Loading...
1#include <linux/bitops.h>
2#include <linux/slab.h>
3#include <linux/bio.h>
4#include <linux/mm.h>
5#include <linux/pagemap.h>
6#include <linux/page-flags.h>
7#include <linux/module.h>
8#include <linux/spinlock.h>
9#include <linux/blkdev.h>
10#include <linux/swap.h>
11#include <linux/writeback.h>
12#include <linux/pagevec.h>
13#include <linux/prefetch.h>
14#include <linux/cleancache.h>
15#include "extent_io.h"
16#include "extent_map.h"
17#include "compat.h"
18#include "ctree.h"
19#include "btrfs_inode.h"
20
21static struct kmem_cache *extent_state_cache;
22static struct kmem_cache *extent_buffer_cache;
23
24static LIST_HEAD(buffers);
25static LIST_HEAD(states);
26
27#define LEAK_DEBUG 0
28#if LEAK_DEBUG
29static DEFINE_SPINLOCK(leak_lock);
30#endif
31
32#define BUFFER_LRU_MAX 64
33
34struct tree_entry {
35 u64 start;
36 u64 end;
37 struct rb_node rb_node;
38};
39
40struct extent_page_data {
41 struct bio *bio;
42 struct extent_io_tree *tree;
43 get_extent_t *get_extent;
44
45 /* tells writepage not to lock the state bits for this range
46 * it still does the unlocking
47 */
48 unsigned int extent_locked:1;
49
50 /* tells the submit_bio code to use a WRITE_SYNC */
51 unsigned int sync_io:1;
52};
53
54int __init extent_io_init(void)
55{
56 extent_state_cache = kmem_cache_create("extent_state",
57 sizeof(struct extent_state), 0,
58 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
59 if (!extent_state_cache)
60 return -ENOMEM;
61
62 extent_buffer_cache = kmem_cache_create("extent_buffers",
63 sizeof(struct extent_buffer), 0,
64 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
65 if (!extent_buffer_cache)
66 goto free_state_cache;
67 return 0;
68
69free_state_cache:
70 kmem_cache_destroy(extent_state_cache);
71 return -ENOMEM;
72}
73
74void extent_io_exit(void)
75{
76 struct extent_state *state;
77 struct extent_buffer *eb;
78
79 while (!list_empty(&states)) {
80 state = list_entry(states.next, struct extent_state, leak_list);
81 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
82 "state %lu in tree %p refs %d\n",
83 (unsigned long long)state->start,
84 (unsigned long long)state->end,
85 state->state, state->tree, atomic_read(&state->refs));
86 list_del(&state->leak_list);
87 kmem_cache_free(extent_state_cache, state);
88
89 }
90
91 while (!list_empty(&buffers)) {
92 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
93 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
94 "refs %d\n", (unsigned long long)eb->start,
95 eb->len, atomic_read(&eb->refs));
96 list_del(&eb->leak_list);
97 kmem_cache_free(extent_buffer_cache, eb);
98 }
99 if (extent_state_cache)
100 kmem_cache_destroy(extent_state_cache);
101 if (extent_buffer_cache)
102 kmem_cache_destroy(extent_buffer_cache);
103}
104
105void extent_io_tree_init(struct extent_io_tree *tree,
106 struct address_space *mapping)
107{
108 tree->state = RB_ROOT;
109 INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
110 tree->ops = NULL;
111 tree->dirty_bytes = 0;
112 spin_lock_init(&tree->lock);
113 spin_lock_init(&tree->buffer_lock);
114 tree->mapping = mapping;
115}
116
117static struct extent_state *alloc_extent_state(gfp_t mask)
118{
119 struct extent_state *state;
120#if LEAK_DEBUG
121 unsigned long flags;
122#endif
123
124 state = kmem_cache_alloc(extent_state_cache, mask);
125 if (!state)
126 return state;
127 state->state = 0;
128 state->private = 0;
129 state->tree = NULL;
130#if LEAK_DEBUG
131 spin_lock_irqsave(&leak_lock, flags);
132 list_add(&state->leak_list, &states);
133 spin_unlock_irqrestore(&leak_lock, flags);
134#endif
135 atomic_set(&state->refs, 1);
136 init_waitqueue_head(&state->wq);
137 return state;
138}
139
140void free_extent_state(struct extent_state *state)
141{
142 if (!state)
143 return;
144 if (atomic_dec_and_test(&state->refs)) {
145#if LEAK_DEBUG
146 unsigned long flags;
147#endif
148 WARN_ON(state->tree);
149#if LEAK_DEBUG
150 spin_lock_irqsave(&leak_lock, flags);
151 list_del(&state->leak_list);
152 spin_unlock_irqrestore(&leak_lock, flags);
153#endif
154 kmem_cache_free(extent_state_cache, state);
155 }
156}
157
158static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
159 struct rb_node *node)
160{
161 struct rb_node **p = &root->rb_node;
162 struct rb_node *parent = NULL;
163 struct tree_entry *entry;
164
165 while (*p) {
166 parent = *p;
167 entry = rb_entry(parent, struct tree_entry, rb_node);
168
169 if (offset < entry->start)
170 p = &(*p)->rb_left;
171 else if (offset > entry->end)
172 p = &(*p)->rb_right;
173 else
174 return parent;
175 }
176
177 entry = rb_entry(node, struct tree_entry, rb_node);
178 rb_link_node(node, parent, p);
179 rb_insert_color(node, root);
180 return NULL;
181}
182
183static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
184 struct rb_node **prev_ret,
185 struct rb_node **next_ret)
186{
187 struct rb_root *root = &tree->state;
188 struct rb_node *n = root->rb_node;
189 struct rb_node *prev = NULL;
190 struct rb_node *orig_prev = NULL;
191 struct tree_entry *entry;
192 struct tree_entry *prev_entry = NULL;
193
194 while (n) {
195 entry = rb_entry(n, struct tree_entry, rb_node);
196 prev = n;
197 prev_entry = entry;
198
199 if (offset < entry->start)
200 n = n->rb_left;
201 else if (offset > entry->end)
202 n = n->rb_right;
203 else
204 return n;
205 }
206
207 if (prev_ret) {
208 orig_prev = prev;
209 while (prev && offset > prev_entry->end) {
210 prev = rb_next(prev);
211 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
212 }
213 *prev_ret = prev;
214 prev = orig_prev;
215 }
216
217 if (next_ret) {
218 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
219 while (prev && offset < prev_entry->start) {
220 prev = rb_prev(prev);
221 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
222 }
223 *next_ret = prev;
224 }
225 return NULL;
226}
227
228static inline struct rb_node *tree_search(struct extent_io_tree *tree,
229 u64 offset)
230{
231 struct rb_node *prev = NULL;
232 struct rb_node *ret;
233
234 ret = __etree_search(tree, offset, &prev, NULL);
235 if (!ret)
236 return prev;
237 return ret;
238}
239
240static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
241 struct extent_state *other)
242{
243 if (tree->ops && tree->ops->merge_extent_hook)
244 tree->ops->merge_extent_hook(tree->mapping->host, new,
245 other);
246}
247
248/*
249 * utility function to look for merge candidates inside a given range.
250 * Any extents with matching state are merged together into a single
251 * extent in the tree. Extents with EXTENT_IO in their state field
252 * are not merged because the end_io handlers need to be able to do
253 * operations on them without sleeping (or doing allocations/splits).
254 *
255 * This should be called with the tree lock held.
256 */
257static void merge_state(struct extent_io_tree *tree,
258 struct extent_state *state)
259{
260 struct extent_state *other;
261 struct rb_node *other_node;
262
263 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
264 return;
265
266 other_node = rb_prev(&state->rb_node);
267 if (other_node) {
268 other = rb_entry(other_node, struct extent_state, rb_node);
269 if (other->end == state->start - 1 &&
270 other->state == state->state) {
271 merge_cb(tree, state, other);
272 state->start = other->start;
273 other->tree = NULL;
274 rb_erase(&other->rb_node, &tree->state);
275 free_extent_state(other);
276 }
277 }
278 other_node = rb_next(&state->rb_node);
279 if (other_node) {
280 other = rb_entry(other_node, struct extent_state, rb_node);
281 if (other->start == state->end + 1 &&
282 other->state == state->state) {
283 merge_cb(tree, state, other);
284 state->end = other->end;
285 other->tree = NULL;
286 rb_erase(&other->rb_node, &tree->state);
287 free_extent_state(other);
288 }
289 }
290}
291
292static void set_state_cb(struct extent_io_tree *tree,
293 struct extent_state *state, int *bits)
294{
295 if (tree->ops && tree->ops->set_bit_hook)
296 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
297}
298
299static void clear_state_cb(struct extent_io_tree *tree,
300 struct extent_state *state, int *bits)
301{
302 if (tree->ops && tree->ops->clear_bit_hook)
303 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
304}
305
306static void set_state_bits(struct extent_io_tree *tree,
307 struct extent_state *state, int *bits);
308
309/*
310 * insert an extent_state struct into the tree. 'bits' are set on the
311 * struct before it is inserted.
312 *
313 * This may return -EEXIST if the extent is already there, in which case the
314 * state struct is freed.
315 *
316 * The tree lock is not taken internally. This is a utility function and
317 * probably isn't what you want to call (see set/clear_extent_bit).
318 */
319static int insert_state(struct extent_io_tree *tree,
320 struct extent_state *state, u64 start, u64 end,
321 int *bits)
322{
323 struct rb_node *node;
324
325 if (end < start) {
326 printk(KERN_ERR "btrfs end < start %llu %llu\n",
327 (unsigned long long)end,
328 (unsigned long long)start);
329 WARN_ON(1);
330 }
331 state->start = start;
332 state->end = end;
333
334 set_state_bits(tree, state, bits);
335
336 node = tree_insert(&tree->state, end, &state->rb_node);
337 if (node) {
338 struct extent_state *found;
339 found = rb_entry(node, struct extent_state, rb_node);
340 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
341 "%llu %llu\n", (unsigned long long)found->start,
342 (unsigned long long)found->end,
343 (unsigned long long)start, (unsigned long long)end);
344 return -EEXIST;
345 }
346 state->tree = tree;
347 merge_state(tree, state);
348 return 0;
349}
350
351static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
352 u64 split)
353{
354 if (tree->ops && tree->ops->split_extent_hook)
355 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
356}
357
358/*
359 * split a given extent state struct in two, inserting the preallocated
360 * struct 'prealloc' as the newly created second half. 'split' indicates an
361 * offset inside 'orig' where it should be split.
362 *
363 * Before calling,
364 * the tree has 'orig' at [orig->start, orig->end]. After calling, there
365 * are two extent state structs in the tree:
366 * prealloc: [orig->start, split - 1]
367 * orig: [ split, orig->end ]
368 *
369 * The tree locks are not taken by this function. They need to be held
370 * by the caller.
371 */
372static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
373 struct extent_state *prealloc, u64 split)
374{
375 struct rb_node *node;
376
377 split_cb(tree, orig, split);
378
379 prealloc->start = orig->start;
380 prealloc->end = split - 1;
381 prealloc->state = orig->state;
382 orig->start = split;
383
384 node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
385 if (node) {
386 free_extent_state(prealloc);
387 return -EEXIST;
388 }
389 prealloc->tree = tree;
390 return 0;
391}
392
393/*
394 * utility function to clear some bits in an extent state struct.
395 * it will optionally wake up any one waiting on this state (wake == 1), or
396 * forcibly remove the state from the tree (delete == 1).
397 *
398 * If no bits are set on the state struct after clearing things, the
399 * struct is freed and removed from the tree
400 */
401static int clear_state_bit(struct extent_io_tree *tree,
402 struct extent_state *state,
403 int *bits, int wake)
404{
405 int bits_to_clear = *bits & ~EXTENT_CTLBITS;
406 int ret = state->state & bits_to_clear;
407
408 if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
409 u64 range = state->end - state->start + 1;
410 WARN_ON(range > tree->dirty_bytes);
411 tree->dirty_bytes -= range;
412 }
413 clear_state_cb(tree, state, bits);
414 state->state &= ~bits_to_clear;
415 if (wake)
416 wake_up(&state->wq);
417 if (state->state == 0) {
418 if (state->tree) {
419 rb_erase(&state->rb_node, &tree->state);
420 state->tree = NULL;
421 free_extent_state(state);
422 } else {
423 WARN_ON(1);
424 }
425 } else {
426 merge_state(tree, state);
427 }
428 return ret;
429}
430
431static struct extent_state *
432alloc_extent_state_atomic(struct extent_state *prealloc)
433{
434 if (!prealloc)
435 prealloc = alloc_extent_state(GFP_ATOMIC);
436
437 return prealloc;
438}
439
440/*
441 * clear some bits on a range in the tree. This may require splitting
442 * or inserting elements in the tree, so the gfp mask is used to
443 * indicate which allocations or sleeping are allowed.
444 *
445 * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
446 * the given range from the tree regardless of state (ie for truncate).
447 *
448 * the range [start, end] is inclusive.
449 *
450 * This takes the tree lock, and returns < 0 on error, > 0 if any of the
451 * bits were already set, or zero if none of the bits were already set.
452 */
453int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
454 int bits, int wake, int delete,
455 struct extent_state **cached_state,
456 gfp_t mask)
457{
458 struct extent_state *state;
459 struct extent_state *cached;
460 struct extent_state *prealloc = NULL;
461 struct rb_node *next_node;
462 struct rb_node *node;
463 u64 last_end;
464 int err;
465 int set = 0;
466 int clear = 0;
467
468 if (delete)
469 bits |= ~EXTENT_CTLBITS;
470 bits |= EXTENT_FIRST_DELALLOC;
471
472 if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
473 clear = 1;
474again:
475 if (!prealloc && (mask & __GFP_WAIT)) {
476 prealloc = alloc_extent_state(mask);
477 if (!prealloc)
478 return -ENOMEM;
479 }
480
481 spin_lock(&tree->lock);
482 if (cached_state) {
483 cached = *cached_state;
484
485 if (clear) {
486 *cached_state = NULL;
487 cached_state = NULL;
488 }
489
490 if (cached && cached->tree && cached->start <= start &&
491 cached->end > start) {
492 if (clear)
493 atomic_dec(&cached->refs);
494 state = cached;
495 goto hit_next;
496 }
497 if (clear)
498 free_extent_state(cached);
499 }
500 /*
501 * this search will find the extents that end after
502 * our range starts
503 */
504 node = tree_search(tree, start);
505 if (!node)
506 goto out;
507 state = rb_entry(node, struct extent_state, rb_node);
508hit_next:
509 if (state->start > end)
510 goto out;
511 WARN_ON(state->end < start);
512 last_end = state->end;
513
514 /*
515 * | ---- desired range ---- |
516 * | state | or
517 * | ------------- state -------------- |
518 *
519 * We need to split the extent we found, and may flip
520 * bits on second half.
521 *
522 * If the extent we found extends past our range, we
523 * just split and search again. It'll get split again
524 * the next time though.
525 *
526 * If the extent we found is inside our range, we clear
527 * the desired bit on it.
528 */
529
530 if (state->start < start) {
531 prealloc = alloc_extent_state_atomic(prealloc);
532 BUG_ON(!prealloc);
533 err = split_state(tree, state, prealloc, start);
534 BUG_ON(err == -EEXIST);
535 prealloc = NULL;
536 if (err)
537 goto out;
538 if (state->end <= end) {
539 set |= clear_state_bit(tree, state, &bits, wake);
540 if (last_end == (u64)-1)
541 goto out;
542 start = last_end + 1;
543 }
544 goto search_again;
545 }
546 /*
547 * | ---- desired range ---- |
548 * | state |
549 * We need to split the extent, and clear the bit
550 * on the first half
551 */
552 if (state->start <= end && state->end > end) {
553 prealloc = alloc_extent_state_atomic(prealloc);
554 BUG_ON(!prealloc);
555 err = split_state(tree, state, prealloc, end + 1);
556 BUG_ON(err == -EEXIST);
557 if (wake)
558 wake_up(&state->wq);
559
560 set |= clear_state_bit(tree, prealloc, &bits, wake);
561
562 prealloc = NULL;
563 goto out;
564 }
565
566 if (state->end < end && prealloc && !need_resched())
567 next_node = rb_next(&state->rb_node);
568 else
569 next_node = NULL;
570
571 set |= clear_state_bit(tree, state, &bits, wake);
572 if (last_end == (u64)-1)
573 goto out;
574 start = last_end + 1;
575 if (start <= end && next_node) {
576 state = rb_entry(next_node, struct extent_state,
577 rb_node);
578 if (state->start == start)
579 goto hit_next;
580 }
581 goto search_again;
582
583out:
584 spin_unlock(&tree->lock);
585 if (prealloc)
586 free_extent_state(prealloc);
587
588 return set;
589
590search_again:
591 if (start > end)
592 goto out;
593 spin_unlock(&tree->lock);
594 if (mask & __GFP_WAIT)
595 cond_resched();
596 goto again;
597}
598
599static int wait_on_state(struct extent_io_tree *tree,
600 struct extent_state *state)
601 __releases(tree->lock)
602 __acquires(tree->lock)
603{
604 DEFINE_WAIT(wait);
605 prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
606 spin_unlock(&tree->lock);
607 schedule();
608 spin_lock(&tree->lock);
609 finish_wait(&state->wq, &wait);
610 return 0;
611}
612
613/*
614 * waits for one or more bits to clear on a range in the state tree.
615 * The range [start, end] is inclusive.
616 * The tree lock is taken by this function
617 */
618int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
619{
620 struct extent_state *state;
621 struct rb_node *node;
622
623 spin_lock(&tree->lock);
624again:
625 while (1) {
626 /*
627 * this search will find all the extents that end after
628 * our range starts
629 */
630 node = tree_search(tree, start);
631 if (!node)
632 break;
633
634 state = rb_entry(node, struct extent_state, rb_node);
635
636 if (state->start > end)
637 goto out;
638
639 if (state->state & bits) {
640 start = state->start;
641 atomic_inc(&state->refs);
642 wait_on_state(tree, state);
643 free_extent_state(state);
644 goto again;
645 }
646 start = state->end + 1;
647
648 if (start > end)
649 break;
650
651 cond_resched_lock(&tree->lock);
652 }
653out:
654 spin_unlock(&tree->lock);
655 return 0;
656}
657
658static void set_state_bits(struct extent_io_tree *tree,
659 struct extent_state *state,
660 int *bits)
661{
662 int bits_to_set = *bits & ~EXTENT_CTLBITS;
663
664 set_state_cb(tree, state, bits);
665 if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
666 u64 range = state->end - state->start + 1;
667 tree->dirty_bytes += range;
668 }
669 state->state |= bits_to_set;
670}
671
672static void cache_state(struct extent_state *state,
673 struct extent_state **cached_ptr)
674{
675 if (cached_ptr && !(*cached_ptr)) {
676 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
677 *cached_ptr = state;
678 atomic_inc(&state->refs);
679 }
680 }
681}
682
683static void uncache_state(struct extent_state **cached_ptr)
684{
685 if (cached_ptr && (*cached_ptr)) {
686 struct extent_state *state = *cached_ptr;
687 *cached_ptr = NULL;
688 free_extent_state(state);
689 }
690}
691
692/*
693 * set some bits on a range in the tree. This may require allocations or
694 * sleeping, so the gfp mask is used to indicate what is allowed.
