btrfs: document more parameters of submit_extent_page
[linux-block.git] / fs / btrfs / extent_io.c
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/bitops.h>
3 #include <linux/slab.h>
4 #include <linux/bio.h>
5 #include <linux/mm.h>
6 #include <linux/pagemap.h>
7 #include <linux/page-flags.h>
8 #include <linux/spinlock.h>
9 #include <linux/blkdev.h>
10 #include <linux/swap.h>
11 #include <linux/writeback.h>
12 #include <linux/pagevec.h>
13 #include <linux/prefetch.h>
14 #include <linux/cleancache.h>
15 #include "extent_io.h"
16 #include "extent_map.h"
17 #include "ctree.h"
18 #include "btrfs_inode.h"
19 #include "volumes.h"
20 #include "check-integrity.h"
21 #include "locking.h"
22 #include "rcu-string.h"
23 #include "backref.h"
24 #include "disk-io.h"
25
26 static struct kmem_cache *extent_state_cache;
27 static struct kmem_cache *extent_buffer_cache;
28 static struct bio_set *btrfs_bioset;
29
30 static inline bool extent_state_in_tree(const struct extent_state *state)
31 {
32         return !RB_EMPTY_NODE(&state->rb_node);
33 }
34
35 #ifdef CONFIG_BTRFS_DEBUG
36 static LIST_HEAD(buffers);
37 static LIST_HEAD(states);
38
39 static DEFINE_SPINLOCK(leak_lock);
40
41 static inline
42 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
43 {
44         unsigned long flags;
45
46         spin_lock_irqsave(&leak_lock, flags);
47         list_add(new, head);
48         spin_unlock_irqrestore(&leak_lock, flags);
49 }
50
51 static inline
52 void btrfs_leak_debug_del(struct list_head *entry)
53 {
54         unsigned long flags;
55
56         spin_lock_irqsave(&leak_lock, flags);
57         list_del(entry);
58         spin_unlock_irqrestore(&leak_lock, flags);
59 }
60
61 static inline
62 void btrfs_leak_debug_check(void)
63 {
64         struct extent_state *state;
65         struct extent_buffer *eb;
66
67         while (!list_empty(&states)) {
68                 state = list_entry(states.next, struct extent_state, leak_list);
69                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
70                        state->start, state->end, state->state,
71                        extent_state_in_tree(state),
72                        refcount_read(&state->refs));
73                 list_del(&state->leak_list);
74                 kmem_cache_free(extent_state_cache, state);
75         }
76
77         while (!list_empty(&buffers)) {
78                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
79                 pr_err("BTRFS: buffer leak start %llu len %lu refs %d bflags %lu\n",
80                        eb->start, eb->len, atomic_read(&eb->refs), eb->bflags);
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         if (tree->ops && tree->ops->check_extent_io_range)
92                 tree->ops->check_extent_io_range(tree->private_data, caller,
93                                                  start, end);
94 }
95 #else
96 #define btrfs_leak_debug_add(new, head) do {} while (0)
97 #define btrfs_leak_debug_del(entry)     do {} while (0)
98 #define btrfs_leak_debug_check()        do {} while (0)
99 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
100 #endif
101
102 #define BUFFER_LRU_MAX 64
103
104 struct tree_entry {
105         u64 start;
106         u64 end;
107         struct rb_node rb_node;
108 };
109
110 struct extent_page_data {
111         struct bio *bio;
112         struct extent_io_tree *tree;
113         /* tells writepage not to lock the state bits for this range
114          * it still does the unlocking
115          */
116         unsigned int extent_locked:1;
117
118         /* tells the submit_bio code to use REQ_SYNC */
119         unsigned int sync_io:1;
120 };
121
122 static void add_extent_changeset(struct extent_state *state, unsigned bits,
123                                  struct extent_changeset *changeset,
124                                  int set)
125 {
126         int ret;
127
128         if (!changeset)
129                 return;
130         if (set && (state->state & bits) == bits)
131                 return;
132         if (!set && (state->state & bits) == 0)
133                 return;
134         changeset->bytes_changed += state->end - state->start + 1;
135         ret = ulist_add(&changeset->range_changed, state->start, state->end,
136                         GFP_ATOMIC);
137         /* ENOMEM */
138         BUG_ON(ret < 0);
139 }
140
141 static void flush_write_bio(struct extent_page_data *epd);
142
143 static inline struct btrfs_fs_info *
144 tree_fs_info(struct extent_io_tree *tree)
145 {
146         if (tree->ops)
147                 return tree->ops->tree_fs_info(tree->private_data);
148         return NULL;
149 }
150
151 int __init extent_io_init(void)
152 {
153         extent_state_cache = kmem_cache_create("btrfs_extent_state",
154                         sizeof(struct extent_state), 0,
155                         SLAB_MEM_SPREAD, NULL);
156         if (!extent_state_cache)
157                 return -ENOMEM;
158
159         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
160                         sizeof(struct extent_buffer), 0,
161                         SLAB_MEM_SPREAD, NULL);
162         if (!extent_buffer_cache)
163                 goto free_state_cache;
164
165         btrfs_bioset = bioset_create(BIO_POOL_SIZE,
166                                      offsetof(struct btrfs_io_bio, bio),
167                                      BIOSET_NEED_BVECS);
168         if (!btrfs_bioset)
169                 goto free_buffer_cache;
170
171         if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
172                 goto free_bioset;
173
174         return 0;
175
176 free_bioset:
177         bioset_free(btrfs_bioset);
178         btrfs_bioset = NULL;
179
180 free_buffer_cache:
181         kmem_cache_destroy(extent_buffer_cache);
182         extent_buffer_cache = NULL;
183
184 free_state_cache:
185         kmem_cache_destroy(extent_state_cache);
186         extent_state_cache = NULL;
187         return -ENOMEM;
188 }
189
190 void __cold extent_io_exit(void)
191 {
192         btrfs_leak_debug_check();
193
194         /*
195          * Make sure all delayed rcu free are flushed before we
196          * destroy caches.
197          */
198         rcu_barrier();
199         kmem_cache_destroy(extent_state_cache);
200         kmem_cache_destroy(extent_buffer_cache);
201         if (btrfs_bioset)
202                 bioset_free(btrfs_bioset);
203 }
204
205 void extent_io_tree_init(struct extent_io_tree *tree,
206                          void *private_data)
207 {
208         tree->state = RB_ROOT;
209         tree->ops = NULL;
210         tree->dirty_bytes = 0;
211         spin_lock_init(&tree->lock);
212         tree->private_data = private_data;
213 }
214
215 static struct extent_state *alloc_extent_state(gfp_t mask)
216 {
217         struct extent_state *state;
218
219         /*
220          * The given mask might be not appropriate for the slab allocator,
221          * drop the unsupported bits
222          */
223         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
224         state = kmem_cache_alloc(extent_state_cache, mask);
225         if (!state)
226                 return state;
227         state->state = 0;
228         state->failrec = NULL;
229         RB_CLEAR_NODE(&state->rb_node);
230         btrfs_leak_debug_add(&state->leak_list, &states);
231         refcount_set(&state->refs, 1);
232         init_waitqueue_head(&state->wq);
233         trace_alloc_extent_state(state, mask, _RET_IP_);
234         return state;
235 }
236
237 void free_extent_state(struct extent_state *state)
238 {
239         if (!state)
240                 return;
241         if (refcount_dec_and_test(&state->refs)) {
242                 WARN_ON(extent_state_in_tree(state));
243                 btrfs_leak_debug_del(&state->leak_list);
244                 trace_free_extent_state(state, _RET_IP_);
245                 kmem_cache_free(extent_state_cache, state);
246         }
247 }
248
249 static struct rb_node *tree_insert(struct rb_root *root,
250                                    struct rb_node *search_start,
251                                    u64 offset,
252                                    struct rb_node *node,
253                                    struct rb_node ***p_in,
254                                    struct rb_node **parent_in)
255 {
256         struct rb_node **p;
257         struct rb_node *parent = NULL;
258         struct tree_entry *entry;
259
260         if (p_in && parent_in) {
261                 p = *p_in;
262                 parent = *parent_in;
263                 goto do_insert;
264         }
265
266         p = search_start ? &search_start : &root->rb_node;
267         while (*p) {
268                 parent = *p;
269                 entry = rb_entry(parent, struct tree_entry, rb_node);
270
271                 if (offset < entry->start)
272                         p = &(*p)->rb_left;
273                 else if (offset > entry->end)
274                         p = &(*p)->rb_right;
275                 else
276                         return parent;
277         }
278
279 do_insert:
280         rb_link_node(node, parent, p);
281         rb_insert_color(node, root);
282         return NULL;
283 }
284
285 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
286                                       struct rb_node **prev_ret,
287                                       struct rb_node **next_ret,
288                                       struct rb_node ***p_ret,
289                                       struct rb_node **parent_ret)
290 {
291         struct rb_root *root = &tree->state;
292         struct rb_node **n = &root->rb_node;
293         struct rb_node *prev = NULL;
294         struct rb_node *orig_prev = NULL;
295         struct tree_entry *entry;
296         struct tree_entry *prev_entry = NULL;
297
298         while (*n) {
299                 prev = *n;
300                 entry = rb_entry(prev, struct tree_entry, rb_node);
301                 prev_entry = entry;
302
303                 if (offset < entry->start)
304                         n = &(*n)->rb_left;
305                 else if (offset > entry->end)
306                         n = &(*n)->rb_right;
307                 else
308                         return *n;
309         }
310
311         if (p_ret)
312                 *p_ret = n;
313         if (parent_ret)
314                 *parent_ret = prev;
315
316         if (prev_ret) {
317                 orig_prev = prev;
318                 while (prev && offset > prev_entry->end) {
319                         prev = rb_next(prev);
320                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
321                 }
322                 *prev_ret = prev;
323                 prev = orig_prev;
324         }
325
326         if (next_ret) {
327                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
328                 while (prev && offset < prev_entry->start) {
329                         prev = rb_prev(prev);
330                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
331                 }
332                 *next_ret = prev;
333         }
334         return NULL;
335 }
336
337 static inline struct rb_node *
338 tree_search_for_insert(struct extent_io_tree *tree,
339                        u64 offset,
340                        struct rb_node ***p_ret,
341                        struct rb_node **parent_ret)
342 {
343         struct rb_node *prev = NULL;
344         struct rb_node *ret;
345
346         ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
347         if (!ret)
348                 return prev;
349         return ret;
350 }
351
352 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
353                                           u64 offset)
354 {
355         return tree_search_for_insert(tree, offset, NULL, NULL);
356 }
357
358 static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
359                      struct extent_state *other)
360 {
361         if (tree->ops && tree->ops->merge_extent_hook)
362                 tree->ops->merge_extent_hook(tree->private_data, new, other);
363 }
364
365 /*
366  * utility function to look for merge candidates inside a given range.
367  * Any extents with matching state are merged together into a single
368  * extent in the tree.  Extents with EXTENT_IO in their state field
369  * are not merged because the end_io handlers need to be able to do
370  * operations on them without sleeping (or doing allocations/splits).
371  *
372  * This should be called with the tree lock held.
373  */
374 static void merge_state(struct extent_io_tree *tree,
375                         struct extent_state *state)
376 {
377         struct extent_state *other;
378         struct rb_node *other_node;
379
380         if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
381                 return;
382
383         other_node = rb_prev(&state->rb_node);
384         if (other_node) {
385                 other = rb_entry(other_node, struct extent_state, rb_node);
386                 if (other->end == state->start - 1 &&
387                     other->state == state->state) {
388                         merge_cb(tree, state, other);
389                         state->start = other->start;
390                         rb_erase(&other->rb_node, &tree->state);
391                         RB_CLEAR_NODE(&other->rb_node);
392                         free_extent_state(other);
393                 }
394         }
395         other_node = rb_next(&state->rb_node);
396         if (other_node) {
397                 other = rb_entry(other_node, struct extent_state, rb_node);
398                 if (other->start == state->end + 1 &&
399                     other->state == state->state) {
400                         merge_cb(tree, state, other);
401                         state->end = other->end;
402                         rb_erase(&other->rb_node, &tree->state);
403                         RB_CLEAR_NODE(&other->rb_node);
404                         free_extent_state(other);
405                 }
406         }
407 }
408
409 static void set_state_cb(struct extent_io_tree *tree,
410                          struct extent_state *state, unsigned *bits)
411 {
412         if (tree->ops && tree->ops->set_bit_hook)
413                 tree->ops->set_bit_hook(tree->private_data, state, bits);
414 }
415
416 static void clear_state_cb(struct extent_io_tree *tree,
417                            struct extent_state *state, unsigned *bits)
418 {
419         if (tree->ops && tree->ops->clear_bit_hook)
420                 tree->ops->clear_bit_hook(tree->private_data, state, bits);
421 }
422
423 static void set_state_bits(struct extent_io_tree *tree,
424                            struct extent_state *state, unsigned *bits,
425                            struct extent_changeset *changeset);
426
427 /*
428  * insert an extent_state struct into the tree.  'bits' are set on the
429  * struct before it is inserted.
430  *
431  * This may return -EEXIST if the extent is already there, in which case the
432  * state struct is freed.
433  *
434  * The tree lock is not taken internally.  This is a utility function and
435  * probably isn't what you want to call (see set/clear_extent_bit).
436  */
437 static int insert_state(struct extent_io_tree *tree,
438                         struct extent_state *state, u64 start, u64 end,
439                         struct rb_node ***p,
440                         struct rb_node **parent,
441                         unsigned *bits, struct extent_changeset *changeset)
442 {
443         struct rb_node *node;
444
445         if (end < start)
446                 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
447                        end, start);
448         state->start = start;
449         state->end = end;
450
451         set_state_bits(tree, state, bits, changeset);
452
453         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
454         if (node) {
455                 struct extent_state *found;
456                 found = rb_entry(node, struct extent_state, rb_node);
457                 pr_err("BTRFS: found node %llu %llu on insert of %llu %llu\n",
458                        found->start, found->end, start, end);
459                 return -EEXIST;
460         }
461         merge_state(tree, state);
462         return 0;
463 }
464
465 static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
466                      u64 split)
467 {
468         if (tree->ops && tree->ops->split_extent_hook)
469                 tree->ops->split_extent_hook(tree->private_data, orig, split);
470 }
471
472 /*
473  * split a given extent state struct in two, inserting the preallocated
474  * struct 'prealloc' as the newly created second half.  'split' indicates an
475  * offset inside 'orig' where it should be split.
476  *
477  * Before calling,
478  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
479  * are two extent state structs in the tree:
480  * prealloc: [orig->start, split - 1]
481  * orig: [ split, orig->end ]
482  *
483  * The tree locks are not taken by this function. They need to be held
484  * by the caller.
485  */
486 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
487                        struct extent_state *prealloc, u64 split)
488 {
489         struct rb_node *node;
490
491         split_cb(tree, orig, split);
492
493         prealloc->start = orig->start;
494         prealloc->end = split - 1;
495         prealloc->state = orig->state;
496         orig->start = split;
497
498         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
499                            &prealloc->rb_node, NULL, NULL);
500         if (node) {
501                 free_extent_state(prealloc);
502                 return -EEXIST;
503         }
504         return 0;
505 }
506
507 static struct extent_state *next_state(struct extent_state *state)
508 {
509         struct rb_node *next = rb_next(&state->rb_node);
510         if (next)
511                 return rb_entry(next, struct extent_state, rb_node);
512         else
513                 return NULL;
514 }
515
516 /*
517  * utility function to clear some bits in an extent state struct.
518  * it will optionally wake up any one waiting on this state (wake == 1).
519  *
520  * If no bits are set on the state struct after clearing things, the
521  * struct is freed and removed from the tree
522  */
523 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
524                                             struct extent_state *state,
525                                             unsigned *bits, int wake,
526                                             struct extent_changeset *changeset)
527 {
528         struct extent_state *next;
529         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
530
531         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
532                 u64 range = state->end - state->start + 1;
533                 WARN_ON(range > tree->dirty_bytes);
534                 tree->dirty_bytes -= range;
535         }
536         clear_state_cb(tree, state, bits);
537         add_extent_changeset(state, bits_to_clear, changeset, 0);
538         state->state &= ~bits_to_clear;
539         if (wake)
540                 wake_up(&state->wq);
541         if (state->state == 0) {
542                 next = next_state(state);
543                 if (extent_state_in_tree(state)) {
544                         rb_erase(&state->rb_node, &tree->state);
545                         RB_CLEAR_NODE(&state->rb_node);
546                         free_extent_state(state);
547                 } else {
548                         WARN_ON(1);
549                 }
550         } else {
551                 merge_state(tree, state);
552                 next = next_state(state);
553         }
554         return next;
555 }
556
557 static struct extent_state *
558 alloc_extent_state_atomic(struct extent_state *prealloc)
559 {
560         if (!prealloc)
561                 prealloc = alloc_extent_state(GFP_ATOMIC);
562
563         return prealloc;
564 }
565
566 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
567 {
568         btrfs_panic(tree_fs_info(tree), err,
569                     "Locking error: Extent tree was modified by another thread while locked.");
570 }
571
572 /*
573  * clear some bits on a range in the tree.  This may require splitting
574  * or inserting elements in the tree, so the gfp mask is used to
575  * indicate which allocations or sleeping are allowed.
576  *
577  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
578  * the given range from the tree regardless of state (ie for truncate).
579  *
580  * the range [start, end] is inclusive.
581  *
582  * This takes the tree lock, and returns 0 on success and < 0 on error.
583  */
584 int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
585                               unsigned bits, int wake, int delete,
586                               struct extent_state **cached_state,
587                               gfp_t mask, struct extent_changeset *changeset)
588 {
589         struct extent_state *state;
590         struct extent_state *cached;
591         struct extent_state *prealloc = NULL;
592         struct rb_node *node;
593         u64 last_end;
594         int err;
595         int clear = 0;
596
597         btrfs_debug_check_extent_io_range(tree, start, end);
598
599         if (bits & EXTENT_DELALLOC)
600                 bits |= EXTENT_NORESERVE;
601
602         if (delete)
603                 bits |= ~EXTENT_CTLBITS;
604         bits |= EXTENT_FIRST_DELALLOC;
605
606         if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
607                 clear = 1;
608 again:
609         if (!prealloc && gfpflags_allow_blocking(mask)) {
610                 /*
611                  * Don't care for allocation failure here because we might end
612                  * up not needing the pre-allocated extent state at all, which
613                  * is the case if we only have in the tree extent states that
614                  * cover our input range and don't cover too any other range.
615                  * If we end up needing a new extent state we allocate it later.
616                  */
617                 prealloc = alloc_extent_state(mask);
618         }
619
620         spin_lock(&tree->lock);
621         if (cached_state) {
622                 cached = *cached_state;
623
624                 if (clear) {
625                         *cached_state = NULL;
626                         cached_state = NULL;
627                 }
628
629                 if (cached && extent_state_in_tree(cached) &&
630                     cached->start <= start && cached->end > start) {
631                         if (clear)
632                                 refcount_dec(&cached->refs);
633                         state = cached;
634                         goto hit_next;
635                 }
636                 if (clear)
637                         free_extent_state(cached);
638         }
639         /*
640          * this search will find the extents that end after
641          * our range starts
642          */
643         node = tree_search(tree, start);
644         if (!node)
645                 goto out;
646         state = rb_entry(node, struct extent_state, rb_node);
647 hit_next:
648         if (state->start > end)
649                 goto out;
650         WARN_ON(state->end < start);
651         last_end = state->end;
652
653         /* the state doesn't have the wanted bits, go ahead */
654         if (!(state->state & bits)) {
655                 state = next_state(state);
656                 goto next;
657         }
658
659         /*
660          *     | ---- desired range ---- |
661          *  | state | or
662          *  | ------------- state -------------- |
663          *
664          * We need to split the extent we found, and may flip
665          * bits on second half.
666          *
667          * If the extent we found extends past our range, we
668          * just split and search again.  It'll get split again
669          * the next time though.
670          *
671          * If the extent we found is inside our range, we clear
672          * the desired bit on it.
673          */
674
675         if (state->start < start) {
676                 prealloc = alloc_extent_state_atomic(prealloc);
677                 BUG_ON(!prealloc);
678                 err = split_state(tree, state, prealloc, start);
679                 if (err)
680                         extent_io_tree_panic(tree, err);
681
682                 prealloc = NULL;
683                 if (err)
684                         goto out;
685                 if (state->end <= end) {
686                         state = clear_state_bit(tree, state, &bits, wake,
687                                                 changeset);
688                         goto next;
689                 }
690                 goto search_again;
691         }
692         /*
693          * | ---- desired range ---- |
694          *                        | state |
695          * We need to split the extent, and clear the bit
696          * on the first half
697          */
698         if (state->start <= end && state->end > end) {
699                 prealloc = alloc_extent_state_atomic(prealloc);
700                 BUG_ON(!prealloc);
701                 err = split_state(tree, state, prealloc, end + 1);
702                 if (err)
703                         extent_io_tree_panic(tree, err);
704
705                 if (wake)
706                         wake_up(&state->wq);
707
708                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
709
710                 prealloc = NULL;
711                 goto out;
712         }
713
714         state = clear_state_bit(tree, state, &bits, wake, changeset);
715 next:
716         if (last_end == (u64)-1)
717                 goto out;
718         start = last_end + 1;
719         if (start <= end && state && !need_resched())
720                 goto hit_next;
721
722 search_again:
723         if (start > end)
724                 goto out;
725         spin_unlock(&tree->lock);
726         if (gfpflags_allow_blocking(mask))
727                 cond_resched();
728         goto again;
729
730 out:
731         spin_unlock(&tree->lock);
732         if (prealloc)
733                 free_extent_state(prealloc);
734
735         return 0;
736
737 }
738
739 static void wait_on_state(struct extent_io_tree *tree,
740                           struct extent_state *state)
741                 __releases(tree->lock)
742                 __acquires(tree->lock)
743 {
744         DEFINE_WAIT(wait);
745         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
746         spin_unlock(&tree->lock);
747         schedule();
748         spin_lock(&tree->lock);
749         finish_wait(&state->wq, &wait);
750 }
751
752 /*
753  * waits for one or more bits to clear on a range in the state tree.
754  * The range [start, end] is inclusive.
755  * The tree lock is taken by this function
756  */
757 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
758                             unsigned long bits)
759 {
760         struct extent_state *state;
761         struct rb_node *node;
762
763         btrfs_debug_check_extent_io_range(tree, start, end);
764
765         spin_lock(&tree->lock);
766 again:
767         while (1) {
768                 /*
769                  * this search will find all the extents that end after
770                  * our range starts
771                  */
772                 node = tree_search(tree, start);
773 process_node:
774                 if (!node)
775                         break;
776
777                 state = rb_entry(node, struct extent_state, rb_node);
778
779                 if (state->start > end)
780                         goto out;
781
782                 if (state->state & bits) {
783                         start = state->start;
784                         refcount_inc(&state->refs);
785                         wait_on_state(tree, state);
786                         free_extent_state(state);
787                         goto again;
788                 }
789                 start = state->end + 1;
790
791                 if (start > end)
792                         break;
793
794                 if (!cond_resched_lock(&tree->lock)) {
795                         node = rb_next(node);
796                         goto process_node;
797                 }
798         }
799 out:
800         spin_unlock(&tree->lock);
801 }
802
803 static void set_state_bits(struct extent_io_tree *tree,
804                            struct extent_state *state,
805                            unsigned *bits, struct extent_changeset *changeset)
806 {
807         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
808
809         set_state_cb(tree, state, bits);
810         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
811                 u64 range = state->end - state->start + 1;
812                 tree->dirty_bytes += range;
813         }
814         add_extent_changeset(state, bits_to_set, changeset, 1);
815         state->state |= bits_to_set;
816 }
817
818 static void cache_state_if_flags(struct extent_state *state,
819                                  struct extent_state **cached_ptr,
820                                  unsigned flags)
821 {
822         if (cached_ptr && !(*cached_ptr)) {
823                 if (!flags || (state->state & flags)) {
824                         *cached_ptr = state;
825                         refcount_inc(&state->refs);
826                 }
827         }
828 }
829
830 static void cache_state(struct extent_state *state,
831                         struct extent_state **cached_ptr)
832 {
833         return cache_state_if_flags(state, cached_ptr,
834                                     EXTENT_IOBITS | EXTENT_BOUNDARY);
835 }
836
837 /*
838  * set some bits on a range in the tree.  This may require allocations or
839  * sleeping, so the gfp mask is used to indicate what is allowed.
