btrfs: fix check_shared for fiemap ioctl
[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 #include "transaction.h"
24
25 static struct kmem_cache *extent_state_cache;
26 static struct kmem_cache *extent_buffer_cache;
27 static struct bio_set *btrfs_bioset;
28
29 static inline bool extent_state_in_tree(const struct extent_state *state)
30 {
31         return !RB_EMPTY_NODE(&state->rb_node);
32 }
33
34 #ifdef CONFIG_BTRFS_DEBUG
35 static LIST_HEAD(buffers);
36 static LIST_HEAD(states);
37
38 static DEFINE_SPINLOCK(leak_lock);
39
40 static inline
41 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
42 {
43         unsigned long flags;
44
45         spin_lock_irqsave(&leak_lock, flags);
46         list_add(new, head);
47         spin_unlock_irqrestore(&leak_lock, flags);
48 }
49
50 static inline
51 void btrfs_leak_debug_del(struct list_head *entry)
52 {
53         unsigned long flags;
54
55         spin_lock_irqsave(&leak_lock, flags);
56         list_del(entry);
57         spin_unlock_irqrestore(&leak_lock, flags);
58 }
59
60 static inline
61 void btrfs_leak_debug_check(void)
62 {
63         struct extent_state *state;
64         struct extent_buffer *eb;
65
66         while (!list_empty(&states)) {
67                 state = list_entry(states.next, struct extent_state, leak_list);
68                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
69                        state->start, state->end, state->state,
70                        extent_state_in_tree(state),
71                        atomic_read(&state->refs));
72                 list_del(&state->leak_list);
73                 kmem_cache_free(extent_state_cache, state);
74         }
75
76         while (!list_empty(&buffers)) {
77                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
78                 printk(KERN_ERR "BTRFS: buffer leak start %llu len %lu "
79                        "refs %d\n",
80                        eb->start, eb->len, atomic_read(&eb->refs));
81                 list_del(&eb->leak_list);
82                 kmem_cache_free(extent_buffer_cache, eb);
83         }
84 }
85
86 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
87         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
88 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
89                 struct extent_io_tree *tree, u64 start, u64 end)
90 {
91         struct inode *inode;
92         u64 isize;
93
94         if (!tree->mapping)
95                 return;
96
97         inode = tree->mapping->host;
98         isize = i_size_read(inode);
99         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
100                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
101                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
102                                 caller, btrfs_ino(inode), isize, start, end);
103         }
104 }
105 #else
106 #define btrfs_leak_debug_add(new, head) do {} while (0)
107 #define btrfs_leak_debug_del(entry)     do {} while (0)
108 #define btrfs_leak_debug_check()        do {} while (0)
109 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
110 #endif
111
112 #define BUFFER_LRU_MAX 64
113
114 struct tree_entry {
115         u64 start;
116         u64 end;
117         struct rb_node rb_node;
118 };
119
120 struct extent_page_data {
121         struct bio *bio;
122         struct extent_io_tree *tree;
123         get_extent_t *get_extent;
124         unsigned long bio_flags;
125
126         /* tells writepage not to lock the state bits for this range
127          * it still does the unlocking
128          */
129         unsigned int extent_locked:1;
130
131         /* tells the submit_bio code to use a WRITE_SYNC */
132         unsigned int sync_io:1;
133 };
134
135 static void add_extent_changeset(struct extent_state *state, unsigned bits,
136                                  struct extent_changeset *changeset,
137                                  int set)
138 {
139         int ret;
140
141         if (!changeset)
142                 return;
143         if (set && (state->state & bits) == bits)
144                 return;
145         if (!set && (state->state & bits) == 0)
146                 return;
147         changeset->bytes_changed += state->end - state->start + 1;
148         ret = ulist_add(changeset->range_changed, state->start, state->end,
149                         GFP_ATOMIC);
150         /* ENOMEM */
151         BUG_ON(ret < 0);
152 }
153
154 static noinline void flush_write_bio(void *data);
155 static inline struct btrfs_fs_info *
156 tree_fs_info(struct extent_io_tree *tree)
157 {
158         if (!tree->mapping)
159                 return NULL;
160         return btrfs_sb(tree->mapping->host->i_sb);
161 }
162
163 int __init extent_io_init(void)
164 {
165         extent_state_cache = kmem_cache_create("btrfs_extent_state",
166                         sizeof(struct extent_state), 0,
167                         SLAB_MEM_SPREAD, NULL);
168         if (!extent_state_cache)
169                 return -ENOMEM;
170
171         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
172                         sizeof(struct extent_buffer), 0,
173                         SLAB_MEM_SPREAD, NULL);
174         if (!extent_buffer_cache)
175                 goto free_state_cache;
176
177         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
178                                      offsetof(struct btrfs_io_bio, bio));
179         if (!btrfs_bioset)
180                 goto free_buffer_cache;
181
182         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
183                 goto free_bioset;
184
185         return 0;
186
187 free_bioset:
188         bioset_free(btrfs_bioset);
189         btrfs_bioset = NULL;
190
191 free_buffer_cache:
192         kmem_cache_destroy(extent_buffer_cache);
193         extent_buffer_cache = NULL;
194
195 free_state_cache:
196         kmem_cache_destroy(extent_state_cache);
197         extent_state_cache = NULL;
198         return -ENOMEM;
199 }
200
201 void extent_io_exit(void)
202 {
203         btrfs_leak_debug_check();
204
205         /*
206          * Make sure all delayed rcu free are flushed before we
207          * destroy caches.
208          */
209         rcu_barrier();
210         kmem_cache_destroy(extent_state_cache);
211         kmem_cache_destroy(extent_buffer_cache);
212         if (btrfs_bioset)
213                 bioset_free(btrfs_bioset);
214 }
215
216 void extent_io_tree_init(struct extent_io_tree *tree,
217                          struct address_space *mapping)
218 {
219         tree->state = RB_ROOT;
220         tree->ops = NULL;
221         tree->dirty_bytes = 0;
222         spin_lock_init(&tree->lock);
223         tree->mapping = mapping;
224 }
225
226 static struct extent_state *alloc_extent_state(gfp_t mask)
227 {
228         struct extent_state *state;
229
230         state = kmem_cache_alloc(extent_state_cache, mask);
231         if (!state)
232                 return state;
233         state->state = 0;
234         state->failrec = NULL;
235         RB_CLEAR_NODE(&state->rb_node);
236         btrfs_leak_debug_add(&state->leak_list, &states);
237         atomic_set(&state->refs, 1);
238         init_waitqueue_head(&state->wq);
239         trace_alloc_extent_state(state, mask, _RET_IP_);
240         return state;
241 }
242
243 void free_extent_state(struct extent_state *state)
244 {
245         if (!state)
246                 return;
247         if (atomic_dec_and_test(&state->refs)) {
248                 WARN_ON(extent_state_in_tree(state));
249                 btrfs_leak_debug_del(&state->leak_list);
250                 trace_free_extent_state(state, _RET_IP_);
251                 kmem_cache_free(extent_state_cache, state);
252         }
253 }
254
255 static struct rb_node *tree_insert(struct rb_root *root,
256                                    struct rb_node *search_start,
257                                    u64 offset,
258                                    struct rb_node *node,
259                                    struct rb_node ***p_in,
260                                    struct rb_node **parent_in)
261 {
262         struct rb_node **p;
263         struct rb_node *parent = NULL;
264         struct tree_entry *entry;
265
266         if (p_in && parent_in) {
267                 p = *p_in;
268                 parent = *parent_in;
269                 goto do_insert;
270         }
271
272         p = search_start ? &search_start : &root->rb_node;
273         while (*p) {
274                 parent = *p;
275                 entry = rb_entry(parent, struct tree_entry, rb_node);
276
277                 if (offset < entry->start)
278                         p = &(*p)->rb_left;
279                 else if (offset > entry->end)
280                         p = &(*p)->rb_right;
281                 else
282                         return parent;
283         }
284
285 do_insert:
286         rb_link_node(node, parent, p);
287         rb_insert_color(node, root);
288         return NULL;
289 }
290
291 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
292                                       struct rb_node **prev_ret,
293                                       struct rb_node **next_ret,
294                                       struct rb_node ***p_ret,
295                                       struct rb_node **parent_ret)
296 {
297         struct rb_root *root = &tree->state;
298         struct rb_node **n = &root->rb_node;
299         struct rb_node *prev = NULL;
300         struct rb_node *orig_prev = NULL;
301         struct tree_entry *entry;
302         struct tree_entry *prev_entry = NULL;
303
304         while (*n) {
305                 prev = *n;
306                 entry = rb_entry(prev, struct tree_entry, rb_node);
307                 prev_entry = entry;
308
309                 if (offset < entry->start)
310                         n = &(*n)->rb_left;
311                 else if (offset > entry->end)
312                         n = &(*n)->rb_right;
313                 else
314                         return *n;
315         }
316
317         if (p_ret)
318                 *p_ret = n;
319         if (parent_ret)
320                 *parent_ret = prev;
321
322         if (prev_ret) {
323                 orig_prev = prev;
324                 while (prev && offset > prev_entry->end) {
325                         prev = rb_next(prev);
326                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
327                 }
328                 *prev_ret = prev;
329                 prev = orig_prev;
330         }
331
332         if (next_ret) {
333                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
334                 while (prev && offset < prev_entry->start) {
335                         prev = rb_prev(prev);
336                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
337                 }
338                 *next_ret = prev;
339         }
340         return NULL;
341 }
342
343 static inline struct rb_node *
344 tree_search_for_insert(struct extent_io_tree *tree,
345                        u64 offset,
346                        struct rb_node ***p_ret,
347                        struct rb_node **parent_ret)
348 {
349         struct rb_node *prev = NULL;
350         struct rb_node *ret;
351
352         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
353         if (!ret)
354                 return prev;
355         return ret;
356 }
357
358 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
359                                           u64 offset)
360 {
361         return tree_search_for_insert(tree, offset, NULL, NULL);
362 }
363
364 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
365                      struct extent_state *other)
366 {
367         if (tree->ops && tree->ops->merge_extent_hook)
368                 tree->ops->merge_extent_hook(tree->mapping->host, new,
369                                              other);
370 }
371
372 /*
373  * utility function to look for merge candidates inside a given range.
374  * Any extents with matching state are merged together into a single
375  * extent in the tree.  Extents with EXTENT_IO in their state field
376  * are not merged because the end_io handlers need to be able to do
377  * operations on them without sleeping (or doing allocations/splits).
378  *
379  * This should be called with the tree lock held.
380  */
381 static void merge_state(struct extent_io_tree *tree,
382                         struct extent_state *state)
383 {
384         struct extent_state *other;
385         struct rb_node *other_node;
386
387         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
388                 return;
389
390         other_node = rb_prev(&state->rb_node);
391         if (other_node) {
392                 other = rb_entry(other_node, struct extent_state, rb_node);
393                 if (other->end == state->start - 1 &&
394                     other->state == state->state) {
395                         merge_cb(tree, state, other);
396                         state->start = other->start;
397                         rb_erase(&other->rb_node, &tree->state);
398                         RB_CLEAR_NODE(&other->rb_node);
399                         free_extent_state(other);
400                 }
401         }
402         other_node = rb_next(&state->rb_node);
403         if (other_node) {
404                 other = rb_entry(other_node, struct extent_state, rb_node);
405                 if (other->start == state->end + 1 &&
406                     other->state == state->state) {
407                         merge_cb(tree, state, other);
408                         state->end = other->end;
409                         rb_erase(&other->rb_node, &tree->state);
410                         RB_CLEAR_NODE(&other->rb_node);
411                         free_extent_state(other);
412                 }
413         }
414 }
415
416 static void set_state_cb(struct extent_io_tree *tree,
417                          struct extent_state *state, unsigned *bits)
418 {
419         if (tree->ops && tree->ops->set_bit_hook)
420                 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
421 }
422
423 static void clear_state_cb(struct extent_io_tree *tree,
424                            struct extent_state *state, unsigned *bits)
425 {
426         if (tree->ops && tree->ops->clear_bit_hook)
427                 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
428 }
429
430 static void set_state_bits(struct extent_io_tree *tree,
431                            struct extent_state *state, unsigned *bits,
432                            struct extent_changeset *changeset);
433
434 /*
435  * insert an extent_state struct into the tree.  'bits' are set on the
436  * struct before it is inserted.
