276c9250ca230d6b1924c0bf3c2b2beaf9c93b75
[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 struct tree_entry {
113         u64 start;
114         u64 end;
115         struct rb_node rb_node;
116 };
117
118 struct extent_page_data {
119         struct bio *bio;
120         struct extent_io_tree *tree;
121         /* tells writepage not to lock the state bits for this range
122          * it still does the unlocking
123          */
124         unsigned int extent_locked:1;
125
126         /* tells the submit_bio code to use REQ_SYNC */
127         unsigned int sync_io:1;
128 };
129
130 static int add_extent_changeset(struct extent_state *state, unsigned bits,
131                                  struct extent_changeset *changeset,
132                                  int set)
133 {
134         int ret;
135
136         if (!changeset)
137                 return 0;
138         if (set && (state->state & bits) == bits)
139                 return 0;
140         if (!set && (state->state & bits) == 0)
141                 return 0;
142         changeset->bytes_changed += state->end - state->start + 1;
143         ret = ulist_add(&changeset->range_changed, state->start, state->end,
144                         GFP_ATOMIC);
145         return ret;
146 }
147
148 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
149                                        unsigned long bio_flags)
150 {
151         blk_status_t ret = 0;
152         struct bio_vec *bvec = bio_last_bvec_all(bio);
153         struct bio_vec bv;
154         struct extent_io_tree *tree = bio->bi_private;
155         u64 start;
156
157         mp_bvec_last_segment(bvec, &bv);
158         start = page_offset(bv.bv_page) + bv.bv_offset;
159
160         bio->bi_private = NULL;
161
162         if (tree->ops)
163                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
164                                            mirror_num, bio_flags, start);
165         else
166                 btrfsic_submit_bio(bio);
167
168         return blk_status_to_errno(ret);
169 }
170
171 /* Cleanup unsubmitted bios */
172 static void end_write_bio(struct extent_page_data *epd, int ret)
173 {
174         if (epd->bio) {
175                 epd->bio->bi_status = errno_to_blk_status(ret);
176                 bio_endio(epd->bio);
177                 epd->bio = NULL;
178         }
179 }
180
181 /*
182  * Submit bio from extent page data via submit_one_bio
183  *
184  * Return 0 if everything is OK.
185  * Return <0 for error.
186  */
187 static int __must_check flush_write_bio(struct extent_page_data *epd)
188 {
189         int ret = 0;
190
191         if (epd->bio) {
192                 ret = submit_one_bio(epd->bio, 0, 0);
193                 /*
194                  * Clean up of epd->bio is handled by its endio function.
195                  * And endio is either triggered by successful bio execution
196                  * or the error handler of submit bio hook.
197                  * So at this point, no matter what happened, we don't need
198                  * to clean up epd->bio.
199                  */
200                 epd->bio = NULL;
201         }
202         return ret;
203 }
204
205 int __init extent_io_init(void)
206 {
207         extent_state_cache = kmem_cache_create("btrfs_extent_state",
208                         sizeof(struct extent_state), 0,
209                         SLAB_MEM_SPREAD, NULL);
210         if (!extent_state_cache)
211                 return -ENOMEM;
212
213         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
214                         sizeof(struct extent_buffer), 0,
215                         SLAB_MEM_SPREAD, NULL);
216         if (!extent_buffer_cache)
217                 goto free_state_cache;
218
219         if (bioset_init(&btrfs_bioset, BIO_POOL_SIZE,
220                         offsetof(struct btrfs_io_bio, bio),
221                         BIOSET_NEED_BVECS))
222                 goto free_buffer_cache;
223
224         if (bioset_integrity_create(&btrfs_bioset, BIO_POOL_SIZE))
225                 goto free_bioset;
226
227         return 0;
228
229 free_bioset:
230         bioset_exit(&btrfs_bioset);
231
232 free_buffer_cache:
233         kmem_cache_destroy(extent_buffer_cache);
234         extent_buffer_cache = NULL;
235
236 free_state_cache:
237         kmem_cache_destroy(extent_state_cache);
238         extent_state_cache = NULL;
239         return -ENOMEM;
240 }
241
242 void __cold extent_io_exit(void)
243 {
244         btrfs_leak_debug_check();
245
246         /*
247          * Make sure all delayed rcu free are flushed before we
248          * destroy caches.
249          */
250         rcu_barrier();
251         kmem_cache_destroy(extent_state_cache);
252         kmem_cache_destroy(extent_buffer_cache);
253         bioset_exit(&btrfs_bioset);
254 }
255
256 void extent_io_tree_init(struct btrfs_fs_info *fs_info,
257                          struct extent_io_tree *tree, unsigned int owner,
258                          void *private_data)
259 {
260         tree->fs_info = fs_info;
261         tree->state = RB_ROOT;
262         tree->ops = NULL;
263         tree->dirty_bytes = 0;
264         spin_lock_init(&tree->lock);
265         tree->private_data = private_data;
266         tree->owner = owner;
267 }
268
269 static struct extent_state *alloc_extent_state(gfp_t mask)
270 {
271         struct extent_state *state;
272
273         /*
274          * The given mask might be not appropriate for the slab allocator,
275          * drop the unsupported bits
276          */
277         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
278         state = kmem_cache_alloc(extent_state_cache, mask);
279         if (!state)
280                 return state;
281         state->state = 0;
282         state->failrec = NULL;
283         RB_CLEAR_NODE(&state->rb_node);
284         btrfs_leak_debug_add(&state->leak_list, &states);
285         refcount_set(&state->refs, 1);
286         init_waitqueue_head(&state->wq);
287         trace_alloc_extent_state(state, mask, _RET_IP_);
288         return state;
289 }
290
291 void free_extent_state(struct extent_state *state)
292 {
293         if (!state)
294                 return;
295         if (refcount_dec_and_test(&state->refs)) {
296                 WARN_ON(extent_state_in_tree(state));
297                 btrfs_leak_debug_del(&state->leak_list);
298                 trace_free_extent_state(state, _RET_IP_);
299                 kmem_cache_free(extent_state_cache, state);
300         }
301 }
302
303 static struct rb_node *tree_insert(struct rb_root *root,
304                                    struct rb_node *search_start,
305                                    u64 offset,
306                                    struct rb_node *node,
307                                    struct rb_node ***p_in,
308                                    struct rb_node **parent_in)
309 {
310         struct rb_node **p;
311         struct rb_node *parent = NULL;
312         struct tree_entry *entry;
313
314         if (p_in && parent_in) {
315                 p = *p_in;
316                 parent = *parent_in;
317                 goto do_insert;
318         }
319
320         p = search_start ? &search_start : &root->rb_node;
321         while (*p) {
322                 parent = *p;
323                 entry = rb_entry(parent, struct tree_entry, rb_node);
324
325                 if (offset < entry->start)
326                         p = &(*p)->rb_left;
327                 else if (offset > entry->end)
328                         p = &(*p)->rb_right;
329                 else
330                         return parent;
331         }
332
333 do_insert:
334         rb_link_node(node, parent, p);
335         rb_insert_color(node, root);
336         return NULL;
337 }
338
339 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
340                                       struct rb_node **next_ret,
341                                       struct rb_node **prev_ret,
342                                       struct rb_node ***p_ret,
343                                       struct rb_node **parent_ret)
344 {
345         struct rb_root *root = &tree->state;
346         struct rb_node **n = &root->rb_node;
347         struct rb_node *prev = NULL;
348         struct rb_node *orig_prev = NULL;
349         struct tree_entry *entry;
350         struct tree_entry *prev_entry = NULL;
351
352         while (*n) {
353                 prev = *n;
354                 entry = rb_entry(prev, struct tree_entry, rb_node);
355                 prev_entry = entry;
356
357                 if (offset < entry->start)
358                         n = &(*n)->rb_left;
359                 else if (offset > entry->end)
360                         n = &(*n)->rb_right;
361                 else
362                         return *n;
363         }
364
365         if (p_ret)
366                 *p_ret = n;
367         if (parent_ret)
368                 *parent_ret = prev;
369
370         if (next_ret) {
371                 orig_prev = prev;
372                 while (prev && offset > prev_entry->end) {
373                         prev = rb_next(prev);
374                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
375                 }
376                 *next_ret = prev;
377                 prev = orig_prev;
378         }
379
380         if (prev_ret) {
381                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
382                 while (prev && offset < prev_entry->start) {
383                         prev = rb_prev(prev);
384                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
385                 }
386                 *prev_ret = prev;
387         }
388         return NULL;
389 }
390
391 static inline struct rb_node *
392 tree_search_for_insert(struct extent_io_tree *tree,
393                        u64 offset,
394                        struct rb_node ***p_ret,
395                        struct rb_node **parent_ret)
396 {
397         struct rb_node *next= NULL;
398         struct rb_node *ret;
399
400         ret = __etree_search(tree, offset, &next, NULL, p_ret, parent_ret);
401         if (!ret)
402                 return next;
403         return ret;
404 }
405
406 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
407                                           u64 offset)
408 {
409         return tree_search_for_insert(tree, offset, NULL, NULL);
410 }
411
412 /*
413  * utility function to look for merge candidates inside a given range.
414  * Any extents with matching state are merged together into a single
415  * extent in the tree.  Extents with EXTENT_IO in their state field
416  * are not merged because the end_io handlers need to be able to do
417  * operations on them without sleeping (or doing allocations/splits).
418  *
419  * This should be called with the tree lock held.
420  */
421 static void merge_state(struct extent_io_tree *tree,
422                         struct extent_state *state)
423 {
424         struct extent_state *other;
425         struct rb_node *other_node;
426
427         if (state->state & (EXTENT_LOCKED | EXTENT_BOUNDARY))
428                 return;
429
430         other_node = rb_prev(&state->rb_node);
431         if (other_node) {
432                 other = rb_entry(other_node, struct extent_state, rb_node);
433                 if (other->end == state->start - 1 &&
434                     other->state == state->state) {
435                         if (tree->private_data &&
436                             is_data_inode(tree->private_data))
437                                 btrfs_merge_delalloc_extent(tree->private_data,
438                                                             state, other);
439                         state->start = other->start;
440                         rb_erase(&other->rb_node, &tree->state);
441                         RB_CLEAR_NODE(&other->rb_node);
442                         free_extent_state(other);
443                 }
444         }
445         other_node = rb_next(&state->rb_node);
446         if (other_node) {
447                 other = rb_entry(other_node, struct extent_state, rb_node);
448                 if (other->start == state->end + 1 &&
449                     other->state == state->state) {
450                         if (tree->private_data &&
451                             is_data_inode(tree->private_data))
452                                 btrfs_merge_delalloc_extent(tree->private_data,
453                                                             state, other);
454                         state->end = other->end;
455                         rb_erase(&other->rb_node, &tree->state);
456                         RB_CLEAR_NODE(&other->rb_node);
457                         free_extent_state(other);
458                 }
459         }
460 }
461
462 static void set_state_bits(struct extent_io_tree *tree,
463                            struct extent_state *state, unsigned *bits,
464                            struct extent_changeset *changeset);
465
466 /*
467  * insert an extent_state struct into the tree.  'bits' are set on the
468  * struct before it is inserted.
469  *
470  * This may return -EEXIST if the extent is already there, in which case the
471  * state struct is freed.
472  *
473  * The tree lock is not taken internally.  This is a utility function and
474  * probably isn't what you want to call (see set/clear_extent_bit).
475  */
476 static int insert_state(struct extent_io_tree *tree,
477                         struct extent_state *state, u64 start, u64 end,
478                         struct rb_node ***p,
479                         struct rb_node **parent,
480                         unsigned *bits, struct extent_changeset *changeset)
481 {
482         struct rb_node *node;
483
484         if (end < start)
485                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
486                        end, start);
487         state->start = start;
488         state->end = end;
489
490         set_state_bits(tree, state, bits, changeset);
491
492         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
493         if (node) {
494                 struct extent_state *found;
495                 found = rb_entry(node, struct extent_state, rb_node);
496                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
497                        found->start, found->end, start, end);
498                 return -EEXIST;
499         }
500         merge_state(tree, state);
501         return 0;
502 }
503
504 /*
505  * split a given extent state struct in two, inserting the preallocated
506  * struct 'prealloc' as the newly created second half.  'split' indicates an
507  * offset inside 'orig' where it should be split.