840  *
841  * If any of the exclusive bits are set, this will fail with -EEXIST if some
842  * part of the range already has the desired bits set.  The start of the
843  * existing range is returned in failed_start in this case.
844  *
845  * [start, end] is inclusive This takes the tree lock.
846  */
847
848 static int __must_check
849 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
850                  unsigned bits, unsigned exclusive_bits,
851                  u64 *failed_start, struct extent_state **cached_state,
852                  gfp_t mask, struct extent_changeset *changeset)
853 {
854         struct extent_state *state;
855         struct extent_state *prealloc = NULL;
856         struct rb_node *node;
857         struct rb_node **p;
858         struct rb_node *parent;
859         int err = 0;
860         u64 last_start;
861         u64 last_end;
862
863         btrfs_debug_check_extent_io_range(tree, start, end);
864
865         bits |= EXTENT_FIRST_DELALLOC;
866 again:
867         if (!prealloc && gfpflags_allow_blocking(mask)) {
868                 /*
869                  * Don't care for allocation failure here because we might end
870                  * up not needing the pre-allocated extent state at all, which
871                  * is the case if we only have in the tree extent states that
872                  * cover our input range and don't cover too any other range.
873                  * If we end up needing a new extent state we allocate it later.
874                  */
875                 prealloc = alloc_extent_state(mask);
876         }
877
878         spin_lock(&tree->lock);
879         if (cached_state && *cached_state) {
880                 state = *cached_state;
881                 if (state->start <= start && state->end > start &&
882                     extent_state_in_tree(state)) {
883                         node = &state->rb_node;
884                         goto hit_next;
885                 }
886         }
887         /*
888          * this search will find all the extents that end after
889          * our range starts.
890          */
891         node = tree_search_for_insert(tree, start, &p, &parent);
892         if (!node) {
893                 prealloc = alloc_extent_state_atomic(prealloc);
894                 BUG_ON(!prealloc);
895                 err = insert_state(tree, prealloc, start, end,
896                                    &p, &parent, &bits, changeset);
897                 if (err)
898                         extent_io_tree_panic(tree, err);
899
900                 cache_state(prealloc, cached_state);
901                 prealloc = NULL;
902                 goto out;
903         }
904         state = rb_entry(node, struct extent_state, rb_node);
905 hit_next:
906         last_start = state->start;
907         last_end = state->end;
908
909         /*
910          * | ---- desired range ---- |
911          * | state |
912          *
913          * Just lock what we found and keep going
914          */
915         if (state->start == start && state->end <= end) {
916                 if (state->state & exclusive_bits) {
917                         *failed_start = state->start;
918                         err = -EEXIST;
919                         goto out;
920                 }
921
922                 set_state_bits(tree, state, &bits, changeset);
923                 cache_state(state, cached_state);
924                 merge_state(tree, state);
925                 if (last_end == (u64)-1)
926                         goto out;
927                 start = last_end + 1;
928                 state = next_state(state);
929                 if (start < end && state && state->start == start &&
930                     !need_resched())
931                         goto hit_next;
932                 goto search_again;
933         }
934
935         /*
936          *     | ---- desired range ---- |
937          * | state |
938          *   or
939          * | ------------- state -------------- |
940          *
941          * We need to split the extent we found, and may flip bits on
942          * second half.
943          *
944          * If the extent we found extends past our
945          * range, we just split and search again.  It'll get split
946          * again the next time though.
947          *
948          * If the extent we found is inside our range, we set the
949          * desired bit on it.
950          */
951         if (state->start < start) {
952                 if (state->state & exclusive_bits) {
953                         *failed_start = start;
954                         err = -EEXIST;
955                         goto out;
956                 }
957
958                 prealloc = alloc_extent_state_atomic(prealloc);
959                 BUG_ON(!prealloc);
960                 err = split_state(tree, state, prealloc, start);
961                 if (err)
962                         extent_io_tree_panic(tree, err);
963
964                 prealloc = NULL;
965                 if (err)
966                         goto out;
967                 if (state->end <= end) {
968                         set_state_bits(tree, state, &bits, changeset);
969                         cache_state(state, cached_state);
970                         merge_state(tree, state);
971                         if (last_end == (u64)-1)
972                                 goto out;
973                         start = last_end + 1;
974                         state = next_state(state);
975                         if (start < end && state && state->start == start &&
976                             !need_resched())
977                                 goto hit_next;
978                 }
979                 goto search_again;
980         }
981         /*
982          * | ---- desired range ---- |
983          *     | state | or               | state |
984          *
985          * There's a hole, we need to insert something in it and
986          * ignore the extent we found.
987          */
988         if (state->start > start) {
989                 u64 this_end;
990                 if (end < last_start)
991                         this_end = end;
992                 else
993                         this_end = last_start - 1;
994
995                 prealloc = alloc_extent_state_atomic(prealloc);
996                 BUG_ON(!prealloc);
997
998                 /*
999                  * Avoid to free 'prealloc' if it can be merged with
1000                  * the later extent.
1001                  */
1002                 err = insert_state(tree, prealloc, start, this_end,
1003                                    NULL, NULL, &bits, changeset);
1004                 if (err)
1005                         extent_io_tree_panic(tree, err);
1006
1007                 cache_state(prealloc, cached_state);
1008                 prealloc = NULL;
1009                 start = this_end + 1;
1010                 goto search_again;
1011         }
1012         /*
1013          * | ---- desired range ---- |
1014          *                        | state |
1015          * We need to split the extent, and set the bit
1016          * on the first half
1017          */
1018         if (state->start <= end && state->end > end) {
1019                 if (state->state & exclusive_bits) {
1020                         *failed_start = start;
1021                         err = -EEXIST;
1022                         goto out;
1023                 }
1024
1025                 prealloc = alloc_extent_state_atomic(prealloc);
1026                 BUG_ON(!prealloc);
1027                 err = split_state(tree, state, prealloc, end + 1);
1028                 if (err)
1029                         extent_io_tree_panic(tree, err);
1030
1031                 set_state_bits(tree, prealloc, &bits, changeset);
1032                 cache_state(prealloc, cached_state);
1033                 merge_state(tree, prealloc);
1034                 prealloc = NULL;
1035                 goto out;
1036         }
1037
1038 search_again:
1039         if (start > end)
1040                 goto out;
1041         spin_unlock(&tree->lock);
1042         if (gfpflags_allow_blocking(mask))
1043                 cond_resched();
1044         goto again;
1045
1046 out:
1047         spin_unlock(&tree->lock);
1048         if (prealloc)
1049                 free_extent_state(prealloc);
1050
1051         return err;
1052
1053 }
1054
1055 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1056                    unsigned bits, u64 * failed_start,
1057                    struct extent_state **cached_state, gfp_t mask)
1058 {
1059         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1060                                 cached_state, mask, NULL);
1061 }
1062
1063
1064 /**
1065  * convert_extent_bit - convert all bits in a given range from one bit to
1066  *                      another
1067  * @tree:       the io tree to search
1068  * @start:      the start offset in bytes
1069  * @end:        the end offset in bytes (inclusive)
1070  * @bits:       the bits to set in this range
1071  * @clear_bits: the bits to clear in this range
1072  * @cached_state:       state that we're going to cache
1073  *
1074  * This will go through and set bits for the given range.  If any states exist
1075  * already in this range they are set with the given bit and cleared of the
1076  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1077  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1078  * boundary bits like LOCK.
1079  *
1080  * All allocations are done with GFP_NOFS.
1081  */
1082 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1083                        unsigned bits, unsigned clear_bits,
1084                        struct extent_state **cached_state)
1085 {
1086         struct extent_state *state;
1087         struct extent_state *prealloc = NULL;
1088         struct rb_node *node;
1089         struct rb_node **p;
1090         struct rb_node *parent;
1091         int err = 0;
1092         u64 last_start;
1093         u64 last_end;
1094         bool first_iteration = true;
1095
1096         btrfs_debug_check_extent_io_range(tree, start, end);
1097
1098 again:
1099         if (!prealloc) {
1100                 /*
1101                  * Best effort, don't worry if extent state allocation fails
1102                  * here for the first iteration. We might have a cached state
1103                  * that matches exactly the target range, in which case no
1104                  * extent state allocations are needed. We'll only know this
1105                  * after locking the tree.
1106                  */
1107                 prealloc = alloc_extent_state(GFP_NOFS);
1108                 if (!prealloc && !first_iteration)
1109                         return -ENOMEM;
1110         }
1111
1112         spin_lock(&tree->lock);
1113         if (cached_state && *cached_state) {
1114                 state = *cached_state;
1115                 if (state->start <= start && state->end > start &&
1116                     extent_state_in_tree(state)) {
1117                         node = &state->rb_node;
1118                         goto hit_next;
1119                 }
1120         }
1121
1122         /*
1123          * this search will find all the extents that end after
1124          * our range starts.
1125          */
1126         node = tree_search_for_insert(tree, start, &p, &parent);
1127         if (!node) {
1128                 prealloc = alloc_extent_state_atomic(prealloc);
1129                 if (!prealloc) {
1130                         err = -ENOMEM;
1131                         goto out;
1132                 }
1133                 err = insert_state(tree, prealloc, start, end,
1134                                    &p, &parent, &bits, NULL);
1135                 if (err)
1136                         extent_io_tree_panic(tree, err);
1137                 cache_state(prealloc, cached_state);
1138                 prealloc = NULL;
1139                 goto out;
1140         }
1141         state = rb_entry(node, struct extent_state, rb_node);
1142 hit_next:
1143         last_start = state->start;
1144         last_end = state->end;
1145
1146         /*
1147          * | ---- desired range ---- |
1148          * | state |
1149          *
1150          * Just lock what we found and keep going
1151          */
1152         if (state->start == start && state->end <= end) {
1153                 set_state_bits(tree, state, &bits, NULL);
1154                 cache_state(state, cached_state);
1155                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1156                 if (last_end == (u64)-1)
1157                         goto out;
1158                 start = last_end + 1;
1159                 if (start < end && state && state->start == start &&
1160                     !need_resched())
1161                         goto hit_next;
1162                 goto search_again;
1163         }
1164
1165         /*
1166          *     | ---- desired range ---- |
1167          * | state |
1168          *   or
1169          * | ------------- state -------------- |
1170          *
1171          * We need to split the extent we found, and may flip bits on
1172          * second half.
1173          *
1174          * If the extent we found extends past our
1175          * range, we just split and search again.  It'll get split
1176          * again the next time though.
1177          *
1178          * If the extent we found is inside our range, we set the
1179          * desired bit on it.
1180          */
1181         if (state->start < start) {
1182                 prealloc = alloc_extent_state_atomic(prealloc);
1183                 if (!prealloc) {
1184                         err = -ENOMEM;
1185                         goto out;
1186                 }
1187                 err = split_state(tree, state, prealloc, start);
1188                 if (err)
1189                         extent_io_tree_panic(tree, err);
1190                 prealloc = NULL;
1191                 if (err)
1192                         goto out;
1193                 if (state->end <= end) {
1194                         set_state_bits(tree, state, &bits, NULL);
1195                         cache_state(state, cached_state);
1196                         state = clear_state_bit(tree, state, &clear_bits, 0,
1197                                                 NULL);
1198                         if (last_end == (u64)-1)
1199                                 goto out;
1200                         start = last_end + 1;
1201                         if (start < end && state && state->start == start &&
1202                             !need_resched())
1203                                 goto hit_next;
1204                 }
1205                 goto search_again;
1206         }
1207         /*
1208          * | ---- desired range ---- |
1209          *     | state | or               | state |
1210          *
1211          * There's a hole, we need to insert something in it and
1212          * ignore the extent we found.
1213          */
1214         if (state->start > start) {
1215                 u64 this_end;
1216                 if (end < last_start)
1217                         this_end = end;
1218                 else
1219                         this_end = last_start - 1;
1220
1221                 prealloc = alloc_extent_state_atomic(prealloc);
1222                 if (!prealloc) {
1223                         err = -ENOMEM;
1224                         goto out;
1225                 }
1226
1227                 /*
1228                  * Avoid to free 'prealloc' if it can be merged with
1229                  * the later extent.
1230                  */
1231                 err = insert_state(tree, prealloc, start, this_end,
1232                                    NULL, NULL, &bits, NULL);
1233                 if (err)
1234                         extent_io_tree_panic(tree, err);
1235                 cache_state(prealloc, cached_state);
1236                 prealloc = NULL;
1237                 start = this_end + 1;
1238                 goto search_again;
1239         }
1240         /*
1241          * | ---- desired range ---- |
1242          *                        | state |
1243          * We need to split the extent, and set the bit
1244          * on the first half
1245          */
1246         if (state->start <= end && state->end > end) {
1247                 prealloc = alloc_extent_state_atomic(prealloc);
1248                 if (!prealloc) {
1249                         err = -ENOMEM;
1250                         goto out;
1251                 }
1252
1253                 err = split_state(tree, state, prealloc, end + 1);
1254                 if (err)
1255                         extent_io_tree_panic(tree, err);
1256
1257                 set_state_bits(tree, prealloc, &bits, NULL);
1258                 cache_state(prealloc, cached_state);
1259                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1260                 prealloc = NULL;
1261                 goto out;
1262         }
1263
1264 search_again:
1265         if (start > end)
1266                 goto out;
1267         spin_unlock(&tree->lock);
1268         cond_resched();
1269         first_iteration = false;
1270         goto again;
1271
1272 out:
1273         spin_unlock(&tree->lock);
1274         if (prealloc)
1275                 free_extent_state(prealloc);
1276
1277         return err;
1278 }
1279
1280 /* wrappers around set/clear extent bit */
1281 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1282                            unsigned bits, struct extent_changeset *changeset)
1283 {
1284         /*
1285          * We don't support EXTENT_LOCKED yet, as current changeset will
1286          * record any bits changed, so for EXTENT_LOCKED case, it will
1287          * either fail with -EEXIST or changeset will record the whole
1288          * range.
1289          */
1290         BUG_ON(bits & EXTENT_LOCKED);
1291
1292         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1293                                 changeset);
1294 }
1295
1296 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1297                      unsigned bits, int wake, int delete,
1298                      struct extent_state **cached)
1299 {
1300         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1301                                   cached, GFP_NOFS, NULL);
1302 }
1303
1304 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1305                 unsigned bits, struct extent_changeset *changeset)
1306 {
1307         /*
1308          * Don't support EXTENT_LOCKED case, same reason as
1309          * set_record_extent_bits().
1310          */
1311         BUG_ON(bits & EXTENT_LOCKED);
1312
1313         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1314                                   changeset);
1315 }
1316
1317 /*
1318  * either insert or lock state struct between start and end use mask to tell
1319  * us if waiting is desired.
1320  */
1321 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1322                      struct extent_state **cached_state)
1323 {
1324         int err;
1325         u64 failed_start;
1326
1327         while (1) {
1328                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1329                                        EXTENT_LOCKED, &failed_start,
1330                                        cached_state, GFP_NOFS, NULL);
1331                 if (err == -EEXIST) {
1332                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1333                         start = failed_start;
1334                 } else
1335                         break;
1336                 WARN_ON(start > end);
1337         }
1338         return err;
1339 }
1340
1341 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1342 {
1343         int err;
1344         u64 failed_start;
1345
1346         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1347                                &failed_start, NULL, GFP_NOFS, NULL);
1348         if (err == -EEXIST) {
1349                 if (failed_start > start)
1350                         clear_extent_bit(tree, start, failed_start - 1,
1351                                          EXTENT_LOCKED, 1, 0, NULL);
1352                 return 0;
1353         }
1354         return 1;
1355 }
1356
1357 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1358 {
1359         unsigned long index = start >> PAGE_SHIFT;
1360         unsigned long end_index = end >> PAGE_SHIFT;
1361         struct page *page;
1362
1363         while (index <= end_index) {
1364                 page = find_get_page(inode->i_mapping, index);
1365                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1366                 clear_page_dirty_for_io(page);
1367                 put_page(page);
1368                 index++;
1369         }
1370 }
1371
1372 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1373 {
1374         unsigned long index = start >> PAGE_SHIFT;
1375         unsigned long end_index = end >> PAGE_SHIFT;
1376         struct page *page;
1377
1378         while (index <= end_index) {
1379                 page = find_get_page(inode->i_mapping, index);
1380                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1381                 __set_page_dirty_nobuffers(page);
1382                 account_page_redirty(page);
1383                 put_page(page);
1384                 index++;
1385         }
1386 }
1387
1388 /*
1389  * helper function to set both pages and extents in the tree writeback
1390  */
1391 static void set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1392 {
1393         tree->ops->set_range_writeback(tree->private_data, start, end);
1394 }
1395
1396 /* find the first state struct with 'bits' set after 'start', and
1397  * return it.  tree->lock must be held.  NULL will returned if
1398  * nothing was found after 'start'
1399  */
1400 static struct extent_state *
1401 find_first_extent_bit_state(struct extent_io_tree *tree,
1402                             u64 start, unsigned bits)
1403 {
1404         struct rb_node *node;
1405         struct extent_state *state;
1406
1407         /*
1408          * this search will find all the extents that end after
1409          * our range starts.
1410          */
1411         node = tree_search(tree, start);
1412         if (!node)
1413                 goto out;
1414
1415         while (1) {
1416                 state = rb_entry(node, struct extent_state, rb_node);
1417                 if (state->end >= start && (state->state & bits))
1418                         return state;
1419
1420                 node = rb_next(node);
1421                 if (!node)
1422                         break;
1423         }
1424 out:
1425         return NULL;
1426 }
1427
1428 /*
1429  * find the first offset in the io tree with 'bits' set. zero is
1430  * returned if we find something, and *start_ret and *end_ret are
1431  * set to reflect the state struct that was found.
1432  *
1433  * If nothing was found, 1 is returned. If found something, return 0.
1434  */
1435 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1436                           u64 *start_ret, u64 *end_ret, unsigned bits,
1437                           struct extent_state **cached_state)
1438 {
1439         struct extent_state *state;
1440         struct rb_node *n;
1441         int ret = 1;
1442
1443         spin_lock(&tree->lock);
1444         if (cached_state && *cached_state) {
1445                 state = *cached_state;
1446                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1447                         n = rb_next(&state->rb_node);
1448                         while (n) {
1449                                 state = rb_entry(n, struct extent_state,
1450                                                  rb_node);
1451                                 if (state->state & bits)
1452                                         goto got_it;
1453                                 n = rb_next(n);
1454                         }
1455                         free_extent_state(*cached_state);
1456                         *cached_state = NULL;
1457                         goto out;
1458                 }
1459                 free_extent_state(*cached_state);
1460                 *cached_state = NULL;
1461         }
1462
1463         state = find_first_extent_bit_state(tree, start, bits);
1464 got_it:
1465         if (state) {
1466                 cache_state_if_flags(state, cached_state, 0);
1467                 *start_ret = state->start;
1468                 *end_ret = state->end;
1469                 ret = 0;
1470         }
1471 out:
1472         spin_unlock(&tree->lock);
1473         return ret;
1474 }
1475
1476 /*
1477  * find a contiguous range of bytes in the file marked as delalloc, not
1478  * more than 'max_bytes'.  start and end are used to return the range,
1479  *
1480  * 1 is returned if we find something, 0 if nothing was in the tree
1481  */
1482 static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1483                                         u64 *start, u64 *end, u64 max_bytes,
1484                                         struct extent_state **cached_state)
1485 {
1486         struct rb_node *node;
1487         struct extent_state *state;
1488         u64 cur_start = *start;
1489         u64 found = 0;
1490         u64 total_bytes = 0;
1491
1492         spin_lock(&tree->lock);
1493
1494         /*
1495          * this search will find all the extents that end after
1496          * our range starts.