437  *
438  * This may return -EEXIST if the extent is already there, in which case the
439  * state struct is freed.
440  *
441  * The tree lock is not taken internally.  This is a utility function and
442  * probably isn't what you want to call (see set/clear_extent_bit).
443  */
444 static int insert_state(struct extent_io_tree *tree,
445                         struct extent_state *state, u64 start, u64 end,
446                         struct rb_node ***p,
447                         struct rb_node **parent,
448                         unsigned *bits, struct extent_changeset *changeset)
449 {
450         struct rb_node *node;
451
452         if (end < start)
453                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
454                        end, start);
455         state->start = start;
456         state->end = end;
457
458         set_state_bits(tree, state, bits, changeset);
459
460         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
461         if (node) {
462                 struct extent_state *found;
463                 found = rb_entry(node, struct extent_state, rb_node);
464                 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
465                        "%llu %llu\n",
466                        found->start, found->end, start, end);
467                 return -EEXIST;
468         }
469         merge_state(tree, state);
470         return 0;
471 }
472
473 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
474                      u64 split)
475 {
476         if (tree->ops && tree->ops->split_extent_hook)
477                 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
478 }
479
480 /*
481  * split a given extent state struct in two, inserting the preallocated
482  * struct 'prealloc' as the newly created second half.  'split' indicates an
483  * offset inside 'orig' where it should be split.
484  *
485  * Before calling,
486  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
487  * are two extent state structs in the tree:
488  * prealloc: [orig->start, split - 1]
489  * orig: [ split, orig->end ]
490  *
491  * The tree locks are not taken by this function. They need to be held
492  * by the caller.
493  */
494 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
495                        struct extent_state *prealloc, u64 split)
496 {
497         struct rb_node *node;
498
499         split_cb(tree, orig, split);
500
501         prealloc->start = orig->start;
502         prealloc->end = split - 1;
503         prealloc->state = orig->state;
504         orig->start = split;
505
506         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
507                            &prealloc->rb_node, NULL, NULL);
508         if (node) {
509                 free_extent_state(prealloc);
510                 return -EEXIST;
511         }
512         return 0;
513 }
514
515 static struct extent_state *next_state(struct extent_state *state)
516 {
517         struct rb_node *next = rb_next(&state->rb_node);
518         if (next)
519                 return rb_entry(next, struct extent_state, rb_node);
520         else
521                 return NULL;
522 }
523
524 /*
525  * utility function to clear some bits in an extent state struct.
526  * it will optionally wake up any one waiting on this state (wake == 1).
527  *
528  * If no bits are set on the state struct after clearing things, the
529  * struct is freed and removed from the tree
530  */
531 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
532                                             struct extent_state *state,
533                                             unsigned *bits, int wake,
534                                             struct extent_changeset *changeset)
535 {
536         struct extent_state *next;
537         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
538
539         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
540                 u64 range = state->end - state->start + 1;
541                 WARN_ON(range > tree->dirty_bytes);
542                 tree->dirty_bytes -= range;
543         }
544         clear_state_cb(tree, state, bits);
545         add_extent_changeset(state, bits_to_clear, changeset, 0);
546         state->state &= ~bits_to_clear;
547         if (wake)
548                 wake_up(&state->wq);
549         if (state->state == 0) {
550                 next = next_state(state);
551                 if (extent_state_in_tree(state)) {
552                         rb_erase(&state->rb_node, &tree->state);
553                         RB_CLEAR_NODE(&state->rb_node);
554                         free_extent_state(state);
555                 } else {
556                         WARN_ON(1);
557                 }
558         } else {
559                 merge_state(tree, state);
560                 next = next_state(state);
561         }
562         return next;
563 }
564
565 static struct extent_state *
566 alloc_extent_state_atomic(struct extent_state *prealloc)
567 {
568         if (!prealloc)
569                 prealloc = alloc_extent_state(GFP_ATOMIC);
570
571         return prealloc;
572 }
573
574 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
575 {
576         btrfs_panic(tree_fs_info(tree), err, "Locking error: "
577                     "Extent tree was modified by another "
578                     "thread while locked.");
579 }
580
581 /*
582  * clear some bits on a range in the tree.  This may require splitting
583  * or inserting elements in the tree, so the gfp mask is used to
584  * indicate which allocations or sleeping are allowed.
585  *
586  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
587  * the given range from the tree regardless of state (ie for truncate).
588  *
589  * the range [start, end] is inclusive.
590  *
591  * This takes the tree lock, and returns 0 on success and < 0 on error.
592  */
593 static int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
594                               unsigned bits, int wake, int delete,
595                               struct extent_state **cached_state,
596                               gfp_t mask, struct extent_changeset *changeset)
597 {
598         struct extent_state *state;
599         struct extent_state *cached;
600         struct extent_state *prealloc = NULL;
601         struct rb_node *node;
602         u64 last_end;
603         int err;
604         int clear = 0;
605
606         btrfs_debug_check_extent_io_range(tree, start, end);
607
608         if (bits & EXTENT_DELALLOC)
609                 bits |= EXTENT_NORESERVE;
610
611         if (delete)
612                 bits |= ~EXTENT_CTLBITS;
613         bits |= EXTENT_FIRST_DELALLOC;
614
615         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
616                 clear = 1;
617 again:
618         if (!prealloc && gfpflags_allow_blocking(mask)) {
619                 /*
620                  * Don't care for allocation failure here because we might end
621                  * up not needing the pre-allocated extent state at all, which
622                  * is the case if we only have in the tree extent states that
623                  * cover our input range and don't cover too any other range.
624                  * If we end up needing a new extent state we allocate it later.
625                  */
626                 prealloc = alloc_extent_state(mask);
627         }
628
629         spin_lock(&tree->lock);
630         if (cached_state) {
631                 cached = *cached_state;
632
633                 if (clear) {
634                         *cached_state = NULL;
635                         cached_state = NULL;
636                 }
637
638                 if (cached && extent_state_in_tree(cached) &&
639                     cached->start <= start && cached->end > start) {
640                         if (clear)
641                                 atomic_dec(&cached->refs);
642                         state = cached;
643                         goto hit_next;
644                 }
645                 if (clear)
646                         free_extent_state(cached);
647         }
648         /*
649          * this search will find the extents that end after
650          * our range starts
651          */
652         node = tree_search(tree, start);
653         if (!node)
654                 goto out;
655         state = rb_entry(node, struct extent_state, rb_node);
656 hit_next:
657         if (state->start > end)
658                 goto out;
659         WARN_ON(state->end < start);
660         last_end = state->end;
661
662         /* the state doesn't have the wanted bits, go ahead */
663         if (!(state->state & bits)) {
664                 state = next_state(state);
665                 goto next;
666         }
667
668         /*
669          *     | ---- desired range ---- |
670          *  | state | or
671          *  | ------------- state -------------- |
672          *
673          * We need to split the extent we found, and may flip
674          * bits on second half.
675          *
676          * If the extent we found extends past our range, we
677          * just split and search again.  It'll get split again
678          * the next time though.
679          *
680          * If the extent we found is inside our range, we clear
681          * the desired bit on it.
682          */
683
684         if (state->start < start) {
685                 prealloc = alloc_extent_state_atomic(prealloc);
686                 BUG_ON(!prealloc);
687                 err = split_state(tree, state, prealloc, start);
688                 if (err)
689                         extent_io_tree_panic(tree, err);
690
691                 prealloc = NULL;
692                 if (err)
693                         goto out;
694                 if (state->end <= end) {
695                         state = clear_state_bit(tree, state, &bits, wake,
696                                                 changeset);
697                         goto next;
698                 }
699                 goto search_again;
700         }
701         /*
702          * | ---- desired range ---- |
703          *                        | state |
704          * We need to split the extent, and clear the bit
705          * on the first half
706          */
707         if (state->start <= end && state->end > end) {
708                 prealloc = alloc_extent_state_atomic(prealloc);
709                 BUG_ON(!prealloc);
710                 err = split_state(tree, state, prealloc, end + 1);
711                 if (err)
712                         extent_io_tree_panic(tree, err);
713
714                 if (wake)
715                         wake_up(&state->wq);
716
717                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
718
719                 prealloc = NULL;
720                 goto out;
721         }
722
723         state = clear_state_bit(tree, state, &bits, wake, changeset);
724 next:
725         if (last_end == (u64)-1)
726                 goto out;
727         start = last_end + 1;
728         if (start <= end && state && !need_resched())
729                 goto hit_next;
730
731 search_again:
732         if (start > end)
733                 goto out;
734         spin_unlock(&tree->lock);
735         if (gfpflags_allow_blocking(mask))
736                 cond_resched();
737         goto again;
738
739 out:
740         spin_unlock(&tree->lock);
741         if (prealloc)
742                 free_extent_state(prealloc);
743
744         return 0;
745
746 }
747
748 static void wait_on_state(struct extent_io_tree *tree,
749                           struct extent_state *state)
750                 __releases(tree->lock)
751                 __acquires(tree->lock)
752 {
753         DEFINE_WAIT(wait);
754         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
755         spin_unlock(&tree->lock);
756         schedule();
757         spin_lock(&tree->lock);
758         finish_wait(&state->wq, &wait);
759 }
760
761 /*
762  * waits for one or more bits to clear on a range in the state tree.
763  * The range [start, end] is inclusive.
764  * The tree lock is taken by this function
765  */
766 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
767                             unsigned long bits)
768 {
769         struct extent_state *state;
770         struct rb_node *node;
771
772         btrfs_debug_check_extent_io_range(tree, start, end);
773
774         spin_lock(&tree->lock);
775 again:
776         while (1) {
777                 /*
778                  * this search will find all the extents that end after
779                  * our range starts
780                  */
781                 node = tree_search(tree, start);
782 process_node:
783                 if (!node)
784                         break;
785
786                 state = rb_entry(node, struct extent_state, rb_node);
787
788                 if (state->start > end)
789                         goto out;
790
791                 if (state->state & bits) {
792                         start = state->start;
793                         atomic_inc(&state->refs);
794                         wait_on_state(tree, state);
795                         free_extent_state(state);
796                         goto again;
797                 }
798                 start = state->end + 1;
799
800                 if (start > end)
801                         break;
802
803                 if (!cond_resched_lock(&tree->lock)) {
804                         node = rb_next(node);
805                         goto process_node;
806                 }
807         }
808 out:
809         spin_unlock(&tree->lock);
810 }
811
812 static void set_state_bits(struct extent_io_tree *tree,
813                            struct extent_state *state,
814                            unsigned *bits, struct extent_changeset *changeset)
815 {
816         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
817
818         set_state_cb(tree, state, bits);
819         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
820                 u64 range = state->end - state->start + 1;
821                 tree->dirty_bytes += range;
822         }
823         add_extent_changeset(state, bits_to_set, changeset, 1);
824         state->state |= bits_to_set;
825 }
826
827 static void cache_state_if_flags(struct extent_state *state,
828                                  struct extent_state **cached_ptr,
829                                  unsigned flags)
830 {
831         if (cached_ptr && !(*cached_ptr)) {
832                 if (!flags || (state->state & flags)) {
833                         *cached_ptr = state;
834                         atomic_inc(&state->refs);
835                 }
836         }
837 }
838
839 static void cache_state(struct extent_state *state,
840                         struct extent_state **cached_ptr)
841 {
842         return cache_state_if_flags(state, cached_ptr,
843                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
844 }
845
846 /*
847  * set some bits on a range in the tree.  This may require allocations or
848  * sleeping, so the gfp mask is used to indicate what is allowed.
