Btrfs: incremental send, fix wrong unlink path after renaming file
[sfrench/cifs-2.6.git] / fs / btrfs / extent_io.c
1 #include <linux/bitops.h>
2 #include <linux/slab.h>
3 #include <linux/bio.h>
4 #include <linux/mm.h>
5 #include <linux/pagemap.h>
6 #include <linux/page-flags.h>
7 #include <linux/spinlock.h>
8 #include <linux/blkdev.h>
9 #include <linux/swap.h>
10 #include <linux/writeback.h>
11 #include <linux/pagevec.h>
12 #include <linux/prefetch.h>
13 #include <linux/cleancache.h>
14 #include "extent_io.h"
15 #include "extent_map.h"
16 #include "ctree.h"
17 #include "btrfs_inode.h"
18 #include "volumes.h"
19 #include "check-integrity.h"
20 #include "locking.h"
21 #include "rcu-string.h"
22 #include "backref.h"
23
24 static struct kmem_cache *extent_state_cache;
25 static struct kmem_cache *extent_buffer_cache;
26 static struct bio_set *btrfs_bioset;
27
28 static inline bool extent_state_in_tree(const struct extent_state *state)
29 {
30         return !RB_EMPTY_NODE(&state->rb_node);
31 }
32
33 #ifdef CONFIG_BTRFS_DEBUG
34 static LIST_HEAD(buffers);
35 static LIST_HEAD(states);
36
37 static DEFINE_SPINLOCK(leak_lock);
38
39 static inline
40 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
41 {
42         unsigned long flags;
43
44         spin_lock_irqsave(&leak_lock, flags);
45         list_add(new, head);
46         spin_unlock_irqrestore(&leak_lock, flags);
47 }
48
49 static inline
50 void btrfs_leak_debug_del(struct list_head *entry)
51 {
52         unsigned long flags;
53
54         spin_lock_irqsave(&leak_lock, flags);
55         list_del(entry);
56         spin_unlock_irqrestore(&leak_lock, flags);
57 }
58
59 static inline
60 void btrfs_leak_debug_check(void)
61 {
62         struct extent_state *state;
63         struct extent_buffer *eb;
64
65         while (!list_empty(&states)) {
66                 state = list_entry(states.next, struct extent_state, leak_list);
67                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
68                        state->start, state->end, state->state,
69                        extent_state_in_tree(state),
70                        refcount_read(&state->refs));
71                 list_del(&state->leak_list);
72                 kmem_cache_free(extent_state_cache, state);
73         }
74
75         while (!list_empty(&buffers)) {
76                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
77                 pr_err("BTRFS: buffer leak start %llu len %lu refs %d\n",
78                        eb->start, eb->len, atomic_read(&eb->refs));
79                 list_del(&eb->leak_list);
80                 kmem_cache_free(extent_buffer_cache, eb);
81         }
82 }
83
84 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
85         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
86 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
87                 struct extent_io_tree *tree, u64 start, u64 end)
88 {
89         if (tree->ops && tree->ops->check_extent_io_range)
90                 tree->ops->check_extent_io_range(tree->private_data, caller,
91                                                  start, end);
92 }
93 #else
94 #define btrfs_leak_debug_add(new, head) do {} while (0)
95 #define btrfs_leak_debug_del(entry)     do {} while (0)
96 #define btrfs_leak_debug_check()        do {} while (0)
97 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
98 #endif
99
100 #define BUFFER_LRU_MAX 64
101
102 struct tree_entry {
103         u64 start;
104         u64 end;
105         struct rb_node rb_node;
106 };
107
108 struct extent_page_data {
109         struct bio *bio;
110         struct extent_io_tree *tree;
111         get_extent_t *get_extent;
112
113         /* tells writepage not to lock the state bits for this range
114          * it still does the unlocking
115          */
116         unsigned int extent_locked:1;
117
118         /* tells the submit_bio code to use REQ_SYNC */
119         unsigned int sync_io:1;
120 };
121
122 static void add_extent_changeset(struct extent_state *state, unsigned bits,
123                                  struct extent_changeset *changeset,
124                                  int set)
125 {
126         int ret;
127
128         if (!changeset)
129                 return;
130         if (set && (state->state & bits) == bits)
131                 return;
132         if (!set && (state->state & bits) == 0)
133                 return;
134         changeset->bytes_changed += state->end - state->start + 1;
135         ret = ulist_add(&changeset->range_changed, state->start, state->end,
136                         GFP_ATOMIC);
137         /* ENOMEM */
138         BUG_ON(ret < 0);
139 }
140
141 static noinline void flush_write_bio(void *data);
142 static inline struct btrfs_fs_info *
143 tree_fs_info(struct extent_io_tree *tree)
144 {
145         if (tree->ops)
146                 return tree->ops->tree_fs_info(tree->private_data);
147         return NULL;
148 }
149
150 int __init extent_io_init(void)
151 {
152         extent_state_cache = kmem_cache_create("btrfs_extent_state",
153                         sizeof(struct extent_state), 0,
154                         SLAB_MEM_SPREAD, NULL);
155         if (!extent_state_cache)
156                 return -ENOMEM;
157
158         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
159                         sizeof(struct extent_buffer), 0,
160                         SLAB_MEM_SPREAD, NULL);
161         if (!extent_buffer_cache)
162                 goto free_state_cache;
163
164         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
165                                      offsetof(struct btrfs_io_bio, bio),
166                                      BIOSET_NEED_BVECS);
167         if (!btrfs_bioset)
168                 goto free_buffer_cache;
169
170         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
171                 goto free_bioset;
172
173         return 0;
174
175 free_bioset:
176         bioset_free(btrfs_bioset);
177         btrfs_bioset = NULL;
178
179 free_buffer_cache:
180         kmem_cache_destroy(extent_buffer_cache);
181         extent_buffer_cache = NULL;
182
183 free_state_cache:
184         kmem_cache_destroy(extent_state_cache);
185         extent_state_cache = NULL;
186         return -ENOMEM;
187 }
188
189 void extent_io_exit(void)
190 {
191         btrfs_leak_debug_check();
192
193         /*
194          * Make sure all delayed rcu free are flushed before we
195          * destroy caches.
196          */
197         rcu_barrier();
198         kmem_cache_destroy(extent_state_cache);
199         kmem_cache_destroy(extent_buffer_cache);
200         if (btrfs_bioset)
201                 bioset_free(btrfs_bioset);
202 }
203
204 void extent_io_tree_init(struct extent_io_tree *tree,
205                          void *private_data)
206 {
207         tree->state = RB_ROOT;
208         tree->ops = NULL;
209         tree->dirty_bytes = 0;
210         spin_lock_init(&tree->lock);
211         tree->private_data = private_data;
212 }
213
214 static struct extent_state *alloc_extent_state(gfp_t mask)
215 {
216         struct extent_state *state;
217
218         /*
219          * The given mask might be not appropriate for the slab allocator,
220          * drop the unsupported bits
221          */
222         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
223         state = kmem_cache_alloc(extent_state_cache, mask);
224         if (!state)
225                 return state;
226         state->state = 0;
227         state->failrec = NULL;
228         RB_CLEAR_NODE(&state->rb_node);
229         btrfs_leak_debug_add(&state->leak_list, &states);
230         refcount_set(&state->refs, 1);
231         init_waitqueue_head(&state->wq);
232         trace_alloc_extent_state(state, mask, _RET_IP_);
233         return state;
234 }
235
236 void free_extent_state(struct extent_state *state)
237 {
238         if (!state)
239                 return;
240         if (refcount_dec_and_test(&state->refs)) {
241                 WARN_ON(extent_state_in_tree(state));
242                 btrfs_leak_debug_del(&state->leak_list);
243                 trace_free_extent_state(state, _RET_IP_);
244                 kmem_cache_free(extent_state_cache, state);
245         }
246 }
247
248 static struct rb_node *tree_insert(struct rb_root *root,
249                                    struct rb_node *search_start,
250                                    u64 offset,
251                                    struct rb_node *node,
252                                    struct rb_node ***p_in,
253                                    struct rb_node **parent_in)
254 {
255         struct rb_node **p;
256         struct rb_node *parent = NULL;
257         struct tree_entry *entry;
258
259         if (p_in && parent_in) {
260                 p = *p_in;
261                 parent = *parent_in;
262                 goto do_insert;
263         }
264
265         p = search_start ? &search_start : &root->rb_node;
266         while (*p) {
267                 parent = *p;
268                 entry = rb_entry(parent, struct tree_entry, rb_node);
269
270                 if (offset < entry->start)
271                         p = &(*p)->rb_left;
272                 else if (offset > entry->end)
273                         p = &(*p)->rb_right;
274                 else
275                         return parent;
276         }
277
278 do_insert:
279         rb_link_node(node, parent, p);
280         rb_insert_color(node, root);
281         return NULL;
282 }
283
284 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
285                                       struct rb_node **prev_ret,
286                                       struct rb_node **next_ret,
287                                       struct rb_node ***p_ret,
288                                       struct rb_node **parent_ret)
289 {
290         struct rb_root *root = &tree->state;
291         struct rb_node **n = &root->rb_node;
292         struct rb_node *prev = NULL;
293         struct rb_node *orig_prev = NULL;
294         struct tree_entry *entry;
295         struct tree_entry *prev_entry = NULL;
296
297         while (*n) {
298                 prev = *n;
299                 entry = rb_entry(prev, struct tree_entry, rb_node);
300                 prev_entry = entry;
301
302                 if (offset < entry->start)
303                         n = &(*n)->rb_left;
304                 else if (offset > entry->end)
305                         n = &(*n)->rb_right;
306                 else
307                         return *n;
308         }
309
310         if (p_ret)
311                 *p_ret = n;
312         if (parent_ret)
313                 *parent_ret = prev;
314
315         if (prev_ret) {
316                 orig_prev = prev;
317                 while (prev && offset > prev_entry->end) {
318                         prev = rb_next(prev);
319                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
320                 }
321                 *prev_ret = prev;
322                 prev = orig_prev;
323         }
324
325         if (next_ret) {
326                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
327                 while (prev && offset < prev_entry->start) {
328                         prev = rb_prev(prev);
329                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
330                 }
331                 *next_ret = prev;
332         }
333         return NULL;
334 }
335
336 static inline struct rb_node *
337 tree_search_for_insert(struct extent_io_tree *tree,
338                        u64 offset,
339                        struct rb_node ***p_ret,
340                        struct rb_node **parent_ret)
341 {
342         struct rb_node *prev = NULL;
343         struct rb_node *ret;
344
345         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
346         if (!ret)
347                 return prev;
348         return ret;
349 }
350
351 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
352                                           u64 offset)
353 {
354         return tree_search_for_insert(tree, offset, NULL, NULL);
355 }
356
357 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
358                      struct extent_state *other)
359 {
360         if (tree->ops && tree->ops->merge_extent_hook)
361                 tree->ops->merge_extent_hook(tree->private_data, new, other);
362 }
363
364 /*
365  * utility function to look for merge candidates inside a given range.
366  * Any extents with matching state are merged together into a single
367  * extent in the tree.  Extents with EXTENT_IO in their state field
368  * are not merged because the end_io handlers need to be able to do
369  * operations on them without sleeping (or doing allocations/splits).
370  *
371  * This should be called with the tree lock held.