1497          */
1498         node = tree_search(tree, cur_start);
1499         if (!node) {
1500                 if (!found)
1501                         *end = (u64)-1;
1502                 goto out;
1503         }
1504
1505         while (1) {
1506                 state = rb_entry(node, struct extent_state, rb_node);
1507                 if (found && (state->start != cur_start ||
1508                               (state->state & EXTENT_BOUNDARY))) {
1509                         goto out;
1510                 }
1511                 if (!(state->state & EXTENT_DELALLOC)) {
1512                         if (!found)
1513                                 *end = state->end;
1514                         goto out;
1515                 }
1516                 if (!found) {
1517                         *start = state->start;
1518                         *cached_state = state;
1519                         refcount_inc(&state->refs);
1520                 }
1521                 found++;
1522                 *end = state->end;
1523                 cur_start = state->end + 1;
1524                 node = rb_next(node);
1525                 total_bytes += state->end - state->start + 1;
1526                 if (total_bytes >= max_bytes)
1527                         break;
1528                 if (!node)
1529                         break;
1530         }
1531 out:
1532         spin_unlock(&tree->lock);
1533         return found;
1534 }
1535
1536 static int __process_pages_contig(struct address_space *mapping,
1537                                   struct page *locked_page,
1538                                   pgoff_t start_index, pgoff_t end_index,
1539                                   unsigned long page_ops, pgoff_t *index_ret);
1540
1541 static noinline void __unlock_for_delalloc(struct inode *inode,
1542                                            struct page *locked_page,
1543                                            u64 start, u64 end)
1544 {
1545         unsigned long index = start >> PAGE_SHIFT;
1546         unsigned long end_index = end >> PAGE_SHIFT;
1547
1548         ASSERT(locked_page);
1549         if (index == locked_page->index && end_index == index)
1550                 return;
1551
1552         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1553                                PAGE_UNLOCK, NULL);
1554 }
1555
1556 static noinline int lock_delalloc_pages(struct inode *inode,
1557                                         struct page *locked_page,
1558                                         u64 delalloc_start,
1559                                         u64 delalloc_end)
1560 {
1561         unsigned long index = delalloc_start >> PAGE_SHIFT;
1562         unsigned long index_ret = index;
1563         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1564         int ret;
1565
1566         ASSERT(locked_page);
1567         if (index == locked_page->index && index == end_index)
1568                 return 0;
1569
1570         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1571                                      end_index, PAGE_LOCK, &index_ret);
1572         if (ret == -EAGAIN)
1573                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1574                                       (u64)index_ret << PAGE_SHIFT);
1575         return ret;
1576 }
1577
1578 /*
1579  * find a contiguous range of bytes in the file marked as delalloc, not
1580  * more than 'max_bytes'.  start and end are used to return the range,
1581  *
1582  * 1 is returned if we find something, 0 if nothing was in the tree
1583  */
1584 STATIC u64 find_lock_delalloc_range(struct inode *inode,
1585                                     struct extent_io_tree *tree,
1586                                     struct page *locked_page, u64 *start,
1587                                     u64 *end, u64 max_bytes)
1588 {
1589         u64 delalloc_start;
1590         u64 delalloc_end;
1591         u64 found;
1592         struct extent_state *cached_state = NULL;
1593         int ret;
1594         int loops = 0;
1595
1596 again:
1597         /* step one, find a bunch of delalloc bytes starting at start */
1598         delalloc_start = *start;
1599         delalloc_end = 0;
1600         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1601                                     max_bytes, &cached_state);
1602         if (!found || delalloc_end <= *start) {
1603                 *start = delalloc_start;
1604                 *end = delalloc_end;
1605                 free_extent_state(cached_state);
1606                 return 0;
1607         }
1608
1609         /*
1610          * start comes from the offset of locked_page.  We have to lock
1611          * pages in order, so we can't process delalloc bytes before
1612          * locked_page
1613          */
1614         if (delalloc_start < *start)
1615                 delalloc_start = *start;
1616
1617         /*
1618          * make sure to limit the number of pages we try to lock down
1619          */
1620         if (delalloc_end + 1 - delalloc_start > max_bytes)
1621                 delalloc_end = delalloc_start + max_bytes - 1;
1622
1623         /* step two, lock all the pages after the page that has start */
1624         ret = lock_delalloc_pages(inode, locked_page,
1625                                   delalloc_start, delalloc_end);
1626         if (ret == -EAGAIN) {
1627                 /* some of the pages are gone, lets avoid looping by
1628                  * shortening the size of the delalloc range we're searching
1629                  */
1630                 free_extent_state(cached_state);
1631                 cached_state = NULL;
1632                 if (!loops) {
1633                         max_bytes = PAGE_SIZE;
1634                         loops = 1;
1635                         goto again;
1636                 } else {
1637                         found = 0;
1638                         goto out_failed;
1639                 }
1640         }
1641         BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1642
1643         /* step three, lock the state bits for the whole range */
1644         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1645
1646         /* then test to make sure it is all still delalloc */
1647         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1648                              EXTENT_DELALLOC, 1, cached_state);
1649         if (!ret) {
1650                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1651                                      &cached_state);
1652                 __unlock_for_delalloc(inode, locked_page,
1653                               delalloc_start, delalloc_end);
1654                 cond_resched();
1655                 goto again;
1656         }
1657         free_extent_state(cached_state);
1658         *start = delalloc_start;
1659         *end = delalloc_end;
1660 out_failed:
1661         return found;
1662 }
1663
1664 static int __process_pages_contig(struct address_space *mapping,
1665                                   struct page *locked_page,
1666                                   pgoff_t start_index, pgoff_t end_index,
1667                                   unsigned long page_ops, pgoff_t *index_ret)
1668 {
1669         unsigned long nr_pages = end_index - start_index + 1;
1670         unsigned long pages_locked = 0;
1671         pgoff_t index = start_index;
1672         struct page *pages[16];
1673         unsigned ret;
1674         int err = 0;
1675         int i;
1676
1677         if (page_ops & PAGE_LOCK) {
1678                 ASSERT(page_ops == PAGE_LOCK);
1679                 ASSERT(index_ret && *index_ret == start_index);
1680         }
1681
1682         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1683                 mapping_set_error(mapping, -EIO);
1684
1685         while (nr_pages > 0) {
1686                 ret = find_get_pages_contig(mapping, index,
1687                                      min_t(unsigned long,
1688                                      nr_pages, ARRAY_SIZE(pages)), pages);
1689                 if (ret == 0) {
1690                         /*
1691                          * Only if we're going to lock these pages,
1692                          * can we find nothing at @index.
1693                          */
1694                         ASSERT(page_ops & PAGE_LOCK);
1695                         err = -EAGAIN;
1696                         goto out;
1697                 }
1698
1699                 for (i = 0; i < ret; i++) {
1700                         if (page_ops & PAGE_SET_PRIVATE2)
1701                                 SetPagePrivate2(pages[i]);
1702
1703                         if (pages[i] == locked_page) {
1704                                 put_page(pages[i]);
1705                                 pages_locked++;
1706                                 continue;
1707                         }
1708                         if (page_ops & PAGE_CLEAR_DIRTY)
1709                                 clear_page_dirty_for_io(pages[i]);
1710                         if (page_ops & PAGE_SET_WRITEBACK)
1711                                 set_page_writeback(pages[i]);
1712                         if (page_ops & PAGE_SET_ERROR)
1713                                 SetPageError(pages[i]);
1714                         if (page_ops & PAGE_END_WRITEBACK)
1715                                 end_page_writeback(pages[i]);
1716                         if (page_ops & PAGE_UNLOCK)
1717                                 unlock_page(pages[i]);
1718                         if (page_ops & PAGE_LOCK) {
1719                                 lock_page(pages[i]);
1720                                 if (!PageDirty(pages[i]) ||
1721                                     pages[i]->mapping != mapping) {
1722                                         unlock_page(pages[i]);
1723                                         put_page(pages[i]);
1724                                         err = -EAGAIN;
1725                                         goto out;
1726                                 }
1727                         }
1728                         put_page(pages[i]);
1729                         pages_locked++;
1730                 }
1731                 nr_pages -= ret;
1732                 index += ret;
1733                 cond_resched();
1734         }
1735 out:
1736         if (err && index_ret)
1737                 *index_ret = start_index + pages_locked - 1;
1738         return err;
1739 }
1740
1741 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1742                                  u64 delalloc_end, struct page *locked_page,
1743                                  unsigned clear_bits,
1744                                  unsigned long page_ops)
1745 {
1746         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1747                          NULL);
1748
1749         __process_pages_contig(inode->i_mapping, locked_page,
1750                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1751                                page_ops, NULL);
1752 }
1753
1754 /*
1755  * count the number of bytes in the tree that have a given bit(s)
1756  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1757  * cached.  The total number found is returned.
1758  */
1759 u64 count_range_bits(struct extent_io_tree *tree,
1760                      u64 *start, u64 search_end, u64 max_bytes,
1761                      unsigned bits, int contig)
1762 {
1763         struct rb_node *node;
1764         struct extent_state *state;
1765         u64 cur_start = *start;
1766         u64 total_bytes = 0;
1767         u64 last = 0;
1768         int found = 0;
1769
1770         if (WARN_ON(search_end <= cur_start))
1771                 return 0;
1772
1773         spin_lock(&tree->lock);
1774         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1775                 total_bytes = tree->dirty_bytes;
1776                 goto out;
1777         }
1778         /*
1779          * this search will find all the extents that end after
1780          * our range starts.
1781          */
1782         node = tree_search(tree, cur_start);
1783         if (!node)
1784                 goto out;
1785
1786         while (1) {
1787                 state = rb_entry(node, struct extent_state, rb_node);
1788                 if (state->start > search_end)
1789                         break;
1790                 if (contig && found && state->start > last + 1)
1791                         break;
1792                 if (state->end >= cur_start && (state->state & bits) == bits) {
1793                         total_bytes += min(search_end, state->end) + 1 -
1794                                        max(cur_start, state->start);
1795                         if (total_bytes >= max_bytes)
1796                                 break;
1797                         if (!found) {
1798                                 *start = max(cur_start, state->start);
1799                                 found = 1;
1800                         }
1801                         last = state->end;
1802                 } else if (contig && found) {
1803                         break;
1804                 }
1805                 node = rb_next(node);
1806                 if (!node)
1807                         break;
1808         }
1809 out:
1810         spin_unlock(&tree->lock);
1811         return total_bytes;
1812 }
1813
1814 /*
1815  * set the private field for a given byte offset in the tree.  If there isn't
1816  * an extent_state there already, this does nothing.
1817  */
1818 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
1819                 struct io_failure_record *failrec)
1820 {
1821         struct rb_node *node;
1822         struct extent_state *state;
1823         int ret = 0;
1824
1825         spin_lock(&tree->lock);
1826         /*
1827          * this search will find all the extents that end after
1828          * our range starts.
1829          */
1830         node = tree_search(tree, start);
1831         if (!node) {
1832                 ret = -ENOENT;
1833                 goto out;
1834         }
1835         state = rb_entry(node, struct extent_state, rb_node);
1836         if (state->start != start) {
1837                 ret = -ENOENT;
1838                 goto out;
1839         }
1840         state->failrec = failrec;
1841 out:
1842         spin_unlock(&tree->lock);
1843         return ret;
1844 }
1845
1846 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
1847                 struct io_failure_record **failrec)
1848 {
1849         struct rb_node *node;
1850         struct extent_state *state;
1851         int ret = 0;
1852
1853         spin_lock(&tree->lock);
1854         /*
1855          * this search will find all the extents that end after
1856          * our range starts.
1857          */
1858         node = tree_search(tree, start);
1859         if (!node) {
1860                 ret = -ENOENT;
1861                 goto out;
1862         }
1863         state = rb_entry(node, struct extent_state, rb_node);
1864         if (state->start != start) {
1865                 ret = -ENOENT;
1866                 goto out;
1867         }
1868         *failrec = state->failrec;
1869 out:
1870         spin_unlock(&tree->lock);
1871         return ret;
1872 }
1873
1874 /*
1875  * searches a range in the state tree for a given mask.
1876  * If 'filled' == 1, this returns 1 only if every extent in the tree
1877  * has the bits set.  Otherwise, 1 is returned if any bit in the
1878  * range is found set.
1879  */
1880 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1881                    unsigned bits, int filled, struct extent_state *cached)
1882 {
1883         struct extent_state *state = NULL;
1884         struct rb_node *node;
1885         int bitset = 0;
1886
1887         spin_lock(&tree->lock);
1888         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
1889             cached->end > start)
1890                 node = &cached->rb_node;
1891         else
1892                 node = tree_search(tree, start);
1893         while (node && start <= end) {
1894                 state = rb_entry(node, struct extent_state, rb_node);
1895
1896                 if (filled && state->start > start) {
1897                         bitset = 0;
1898                         break;
1899                 }
1900
1901                 if (state->start > end)
1902                         break;
1903
1904                 if (state->state & bits) {
1905                         bitset = 1;
1906                         if (!filled)
1907                                 break;
1908                 } else if (filled) {
1909                         bitset = 0;
1910                         break;
1911                 }
1912
1913                 if (state->end == (u64)-1)
1914                         break;
1915
1916                 start = state->end + 1;
1917                 if (start > end)
1918                         break;
1919                 node = rb_next(node);
1920                 if (!node) {
1921                         if (filled)
1922                                 bitset = 0;
1923                         break;
1924                 }
1925         }
1926         spin_unlock(&tree->lock);
1927         return bitset;
1928 }
1929
1930 /*
1931  * helper function to set a given page up to date if all the
1932  * extents in the tree for that page are up to date
1933  */
1934 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1935 {
1936         u64 start = page_offset(page);
1937         u64 end = start + PAGE_SIZE - 1;
1938         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1939                 SetPageUptodate(page);
1940 }
1941
1942 int free_io_failure(struct extent_io_tree *failure_tree,
1943                     struct extent_io_tree *io_tree,
1944                     struct io_failure_record *rec)
1945 {
1946         int ret;
1947         int err = 0;
1948
1949         set_state_failrec(failure_tree, rec->start, NULL);
1950         ret = clear_extent_bits(failure_tree, rec->start,
1951                                 rec->start + rec->len - 1,
1952                                 EXTENT_LOCKED | EXTENT_DIRTY);
1953         if (ret)
1954                 err = ret;
1955
1956         ret = clear_extent_bits(io_tree, rec->start,
1957                                 rec->start + rec->len - 1,
1958                                 EXTENT_DAMAGED);
1959         if (ret && !err)
1960                 err = ret;
1961
1962         kfree(rec);
1963         return err;
1964 }
1965
1966 /*
1967  * this bypasses the standard btrfs submit functions deliberately, as
1968  * the standard behavior is to write all copies in a raid setup. here we only
1969  * want to write the one bad copy. so we do the mapping for ourselves and issue
1970  * submit_bio directly.
1971  * to avoid any synchronization issues, wait for the data after writing, which
1972  * actually prevents the read that triggered the error from finishing.
1973  * currently, there can be no more than two copies of every data bit. thus,
1974  * exactly one rewrite is required.
1975  */
1976 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
1977                       u64 length, u64 logical, struct page *page,
1978                       unsigned int pg_offset, int mirror_num)
1979 {
1980         struct bio *bio;
1981         struct btrfs_device *dev;
1982         u64 map_length = 0;
1983         u64 sector;
1984         struct btrfs_bio *bbio = NULL;
1985         int ret;
1986
1987         ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
1988         BUG_ON(!mirror_num);
1989
1990         bio = btrfs_io_bio_alloc(1);
1991         bio->bi_iter.bi_size = 0;
1992         map_length = length;
1993
1994         /*
1995          * Avoid races with device replace and make sure our bbio has devices
1996          * associated to its stripes that don't go away while we are doing the
1997          * read repair operation.
1998          */
1999         btrfs_bio_counter_inc_blocked(fs_info);
2000         if (btrfs_is_parity_mirror(fs_info, logical, length)) {
2001                 /*
2002                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2003                  * to update all raid stripes, but here we just want to correct
2004                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2005                  * stripe's dev and sector.
2006                  */
2007                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2008                                       &map_length, &bbio, 0);
2009                 if (ret) {
2010                         btrfs_bio_counter_dec(fs_info);
2011                         bio_put(bio);
2012                         return -EIO;
2013                 }
2014                 ASSERT(bbio->mirror_num == 1);
2015         } else {
2016                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2017                                       &map_length, &bbio, mirror_num);
2018                 if (ret) {
2019                         btrfs_bio_counter_dec(fs_info);
2020                         bio_put(bio);
2021                         return -EIO;
2022                 }
2023                 BUG_ON(mirror_num != bbio->mirror_num);
2024         }
2025
2026         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2027         bio->bi_iter.bi_sector = sector;
2028         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2029         btrfs_put_bbio(bbio);
2030         if (!dev || !dev->bdev ||
2031             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2032                 btrfs_bio_counter_dec(fs_info);
2033                 bio_put(bio);
2034                 return -EIO;
2035         }
2036         bio_set_dev(bio, dev->bdev);
2037         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2038         bio_add_page(bio, page, length, pg_offset);
2039
2040         if (btrfsic_submit_bio_wait(bio)) {
2041                 /* try to remap that extent elsewhere? */
2042                 btrfs_bio_counter_dec(fs_info);
2043                 bio_put(bio);
2044                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2045                 return -EIO;
2046         }
2047
2048         btrfs_info_rl_in_rcu(fs_info,
2049                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2050                                   ino, start,
2051                                   rcu_str_deref(dev->name), sector);
2052         btrfs_bio_counter_dec(fs_info);
2053         bio_put(bio);
2054         return 0;
2055 }
2056
2057 int repair_eb_io_failure(struct btrfs_fs_info *fs_info,
2058                          struct extent_buffer *eb, int mirror_num)
2059 {
2060         u64 start = eb->start;
2061         unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2062         int ret = 0;
2063
2064         if (sb_rdonly(fs_info->sb))
2065                 return -EROFS;
2066
2067         for (i = 0; i < num_pages; i++) {
2068                 struct page *p = eb->pages[i];
2069
2070                 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2071                                         start - page_offset(p), mirror_num);
2072                 if (ret)
2073                         break;
2074                 start += PAGE_SIZE;
2075         }
2076
2077         return ret;
2078 }
2079
2080 /*
2081  * each time an IO finishes, we do a fast check in the IO failure tree
2082  * to see if we need to process or clean up an io_failure_record
2083  */
2084 int clean_io_failure(struct btrfs_fs_info *fs_info,
2085                      struct extent_io_tree *failure_tree,
2086                      struct extent_io_tree *io_tree, u64 start,
2087                      struct page *page, u64 ino, unsigned int pg_offset)
2088 {
2089         u64 private;
2090         struct io_failure_record *failrec;
2091         struct extent_state *state;
2092         int num_copies;
2093         int ret;
2094
2095         private = 0;
2096         ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2097                                EXTENT_DIRTY, 0);
2098         if (!ret)
2099                 return 0;
2100
2101         ret = get_state_failrec(failure_tree, start, &failrec);
2102         if (ret)
2103                 return 0;
2104
2105         BUG_ON(!failrec->this_mirror);
2106
2107         if (failrec->in_validation) {
2108                 /* there was no real error, just free the record */
2109                 btrfs_debug(fs_info,
2110                         "clean_io_failure: freeing dummy error at %llu",
2111                         failrec->start);
2112                 goto out;
2113         }
2114         if (sb_rdonly(fs_info->sb))
2115                 goto out;
2116
2117         spin_lock(&io_tree->lock);
2118         state = find_first_extent_bit_state(io_tree,
2119                                             failrec->start,
2120                                             EXTENT_LOCKED);
2121         spin_unlock(&io_tree->lock);
2122
2123         if (state && state->start <= failrec->start &&
2124             state->end >= failrec->start + failrec->len - 1) {
2125                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2126                                               failrec->len);
2127                 if (num_copies > 1)  {
2128                         repair_io_failure(fs_info, ino, start, failrec->len,
2129                                           failrec->logical, page, pg_offset,
2130                                           failrec->failed_mirror);
2131                 }
2132         }
2133
2134 out:
2135         free_io_failure(failure_tree, io_tree, failrec);
2136
2137         return 0;
2138 }
2139
2140 /*
2141  * Can be called when
2142  * - hold extent lock
2143  * - under ordered extent
2144  * - the inode is freeing
2145  */
2146 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2147 {
2148         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2149         struct io_failure_record *failrec;
2150         struct extent_state *state, *next;
2151
2152         if (RB_EMPTY_ROOT(&failure_tree->state))
2153                 return;
2154
2155         spin_lock(&failure_tree->lock);
2156         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2157         while (state) {
2158                 if (state->start > end)
2159                         break;
2160
2161                 ASSERT(state->end <= end);
2162
2163                 next = next_state(state);
2164
2165                 failrec = state->failrec;
2166                 free_extent_state(state);
2167                 kfree(failrec);
2168
2169                 state = next;
2170         }
2171         spin_unlock(&failure_tree->lock);
2172 }
2173
2174 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2175                 struct io_failure_record **failrec_ret)
2176 {
2177         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2178         struct io_failure_record *failrec;
2179         struct extent_map *em;
2180         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2181         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2182         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2183         int ret;
2184         u64 logical;
2185
2186         ret = get_state_failrec(failure_tree, start, &failrec);
2187         if (ret) {
2188                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2189                 if (!failrec)
2190                         return -ENOMEM;
2191
2192                 failrec->start = start;
2193                 failrec->len = end - start + 1;
2194                 failrec->this_mirror = 0;
2195                 failrec->bio_flags = 0;
2196                 failrec->in_validation = 0;
2197
2198                 read_lock(&em_tree->lock);
2199                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2200                 if (!em) {
2201                         read_unlock(&em_tree->lock);
2202                         kfree(failrec);
2203                         return -EIO;
2204                 }
2205
2206                 if (em->start > start || em->start + em->len <= start) {
2207                         free_extent_map(em);
2208                         em = NULL;
2209                 }
2210                 read_unlock(&em_tree->lock);
2211                 if (!em) {
2212                         kfree(failrec);
2213                         return -EIO;
2214                 }
2215
2216                 logical = start - em->start;
2217                 logical = em->block_start + logical;
2218                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2219                         logical = em->block_start;
2220                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2221                         extent_set_compress_type(&failrec->bio_flags,
2222                                                  em->compress_type);
2223                 }
2224
2225                 btrfs_debug(fs_info,
2226                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2227                         logical, start, failrec->len);
2228
2229                 failrec->logical = logical;
2230                 free_extent_map(em);
2231
2232                 /* set the bits in the private failure tree */
2233                 ret = set_extent_bits(failure_tree, start, end,
2234                                         EXTENT_LOCKED | EXTENT_DIRTY);
2235                 if (ret >= 0)
2236                         ret = set_state_failrec(failure_tree, start, failrec);
2237                 /* set the bits in the inode's tree */
2238                 if (ret >= 0)
2239                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2240                 if (ret < 0) {
2241                         kfree(failrec);
2242                         return ret;
2243                 }
2244         } else {
2245                 btrfs_debug(fs_info,
2246                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2247                         failrec->logical, failrec->start, failrec->len,
2248                         failrec->in_validation);
2249                 /*
2250                  * when data can be on disk more than twice, add to failrec here
2251                  * (e.g. with a list for failed_mirror) to make
2252                  * clean_io_failure() clean all those errors at once.
2253                  */
2254         }
2255
2256         *failrec_ret = failrec;
2257
2258         return 0;
2259 }
2260
2261 bool btrfs_check_repairable(struct inode *inode, unsigned failed_bio_pages,
2262                            struct io_failure_record *failrec, int failed_mirror)
2263 {
2264         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2265         int num_copies;
2266
2267         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2268         if (num_copies == 1) {
2269                 /*
2270                  * we only have a single copy of the data, so don't bother with
2271                  * all the retry and error correction code that follows. no
2272                  * matter what the error is, it is very likely to persist.
2273                  */
2274                 btrfs_debug(fs_info,
2275                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2276                         num_copies, failrec->this_mirror, failed_mirror);
2277                 return false;
2278         }
2279
2280         /*
2281          * there are two premises:
2282          *      a) deliver good data to the caller
2283          *      b) correct the bad sectors on disk
2284          */
2285         if (failed_bio_pages > 1) {
2286                 /*
2287                  * to fulfill b), we need to know the exact failing sectors, as
2288                  * we don't want to rewrite any more than the failed ones. thus,
2289                  * we need separate read requests for the failed bio
2290                  *
2291                  * if the following BUG_ON triggers, our validation request got
2292                  * merged. we need separate requests for our algorithm to work.
2293                  */
2294                 BUG_ON(failrec->in_validation);
2295                 failrec->in_validation = 1;
2296                 failrec->this_mirror = failed_mirror;
2297         } else {
2298                 /*
2299                  * we're ready to fulfill a) and b) alongside. get a good copy
2300                  * of the failed sector and if we succeed, we have setup
2301                  * everything for repair_io_failure to do the rest for us.