849  *
850  * If any of the exclusive bits are set, this will fail with -EEXIST if some
851  * part of the range already has the desired bits set.  The start of the
852  * existing range is returned in failed_start in this case.
853  *
854  * [start, end] is inclusive This takes the tree lock.
855  */
856
857 static int __must_check
858 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
859                  unsigned bits, unsigned exclusive_bits,
860                  u64 *failed_start, struct extent_state **cached_state,
861                  gfp_t mask, struct extent_changeset *changeset)
862 {
863         struct extent_state *state;
864         struct extent_state *prealloc = NULL;
865         struct rb_node *node;
866         struct rb_node **p;
867         struct rb_node *parent;
868         int err = 0;
869         u64 last_start;
870         u64 last_end;
871
872         btrfs_debug_check_extent_io_range(tree, start, end);
873
874         bits |= EXTENT_FIRST_DELALLOC;
875 again:
876         if (!prealloc && gfpflags_allow_blocking(mask)) {
877                 /*
878                  * Don't care for allocation failure here because we might end
879                  * up not needing the pre-allocated extent state at all, which
880                  * is the case if we only have in the tree extent states that
881                  * cover our input range and don't cover too any other range.
882                  * If we end up needing a new extent state we allocate it later.
883                  */
884                 prealloc = alloc_extent_state(mask);
885         }
886
887         spin_lock(&tree->lock);
888         if (cached_state && *cached_state) {
889                 state = *cached_state;
890                 if (state->start <= start && state->end > start &&
891                     extent_state_in_tree(state)) {
892                         node = &state->rb_node;
893                         goto hit_next;
894                 }
895         }
896         /*
897          * this search will find all the extents that end after
898          * our range starts.
899          */
900         node = tree_search_for_insert(tree, start, &p, &parent);
901         if (!node) {
902                 prealloc = alloc_extent_state_atomic(prealloc);
903                 BUG_ON(!prealloc);
904                 err = insert_state(tree, prealloc, start, end,
905                                    &p, &parent, &bits, changeset);
906                 if (err)
907                         extent_io_tree_panic(tree, err);
908
909                 cache_state(prealloc, cached_state);
910                 prealloc = NULL;
911                 goto out;
912         }
913         state = rb_entry(node, struct extent_state, rb_node);
914 hit_next:
915         last_start = state->start;
916         last_end = state->end;
917
918         /*
919          * | ---- desired range ---- |
920          * | state |
921          *
922          * Just lock what we found and keep going
923          */
924         if (state->start == start && state->end <= end) {
925                 if (state->state & exclusive_bits) {
926                         *failed_start = state->start;
927                         err = -EEXIST;
928                         goto out;
929                 }
930
931                 set_state_bits(tree, state, &bits, changeset);
932                 cache_state(state, cached_state);
933                 merge_state(tree, state);
934                 if (last_end == (u64)-1)
935                         goto out;
936                 start = last_end + 1;
937                 state = next_state(state);
938                 if (start < end && state && state->start == start &&
939                     !need_resched())
940                         goto hit_next;
941                 goto search_again;
942         }
943
944         /*
945          *     | ---- desired range ---- |
946          * | state |
947          *   or
948          * | ------------- state -------------- |
949          *
950          * We need to split the extent we found, and may flip bits on
951          * second half.
952          *
953          * If the extent we found extends past our
954          * range, we just split and search again.  It'll get split
955          * again the next time though.
956          *
957          * If the extent we found is inside our range, we set the
958          * desired bit on it.
959          */
960         if (state->start < start) {
961                 if (state->state & exclusive_bits) {
962                         *failed_start = start;
963                         err = -EEXIST;
964                         goto out;
965                 }
966
967                 prealloc = alloc_extent_state_atomic(prealloc);
968                 BUG_ON(!prealloc);
969                 err = split_state(tree, state, prealloc, start);
970                 if (err)
971                         extent_io_tree_panic(tree, err);
972
973                 prealloc = NULL;
974                 if (err)
975                         goto out;
976                 if (state->end <= end) {
977                         set_state_bits(tree, state, &bits, changeset);
978                         cache_state(state, cached_state);
979                         merge_state(tree, state);
980                         if (last_end == (u64)-1)
981                                 goto out;
982                         start = last_end + 1;
983                         state = next_state(state);
984                         if (start < end && state && state->start == start &&
985                             !need_resched())
986                                 goto hit_next;
987                 }
988                 goto search_again;
989         }
990         /*
991          * | ---- desired range ---- |
992          *     | state | or               | state |
993          *
994          * There's a hole, we need to insert something in it and
995          * ignore the extent we found.
996          */
997         if (state->start > start) {
998                 u64 this_end;
999                 if (end < last_start)
1000                         this_end = end;
1001                 else
1002                         this_end = last_start - 1;
1003
1004                 prealloc = alloc_extent_state_atomic(prealloc);
1005                 BUG_ON(!prealloc);
1006
1007                 /*
1008                  * Avoid to free 'prealloc' if it can be merged with
1009                  * the later extent.
1010                  */
1011                 err = insert_state(tree, prealloc, start, this_end,
1012                                    NULL, NULL, &bits, changeset);
1013                 if (err)
1014                         extent_io_tree_panic(tree, err);
1015
1016                 cache_state(prealloc, cached_state);
1017                 prealloc = NULL;
1018                 start = this_end + 1;
1019                 goto search_again;
1020         }
1021         /*
1022          * | ---- desired range ---- |
1023          *                        | state |
1024          * We need to split the extent, and set the bit
1025          * on the first half
1026          */
1027         if (state->start <= end && state->end > end) {
1028                 if (state->state & exclusive_bits) {
1029                         *failed_start = start;
1030                         err = -EEXIST;
1031                         goto out;
1032                 }
1033
1034                 prealloc = alloc_extent_state_atomic(prealloc);
1035                 BUG_ON(!prealloc);
1036                 err = split_state(tree, state, prealloc, end + 1);
1037                 if (err)
1038                         extent_io_tree_panic(tree, err);
1039
1040                 set_state_bits(tree, prealloc, &bits, changeset);
1041                 cache_state(prealloc, cached_state);
1042                 merge_state(tree, prealloc);
1043                 prealloc = NULL;
1044                 goto out;
1045         }
1046
1047 search_again:
1048         if (start > end)
1049                 goto out;
1050         spin_unlock(&tree->lock);
1051         if (gfpflags_allow_blocking(mask))
1052                 cond_resched();
1053         goto again;
1054
1055 out:
1056         spin_unlock(&tree->lock);
1057         if (prealloc)
1058                 free_extent_state(prealloc);
1059
1060         return err;
1061
1062 }
1063
1064 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1065                    unsigned bits, u64 * failed_start,
1066                    struct extent_state **cached_state, gfp_t mask)
1067 {
1068         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1069                                 cached_state, mask, NULL);
1070 }
1071
1072
1073 /**
1074  * convert_extent_bit - convert all bits in a given range from one bit to
1075  *                      another
1076  * @tree:       the io tree to search
1077  * @start:      the start offset in bytes
1078  * @end:        the end offset in bytes (inclusive)
1079  * @bits:       the bits to set in this range
1080  * @clear_bits: the bits to clear in this range
1081  * @cached_state:       state that we're going to cache
1082  *
1083  * This will go through and set bits for the given range.  If any states exist
1084  * already in this range they are set with the given bit and cleared of the
1085  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1086  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1087  * boundary bits like LOCK.
1088  *
1089  * All allocations are done with GFP_NOFS.
1090  */
1091 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1092                        unsigned bits, unsigned clear_bits,
1093                        struct extent_state **cached_state)
1094 {
1095         struct extent_state *state;
1096         struct extent_state *prealloc = NULL;
1097         struct rb_node *node;
1098         struct rb_node **p;
1099         struct rb_node *parent;
1100         int err = 0;
1101         u64 last_start;
1102         u64 last_end;
1103         bool first_iteration = true;
1104
1105         btrfs_debug_check_extent_io_range(tree, start, end);
1106
1107 again:
1108         if (!prealloc) {
1109                 /*
1110                  * Best effort, don't worry if extent state allocation fails
1111                  * here for the first iteration. We might have a cached state
1112                  * that matches exactly the target range, in which case no
1113                  * extent state allocations are needed. We'll only know this
1114                  * after locking the tree.
1115                  */
1116                 prealloc = alloc_extent_state(GFP_NOFS);
1117                 if (!prealloc && !first_iteration)
1118                         return -ENOMEM;
1119         }
1120
1121         spin_lock(&tree->lock);
1122         if (cached_state && *cached_state) {
1123                 state = *cached_state;
1124                 if (state->start <= start && state->end > start &&
1125                     extent_state_in_tree(state)) {
1126                         node = &state->rb_node;
1127                         goto hit_next;
1128                 }
1129         }
1130
1131         /*
1132          * this search will find all the extents that end after
1133          * our range starts.
1134          */
1135         node = tree_search_for_insert(tree, start, &p, &parent);
1136         if (!node) {
1137                 prealloc = alloc_extent_state_atomic(prealloc);
1138                 if (!prealloc) {
1139                         err = -ENOMEM;
1140                         goto out;
1141                 }
1142                 err = insert_state(tree, prealloc, start, end,
1143                                    &p, &parent, &bits, NULL);
1144                 if (err)
1145                         extent_io_tree_panic(tree, err);
1146                 cache_state(prealloc, cached_state);
1147                 prealloc = NULL;
1148                 goto out;
1149         }
1150         state = rb_entry(node, struct extent_state, rb_node);
1151 hit_next:
1152         last_start = state->start;
1153         last_end = state->end;
1154
1155         /*
1156          * | ---- desired range ---- |
1157          * | state |
1158          *
1159          * Just lock what we found and keep going
1160          */
1161         if (state->start == start && state->end <= end) {
1162                 set_state_bits(tree, state, &bits, NULL);
1163                 cache_state(state, cached_state);
1164                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1165                 if (last_end == (u64)-1)
1166                         goto out;
1167                 start = last_end + 1;
1168                 if (start < end && state && state->start == start &&
1169                     !need_resched())
1170                         goto hit_next;
1171                 goto search_again;
1172         }
1173
1174         /*
1175          *     | ---- desired range ---- |
1176          * | state |
1177          *   or
1178          * | ------------- state -------------- |
1179          *
1180          * We need to split the extent we found, and may flip bits on
1181          * second half.
1182          *
1183          * If the extent we found extends past our
1184          * range, we just split and search again.  It'll get split
1185          * again the next time though.
1186          *
1187          * If the extent we found is inside our range, we set the
1188          * desired bit on it.
1189          */
1190         if (state->start < start) {
1191                 prealloc = alloc_extent_state_atomic(prealloc);
1192                 if (!prealloc) {
1193                         err = -ENOMEM;
1194                         goto out;
1195                 }
1196                 err = split_state(tree, state, prealloc, start);
1197                 if (err)
1198                         extent_io_tree_panic(tree, err);
1199                 prealloc = NULL;
1200                 if (err)
1201                         goto out;
1202                 if (state->end <= end) {
1203                         set_state_bits(tree, state, &bits, NULL);
1204                         cache_state(state, cached_state);
1205                         state = clear_state_bit(tree, state, &clear_bits, 0,
1206                                                 NULL);
1207                         if (last_end == (u64)-1)
1208                                 goto out;
1209                         start = last_end + 1;
1210                         if (start < end && state && state->start == start &&
1211                             !need_resched())
1212                                 goto hit_next;
1213                 }
1214                 goto search_again;
1215         }
1216         /*
1217          * | ---- desired range ---- |
1218          *     | state | or               | state |
1219          *
1220          * There's a hole, we need to insert something in it and
1221          * ignore the extent we found.
