Merge tag 'for-5.4-rc1-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave...
[sfrench/cifs-2.6.git] / fs / btrfs / extent_io.c
1 // SPDX-License-Identifier: GPL-2.0
2
3 #include <linux/bitops.h>
4 #include <linux/slab.h>
5 #include <linux/bio.h>
6 #include <linux/mm.h>
7 #include <linux/pagemap.h>
8 #include <linux/page-flags.h>
9 #include <linux/spinlock.h>
10 #include <linux/blkdev.h>
11 #include <linux/swap.h>
12 #include <linux/writeback.h>
13 #include <linux/pagevec.h>
14 #include <linux/prefetch.h>
15 #include <linux/cleancache.h>
16 #include "extent_io.h"
17 #include "extent_map.h"
18 #include "ctree.h"
19 #include "btrfs_inode.h"
20 #include "volumes.h"
21 #include "check-integrity.h"
22 #include "locking.h"
23 #include "rcu-string.h"
24 #include "backref.h"
25 #include "disk-io.h"
26
27 static struct kmem_cache *extent_state_cache;
28 static struct kmem_cache *extent_buffer_cache;
29 static struct bio_set btrfs_bioset;
30
31 static inline bool extent_state_in_tree(const struct extent_state *state)
32 {
33         return !RB_EMPTY_NODE(&state->rb_node);
34 }
35
36 #ifdef CONFIG_BTRFS_DEBUG
37 static LIST_HEAD(buffers);
38 static LIST_HEAD(states);
39
40 static DEFINE_SPINLOCK(leak_lock);
41
42 static inline
43 void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
44 {
45         unsigned long flags;
46
47         spin_lock_irqsave(&leak_lock, flags);
48         list_add(new, head);
49         spin_unlock_irqrestore(&leak_lock, flags);
50 }
51
52 static inline
53 void btrfs_leak_debug_del(struct list_head *entry)
54 {
55         unsigned long flags;
56
57         spin_lock_irqsave(&leak_lock, flags);
58         list_del(entry);
59         spin_unlock_irqrestore(&leak_lock, flags);
60 }
61
62 static inline
63 void btrfs_leak_debug_check(void)
64 {
65         struct extent_state *state;
66         struct extent_buffer *eb;
67
68         while (!list_empty(&states)) {
69                 state = list_entry(states.next, struct extent_state, leak_list);
70                 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
71                        state->start, state->end, state->state,
72                        extent_state_in_tree(state),
73                        refcount_read(&state->refs));
74                 list_del(&state->leak_list);
75                 kmem_cache_free(extent_state_cache, state);
76         }
77
78         while (!list_empty(&buffers)) {
79                 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
80                 pr_err("BTRFS: buffer leak start %llu len %lu refs %d bflags %lu\n",
81                        eb->start, eb->len, atomic_read(&eb->refs), eb->bflags);
82                 list_del(&eb->leak_list);
83                 kmem_cache_free(extent_buffer_cache, eb);
84         }
85 }
86
87 #define btrfs_debug_check_extent_io_range(tree, start, end)             \
88         __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
89 static inline void __btrfs_debug_check_extent_io_range(const char *caller,
90                 struct extent_io_tree *tree, u64 start, u64 end)
91 {
92         struct inode *inode = tree->private_data;
93         u64 isize;
94
95         if (!inode || !is_data_inode(inode))
96                 return;
97
98         isize = i_size_read(inode);
99         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
100                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
101                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
102                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
103         }
104 }
105 #else
106 #define btrfs_leak_debug_add(new, head) do {} while (0)
107 #define btrfs_leak_debug_del(entry)     do {} while (0)
108 #define btrfs_leak_debug_check()        do {} while (0)
109 #define btrfs_debug_check_extent_io_range(c, s, e)      do {} while (0)
110 #endif
111
112 struct tree_entry {
113         u64 start;
114         u64 end;
115         struct rb_node rb_node;
116 };
117
118 struct extent_page_data {
119         struct bio *bio;
120         struct extent_io_tree *tree;
121         /* tells writepage not to lock the state bits for this range
122          * it still does the unlocking
123          */
124         unsigned int extent_locked:1;
125
126         /* tells the submit_bio code to use REQ_SYNC */
127         unsigned int sync_io:1;
128 };
129
130 static int add_extent_changeset(struct extent_state *state, unsigned bits,
131                                  struct extent_changeset *changeset,
132                                  int set)
133 {
134         int ret;
135
136         if (!changeset)
137                 return 0;
138         if (set && (state->state & bits) == bits)
139                 return 0;
140         if (!set && (state->state & bits) == 0)
141                 return 0;
142         changeset->bytes_changed += state->end - state->start + 1;
143         ret = ulist_add(&changeset->range_changed, state->start, state->end,
144                         GFP_ATOMIC);
145         return ret;
146 }
147
148 static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
149                                        unsigned long bio_flags)
150 {
151         blk_status_t ret = 0;
152         struct extent_io_tree *tree = bio->bi_private;
153
154         bio->bi_private = NULL;
155
156         if (tree->ops)
157                 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
158                                                  mirror_num, bio_flags);
159         else
160                 btrfsic_submit_bio(bio);
161
162         return blk_status_to_errno(ret);
163 }
164
165 /* Cleanup unsubmitted bios */
166 static void end_write_bio(struct extent_page_data *epd, int ret)
167 {
168         if (epd->bio) {
169                 epd->bio->bi_status = errno_to_blk_status(ret);
170                 bio_endio(epd->bio);
171                 epd->bio = NULL;
172         }
173 }
174
175 /*
176  * Submit bio from extent page data via submit_one_bio
177  *
178  * Return 0 if everything is OK.
179  * Return <0 for error.
180  */
181 static int __must_check flush_write_bio(struct extent_page_data *epd)
182 {
183         int ret = 0;
184
185         if (epd->bio) {
186                 ret = submit_one_bio(epd->bio, 0, 0);
187                 /*
188                  * Clean up of epd->bio is handled by its endio function.
189                  * And endio is either triggered by successful bio execution
190                  * or the error handler of submit bio hook.
191                  * So at this point, no matter what happened, we don't need
192                  * to clean up epd->bio.
193                  */
194                 epd->bio = NULL;
195         }
196         return ret;
197 }
198
199 int __init extent_io_init(void)
200 {
201         extent_state_cache = kmem_cache_create("btrfs_extent_state",
202                         sizeof(struct extent_state), 0,
203                         SLAB_MEM_SPREAD, NULL);
204         if (!extent_state_cache)
205                 return -ENOMEM;
206
207         extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
208                         sizeof(struct extent_buffer), 0,
209                         SLAB_MEM_SPREAD, NULL);
210         if (!extent_buffer_cache)
211                 goto free_state_cache;
212
213         if (bioset_init(&btrfs_bioset, BIO_POOL_SIZE,
214                         offsetof(struct btrfs_io_bio, bio),
215                         BIOSET_NEED_BVECS))
216                 goto free_buffer_cache;
217
218         if (bioset_integrity_create(&btrfs_bioset, BIO_POOL_SIZE))
219                 goto free_bioset;
220
221         return 0;
222
223 free_bioset:
224         bioset_exit(&btrfs_bioset);
225
226 free_buffer_cache:
227         kmem_cache_destroy(extent_buffer_cache);
228         extent_buffer_cache = NULL;
229
230 free_state_cache:
231         kmem_cache_destroy(extent_state_cache);
232         extent_state_cache = NULL;
233         return -ENOMEM;
234 }
235
236 void __cold extent_io_exit(void)
237 {
238         btrfs_leak_debug_check();
239
240         /*
241          * Make sure all delayed rcu free are flushed before we
242          * destroy caches.
243          */
244         rcu_barrier();
245         kmem_cache_destroy(extent_state_cache);
246         kmem_cache_destroy(extent_buffer_cache);
247         bioset_exit(&btrfs_bioset);
248 }
249
250 void extent_io_tree_init(struct btrfs_fs_info *fs_info,
251                          struct extent_io_tree *tree, unsigned int owner,
252                          void *private_data)
253 {
254         tree->fs_info = fs_info;
255         tree->state = RB_ROOT;
256         tree->ops = NULL;
257         tree->dirty_bytes = 0;
258         spin_lock_init(&tree->lock);
259         tree->private_data = private_data;
260         tree->owner = owner;
261 }
262
263 void extent_io_tree_release(struct extent_io_tree *tree)
264 {
265         spin_lock(&tree->lock);
266         /*
267          * Do a single barrier for the waitqueue_active check here, the state
268          * of the waitqueue should not change once extent_io_tree_release is
269          * called.
270          */
271         smp_mb();
272         while (!RB_EMPTY_ROOT(&tree->state)) {
273                 struct rb_node *node;
274                 struct extent_state *state;
275
276                 node = rb_first(&tree->state);
277                 state = rb_entry(node, struct extent_state, rb_node);
278                 rb_erase(&state->rb_node, &tree->state);
279                 RB_CLEAR_NODE(&state->rb_node);
280                 /*
281                  * btree io trees aren't supposed to have tasks waiting for
282                  * changes in the flags of extent states ever.
283                  */
284                 ASSERT(!waitqueue_active(&state->wq));
285                 free_extent_state(state);
286
287                 cond_resched_lock(&tree->lock);
288         }
289         spin_unlock(&tree->lock);
290 }
291
292 static struct extent_state *alloc_extent_state(gfp_t mask)
293 {
294         struct extent_state *state;
295
296         /*
297          * The given mask might be not appropriate for the slab allocator,
298          * drop the unsupported bits
299          */
300         mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
301         state = kmem_cache_alloc(extent_state_cache, mask);
302         if (!state)
303                 return state;
304         state->state = 0;
305         state->failrec = NULL;
306         RB_CLEAR_NODE(&state->rb_node);
307         btrfs_leak_debug_add(&state->leak_list, &states);
308         refcount_set(&state->refs, 1);
309         init_waitqueue_head(&state->wq);
310         trace_alloc_extent_state(state, mask, _RET_IP_);
311         return state;
312 }
313
314 void free_extent_state(struct extent_state *state)
315 {
316         if (!state)
317                 return;
318         if (refcount_dec_and_test(&state->refs)) {
319                 WARN_ON(extent_state_in_tree(state));
320                 btrfs_leak_debug_del(&state->leak_list);
321                 trace_free_extent_state(state, _RET_IP_);
322                 kmem_cache_free(extent_state_cache, state);
323         }
324 }
325
326 static struct rb_node *tree_insert(struct rb_root *root,
327                                    struct rb_node *search_start,
328                                    u64 offset,
329                                    struct rb_node *node,
330                                    struct rb_node ***p_in,
331                                    struct rb_node **parent_in)
332 {
333         struct rb_node **p;
334         struct rb_node *parent = NULL;
335         struct tree_entry *entry;
336
337         if (p_in && parent_in) {
338                 p = *p_in;
339                 parent = *parent_in;
340                 goto do_insert;
341         }
342
343         p = search_start ? &search_start : &root->rb_node;
344         while (*p) {
345                 parent = *p;
346                 entry = rb_entry(parent, struct tree_entry, rb_node);
347
348                 if (offset < entry->start)
349                         p = &(*p)->rb_left;
350                 else if (offset > entry->end)
351                         p = &(*p)->rb_right;
352                 else
353                         return parent;
354         }
355
356 do_insert:
357         rb_link_node(node, parent, p);
358         rb_insert_color(node, root);
359         return NULL;
360 }
361
362 /**
363  * __etree_search - searche @tree for an entry that contains @offset. Such
364  * entry would have entry->start <= offset && entry->end >= offset.
365  *
366  * @tree - the tree to search
367  * @offset - offset that should fall within an entry in @tree
368  * @next_ret - pointer to the first entry whose range ends after @offset
369  * @prev - pointer to the first entry whose range begins before @offset
370  * @p_ret - pointer where new node should be anchored (used when inserting an
371  *          entry in the tree)
372  * @parent_ret - points to entry which would have been the parent of the entry,
373  *               containing @offset
374  *
375  * This function returns a pointer to the entry that contains @offset byte
376  * address. If no such entry exists, then NULL is returned and the other
377  * pointer arguments to the function are filled, otherwise the found entry is
378  * returned and other pointers are left untouched.