508  *
509  * Before calling,
510  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
511  * are two extent state structs in the tree:
512  * prealloc: [orig->start, split - 1]
513  * orig: [ split, orig->end ]
514  *
515  * The tree locks are not taken by this function. They need to be held
516  * by the caller.
517  */
518 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
519                        struct extent_state *prealloc, u64 split)
520 {
521         struct rb_node *node;
522
523         if (tree->private_data && is_data_inode(tree->private_data))
524                 btrfs_split_delalloc_extent(tree->private_data, orig, split);
525
526         prealloc->start = orig->start;
527         prealloc->end = split - 1;
528         prealloc->state = orig->state;
529         orig->start = split;
530
531         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
532                            &prealloc->rb_node, NULL, NULL);
533         if (node) {
534                 free_extent_state(prealloc);
535                 return -EEXIST;
536         }
537         return 0;
538 }
539
540 static struct extent_state *next_state(struct extent_state *state)
541 {
542         struct rb_node *next = rb_next(&state->rb_node);
543         if (next)
544                 return rb_entry(next, struct extent_state, rb_node);
545         else
546                 return NULL;
547 }
548
549 /*
550  * utility function to clear some bits in an extent state struct.
551  * it will optionally wake up anyone waiting on this state (wake == 1).
552  *
553  * If no bits are set on the state struct after clearing things, the
554  * struct is freed and removed from the tree
555  */
556 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
557                                             struct extent_state *state,
558                                             unsigned *bits, int wake,
559                                             struct extent_changeset *changeset)
560 {
561         struct extent_state *next;
562         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
563         int ret;
564
565         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
566                 u64 range = state->end - state->start + 1;
567                 WARN_ON(range > tree->dirty_bytes);
568                 tree->dirty_bytes -= range;
569         }
570
571         if (tree->private_data && is_data_inode(tree->private_data))
572                 btrfs_clear_delalloc_extent(tree->private_data, state, bits);
573
574         ret = add_extent_changeset(state, bits_to_clear, changeset, 0);
575         BUG_ON(ret < 0);
576         state->state &= ~bits_to_clear;
577         if (wake)
578                 wake_up(&state->wq);
579         if (state->state == 0) {
580                 next = next_state(state);
581                 if (extent_state_in_tree(state)) {
582                         rb_erase(&state->rb_node, &tree->state);
583                         RB_CLEAR_NODE(&state->rb_node);
584                         free_extent_state(state);
585                 } else {
586                         WARN_ON(1);
587                 }
588         } else {
589                 merge_state(tree, state);
590                 next = next_state(state);
591         }
592         return next;
593 }
594
595 static struct extent_state *
596 alloc_extent_state_atomic(struct extent_state *prealloc)
597 {
598         if (!prealloc)
599                 prealloc = alloc_extent_state(GFP_ATOMIC);
600
601         return prealloc;
602 }
603
604 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
605 {
606         struct inode *inode = tree->private_data;
607
608         btrfs_panic(btrfs_sb(inode->i_sb), err,
609         "locking error: extent tree was modified by another thread while locked");
610 }
611
612 /*
613  * clear some bits on a range in the tree.  This may require splitting
614  * or inserting elements in the tree, so the gfp mask is used to
615  * indicate which allocations or sleeping are allowed.
616  *
617  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
618  * the given range from the tree regardless of state (ie for truncate).
619  *
620  * the range [start, end] is inclusive.
621  *
622  * This takes the tree lock, and returns 0 on success and < 0 on error.
623  */
624 int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
625                               unsigned bits, int wake, int delete,
626                               struct extent_state **cached_state,
627                               gfp_t mask, struct extent_changeset *changeset)
628 {
629         struct extent_state *state;
630         struct extent_state *cached;
631         struct extent_state *prealloc = NULL;
632         struct rb_node *node;
633         u64 last_end;
634         int err;
635         int clear = 0;
636
637         btrfs_debug_check_extent_io_range(tree, start, end);
638         trace_btrfs_clear_extent_bit(tree, start, end - start + 1, bits);
639
640         if (bits & EXTENT_DELALLOC)
641                 bits |= EXTENT_NORESERVE;
642
643         if (delete)
644                 bits |= ~EXTENT_CTLBITS;
645
646         if (bits & (EXTENT_LOCKED | EXTENT_BOUNDARY))
647                 clear = 1;
648 again:
649         if (!prealloc && gfpflags_allow_blocking(mask)) {
650                 /*
651                  * Don't care for allocation failure here because we might end
652                  * up not needing the pre-allocated extent state at all, which
653                  * is the case if we only have in the tree extent states that
654                  * cover our input range and don't cover too any other range.
655                  * If we end up needing a new extent state we allocate it later.
656                  */
657                 prealloc = alloc_extent_state(mask);
658         }
659
660         spin_lock(&tree->lock);
661         if (cached_state) {
662                 cached = *cached_state;
663
664                 if (clear) {
665                         *cached_state = NULL;
666                         cached_state = NULL;
667                 }
668
669                 if (cached && extent_state_in_tree(cached) &&
670                     cached->start <= start && cached->end > start) {
671                         if (clear)
672                                 refcount_dec(&cached->refs);
673                         state = cached;
674                         goto hit_next;
675                 }
676                 if (clear)
677                         free_extent_state(cached);
678         }
679         /*
680          * this search will find the extents that end after
681          * our range starts
682          */
683         node = tree_search(tree, start);
684         if (!node)
685                 goto out;
686         state = rb_entry(node, struct extent_state, rb_node);
687 hit_next:
688         if (state->start > end)
689                 goto out;
690         WARN_ON(state->end < start);
691         last_end = state->end;
692
693         /* the state doesn't have the wanted bits, go ahead */
694         if (!(state->state & bits)) {
695                 state = next_state(state);
696                 goto next;
697         }
698
699         /*
700          *     | ---- desired range ---- |
701          *  | state | or
702          *  | ------------- state -------------- |
703          *
704          * We need to split the extent we found, and may flip
705          * bits on second half.
706          *
707          * If the extent we found extends past our range, we
708          * just split and search again.  It'll get split again
709          * the next time though.
710          *
711          * If the extent we found is inside our range, we clear
712          * the desired bit on it.
713          */
714
715         if (state->start < start) {
716                 prealloc = alloc_extent_state_atomic(prealloc);
717                 BUG_ON(!prealloc);
718                 err = split_state(tree, state, prealloc, start);
719                 if (err)
720                         extent_io_tree_panic(tree, err);
721
722                 prealloc = NULL;
723                 if (err)
724                         goto out;
725                 if (state->end <= end) {
726                         state = clear_state_bit(tree, state, &bits, wake,
727                                                 changeset);
728                         goto next;
729                 }
730                 goto search_again;
731         }
732         /*
733          * | ---- desired range ---- |
734          *                        | state |
735          * We need to split the extent, and clear the bit
736          * on the first half
737          */
738         if (state->start <= end && state->end > end) {
739                 prealloc = alloc_extent_state_atomic(prealloc);
740                 BUG_ON(!prealloc);
741                 err = split_state(tree, state, prealloc, end + 1);
742                 if (err)
743                         extent_io_tree_panic(tree, err);
744
745                 if (wake)
746                         wake_up(&state->wq);
747
748                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
749
750                 prealloc = NULL;
751                 goto out;
752         }
753
754         state = clear_state_bit(tree, state, &bits, wake, changeset);
755 next:
756         if (last_end == (u64)-1)
757                 goto out;
758         start = last_end + 1;
759         if (start <= end && state && !need_resched())
760                 goto hit_next;
761
762 search_again:
763         if (start > end)
764                 goto out;
765         spin_unlock(&tree->lock);
766         if (gfpflags_allow_blocking(mask))
767                 cond_resched();
768         goto again;
769
770 out:
771         spin_unlock(&tree->lock);
772         if (prealloc)
773                 free_extent_state(prealloc);
774
775         return 0;
776
777 }
778
779 static void wait_on_state(struct extent_io_tree *tree,
780                           struct extent_state *state)
781                 __releases(tree->lock)
782                 __acquires(tree->lock)
783 {
784         DEFINE_WAIT(wait);
785         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
786         spin_unlock(&tree->lock);
787         schedule();
788         spin_lock(&tree->lock);
789         finish_wait(&state->wq, &wait);
790 }
791
792 /*
793  * waits for one or more bits to clear on a range in the state tree.
794  * The range [start, end] is inclusive.
795  * The tree lock is taken by this function
796  */
797 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
798                             unsigned long bits)
799 {
800         struct extent_state *state;
801         struct rb_node *node;
802
803         btrfs_debug_check_extent_io_range(tree, start, end);
804
805         spin_lock(&tree->lock);
806 again:
807         while (1) {
808                 /*
809                  * this search will find all the extents that end after
810                  * our range starts
811                  */
812                 node = tree_search(tree, start);
813 process_node:
814                 if (!node)
815                         break;
816
817                 state = rb_entry(node, struct extent_state, rb_node);
818
819                 if (state->start > end)
820                         goto out;
821
822                 if (state->state & bits) {
823                         start = state->start;
824                         refcount_inc(&state->refs);
825                         wait_on_state(tree, state);
826                         free_extent_state(state);
827                         goto again;
828                 }
829                 start = state->end + 1;
830
831                 if (start > end)
832                         break;
833
834                 if (!cond_resched_lock(&tree->lock)) {
835                         node = rb_next(node);
836                         goto process_node;
837                 }
838         }
839 out:
840         spin_unlock(&tree->lock);
841 }
842
843 static void set_state_bits(struct extent_io_tree *tree,
844                            struct extent_state *state,
845                            unsigned *bits, struct extent_changeset *changeset)
846 {
847         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
848         int ret;
849
850         if (tree->private_data && is_data_inode(tree->private_data))
851                 btrfs_set_delalloc_extent(tree->private_data, state, bits);
852
853         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
854                 u64 range = state->end - state->start + 1;
855                 tree->dirty_bytes += range;
856         }
857         ret = add_extent_changeset(state, bits_to_set, changeset, 1);
858         BUG_ON(ret < 0);
859         state->state |= bits_to_set;
860 }
861
862 static void cache_state_if_flags(struct extent_state *state,
863                                  struct extent_state **cached_ptr,
864                                  unsigned flags)
865 {
866         if (cached_ptr && !(*cached_ptr)) {
867                 if (!flags || (state->state & flags)) {
868                         *cached_ptr = state;
869                         refcount_inc(&state->refs);
870                 }
871         }
872 }
873
874 static void cache_state(struct extent_state *state,
875                         struct extent_state **cached_ptr)
876 {
877         return cache_state_if_flags(state, cached_ptr,
878                                     EXTENT_LOCKED | EXTENT_BOUNDARY);
879 }
880
881 /*
882  * set some bits on a range in the tree.  This may require allocations or
883  * sleeping, so the gfp mask is used to indicate what is allowed.
884  *
885  * If any of the exclusive bits are set, this will fail with -EEXIST if some
886  * part of the range already has the desired bits set.  The start of the
887  * existing range is returned in failed_start in this case.
888  *
889  * [start, end] is inclusive This takes the tree lock.
890  */
891
892 static int __must_check
893 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
894                  unsigned bits, unsigned exclusive_bits,
895                  u64 *failed_start, struct extent_state **cached_state,
896                  gfp_t mask, struct extent_changeset *changeset)
897 {
898         struct extent_state *state;
899         struct extent_state *prealloc = NULL;
900         struct rb_node *node;
901         struct rb_node **p;
902         struct rb_node *parent;
903         int err = 0;
904         u64 last_start;
905         u64 last_end;
906
907         btrfs_debug_check_extent_io_range(tree, start, end);
908         trace_btrfs_set_extent_bit(tree, start, end - start + 1, bits);
909
910 again:
911         if (!prealloc && gfpflags_allow_blocking(mask)) {
912                 /*
913                  * Don't care for allocation failure here because we might end
914                  * up not needing the pre-allocated extent state at all, which
915                  * is the case if we only have in the tree extent states that
916                  * cover our input range and don't cover too any other range.