1222          */
1223         if (state->start > start) {
1224                 u64 this_end;
1225                 if (end < last_start)
1226                         this_end = end;
1227                 else
1228                         this_end = last_start - 1;
1229
1230                 prealloc = alloc_extent_state_atomic(prealloc);
1231                 if (!prealloc) {
1232                         err = -ENOMEM;
1233                         goto out;
1234                 }
1235
1236                 /*
1237                  * Avoid to free 'prealloc' if it can be merged with
1238                  * the later extent.
1239                  */
1240                 err = insert_state(tree, prealloc, start, this_end,
1241                                    NULL, NULL, &bits, NULL);
1242                 if (err)
1243                         extent_io_tree_panic(tree, err);
1244                 cache_state(prealloc, cached_state);
1245                 prealloc = NULL;
1246                 start = this_end + 1;
1247                 goto search_again;
1248         }
1249         /*
1250          * | ---- desired range ---- |
1251          *                        | state |
1252          * We need to split the extent, and set the bit
1253          * on the first half
1254          */
1255         if (state->start <= end && state->end > end) {
1256                 prealloc = alloc_extent_state_atomic(prealloc);
1257                 if (!prealloc) {
1258                         err = -ENOMEM;
1259                         goto out;
1260                 }
1261
1262                 err = split_state(tree, state, prealloc, end + 1);
1263                 if (err)
1264                         extent_io_tree_panic(tree, err);
1265
1266                 set_state_bits(tree, prealloc, &bits, NULL);
1267                 cache_state(prealloc, cached_state);
1268                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1269                 prealloc = NULL;
1270                 goto out;
1271         }
1272
1273 search_again:
1274         if (start > end)
1275                 goto out;
1276         spin_unlock(&tree->lock);
1277         cond_resched();
1278         first_iteration = false;
1279         goto again;
1280
1281 out:
1282         spin_unlock(&tree->lock);
1283         if (prealloc)
1284                 free_extent_state(prealloc);
1285
1286         return err;
1287 }
1288
1289 /* wrappers around set/clear extent bit */
1290 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1291                            unsigned bits, struct extent_changeset *changeset)
1292 {
1293         /*
1294          * We don't support EXTENT_LOCKED yet, as current changeset will
1295          * record any bits changed, so for EXTENT_LOCKED case, it will
1296          * either fail with -EEXIST or changeset will record the whole
1297          * range.
1298          */
1299         BUG_ON(bits & EXTENT_LOCKED);
1300
1301         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1302                                 changeset);
1303 }
1304
1305 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1306                      unsigned bits, int wake, int delete,
1307                      struct extent_state **cached, gfp_t mask)
1308 {
1309         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1310                                   cached, mask, NULL);
1311 }
1312
1313 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1314                 unsigned bits, struct extent_changeset *changeset)
1315 {
1316         /*
1317          * Don't support EXTENT_LOCKED case, same reason as
1318          * set_record_extent_bits().
1319          */
1320         BUG_ON(bits & EXTENT_LOCKED);
1321
1322         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1323                                   changeset);
1324 }
1325
1326 /*
1327  * either insert or lock state struct between start and end use mask to tell
1328  * us if waiting is desired.
1329  */
1330 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1331                      struct extent_state **cached_state)
1332 {
1333         int err;
1334         u64 failed_start;
1335
1336         while (1) {
1337                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1338                                        EXTENT_LOCKED, &failed_start,
1339                                        cached_state, GFP_NOFS, NULL);
1340                 if (err == -EEXIST) {
1341                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1342                         start = failed_start;
1343                 } else
1344                         break;
1345                 WARN_ON(start > end);
1346         }
1347         return err;
1348 }
1349
1350 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1351 {
1352         int err;
1353         u64 failed_start;
1354
1355         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1356                                &failed_start, NULL, GFP_NOFS, NULL);
1357         if (err == -EEXIST) {
1358                 if (failed_start > start)
1359                         clear_extent_bit(tree, start, failed_start - 1,
1360                                          EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1361                 return 0;
1362         }
1363         return 1;
1364 }
1365
1366 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1367 {
1368         unsigned long index = start >> PAGE_SHIFT;
1369         unsigned long end_index = end >> PAGE_SHIFT;
1370         struct page *page;
1371
1372         while (index <= end_index) {
1373                 page = find_get_page(inode->i_mapping, index);
1374                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1375                 clear_page_dirty_for_io(page);
1376                 put_page(page);
1377                 index++;
1378         }
1379 }
1380
1381 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1382 {
1383         unsigned long index = start >> PAGE_SHIFT;
1384         unsigned long end_index = end >> PAGE_SHIFT;
1385         struct page *page;
1386
1387         while (index <= end_index) {
1388                 page = find_get_page(inode->i_mapping, index);
1389                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1390                 __set_page_dirty_nobuffers(page);
1391                 account_page_redirty(page);
1392                 put_page(page);
1393                 index++;
1394         }
1395 }
1396
1397 /*
1398  * helper function to set both pages and extents in the tree writeback
1399  */
1400 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1401 {
1402         unsigned long index = start >> PAGE_SHIFT;
1403         unsigned long end_index = end >> PAGE_SHIFT;
1404         struct page *page;
1405
1406         while (index <= end_index) {
1407                 page = find_get_page(tree->mapping, index);
1408                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1409                 set_page_writeback(page);
1410                 put_page(page);
1411                 index++;
1412         }
1413 }
1414
1415 /* find the first state struct with 'bits' set after 'start', and
1416  * return it.  tree->lock must be held.  NULL will returned if
1417  * nothing was found after 'start'
1418  */
1419 static struct extent_state *
1420 find_first_extent_bit_state(struct extent_io_tree *tree,
1421                             u64 start, unsigned bits)
1422 {
1423         struct rb_node *node;
1424         struct extent_state *state;
1425
1426         /*
1427          * this search will find all the extents that end after
1428          * our range starts.
1429          */
1430         node = tree_search(tree, start);
1431         if (!node)
1432                 goto out;
1433
1434         while (1) {
1435                 state = rb_entry(node, struct extent_state, rb_node);
1436                 if (state->end >= start && (state->state & bits))
1437                         return state;
1438
1439                 node = rb_next(node);
1440                 if (!node)
1441                         break;
1442         }
1443 out:
1444         return NULL;
1445 }
1446
1447 /*
1448  * find the first offset in the io tree with 'bits' set. zero is
1449  * returned if we find something, and *start_ret and *end_ret are
1450  * set to reflect the state struct that was found.
1451  *
1452  * If nothing was found, 1 is returned. If found something, return 0.
1453  */
1454 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1455                           u64 *start_ret, u64 *end_ret, unsigned bits,
1456                           struct extent_state **cached_state)
1457 {
1458         struct extent_state *state;
1459         struct rb_node *n;
1460         int ret = 1;
1461
1462         spin_lock(&tree->lock);
1463         if (cached_state && *cached_state) {
1464                 state = *cached_state;
1465                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1466                         n = rb_next(&state->rb_node);
1467                         while (n) {
1468                                 state = rb_entry(n, struct extent_state,
1469                                                  rb_node);
1470                                 if (state->state & bits)
1471                                         goto got_it;
1472                                 n = rb_next(n);
1473                         }
1474                         free_extent_state(*cached_state);
1475                         *cached_state = NULL;
1476                         goto out;
1477                 }
1478                 free_extent_state(*cached_state);
1479                 *cached_state = NULL;
1480         }
1481
1482         state = find_first_extent_bit_state(tree, start, bits);
1483 got_it:
1484         if (state) {
1485                 cache_state_if_flags(state, cached_state, 0);
1486                 *start_ret = state->start;
1487                 *end_ret = state->end;
1488                 ret = 0;
1489         }
1490 out:
1491         spin_unlock(&tree->lock);
1492         return ret;
1493 }
1494
1495 /*
1496  * find a contiguous range of bytes in the file marked as delalloc, not
1497  * more than 'max_bytes'.  start and end are used to return the range,
1498  *
1499  * 1 is returned if we find something, 0 if nothing was in the tree
1500  */
1501 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1502                                         u64 *start, u64 *end, u64 max_bytes,
1503                                         struct extent_state **cached_state)
1504 {
1505         struct rb_node *node;
1506         struct extent_state *state;
1507         u64 cur_start = *start;
1508         u64 found = 0;
1509         u64 total_bytes = 0;
1510
1511         spin_lock(&tree->lock);
1512
1513         /*
1514          * this search will find all the extents that end after
1515          * our range starts.