2302                  */
2303                 if (failrec->in_validation) {
2304                         BUG_ON(failrec->this_mirror != failed_mirror);
2305                         failrec->in_validation = 0;
2306                         failrec->this_mirror = 0;
2307                 }
2308                 failrec->failed_mirror = failed_mirror;
2309                 failrec->this_mirror++;
2310                 if (failrec->this_mirror == failed_mirror)
2311                         failrec->this_mirror++;
2312         }
2313
2314         if (failrec->this_mirror > num_copies) {
2315                 btrfs_debug(fs_info,
2316                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2317                         num_copies, failrec->this_mirror, failed_mirror);
2318                 return false;
2319         }
2320
2321         return true;
2322 }
2323
2324
2325 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2326                                     struct io_failure_record *failrec,
2327                                     struct page *page, int pg_offset, int icsum,
2328                                     bio_end_io_t *endio_func, void *data)
2329 {
2330         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2331         struct bio *bio;
2332         struct btrfs_io_bio *btrfs_failed_bio;
2333         struct btrfs_io_bio *btrfs_bio;
2334
2335         bio = btrfs_io_bio_alloc(1);
2336         bio->bi_end_io = endio_func;
2337         bio->bi_iter.bi_sector = failrec->logical >> 9;
2338         bio_set_dev(bio, fs_info->fs_devices->latest_bdev);
2339         bio->bi_iter.bi_size = 0;
2340         bio->bi_private = data;
2341
2342         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2343         if (btrfs_failed_bio->csum) {
2344                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2345
2346                 btrfs_bio = btrfs_io_bio(bio);
2347                 btrfs_bio->csum = btrfs_bio->csum_inline;
2348                 icsum *= csum_size;
2349                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2350                        csum_size);
2351         }
2352
2353         bio_add_page(bio, page, failrec->len, pg_offset);
2354
2355         return bio;
2356 }
2357
2358 /*
2359  * this is a generic handler for readpage errors (default
2360  * readpage_io_failed_hook). if other copies exist, read those and write back
2361  * good data to the failed position. does not investigate in remapping the
2362  * failed extent elsewhere, hoping the device will be smart enough to do this as
2363  * needed
2364  */
2365
2366 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2367                               struct page *page, u64 start, u64 end,
2368                               int failed_mirror)
2369 {
2370         struct io_failure_record *failrec;
2371         struct inode *inode = page->mapping->host;
2372         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2373         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2374         struct bio *bio;
2375         int read_mode = 0;
2376         blk_status_t status;
2377         int ret;
2378         unsigned failed_bio_pages = bio_pages_all(failed_bio);
2379
2380         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2381
2382         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2383         if (ret)
2384                 return ret;
2385
2386         if (!btrfs_check_repairable(inode, failed_bio_pages, failrec,
2387                                     failed_mirror)) {
2388                 free_io_failure(failure_tree, tree, failrec);
2389                 return -EIO;
2390         }
2391
2392         if (failed_bio_pages > 1)
2393                 read_mode |= REQ_FAILFAST_DEV;
2394
2395         phy_offset >>= inode->i_sb->s_blocksize_bits;
2396         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2397                                       start - page_offset(page),
2398                                       (int)phy_offset, failed_bio->bi_end_io,
2399                                       NULL);
2400         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
2401
2402         btrfs_debug(btrfs_sb(inode->i_sb),
2403                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2404                 read_mode, failrec->this_mirror, failrec->in_validation);
2405
2406         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2407                                          failrec->bio_flags, 0);
2408         if (status) {
2409                 free_io_failure(failure_tree, tree, failrec);
2410                 bio_put(bio);
2411                 ret = blk_status_to_errno(status);
2412         }
2413
2414         return ret;
2415 }
2416
2417 /* lots and lots of room for performance fixes in the end_bio funcs */
2418
2419 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2420 {
2421         int uptodate = (err == 0);
2422         struct extent_io_tree *tree;
2423         int ret = 0;
2424
2425         tree = &BTRFS_I(page->mapping->host)->io_tree;
2426
2427         if (tree->ops && tree->ops->writepage_end_io_hook)
2428                 tree->ops->writepage_end_io_hook(page, start, end, NULL,
2429                                 uptodate);
2430
2431         if (!uptodate) {
2432                 ClearPageUptodate(page);
2433                 SetPageError(page);
2434                 ret = err < 0 ? err : -EIO;
2435                 mapping_set_error(page->mapping, ret);
2436         }
2437 }
2438
2439 /*
2440  * after a writepage IO is done, we need to:
2441  * clear the uptodate bits on error
2442  * clear the writeback bits in the extent tree for this IO
2443  * end_page_writeback if the page has no more pending IO
2444  *
2445  * Scheduling is not allowed, so the extent state tree is expected
2446  * to have one and only one object corresponding to this IO.
2447  */
2448 static void end_bio_extent_writepage(struct bio *bio)
2449 {
2450         int error = blk_status_to_errno(bio->bi_status);
2451         struct bio_vec *bvec;
2452         u64 start;
2453         u64 end;
2454         int i;
2455
2456         ASSERT(!bio_flagged(bio, BIO_CLONED));
2457         bio_for_each_segment_all(bvec, bio, i) {
2458                 struct page *page = bvec->bv_page;
2459                 struct inode *inode = page->mapping->host;
2460                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2461
2462                 /* We always issue full-page reads, but if some block
2463                  * in a page fails to read, blk_update_request() will
2464                  * advance bv_offset and adjust bv_len to compensate.
2465                  * Print a warning for nonzero offsets, and an error
2466                  * if they don't add up to a full page.  */
2467                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2468                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2469                                 btrfs_err(fs_info,
2470                                    "partial page write in btrfs with offset %u and length %u",
2471                                         bvec->bv_offset, bvec->bv_len);
2472                         else
2473                                 btrfs_info(fs_info,
2474                                    "incomplete page write in btrfs with offset %u and length %u",
2475                                         bvec->bv_offset, bvec->bv_len);
2476                 }
2477
2478                 start = page_offset(page);
2479                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2480
2481                 end_extent_writepage(page, error, start, end);
2482                 end_page_writeback(page);
2483         }
2484
2485         bio_put(bio);
2486 }
2487
2488 static void
2489 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2490                               int uptodate)
2491 {
2492         struct extent_state *cached = NULL;
2493         u64 end = start + len - 1;
2494
2495         if (uptodate && tree->track_uptodate)
2496                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2497         unlock_extent_cached_atomic(tree, start, end, &cached);
2498 }
2499
2500 /*
2501  * after a readpage IO is done, we need to:
2502  * clear the uptodate bits on error
2503  * set the uptodate bits if things worked
2504  * set the page up to date if all extents in the tree are uptodate
2505  * clear the lock bit in the extent tree
2506  * unlock the page if there are no other extents locked for it
2507  *
2508  * Scheduling is not allowed, so the extent state tree is expected
2509  * to have one and only one object corresponding to this IO.
2510  */
2511 static void end_bio_extent_readpage(struct bio *bio)
2512 {
2513         struct bio_vec *bvec;
2514         int uptodate = !bio->bi_status;
2515         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2516         struct extent_io_tree *tree, *failure_tree;
2517         u64 offset = 0;
2518         u64 start;
2519         u64 end;
2520         u64 len;
2521         u64 extent_start = 0;
2522         u64 extent_len = 0;
2523         int mirror;
2524         int ret;
2525         int i;
2526
2527         ASSERT(!bio_flagged(bio, BIO_CLONED));
2528         bio_for_each_segment_all(bvec, bio, i) {
2529                 struct page *page = bvec->bv_page;
2530                 struct inode *inode = page->mapping->host;
2531                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2532
2533                 btrfs_debug(fs_info,
2534                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2535                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2536                         io_bio->mirror_num);
2537                 tree = &BTRFS_I(inode)->io_tree;
2538                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2539
2540                 /* We always issue full-page reads, but if some block
2541                  * in a page fails to read, blk_update_request() will
2542                  * advance bv_offset and adjust bv_len to compensate.
2543                  * Print a warning for nonzero offsets, and an error
2544                  * if they don't add up to a full page.  */
2545                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2546                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2547                                 btrfs_err(fs_info,
2548                                         "partial page read in btrfs with offset %u and length %u",
2549                                         bvec->bv_offset, bvec->bv_len);
2550                         else
2551                                 btrfs_info(fs_info,
2552                                         "incomplete page read in btrfs with offset %u and length %u",
2553                                         bvec->bv_offset, bvec->bv_len);
2554                 }
2555
2556                 start = page_offset(page);
2557                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2558                 len = bvec->bv_len;
2559
2560                 mirror = io_bio->mirror_num;
2561                 if (likely(uptodate && tree->ops)) {
2562                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2563                                                               page, start, end,
2564                                                               mirror);
2565                         if (ret)
2566                                 uptodate = 0;
2567                         else
2568                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2569                                                  failure_tree, tree, start,
2570                                                  page,
2571                                                  btrfs_ino(BTRFS_I(inode)), 0);
2572                 }
2573
2574                 if (likely(uptodate))
2575                         goto readpage_ok;
2576
2577                 if (tree->ops) {
2578                         ret = tree->ops->readpage_io_failed_hook(page, mirror);
2579                         if (ret == -EAGAIN) {
2580                                 /*
2581                                  * Data inode's readpage_io_failed_hook() always
2582                                  * returns -EAGAIN.
2583                                  *
2584                                  * The generic bio_readpage_error handles errors
2585                                  * the following way: If possible, new read
2586                                  * requests are created and submitted and will
2587                                  * end up in end_bio_extent_readpage as well (if
2588                                  * we're lucky, not in the !uptodate case). In
2589                                  * that case it returns 0 and we just go on with
2590                                  * the next page in our bio. If it can't handle
2591                                  * the error it will return -EIO and we remain
2592                                  * responsible for that page.
2593                                  */
2594                                 ret = bio_readpage_error(bio, offset, page,
2595                                                          start, end, mirror);
2596                                 if (ret == 0) {
2597                                         uptodate = !bio->bi_status;
2598                                         offset += len;
2599                                         continue;
2600                                 }
2601                         }
2602
2603                         /*
2604                          * metadata's readpage_io_failed_hook() always returns
2605                          * -EIO and fixes nothing.  -EIO is also returned if
2606                          * data inode error could not be fixed.
2607                          */
2608                         ASSERT(ret == -EIO);
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, blk_status_to_errno(bio->bi_status));
2656         bio_put(bio);
2657 }
2658
2659 /*
2660  * Initialize the members up to but not including 'bio'. Use after allocating a
2661  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2662  * 'bio' because use of __GFP_ZERO is not supported.
2663  */
2664 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2665 {
2666         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2667 }
2668
2669 /*
2670  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2671  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2672  * for the appropriate container_of magic
2673  */
2674 struct bio *btrfs_bio_alloc(struct block_device *bdev, u64 first_byte)
2675 {
2676         struct bio *bio;
2677
2678         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, btrfs_bioset);
2679         bio_set_dev(bio, bdev);
2680         bio->bi_iter.bi_sector = first_byte >> 9;
2681         btrfs_io_bio_init(btrfs_io_bio(bio));
2682         return bio;
2683 }
2684
2685 struct bio *btrfs_bio_clone(struct bio *bio)
2686 {
2687         struct btrfs_io_bio *btrfs_bio;
2688         struct bio *new;
2689
2690         /* Bio allocation backed by a bioset does not fail */
2691         new = bio_clone_fast(bio, GFP_NOFS, btrfs_bioset);
2692         btrfs_bio = btrfs_io_bio(new);
2693         btrfs_io_bio_init(btrfs_bio);
2694         btrfs_bio->iter = bio->bi_iter;
2695         return new;
2696 }
2697
2698 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2699 {
2700         struct bio *bio;
2701
2702         /* Bio allocation backed by a bioset does not fail */
2703         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, btrfs_bioset);
2704         btrfs_io_bio_init(btrfs_io_bio(bio));
2705         return bio;
2706 }
2707
2708 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2709 {
2710         struct bio *bio;
2711         struct btrfs_io_bio *btrfs_bio;
2712
2713         /* this will never fail when it's backed by a bioset */
2714         bio = bio_clone_fast(orig, GFP_NOFS, btrfs_bioset);
2715         ASSERT(bio);
2716
2717         btrfs_bio = btrfs_io_bio(bio);
2718         btrfs_io_bio_init(btrfs_bio);
2719
2720         bio_trim(bio, offset >> 9, size >> 9);
2721         btrfs_bio->iter = bio->bi_iter;
2722         return bio;
2723 }
2724
2725 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
2726                                        unsigned long bio_flags)
2727 {
2728         blk_status_t ret = 0;
2729         struct bio_vec *bvec = bio_last_bvec_all(bio);
2730         struct page *page = bvec->bv_page;
2731         struct extent_io_tree *tree = bio->bi_private;
2732         u64 start;
2733
2734         start = page_offset(page) + bvec->bv_offset;
2735
2736         bio->bi_private = NULL;
2737
2738         if (tree->ops)
2739                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
2740                                            mirror_num, bio_flags, start);
2741         else
2742                 btrfsic_submit_bio(bio);
2743
2744         return blk_status_to_errno(ret);
2745 }
2746
2747 /*
2748  * @opf:        bio REQ_OP_* and REQ_* flags as one value
2749  * @tree:       tree so we can call our merge_bio hook
2750  * @wbc:        optional writeback control for io accounting
2751  * @page:       page to add to the bio
2752  * @pg_offset:  offset of the new bio or to check whether we are adding
2753  *              a contiguous page to the previous one
2754  * @size:       portion of page that we want to write
2755  * @offset:     starting offset in the page
2756  * @bdev:       attach newly created bios to this bdev
2757  * @bio_ret:    must be valid pointer, newly allocated bio will be stored there
2758  * @end_io_func:     end_io callback for new bio
2759  * @mirror_num:      desired mirror to read/write
2760  * @prev_bio_flags:  flags of previous bio to see if we can merge the current one
2761  * @bio_flags:  flags of the current bio to see if we can merge them
2762  */
2763 static int submit_extent_page(unsigned int opf, struct extent_io_tree *tree,
2764                               struct writeback_control *wbc,
2765                               struct page *page, u64 offset,
2766                               size_t size, unsigned long pg_offset,
2767                               struct block_device *bdev,
2768                               struct bio **bio_ret,
2769                               bio_end_io_t end_io_func,
2770                               int mirror_num,
2771                               unsigned long prev_bio_flags,
2772                               unsigned long bio_flags,
2773                               bool force_bio_submit)
2774 {
2775         int ret = 0;
2776         struct bio *bio;
2777         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2778         sector_t sector = offset >> 9;
2779
2780         ASSERT(bio_ret);
2781
2782         if (*bio_ret) {
2783                 bool contig;
2784                 bool can_merge = true;
2785
2786                 bio = *bio_ret;
2787                 if (prev_bio_flags & EXTENT_BIO_COMPRESSED)
2788                         contig = bio->bi_iter.bi_sector == sector;
2789                 else
2790                         contig = bio_end_sector(bio) == sector;
2791
2792                 if (tree->ops && tree->ops->merge_bio_hook(page, offset,
2793                                         page_size, bio, bio_flags))
2794                         can_merge = false;
2795
2796                 if (prev_bio_flags != bio_flags || !contig || !can_merge ||
2797                     force_bio_submit ||
2798                     bio_add_page(bio, page, page_size, pg_offset) < page_size) {
2799                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2800                         if (ret < 0) {
2801                                 *bio_ret = NULL;
2802                                 return ret;
2803                         }
2804                         bio = NULL;
2805                 } else {
2806                         if (wbc)
2807                                 wbc_account_io(wbc, page, page_size);
2808                         return 0;
2809                 }
2810         }
2811
2812         bio = btrfs_bio_alloc(bdev, offset);
2813         bio_add_page(bio, page, page_size, pg_offset);
2814         bio->bi_end_io = end_io_func;
2815         bio->bi_private = tree;
2816         bio->bi_write_hint = page->mapping->host->i_write_hint;
2817         bio->bi_opf = opf;
2818         if (wbc) {
2819                 wbc_init_bio(wbc, bio);
2820                 wbc_account_io(wbc, page, page_size);
2821         }
2822
2823         *bio_ret = bio;
2824
2825         return ret;
2826 }
2827
2828 static void attach_extent_buffer_page(struct extent_buffer *eb,
2829                                       struct page *page)
2830 {
2831         if (!PagePrivate(page)) {
2832                 SetPagePrivate(page);
2833                 get_page(page);
2834                 set_page_private(page, (unsigned long)eb);
2835         } else {
2836                 WARN_ON(page->private != (unsigned long)eb);
2837         }
2838 }
2839
2840 void set_page_extent_mapped(struct page *page)
2841 {
2842         if (!PagePrivate(page)) {
2843                 SetPagePrivate(page);
2844                 get_page(page);
2845                 set_page_private(page, EXTENT_PAGE_PRIVATE);
2846         }
2847 }
2848
2849 static struct extent_map *
2850 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2851                  u64 start, u64 len, get_extent_t *get_extent,
2852                  struct extent_map **em_cached)
2853 {
2854         struct extent_map *em;
2855
2856         if (em_cached && *em_cached) {
2857                 em = *em_cached;
2858                 if (extent_map_in_tree(em) && start >= em->start &&
2859                     start < extent_map_end(em)) {
2860                         refcount_inc(&em->refs);
2861                         return em;
2862                 }
2863
2864                 free_extent_map(em);
2865                 *em_cached = NULL;
2866         }
2867
2868         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
2869         if (em_cached && !IS_ERR_OR_NULL(em)) {
2870                 BUG_ON(*em_cached);
2871                 refcount_inc(&em->refs);
2872                 *em_cached = em;
2873         }
2874         return em;
2875 }
2876 /*
2877  * basic readpage implementation.  Locked extent state structs are inserted
2878  * into the tree that are removed when the IO is done (by the end_io
2879  * handlers)
2880  * XXX JDM: This needs looking at to ensure proper page locking
2881  * return 0 on success, otherwise return error
2882  */
2883 static int __do_readpage(struct extent_io_tree *tree,
2884                          struct page *page,
2885                          get_extent_t *get_extent,
2886                          struct extent_map **em_cached,
2887                          struct bio **bio, int mirror_num,
2888                          unsigned long *bio_flags, unsigned int read_flags,
2889                          u64 *prev_em_start)
2890 {
2891         struct inode *inode = page->mapping->host;
2892         u64 start = page_offset(page);
2893         const u64 end = start + PAGE_SIZE - 1;
2894         u64 cur = start;
2895         u64 extent_offset;
2896         u64 last_byte = i_size_read(inode);
2897         u64 block_start;
2898         u64 cur_end;
2899         struct extent_map *em;
2900         struct block_device *bdev;
2901         int ret = 0;
2902         int nr = 0;
2903         size_t pg_offset = 0;
2904         size_t iosize;
2905         size_t disk_io_size;
2906         size_t blocksize = inode->i_sb->s_blocksize;
2907         unsigned long this_bio_flag = 0;
2908
2909         set_page_extent_mapped(page);
2910
2911         if (!PageUptodate(page)) {
2912                 if (cleancache_get_page(page) == 0) {
2913                         BUG_ON(blocksize != PAGE_SIZE);
2914                         unlock_extent(tree, start, end);
2915                         goto out;
2916                 }
2917         }
2918
2919         if (page->index == last_byte >> PAGE_SHIFT) {
2920                 char *userpage;
2921                 size_t zero_offset = last_byte & (PAGE_SIZE - 1);
2922
2923                 if (zero_offset) {
2924                         iosize = PAGE_SIZE - zero_offset;
2925                         userpage = kmap_atomic(page);
2926                         memset(userpage + zero_offset, 0, iosize);
2927                         flush_dcache_page(page);
2928                         kunmap_atomic(userpage);
2929                 }
2930         }
2931         while (cur <= end) {
2932                 bool force_bio_submit = false;
2933                 u64 offset;
2934
2935                 if (cur >= last_byte) {
2936                         char *userpage;
2937                         struct extent_state *cached = NULL;
2938
2939                         iosize = PAGE_SIZE - pg_offset;
2940                         userpage = kmap_atomic(page);
2941                         memset(userpage + pg_offset, 0, iosize);
2942                         flush_dcache_page(page);
2943                         kunmap_atomic(userpage);
2944                         set_extent_uptodate(tree, cur, cur + iosize - 1,
2945                                             &cached, GFP_NOFS);
2946                         unlock_extent_cached(tree, cur,
2947                                              cur + iosize - 1, &cached);
2948                         break;
2949                 }
2950                 em = __get_extent_map(inode, page, pg_offset, cur,
2951                                       end - cur + 1, get_extent, em_cached);
2952                 if (IS_ERR_OR_NULL(em)) {
2953                         SetPageError(page);
2954                         unlock_extent(tree, cur, end);
2955                         break;
2956                 }
2957                 extent_offset = cur - em->start;
2958                 BUG_ON(extent_map_end(em) <= cur);
2959                 BUG_ON(end < cur);
2960
2961                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2962                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
2963                         extent_set_compress_type(&this_bio_flag,
2964                                                  em->compress_type);
2965                 }
2966
2967                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2968                 cur_end = min(extent_map_end(em) - 1, end);
2969                 iosize = ALIGN(iosize, blocksize);
2970                 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2971                         disk_io_size = em->block_len;
2972                         offset = em->block_start;
2973                 } else {
2974                         offset = em->block_start + extent_offset;
2975                         disk_io_size = iosize;
2976                 }
2977                 bdev = em->bdev;
2978                 block_start = em->block_start;
2979                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2980                         block_start = EXTENT_MAP_HOLE;
2981
2982                 /*
2983                  * If we have a file range that points to a compressed extent
2984                  * and it's followed by a consecutive file range that points to
2985                  * to the same compressed extent (possibly with a different
2986                  * offset and/or length, so it either points to the whole extent
2987                  * or only part of it), we must make sure we do not submit a
2988                  * single bio to populate the pages for the 2 ranges because
2989                  * this makes the compressed extent read zero out the pages
2990                  * belonging to the 2nd range. Imagine the following scenario:
2991                  *
2992                  *  File layout
2993                  *  [0 - 8K]                     [8K - 24K]
2994                  *    |                               |
2995                  *    |                               |
2996                  * points to extent X,         points to extent X,
2997                  * offset 4K, length of 8K     offset 0, length 16K
2998                  *
2999                  * [extent X, compressed length = 4K uncompressed length = 16K]
3000                  *
3001                  * If the bio to read the compressed extent covers both ranges,
3002                  * it will decompress extent X into the pages belonging to the
3003                  * first range and then it will stop, zeroing out the remaining
3004                  * pages that belong to the other range that points to extent X.
3005                  * So here we make sure we submit 2 bios, one for the first
3006                  * range and another one for the third range. Both will target
3007                  * the same physical extent from disk, but we can't currently
3008                  * make the compressed bio endio callback populate the pages
3009                  * for both ranges because each compressed bio is tightly
3010                  * coupled with a single extent map, and each range can have
3011                  * an extent map with a different offset value relative to the
3012                  * uncompressed data of our extent and different lengths. This
3013                  * is a corner case so we prioritize correctness over
3014                  * non-optimal behavior (submitting 2 bios for the same extent).