917                  * If we end up needing a new extent state we allocate it later.
918                  */
919                 prealloc = alloc_extent_state(mask);
920         }
921
922         spin_lock(&tree->lock);
923         if (cached_state && *cached_state) {
924                 state = *cached_state;
925                 if (state->start <= start && state->end > start &&
926                     extent_state_in_tree(state)) {
927                         node = &state->rb_node;
928                         goto hit_next;
929                 }
930         }
931         /*
932          * this search will find all the extents that end after
933          * our range starts.
934          */
935         node = tree_search_for_insert(tree, start, &p, &parent);
936         if (!node) {
937                 prealloc = alloc_extent_state_atomic(prealloc);
938                 BUG_ON(!prealloc);
939                 err = insert_state(tree, prealloc, start, end,
940                                    &p, &parent, &bits, changeset);
941                 if (err)
942                         extent_io_tree_panic(tree, err);
943
944                 cache_state(prealloc, cached_state);
945                 prealloc = NULL;
946                 goto out;
947         }
948         state = rb_entry(node, struct extent_state, rb_node);
949 hit_next:
950         last_start = state->start;
951         last_end = state->end;
952
953         /*
954          * | ---- desired range ---- |
955          * | state |
956          *
957          * Just lock what we found and keep going
958          */
959         if (state->start == start && state->end <= end) {
960                 if (state->state & exclusive_bits) {
961                         *failed_start = state->start;
962                         err = -EEXIST;
963                         goto out;
964                 }
965
966                 set_state_bits(tree, state, &bits, changeset);
967                 cache_state(state, cached_state);
968                 merge_state(tree, state);
969                 if (last_end == (u64)-1)
970                         goto out;
971                 start = last_end + 1;
972                 state = next_state(state);
973                 if (start < end && state && state->start == start &&
974                     !need_resched())
975                         goto hit_next;
976                 goto search_again;
977         }
978
979         /*
980          *     | ---- desired range ---- |
981          * | state |
982          *   or
983          * | ------------- state -------------- |
984          *
985          * We need to split the extent we found, and may flip bits on
986          * second half.
987          *
988          * If the extent we found extends past our
989          * range, we just split and search again.  It'll get split
990          * again the next time though.
991          *
992          * If the extent we found is inside our range, we set the
993          * desired bit on it.
994          */
995         if (state->start < start) {
996                 if (state->state & exclusive_bits) {
997                         *failed_start = start;
998                         err = -EEXIST;
999                         goto out;
1000                 }
1001
1002                 prealloc = alloc_extent_state_atomic(prealloc);
1003                 BUG_ON(!prealloc);
1004                 err = split_state(tree, state, prealloc, start);
1005                 if (err)
1006                         extent_io_tree_panic(tree, err);
1007
1008                 prealloc = NULL;
1009                 if (err)
1010                         goto out;
1011                 if (state->end <= end) {
1012                         set_state_bits(tree, state, &bits, changeset);
1013                         cache_state(state, cached_state);
1014                         merge_state(tree, state);
1015                         if (last_end == (u64)-1)
1016                                 goto out;
1017                         start = last_end + 1;
1018                         state = next_state(state);
1019                         if (start < end && state && state->start == start &&
1020                             !need_resched())
1021                                 goto hit_next;
1022                 }
1023                 goto search_again;
1024         }
1025         /*
1026          * | ---- desired range ---- |
1027          *     | state | or               | state |
1028          *
1029          * There's a hole, we need to insert something in it and
1030          * ignore the extent we found.
1031          */
1032         if (state->start > start) {
1033                 u64 this_end;
1034                 if (end < last_start)
1035                         this_end = end;
1036                 else
1037                         this_end = last_start - 1;
1038
1039                 prealloc = alloc_extent_state_atomic(prealloc);
1040                 BUG_ON(!prealloc);
1041
1042                 /*
1043                  * Avoid to free 'prealloc' if it can be merged with
1044                  * the later extent.
1045                  */
1046                 err = insert_state(tree, prealloc, start, this_end,
1047                                    NULL, NULL, &bits, changeset);
1048                 if (err)
1049                         extent_io_tree_panic(tree, err);
1050
1051                 cache_state(prealloc, cached_state);
1052                 prealloc = NULL;
1053                 start = this_end + 1;
1054                 goto search_again;
1055         }
1056         /*
1057          * | ---- desired range ---- |
1058          *                        | state |
1059          * We need to split the extent, and set the bit
1060          * on the first half
1061          */
1062         if (state->start <= end && state->end > end) {
1063                 if (state->state & exclusive_bits) {
1064                         *failed_start = start;
1065                         err = -EEXIST;
1066                         goto out;
1067                 }
1068
1069                 prealloc = alloc_extent_state_atomic(prealloc);
1070                 BUG_ON(!prealloc);
1071                 err = split_state(tree, state, prealloc, end + 1);
1072                 if (err)
1073                         extent_io_tree_panic(tree, err);
1074
1075                 set_state_bits(tree, prealloc, &bits, changeset);
1076                 cache_state(prealloc, cached_state);
1077                 merge_state(tree, prealloc);
1078                 prealloc = NULL;
1079                 goto out;
1080         }
1081
1082 search_again:
1083         if (start > end)
1084                 goto out;
1085         spin_unlock(&tree->lock);
1086         if (gfpflags_allow_blocking(mask))
1087                 cond_resched();
1088         goto again;
1089
1090 out:
1091         spin_unlock(&tree->lock);
1092         if (prealloc)
1093                 free_extent_state(prealloc);
1094
1095         return err;
1096
1097 }
1098
1099 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1100                    unsigned bits, u64 * failed_start,
1101                    struct extent_state **cached_state, gfp_t mask)
1102 {
1103         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1104                                 cached_state, mask, NULL);
1105 }
1106
1107
1108 /**
1109  * convert_extent_bit - convert all bits in a given range from one bit to
1110  *                      another
1111  * @tree:       the io tree to search
1112  * @start:      the start offset in bytes
1113  * @end:        the end offset in bytes (inclusive)
1114  * @bits:       the bits to set in this range
1115  * @clear_bits: the bits to clear in this range
1116  * @cached_state:       state that we're going to cache
1117  *
1118  * This will go through and set bits for the given range.  If any states exist
1119  * already in this range they are set with the given bit and cleared of the
1120  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1121  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1122  * boundary bits like LOCK.
1123  *
1124  * All allocations are done with GFP_NOFS.
1125  */
1126 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1127                        unsigned bits, unsigned clear_bits,
1128                        struct extent_state **cached_state)
1129 {
1130         struct extent_state *state;
1131         struct extent_state *prealloc = NULL;
1132         struct rb_node *node;
1133         struct rb_node **p;
1134         struct rb_node *parent;
1135         int err = 0;
1136         u64 last_start;
1137         u64 last_end;
1138         bool first_iteration = true;
1139
1140         btrfs_debug_check_extent_io_range(tree, start, end);
1141         trace_btrfs_convert_extent_bit(tree, start, end - start + 1, bits,
1142                                        clear_bits);
1143
1144 again:
1145         if (!prealloc) {
1146                 /*
1147                  * Best effort, don't worry if extent state allocation fails
1148                  * here for the first iteration. We might have a cached state
1149                  * that matches exactly the target range, in which case no
1150                  * extent state allocations are needed. We'll only know this
1151                  * after locking the tree.
1152                  */
1153                 prealloc = alloc_extent_state(GFP_NOFS);
1154                 if (!prealloc && !first_iteration)
1155                         return -ENOMEM;
1156         }
1157
1158         spin_lock(&tree->lock);
1159         if (cached_state && *cached_state) {
1160                 state = *cached_state;
1161                 if (state->start <= start && state->end > start &&
1162                     extent_state_in_tree(state)) {
1163                         node = &state->rb_node;
1164                         goto hit_next;
1165                 }
1166         }
1167
1168         /*
1169          * this search will find all the extents that end after
1170          * our range starts.
1171          */
1172         node = tree_search_for_insert(tree, start, &p, &parent);
1173         if (!node) {
1174                 prealloc = alloc_extent_state_atomic(prealloc);
1175                 if (!prealloc) {
1176                         err = -ENOMEM;
1177                         goto out;
1178                 }
1179                 err = insert_state(tree, prealloc, start, end,
1180                                    &p, &parent, &bits, NULL);
1181                 if (err)
1182                         extent_io_tree_panic(tree, err);
1183                 cache_state(prealloc, cached_state);
1184                 prealloc = NULL;
1185                 goto out;
1186         }
1187         state = rb_entry(node, struct extent_state, rb_node);
1188 hit_next:
1189         last_start = state->start;
1190         last_end = state->end;
1191
1192         /*
1193          * | ---- desired range ---- |
1194          * | state |
1195          *
1196          * Just lock what we found and keep going
1197          */
1198         if (state->start == start && state->end <= end) {
1199                 set_state_bits(tree, state, &bits, NULL);
1200                 cache_state(state, cached_state);
1201                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1202                 if (last_end == (u64)-1)
1203                         goto out;
1204                 start = last_end + 1;
1205                 if (start < end && state && state->start == start &&
1206                     !need_resched())
1207                         goto hit_next;
1208                 goto search_again;
1209         }
1210
1211         /*
1212          *     | ---- desired range ---- |
1213          * | state |
1214          *   or
1215          * | ------------- state -------------- |
1216          *
1217          * We need to split the extent we found, and may flip bits on
1218          * second half.
1219          *
1220          * If the extent we found extends past our
1221          * range, we just split and search again.  It'll get split
1222          * again the next time though.
1223          *
1224          * If the extent we found is inside our range, we set the
1225          * desired bit on it.
1226          */
1227         if (state->start < start) {
1228                 prealloc = alloc_extent_state_atomic(prealloc);
1229                 if (!prealloc) {
1230                         err = -ENOMEM;
1231                         goto out;
1232                 }
1233                 err = split_state(tree, state, prealloc, start);
1234                 if (err)
1235                         extent_io_tree_panic(tree, err);
1236                 prealloc = NULL;
1237                 if (err)
1238                         goto out;
1239                 if (state->end <= end) {
1240                         set_state_bits(tree, state, &bits, NULL);
1241                         cache_state(state, cached_state);
1242                         state = clear_state_bit(tree, state, &clear_bits, 0,
1243                                                 NULL);
1244                         if (last_end == (u64)-1)
1245                                 goto out;
1246                         start = last_end + 1;
1247                         if (start < end && state && state->start == start &&
1248                             !need_resched())
1249                                 goto hit_next;
1250                 }
1251                 goto search_again;
1252         }
1253         /*
1254          * | ---- desired range ---- |
1255          *     | state | or               | state |
1256          *
1257          * There's a hole, we need to insert something in it and
1258          * ignore the extent we found.
1259          */
1260         if (state->start > start) {
1261                 u64 this_end;
1262                 if (end < last_start)
1263                         this_end = end;
1264                 else
1265                         this_end = last_start - 1;
1266
1267                 prealloc = alloc_extent_state_atomic(prealloc);
1268                 if (!prealloc) {
1269                         err = -ENOMEM;
1270                         goto out;
1271                 }
1272
1273                 /*
1274                  * Avoid to free 'prealloc' if it can be merged with
1275                  * the later extent.