379  */
380 static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
381                                       struct rb_node **next_ret,
382                                       struct rb_node **prev_ret,
383                                       struct rb_node ***p_ret,
384                                       struct rb_node **parent_ret)
385 {
386         struct rb_root *root = &tree->state;
387         struct rb_node **n = &root->rb_node;
388         struct rb_node *prev = NULL;
389         struct rb_node *orig_prev = NULL;
390         struct tree_entry *entry;
391         struct tree_entry *prev_entry = NULL;
392
393         while (*n) {
394                 prev = *n;
395                 entry = rb_entry(prev, struct tree_entry, rb_node);
396                 prev_entry = entry;
397
398                 if (offset < entry->start)
399                         n = &(*n)->rb_left;
400                 else if (offset > entry->end)
401                         n = &(*n)->rb_right;
402                 else
403                         return *n;
404         }
405
406         if (p_ret)
407                 *p_ret = n;
408         if (parent_ret)
409                 *parent_ret = prev;
410
411         if (next_ret) {
412                 orig_prev = prev;
413                 while (prev && offset > prev_entry->end) {
414                         prev = rb_next(prev);
415                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
416                 }
417                 *next_ret = prev;
418                 prev = orig_prev;
419         }
420
421         if (prev_ret) {
422                 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
423                 while (prev && offset < prev_entry->start) {
424                         prev = rb_prev(prev);
425                         prev_entry = rb_entry(prev, struct tree_entry, rb_node);
426                 }
427                 *prev_ret = prev;
428         }
429         return NULL;
430 }
431
432 static inline struct rb_node *
433 tree_search_for_insert(struct extent_io_tree *tree,
434                        u64 offset,
435                        struct rb_node ***p_ret,
436                        struct rb_node **parent_ret)
437 {
438         struct rb_node *next= NULL;
439         struct rb_node *ret;
440
441         ret = __etree_search(tree, offset, &next, NULL, p_ret, parent_ret);
442         if (!ret)
443                 return next;
444         return ret;
445 }
446
447 static inline struct rb_node *tree_search(struct extent_io_tree *tree,
448                                           u64 offset)
449 {
450         return tree_search_for_insert(tree, offset, NULL, NULL);
451 }
452
453 /*
454  * utility function to look for merge candidates inside a given range.
455  * Any extents with matching state are merged together into a single
456  * extent in the tree.  Extents with EXTENT_IO in their state field
457  * are not merged because the end_io handlers need to be able to do
458  * operations on them without sleeping (or doing allocations/splits).
459  *
460  * This should be called with the tree lock held.
461  */
462 static void merge_state(struct extent_io_tree *tree,
463                         struct extent_state *state)
464 {
465         struct extent_state *other;
466         struct rb_node *other_node;
467
468         if (state->state & (EXTENT_LOCKED | EXTENT_BOUNDARY))
469                 return;
470
471         other_node = rb_prev(&state->rb_node);
472         if (other_node) {
473                 other = rb_entry(other_node, struct extent_state, rb_node);
474                 if (other->end == state->start - 1 &&
475                     other->state == state->state) {
476                         if (tree->private_data &&
477                             is_data_inode(tree->private_data))
478                                 btrfs_merge_delalloc_extent(tree->private_data,
479                                                             state, other);
480                         state->start = other->start;
481                         rb_erase(&other->rb_node, &tree->state);
482                         RB_CLEAR_NODE(&other->rb_node);
483                         free_extent_state(other);
484                 }
485         }
486         other_node = rb_next(&state->rb_node);
487         if (other_node) {
488                 other = rb_entry(other_node, struct extent_state, rb_node);
489                 if (other->start == state->end + 1 &&
490                     other->state == state->state) {
491                         if (tree->private_data &&
492                             is_data_inode(tree->private_data))
493                                 btrfs_merge_delalloc_extent(tree->private_data,
494                                                             state, other);
495                         state->end = other->end;
496                         rb_erase(&other->rb_node, &tree->state);
497                         RB_CLEAR_NODE(&other->rb_node);
498                         free_extent_state(other);
499                 }
500         }
501 }
502
503 static void set_state_bits(struct extent_io_tree *tree,
504                            struct extent_state *state, unsigned *bits,
505                            struct extent_changeset *changeset);
506
507 /*
508  * insert an extent_state struct into the tree.  'bits' are set on the
509  * struct before it is inserted.
510  *
511  * This may return -EEXIST if the extent is already there, in which case the
512  * state struct is freed.
513  *
514  * The tree lock is not taken internally.  This is a utility function and
515  * probably isn't what you want to call (see set/clear_extent_bit).
516  */
517 static int insert_state(struct extent_io_tree *tree,
518                         struct extent_state *state, u64 start, u64 end,
519                         struct rb_node ***p,
520                         struct rb_node **parent,
521                         unsigned *bits, struct extent_changeset *changeset)
522 {
523         struct rb_node *node;
524
525         if (end < start) {
526                 btrfs_err(tree->fs_info,
527                         "insert state: end < start %llu %llu", end, start);
528                 WARN_ON(1);
529         }
530         state->start = start;
531         state->end = end;
532
533         set_state_bits(tree, state, bits, changeset);
534
535         node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
536         if (node) {
537                 struct extent_state *found;
538                 found = rb_entry(node, struct extent_state, rb_node);
539                 btrfs_err(tree->fs_info,
540                        "found node %llu %llu on insert of %llu %llu",
541                        found->start, found->end, start, end);
542                 return -EEXIST;
543         }
544         merge_state(tree, state);
545         return 0;
546 }
547
548 /*
549  * split a given extent state struct in two, inserting the preallocated
550  * struct 'prealloc' as the newly created second half.  'split' indicates an
551  * offset inside 'orig' where it should be split.
552  *
553  * Before calling,
554  * the tree has 'orig' at [orig->start, orig->end].  After calling, there
555  * are two extent state structs in the tree:
556  * prealloc: [orig->start, split - 1]
557  * orig: [ split, orig->end ]
558  *
559  * The tree locks are not taken by this function. They need to be held
560  * by the caller.
561  */
562 static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
563                        struct extent_state *prealloc, u64 split)
564 {
565         struct rb_node *node;
566
567         if (tree->private_data && is_data_inode(tree->private_data))
568                 btrfs_split_delalloc_extent(tree->private_data, orig, split);
569
570         prealloc->start = orig->start;
571         prealloc->end = split - 1;
572         prealloc->state = orig->state;
573         orig->start = split;
574
575         node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
576                            &prealloc->rb_node, NULL, NULL);
577         if (node) {
578                 free_extent_state(prealloc);
579                 return -EEXIST;
580         }
581         return 0;
582 }
583
584 static struct extent_state *next_state(struct extent_state *state)
585 {
586         struct rb_node *next = rb_next(&state->rb_node);
587         if (next)
588                 return rb_entry(next, struct extent_state, rb_node);
589         else
590                 return NULL;
591 }
592
593 /*
594  * utility function to clear some bits in an extent state struct.
595  * it will optionally wake up anyone waiting on this state (wake == 1).
596  *
597  * If no bits are set on the state struct after clearing things, the
598  * struct is freed and removed from the tree
599  */
600 static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
601                                             struct extent_state *state,
602                                             unsigned *bits, int wake,
603                                             struct extent_changeset *changeset)
604 {
605         struct extent_state *next;
606         unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
607         int ret;
608
609         if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
610                 u64 range = state->end - state->start + 1;
611                 WARN_ON(range > tree->dirty_bytes);
612                 tree->dirty_bytes -= range;
613         }
614
615         if (tree->private_data && is_data_inode(tree->private_data))
616                 btrfs_clear_delalloc_extent(tree->private_data, state, bits);
617
618         ret = add_extent_changeset(state, bits_to_clear, changeset, 0);
619         BUG_ON(ret < 0);
620         state->state &= ~bits_to_clear;
621         if (wake)
622                 wake_up(&state->wq);
623         if (state->state == 0) {
624                 next = next_state(state);
625                 if (extent_state_in_tree(state)) {
626                         rb_erase(&state->rb_node, &tree->state);
627                         RB_CLEAR_NODE(&state->rb_node);
628                         free_extent_state(state);
629                 } else {
630                         WARN_ON(1);
631                 }
632         } else {
633                 merge_state(tree, state);
634                 next = next_state(state);
635         }
636         return next;
637 }
638
639 static struct extent_state *
640 alloc_extent_state_atomic(struct extent_state *prealloc)
641 {
642         if (!prealloc)
643                 prealloc = alloc_extent_state(GFP_ATOMIC);
644
645         return prealloc;
646 }
647
648 static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
649 {
650         struct inode *inode = tree->private_data;
651
652         btrfs_panic(btrfs_sb(inode->i_sb), err,
653         "locking error: extent tree was modified by another thread while locked");
654 }
655
656 /*
657  * clear some bits on a range in the tree.  This may require splitting
658  * or inserting elements in the tree, so the gfp mask is used to
659  * indicate which allocations or sleeping are allowed.
660  *
661  * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
662  * the given range from the tree regardless of state (ie for truncate).
663  *
664  * the range [start, end] is inclusive.
665  *
666  * This takes the tree lock, and returns 0 on success and < 0 on error.
667  */
668 int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
669                               unsigned bits, int wake, int delete,
670                               struct extent_state **cached_state,
671                               gfp_t mask, struct extent_changeset *changeset)
672 {
673         struct extent_state *state;
674         struct extent_state *cached;
675         struct extent_state *prealloc = NULL;
676         struct rb_node *node;
677         u64 last_end;
678         int err;
679         int clear = 0;
680
681         btrfs_debug_check_extent_io_range(tree, start, end);
682         trace_btrfs_clear_extent_bit(tree, start, end - start + 1, bits);
683
684         if (bits & EXTENT_DELALLOC)
685                 bits |= EXTENT_NORESERVE;
686
687         if (delete)
688                 bits |= ~EXTENT_CTLBITS;
689
690         if (bits & (EXTENT_LOCKED | EXTENT_BOUNDARY))
691                 clear = 1;
692 again:
693         if (!prealloc && gfpflags_allow_blocking(mask)) {
694                 /*
695                  * Don't care for allocation failure here because we might end
696                  * up not needing the pre-allocated extent state at all, which
697                  * is the case if we only have in the tree extent states that
698                  * cover our input range and don't cover too any other range.
699                  * If we end up needing a new extent state we allocate it later.
700                  */
701                 prealloc = alloc_extent_state(mask);
702         }
703
704         spin_lock(&tree->lock);
705         if (cached_state) {
706                 cached = *cached_state;
707
708                 if (clear) {
709                         *cached_state = NULL;
710                         cached_state = NULL;
711                 }
712
713                 if (cached && extent_state_in_tree(cached) &&
714                     cached->start <= start && cached->end > start) {
715                         if (clear)
716                                 refcount_dec(&cached->refs);
717                         state = cached;
718                         goto hit_next;
719                 }
720                 if (clear)
721                         free_extent_state(cached);
722         }
723         /*
724          * this search will find the extents that end after
725          * our range starts
726          */
727         node = tree_search(tree, start);
728         if (!node)
729                 goto out;
730         state = rb_entry(node, struct extent_state, rb_node);
731 hit_next:
732         if (state->start > end)
733                 goto out;
734         WARN_ON(state->end < start);
735         last_end = state->end;
736
737         /* the state doesn't have the wanted bits, go ahead */
738         if (!(state->state & bits)) {
739                 state = next_state(state);
740                 goto next;
741         }
742
743         /*
744          *     | ---- desired range ---- |
745          *  | state | or
746          *  | ------------- state -------------- |
747          *
748          * We need to split the extent we found, and may flip
749          * bits on second half.
750          *
751          * If the extent we found extends past our range, we
752          * just split and search again.  It'll get split again
753          * the next time though.
754          *
755          * If the extent we found is inside our range, we clear
756          * the desired bit on it.
757          */
758
759         if (state->start < start) {
760                 prealloc = alloc_extent_state_atomic(prealloc);
761                 BUG_ON(!prealloc);
762                 err = split_state(tree, state, prealloc, start);
763                 if (err)
764                         extent_io_tree_panic(tree, err);
765
766                 prealloc = NULL;
767                 if (err)
768                         goto out;
769                 if (state->end <= end) {
770                         state = clear_state_bit(tree, state, &bits, wake,
771                                                 changeset);
772                         goto next;
773                 }
774                 goto search_again;
775         }
776         /*
777          * | ---- desired range ---- |
778          *                        | state |
779          * We need to split the extent, and clear the bit
780          * on the first half
781          */
782         if (state->start <= end && state->end > end) {
783                 prealloc = alloc_extent_state_atomic(prealloc);
784                 BUG_ON(!prealloc);
785                 err = split_state(tree, state, prealloc, end + 1);
786                 if (err)
787                         extent_io_tree_panic(tree, err);
788
789                 if (wake)
790                         wake_up(&state->wq);
791
792                 clear_state_bit(tree, prealloc, &bits, wake, changeset);
793
794                 prealloc = NULL;
795                 goto out;
796         }
797
798         state = clear_state_bit(tree, state, &bits, wake, changeset);
799 next:
800         if (last_end == (u64)-1)
801                 goto out;
802         start = last_end + 1;
803         if (start <= end && state && !need_resched())
804                 goto hit_next;
805
806 search_again:
807         if (start > end)
808                 goto out;
809         spin_unlock(&tree->lock);
810         if (gfpflags_allow_blocking(mask))
811                 cond_resched();
812         goto again;
813
814 out:
815         spin_unlock(&tree->lock);
816         if (prealloc)
817                 free_extent_state(prealloc);
818
819         return 0;
820
821 }
822
823 static void wait_on_state(struct extent_io_tree *tree,
824                           struct extent_state *state)
825                 __releases(tree->lock)
826                 __acquires(tree->lock)
827 {
828         DEFINE_WAIT(wait);
829         prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
830         spin_unlock(&tree->lock);
831         schedule();
832         spin_lock(&tree->lock);
833         finish_wait(&state->wq, &wait);
834 }
835
836 /*
837  * waits for one or more bits to clear on a range in the state tree.
