfs_bioset);
2660         btrfs_bio = btrfs_io_bio(new);
2661         btrfs_io_bio_init(btrfs_bio);
2662         btrfs_bio->iter = bio->bi_iter;
2663         return new;
2664 }
2665
2666 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2667 {
2668         struct bio *bio;
2669
2670         /* Bio allocation backed by a bioset does not fail */
2671         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, &btrfs_bioset);
2672         btrfs_io_bio_init(btrfs_io_bio(bio));
2673         return bio;
2674 }
2675
2676 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2677 {
2678         struct bio *bio;
2679         struct btrfs_io_bio *btrfs_bio;
2680
2681         /* this will never fail when it's backed by a bioset */
2682         bio = bio_clone_fast(orig, GFP_NOFS, &btrfs_bioset);
2683         ASSERT(bio);
2684
2685         btrfs_bio = btrfs_io_bio(bio);
2686         btrfs_io_bio_init(btrfs_bio);
2687
2688         bio_trim(bio, offset >> 9, size >> 9);
2689         btrfs_bio->iter = bio->bi_iter;
2690         return bio;
2691 }
2692
2693 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2694                                        unsigned long bio_flags)
2695 {
2696         blk_status_t ret = 0;
2697         struct bio_vec *bvec = bio_last_bvec_all(bio);
2698         struct page *page = bvec->bv_page;
2699         struct extent_io_tree *tree = bio->bi_private;
2700         u64 start;
2701
2702         start = page_offset(page) + bvec->bv_offset;
2703
2704         bio->bi_private = NULL;
2705
2706         if (tree->ops)
2707                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
2708                                            mirror_num, bio_flags, start);
2709         else
2710                 btrfsic_submit_bio(bio);
2711
2712         return blk_status_to_errno(ret);
2713 }
2714
2715 /*
2716  * @opf:        bio REQ_OP_* and REQ_* flags as one value
2717  * @tree:       tree so we can call our merge_bio hook
2718  * @wbc:        optional writeback control for io accounting
2719  * @page:       page to add to the bio
2720  * @pg_offset:  offset of the new bio or to check whether we are adding
2721  *              a contiguous page to the previous one
2722  * @size:       portion of page that we want to write
2723  * @offset:     starting offset in the page
2724  * @bdev:       attach newly created bios to this bdev
2725  * @bio_ret:    must be valid pointer, newly allocated bio will be stored there
2726  * @end_io_func:     end_io callback for new bio
2727  * @mirror_num:      desired mirror to read/write
2728  * @prev_bio_flags:  flags of previous bio to see if we can merge the current one
2729  * @bio_flags:  flags of the current bio to see if we can merge them
2730  */
2731 static int submit_extent_page(unsigned int opf, struct extent_io_tree *tree,
2732                               struct writeback_control *wbc,
2733                               struct page *page, u64 offset,
2734                               size_t size, unsigned long pg_offset,
2735                               struct block_device *bdev,
2736                               struct bio **bio_ret,
2737                               bio_end_io_t end_io_func,
2738                               int mirror_num,
2739                               unsigned long prev_bio_flags,
2740                               unsigned long bio_flags,
2741                               bool force_bio_submit)
2742 {
2743         int ret = 0;
2744         struct bio *bio;
2745         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2746         sector_t sector = offset >> 9;
2747
2748         ASSERT(bio_ret);
2749
2750         if (*bio_ret) {
2751                 bool contig;
2752                 bool can_merge = true;
2753
2754                 bio = *bio_ret;
2755                 if (prev_bio_flags & EXTENT_BIO_COMPRESSED)
2756                         contig = bio->bi_iter.bi_sector == sector;
2757                 else
2758                         contig = bio_end_sector(bio) == sector;
2759
2760                 ASSERT(tree->ops);
2761                 if (btrfs_bio_fits_in_stripe(page, page_size, bio, bio_flags))
2762                         can_merge = false;
2763
2764                 if (prev_bio_flags != bio_flags || !contig || !can_merge ||
2765                     force_bio_submit ||
2766                     bio_add_page(bio, page, page_size, pg_offset) < page_size) {
2767                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2768                         if (ret < 0) {
2769                                 *bio_ret = NULL;
2770                                 return ret;
2771                         }
2772                         bio = NULL;
2773                 } else {
2774                         if (wbc)