695 *
696 * If any of the exclusive bits are set, this will fail with -EEXIST if some
697 * part of the range already has the desired bits set. The start of the
698 * existing range is returned in failed_start in this case.
699 *
700 * [start, end] is inclusive This takes the tree lock.
701 */
702
703int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
704 int bits, int exclusive_bits, u64 *failed_start,
705 struct extent_state **cached_state, gfp_t mask)
706{
707 struct extent_state *state;
708 struct extent_state *prealloc = NULL;
709 struct rb_node *node;
710 int err = 0;
711 u64 last_start;
712 u64 last_end;
713
714 bits |= EXTENT_FIRST_DELALLOC;
715again:
716 if (!prealloc && (mask & __GFP_WAIT)) {
717 prealloc = alloc_extent_state(mask);
718 BUG_ON(!prealloc);
719 }
720
721 spin_lock(&tree->lock);
722 if (cached_state && *cached_state) {
723 state = *cached_state;
724 if (state->start <= start && state->end > start &&
725 state->tree) {
726 node = &state->rb_node;
727 goto hit_next;
728 }
729 }
730 /*
731 * this search will find all the extents that end after
732 * our range starts.
733 */
734 node = tree_search(tree, start);
735 if (!node) {
736 prealloc = alloc_extent_state_atomic(prealloc);
737 BUG_ON(!prealloc);
738 err = insert_state(tree, prealloc, start, end, &bits);
739 prealloc = NULL;
740 BUG_ON(err == -EEXIST);
741 goto out;
742 }
743 state = rb_entry(node, struct extent_state, rb_node);
744hit_next:
745 last_start = state->start;
746 last_end = state->end;
747
748 /*
749 * | ---- desired range ---- |
750 * | state |
751 *
752 * Just lock what we found and keep going
753 */
754 if (state->start == start && state->end <= end) {
755 struct rb_node *next_node;
756 if (state->state & exclusive_bits) {
757 *failed_start = state->start;
758 err = -EEXIST;
759 goto out;
760 }
761
762 set_state_bits(tree, state, &bits);
763
764 cache_state(state, cached_state);
765 merge_state(tree, state);
766 if (last_end == (u64)-1)
767 goto out;
768
769 start = last_end + 1;
770 next_node = rb_next(&state->rb_node);
771 if (next_node && start < end && prealloc && !need_resched()) {
772 state = rb_entry(next_node, struct extent_state,
773 rb_node);
774 if (state->start == start)
775 goto hit_next;
776 }
777 goto search_again;
778 }
779
780 /*
781 * | ---- desired range ---- |
782 * | state |
783 * or
784 * | ------------- state -------------- |
785 *
786 * We need to split the extent we found, and may flip bits on
787 * second half.
788 *
789 * If the extent we found extends past our
790 * range, we just split and search again. It'll get split
791 * again the next time though.
792 *
793 * If the extent we found is inside our range, we set the
794 * desired bit on it.
795 */
796 if (state->start < start) {
797 if (state->state & exclusive_bits) {
798 *failed_start = start;
799 err = -EEXIST;
800 goto out;
801 }
802
803 prealloc = alloc_extent_state_atomic(prealloc);
804 BUG_ON(!prealloc);
805 err = split_state(tree, state, prealloc, start);
806 BUG_ON(err == -EEXIST);
807 prealloc = NULL;
808 if (err)
809 goto out;
810 if (state->end <= end) {
811 set_state_bits(tree, state, &bits);
812 cache_state(state, cached_state);
813 merge_state(tree, state);
814 if (last_end == (u64)-1)
815 goto out;
816 start = last_end + 1;
817 }
818 goto search_again;
819 }
820 /*
821 * | ---- desired range ---- |
822 * | state | or | state |
823 *
824 * There's a hole, we need to insert something in it and
825 * ignore the extent we found.
826 */
827 if (state->start > start) {
828 u64 this_end;
829 if (end < last_start)
830 this_end = end;
831 else
832 this_end = last_start - 1;
833
834 prealloc = alloc_extent_state_atomic(prealloc);
835 BUG_ON(!prealloc);
836
837 /*
838 * Avoid to free 'prealloc' if it can be merged with
839 * the later extent.
840 */
841 err = insert_state(tree, prealloc, start, this_end,
842 &bits);
843 BUG_ON(err == -EEXIST);
844 if (err) {
845 free_extent_state(prealloc);
846 prealloc = NULL;
847 goto out;
848 }
849 cache_state(prealloc, cached_state);
850 prealloc = NULL;
851 start = this_end + 1;
852 goto search_again;
853 }
854 /*
855 * | ---- desired range ---- |
856 * | state |
857 * We need to split the extent, and set the bit
858 * on the first half
859 */
860 if (state->start <= end && state->end > end) {
861 if (state->state & exclusive_bits) {
862 *failed_start = start;
863 err = -EEXIST;
864 goto out;
865 }
866
867 prealloc = alloc_extent_state_atomic(prealloc);
868 BUG_ON(!prealloc);
869 err = split_state(tree, state, prealloc, end + 1);
870 BUG_ON(err == -EEXIST);
871
872 set_state_bits(tree, prealloc, &bits);
873 cache_state(prealloc, cached_state);
874 merge_state(tree, prealloc);
875 prealloc = NULL;
876 goto out;
877 }
878
879 goto search_again;
880
881out:
882 spin_unlock(&tree->lock);
883 if (prealloc)
884 free_extent_state(prealloc);
885
886 return err;
887
888search_again:
889 if (start > end)
890 goto out;
891 spin_unlock(&tree->lock);
892 if (mask & __GFP_WAIT)
893 cond_resched();
894 goto again;
895}
896
897/* wrappers around set/clear extent bit */
898int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
899 gfp_t mask)
900{
901 return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
902 NULL, mask);
903}
904
905int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
906 int bits, gfp_t mask)
907{
908 return set_extent_bit(tree, start, end, bits, 0, NULL,
909 NULL, mask);
910}
911
912int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
913 int bits, gfp_t mask)
914{
915 return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
916}
917
918int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
919 struct extent_state **cached_state, gfp_t mask)
920{
921 return set_extent_bit(tree, start, end,
922 EXTENT_DELALLOC | EXTENT_DIRTY | EXTENT_UPTODATE,
923 0, NULL, cached_state, mask);
924}
925
926int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
927 gfp_t mask)
928{
929 return clear_extent_bit(tree, start, end,
930 EXTENT_DIRTY | EXTENT_DELALLOC |
931 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
932}
933
934int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
935 gfp_t mask)
936{
937 return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
938 NULL, mask);
939}
940
941int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
942 struct extent_state **cached_state, gfp_t mask)
943{
944 return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
945 NULL, cached_state, mask);
946}
947
948static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
949 u64 end, struct extent_state **cached_state,
950 gfp_t mask)
951{
952 return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
953 cached_state, mask);
954}
955
956/*
957 * either insert or lock state struct between start and end use mask to tell
958 * us if waiting is desired.
959 */
960int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
961 int bits, struct extent_state **cached_state, gfp_t mask)
962{
963 int err;
964 u64 failed_start;
965 while (1) {
966 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
967 EXTENT_LOCKED, &failed_start,
968 cached_state, mask);
969 if (err == -EEXIST && (mask & __GFP_WAIT)) {
970 wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
971 start = failed_start;
972 } else {
973 break;
974 }
975 WARN_ON(start > end);
976 }
977 return err;
978}
979
980int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
981{
982 return lock_extent_bits(tree, start, end, 0, NULL, mask);
983}
984
985int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
986 gfp_t mask)
987{
988 int err;
989 u64 failed_start;
990
991 err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
992 &failed_start, NULL, mask);
993 if (err == -EEXIST) {
994 if (failed_start > start)
995 clear_extent_bit(tree, start, failed_start - 1,
996 EXTENT_LOCKED, 1, 0, NULL, mask);
997 return 0;
998 }
999 return 1;
1000}
1001
1002int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1003 struct extent_state **cached, gfp_t mask)
1004{
1005 return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1006 mask);
1007}
1008
1009int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1010{
1011 return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1012 mask);
1013}
1014
1015/*
1016 * helper function to set both pages and extents in the tree writeback
1017 */
1018static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1019{
1020 unsigned long index = start >> PAGE_CACHE_SHIFT;
1021 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1022 struct page *page;
1023
1024 while (index <= end_index) {
1025 page = find_get_page(tree->mapping, index);
1026 BUG_ON(!page);
1027 set_page_writeback(page);
1028 page_cache_release(page);
1029 index++;
1030 }
1031 return 0;
1032}
1033
1034/* find the first state struct with 'bits' set after 'start', and
1035 * return it. tree->lock must be held. NULL will returned if
1036 * nothing was found after 'start'
1037 */
1038struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1039 u64 start, int bits)
1040{
1041 struct rb_node *node;
1042 struct extent_state *state;
1043
1044 /*
1045 * this search will find all the extents that end after
1046 * our range starts.
1047 */
1048 node = tree_search(tree, start);
1049 if (!node)
1050 goto out;
1051
1052 while (1) {
1053 state = rb_entry(node, struct extent_state, rb_node);
1054 if (state->end >= start && (state->state & bits))
1055 return state;
1056
1057 node = rb_next(node);
1058 if (!node)
1059 break;
1060 }
1061out:
1062 return NULL;
1063}
1064
1065/*
1066 * find the first offset in the io tree with 'bits' set. zero is
1067 * returned if we find something, and *start_ret and *end_ret are
1068 * set to reflect the state struct that was found.
1069 *
1070 * If nothing was found, 1 is returned, < 0 on error
1071 */
1072int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1073 u64 *start_ret, u64 *end_ret, int bits)
1074{
1075 struct extent_state *state;
1076 int ret = 1;
1077
1078 spin_lock(&tree->lock);
1079 state = find_first_extent_bit_state(tree, start, bits);
1080 if (state) {
1081 *start_ret = state->start;
1082 *end_ret = state->end;
1083 ret = 0;
1084 }
1085 spin_unlock(&tree->lock);
1086 return ret;
1087}
1088
1089/*
1090 * find a contiguous range of bytes in the file marked as delalloc, not
1091 * more than 'max_bytes'. start and end are used to return the range,
1092 *
1093 * 1 is returned if we find something, 0 if nothing was in the tree
1094 */
1095static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1096 u64 *start, u64 *end, u64 max_bytes,
1097 struct extent_state **cached_state)
1098{
1099 struct rb_node *node;
1100 struct extent_state *state;
1101 u64 cur_start = *start;
1102 u64 found = 0;
1103 u64 total_bytes = 0;
1104
1105 spin_lock(&tree->lock);
1106
1107 /*
1108 * this search will find all the extents that end after
1109 * our range starts.
1110 */
1111 node = tree_search(tree, cur_start);
1112 if (!node) {
1113 if (!found)
1114 *end = (u64)-1;
1115 goto out;
1116 }
1117
1118 while (1) {
1119 state = rb_entry(node, struct extent_state, rb_node);
1120 if (found && (state->start != cur_start ||
1121 (state->state & EXTENT_BOUNDARY))) {
1122 goto out;
1123 }
1124 if (!(state->state & EXTENT_DELALLOC)) {
1125 if (!found)
1126 *end = state->end;
1127 goto out;
1128 }
1129 if (!found) {
1130 *start = state->start;
1131 *cached_state = state;
1132 atomic_inc(&state->refs);
1133 }
1134 found++;
1135 *end = state->end;
1136 cur_start = state->end + 1;
1137 node = rb_next(node);
1138 if (!node)
1139 break;
1140 total_bytes += state->end - state->start + 1;
1141 if (total_bytes >= max_bytes)
1142 break;
1143 }
1144out:
1145 spin_unlock(&tree->lock);
1146 return found;
1147}
1148
1149static noinline int __unlock_for_delalloc(struct inode *inode,
1150 struct page *locked_page,
1151 u64 start, u64 end)
1152{
1153 int ret;
1154 struct page *pages[16];
1155 unsigned long index = start >> PAGE_CACHE_SHIFT;
1156 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1157 unsigned long nr_pages = end_index - index + 1;
1158 int i;
1159
1160 if (index == locked_page->index && end_index == index)
1161 return 0;
1162
1163 while (nr_pages > 0) {
1164 ret = find_get_pages_contig(inode->i_mapping, index,
1165 min_t(unsigned long, nr_pages,
1166 ARRAY_SIZE(pages)), pages);
1167 for (i = 0; i < ret; i++) {
1168 if (pages[i] != locked_page)
1169 unlock_page(pages[i]);
1170 page_cache_release(pages[i]);
1171 }
1172 nr_pages -= ret;
1173 index += ret;
1174 cond_resched();
1175 }
1176 return 0;
1177}
1178
1179static noinline int lock_delalloc_pages(struct inode *inode,
1180 struct page *locked_page,
1181 u64 delalloc_start,
1182 u64 delalloc_end)
1183{
1184 unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1185 unsigned long start_index = index;
1186 unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1187 unsigned long pages_locked = 0;
1188 struct page *pages[16];
1189 unsigned long nrpages;
1190 int ret;
1191 int i;
1192
1193 /* the caller is responsible for locking the start index */
1194 if (index == locked_page->index && index == end_index)
1195 return 0;
1196
1197 /* skip the page at the start index */
1198 nrpages = end_index - index + 1;
1199 while (nrpages > 0) {
1200 ret = find_get_pages_contig(inode->i_mapping, index,
1201 min_t(unsigned long,
1202 nrpages, ARRAY_SIZE(pages)), pages);
1203 if (ret == 0) {
1204 ret = -EAGAIN;
1205 goto done;
1206 }
1207 /* now we have an array of pages, lock them all */
1208 for (i = 0; i < ret; i++) {
1209 /*
1210 * the caller is taking responsibility for
1211 * locked_page
1212 */
1213 if (pages[i] != locked_page) {
1214 lock_page(pages[i]);
1215 if (!PageDirty(pages[i]) ||
1216 pages[i]->mapping != inode->i_mapping) {
1217 ret = -EAGAIN;
1218 unlock_page(pages[i]);
1219 page_cache_release(pages[i]);
1220 goto done;
1221 }
1222 }
1223 page_cache_release(pages[i]);
1224 pages_locked++;
1225 }
1226 nrpages -= ret;
1227 index += ret;
1228 cond_resched();
1229 }
1230 ret = 0;
1231done:
1232 if (ret && pages_locked) {
1233 __unlock_for_delalloc(inode, locked_page,
1234 delalloc_start,
1235 ((u64)(start_index + pages_locked - 1)) <<
1236 PAGE_CACHE_SHIFT);
1237 }
1238 return ret;
1239}
1240
1241/*
1242 * find a contiguous range of bytes in the file marked as delalloc, not
1243 * more than 'max_bytes'. start and end are used to return the range,
1244 *
1245 * 1 is returned if we find something, 0 if nothing was in the tree
1246 */
1247static noinline u64 find_lock_delalloc_range(struct inode *inode,
1248 struct extent_io_tree *tree,
1249 struct page *locked_page,
1250 u64 *start, u64 *end,
1251 u64 max_bytes)
1252{
1253 u64 delalloc_start;
1254 u64 delalloc_end;
1255 u64 found;
1256 struct extent_state *cached_state = NULL;
1257 int ret;
1258 int loops = 0;
1259
1260again:
1261 /* step one, find a bunch of delalloc bytes starting at start */
1262 delalloc_start = *start;
1263 delalloc_end = 0;
1264 found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1265 max_bytes, &cached_state);
1266 if (!found || delalloc_end <= *start) {
1267 *start = delalloc_start;
1268 *end = delalloc_end;
1269 free_extent_state(cached_state);
1270 return found;
1271 }
1272
1273 /*
1274 * start comes from the offset of locked_page. We have to lock
1275 * pages in order, so we can't process delalloc bytes before
1276 * locked_page
1277 */
1278 if (delalloc_start < *start)
1279 delalloc_start = *start;
1280
1281 /*
1282 * make sure to limit the number of pages we try to lock down
1283 * if we're looping.
1284 */
1285 if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1286 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1287
1288 /* step two, lock all the pages after the page that has start */
1289 ret = lock_delalloc_pages(inode, locked_page,
1290 delalloc_start, delalloc_end);
1291 if (ret == -EAGAIN) {
1292 /* some of the pages are gone, lets avoid looping by
1293 * shortening the size of the delalloc range we're searching
1294 */
1295 free_extent_state(cached_state);
1296 if (!loops) {
1297 unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1298 max_bytes = PAGE_CACHE_SIZE - offset;
1299 loops = 1;
1300 goto again;
1301 } else {
1302 found = 0;
1303 goto out_failed;
1304 }
1305 }
1306 BUG_ON(ret);
1307
1308 /* step three, lock the state bits for the whole range */
1309 lock_extent_bits(tree, delalloc_start, delalloc_end,
1310 0, &cached_state, GFP_NOFS);
1311
1312 /* then test to make sure it is all still delalloc */
1313 ret = test_range_bit(tree, delalloc_start, delalloc_end,
1314 EXTENT_DELALLOC, 1, cached_state);
1315 if (!ret) {
1316 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1317 &cached_state, GFP_NOFS);
1318 __unlock_for_delalloc(inode, locked_page,
1319 delalloc_start, delalloc_end);
1320 cond_resched();
1321 goto again;
1322 }
1323 free_extent_state(cached_state);
1324 *start = delalloc_start;
1325 *end = delalloc_end;
1326out_failed:
1327 return found;
1328}
1329
1330int extent_clear_unlock_delalloc(struct inode *inode,
1331 struct extent_io_tree *tree,
1332 u64 start, u64 end, struct page *locked_page,
1333 unsigned long op)
1334{
1335 int ret;
1336 struct page *pages[16];
1337 unsigned long index = start >> PAGE_CACHE_SHIFT;
1338 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1339 unsigned long nr_pages = end_index - index + 1;
1340 int i;
1341 int clear_bits = 0;
1342
1343 if (op & EXTENT_CLEAR_UNLOCK)
1344 clear_bits |= EXTENT_LOCKED;
1345 if (op & EXTENT_CLEAR_DIRTY)
1346 clear_bits |= EXTENT_DIRTY;
1347
1348 if (op & EXTENT_CLEAR_DELALLOC)
1349 clear_bits |= EXTENT_DELALLOC;
1350
1351 clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1352 if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1353 EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1354 EXTENT_SET_PRIVATE2)))
1355 return 0;
1356
1357 while (nr_pages > 0) {
1358 ret = find_get_pages_contig(inode->i_mapping, index,
1359 min_t(unsigned long,
1360 nr_pages, ARRAY_SIZE(pages)), pages);
1361 for (i = 0; i < ret; i++) {
1362
1363 if (op & EXTENT_SET_PRIVATE2)
1364 SetPagePrivate2(pages[i]);
1365
1366 if (pages[i] == locked_page) {
1367 page_cache_release(pages[i]);
1368 continue;
1369 }
1370 if (op & EXTENT_CLEAR_DIRTY)
1371 clear_page_dirty_for_io(pages[i]);
1372 if (op & EXTENT_SET_WRITEBACK)
1373 set_page_writeback(pages[i]);
1374 if (op & EXTENT_END_WRITEBACK)
1375 end_page_writeback(pages[i]);
1376 if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1377 unlock_page(pages[i]);
1378 page_cache_release(pages[i]);
1379 }
1380 nr_pages -= ret;
1381 index += ret;
1382 cond_resched();
1383 }
1384 return 0;
1385}
1386
1387/*
1388 * count the number of bytes in the tree that have a given bit(s)
1389 * set. This can be fairly slow, except for EXTENT_DIRTY which is
1390 * cached. The total number found is returned.