3015                  */
3016                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3017                     prev_em_start && *prev_em_start != (u64)-1 &&
3018                     *prev_em_start != em->orig_start)
3019                         force_bio_submit = true;
3020
3021                 if (prev_em_start)
3022                         *prev_em_start = em->orig_start;
3023
3024                 free_extent_map(em);
3025                 em = NULL;
3026
3027                 /* we've found a hole, just zero and go on */
3028                 if (block_start == EXTENT_MAP_HOLE) {
3029                         char *userpage;
3030                         struct extent_state *cached = NULL;
3031
3032                         userpage = kmap_atomic(page);
3033                         memset(userpage + pg_offset, 0, iosize);
3034                         flush_dcache_page(page);
3035                         kunmap_atomic(userpage);
3036
3037                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3038                                             &cached, GFP_NOFS);
3039                         unlock_extent_cached(tree, cur,
3040                                              cur + iosize - 1, &cached);
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                 ret = submit_extent_page(REQ_OP_READ | read_flags, tree, NULL,
3066                                          page, offset, disk_io_size,
3067                                          pg_offset, bdev, bio,
3068                                          end_bio_extent_readpage, mirror_num,
3069                                          *bio_flags,
3070                                          this_bio_flag,
3071                                          force_bio_submit);
3072                 if (!ret) {
3073                         nr++;
3074                         *bio_flags = this_bio_flag;
3075                 } else {
3076                         SetPageError(page);
3077                         unlock_extent(tree, cur, cur + iosize - 1);
3078                         goto out;
3079                 }
3080                 cur = cur + iosize;
3081                 pg_offset += iosize;
3082         }
3083 out:
3084         if (!nr) {
3085                 if (!PageError(page))
3086                         SetPageUptodate(page);
3087                 unlock_page(page);
3088         }
3089         return ret;
3090 }
3091
3092 static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
3093                                              struct page *pages[], int nr_pages,
3094                                              u64 start, u64 end,
3095                                              struct extent_map **em_cached,
3096                                              struct bio **bio,
3097                                              unsigned long *bio_flags,
3098                                              u64 *prev_em_start)
3099 {
3100         struct inode *inode;
3101         struct btrfs_ordered_extent *ordered;
3102         int index;
3103
3104         inode = pages[0]->mapping->host;
3105         while (1) {
3106                 lock_extent(tree, start, end);
3107                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3108                                                      end - start + 1);
3109                 if (!ordered)
3110                         break;
3111                 unlock_extent(tree, start, end);
3112                 btrfs_start_ordered_extent(inode, ordered, 1);
3113                 btrfs_put_ordered_extent(ordered);
3114         }
3115
3116         for (index = 0; index < nr_pages; index++) {
3117                 __do_readpage(tree, pages[index], btrfs_get_extent, em_cached,
3118                                 bio, 0, bio_flags, 0, prev_em_start);
3119                 put_page(pages[index]);
3120         }
3121 }
3122
3123 static void __extent_readpages(struct extent_io_tree *tree,
3124                                struct page *pages[],
3125                                int nr_pages,
3126                                struct extent_map **em_cached,
3127                                struct bio **bio, unsigned long *bio_flags,
3128                                u64 *prev_em_start)
3129 {
3130         u64 start = 0;
3131         u64 end = 0;
3132         u64 page_start;
3133         int index;
3134         int first_index = 0;
3135
3136         for (index = 0; index < nr_pages; index++) {
3137                 page_start = page_offset(pages[index]);
3138                 if (!end) {
3139                         start = page_start;
3140                         end = start + PAGE_SIZE - 1;
3141                         first_index = index;
3142                 } else if (end + 1 == page_start) {
3143                         end += PAGE_SIZE;
3144                 } else {
3145                         __do_contiguous_readpages(tree, &pages[first_index],
3146                                                   index - first_index, start,
3147                                                   end, em_cached,
3148                                                   bio, bio_flags,
3149                                                   prev_em_start);
3150                         start = page_start;
3151                         end = start + PAGE_SIZE - 1;
3152                         first_index = index;
3153                 }
3154         }
3155
3156         if (end)
3157                 __do_contiguous_readpages(tree, &pages[first_index],
3158                                           index - first_index, start,
3159                                           end, em_cached, bio,
3160                                           bio_flags, prev_em_start);
3161 }
3162
3163 static int __extent_read_full_page(struct extent_io_tree *tree,
3164                                    struct page *page,
3165                                    get_extent_t *get_extent,
3166                                    struct bio **bio, int mirror_num,
3167                                    unsigned long *bio_flags,
3168                                    unsigned int read_flags)
3169 {
3170         struct inode *inode = page->mapping->host;
3171         struct btrfs_ordered_extent *ordered;
3172         u64 start = page_offset(page);
3173         u64 end = start + PAGE_SIZE - 1;
3174         int ret;
3175
3176         while (1) {
3177                 lock_extent(tree, start, end);
3178                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
3179                                                 PAGE_SIZE);
3180                 if (!ordered)
3181                         break;
3182                 unlock_extent(tree, start, end);
3183                 btrfs_start_ordered_extent(inode, ordered, 1);
3184                 btrfs_put_ordered_extent(ordered);
3185         }
3186
3187         ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3188                             bio_flags, read_flags, NULL);
3189         return ret;
3190 }
3191
3192 int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3193                             get_extent_t *get_extent, int mirror_num)
3194 {
3195         struct bio *bio = NULL;
3196         unsigned long bio_flags = 0;
3197         int ret;
3198
3199         ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3200                                       &bio_flags, 0);
3201         if (bio)
3202                 ret = submit_one_bio(bio, mirror_num, bio_flags);
3203         return ret;
3204 }
3205
3206 static void update_nr_written(struct writeback_control *wbc,
3207                               unsigned long nr_written)
3208 {
3209         wbc->nr_to_write -= nr_written;
3210 }
3211
3212 /*
3213  * helper for __extent_writepage, doing all of the delayed allocation setup.
3214  *
3215  * This returns 1 if our fill_delalloc function did all the work required
3216  * to write the page (copy into inline extent).  In this case the IO has
3217  * been started and the page is already unlocked.
3218  *
3219  * This returns 0 if all went well (page still locked)
3220  * This returns < 0 if there were errors (page still locked)
3221  */
3222 static noinline_for_stack int writepage_delalloc(struct inode *inode,
3223                               struct page *page, struct writeback_control *wbc,
3224                               struct extent_page_data *epd,
3225                               u64 delalloc_start,
3226                               unsigned long *nr_written)
3227 {
3228         struct extent_io_tree *tree = epd->tree;
3229         u64 page_end = delalloc_start + PAGE_SIZE - 1;
3230         u64 nr_delalloc;
3231         u64 delalloc_to_write = 0;
3232         u64 delalloc_end = 0;
3233         int ret;
3234         int page_started = 0;
3235
3236         if (epd->extent_locked || !tree->ops || !tree->ops->fill_delalloc)
3237                 return 0;
3238
3239         while (delalloc_end < page_end) {
3240                 nr_delalloc = find_lock_delalloc_range(inode, tree,
3241                                                page,
3242                                                &delalloc_start,
3243                                                &delalloc_end,
3244                                                BTRFS_MAX_EXTENT_SIZE);
3245                 if (nr_delalloc == 0) {
3246                         delalloc_start = delalloc_end + 1;
3247                         continue;
3248                 }
3249                 ret = tree->ops->fill_delalloc(inode, page,
3250                                                delalloc_start,
3251                                                delalloc_end,
3252                                                &page_started,
3253                                                nr_written, wbc);
3254                 /* File system has been set read-only */
3255                 if (ret) {
3256                         SetPageError(page);
3257                         /* fill_delalloc should be return < 0 for error
3258                          * but just in case, we use > 0 here meaning the
3259                          * IO is started, so we don't want to return > 0
3260                          * unless things are going well.
3261                          */
3262                         ret = ret < 0 ? ret : -EIO;
3263                         goto done;
3264                 }
3265                 /*
3266                  * delalloc_end is already one less than the total length, so
3267                  * we don't subtract one from PAGE_SIZE
3268                  */
3269                 delalloc_to_write += (delalloc_end - delalloc_start +
3270                                       PAGE_SIZE) >> PAGE_SHIFT;
3271                 delalloc_start = delalloc_end + 1;
3272         }
3273         if (wbc->nr_to_write < delalloc_to_write) {
3274                 int thresh = 8192;
3275
3276                 if (delalloc_to_write < thresh * 2)
3277                         thresh = delalloc_to_write;
3278                 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3279                                          thresh);
3280         }
3281
3282         /* did the fill delalloc function already unlock and start
3283          * the IO?
3284          */
3285         if (page_started) {
3286                 /*
3287                  * we've unlocked the page, so we can't update
3288                  * the mapping's writeback index, just update
3289                  * nr_to_write.
3290                  */
3291                 wbc->nr_to_write -= *nr_written;
3292                 return 1;
3293         }
3294
3295         ret = 0;
3296
3297 done:
3298         return ret;
3299 }
3300
3301 /*
3302  * helper for __extent_writepage.  This calls the writepage start hooks,
3303  * and does the loop to map the page into extents and bios.
3304  *
3305  * We return 1 if the IO is started and the page is unlocked,
3306  * 0 if all went well (page still locked)
3307  * < 0 if there were errors (page still locked)
3308  */
3309 static noinline_for_stack int __extent_writepage_io(struct inode *inode,
3310                                  struct page *page,
3311                                  struct writeback_control *wbc,
3312                                  struct extent_page_data *epd,
3313                                  loff_t i_size,
3314                                  unsigned long nr_written,
3315                                  unsigned int write_flags, int *nr_ret)
3316 {
3317         struct extent_io_tree *tree = epd->tree;
3318         u64 start = page_offset(page);
3319         u64 page_end = start + PAGE_SIZE - 1;
3320         u64 end;
3321         u64 cur = start;
3322         u64 extent_offset;
3323         u64 block_start;
3324         u64 iosize;
3325         struct extent_map *em;
3326         struct block_device *bdev;
3327         size_t pg_offset = 0;
3328         size_t blocksize;
3329         int ret = 0;
3330         int nr = 0;
3331         bool compressed;
3332
3333         if (tree->ops && tree->ops->writepage_start_hook) {
3334                 ret = tree->ops->writepage_start_hook(page, start,
3335                                                       page_end);
3336                 if (ret) {
3337                         /* Fixup worker will requeue */
3338                         if (ret == -EBUSY)
3339                                 wbc->pages_skipped++;
3340                         else
3341                                 redirty_page_for_writepage(wbc, page);
3342
3343                         update_nr_written(wbc, nr_written);
3344                         unlock_page(page);
3345                         return 1;
3346                 }
3347         }
3348
3349         /*
3350          * we don't want to touch the inode after unlocking the page,
3351          * so we update the mapping writeback index now
3352          */
3353         update_nr_written(wbc, nr_written + 1);
3354
3355         end = page_end;
3356         if (i_size <= start) {
3357                 if (tree->ops && tree->ops->writepage_end_io_hook)
3358                         tree->ops->writepage_end_io_hook(page, start,
3359                                                          page_end, NULL, 1);
3360                 goto done;
3361         }
3362
3363         blocksize = inode->i_sb->s_blocksize;
3364
3365         while (cur <= end) {
3366                 u64 em_end;
3367                 u64 offset;
3368
3369                 if (cur >= i_size) {
3370                         if (tree->ops && tree->ops->writepage_end_io_hook)
3371                                 tree->ops->writepage_end_io_hook(page, cur,
3372                                                          page_end, NULL, 1);
3373                         break;
3374                 }
3375                 em = btrfs_get_extent(BTRFS_I(inode), page, pg_offset, cur,
3376                                      end - cur + 1, 1);
3377                 if (IS_ERR_OR_NULL(em)) {
3378                         SetPageError(page);
3379                         ret = PTR_ERR_OR_ZERO(em);
3380                         break;
3381                 }
3382
3383                 extent_offset = cur - em->start;
3384                 em_end = extent_map_end(em);
3385                 BUG_ON(em_end <= cur);
3386                 BUG_ON(end < cur);
3387                 iosize = min(em_end - cur, end - cur + 1);
3388                 iosize = ALIGN(iosize, blocksize);
3389                 offset = em->block_start + extent_offset;
3390                 bdev = em->bdev;
3391                 block_start = em->block_start;
3392                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3393                 free_extent_map(em);
3394                 em = NULL;
3395
3396                 /*
3397                  * compressed and inline extents are written through other
3398                  * paths in the FS
3399                  */
3400                 if (compressed || block_start == EXTENT_MAP_HOLE ||
3401                     block_start == EXTENT_MAP_INLINE) {
3402                         /*
3403                          * end_io notification does not happen here for
3404                          * compressed extents
3405                          */
3406                         if (!compressed && tree->ops &&
3407                             tree->ops->writepage_end_io_hook)
3408                                 tree->ops->writepage_end_io_hook(page, cur,
3409                                                          cur + iosize - 1,
3410                                                          NULL, 1);
3411                         else if (compressed) {
3412                                 /* we don't want to end_page_writeback on
3413                                  * a compressed extent.  this happens
3414                                  * elsewhere
3415                                  */
3416                                 nr++;
3417                         }
3418
3419                         cur += iosize;
3420                         pg_offset += iosize;
3421                         continue;
3422                 }
3423
3424                 set_range_writeback(tree, cur, cur + iosize - 1);
3425                 if (!PageWriteback(page)) {
3426                         btrfs_err(BTRFS_I(inode)->root->fs_info,
3427                                    "page %lu not writeback, cur %llu end %llu",
3428                                page->index, cur, end);
3429                 }
3430
3431                 ret = submit_extent_page(REQ_OP_WRITE | write_flags, tree, wbc,
3432                                          page, offset, iosize, pg_offset,
3433                                          bdev, &epd->bio,
3434                                          end_bio_extent_writepage,
3435                                          0, 0, 0, false);
3436                 if (ret) {
3437                         SetPageError(page);
3438                         if (PageWriteback(page))
3439                                 end_page_writeback(page);
3440                 }
3441
3442                 cur = cur + iosize;
3443                 pg_offset += iosize;
3444                 nr++;
3445         }
3446 done:
3447         *nr_ret = nr;
3448         return ret;
3449 }
3450
3451 /*
3452  * the writepage semantics are similar to regular writepage.  extent
3453  * records are inserted to lock ranges in the tree, and as dirty areas
3454  * are found, they are marked writeback.  Then the lock bits are removed
3455  * and the end_io handler clears the writeback ranges
3456  */
3457 static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3458                               struct extent_page_data *epd)
3459 {
3460         struct inode *inode = page->mapping->host;
3461         u64 start = page_offset(page);
3462         u64 page_end = start + PAGE_SIZE - 1;
3463         int ret;
3464         int nr = 0;
3465         size_t pg_offset = 0;
3466         loff_t i_size = i_size_read(inode);
3467         unsigned long end_index = i_size >> PAGE_SHIFT;
3468         unsigned int write_flags = 0;
3469         unsigned long nr_written = 0;
3470
3471         write_flags = wbc_to_write_flags(wbc);
3472
3473         trace___extent_writepage(page, inode, wbc);
3474
3475         WARN_ON(!PageLocked(page));
3476
3477         ClearPageError(page);
3478
3479         pg_offset = i_size & (PAGE_SIZE - 1);
3480         if (page->index > end_index ||
3481            (page->index == end_index && !pg_offset)) {
3482                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3483                 unlock_page(page);
3484                 return 0;
3485         }
3486
3487         if (page->index == end_index) {
3488                 char *userpage;
3489
3490                 userpage = kmap_atomic(page);
3491                 memset(userpage + pg_offset, 0,
3492                        PAGE_SIZE - pg_offset);
3493                 kunmap_atomic(userpage);
3494                 flush_dcache_page(page);
3495         }
3496
3497         pg_offset = 0;
3498
3499         set_page_extent_mapped(page);
3500
3501         ret = writepage_delalloc(inode, page, wbc, epd, start, &nr_written);
3502         if (ret == 1)
3503                 goto done_unlocked;
3504         if (ret)
3505                 goto done;
3506
3507         ret = __extent_writepage_io(inode, page, wbc, epd,
3508                                     i_size, nr_written, write_flags, &nr);
3509         if (ret == 1)
3510                 goto done_unlocked;
3511
3512 done:
3513         if (nr == 0) {
3514                 /* make sure the mapping tag for page dirty gets cleared */
3515                 set_page_writeback(page);
3516                 end_page_writeback(page);
3517         }
3518         if (PageError(page)) {
3519                 ret = ret < 0 ? ret : -EIO;
3520                 end_extent_writepage(page, ret, start, page_end);
3521         }
3522         unlock_page(page);
3523         return ret;
3524
3525 done_unlocked:
3526         return 0;
3527 }
3528
3529 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3530 {
3531         wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3532                        TASK_UNINTERRUPTIBLE);
3533 }
3534
3535 static noinline_for_stack int
3536 lock_extent_buffer_for_io(struct extent_buffer *eb,
3537                           struct btrfs_fs_info *fs_info,
3538                           struct extent_page_data *epd)
3539 {
3540         unsigned long i, num_pages;
3541         int flush = 0;
3542         int ret = 0;
3543
3544         if (!btrfs_try_tree_write_lock(eb)) {
3545                 flush = 1;
3546                 flush_write_bio(epd);
3547                 btrfs_tree_lock(eb);
3548         }
3549
3550         if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3551                 btrfs_tree_unlock(eb);
3552                 if (!epd->sync_io)
3553                         return 0;
3554                 if (!flush) {
3555                         flush_write_bio(epd);
3556                         flush = 1;
3557                 }
3558                 while (1) {
3559                         wait_on_extent_buffer_writeback(eb);
3560                         btrfs_tree_lock(eb);
3561                         if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3562                                 break;
3563                         btrfs_tree_unlock(eb);
3564                 }
3565         }
3566
3567         /*
3568          * We need to do this to prevent races in people who check if the eb is
3569          * under IO since we can end up having no IO bits set for a short period
3570          * of time.
3571          */
3572         spin_lock(&eb->refs_lock);
3573         if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3574                 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3575                 spin_unlock(&eb->refs_lock);
3576                 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3577                 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
3578                                          -eb->len,
3579                                          fs_info->dirty_metadata_batch);
3580                 ret = 1;
3581         } else {
3582                 spin_unlock(&eb->refs_lock);
3583         }
3584
3585         btrfs_tree_unlock(eb);
3586
3587         if (!ret)
3588                 return ret;
3589
3590         num_pages = num_extent_pages(eb->start, eb->len);
3591         for (i = 0; i < num_pages; i++) {
3592                 struct page *p = eb->pages[i];
3593
3594                 if (!trylock_page(p)) {
3595                         if (!flush) {
3596                                 flush_write_bio(epd);
3597                                 flush = 1;
3598                         }
3599                         lock_page(p);
3600                 }
3601         }
3602
3603         return ret;
3604 }
3605
3606 static void end_extent_buffer_writeback(struct extent_buffer *eb)
3607 {
3608         clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3609         smp_mb__after_atomic();
3610         wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3611 }
3612
3613 static void set_btree_ioerr(struct page *page)
3614 {
3615         struct extent_buffer *eb = (struct extent_buffer *)page->private;
3616
3617         SetPageError(page);
3618         if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3619                 return;
3620
3621         /*
3622          * If writeback for a btree extent that doesn't belong to a log tree
3623          * failed, increment the counter transaction->eb_write_errors.
3624          * We do this because while the transaction is running and before it's
3625          * committing (when we call filemap_fdata[write|wait]_range against
3626          * the btree inode), we might have
3627          * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3628          * returns an error or an error happens during writeback, when we're
3629          * committing the transaction we wouldn't know about it, since the pages
3630          * can be no longer dirty nor marked anymore for writeback (if a
3631          * subsequent modification to the extent buffer didn't happen before the
3632          * transaction commit), which makes filemap_fdata[write|wait]_range not
3633          * able to find the pages tagged with SetPageError at transaction
3634          * commit time. So if this happens we must abort the transaction,
3635          * otherwise we commit a super block with btree roots that point to
3636          * btree nodes/leafs whose content on disk is invalid - either garbage
3637          * or the content of some node/leaf from a past generation that got
3638          * cowed or deleted and is no longer valid.
3639          *
3640          * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3641          * not be enough - we need to distinguish between log tree extents vs
3642          * non-log tree extents, and the next filemap_fdatawait_range() call
3643          * will catch and clear such errors in the mapping - and that call might
3644          * be from a log sync and not from a transaction commit. Also, checking
3645          * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3646          * not done and would not be reliable - the eb might have been released
3647          * from memory and reading it back again means that flag would not be
3648          * set (since it's a runtime flag, not persisted on disk).
3649          *
3650          * Using the flags below in the btree inode also makes us achieve the
3651          * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3652          * writeback for all dirty pages and before filemap_fdatawait_range()
3653          * is called, the writeback for all dirty pages had already finished
3654          * with errors - because we were not using AS_EIO/AS_ENOSPC,
3655          * filemap_fdatawait_range() would return success, as it could not know
3656          * that writeback errors happened (the pages were no longer tagged for
3657          * writeback).
3658          */
3659         switch (eb->log_index) {
3660         case -1:
3661                 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3662                 break;
3663         case 0:
3664                 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3665                 break;
3666         case 1:
3667                 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3668                 break;
3669         default:
3670                 BUG(); /* unexpected, logic error */
3671         }
3672 }
3673
3674 static void end_bio_extent_buffer_writepage(struct bio *bio)
3675 {
3676         struct bio_vec *bvec;
3677         struct extent_buffer *eb;
3678         int i, done;
3679
3680         ASSERT(!bio_flagged(bio, BIO_CLONED));
3681         bio_for_each_segment_all(bvec, bio, i) {
3682                 struct page *page = bvec->bv_page;
3683
3684                 eb = (struct extent_buffer *)page->private;
3685                 BUG_ON(!eb);
3686                 done = atomic_dec_and_test(&eb->io_pages);
3687
3688                 if (bio->bi_status ||
3689                     test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3690                         ClearPageUptodate(page);
3691                         set_btree_ioerr(page);
3692                 }
3693
3694                 end_page_writeback(page);
3695
3696                 if (!done)
3697                         continue;
3698
3699                 end_extent_buffer_writeback(eb);
3700         }
3701
3702         bio_put(bio);
3703 }
3704
3705 static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3706                         struct btrfs_fs_info *fs_info,
3707                         struct writeback_control *wbc,
3708                         struct extent_page_data *epd)
3709 {
3710         struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3711         struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3712         u64 offset = eb->start;
3713         u32 nritems;
3714         unsigned long i, num_pages;
3715         unsigned long start, end;
3716         unsigned int write_flags = wbc_to_write_flags(wbc) | REQ_META;
3717         int ret = 0;
3718
3719         clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3720         num_pages = num_extent_pages(eb->start, eb->len);
3721         atomic_set(&eb->io_pages, num_pages);
3722
3723         /* set btree blocks beyond nritems with 0 to avoid stale content. */
3724         nritems = btrfs_header_nritems(eb);
3725         if (btrfs_header_level(eb) > 0) {
3726                 end = btrfs_node_key_ptr_offset(nritems);
3727
3728                 memzero_extent_buffer(eb, end, eb->len - end);
3729         } else {
3730                 /*
3731                  * leaf:
3732                  * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3733                  */
3734                 start = btrfs_item_nr_offset(nritems);
3735                 end = BTRFS_LEAF_DATA_OFFSET + leaf_data_end(fs_info, eb);
3736                 memzero_extent_buffer(eb, start, end - start);
3737         }
3738
3739         for (i = 0; i < num_pages; i++) {
3740                 struct page *p = eb->pages[i];
3741
3742                 clear_page_dirty_for_io(p);
3743                 set_page_writeback(p);
3744                 ret = submit_extent_page(REQ_OP_WRITE | write_flags, tree, wbc,
3745                                          p, offset, PAGE_SIZE, 0, bdev,
3746                                          &epd->bio,
3747                                          end_bio_extent_buffer_writepage,
3748                                          0, 0, 0, false);
3749                 if (ret) {
3750                         set_btree_ioerr(p);
3751                         if (PageWriteback(p))
3752                                 end_page_writeback(p);
3753                         if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3754                                 end_extent_buffer_writeback(eb);
3755                         ret = -EIO;
3756                         break;
3757                 }
3758                 offset += PAGE_SIZE;
3759                 update_nr_written(wbc, 1);
3760                 unlock_page(p);
3761         }
3762
3763         if (unlikely(ret)) {
3764                 for (; i < num_pages; i++) {
3765                         struct page *p = eb->pages[i];
3766                         clear_page_dirty_for_io(p);
3767                         unlock_page(p);
3768                 }
3769         }
3770
3771         return ret;
3772 }
3773
3774 int btree_write_cache_pages(struct address_space *mapping,
3775                                    struct writeback_control *wbc)
3776 {
3777         struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3778         struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3779         struct extent_buffer *eb, *prev_eb = NULL;
3780         struct extent_page_data epd = {
3781                 .bio = NULL,
3782                 .tree = tree,
3783                 .extent_locked = 0,
3784                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3785         };
3786         int ret = 0;
3787         int done = 0;
3788         int nr_to_write_done = 0;
3789         struct pagevec pvec;
3790         int nr_pages;
3791         pgoff_t index;
3792         pgoff_t end;            /* Inclusive */
3793         int scanned = 0;
3794         int tag;
3795
3796         pagevec_init(&pvec);
3797         if (wbc->range_cyclic) {
3798                 index = mapping->writeback_index; /* Start from prev offset */
3799                 end = -1;
3800         } else {
3801                 index = wbc->range_start >> PAGE_SHIFT;
3802                 end = wbc->range_end >> PAGE_SHIFT;
3803                 scanned = 1;
3804         }
3805         if (wbc->sync_mode == WB_SYNC_ALL)
3806                 tag = PAGECACHE_TAG_TOWRITE;
3807         else
3808                 tag = PAGECACHE_TAG_DIRTY;
3809 retry:
3810         if (wbc->sync_mode == WB_SYNC_ALL)
3811                 tag_pages_for_writeback(mapping, index, end);
3812         while (!done && !nr_to_write_done && (index <= end) &&
3813                (nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
3814                         tag))) {
3815                 unsigned i;
3816
3817                 scanned = 1;
3818                 for (i = 0; i < nr_pages; i++) {
3819                         struct page *page = pvec.pages[i];
3820
3821                         if (!PagePrivate(page))
3822                                 continue;
3823
3824                         spin_lock(&mapping->private_lock);
3825                         if (!PagePrivate(page)) {
3826                                 spin_unlock(&mapping->private_lock);
3827                                 continue;
3828                         }
3829
3830                         eb = (struct extent_buffer *)page->private;
3831
3832                         /*
3833                          * Shouldn't happen and normally this would be a BUG_ON
3834                          * but no sense in crashing the users box for something
3835                          * we can survive anyway.