2775                                 wbc_account_io(wbc, page, page_size);
2776                         return 0;
2777                 }
2778         }
2779
2780         bio = btrfs_bio_alloc(bdev, offset);
2781         bio_add_page(bio, page, page_size, pg_offset);
2782         bio->bi_end_io = end_io_func;
2783         bio->bi_private = tree;
2784         bio->bi_write_hint = page->mapping->host->i_write_hint;
2785         bio->bi_opf = opf;
2786         if (wbc) {
2787                 wbc_init_bio(wbc, bio);
2788                 wbc_account_io(wbc, page, page_size);
2789         }
2790
2791         *bio_ret = bio;
2792
2793         return ret;
2794 }
2795
2796 static void attach_extent_buffer_page(struct extent_buffer *eb,
2797                                       struct page *page)
2798 {
2799         if (!PagePrivate(page)) {
2800                 SetPagePrivate(page);
2801                 get_page(page);
2802                 set_page_private(page, (unsigned long)eb);
2803         } else {
2804                 WARN_ON(page->private != (unsigned long)eb);
2805         }
2806 }
2807
2808 void set_page_extent_mapped(struct page *page)
2809 {
2810         if (!PagePrivate(page)) {
2811                 SetPagePrivate(page);
2812                 get_page(page);
2813                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2814         }
2815 }
2816
2817 static struct extent_map *
2818 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2819                  u64 start, u64 len, get_extent_t *get_extent,
2820                  struct extent_map **em_cached)
2821 {
2822         struct extent_map *em;
2823
2824         if (em_cached && *em_cached) {
2825                 em = *em_cached;
2826                 if (extent_map_in_tree(em) && start >= em->start &&
2827                     start < extent_map_end(em)) {
2828                         refcount_inc(&em->refs);
2829                         return em;
2830                 }
2831
2832                 free_extent_map(em);
2833                 *em_cached = NULL;
2834         }
2835
2836         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2837         if (em_cached && !IS_ERR_OR_NULL(em)) {
2838                 BUG_ON(*em_cached);
2839                 refcount_inc(&em->refs);
2840                 *em_cached = em;
2841         }
2842         return em;
2843 }
2844 /*
2845  * basic readpage implementation.  Locked extent state structs are inserted
2846  * into the tree that are removed when the IO is done (by the end_io
2847  * handlers)
2848  * XXX JDM: This needs looking at to ensure proper page locking
2849  * return 0 on success, otherwise return error
2850  */
2851 static int __do_readpage(struct extent_io_tree *tree,
2852                          struct page *page,
2853                          get_extent_t *get_extent,
2854                          struct extent_map **em_cached,
2855                          struct bio **bio, int mirror_num,
2856                          unsigned long *bio_flags, unsigned int read_flags,
2857                          u64 *prev_em_start)
2858 {
2859         struct inode *inode = page->mapping->host;
2860         u64 start = page_offset(page);
2861         const u64 end = start + PAGE_SIZE - 1;
2862         u64 cur = start;
2863         u64 extent_offset;
2864         u64 last_byte = i_size_read(inode);
2865         u64 block_start;
2866         u64 cur_end;
2867         struct extent_map *em;
2868         struct block_device *bdev;
2869         int ret = 0;
2870         int nr = 0;
2871         size_t pg_offset = 0;
2872         size_t iosize;
2873         size_t disk_io_size;
2874         size_t blocksize = inode->i_sb->s_blocksize;
2875         unsigned long this_bio_flag = 0;
2876
2877         set_page_extent_mapped(page);
2878
2879         if (!PageUptodate(page)) {
2880                 if (cleancache_get_page(page) == 0) {
2881                         BUG_ON(blocksize != PAGE_SIZE);
2882                         unlock_extent(tree, start, end);
2883                         goto out;
2884                 }
2885         }
2886
2887         if (page->index == last_byte >> PAGE_SHIFT) {
2888                 char *userpage;
2889                 size_t zero_offset = offset_in_page(last_byte);
2890
2891                 if (zero_offset) {
2892                         iosize = PAGE_SIZE - zero_offset;
2893                         userpage = kmap_atomic(page);
2894                         memset(userpage + zero_offset, 0, iosize);
2895                         flush_dcache_page(page);
2896                         kunmap_atomic(userpage);
2897                 }
2898         }
2899         while (cur <= end) {
2900                 bool force_bio_submit = false;
2901                 u64 offset;
2902