1391 */
1392u64 count_range_bits(struct extent_io_tree *tree,
1393 u64 *start, u64 search_end, u64 max_bytes,
1394 unsigned long bits, int contig)
1395{
1396 struct rb_node *node;
1397 struct extent_state *state;
1398 u64 cur_start = *start;
1399 u64 total_bytes = 0;
1400 u64 last = 0;
1401 int found = 0;
1402
1403 if (search_end <= cur_start) {
1404 WARN_ON(1);
1405 return 0;
1406 }
1407
1408 spin_lock(&tree->lock);
1409 if (cur_start == 0 && bits == EXTENT_DIRTY) {
1410 total_bytes = tree->dirty_bytes;
1411 goto out;
1412 }
1413 /*
1414 * this search will find all the extents that end after
1415 * our range starts.
1416 */
1417 node = tree_search(tree, cur_start);
1418 if (!node)
1419 goto out;
1420
1421 while (1) {
1422 state = rb_entry(node, struct extent_state, rb_node);
1423 if (state->start > search_end)
1424 break;
1425 if (contig && found && state->start > last + 1)
1426 break;
1427 if (state->end >= cur_start && (state->state & bits) == bits) {
1428 total_bytes += min(search_end, state->end) + 1 -
1429 max(cur_start, state->start);
1430 if (total_bytes >= max_bytes)
1431 break;
1432 if (!found) {
1433 *start = max(cur_start, state->start);
1434 found = 1;
1435 }
1436 last = state->end;
1437 } else if (contig && found) {
1438 break;
1439 }
1440 node = rb_next(node);
1441 if (!node)
1442 break;
1443 }
1444out:
1445 spin_unlock(&tree->lock);
1446 return total_bytes;
1447}
1448
1449/*
1450 * set the private field for a given byte offset in the tree. If there isn't
1451 * an extent_state there already, this does nothing.
1452 */
1453int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1454{
1455 struct rb_node *node;
1456 struct extent_state *state;
1457 int ret = 0;
1458
1459 spin_lock(&tree->lock);
1460 /*
1461 * this search will find all the extents that end after
1462 * our range starts.
1463 */
1464 node = tree_search(tree, start);
1465 if (!node) {
1466 ret = -ENOENT;
1467 goto out;
1468 }
1469 state = rb_entry(node, struct extent_state, rb_node);
1470 if (state->start != start) {
1471 ret = -ENOENT;
1472 goto out;
1473 }
1474 state->private = private;
1475out:
1476 spin_unlock(&tree->lock);
1477 return ret;
1478}
1479
1480int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1481{
1482 struct rb_node *node;
1483 struct extent_state *state;
1484 int ret = 0;
1485
1486 spin_lock(&tree->lock);
1487 /*
1488 * this search will find all the extents that end after
1489 * our range starts.
1490 */
1491 node = tree_search(tree, start);
1492 if (!node) {
1493 ret = -ENOENT;
1494 goto out;
1495 }
1496 state = rb_entry(node, struct extent_state, rb_node);
1497 if (state->start != start) {
1498 ret = -ENOENT;
1499 goto out;
1500 }
1501 *private = state->private;
1502out:
1503 spin_unlock(&tree->lock);
1504 return ret;
1505}
1506
1507/*
1508 * searches a range in the state tree for a given mask.
1509 * If 'filled' == 1, this returns 1 only if every extent in the tree
1510 * has the bits set. Otherwise, 1 is returned if any bit in the
1511 * range is found set.
1512 */
1513int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1514 int bits, int filled, struct extent_state *cached)
1515{
1516 struct extent_state *state = NULL;
1517 struct rb_node *node;
1518 int bitset = 0;
1519
1520 spin_lock(&tree->lock);
1521 if (cached && cached->tree && cached->start <= start &&
1522 cached->end > start)
1523 node = &cached->rb_node;
1524 else
1525 node = tree_search(tree, start);
1526 while (node && start <= end) {
1527 state = rb_entry(node, struct extent_state, rb_node);
1528
1529 if (filled && state->start > start) {
1530 bitset = 0;
1531 break;
1532 }
1533
1534 if (state->start > end)
1535 break;
1536
1537 if (state->state & bits) {
1538 bitset = 1;
1539 if (!filled)
1540 break;
1541 } else if (filled) {
1542 bitset = 0;
1543 break;
1544 }
1545
1546 if (state->end == (u64)-1)
1547 break;
1548
1549 start = state->end + 1;
1550 if (start > end)
1551 break;
1552 node = rb_next(node);
1553 if (!node) {
1554 if (filled)
1555 bitset = 0;
1556 break;
1557 }
1558 }
1559 spin_unlock(&tree->lock);
1560 return bitset;
1561}
1562
1563/*
1564 * helper function to set a given page up to date if all the
1565 * extents in the tree for that page are up to date
1566 */
1567static int check_page_uptodate(struct extent_io_tree *tree,
1568 struct page *page)
1569{
1570 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1571 u64 end = start + PAGE_CACHE_SIZE - 1;
1572 if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1573 SetPageUptodate(page);
1574 return 0;
1575}
1576
1577/*
1578 * helper function to unlock a page if all the extents in the tree
1579 * for that page are unlocked
1580 */
1581static int check_page_locked(struct extent_io_tree *tree,
1582 struct page *page)
1583{
1584 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1585 u64 end = start + PAGE_CACHE_SIZE - 1;
1586 if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1587 unlock_page(page);
1588 return 0;
1589}
1590
1591/*
1592 * helper function to end page writeback if all the extents
1593 * in the tree for that page are done with writeback
1594 */
1595static int check_page_writeback(struct extent_io_tree *tree,
1596 struct page *page)
1597{
1598 end_page_writeback(page);
1599 return 0;
1600}
1601
1602/* lots and lots of room for performance fixes in the end_bio funcs */
1603
1604/*
1605 * after a writepage IO is done, we need to:
1606 * clear the uptodate bits on error
1607 * clear the writeback bits in the extent tree for this IO
1608 * end_page_writeback if the page has no more pending IO
1609 *
1610 * Scheduling is not allowed, so the extent state tree is expected
1611 * to have one and only one object corresponding to this IO.
1612 */
1613static void end_bio_extent_writepage(struct bio *bio, int err)
1614{
1615 int uptodate = err == 0;
1616 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1617 struct extent_io_tree *tree;
1618 u64 start;
1619 u64 end;
1620 int whole_page;
1621 int ret;
1622
1623 do {
1624 struct page *page = bvec->bv_page;
1625 tree = &BTRFS_I(page->mapping->host)->io_tree;
1626
1627 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1628 bvec->bv_offset;
1629 end = start + bvec->bv_len - 1;
1630
1631 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1632 whole_page = 1;
1633 else
1634 whole_page = 0;
1635
1636 if (--bvec >= bio->bi_io_vec)
1637 prefetchw(&bvec->bv_page->flags);
1638 if (tree->ops && tree->ops->writepage_end_io_hook) {
1639 ret = tree->ops->writepage_end_io_hook(page, start,
1640 end, NULL, uptodate);
1641 if (ret)
1642 uptodate = 0;
1643 }
1644
1645 if (!uptodate && tree->ops &&
1646 tree->ops->writepage_io_failed_hook) {
1647 ret = tree->ops->writepage_io_failed_hook(bio, page,
1648 start, end, NULL);
1649 if (ret == 0) {
1650 uptodate = (err == 0);
1651 continue;
1652 }
1653 }
1654
1655 if (!uptodate) {
1656 clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
1657 ClearPageUptodate(page);
1658 SetPageError(page);
1659 }
1660
1661 if (whole_page)
1662 end_page_writeback(page);
1663 else
1664 check_page_writeback(tree, page);
1665 } while (bvec >= bio->bi_io_vec);
1666
1667 bio_put(bio);
1668}
1669
1670/*
1671 * after a readpage IO is done, we need to:
1672 * clear the uptodate bits on error
1673 * set the uptodate bits if things worked
1674 * set the page up to date if all extents in the tree are uptodate
1675 * clear the lock bit in the extent tree
1676 * unlock the page if there are no other extents locked for it
1677 *
1678 * Scheduling is not allowed, so the extent state tree is expected
1679 * to have one and only one object corresponding to this IO.
1680 */
1681static void end_bio_extent_readpage(struct bio *bio, int err)
1682{
1683 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1684 struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
1685 struct bio_vec *bvec = bio->bi_io_vec;
1686 struct extent_io_tree *tree;
1687 u64 start;
1688 u64 end;
1689 int whole_page;
1690 int ret;
1691
1692 if (err)
1693 uptodate = 0;
1694
1695 do {
1696 struct page *page = bvec->bv_page;
1697 struct extent_state *cached = NULL;
1698 struct extent_state *state;
1699
1700 tree = &BTRFS_I(page->mapping->host)->io_tree;
1701
1702 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1703 bvec->bv_offset;
1704 end = start + bvec->bv_len - 1;
1705
1706 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1707 whole_page = 1;
1708 else
1709 whole_page = 0;
1710
1711 if (++bvec <= bvec_end)
1712 prefetchw(&bvec->bv_page->flags);
1713
1714 spin_lock(&tree->lock);
1715 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
1716 if (state && state->start == start) {
1717 /*
1718 * take a reference on the state, unlock will drop
1719 * the ref
1720 */
1721 cache_state(state, &cached);
1722 }
1723 spin_unlock(&tree->lock);
1724
1725 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1726 ret = tree->ops->readpage_end_io_hook(page, start, end,
1727 state);
1728 if (ret)
1729 uptodate = 0;
1730 }
1731 if (!uptodate && tree->ops &&
1732 tree->ops->readpage_io_failed_hook) {
1733 ret = tree->ops->readpage_io_failed_hook(bio, page,
1734 start, end, NULL);
1735 if (ret == 0) {
1736 uptodate =
1737 test_bit(BIO_UPTODATE, &bio->bi_flags);
1738 if (err)
1739 uptodate = 0;
1740 uncache_state(&cached);
1741 continue;
1742 }
1743 }
1744
1745 if (uptodate) {
1746 set_extent_uptodate(tree, start, end, &cached,
1747 GFP_ATOMIC);
1748 }
1749 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
1750
1751 if (whole_page) {
1752 if (uptodate) {
1753 SetPageUptodate(page);
1754 } else {
1755 ClearPageUptodate(page);
1756 SetPageError(page);
1757 }
1758 unlock_page(page);
1759 } else {
1760 if (uptodate) {
1761 check_page_uptodate(tree, page);
1762 } else {
1763 ClearPageUptodate(page);
1764 SetPageError(page);
1765 }
1766 check_page_locked(tree, page);
1767 }
1768 } while (bvec <= bvec_end);
1769
1770 bio_put(bio);
1771}
1772
1773struct bio *
1774btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1775 gfp_t gfp_flags)
1776{
1777 struct bio *bio;
1778
1779 bio = bio_alloc(gfp_flags, nr_vecs);
1780
1781 if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1782 while (!bio && (nr_vecs /= 2))
1783 bio = bio_alloc(gfp_flags, nr_vecs);
1784 }
1785
1786 if (bio) {
1787 bio->bi_size = 0;
1788 bio->bi_bdev = bdev;
1789 bio->bi_sector = first_sector;
1790 }
1791 return bio;
1792}
1793
1794static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1795 unsigned long bio_flags)
1796{
1797 int ret = 0;
1798 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1799 struct page *page = bvec->bv_page;
1800 struct extent_io_tree *tree = bio->bi_private;
1801 u64 start;
1802
1803 start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1804
1805 bio->bi_private = NULL;
1806
1807 bio_get(bio);
1808
1809 if (tree->ops && tree->ops->submit_bio_hook)
1810 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1811 mirror_num, bio_flags, start);
1812 else
1813 submit_bio(rw, bio);
1814 if (bio_flagged(bio, BIO_EOPNOTSUPP))
1815 ret = -EOPNOTSUPP;
1816 bio_put(bio);
1817 return ret;
1818}
1819
1820static int submit_extent_page(int rw, struct extent_io_tree *tree,
1821 struct page *page, sector_t sector,
1822 size_t size, unsigned long offset,
1823 struct block_device *bdev,
1824 struct bio **bio_ret,
1825 unsigned long max_pages,
1826 bio_end_io_t end_io_func,
1827 int mirror_num,
1828 unsigned long prev_bio_flags,
1829 unsigned long bio_flags)
1830{
1831 int ret = 0;
1832 struct bio *bio;
1833 int nr;
1834 int contig = 0;
1835 int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1836 int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1837 size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1838
1839 if (bio_ret && *bio_ret) {
1840 bio = *bio_ret;
1841 if (old_compressed)
1842 contig = bio->bi_sector == sector;
1843 else
1844 contig = bio->bi_sector + (bio->bi_size >> 9) ==
1845 sector;
1846
1847 if (prev_bio_flags != bio_flags || !contig ||
1848 (tree->ops && tree->ops->merge_bio_hook &&
1849 tree->ops->merge_bio_hook(page, offset, page_size, bio,
1850 bio_flags)) ||
1851 bio_add_page(bio, page, page_size, offset) < page_size) {
1852 ret = submit_one_bio(rw, bio, mirror_num,
1853 prev_bio_flags);
1854 bio = NULL;
1855 } else {
1856 return 0;
1857 }
1858 }
1859 if (this_compressed)
1860 nr = BIO_MAX_PAGES;
1861 else
1862 nr = bio_get_nr_vecs(bdev);
1863
1864 bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1865 if (!bio)
1866 return -ENOMEM;
1867
1868 bio_add_page(bio, page, page_size, offset);
1869 bio->bi_end_io = end_io_func;
1870 bio->bi_private = tree;
1871
1872 if (bio_ret)
1873 *bio_ret = bio;
1874 else
1875 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1876
1877 return ret;
1878}
1879
1880void set_page_extent_mapped(struct page *page)
1881{
1882 if (!PagePrivate(page)) {
1883 SetPagePrivate(page);
1884 page_cache_get(page);
1885 set_page_private(page, EXTENT_PAGE_PRIVATE);
1886 }
1887}
1888
1889static void set_page_extent_head(struct page *page, unsigned long len)
1890{
1891 WARN_ON(!PagePrivate(page));
1892 set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1893}
1894
1895/*
1896 * basic readpage implementation. Locked extent state structs are inserted
1897 * into the tree that are removed when the IO is done (by the end_io
1898 * handlers)
1899 */
1900static int __extent_read_full_page(struct extent_io_tree *tree,
1901 struct page *page,
1902 get_extent_t *get_extent,
1903 struct bio **bio, int mirror_num,
1904 unsigned long *bio_flags)
1905{
1906 struct inode *inode = page->mapping->host;
1907 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1908 u64 page_end = start + PAGE_CACHE_SIZE - 1;
1909 u64 end;
1910 u64 cur = start;
1911 u64 extent_offset;
1912 u64 last_byte = i_size_read(inode);
1913 u64 block_start;
1914 u64 cur_end;
1915 sector_t sector;
1916 struct extent_map *em;
1917 struct block_device *bdev;
1918 struct btrfs_ordered_extent *ordered;
1919 int ret;
1920 int nr = 0;
1921 size_t pg_offset = 0;
1922 size_t iosize;
1923 size_t disk_io_size;
1924 size_t blocksize = inode->i_sb->s_blocksize;
1925 unsigned long this_bio_flag = 0;
1926
1927 set_page_extent_mapped(page);
1928
1929 if (!PageUptodate(page)) {
1930 if (cleancache_get_page(page) == 0) {
1931 BUG_ON(blocksize != PAGE_SIZE);
1932 goto out;
1933 }
1934 }
1935
1936 end = page_end;
1937 while (1) {
1938 lock_extent(tree, start, end, GFP_NOFS);
1939 ordered = btrfs_lookup_ordered_extent(inode, start);
1940 if (!ordered)
1941 break;
1942 unlock_extent(tree, start, end, GFP_NOFS);
1943 btrfs_start_ordered_extent(inode, ordered, 1);
1944 btrfs_put_ordered_extent(ordered);
1945 }
1946
1947 if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
1948 char *userpage;
1949 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
1950
1951 if (zero_offset) {
1952 iosize = PAGE_CACHE_SIZE - zero_offset;
1953 userpage = kmap_atomic(page, KM_USER0);
1954 memset(userpage + zero_offset, 0, iosize);
1955 flush_dcache_page(page);
1956 kunmap_atomic(userpage, KM_USER0);
1957 }
1958 }
1959 while (cur <= end) {
1960 if (cur >= last_byte) {
1961 char *userpage;
1962 struct extent_state *cached = NULL;
1963
1964 iosize = PAGE_CACHE_SIZE - pg_offset;
1965 userpage = kmap_atomic(page, KM_USER0);
1966 memset(userpage + pg_offset, 0, iosize);
1967 flush_dcache_page(page);
1968 kunmap_atomic(userpage, KM_USER0);
1969 set_extent_uptodate(tree, cur, cur + iosize - 1,
1970 &cached, GFP_NOFS);
1971 unlock_extent_cached(tree, cur, cur + iosize - 1,
1972 &cached, GFP_NOFS);
1973 break;
1974 }
1975 em = get_extent(inode, page, pg_offset, cur,
1976 end - cur + 1, 0);
1977 if (IS_ERR_OR_NULL(em)) {
1978 SetPageError(page);
1979 unlock_extent(tree, cur, end, GFP_NOFS);
1980 break;
1981 }
1982 extent_offset = cur - em->start;
1983 BUG_ON(extent_map_end(em) <= cur);
1984 BUG_ON(end < cur);
1985
1986 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
1987 this_bio_flag = EXTENT_BIO_COMPRESSED;
1988 extent_set_compress_type(&this_bio_flag,
1989 em->compress_type);
1990 }
1991
1992 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1993 cur_end = min(extent_map_end(em) - 1, end);
1994 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1995 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
1996 disk_io_size = em->block_len;
1997 sector = em->block_start >> 9;
1998 } else {
1999 sector = (em->block_start + extent_offset) >> 9;
2000 disk_io_size = iosize;
2001 }
2002 bdev = em->bdev;
2003 block_start = em->block_start;
2004 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2005 block_start = EXTENT_MAP_HOLE;
2006 free_extent_map(em);
2007 em = NULL;
2008
2009 /* we've found a hole, just zero and go on */
2010 if (block_start == EXTENT_MAP_HOLE) {
2011 char *userpage;
2012 struct extent_state *cached = NULL;
2013
2014 userpage = kmap_atomic(page, KM_USER0);
2015 memset(userpage + pg_offset, 0, iosize);
2016 flush_dcache_page(page);
2017 kunmap_atomic(userpage, KM_USER0);
2018
2019 set_extent_uptodate(tree, cur, cur + iosize - 1,
2020 &cached, GFP_NOFS);
2021 unlock_extent_cached(tree, cur, cur + iosize - 1,
2022 &cached, GFP_NOFS);
2023 cur = cur + iosize;
2024 pg_offset += iosize;
2025 continue;
2026 }
2027 /* the get_extent function already copied into the page */
2028 if (test_range_bit(tree, cur, cur_end,
2029 EXTENT_UPTODATE, 1, NULL)) {
2030 check_page_uptodate(tree, page);
2031 unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2032 cur = cur + iosize;
2033 pg_offset += iosize;
2034 continue;
2035 }
2036 /* we have an inline extent but it didn't get marked up
2037 * to date. Error out
2038 */
2039 if (block_start == EXTENT_MAP_INLINE) {
2040 SetPageError(page);
2041 unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2042 cur = cur + iosize;
2043 pg_offset += iosize;
2044 continue;
2045 }
2046
2047 ret = 0;
2048 if (tree->ops && tree->ops->readpage_io_hook) {
2049 ret = tree->ops->readpage_io_hook(page, cur,
2050 cur + iosize - 1);
2051 }
2052 if (!ret) {
2053 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2054 pnr -= page->index;
2055 ret = submit_extent_page(READ, tree, page,
2056 sector, disk_io_size, pg_offset,
2057 bdev, bio, pnr,
2058 end_bio_extent_readpage, mirror_num,
2059 *bio_flags,
2060 this_bio_flag);
2061 nr++;
2062 *bio_flags = this_bio_flag;
2063 }
2064 if (ret)
2065 SetPageError(page);
2066 cur = cur + iosize;
2067 pg_offset += iosize;
2068 }
2069out:
2070 if (!nr) {
2071 if (!PageError(page))
2072 SetPageUptodate(page);
2073 unlock_page(page);
2074 }
2075 return 0;
2076}
2077
2078int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2079 get_extent_t *get_extent)
2080{
2081 struct bio *bio = NULL;
2082 unsigned long bio_flags = 0;
2083 int ret;
2084
2085 ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2086 &bio_flags);
2087 if (bio)
2088 ret = submit_one_bio(READ, bio, 0, bio_flags);
2089 return ret;
2090}
2091
2092static noinline void update_nr_written(struct page *page,
2093 struct writeback_control *wbc,
2094 unsigned long nr_written)
2095{
2096 wbc->nr_to_write -= nr_written;
2097 if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2098 wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2099 page->mapping->writeback_index = page->index + nr_written;
2100}
2101
2102/*
2103 * the writepage semantics are similar to regular writepage. extent
2104 * records are inserted to lock ranges in the tree, and as dirty areas
2105 * are found, they are marked writeback. Then the lock bits are removed
2106 * and the end_io handler clears the writeback ranges
2107 */
2108static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2109 void *data)
2110{
2111 struct inode *inode = page->mapping->host;
2112 struct extent_page_data *epd = data;
2113 struct extent_io_tree *tree = epd->tree;
2114 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2115 u64 delalloc_start;
2116 u64 page_end = start + PAGE_CACHE_SIZE - 1;
2117 u64 end;
2118 u64 cur = start;
2119 u64 extent_offset;
2120 u64 last_byte = i_size_read(inode);
2121 u64 block_start;
2122 u64 iosize;
2123 sector_t sector;
2124 struct extent_state *cached_state = NULL;
2125 struct extent_map *em;
2126 struct block_device *bdev;
2127 int ret;
2128 int nr = 0;
2129 size_t pg_offset = 0;
2130 size_t blocksize;
2131 loff_t i_size = i_size_read(inode);
2132 unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2133 u64 nr_delalloc;
2134 u64 delalloc_end;
2135 int page_started;
2136 int compressed;
2137 int write_flags;
2138 unsigned long nr_written = 0;
2139
2140 if (wbc->sync_mode == WB_SYNC_ALL)
2141 write_flags = WRITE_SYNC;
2142 else
2143 write_flags = WRITE;
2144
2145 trace___extent_writepage(page, inode, wbc);
2146
2147 WARN_ON(!PageLocked(page));
2148 pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2149 if (page->index > end_index ||
2150 (page->index == end_index && !pg_offset)) {
2151 page->mapping->a_ops->invalidatepage(page, 0);
2152 unlock_page(page);
2153 return 0;
2154 }
2155
2156 if (page->index == end_index) {
2157 char *userpage;
2158
2159 userpage = kmap_atomic(page, KM_USER0);
2160 memset(userpage + pg_offset, 0,
2161 PAGE_CACHE_SIZE - pg_offset);
2162 kunmap_atomic(userpage, KM_USER0);
2163 flush_dcache_page(page);
2164 }
2165 pg_offset = 0;
2166
2167 set_page_extent_mapped(page);
2168
2169 delalloc_start = start;
2170 delalloc_end = 0;
2171 page_started = 0;
2172 if (!epd->extent_locked) {
2173 u64 delalloc_to_write = 0;
2174 /*
2175 * make sure the wbc mapping index is at least updated
2176 * to this page.