3836                          */
3837                         if (WARN_ON(!eb)) {
3838                                 spin_unlock(&mapping->private_lock);
3839                                 continue;
3840                         }
3841
3842                         if (eb == prev_eb) {
3843                                 spin_unlock(&mapping->private_lock);
3844                                 continue;
3845                         }
3846
3847                         ret = atomic_inc_not_zero(&eb->refs);
3848                         spin_unlock(&mapping->private_lock);
3849                         if (!ret)
3850                                 continue;
3851
3852                         prev_eb = eb;
3853                         ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3854                         if (!ret) {
3855                                 free_extent_buffer(eb);
3856                                 continue;
3857                         }
3858
3859                         ret = write_one_eb(eb, fs_info, wbc, &epd);
3860                         if (ret) {
3861                                 done = 1;
3862                                 free_extent_buffer(eb);
3863                                 break;
3864                         }
3865                         free_extent_buffer(eb);
3866
3867                         /*
3868                          * the filesystem may choose to bump up nr_to_write.
3869                          * We have to make sure to honor the new nr_to_write
3870                          * at any time
3871                          */
3872                         nr_to_write_done = wbc->nr_to_write <= 0;
3873                 }
3874                 pagevec_release(&pvec);
3875                 cond_resched();
3876         }
3877         if (!scanned && !done) {
3878                 /*
3879                  * We hit the last page and there is more work to be done: wrap
3880                  * back to the start of the file
3881                  */
3882                 scanned = 1;
3883                 index = 0;
3884                 goto retry;
3885         }
3886         flush_write_bio(&epd);
3887         return ret;
3888 }
3889
3890 /**
3891  * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3892  * @mapping: address space structure to write
3893  * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3894  * @data: data passed to __extent_writepage function
3895  *
3896  * If a page is already under I/O, write_cache_pages() skips it, even
3897  * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
3898  * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
3899  * and msync() need to guarantee that all the data which was dirty at the time
3900  * the call was made get new I/O started against them.  If wbc->sync_mode is
3901  * WB_SYNC_ALL then we were called for data integrity and we must wait for
3902  * existing IO to complete.
3903  */
3904 static int extent_write_cache_pages(struct address_space *mapping,
3905                              struct writeback_control *wbc,
3906                              struct extent_page_data *epd)
3907 {
3908         struct inode *inode = mapping->host;
3909         int ret = 0;
3910         int done = 0;
3911         int nr_to_write_done = 0;
3912         struct pagevec pvec;
3913         int nr_pages;
3914         pgoff_t index;
3915         pgoff_t end;            /* Inclusive */
3916         pgoff_t done_index;
3917         int range_whole = 0;
3918         int scanned = 0;
3919         int tag;
3920
3921         /*
3922          * We have to hold onto the inode so that ordered extents can do their
3923          * work when the IO finishes.  The alternative to this is failing to add
3924          * an ordered extent if the igrab() fails there and that is a huge pain
3925          * to deal with, so instead just hold onto the inode throughout the
3926          * writepages operation.  If it fails here we are freeing up the inode
3927          * anyway and we'd rather not waste our time writing out stuff that is
3928          * going to be truncated anyway.
3929          */
3930         if (!igrab(inode))
3931                 return 0;
3932
3933         pagevec_init(&pvec);
3934         if (wbc->range_cyclic) {
3935                 index = mapping->writeback_index; /* Start from prev offset */
3936                 end = -1;
3937         } else {
3938                 index = wbc->range_start >> PAGE_SHIFT;
3939                 end = wbc->range_end >> PAGE_SHIFT;
3940                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3941                         range_whole = 1;
3942                 scanned = 1;
3943         }
3944         if (wbc->sync_mode == WB_SYNC_ALL)
3945                 tag = PAGECACHE_TAG_TOWRITE;
3946         else
3947                 tag = PAGECACHE_TAG_DIRTY;
3948 retry:
3949         if (wbc->sync_mode == WB_SYNC_ALL)
3950                 tag_pages_for_writeback(mapping, index, end);
3951         done_index = index;
3952         while (!done && !nr_to_write_done && (index <= end) &&
3953                         (nr_pages = pagevec_lookup_range_tag(&pvec, mapping,
3954                                                 &index, end, tag))) {
3955                 unsigned i;
3956
3957                 scanned = 1;
3958                 for (i = 0; i < nr_pages; i++) {
3959                         struct page *page = pvec.pages[i];
3960
3961                         done_index = page->index;
3962                         /*
3963                          * At this point we hold neither mapping->tree_lock nor
3964                          * lock on the page itself: the page may be truncated or
3965                          * invalidated (changing page->mapping to NULL), or even
3966                          * swizzled back from swapper_space to tmpfs file
3967                          * mapping
3968                          */
3969                         if (!trylock_page(page)) {
3970                                 flush_write_bio(epd);
3971                                 lock_page(page);
3972                         }
3973
3974                         if (unlikely(page->mapping != mapping)) {
3975                                 unlock_page(page);
3976                                 continue;
3977                         }
3978
3979                         if (wbc->sync_mode != WB_SYNC_NONE) {
3980                                 if (PageWriteback(page))
3981                                         flush_write_bio(epd);
3982                                 wait_on_page_writeback(page);
3983                         }
3984
3985                         if (PageWriteback(page) ||
3986                             !clear_page_dirty_for_io(page)) {
3987                                 unlock_page(page);
3988                                 continue;
3989                         }
3990
3991                         ret = __extent_writepage(page, wbc, epd);
3992
3993                         if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3994                                 unlock_page(page);
3995                                 ret = 0;
3996                         }
3997                         if (ret < 0) {
3998                                 /*
3999                                  * done_index is set past this page,
4000                                  * so media errors will not choke
4001                                  * background writeout for the entire
4002                                  * file. This has consequences for
4003                                  * range_cyclic semantics (ie. it may
4004                                  * not be suitable for data integrity
4005                                  * writeout).
4006                                  */
4007                                 done_index = page->index + 1;
4008                                 done = 1;
4009                                 break;
4010                         }
4011
4012                         /*
4013                          * the filesystem may choose to bump up nr_to_write.
4014                          * We have to make sure to honor the new nr_to_write
4015                          * at any time
4016                          */
4017                         nr_to_write_done = wbc->nr_to_write <= 0;
4018                 }
4019                 pagevec_release(&pvec);
4020                 cond_resched();
4021         }
4022         if (!scanned && !done) {
4023                 /*
4024                  * We hit the last page and there is more work to be done: wrap
4025                  * back to the start of the file
4026                  */
4027                 scanned = 1;
4028                 index = 0;
4029                 goto retry;
4030         }
4031
4032         if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4033                 mapping->writeback_index = done_index;
4034
4035         btrfs_add_delayed_iput(inode);
4036         return ret;
4037 }
4038
4039 static void flush_write_bio(struct extent_page_data *epd)
4040 {
4041         if (epd->bio) {
4042                 int ret;
4043
4044                 ret = submit_one_bio(epd->bio, 0, 0);
4045                 BUG_ON(ret < 0); /* -ENOMEM */
4046                 epd->bio = NULL;
4047         }
4048 }
4049
4050 int extent_write_full_page(struct page *page, struct writeback_control *wbc)
4051 {
4052         int ret;
4053         struct extent_page_data epd = {
4054                 .bio = NULL,
4055                 .tree = &BTRFS_I(page->mapping->host)->io_tree,
4056                 .extent_locked = 0,
4057                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4058         };
4059
4060         ret = __extent_writepage(page, wbc, &epd);
4061
4062         flush_write_bio(&epd);
4063         return ret;
4064 }
4065
4066 int extent_write_locked_range(struct inode *inode, u64 start, u64 end,
4067                               int mode)
4068 {
4069         int ret = 0;
4070         struct address_space *mapping = inode->i_mapping;
4071         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
4072         struct page *page;
4073         unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4074                 PAGE_SHIFT;
4075
4076         struct extent_page_data epd = {
4077                 .bio = NULL,
4078                 .tree = tree,
4079                 .extent_locked = 1,
4080                 .sync_io = mode == WB_SYNC_ALL,
4081         };
4082         struct writeback_control wbc_writepages = {
4083                 .sync_mode      = mode,
4084                 .nr_to_write    = nr_pages * 2,
4085                 .range_start    = start,
4086                 .range_end      = end + 1,
4087         };
4088
4089         while (start <= end) {
4090                 page = find_get_page(mapping, start >> PAGE_SHIFT);
4091                 if (clear_page_dirty_for_io(page))
4092                         ret = __extent_writepage(page, &wbc_writepages, &epd);
4093                 else {
4094                         if (tree->ops && tree->ops->writepage_end_io_hook)
4095                                 tree->ops->writepage_end_io_hook(page, start,
4096                                                  start + PAGE_SIZE - 1,
4097                                                  NULL, 1);
4098                         unlock_page(page);
4099                 }
4100                 put_page(page);
4101                 start += PAGE_SIZE;
4102         }
4103
4104         flush_write_bio(&epd);
4105         return ret;
4106 }
4107
4108 int extent_writepages(struct extent_io_tree *tree,
4109                       struct address_space *mapping,
4110                       struct writeback_control *wbc)
4111 {
4112         int ret = 0;
4113         struct extent_page_data epd = {
4114                 .bio = NULL,
4115                 .tree = tree,
4116                 .extent_locked = 0,
4117                 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4118         };
4119
4120         ret = extent_write_cache_pages(mapping, wbc, &epd);
4121         flush_write_bio(&epd);
4122         return ret;
4123 }
4124
4125 int extent_readpages(struct extent_io_tree *tree,
4126                      struct address_space *mapping,
4127                      struct list_head *pages, unsigned nr_pages)
4128 {
4129         struct bio *bio = NULL;
4130         unsigned page_idx;
4131         unsigned long bio_flags = 0;
4132         struct page *pagepool[16];
4133         struct page *page;
4134         struct extent_map *em_cached = NULL;
4135         int nr = 0;
4136         u64 prev_em_start = (u64)-1;
4137
4138         for (page_idx = 0; page_idx < nr_pages; page_idx++) {
4139                 page = list_entry(pages->prev, struct page, lru);
4140
4141                 prefetchw(&page->flags);
4142                 list_del(&page->lru);
4143                 if (add_to_page_cache_lru(page, mapping,
4144                                         page->index,
4145                                         readahead_gfp_mask(mapping))) {
4146                         put_page(page);
4147                         continue;
4148                 }
4149
4150                 pagepool[nr++] = page;
4151                 if (nr < ARRAY_SIZE(pagepool))
4152                         continue;
4153                 __extent_readpages(tree, pagepool, nr, &em_cached, &bio,
4154                                 &bio_flags, &prev_em_start);
4155                 nr = 0;
4156         }
4157         if (nr)
4158                 __extent_readpages(tree, pagepool, nr, &em_cached, &bio,
4159                                 &bio_flags, &prev_em_start);
4160
4161         if (em_cached)
4162                 free_extent_map(em_cached);
4163
4164         BUG_ON(!list_empty(pages));
4165         if (bio)
4166                 return submit_one_bio(bio, 0, bio_flags);
4167         return 0;
4168 }
4169
4170 /*
4171  * basic invalidatepage code, this waits on any locked or writeback
4172  * ranges corresponding to the page, and then deletes any extent state
4173  * records from the tree
4174  */
4175 int extent_invalidatepage(struct extent_io_tree *tree,
4176                           struct page *page, unsigned long offset)
4177 {
4178         struct extent_state *cached_state = NULL;
4179         u64 start = page_offset(page);
4180         u64 end = start + PAGE_SIZE - 1;
4181         size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4182
4183         start += ALIGN(offset, blocksize);
4184         if (start > end)
4185                 return 0;
4186
4187         lock_extent_bits(tree, start, end, &cached_state);
4188         wait_on_page_writeback(page);
4189         clear_extent_bit(tree, start, end,
4190                          EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
4191                          EXTENT_DO_ACCOUNTING,
4192                          1, 1, &cached_state);
4193         return 0;
4194 }
4195
4196 /*
4197  * a helper for releasepage, this tests for areas of the page that
4198  * are locked or under IO and drops the related state bits if it is safe
4199  * to drop the page.
4200  */
4201 static int try_release_extent_state(struct extent_map_tree *map,
4202                                     struct extent_io_tree *tree,
4203                                     struct page *page, gfp_t mask)
4204 {
4205         u64 start = page_offset(page);
4206         u64 end = start + PAGE_SIZE - 1;
4207         int ret = 1;
4208
4209         if (test_range_bit(tree, start, end,
4210                            EXTENT_IOBITS, 0, NULL))
4211                 ret = 0;
4212         else {
4213                 /*
4214                  * at this point we can safely clear everything except the
4215                  * locked bit and the nodatasum bit
4216                  */
4217                 ret = __clear_extent_bit(tree, start, end,
4218                                  ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4219                                  0, 0, NULL, mask, NULL);
4220
4221                 /* if clear_extent_bit failed for enomem reasons,
4222                  * we can't allow the release to continue.
4223                  */
4224                 if (ret < 0)
4225                         ret = 0;
4226                 else
4227                         ret = 1;
4228         }
4229         return ret;
4230 }
4231
4232 /*
4233  * a helper for releasepage.  As long as there are no locked extents
4234  * in the range corresponding to the page, both state records and extent
4235  * map records are removed
4236  */
4237 int try_release_extent_mapping(struct extent_map_tree *map,
4238                                struct extent_io_tree *tree, struct page *page,
4239                                gfp_t mask)
4240 {
4241         struct extent_map *em;
4242         u64 start = page_offset(page);
4243         u64 end = start + PAGE_SIZE - 1;
4244
4245         if (gfpflags_allow_blocking(mask) &&
4246             page->mapping->host->i_size > SZ_16M) {
4247                 u64 len;
4248                 while (start <= end) {
4249                         len = end - start + 1;
4250                         write_lock(&map->lock);
4251                         em = lookup_extent_mapping(map, start, len);
4252                         if (!em) {
4253                                 write_unlock(&map->lock);
4254                                 break;
4255                         }
4256                         if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4257                             em->start != start) {
4258                                 write_unlock(&map->lock);
4259                                 free_extent_map(em);
4260                                 break;
4261                         }
4262                         if (!test_range_bit(tree, em->start,
4263                                             extent_map_end(em) - 1,
4264                                             EXTENT_LOCKED | EXTENT_WRITEBACK,
4265                                             0, NULL)) {
4266                                 remove_extent_mapping(map, em);
4267                                 /* once for the rb tree */
4268                                 free_extent_map(em);
4269                         }
4270                         start = extent_map_end(em);
4271                         write_unlock(&map->lock);
4272
4273                         /* once for us */
4274                         free_extent_map(em);
4275                 }
4276         }
4277         return try_release_extent_state(map, tree, page, mask);
4278 }
4279
4280 /*
4281  * helper function for fiemap, which doesn't want to see any holes.
4282  * This maps until we find something past 'last'
4283  */
4284 static struct extent_map *get_extent_skip_holes(struct inode *inode,
4285                                                 u64 offset, u64 last)
4286 {
4287         u64 sectorsize = btrfs_inode_sectorsize(inode);
4288         struct extent_map *em;
4289         u64 len;
4290
4291         if (offset >= last)
4292                 return NULL;
4293
4294         while (1) {
4295                 len = last - offset;
4296                 if (len == 0)
4297                         break;
4298                 len = ALIGN(len, sectorsize);
4299                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0, offset,
4300                                 len, 0);
4301                 if (IS_ERR_OR_NULL(em))
4302                         return em;
4303
4304                 /* if this isn't a hole return it */
4305                 if (em->block_start != EXTENT_MAP_HOLE)
4306                         return em;
4307
4308                 /* this is a hole, advance to the next extent */
4309                 offset = extent_map_end(em);
4310                 free_extent_map(em);
4311                 if (offset >= last)
4312                         break;
4313         }
4314         return NULL;
4315 }
4316
4317 /*
4318  * To cache previous fiemap extent
4319  *
4320  * Will be used for merging fiemap extent
4321  */
4322 struct fiemap_cache {
4323         u64 offset;
4324         u64 phys;
4325         u64 len;
4326         u32 flags;
4327         bool cached;
4328 };
4329
4330 /*
4331  * Helper to submit fiemap extent.
4332  *
4333  * Will try to merge current fiemap extent specified by @offset, @phys,
4334  * @len and @flags with cached one.
4335  * And only when we fails to merge, cached one will be submitted as
4336  * fiemap extent.
4337  *
4338  * Return value is the same as fiemap_fill_next_extent().
4339  */
4340 static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
4341                                 struct fiemap_cache *cache,
4342                                 u64 offset, u64 phys, u64 len, u32 flags)
4343 {
4344         int ret = 0;
4345
4346         if (!cache->cached)
4347                 goto assign;
4348
4349         /*
4350          * Sanity check, extent_fiemap() should have ensured that new
4351          * fiemap extent won't overlap with cahced one.
4352          * Not recoverable.
4353          *
4354          * NOTE: Physical address can overlap, due to compression
4355          */
4356         if (cache->offset + cache->len > offset) {
4357                 WARN_ON(1);
4358                 return -EINVAL;
4359         }
4360
4361         /*
4362          * Only merges fiemap extents if
4363          * 1) Their logical addresses are continuous
4364          *
4365          * 2) Their physical addresses are continuous
4366          *    So truly compressed (physical size smaller than logical size)
4367          *    extents won't get merged with each other
4368          *
4369          * 3) Share same flags except FIEMAP_EXTENT_LAST
4370          *    So regular extent won't get merged with prealloc extent
4371          */
4372         if (cache->offset + cache->len  == offset &&
4373             cache->phys + cache->len == phys  &&
4374             (cache->flags & ~FIEMAP_EXTENT_LAST) ==
4375                         (flags & ~FIEMAP_EXTENT_LAST)) {
4376                 cache->len += len;
4377                 cache->flags |= flags;
4378                 goto try_submit_last;
4379         }
4380
4381         /* Not mergeable, need to submit cached one */
4382         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4383                                       cache->len, cache->flags);
4384         cache->cached = false;
4385         if (ret)
4386                 return ret;
4387 assign:
4388         cache->cached = true;
4389         cache->offset = offset;
4390         cache->phys = phys;
4391         cache->len = len;
4392         cache->flags = flags;
4393 try_submit_last:
4394         if (cache->flags & FIEMAP_EXTENT_LAST) {
4395                 ret = fiemap_fill_next_extent(fieinfo, cache->offset,
4396                                 cache->phys, cache->len, cache->flags);
4397                 cache->cached = false;
4398         }
4399         return ret;
4400 }
4401
4402 /*
4403  * Emit last fiemap cache
4404  *
4405  * The last fiemap cache may still be cached in the following case:
4406  * 0                  4k                    8k
4407  * |<- Fiemap range ->|
4408  * |<------------  First extent ----------->|
4409  *
4410  * In this case, the first extent range will be cached but not emitted.
4411  * So we must emit it before ending extent_fiemap().
4412  */
4413 static int emit_last_fiemap_cache(struct btrfs_fs_info *fs_info,
4414                                   struct fiemap_extent_info *fieinfo,
4415                                   struct fiemap_cache *cache)
4416 {
4417         int ret;
4418
4419         if (!cache->cached)
4420                 return 0;
4421
4422         ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4423                                       cache->len, cache->flags);
4424         cache->cached = false;
4425         if (ret > 0)
4426                 ret = 0;
4427         return ret;
4428 }
4429
4430 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4431                 __u64 start, __u64 len)
4432 {
4433         int ret = 0;
4434         u64 off = start;
4435         u64 max = start + len;
4436         u32 flags = 0;
4437         u32 found_type;
4438         u64 last;
4439         u64 last_for_get_extent = 0;
4440         u64 disko = 0;
4441         u64 isize = i_size_read(inode);
4442         struct btrfs_key found_key;
4443         struct extent_map *em = NULL;
4444         struct extent_state *cached_state = NULL;
4445         struct btrfs_path *path;
4446         struct btrfs_root *root = BTRFS_I(inode)->root;
4447         struct fiemap_cache cache = { 0 };
4448         int end = 0;
4449         u64 em_start = 0;
4450         u64 em_len = 0;
4451         u64 em_end = 0;
4452
4453         if (len == 0)
4454                 return -EINVAL;
4455
4456         path = btrfs_alloc_path();
4457         if (!path)
4458                 return -ENOMEM;
4459         path->leave_spinning = 1;
4460
4461         start = round_down(start, btrfs_inode_sectorsize(inode));
4462         len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4463
4464         /*
4465          * lookup the last file extent.  We're not using i_size here
4466          * because there might be preallocation past i_size
4467          */
4468         ret = btrfs_lookup_file_extent(NULL, root, path,
4469                         btrfs_ino(BTRFS_I(inode)), -1, 0);
4470         if (ret < 0) {
4471                 btrfs_free_path(path);
4472                 return ret;
4473         } else {
4474                 WARN_ON(!ret);
4475                 if (ret == 1)
4476                         ret = 0;
4477         }
4478
4479         path->slots[0]--;
4480         btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4481         found_type = found_key.type;
4482
4483         /* No extents, but there might be delalloc bits */
4484         if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4485             found_type != BTRFS_EXTENT_DATA_KEY) {
4486                 /* have to trust i_size as the end */
4487                 last = (u64)-1;
4488                 last_for_get_extent = isize;
4489         } else {
4490                 /*
4491                  * remember the start of the last extent.  There are a
4492                  * bunch of different factors that go into the length of the
4493                  * extent, so its much less complex to remember where it started
4494                  */
4495                 last = found_key.offset;
4496                 last_for_get_extent = last + 1;
4497         }
4498         btrfs_release_path(path);
4499
4500         /*
4501          * we might have some extents allocated but more delalloc past those
4502          * extents.  so, we trust isize unless the start of the last extent is
4503          * beyond isize
4504          */
4505         if (last < isize) {
4506                 last = (u64)-1;
4507                 last_for_get_extent = isize;
4508         }
4509
4510         lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4511                          &cached_state);
4512
4513         em = get_extent_skip_holes(inode, start, last_for_get_extent);
4514         if (!em)
4515                 goto out;
4516         if (IS_ERR(em)) {
4517                 ret = PTR_ERR(em);
4518                 goto out;
4519         }
4520
4521         while (!end) {
4522                 u64 offset_in_extent = 0;
4523
4524                 /* break if the extent we found is outside the range */
4525                 if (em->start >= max || extent_map_end(em) < off)
4526                         break;
4527
4528                 /*
4529                  * get_extent may return an extent that starts before our
4530                  * requested range.  We have to make sure the ranges
4531                  * we return to fiemap always move forward and don't
4532                  * overlap, so adjust the offsets here
4533                  */
4534                 em_start = max(em->start, off);
4535
4536                 /*
4537                  * record the offset from the start of the extent
4538                  * for adjusting the disk offset below.  Only do this if the
4539                  * extent isn't compressed since our in ram offset may be past
4540                  * what we have actually allocated on disk.