1276                  */
1277                 err = insert_state(tree, prealloc, start, this_end,
1278                                    NULL, NULL, &bits, NULL);
1279                 if (err)
1280                         extent_io_tree_panic(tree, err);
1281                 cache_state(prealloc, cached_state);
1282                 prealloc = NULL;
1283                 start = this_end + 1;
1284                 goto search_again;
1285         }
1286         /*
1287          * | ---- desired range ---- |
1288          *                        | state |
1289          * We need to split the extent, and set the bit
1290          * on the first half
1291          */
1292         if (state->start <= end && state->end > end) {
1293                 prealloc = alloc_extent_state_atomic(prealloc);
1294                 if (!prealloc) {
1295                         err = -ENOMEM;
1296                         goto out;
1297                 }
1298
1299                 err = split_state(tree, state, prealloc, end + 1);
1300                 if (err)
1301                         extent_io_tree_panic(tree, err);
1302
1303                 set_state_bits(tree, prealloc, &bits, NULL);
1304                 cache_state(prealloc, cached_state);
1305                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1306                 prealloc = NULL;
1307                 goto out;
1308         }
1309
1310 search_again:
1311         if (start > end)
1312                 goto out;
1313         spin_unlock(&tree->lock);
1314         cond_resched();
1315         first_iteration = false;
1316         goto again;
1317
1318 out:
1319         spin_unlock(&tree->lock);
1320         if (prealloc)
1321                 free_extent_state(prealloc);
1322
1323         return err;
1324 }
1325
1326 /* wrappers around set/clear extent bit */
1327 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1328                            unsigned bits, struct extent_changeset *changeset)
1329 {
1330         /*
1331          * We don't support EXTENT_LOCKED yet, as current changeset will
1332          * record any bits changed, so for EXTENT_LOCKED case, it will
1333          * either fail with -EEXIST or changeset will record the whole
1334          * range.
1335          */
1336         BUG_ON(bits & EXTENT_LOCKED);
1337
1338         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1339                                 changeset);
1340 }
1341
1342 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1343                      unsigned bits, int wake, int delete,
1344                      struct extent_state **cached)
1345 {
1346         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1347                                   cached, GFP_NOFS, NULL);
1348 }
1349
1350 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1351                 unsigned bits, struct extent_changeset *changeset)
1352 {
1353         /*
1354          * Don't support EXTENT_LOCKED case, same reason as
1355          * set_record_extent_bits().
1356          */
1357         BUG_ON(bits & EXTENT_LOCKED);
1358
1359         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1360                                   changeset);
1361 }
1362
1363 /*
1364  * either insert or lock state struct between start and end use mask to tell
1365  * us if waiting is desired.
1366  */
1367 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1368                      struct extent_state **cached_state)
1369 {
1370         int err;
1371         u64 failed_start;
1372
1373         while (1) {
1374                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1375                                        EXTENT_LOCKED, &failed_start,
1376                                        cached_state, GFP_NOFS, NULL);
1377                 if (err == -EEXIST) {
1378                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1379                         start = failed_start;
1380                 } else
1381                         break;
1382                 WARN_ON(start > end);
1383         }
1384         return err;
1385 }
1386
1387 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1388 {
1389         int err;
1390         u64 failed_start;
1391
1392         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1393                                &failed_start, NULL, GFP_NOFS, NULL);
1394         if (err == -EEXIST) {
1395                 if (failed_start > start)
1396                         clear_extent_bit(tree, start, failed_start - 1,
1397                                          EXTENT_LOCKED, 1, 0, NULL);
1398                 return 0;
1399         }
1400         return 1;
1401 }
1402
1403 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1404 {
1405         unsigned long index = start >> PAGE_SHIFT;
1406         unsigned long end_index = end >> PAGE_SHIFT;
1407         struct page *page;
1408
1409         while (index <= end_index) {
1410                 page = find_get_page(inode->i_mapping, index);
1411                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1412                 clear_page_dirty_for_io(page);
1413                 put_page(page);
1414                 index++;
1415         }
1416 }
1417
1418 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1419 {
1420         unsigned long index = start >> PAGE_SHIFT;
1421         unsigned long end_index = end >> PAGE_SHIFT;
1422         struct page *page;
1423
1424         while (index <= end_index) {
1425                 page = find_get_page(inode->i_mapping, index);
1426                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1427                 __set_page_dirty_nobuffers(page);
1428                 account_page_redirty(page);
1429                 put_page(page);
1430                 index++;
1431         }
1432 }
1433
1434 /* find the first state struct with 'bits' set after 'start', and
1435  * return it.  tree->lock must be held.  NULL will returned if
1436  * nothing was found after 'start'
1437  */
1438 static struct extent_state *
1439 find_first_extent_bit_state(struct extent_io_tree *tree,
1440                             u64 start, unsigned bits)
1441 {
1442         struct rb_node *node;
1443         struct extent_state *state;
1444
1445         /*
1446          * this search will find all the extents that end after
1447          * our range starts.
1448          */
1449         node = tree_search(tree, start);
1450         if (!node)
1451                 goto out;
1452
1453         while (1) {
1454                 state = rb_entry(node, struct extent_state, rb_node);
1455                 if (state->end >= start && (state->state & bits))
1456                         return state;
1457
1458                 node = rb_next(node);
1459                 if (!node)
1460                         break;
1461         }
1462 out:
1463         return NULL;
1464 }
1465
1466 /*
1467  * find the first offset in the io tree with 'bits' set. zero is
1468  * returned if we find something, and *start_ret and *end_ret are
1469  * set to reflect the state struct that was found.
1470  *
1471  * If nothing was found, 1 is returned. If found something, return 0.
1472  */
1473 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1474                           u64 *start_ret, u64 *end_ret, unsigned bits,
1475                           struct extent_state **cached_state)
1476 {
1477         struct extent_state *state;
1478         int ret = 1;
1479
1480         spin_lock(&tree->lock);
1481         if (cached_state && *cached_state) {
1482                 state = *cached_state;
1483                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1484                         while ((state = next_state(state)) != NULL) {
1485                                 if (state->state & bits)
1486                                         goto got_it;
1487                         }
1488                         free_extent_state(*cached_state);
1489                         *cached_state = NULL;
1490                         goto out;
1491                 }
1492                 free_extent_state(*cached_state);
1493                 *cached_state = NULL;
1494         }
1495
1496         state = find_first_extent_bit_state(tree, start, bits);
1497 got_it:
1498         if (state) {
1499                 cache_state_if_flags(state, cached_state, 0);
1500                 *start_ret = state->start;
1501                 *end_ret = state->end;
1502                 ret = 0;
1503         }
1504 out:
1505         spin_unlock(&tree->lock);
1506         return ret;
1507 }
1508
1509 /*
1510  * find a contiguous range of bytes in the file marked as delalloc, not
1511  * more than 'max_bytes'.  start and end are used to return the range,
1512  *
1513  * true is returned if we find something, false if nothing was in the tree
1514  */
1515 static noinline bool find_delalloc_range(struct extent_io_tree *tree,
1516                                         u64 *start, u64 *end, u64 max_bytes,
1517                                         struct extent_state **cached_state)
1518 {
1519         struct rb_node *node;
1520         struct extent_state *state;
1521         u64 cur_start = *start;
1522         bool found = false;
1523         u64 total_bytes = 0;
1524
1525         spin_lock(&tree->lock);
1526
1527         /*
1528          * this search will find all the extents that end after
1529          * our range starts.
1530          */
1531         node = tree_search(tree, cur_start);
1532         if (!node) {
1533                 *end = (u64)-1;
1534                 goto out;
1535         }
1536
1537         while (1) {
1538                 state = rb_entry(node, struct extent_state, rb_node);
1539                 if (found && (state->start != cur_start ||
1540                               (state->state & EXTENT_BOUNDARY))) {
1541                         goto out;
1542                 }
1543                 if (!(state->state & EXTENT_DELALLOC)) {
1544                         if (!found)
1545                                 *end = state->end;
1546                         goto out;
1547                 }
1548                 if (!found) {
1549                         *start = state->start;
1550                         *cached_state = state;
1551                         refcount_inc(&state->refs);
1552                 }
1553                 found = true;
1554                 *end = state->end;
1555                 cur_start = state->end + 1;
1556                 node = rb_next(node);
1557                 total_bytes += state->end - state->start + 1;
1558                 if (total_bytes >= max_bytes)
1559                         break;
1560                 if (!node)
1561                         break;
1562         }
1563 out:
1564         spin_unlock(&tree->lock);
1565         return found;
1566 }
1567
1568 static int __process_pages_contig(struct address_space *mapping,
1569                                   struct page *locked_page,
1570                                   pgoff_t start_index, pgoff_t end_index,
1571                                   unsigned long page_ops, pgoff_t *index_ret);
1572
1573 static noinline void __unlock_for_delalloc(struct inode *inode,
1574                                            struct page *locked_page,
1575                                            u64 start, u64 end)
1576 {
1577         unsigned long index = start >> PAGE_SHIFT;
1578         unsigned long end_index = end >> PAGE_SHIFT;
1579
1580         ASSERT(locked_page);
1581         if (index == locked_page->index && end_index == index)
1582                 return;
1583
1584         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1585                                PAGE_UNLOCK, NULL);
1586 }
1587
1588 static noinline int lock_delalloc_pages(struct inode *inode,
1589                                         struct page *locked_page,
1590                                         u64 delalloc_start,
1591                                         u64 delalloc_end)
1592 {
1593         unsigned long index = delalloc_start >> PAGE_SHIFT;
1594         unsigned long index_ret = index;
1595         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1596         int ret;
1597
1598         ASSERT(locked_page);
1599         if (index == locked_page->index && index == end_index)
1600                 return 0;
1601
1602         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1603                                      end_index, PAGE_LOCK, &index_ret);
1604         if (ret == -EAGAIN)
1605                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1606                                       (u64)index_ret << PAGE_SHIFT);
1607         return ret;
1608 }
1609
1610 /*
1611  * Find and lock a contiguous range of bytes in the file marked as delalloc, no
1612  * more than @max_bytes.  @Start and @end are used to return the range,
1613  *
1614  * Return: true if we find something
1615  *         false if nothing was in the tree
1616  */
1617 EXPORT_FOR_TESTS
1618 noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
1619                                     struct extent_io_tree *tree,
1620                                     struct page *locked_page, u64 *start,
1621                                     u64 *end)
1622 {
1623         u64 max_bytes = BTRFS_MAX_EXTENT_SIZE;
1624         u64 delalloc_start;
1625         u64 delalloc_end;
1626         bool found;
1627         struct extent_state *cached_state = NULL;
1628         int ret;
1629         int loops = 0;
1630
1631 again:
1632         /* step one, find a bunch of delalloc bytes starting at start */
1633         delalloc_start = *start;
1634         delalloc_end = 0;
1635         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1636                                     max_bytes, &cached_state);
1637         if (!found || delalloc_end <= *start) {
1638                 *start = delalloc_start;
1639                 *end = delalloc_end;
1640                 free_extent_state(cached_state);
1641                 return false;
1642         }
1643
1644         /*
1645          * start comes from the offset of locked_page.  We have to lock
1646          * pages in order, so we can't process delalloc bytes before
1647          * locked_page
1648          */
1649         if (delalloc_start < *start)
1650                 delalloc_start = *start;
1651
1652         /*
1653          * make sure to limit the number of pages we try to lock down
1654          */
1655         if (delalloc_end + 1 - delalloc_start > max_bytes)
1656                 delalloc_end = delalloc_start + max_bytes - 1;
1657
1658         /* step two, lock all the pages after the page that has start */
1659         ret = lock_delalloc_pages(inode, locked_page,
1660                                   delalloc_start, delalloc_end);
1661         ASSERT(!ret || ret == -EAGAIN);
1662         if (ret == -EAGAIN) {
1663                 /* some of the pages are gone, lets avoid looping by
1664                  * shortening the size of the delalloc range we're searching
1665                  */
1666                 free_extent_state(cached_state);
1667                 cached_state = NULL;
1668                 if (!loops) {
1669                         max_bytes = PAGE_SIZE;
1670                         loops = 1;
1671                         goto again;
1672                 } else {
1673                         found = false;
1674                         goto out_failed;
1675                 }
1676         }
1677
1678         /* step three, lock the state bits for the whole range */
1679         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1680
1681         /* then test to make sure it is all still delalloc */
1682         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1683                              EXTENT_DELALLOC, 1, cached_state);
1684         if (!ret) {
1685                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1686                                      &cached_state);
1687                 __unlock_for_delalloc(inode, locked_page,
1688                               delalloc_start, delalloc_end);
1689                 cond_resched();
1690                 goto again;
1691         }
1692         free_extent_state(cached_state);
1693         *start = delalloc_start;
1694         *end = delalloc_end;
1695 out_failed:
1696         return found;
1697 }
1698
1699 static int __process_pages_contig(struct address_space *mapping,
1700                                   struct page *locked_page,
1701                                   pgoff_t start_index, pgoff_t end_index,
1702                                   unsigned long page_ops, pgoff_t *index_ret)
1703 {
1704         unsigned long nr_pages = end_index - start_index + 1;
1705         unsigned long pages_locked = 0;
1706         pgoff_t index = start_index;
1707         struct page *pages[16];
1708         unsigned ret;
1709         int err = 0;
1710         int i;
1711
1712         if (page_ops & PAGE_LOCK) {
1713                 ASSERT(page_ops == PAGE_LOCK);
1714                 ASSERT(index_ret && *index_ret == start_index);
1715         }
1716
1717         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1718                 mapping_set_error(mapping, -EIO);
1719
1720         while (nr_pages > 0) {
1721                 ret = find_get_pages_contig(mapping, index,
1722                                      min_t(unsigned long,
1723                                      nr_pages, ARRAY_SIZE(pages)), pages);
1724                 if (ret == 0) {
1725                         /*
1726                          * Only if we're going to lock these pages,
1727                          * can we find nothing at @index.