838  * The range [start, end] is inclusive.
839  * The tree lock is taken by this function
840  */
841 static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
842                             unsigned long bits)
843 {
844         struct extent_state *state;
845         struct rb_node *node;
846
847         btrfs_debug_check_extent_io_range(tree, start, end);
848
849         spin_lock(&tree->lock);
850 again:
851         while (1) {
852                 /*
853                  * this search will find all the extents that end after
854                  * our range starts
855                  */
856                 node = tree_search(tree, start);
857 process_node:
858                 if (!node)
859                         break;
860
861                 state = rb_entry(node, struct extent_state, rb_node);
862
863                 if (state->start > end)
864                         goto out;
865
866                 if (state->state & bits) {
867                         start = state->start;
868                         refcount_inc(&state->refs);
869                         wait_on_state(tree, state);
870                         free_extent_state(state);
871                         goto again;
872                 }
873                 start = state->end + 1;
874
875                 if (start > end)
876                         break;
877
878                 if (!cond_resched_lock(&tree->lock)) {
879                         node = rb_next(node);
880                         goto process_node;
881                 }
882         }
883 out:
884         spin_unlock(&tree->lock);
885 }
886
887 static void set_state_bits(struct extent_io_tree *tree,
888                            struct extent_state *state,
889                            unsigned *bits, struct extent_changeset *changeset)
890 {
891         unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
892         int ret;
893
894         if (tree->private_data && is_data_inode(tree->private_data))
895                 btrfs_set_delalloc_extent(tree->private_data, state, bits);
896
897         if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
898                 u64 range = state->end - state->start + 1;
899                 tree->dirty_bytes += range;
900         }
901         ret = add_extent_changeset(state, bits_to_set, changeset, 1);
902         BUG_ON(ret < 0);
903         state->state |= bits_to_set;
904 }
905
906 static void cache_state_if_flags(struct extent_state *state,
907                                  struct extent_state **cached_ptr,
908                                  unsigned flags)
909 {
910         if (cached_ptr && !(*cached_ptr)) {
911                 if (!flags || (state->state & flags)) {
912                         *cached_ptr = state;
913                         refcount_inc(&state->refs);
914                 }
915         }
916 }
917
918 static void cache_state(struct extent_state *state,
919                         struct extent_state **cached_ptr)
920 {
921         return cache_state_if_flags(state, cached_ptr,
922                                     EXTENT_LOCKED | EXTENT_BOUNDARY);
923 }
924
925 /*
926  * set some bits on a range in the tree.  This may require allocations or
927  * sleeping, so the gfp mask is used to indicate what is allowed.
928  *
929  * If any of the exclusive bits are set, this will fail with -EEXIST if some
930  * part of the range already has the desired bits set.  The start of the
931  * existing range is returned in failed_start in this case.
932  *
933  * [start, end] is inclusive This takes the tree lock.
934  */
935
936 static int __must_check
937 __set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
938                  unsigned bits, unsigned exclusive_bits,
939                  u64 *failed_start, struct extent_state **cached_state,
940                  gfp_t mask, struct extent_changeset *changeset)
941 {
942         struct extent_state *state;
943         struct extent_state *prealloc = NULL;
944         struct rb_node *node;
945         struct rb_node **p;
946         struct rb_node *parent;
947         int err = 0;
948         u64 last_start;
949         u64 last_end;
950
951         btrfs_debug_check_extent_io_range(tree, start, end);
952         trace_btrfs_set_extent_bit(tree, start, end - start + 1, bits);
953
954 again:
955         if (!prealloc && gfpflags_allow_blocking(mask)) {
956                 /*
957                  * Don't care for allocation failure here because we might end
958                  * up not needing the pre-allocated extent state at all, which
959                  * is the case if we only have in the tree extent states that
960                  * cover our input range and don't cover too any other range.
961                  * If we end up needing a new extent state we allocate it later.
962                  */
963                 prealloc = alloc_extent_state(mask);
964         }
965
966         spin_lock(&tree->lock);
967         if (cached_state && *cached_state) {
968                 state = *cached_state;
969                 if (state->start <= start && state->end > start &&
970                     extent_state_in_tree(state)) {
971                         node = &state->rb_node;
972                         goto hit_next;
973                 }
974         }
975         /*
976          * this search will find all the extents that end after
977          * our range starts.
978          */
979         node = tree_search_for_insert(tree, start, &p, &parent);
980         if (!node) {
981                 prealloc = alloc_extent_state_atomic(prealloc);
982                 BUG_ON(!prealloc);
983                 err = insert_state(tree, prealloc, start, end,
984                                    &p, &parent, &bits, changeset);
985                 if (err)
986                         extent_io_tree_panic(tree, err);
987
988                 cache_state(prealloc, cached_state);
989                 prealloc = NULL;
990                 goto out;
991         }
992         state = rb_entry(node, struct extent_state, rb_node);
993 hit_next:
994         last_start = state->start;
995         last_end = state->end;
996
997         /*
998          * | ---- desired range ---- |
999          * | state |
1000          *
1001          * Just lock what we found and keep going
1002          */
1003         if (state->start == start && state->end <= end) {
1004                 if (state->state & exclusive_bits) {
1005                         *failed_start = state->start;
1006                         err = -EEXIST;
1007                         goto out;
1008                 }
1009
1010                 set_state_bits(tree, state, &bits, changeset);
1011                 cache_state(state, cached_state);
1012                 merge_state(tree, state);
1013                 if (last_end == (u64)-1)
1014                         goto out;
1015                 start = last_end + 1;
1016                 state = next_state(state);
1017                 if (start < end && state && state->start == start &&
1018                     !need_resched())
1019                         goto hit_next;
1020                 goto search_again;
1021         }
1022
1023         /*
1024          *     | ---- desired range ---- |
1025          * | state |
1026          *   or
1027          * | ------------- state -------------- |
1028          *
1029          * We need to split the extent we found, and may flip bits on
1030          * second half.
1031          *
1032          * If the extent we found extends past our
1033          * range, we just split and search again.  It'll get split
1034          * again the next time though.
1035          *
1036          * If the extent we found is inside our range, we set the
1037          * desired bit on it.
1038          */
1039         if (state->start < start) {
1040                 if (state->state & exclusive_bits) {
1041                         *failed_start = start;
1042                         err = -EEXIST;
1043                         goto out;
1044                 }
1045
1046                 prealloc = alloc_extent_state_atomic(prealloc);
1047                 BUG_ON(!prealloc);
1048                 err = split_state(tree, state, prealloc, start);
1049                 if (err)
1050                         extent_io_tree_panic(tree, err);
1051
1052                 prealloc = NULL;
1053                 if (err)
1054                         goto out;
1055                 if (state->end <= end) {
1056                         set_state_bits(tree, state, &bits, changeset);
1057                         cache_state(state, cached_state);
1058                         merge_state(tree, state);
1059                         if (last_end == (u64)-1)
1060                                 goto out;
1061                         start = last_end + 1;
1062                         state = next_state(state);
1063                         if (start < end && state && state->start == start &&
1064                             !need_resched())
1065                                 goto hit_next;
1066                 }
1067                 goto search_again;
1068         }
1069         /*
1070          * | ---- desired range ---- |
1071          *     | state | or               | state |
1072          *
1073          * There's a hole, we need to insert something in it and
1074          * ignore the extent we found.
1075          */
1076         if (state->start > start) {
1077                 u64 this_end;
1078                 if (end < last_start)
1079                         this_end = end;
1080                 else
1081                         this_end = last_start - 1;
1082
1083                 prealloc = alloc_extent_state_atomic(prealloc);
1084                 BUG_ON(!prealloc);
1085
1086                 /*
1087                  * Avoid to free 'prealloc' if it can be merged with
1088                  * the later extent.
1089                  */
1090                 err = insert_state(tree, prealloc, start, this_end,
1091                                    NULL, NULL, &bits, changeset);
1092                 if (err)
1093                         extent_io_tree_panic(tree, err);
1094
1095                 cache_state(prealloc, cached_state);
1096                 prealloc = NULL;
1097                 start = this_end + 1;
1098                 goto search_again;
1099         }
1100         /*
1101          * | ---- desired range ---- |
1102          *                        | state |
1103          * We need to split the extent, and set the bit
1104          * on the first half
1105          */
1106         if (state->start <= end && state->end > end) {
1107                 if (state->state & exclusive_bits) {
1108                         *failed_start = start;
1109                         err = -EEXIST;
1110                         goto out;
1111                 }
1112
1113                 prealloc = alloc_extent_state_atomic(prealloc);
1114                 BUG_ON(!prealloc);
1115                 err = split_state(tree, state, prealloc, end + 1);
1116                 if (err)
1117                         extent_io_tree_panic(tree, err);
1118
1119                 set_state_bits(tree, prealloc, &bits, changeset);
1120                 cache_state(prealloc, cached_state);
1121                 merge_state(tree, prealloc);
1122                 prealloc = NULL;
1123                 goto out;
1124         }
1125
1126 search_again:
1127         if (start > end)
1128                 goto out;
1129         spin_unlock(&tree->lock);
1130         if (gfpflags_allow_blocking(mask))
1131                 cond_resched();
1132         goto again;
1133
1134 out:
1135         spin_unlock(&tree->lock);
1136         if (prealloc)
1137                 free_extent_state(prealloc);
1138
1139         return err;
1140
1141 }
1142
1143 int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1144                    unsigned bits, u64 * failed_start,
1145                    struct extent_state **cached_state, gfp_t mask)
1146 {
1147         return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1148                                 cached_state, mask, NULL);
1149 }
1150
1151
1152 /**
1153  * convert_extent_bit - convert all bits in a given range from one bit to
1154  *                      another
1155  * @tree:       the io tree to search
1156  * @start:      the start offset in bytes
1157  * @end:        the end offset in bytes (inclusive)
1158  * @bits:       the bits to set in this range
1159  * @clear_bits: the bits to clear in this range
1160  * @cached_state:       state that we're going to cache
1161  *
1162  * This will go through and set bits for the given range.  If any states exist
1163  * already in this range they are set with the given bit and cleared of the
1164  * clear_bits.  This is only meant to be used by things that are mergeable, ie
1165  * converting from say DELALLOC to DIRTY.  This is not meant to be used with
1166  * boundary bits like LOCK.
1167  *
1168  * All allocations are done with GFP_NOFS.
1169  */
1170 int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1171                        unsigned bits, unsigned clear_bits,
1172                        struct extent_state **cached_state)
1173 {
1174         struct extent_state *state;
1175         struct extent_state *prealloc = NULL;
1176         struct rb_node *node;
1177         struct rb_node **p;
1178         struct rb_node *parent;
1179         int err = 0;
1180         u64 last_start;
1181         u64 last_end;
1182         bool first_iteration = true;
1183
1184         btrfs_debug_check_extent_io_range(tree, start, end);
1185         trace_btrfs_convert_extent_bit(tree, start, end - start + 1, bits,
1186                                        clear_bits);
1187
1188 again:
1189         if (!prealloc) {
1190                 /*
1191                  * Best effort, don't worry if extent state allocation fails
1192                  * here for the first iteration. We might have a cached state
1193                  * that matches exactly the target range, in which case no
1194                  * extent state allocations are needed. We'll only know this
1195                  * after locking the tree.
1196                  */
1197                 prealloc = alloc_extent_state(GFP_NOFS);
1198                 if (!prealloc && !first_iteration)
1199                         return -ENOMEM;
1200         }
1201
1202         spin_lock(&tree->lock);
1203         if (cached_state && *cached_state) {
1204                 state = *cached_state;
1205                 if (state->start <= start && state->end > start &&
1206                     extent_state_in_tree(state)) {
1207                         node = &state->rb_node;
1208                         goto hit_next;
1209                 }
1210         }
1211
1212         /*
1213          * this search will find all the extents that end after
1214          * our range starts.