1516          */
1517         node = tree_search(tree, cur_start);
1518         if (!node) {
1519                 if (!found)
1520                         *end = (u64)-1;
1521                 goto out;
1522         }
1523
1524         while (1) {
1525                 state = rb_entry(node, struct extent_state, rb_node);
1526                 if (found && (state->start != cur_start ||
1527                               (state->state & EXTENT_BOUNDARY))) {
1528                         goto out;
1529                 }
1530                 if (!(state->state & EXTENT_DELALLOC)) {
1531                         if (!found)
1532                                 *end = state->end;
1533                         goto out;
1534                 }
1535                 if (!found) {
1536                         *start = state->start;
1537                         *cached_state = state;
1538                         atomic_inc(&state->refs);
1539                 }
1540                 found++;
1541                 *end = state->end;
1542                 cur_start = state->end + 1;
1543                 node = rb_next(node);
1544                 total_bytes += state->end - state->start + 1;
1545                 if (total_bytes >= max_bytes)
1546                         break;
1547                 if (!node)
1548                         break;
1549         }
1550 out:
1551         spin_unlock(&tree->lock);
1552         return found;
1553 }
1554
1555 static noinline void __unlock_for_delalloc(struct inode *inode,
1556                                            struct page *locked_page,
1557                                            u64 start, u64 end)
1558 {
1559         int ret;
1560         struct page *pages[16];
1561         unsigned long index = start >> PAGE_SHIFT;
1562         unsigned long end_index = end >> PAGE_SHIFT;
1563         unsigned long nr_pages = end_index - index + 1;
1564         int i;
1565
1566         if (index == locked_page->index && end_index == index)
1567                 return;
1568
1569         while (nr_pages > 0) {
1570                 ret = find_get_pages_contig(inode->i_mapping, index,
1571                                      min_t(unsigned long, nr_pages,
1572                                      ARRAY_SIZE(pages)), pages);
1573                 for (i = 0; i < ret; i++) {
1574                         if (pages[i] != locked_page)
1575                                 unlock_page(pages[i]);
1576                         put_page(pages[i]);
1577                 }
1578                 nr_pages -= ret;
1579                 index += ret;
1580                 cond_resched();
1581         }
1582 }
1583
1584 static noinline int lock_delalloc_pages(struct inode *inode,
1585                                         struct page *locked_page,
1586                                         u64 delalloc_start,
1587                                         u64 delalloc_end)
1588 {
1589         unsigned long index = delalloc_start >> PAGE_SHIFT;
1590         unsigned long start_index = index;
1591         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1592         unsigned long pages_locked = 0;
1593         struct page *pages[16];
1594         unsigned long nrpages;
1595         int ret;
1596         int i;
1597
1598         /* the caller is responsible for locking the start index */
1599         if (index == locked_page->index && index == end_index)
1600                 return 0;
1601
1602         /* skip the page at the start index */
1603         nrpages = end_index - index + 1;
1604         while (nrpages > 0) {
1605                 ret = find_get_pages_contig(inode->i_mapping, index,
1606                                      min_t(unsigned long,
1607                                      nrpages, ARRAY_SIZE(pages)), pages);
1608                 if (ret == 0) {
1609                         ret = -EAGAIN;
1610                         goto done;
1611                 }
1612                 /* now we have an array of pages, lock them all */
1613                 for (i = 0; i < ret; i++) {
1614                         /*
1615                          * the caller is taking responsibility for
1616                          * locked_page
1617                          */
1618                         if (pages[i] != locked_page) {
1619                                 lock_page(pages[i]);
1620                                 if (!PageDirty(pages[i]) ||
1621                                     pages[i]->mapping != inode->i_mapping) {
1622                                         ret = -EAGAIN;
1623                                         unlock_page(pages[i]);
1624                                         put_page(pages[i]);
1625                                         goto done;
1626                                 }
1627                         }
1628                         put_page(pages[i]);
1629                         pages_locked++;
1630                 }
1631                 nrpages -= ret;
1632                 index += ret;
1633                 cond_resched();
1634         }
1635         ret = 0;
1636 done:
1637         if (ret && pages_locked) {
1638                 __unlock_for_delalloc(inode, locked_page,
1639                               delalloc_start,
1640                               ((u64)(start_index + pages_locked - 1)) <<
1641                               PAGE_SHIFT);
1642         }
1643         return ret;
1644 }
1645
1646 /*
1647  * find a contiguous range of bytes in the file marked as delalloc, not
1648  * more than 'max_bytes'.  start and end are used to return the range,
1649  *
1650  * 1 is returned if we find something, 0 if nothing was in the tree
1651  */
1652 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1653                                     struct extent_io_tree *tree,
1654                                     struct page *locked_page, u64 *start,
1655                                     u64 *end, u64 max_bytes)
1656 {
1657         u64 delalloc_start;
1658         u64 delalloc_end;
1659         u64 found;
1660         struct extent_state *cached_state = NULL;
1661         int ret;
1662         int loops = 0;
1663
1664 again:
1665         /* step one, find a bunch of delalloc bytes starting at start */
1666         delalloc_start = *start;
1667         delalloc_end = 0;
1668         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1669                                     max_bytes, &cached_state);
1670         if (!found || delalloc_end <= *start) {
1671                 *start = delalloc_start;
1672                 *end = delalloc_end;
1673                 free_extent_state(cached_state);
1674                 return 0;
1675         }
1676
1677         /*
1678          * start comes from the offset of locked_page.  We have to lock
1679          * pages in order, so we can't process delalloc bytes before
1680          * locked_page
1681          */
1682         if (delalloc_start < *start)
1683                 delalloc_start = *start;
1684
1685         /*
1686          * make sure to limit the number of pages we try to lock down
1687          */
1688         if (delalloc_end + 1 - delalloc_start > max_bytes)
1689                 delalloc_end = delalloc_start + max_bytes - 1;
1690
1691         /* step two, lock all the pages after the page that has start */
1692         ret = lock_delalloc_pages(inode, locked_page,
1693                                   delalloc_start, delalloc_end);
1694         if (ret == -EAGAIN) {
1695                 /* some of the pages are gone, lets avoid looping by
1696                  * shortening the size of the delalloc range we're searching
1697                  */
1698                 free_extent_state(cached_state);
1699                 cached_state = NULL;
1700                 if (!loops) {
1701                         max_bytes = PAGE_SIZE;
1702                         loops = 1;
1703                         goto again;
1704                 } else {
1705                         found = 0;
1706                         goto out_failed;
1707                 }
1708         }
1709         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1710
1711         /* step three, lock the state bits for the whole range */
1712         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1713
1714         /* then test to make sure it is all still delalloc */
1715         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1716                              EXTENT_DELALLOC, 1, cached_state);
1717         if (!ret) {
1718                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1719                                      &cached_state, GFP_NOFS);
1720                 __unlock_for_delalloc(inode, locked_page,
1721                               delalloc_start, delalloc_end);
1722                 cond_resched();
1723                 goto again;
1724         }
1725         free_extent_state(cached_state);
1726         *start = delalloc_start;
1727         *end = delalloc_end;
1728 out_failed:
1729         return found;
1730 }
1731
1732 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1733                                  struct page *locked_page,
1734                                  unsigned clear_bits,
1735                                  unsigned long page_ops)
1736 {
1737         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1738         int ret;
1739         struct page *pages[16];
1740         unsigned long index = start >> PAGE_SHIFT;
1741         unsigned long end_index = end >> PAGE_SHIFT;
1742         unsigned long nr_pages = end_index - index + 1;
1743         int i;
1744
1745         clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1746         if (page_ops == 0)
1747                 return;
1748
1749         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1750                 mapping_set_error(inode->i_mapping, -EIO);
1751
1752         while (nr_pages > 0) {
1753                 ret = find_get_pages_contig(inode->i_mapping, index,
1754                                      min_t(unsigned long,
1755                                      nr_pages, ARRAY_SIZE(pages)), pages);
1756                 for (i = 0; i < ret; i++) {
1757
1758                         if (page_ops & PAGE_SET_PRIVATE2)
1759                                 SetPagePrivate2(pages[i]);
1760
1761                         if (pages[i] == locked_page) {
1762                                 put_page(pages[i]);
1763                                 continue;
1764                         }
1765                         if (page_ops & PAGE_CLEAR_DIRTY)
1766                                 clear_page_dirty_for_io(pages[i]);
1767                         if (page_ops & PAGE_SET_WRITEBACK)
1768                                 set_page_writeback(pages[i]);
1769                         if (page_ops & PAGE_SET_ERROR)
1770                                 SetPageError(pages[i]);
1771                         if (page_ops & PAGE_END_WRITEBACK)
1772                                 end_page_writeback(pages[i]);
1773                         if (page_ops & PAGE_UNLOCK)
1774                                 unlock_page(pages[i]);
1775                         put_page(pages[i]);
1776                 }
1777                 nr_pages -= ret;
1778                 index += ret;
1779                 cond_resched();
1780         }
1781 }
1782
1783 /*
1784  * count the number of bytes in the tree that have a given bit(s)
1785  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1786  * cached.  The total number found is returned.
1787  */
1788 u64 count_range_bits(struct extent_io_tree *tree,
1789                      u64 *start, u64 search_end, u64 max_bytes,
1790                      unsigned bits, int contig)
1791 {
1792         struct rb_node *node;
1793         struct extent_state *state;
1794         u64 cur_start = *start;
1795         u64 total_bytes = 0;
1796         u64 last = 0;
1797         int found = 0;
1798
1799         if (WARN_ON(search_end <= cur_start))
1800                 return 0;
1801
1802         spin_lock(&tree->lock);
1803         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1804                 total_bytes = tree->dirty_bytes;
1805                 goto out;
1806         }
1807         /*
1808          * this search will find all the extents that end after
1809          * our range starts.
1810          */
1811         node = tree_search(tree, cur_start);
1812         if (!node)
1813                 goto out;
1814
1815         while (1) {
1816                 state = rb_entry(node, struct extent_state, rb_node);
1817                 if (state->start > search_end)
1818                         break;
1819                 if (contig && found && state->start > last + 1)
1820                         break;
1821                 if (state->end >= cur_start && (state->state & bits) == bits) {
1822                         total_bytes += min(search_end, state->end) + 1 -
1823                                        max(cur_start, state->start);
1824                         if (total_bytes >= max_bytes)
1825                                 break;
1826                         if (!found) {
1827                                 *start = max(cur_start, state->start);
1828                                 found = 1;
1829                         }
1830                         last = state->end;
1831                 } else if (contig && found) {
1832                         break;
1833                 }
1834                 node = rb_next(node);
1835                 if (!node)
1836                         break;
1837         }
1838 out:
1839         spin_unlock(&tree->lock);
1840         return total_bytes;
1841 }
1842
1843 /*
1844  * set the private field for a given byte offset in the tree.  If there isn't
1845  * an extent_state there already, this does nothing.
1846  */
1847 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1848                 struct io_failure_record *failrec)
1849 {
1850         struct rb_node *node;
1851         struct extent_state *state;
1852         int ret = 0;
1853
1854         spin_lock(&tree->lock);
1855         /*
1856          * this search will find all the extents that end after
1857          * our range starts.
1858          */
1859         node = tree_search(tree, start);
1860         if (!node) {
1861                 ret = -ENOENT;
1862                 goto out;
1863         }
1864         state = rb_entry(node, struct extent_state, rb_node);
1865         if (state->start != start) {
1866                 ret = -ENOENT;
1867                 goto out;
1868         }
1869         state->failrec = failrec;
1870 out:
1871         spin_unlock(&tree->lock);
1872         return ret;
1873 }
1874
1875 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1876                 struct io_failure_record **failrec)
1877 {
1878         struct rb_node *node;
1879         struct extent_state *state;
1880         int ret = 0;
1881
1882         spin_lock(&tree->lock);
1883         /*
1884          * this search will find all the extents that end after
1885          * our range starts.
1886          */
1887         node = tree_search(tree, start);
1888         if (!node) {
1889                 ret = -ENOENT;
1890                 goto out;
1891         }
1892         state = rb_entry(node, struct extent_state, rb_node);
1893         if (state->start != start) {
1894                 ret = -ENOENT;
1895                 goto out;
1896         }
1897         *failrec = state->failrec;
1898 out:
1899         spin_unlock(&tree->lock);
1900         return ret;
1901 }
1902
1903 /*
1904  * searches a range in the state tree for a given mask.
1905  * If 'filled' == 1, this returns 1 only if every extent in the tree
1906  * has the bits set.  Otherwise, 1 is returned if any bit in the
1907  * range is found set.
1908  */
1909 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1910                    unsigned bits, int filled, struct extent_state *cached)
1911 {
1912         struct extent_state *state = NULL;
1913         struct rb_node *node;
1914         int bitset = 0;
1915
1916         spin_lock(&tree->lock);
1917         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1918             cached->end > start)
1919                 node = &cached->rb_node;
1920         else
1921                 node = tree_search(tree, start);
1922         while (node && start <= end) {
1923                 state = rb_entry(node, struct extent_state, rb_node);
1924
1925                 if (filled && state->start > start) {
1926                         bitset = 0;
1927                         break;
1928                 }
1929
1930                 if (state->start > end)
1931                         break;
1932
1933                 if (state->state & bits) {
1934                         bitset = 1;
1935                         if (!filled)
1936                                 break;
1937                 } else if (filled) {
1938                         bitset = 0;
1939                         break;
1940                 }
1941
1942                 if (state->end == (u64)-1)
1943                         break;
1944
1945                 start = state->end + 1;
1946                 if (start > end)
1947                         break;
1948                 node = rb_next(node);
1949                 if (!node) {
1950                         if (filled)
1951                                 bitset = 0;
1952                         break;
1953                 }
1954         }
1955         spin_unlock(&tree->lock);
1956         return bitset;
1957 }
1958
1959 /*
1960  * helper function to set a given page up to date if all the
1961  * extents in the tree for that page are up to date
1962  */
1963 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1964 {
1965         u64 start = page_offset(page);
1966         u64 end = start + PAGE_SIZE - 1;
1967         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1968                 SetPageUptodate(page);
1969 }
1970
1971 int free_io_failure(struct inode *inode, struct io_failure_record *rec)
1972 {
1973         int ret;
1974         int err = 0;
1975         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1976
1977         set_state_failrec(failure_tree, rec->start, NULL);
1978         ret = clear_extent_bits(failure_tree, rec->start,
1979                                 rec->start + rec->len - 1,
1980                                 EXTENT_LOCKED | EXTENT_DIRTY);
1981         if (ret)
1982                 err = ret;
1983
1984         ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1985                                 rec->start + rec->len - 1,
1986                                 EXTENT_DAMAGED);
1987         if (ret && !err)
1988                 err = ret;
1989
1990         kfree(rec);
1991         return err;
1992 }
1993
1994 /*
1995  * this bypasses the standard btrfs submit functions deliberately, as
1996  * the standard behavior is to write all copies in a raid setup. here we only
1997  * want to write the one bad copy. so we do the mapping for ourselves and issue
1998  * submit_bio directly.