1728                          */
1729                         ASSERT(page_ops & PAGE_LOCK);
1730                         err = -EAGAIN;
1731                         goto out;
1732                 }
1733
1734                 for (i = 0; i < ret; i++) {
1735                         if (page_ops & PAGE_SET_PRIVATE2)
1736                                 SetPagePrivate2(pages[i]);
1737
1738                         if (pages[i] == locked_page) {
1739                                 put_page(pages[i]);
1740                                 pages_locked++;
1741                                 continue;
1742                         }
1743                         if (page_ops & PAGE_CLEAR_DIRTY)
1744                                 clear_page_dirty_for_io(pages[i]);
1745                         if (page_ops & PAGE_SET_WRITEBACK)
1746                                 set_page_writeback(pages[i]);
1747                         if (page_ops & PAGE_SET_ERROR)
1748                                 SetPageError(pages[i]);
1749                         if (page_ops & PAGE_END_WRITEBACK)
1750                                 end_page_writeback(pages[i]);
1751                         if (page_ops & PAGE_UNLOCK)
1752                                 unlock_page(pages[i]);
1753                         if (page_ops & PAGE_LOCK) {
1754                                 lock_page(pages[i]);
1755                                 if (!PageDirty(pages[i]) ||
1756                                     pages[i]->mapping != mapping) {
1757                                         unlock_page(pages[i]);
1758                                         put_page(pages[i]);
1759                                         err = -EAGAIN;
1760                                         goto out;
1761                                 }
1762                         }
1763                         put_page(pages[i]);
1764                         pages_locked++;
1765                 }
1766                 nr_pages -= ret;
1767                 index += ret;
1768                 cond_resched();
1769         }
1770 out:
1771         if (err && index_ret)
1772                 *index_ret = start_index + pages_locked - 1;
1773         return err;
1774 }
1775
1776 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1777                                  u64 delalloc_end, struct page *locked_page,
1778                                  unsigned clear_bits,
1779                                  unsigned long page_ops)
1780 {
1781         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1782                          NULL);
1783
1784         __process_pages_contig(inode->i_mapping, locked_page,
1785                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1786                                page_ops, NULL);
1787 }
1788
1789 /*
1790  * count the number of bytes in the tree that have a given bit(s)
1791  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1792  * cached.  The total number found is returned.
1793  */
1794 u64 count_range_bits(struct extent_io_tree *tree,
1795                      u64 *start, u64 search_end, u64 max_bytes,
1796                      unsigned bits, int contig)
1797 {
1798         struct rb_node *node;
1799         struct extent_state *state;
1800         u64 cur_start = *start;
1801         u64 total_bytes = 0;
1802         u64 last = 0;
1803         int found = 0;
1804
1805         if (WARN_ON(search_end <= cur_start))
1806                 return 0;
1807
1808         spin_lock(&tree->lock);
1809         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1810                 total_bytes = tree->dirty_bytes;
1811                 goto out;
1812         }
1813         /*
1814          * this search will find all the extents that end after
1815          * our range starts.
1816          */
1817         node = tree_search(tree, cur_start);
1818         if (!node)
1819                 goto out;
1820
1821         while (1) {
1822                 state = rb_entry(node, struct extent_state, rb_node);
1823                 if (state->start > search_end)
1824                         break;
1825                 if (contig && found && state->start > last + 1)
1826                         break;
1827                 if (state->end >= cur_start && (state->state & bits) == bits) {
1828                         total_bytes += min(search_end, state->end) + 1 -
1829                                        max(cur_start, state->start);
1830                         if (total_bytes >= max_bytes)
1831                                 break;
1832                         if (!found) {
1833                                 *start = max(cur_start, state->start);
1834                                 found = 1;
1835                         }
1836                         last = state->end;
1837                 } else if (contig && found) {
1838                         break;
1839                 }
1840                 node = rb_next(node);
1841                 if (!node)
1842                         break;
1843         }
1844 out:
1845         spin_unlock(&tree->lock);
1846         return total_bytes;
1847 }
1848
1849 /*
1850  * set the private field for a given byte offset in the tree.  If there isn't
1851  * an extent_state there already, this does nothing.
1852  */
1853 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1854                 struct io_failure_record *failrec)
1855 {
1856         struct rb_node *node;
1857         struct extent_state *state;
1858         int ret = 0;
1859
1860         spin_lock(&tree->lock);
1861         /*
1862          * this search will find all the extents that end after
1863          * our range starts.
1864          */
1865         node = tree_search(tree, start);
1866         if (!node) {
1867                 ret = -ENOENT;
1868                 goto out;
1869         }
1870         state = rb_entry(node, struct extent_state, rb_node);
1871         if (state->start != start) {
1872                 ret = -ENOENT;
1873                 goto out;
1874         }
1875         state->failrec = failrec;
1876 out:
1877         spin_unlock(&tree->lock);
1878         return ret;
1879 }
1880
1881 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1882                 struct io_failure_record **failrec)
1883 {
1884         struct rb_node *node;
1885         struct extent_state *state;
1886         int ret = 0;
1887
1888         spin_lock(&tree->lock);
1889         /*
1890          * this search will find all the extents that end after
1891          * our range starts.
1892          */
1893         node = tree_search(tree, start);
1894         if (!node) {
1895                 ret = -ENOENT;
1896                 goto out;
1897         }
1898         state = rb_entry(node, struct extent_state, rb_node);
1899         if (state->start != start) {
1900                 ret = -ENOENT;
1901                 goto out;
1902         }
1903         *failrec = state->failrec;
1904 out:
1905         spin_unlock(&tree->lock);
1906         return ret;
1907 }
1908
1909 /*
1910  * searches a range in the state tree for a given mask.
1911  * If 'filled' == 1, this returns 1 only if every extent in the tree
1912  * has the bits set.  Otherwise, 1 is returned if any bit in the
1913  * range is found set.
1914  */
1915 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1916                    unsigned bits, int filled, struct extent_state *cached)
1917 {
1918         struct extent_state *state = NULL;
1919         struct rb_node *node;
1920         int bitset = 0;
1921
1922         spin_lock(&tree->lock);
1923         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1924             cached->end > start)
1925                 node = &cached->rb_node;
1926         else
1927                 node = tree_search(tree, start);
1928         while (node && start <= end) {
1929                 state = rb_entry(node, struct extent_state, rb_node);
1930
1931                 if (filled && state->start > start) {
1932                         bitset = 0;
1933                         break;
1934                 }
1935
1936                 if (state->start > end)
1937                         break;
1938
1939                 if (state->state & bits) {
1940                         bitset = 1;
1941                         if (!filled)
1942                                 break;
1943                 } else if (filled) {
1944                         bitset = 0;
1945                         break;
1946                 }
1947
1948                 if (state->end == (u64)-1)
1949                         break;
1950
1951                 start = state->end + 1;
1952                 if (start > end)
1953                         break;
1954                 node = rb_next(node);
1955                 if (!node) {
1956                         if (filled)
1957                                 bitset = 0;
1958                         break;
1959                 }
1960         }
1961         spin_unlock(&tree->lock);
1962         return bitset;
1963 }
1964
1965 /*
1966  * helper function to set a given page up to date if all the
1967  * extents in the tree for that page are up to date
1968  */
1969 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1970 {
1971         u64 start = page_offset(page);
1972         u64 end = start + PAGE_SIZE - 1;
1973         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1974                 SetPageUptodate(page);
1975 }
1976
1977 int free_io_failure(struct extent_io_tree *failure_tree,
1978                     struct extent_io_tree *io_tree,
1979                     struct io_failure_record *rec)
1980 {
1981         int ret;
1982         int err = 0;
1983
1984         set_state_failrec(failure_tree, rec->start, NULL);
1985         ret = clear_extent_bits(failure_tree, rec->start,
1986                                 rec->start + rec->len - 1,
1987                                 EXTENT_LOCKED | EXTENT_DIRTY);
1988         if (ret)
1989                 err = ret;
1990
1991         ret = clear_extent_bits(io_tree, rec->start,
1992                                 rec->start + rec->len - 1,
1993                                 EXTENT_DAMAGED);
1994         if (ret && !err)
1995                 err = ret;
1996
1997         kfree(rec);
1998         return err;
1999 }
2000
2001 /*
2002  * this bypasses the standard btrfs submit functions deliberately, as
2003  * the standard behavior is to write all copies in a raid setup. here we only
2004  * want to write the one bad copy. so we do the mapping for ourselves and issue
2005  * submit_bio directly.
2006  * to avoid any synchronization issues, wait for the data after writing, which
2007  * actually prevents the read that triggered the error from finishing.
2008  * currently, there can be no more than two copies of every data bit. thus,
2009  * exactly one rewrite is required.
2010  */
2011 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
2012                       u64 length, u64 logical, struct page *page,
2013                       unsigned int pg_offset, int mirror_num)
2014 {
2015         struct bio *bio;
2016         struct btrfs_device *dev;
2017         u64 map_length = 0;
2018         u64 sector;
2019         struct btrfs_bio *bbio = NULL;
2020         int ret;
2021
2022         ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
2023         BUG_ON(!mirror_num);
2024
2025         bio = btrfs_io_bio_alloc(1);
2026         bio->bi_iter.bi_size = 0;
2027         map_length = length;
2028
2029         /*
2030          * Avoid races with device replace and make sure our bbio has devices
2031          * associated to its stripes that don't go away while we are doing the
2032          * read repair operation.
2033          */
2034         btrfs_bio_counter_inc_blocked(fs_info);
2035         if (btrfs_is_parity_mirror(fs_info, logical, length)) {
2036                 /*
2037                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2038                  * to update all raid stripes, but here we just want to correct
2039                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2040                  * stripe's dev and sector.