4541                  */
4542                 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4543                         offset_in_extent = em_start - em->start;
4544                 em_end = extent_map_end(em);
4545                 em_len = em_end - em_start;
4546                 disko = 0;
4547                 flags = 0;
4548
4549                 /*
4550                  * bump off for our next call to get_extent
4551                  */
4552                 off = extent_map_end(em);
4553                 if (off >= max)
4554                         end = 1;
4555
4556                 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4557                         end = 1;
4558                         flags |= FIEMAP_EXTENT_LAST;
4559                 } else if (em->block_start == EXTENT_MAP_INLINE) {
4560                         flags |= (FIEMAP_EXTENT_DATA_INLINE |
4561                                   FIEMAP_EXTENT_NOT_ALIGNED);
4562                 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4563                         flags |= (FIEMAP_EXTENT_DELALLOC |
4564                                   FIEMAP_EXTENT_UNKNOWN);
4565                 } else if (fieinfo->fi_extents_max) {
4566                         u64 bytenr = em->block_start -
4567                                 (em->start - em->orig_start);
4568
4569                         disko = em->block_start + offset_in_extent;
4570
4571                         /*
4572                          * As btrfs supports shared space, this information
4573                          * can be exported to userspace tools via
4574                          * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
4575                          * then we're just getting a count and we can skip the
4576                          * lookup stuff.
4577                          */
4578                         ret = btrfs_check_shared(root,
4579                                                  btrfs_ino(BTRFS_I(inode)),
4580                                                  bytenr);
4581                         if (ret < 0)
4582                                 goto out_free;
4583                         if (ret)
4584                                 flags |= FIEMAP_EXTENT_SHARED;
4585                         ret = 0;
4586                 }
4587                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4588                         flags |= FIEMAP_EXTENT_ENCODED;
4589                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4590                         flags |= FIEMAP_EXTENT_UNWRITTEN;
4591
4592                 free_extent_map(em);
4593                 em = NULL;
4594                 if ((em_start >= last) || em_len == (u64)-1 ||
4595                    (last == (u64)-1 && isize <= em_end)) {
4596                         flags |= FIEMAP_EXTENT_LAST;
4597                         end = 1;
4598                 }
4599
4600                 /* now scan forward to see if this is really the last extent. */
4601                 em = get_extent_skip_holes(inode, off, last_for_get_extent);
4602                 if (IS_ERR(em)) {
4603                         ret = PTR_ERR(em);
4604                         goto out;
4605                 }
4606                 if (!em) {
4607                         flags |= FIEMAP_EXTENT_LAST;
4608                         end = 1;
4609                 }
4610                 ret = emit_fiemap_extent(fieinfo, &cache, em_start, disko,
4611                                            em_len, flags);
4612                 if (ret) {
4613                         if (ret == 1)
4614                                 ret = 0;
4615                         goto out_free;
4616                 }
4617         }
4618 out_free:
4619         if (!ret)
4620                 ret = emit_last_fiemap_cache(root->fs_info, fieinfo, &cache);
4621         free_extent_map(em);
4622 out:
4623         btrfs_free_path(path);
4624         unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4625                              &cached_state);
4626         return ret;
4627 }
4628
4629 static void __free_extent_buffer(struct extent_buffer *eb)
4630 {
4631         btrfs_leak_debug_del(&eb->leak_list);
4632         kmem_cache_free(extent_buffer_cache, eb);
4633 }
4634
4635 int extent_buffer_under_io(struct extent_buffer *eb)
4636 {
4637         return (atomic_read(&eb->io_pages) ||
4638                 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4639                 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4640 }
4641
4642 /*
4643  * Helper for releasing extent buffer page.
4644  */
4645 static void btrfs_release_extent_buffer_page(struct extent_buffer *eb)
4646 {
4647         unsigned long index;
4648         struct page *page;
4649         int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4650
4651         BUG_ON(extent_buffer_under_io(eb));
4652
4653         index = num_extent_pages(eb->start, eb->len);
4654         if (index == 0)
4655                 return;
4656
4657         do {
4658                 index--;
4659                 page = eb->pages[index];
4660                 if (!page)
4661                         continue;
4662                 if (mapped)
4663                         spin_lock(&page->mapping->private_lock);
4664                 /*
4665                  * We do this since we'll remove the pages after we've
4666                  * removed the eb from the radix tree, so we could race
4667                  * and have this page now attached to the new eb.  So
4668                  * only clear page_private if it's still connected to
4669                  * this eb.
4670                  */
4671                 if (PagePrivate(page) &&
4672                     page->private == (unsigned long)eb) {
4673                         BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4674                         BUG_ON(PageDirty(page));
4675                         BUG_ON(PageWriteback(page));
4676                         /*
4677                          * We need to make sure we haven't be attached
4678                          * to a new eb.
4679                          */
4680                         ClearPagePrivate(page);
4681                         set_page_private(page, 0);
4682                         /* One for the page private */
4683                         put_page(page);
4684                 }
4685
4686                 if (mapped)
4687                         spin_unlock(&page->mapping->private_lock);
4688
4689                 /* One for when we allocated the page */
4690                 put_page(page);
4691         } while (index != 0);
4692 }
4693
4694 /*
4695  * Helper for releasing the extent buffer.
4696  */
4697 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4698 {
4699         btrfs_release_extent_buffer_page(eb);
4700         __free_extent_buffer(eb);
4701 }
4702
4703 static struct extent_buffer *
4704 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4705                       unsigned long len)
4706 {
4707         struct extent_buffer *eb = NULL;
4708
4709         eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4710         eb->start = start;
4711         eb->len = len;
4712         eb->fs_info = fs_info;
4713         eb->bflags = 0;
4714         rwlock_init(&eb->lock);
4715         atomic_set(&eb->write_locks, 0);
4716         atomic_set(&eb->read_locks, 0);
4717         atomic_set(&eb->blocking_readers, 0);
4718         atomic_set(&eb->blocking_writers, 0);
4719         atomic_set(&eb->spinning_readers, 0);
4720         atomic_set(&eb->spinning_writers, 0);
4721         eb->lock_nested = 0;
4722         init_waitqueue_head(&eb->write_lock_wq);
4723         init_waitqueue_head(&eb->read_lock_wq);
4724
4725         btrfs_leak_debug_add(&eb->leak_list, &buffers);
4726
4727         spin_lock_init(&eb->refs_lock);
4728         atomic_set(&eb->refs, 1);
4729         atomic_set(&eb->io_pages, 0);
4730
4731         /*
4732          * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4733          */
4734         BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4735                 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4736         BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4737
4738         return eb;
4739 }
4740
4741 struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4742 {
4743         unsigned long i;
4744         struct page *p;
4745         struct extent_buffer *new;
4746         unsigned long num_pages = num_extent_pages(src->start, src->len);
4747
4748         new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4749         if (new == NULL)
4750                 return NULL;
4751
4752         for (i = 0; i < num_pages; i++) {
4753                 p = alloc_page(GFP_NOFS);
4754                 if (!p) {
4755                         btrfs_release_extent_buffer(new);
4756                         return NULL;
4757                 }
4758                 attach_extent_buffer_page(new, p);
4759                 WARN_ON(PageDirty(p));
4760                 SetPageUptodate(p);
4761                 new->pages[i] = p;
4762                 copy_page(page_address(p), page_address(src->pages[i]));
4763         }
4764
4765         set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4766         set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4767
4768         return new;
4769 }
4770
4771 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4772                                                   u64 start, unsigned long len)
4773 {
4774         struct extent_buffer *eb;
4775         unsigned long num_pages;
4776         unsigned long i;
4777
4778         num_pages = num_extent_pages(start, len);
4779
4780         eb = __alloc_extent_buffer(fs_info, start, len);
4781         if (!eb)
4782                 return NULL;
4783
4784         for (i = 0; i < num_pages; i++) {
4785                 eb->pages[i] = alloc_page(GFP_NOFS);
4786                 if (!eb->pages[i])
4787                         goto err;
4788         }
4789         set_extent_buffer_uptodate(eb);
4790         btrfs_set_header_nritems(eb, 0);
4791         set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4792
4793         return eb;
4794 err:
4795         for (; i > 0; i--)
4796                 __free_page(eb->pages[i - 1]);
4797         __free_extent_buffer(eb);
4798         return NULL;
4799 }
4800
4801 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4802                                                 u64 start)
4803 {
4804         return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4805 }
4806
4807 static void check_buffer_tree_ref(struct extent_buffer *eb)
4808 {
4809         int refs;
4810         /* the ref bit is tricky.  We have to make sure it is set
4811          * if we have the buffer dirty.   Otherwise the
4812          * code to free a buffer can end up dropping a dirty
4813          * page
4814          *
4815          * Once the ref bit is set, it won't go away while the
4816          * buffer is dirty or in writeback, and it also won't
4817          * go away while we have the reference count on the
4818          * eb bumped.
4819          *
4820          * We can't just set the ref bit without bumping the
4821          * ref on the eb because free_extent_buffer might
4822          * see the ref bit and try to clear it.  If this happens
4823          * free_extent_buffer might end up dropping our original
4824          * ref by mistake and freeing the page before we are able
4825          * to add one more ref.
4826          *
4827          * So bump the ref count first, then set the bit.  If someone
4828          * beat us to it, drop the ref we added.
4829          */
4830         refs = atomic_read(&eb->refs);
4831         if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4832                 return;
4833
4834         spin_lock(&eb->refs_lock);
4835         if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4836                 atomic_inc(&eb->refs);
4837         spin_unlock(&eb->refs_lock);
4838 }
4839
4840 static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4841                 struct page *accessed)
4842 {
4843         unsigned long num_pages, i;
4844
4845         check_buffer_tree_ref(eb);
4846
4847         num_pages = num_extent_pages(eb->start, eb->len);
4848         for (i = 0; i < num_pages; i++) {
4849                 struct page *p = eb->pages[i];
4850
4851                 if (p != accessed)
4852                         mark_page_accessed(p);
4853         }
4854 }
4855
4856 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4857                                          u64 start)
4858 {
4859         struct extent_buffer *eb;
4860
4861         rcu_read_lock();
4862         eb = radix_tree_lookup(&fs_info->buffer_radix,
4863                                start >> PAGE_SHIFT);
4864         if (eb && atomic_inc_not_zero(&eb->refs)) {
4865                 rcu_read_unlock();
4866                 /*
4867                  * Lock our eb's refs_lock to avoid races with
4868                  * free_extent_buffer. When we get our eb it might be flagged
4869                  * with EXTENT_BUFFER_STALE and another task running
4870                  * free_extent_buffer might have seen that flag set,
4871                  * eb->refs == 2, that the buffer isn't under IO (dirty and
4872                  * writeback flags not set) and it's still in the tree (flag
4873                  * EXTENT_BUFFER_TREE_REF set), therefore being in the process
4874                  * of decrementing the extent buffer's reference count twice.
4875                  * So here we could race and increment the eb's reference count,
4876                  * clear its stale flag, mark it as dirty and drop our reference
4877                  * before the other task finishes executing free_extent_buffer,
4878                  * which would later result in an attempt to free an extent
4879                  * buffer that is dirty.
4880                  */
4881                 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4882                         spin_lock(&eb->refs_lock);
4883                         spin_unlock(&eb->refs_lock);
4884                 }
4885                 mark_extent_buffer_accessed(eb, NULL);
4886                 return eb;
4887         }
4888         rcu_read_unlock();
4889
4890         return NULL;
4891 }
4892
4893 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4894 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4895                                         u64 start)
4896 {
4897         struct extent_buffer *eb, *exists = NULL;
4898         int ret;
4899
4900         eb = find_extent_buffer(fs_info, start);
4901         if (eb)
4902                 return eb;
4903         eb = alloc_dummy_extent_buffer(fs_info, start);
4904         if (!eb)
4905                 return NULL;
4906         eb->fs_info = fs_info;
4907 again:
4908         ret = radix_tree_preload(GFP_NOFS);
4909         if (ret)
4910                 goto free_eb;
4911         spin_lock(&fs_info->buffer_lock);
4912         ret = radix_tree_insert(&fs_info->buffer_radix,
4913                                 start >> PAGE_SHIFT, eb);
4914         spin_unlock(&fs_info->buffer_lock);
4915         radix_tree_preload_end();
4916         if (ret == -EEXIST) {
4917                 exists = find_extent_buffer(fs_info, start);
4918                 if (exists)
4919                         goto free_eb;
4920                 else
4921                         goto again;
4922         }
4923         check_buffer_tree_ref(eb);
4924         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4925
4926         /*
4927          * We will free dummy extent buffer's if they come into
4928          * free_extent_buffer with a ref count of 2, but if we are using this we
4929          * want the buffers to stay in memory until we're done with them, so
4930          * bump the ref count again.
4931          */
4932         atomic_inc(&eb->refs);
4933         return eb;
4934 free_eb:
4935         btrfs_release_extent_buffer(eb);
4936         return exists;
4937 }
4938 #endif
4939
4940 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4941                                           u64 start)
4942 {
4943         unsigned long len = fs_info->nodesize;
4944         unsigned long num_pages = num_extent_pages(start, len);
4945         unsigned long i;
4946         unsigned long index = start >> PAGE_SHIFT;
4947         struct extent_buffer *eb;
4948         struct extent_buffer *exists = NULL;
4949         struct page *p;
4950         struct address_space *mapping = fs_info->btree_inode->i_mapping;
4951         int uptodate = 1;
4952         int ret;
4953
4954         if (!IS_ALIGNED(start, fs_info->sectorsize)) {
4955                 btrfs_err(fs_info, "bad tree block start %llu", start);
4956                 return ERR_PTR(-EINVAL);
4957         }
4958
4959         eb = find_extent_buffer(fs_info, start);
4960         if (eb)
4961                 return eb;
4962
4963         eb = __alloc_extent_buffer(fs_info, start, len);
4964         if (!eb)
4965                 return ERR_PTR(-ENOMEM);
4966
4967         for (i = 0; i < num_pages; i++, index++) {
4968                 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4969                 if (!p) {
4970                         exists = ERR_PTR(-ENOMEM);
4971                         goto free_eb;
4972                 }
4973
4974                 spin_lock(&mapping->private_lock);
4975                 if (PagePrivate(p)) {
4976                         /*
4977                          * We could have already allocated an eb for this page
4978                          * and attached one so lets see if we can get a ref on
4979                          * the existing eb, and if we can we know it's good and
4980                          * we can just return that one, else we know we can just
4981                          * overwrite page->private.
4982                          */
4983                         exists = (struct extent_buffer *)p->private;
4984                         if (atomic_inc_not_zero(&exists->refs)) {
4985                                 spin_unlock(&mapping->private_lock);
4986                                 unlock_page(p);
4987                                 put_page(p);
4988                                 mark_extent_buffer_accessed(exists, p);
4989                                 goto free_eb;
4990                         }
4991                         exists = NULL;
4992
4993                         /*
4994                          * Do this so attach doesn't complain and we need to
4995                          * drop the ref the old guy had.
4996                          */
4997                         ClearPagePrivate(p);
4998                         WARN_ON(PageDirty(p));
4999                         put_page(p);
5000                 }
5001                 attach_extent_buffer_page(eb, p);
5002                 spin_unlock(&mapping->private_lock);
5003                 WARN_ON(PageDirty(p));
5004                 eb->pages[i] = p;
5005                 if (!PageUptodate(p))
5006                         uptodate = 0;
5007
5008                 /*
5009                  * see below about how we avoid a nasty race with release page
5010                  * and why we unlock later
5011                  */
5012         }
5013         if (uptodate)
5014                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5015 again:
5016         ret = radix_tree_preload(GFP_NOFS);
5017         if (ret) {
5018                 exists = ERR_PTR(ret);
5019                 goto free_eb;
5020         }
5021
5022         spin_lock(&fs_info->buffer_lock);
5023         ret = radix_tree_insert(&fs_info->buffer_radix,
5024                                 start >> PAGE_SHIFT, eb);
5025         spin_unlock(&fs_info->buffer_lock);
5026         radix_tree_preload_end();
5027         if (ret == -EEXIST) {
5028                 exists = find_extent_buffer(fs_info, start);
5029                 if (exists)
5030                         goto free_eb;
5031                 else
5032                         goto again;
5033         }
5034         /* add one reference for the tree */
5035         check_buffer_tree_ref(eb);
5036         set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5037
5038         /*
5039          * there is a race where release page may have
5040          * tried to find this extent buffer in the radix
5041          * but failed.  It will tell the VM it is safe to
5042          * reclaim the, and it will clear the page private bit.
5043          * We must make sure to set the page private bit properly
5044          * after the extent buffer is in the radix tree so
5045          * it doesn't get lost
5046          */
5047         SetPageChecked(eb->pages[0]);
5048         for (i = 1; i < num_pages; i++) {
5049                 p = eb->pages[i];
5050                 ClearPageChecked(p);
5051                 unlock_page(p);
5052         }
5053         unlock_page(eb->pages[0]);
5054         return eb;
5055
5056 free_eb:
5057         WARN_ON(!atomic_dec_and_test(&eb->refs));
5058         for (i = 0; i < num_pages; i++) {
5059                 if (eb->pages[i])
5060                         unlock_page(eb->pages[i]);
5061         }
5062
5063         btrfs_release_extent_buffer(eb);
5064         return exists;
5065 }
5066
5067 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5068 {
5069         struct extent_buffer *eb =
5070                         container_of(head, struct extent_buffer, rcu_head);
5071
5072         __free_extent_buffer(eb);
5073 }
5074
5075 /* Expects to have eb->eb_lock already held */
5076 static int release_extent_buffer(struct extent_buffer *eb)
5077 {
5078         WARN_ON(atomic_read(&eb->refs) == 0);
5079         if (atomic_dec_and_test(&eb->refs)) {
5080                 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5081                         struct btrfs_fs_info *fs_info = eb->fs_info;
5082
5083                         spin_unlock(&eb->refs_lock);
5084
5085                         spin_lock(&fs_info->buffer_lock);
5086                         radix_tree_delete(&fs_info->buffer_radix,
5087                                           eb->start >> PAGE_SHIFT);
5088                         spin_unlock(&fs_info->buffer_lock);
5089                 } else {
5090                         spin_unlock(&eb->refs_lock);
5091                 }
5092
5093                 /* Should be safe to release our pages at this point */
5094                 btrfs_release_extent_buffer_page(eb);
5095 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5096                 if (unlikely(test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))) {
5097                         __free_extent_buffer(eb);
5098                         return 1;
5099                 }
5100 #endif
5101                 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5102                 return 1;
5103         }
5104         spin_unlock(&eb->refs_lock);
5105
5106         return 0;
5107 }
5108
5109 void free_extent_buffer(struct extent_buffer *eb)
5110 {
5111         int refs;
5112         int old;
5113         if (!eb)
5114                 return;
5115
5116         while (1) {
5117                 refs = atomic_read(&eb->refs);
5118                 if (refs <= 3)
5119                         break;
5120                 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5121                 if (old == refs)
5122                         return;
5123         }
5124
5125         spin_lock(&eb->refs_lock);
5126         if (atomic_read(&eb->refs) == 2 &&
5127             test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
5128                 atomic_dec(&eb->refs);
5129
5130         if (atomic_read(&eb->refs) == 2 &&
5131             test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5132             !extent_buffer_under_io(eb) &&
5133             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5134                 atomic_dec(&eb->refs);
5135
5136         /*
5137          * I know this is terrible, but it's temporary until we stop tracking
5138          * the uptodate bits and such for the extent buffers.
5139          */
5140         release_extent_buffer(eb);
5141 }
5142
5143 void free_extent_buffer_stale(struct extent_buffer *eb)
5144 {
5145         if (!eb)
5146                 return;
5147
5148         spin_lock(&eb->refs_lock);
5149         set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5150
5151         if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5152             test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5153                 atomic_dec(&eb->refs);
5154         release_extent_buffer(eb);
5155 }
5156
5157 void clear_extent_buffer_dirty(struct extent_buffer *eb)
5158 {
5159         unsigned long i;
5160         unsigned long num_pages;
5161         struct page *page;
5162
5163         num_pages = num_extent_pages(eb->start, eb->len);
5164
5165         for (i = 0; i < num_pages; i++) {
5166                 page = eb->pages[i];
5167                 if (!PageDirty(page))
5168                         continue;
5169
5170                 lock_page(page);
5171                 WARN_ON(!PagePrivate(page));
5172
5173                 clear_page_dirty_for_io(page);
5174                 spin_lock_irq(&page->mapping->tree_lock);
5175                 if (!PageDirty(page)) {
5176                         radix_tree_tag_clear(&page->mapping->page_tree,
5177                                                 page_index(page),
5178                                                 PAGECACHE_TAG_DIRTY);
5179                 }
5180                 spin_unlock_irq(&page->mapping->tree_lock);
5181                 ClearPageError(page);
5182                 unlock_page(page);
5183         }
5184         WARN_ON(atomic_read(&eb->refs) == 0);
5185 }
5186
5187 int set_extent_buffer_dirty(struct extent_buffer *eb)
5188 {
5189         unsigned long i;
5190         unsigned long num_pages;
5191         int was_dirty = 0;
5192
5193         check_buffer_tree_ref(eb);
5194
5195         was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5196
5197         num_pages = num_extent_pages(eb->start, eb->len);
5198         WARN_ON(atomic_read(&eb->refs) == 0);
5199         WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5200
5201         for (i = 0; i < num_pages; i++)
5202                 set_page_dirty(eb->pages[i]);
5203         return was_dirty;
5204 }
5205
5206 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5207 {
5208         unsigned long i;
5209         struct page *page;
5210         unsigned long num_pages;
5211
5212         clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5213         num_pages = num_extent_pages(eb->start, eb->len);
5214         for (i = 0; i < num_pages; i++) {
5215                 page = eb->pages[i];
5216                 if (page)
5217                         ClearPageUptodate(page);
5218         }
5219 }
5220
5221 void set_extent_buffer_uptodate(struct extent_buffer *eb)
5222 {
5223         unsigned long i;
5224         struct page *page;
5225         unsigned long num_pages;
5226
5227         set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5228         num_pages = num_extent_pages(eb->start, eb->len);
5229         for (i = 0; i < num_pages; i++) {
5230                 page = eb->pages[i];
5231                 SetPageUptodate(page);
5232         }
5233 }
5234
5235 int read_extent_buffer_pages(struct extent_io_tree *tree,
5236                              struct extent_buffer *eb, int wait, int mirror_num)
5237 {
5238         unsigned long i;
5239         struct page *page;
5240         int err;
5241         int ret = 0;
5242         int locked_pages = 0;
5243         int all_uptodate = 1;
5244         unsigned long num_pages;
5245         unsigned long num_reads = 0;
5246         struct bio *bio = NULL;
5247         unsigned long bio_flags = 0;
5248
5249         if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5250                 return 0;
5251
5252         num_pages = num_extent_pages(eb->start, eb->len);
5253         for (i = 0; i < num_pages; i++) {
5254                 page = eb->pages[i];
5255                 if (wait == WAIT_NONE) {
5256                         if (!trylock_page(page))
5257                                 goto unlock_exit;
5258                 } else {
5259                         lock_page(page);
5260                 }
5261                 locked_pages++;
5262         }
5263         /*
5264          * We need to firstly lock all pages to make sure that
5265          * the uptodate bit of our pages won't be affected by
5266          * clear_extent_buffer_uptodate().