1999  * to avoid any synchronization issues, wait for the data after writing, which
2000  * actually prevents the read that triggered the error from finishing.
2001  * currently, there can be no more than two copies of every data bit. thus,
2002  * exactly one rewrite is required.
2003  */
2004 int repair_io_failure(struct inode *inode, u64 start, u64 length, u64 logical,
2005                       struct page *page, unsigned int pg_offset, int mirror_num)
2006 {
2007         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2008         struct bio *bio;
2009         struct btrfs_device *dev;
2010         u64 map_length = 0;
2011         u64 sector;
2012         struct btrfs_bio *bbio = NULL;
2013         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2014         int ret;
2015
2016         ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2017         BUG_ON(!mirror_num);
2018
2019         /* we can't repair anything in raid56 yet */
2020         if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2021                 return 0;
2022
2023         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2024         if (!bio)
2025                 return -EIO;
2026         bio->bi_iter.bi_size = 0;
2027         map_length = length;
2028
2029         /*
2030          * Avoid races with device replace and make sure our bbio has devices
2031          * associated to its stripes that don't go away while we are doing the
2032          * read repair operation.
2033          */
2034         btrfs_bio_counter_inc_blocked(fs_info);
2035         ret = btrfs_map_block(fs_info, WRITE, logical,
2036                               &map_length, &bbio, mirror_num);
2037         if (ret) {
2038                 btrfs_bio_counter_dec(fs_info);
2039                 bio_put(bio);
2040                 return -EIO;
2041         }
2042         BUG_ON(mirror_num != bbio->mirror_num);
2043         sector = bbio->stripes[mirror_num-1].physical >> 9;
2044         bio->bi_iter.bi_sector = sector;
2045         dev = bbio->stripes[mirror_num-1].dev;
2046         btrfs_put_bbio(bbio);
2047         if (!dev || !dev->bdev || !dev->writeable) {
2048                 btrfs_bio_counter_dec(fs_info);
2049                 bio_put(bio);
2050                 return -EIO;
2051         }
2052         bio->bi_bdev = dev->bdev;
2053         bio_set_op_attrs(bio, REQ_OP_WRITE, WRITE_SYNC);
2054         bio_add_page(bio, page, length, pg_offset);
2055
2056         if (btrfsic_submit_bio_wait(bio)) {
2057                 /* try to remap that extent elsewhere? */
2058                 btrfs_bio_counter_dec(fs_info);
2059                 bio_put(bio);
2060                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2061                 return -EIO;
2062         }
2063
2064         btrfs_info_rl_in_rcu(fs_info,
2065                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2066                                   btrfs_ino(inode), start,
2067                                   rcu_str_deref(dev->name), sector);
2068         btrfs_bio_counter_dec(fs_info);
2069         bio_put(bio);
2070         return 0;
2071 }
2072
2073 int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2074                          int mirror_num)
2075 {
2076         u64 start = eb->start;
2077         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2078         int ret = 0;
2079
2080         if (root->fs_info->sb->s_flags & MS_RDONLY)
2081                 return -EROFS;
2082
2083         for (i = 0; i < num_pages; i++) {
2084                 struct page *p = eb->pages[i];
2085
2086                 ret = repair_io_failure(root->fs_info->btree_inode, start,
2087                                         PAGE_SIZE, start, p,
2088                                         start - page_offset(p), mirror_num);
2089                 if (ret)
2090                         break;
2091                 start += PAGE_SIZE;
2092         }
2093
2094         return ret;
2095 }
2096
2097 /*
2098  * each time an IO finishes, we do a fast check in the IO failure tree
2099  * to see if we need to process or clean up an io_failure_record
2100  */
2101 int clean_io_failure(struct inode *inode, u64 start, struct page *page,
2102                      unsigned int pg_offset)
2103 {
2104         u64 private;
2105         struct io_failure_record *failrec;
2106         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2107         struct extent_state *state;
2108         int num_copies;
2109         int ret;
2110
2111         private = 0;
2112         ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2113                                 (u64)-1, 1, EXTENT_DIRTY, 0);
2114         if (!ret)
2115                 return 0;
2116
2117         ret = get_state_failrec(&BTRFS_I(inode)->io_failure_tree, start,
2118                         &failrec);
2119         if (ret)
2120                 return 0;
2121
2122         BUG_ON(!failrec->this_mirror);
2123
2124         if (failrec->in_validation) {
2125                 /* there was no real error, just free the record */
2126                 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2127                          failrec->start);
2128                 goto out;
2129         }
2130         if (fs_info->sb->s_flags & MS_RDONLY)
2131                 goto out;
2132
2133         spin_lock(&BTRFS_I(inode)->io_tree.lock);
2134         state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2135                                             failrec->start,
2136                                             EXTENT_LOCKED);
2137         spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2138
2139         if (state && state->start <= failrec->start &&
2140             state->end >= failrec->start + failrec->len - 1) {
2141                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2142                                               failrec->len);
2143                 if (num_copies > 1)  {
2144                         repair_io_failure(inode, start, failrec->len,
2145                                           failrec->logical, page,
2146                                           pg_offset, failrec->failed_mirror);
2147                 }
2148         }
2149
2150 out:
2151         free_io_failure(inode, failrec);
2152
2153         return 0;
2154 }
2155
2156 /*
2157  * Can be called when
2158  * - hold extent lock
2159  * - under ordered extent
2160  * - the inode is freeing
2161  */
2162 void btrfs_free_io_failure_record(struct inode *inode, u64 start, u64 end)
2163 {
2164         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2165         struct io_failure_record *failrec;
2166         struct extent_state *state, *next;
2167
2168         if (RB_EMPTY_ROOT(&failure_tree->state))
2169                 return;
2170
2171         spin_lock(&failure_tree->lock);
2172         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2173         while (state) {
2174                 if (state->start > end)
2175                         break;
2176
2177                 ASSERT(state->end <= end);
2178
2179                 next = next_state(state);
2180
2181                 failrec = state->failrec;
2182                 free_extent_state(state);
2183                 kfree(failrec);
2184
2185                 state = next;
2186         }
2187         spin_unlock(&failure_tree->lock);
2188 }
2189
2190 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2191                 struct io_failure_record **failrec_ret)
2192 {
2193         struct io_failure_record *failrec;
2194         struct extent_map *em;
2195         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2196         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2197         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2198         int ret;
2199         u64 logical;
2200
2201         ret = get_state_failrec(failure_tree, start, &failrec);
2202         if (ret) {
2203                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2204                 if (!failrec)
2205                         return -ENOMEM;
2206
2207                 failrec->start = start;
2208                 failrec->len = end - start + 1;
2209                 failrec->this_mirror = 0;
2210                 failrec->bio_flags = 0;
2211                 failrec->in_validation = 0;
2212
2213                 read_lock(&em_tree->lock);
2214                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2215                 if (!em) {
2216                         read_unlock(&em_tree->lock);
2217                         kfree(failrec);
2218                         return -EIO;
2219                 }
2220
2221                 if (em->start > start || em->start + em->len <= start) {
2222                         free_extent_map(em);
2223                         em = NULL;
2224                 }
2225                 read_unlock(&em_tree->lock);
2226                 if (!em) {
2227                         kfree(failrec);
2228                         return -EIO;
2229                 }
2230
2231                 logical = start - em->start;
2232                 logical = em->block_start + logical;
2233                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2234                         logical = em->block_start;
2235                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2236                         extent_set_compress_type(&failrec->bio_flags,
2237                                                  em->compress_type);
2238                 }
2239
2240                 pr_debug("Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu\n",
2241                          logical, start, failrec->len);
2242
2243                 failrec->logical = logical;
2244                 free_extent_map(em);
2245
2246                 /* set the bits in the private failure tree */
2247                 ret = set_extent_bits(failure_tree, start, end,
2248                                         EXTENT_LOCKED | EXTENT_DIRTY);
2249                 if (ret >= 0)
2250                         ret = set_state_failrec(failure_tree, start, failrec);
2251                 /* set the bits in the inode's tree */
2252                 if (ret >= 0)
2253                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2254                 if (ret < 0) {
2255                         kfree(failrec);
2256                         return ret;
2257                 }
2258         } else {
2259                 pr_debug("Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d\n",
2260                          failrec->logical, failrec->start, failrec->len,
2261                          failrec->in_validation);
2262                 /*
2263                  * when data can be on disk more than twice, add to failrec here
2264                  * (e.g. with a list for failed_mirror) to make
2265                  * clean_io_failure() clean all those errors at once.
2266                  */
2267         }
2268
2269         *failrec_ret = failrec;
2270
2271         return 0;
2272 }
2273
2274 int btrfs_check_repairable(struct inode *inode, struct bio *failed_bio,
2275                            struct io_failure_record *failrec, int failed_mirror)
2276 {
2277         int num_copies;
2278
2279         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2280                                       failrec->logical, failrec->len);
2281         if (num_copies == 1) {
2282                 /*
2283                  * we only have a single copy of the data, so don't bother with
2284                  * all the retry and error correction code that follows. no
2285                  * matter what the error is, it is very likely to persist.
2286                  */
2287                 pr_debug("Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2288                          num_copies, failrec->this_mirror, failed_mirror);
2289                 return 0;
2290         }
2291
2292         /*
2293          * there are two premises:
2294          *      a) deliver good data to the caller
2295          *      b) correct the bad sectors on disk
2296          */
2297         if (failed_bio->bi_vcnt > 1) {
2298                 /*
2299                  * to fulfill b), we need to know the exact failing sectors, as
2300                  * we don't want to rewrite any more than the failed ones. thus,
2301                  * we need separate read requests for the failed bio
2302                  *
2303                  * if the following BUG_ON triggers, our validation request got
2304                  * merged. we need separate requests for our algorithm to work.
2305                  */
2306                 BUG_ON(failrec->in_validation);
2307                 failrec->in_validation = 1;
2308                 failrec->this_mirror = failed_mirror;
2309         } else {
2310                 /*
2311                  * we're ready to fulfill a) and b) alongside. get a good copy
2312                  * of the failed sector and if we succeed, we have setup
2313                  * everything for repair_io_failure to do the rest for us.