2177 */
2178 update_nr_written(page, wbc, 0);
2179
2180 while (delalloc_end < page_end) {
2181 nr_delalloc = find_lock_delalloc_range(inode, tree,
2182 page,
2183 &delalloc_start,
2184 &delalloc_end,
2185 128 * 1024 * 1024);
2186 if (nr_delalloc == 0) {
2187 delalloc_start = delalloc_end + 1;
2188 continue;
2189 }
2190 tree->ops->fill_delalloc(inode, page, delalloc_start,
2191 delalloc_end, &page_started,
2192 &nr_written);
2193 /*
2194 * delalloc_end is already one less than the total
2195 * length, so we don't subtract one from
2196 * PAGE_CACHE_SIZE
2197 */
2198 delalloc_to_write += (delalloc_end - delalloc_start +
2199 PAGE_CACHE_SIZE) >>
2200 PAGE_CACHE_SHIFT;
2201 delalloc_start = delalloc_end + 1;
2202 }
2203 if (wbc->nr_to_write < delalloc_to_write) {
2204 int thresh = 8192;
2205
2206 if (delalloc_to_write < thresh * 2)
2207 thresh = delalloc_to_write;
2208 wbc->nr_to_write = min_t(u64, delalloc_to_write,
2209 thresh);
2210 }
2211
2212 /* did the fill delalloc function already unlock and start
2213 * the IO?
2214 */
2215 if (page_started) {
2216 ret = 0;
2217 /*
2218 * we've unlocked the page, so we can't update
2219 * the mapping's writeback index, just update
2220 * nr_to_write.
2221 */
2222 wbc->nr_to_write -= nr_written;
2223 goto done_unlocked;
2224 }
2225 }
2226 if (tree->ops && tree->ops->writepage_start_hook) {
2227 ret = tree->ops->writepage_start_hook(page, start,
2228 page_end);
2229 if (ret == -EAGAIN) {
2230 redirty_page_for_writepage(wbc, page);
2231 update_nr_written(page, wbc, nr_written);
2232 unlock_page(page);
2233 ret = 0;
2234 goto done_unlocked;
2235 }
2236 }
2237
2238 /*
2239 * we don't want to touch the inode after unlocking the page,
2240 * so we update the mapping writeback index now
2241 */
2242 update_nr_written(page, wbc, nr_written + 1);
2243
2244 end = page_end;
2245 if (last_byte <= start) {
2246 if (tree->ops && tree->ops->writepage_end_io_hook)
2247 tree->ops->writepage_end_io_hook(page, start,
2248 page_end, NULL, 1);
2249 goto done;
2250 }
2251
2252 blocksize = inode->i_sb->s_blocksize;
2253
2254 while (cur <= end) {
2255 if (cur >= last_byte) {
2256 if (tree->ops && tree->ops->writepage_end_io_hook)
2257 tree->ops->writepage_end_io_hook(page, cur,
2258 page_end, NULL, 1);
2259 break;
2260 }
2261 em = epd->get_extent(inode, page, pg_offset, cur,
2262 end - cur + 1, 1);
2263 if (IS_ERR_OR_NULL(em)) {
2264 SetPageError(page);
2265 break;
2266 }
2267
2268 extent_offset = cur - em->start;
2269 BUG_ON(extent_map_end(em) <= cur);
2270 BUG_ON(end < cur);
2271 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2272 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2273 sector = (em->block_start + extent_offset) >> 9;
2274 bdev = em->bdev;
2275 block_start = em->block_start;
2276 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2277 free_extent_map(em);
2278 em = NULL;
2279
2280 /*
2281 * compressed and inline extents are written through other
2282 * paths in the FS
2283 */
2284 if (compressed || block_start == EXTENT_MAP_HOLE ||
2285 block_start == EXTENT_MAP_INLINE) {
2286 /*
2287 * end_io notification does not happen here for
2288 * compressed extents
2289 */
2290 if (!compressed && tree->ops &&
2291 tree->ops->writepage_end_io_hook)
2292 tree->ops->writepage_end_io_hook(page, cur,
2293 cur + iosize - 1,
2294 NULL, 1);
2295 else if (compressed) {
2296 /* we don't want to end_page_writeback on
2297 * a compressed extent. this happens
2298 * elsewhere
2299 */
2300 nr++;
2301 }
2302
2303 cur += iosize;
2304 pg_offset += iosize;
2305 continue;
2306 }
2307 /* leave this out until we have a page_mkwrite call */
2308 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2309 EXTENT_DIRTY, 0, NULL)) {
2310 cur = cur + iosize;
2311 pg_offset += iosize;
2312 continue;
2313 }
2314
2315 if (tree->ops && tree->ops->writepage_io_hook) {
2316 ret = tree->ops->writepage_io_hook(page, cur,
2317 cur + iosize - 1);
2318 } else {
2319 ret = 0;
2320 }
2321 if (ret) {
2322 SetPageError(page);
2323 } else {
2324 unsigned long max_nr = end_index + 1;
2325
2326 set_range_writeback(tree, cur, cur + iosize - 1);
2327 if (!PageWriteback(page)) {
2328 printk(KERN_ERR "btrfs warning page %lu not "
2329 "writeback, cur %llu end %llu\n",
2330 page->index, (unsigned long long)cur,
2331 (unsigned long long)end);
2332 }
2333
2334 ret = submit_extent_page(write_flags, tree, page,
2335 sector, iosize, pg_offset,
2336 bdev, &epd->bio, max_nr,
2337 end_bio_extent_writepage,
2338 0, 0, 0);
2339 if (ret)
2340 SetPageError(page);
2341 }
2342 cur = cur + iosize;
2343 pg_offset += iosize;
2344 nr++;
2345 }
2346done:
2347 if (nr == 0) {
2348 /* make sure the mapping tag for page dirty gets cleared */
2349 set_page_writeback(page);
2350 end_page_writeback(page);
2351 }
2352 unlock_page(page);
2353
2354done_unlocked:
2355
2356 /* drop our reference on any cached states */
2357 free_extent_state(cached_state);
2358 return 0;
2359}
2360
2361/**
2362 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2363 * @mapping: address space structure to write
2364 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2365 * @writepage: function called for each page
2366 * @data: data passed to writepage function
2367 *
2368 * If a page is already under I/O, write_cache_pages() skips it, even
2369 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2370 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2371 * and msync() need to guarantee that all the data which was dirty at the time
2372 * the call was made get new I/O started against them. If wbc->sync_mode is
2373 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2374 * existing IO to complete.
2375 */
2376static int extent_write_cache_pages(struct extent_io_tree *tree,
2377 struct address_space *mapping,
2378 struct writeback_control *wbc,
2379 writepage_t writepage, void *data,
2380 void (*flush_fn)(void *))
2381{
2382 int ret = 0;
2383 int done = 0;
2384 int nr_to_write_done = 0;
2385 struct pagevec pvec;
2386 int nr_pages;
2387 pgoff_t index;
2388 pgoff_t end; /* Inclusive */
2389 int scanned = 0;
2390 int tag;
2391
2392 pagevec_init(&pvec, 0);
2393 if (wbc->range_cyclic) {
2394 index = mapping->writeback_index; /* Start from prev offset */
2395 end = -1;
2396 } else {
2397 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2398 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2399 scanned = 1;
2400 }
2401 if (wbc->sync_mode == WB_SYNC_ALL)
2402 tag = PAGECACHE_TAG_TOWRITE;
2403 else
2404 tag = PAGECACHE_TAG_DIRTY;
2405retry:
2406 if (wbc->sync_mode == WB_SYNC_ALL)
2407 tag_pages_for_writeback(mapping, index, end);
2408 while (!done && !nr_to_write_done && (index <= end) &&
2409 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2410 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2411 unsigned i;
2412
2413 scanned = 1;
2414 for (i = 0; i < nr_pages; i++) {
2415 struct page *page = pvec.pages[i];
2416
2417 /*
2418 * At this point we hold neither mapping->tree_lock nor
2419 * lock on the page itself: the page may be truncated or
2420 * invalidated (changing page->mapping to NULL), or even
2421 * swizzled back from swapper_space to tmpfs file
2422 * mapping
2423 */
2424 if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2425 tree->ops->write_cache_pages_lock_hook(page);
2426 else
2427 lock_page(page);
2428
2429 if (unlikely(page->mapping != mapping)) {
2430 unlock_page(page);
2431 continue;
2432 }
2433
2434 if (!wbc->range_cyclic && page->index > end) {
2435 done = 1;
2436 unlock_page(page);
2437 continue;
2438 }
2439
2440 if (wbc->sync_mode != WB_SYNC_NONE) {
2441 if (PageWriteback(page))
2442 flush_fn(data);
2443 wait_on_page_writeback(page);
2444 }
2445
2446 if (PageWriteback(page) ||
2447 !clear_page_dirty_for_io(page)) {
2448 unlock_page(page);
2449 continue;
2450 }
2451
2452 ret = (*writepage)(page, wbc, data);
2453
2454 if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2455 unlock_page(page);
2456 ret = 0;
2457 }
2458 if (ret)
2459 done = 1;
2460
2461 /*
2462 * the filesystem may choose to bump up nr_to_write.
2463 * We have to make sure to honor the new nr_to_write
2464 * at any time
2465 */
2466 nr_to_write_done = wbc->nr_to_write <= 0;
2467 }
2468 pagevec_release(&pvec);
2469 cond_resched();
2470 }
2471 if (!scanned && !done) {
2472 /*
2473 * We hit the last page and there is more work to be done: wrap
2474 * back to the start of the file
2475 */
2476 scanned = 1;
2477 index = 0;
2478 goto retry;
2479 }
2480 return ret;
2481}
2482
2483static void flush_epd_write_bio(struct extent_page_data *epd)
2484{
2485 if (epd->bio) {
2486 if (epd->sync_io)
2487 submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2488 else
2489 submit_one_bio(WRITE, epd->bio, 0, 0);
2490 epd->bio = NULL;
2491 }
2492}
2493
2494static noinline void flush_write_bio(void *data)
2495{
2496 struct extent_page_data *epd = data;
2497 flush_epd_write_bio(epd);
2498}
2499
2500int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2501 get_extent_t *get_extent,
2502 struct writeback_control *wbc)
2503{
2504 int ret;
2505 struct extent_page_data epd = {
2506 .bio = NULL,
2507 .tree = tree,
2508 .get_extent = get_extent,
2509 .extent_locked = 0,
2510 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2511 };
2512
2513 ret = __extent_writepage(page, wbc, &epd);
2514
2515 flush_epd_write_bio(&epd);
2516 return ret;
2517}
2518
2519int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2520 u64 start, u64 end, get_extent_t *get_extent,
2521 int mode)
2522{
2523 int ret = 0;
2524 struct address_space *mapping = inode->i_mapping;
2525 struct page *page;
2526 unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2527 PAGE_CACHE_SHIFT;
2528
2529 struct extent_page_data epd = {
2530 .bio = NULL,
2531 .tree = tree,
2532 .get_extent = get_extent,
2533 .extent_locked = 1,
2534 .sync_io = mode == WB_SYNC_ALL,
2535 };
2536 struct writeback_control wbc_writepages = {
2537 .sync_mode = mode,
2538 .nr_to_write = nr_pages * 2,
2539 .range_start = start,
2540 .range_end = end + 1,
2541 };
2542
2543 while (start <= end) {
2544 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2545 if (clear_page_dirty_for_io(page))
2546 ret = __extent_writepage(page, &wbc_writepages, &epd);
2547 else {
2548 if (tree->ops && tree->ops->writepage_end_io_hook)
2549 tree->ops->writepage_end_io_hook(page, start,
2550 start + PAGE_CACHE_SIZE - 1,
2551 NULL, 1);
2552 unlock_page(page);
2553 }
2554 page_cache_release(page);
2555 start += PAGE_CACHE_SIZE;
2556 }
2557
2558 flush_epd_write_bio(&epd);
2559 return ret;
2560}
2561
2562int extent_writepages(struct extent_io_tree *tree,
2563 struct address_space *mapping,
2564 get_extent_t *get_extent,
2565 struct writeback_control *wbc)
2566{
2567 int ret = 0;
2568 struct extent_page_data epd = {
2569 .bio = NULL,
2570 .tree = tree,
2571 .get_extent = get_extent,
2572 .extent_locked = 0,
2573 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2574 };
2575
2576 ret = extent_write_cache_pages(tree, mapping, wbc,
2577 __extent_writepage, &epd,
2578 flush_write_bio);
2579 flush_epd_write_bio(&epd);
2580 return ret;
2581}
2582
2583int extent_readpages(struct extent_io_tree *tree,
2584 struct address_space *mapping,
2585 struct list_head *pages, unsigned nr_pages,
2586 get_extent_t get_extent)
2587{
2588 struct bio *bio = NULL;
2589 unsigned page_idx;
2590 unsigned long bio_flags = 0;
2591
2592 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2593 struct page *page = list_entry(pages->prev, struct page, lru);
2594
2595 prefetchw(&page->flags);
2596 list_del(&page->lru);
2597 if (!add_to_page_cache_lru(page, mapping,
2598 page->index, GFP_NOFS)) {
2599 __extent_read_full_page(tree, page, get_extent,
2600 &bio, 0, &bio_flags);
2601 }
2602 page_cache_release(page);
2603 }
2604 BUG_ON(!list_empty(pages));
2605 if (bio)
2606 submit_one_bio(READ, bio, 0, bio_flags);
2607 return 0;
2608}
2609
2610/*
2611 * basic invalidatepage code, this waits on any locked or writeback
2612 * ranges corresponding to the page, and then deletes any extent state
2613 * records from the tree
2614 */
2615int extent_invalidatepage(struct extent_io_tree *tree,
2616 struct page *page, unsigned long offset)
2617{
2618 struct extent_state *cached_state = NULL;
2619 u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2620 u64 end = start + PAGE_CACHE_SIZE - 1;
2621 size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2622
2623 start += (offset + blocksize - 1) & ~(blocksize - 1);
2624 if (start > end)
2625 return 0;
2626
2627 lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
2628 wait_on_page_writeback(page);
2629 clear_extent_bit(tree, start, end,
2630 EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2631 EXTENT_DO_ACCOUNTING,
2632 1, 1, &cached_state, GFP_NOFS);
2633 return 0;
2634}
2635
2636/*
2637 * a helper for releasepage, this tests for areas of the page that
2638 * are locked or under IO and drops the related state bits if it is safe
2639 * to drop the page.