2041                  */
2042                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2043                                       &map_length, &bbio, 0);
2044                 if (ret) {
2045                         btrfs_bio_counter_dec(fs_info);
2046                         bio_put(bio);
2047                         return -EIO;
2048                 }
2049                 ASSERT(bbio->mirror_num == 1);
2050         } else {
2051                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2052                                       &map_length, &bbio, mirror_num);
2053                 if (ret) {
2054                         btrfs_bio_counter_dec(fs_info);
2055                         bio_put(bio);
2056                         return -EIO;
2057                 }
2058                 BUG_ON(mirror_num != bbio->mirror_num);
2059         }
2060
2061         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2062         bio->bi_iter.bi_sector = sector;
2063         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2064         btrfs_put_bbio(bbio);
2065         if (!dev || !dev->bdev ||
2066             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2067                 btrfs_bio_counter_dec(fs_info);
2068                 bio_put(bio);
2069                 return -EIO;
2070         }
2071         bio_set_dev(bio, dev->bdev);
2072         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2073         bio_add_page(bio, page, length, pg_offset);
2074
2075         if (btrfsic_submit_bio_wait(bio)) {
2076                 /* try to remap that extent elsewhere? */
2077                 btrfs_bio_counter_dec(fs_info);
2078                 bio_put(bio);
2079                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2080                 return -EIO;
2081         }
2082
2083         btrfs_info_rl_in_rcu(fs_info,
2084                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2085                                   ino, start,
2086                                   rcu_str_deref(dev->name), sector);
2087         btrfs_bio_counter_dec(fs_info);
2088         bio_put(bio);
2089         return 0;
2090 }
2091
2092 int btrfs_repair_eb_io_failure(struct extent_buffer *eb, int mirror_num)
2093 {
2094         struct btrfs_fs_info *fs_info = eb->fs_info;
2095         u64 start = eb->start;
2096         int i, num_pages = num_extent_pages(eb);
2097         int ret = 0;
2098
2099         if (sb_rdonly(fs_info->sb))
2100                 return -EROFS;
2101
2102         for (i = 0; i < num_pages; i++) {
2103                 struct page *p = eb->pages[i];
2104
2105                 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2106                                         start - page_offset(p), mirror_num);
2107                 if (ret)
2108                         break;
2109                 start += PAGE_SIZE;
2110         }
2111
2112         return ret;
2113 }
2114
2115 /*
2116  * each time an IO finishes, we do a fast check in the IO failure tree
2117  * to see if we need to process or clean up an io_failure_record
2118  */
2119 int clean_io_failure(struct btrfs_fs_info *fs_info,
2120                      struct extent_io_tree *failure_tree,
2121                      struct extent_io_tree *io_tree, u64 start,
2122                      struct page *page, u64 ino, unsigned int pg_offset)
2123 {
2124         u64 private;
2125         struct io_failure_record *failrec;
2126         struct extent_state *state;
2127         int num_copies;
2128         int ret;
2129
2130         private = 0;
2131         ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2132                                EXTENT_DIRTY, 0);
2133         if (!ret)
2134                 return 0;
2135
2136         ret = get_state_failrec(failure_tree, start, &failrec);
2137         if (ret)
2138                 return 0;
2139
2140         BUG_ON(!failrec->this_mirror);
2141
2142         if (failrec->in_validation) {
2143                 /* there was no real error, just free the record */
2144                 btrfs_debug(fs_info,
2145                         "clean_io_failure: freeing dummy error at %llu",
2146                         failrec->start);
2147                 goto out;
2148         }
2149         if (sb_rdonly(fs_info->sb))
2150                 goto out;
2151
2152         spin_lock(&io_tree->lock);
2153         state = find_first_extent_bit_state(io_tree,
2154                                             failrec->start,
2155                                             EXTENT_LOCKED);
2156         spin_unlock(&io_tree->lock);
2157
2158         if (state && state->start <= failrec->start &&
2159             state->end >= failrec->start + failrec->len - 1) {
2160                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2161                                               failrec->len);
2162                 if (num_copies > 1)  {
2163                         repair_io_failure(fs_info, ino, start, failrec->len,
2164                                           failrec->logical, page, pg_offset,
2165                                           failrec->failed_mirror);
2166                 }
2167         }
2168
2169 out:
2170         free_io_failure(failure_tree, io_tree, failrec);
2171
2172         return 0;
2173 }
2174
2175 /*
2176  * Can be called when
2177  * - hold extent lock
2178  * - under ordered extent
2179  * - the inode is freeing
2180  */
2181 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2182 {
2183         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2184         struct io_failure_record *failrec;
2185         struct extent_state *state, *next;
2186
2187         if (RB_EMPTY_ROOT(&failure_tree->state))
2188                 return;
2189
2190         spin_lock(&failure_tree->lock);
2191         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2192         while (state) {
2193                 if (state->start > end)
2194                         break;
2195
2196                 ASSERT(state->end <= end);
2197
2198                 next = next_state(state);
2199
2200                 failrec = state->failrec;
2201                 free_extent_state(state);
2202                 kfree(failrec);
2203
2204                 state = next;
2205         }
2206         spin_unlock(&failure_tree->lock);
2207 }
2208
2209 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2210                 struct io_failure_record **failrec_ret)
2211 {
2212         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2213         struct io_failure_record *failrec;
2214         struct extent_map *em;
2215         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2216         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2217         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2218         int ret;
2219         u64 logical;
2220
2221         ret = get_state_failrec(failure_tree, start, &failrec);
2222         if (ret) {
2223                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2224                 if (!failrec)
2225                         return -ENOMEM;
2226
2227                 failrec->start = start;
2228                 failrec->len = end - start + 1;
2229                 failrec->this_mirror = 0;
2230                 failrec->bio_flags = 0;
2231                 failrec->in_validation = 0;
2232
2233                 read_lock(&em_tree->lock);
2234                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2235                 if (!em) {
2236                         read_unlock(&em_tree->lock);
2237                         kfree(failrec);
2238                         return -EIO;
2239                 }
2240
2241                 if (em->start > start || em->start + em->len <= start) {
2242                         free_extent_map(em);
2243                         em = NULL;
2244                 }
2245                 read_unlock(&em_tree->lock);
2246                 if (!em) {
2247                         kfree(failrec);
2248                         return -EIO;
2249                 }
2250
2251                 logical = start - em->start;
2252                 logical = em->block_start + logical;
2253                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2254                         logical = em->block_start;
2255                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2256                         extent_set_compress_type(&failrec->bio_flags,
2257                                                  em->compress_type);
2258                 }
2259
2260                 btrfs_debug(fs_info,
2261                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2262                         logical, start, failrec->len);
2263
2264                 failrec->logical = logical;
2265                 free_extent_map(em);
2266
2267                 /* set the bits in the private failure tree */
2268                 ret = set_extent_bits(failure_tree, start, end,
2269                                         EXTENT_LOCKED | EXTENT_DIRTY);
2270                 if (ret >= 0)
2271                         ret = set_state_failrec(failure_tree, start, failrec);
2272                 /* set the bits in the inode's tree */
2273                 if (ret >= 0)
2274                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2275                 if (ret < 0) {
2276                         kfree(failrec);
2277                         return ret;
2278                 }
2279         } else {
2280                 btrfs_debug(fs_info,
2281                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2282                         failrec->logical, failrec->start, failrec->len,
2283                         failrec->in_validation);
2284                 /*
2285                  * when data can be on disk more than twice, add to failrec here
2286                  * (e.g. with a list for failed_mirror) to make
2287                  * clean_io_failure() clean all those errors at once.
2288                  */
2289         }
2290
2291         *failrec_ret = failrec;
2292
2293         return 0;
2294 }
2295
2296 bool btrfs_check_repairable(struct inode *inode, unsigned failed_bio_pages,
2297                            struct io_failure_record *failrec, int failed_mirror)
2298 {
2299         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2300         int num_copies;
2301
2302         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2303         if (num_copies == 1) {
2304                 /*
2305                  * we only have a single copy of the data, so don't bother with
2306                  * all the retry and error correction code that follows. no
2307                  * matter what the error is, it is very likely to persist.
2308                  */
2309                 btrfs_debug(fs_info,
2310                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2311                         num_copies, failrec->this_mirror, failed_mirror);
2312                 return false;
2313         }
2314
2315         /*
2316          * there are two premises:
2317          *      a) deliver good data to the caller
2318          *      b) correct the bad sectors on disk
2319          */
2320         if (failed_bio_pages > 1) {
2321                 /*
2322                  * to fulfill b), we need to know the exact failing sectors, as
2323                  * we don't want to rewrite any more than the failed ones. thus,
2324                  * we need separate read requests for the failed bio
2325                  *
2326                  * if the following BUG_ON triggers, our validation request got
2327                  * merged. we need separate requests for our algorithm to work.
2328                  */
2329                 BUG_ON(failrec->in_validation);
2330                 failrec->in_validation = 1;
2331                 failrec->this_mirror = failed_mirror;
2332         } else {
2333                 /*
2334                  * we're ready to fulfill a) and b) alongside. get a good copy
2335                  * of the failed sector and if we succeed, we have setup
2336                  * everything for repair_io_failure to do the rest for us.
2337                  */
2338                 if (failrec->in_validation) {
2339                         BUG_ON(failrec->this_mirror != failed_mirror);
2340                         failrec->in_validation = 0;
2341                         failrec->this_mirror = 0;
2342                 }
2343                 failrec->failed_mirror = failed_mirror;
2344                 failrec->this_mirror++;
2345                 if (failrec->this_mirror == failed_mirror)
2346                         failrec->this_mirror++;
2347         }
2348
2349         if (failrec->this_mirror > num_copies) {
2350                 btrfs_debug(fs_info,
2351                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2352                         num_copies, failrec->this_mirror, failed_mirror);
2353                 return false;
2354         }
2355
2356         return true;
2357 }
2358
2359
2360 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2361                                     struct io_failure_record *failrec,
2362                                     struct page *page, int pg_offset, int icsum,
2363                                     bio_end_io_t *endio_func, void *data)
2364 {
2365         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2366         struct bio *bio;
2367         struct btrfs_io_bio *btrfs_failed_bio;
2368         struct btrfs_io_bio *btrfs_bio;
2369
2370         bio = btrfs_io_bio_alloc(1);
2371         bio->bi_end_io = endio_func;
2372         bio->bi_iter.bi_sector = failrec->logical >> 9;
2373         bio_set_dev(bio, fs_info->fs_devices->latest_bdev);
2374         bio->bi_iter.bi_size = 0;
2375         bio->bi_private = data;
2376
2377         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2378         if (btrfs_failed_bio->csum) {
2379                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2380
2381                 btrfs_bio = btrfs_io_bio(bio);
2382                 btrfs_bio->csum = btrfs_bio->csum_inline;
2383                 icsum *= csum_size;
2384                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2385                        csum_size);
2386         }
2387
2388         bio_add_page(bio, page, failrec->len, pg_offset);
2389
2390         return bio;
2391 }
2392
2393 /*
2394  * This is a generic handler for readpage errors. If other copies exist, read
2395  * those and write back good data to the failed position. Does not investigate
2396  * in remapping the failed extent elsewhere, hoping the device will be smart
2397  * enough to do this as needed
2398  */
2399 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2400                               struct page *page, u64 start, u64 end,
2401                               int failed_mirror)
2402 {
2403         struct io_failure_record *failrec;
2404         struct inode *inode = page->mapping->host;
2405         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2406         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2407         struct bio *bio;
2408         int read_mode = 0;
2409         blk_status_t status;
2410         int ret;
2411         unsigned failed_bio_pages = failed_bio->bi_iter.bi_size >> PAGE_SHIFT;
2412
2413         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2414
2415         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2416         if (ret)
2417                 return ret;
2418
2419         if (!btrfs_check_repairable(inode, failed_bio_pages, failrec,
2420                                     failed_mirror)) {
2421                 free_io_failure(failure_tree, tree, failrec);
2422                 return -EIO;
2423         }
2424
2425         if (failed_bio_pages > 1)
2426                 read_mode |= REQ_FAILFAST_DEV;
2427
2428         phy_offset >>= inode->i_sb->s_blocksize_bits;
2429         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2430                                       start - page_offset(page),
2431                                       (int)phy_offset, failed_bio->bi_end_io,
2432                                       NULL);
2433         bio->bi_opf = REQ_OP_READ | read_mode;
2434
2435         btrfs_debug(btrfs_sb(inode->i_sb),
2436                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2437                 read_mode, failrec->this_mirror, failrec->in_validation);
2438
2439         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2440                                          failrec->bio_flags, 0);
2441         if (status) {
2442                 free_io_failure(failure_tree, tree, failrec);
2443                 bio_put(bio);
2444                 ret = blk_status_to_errno(status);
2445         }
2446
2447         return ret;
2448 }
2449
2450 /* lots and lots of room for performance fixes in the end_bio funcs */
2451
2452 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2453 {
2454         int uptodate = (err == 0);
2455         int ret = 0;
2456
2457         btrfs_writepage_endio_finish_ordered(page, start, end, uptodate);
2458
2459         if (!uptodate) {
2460                 ClearPageUptodate(page);
2461                 SetPageError(page);
2462                 ret = err < 0 ? err : -EIO;
2463                 mapping_set_error(page->mapping, ret);
2464         }
2465 }
2466
2467 /*
2468  * after a writepage IO is done, we need to:
2469  * clear the uptodate bits on error
2470  * clear the writeback bits in the extent tree for this IO
2471  * end_page_writeback if the page has no more pending IO
2472  *
2473  * Scheduling is not allowed, so the extent state tree is expected
2474  * to have one and only one object corresponding to this IO.