1215          */
1216         node = tree_search_for_insert(tree, start, &p, &parent);
1217         if (!node) {
1218                 prealloc = alloc_extent_state_atomic(prealloc);
1219                 if (!prealloc) {
1220                         err = -ENOMEM;
1221                         goto out;
1222                 }
1223                 err = insert_state(tree, prealloc, start, end,
1224                                    &p, &parent, &bits, NULL);
1225                 if (err)
1226                         extent_io_tree_panic(tree, err);
1227                 cache_state(prealloc, cached_state);
1228                 prealloc = NULL;
1229                 goto out;
1230         }
1231         state = rb_entry(node, struct extent_state, rb_node);
1232 hit_next:
1233         last_start = state->start;
1234         last_end = state->end;
1235
1236         /*
1237          * | ---- desired range ---- |
1238          * | state |
1239          *
1240          * Just lock what we found and keep going
1241          */
1242         if (state->start == start && state->end <= end) {
1243                 set_state_bits(tree, state, &bits, NULL);
1244                 cache_state(state, cached_state);
1245                 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1246                 if (last_end == (u64)-1)
1247                         goto out;
1248                 start = last_end + 1;
1249                 if (start < end && state && state->start == start &&
1250                     !need_resched())
1251                         goto hit_next;
1252                 goto search_again;
1253         }
1254
1255         /*
1256          *     | ---- desired range ---- |
1257          * | state |
1258          *   or
1259          * | ------------- state -------------- |
1260          *
1261          * We need to split the extent we found, and may flip bits on
1262          * second half.
1263          *
1264          * If the extent we found extends past our
1265          * range, we just split and search again.  It'll get split
1266          * again the next time though.
1267          *
1268          * If the extent we found is inside our range, we set the
1269          * desired bit on it.
1270          */
1271         if (state->start < start) {
1272                 prealloc = alloc_extent_state_atomic(prealloc);
1273                 if (!prealloc) {
1274                         err = -ENOMEM;
1275                         goto out;
1276                 }
1277                 err = split_state(tree, state, prealloc, start);
1278                 if (err)
1279                         extent_io_tree_panic(tree, err);
1280                 prealloc = NULL;
1281                 if (err)
1282                         goto out;
1283                 if (state->end <= end) {
1284                         set_state_bits(tree, state, &bits, NULL);
1285                         cache_state(state, cached_state);
1286                         state = clear_state_bit(tree, state, &clear_bits, 0,
1287                                                 NULL);
1288                         if (last_end == (u64)-1)
1289                                 goto out;
1290                         start = last_end + 1;
1291                         if (start < end && state && state->start == start &&
1292                             !need_resched())
1293                                 goto hit_next;
1294                 }
1295                 goto search_again;
1296         }
1297         /*
1298          * | ---- desired range ---- |
1299          *     | state | or               | state |
1300          *
1301          * There's a hole, we need to insert something in it and
1302          * ignore the extent we found.
1303          */
1304         if (state->start > start) {
1305                 u64 this_end;
1306                 if (end < last_start)
1307                         this_end = end;
1308                 else
1309                         this_end = last_start - 1;
1310
1311                 prealloc = alloc_extent_state_atomic(prealloc);
1312                 if (!prealloc) {
1313                         err = -ENOMEM;
1314                         goto out;
1315                 }
1316
1317                 /*
1318                  * Avoid to free 'prealloc' if it can be merged with
1319                  * the later extent.
1320                  */
1321                 err = insert_state(tree, prealloc, start, this_end,
1322                                    NULL, NULL, &bits, NULL);
1323                 if (err)
1324                         extent_io_tree_panic(tree, err);
1325                 cache_state(prealloc, cached_state);
1326                 prealloc = NULL;
1327                 start = this_end + 1;
1328                 goto search_again;
1329         }
1330         /*
1331          * | ---- desired range ---- |
1332          *                        | state |
1333          * We need to split the extent, and set the bit
1334          * on the first half
1335          */
1336         if (state->start <= end && state->end > end) {
1337                 prealloc = alloc_extent_state_atomic(prealloc);
1338                 if (!prealloc) {
1339                         err = -ENOMEM;
1340                         goto out;
1341                 }
1342
1343                 err = split_state(tree, state, prealloc, end + 1);
1344                 if (err)
1345                         extent_io_tree_panic(tree, err);
1346
1347                 set_state_bits(tree, prealloc, &bits, NULL);
1348                 cache_state(prealloc, cached_state);
1349                 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1350                 prealloc = NULL;
1351                 goto out;
1352         }
1353
1354 search_again:
1355         if (start > end)
1356                 goto out;
1357         spin_unlock(&tree->lock);
1358         cond_resched();
1359         first_iteration = false;
1360         goto again;
1361
1362 out:
1363         spin_unlock(&tree->lock);
1364         if (prealloc)
1365                 free_extent_state(prealloc);
1366
1367         return err;
1368 }
1369
1370 /* wrappers around set/clear extent bit */
1371 int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1372                            unsigned bits, struct extent_changeset *changeset)
1373 {
1374         /*
1375          * We don't support EXTENT_LOCKED yet, as current changeset will
1376          * record any bits changed, so for EXTENT_LOCKED case, it will
1377          * either fail with -EEXIST or changeset will record the whole
1378          * range.
1379          */
1380         BUG_ON(bits & EXTENT_LOCKED);
1381
1382         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1383                                 changeset);
1384 }
1385
1386 int set_extent_bits_nowait(struct extent_io_tree *tree, u64 start, u64 end,
1387                            unsigned bits)
1388 {
1389         return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL,
1390                                 GFP_NOWAIT, NULL);
1391 }
1392
1393 int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1394                      unsigned bits, int wake, int delete,
1395                      struct extent_state **cached)
1396 {
1397         return __clear_extent_bit(tree, start, end, bits, wake, delete,
1398                                   cached, GFP_NOFS, NULL);
1399 }
1400
1401 int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1402                 unsigned bits, struct extent_changeset *changeset)
1403 {
1404         /*
1405          * Don't support EXTENT_LOCKED case, same reason as
1406          * set_record_extent_bits().
1407          */
1408         BUG_ON(bits & EXTENT_LOCKED);
1409
1410         return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1411                                   changeset);
1412 }
1413
1414 /*
1415  * either insert or lock state struct between start and end use mask to tell
1416  * us if waiting is desired.
1417  */
1418 int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1419                      struct extent_state **cached_state)
1420 {
1421         int err;
1422         u64 failed_start;
1423
1424         while (1) {
1425                 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1426                                        EXTENT_LOCKED, &failed_start,
1427                                        cached_state, GFP_NOFS, NULL);
1428                 if (err == -EEXIST) {
1429                         wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1430                         start = failed_start;
1431                 } else
1432                         break;
1433                 WARN_ON(start > end);
1434         }
1435         return err;
1436 }
1437
1438 int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1439 {
1440         int err;
1441         u64 failed_start;
1442
1443         err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1444                                &failed_start, NULL, GFP_NOFS, NULL);
1445         if (err == -EEXIST) {
1446                 if (failed_start > start)
1447                         clear_extent_bit(tree, start, failed_start - 1,
1448                                          EXTENT_LOCKED, 1, 0, NULL);
1449                 return 0;
1450         }
1451         return 1;
1452 }
1453
1454 void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1455 {
1456         unsigned long index = start >> PAGE_SHIFT;
1457         unsigned long end_index = end >> PAGE_SHIFT;
1458         struct page *page;
1459
1460         while (index <= end_index) {
1461                 page = find_get_page(inode->i_mapping, index);
1462                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1463                 clear_page_dirty_for_io(page);
1464                 put_page(page);
1465                 index++;
1466         }
1467 }
1468
1469 void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1470 {
1471         unsigned long index = start >> PAGE_SHIFT;
1472         unsigned long end_index = end >> PAGE_SHIFT;
1473         struct page *page;
1474
1475         while (index <= end_index) {
1476                 page = find_get_page(inode->i_mapping, index);
1477                 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1478                 __set_page_dirty_nobuffers(page);
1479                 account_page_redirty(page);
1480                 put_page(page);
1481                 index++;
1482         }
1483 }
1484
1485 /* find the first state struct with 'bits' set after 'start', and
1486  * return it.  tree->lock must be held.  NULL will returned if
1487  * nothing was found after 'start'
1488  */
1489 static struct extent_state *
1490 find_first_extent_bit_state(struct extent_io_tree *tree,
1491                             u64 start, unsigned bits)
1492 {
1493         struct rb_node *node;
1494         struct extent_state *state;
1495
1496         /*
1497          * this search will find all the extents that end after
1498          * our range starts.
1499          */
1500         node = tree_search(tree, start);
1501         if (!node)
1502                 goto out;
1503
1504         while (1) {
1505                 state = rb_entry(node, struct extent_state, rb_node);
1506                 if (state->end >= start && (state->state & bits))
1507                         return state;
1508
1509                 node = rb_next(node);
1510                 if (!node)
1511                         break;
1512         }
1513 out:
1514         return NULL;
1515 }
1516
1517 /*
1518  * find the first offset in the io tree with 'bits' set. zero is
1519  * returned if we find something, and *start_ret and *end_ret are
1520  * set to reflect the state struct that was found.
1521  *
1522  * If nothing was found, 1 is returned. If found something, return 0.
1523  */
1524 int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1525                           u64 *start_ret, u64 *end_ret, unsigned bits,
1526                           struct extent_state **cached_state)
1527 {
1528         struct extent_state *state;
1529         int ret = 1;
1530
1531         spin_lock(&tree->lock);
1532         if (cached_state && *cached_state) {
1533                 state = *cached_state;
1534                 if (state->end == start - 1 && extent_state_in_tree(state)) {
1535                         while ((state = next_state(state)) != NULL) {
1536                                 if (state->state & bits)
1537                                         goto got_it;
1538                         }
1539                         free_extent_state(*cached_state);
1540                         *cached_state = NULL;
1541                         goto out;
1542                 }
1543                 free_extent_state(*cached_state);
1544                 *cached_state = NULL;
1545         }
1546
1547         state = find_first_extent_bit_state(tree, start, bits);
1548 got_it:
1549         if (state) {
1550                 cache_state_if_flags(state, cached_state, 0);
1551                 *start_ret = state->start;
1552                 *end_ret = state->end;
1553                 ret = 0;
1554         }
1555 out:
1556         spin_unlock(&tree->lock);
1557         return ret;
1558 }
1559
1560 /**
1561  * find_first_clear_extent_bit - find the first range that has @bits not set.
1562  * This range could start before @start.
1563  *
1564  * @tree - the tree to search
1565  * @start - the offset at/after which the found extent should start
1566  * @start_ret - records the beginning of the range
1567  * @end_ret - records the end of the range (inclusive)
1568  * @bits - the set of bits which must be unset
1569  *
1570  * Since unallocated range is also considered one which doesn't have the bits
1571  * set it's possible that @end_ret contains -1, this happens in case the range
1572  * spans (last_range_end, end of device]. In this case it's up to the caller to
1573  * trim @end_ret to the appropriate size.
1574  */
1575 void find_first_clear_extent_bit(struct extent_io_tree *tree, u64 start,
1576                                  u64 *start_ret, u64 *end_ret, unsigned bits)
1577 {
1578         struct extent_state *state;
1579         struct rb_node *node, *prev = NULL, *next;
1580
1581         spin_lock(&tree->lock);
1582
1583         /* Find first extent with bits cleared */
1584         while (1) {
1585                 node = __etree_search(tree, start, &next, &prev, NULL, NULL);
1586                 if (!node) {
1587                         node = next;
1588                         if (!node) {
1589                                 /*
1590                                  * We are past the last allocated chunk,
1591                                  * set start at the end of the last extent. The
1592                                  * device alloc tree should never be empty so
1593                                  * prev is always set.
