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