372  */
373 static void merge_state(struct extent_io_tree *tree,
374                         struct extent_state *state)
375 {
376         struct extent_state *other;
377         struct rb_node *other_node;
378
379         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
380                 return;
381
382         other_node = rb_prev(&state->rb_node);
383         if (other_node) {
384                 other = rb_entry(other_node, struct extent_state, rb_node);
385                 if (other->end == state->start - 1 &&
386                     other->state == state->state) {
387                         merge_cb(tree, state, other);
388                         state->start = other->start;
389                         rb_erase(&other->rb_node, &tree->state);
390                         RB_CLEAR_NODE(&other->rb_node);
391                         free_extent_state(other);
392                 }
393         }
394         other_node = rb_next(&state->rb_node);
395         if (other_node) {
396                 other = rb_entry(other_node, struct extent_state, rb_node);
397                 if (other->start == state->end + 1 &&
398                     other->state == state->state) {
399                         merge_cb(tree, state, other);
400                         state->end = other->end;
401                         rb_erase(&other->rb_node, &tree->state);
402                         RB_CLEAR_NODE(&other->rb_node);
403                         free_extent_state(other);
404                 }
405         }
406 }
407
408 static void set_state_cb(struct extent_io_tree *tree,
409                          struct extent_state *state, unsigned *bits)
410 {
411         if (tree->ops && tree->ops->set_bit_hook)
412                 tree->ops->set_bit_hook(tree->private_data, state, bits);
413 }
414
415 static void clear_state_cb(struct extent_io_tree *tree,
416                            struct extent_state *state, unsigned *bits)
417 {
418         if (tree->ops && tree->ops->clear_bit_hook)
419                 tree->ops->clear_bit_hook(tree->private_data, state, bits);
420 }
421
422 static void set_state_bits(struct extent_io_tree *tree,
423                            struct extent_state *state, unsigned *bits,
424                            struct extent_changeset *changeset);
425
426 /*
427  * insert an extent_state struct into the tree.  'bits' are set on the
428  * struct before it is inserted.
429  *
430  * This may return -EEXIST if the extent is already there, in which case the
431  * state struct is freed.
432  *
433  * The tree lock is not taken internally.  This is a utility function and
434  * probably isn't what you want to call (see set/clear_extent_bit).
435  */
436 static int insert_state(struct extent_io_tree *tree,
437                         struct extent_state *state, u64 start, u64 end,
438                         struct rb_node ***p,
439                         struct rb_node **parent,
440                         unsigned *bits, struct extent_changeset *changeset)
441 {
442         struct rb_node *node;
443
444         if (end < start)
445                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
446                        end, start);
447         state->start = start;
448         state->end = end;
449
450         set_state_bits(tree, state, bits, changeset);
451
452         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
453         if (node) {
454                 struct extent_state *found;
455                 found = rb_entry(node, struct extent_state, rb_node);
456                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
457                        found->start, found->end, start, end);
458                 return -EEXIST;
459         }
460         merge_state(tree, state);
461         return 0;
462 }
463
464 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
465                      u64 split)
466 {
467         if (tree->ops && tree->ops->split_extent_hook)
468                 tree->ops->split_extent_hook(tree->private_data, orig, split);
469 }
470
471 /*
472  * split a given extent state struct in two, inserting the preallocated
473  * struct 'prealloc' as the newly created second half.  'split' indicates an
474  * offset inside 'orig' where it should be split.
475  *
476  * Before calling,
477  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
478  * are two extent state structs in the tree:
479  * prealloc: [orig->start, split - 1]
480  * orig: [ split, orig->end ]
481  *
482  * The tree locks are not taken by this function. They need to be held
483  * by the caller.
484  */
485 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
486                        struct extent_state *prealloc, u64 split)
487 {
488         struct rb_node *node;
489
490         split_cb(tree, orig, split);
491
492         prealloc->start = orig->start;
493         prealloc->end = split - 1;
494         prealloc->state = orig->state;
495         orig->start = split;
496
497         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
498                            &prealloc->rb_node, NULL, NULL);
499         if (node) {
500                 free_extent_state(prealloc);
501                 return -EEXIST;
502         }
503         return 0;
504 }
505
506 static struct extent_state *next_state(struct extent_state *state)
507 {
508         struct rb_node *next = rb_next(&state->rb_node);
509         if (next)
510                 return rb_entry(next, struct extent_state, rb_node);
511         else
512                 return NULL;
513 }
514
515 /*
516  * utility function to clear some bits in an extent state struct.
517  * it will optionally wake up any one waiting on this state (wake == 1).
518  *
519  * If no bits are set on the state struct after clearing things, the
520  * struct is freed and removed from the tree
521  */
522 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
523                                             struct extent_state *state,
524                                             unsigned *bits, int wake,
525                                             struct extent_changeset *changeset)
526 {
527         struct extent_state *next;
528         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
529
530         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
531                 u64 range = state->end - state->start + 1;
532                 WARN_ON(range > tree->dirty_bytes);
533                 tree->dirty_bytes -= range;
534         }
535         clear_state_cb(tree, state, bits);
536         add_extent_changeset(state, bits_to_clear, changeset, 0);
537         state->state &= ~bits_to_clear;
538         if (wake)
539                 wake_up(&state->wq);
540         if (state->state == 0) {
541                 next = next_state(state);
542                 if (extent_state_in_tree(state)) {
543                         rb_erase(&state->rb_node, &tree->state);
544                         RB_CLEAR_NODE(&state->rb_node);
545                         free_extent_state(state);
546                 } else {
547                         WARN_ON(1);
548                 }
549         } else {
550                 merge_state(tree, state);
551                 next = next_state(state);
552         }
553         return next;
554 }
555
556 static struct extent_state *
557 alloc_extent_state_atomic(struct extent_state *prealloc)
558 {
559         if (!prealloc)
560                 prealloc = alloc_extent_state(GFP_ATOMIC);
561
562         return prealloc;
563 }
564
565 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
566 {
567         btrfs_panic(tree_fs_info(tree), err,
568                     "Locking error: Extent tree was modified by another thread while locked.");
569 }
570
571 /*
572  * clear some bits on a range in the tree.  This may require splitting
573  * or inserting elements in the tree, so the gfp mask is used to
574  * indicate which allocations or sleeping are allowed.
575  *
576  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
577  * the given range from the tree regardless of state (ie for truncate).
578  *
579  * the range [start, end] is inclusive.
580  *
581  * This takes the tree lock, and returns 0 on success and < 0 on error.
582  */
583 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
584                               unsigned bits, int wake, int delete,
585                               struct extent_state **cached_state,
586                               gfp_t mask, struct extent_changeset *changeset)
587 {
588         struct extent_state *state;
589         struct extent_state *cached;
590         struct extent_state *prealloc = NULL;
591         struct rb_node *node;
592         u64 last_end;
593         int err;
594         int clear = 0;
595
596         btrfs_debug_check_extent_io_range(tree, start, end);
597
598         if (bits & EXTENT_DELALLOC)
599                 bits |= EXTENT_NORESERVE;
600
601         if (delete)
602                 bits |= ~EXTENT_CTLBITS;
603         bits |= EXTENT_FIRST_DELALLOC;
604
605         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
606                 clear = 1;
607 again:
608         if (!prealloc && gfpflags_allow_blocking(mask)) {
609                 /*
610                  * Don't care for allocation failure here because we might end
611                  * up not needing the pre-allocated extent state at all, which
612                  * is the case if we only have in the tree extent states that
613                  * cover our input range and don't cover too any other range.
614                  * If we end up needing a new extent state we allocate it later.
615                  */
616                 prealloc = alloc_extent_state(mask);
617         }
618
619         spin_lock(&tree->lock);
620         if (cached_state) {
621                 cached = *cached_state;
622
623                 if (clear) {
624                         *cached_state = NULL;
625                         cached_state = NULL;
626                 }
627
628                 if (cached && extent_state_in_tree(cached) &&
629                     cached->start <= start && cached->end > start) {
630                         if (clear)
631                                 refcount_dec(&cached->refs);
632                         state = cached;
633                         goto hit_next;
634                 }
635                 if (clear)
636                         free_extent_state(cached);
637         }
638         /*
639          * this search will find the extents that end after
640          * our range starts
641          */
642         node = tree_search(tree, start);
643         if (!node)
644                 goto out;
645         state = rb_entry(node, struct extent_state, rb_node);
646 hit_next:
647         if (state->start > end)
648                 goto out;
649         WARN_ON(state->end < start);
650         last_end = state->end;
651
652         /* the state doesn't have the wanted bits, go ahead */
653         if (!(state->state & bits)) {
654                 state = next_state(state);
655                 goto next;
656         }
657
658         /*
659          *     | ---- desired range ---- |
660          *  | state | or
661          *  | ------------- state -------------- |
662          *
663          * We need to split the extent we found, and may flip
664          * bits on second half.
665          *
666          * If the extent we found extends past our range, we
667          * just split and search again.  It'll get split again
668          * the next time though.
669          *
670          * If the extent we found is inside our range, we clear
671          * the desired bit on it.
672          */
673
674         if (state->start < start) {
675                 prealloc = alloc_extent_state_atomic(prealloc);
676                 BUG_ON(!prealloc);
677                 err = split_state(tree, state, prealloc, start);
678                 if (err)
679                         extent_io_tree_panic(tree, err);
680
681                 prealloc = NULL;
682                 if (err)
683                         goto out;
684                 if (state->end <= end) {
685                         state = clear_state_bit(tree, state, &bits, wake,
686                                                 changeset);
687                         goto next;
688                 }
689                 goto search_again;
690         }
691         /*
692          * | ---- desired range ---- |
693          *                        | state |
694          * We need to split the extent, and clear the bit
695          * on the first half
696          */
697         if (state->start <= end && state->end > end) {
698                 prealloc = alloc_extent_state_atomic(prealloc);
699                 BUG_ON(!prealloc);
700                 err = split_state(tree, state, prealloc, end + 1);
701                 if (err)
702                         extent_io_tree_panic(tree, err);
703
704                 if (wake)
705                         wake_up(&state->wq);
706
707                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
708
709                 prealloc = NULL;
710                 goto out;
711         }
712
713         state = clear_state_bit(tree, state, &bits, wake, changeset);
714 next:
715         if (last_end == (u64)-1)
716                 goto out;
717         start = last_end + 1;
718         if (start <= end && state && !need_resched())
719                 goto hit_next;
720
721 search_again:
722         if (start > end)
723                 goto out;
724         spin_unlock(&tree->lock);
725         if (gfpflags_allow_blocking(mask))
726                 cond_resched();
727         goto again;
728
729 out:
730         spin_unlock(&tree->lock);
731         if (prealloc)
732                 free_extent_state(prealloc);
733
734         return 0;
735
736 }
737
738 static void wait_on_state(struct extent_io_tree *tree,
739                           struct extent_state *state)
740                 __releases(tree->lock)
741                 __acquires(tree->lock)
742 {
743         DEFINE_WAIT(wait);
744         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
745         spin_unlock(&tree->lock);
746         schedule();
747         spin_lock(&tree->lock);
748         finish_wait(&state->wq, &wait);
749 }
750
751 /*
752  * waits for one or more bits to clear on a range in the state tree.
753  * The range [start, end] is inclusive.
754  * The tree lock is taken by this function
755  */
756 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
757                             unsigned long bits)
758 {
759         struct extent_state *state;
760         struct rb_node *node;
761
762         btrfs_debug_check_extent_io_range(tree, start, end);
763
764         spin_lock(&tree->lock);
765 again:
766         while (1) {
767                 /*
768                  * this search will find all the extents that end after
769                  * our range starts
770                  */
771                 node = tree_search(tree, start);
772 process_node:
773                 if (!node)
774                         break;
775
776                 state = rb_entry(node, struct extent_state, rb_node);
777
778                 if (state->start > end)
779                         goto out;
780
781                 if (state->state & bits) {
782                         start = state->start;
783                         refcount_inc(&state->refs);
784                         wait_on_state(tree, state);
785                         free_extent_state(state);
786                         goto again;
787                 }
788                 start = state->end + 1;
789
790                 if (start > end)
791                         break;
792
793                 if (!cond_resched_lock(&tree->lock)) {
794                         node = rb_next(node);
795                         goto process_node;
796                 }
797         }
798 out:
799         spin_unlock(&tree->lock);
800 }
801
802 static void set_state_bits(struct extent_io_tree *tree,
803                            struct extent_state *state,
804                            unsigned *bits, struct extent_changeset *changeset)
805 {
806         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
807
808         set_state_cb(tree, state, bits);
809         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
810                 u64 range = state->end - state->start + 1;
811                 tree->dirty_bytes += range;
812         }
813         add_extent_changeset(state, bits_to_set, changeset, 1);
814         state->state |= bits_to_set;
815 }
816
817 static void cache_state_if_flags(struct extent_state *state,
818                                  struct extent_state **cached_ptr,
819                                  unsigned flags)
820 {
821         if (cached_ptr && !(*cached_ptr)) {
822                 if (!flags || (state->state & flags)) {
823                         *cached_ptr = state;
824                         refcount_inc(&state->refs);
825                 }
826         }
827 }
828
829 static void cache_state(struct extent_state *state,
830                         struct extent_state **cached_ptr)
831 {
832         return cache_state_if_flags(state, cached_ptr,
833                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
834 }
835
836 /*
837  * set some bits on a range in the tree.  This may require allocations or
838  * sleeping, so the gfp mask is used to indicate what is allowed.