2314                  */
2315                 if (failrec->in_validation) {
2316                         BUG_ON(failrec->this_mirror != failed_mirror);
2317                         failrec->in_validation = 0;
2318                         failrec->this_mirror = 0;
2319                 }
2320                 failrec->failed_mirror = failed_mirror;
2321                 failrec->this_mirror++;
2322                 if (failrec->this_mirror == failed_mirror)
2323                         failrec->this_mirror++;
2324         }
2325
2326         if (failrec->this_mirror > num_copies) {
2327                 pr_debug("Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2328                          num_copies, failrec->this_mirror, failed_mirror);
2329                 return 0;
2330         }
2331
2332         return 1;
2333 }
2334
2335
2336 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2337                                     struct io_failure_record *failrec,
2338                                     struct page *page, int pg_offset, int icsum,
2339                                     bio_end_io_t *endio_func, void *data)
2340 {
2341         struct bio *bio;
2342         struct btrfs_io_bio *btrfs_failed_bio;
2343         struct btrfs_io_bio *btrfs_bio;
2344
2345         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2346         if (!bio)
2347                 return NULL;
2348
2349         bio->bi_end_io = endio_func;
2350         bio->bi_iter.bi_sector = failrec->logical >> 9;
2351         bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2352         bio->bi_iter.bi_size = 0;
2353         bio->bi_private = data;
2354
2355         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2356         if (btrfs_failed_bio->csum) {
2357                 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2358                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2359
2360                 btrfs_bio = btrfs_io_bio(bio);
2361                 btrfs_bio->csum = btrfs_bio->csum_inline;
2362                 icsum *= csum_size;
2363                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2364                        csum_size);
2365         }
2366
2367         bio_add_page(bio, page, failrec->len, pg_offset);
2368
2369         return bio;
2370 }
2371
2372 /*
2373  * this is a generic handler for readpage errors (default
2374  * readpage_io_failed_hook). if other copies exist, read those and write back
2375  * good data to the failed position. does not investigate in remapping the
2376  * failed extent elsewhere, hoping the device will be smart enough to do this as
2377  * needed
2378  */
2379
2380 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2381                               struct page *page, u64 start, u64 end,
2382                               int failed_mirror)
2383 {
2384         struct io_failure_record *failrec;
2385         struct inode *inode = page->mapping->host;
2386         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2387         struct bio *bio;
2388         int read_mode;
2389         int ret;
2390
2391         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2392
2393         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2394         if (ret)
2395                 return ret;
2396
2397         ret = btrfs_check_repairable(inode, failed_bio, failrec, failed_mirror);
2398         if (!ret) {
2399                 free_io_failure(inode, failrec);
2400                 return -EIO;
2401         }
2402
2403         if (failed_bio->bi_vcnt > 1)
2404                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2405         else
2406                 read_mode = READ_SYNC;
2407
2408         phy_offset >>= inode->i_sb->s_blocksize_bits;
2409         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2410                                       start - page_offset(page),
2411                                       (int)phy_offset, failed_bio->bi_end_io,
2412                                       NULL);
2413         if (!bio) {
2414                 free_io_failure(inode, failrec);
2415                 return -EIO;
2416         }
2417         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2418
2419         pr_debug("Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d\n",
2420                  read_mode, failrec->this_mirror, failrec->in_validation);
2421
2422         ret = tree->ops->submit_bio_hook(inode, bio, failrec->this_mirror,
2423                                          failrec->bio_flags, 0);
2424         if (ret) {
2425                 free_io_failure(inode, failrec);
2426                 bio_put(bio);
2427         }
2428
2429         return ret;
2430 }
2431
2432 /* lots and lots of room for performance fixes in the end_bio funcs */
2433
2434 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2435 {
2436         int uptodate = (err == 0);
2437         struct extent_io_tree *tree;
2438         int ret = 0;
2439
2440         tree = &BTRFS_I(page->mapping->host)->io_tree;
2441
2442         if (tree->ops && tree->ops->writepage_end_io_hook) {
2443                 ret = tree->ops->writepage_end_io_hook(page, start,
2444                                                end, NULL, uptodate);
2445                 if (ret)
2446                         uptodate = 0;
2447         }
2448
2449         if (!uptodate) {
2450                 ClearPageUptodate(page);
2451                 SetPageError(page);
2452                 ret = ret < 0 ? ret : -EIO;
2453                 mapping_set_error(page->mapping, ret);
2454         }
2455 }
2456
2457 /*
2458  * after a writepage IO is done, we need to:
2459  * clear the uptodate bits on error
2460  * clear the writeback bits in the extent tree for this IO
2461  * end_page_writeback if the page has no more pending IO
2462  *
2463  * Scheduling is not allowed, so the extent state tree is expected
2464  * to have one and only one object corresponding to this IO.
2465  */
2466 static void end_bio_extent_writepage(struct bio *bio)
2467 {
2468         struct bio_vec *bvec;
2469         u64 start;
2470         u64 end;
2471         int i;
2472
2473         bio_for_each_segment_all(bvec, bio, i) {
2474                 struct page *page = bvec->bv_page;
2475
2476                 /* We always issue full-page reads, but if some block
2477                  * in a page fails to read, blk_update_request() will
2478                  * advance bv_offset and adjust bv_len to compensate.
2479                  * Print a warning for nonzero offsets, and an error
2480                  * if they don't add up to a full page.  */
2481                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2482                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2483                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2484                                    "partial page write in btrfs with offset %u and length %u",
2485                                         bvec->bv_offset, bvec->bv_len);
2486                         else
2487                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2488                                    "incomplete page write in btrfs with offset %u and "
2489                                    "length %u",
2490                                         bvec->bv_offset, bvec->bv_len);
2491                 }
2492
2493                 start = page_offset(page);
2494                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2495
2496                 end_extent_writepage(page, bio->bi_error, start, end);
2497                 end_page_writeback(page);
2498         }
2499
2500         bio_put(bio);
2501 }
2502
2503 static void
2504 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2505                               int uptodate)
2506 {
2507         struct extent_state *cached = NULL;
2508         u64 end = start + len - 1;
2509
2510         if (uptodate && tree->track_uptodate)
2511                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2512         unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2513 }
2514
2515 /*
2516  * after a readpage IO is done, we need to:
2517  * clear the uptodate bits on error
2518  * set the uptodate bits if things worked
2519  * set the page up to date if all extents in the tree are uptodate
2520  * clear the lock bit in the extent tree
2521  * unlock the page if there are no other extents locked for it
2522  *
2523  * Scheduling is not allowed, so the extent state tree is expected
2524  * to have one and only one object corresponding to this IO.
2525  */
2526 static void end_bio_extent_readpage(struct bio *bio)
2527 {
2528         struct bio_vec *bvec;
2529         int uptodate = !bio->bi_error;
2530         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2531         struct extent_io_tree *tree;
2532         u64 offset = 0;
2533         u64 start;
2534         u64 end;
2535         u64 len;
2536         u64 extent_start = 0;
2537         u64 extent_len = 0;
2538         int mirror;
2539         int ret;
2540         int i;
2541
2542         bio_for_each_segment_all(bvec, bio, i) {
2543                 struct page *page = bvec->bv_page;
2544                 struct inode *inode = page->mapping->host;
2545
2546                 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2547                          "mirror=%u\n", (u64)bio->bi_iter.bi_sector,
2548                          bio->bi_error, io_bio->mirror_num);
2549                 tree = &BTRFS_I(inode)->io_tree;
2550
2551                 /* We always issue full-page reads, but if some block
2552                  * in a page fails to read, blk_update_request() will
2553                  * advance bv_offset and adjust bv_len to compensate.
2554                  * Print a warning for nonzero offsets, and an error
2555                  * if they don't add up to a full page.  */
2556                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2557                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2558                                 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2559                                    "partial page read in btrfs with offset %u and length %u",
2560                                         bvec->bv_offset, bvec->bv_len);
2561                         else
2562                                 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2563                                    "incomplete page read in btrfs with offset %u and "
2564                                    "length %u",
2565                                         bvec->bv_offset, bvec->bv_len);
2566                 }
2567
2568                 start = page_offset(page);
2569                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2570                 len = bvec->bv_len;
2571
2572                 mirror = io_bio->mirror_num;
2573                 if (likely(uptodate && tree->ops &&
2574                            tree->ops->readpage_end_io_hook)) {
2575                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2576                                                               page, start, end,
2577                                                               mirror);
2578                         if (ret)
2579                                 uptodate = 0;
2580                         else
2581                                 clean_io_failure(inode, start, page, 0);
2582                 }
2583
2584                 if (likely(uptodate))
2585                         goto readpage_ok;
2586
2587                 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2588                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2589                         if (!ret && !bio->bi_error)
2590                                 uptodate = 1;
2591                 } else {
2592                         /*
2593                          * The generic bio_readpage_error handles errors the
2594                          * following way: If possible, new read requests are
2595                          * created and submitted and will end up in
2596                          * end_bio_extent_readpage as well (if we're lucky, not
2597                          * in the !uptodate case). In that case it returns 0 and
2598                          * we just go on with the next page in our bio. If it
2599                          * can't handle the error it will return -EIO and we
2600                          * remain responsible for that page.
2601                          */
2602                         ret = bio_readpage_error(bio, offset, page, start, end,
2603                                                  mirror);
2604                         if (ret == 0) {
2605                                 uptodate = !bio->bi_error;
2606                                 offset += len;
2607                                 continue;
2608                         }
2609                 }
2610 readpage_ok:
2611                 if (likely(uptodate)) {
2612                         loff_t i_size = i_size_read(inode);
2613                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2614                         unsigned off;
2615
2616                         /* Zero out the end if this page straddles i_size */
2617                         off = i_size & (PAGE_SIZE-1);
2618                         if (page->index == end_index && off)
2619                                 zero_user_segment(page, off, PAGE_SIZE);
2620                         SetPageUptodate(page);
2621                 } else {
2622                         ClearPageUptodate(page);
2623                         SetPageError(page);
2624                 }
2625                 unlock_page(page);
2626                 offset += len;
2627
2628                 if (unlikely(!uptodate)) {
2629                         if (extent_len) {
2630                                 endio_readpage_release_extent(tree,
2631                                                               extent_start,
2632                                                               extent_len, 1);
2633                                 extent_start = 0;
2634                                 extent_len = 0;
2635                         }
2636                         endio_readpage_release_extent(tree, start,
2637                                                       end - start + 1, 0);
2638                 } else if (!extent_len) {
2639                         extent_start = start;
2640                         extent_len = end + 1 - start;
2641                 } else if (extent_start + extent_len == start) {
2642                         extent_len += end + 1 - start;
2643                 } else {
2644                         endio_readpage_release_extent(tree, extent_start,
2645                                                       extent_len, uptodate);
2646                         extent_start = start;
2647                         extent_len = end + 1 - start;
2648                 }
2649         }
2650
2651         if (extent_len)
2652                 endio_readpage_release_extent(tree, extent_start, extent_len,
2653                                               uptodate);
2654         if (io_bio->end_io)
2655                 io_bio->end_io(io_bio, bio->bi_error);
2656         bio_put(bio);
2657 }
2658
2659 /*
2660  * this allocates from the btrfs_bioset.  We're returning a bio right now
2661  * but you can call btrfs_io_bio for the appropriate container_of magic
2662  */
2663 struct bio *
2664 btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2665                 gfp_t gfp_flags)
2666 {
2667         struct btrfs_io_bio *btrfs_bio;
2668         struct bio *bio;
2669
2670         bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2671
2672         if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2673                 while (!bio && (nr_vecs /= 2)) {
2674                         bio = bio_alloc_bioset(gfp_flags,
2675                                                nr_vecs, btrfs_bioset);
2676                 }
2677         }
2678
2679         if (bio) {
2680                 bio->bi_bdev = bdev;
2681                 bio->bi_iter.bi_sector = first_sector;
2682                 btrfs_bio = btrfs_io_bio(bio);
2683                 btrfs_bio->csum = NULL;
2684                 btrfs_bio->csum_allocated = NULL;
2685                 btrfs_bio->end_io = NULL;
2686         }
2687         return bio;
2688 }
2689