2640 */
2641int try_release_extent_state(struct extent_map_tree *map,
2642 struct extent_io_tree *tree, struct page *page,
2643 gfp_t mask)
2644{
2645 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2646 u64 end = start + PAGE_CACHE_SIZE - 1;
2647 int ret = 1;
2648
2649 if (test_range_bit(tree, start, end,
2650 EXTENT_IOBITS, 0, NULL))
2651 ret = 0;
2652 else {
2653 if ((mask & GFP_NOFS) == GFP_NOFS)
2654 mask = GFP_NOFS;
2655 /*
2656 * at this point we can safely clear everything except the
2657 * locked bit and the nodatasum bit
2658 */
2659 ret = clear_extent_bit(tree, start, end,
2660 ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2661 0, 0, NULL, mask);
2662
2663 /* if clear_extent_bit failed for enomem reasons,
2664 * we can't allow the release to continue.
2665 */
2666 if (ret < 0)
2667 ret = 0;
2668 else
2669 ret = 1;
2670 }
2671 return ret;
2672}
2673
2674/*
2675 * a helper for releasepage. As long as there are no locked extents
2676 * in the range corresponding to the page, both state records and extent
2677 * map records are removed
2678 */
2679int try_release_extent_mapping(struct extent_map_tree *map,
2680 struct extent_io_tree *tree, struct page *page,
2681 gfp_t mask)
2682{
2683 struct extent_map *em;
2684 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2685 u64 end = start + PAGE_CACHE_SIZE - 1;
2686
2687 if ((mask & __GFP_WAIT) &&
2688 page->mapping->host->i_size > 16 * 1024 * 1024) {
2689 u64 len;
2690 while (start <= end) {
2691 len = end - start + 1;
2692 write_lock(&map->lock);
2693 em = lookup_extent_mapping(map, start, len);
2694 if (IS_ERR_OR_NULL(em)) {
2695 write_unlock(&map->lock);
2696 break;
2697 }
2698 if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2699 em->start != start) {
2700 write_unlock(&map->lock);
2701 free_extent_map(em);
2702 break;
2703 }
2704 if (!test_range_bit(tree, em->start,
2705 extent_map_end(em) - 1,
2706 EXTENT_LOCKED | EXTENT_WRITEBACK,
2707 0, NULL)) {
2708 remove_extent_mapping(map, em);
2709 /* once for the rb tree */
2710 free_extent_map(em);
2711 }
2712 start = extent_map_end(em);
2713 write_unlock(&map->lock);
2714
2715 /* once for us */
2716 free_extent_map(em);
2717 }
2718 }
2719 return try_release_extent_state(map, tree, page, mask);
2720}
2721
2722/*
2723 * helper function for fiemap, which doesn't want to see any holes.
2724 * This maps until we find something past 'last'
2725 */
2726static struct extent_map *get_extent_skip_holes(struct inode *inode,
2727 u64 offset,
2728 u64 last,
2729 get_extent_t *get_extent)
2730{
2731 u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
2732 struct extent_map *em;
2733 u64 len;
2734
2735 if (offset >= last)
2736 return NULL;
2737
2738 while(1) {
2739 len = last - offset;
2740 if (len == 0)
2741 break;
2742 len = (len + sectorsize - 1) & ~(sectorsize - 1);
2743 em = get_extent(inode, NULL, 0, offset, len, 0);
2744 if (IS_ERR_OR_NULL(em))
2745 return em;
2746
2747 /* if this isn't a hole return it */
2748 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
2749 em->block_start != EXTENT_MAP_HOLE) {
2750 return em;
2751 }
2752
2753 /* this is a hole, advance to the next extent */
2754 offset = extent_map_end(em);
2755 free_extent_map(em);
2756 if (offset >= last)
2757 break;
2758 }
2759 return NULL;
2760}
2761
2762int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2763 __u64 start, __u64 len, get_extent_t *get_extent)
2764{
2765 int ret = 0;
2766 u64 off = start;
2767 u64 max = start + len;
2768 u32 flags = 0;
2769 u32 found_type;
2770 u64 last;
2771 u64 last_for_get_extent = 0;
2772 u64 disko = 0;
2773 u64 isize = i_size_read(inode);
2774 struct btrfs_key found_key;
2775 struct extent_map *em = NULL;
2776 struct extent_state *cached_state = NULL;
2777 struct btrfs_path *path;
2778 struct btrfs_file_extent_item *item;
2779 int end = 0;
2780 u64 em_start = 0;
2781 u64 em_len = 0;
2782 u64 em_end = 0;
2783 unsigned long emflags;
2784
2785 if (len == 0)
2786 return -EINVAL;
2787
2788 path = btrfs_alloc_path();
2789 if (!path)
2790 return -ENOMEM;
2791 path->leave_spinning = 1;
2792
2793 /*
2794 * lookup the last file extent. We're not using i_size here
2795 * because there might be preallocation past i_size
2796 */
2797 ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
2798 path, btrfs_ino(inode), -1, 0);
2799 if (ret < 0) {
2800 btrfs_free_path(path);
2801 return ret;
2802 }
2803 WARN_ON(!ret);
2804 path->slots[0]--;
2805 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2806 struct btrfs_file_extent_item);
2807 btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
2808 found_type = btrfs_key_type(&found_key);
2809
2810 /* No extents, but there might be delalloc bits */
2811 if (found_key.objectid != btrfs_ino(inode) ||
2812 found_type != BTRFS_EXTENT_DATA_KEY) {
2813 /* have to trust i_size as the end */
2814 last = (u64)-1;
2815 last_for_get_extent = isize;
2816 } else {
2817 /*
2818 * remember the start of the last extent. There are a
2819 * bunch of different factors that go into the length of the
2820 * extent, so its much less complex to remember where it started
2821 */
2822 last = found_key.offset;
2823 last_for_get_extent = last + 1;
2824 }
2825 btrfs_free_path(path);
2826
2827 /*
2828 * we might have some extents allocated but more delalloc past those
2829 * extents. so, we trust isize unless the start of the last extent is
2830 * beyond isize
2831 */
2832 if (last < isize) {
2833 last = (u64)-1;
2834 last_for_get_extent = isize;
2835 }
2836
2837 lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
2838 &cached_state, GFP_NOFS);
2839
2840 em = get_extent_skip_holes(inode, off, last_for_get_extent,
2841 get_extent);
2842 if (!em)
2843 goto out;
2844 if (IS_ERR(em)) {
2845 ret = PTR_ERR(em);
2846 goto out;
2847 }
2848
2849 while (!end) {
2850 u64 offset_in_extent;
2851
2852 /* break if the extent we found is outside the range */
2853 if (em->start >= max || extent_map_end(em) < off)
2854 break;
2855
2856 /*
2857 * get_extent may return an extent that starts before our
2858 * requested range. We have to make sure the ranges
2859 * we return to fiemap always move forward and don't
2860 * overlap, so adjust the offsets here
2861 */
2862 em_start = max(em->start, off);
2863
2864 /*
2865 * record the offset from the start of the extent
2866 * for adjusting the disk offset below
2867 */
2868 offset_in_extent = em_start - em->start;
2869 em_end = extent_map_end(em);
2870 em_len = em_end - em_start;
2871 emflags = em->flags;
2872 disko = 0;
2873 flags = 0;
2874
2875 /*
2876 * bump off for our next call to get_extent
2877 */
2878 off = extent_map_end(em);
2879 if (off >= max)
2880 end = 1;
2881
2882 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
2883 end = 1;
2884 flags |= FIEMAP_EXTENT_LAST;
2885 } else if (em->block_start == EXTENT_MAP_INLINE) {
2886 flags |= (FIEMAP_EXTENT_DATA_INLINE |
2887 FIEMAP_EXTENT_NOT_ALIGNED);
2888 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
2889 flags |= (FIEMAP_EXTENT_DELALLOC |
2890 FIEMAP_EXTENT_UNKNOWN);
2891 } else {
2892 disko = em->block_start + offset_in_extent;
2893 }
2894 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2895 flags |= FIEMAP_EXTENT_ENCODED;
2896
2897 free_extent_map(em);
2898 em = NULL;
2899 if ((em_start >= last) || em_len == (u64)-1 ||
2900 (last == (u64)-1 && isize <= em_end)) {
2901 flags |= FIEMAP_EXTENT_LAST;
2902 end = 1;
2903 }
2904
2905 /* now scan forward to see if this is really the last extent. */
2906 em = get_extent_skip_holes(inode, off, last_for_get_extent,
2907 get_extent);
2908 if (IS_ERR(em)) {
2909 ret = PTR_ERR(em);
2910 goto out;
2911 }
2912 if (!em) {
2913 flags |= FIEMAP_EXTENT_LAST;
2914 end = 1;
2915 }
2916 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
2917 em_len, flags);
2918 if (ret)
2919 goto out_free;
2920 }
2921out_free:
2922 free_extent_map(em);
2923out:
2924 unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
2925 &cached_state, GFP_NOFS);
2926 return ret;
2927}
2928
2929static inline struct page *extent_buffer_page(struct extent_buffer *eb,
2930 unsigned long i)
2931{
2932 struct page *p;
2933 struct address_space *mapping;
2934
2935 if (i == 0)
2936 return eb->first_page;
2937 i += eb->start >> PAGE_CACHE_SHIFT;
2938 mapping = eb->first_page->mapping;
2939 if (!mapping)
2940 return NULL;
2941
2942 /*
2943 * extent_buffer_page is only called after pinning the page
2944 * by increasing the reference count. So we know the page must
2945 * be in the radix tree.
2946 */
2947 rcu_read_lock();
2948 p = radix_tree_lookup(&mapping->page_tree, i);
2949 rcu_read_unlock();
2950
2951 return p;
2952}
2953
2954static inline unsigned long num_extent_pages(u64 start, u64 len)
2955{
2956 return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
2957 (start >> PAGE_CACHE_SHIFT);
2958}
2959
2960static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
2961 u64 start,
2962 unsigned long len,
2963 gfp_t mask)
2964{
2965 struct extent_buffer *eb = NULL;
2966#if LEAK_DEBUG
2967 unsigned long flags;
2968#endif
2969
2970 eb = kmem_cache_zalloc(extent_buffer_cache, mask);
2971 if (eb == NULL)
2972 return NULL;
2973 eb->start = start;
2974 eb->len = len;
2975 rwlock_init(&eb->lock);
2976 atomic_set(&eb->write_locks, 0);
2977 atomic_set(&eb->read_locks, 0);
2978 atomic_set(&eb->blocking_readers, 0);
2979 atomic_set(&eb->blocking_writers, 0);
2980 atomic_set(&eb->spinning_readers, 0);
2981 atomic_set(&eb->spinning_writers, 0);
2982 init_waitqueue_head(&eb->write_lock_wq);
2983 init_waitqueue_head(&eb->read_lock_wq);
2984
2985#if LEAK_DEBUG
2986 spin_lock_irqsave(&leak_lock, flags);
2987 list_add(&eb->leak_list, &buffers);
2988 spin_unlock_irqrestore(&leak_lock, flags);
2989#endif
2990 atomic_set(&eb->refs, 1);
2991
2992 return eb;
2993}
2994
2995static void __free_extent_buffer(struct extent_buffer *eb)
2996{
2997#if LEAK_DEBUG
2998 unsigned long flags;
2999 spin_lock_irqsave(&leak_lock, flags);
3000 list_del(&eb->leak_list);
3001 spin_unlock_irqrestore(&leak_lock, flags);
3002#endif
3003 kmem_cache_free(extent_buffer_cache, eb);
3004}
3005
3006/*
3007 * Helper for releasing extent buffer page.
3008 */
3009static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3010 unsigned long start_idx)
3011{
3012 unsigned long index;
3013 struct page *page;
3014
3015 if (!eb->first_page)
3016 return;
3017
3018 index = num_extent_pages(eb->start, eb->len);
3019 if (start_idx >= index)
3020 return;
3021
3022 do {
3023 index--;
3024 page = extent_buffer_page(eb, index);
3025 if (page)
3026 page_cache_release(page);
3027 } while (index != start_idx);
3028}
3029
3030/*
3031 * Helper for releasing the extent buffer.
3032 */
3033static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3034{
3035 btrfs_release_extent_buffer_page(eb, 0);
3036 __free_extent_buffer(eb);
3037}
3038
3039struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3040 u64 start, unsigned long len,
3041 struct page *page0)
3042{
3043 unsigned long num_pages = num_extent_pages(start, len);
3044 unsigned long i;
3045 unsigned long index = start >> PAGE_CACHE_SHIFT;
3046 struct extent_buffer *eb;
3047 struct extent_buffer *exists = NULL;
3048 struct page *p;
3049 struct address_space *mapping = tree->mapping;
3050 int uptodate = 1;
3051 int ret;
3052
3053 rcu_read_lock();
3054 eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3055 if (eb && atomic_inc_not_zero(&eb->refs)) {
3056 rcu_read_unlock();
3057 mark_page_accessed(eb->first_page);
3058 return eb;
3059 }
3060 rcu_read_unlock();
3061
3062 eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
3063 if (!eb)
3064 return NULL;
3065
3066 if (page0) {
3067 eb->first_page = page0;
3068 i = 1;
3069 index++;
3070 page_cache_get(page0);
3071 mark_page_accessed(page0);
3072 set_page_extent_mapped(page0);
3073 set_page_extent_head(page0, len);
3074 uptodate = PageUptodate(page0);
3075 } else {
3076 i = 0;
3077 }
3078 for (; i < num_pages; i++, index++) {
3079 p = find_or_create_page(mapping, index, GFP_NOFS);
3080 if (!p) {
3081 WARN_ON(1);
3082 goto free_eb;
3083 }
3084 set_page_extent_mapped(p);
3085 mark_page_accessed(p);
3086 if (i == 0) {
3087 eb->first_page = p;
3088 set_page_extent_head(p, len);
3089 } else {
3090 set_page_private(p, EXTENT_PAGE_PRIVATE);
3091 }
3092 if (!PageUptodate(p))
3093 uptodate = 0;
3094
3095 /*
3096 * see below about how we avoid a nasty race with release page
3097 * and why we unlock later
3098 */
3099 if (i != 0)
3100 unlock_page(p);
3101 }
3102 if (uptodate)
3103 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3104
3105 ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
3106 if (ret)
3107 goto free_eb;
3108
3109 spin_lock(&tree->buffer_lock);
3110 ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
3111 if (ret == -EEXIST) {
3112 exists = radix_tree_lookup(&tree->buffer,
3113 start >> PAGE_CACHE_SHIFT);
3114 /* add one reference for the caller */
3115 atomic_inc(&exists->refs);
3116 spin_unlock(&tree->buffer_lock);
3117 radix_tree_preload_end();
3118 goto free_eb;
3119 }
3120 /* add one reference for the tree */
3121 atomic_inc(&eb->refs);
3122 spin_unlock(&tree->buffer_lock);
3123 radix_tree_preload_end();
3124
3125 /*
3126 * there is a race where release page may have
3127 * tried to find this extent buffer in the radix
3128 * but failed. It will tell the VM it is safe to
3129 * reclaim the, and it will clear the page private bit.