2903                 if (cur >= last_byte) {
2904                         char *userpage;
2905                         struct extent_state *cached = NULL;
2906
2907                         iosize = PAGE_SIZE - pg_offset;
2908                         userpage = kmap_atomic(page);
2909                         memset(userpage + pg_offset, 0, iosize);
2910                         flush_dcache_page(page);
2911                         kunmap_atomic(userpage);
2912                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2913                                             &cached, GFP_NOFS);
2914                         unlock_extent_cached(tree, cur,
2915                                              cur + iosize - 1, &cached);
2916                         break;
2917                 }
2918                 em = __get_extent_map(inode, page, pg_offset, cur,
2919                                       end - cur + 1, get_extent, em_cached);
2920                 if (IS_ERR_OR_NULL(em)) {
2921                         SetPageError(page);
2922                         unlock_extent(tree, cur, end);
2923                         break;
2924                 }
2925                 extent_offset = cur - em->start;
2926                 BUG_ON(extent_map_end(em) <= cur);
2927                 BUG_ON(end < cur);
2928
2929                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2930                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2931                         extent_set_compress_type(&this_bio_flag,
2932                                                  em->compress_type);
2933                 }
2934
2935                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2936                 cur_end = min(extent_map_end(em) - 1, end);
2937                 iosize = ALIGN(iosize, blocksize);
2938                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2939                         disk_io_size = em->block_len;
2940                         offset = em->block_start;
2941                 } else {
2942                         offset = em->block_start + extent_offset;
2943                         disk_io_size = iosize;
2944                 }
2945                 bdev = em->bdev;
2946                 block_start = em->block_start;
2947                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2948                         block_start = EXTENT_MAP_HOLE;
2949
2950                 /*
2951                  * If we have a file range that points to a compressed extent
2952                  * and it's followed by a consecutive file range that points to
2953                  * to the same compressed extent (possibly with a different
2954                  * offset and/or length, so it either points to the whole extent
2955                  * or only part of it), we must make sure we do not submit a
2956                  * single bio to populate the pages for the 2 ranges because
2957                  * this makes the compressed extent read zero out the pages
2958                  * belonging to the 2nd range. Imagine the following scenario:
2959                  *
2960                  *  File layout
2961                  *  [0 - 8K]                     [8K - 24K]
2962                  *    |                               |
2963                  *    |                               |
2964                  * points to extent X,         points to extent X,
2965                  * offset 4K, length of 8K     offset 0, length 16K
2966                  *
2967                  * [extent X, compressed length = 4K uncompressed length = 16K]
2968                  *
2969                  * If the bio to read the compressed extent covers both ranges,
2970                  * it will decompress extent X into the pages belonging to the
2971                  * first range and then it will stop, zeroing out the remaining
2972                  * pages that belong to the other range that points to extent X.
2973                  * So here we make sure we submit 2 bios, one for the first
2974                  * range and another one for the third range. Both will target
2975                  * the same physical extent from disk, but we can't currently
2976                  * make the compressed bio endio callback populate the pages
2977                  * for both ranges because each compressed bio is tightly
2978                  * coupled with a single extent map, and each range can have
2979                  * an extent map with a different offset value relative to the
2980                  * uncompressed data of our extent and different lengths. This
2981                  * is a corner case so we prioritize correctness over
2982                  * non-optimal behavior (submitting 2 bios for the same extent).