2475  */
2476 static void end_bio_extent_writepage(struct bio *bio)
2477 {
2478         int error = blk_status_to_errno(bio->bi_status);
2479         struct bio_vec *bvec;
2480         u64 start;
2481         u64 end;
2482         int i;
2483         struct bvec_iter_all iter_all;
2484
2485         ASSERT(!bio_flagged(bio, BIO_CLONED));
2486         bio_for_each_segment_all(bvec, bio, i, iter_all) {
2487                 struct page *page = bvec->bv_page;
2488                 struct inode *inode = page->mapping->host;
2489                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2490
2491                 /* We always issue full-page reads, but if some block
2492                  * in a page fails to read, blk_update_request() will
2493                  * advance bv_offset and adjust bv_len to compensate.
2494                  * Print a warning for nonzero offsets, and an error
2495                  * if they don't add up to a full page.  */
2496                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2497                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2498                                 btrfs_err(fs_info,
2499                                    "partial page write in btrfs with offset %u and length %u",
2500                                         bvec->bv_offset, bvec->bv_len);
2501                         else
2502                                 btrfs_info(fs_info,
2503                                    "incomplete page write in btrfs with offset %u and length %u",
2504                                         bvec->bv_offset, bvec->bv_len);
2505                 }
2506
2507                 start = page_offset(page);
2508                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2509
2510                 end_extent_writepage(page, error, start, end);
2511                 end_page_writeback(page);
2512         }
2513
2514         bio_put(bio);
2515 }
2516
2517 static void
2518 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2519                               int uptodate)
2520 {
2521         struct extent_state *cached = NULL;
2522         u64 end = start + len - 1;
2523
2524         if (uptodate && tree->track_uptodate)
2525                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2526         unlock_extent_cached_atomic(tree, start, end, &cached);
2527 }
2528
2529 /*
2530  * after a readpage IO is done, we need to:
2531  * clear the uptodate bits on error
2532  * set the uptodate bits if things worked
2533  * set the page up to date if all extents in the tree are uptodate
2534  * clear the lock bit in the extent tree
2535  * unlock the page if there are no other extents locked for it
2536  *
2537  * Scheduling is not allowed, so the extent state tree is expected
2538  * to have one and only one object corresponding to this IO.
2539  */
2540 static void end_bio_extent_readpage(struct bio *bio)
2541 {
2542         struct bio_vec *bvec;
2543         int uptodate = !bio->bi_status;
2544         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2545         struct extent_io_tree *tree, *failure_tree;
2546         u64 offset = 0;
2547         u64 start;
2548         u64 end;
2549         u64 len;
2550         u64 extent_start = 0;
2551         u64 extent_len = 0;
2552         int mirror;
2553         int ret;
2554         int i;
2555         struct bvec_iter_all iter_all;
2556
2557         ASSERT(!bio_flagged(bio, BIO_CLONED));
2558         bio_for_each_segment_all(bvec, bio, i, iter_all) {
2559                 struct page *page = bvec->bv_page;
2560                 struct inode *inode = page->mapping->host;
2561                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2562                 bool data_inode = btrfs_ino(BTRFS_I(inode))
2563                         != BTRFS_BTREE_INODE_OBJECTID;
2564
2565                 btrfs_debug(fs_info,
2566                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2567                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2568                         io_bio->mirror_num);
2569                 tree = &BTRFS_I(inode)->io_tree;
2570                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2571
2572                 /* We always issue full-page reads, but if some block
2573                  * in a page fails to read, blk_update_request() will
2574                  * advance bv_offset and adjust bv_len to compensate.
2575                  * Print a warning for nonzero offsets, and an error
2576                  * if they don't add up to a full page.  */
2577                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2578                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2579                                 btrfs_err(fs_info,
2580                                         "partial page read in btrfs with offset %u and length %u",
2581                                         bvec->bv_offset, bvec->bv_len);
2582                         else
2583                                 btrfs_info(fs_info,
2584                                         "incomplete page read in btrfs with offset %u and length %u",
2585                                         bvec->bv_offset, bvec->bv_len);
2586                 }
2587
2588                 start = page_offset(page);
2589                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2590                 len = bvec->bv_len;
2591
2592                 mirror = io_bio->mirror_num;
2593                 if (likely(uptodate)) {
2594                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2595                                                               page, start, end,
2596                                                               mirror);
2597                         if (ret)
2598                                 uptodate = 0;
2599                         else
2600                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2601                                                  failure_tree, tree, start,
2602                                                  page,
2603                                                  btrfs_ino(BTRFS_I(inode)), 0);
2604                 }
2605
2606                 if (likely(uptodate))
2607                         goto readpage_ok;
2608
2609                 if (data_inode) {
2610
2611                         /*
2612                          * The generic bio_readpage_error handles errors the
2613                          * following way: If possible, new read requests are
2614                          * created and submitted and will end up in
2615                          * end_bio_extent_readpage as well (if we're lucky,
2616                          * not in the !uptodate case). In that case it returns
2617                          * 0 and we just go on with the next page in our bio.
2618                          * If it can't handle the error it will return -EIO and
2619                          * we remain responsible for that page.
2620                          */
2621                         ret = bio_readpage_error(bio, offset, page, start, end,
2622                                                  mirror);
2623                         if (ret == 0) {
2624                                 uptodate = !bio->bi_status;
2625                                 offset += len;
2626                                 continue;
2627                         }
2628                 } else {
2629                         struct extent_buffer *eb;
2630
2631                         eb = (struct extent_buffer *)page->private;
2632                         set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
2633                         eb->read_mirror = mirror;
2634                         atomic_dec(&eb->io_pages);
2635                         if (test_and_clear_bit(EXTENT_BUFFER_READAHEAD,
2636                                                &eb->bflags))
2637                                 btree_readahead_hook(eb, -EIO);
2638                 }
2639 readpage_ok:
2640                 if (likely(uptodate)) {
2641                         loff_t i_size = i_size_read(inode);
2642                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2643                         unsigned off;
2644
2645                         /* Zero out the end if this page straddles i_size */
2646                         off = offset_in_page(i_size);
2647                         if (page->index == end_index && off)
2648                                 zero_user_segment(page, off, PAGE_SIZE);
2649                         SetPageUptodate(page);
2650                 } else {
2651                         ClearPageUptodate(page);
2652                         SetPageError(page);
2653                 }
2654                 unlock_page(page);
2655                 offset += len;
2656
2657                 if (unlikely(!uptodate)) {
2658                         if (extent_len) {
2659                                 endio_readpage_release_extent(tree,
2660                                                               extent_start,
2661                                                               extent_len, 1);
2662                                 extent_start = 0;
2663                                 extent_len = 0;
2664                         }
2665                         endio_readpage_release_extent(tree, start,
2666                                                       end - start + 1, 0);
2667                 } else if (!extent_len) {
2668                         extent_start = start;
2669                         extent_len = end + 1 - start;
2670                 } else if (extent_start + extent_len == start) {
2671                         extent_len += end + 1 - start;
2672                 } else {
2673                         endio_readpage_release_extent(tree, extent_start,
2674                                                       extent_len, uptodate);
2675                         extent_start = start;
2676                         extent_len = end + 1 - start;
2677                 }
2678         }
2679
2680         if (extent_len)
2681                 endio_readpage_release_extent(tree, extent_start, extent_len,
2682                                               uptodate);
2683         btrfs_io_bio_free_csum(io_bio);
2684         bio_put(bio);
2685 }
2686
2687 /*
2688  * Initialize the members up to but not including 'bio'. Use after allocating a
2689  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2690  * 'bio' because use of __GFP_ZERO is not supported.
2691  */
2692 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2693 {
2694         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2695 }
2696
2697 /*
2698  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2699  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2700  * for the appropriate container_of magic
2701  */
2702 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte)
2703 {
2704         struct bio *bio;
2705
2706         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, &btrfs_bioset);
2707         bio_set_dev(bio, bdev);
2708         bio->bi_iter.bi_sector = first_byte >> 9;
2709         btrfs_io_bio_init(btrfs_io_bio(bio));
2710         return bio;
2711 }
2712
2713 struct bio *btrfs_bio_clone(struct bio *bio)
2714 {
2715         struct btrfs_io_bio *btrfs_bio;
2716         struct bio *new;
2717
2718         /* Bio allocation backed by a bioset does not fail */
2719         new = bio_clone_fast(bio, GFP_NOFS, &btrfs_bioset);
2720         btrfs_bio = btrfs_io_bio(new);
2721         btrfs_io_bio_init(btrfs_bio);
2722         btrfs_bio->iter = bio->bi_iter;
2723         return new;
2724 }
2725
2726 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2727 {
2728         struct bio *bio;
2729
2730         /* Bio allocation backed by a bioset does not fail */
2731         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, &btrfs_bioset);
2732         btrfs_io_bio_init(btrfs_io_bio(bio));
2733         return bio;
2734 }
2735
2736 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2737 {
2738         struct bio *bio;
2739         struct btrfs_io_bio *btrfs_bio;
2740
2741         /* this will never fail when it's backed by a bioset */
2742         bio = bio_clone_fast(orig, GFP_NOFS, &btrfs_bioset);
2743         ASSERT(bio);
2744
2745         btrfs_bio = btrfs_io_bio(bio);
2746         btrfs_io_bio_init(btrfs_bio);
2747
2748         bio_trim(bio, offset >> 9, size >> 9);
2749         btrfs_bio->iter = bio->bi_iter;
2750         return bio;
2751 }
2752
2753 /*
2754  * @opf:        bio REQ_OP_* and REQ_* flags as one value
2755  * @tree:       tree so we can call our merge_bio hook
2756  * @wbc:        optional writeback control for io accounting
2757  * @page:       page to add to the bio
2758  * @pg_offset:  offset of the new bio or to check whether we are adding
2759  *              a contiguous page to the previous one
2760  * @size:       portion of page that we want to write
2761  * @offset:     starting offset in the page
2762  * @bdev:       attach newly created bios to this bdev
2763  * @bio_ret:    must be valid pointer, newly allocated bio will be stored there
2764  * @end_io_func:     end_io callback for new bio
2765  * @mirror_num:      desired mirror to read/write
2766  * @prev_bio_flags:  flags of previous bio to see if we can merge the current one
2767  * @bio_flags:  flags of the current bio to see if we can merge them
2768  */
2769 static int submit_extent_page(unsigned int opf, struct extent_io_tree *tree,
2770                               struct writeback_control *wbc,
2771                               struct page *page, u64 offset,
2772                               size_t size, unsigned long pg_offset,
2773                               struct block_device *bdev,
2774                               struct bio **bio_ret,
2775                               bio_end_io_t end_io_func,
2776                               int mirror_num,
2777                               unsigned long prev_bio_flags,
2778                               unsigned long bio_flags,
2779                               bool force_bio_submit)
2780 {
2781         int ret = 0;
2782         struct bio *bio;
2783         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2784         sector_t sector = offset >> 9;
2785
2786         ASSERT(bio_ret);
2787
2788         if (*bio_ret) {
2789                 bool contig;
2790                 bool can_merge = true;
2791
2792                 bio = *bio_ret;
2793                 if (prev_bio_flags & EXTENT_BIO_COMPRESSED)
2794                         contig = bio->bi_iter.bi_sector == sector;
2795                 else
2796                         contig = bio_end_sector(bio) == sector;
2797
2798                 ASSERT(tree->ops);
2799                 if (btrfs_bio_fits_in_stripe(page, page_size, bio, bio_flags))
2800                         can_merge = false;
2801
2802                 if (prev_bio_flags != bio_flags || !contig || !can_merge ||