1594                                  */
1595                                 ASSERT(prev);
1596                                 state = rb_entry(prev, struct extent_state, rb_node);
1597                                 *start_ret = state->end + 1;
1598                                 *end_ret = -1;
1599                                 goto out;
1600                         }
1601                 }
1602                 /*
1603                  * At this point 'node' either contains 'start' or start is
1604                  * before 'node'
1605                  */
1606                 state = rb_entry(node, struct extent_state, rb_node);
1607
1608                 if (in_range(start, state->start, state->end - state->start + 1)) {
1609                         if (state->state & bits) {
1610                                 /*
1611                                  * |--range with bits sets--|
1612                                  *    |
1613                                  *    start
1614                                  */
1615                                 start = state->end + 1;
1616                         } else {
1617                                 /*
1618                                  * 'start' falls within a range that doesn't
1619                                  * have the bits set, so take its start as
1620                                  * the beginning of the desired range
1621                                  *
1622                                  * |--range with bits cleared----|
1623                                  *      |
1624                                  *      start
1625                                  */
1626                                 *start_ret = state->start;
1627                                 break;
1628                         }
1629                 } else {
1630                         /*
1631                          * |---prev range---|---hole/unset---|---node range---|
1632                          *                          |
1633                          *                        start
1634                          *
1635                          *                        or
1636                          *
1637                          * |---hole/unset--||--first node--|
1638                          * 0   |
1639                          *    start
1640                          */
1641                         if (prev) {
1642                                 state = rb_entry(prev, struct extent_state,
1643                                                  rb_node);
1644                                 *start_ret = state->end + 1;
1645                         } else {
1646                                 *start_ret = 0;
1647                         }
1648                         break;
1649                 }
1650         }
1651
1652         /*
1653          * Find the longest stretch from start until an entry which has the
1654          * bits set
1655          */
1656         while (1) {
1657                 state = rb_entry(node, struct extent_state, rb_node);
1658                 if (state->end >= start && !(state->state & bits)) {
1659                         *end_ret = state->end;
1660                 } else {
1661                         *end_ret = state->start - 1;
1662                         break;
1663                 }
1664
1665                 node = rb_next(node);
1666                 if (!node)
1667                         break;
1668         }
1669 out:
1670         spin_unlock(&tree->lock);
1671 }
1672
1673 /*
1674  * find a contiguous range of bytes in the file marked as delalloc, not
1675  * more than 'max_bytes'.  start and end are used to return the range,
1676  *
1677  * true is returned if we find something, false if nothing was in the tree
1678  */
1679 static noinline bool find_delalloc_range(struct extent_io_tree *tree,
1680                                         u64 *start, u64 *end, u64 max_bytes,
1681                                         struct extent_state **cached_state)
1682 {
1683         struct rb_node *node;
1684         struct extent_state *state;
1685         u64 cur_start = *start;
1686         bool found = false;
1687         u64 total_bytes = 0;
1688
1689         spin_lock(&tree->lock);
1690
1691         /*
1692          * this search will find all the extents that end after
1693          * our range starts.
1694          */
1695         node = tree_search(tree, cur_start);
1696         if (!node) {
1697                 *end = (u64)-1;
1698                 goto out;
1699         }
1700
1701         while (1) {
1702                 state = rb_entry(node, struct extent_state, rb_node);
1703                 if (found && (state->start != cur_start ||
1704                               (state->state & EXTENT_BOUNDARY))) {
1705                         goto out;
1706                 }
1707                 if (!(state->state & EXTENT_DELALLOC)) {
1708                         if (!found)
1709                                 *end = state->end;
1710                         goto out;
1711                 }
1712                 if (!found) {
1713                         *start = state->start;
1714                         *cached_state = state;
1715                         refcount_inc(&state->refs);
1716                 }
1717                 found = true;
1718                 *end = state->end;
1719                 cur_start = state->end + 1;
1720                 node = rb_next(node);
1721                 total_bytes += state->end - state->start + 1;
1722                 if (total_bytes >= max_bytes)
1723                         break;
1724                 if (!node)
1725                         break;
1726         }
1727 out:
1728         spin_unlock(&tree->lock);
1729         return found;
1730 }
1731
1732 static int __process_pages_contig(struct address_space *mapping,
1733                                   struct page *locked_page,
1734                                   pgoff_t start_index, pgoff_t end_index,
1735                                   unsigned long page_ops, pgoff_t *index_ret);
1736
1737 static noinline void __unlock_for_delalloc(struct inode *inode,
1738                                            struct page *locked_page,
1739                                            u64 start, u64 end)
1740 {
1741         unsigned long index = start >> PAGE_SHIFT;
1742         unsigned long end_index = end >> PAGE_SHIFT;
1743
1744         ASSERT(locked_page);
1745         if (index == locked_page->index && end_index == index)
1746                 return;
1747
1748         __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1749                                PAGE_UNLOCK, NULL);
1750 }
1751
1752 static noinline int lock_delalloc_pages(struct inode *inode,
1753                                         struct page *locked_page,
1754                                         u64 delalloc_start,
1755                                         u64 delalloc_end)
1756 {
1757         unsigned long index = delalloc_start >> PAGE_SHIFT;
1758         unsigned long index_ret = index;
1759         unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1760         int ret;
1761
1762         ASSERT(locked_page);
1763         if (index == locked_page->index && index == end_index)
1764                 return 0;
1765
1766         ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1767                                      end_index, PAGE_LOCK, &index_ret);
1768         if (ret == -EAGAIN)
1769                 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1770                                       (u64)index_ret << PAGE_SHIFT);
1771         return ret;
1772 }
1773
1774 /*
1775  * Find and lock a contiguous range of bytes in the file marked as delalloc, no
1776  * more than @max_bytes.  @Start and @end are used to return the range,
1777  *
1778  * Return: true if we find something
1779  *         false if nothing was in the tree
1780  */
1781 EXPORT_FOR_TESTS
1782 noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
1783                                     struct page *locked_page, u64 *start,
1784                                     u64 *end)
1785 {
1786         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1787         u64 max_bytes = BTRFS_MAX_EXTENT_SIZE;
1788         u64 delalloc_start;
1789         u64 delalloc_end;
1790         bool found;
1791         struct extent_state *cached_state = NULL;
1792         int ret;
1793         int loops = 0;
1794
1795 again:
1796         /* step one, find a bunch of delalloc bytes starting at start */
1797         delalloc_start = *start;
1798         delalloc_end = 0;
1799         found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1800                                     max_bytes, &cached_state);
1801         if (!found || delalloc_end <= *start) {
1802                 *start = delalloc_start;
1803                 *end = delalloc_end;
1804                 free_extent_state(cached_state);
1805                 return false;
1806         }
1807
1808         /*
1809          * start comes from the offset of locked_page.  We have to lock
1810          * pages in order, so we can't process delalloc bytes before
1811          * locked_page
1812          */
1813         if (delalloc_start < *start)
1814                 delalloc_start = *start;
1815
1816         /*
1817          * make sure to limit the number of pages we try to lock down
1818          */
1819         if (delalloc_end + 1 - delalloc_start > max_bytes)
1820                 delalloc_end = delalloc_start + max_bytes - 1;
1821
1822         /* step two, lock all the pages after the page that has start */
1823         ret = lock_delalloc_pages(inode, locked_page,
1824                                   delalloc_start, delalloc_end);
1825         ASSERT(!ret || ret == -EAGAIN);
1826         if (ret == -EAGAIN) {
1827                 /* some of the pages are gone, lets avoid looping by
1828                  * shortening the size of the delalloc range we're searching
1829                  */
1830                 free_extent_state(cached_state);
1831                 cached_state = NULL;
1832                 if (!loops) {
1833                         max_bytes = PAGE_SIZE;
1834                         loops = 1;
1835                         goto again;
1836                 } else {
1837                         found = false;
1838                         goto out_failed;
1839                 }
1840         }
1841
1842         /* step three, lock the state bits for the whole range */
1843         lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1844
1845         /* then test to make sure it is all still delalloc */
1846         ret = test_range_bit(tree, delalloc_start, delalloc_end,
1847                              EXTENT_DELALLOC, 1, cached_state);
1848         if (!ret) {
1849                 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1850                                      &cached_state);
1851                 __unlock_for_delalloc(inode, locked_page,
1852                               delalloc_start, delalloc_end);
1853                 cond_resched();
1854                 goto again;
1855         }
1856         free_extent_state(cached_state);
1857         *start = delalloc_start;
1858         *end = delalloc_end;
1859 out_failed:
1860         return found;
1861 }
1862
1863 static int __process_pages_contig(struct address_space *mapping,
1864                                   struct page *locked_page,
1865                                   pgoff_t start_index, pgoff_t end_index,
1866                                   unsigned long page_ops, pgoff_t *index_ret)
1867 {
1868         unsigned long nr_pages = end_index - start_index + 1;
1869         unsigned long pages_locked = 0;
1870         pgoff_t index = start_index;
1871         struct page *pages[16];
1872         unsigned ret;
1873         int err = 0;
1874         int i;
1875
1876         if (page_ops & PAGE_LOCK) {
1877                 ASSERT(page_ops == PAGE_LOCK);
1878                 ASSERT(index_ret && *index_ret == start_index);
1879         }
1880
1881         if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1882                 mapping_set_error(mapping, -EIO);
1883
1884         while (nr_pages > 0) {
1885                 ret = find_get_pages_contig(mapping, index,
1886                                      min_t(unsigned long,
1887                                      nr_pages, ARRAY_SIZE(pages)), pages);
1888                 if (ret == 0) {
1889                         /*
1890                          * Only if we're going to lock these pages,
1891                          * can we find nothing at @index.
1892                          */
1893                         ASSERT(page_ops & PAGE_LOCK);
1894                         err = -EAGAIN;
1895                         goto out;
1896                 }
1897
1898                 for (i = 0; i < ret; i++) {
1899                         if (page_ops & PAGE_SET_PRIVATE2)
1900                                 SetPagePrivate2(pages[i]);
1901
1902                         if (pages[i] == locked_page) {
1903                                 put_page(pages[i]);
1904                                 pages_locked++;
1905                                 continue;
1906                         }
1907                         if (page_ops & PAGE_CLEAR_DIRTY)
1908                                 clear_page_dirty_for_io(pages[i]);
1909                         if (page_ops & PAGE_SET_WRITEBACK)
1910                                 set_page_writeback(pages[i]);
1911                         if (page_ops & PAGE_SET_ERROR)
1912                                 SetPageError(pages[i]);
1913                         if (page_ops & PAGE_END_WRITEBACK)
1914                                 end_page_writeback(pages[i]);
1915                         if (page_ops & PAGE_UNLOCK)
1916                                 unlock_page(pages[i]);
1917                         if (page_ops & PAGE_LOCK) {
1918                                 lock_page(pages[i]);
1919                                 if (!PageDirty(pages[i]) ||
1920                                     pages[i]->mapping != mapping) {
1921                                         unlock_page(pages[i]);
1922                                         put_page(pages[i]);
1923                                         err = -EAGAIN;
1924                                         goto out;
1925                                 }
1926                         }
1927                         put_page(pages[i]);
1928                         pages_locked++;
1929                 }
1930                 nr_pages -= ret;
1931                 index += ret;
1932                 cond_resched();
1933         }
1934 out:
1935         if (err && index_ret)
1936                 *index_ret = start_index + pages_locked - 1;
1937         return err;
1938 }
1939
1940 void extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1941                                   struct page *locked_page,
1942                                   unsigned clear_bits,
1943                                   unsigned long page_ops)
1944 {
1945         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, clear_bits, 1, 0,
1946                          NULL);
1947
1948         __process_pages_contig(inode->i_mapping, locked_page,
1949                                start >> PAGE_SHIFT, end >> PAGE_SHIFT,
1950                                page_ops, NULL);
1951 }
1952
1953 /*
1954  * count the number of bytes in the tree that have a given bit(s)
1955  * set.  This can be fairly slow, except for EXTENT_DIRTY which is
1956  * cached.  The total number found is returned.
1957  */
1958 u64 count_range_bits(struct extent_io_tree *tree,
1959                      u64 *start, u64 search_end, u64 max_bytes,
1960                      unsigned bits, int contig)
1961 {
1962         struct rb_node *node;
1963         struct extent_state *state;
1964         u64 cur_start = *start;
1965         u64 total_bytes = 0;
1966         u64 last = 0;
1967         int found = 0;
1968
1969         if (WARN_ON(search_end <= cur_start))
1970                 return 0;
1971
1972         spin_lock(&tree->lock);
1973         if (cur_start == 0 && bits == EXTENT_DIRTY) {
1974                 total_bytes = tree->dirty_bytes;
1975                 goto out;
1976         }
1977         /*
1978          * this search will find all the extents that end after
1979          * our range starts.
1980          */
1981         node = tree_search(tree, cur_start);
1982         if (!node)
1983                 goto out;
1984
1985         while (1) {
1986                 state = rb_entry(node, struct extent_state, rb_node);
1987                 if (state->start > search_end)
1988                         break;
1989                 if (contig && found && state->start > last + 1)
1990                         break;
1991                 if (state->end >= cur_start && (state->state & bits) == bits) {
1992                         total_bytes += min(search_end, state->end) + 1 -
1993                                        max(cur_start, state->start);
1994                         if (total_bytes >= max_bytes)
1995                                 break;
1996                         if (!found) {
1997                                 *start = max(cur_start, state->start);
1998                                 found = 1;
1999                         }
2000                         last = state->end;
2001                 } else if (contig && found) {
2002                         break;
2003                 }
2004                 node = rb_next(node);
2005                 if (!node)
2006                         break;
2007         }
2008 out:
2009         spin_unlock(&tree->lock);
2010         return total_bytes;
2011 }
2012
2013 /*
2014  * set the private field for a given byte offset in the tree.  If there isn't
2015  * an extent_state there already, this does nothing.