5267          */
5268         for (i = 0; i < num_pages; i++) {
5269                 page = eb->pages[i];
5270                 if (!PageUptodate(page)) {
5271                         num_reads++;
5272                         all_uptodate = 0;
5273                 }
5274         }
5275
5276         if (all_uptodate) {
5277                 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5278                 goto unlock_exit;
5279         }
5280
5281         clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5282         eb->read_mirror = 0;
5283         atomic_set(&eb->io_pages, num_reads);
5284         for (i = 0; i < num_pages; i++) {
5285                 page = eb->pages[i];
5286
5287                 if (!PageUptodate(page)) {
5288                         if (ret) {
5289                                 atomic_dec(&eb->io_pages);
5290                                 unlock_page(page);
5291                                 continue;
5292                         }
5293
5294                         ClearPageError(page);
5295                         err = __extent_read_full_page(tree, page,
5296                                                       btree_get_extent, &bio,
5297                                                       mirror_num, &bio_flags,
5298                                                       REQ_META);
5299                         if (err) {
5300                                 ret = err;
5301                                 /*
5302                                  * We use &bio in above __extent_read_full_page,
5303                                  * so we ensure that if it returns error, the
5304                                  * current page fails to add itself to bio and
5305                                  * it's been unlocked.
5306                                  *
5307                                  * We must dec io_pages by ourselves.
5308                                  */
5309                                 atomic_dec(&eb->io_pages);
5310                         }
5311                 } else {
5312                         unlock_page(page);
5313                 }
5314         }
5315
5316         if (bio) {
5317                 err = submit_one_bio(bio, mirror_num, bio_flags);
5318                 if (err)
5319                         return err;
5320         }
5321
5322         if (ret || wait != WAIT_COMPLETE)
5323                 return ret;
5324
5325         for (i = 0; i < num_pages; i++) {
5326                 page = eb->pages[i];
5327                 wait_on_page_locked(page);
5328                 if (!PageUptodate(page))
5329                         ret = -EIO;
5330         }
5331
5332         return ret;
5333
5334 unlock_exit:
5335         while (locked_pages > 0) {
5336                 locked_pages--;
5337                 page = eb->pages[locked_pages];
5338                 unlock_page(page);
5339         }
5340         return ret;
5341 }
5342
5343 void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
5344                         unsigned long start, unsigned long len)
5345 {
5346         size_t cur;
5347         size_t offset;
5348         struct page *page;
5349         char *kaddr;
5350         char *dst = (char *)dstv;
5351         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5352         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5353
5354         if (start + len > eb->len) {
5355                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5356                      eb->start, eb->len, start, len);
5357                 memset(dst, 0, len);
5358                 return;
5359         }
5360
5361         offset = (start_offset + start) & (PAGE_SIZE - 1);
5362
5363         while (len > 0) {
5364                 page = eb->pages[i];
5365
5366                 cur = min(len, (PAGE_SIZE - offset));
5367                 kaddr = page_address(page);
5368                 memcpy(dst, kaddr + offset, cur);
5369
5370                 dst += cur;
5371                 len -= cur;
5372                 offset = 0;
5373                 i++;
5374         }
5375 }
5376
5377 int read_extent_buffer_to_user(const struct extent_buffer *eb,
5378                                void __user *dstv,
5379                                unsigned long start, unsigned long len)
5380 {
5381         size_t cur;
5382         size_t offset;
5383         struct page *page;
5384         char *kaddr;
5385         char __user *dst = (char __user *)dstv;
5386         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5387         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5388         int ret = 0;
5389
5390         WARN_ON(start > eb->len);
5391         WARN_ON(start + len > eb->start + eb->len);
5392
5393         offset = (start_offset + start) & (PAGE_SIZE - 1);
5394
5395         while (len > 0) {
5396                 page = eb->pages[i];
5397
5398                 cur = min(len, (PAGE_SIZE - offset));
5399                 kaddr = page_address(page);
5400                 if (copy_to_user(dst, kaddr + offset, cur)) {
5401                         ret = -EFAULT;
5402                         break;
5403                 }
5404
5405                 dst += cur;
5406                 len -= cur;
5407                 offset = 0;
5408                 i++;
5409         }
5410
5411         return ret;
5412 }
5413
5414 /*
5415  * return 0 if the item is found within a page.
5416  * return 1 if the item spans two pages.
5417  * return -EINVAL otherwise.
5418  */
5419 int map_private_extent_buffer(const struct extent_buffer *eb,
5420                               unsigned long start, unsigned long min_len,
5421                               char **map, unsigned long *map_start,
5422                               unsigned long *map_len)
5423 {
5424         size_t offset = start & (PAGE_SIZE - 1);
5425         char *kaddr;
5426         struct page *p;
5427         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5428         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5429         unsigned long end_i = (start_offset + start + min_len - 1) >>
5430                 PAGE_SHIFT;
5431
5432         if (start + min_len > eb->len) {
5433                 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5434                        eb->start, eb->len, start, min_len);
5435                 return -EINVAL;
5436         }
5437
5438         if (i != end_i)
5439                 return 1;
5440
5441         if (i == 0) {
5442                 offset = start_offset;
5443                 *map_start = 0;
5444         } else {
5445                 offset = 0;
5446                 *map_start = ((u64)i << PAGE_SHIFT) - start_offset;
5447         }
5448
5449         p = eb->pages[i];
5450         kaddr = page_address(p);
5451         *map = kaddr + offset;
5452         *map_len = PAGE_SIZE - offset;
5453         return 0;
5454 }
5455
5456 int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
5457                          unsigned long start, unsigned long len)
5458 {
5459         size_t cur;
5460         size_t offset;
5461         struct page *page;
5462         char *kaddr;
5463         char *ptr = (char *)ptrv;
5464         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5465         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5466         int ret = 0;
5467
5468         WARN_ON(start > eb->len);
5469         WARN_ON(start + len > eb->start + eb->len);
5470
5471         offset = (start_offset + start) & (PAGE_SIZE - 1);
5472
5473         while (len > 0) {
5474                 page = eb->pages[i];
5475
5476                 cur = min(len, (PAGE_SIZE - offset));
5477
5478                 kaddr = page_address(page);
5479                 ret = memcmp(ptr, kaddr + offset, cur);
5480                 if (ret)
5481                         break;
5482
5483                 ptr += cur;
5484                 len -= cur;
5485                 offset = 0;
5486                 i++;
5487         }
5488         return ret;
5489 }
5490
5491 void write_extent_buffer_chunk_tree_uuid(struct extent_buffer *eb,
5492                 const void *srcv)
5493 {
5494         char *kaddr;
5495
5496         WARN_ON(!PageUptodate(eb->pages[0]));
5497         kaddr = page_address(eb->pages[0]);
5498         memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5499                         BTRFS_FSID_SIZE);
5500 }
5501
5502 void write_extent_buffer_fsid(struct extent_buffer *eb, const void *srcv)
5503 {
5504         char *kaddr;
5505
5506         WARN_ON(!PageUptodate(eb->pages[0]));
5507         kaddr = page_address(eb->pages[0]);
5508         memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5509                         BTRFS_FSID_SIZE);
5510 }
5511
5512 void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5513                          unsigned long start, unsigned long len)
5514 {
5515         size_t cur;
5516         size_t offset;
5517         struct page *page;
5518         char *kaddr;
5519         char *src = (char *)srcv;
5520         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5521         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5522
5523         WARN_ON(start > eb->len);
5524         WARN_ON(start + len > eb->start + eb->len);
5525
5526         offset = (start_offset + start) & (PAGE_SIZE - 1);
5527
5528         while (len > 0) {
5529                 page = eb->pages[i];
5530                 WARN_ON(!PageUptodate(page));
5531
5532                 cur = min(len, PAGE_SIZE - offset);
5533                 kaddr = page_address(page);
5534                 memcpy(kaddr + offset, src, cur);
5535
5536                 src += cur;
5537                 len -= cur;
5538                 offset = 0;
5539                 i++;
5540         }
5541 }
5542
5543 void memzero_extent_buffer(struct extent_buffer *eb, unsigned long start,
5544                 unsigned long len)
5545 {
5546         size_t cur;
5547         size_t offset;
5548         struct page *page;
5549         char *kaddr;
5550         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5551         unsigned long i = (start_offset + start) >> PAGE_SHIFT;
5552
5553         WARN_ON(start > eb->len);
5554         WARN_ON(start + len > eb->start + eb->len);
5555
5556         offset = (start_offset + start) & (PAGE_SIZE - 1);
5557
5558         while (len > 0) {
5559                 page = eb->pages[i];
5560                 WARN_ON(!PageUptodate(page));
5561
5562                 cur = min(len, PAGE_SIZE - offset);
5563                 kaddr = page_address(page);
5564                 memset(kaddr + offset, 0, cur);
5565
5566                 len -= cur;
5567                 offset = 0;
5568                 i++;
5569         }
5570 }
5571
5572 void copy_extent_buffer_full(struct extent_buffer *dst,
5573                              struct extent_buffer *src)
5574 {
5575         int i;
5576         unsigned num_pages;
5577
5578         ASSERT(dst->len == src->len);
5579
5580         num_pages = num_extent_pages(dst->start, dst->len);
5581         for (i = 0; i < num_pages; i++)
5582                 copy_page(page_address(dst->pages[i]),
5583                                 page_address(src->pages[i]));
5584 }
5585
5586 void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5587                         unsigned long dst_offset, unsigned long src_offset,
5588                         unsigned long len)
5589 {
5590         u64 dst_len = dst->len;
5591         size_t cur;
5592         size_t offset;
5593         struct page *page;
5594         char *kaddr;
5595         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5596         unsigned long i = (start_offset + dst_offset) >> PAGE_SHIFT;
5597
5598         WARN_ON(src->len != dst_len);
5599
5600         offset = (start_offset + dst_offset) &
5601                 (PAGE_SIZE - 1);
5602
5603         while (len > 0) {
5604                 page = dst->pages[i];
5605                 WARN_ON(!PageUptodate(page));
5606
5607                 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5608
5609                 kaddr = page_address(page);
5610                 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5611
5612                 src_offset += cur;
5613                 len -= cur;
5614                 offset = 0;
5615                 i++;
5616         }
5617 }
5618
5619 void le_bitmap_set(u8 *map, unsigned int start, int len)
5620 {
5621         u8 *p = map + BIT_BYTE(start);
5622         const unsigned int size = start + len;
5623         int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5624         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
5625
5626         while (len - bits_to_set >= 0) {
5627                 *p |= mask_to_set;
5628                 len -= bits_to_set;
5629                 bits_to_set = BITS_PER_BYTE;
5630                 mask_to_set = ~0;
5631                 p++;
5632         }
5633         if (len) {
5634                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5635                 *p |= mask_to_set;
5636         }
5637 }
5638
5639 void le_bitmap_clear(u8 *map, unsigned int start, int len)
5640 {
5641         u8 *p = map + BIT_BYTE(start);
5642         const unsigned int size = start + len;
5643         int bits_to_clear = BITS_PER_BYTE - (start % BITS_PER_BYTE);
5644         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(start);
5645
5646         while (len - bits_to_clear >= 0) {
5647                 *p &= ~mask_to_clear;
5648                 len -= bits_to_clear;
5649                 bits_to_clear = BITS_PER_BYTE;
5650                 mask_to_clear = ~0;
5651                 p++;
5652         }
5653         if (len) {
5654                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5655                 *p &= ~mask_to_clear;
5656         }
5657 }
5658
5659 /*
5660  * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5661  * given bit number
5662  * @eb: the extent buffer
5663  * @start: offset of the bitmap item in the extent buffer
5664  * @nr: bit number
5665  * @page_index: return index of the page in the extent buffer that contains the
5666  * given bit number
5667  * @page_offset: return offset into the page given by page_index
5668  *
5669  * This helper hides the ugliness of finding the byte in an extent buffer which
5670  * contains a given bit.
5671  */
5672 static inline void eb_bitmap_offset(struct extent_buffer *eb,
5673                                     unsigned long start, unsigned long nr,
5674                                     unsigned long *page_index,
5675                                     size_t *page_offset)
5676 {
5677         size_t start_offset = eb->start & ((u64)PAGE_SIZE - 1);
5678         size_t byte_offset = BIT_BYTE(nr);
5679         size_t offset;
5680
5681         /*
5682          * The byte we want is the offset of the extent buffer + the offset of
5683          * the bitmap item in the extent buffer + the offset of the byte in the
5684          * bitmap item.
5685          */
5686         offset = start_offset + start + byte_offset;
5687
5688         *page_index = offset >> PAGE_SHIFT;
5689         *page_offset = offset & (PAGE_SIZE - 1);
5690 }
5691
5692 /**
5693  * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5694  * @eb: the extent buffer
5695  * @start: offset of the bitmap item in the extent buffer
5696  * @nr: bit number to test
5697  */
5698 int extent_buffer_test_bit(struct extent_buffer *eb, unsigned long start,
5699                            unsigned long nr)
5700 {
5701         u8 *kaddr;
5702         struct page *page;
5703         unsigned long i;
5704         size_t offset;
5705
5706         eb_bitmap_offset(eb, start, nr, &i, &offset);
5707         page = eb->pages[i];
5708         WARN_ON(!PageUptodate(page));
5709         kaddr = page_address(page);
5710         return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5711 }
5712
5713 /**
5714  * extent_buffer_bitmap_set - set an area of a bitmap
5715  * @eb: the extent buffer
5716  * @start: offset of the bitmap item in the extent buffer
5717  * @pos: bit number of the first bit
5718  * @len: number of bits to set
5719  */
5720 void extent_buffer_bitmap_set(struct extent_buffer *eb, unsigned long start,
5721                               unsigned long pos, unsigned long len)
5722 {
5723         u8 *kaddr;
5724         struct page *page;
5725         unsigned long i;
5726         size_t offset;
5727         const unsigned int size = pos + len;
5728         int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5729         u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5730
5731         eb_bitmap_offset(eb, start, pos, &i, &offset);
5732         page = eb->pages[i];
5733         WARN_ON(!PageUptodate(page));
5734         kaddr = page_address(page);
5735
5736         while (len >= bits_to_set) {
5737                 kaddr[offset] |= mask_to_set;
5738                 len -= bits_to_set;
5739                 bits_to_set = BITS_PER_BYTE;
5740                 mask_to_set = ~0;
5741                 if (++offset >= PAGE_SIZE && len > 0) {
5742                         offset = 0;
5743                         page = eb->pages[++i];
5744                         WARN_ON(!PageUptodate(page));
5745                         kaddr = page_address(page);
5746                 }
5747         }
5748         if (len) {
5749                 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5750                 kaddr[offset] |= mask_to_set;
5751         }
5752 }
5753
5754
5755 /**
5756  * extent_buffer_bitmap_clear - clear an area of a bitmap
5757  * @eb: the extent buffer
5758  * @start: offset of the bitmap item in the extent buffer
5759  * @pos: bit number of the first bit
5760  * @len: number of bits to clear
5761  */
5762 void extent_buffer_bitmap_clear(struct extent_buffer *eb, unsigned long start,
5763                                 unsigned long pos, unsigned long len)
5764 {
5765         u8 *kaddr;
5766         struct page *page;
5767         unsigned long i;
5768         size_t offset;
5769         const unsigned int size = pos + len;
5770         int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5771         u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5772
5773         eb_bitmap_offset(eb, start, pos, &i, &offset);
5774         page = eb->pages[i];
5775         WARN_ON(!PageUptodate(page));
5776         kaddr = page_address(page);
5777
5778         while (len >= bits_to_clear) {
5779                 kaddr[offset] &= ~mask_to_clear;
5780                 len -= bits_to_clear;
5781                 bits_to_clear = BITS_PER_BYTE;
5782                 mask_to_clear = ~0;
5783                 if (++offset >= PAGE_SIZE && len > 0) {
5784                         offset = 0;
5785                         page = eb->pages[++i];
5786                         WARN_ON(!PageUptodate(page));
5787                         kaddr = page_address(page);
5788                 }
5789         }
5790         if (len) {
5791                 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5792                 kaddr[offset] &= ~mask_to_clear;
5793         }
5794 }
5795
5796 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5797 {
5798         unsigned long distance = (src > dst) ? src - dst : dst - src;
5799         return distance < len;
5800 }
5801
5802 static void copy_pages(struct page *dst_page, struct page *src_page,
5803                        unsigned long dst_off, unsigned long src_off,
5804                        unsigned long len)
5805 {
5806         char *dst_kaddr = page_address(dst_page);
5807         char *src_kaddr;
5808         int must_memmove = 0;
5809
5810         if (dst_page != src_page) {
5811                 src_kaddr = page_address(src_page);
5812         } else {
5813                 src_kaddr = dst_kaddr;
5814                 if (areas_overlap(src_off, dst_off, len))
5815                         must_memmove = 1;
5816         }
5817
5818         if (must_memmove)
5819                 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5820         else
5821                 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5822 }
5823
5824 void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5825                            unsigned long src_offset, unsigned long len)
5826 {
5827         struct btrfs_fs_info *fs_info = dst->fs_info;
5828         size_t cur;
5829         size_t dst_off_in_page;
5830         size_t src_off_in_page;
5831         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5832         unsigned long dst_i;
5833         unsigned long src_i;
5834
5835         if (src_offset + len > dst->len) {
5836                 btrfs_err(fs_info,
5837                         "memmove bogus src_offset %lu move len %lu dst len %lu",
5838                          src_offset, len, dst->len);
5839                 BUG_ON(1);
5840         }
5841         if (dst_offset + len > dst->len) {
5842                 btrfs_err(fs_info,
5843                         "memmove bogus dst_offset %lu move len %lu dst len %lu",
5844                          dst_offset, len, dst->len);
5845                 BUG_ON(1);
5846         }
5847
5848         while (len > 0) {
5849                 dst_off_in_page = (start_offset + dst_offset) &
5850                         (PAGE_SIZE - 1);
5851                 src_off_in_page = (start_offset + src_offset) &
5852                         (PAGE_SIZE - 1);
5853
5854                 dst_i = (start_offset + dst_offset) >> PAGE_SHIFT;
5855                 src_i = (start_offset + src_offset) >> PAGE_SHIFT;
5856
5857                 cur = min(len, (unsigned long)(PAGE_SIZE -
5858                                                src_off_in_page));
5859                 cur = min_t(unsigned long, cur,
5860                         (unsigned long)(PAGE_SIZE - dst_off_in_page));
5861
5862                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5863                            dst_off_in_page, src_off_in_page, cur);
5864
5865                 src_offset += cur;
5866                 dst_offset += cur;
5867                 len -= cur;
5868         }
5869 }
5870
5871 void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5872                            unsigned long src_offset, unsigned long len)
5873 {
5874         struct btrfs_fs_info *fs_info = dst->fs_info;
5875         size_t cur;
5876         size_t dst_off_in_page;
5877         size_t src_off_in_page;
5878         unsigned long dst_end = dst_offset + len - 1;
5879         unsigned long src_end = src_offset + len - 1;
5880         size_t start_offset = dst->start & ((u64)PAGE_SIZE - 1);
5881         unsigned long dst_i;
5882         unsigned long src_i;
5883
5884         if (src_offset + len > dst->len) {
5885                 btrfs_err(fs_info,
5886                           "memmove bogus src_offset %lu move len %lu len %lu",
5887                           src_offset, len, dst->len);
5888                 BUG_ON(1);
5889         }
5890         if (dst_offset + len > dst->len) {
5891                 btrfs_err(fs_info,
5892                           "memmove bogus dst_offset %lu move len %lu len %lu",
5893                           dst_offset, len, dst->len);
5894                 BUG_ON(1);
5895         }
5896         if (dst_offset < src_offset) {
5897                 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5898                 return;
5899         }
5900         while (len > 0) {
5901                 dst_i = (start_offset + dst_end) >> PAGE_SHIFT;
5902                 src_i = (start_offset + src_end) >> PAGE_SHIFT;
5903
5904                 dst_off_in_page = (start_offset + dst_end) &
5905                         (PAGE_SIZE - 1);
5906                 src_off_in_page = (start_offset + src_end) &
5907                         (PAGE_SIZE - 1);
5908
5909                 cur = min_t(unsigned long, len, src_off_in_page + 1);
5910                 cur = min(cur, dst_off_in_page + 1);
5911                 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5912                            dst_off_in_page - cur + 1,
5913                            src_off_in_page - cur + 1, cur);
5914
5915                 dst_end -= cur;
5916                 src_end -= cur;
5917                 len -= cur;
5918         }
5919 }
5920
5921 int try_release_extent_buffer(struct page *page)
5922 {
5923         struct extent_buffer *eb;
5924
5925         /*
5926          * We need to make sure nobody is attaching this page to an eb right
5927          * now.
5928          */
5929         spin_lock(&page->mapping->private_lock);
5930         if (!PagePrivate(page)) {
5931                 spin_unlock(&page->mapping->private_lock);
5932                 return 1;
5933         }
5934
5935         eb = (struct extent_buffer *)page->private;
5936         BUG_ON(!eb);
5937
5938         /*
5939          * This is a little awful but should be ok, we need to make sure that
5940          * the eb doesn't disappear out from under us while we're looking at
5941          * this page.
5942          */
5943         spin_lock(&eb->refs_lock);
5944         if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5945                 spin_unlock(&eb->refs_lock);
5946                 spin_unlock(&page->mapping->private_lock);
5947                 return 0;
5948         }
5949         spin_unlock(&page->mapping->private_lock);
5950
5951         /*
5952          * If tree ref isn't set then we know the ref on this eb is a real ref,
5953          * so just return, this page will likely be freed soon anyway.
5954          */
5955         if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5956                 spin_unlock(&eb->refs_lock);
5957                 return 0;
5958         }
5959
5960         return release_extent_buffer(eb);
5961 }