839  *
840  * If any of the exclusive bits are set, this will fail with -EEXIST if some
841  * part of the range already has the desired bits set.  The start of the
842  * existing range is returned in failed_start in this case.
843  *
844  * [start, end] is inclusive This takes the tree lock.
845  */
846
847 static int __must_check
848 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
849                  unsigned bits, unsigned exclusive_bits,
850                  u64 *failed_start, struct extent_state **cached_state,
851                  gfp_t mask, struct extent_changeset *changeset)
852 {
853         struct extent_state *state;
854         struct extent_state *prealloc = NULL;
855         struct rb_node *node;
856         struct rb_node **p;
857         struct rb_node *parent;
858         int err = 0;
859         u64 last_start;
860         u64 last_end;
861
862         btrfs_debug_check_extent_io_range(tree, start, end);
863
864         bits |= EXTENT_FIRST_DELALLOC;
865 again:
866         if (!prealloc && gfpflags_allow_blocking(mask)) {
867                 /*
868                  * Don't care for allocation failure here because we might end
869                  * up not needing the pre-allocated extent state at all, which
870                  * is the case if we only have in the tree extent states that
871                  * cover our input range and don't cover too any other range.
872                  * If we end up needing a new extent state we allocate it later.
873                  */
874                 prealloc = alloc_extent_state(mask);
875         }
876
877         spin_lock(&tree->lock);
878         if (cached_state && *cached_state) {
879                 state = *cached_state;
880                 if (state->start <= start && state->end > start &&
881                     extent_state_in_tree(state)) {
882                         node = &state->rb_node;
883                         goto hit_next;
884                 }
885         }
886         /*
887          * this search will find all the extents that end after
888          * our range starts.
889          */
890         node = tree_search_for_insert(tree, start, &p, &parent);
891         if (!node) {
892                 prealloc = alloc_extent_state_atomic(prealloc);
893                 BUG_ON(!prealloc);
894                 err = insert_state(tree, prealloc, start, end,
895                                    &p, &parent, &bits, changeset);
896                 if (err)
897                         extent_io_tree_panic(tree, err);
898
899                 cache_state(prealloc, cached_state);
900                 prealloc = NULL;
901                 goto out;
902         }
903         state = rb_entry(node, struct extent_state, rb_node);
904 hit_next:
905         last_start = state->start;
906         last_end = state->end;
907
908         /*
909          * | ---- desired range ---- |
910          * | state |
911          *
912          * Just lock what we found and keep going
913          */
914         if (state->start == start && state->end <= end) {
915                 if (state->state & exclusive_bits) {
916                         *failed_start = state->start;
917                         err = -EEXIST;
918                         goto out;
919                 }
920
921                 set_state_bits(tree, state, &bits, changeset);
922                 cache_state(state, cached_state);
923                 merge_state(tree, state);
924                 if (last_end == (u64)-1)
925                         goto out;
926                 start = last_end + 1;
927                 state = next_state(state);
928                 if (start < end && state && state->start == start &&
929                     !need_resched())
930                         goto hit_next;
931                 goto search_again;
932         }
933
934         /*
935          *     | ---- desired range ---- |
936          * | state |
937          *   or
938          * | ------------- state -------------- |
939          *
940          * We need to split the extent we found, and may flip bits on
941          * second half.
942          *
943          * If the extent we found extends past our
944          * range, we just split and search again.  It'll get split
945          * again the next time though.
946          *
947          * If the extent we found is inside our range, we set the
948          * desired bit on it.
949          */
950         if (state->start < start) {
951                 if (state->state & exclusive_bits) {
952                         *failed_start = start;
953                         err = -EEXIST;
954                         goto out;
955                 }
956
957                 prealloc = alloc_extent_state_atomic(prealloc);
958                 BUG_ON(!prealloc);
959                 err = split_state(tree, state, prealloc, start);
960                 if (err)
961                         extent_io_tree_panic(tree, err);
962
963                 prealloc = NULL;
964                 if (err)
965                         goto out;
966                 if (state->end <= end) {
967                         set_state_bits(tree, state, &bits, changeset);
968                         cache_state(state, cached_state);
969                         merge_state(tree, state);
970                         if (last_end == (u64)-1)
971                                 goto out;
972                         start = last_end + 1;
973                         state = next_state(state);
974                         if (start < end && state && state->start == start &&
975                             !need_resched())
976                                 goto hit_next;
977                 }
978                 goto search_again;
979         }
980         /*
981          * | ---- desired range ---- |
982          *     | state | or               | state |
983          *
984          * There's a hole, we need to insert something in it and
985          * ignore the extent we found.
986          */
987         if (state->start > start) {
988                 u64 this_end;
989                 if (end < last_start)
990                         this_end = end;
991                 else
992                         this_end = last_start - 1;
993
994                 prealloc = alloc_extent_state_atomic(prealloc);
995                 BUG_ON(!prealloc);
996
997                 /*
998                  * Avoid to free 'prealloc' if it can be merged with
999                  * the later extent.
1000                  */
1001                 err = insert_state(tree, prealloc, start, this_end,
1002                                    NULL, NULL, &bits, changeset);
1003                 if (err)
1004                         extent_io_tree_panic(tree, err);
1005
1006                 cache_state(prealloc, cached_state);
1007                 prealloc = NULL;
1008                 start = this_end + 1;
1009                 goto search_again;
1010         }
1011         /*
1012          * | ---- desired range ---- |
1013          *                        | state |
1014          * We need to split the extent, and set the bit
1015          * on the first half
1016          */
1017         if (state->start <= end && state->end > end) {
1018                 if (state->state & exclusive_bits) {
1019                         *failed_start = start;
1020                         err = -EEXIST;
1021                         goto out;
1022                 }
1023
1024                 prealloc = alloc_extent_state_atomic(prealloc);
1025                 BUG_ON(!prealloc);
1026                 err = split_state(tree, state, prealloc, end + 1);
1027                 if (err)
1028                         extent_io_tree_panic(tree, err);
1029
1030                 set_state_bits(tree, prealloc, &bits, changeset);
1031                 cache_state(prealloc, cached_state);
1032                 merge_state(tree, prealloc);
1033                 prealloc = NULL;
1034                 goto out;
1035         }
1036
1037 search_again:
1038         if (start > end)
1039                 goto out;
1040         spin_unlock(&tree->lock);
1041         if (gfpflags_allow_blocking(mask))
1042                 cond_resched();
1043         goto again;
1044
1045 out:
1046         spin_unlock(&tree->lock);
1047         if (prealloc)
1048                 free_extent_state(prealloc);
1049
1050         return err;
1051
1052 }
1053
1054 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1055                    unsigned bits, u64 * failed_start,
1056                    struct extent_state **cached_state, gfp_t mask)
1057 {
1058         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1059                                 cached_state, mask, NULL);
1060 }
1061
1062
1063 /**
1064  * convert_extent_bit - convert all bits in a given range from one bit to
1065  *                      another
1066  * @tree:       the io tree to search
1067  * @start:      the start offset in bytes
1068  * @end:        the end offset in bytes (inclusive)
1069  * @bits:       the bits to set in this range
1070  * @clear_bits: the bits to clear in this range
1071  * @cached_state:       state that we're going to cache
1072  *
1073  * This will go through and set bits for the given range.  If any states exist
1074  * already in this range they are set with the given bit and cleared of the
1075  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1076  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1077  * boundary bits like LOCK.
1078  *
1079  * All allocations are done with GFP_NOFS.
1080  */
1081 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1082                        unsigned bits, unsigned clear_bits,
1083                        struct extent_state **cached_state)
1084 {
1085         struct extent_state *state;
1086         struct extent_state *prealloc = NULL;
1087         struct rb_node *node;
1088         struct rb_node **p;
1089         struct rb_node *parent;
1090         int err = 0;
1091         u64 last_start;
1092         u64 last_end;
1093         bool first_iteration = true;
1094
1095         btrfs_debug_check_extent_io_range(tree, start, end);
1096
1097 again:
1098         if (!prealloc) {
1099                 /*
1100                  * Best effort, don't worry if extent state allocation fails
1101                  * here for the first iteration. We might have a cached state
1102                  * that matches exactly the target range, in which case no
1103                  * extent state allocations are needed. We'll only know this
1104                  * after locking the tree.
1105                  */
1106                 prealloc = alloc_extent_state(GFP_NOFS);
1107                 if (!prealloc && !first_iteration)
1108                         return -ENOMEM;
1109         }
1110
1111         spin_lock(&tree->lock);
1112         if (cached_state && *cached_state) {
1113                 state = *cached_state;
1114                 if (state->start <= start && state->end > start &&
1115                     extent_state_in_tree(state)) {
1116                         node = &state->rb_node;
1117                         goto hit_next;
1118                 }
1119         }
1120
1121         /*
1122          * this search will find all the extents that end after
1123          * our range starts.
1124          */
1125         node = tree_search_for_insert(tree, start, &p, &parent);
1126         if (!node) {
1127                 prealloc = alloc_extent_state_atomic(prealloc);
1128                 if (!prealloc) {
1129                         err = -ENOMEM;
1130                         goto out;
1131                 }
1132                 err = insert_state(tree, prealloc, start, end,
1133                                    &p, &parent, &bits, NULL);
1134                 if (err)
1135                         extent_io_tree_panic(tree, err);
1136                 cache_state(prealloc, cached_state);
1137                 prealloc = NULL;
1138                 goto out;
1139         }
1140         state = rb_entry(node, struct extent_state, rb_node);
1141 hit_next:
1142         last_start = state->start;
1143         last_end = state->end;
1144
1145         /*
1146          * | ---- desired range ---- |
1147          * | state |
1148          *
1149          * Just lock what we found and keep going
1150          */
1151         if (state->start == start && state->end <= end) {
1152                 set_state_bits(tree, state, &bits, NULL);
1153                 cache_state(state, cached_state);
1154                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1155                 if (last_end == (u64)-1)
1156                         goto out;
1157                 start = last_end + 1;
1158                 if (start < end && state && state->start == start &&
1159                     !need_resched())
1160                         goto hit_next;
1161                 goto search_again;
1162         }
1163
1164         /*
1165          *     | ---- desired range ---- |
1166          * | state |
1167          *   or
1168          * | ------------- state -------------- |
1169          *
1170          * We need to split the extent we found, and may flip bits on
1171          * second half.
1172          *
1173          * If the extent we found extends past our
1174          * range, we just split and search again.  It'll get split
1175          * again the next time though.
1176          *
1177          * If the extent we found is inside our range, we set the
1178          * desired bit on it.
1179          */
1180         if (state->start < start) {
1181                 prealloc = alloc_extent_state_atomic(prealloc);
1182                 if (!prealloc) {
1183                         err = -ENOMEM;
1184                         goto out;
1185                 }
1186                 err = split_state(tree, state, prealloc, start);
1187                 if (err)
1188                         extent_io_tree_panic(tree, err);
1189                 prealloc = NULL;
1190                 if (err)
1191                         goto out;
1192                 if (state->end <= end) {
1193                         set_state_bits(tree, state, &bits, NULL);
1194                         cache_state(state, cached_state);
1195                         state = clear_state_bit(tree, state, &clear_bits, 0,
1196                                                 NULL);
1197                         if (last_end == (u64)-1)
1198                                 goto out;
1199                         start = last_end + 1;
1200                         if (start < end && state && state->start == start &&
1201                             !need_resched())
1202                                 goto hit_next;
1203                 }
1204                 goto search_again;
1205         }
1206         /*
1207          * | ---- desired range ---- |
1208          *     | state | or               | state |
1209          *
1210          * There's a hole, we need to insert something in it and
1211          * ignore the extent we found.