2016  */
2017 static noinline int set_state_failrec(struct extent_io_tree *tree, u64 start,
2018                 struct io_failure_record *failrec)
2019 {
2020         struct rb_node *node;
2021         struct extent_state *state;
2022         int ret = 0;
2023
2024         spin_lock(&tree->lock);
2025         /*
2026          * this search will find all the extents that end after
2027          * our range starts.
2028          */
2029         node = tree_search(tree, start);
2030         if (!node) {
2031                 ret = -ENOENT;
2032                 goto out;
2033         }
2034         state = rb_entry(node, struct extent_state, rb_node);
2035         if (state->start != start) {
2036                 ret = -ENOENT;
2037                 goto out;
2038         }
2039         state->failrec = failrec;
2040 out:
2041         spin_unlock(&tree->lock);
2042         return ret;
2043 }
2044
2045 static noinline int get_state_failrec(struct extent_io_tree *tree, u64 start,
2046                 struct io_failure_record **failrec)
2047 {
2048         struct rb_node *node;
2049         struct extent_state *state;
2050         int ret = 0;
2051
2052         spin_lock(&tree->lock);
2053         /*
2054          * this search will find all the extents that end after
2055          * our range starts.
2056          */
2057         node = tree_search(tree, start);
2058         if (!node) {
2059                 ret = -ENOENT;
2060                 goto out;
2061         }
2062         state = rb_entry(node, struct extent_state, rb_node);
2063         if (state->start != start) {
2064                 ret = -ENOENT;
2065                 goto out;
2066         }
2067         *failrec = state->failrec;
2068 out:
2069         spin_unlock(&tree->lock);
2070         return ret;
2071 }
2072
2073 /*
2074  * searches a range in the state tree for a given mask.
2075  * If 'filled' == 1, this returns 1 only if every extent in the tree
2076  * has the bits set.  Otherwise, 1 is returned if any bit in the
2077  * range is found set.
2078  */
2079 int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
2080                    unsigned bits, int filled, struct extent_state *cached)
2081 {
2082         struct extent_state *state = NULL;
2083         struct rb_node *node;
2084         int bitset = 0;
2085
2086         spin_lock(&tree->lock);
2087         if (cached && extent_state_in_tree(cached) && cached->start <= start &&
2088             cached->end > start)
2089                 node = &cached->rb_node;
2090         else
2091                 node = tree_search(tree, start);
2092         while (node && start <= end) {
2093                 state = rb_entry(node, struct extent_state, rb_node);
2094
2095                 if (filled && state->start > start) {
2096                         bitset = 0;
2097                         break;
2098                 }
2099
2100                 if (state->start > end)
2101                         break;
2102
2103                 if (state->state & bits) {
2104                         bitset = 1;
2105                         if (!filled)
2106                                 break;
2107                 } else if (filled) {
2108                         bitset = 0;
2109                         break;
2110                 }
2111
2112                 if (state->end == (u64)-1)
2113                         break;
2114
2115                 start = state->end + 1;
2116                 if (start > end)
2117                         break;
2118                 node = rb_next(node);
2119                 if (!node) {
2120                         if (filled)
2121                                 bitset = 0;
2122                         break;
2123                 }
2124         }
2125         spin_unlock(&tree->lock);
2126         return bitset;
2127 }
2128
2129 /*
2130  * helper function to set a given page up to date if all the
2131  * extents in the tree for that page are up to date
2132  */
2133 static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
2134 {
2135         u64 start = page_offset(page);
2136         u64 end = start + PAGE_SIZE - 1;
2137         if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
2138                 SetPageUptodate(page);
2139 }
2140
2141 int free_io_failure(struct extent_io_tree *failure_tree,
2142                     struct extent_io_tree *io_tree,
2143                     struct io_failure_record *rec)
2144 {
2145         int ret;
2146         int err = 0;
2147
2148         set_state_failrec(failure_tree, rec->start, NULL);
2149         ret = clear_extent_bits(failure_tree, rec->start,
2150                                 rec->start + rec->len - 1,
2151                                 EXTENT_LOCKED | EXTENT_DIRTY);
2152         if (ret)
2153                 err = ret;
2154
2155         ret = clear_extent_bits(io_tree, rec->start,
2156                                 rec->start + rec->len - 1,
2157                                 EXTENT_DAMAGED);
2158         if (ret && !err)
2159                 err = ret;
2160
2161         kfree(rec);
2162         return err;
2163 }
2164
2165 /*
2166  * this bypasses the standard btrfs submit functions deliberately, as
2167  * the standard behavior is to write all copies in a raid setup. here we only
2168  * want to write the one bad copy. so we do the mapping for ourselves and issue
2169  * submit_bio directly.
2170  * to avoid any synchronization issues, wait for the data after writing, which
2171  * actually prevents the read that triggered the error from finishing.
2172  * currently, there can be no more than two copies of every data bit. thus,
2173  * exactly one rewrite is required.
2174  */
2175 int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
2176                       u64 length, u64 logical, struct page *page,
2177                       unsigned int pg_offset, int mirror_num)
2178 {
2179         struct bio *bio;
2180         struct btrfs_device *dev;
2181         u64 map_length = 0;
2182         u64 sector;
2183         struct btrfs_bio *bbio = NULL;
2184         int ret;
2185
2186         ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
2187         BUG_ON(!mirror_num);
2188
2189         bio = btrfs_io_bio_alloc(1);
2190         bio->bi_iter.bi_size = 0;
2191         map_length = length;
2192
2193         /*
2194          * Avoid races with device replace and make sure our bbio has devices
2195          * associated to its stripes that don't go away while we are doing the
2196          * read repair operation.
2197          */
2198         btrfs_bio_counter_inc_blocked(fs_info);
2199         if (btrfs_is_parity_mirror(fs_info, logical, length)) {
2200                 /*
2201                  * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2202                  * to update all raid stripes, but here we just want to correct
2203                  * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2204                  * stripe's dev and sector.
2205                  */
2206                 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2207                                       &map_length, &bbio, 0);
2208                 if (ret) {
2209                         btrfs_bio_counter_dec(fs_info);
2210                         bio_put(bio);
2211                         return -EIO;
2212                 }
2213                 ASSERT(bbio->mirror_num == 1);
2214         } else {
2215                 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2216                                       &map_length, &bbio, mirror_num);
2217                 if (ret) {
2218                         btrfs_bio_counter_dec(fs_info);
2219                         bio_put(bio);
2220                         return -EIO;
2221                 }
2222                 BUG_ON(mirror_num != bbio->mirror_num);
2223         }
2224
2225         sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2226         bio->bi_iter.bi_sector = sector;
2227         dev = bbio->stripes[bbio->mirror_num - 1].dev;
2228         btrfs_put_bbio(bbio);
2229         if (!dev || !dev->bdev ||
2230             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2231                 btrfs_bio_counter_dec(fs_info);
2232                 bio_put(bio);
2233                 return -EIO;
2234         }
2235         bio_set_dev(bio, dev->bdev);
2236         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2237         bio_add_page(bio, page, length, pg_offset);
2238
2239         if (btrfsic_submit_bio_wait(bio)) {
2240                 /* try to remap that extent elsewhere? */
2241                 btrfs_bio_counter_dec(fs_info);
2242                 bio_put(bio);
2243                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2244                 return -EIO;
2245         }
2246
2247         btrfs_info_rl_in_rcu(fs_info,
2248                 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2249                                   ino, start,
2250                                   rcu_str_deref(dev->name), sector);
2251         btrfs_bio_counter_dec(fs_info);
2252         bio_put(bio);
2253         return 0;
2254 }
2255
2256 int btrfs_repair_eb_io_failure(struct extent_buffer *eb, int mirror_num)
2257 {
2258         struct btrfs_fs_info *fs_info = eb->fs_info;
2259         u64 start = eb->start;
2260         int i, num_pages = num_extent_pages(eb);
2261         int ret = 0;
2262
2263         if (sb_rdonly(fs_info->sb))
2264                 return -EROFS;
2265
2266         for (i = 0; i < num_pages; i++) {
2267                 struct page *p = eb->pages[i];
2268
2269                 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2270                                         start - page_offset(p), mirror_num);
2271                 if (ret)
2272                         break;
2273                 start += PAGE_SIZE;
2274         }
2275
2276         return ret;
2277 }
2278
2279 /*
2280  * each time an IO finishes, we do a fast check in the IO failure tree
2281  * to see if we need to process or clean up an io_failure_record
2282  */
2283 int clean_io_failure(struct btrfs_fs_info *fs_info,
2284                      struct extent_io_tree *failure_tree,
2285                      struct extent_io_tree *io_tree, u64 start,
2286                      struct page *page, u64 ino, unsigned int pg_offset)
2287 {
2288         u64 private;
2289         struct io_failure_record *failrec;
2290         struct extent_state *state;
2291         int num_copies;
2292         int ret;
2293
2294         private = 0;
2295         ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2296                                EXTENT_DIRTY, 0);
2297         if (!ret)
2298                 return 0;
2299
2300         ret = get_state_failrec(failure_tree, start, &failrec);
2301         if (ret)
2302                 return 0;
2303
2304         BUG_ON(!failrec->this_mirror);
2305
2306         if (failrec->in_validation) {
2307                 /* there was no real error, just free the record */
2308                 btrfs_debug(fs_info,
2309                         "clean_io_failure: freeing dummy error at %llu",
2310                         failrec->start);
2311                 goto out;
2312         }
2313         if (sb_rdonly(fs_info->sb))
2314                 goto out;
2315
2316         spin_lock(&io_tree->lock);
2317         state = find_first_extent_bit_state(io_tree,
2318                                             failrec->start,
2319                                             EXTENT_LOCKED);
2320         spin_unlock(&io_tree->lock);
2321
2322         if (state && state->start <= failrec->start &&
2323             state->end >= failrec->start + failrec->len - 1) {
2324                 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2325                                               failrec->len);
2326                 if (num_copies > 1)  {
2327                         repair_io_failure(fs_info, ino, start, failrec->len,
2328                                           failrec->logical, page, pg_offset,
2329                                           failrec->failed_mirror);
2330                 }
2331         }
2332
2333 out:
2334         free_io_failure(failure_tree, io_tree, failrec);
2335
2336         return 0;
2337 }
2338
2339 /*
2340  * Can be called when
2341  * - hold extent lock
2342  * - under ordered extent
2343  * - the inode is freeing
2344  */
2345 void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2346 {
2347         struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2348         struct io_failure_record *failrec;
2349         struct extent_state *state, *next;
2350
2351         if (RB_EMPTY_ROOT(&failure_tree->state))
2352                 return;
2353
2354         spin_lock(&failure_tree->lock);
2355         state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2356         while (state) {
2357                 if (state->start > end)
2358                         break;
2359
2360                 ASSERT(state->end <= end);
2361
2362                 next = next_state(state);
2363
2364                 failrec = state->failrec;
2365                 free_extent_state(state);
2366                 kfree(failrec);
2367
2368                 state = next;
2369         }
2370         spin_unlock(&failure_tree->lock);
2371 }
2372
2373 int btrfs_get_io_failure_record(struct inode *inode, u64 start, u64 end,
2374                 struct io_failure_record **failrec_ret)
2375 {
2376         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2377         struct io_failure_record *failrec;
2378         struct extent_map *em;
2379         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2380         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2381         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2382         int ret;
2383         u64 logical;
2384
2385         ret = get_state_failrec(failure_tree, start, &failrec);
2386         if (ret) {
2387                 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2388                 if (!failrec)
2389                         return -ENOMEM;
2390
2391                 failrec->start = start;
2392                 failrec->len = end - start + 1;
2393                 failrec->this_mirror = 0;
2394                 failrec->bio_flags = 0;
2395                 failrec->in_validation = 0;
2396
2397                 read_lock(&em_tree->lock);
2398                 em = lookup_extent_mapping(em_tree, start, failrec->len);
2399                 if (!em) {
2400                         read_unlock(&em_tree->lock);
2401                         kfree(failrec);
2402                         return -EIO;
2403                 }
2404
2405                 if (em->start > start || em->start + em->len <= start) {
2406                         free_extent_map(em);
2407                         em = NULL;
2408                 }
2409                 read_unlock(&em_tree->lock);
2410                 if (!em) {
2411                         kfree(failrec);
2412                         return -EIO;
2413                 }
2414
2415                 logical = start - em->start;
2416                 logical = em->block_start + logical;
2417                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2418                         logical = em->block_start;
2419                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2420                         extent_set_compress_type(&failrec->bio_flags,
2421                                                  em->compress_type);
2422                 }
2423
2424                 btrfs_debug(fs_info,
2425                         "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2426                         logical, start, failrec->len);
2427
2428                 failrec->logical = logical;
2429                 free_extent_map(em);
2430
2431                 /* set the bits in the private failure tree */
2432                 ret = set_extent_bits(failure_tree, start, end,
2433                                         EXTENT_LOCKED | EXTENT_DIRTY);
2434                 if (ret >= 0)
2435                         ret = set_state_failrec(failure_tree, start, failrec);
2436                 /* set the bits in the inode's tree */
2437                 if (ret >= 0)
2438                         ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2439                 if (ret < 0) {
2440                         kfree(failrec);
2441                         return ret;
2442                 }
2443         } else {
2444                 btrfs_debug(fs_info,
2445                         "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2446                         failrec->logical, failrec->start, failrec->len,
2447                         failrec->in_validation);
2448                 /*
2449                  * when data can be on disk more than twice, add to failrec here
2450                  * (e.g. with a list for failed_mirror) to make
2451                  * clean_io_failure() clean all those errors at once.