3130 * We must make sure to set the page private bit properly
3131 * after the extent buffer is in the radix tree so
3132 * it doesn't get lost
3133 */
3134 set_page_extent_mapped(eb->first_page);
3135 set_page_extent_head(eb->first_page, eb->len);
3136 if (!page0)
3137 unlock_page(eb->first_page);
3138 return eb;
3139
3140free_eb:
3141 if (eb->first_page && !page0)
3142 unlock_page(eb->first_page);
3143
3144 if (!atomic_dec_and_test(&eb->refs))
3145 return exists;
3146 btrfs_release_extent_buffer(eb);
3147 return exists;
3148}
3149
3150struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3151 u64 start, unsigned long len)
3152{
3153 struct extent_buffer *eb;
3154
3155 rcu_read_lock();
3156 eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3157 if (eb && atomic_inc_not_zero(&eb->refs)) {
3158 rcu_read_unlock();
3159 mark_page_accessed(eb->first_page);
3160 return eb;
3161 }
3162 rcu_read_unlock();
3163
3164 return NULL;
3165}
3166
3167void free_extent_buffer(struct extent_buffer *eb)
3168{
3169 if (!eb)
3170 return;
3171
3172 if (!atomic_dec_and_test(&eb->refs))
3173 return;
3174
3175 WARN_ON(1);
3176}
3177
3178int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3179 struct extent_buffer *eb)
3180{
3181 unsigned long i;
3182 unsigned long num_pages;
3183 struct page *page;
3184
3185 num_pages = num_extent_pages(eb->start, eb->len);
3186
3187 for (i = 0; i < num_pages; i++) {
3188 page = extent_buffer_page(eb, i);
3189 if (!PageDirty(page))
3190 continue;
3191
3192 lock_page(page);
3193 WARN_ON(!PagePrivate(page));
3194
3195 set_page_extent_mapped(page);
3196 if (i == 0)
3197 set_page_extent_head(page, eb->len);
3198
3199 clear_page_dirty_for_io(page);
3200 spin_lock_irq(&page->mapping->tree_lock);
3201 if (!PageDirty(page)) {
3202 radix_tree_tag_clear(&page->mapping->page_tree,
3203 page_index(page),
3204 PAGECACHE_TAG_DIRTY);
3205 }
3206 spin_unlock_irq(&page->mapping->tree_lock);
3207 unlock_page(page);
3208 }
3209 return 0;
3210}
3211
3212int set_extent_buffer_dirty(struct extent_io_tree *tree,
3213 struct extent_buffer *eb)
3214{
3215 unsigned long i;
3216 unsigned long num_pages;
3217 int was_dirty = 0;
3218
3219 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3220 num_pages = num_extent_pages(eb->start, eb->len);
3221 for (i = 0; i < num_pages; i++)
3222 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3223 return was_dirty;
3224}
3225
3226static int __eb_straddles_pages(u64 start, u64 len)
3227{
3228 if (len < PAGE_CACHE_SIZE)
3229 return 1;
3230 if (start & (PAGE_CACHE_SIZE - 1))
3231 return 1;
3232 if ((start + len) & (PAGE_CACHE_SIZE - 1))
3233 return 1;
3234 return 0;
3235}
3236
3237static int eb_straddles_pages(struct extent_buffer *eb)
3238{
3239 return __eb_straddles_pages(eb->start, eb->len);
3240}
3241
3242int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3243 struct extent_buffer *eb,
3244 struct extent_state **cached_state)
3245{
3246 unsigned long i;
3247 struct page *page;
3248 unsigned long num_pages;
3249
3250 num_pages = num_extent_pages(eb->start, eb->len);
3251 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3252
3253 if (eb_straddles_pages(eb)) {
3254 clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3255 cached_state, GFP_NOFS);
3256 }
3257 for (i = 0; i < num_pages; i++) {
3258 page = extent_buffer_page(eb, i);
3259 if (page)
3260 ClearPageUptodate(page);
3261 }
3262 return 0;
3263}
3264
3265int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3266 struct extent_buffer *eb)
3267{
3268 unsigned long i;
3269 struct page *page;
3270 unsigned long num_pages;
3271
3272 num_pages = num_extent_pages(eb->start, eb->len);
3273
3274 if (eb_straddles_pages(eb)) {
3275 set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3276 NULL, GFP_NOFS);
3277 }
3278 for (i = 0; i < num_pages; i++) {
3279 page = extent_buffer_page(eb, i);
3280 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3281 ((i == num_pages - 1) &&
3282 ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3283 check_page_uptodate(tree, page);
3284 continue;
3285 }
3286 SetPageUptodate(page);
3287 }
3288 return 0;
3289}
3290
3291int extent_range_uptodate(struct extent_io_tree *tree,
3292 u64 start, u64 end)
3293{
3294 struct page *page;
3295 int ret;
3296 int pg_uptodate = 1;
3297 int uptodate;
3298 unsigned long index;
3299
3300 if (__eb_straddles_pages(start, end - start + 1)) {
3301 ret = test_range_bit(tree, start, end,
3302 EXTENT_UPTODATE, 1, NULL);
3303 if (ret)
3304 return 1;
3305 }
3306 while (start <= end) {
3307 index = start >> PAGE_CACHE_SHIFT;
3308 page = find_get_page(tree->mapping, index);
3309 uptodate = PageUptodate(page);
3310 page_cache_release(page);
3311 if (!uptodate) {
3312 pg_uptodate = 0;
3313 break;
3314 }
3315 start += PAGE_CACHE_SIZE;
3316 }
3317 return pg_uptodate;
3318}
3319
3320int extent_buffer_uptodate(struct extent_io_tree *tree,
3321 struct extent_buffer *eb,
3322 struct extent_state *cached_state)
3323{
3324 int ret = 0;
3325 unsigned long num_pages;
3326 unsigned long i;
3327 struct page *page;
3328 int pg_uptodate = 1;
3329
3330 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3331 return 1;
3332
3333 if (eb_straddles_pages(eb)) {
3334 ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3335 EXTENT_UPTODATE, 1, cached_state);
3336 if (ret)
3337 return ret;
3338 }
3339
3340 num_pages = num_extent_pages(eb->start, eb->len);
3341 for (i = 0; i < num_pages; i++) {
3342 page = extent_buffer_page(eb, i);
3343 if (!PageUptodate(page)) {
3344 pg_uptodate = 0;
3345 break;
3346 }
3347 }
3348 return pg_uptodate;
3349}
3350
3351int read_extent_buffer_pages(struct extent_io_tree *tree,
3352 struct extent_buffer *eb,
3353 u64 start, int wait,
3354 get_extent_t *get_extent, int mirror_num)
3355{
3356 unsigned long i;
3357 unsigned long start_i;
3358 struct page *page;
3359 int err;
3360 int ret = 0;
3361 int locked_pages = 0;
3362 int all_uptodate = 1;
3363 int inc_all_pages = 0;
3364 unsigned long num_pages;
3365 struct bio *bio = NULL;
3366 unsigned long bio_flags = 0;
3367
3368 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3369 return 0;
3370
3371 if (eb_straddles_pages(eb)) {
3372 if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3373 EXTENT_UPTODATE, 1, NULL)) {
3374 return 0;
3375 }
3376 }
3377
3378 if (start) {
3379 WARN_ON(start < eb->start);
3380 start_i = (start >> PAGE_CACHE_SHIFT) -
3381 (eb->start >> PAGE_CACHE_SHIFT);
3382 } else {
3383 start_i = 0;
3384 }
3385
3386 num_pages = num_extent_pages(eb->start, eb->len);
3387 for (i = start_i; i < num_pages; i++) {
3388 page = extent_buffer_page(eb, i);
3389 if (!wait) {
3390 if (!trylock_page(page))
3391 goto unlock_exit;
3392 } else {
3393 lock_page(page);
3394 }
3395 locked_pages++;
3396 if (!PageUptodate(page))
3397 all_uptodate = 0;
3398 }
3399 if (all_uptodate) {
3400 if (start_i == 0)
3401 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3402 goto unlock_exit;
3403 }
3404
3405 for (i = start_i; i < num_pages; i++) {
3406 page = extent_buffer_page(eb, i);
3407
3408 WARN_ON(!PagePrivate(page));
3409
3410 set_page_extent_mapped(page);
3411 if (i == 0)
3412 set_page_extent_head(page, eb->len);
3413
3414 if (inc_all_pages)
3415 page_cache_get(page);
3416 if (!PageUptodate(page)) {
3417 if (start_i == 0)
3418 inc_all_pages = 1;
3419 ClearPageError(page);
3420 err = __extent_read_full_page(tree, page,
3421 get_extent, &bio,
3422 mirror_num, &bio_flags);
3423 if (err)
3424 ret = err;
3425 } else {
3426 unlock_page(page);
3427 }
3428 }
3429
3430 if (bio)
3431 submit_one_bio(READ, bio, mirror_num, bio_flags);
3432
3433 if (ret || !wait)
3434 return ret;
3435
3436 for (i = start_i; i < num_pages; i++) {
3437 page = extent_buffer_page(eb, i);
3438 wait_on_page_locked(page);
3439 if (!PageUptodate(page))
3440 ret = -EIO;
3441 }
3442
3443 if (!ret)
3444 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3445 return ret;
3446
3447unlock_exit:
3448 i = start_i;
3449 while (locked_pages > 0) {
3450 page = extent_buffer_page(eb, i);
3451 i++;
3452 unlock_page(page);
3453 locked_pages--;
3454 }
3455 return ret;
3456}
3457
3458void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3459 unsigned long start,
3460 unsigned long len)
3461{
3462 size_t cur;
3463 size_t offset;
3464 struct page *page;
3465 char *kaddr;
3466 char *dst = (char *)dstv;
3467 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3468 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3469
3470 WARN_ON(start > eb->len);
3471 WARN_ON(start + len > eb->start + eb->len);
3472
3473 offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3474
3475 while (len > 0) {
3476 page = extent_buffer_page(eb, i);
3477
3478 cur = min(len, (PAGE_CACHE_SIZE - offset));
3479 kaddr = page_address(page);
3480 memcpy(dst, kaddr + offset, cur);
3481
3482 dst += cur;
3483 len -= cur;
3484 offset = 0;
3485 i++;
3486 }
3487}
3488
3489int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3490 unsigned long min_len, char **map,
3491 unsigned long *map_start,
3492 unsigned long *map_len)
3493{
3494 size_t offset = start & (PAGE_CACHE_SIZE - 1);
3495 char *kaddr;
3496 struct page *p;
3497 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3498 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3499 unsigned long end_i = (start_offset + start + min_len - 1) >>
3500 PAGE_CACHE_SHIFT;
3501
3502 if (i != end_i)
3503 return -EINVAL;
3504
3505 if (i == 0) {
3506 offset = start_offset;
3507 *map_start = 0;
3508 } else {
3509 offset = 0;
3510 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3511 }
3512
3513 if (start + min_len > eb->len) {
3514 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
3515 "wanted %lu %lu\n", (unsigned long long)eb->start,
3516 eb->len, start, min_len);
3517 WARN_ON(1);
3518 return -EINVAL;
3519 }
3520
3521 p = extent_buffer_page(eb, i);
3522 kaddr = page_address(p);
3523 *map = kaddr + offset;
3524 *map_len = PAGE_CACHE_SIZE - offset;
3525 return 0;
3526}
3527
3528int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3529 unsigned long start,
3530 unsigned long len)
3531{
3532 size_t cur;
3533 size_t offset;
3534 struct page *page;
3535 char *kaddr;
3536 char *ptr = (char *)ptrv;
3537 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3538 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3539 int ret = 0;
3540
3541 WARN_ON(start > eb->len);
3542 WARN_ON(start + len > eb->start + eb->len);
3543
3544 offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3545
3546 while (len > 0) {
3547 page = extent_buffer_page(eb, i);
3548
3549 cur = min(len, (PAGE_CACHE_SIZE - offset));
3550
3551 kaddr = page_address(page);
3552 ret = memcmp(ptr, kaddr + offset, cur);
3553 if (ret)
3554 break;
3555
3556 ptr += cur;
3557 len -= cur;
3558 offset = 0;
3559 i++;
3560 }
3561 return ret;
3562}
3563
3564void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3565 unsigned long start, unsigned long len)
3566{
3567 size_t cur;
3568 size_t offset;
3569 struct page *page;
3570 char *kaddr;
3571 char *src = (char *)srcv;
3572 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3573 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3574
3575 WARN_ON(start > eb->len);
3576 WARN_ON(start + len > eb->start + eb->len);
3577
3578 offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3579
3580 while (len > 0) {
3581 page = extent_buffer_page(eb, i);
3582 WARN_ON(!PageUptodate(page));
3583
3584 cur = min(len, PAGE_CACHE_SIZE - offset);
3585 kaddr = page_address(page);
3586 memcpy(kaddr + offset, src, cur);
3587
3588 src += cur;
3589 len -= cur;
3590 offset = 0;
3591 i++;
3592 }
3593}
3594
3595void memset_extent_buffer(struct extent_buffer *eb, char c,
3596 unsigned long start, unsigned long len)
3597{
3598 size_t cur;
3599 size_t offset;
3600 struct page *page;
3601 char *kaddr;
3602 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3603 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3604
3605 WARN_ON(start > eb->len);
3606 WARN_ON(start + len > eb->start + eb->len);
3607
3608 offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3609
3610 while (len > 0) {
3611 page = extent_buffer_page(eb, i);
3612 WARN_ON(!PageUptodate(page));
3613
3614 cur = min(len, PAGE_CACHE_SIZE - offset);
3615 kaddr = page_address(page);
3616 memset(kaddr + offset, c, cur);
3617
3618 len -= cur;
3619 offset = 0;
3620 i++;
3621 }
3622}
3623
3624void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3625 unsigned long dst_offset, unsigned long src_offset,
3626 unsigned long len)
3627{
3628 u64 dst_len = dst->len;
3629 size_t cur;
3630 size_t offset;
3631 struct page *page;
3632 char *kaddr;
3633 size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3634 unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3635
3636 WARN_ON(src->len != dst_len);
3637
3638 offset = (start_offset + dst_offset) &
3639 ((unsigned long)PAGE_CACHE_SIZE - 1);
3640
3641 while (len > 0) {
3642 page = extent_buffer_page(dst, i);
3643 WARN_ON(!PageUptodate(page));
3644
3645 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3646
3647 kaddr = page_address(page);
3648 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3649
3650 src_offset += cur;
3651 len -= cur;
3652 offset = 0;
3653 i++;
3654 }
3655}
3656
3657static void move_pages(struct page *dst_page, struct page *src_page,
3658 unsigned long dst_off, unsigned long src_off,
3659 unsigned long len)
3660{
3661 char *dst_kaddr = page_address(dst_page);
3662 if (dst_page == src_page) {
3663 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3664 } else {
3665 char *src_kaddr = page_address(src_page);
3666 char *p = dst_kaddr + dst_off + len;
3667 char *s = src_kaddr + src_off + len;
3668
3669 while (len--)
3670 *--p = *--s;
3671 }
3672}
3673
3674static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
3675{
3676 unsigned long distance = (src > dst) ? src - dst : dst - src;
3677 return distance < len;
3678}
3679
3680static void copy_pages(struct page *dst_page, struct page *src_page,
3681 unsigned long dst_off, unsigned long src_off,
3682 unsigned long len)
3683{
3684 char *dst_kaddr = page_address(dst_page);
3685 char *src_kaddr;
3686
3687 if (dst_page != src_page) {
3688 src_kaddr = page_address(src_page);
3689 } else {
3690 src_kaddr = dst_kaddr;
3691 BUG_ON(areas_overlap(src_off, dst_off, len));
3692 }
3693
3694 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3695}
3696
3697void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3698 unsigned long src_offset, unsigned long len)
3699{
3700 size_t cur;
3701 size_t dst_off_in_page;
3702 size_t src_off_in_page;
3703 size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3704 unsigned long dst_i;
3705 unsigned long src_i;
3706
3707 if (src_offset + len > dst->len) {
3708 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3709 "len %lu dst len %lu\n", src_offset, len, dst->len);
3710 BUG_ON(1);
3711 }
3712 if (dst_offset + len > dst->len) {
3713 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3714 "len %lu dst len %lu\n", dst_offset, len, dst->len);
3715 BUG_ON(1);
3716 }
3717
3718 while (len > 0) {
3719 dst_off_in_page = (start_offset + dst_offset) &
3720 ((unsigned long)PAGE_CACHE_SIZE - 1);
3721 src_off_in_page = (start_offset + src_offset) &
3722 ((unsigned long)PAGE_CACHE_SIZE - 1);
3723
3724 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3725 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3726
3727 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3728 src_off_in_page));
3729 cur = min_t(unsigned long, cur,
3730 (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3731
3732 copy_pages(extent_buffer_page(dst, dst_i),
3733 extent_buffer_page(dst, src_i),
3734 dst_off_in_page, src_off_in_page, cur);
3735
3736 src_offset += cur;
3737 dst_offset += cur;
3738 len -= cur;
3739 }
3740}
3741
3742void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3743 unsigned long src_offset, unsigned long len)
3744{
3745 size_t cur;
3746 size_t dst_off_in_page;
3747 size_t src_off_in_page;
3748 unsigned long dst_end = dst_offset + len - 1;
3749 unsigned long src_end = src_offset + len - 1;
3750 size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3751 unsigned long dst_i;
3752 unsigned long src_i;
3753
3754 if (src_offset + len > dst->len) {
3755 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3756 "len %lu len %lu\n", src_offset, len, dst->len);
3757 BUG_ON(1);
3758 }
3759 if (dst_offset + len > dst->len) {
3760 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3761 "len %lu len %lu\n", dst_offset, len, dst->len);
3762 BUG_ON(1);
3763 }
3764 if (!areas_overlap(src_offset, dst_offset, len)) {
3765 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3766 return;
3767 }
3768 while (len > 0) {
3769 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3770 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3771
3772 dst_off_in_page = (start_offset + dst_end) &
3773 ((unsigned long)PAGE_CACHE_SIZE - 1);
3774 src_off_in_page = (start_offset + src_end) &
3775 ((unsigned long)PAGE_CACHE_SIZE - 1);
3776
3777 cur = min_t(unsigned long, len, src_off_in_page + 1);
3778 cur = min(cur, dst_off_in_page + 1);
3779 move_pages(extent_buffer_page(dst, dst_i),
3780 extent_buffer_page(dst, src_i),
3781 dst_off_in_page - cur + 1,
3782 src_off_in_page - cur + 1, cur);
3783
3784 dst_end -= cur;
3785 src_end -= cur;
3786 len -= cur;
3787 }
3788}
3789
3790static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3791{
3792 struct extent_buffer *eb =
3793 container_of(head, struct extent_buffer, rcu_head);
3794
3795 btrfs_release_extent_buffer(eb);
3796}
3797
3798int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
3799{
3800 u64 start = page_offset(page);
3801 struct extent_buffer *eb;
3802 int ret = 1;
3803
3804 spin_lock(&tree->buffer_lock);
3805 eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3806 if (!eb) {
3807 spin_unlock(&tree->buffer_lock);
3808 return ret;
3809 }
3810
3811 if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3812 ret = 0;
3813 goto out;
3814 }
3815
3816 /*
3817 * set @eb->refs to 0 if it is already 1, and then release the @eb.
3818 * Or go back.
3819 */
3820 if (atomic_cmpxchg(&eb->refs, 1, 0) != 1) {
3821 ret = 0;
3822 goto out;
3823 }
3824
3825 radix_tree_delete(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3826out:
3827 spin_unlock(&tree->buffer_lock);
3828
3829 /* at this point we can safely release the extent buffer */
3830 if (atomic_read(&eb->refs) == 0)
3831 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3832 return ret;
3833}
1#include <linux/bitops.h>
2#include <linux/slab.h>
3#include <linux/bio.h>
4#include <linux/mm.h>
5#include <linux/pagemap.h>
6#include <linux/page-flags.h>
7#include <linux/spinlock.h>
8#include <linux/blkdev.h>
9#include <linux/swap.h>
10#include <linux/writeback.h>
11#include <linux/pagevec.h>
12#include <linux/prefetch.h>
13#include <linux/cleancache.h>
14#include "extent_io.h"
15#include "extent_map.h"
16#include "ctree.h"
17#include "btrfs_inode.h"
18#include "volumes.h"
19#include "check-integrity.h"
20#include "locking.h"
21#include "rcu-string.h"
22#include "backref.h"
23
24static struct kmem_cache *extent_state_cache;
25static struct kmem_cache *extent_buffer_cache;
26static struct bio_set *btrfs_bioset;
27
28#ifdef CONFIG_BTRFS_DEBUG
29static LIST_HEAD(buffers);
30static LIST_HEAD(states);
31
32static DEFINE_SPINLOCK(leak_lock);
33
34static inline
35void btrfs_leak_debug_add(struct list_head *new, struct list_head *head)
36{
37 unsigned long flags;
38
39 spin_lock_irqsave(&leak_lock, flags);
40 list_add(new, head);
41 spin_unlock_irqrestore(&leak_lock, flags);
42}
43
44static inline
45void btrfs_leak_debug_del(struct list_head *entry)
46{
47 unsigned long flags;
48
49 spin_lock_irqsave(&leak_lock, flags);
50 list_del(entry);
51 spin_unlock_irqrestore(&leak_lock, flags);
52}
53
54static inline
55void btrfs_leak_debug_check(void)
56{
57 struct extent_state *state;
58 struct extent_buffer *eb;
59
60 while (!list_empty(&states)) {
61 state = list_entry(states.next, struct extent_state, leak_list);
62 printk(KERN_ERR "BTRFS: state leak: start %llu end %llu "
63 "state %lu in tree %p refs %d\n",
64 state->start, state->end, state->state, state->tree,
65 atomic_read(&state->refs));
66 list_del(&state->leak_list);
67 kmem_cache_free(extent_state_cache, state);
68 }
69
70 while (!list_empty(&buffers)) {
71 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
72 printk(KERN_ERR "BTRFS: buffer leak start %llu len %lu "
73 "refs %d\n",
74 eb->start, eb->len, atomic_read(&eb->refs));
75 list_del(&eb->leak_list);
76 kmem_cache_free(extent_buffer_cache, eb);
77 }
78}
79
80#define btrfs_debug_check_extent_io_range(tree, start, end) \
81 __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
82static inline void __btrfs_debug_check_extent_io_range(const char *caller,
83 struct extent_io_tree *tree, u64 start, u64 end)
84{
85 struct inode *inode;
86 u64 isize;
87
88 if (!tree->mapping)
89 return;
90
91 inode = tree->mapping->host;
92 isize = i_size_read(inode);
93 if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
94 printk_ratelimited(KERN_DEBUG
95 "BTRFS: %s: ino %llu isize %llu odd range [%llu,%llu]\n",
96 caller, btrfs_ino(inode), isize, start, end);
97 }
98}
99#else
100#define btrfs_leak_debug_add(new, head) do {} while (0)
101#define btrfs_leak_debug_del(entry) do {} while (0)
102#define btrfs_leak_debug_check() do {} while (0)
103#define btrfs_debug_check_extent_io_range(c, s, e) do {} while (0)
104#endif
105
106#define BUFFER_LRU_MAX 64
107
108struct tree_entry {
109 u64 start;
110 u64 end;
111 struct rb_node rb_node;
112};
113
114struct extent_page_data {
115 struct bio *bio;
116 struct extent_io_tree *tree;
117 get_extent_t *get_extent;
118 unsigned long bio_flags;
119
120 /* tells writepage not to lock the state bits for this range
121 * it still does the unlocking
122 */
123 unsigned int extent_locked:1;
124
125 /* tells the submit_bio code to use a WRITE_SYNC */
126 unsigned int sync_io:1;
127};
128
129static noinline void flush_write_bio(void *data);
130static inline struct btrfs_fs_info *
131tree_fs_info(struct extent_io_tree *tree)
132{
133 if (!tree->mapping)
134 return NULL;
135 return btrfs_sb(tree->mapping->host->i_sb);
136}
137
138int __init extent_io_init(void)
139{
140 extent_state_cache = kmem_cache_create("btrfs_extent_state",
141 sizeof(struct extent_state), 0,
142 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
143 if (!extent_state_cache)
144 return -ENOMEM;
145
146 extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
147 sizeof(struct extent_buffer), 0,
148 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
149 if (!extent_buffer_cache)
150 goto free_state_cache;
151
152 btrfs_bioset = bioset_create(BIO_POOL_SIZE,
153 offsetof(struct btrfs_io_bio, bio));
154 if (!btrfs_bioset)
155 goto free_buffer_cache;
156
157 if (bioset_integrity_create(btrfs_bioset, BIO_POOL_SIZE))
158 goto free_bioset;
159
160 return 0;
161
162free_bioset:
163 bioset_free(btrfs_bioset);
164 btrfs_bioset = NULL;
165
166free_buffer_cache:
167 kmem_cache_destroy(extent_buffer_cache);
168 extent_buffer_cache = NULL;
169
170free_state_cache:
171 kmem_cache_destroy(extent_state_cache);
172 extent_state_cache = NULL;
173 return -ENOMEM;
174}
175
176void extent_io_exit(void)
177{
178 btrfs_leak_debug_check();
179
180 /*
181 * Make sure all delayed rcu free are flushed before we
182 * destroy caches.