1212          */
1213         if (state->start > start) {
1214                 u64 this_end;
1215                 if (end < last_start)
1216                         this_end = end;
1217                 else
1218                         this_end = last_start - 1;
1219
1220                 prealloc = alloc_extent_state_atomic(prealloc);
1221                 if (!prealloc) {
1222                         err = -ENOMEM;
1223                         goto out;
1224                 }
1225
1226                 /*
1227                  * Avoid to free 'prealloc' if it can be merged with
1228                  * the later extent.
1229                  */
1230                 err = insert_state(tree, prealloc, start, this_end,
1231                                    NULL, NULL, &bits, NULL);
1232                 if (err)
1233                         extent_io_tree_panic(tree, err);
1234                 cache_state(prealloc, cached_state);
1235                 prealloc = NULL;
1236                 start = this_end + 1;
1237                 goto search_again;
1238         }
1239         /*
1240          * | ---- desired range ---- |
1241          *                        | state |
1242          * We need to split the extent, and set the bit
1243          * on the first half
1244          */
1245         if (state->start <= end && state->end > end) {
1246                 prealloc = alloc_extent_state_atomic(prealloc);
1247                 if (!prealloc) {
1248                         err = -ENOMEM;
1249                         goto out;
1250                 }
1251
1252                 err = split_state(tree, state, prealloc, end + 1);
1253                 if (err)
1254                         extent_io_tree_panic(tree, err);
1255
1256                 set_state_bits(tree, prealloc, &bits, NULL);
1257                 cache_state(prealloc, cached_state);
1258                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1259                 prealloc = NULL;
1260                 goto out;
1261         }
1262
1263 search_again:
1264         if (start > end)
1265                 goto out;
1266         spin_unlock(&tree->lock);
1267         cond_resched();
1268         first_iteration = false;
1269         goto again;
1270
1271 out:
1272         spin_unlock(&tree->lock);
1273         if (prealloc)
1274                 free_extent_state(prealloc);
1275
1276         return err;
1277 }
1278
1279 /* wrappers around set/clear extent bit */
1280 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1281                            unsigned bits, struct extent_changeset *changeset)
1282 {
1283         /*
1284          * We don't support EXTENT_LOCKED yet, as current changeset will
1285          * record any bits changed, so for EXTENT_LOCKED case, it will
1286          * either fail with -EEXIST or changeset will record the whole
1287          * range.
1288          */
1289         BUG_ON(bits & EXTENT_LOCKED);
1290
1291         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1292                                 changeset);
1293 }
1294
1295 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1296                      unsigned bits, int wake, int delete,
1297                      struct extent_state **cached, gfp_t mask)
1298 {
1299         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1300                                   cached, mask, NULL);
1301 }
1302
1303 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1304                 unsigned bits, struct extent_changeset *changeset)
1305 {
1306         /*
1307          * Don't support EXTENT_LOCKED case, same reason as
1308          * set_record_extent_bits().
1309          */
1310         BUG_ON(bits & EXTENT_LOCKED);
1311
1312         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1313                                   changeset);
1314 }
1315
1316 /*
1317  * either insert or lock state struct between start and end use mask to tell
1318  * us if waiting is desired.
1319  */
1320 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1321                      struct extent_state **cached_state)
1322 {
1323         int err;
1324         u64 failed_start;
1325
1326         while (1) {
1327                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1328                                        EXTENT_LOCKED, &failed_start,
1329                                        cached_state, GFP_NOFS, NULL);
1330                 if (err == -EEXIST) {
1331                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1332                         start = failed_start;
1333                 } else
1334                         break;
1335                 WARN_ON(start > end);
1336         }
1337         return err;
1338 }
1339
1340 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1341 {
1342         int err;
1343         u64 failed_start;
1344
1345         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1346                                &failed_start, NULL, GFP_NOFS, NULL);
1347         if (err == -EEXIST) {
1348                 if (failed_start > start)
1349                         clear_extent_bit(tree, start, failed_start - 1,
1350                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1351                 return 0;
1352         }
1353         return 1;
1354 }
1355
1356 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1357 {
1358         unsigned long index = start >> PAGE_SHIFT;
1359         unsigned long end_index = end >> PAGE_SHIFT;
1360         struct page *page;
1361
1362         while (index <= end_index) {
1363                 page = find_get_page(inode->i_mapping, index);
1364                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1365                 clear_page_dirty_for_io(page);
1366                 put_page(page);
1367                 index++;
1368         }
1369 }
1370
1371 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1372 {
1373         unsigned long index = start >> PAGE_SHIFT;
1374         unsigned long end_index = end >> PAGE_SHIFT;
1375         struct page *page;
1376
1377         while (index <= end_index) {
1378                 page = find_get_page(inode->i_mapping, index);
1379                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1380                 __set_page_dirty_nobuffers(page);
1381                 account_page_redirty(page);
1382                 put_page(page);
1383                 index++;
1384         }
1385 }
1386
1387 /*
1388  * helper function to set both pages and extents in the tree writeback
1389  */
1390 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1391 {
1392         tree->ops->set_range_writeback(tree->private_data, start, end);
1393 }
1394
1395 /* find the first state struct with 'bits' set after 'start', and
1396  * return it.  tree->lock must be held.  NULL will returned if
1397  * nothing was found after 'start'
1398  */
1399 static struct extent_state *
1400 find_first_extent_bit_state(struct extent_io_tree *tree,
1401                             u64 start, unsigned bits)
1402 {
1403         struct rb_node *node;
1404         struct extent_state *state;
1405
1406         /*
1407          * this search will find all the extents that end after
1408          * our range starts.
1409          */
1410         node = tree_search(tree, start);
1411         if (!node)
1412                 goto out;
1413
1414         while (1) {
1415                 state = rb_entry(node, struct extent_state, rb_node);
1416                 if (state->end >= start && (state->state & bits))
1417                         return state;
1418
1419                 node = rb_next(node);
1420                 if (!node)
1421                         break;
1422         }
1423 out:
1424         return NULL;
1425 }
1426
1427 /*
1428  * find the first offset in the io tree with 'bits' set. zero is
1429  * returned if we find something, and *start_ret and *end_ret are
1430  * set to reflect the state struct that was found.
1431  *
1432  * If nothing was found, 1 is returned. If found something, return 0.
1433  */
1434 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1435                           u64 *start_ret, u64 *end_ret, unsigned bits,
1436                           struct extent_state **cached_state)
1437 {
1438         struct extent_state *state;
1439         struct rb_node *n;
1440         int ret = 1;
1441
1442         spin_lock(&tree->lock);
1443         if (cached_state && *cached_state) {
1444                 state = *cached_state;
1445                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1446                         n = rb_next(&state->rb_node);
1447                         while (n) {
1448                                 state = rb_entry(n, struct extent_state,
1449                                                  rb_node);
1450                                 if (state->state & bits)
1451                                         goto got_it;
1452                                 n = rb_next(n);
1453                         }
1454                         free_extent_state(*cached_state);
1455                         *cached_state = NULL;
1456                         goto out;
1457                 }
1458                 free_extent_state(*cached_state);
1459                 *cached_state = NULL;
1460         }
1461
1462         state = find_first_extent_bit_state(tree, start, bits);
1463 got_it:
1464         if (state) {
1465                 cache_state_if_flags(state, cached_state, 0);
1466                 *start_ret = state->start;
1467                 *end_ret = state->end;
1468                 ret = 0;
1469         }
1470 out:
1471         spin_unlock(&tree->lock);
1472         return ret;
1473 }
1474
1475 /*
1476  * find a contiguous range of bytes in the file marked as delalloc, not
1477  * more than 'max_bytes'.  start and end are used to return the range,
1478  *
1479  * 1 is returned if we find something, 0 if nothing was in the tree
1480  */
1481 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1482                                         u64 *start, u64 *end, u64 max_bytes,
1483                                         struct extent_state **cached_state)
1484 {
1485         struct rb_node *node;
1486         struct extent_state *state;
1487         u64 cur_start = *start;
1488         u64 found = 0;
1489         u64 total_bytes = 0;
1490
1491         spin_lock(&tree->lock);
1492
1493         /*
1494          * this search will find all the extents that end after
1495          * our range starts.
1496          */
1497         node = tree_search(tree, cur_start);
1498         if (!node) {
1499                 if (!found)
1500                         *end = (u64)-1;
1501                 goto out;
1502         }
1503
1504         while (1) {
1505                 state = rb_entry(node, struct extent_state, rb_node);
1506                 if (found && (state->start != cur_start ||
1507                               (state->state & EXTENT_BOUNDARY))) {
1508                         goto out;
1509                 }
1510                 if (!(state->state & EXTENT_DELALLOC)) {
1511                         if (!found)
1512                                 *end = state->end;
1513                         goto out;
1514                 }
1515                 if (!found) {
1516                         *start = state->start;
1517                         *cached_state = state;
1518                         refcount_inc(&state->refs);
1519                 }
1520                 found++;
1521                 *end = state->end;
1522                 cur_start = state->end + 1;
1523                 node = rb_next(node);
1524                 total_bytes += state->end - state->start + 1;
1525                 if (total_bytes >= max_bytes)
1526                         break;
1527                 if (!node)
1528                         break;
1529         }
1530 out:
1531         spin_unlock(&tree->lock);
1532         return found;
1533 }
1534
1535 static int __process_pages_contig(struct address_space *mapping,
1536                                   struct page *locked_page,
1537                                   pgoff_t start_index, pgoff_t end_index,
1538                                   unsigned long page_ops, pgoff_t *index_ret);
1539
1540 static noinline void __unlock_for_delalloc(struct inode *inode,
1541                                            struct page *locked_page,
1542                                            u64 start, u64 end)
1543 {
1544         unsigned long index = start >> PAGE_SHIFT;
1545         unsigned long end_index = end >> PAGE_SHIFT;
1546
1547         ASSERT(locked_page);
1548         if (index == locked_page->index && end_index == index)
1549                 return;
1550
1551         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1552                                PAGE_UNLOCK, NULL);
1553 }
1554
1555 static noinline int lock_delalloc_pages(struct inode *inode,
1556                                         struct page *locked_page,
1557                                         u64 delalloc_start,
1558                                         u64 delalloc_end)
1559 {
1560         unsigned long index = delalloc_start >> PAGE_SHIFT;
1561         unsigned long index_ret = index;
1562         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1563         int ret;
1564
1565         ASSERT(locked_page);
1566         if (index == locked_page->index && index == end_index)
1567                 return 0;
1568
1569         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1570                                      end_index, PAGE_LOCK, &index_ret);
1571         if (ret == -EAGAIN)
1572                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1573                                       (u64)index_ret << PAGE_SHIFT);
1574         return ret;
1575 }
1576
1577 /*
1578  * find a contiguous range of bytes in the file marked as delalloc, not
1579  * more than 'max_bytes'.  start and end are used to return the range,
1580  *
1581  * 1 is returned if we find something, 0 if nothing was in the tree
1582  */
1583 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1584                                     struct extent_io_tree *tree,
1585                                     struct page *locked_page, u64 *start,
1586                                     u64 *end, u64 max_bytes)
1587 {
1588         u64 delalloc_start;
1589         u64 delalloc_end;
1590         u64 found;
1591         struct extent_state *cached_state = NULL;
1592         int ret;
1593         int loops = 0;
1594
1595 again:
1596         /* step one, find a bunch of delalloc bytes starting at start */
1597         delalloc_start = *start;
1598         delalloc_end = 0;
1599         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1600                                     max_bytes, &cached_state);
1601         if (!found || delalloc_end <= *start) {
1602                 *start = delalloc_start;
1603                 *end = delalloc_end;
1604                 free_extent_state(cached_state);
1605                 return 0;
1606         }
1607
1608         /*
1609          * start comes from the offset of locked_page.  We have to lock
1610          * pages in order, so we can't process delalloc bytes before
1611          * locked_page
1612          */
1613         if (delalloc_start < *start)
1614                 delalloc_start = *start;
1615
1616         /*
1617          * make sure to limit the number of pages we try to lock down
1618          */
1619         if (delalloc_end + 1 - delalloc_start > max_bytes)
1620                 delalloc_end = delalloc_start + max_bytes - 1;
1621
1622         /* step two, lock all the pages after the page that has start */
1623         ret = lock_delalloc_pages(inode, locked_page,
1624                                   delalloc_start, delalloc_end);
1625         if (ret == -EAGAIN) {
1626                 /* some of the pages are gone, lets avoid looping by
1627                  * shortening the size of the delalloc range we're searching
1628                  */
1629                 free_extent_state(cached_state);
1630                 cached_state = NULL;
1631                 if (!loops) {
1632                         max_bytes = PAGE_SIZE;
1633                         loops = 1;
1634                         goto again;
1635                 } else {
1636                         found = 0;
1637                         goto out_failed;
1638                 }
1639         }
1640         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1641
1642         /* step three, lock the state bits for the whole range */
1643         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1644
1645         /* then test to make sure it is all still delalloc */
1646         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1647                              EXTENT_DELALLOC, 1, cached_state);
1648         if (!ret) {
1649                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1650                                      &cached_state, GFP_NOFS);
1651                 __unlock_for_delalloc(inode, locked_page,
1652                               delalloc_start, delalloc_end);
1653                 cond_resched();
1654                 goto again;
1655         }
1656         free_extent_state(cached_state);
1657         *start = delalloc_start;
1658         *end = delalloc_end;
1659 out_failed:
1660         return found;
1661 }
1662
1663 static int __process_pages_contig(struct address_space *mapping,
1664                                   struct page *locked_page,
1665                                   pgoff_t start_index, pgoff_t end_index,
1666                                   unsigned long page_ops, pgoff_t *index_ret)
1667 {
1668         unsigned long nr_pages = end_index - start_index + 1;
1669         unsigned long pages_locked = 0;
1670         pgoff_t index = start_index;
1671         struct page *pages[16];
1672         unsigned ret;
1673         int err = 0;
1674         int i;
1675
1676         if (page_ops & PAGE_LOCK) {
1677                 ASSERT(page_ops == PAGE_LOCK);
1678                 ASSERT(index_ret && *index_ret == start_index);
1679         }
1680
1681         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1682                 mapping_set_error(mapping, -EIO);
1683
1684         while (nr_pages > 0) {
1685                 ret = find_get_pages_contig(mapping, index,
1686                                      min_t(unsigned long,
1687                                      nr_pages, ARRAY_SIZE(pages)), pages);
1688                 if (ret == 0) {
1689                         /*
1690                          * Only if we're going to lock these pages,
1691                          * can we find nothing at @index.