2803                     force_bio_submit ||
2804                     bio_add_page(bio, page, page_size, pg_offset) < page_size) {
2805                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2806                         if (ret < 0) {
2807                                 *bio_ret = NULL;
2808                                 return ret;
2809                         }
2810                         bio = NULL;
2811                 } else {
2812                         if (wbc)
2813                                 wbc_account_io(wbc, page, page_size);
2814                         return 0;
2815                 }
2816         }
2817
2818         bio = btrfs_bio_alloc(bdev, offset);
2819         bio_add_page(bio, page, page_size, pg_offset);
2820         bio->bi_end_io = end_io_func;
2821         bio->bi_private = tree;
2822         bio->bi_write_hint = page->mapping->host->i_write_hint;
2823         bio->bi_opf = opf;
2824         if (wbc) {
2825                 wbc_init_bio(wbc, bio);
2826                 wbc_account_io(wbc, page, page_size);
2827         }
2828
2829         *bio_ret = bio;
2830
2831         return ret;
2832 }
2833
2834 static void attach_extent_buffer_page(struct extent_buffer *eb,
2835                                       struct page *page)
2836 {
2837         if (!PagePrivate(page)) {
2838                 SetPagePrivate(page);
2839                 get_page(page);
2840                 set_page_private(page, (unsigned long)eb);
2841         } else {
2842                 WARN_ON(page->private != (unsigned long)eb);
2843         }
2844 }
2845
2846 void set_page_extent_mapped(struct page *page)
2847 {
2848         if (!PagePrivate(page)) {
2849                 SetPagePrivate(page);
2850                 get_page(page);
2851                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2852         }
2853 }
2854
2855 static struct extent_map *
2856 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2857                  u64 start, u64 len, get_extent_t *get_extent,
2858                  struct extent_map **em_cached)
2859 {
2860         struct extent_map *em;
2861
2862         if (em_cached && *em_cached) {
2863                 em = *em_cached;
2864                 if (extent_map_in_tree(em) && start >= em->start &&
2865                     start < extent_map_end(em)) {
2866                         refcount_inc(&em->refs);
2867                         return em;
2868                 }
2869
2870                 free_extent_map(em);
2871                 *em_cached = NULL;
2872         }
2873
2874         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2875         if (em_cached && !IS_ERR_OR_NULL(em)) {
2876                 BUG_ON(*em_cached);
2877                 refcount_inc(&em->refs);
2878                 *em_cached = em;
2879         }
2880         return em;
2881 }
2882 /*
2883  * basic readpage implementation.  Locked extent state structs are inserted
2884  * into the tree that are removed when the IO is done (by the end_io
2885  * handlers)
2886  * XXX JDM: This needs looking at to ensure proper page locking
2887  * return 0 on success, otherwise return error
2888  */
2889 static int __do_readpage(struct extent_io_tree *tree,
2890                          struct page *page,
2891                          get_extent_t *get_extent,
2892                          struct extent_map **em_cached,
2893                          struct bio **bio, int mirror_num,
2894                          unsigned long *bio_flags, unsigned int read_flags,
2895                          u64 *prev_em_start)
2896 {
2897         struct inode *inode = page->mapping->host;
2898         u64 start = page_offset(page);
2899         const u64 end = start + PAGE_SIZE - 1;
2900         u64 cur = start;
2901         u64 extent_offset;
2902         u64 last_byte = i_size_read(inode);
2903         u64 block_start;
2904         u64 cur_end;
2905         struct extent_map *em;
2906         struct block_device *bdev;
2907         int ret = 0;
2908         int nr = 0;
2909         size_t pg_offset = 0;
2910         size_t iosize;
2911         size_t disk_io_size;
2912         size_t blocksize = inode->i_sb->s_blocksize;
2913         unsigned long this_bio_flag = 0;
2914
2915         set_page_extent_mapped(page);
2916
2917         if (!PageUptodate(page)) {
2918                 if (cleancache_get_page(page) == 0) {
2919                         BUG_ON(blocksize != PAGE_SIZE);
2920                         unlock_extent(tree, start, end);
2921                         goto out;
2922                 }
2923         }
2924
2925         if (page->index == last_byte >> PAGE_SHIFT) {
2926                 char *userpage;
2927                 size_t zero_offset = offset_in_page(last_byte);
2928
2929                 if (zero_offset) {
2930                         iosize = PAGE_SIZE - zero_offset;
2931                         userpage = kmap_atomic(page);
2932                         memset(userpage + zero_offset, 0, iosize);
2933                         flush_dcache_page(page);
2934                         kunmap_atomic(userpage);
2935                 }
2936         }
2937         while (cur <= end) {
2938                 bool force_bio_submit = false;
2939                 u64 offset;
2940
2941                 if (cur >= last_byte) {
2942                         char *userpage;
2943                         struct extent_state *cached = NULL;
2944
2945                         iosize = PAGE_SIZE - pg_offset;
2946                         userpage = kmap_atomic(page);
2947                         memset(userpage + pg_offset, 0, iosize);
2948                         flush_dcache_page(page);
2949                         kunmap_atomic(userpage);
2950                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2951                                             &cached, GFP_NOFS);
2952                         unlock_extent_cached(tree, cur,
2953                                              cur + iosize - 1, &cached);
2954                         break;
2955                 }
2956                 em = __get_extent_map(inode, page, pg_offset, cur,
2957                                       end - cur + 1, get_extent, em_cached);
2958                 if (IS_ERR_OR_NULL(em)) {
2959                         SetPageError(page);
2960                         unlock_extent(tree, cur, end);
2961                         break;
2962                 }
2963                 extent_offset = cur - em->start;
2964                 BUG_ON(extent_map_end(em) <= cur);
2965                 BUG_ON(end < cur);
2966
2967                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2968                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2969                         extent_set_compress_type(&this_bio_flag,
2970                                                  em->compress_type);
2971                 }
2972
2973                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2974                 cur_end = min(extent_map_end(em) - 1, end);
2975                 iosize = ALIGN(iosize, blocksize);
2976                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2977                         disk_io_size = em->block_len;
2978                         offset = em->block_start;
2979                 } else {
2980                         offset = em->block_start + extent_offset;
2981                         disk_io_size = iosize;
2982                 }
2983                 bdev = em->bdev;
2984                 block_start = em->block_start;
2985                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2986                         block_start = EXTENT_MAP_HOLE;
2987
2988                 /*
2989                  * If we have a file range that points to a compressed extent
2990                  * and it's followed by a consecutive file range that points to
2991                  * to the same compressed extent (possibly with a different
2992                  * offset and/or length, so it either points to the whole extent
2993                  * or only part of it), we must make sure we do not submit a
2994                  * single bio to populate the pages for the 2 ranges because
2995                  * this makes the compressed extent read zero out the pages
2996                  * belonging to the 2nd range. Imagine the following scenario:
2997                  *
2998                  *  File layout
2999                  *  [0 - 8K]                     [8K - 24K]
3000                  *    |                               |
3001                  *    |                               |
3002                  * points to extent X,         points to extent X,
3003                  * offset 4K, length of 8K     offset 0, length 16K
3004                  *
3005                  * [extent X, compressed length = 4K uncompressed length = 16K]
3006                  *
3007                  * If the bio to read the compressed extent covers both ranges,
3008                  * it will decompress extent X into the pages belonging to the
3009                  * first range and then it will stop, zeroing out the remaining
3010                  * pages that belong to the other range that points to extent X.
3011                  * So here we make sure we submit 2 bios, one for the first
3012                  * range and another one for the third range. Both will target
3013                  * the same physical extent from disk, but we can't currently
3014                  * make the compressed bio endio callback populate the pages
3015                  * for both ranges because each compressed bio is tightly
3016                  * coupled with a single extent map, and each range can have
3017                  * an extent map with a different offset value relative to the
3018                  * uncompressed data of our extent and different lengths. This
3019                  * is a corner case so we prioritize correctness over
3020                  * non-optimal behavior (submitting 2 bios for the same extent).
3021                  */
3022                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3023                     prev_em_start && *prev_em_start != (u64)-1 &&
3024                     *prev_em_start != em->start)
3025                         force_bio_submit = true;
3026
3027                 if (prev_em_start)
3028                         *prev_em_start = em->start;
3029
3030                 free_extent_map(em);
3031                 em = NULL;
3032
3033                 /* we've found a hole, just zero and go on */
3034                 if (block_start == EXTENT_MAP_HOLE) {
3035                         char *userpage;
3036                         struct extent_state *cached = NULL;
3037
3038                         userpage = kmap_atomic(page);
3039                         memset(userpage + pg_offset, 0, iosize);
3040                         flush_dcache_page(page);
3041                         kunmap_atomic(userpage);
3042
3043                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3044                                             &cached, GFP_NOFS);
3045                         unlock_extent_cached(tree, cur,
3046                                              cur + iosize - 1, &cached);
3047                         cur = cur + iosize;
3048                         pg_offset += iosize;
3049                         continue;
3050                 }
3051                 /* the get_extent function already copied into the page */
3052                 if (test_range_bit(tree, cur, cur_end,
3053                                    EXTENT_UPTODATE, 1, NULL)) {
3054                         check_page_uptodate(tree, page);
3055                         unlock_extent(tree, cur, cur + iosize - 1);
3056                         cur = cur + iosize;
3057                         pg_offset += iosize;
3058                         continue;
3059                 }
3060                 /* we have an inline extent but it didn't get marked up
3061                  * to date.  Error out
3062                  */
3063                 if (block_start == EXTENT_MAP_INLINE) {
3064                         SetPageError(page);
3065                         unlock_extent(tree, cur, cur + iosize - 1);
3066                         cur = cur + iosize;
3067                         pg_offset += iosize;
3068                         continue;
3069                 }
3070
3071                 ret = submit_extent_page(REQ_OP_READ | read_flags, tree, NULL,
3072                                          page, offset, disk_io_size,
3073                                          pg_offset, bdev, bio,
3074                                          end_bio_extent_readpage, mirror_num,
3075                                          *bio_flags,
3076                                          this_bio_flag,
3077                                          force_bio_submit);
3078                 if (!ret) {
3079                         nr++;
3080                         *bio_flags = this_bio_flag;
3081                 } else {
3082                         SetPageError(page);
3083                         unlock_extent(tree, cur, cur + iosize - 1);
3084                         goto out;
3085                 }
3086                 cur = cur + iosize;
3087                 pg_offset += iosize;
3088         }
3089 out:
3090         if (!nr) {
3091                 if (!PageError(page))
3092                         SetPageUptodate(page);
3093                 unlock_page(page);
3094         }
3095         return ret;
3096 }
3097
3098 static inline void contiguous_readpages(struct extent_io_tree *tree,
3099                                              struct page *pages[], int nr_pages,
3100                                              u64 start, u64 end,
3101                                              struct extent_map **em_cached,
3102                                              struct bio **bio,
3103                                              unsigned long *bio_flags,
3104                                              u64 *prev_em_start)
3105 {
3106         struct inode *inode;
3107         struct btrfs_ordered_extent *ordered;
3108         int index;
3109
3110         inode = pages[0]->mapping->host;
3111         while (1) {
3112                 lock_extent(tree, start, end);
3113                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3114                                                      end - start + 1);
3115                 if (!ordered)
3116                         break;
3117                 unlock_extent(tree, start, end);
3118                 btrfs_start_ordered_extent(inode, ordered, 1);
3119                 btrfs_put_ordered_extent(ordered);
3120         }
3121
3122         for (index = 0; index < nr_pages; index++) {
3123                 __do_readpage(tree, pages[index], btrfs_get_extent, em_cached,
3124                                 bio, 0, bio_flags, REQ_RAHEAD, prev_em_start);
3125                 put_page(pages[index]);
3126         }
3127 }