183 */
184 rcu_barrier();
185 if (extent_state_cache)
186 kmem_cache_destroy(extent_state_cache);
187 if (extent_buffer_cache)
188 kmem_cache_destroy(extent_buffer_cache);
189 if (btrfs_bioset)
190 bioset_free(btrfs_bioset);
191}
192
193void extent_io_tree_init(struct extent_io_tree *tree,
194 struct address_space *mapping)
195{
196 tree->state = RB_ROOT;
197 tree->ops = NULL;
198 tree->dirty_bytes = 0;
199 spin_lock_init(&tree->lock);
200 tree->mapping = mapping;
201}
202
203static struct extent_state *alloc_extent_state(gfp_t mask)
204{
205 struct extent_state *state;
206
207 state = kmem_cache_alloc(extent_state_cache, mask);
208 if (!state)
209 return state;
210 state->state = 0;
211 state->private = 0;
212 state->tree = NULL;
213 btrfs_leak_debug_add(&state->leak_list, &states);
214 atomic_set(&state->refs, 1);
215 init_waitqueue_head(&state->wq);
216 trace_alloc_extent_state(state, mask, _RET_IP_);
217 return state;
218}
219
220void free_extent_state(struct extent_state *state)
221{
222 if (!state)
223 return;
224 if (atomic_dec_and_test(&state->refs)) {
225 WARN_ON(state->tree);
226 btrfs_leak_debug_del(&state->leak_list);
227 trace_free_extent_state(state, _RET_IP_);
228 kmem_cache_free(extent_state_cache, state);
229 }
230}
231
232static struct rb_node *tree_insert(struct rb_root *root,
233 struct rb_node *search_start,
234 u64 offset,
235 struct rb_node *node,
236 struct rb_node ***p_in,
237 struct rb_node **parent_in)
238{
239 struct rb_node **p;
240 struct rb_node *parent = NULL;
241 struct tree_entry *entry;
242
243 if (p_in && parent_in) {
244 p = *p_in;
245 parent = *parent_in;
246 goto do_insert;
247 }
248
249 p = search_start ? &search_start : &root->rb_node;
250 while (*p) {
251 parent = *p;
252 entry = rb_entry(parent, struct tree_entry, rb_node);
253
254 if (offset < entry->start)
255 p = &(*p)->rb_left;
256 else if (offset > entry->end)
257 p = &(*p)->rb_right;
258 else
259 return parent;
260 }
261
262do_insert:
263 rb_link_node(node, parent, p);
264 rb_insert_color(node, root);
265 return NULL;
266}
267
268static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
269 struct rb_node **prev_ret,
270 struct rb_node **next_ret,
271 struct rb_node ***p_ret,
272 struct rb_node **parent_ret)
273{
274 struct rb_root *root = &tree->state;
275 struct rb_node **n = &root->rb_node;
276 struct rb_node *prev = NULL;
277 struct rb_node *orig_prev = NULL;
278 struct tree_entry *entry;
279 struct tree_entry *prev_entry = NULL;
280
281 while (*n) {
282 prev = *n;
283 entry = rb_entry(prev, struct tree_entry, rb_node);
284 prev_entry = entry;
285
286 if (offset < entry->start)
287 n = &(*n)->rb_left;
288 else if (offset > entry->end)
289 n = &(*n)->rb_right;
290 else
291 return *n;
292 }
293
294 if (p_ret)
295 *p_ret = n;
296 if (parent_ret)
297 *parent_ret = prev;
298
299 if (prev_ret) {
300 orig_prev = prev;
301 while (prev && offset > prev_entry->end) {
302 prev = rb_next(prev);
303 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
304 }
305 *prev_ret = prev;
306 prev = orig_prev;
307 }
308
309 if (next_ret) {
310 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
311 while (prev && offset < prev_entry->start) {
312 prev = rb_prev(prev);
313 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
314 }
315 *next_ret = prev;
316 }
317 return NULL;
318}
319
320static inline struct rb_node *
321tree_search_for_insert(struct extent_io_tree *tree,
322 u64 offset,
323 struct rb_node ***p_ret,
324 struct rb_node **parent_ret)
325{
326 struct rb_node *prev = NULL;
327 struct rb_node *ret;
328
329 ret = __etree_search(tree, offset, &prev, NULL, p_ret, parent_ret);
330 if (!ret)
331 return prev;
332 return ret;
333}
334
335static inline struct rb_node *tree_search(struct extent_io_tree *tree,
336 u64 offset)
337{
338 return tree_search_for_insert(tree, offset, NULL, NULL);
339}
340
341static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
342 struct extent_state *other)
343{
344 if (tree->ops && tree->ops->merge_extent_hook)
345 tree->ops->merge_extent_hook(tree->mapping->host, new,
346 other);
347}
348
349/*
350 * utility function to look for merge candidates inside a given range.
351 * Any extents with matching state are merged together into a single
352 * extent in the tree. Extents with EXTENT_IO in their state field
353 * are not merged because the end_io handlers need to be able to do
354 * operations on them without sleeping (or doing allocations/splits).
355 *
356 * This should be called with the tree lock held.
357 */
358static void merge_state(struct extent_io_tree *tree,
359 struct extent_state *state)
360{
361 struct extent_state *other;
362 struct rb_node *other_node;
363
364 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
365 return;
366
367 other_node = rb_prev(&state->rb_node);
368 if (other_node) {
369 other = rb_entry(other_node, struct extent_state, rb_node);
370 if (other->end == state->start - 1 &&
371 other->state == state->state) {
372 merge_cb(tree, state, other);
373 state->start = other->start;
374 other->tree = NULL;
375 rb_erase(&other->rb_node, &tree->state);
376 free_extent_state(other);
377 }
378 }
379 other_node = rb_next(&state->rb_node);
380 if (other_node) {
381 other = rb_entry(other_node, struct extent_state, rb_node);
382 if (other->start == state->end + 1 &&
383 other->state == state->state) {
384 merge_cb(tree, state, other);
385 state->end = other->end;
386 other->tree = NULL;
387 rb_erase(&other->rb_node, &tree->state);
388 free_extent_state(other);
389 }
390 }
391}
392
393static void set_state_cb(struct extent_io_tree *tree,
394 struct extent_state *state, unsigned long *bits)
395{
396 if (tree->ops && tree->ops->set_bit_hook)
397 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
398}
399
400static void clear_state_cb(struct extent_io_tree *tree,
401 struct extent_state *state, unsigned long *bits)
402{
403 if (tree->ops && tree->ops->clear_bit_hook)
404 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
405}
406
407static void set_state_bits(struct extent_io_tree *tree,
408 struct extent_state *state, unsigned long *bits);
409
410/*
411 * insert an extent_state struct into the tree. 'bits' are set on the
412 * struct before it is inserted.
413 *
414 * This may return -EEXIST if the extent is already there, in which case the
415 * state struct is freed.
416 *
417 * The tree lock is not taken internally. This is a utility function and
418 * probably isn't what you want to call (see set/clear_extent_bit).
419 */
420static int insert_state(struct extent_io_tree *tree,
421 struct extent_state *state, u64 start, u64 end,
422 struct rb_node ***p,
423 struct rb_node **parent,
424 unsigned long *bits)
425{
426 struct rb_node *node;
427
428 if (end < start)
429 WARN(1, KERN_ERR "BTRFS: end < start %llu %llu\n",
430 end, start);
431 state->start = start;
432 state->end = end;
433
434 set_state_bits(tree, state, bits);
435
436 node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
437 if (node) {
438 struct extent_state *found;
439 found = rb_entry(node, struct extent_state, rb_node);
440 printk(KERN_ERR "BTRFS: found node %llu %llu on insert of "
441 "%llu %llu\n",
442 found->start, found->end, start, end);
443 return -EEXIST;
444 }
445 state->tree = tree;
446 merge_state(tree, state);
447 return 0;
448}
449
450static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
451 u64 split)
452{
453 if (tree->ops && tree->ops->split_extent_hook)
454 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
455}
456
457/*
458 * split a given extent state struct in two, inserting the preallocated
459 * struct 'prealloc' as the newly created second half. 'split' indicates an
460 * offset inside 'orig' where it should be split.
461 *
462 * Before calling,
463 * the tree has 'orig' at [orig->start, orig->end]. After calling, there
464 * are two extent state structs in the tree:
465 * prealloc: [orig->start, split - 1]
466 * orig: [ split, orig->end ]
467 *
468 * The tree locks are not taken by this function. They need to be held
469 * by the caller.
470 */
471static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
472 struct extent_state *prealloc, u64 split)
473{
474 struct rb_node *node;
475
476 split_cb(tree, orig, split);
477
478 prealloc->start = orig->start;
479 prealloc->end = split - 1;
480 prealloc->state = orig->state;
481 orig->start = split;
482
483 node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
484 &prealloc->rb_node, NULL, NULL);
485 if (node) {
486 free_extent_state(prealloc);
487 return -EEXIST;
488 }
489 prealloc->tree = tree;
490 return 0;
491}
492
493static struct extent_state *next_state(struct extent_state *state)
494{
495 struct rb_node *next = rb_next(&state->rb_node);
496 if (next)
497 return rb_entry(next, struct extent_state, rb_node);
498 else
499 return NULL;
500}
501
502/*
503 * utility function to clear some bits in an extent state struct.
504 * it will optionally wake up any one waiting on this state (wake == 1).
505 *
506 * If no bits are set on the state struct after clearing things, the
507 * struct is freed and removed from the tree
508 */
509static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
510 struct extent_state *state,
511 unsigned long *bits, int wake)
512{
513 struct extent_state *next;
514 unsigned long bits_to_clear = *bits & ~EXTENT_CTLBITS;
515
516 if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
517 u64 range = state->end - state->start + 1;
518 WARN_ON(range > tree->dirty_bytes);
519 tree->dirty_bytes -= range;
520 }
521 clear_state_cb(tree, state, bits);
522 state->state &= ~bits_to_clear;
523 if (wake)
524 wake_up(&state->wq);
525 if (state->state == 0) {
526 next = next_state(state);
527 if (state->tree) {
528 rb_erase(&state->rb_node, &tree->state);
529 state->tree = NULL;
530 free_extent_state(state);
531 } else {
532 WARN_ON(1);
533 }
534 } else {
535 merge_state(tree, state);
536 next = next_state(state);
537 }
538 return next;
539}
540
541static struct extent_state *
542alloc_extent_state_atomic(struct extent_state *prealloc)
543{
544 if (!prealloc)
545 prealloc = alloc_extent_state(GFP_ATOMIC);
546
547 return prealloc;
548}
549
550static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
551{
552 btrfs_panic(tree_fs_info(tree), err, "Locking error: "
553 "Extent tree was modified by another "
554 "thread while locked.");
555}
556
557/*
558 * clear some bits on a range in the tree. This may require splitting
559 * or inserting elements in the tree, so the gfp mask is used to
560 * indicate which allocations or sleeping are allowed.
561 *
562 * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
563 * the given range from the tree regardless of state (ie for truncate).
564 *
565 * the range [start, end] is inclusive.
566 *
567 * This takes the tree lock, and returns 0 on success and < 0 on error.
568 */
569int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
570 unsigned long bits, int wake, int delete,
571 struct extent_state **cached_state,
572 gfp_t mask)
573{
574 struct extent_state *state;
575 struct extent_state *cached;
576 struct extent_state *prealloc = NULL;
577 struct rb_node *node;
578 u64 last_end;
579 int err;
580 int clear = 0;
581
582 btrfs_debug_check_extent_io_range(tree, start, end);
583
584 if (bits & EXTENT_DELALLOC)
585 bits |= EXTENT_NORESERVE;
586
587 if (delete)
588 bits |= ~EXTENT_CTLBITS;
589 bits |= EXTENT_FIRST_DELALLOC;
590
591 if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
592 clear = 1;
593again:
594 if (!prealloc && (mask & __GFP_WAIT)) {
595 prealloc = alloc_extent_state(mask);
596 if (!prealloc)
597 return -ENOMEM;
598 }
599
600 spin_lock(&tree->lock);
601 if (cached_state) {
602 cached = *cached_state;
603
604 if (clear) {
605 *cached_state = NULL;
606 cached_state = NULL;
607 }
608
609 if (cached && cached->tree && cached->start <= start &&
610 cached->end > start) {
611 if (clear)
612 atomic_dec(&cached->refs);
613 state = cached;
614 goto hit_next;
615 }
616 if (clear)
617 free_extent_state(cached);
618 }
619 /*
620 * this search will find the extents that end after
621 * our range starts
622 */
623 node = tree_search(tree, start);
624 if (!node)
625 goto out;
626 state = rb_entry(node, struct extent_state, rb_node);
627hit_next:
628 if (state->start > end)
629 goto out;
630 WARN_ON(state->end < start);
631 last_end = state->end;
632
633 /* the state doesn't have the wanted bits, go ahead */
634 if (!(state->state & bits)) {
635 state = next_state(state);
636 goto next;
637 }
638
639 /*
640 * | ---- desired range ---- |
641 * | state | or
642 * | ------------- state -------------- |
643 *
644 * We need to split the extent we found, and may flip
645 * bits on second half.
646 *
647 * If the extent we found extends past our range, we
648 * just split and search again. It'll get split again
649 * the next time though.
650 *
651 * If the extent we found is inside our range, we clear
652 * the desired bit on it.
653 */
654
655 if (state->start < start) {
656 prealloc = alloc_extent_state_atomic(prealloc);
657 BUG_ON(!prealloc);
658 err = split_state(tree, state, prealloc, start);
659 if (err)
660 extent_io_tree_panic(tree, err);
661
662 prealloc = NULL;
663 if (err)
664 goto out;
665 if (state->end <= end) {
666 state = clear_state_bit(tree, state, &bits, wake);
667 goto next;
668 }
669 goto search_again;
670 }
671 /*
672 * | ---- desired range ---- |
673 * | state |
674 * We need to split the extent, and clear the bit
675 * on the first half
676 */
677 if (state->start <= end && state->end > end) {
678 prealloc = alloc_extent_state_atomic(prealloc);
679 BUG_ON(!prealloc);
680 err = split_state(tree, state, prealloc, end + 1);
681 if (err)
682 extent_io_tree_panic(tree, err);
683
684 if (wake)
685 wake_up(&state->wq);
686
687 clear_state_bit(tree, prealloc, &bits, wake);
688
689 prealloc = NULL;
690 goto out;
691 }
692
693 state = clear_state_bit(tree, state, &bits, wake);
694next:
695 if (last_end == (u64)-1)
696 goto out;
697 start = last_end + 1;
698 if (start <= end && state && !need_resched())
699 goto hit_next;
700 goto search_again;
701
702out:
703 spin_unlock(&tree->lock);
704 if (prealloc)
705 free_extent_state(prealloc);
706
707 return 0;
708
709search_again:
710 if (start > end)
711 goto out;
712 spin_unlock(&tree->lock);
713 if (mask & __GFP_WAIT)
714 cond_resched();
715 goto again;
716}
717
718static void wait_on_state(struct extent_io_tree *tree,
719 struct extent_state *state)
720 __releases(tree->lock)
721 __acquires(tree->lock)
722{
723 DEFINE_WAIT(wait);
724 prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
725 spin_unlock(&tree->lock);
726 schedule();
727 spin_lock(&tree->lock);
728 finish_wait(&state->wq, &wait);
729}
730
731/*
732 * waits for one or more bits to clear on a range in the state tree.
733 * The range [start, end] is inclusive.