1692                          */
1693                         ASSERT(page_ops & PAGE_LOCK);
1694                         err = -EAGAIN;
1695                         goto out;
1696                 }
1697
1698                 for (i = 0; i < ret; i++) {
1699                         if (page_ops & PAGE_SET_PRIVATE2)
1700                                 SetPagePrivate2(pages[i]);
1701
1702                         if (pages[i] == locked_page) {
1703                                 put_page(pages[i]);
1704                                 pages_locked++;
1705                                 continue;
1706                         }
1707                         if (page_ops & PAGE_CLEAR_DIRTY)
1708                                 clear_page_dirty_for_io(pages[i]);
1709                         if (page_ops & PAGE_SET_WRITEBACK)
1710                                 set_page_writeback(pages[i]);
1711                         if (page_ops & PAGE_SET_ERROR)
1712                                 SetPageError(pages[i]);
1713                         if (page_ops & PAGE_END_WRITEBACK)
1714                                 end_page_writeback(pages[i]);
1715                         if (page_ops & PAGE_UNLOCK)
1716                                 unlock_page(pages[i]);
1717                         if (page_ops & PAGE_LOCK) {
1718                                 lock_page(pages[i]);
1719                                 if (!PageDirty(pages[i]) ||
1720                                     pages[i]->mapping != mapping) {
1721                                         unlock_page(pages[i]);
1722                                         put_page(pages[i]);
1723                                         err = -EAGAIN;
1724                                         goto out;
1725                                 }
1726                         }
1727                         put_page(pages[i]);
1728                         pages_locked++;
1729                 }
1730                 nr_pages -= ret;
1731                 index += ret;
1732                 cond_resched();
1733         }
1734 out:
1735         if (err && index_ret)
1736                 *index_ret = start_index + pages_locked - 1;
1737         return err;
1738 }
1739
1740 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1741                                  u64 delalloc_end, struct page *locked_page,
1742                                  unsigned clear_bits,
1743                                  unsigned long page_ops)
1744 {
1745         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1746                          NULL, GFP_NOFS);
1747
1748         __process_pages_contig(inode->i_mapping, locked_page,
1749                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1750                                page_ops, NULL);
1751 }
1752
1753 /*
1754  * count the number of bytes in the tree that have a given bit(s)
1755  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1756  * cached.  The total number found is returned.
1757  */
1758 u64 count_range_bits(struct extent_io_tree *tree,
1759                      u64 *start, u64 search_end, u64 max_bytes,
1760                      unsigned bits, int contig)
1761 {
1762         struct rb_node *node;
1763         struct extent_state *state;
1764         u64 cur_start = *start;
1765         u64 total_bytes = 0;
1766         u64 last = 0;
1767         int found = 0;
1768
1769         if (WARN_ON(search_end <= cur_start))
1770                 return 0;
1771
1772         spin_lock(&tree->lock);
1773         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1774                 total_bytes = tree->dirty_bytes;
1775                 goto out;
1776         }
1777         /*
1778          * this search will find all the extents that end after
1779          * our range starts.
1780          */
1781         node = tree_search(tree, cur_start);
1782         if (!node)
1783                 goto out;
1784
1785         while (1) {
1786                 state = rb_entry(node, struct extent_state, rb_node);
1787                 if (state->start > search_end)
1788                         break;
1789                 if (contig && found && state->start > last + 1)
1790                         break;
1791                 if (state->end >= cur_start && (state->state & bits) == bits) {
1792                         total_bytes += min(search_end, state->end) + 1 -
1793                                        max(cur_start, state->start);
1794                         if (total_bytes >= max_bytes)
1795                                 break;
1796                         if (!found) {
1797                                 *start = max(cur_start, state->start);
1798                                 found = 1;
1799                         }
1800                         last = state->end;
1801                 } else if (contig && found) {
1802                         break;
1803                 }
1804                 node = rb_next(node);
1805                 if (!node)
1806                         break;
1807         }
1808 out:
1809         spin_unlock(&tree->lock);
1810         return total_bytes;
1811 }
1812
1813 /*
1814  * set the private field for a given byte offset in the tree.  If there isn't
1815  * an extent_state there already, this does nothing.
1816  */
1817 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1818                 struct io_failure_record *failrec)
1819 {
1820         struct rb_node *node;
1821         struct extent_state *state;
1822         int ret = 0;
1823
1824         spin_lock(&tree->lock);
1825         /*
1826          * this search will find all the extents that end after
1827          * our range starts.
1828          */
1829         node = tree_search(tree, start);
1830         if (!node) {
1831                 ret = -ENOENT;
1832                 goto out;
1833         }
1834         state = rb_entry(node, struct extent_state, rb_node);
1835         if (state->start != start) {
1836                 ret = -ENOENT;
1837                 goto out;
1838         }
1839         state->failrec = failrec;
1840 out:
1841         spin_unlock(&tree->lock);
1842         return ret;
1843 }
1844
1845 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1846                 struct io_failure_record **failrec)
1847 {
1848         struct rb_node *node;
1849         struct extent_state *state;
1850         int ret = 0;
1851
1852         spin_lock(&tree->lock);
1853         /*
1854          * this search will find all the extents that end after
1855          * our range starts.
1856          */
1857         node = tree_search(tree, start);
1858         if (!node) {
1859                 ret = -ENOENT;
1860                 goto out;
1861         }
1862         state = rb_entry(node, struct extent_state, rb_node);
1863         if (state->start != start) {
1864                 ret = -ENOENT;
1865                 goto out;
1866         }
1867         *failrec = state->failrec;
1868 out:
1869         spin_unlock(&tree->lock);
1870         return ret;
1871 }
1872
1873 /*
1874  * searches a range in the state tree for a given mask.
1875  * If 'filled' == 1, this returns 1 only if every extent in the tree
1876  * has the bits set.  Otherwise, 1 is returned if any bit in the
1877  * range is found set.
1878  */
1879 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1880                    unsigned bits, int filled, struct extent_state *cached)
1881 {
1882         struct extent_state *state = NULL;
1883         struct rb_node *node;
1884         int bitset = 0;
1885
1886         spin_lock(&tree->lock);
1887         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1888             cached->end > start)
1889                 node = &cached->rb_node;
1890         else
1891                 node = tree_search(tree, start);
1892         while (node && start <= end) {
1893                 state = rb_entry(node, struct extent_state, rb_node);
1894
1895                 if (filled && state->start > start) {
1896                         bitset = 0;
1897                         break;
1898                 }
1899
1900                 if (state->start > end)
1901                         break;
1902
1903                 if (state->state & bits) {
1904                         bitset = 1;
1905                         if (!filled)
1906                                 break;
1907                 } else if (filled) {
1908                         bitset = 0;
1909                         break;
1910                 }
1911
1912                 if (state->end == (u64)-1)
1913                         break;
1914
1915                 start = state->end + 1;
1916                 if (start > end)
1917                         break;
1918                 node = rb_next(node);
1919                 if (!node) {
1920                         if (filled)
1921                                 bitset = 0;
1922                         break;
1923                 }
1924         }
1925         spin_unlock(&tree->lock);
1926         return bitset;
1927 }
1928
1929 /*
1930  * helper function to set a given page up to date if all the
1931  * extents in the tree for that page are up to date
1932  */
1933 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1934 {
1935         u64 start = page_offset(page);
1936         u64 end = start + PAGE_SIZE - 1;
1937         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1938                 SetPageUptodate(page);
1939 }
1940
1941 int free_io_failure(struct extent_io_tree *failure_tree,
1942                     struct extent_io_tree *io_tree,
1943                     struct io_failure_record *rec)
1944 {
1945         int ret;
1946         int err = 0;
1947
1948         set_state_failrec(failure_tree, rec->start, NULL);
1949         ret = clear_extent_bits(failure_tree, rec->start,
1950                                 rec->start + rec->len - 1,
1951                                 EXTENT_LOCKED | EXTENT_DIRTY);
1952         if (ret)
1953                 err = ret;
1954
1955         ret = clear_extent_bits(io_tree, rec->start,
1956                                 rec->start + rec->len - 1,
1957                                 EXTENT_DAMAGED);
1958         if (ret && !err)
1959                 err = ret;
1960
1961         kfree(rec);
1962         return err;
1963 }
1964
1965 /*
1966  * this bypasses the standard btrfs submit functions deliberately, as
1967  * the standard behavior is to write all copies in a raid setup. here we only
1968  * want to write the one bad copy. so we do the mapping for ourselves and issue
1969  * submit_bio directly.
1970  * to avoid any synchronization issues, wait for the data after writing, which
1971  * actually prevents the read that triggered the error from finishing.
1972  * currently, there can be no more than two copies of every data bit. thus,
1973  * exactly one rewrite is required.
1974  */
1975 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
1976                       u64 length, u64 logical, struct page *page,
1977                       unsigned int pg_offset, int mirror_num)
1978 {
1979         struct bio *bio;
1980         struct btrfs_device *dev;
1981         u64 map_length = 0;
1982         u64 sector;
1983         struct btrfs_bio *bbio = NULL;
1984         int ret;
1985
1986         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
1987         BUG_ON(!mirror_num);
1988
1989         bio = btrfs_io_bio_alloc(1);
1990         bio->bi_iter.bi_size = 0;
1991         map_length = length;
1992
1993         /*
1994          * Avoid races with device replace and make sure our bbio has devices
1995          * associated to its stripes that don't go away while we are doing the
1996          * read repair operation.
1997          */
1998         btrfs_bio_counter_inc_blocked(fs_info);
1999         if (btrfs_is_parity_mirror(fs_info, logical, length)) {
2000                 /*
2001                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2002                  * to update all raid stripes, but here we just want to correct
2003                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2004                  * stripe's dev and sector.