2452                  */
2453         }
2454
2455         *failrec_ret = failrec;
2456
2457         return 0;
2458 }
2459
2460 bool btrfs_check_repairable(struct inode *inode, unsigned failed_bio_pages,
2461                            struct io_failure_record *failrec, int failed_mirror)
2462 {
2463         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2464         int num_copies;
2465
2466         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2467         if (num_copies == 1) {
2468                 /*
2469                  * we only have a single copy of the data, so don't bother with
2470                  * all the retry and error correction code that follows. no
2471                  * matter what the error is, it is very likely to persist.
2472                  */
2473                 btrfs_debug(fs_info,
2474                         "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2475                         num_copies, failrec->this_mirror, failed_mirror);
2476                 return false;
2477         }
2478
2479         /*
2480          * there are two premises:
2481          *      a) deliver good data to the caller
2482          *      b) correct the bad sectors on disk
2483          */
2484         if (failed_bio_pages > 1) {
2485                 /*
2486                  * to fulfill b), we need to know the exact failing sectors, as
2487                  * we don't want to rewrite any more than the failed ones. thus,
2488                  * we need separate read requests for the failed bio
2489                  *
2490                  * if the following BUG_ON triggers, our validation request got
2491                  * merged. we need separate requests for our algorithm to work.
2492                  */
2493                 BUG_ON(failrec->in_validation);
2494                 failrec->in_validation = 1;
2495                 failrec->this_mirror = failed_mirror;
2496         } else {
2497                 /*
2498                  * we're ready to fulfill a) and b) alongside. get a good copy
2499                  * of the failed sector and if we succeed, we have setup
2500                  * everything for repair_io_failure to do the rest for us.
2501                  */
2502                 if (failrec->in_validation) {
2503                         BUG_ON(failrec->this_mirror != failed_mirror);
2504                         failrec->in_validation = 0;
2505                         failrec->this_mirror = 0;
2506                 }
2507                 failrec->failed_mirror = failed_mirror;
2508                 failrec->this_mirror++;
2509                 if (failrec->this_mirror == failed_mirror)
2510                         failrec->this_mirror++;
2511         }
2512
2513         if (failrec->this_mirror > num_copies) {
2514                 btrfs_debug(fs_info,
2515                         "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2516                         num_copies, failrec->this_mirror, failed_mirror);
2517                 return false;
2518         }
2519
2520         return true;
2521 }
2522
2523
2524 struct bio *btrfs_create_repair_bio(struct inode *inode, struct bio *failed_bio,
2525                                     struct io_failure_record *failrec,
2526                                     struct page *page, int pg_offset, int icsum,
2527                                     bio_end_io_t *endio_func, void *data)
2528 {
2529         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2530         struct bio *bio;
2531         struct btrfs_io_bio *btrfs_failed_bio;
2532         struct btrfs_io_bio *btrfs_bio;
2533
2534         bio = btrfs_io_bio_alloc(1);
2535         bio->bi_end_io = endio_func;
2536         bio->bi_iter.bi_sector = failrec->logical >> 9;
2537         bio_set_dev(bio, fs_info->fs_devices->latest_bdev);
2538         bio->bi_iter.bi_size = 0;
2539         bio->bi_private = data;
2540
2541         btrfs_failed_bio = btrfs_io_bio(failed_bio);
2542         if (btrfs_failed_bio->csum) {
2543                 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2544
2545                 btrfs_bio = btrfs_io_bio(bio);
2546                 btrfs_bio->csum = btrfs_bio->csum_inline;
2547                 icsum *= csum_size;
2548                 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + icsum,
2549                        csum_size);
2550         }
2551
2552         bio_add_page(bio, page, failrec->len, pg_offset);
2553
2554         return bio;
2555 }
2556
2557 /*
2558  * This is a generic handler for readpage errors. If other copies exist, read
2559  * those and write back good data to the failed position. Does not investigate
2560  * in remapping the failed extent elsewhere, hoping the device will be smart
2561  * enough to do this as needed
2562  */
2563 static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2564                               struct page *page, u64 start, u64 end,
2565                               int failed_mirror)
2566 {
2567         struct io_failure_record *failrec;
2568         struct inode *inode = page->mapping->host;
2569         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2570         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2571         struct bio *bio;
2572         int read_mode = 0;
2573         blk_status_t status;
2574         int ret;
2575         unsigned failed_bio_pages = failed_bio->bi_iter.bi_size >> PAGE_SHIFT;
2576
2577         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2578
2579         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
2580         if (ret)
2581                 return ret;
2582
2583         if (!btrfs_check_repairable(inode, failed_bio_pages, failrec,
2584                                     failed_mirror)) {
2585                 free_io_failure(failure_tree, tree, failrec);
2586                 return -EIO;
2587         }
2588
2589         if (failed_bio_pages > 1)
2590                 read_mode |= REQ_FAILFAST_DEV;
2591
2592         phy_offset >>= inode->i_sb->s_blocksize_bits;
2593         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
2594                                       start - page_offset(page),
2595                                       (int)phy_offset, failed_bio->bi_end_io,
2596                                       NULL);
2597         bio->bi_opf = REQ_OP_READ | read_mode;
2598
2599         btrfs_debug(btrfs_sb(inode->i_sb),
2600                 "Repair Read Error: submitting new read[%#x] to this_mirror=%d, in_validation=%d",
2601                 read_mode, failrec->this_mirror, failrec->in_validation);
2602
2603         status = tree->ops->submit_bio_hook(tree->private_data, bio, failrec->this_mirror,
2604                                          failrec->bio_flags);
2605         if (status) {
2606                 free_io_failure(failure_tree, tree, failrec);
2607                 bio_put(bio);
2608                 ret = blk_status_to_errno(status);
2609         }
2610
2611         return ret;
2612 }
2613
2614 /* lots and lots of room for performance fixes in the end_bio funcs */
2615
2616 void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2617 {
2618         int uptodate = (err == 0);
2619         int ret = 0;
2620
2621         btrfs_writepage_endio_finish_ordered(page, start, end, uptodate);
2622
2623         if (!uptodate) {
2624                 ClearPageUptodate(page);
2625                 SetPageError(page);
2626                 ret = err < 0 ? err : -EIO;
2627                 mapping_set_error(page->mapping, ret);
2628         }
2629 }
2630
2631 /*
2632  * after a writepage IO is done, we need to:
2633  * clear the uptodate bits on error
2634  * clear the writeback bits in the extent tree for this IO
2635  * end_page_writeback if the page has no more pending IO
2636  *
2637  * Scheduling is not allowed, so the extent state tree is expected
2638  * to have one and only one object corresponding to this IO.
2639  */
2640 static void end_bio_extent_writepage(struct bio *bio)
2641 {
2642         int error = blk_status_to_errno(bio->bi_status);
2643         struct bio_vec *bvec;
2644         u64 start;
2645         u64 end;
2646         struct bvec_iter_all iter_all;
2647
2648         ASSERT(!bio_flagged(bio, BIO_CLONED));
2649         bio_for_each_segment_all(bvec, bio, iter_all) {
2650                 struct page *page = bvec->bv_page;
2651                 struct inode *inode = page->mapping->host;
2652                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2653
2654                 /* We always issue full-page reads, but if some block
2655                  * in a page fails to read, blk_update_request() will
2656                  * advance bv_offset and adjust bv_len to compensate.
2657                  * Print a warning for nonzero offsets, and an error
2658                  * if they don't add up to a full page.  */
2659                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2660                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2661                                 btrfs_err(fs_info,
2662                                    "partial page write in btrfs with offset %u and length %u",
2663                                         bvec->bv_offset, bvec->bv_len);
2664                         else
2665                                 btrfs_info(fs_info,
2666                                    "incomplete page write in btrfs with offset %u and length %u",
2667                                         bvec->bv_offset, bvec->bv_len);
2668                 }
2669
2670                 start = page_offset(page);
2671                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2672
2673                 end_extent_writepage(page, error, start, end);
2674                 end_page_writeback(page);
2675         }
2676
2677         bio_put(bio);
2678 }
2679
2680 static void
2681 endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2682                               int uptodate)
2683 {
2684         struct extent_state *cached = NULL;
2685         u64 end = start + len - 1;
2686
2687         if (uptodate && tree->track_uptodate)
2688                 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2689         unlock_extent_cached_atomic(tree, start, end, &cached);
2690 }
2691
2692 /*
2693  * after a readpage IO is done, we need to:
2694  * clear the uptodate bits on error
2695  * set the uptodate bits if things worked
2696  * set the page up to date if all extents in the tree are uptodate
2697  * clear the lock bit in the extent tree
2698  * unlock the page if there are no other extents locked for it
2699  *
2700  * Scheduling is not allowed, so the extent state tree is expected
2701  * to have one and only one object corresponding to this IO.
2702  */
2703 static void end_bio_extent_readpage(struct bio *bio)
2704 {
2705         struct bio_vec *bvec;
2706         int uptodate = !bio->bi_status;
2707         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2708         struct extent_io_tree *tree, *failure_tree;
2709         u64 offset = 0;
2710         u64 start;
2711         u64 end;
2712         u64 len;
2713         u64 extent_start = 0;
2714         u64 extent_len = 0;
2715         int mirror;
2716         int ret;
2717         struct bvec_iter_all iter_all;
2718
2719         ASSERT(!bio_flagged(bio, BIO_CLONED));
2720         bio_for_each_segment_all(bvec, bio, iter_all) {
2721                 struct page *page = bvec->bv_page;
2722                 struct inode *inode = page->mapping->host;
2723                 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2724                 bool data_inode = btrfs_ino(BTRFS_I(inode))
2725                         != BTRFS_BTREE_INODE_OBJECTID;
2726
2727                 btrfs_debug(fs_info,
2728                         "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2729                         (u64)bio->bi_iter.bi_sector, bio->bi_status,
2730                         io_bio->mirror_num);
2731                 tree = &BTRFS_I(inode)->io_tree;
2732                 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2733
2734                 /* We always issue full-page reads, but if some block
2735                  * in a page fails to read, blk_update_request() will
2736                  * advance bv_offset and adjust bv_len to compensate.
2737                  * Print a warning for nonzero offsets, and an error
2738                  * if they don't add up to a full page.  */
2739                 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2740                         if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2741                                 btrfs_err(fs_info,
2742                                         "partial page read in btrfs with offset %u and length %u",
2743                                         bvec->bv_offset, bvec->bv_len);
2744                         else
2745                                 btrfs_info(fs_info,
2746                                         "incomplete page read in btrfs with offset %u and length %u",
2747                                         bvec->bv_offset, bvec->bv_len);
2748                 }
2749
2750                 start = page_offset(page);
2751                 end = start + bvec->bv_offset + bvec->bv_len - 1;
2752                 len = bvec->bv_len;
2753
2754                 mirror = io_bio->mirror_num;
2755                 if (likely(uptodate)) {
2756                         ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2757                                                               page, start, end,
2758                                                               mirror);
2759                         if (ret)
2760                                 uptodate = 0;
2761                         else
2762                                 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2763                                                  failure_tree, tree, start,
2764                                                  page,
2765                                                  btrfs_ino(BTRFS_I(inode)), 0);
2766                 }
2767
2768                 if (likely(uptodate))
2769                         goto readpage_ok;
2770
2771                 if (data_inode) {
2772
2773                         /*
2774                          * The generic bio_readpage_error handles errors the
2775                          * following way: If possible, new read requests are
2776                          * created and submitted and will end up in
2777                          * end_bio_extent_readpage as well (if we're lucky,
2778                          * not in the !uptodate case). In that case it returns
2779                          * 0 and we just go on with the next page in our bio.
2780                          * If it can't handle the error it will return -EIO and
2781                          * we remain responsible for that page.