2983                  */
2984                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
2985                     prev_em_start && *prev_em_start != (u64)-1 &&
2986                     *prev_em_start != em->orig_start)
2987                         force_bio_submit = true;
2988
2989                 if (prev_em_start)
2990                         *prev_em_start = em->orig_start;
2991
2992                 free_extent_map(em);
2993                 em = NULL;
2994
2995                 /* we've found a hole, just zero and go on */
2996                 if (block_start == EXTENT_MAP_HOLE) {
2997                         char *userpage;
2998                         struct extent_state *cached = NULL;
2999
3000                         userpage = kmap_atomic(page);
3001                         memset(userpage + pg_offset, 0, iosize);
3002                         flush_dcache_page(page);
3003                         kunmap_atomic(userpage);
3004
3005                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3006                                             &cached, GFP_NOFS);
3007                         unlock_extent_cached(tree, cur,
3008                                              cur + iosize - 1, &cached);
3009                         cur = cur + iosize;
3010                         pg_offset += iosize;
3011                         continue;
3012                 }
3013                 /* the get_extent function already copied into the page */
3014                 if (test_range_bit(tree, cur, cur_end,
3015                                    EXTENT_UPTODATE, 1, NULL)) {
3016                         check_page_uptodate(tree, page);
3017                         unlock_extent(tree, cur, cur + iosize - 1);
3018                         cur = cur + iosize;
3019                         pg_offset += iosize;
3020                         continue;
3021                 }
3022                 /* we have an inline extent but it didn't get marked up
3023                  * to date.  Error out
3024                  */
3025                 if (block_start == EXTENT_MAP_INLINE) {
3026                         SetPageError(page);
3027                         unlock_extent(tree, cur, cur + iosize - 1);
3028                         cur = cur + iosize;
3029                         pg_offset += iosize;
3030                         continue;
3031                 }
3032
3033                 ret = submit_extent_page(REQ_OP_READ | read_flags, tree, NULL,
3034                                          page, offset, disk_io_size,
3035                                          pg_offset, bdev, bio,
3036                                          end_bio_extent_readpage, mirror_num,
3037                                          *bio_flags,
3038                                          this_bio_flag,
3039                                          force_bio_submit);
3040                 if (!ret) {
3041                         nr++;
3042                         *bio_flags = this_bio_flag;
3043                 } else {
3044                         SetPageError(page);
3045                         unlock_extent(tree, cur, cur + iosize - 1);
3046                         goto out;
3047                 }
3048                 cur = cur + iosize;
3049                 pg_offset += iosize;
3050         }
3051 out:
3052         if (!nr) {
3053                 if (!PageError(page))
3054                         SetPageUptodate(page);
3055                 unlock_page(page);
3056         }
3057         return ret;
3058 }
3059
3060 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3061                                              struct page *pages[], int nr_pages,
3062                                              u64 start, u64 end,
3063                                              struct extent_map **em_cached,
3064                                              struct bio **bio,
3065                                              unsigned long *bio_flags,
3066                                              u64 *prev_em_start)
3067 {
3068         struct inode *inode;
3069         struct btrfs_ordered_extent *ordered;
3070         int index;
3071
3072         inode = pages[0]->mapping->host;
3073         while (1) {
3074                 lock_extent(tree, start, end);
3075                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3076                                                      end - start + 1);
3077                 if (!ordered)
3078                         break;
3079                 unlock_extent(tree, start, end);
3080                 btrfs_start_ordered_extent(inode, ordered, 1);
3081                 btrfs_put_ordered_extent(ordered);
3082         }
3083
3084         for (index = 0; index < nr_pages; index++) {
3085                 __do_readpage(tree, pages[index], btrfs_get_extent, em_cached,
3086                                 bio, 0, bio_flags, REQ_RAHEAD, prev_em_start);
3087                 put_page(pages[index]);
3088         }
3089 }
3090
3091 static void __extent_readpages(struct extent_io_tree *tree,
3092                                struct page *pages[],
3093                                int nr_pages,
3094                                struct extent_map **em_cached,
3095                                struct bio **bio, unsigned long *bio_flags,
3096                                u64 *prev_em_start)
3097 {
3098         u64 start = 0;
3099         u64 end = 0;
3100         u64 page_start;
3101         int index;
3102         int first_index = 0;
3103
3104         for (index = 0; index < nr_pages; index++) {
3105                 page_start = page_offset(pages[index]);
3106                 if (!end) {
3107                         start = page_start;
3108                         end = start + PAGE_SIZE - 1;
3109                         first_index = index;
3110                 } else if (end + 1 == page_start) {
3111                         end += PAGE_SIZE;
3112                 } else {
3113                         __do_contiguous_readpages(tree, &pages[first_index],
3114                                                   index - first_index, start,
3115                                                   end, em_cached,
3116                                                   bio, bio_flags,
3117                                                   prev_em_start);
3118                         start = page_start;