2005                  */
2006                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2007                                       &map_length, &bbio, 0);
2008                 if (ret) {
2009                         btrfs_bio_counter_dec(fs_info);
2010                         bio_put(bio);
2011                         return -EIO;
2012                 }
2013                 ASSERT(bbio->mirror_num == 1);
2014         } else {
2015                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2016                                       &map_length, &bbio, mirror_num);
2017                 if (ret) {
2018                         btrfs_bio_counter_dec(fs_info);
2019                         bio_put(bio);
2020                         return -EIO;
2021                 }
2022                 BUG_ON(mirror_num != bbio->mirror_num);
2023         }
2024
2025         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2026         bio->bi_iter.bi_sector = sector;
2027         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2028         btrfs_put_bbio(bbio);
2029         if (!dev || !dev->bdev || !dev->writeable) {
2030                 btrfs_bio_counter_dec(fs_info);
2031                 bio_put(bio);
2032                 return -EIO;
2033         }
2034         bio_set_dev(bio, dev->bdev);
2035         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2036         bio_add_page(bio, page, length, pg_offset);
2037
2038         if (btrfsic_submit_bio_wait(bio)) {
2039                 /* try to remap that extent elsewhere? */
2040                 btrfs_bio_counter_dec(fs_info);
2041                 bio_put(bio);
2042                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2043                 return -EIO;
2044         }
2045
2046         btrfs_info_rl_in_rcu(fs_info,
2047                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2048                                   ino, start,
2049                                   rcu_str_deref(dev->name), sector);
2050         btrfs_bio_counter_dec(fs_info);
2051         bio_put(bio);
2052         return 0;
2053 }
2054
2055 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2056                          struct extent_buffer *eb, int mirror_num)
2057 {
2058         u64 start = eb->start;
2059         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2060         int ret = 0;
2061
2062         if (sb_rdonly(fs_info->sb))
2063                 return -EROFS;
2064
2065         for (i = 0; i < num_pages; i++) {
2066                 struct page *p = eb->pages[i];
2067
2068                 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2069                                         start - page_offset(p), mirror_num);
2070                 if (ret)
2071                         break;
2072                 start += PAGE_SIZE;
2073         }
2074
2075         return ret;
2076 }
2077
2078 /*
2079  * each time an IO finishes, we do a fast check in the IO failure tree
2080  * to see if we need to process or clean up an io_failure_record
2081  */
2082 int clean_io_failure(struct btrfs_fs_info *fs_info,
2083                      struct extent_io_tree *failure_tree,
2084                      struct extent_io_tree *io_tree, u64 start,
2085                      struct page *page, u64 ino, unsigned int pg_offset)
2086 {
2087         u64 private;
2088         struct io_failure_record *failrec;
2089         struct extent_state *state;
2090         int num_copies;
2091         int ret;
2092
2093         private = 0;
2094         ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2095                                EXTENT_DIRTY, 0);
2096         if (!ret)
2097                 return 0;
2098
2099         ret = get_state_failrec(failure_tree, start, &failrec);
2100         if (ret)
2101                 return 0;
2102
2103         BUG_ON(!failrec->this_mirror);
2104
2105         if (failrec->in_validation) {
2106                 /* there was no real error, just free the record */
2107                 btrfs_debug(fs_info,
2108                         "clean_io_failure: freeing dummy error at %llu",
2109                         failrec->start);
2110                 goto out;
2111         }
2112         if (sb_rdonly(fs_info->sb))
2113                 goto out;
2114
2115         spin_lock(&io_tree->lock);
2116         state = find_first_extent_bit_state(io_tree,
2117                                             failrec->start,
2118                                             EXTENT_LOCKED);
2119         spin_unlock(&io_tree->lock);
2120
2121         if (state && state->start <= failrec->start &&
2122             state->end >= failrec->start + failrec->len - 1) {
2123                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2124                                               failrec->len);
2125                 if (num_copies > 1)  {
2126                         repair_io_failure(fs_info, ino, start, failrec->len,
2127                                           failrec->logical, page, pg_offset,
2128                                           failrec->failed_mirror);
2129                 }
2130         }
2131
2132 out:
2133         free_io_failure(failure_tree, io_tree, failrec);
2134
2135         return 0;
2136 }
2137
2138 /*
2139  * Can be called when
2140  * - hold extent lock
2141  * - under ordered extent
2142  * - the inode is freeing
2143  */
2144 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2145 {
2146         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2147         struct io_failure_record *failrec;
2148         struct extent_state *state, *next;
2149
2150         if (RB_EMPTY_ROOT(&failure_tree->state))
2151                 return;
2152
2153         spin_lock(&failure_tree->lock);
2154         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2155         while (state) {
2156                 if (state->start > end)
2157                         break;
2158
2159                 ASSERT(state->end <= end);
2160
2161                 next = next_state(state);
2162
2163                 failrec = state->failrec;
2164                 free_extent_state(state);
2165                 kfree(failrec);
2166
2167                 state = next;
2168         }
2169         spin_unlock(&failure_tree->lock);
2170 }
2171
2172 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2173                 struct io_failure_record **failrec_ret)
2174 {
2175         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2176         struct io_failure_record *failrec;
2177         struct extent_map *em;
2178         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2179         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2180         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2181         int ret;
2182         u64 logical;
2183
2184         ret = get_state_failrec(failure_tree, start, &failrec);
2185         if (ret) {
2186                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2187                 if (!failrec)
2188                         return -ENOMEM;
2189
2190                 failrec->start = start;
2191                 failrec->len = end - start + 1;
2192                 failrec->this_mirror = 0;
2193                 failrec->bio_flags = 0;
2194                 failrec->in_validation = 0;
2195
2196                 read_lock(&em_tree->lock);
2197                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2198                 if (!em) {
2199                         read_unlock(&em_tree->lock);
2200                         kfree(failrec);
2201                         return -EIO;
2202                 }
2203
2204                 if (em->start > start || em->start + em->len <= start) {
2205                         free_extent_map(em);
2206                         em = NULL;
2207                 }
2208                 read_unlock(&em_tree->lock);
2209                 if (!em) {
2210                         kfree(failrec);
2211                         return -EIO;
2212                 }
2213
2214                 logical = start - em->start;
2215                 logical = em->block_start + logical;
2216                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2217                         logical = em->block_start;
2218                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2219                         extent_set_compress_type(&failrec->bio_flags,
2220                                                  em->compress_type);
2221                 }
2222
2223                 btrfs_debug(fs_info,
2224                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2225                         logical, start, failrec->len);
2226
2227                 failrec->logical = logical;
2228                 free_extent_map(em);
2229
2230                 /* set the bits in the private failure tree */
2231                 ret = set_extent_bits(failure_tree, start, end,
2232                                         EXTENT_LOCKED | EXTENT_DIRTY);
2233                 if (ret >= 0)
2234                         ret = set_state_failrec(failure_tree, start, failrec);
2235                 /* set the bits in the inode's tree */
2236                 if (ret >= 0)
2237                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2238                 if (ret < 0) {
2239                         kfree(failrec);
2240                         return ret;
2241                 }
2242         } else {
2243                 btrfs_debug(fs_info,
2244                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2245                         failrec->logical, failrec->start, failrec->len,
2246                         failrec->in_validation);
2247                 /*
2248                  * when data can be on disk more than twice, add to failrec here
2249                  * (e.g. with a list for failed_mirror) to make
2250                  * clean_io_failure() clean all those errors at once.
2251                  */
2252         }
2253
2254         *failrec_ret = failrec;
2255
2256         return 0;
2257 }
2258
2259 bool btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2260                            struct io_failure_record *failrec, int failed_mirror)
2261 {
2262         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2263         int num_copies;
2264
2265         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2266         if (num_copies == 1) {
2267                 /*
2268                  * we only have a single copy of the data, so don't bother with
2269                  * all the retry and error correction code that follows. no
2270                  * matter what the error is, it is very likely to persist.
2271                  */
2272                 btrfs_debug(fs_info,
2273                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2274                         num_copies, failrec->this_mirror, failed_mirror);
2275                 return false;
2276         }
2277
2278         /*
2279          * there are two premises:
2280          *      a) deliver good data to the caller
2281          *      b) correct the bad sectors on disk
2282          */
2283         if (failed_bio->bi_vcnt > 1) {
2284                 /*
2285                  * to fulfill b), we need to know the exact failing sectors, as
2286                  * we don't want to rewrite any more than the failed ones. thus,
2287                  * we need separate read requests for the failed bio
2288                  *
2289                  * if the following BUG_ON triggers, our validation request got
2290                  * merged. we need separate requests for our algorithm to work.
2291                  */
2292                 BUG_ON(failrec->in_validation);
2293                 failrec->in_validation = 1;
2294                 failrec->this_mirror = failed_mirror;
2295         } else {
2296                 /*
2297                  * we're ready to fulfill a) and b) alongside. get a good copy
2298                  * of the failed sector and if we succeed, we have setup
2299                  * everything for repair_io_failure to do the rest for us.
2300                  */
2301                 if (failrec->in_validation) {
2302                         BUG_ON(failrec->this_mirror != failed_mirror);
2303                         failrec->in_validation = 0;
2304                         failrec->this_mirror = 0;
2305                 }
2306                 failrec->failed_mirror = failed_mirror;
2307                 failrec->this_mirror++;
2308                 if (failrec->this_mirror == failed_mirror)
2309                         failrec->this_mirror++;
2310         }
2311
2312         if (failrec->this_mirror > num_copies) {
2313                 btrfs_debug(fs_info,
2314                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2315                         num_copies, failrec->this_mirror, failed_mirror);
2316                 return false;
2317         }
2318
2319         return true;
2320 }
2321
2322
2323 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2324                                     struct io_failure_record *failrec,
2325                                     struct page *page, int pg_offset, int icsum,
2326                                     bio_end_io_t *endio_func, void *data)
2327 {
2328         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2329         struct bio *bio;
2330         struct btrfs_io_bio *btrfs_failed_bio;
2331         struct btrfs_io_bio *btrfs_bio;
2332
2333         bio = btrfs_io_bio_alloc(1);
2334         bio->bi_end_io = endio_func;
2335         bio->bi_iter.bi_sector = failrec->logical >> 9;
2336         bio_set_dev(bio, fs_info->fs_devices->latest_bdev);
2337         bio->bi_iter.bi_size = 0;
2338         bio->bi_private = data;
2339
2340         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2341         if (btrfs_failed_bio->csum) {
2342                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2343
2344                 btrfs_bio = btrfs_io_bio(bio);
2345                 btrfs_bio->csum = btrfs_bio->csum_inline;
2346                 icsum *= csum_size;
2347                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2348                        csum_size);
2349         }
2350
2351         bio_add_page(bio, page, failrec->len, pg_offset);
2352
2353         return bio;
2354 }
2355
2356 /*
2357  * this is a generic handler for readpage errors (default
2358  * readpage_io_failed_hook). if other copies exist, read those and write back
2359  * good data to the failed position. does not investigate in remapping the
2360  * failed extent elsewhere, hoping the device will be smart enough to do this as
2361  * needed
2362  */
2363
2364 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2365                               struct page *page, u64 start, u64 end,
2366                               int failed_mirror)
2367 {
2368         struct io_failure_record *failrec;
2369         struct inode *inode = page->mapping->host;
2370         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2371         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2372         struct bio *bio;
2373         int read_mode = 0;
2374         blk_status_t status;
2375         int ret;
2376
2377         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2378
2379         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2380         if (ret)
2381                 return ret;
2382
2383         if (!btrfs_check_repairable(inode, failed_bio, failrec,
2384                                     failed_mirror)) {
2385                 free_io_failure(failure_tree, tree, failrec);
2386                 return -EIO;
2387         }
2388
2389         if (failed_bio->bi_vcnt > 1)
2390                 read_mode |= REQ_FAILFAST_DEV;
2391
2392         phy_offset >>= inode->i_sb->s_blocksize_bits;
2393         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2394                                       start - page_offset(page),
2395                                       (int)phy_offset, failed_bio->bi_end_io,
2396                                       NULL);
2397         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2398
2399         btrfs_debug(btrfs_sb(inode->i_sb),
2400                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2401                 read_mode, failrec->this_mirror, failrec->in_validation);
2402
2403         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2404                                          failrec->bio_flags, 0);
2405         if (status) {
2406                 free_io_failure(failure_tree, tree, failrec);
2407                 bio_put(bio);
2408                 ret = blk_status_to_errno(status);
2409         }
2410
2411         return ret;
2412 }
2413
2414 /* lots and lots of room for performance fixes in the end_bio funcs */
2415
2416 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2417 {
2418         int uptodate = (err == 0);
2419         struct extent_io_tree *tree;
2420         int ret = 0;
2421
2422         tree = &BTRFS_I(page->mapping->host)->io_tree;
2423
2424         if (tree->ops && tree->ops->writepage_end_io_hook)
2425                 tree->ops->writepage_end_io_hook(page, start, end, NULL,
2426                                 uptodate);
2427
2428         if (!uptodate) {
2429                 ClearPageUptodate(page);
2430                 SetPageError(page);
2431                 ret = err < 0 ? err : -EIO;
2432                 mapping_set_error(page->mapping, ret);
2433         }
2434 }
2435
2436 /*
2437  * after a writepage IO is done, we need to:
2438  * clear the uptodate bits on error
2439  * clear the writeback bits in the extent tree for this IO
2440  * end_page_writeback if the page has no more pending IO
2441  *
2442  * Scheduling is not allowed, so the extent state tree is expected
2443  * to have one and only one object corresponding to this IO.