2782                          */
2783                         ret = bio_readpage_error(bio, offset, page, start, end,
2784                                                  mirror);
2785                         if (ret == 0) {
2786                                 uptodate = !bio->bi_status;
2787                                 offset += len;
2788                                 continue;
2789                         }
2790                 } else {
2791                         struct extent_buffer *eb;
2792
2793                         eb = (struct extent_buffer *)page->private;
2794                         set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
2795                         eb->read_mirror = mirror;
2796                         atomic_dec(&eb->io_pages);
2797                         if (test_and_clear_bit(EXTENT_BUFFER_READAHEAD,
2798                                                &eb->bflags))
2799                                 btree_readahead_hook(eb, -EIO);
2800                 }
2801 readpage_ok:
2802                 if (likely(uptodate)) {
2803                         loff_t i_size = i_size_read(inode);
2804                         pgoff_t end_index = i_size >> PAGE_SHIFT;
2805                         unsigned off;
2806
2807                         /* Zero out the end if this page straddles i_size */
2808                         off = offset_in_page(i_size);
2809                         if (page->index == end_index && off)
2810                                 zero_user_segment(page, off, PAGE_SIZE);
2811                         SetPageUptodate(page);
2812                 } else {
2813                         ClearPageUptodate(page);
2814                         SetPageError(page);
2815                 }
2816                 unlock_page(page);
2817                 offset += len;
2818
2819                 if (unlikely(!uptodate)) {
2820                         if (extent_len) {
2821                                 endio_readpage_release_extent(tree,
2822                                                               extent_start,
2823                                                               extent_len, 1);
2824                                 extent_start = 0;
2825                                 extent_len = 0;
2826                         }
2827                         endio_readpage_release_extent(tree, start,
2828                                                       end - start + 1, 0);
2829                 } else if (!extent_len) {
2830                         extent_start = start;
2831                         extent_len = end + 1 - start;
2832                 } else if (extent_start + extent_len == start) {
2833                         extent_len += end + 1 - start;
2834                 } else {
2835                         endio_readpage_release_extent(tree, extent_start,
2836                                                       extent_len, uptodate);
2837                         extent_start = start;
2838                         extent_len = end + 1 - start;
2839                 }
2840         }
2841
2842         if (extent_len)
2843                 endio_readpage_release_extent(tree, extent_start, extent_len,
2844                                               uptodate);
2845         btrfs_io_bio_free_csum(io_bio);
2846         bio_put(bio);
2847 }
2848
2849 /*
2850  * Initialize the members up to but not including 'bio'. Use after allocating a
2851  * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2852  * 'bio' because use of __GFP_ZERO is not supported.
2853  */
2854 static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2855 {
2856         memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2857 }
2858
2859 /*
2860  * The following helpers allocate a bio. As it's backed by a bioset, it'll
2861  * never fail.  We're returning a bio right now but you can call btrfs_io_bio
2862  * for the appropriate container_of magic
2863  */
2864 struct bio *btrfs_bio_alloc(u64 first_byte)
2865 {
2866         struct bio *bio;
2867
2868         bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, &btrfs_bioset);
2869         bio->bi_iter.bi_sector = first_byte >> 9;
2870         btrfs_io_bio_init(btrfs_io_bio(bio));
2871         return bio;
2872 }
2873
2874 struct bio *btrfs_bio_clone(struct bio *bio)
2875 {
2876         struct btrfs_io_bio *btrfs_bio;
2877         struct bio *new;
2878
2879         /* Bio allocation backed by a bioset does not fail */
2880         new = bio_clone_fast(bio, GFP_NOFS, &btrfs_bioset);
2881         btrfs_bio = btrfs_io_bio(new);
2882         btrfs_io_bio_init(btrfs_bio);
2883         btrfs_bio->iter = bio->bi_iter;
2884         return new;
2885 }
2886
2887 struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2888 {
2889         struct bio *bio;
2890
2891         /* Bio allocation backed by a bioset does not fail */
2892         bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, &btrfs_bioset);
2893         btrfs_io_bio_init(btrfs_io_bio(bio));
2894         return bio;
2895 }
2896
2897 struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2898 {
2899         struct bio *bio;
2900         struct btrfs_io_bio *btrfs_bio;
2901
2902         /* this will never fail when it's backed by a bioset */
2903         bio = bio_clone_fast(orig, GFP_NOFS, &btrfs_bioset);
2904         ASSERT(bio);
2905
2906         btrfs_bio = btrfs_io_bio(bio);
2907         btrfs_io_bio_init(btrfs_bio);
2908
2909         bio_trim(bio, offset >> 9, size >> 9);
2910         btrfs_bio->iter = bio->bi_iter;
2911         return bio;
2912 }
2913
2914 /*
2915  * @opf:        bio REQ_OP_* and REQ_* flags as one value
2916  * @tree:       tree so we can call our merge_bio hook
2917  * @wbc:        optional writeback control for io accounting
2918  * @page:       page to add to the bio
2919  * @pg_offset:  offset of the new bio or to check whether we are adding
2920  *              a contiguous page to the previous one
2921  * @size:       portion of page that we want to write
2922  * @offset:     starting offset in the page
2923  * @bdev:       attach newly created bios to this bdev
2924  * @bio_ret:    must be valid pointer, newly allocated bio will be stored there
2925  * @end_io_func:     end_io callback for new bio
2926  * @mirror_num:      desired mirror to read/write
2927  * @prev_bio_flags:  flags of previous bio to see if we can merge the current one
2928  * @bio_flags:  flags of the current bio to see if we can merge them
2929  */
2930 static int submit_extent_page(unsigned int opf, struct extent_io_tree *tree,
2931                               struct writeback_control *wbc,
2932                               struct page *page, u64 offset,
2933                               size_t size, unsigned long pg_offset,
2934                               struct block_device *bdev,
2935                               struct bio **bio_ret,
2936                               bio_end_io_t end_io_func,
2937                               int mirror_num,
2938                               unsigned long prev_bio_flags,
2939                               unsigned long bio_flags,
2940                               bool force_bio_submit)
2941 {
2942         int ret = 0;
2943         struct bio *bio;
2944         size_t page_size = min_t(size_t, size, PAGE_SIZE);
2945         sector_t sector = offset >> 9;
2946
2947         ASSERT(bio_ret);
2948
2949         if (*bio_ret) {
2950                 bool contig;
2951                 bool can_merge = true;
2952
2953                 bio = *bio_ret;
2954                 if (prev_bio_flags & EXTENT_BIO_COMPRESSED)
2955                         contig = bio->bi_iter.bi_sector == sector;
2956                 else
2957                         contig = bio_end_sector(bio) == sector;
2958
2959                 ASSERT(tree->ops);
2960                 if (btrfs_bio_fits_in_stripe(page, page_size, bio, bio_flags))
2961                         can_merge = false;
2962
2963                 if (prev_bio_flags != bio_flags || !contig || !can_merge ||
2964                     force_bio_submit ||
2965                     bio_add_page(bio, page, page_size, pg_offset) < page_size) {
2966                         ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
2967                         if (ret < 0) {
2968                                 *bio_ret = NULL;
2969                                 return ret;
2970                         }
2971                         bio = NULL;
2972                 } else {
2973                         if (wbc)
2974                                 wbc_account_cgroup_owner(wbc, page, page_size);
2975                         return 0;
2976                 }
2977         }
2978
2979         bio = btrfs_bio_alloc(offset);
2980         bio_set_dev(bio, bdev);
2981         bio_add_page(bio, page, page_size, pg_offset);
2982         bio->bi_end_io = end_io_func;
2983         bio->bi_private = tree;
2984         bio->bi_write_hint = page->mapping->host->i_write_hint;
2985         bio->bi_opf = opf;
2986         if (wbc) {
2987                 wbc_init_bio(wbc, bio);
2988                 wbc_account_cgroup_owner(wbc, page, page_size);
2989         }
2990
2991         *bio_ret = bio;
2992
2993         return ret;
2994 }
2995
2996 static void attach_extent_buffer_page(struct extent_buffer *eb,
2997                                       struct page *page)
2998 {
2999         if (!PagePrivate(page)) {
3000                 SetPagePrivate(page);
3001                 get_page(page);
3002                 set_page_private(page, (unsigned long)eb);
3003         } else {
3004                 WARN_ON(page->private != (unsigned long)eb);
3005         }
3006 }
3007
3008 void set_page_extent_mapped(struct page *page)
3009 {
3010         if (!PagePrivate(page)) {
3011                 SetPagePrivate(page);
3012                 get_page(page);
3013                 set_page_private(page, EXTENT_PAGE_PRIVATE);
3014         }
3015 }
3016
3017 static struct extent_map *
3018 __get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
3019                  u64 start, u64 len, get_extent_t *get_extent,
3020                  struct extent_map **em_cached)
3021 {
3022         struct extent_map *em;
3023
3024         if (em_cached && *em_cached) {
3025                 em = *em_cached;
3026                 if (extent_map_in_tree(em) && start >= em->start &&
3027                     start < extent_map_end(em)) {
3028                         refcount_inc(&em->refs);
3029                         return em;
3030                 }
3031
3032                 free_extent_map(em);
3033                 *em_cached = NULL;
3034         }
3035
3036         em = get_extent(BTRFS_I(inode), page, pg_offset, start, len, 0);
3037         if (em_cached && !IS_ERR_OR_NULL(em)) {
3038                 BUG_ON(*em_cached);
3039                 refcount_inc(&em->refs);
3040                 *em_cached = em;
3041         }
3042         return em;
3043 }
3044 /*
3045  * basic readpage implementation.  Locked extent state structs are inserted
3046  * into the tree that are removed when the IO is done (by the end_io
3047  * handlers)
3048  * XXX JDM: This needs looking at to ensure proper page locking
3049  * return 0 on success, otherwise return error
3050  */
3051 static int __do_readpage(struct extent_io_tree *tree,
3052                          struct page *page,
3053                          get_extent_t *get_extent,
3054                          struct extent_map **em_cached,
3055                          struct bio **bio, int mirror_num,
3056                          unsigned long *bio_flags, unsigned int read_flags,
3057                          u64 *prev_em_start)
3058 {
3059         struct inode *inode = page->mapping->host;
3060         u64 start = page_offset(page);
3061         const u64 end = start + PAGE_SIZE - 1;
3062         u64 cur = start;
3063         u64 extent_offset;
3064         u64 last_byte = i_size_read(inode);
3065         u64 block_start;
3066         u64 cur_end;
3067         struct extent_map *em;
3068         struct block_device *bdev;
3069         int ret = 0;
3070         int nr = 0;
3071         size_t pg_offset = 0;
3072         size_t iosize;
3073         size_t disk_io_size;
3074         size_t blocksize = inode->i_sb->s_blocksize;
3075         unsigned long this_bio_flag = 0;
3076
3077         set_page_extent_mapped(page);
3078
3079         if (!PageUptodate(page)) {
3080                 if (cleancache_get_page(page) == 0) {
3081                         BUG_ON(blocksize != PAGE_SIZE);
3082                         unlock_extent(tree, start, end);
3083                         goto out;
3084                 }
3085         }
3086
3087         if (page->index == last_byte >> PAGE_SHIFT) {
3088                 char *userpage;
3089                 size_t zero_offset = offset_in_page(last_byte);
3090
3091                 if (zero_offset) {
3092                         iosize = PAGE_SIZE - zero_offset;
3093                         userpage = kmap_atomic(page);
3094                         memset(userpage + zero_offset, 0, iosize);
3095                         flush_dcache_page(page);
3096                         kunmap_atomic(userpage);
3097                 }
3098         }
3099         while (cur <= end) {
3100                 bool force_bio_submit = false;
3101                 u64 offset;
3102
3103                 if (cur >= last_byte) {
3104                         char *userpage;
3105                         struct extent_state *cached = NULL;
3106
3107                         iosize = PAGE_SIZE - pg_offset;
3108                         userpage = kmap_atomic(page);
3109                         memset(userpage + pg_offset, 0, iosize);
3110                         flush_dcache_page(page);
3111                         kunmap_atomic(userpage);
3112                         set_extent_uptodate(tree, cur, cur + iosize - 1,
3113                                             &cached, GFP_NOFS);
3114                         unlock_extent_cached(tree, cur,
3115                                              cur + iosize - 1, &cached);
3116                         break;
3117                 }
3118                 em = __get_extent_map(inode, page, pg_offset, cur,
3119                                       end - cur + 1, get_extent, em_cached);
3120                 if (IS_ERR_OR_NULL(em)) {
3121                         SetPageError(page);
3122                         unlock_extent(tree, cur, end);
3123                         break;
3124                 }
3125                 extent_offset = cur - em->start;
3126                 BUG_ON(extent_map_end(em) <= cur);
3127                 BUG_ON(end < cur);
3128
3129                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
3130                         this_bio_flag |= EXTENT_BIO_COMPRESSED;
3131                         extent_set_compress_type(&this_bio_flag,
3132                                                  em->compress_type);
3133                 }
3134
3135                 iosize = min(extent_map_end(em) - cur, end - cur + 1);
<