2690 struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2691 {
2692         struct btrfs_io_bio *btrfs_bio;
2693         struct bio *new;
2694
2695         new = bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2696         if (new) {
2697                 btrfs_bio = btrfs_io_bio(new);
2698                 btrfs_bio->csum = NULL;
2699                 btrfs_bio->csum_allocated = NULL;
2700                 btrfs_bio->end_io = NULL;
2701         }
2702         return new;
2703 }
2704
2705 /* this also allocates from the btrfs_bioset */
2706 struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2707 {
2708         struct btrfs_io_bio *btrfs_bio;
2709         struct bio *bio;
2710
2711         bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2712         if (bio) {
2713                 btrfs_bio = btrfs_io_bio(bio);
2714                 btrfs_bio->csum = NULL;
2715                 btrfs_bio->csum_allocated = NULL;
2716                 btrfs_bio->end_io = NULL;
2717         }
2718         return bio;
2719 }
2720
2721
2722 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2723                                        unsigned long bio_flags)
2724 {
2725         int 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 && tree->ops->submit_bio_hook)
2737                 ret = tree->ops->submit_bio_hook(page->mapping->host, bio,
2738                                            mirror_num, bio_flags, start);
2739         else
2740                 btrfsic_submit_bio(bio);
2741
2742         bio_put(bio);
2743         return 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 && tree->ops->merge_bio_hook)
2752                 ret = tree->ops->merge_bio_hook(page, offset, size, bio,
2753                                                 bio_flags);
2754         return ret;
2755
2756 }
2757
2758 static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree,
2759                               struct writeback_control *wbc,
2760                               struct page *page, sector_t sector,
2761                               size_t size, unsigned long offset,
2762                               struct block_device *bdev,
2763                               struct bio **bio_ret,
2764                               unsigned long max_pages,
2765                               bio_end_io_t end_io_func,
2766                               int mirror_num,
2767                               unsigned long prev_bio_flags,
2768                               unsigned long bio_flags,
2769                               bool force_bio_submit)
2770 {
2771         int ret = 0;
2772         struct bio *bio;
2773         int contig = 0;
2774         int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2775         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2776
2777         if (bio_ret && *bio_ret) {
2778                 bio = *bio_ret;
2779                 if (old_compressed)
2780                         contig = bio->bi_iter.bi_sector == sector;
2781                 else
2782                         contig = bio_end_sector(bio) == sector;
2783
2784                 if (prev_bio_flags != bio_flags || !contig ||
2785                     force_bio_submit ||
2786                     merge_bio(tree, page, offset, page_size, bio, bio_flags) ||
2787                     bio_add_page(bio, page, page_size, offset) < page_size) {
2788                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2789                         if (ret < 0) {
2790                                 *bio_ret = NULL;
2791                                 return ret;
2792                         }
2793                         bio = NULL;
2794                 } else {
2795                         if (wbc)
2796                                 wbc_account_io(wbc, page, page_size);
2797                         return 0;
2798                 }
2799         }
2800
2801         bio = btrfs_bio_alloc(bdev, sector, BIO_MAX_PAGES,
2802                         GFP_NOFS | __GFP_HIGH);
2803         if (!bio)
2804                 return -ENOMEM;
2805
2806         bio_add_page(bio, page, page_size, offset);
2807         bio->bi_end_io = end_io_func;
2808         bio->bi_private = tree;
2809         bio_set_op_attrs(bio, op, op_flags);
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                         atomic_inc(&em->refs);
2856                         return em;
2857                 }
2858
2859                 free_extent_map(em);
2860                 *em_cached = NULL;
2861         }
2862
2863         em = get_extent(inode, page, pg_offset, start, len, 0);
2864         if (em_cached && !IS_ERR_OR_NULL(em)) {
2865                 BUG_ON(*em_cached);
2866                 atomic_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, 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         sector_t sector;
2896         struct extent_map *em;
2897         struct block_device *bdev;
2898         int ret = 0;
2899         int nr = 0;
2900         size_t pg_offset = 0;
2901         size_t iosize;
2902         size_t disk_io_size;
2903         size_t blocksize = inode->i_sb->s_blocksize;
2904         unsigned long this_bio_flag = 0;
2905
2906         set_page_extent_mapped(page);
2907
2908         end = page_end;
2909         if (!PageUptodate(page)) {
2910                 if (cleancache_get_page(page) == 0) {
2911                         BUG_ON(blocksize != PAGE_SIZE);
2912                         unlock_extent(tree, start, end);
2913                         goto out;
2914                 }
2915         }
2916
2917         if (page->index == last_byte >> PAGE_SHIFT) {
2918                 char *userpage;
2919                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2920
2921                 if (zero_offset) {
2922                         iosize = PAGE_SIZE - zero_offset;
2923                         userpage = kmap_atomic(page);
2924                         memset(userpage + zero_offset, 0, iosize);
2925                         flush_dcache_page(page);
2926                         kunmap_atomic(userpage);
2927                 }
2928         }
2929         while (cur <= end) {
2930                 unsigned long pnr = (last_byte >> PAGE_SHIFT) + 1;
2931                 bool force_bio_submit = false;
2932
2933                 if (cur >= last_byte) {
2934                         char *userpage;
2935                         struct extent_state *cached = NULL;
2936
2937                         iosize = PAGE_SIZE - pg_offset;
2938                         userpage = kmap_atomic(page);
2939                         memset(userpage + pg_offset, 0, iosize);
2940                         flush_dcache_page(page);
2941                         kunmap_atomic(userpage);
2942                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2943                                             &cached, GFP_NOFS);
2944                         unlock_extent_cached(tree, cur,
2945                                              cur + iosize - 1,
2946                                              &cached, GFP_NOFS);
2947                         break;
2948                 }
2949                 em = __get_extent_map(inode, page, pg_offset, cur,
2950                                       end - cur + 1, get_extent, em_cached);
2951                 if (IS_ERR_OR_NULL(em)) {
2952                         SetPageError(page);
2953                         unlock_extent(tree, cur, end);
2954                         break;
2955                 }
2956                 extent_offset = cur - em->start;
2957                 BUG_ON(extent_map_end(em) <= cur);
2958                 BUG_ON(end < cur);
2959
2960                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2961                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2962                         extent_set_compress_type(&this_bio_flag,
2963                                                  em->compress_type);
2964                 }
2965
2966                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2967                 cur_end = min(extent_map_end(em) - 1, end);
2968                 iosize = ALIGN(iosize, blocksize);
2969                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2970                         disk_io_size = em->block_len;
2971                         sector = em->block_start >> 9;
2972                 } else {
2973                         sector = (em->block_start + extent_offset) >> 9;
2974                         disk_io_size = iosize;
2975                 }
2976                 bdev = em->bdev;
2977                 block_start = em->block_start;
2978                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2979                         block_start = EXTENT_MAP_HOLE;
2980
2981                 /*
2982                  * If we have a file range that points to a compressed extent
2983                  * and it's followed by a consecutive file range that points to
2984                  * to the same compressed extent (possibly with a different
2985                  * offset and/or length, so it either points to the whole extent
2986                  * or only part of it), we must make sure we do not submit a
2987                  * single bio to populate the pages for the 2 ranges because
2988                  * this makes the compressed extent read zero out the pages
2989                  * belonging to the 2nd range. Imagine the following scenario:
2990                  *
2991                  *  File layout
2992                  *  [0 - 8K]                     [8K - 24K]
2993                  *    |                               |
2994                  *    |                               |
2995                  * points to extent X,         points to extent X,
2996                  * offset 4K, length of 8K     offset 0, length 16K
2997                  *
2998                  * [extent X, compressed length = 4K uncompressed length = 16K]
2999                  *
3000                  * If the bio to read the compressed extent covers both ranges,
3001                  * it will decompress extent X into the pages belonging to the
3002                  * first range and then it will stop, zeroing out the remaining
3003                  * pages that belong to the other range that points to extent X.
3004                  * So here we make sure we submit 2 bios, one for the first
3005                  * range and another one for the third range. Both will target
3006                  * the same physical extent from disk, but we can't currently
3007                  * make the compressed bio endio callback populate the pages
3008                  * for both ranges because each compressed bio is tightly
3009                  * coupled with a single extent map, and each range can have
3010                  * an extent map with a different offset value relative to the
3011                  * uncompressed data of our extent and different lengths. This
3012                  * is a corner case so we prioritize correctness over
3013                  * non-optimal behavior (submitting 2 bios for the same extent).
3014                  */
3015                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3016                     prev_em_start && *prev_em_start != (u64)-1 &&
3017                     *prev_em_start != em->orig_start)
3018                         force_bio_submit = true;
3019
3020                 if (prev_em_start)
3021                         *prev_em_start = em->orig_start;
3022
3023                 free_extent_map(em);
3024                 em = NULL;
3025
3026                 /* we've found a hole, just zero and go on */
3027                 if (block_start == EXTENT_MAP_HOLE) {
3028                         char *userpage;
3029                         struct extent_state *cached = NULL;
3030
3031                         userpage = kmap_atomic(page);
3032                         memset(userpage + pg_offset, 0, iosize);
3033                         flush_dcache_page(page);
3034                         kunmap_atomic(userpage);
3035
3036                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3037                                             &cached, GFP_NOFS);
3038                         unlock_extent_cached(tree, cur,
3039                                              cur + iosize - 1,
3040                                              &cached, GFP_NOFS);
3041                         cur = cur + iosize;
3042                         pg_offset += iosize;
3043                         continue;
3044                 }
3045                 /* the get_extent function already copied into the page */
3046                 if (test_range_bit(tree, cur, cur_end,
3047                                    EXTENT_UPTODATE, 1, NULL)) {
3048                         check_page_uptodate(tree, page);
3049                         unlock_extent(tree, cur, cur + iosize - 1);
3050                         cur = cur + iosize;
3051                         pg_offset += iosize;
3052                         continue;
3053                 }
3054                 /* we have an inline extent but it didn't get marked up
3055                  * to date.  Error out
3056                  */
3057                 if (block_start == EXTENT_MAP_INLINE) {
3058                         SetPageError(page);
3059                         unlock_extent(tree, cur, cur + iosize - 1);
3060                         cur = cur + iosize;
3061                         pg_offset += iosize;
3062                         continue;
3063                 }
3064
3065                 pnr -= page->index;
3066                 ret = submit_extent_page(REQ_OP_READ, read_flags, tree, NULL,
3067                                          page, sector, disk_io_size, pg_offset,
3068                                          bdev, bio, pnr,
3069                                          end_bio_extent_readpage, mirror_num,
3070                                          *bio_flags,
3071                                          this_bio_flag,
3072                                          force_bio_submit);
3073                 if (!ret) {
3074                         nr++;
3075                         *bio_flags = this_bio_flag;
3076                 } else {
3077                         SetPageError(page);
3078                         unlock_extent(tree, cur, cur + iosize - 1);
3079                         goto out;
3080                 }
3081                 cur = cur + iosize;
3082                 pg_offset += iosize;
3083         }
3084 out:
3085         if (!nr) {
3086                 if (!PageError(page))
3087                         SetPageUptodate(page);
3088                 unlock_page(page);
3089         }
3090         return ret;
3091 }
3092
3093 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3094                                              struct page *pages[], int nr_pages,
3095                                              u64 start, u64 end,
3096                                              get_extent_t *get_extent,
3097                                              struct extent_map **em_cached,
3098                                              struct bio **bio, int mirror_num,
3099                                              unsigned long *bio_flags,
3100                                              u64 *prev_em_start)
3101 {
3102         struct inode *inode;
3103         struct btrfs_ordered_extent *ordered;
3104         int index;
3105
3106         inode = pages[0]->mapping->host;
3107         while (1) {
3108                 lock_extent(tree, start, end);
3109                 ordered = btrfs_lookup_ordered_range(inode, start,
3110                                                      end - start + 1);
3111                 if (!ordered)
3112                         break;
3113                 unlock_extent(tree, start, end);
3114                 btrfs_start_ordered_extent(inode, ordered, 1);
3115                 btrfs_put_ordered_extent(ordered);
3116         }
3117
3118         for (index = 0; index < nr_pages; index++) {
3119                 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
3120                               mirror_num, bio_flags, 0, prev_em_start);
3121                 put_page(pages[index]);
3122         }
3123 }
3124
3125 static void __extent_readpages(struct extent_io_tree *tree,
3126                                struct page *pages[],
3127                                int nr_pages, get_extent_t *get_extent,