2444  */
2445 static void end_bio_extent_writepage(struct bio *bio)
2446 {
2447         int error = blk_status_to_errno(bio->bi_status);
2448         struct bio_vec *bvec;
2449         u64 start;
2450         u64 end;
2451         int i;
2452
2453         ASSERT(!bio_flagged(bio, BIO_CLONED));
2454         bio_for_each_segment_all(bvec, bio, i) {
2455                 struct page *page = bvec->bv_page;
2456                 struct inode *inode = page->mapping->host;
2457                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2458
2459                 /* We always issue full-page reads, but if some block
2460                  * in a page fails to read, blk_update_request() will
2461                  * advance bv_offset and adjust bv_len to compensate.
2462                  * Print a warning for nonzero offsets, and an error
2463                  * if they don't add up to a full page.  */
2464                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2465                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2466                                 btrfs_err(fs_info,
2467                                    "partial page write in btrfs with offset %u and length %u",
2468                                         bvec->bv_offset, bvec->bv_len);
2469                         else
2470                                 btrfs_info(fs_info,
2471                                    "incomplete page write in btrfs with offset %u and length %u",
2472                                         bvec->bv_offset, bvec->bv_len);
2473                 }
2474
2475                 start = page_offset(page);
2476                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2477
2478                 end_extent_writepage(page, error, start, end);
2479                 end_page_writeback(page);
2480         }
2481
2482         bio_put(bio);
2483 }
2484
2485 static void
2486 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2487                               int uptodate)
2488 {
2489         struct extent_state *cached = NULL;
2490         u64 end = start + len - 1;
2491
2492         if (uptodate && tree->track_uptodate)
2493                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2494         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2495 }
2496
2497 /*
2498  * after a readpage IO is done, we need to:
2499  * clear the uptodate bits on error
2500  * set the uptodate bits if things worked
2501  * set the page up to date if all extents in the tree are uptodate
2502  * clear the lock bit in the extent tree
2503  * unlock the page if there are no other extents locked for it
2504  *
2505  * Scheduling is not allowed, so the extent state tree is expected
2506  * to have one and only one object corresponding to this IO.
2507  */
2508 static void end_bio_extent_readpage(struct bio *bio)
2509 {
2510         struct bio_vec *bvec;
2511         int uptodate = !bio->bi_status;
2512         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2513         struct extent_io_tree *tree, *failure_tree;
2514         u64 offset = 0;
2515         u64 start;
2516         u64 end;
2517         u64 len;
2518         u64 extent_start = 0;
2519         u64 extent_len = 0;
2520         int mirror;
2521         int ret;
2522         int i;
2523
2524         ASSERT(!bio_flagged(bio, BIO_CLONED));
2525         bio_for_each_segment_all(bvec, bio, i) {
2526                 struct page *page = bvec->bv_page;
2527                 struct inode *inode = page->mapping->host;
2528                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2529
2530                 btrfs_debug(fs_info,
2531                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2532                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2533                         io_bio->mirror_num);
2534                 tree = &BTRFS_I(inode)->io_tree;
2535                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2536
2537                 /* We always issue full-page reads, but if some block
2538                  * in a page fails to read, blk_update_request() will
2539                  * advance bv_offset and adjust bv_len to compensate.
2540                  * Print a warning for nonzero offsets, and an error
2541                  * if they don't add up to a full page.  */
2542                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2543                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2544                                 btrfs_err(fs_info,
2545                                         "partial page read in btrfs with offset %u and length %u",
2546                                         bvec->bv_offset, bvec->bv_len);
2547                         else
2548                                 btrfs_info(fs_info,
2549                                         "incomplete page read in btrfs with offset %u and length %u",
2550                                         bvec->bv_offset, bvec->bv_len);
2551                 }
2552
2553                 start = page_offset(page);
2554                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2555                 len = bvec->bv_len;
2556
2557                 mirror = io_bio->mirror_num;
2558                 if (likely(uptodate && tree->ops)) {
2559                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2560                                                               page, start, end,
2561                                                               mirror);
2562                         if (ret)
2563                                 uptodate = 0;
2564                         else
2565                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2566                                                  failure_tree, tree, start,
2567                                                  page,
2568                                                  btrfs_ino(BTRFS_I(inode)), 0);
2569                 }
2570
2571                 if (likely(uptodate))
2572                         goto readpage_ok;
2573
2574                 if (tree->ops) {
2575                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2576                         if (ret == -EAGAIN) {
2577                                 /*
2578                                  * Data inode's readpage_io_failed_hook() always
2579                                  * returns -EAGAIN.
2580                                  *
2581                                  * The generic bio_readpage_error handles errors
2582                                  * the following way: If possible, new read
2583                                  * requests are created and submitted and will
2584                                  * end up in end_bio_extent_readpage as well (if
2585                                  * we're lucky, not in the !uptodate case). In
2586                                  * that case it returns 0 and we just go on with
2587                                  * the next page in our bio. If it can't handle
2588                                  * the error it will return -EIO and we remain
2589                                  * responsible for that page.
2590                                  */
2591                                 ret = bio_readpage_error(bio, offset, page,
2592                                                          start, end, mirror);
2593                                 if (ret == 0) {
2594                                         uptodate = !bio->bi_status;
2595                                         offset += len;
2596                                         continue;
2597                                 }
2598                         }
2599
2600                         /*
2601                          * metadata's readpage_io_failed_hook() always returns
2602                          * -EIO and fixes nothing.  -EIO is also returned if
2603                          * data inode error could not be fixed.
2604                          */
2605                         ASSERT(ret == -EIO);
2606                 }
2607 readpage_ok:
2608                 if (likely(uptodate)) {
2609                         loff_t i_size = i_size_read(inode);
2610                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2611                         unsigned off;
2612
2613                         /* Zero out the end if this page straddles i_size */
2614                         off = i_size & (PAGE_SIZE-1);
2615                         if (page->index == end_index && off)
2616                                 zero_user_segment(page, off, PAGE_SIZE);
2617                         SetPageUptodate(page);
2618                 } else {
2619                         ClearPageUptodate(page);
2620                         SetPageError(page);
2621                 }
2622                 unlock_page(page);
2623                 offset += len;
2624
2625                 if (unlikely(!uptodate)) {
2626                         if (extent_len) {
2627                                 endio_readpage_release_extent(tree,
2628                                                               extent_start,
2629                                                               extent_len, 1);
2630                                 extent_start = 0;
2631                                 extent_len = 0;
2632                         }
2633                         endio_readpage_release_extent(tree, start,
2634                                                       end - start + 1, 0);
2635                 } else if (!extent_len) {
2636                         extent_start = start;
2637                         extent_len = end + 1 - start;
2638                 } else if (extent_start + extent_len == start) {
2639                         extent_len += end + 1 - start;
2640                 } else {
2641                         endio_readpage_release_extent(tree, extent_start,
2642                                                       extent_len, uptodate);
2643                         extent_start = start;
2644                         extent_len = end + 1 - start;
2645                 }
2646         }
2647
2648         if (extent_len)
2649                 endio_readpage_release_extent(tree, extent_start, extent_len,
2650                                               uptodate);
2651         if (io_bio->end_io)
2652                 io_bio->end_io(io_bio, blk_status_to_errno(bio->bi_status));
2653         bio_put(bio);
2654 }
2655
2656 /*
2657  * Initialize the members up to but not including 'bio'. Use after allocating a
2658  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2659  * 'bio' because use of __GFP_ZERO is not supported.
2660  */
2661 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2662 {
2663         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2664 }
2665
2666 /*
2667  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2668  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2669  * for the appropriate container_of magic
2670  */
2671 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte)
2672 {
2673         struct bio *bio;
2674
2675         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, btrfs_bioset);
2676         bio_set_dev(bio, bdev);
2677         bio->bi_iter.bi_sector = first_byte >> 9;
2678         btrfs_io_bio_init(btrfs_io_bio(bio));
2679         return bio;
2680 }
2681
2682 struct bio *btrfs_bio_clone(struct bio *bio)
2683 {
2684         struct btrfs_io_bio *btrfs_bio;
2685         struct bio *new;
2686
2687         /* Bio allocation backed by a bioset does not fail */
2688         new = bio_clone_fast(bio, GFP_NOFS, btrfs_bioset);
2689         btrfs_bio = btrfs_io_bio(new);
2690         btrfs_io_bio_init(btrfs_bio);
2691         btrfs_bio->iter = bio->bi_iter;
2692         return new;
2693 }
2694
2695 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2696 {
2697         struct bio *bio;
2698
2699         /* Bio allocation backed by a bioset does not fail */
2700         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, btrfs_bioset);
2701         btrfs_io_bio_init(btrfs_io_bio(bio));
2702         return bio;
2703 }
2704
2705 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2706 {
2707         struct bio *bio;
2708         struct btrfs_io_bio *btrfs_bio;
2709
2710         /* this will never fail when it's backed by a bioset */
2711         bio = bio_clone_fast(orig, GFP_NOFS, btrfs_bioset);
2712         ASSERT(bio);
2713
2714         btrfs_bio = btrfs_io_bio(bio);
2715         btrfs_io_bio_init(btrfs_bio);
2716
2717         bio_trim(bio, offset >> 9, size >> 9);
2718         btrfs_bio->iter = bio->bi_iter;
2719         return bio;
2720 }
2721
2722 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2723                                        unsigned long bio_flags)
2724 {
2725         blk_status_t ret = 0;
2726         struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2727         struct page *page = bvec->bv_page;
2728         struct extent_io_tree *tree = bio->bi_private;
2729         u64 start;
2730
2731         start = page_offset(page) + bvec->bv_offset;
2732
2733         bio->bi_private = NULL;
2734         bio_get(bio);
2735
2736         if (tree->ops)
2737                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
2738                                            mirror_num, bio_flags, start);
2739         else
2740                 btrfsic_submit_bio(bio);
2741
2742         bio_put(bio);
2743         return blk_status_to_errno(ret);
2744 }
2745
2746 static int merge_bio(struct extent_io_tree *tree, struct page *page,
2747                      unsigned long offset, size_t size, struct bio *bio,
2748                      unsigned long bio_flags)
2749 {
2750         int ret = 0;
2751         if (tree->ops)
2752                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2753                                                 bio_flags);
2754         return ret;
2755
2756 }
2757
2758 /*
2759  * @opf:        bio REQ_OP_* and REQ_* flags as one value
2760  */
2761 static int submit_extent_page(unsigned int opf, struct extent_io_tree *tree,
2762                               struct writeback_control *wbc,
2763                               struct page *page, u64 offset,
2764                               size_t size, unsigned long pg_offset,
2765                               struct block_device *bdev,
2766                               struct bio **bio_ret,
2767                               bio_end_io_t end_io_func,
2768                               int mirror_num,
2769                               unsigned long prev_bio_flags,
2770                               unsigned long bio_flags,
2771                               bool force_bio_submit)
2772 {
2773         int ret = 0;
2774         struct bio *bio;
2775         int contig = 0;
2776         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2777         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2778         sector_t sector = offset >> 9;
2779
2780         if (bio_ret && *bio_ret) {
2781                 bio = *bio_ret;
2782                 if (old_compressed)
2783                         contig = bio->bi_iter.bi_sector == sector;
2784                 else
2785                         contig = bio_end_sector(bio) == sector;
2786
2787                 if (prev_bio_flags != bio_flags || !contig ||
2788                     force_bio_submit ||
2789                     merge_bio(tree, page, pg_offset, page_size, bio, bio_flags) ||
2790                     bio_add_page(bio, page, page_size, pg_offset) < page_size) {
2791                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2792                         if (ret < 0) {
2793                                 *bio_ret = NULL;
2794                                 return ret;
2795                         }
2796                         bio = NULL;
2797                 } else {
2798                         if (wbc)
2799                                 wbc_account_io(wbc, page, page_size);
2800                         return 0;
2801                 }
2802         }
2803
2804         bio = btrfs_bio_alloc(bdev, offset);
2805         bio_add_page(bio, page, page_size, pg_offset);
2806         bio->bi_end_io = end_io_func;
2807         bio->bi_private = tree;
2808         bio->bi_write_hint = page->mapping->host->i_write_hint;
2809         bio->bi_opf = opf;
2810         if (wbc) {
2811                 wbc_init_bio(wbc, bio);
2812                 wbc_account_io(wbc, page, page_size);
2813         }
2814
2815         if (bio_ret)
2816                 *bio_ret = bio;
2817         else
2818                 ret = submit_one_bio(bio, mirror_num, bio_flags);
2819
2820         return ret;
2821 }
2822
2823 static void attach_extent_buffer_page(struct extent_buffer *eb,
2824                                       struct page *page)
2825 {
2826         if (!PagePrivate(page)) {
2827                 SetPagePrivate(page);
2828                 get_page(page);
2829                 set_page_private(page, (unsigned long)eb);
2830         } else {
2831                 WARN_ON(page->private != (unsigned long)eb);
2832         }
2833 }
2834
2835 void set_page_extent_mapped(struct page *page)
2836 {
2837         if (!PagePrivate(page)) {
2838                 SetPagePrivate(page);
2839                 get_page(page);
2840                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2841         }
2842 }
2843
2844 static struct extent_map *
2845 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2846                  u64 start, u64 len, get_extent_t *get_extent,
2847                  struct extent_map **em_cached)
2848 {
2849         struct extent_map *em;
2850
2851         if (em_cached && *em_cached) {
2852                 em = *em_cached;
2853                 if (extent_map_in_tree(em) && start >= em->start &&
2854                     start < extent_map_end(em)) {
2855                         refcount_inc(&em->refs);
2856                         return em;
2857                 }
2858
2859                 free_extent_map(em);
2860                 *em_cached = NULL;
2861         }
2862
2863         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2864         if (em_cached && !IS_ERR_OR_NULL(em)) {
2865                 BUG_ON(*em_cached);
2866                 refcount_inc(&em->refs);
2867                 *em_cached = em;
2868         }
2869         return em;
2870 }
2871 /*
2872  * basic readpage implementation.  Locked extent state structs are inserted
2873  * into the tree that are removed when the IO is done (by the end_io
2874  * handlers)
2875  * XXX JDM: This needs looking at to ensure proper page locking
2876  * return 0 on success, otherwise return error
2877  */
2878 static int __do_readpage(struct extent_io_tree *tree,
2879                          struct page *page,
2880                          get_extent_t *get_extent,
2881                          struct extent_map **em_cached,
2882                          struct bio **bio, int mirror_num,
2883                          unsigned long *bio_flags, unsigned int read_flags,
2884                          u64 *prev_em_start)
2885 {
2886         struct inode *inode = page->mapping->host;
2887         u64 start = page_offset(page);
2888         u64 page_end = start + PAGE_SIZE - 1;
2889         u64 end;
2890         u64 cur = start;
2891         u64 extent_offset;
2892         u64 last_byte = i_size_read(inode);
2893         u64 block_start;
2894         u64 cur_end;
2895         struct extent_map *em;
2896         struct block_device *bdev;
2897         int ret = 0;
2898         int nr = 0;
2899         size_t pg_offset = 0;
2900         size_t iosize;
2901         size_t disk_io_size;
2902         size_t blocksize = inode->i_sb->s_blocksize;
2903         unsigned long this_bio_flag = 0;
2904
2905         set_page_extent_mapped(page);
2906
2907         end = page_end;
2908         if (!PageUptodate(page)) {
2909                 if (cleancache_get_page(page) == 0) {
2910                         BUG_ON(blocksize != PAGE_SIZE);
2911                         unlock_extent(tree, start, end);
2912                         goto out;
2913                 }
2914         }
2915
2916         if (page->index == last_byte >> PAGE_SHIFT) {
2917                 char *userpage;
2918                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2919
2920                 if (zero_offset) {
2921                         iosize = PAGE_SIZE - zero_offset;
2922                         userpage = kmap_atomic(page);
2923                         memset(userpage + zero_offset, 0, iosize);
2924                         flush_dcache_page(page);
2925                         kunmap_atomic(userpage);
2926                 }
2927         }
2928         while (cur <= end) {
2929                 bool force_bio_submit = false;
2930                 u64 offset;
2931
2932                 if (cur >= last_byte) {
2933                         char *userpage;
2934                         struct extent_state *cached = NULL;
2935
2936                         iosize = PAGE_SIZE - pg_offset;
2937                         userpage = kmap_atomic(page);
2938                         memset(userpage + pg_offset, 0, iosize);
2939                         flush_dcache_page(page);
2940                         kunmap_atomic(userpage);
2941                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2942                                             &cached, GFP_NOFS);
2943                         unlock_extent_cached(tree, cur,
2944                                              cur + iosize - 1,
2945                                              &cached, GFP_NOFS);
2946                         break;
2947                 }
2948                 em = __get_extent_map(inode, page, pg_offset, cur,
2949                                       end - cur + 1, get_extent, em_cached);
2950                 if (IS_ERR_OR_NULL(em)) {
2951                         SetPageError(page);
2952                         unlock_extent(tree, cur, end);
2953                         break;
2954                 }
2955                 extent_offset = cur - em->start;
2956                 BUG_ON(extent_map_end(em) <= cur);
2957                 BUG_ON(end < cur);
2958
2959                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2960                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2961                         extent_set_compress_type(&this_bio_flag,
2962                                                  em->compress_type);
2963                 }
2964
2965                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2966                 cur_end = min(extent_map_end(em) - 1, end);
2967                 iosize = ALIGN(iosize, blocksize);
2968                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2969                         disk_io_size = em->block_len;
2970                         offset = em->block_start;
2971                 } else {
2972                         offset = em->block_start + extent_offset;
2973                         disk_io_size = iosize;
2974                 }
2975                 bdev = em->bdev;
2976                 block_start = em->block_start;
2977                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2978                         block_start = EXTENT_MAP_HOLE;
2979
2980                 /*
2981                  * If we have a file range that points to a compressed extent
2982                  * and it's followed by a consecutive file range that points to
2983                  * to the same compressed extent (possibly with a different
2984                  * offset and/or length, so it either points to the whole extent
2985                  * or only part of it), we must make sure we do not submit a
2986                  * single bio to populate the pages for the 2 ranges because
2987                  * this makes the compressed extent read zero out the pages
2988                  * belonging to the 2nd range. Imagine the following scenario:
2989                  *
2990                  *  File layout
2991                  *  [0 - 8K]                     [8K - 24K]
2992                  *    |                               |
2993                  *    |                               |
2994                  * points to extent X,         points to extent X,
2995                  * offset 4K, length of 8K     offset 0, length 16K
2996                  *
2997                  * [extent X, compressed length = 4K uncompressed length = 16K]
2998                  *
2999                  * If the bio to read the compressed extent covers both ranges,
3000                  * it will decompress extent X into the pages belonging to the
3001                  * first range and then it will stop, zeroing out the remaining
3002                  * pages that belong to the other range that points to extent X.
3003                  * So here we make sure we submit 2 bios, one for the first
3004                  * range and another one for the third range. Both will target
3005                  * the same physical extent from disk, but we can't currently
3006                  * make the compressed bio endio callback populate the pages
3007                  * for both ranges because each compressed bio is tightly
3008                  * coupled with a single extent map, and each range can have
3009                  * an extent map with a different offset value relative to the
3010                  * uncompressed data of our extent and different lengths. This
3011                  * is a corner case so we prioritize correctness over
3012                  * non-optimal behavior (submitting 2 bios for the same extent).
3013                  */
3014                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3015                     prev_em_start && *prev_em_start != (u64)-1 &&
3016                     *prev_em_start != em->orig_start)
3017                         force_bio_submit = true;
3018
3019                 if (prev_em_start)
3020                         *prev_em_start = em->orig_start;
3021
3022                 free_extent_map(em);
3023                 em = NULL;
3024
3025                 /* we've found a hole, just zero and go on */
3026                 if (block_start == EXTENT_MAP_HOLE) {
3027                         char *userpage;
3028                         struct extent_state *cached = NULL;
3029
3030                         userpage = kmap_atomic(page);
3031                         memset(userpage + pg_offset, 0, iosize);
3032                         flush_dcache_page(page);
3033                         kunmap_atomic(userpage);
3034
3035                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3036                                             &cached, GFP_NOFS);
3037                         unlock_extent_cached(tree, cur,
3038                                              cur + iosize - 1,
3039                                              &cached, GFP_NOFS);
3040                         cur = cur + iosize;
3041                         pg_offset += iosize;
3042                         continue;
3043                 }
3044                 /* the get_extent function already copied into the page */
3045                 if (test_range_bit(tree, cur, cur_end,
3046                                    EXTENT_UPTODATE, 1, NULL)) {
3047                         check_page_uptodate(tree, page);
3048                         unlock_extent(tree, cur, cur + iosize - 1);
3049                         cur = cur + iosize;
3050                         pg_offset += iosize;
3051                         continue;
3052                 }
3053                 /* we have an inline extent but it didn't get marked up
3054                  * to date.  Error out
3055                  */
3056                 if (block_start == EXTENT_MAP_INLINE) {
3057                         SetPageError(page);
3058                         unlock_extent(tree, cur, cur + iosize - 1);
3059                         cur = cur + iosize;
3060                         pg_offset += iosize;
3061                         continue;
3062                 }
3063
3064                 ret = submit_extent_page(REQ_OP_READ | read_flags, tree, NULL,
3065                                          page, offset, disk_io_size,
3066                                          pg_offset, bdev, bio,
3067                                          end_bio_extent_readpage, mirror_num,
3068                                          *bio_flags,
3069                                          this_bio_flag,
3070                                          force_bio_submit);
3071                 if (!ret) {
3072                         nr++;
3073                         *bio_flags = this_bio_flag;
3074                 } else {
3075                         SetPageError(page);
3076                         unlock_extent(tree, cur, cur + iosize - 1);
3077                         goto out;
3078                 }
3079                 cur = cur + iosize;
3080                 pg_offset += iosize;
3081         }
3082 out:
3083         if (!nr) {
3084                 if (!PageError(page))
3085                         SetPageUptodate(page);
3086                 unlock_page(page);
3087         }
3088         return ret;
3089 }
3090
3091 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3092                                              struct page *pages[], int nr_pages,
3093                                              u64 start, u64 end,
3094                                              get_extent_t *get_extent,
3095                                              struct extent_map **em_cached,
3096                                              struct bio **bio, int mirror_num,
3097                                              unsigned long *bio_flags,
3098                                              u64 *prev_em_start)
3099 {
3100         struct inode *inode;
3101         struct btrfs_ordered_extent *ordered;
3102         int index;
3103
3104         inode = pages[0]->mapping->host;
3105         while (1) {
3106                 lock_extent(tree, start, end);
3107                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3108                                                      end - start + 1);
3109                 if (!ordered)
3110                         break;
3111                 unlock_extent(tree, start, end);
3112                 btrfs_start_ordered_extent(inode, ordered, 1);
3113                 btrfs_put_ordered_extent(ordered);
3114         }
3115
3116         for (index = 0; index < nr_pages; index++) {
3117                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3118                               mirror_num, bio_flags, 0, prev_em_start);
3119                 put_page(pages[index]);
3120         }
3121 }
3122