734 * The tree lock is taken by this function
735 */
736static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
737 unsigned long bits)
738{
739 struct extent_state *state;
740 struct rb_node *node;
741
742 btrfs_debug_check_extent_io_range(tree, start, end);
743
744 spin_lock(&tree->lock);
745again:
746 while (1) {
747 /*
748 * this search will find all the extents that end after
749 * our range starts
750 */
751 node = tree_search(tree, start);
752process_node:
753 if (!node)
754 break;
755
756 state = rb_entry(node, struct extent_state, rb_node);
757
758 if (state->start > end)
759 goto out;
760
761 if (state->state & bits) {
762 start = state->start;
763 atomic_inc(&state->refs);
764 wait_on_state(tree, state);
765 free_extent_state(state);
766 goto again;
767 }
768 start = state->end + 1;
769
770 if (start > end)
771 break;
772
773 if (!cond_resched_lock(&tree->lock)) {
774 node = rb_next(node);
775 goto process_node;
776 }
777 }
778out:
779 spin_unlock(&tree->lock);
780}
781
782static void set_state_bits(struct extent_io_tree *tree,
783 struct extent_state *state,
784 unsigned long *bits)
785{
786 unsigned long bits_to_set = *bits & ~EXTENT_CTLBITS;
787
788 set_state_cb(tree, state, bits);
789 if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
790 u64 range = state->end - state->start + 1;
791 tree->dirty_bytes += range;
792 }
793 state->state |= bits_to_set;
794}
795
796static void cache_state(struct extent_state *state,
797 struct extent_state **cached_ptr)
798{
799 if (cached_ptr && !(*cached_ptr)) {
800 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
801 *cached_ptr = state;
802 atomic_inc(&state->refs);
803 }
804 }
805}
806
807/*
808 * set some bits on a range in the tree. This may require allocations or
809 * sleeping, so the gfp mask is used to indicate what is allowed.
810 *
811 * If any of the exclusive bits are set, this will fail with -EEXIST if some
812 * part of the range already has the desired bits set. The start of the
813 * existing range is returned in failed_start in this case.
814 *
815 * [start, end] is inclusive This takes the tree lock.
816 */
817
818static int __must_check
819__set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
820 unsigned long bits, unsigned long exclusive_bits,
821 u64 *failed_start, struct extent_state **cached_state,
822 gfp_t mask)
823{
824 struct extent_state *state;
825 struct extent_state *prealloc = NULL;
826 struct rb_node *node;
827 struct rb_node **p;
828 struct rb_node *parent;
829 int err = 0;
830 u64 last_start;
831 u64 last_end;
832
833 btrfs_debug_check_extent_io_range(tree, start, end);
834
835 bits |= EXTENT_FIRST_DELALLOC;
836again:
837 if (!prealloc && (mask & __GFP_WAIT)) {
838 prealloc = alloc_extent_state(mask);
839 BUG_ON(!prealloc);
840 }
841
842 spin_lock(&tree->lock);
843 if (cached_state && *cached_state) {
844 state = *cached_state;
845 if (state->start <= start && state->end > start &&
846 state->tree) {
847 node = &state->rb_node;
848 goto hit_next;
849 }
850 }
851 /*
852 * this search will find all the extents that end after
853 * our range starts.
854 */
855 node = tree_search_for_insert(tree, start, &p, &parent);
856 if (!node) {
857 prealloc = alloc_extent_state_atomic(prealloc);
858 BUG_ON(!prealloc);
859 err = insert_state(tree, prealloc, start, end,
860 &p, &parent, &bits);
861 if (err)
862 extent_io_tree_panic(tree, err);
863
864 cache_state(prealloc, cached_state);
865 prealloc = NULL;
866 goto out;
867 }
868 state = rb_entry(node, struct extent_state, rb_node);
869hit_next:
870 last_start = state->start;
871 last_end = state->end;
872
873 /*
874 * | ---- desired range ---- |
875 * | state |
876 *
877 * Just lock what we found and keep going
878 */
879 if (state->start == start && state->end <= end) {
880 if (state->state & exclusive_bits) {
881 *failed_start = state->start;
882 err = -EEXIST;
883 goto out;
884 }
885
886 set_state_bits(tree, state, &bits);
887 cache_state(state, cached_state);
888 merge_state(tree, state);
889 if (last_end == (u64)-1)
890 goto out;
891 start = last_end + 1;
892 state = next_state(state);
893 if (start < end && state && state->start == start &&
894 !need_resched())
895 goto hit_next;
896 goto search_again;
897 }
898
899 /*
900 * | ---- desired range ---- |
901 * | state |
902 * or
903 * | ------------- state -------------- |
904 *
905 * We need to split the extent we found, and may flip bits on
906 * second half.
907 *
908 * If the extent we found extends past our
909 * range, we just split and search again. It'll get split
910 * again the next time though.
911 *
912 * If the extent we found is inside our range, we set the
913 * desired bit on it.
914 */
915 if (state->start < start) {
916 if (state->state & exclusive_bits) {
917 *failed_start = start;
918 err = -EEXIST;
919 goto out;
920 }
921
922 prealloc = alloc_extent_state_atomic(prealloc);
923 BUG_ON(!prealloc);
924 err = split_state(tree, state, prealloc, start);
925 if (err)
926 extent_io_tree_panic(tree, err);
927
928 prealloc = NULL;
929 if (err)
930 goto out;
931 if (state->end <= end) {
932 set_state_bits(tree, state, &bits);
933 cache_state(state, cached_state);
934 merge_state(tree, state);
935 if (last_end == (u64)-1)
936 goto out;
937 start = last_end + 1;
938 state = next_state(state);
939 if (start < end && state && state->start == start &&
940 !need_resched())
941 goto hit_next;
942 }
943 goto search_again;
944 }
945 /*
946 * | ---- desired range ---- |
947 * | state | or | state |
948 *
949 * There's a hole, we need to insert something in it and
950 * ignore the extent we found.
951 */
952 if (state->start > start) {
953 u64 this_end;
954 if (end < last_start)
955 this_end = end;
956 else
957 this_end = last_start - 1;
958
959 prealloc = alloc_extent_state_atomic(prealloc);
960 BUG_ON(!prealloc);
961
962 /*
963 * Avoid to free 'prealloc' if it can be merged with
964 * the later extent.
965 */
966 err = insert_state(tree, prealloc, start, this_end,
967 NULL, NULL, &bits);
968 if (err)
969 extent_io_tree_panic(tree, err);
970
971 cache_state(prealloc, cached_state);
972 prealloc = NULL;
973 start = this_end + 1;
974 goto search_again;
975 }
976 /*
977 * | ---- desired range ---- |
978 * | state |
979 * We need to split the extent, and set the bit
980 * on the first half
981 */
982 if (state->start <= end && state->end > end) {
983 if (state->state & exclusive_bits) {
984 *failed_start = start;
985 err = -EEXIST;
986 goto out;
987 }
988
989 prealloc = alloc_extent_state_atomic(prealloc);
990 BUG_ON(!prealloc);
991 err = split_state(tree, state, prealloc, end + 1);
992 if (err)
993 extent_io_tree_panic(tree, err);
994
995 set_state_bits(tree, prealloc, &bits);
996 cache_state(prealloc, cached_state);
997 merge_state(tree, prealloc);
998 prealloc = NULL;
999 goto out;
1000 }
1001
1002 goto search_again;
1003
1004out:
1005 spin_unlock(&tree->lock);
1006 if (prealloc)
1007 free_extent_state(prealloc);
1008
1009 return err;
1010
1011search_again:
1012 if (start > end)
1013 goto out;
1014 spin_unlock(&tree->lock);
1015 if (mask & __GFP_WAIT)
1016 cond_resched();
1017 goto again;
1018}
1019
1020int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1021 unsigned long bits, u64 * failed_start,
1022 struct extent_state **cached_state, gfp_t mask)
1023{
1024 return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1025 cached_state, mask);
1026}
1027
1028
1029/**
1030 * convert_extent_bit - convert all bits in a given range from one bit to
1031 * another
1032 * @tree: the io tree to search
1033 * @start: the start offset in bytes
1034 * @end: the end offset in bytes (inclusive)
1035 * @bits: the bits to set in this range
1036 * @clear_bits: the bits to clear in this range
1037 * @cached_state: state that we're going to cache
1038 * @mask: the allocation mask
1039 *
1040 * This will go through and set bits for the given range. If any states exist
1041 * already in this range they are set with the given bit and cleared of the
1042 * clear_bits. This is only meant to be used by things that are mergeable, ie
1043 * converting from say DELALLOC to DIRTY. This is not meant to be used with
1044 * boundary bits like LOCK.
1045 */
1046int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1047 unsigned long bits, unsigned long clear_bits,
1048 struct extent_state **cached_state, gfp_t mask)
1049{
1050 struct extent_state *state;
1051 struct extent_state *prealloc = NULL;
1052 struct rb_node *node;
1053 struct rb_node **p;
1054 struct rb_node *parent;
1055 int err = 0;
1056 u64 last_start;
1057 u64 last_end;
1058
1059 btrfs_debug_check_extent_io_range(tree, start, end);
1060
1061again:
1062 if (!prealloc && (mask & __GFP_WAIT)) {
1063 prealloc = alloc_extent_state(mask);
1064 if (!prealloc)
1065 return -ENOMEM;
1066 }
1067
1068 spin_lock(&tree->lock);
1069 if (cached_state && *cached_state) {
1070 state = *cached_state;
1071 if (state->start <= start && state->end > start &&
1072 state->tree) {
1073 node = &state->rb_node;
1074 goto hit_next;
1075 }
1076 }
1077
1078 /*
1079 * this search will find all the extents that end after
1080 * our range starts.
1081 */
1082 node = tree_search_for_insert(tree, start, &p, &parent);
1083 if (!node) {
1084 prealloc = alloc_extent_state_atomic(prealloc);
1085 if (!prealloc) {
1086 err = -ENOMEM;
1087 goto out;
1088 }
1089 err = insert_state(tree, prealloc, start, end,
1090 &p, &parent, &bits);
1091 if (err)
1092 extent_io_tree_panic(tree, err);
1093 cache_state(prealloc, cached_state);
1094 prealloc = NULL;
1095 goto out;
1096 }
1097 state = rb_entry(node, struct extent_state, rb_node);
1098hit_next:
1099 last_start = state->start;
1100 last_end = state->end;
1101
1102 /*
1103 * | ---- desired range ---- |
1104 * | state |
1105 *
1106 * Just lock what we found and keep going
1107 */
1108 if (state->start == start && state->end <= end) {
1109 set_state_bits(tree, state, &bits);
1110 cache_state(state, cached_state);
1111 state = clear_state_bit(tree, state, &clear_bits, 0);
1112 if (last_end == (u64)-1)
1113 goto out;
1114 start = last_end + 1;
1115 if (start < end && state && state->start == start &&
1116 !need_resched())
1117 goto hit_next;
1118 goto search_again;
1119 }
1120
1121 /*
1122 * | ---- desired range ---- |
1123 * | state |
1124 * or
1125 * | ------------- state -------------- |
1126 *
1127 * We need to split the extent we found, and may flip bits on
1128 * second half.
1129 *
1130 * If the extent we found extends past our
1131 * range, we just split and search again. It'll get split
1132 * again the next time though.
1133 *
1134 * If the extent we found is inside our range, we set the
1135 * desired bit on it.
1136 */
1137 if (state->start < start) {
1138 prealloc = alloc_extent_state_atomic(prealloc);
1139 if (!prealloc) {
1140 err = -ENOMEM;
1141 goto out;
1142 }
1143 err = split_state(tree, state, prealloc, start);
1144 if (err)
1145 extent_io_tree_panic(tree, err);
1146 prealloc = NULL;
1147 if (err)
1148 goto out;
1149 if (state->end <= end) {
1150 set_state_bits(tree, state, &bits);
1151 cache_state(state, cached_state);
1152 state = clear_state_bit(tree, state, &clear_bits, 0);
1153 if (last_end == (u64)-1)
1154 goto out;
1155 start = last_end + 1;
1156 if (start < end && state && state->start == start &&
1157 !need_resched())
1158 goto hit_next;
1159 }
1160 goto search_again;
1161 }
1162 /*
1163 * | ---- desired range ---- |
1164 * | state | or | state |
1165 *
1166 * There's a hole, we need to insert something in it and
1167 * ignore the extent we found.
1168 */
1169 if (state->start > start) {
1170 u64 this_end;
1171 if (end < last_start)
1172 this_end = end;
1173 else
1174 this_end = last_start - 1;
1175
1176 prealloc = alloc_extent_state_atomic(prealloc);
1177 if (!prealloc) {
1178 err = -ENOMEM;
1179 goto out;
1180 }
1181
1182 /*
1183 * Avoid to free 'prealloc' if it can be merged with
1184 * the later extent.
1185 */
1186 err = insert_state(tree, prealloc, start, this_end,
1187 NULL, NULL, &bits);
1188 if (err)
1189 extent_io_tree_panic(tree, err);
1190 cache_state(prealloc, cached_state);
1191 prealloc = NULL;
1192 start = this_end + 1;
1193 goto search_again;
1194 }
1195 /*
1196 * | ---- desired range ---- |
1197 * | state |
1198 * We need to split the extent, and set the bit
1199 * on the first half
1200 */
1201 if (state->start <= end && state->end > end) {
1202 prealloc = alloc_extent_state_atomic(prealloc);
1203 if (!prealloc) {
1204 err = -ENOMEM;
1205 goto out;
1206 }
1207
1208 err = split_state(tree, state, prealloc, end + 1);
1209 if (err)
1210 extent_io_tree_panic(tree, err);
1211
1212 set_state_bits(tree, prealloc, &bits);
1213 cache_state(prealloc, cached_state);
1214 clear_state_bit(tree, prealloc, &clear_bits, 0);
1215 prealloc = NULL;
1216 goto out;
1217 }
1218
1219 goto search_again;
1220
1221out:
1222 spin_unlock(&tree->lock);
1223 if (prealloc)
1224 free_extent_state(prealloc);
1225
1226 return err;
1227
1228search_again:
1229 if (start > end)
1230 goto out;
1231 spin_unlock(&tree->lock);
1232 if (mask & __GFP_WAIT)
1233 cond_resched();
1234 goto again;
1235}
1236
1237/* wrappers around set/clear extent bit */
1238int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1239 gfp_t mask)
1240{
1241 return set_extent_bit(tree, start, end, EXTENT_DIRTY, NULL,
1242 NULL, mask);
1243}
1244
1245int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1246 unsigned long bits, gfp_t mask)
1247{
1248 return set_extent_bit(tree, start, end, bits, NULL,
1249 NULL, mask);
1250}
1251
1252int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1253 unsigned long bits, gfp_t mask)
1254{
1255 return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
1256}
1257
1258int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
1259 struct extent_state **cached_state, gfp_t mask)
1260{
1261 return set_extent_bit(tree, start, end,
1262 EXTENT_DELALLOC | EXTENT_UPTODATE,
1263 NULL, cached_state, mask);
1264}
1265
1266int set_extent_defrag(struct extent_io_tree *tree, u64 start, u64 end,
1267 struct extent_state **cached_state, gfp_t mask)
1268{
1269 return set_extent_bit(tree, start, end,
1270 EXTENT_DELALLOC | EXTENT_UPTODATE | EXTENT_DEFRAG,
1271 NULL, cached_state, mask);
1272}
1273
1274int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
1275 gfp_t mask)
1276{
1277 return clear_extent_bit(tree, start, end,
1278 EXTENT_DIRTY | EXTENT_DELALLOC |
1279 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
1280}
1281
1282int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
1283 gfp_t mask)
1284{
1285 return set_extent_bit(tree, start, end, EXTENT_NEW, NULL,
1286 NULL, mask);
1287}
1288
1289int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1290 struct extent_state **cached_state, gfp_t mask)
1291{
1292 return set_extent_bit(tree, start, end, EXTENT_UPTODATE, NULL,
1293 cached_state, mask);
1294}
1295
1296int clear_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
1297 struct extent_state **cached_state, gfp_t mask)
1298{
1299 return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
1300 cached_state, mask);
1301}
1302
1303/*
1304 * either insert or lock state struct between start and end use mask to tell
1305 * us if waiting is desired.
1306 */
1307int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1308 unsigned long bits, struct extent_state **cached_state)
1309{
1310 int err;
1311 u64 failed_start;
1312 while (1) {
1313 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
1314 EXTENT_LOCKED, &failed_start,
1315 cached_state, GFP_NOFS);
1316 if (err == -EEXIST) {
1317 wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1318 start = failed_start;
1319 } else
1320 break;
1321 WARN_ON(start > end);
1322 }
1323 return err;
1324}
1325
1326int lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1327{
1328 return lock_extent_bits(tree, start, end, 0, NULL);
1329}
1330
1331int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1332{
1333 int err;
1334 u64 failed_start;
1335
1336 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1337 &failed_start, NULL, GFP_NOFS);
1338 if (err == -EEXIST) {
1339 if (failed_start > start)
1340 clear_extent_bit(tree, start, failed_start - 1,
1341 EXTENT_LOCKED, 1, 0, NULL, GFP_NOFS);
1342 return 0;
1343 }
1344 return 1;
1345}
1346
1347int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1348 struct extent_state **cached, gfp_t mask)
1349{
1350 return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1351 mask);
1352}
1353
1354int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1355{
1356 return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1357 GFP_NOFS);
1358}
1359
1360int extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1361{
1362 unsigned long index = start >> PAGE_CACHE_SHIFT;
1363 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1364 struct page *page;
1365
1366 while (index <= end_index) {
1367 page = find_get_page(inode->i_mapping, index);
1368 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1369 clear_page_dirty_for_io(page);
1370 page_cache_release(page);
1371 index++;
1372 }
1373 return 0;
1374}
1375
1376int extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1377{
1378 unsigned long index = start >> PAGE_CACHE_SHIFT;
1379 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1380 struct page *page;
1381
1382 while (index <= end_index) {
1383 page = find_get_page(inode->i_mapping, index);
1384 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1385 account_page_redirty(page);
1386 __set_page_dirty_nobuffers(page);
1387 page_cache_release(page);
1388 index++;
1389 }
1390 return 0;
1391}
1392
1393/*
1394 * helper function to set both pages and extents in the tree writeback
1395 */
1396static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1397{
1398 unsigned long index = start >> PAGE_CACHE_SHIFT;
1399 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1400 struct page *page;
1401
1402 while (index <= end_index) {
1403 page = find_get_page(tree->mapping, index);
1404 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1405 set_page_writeback(page);
1406 page_cache_release(page);
1407 index++;
1408 }
1409 return 0;
1410}
1411
1412/* find the first state struct with 'bits' set after 'start', and
1413 * return it. tree->lock must be held. NULL will returned if
1414 * nothing was found after 'start'
1415 */
1416static struct extent_state *
1417find_first_extent_bit_state(struct extent_io_tree *tree,
1418 u64 start, unsigned long bits)
1419{
1420 struct rb_node *node;
1421 struct extent_state *state;
1422
1423 /*
1424 * this search will find all the extents that end after
1425 * our range starts.
1426 */
1427 node = tree_search(tree, start);
1428 if (!node)
1429 goto out;
1430
1431 while (1) {
1432 state = rb_entry(node, struct extent_state, rb_node);
1433 if (state->end >= start && (state->state & bits))
1434 return state;
1435
1436 node = rb_next(node);
1437 if (!node)
1438 break;
1439 }
1440out:
1441 return NULL;
1442}
1443
1444/*
1445 * find the first offset in the io tree with 'bits' set. zero is
1446 * returned if we find something, and *start_ret and *end_ret are
1447 * set to reflect the state struct that was found.
1448 *
1449 * If nothing was found, 1 is returned. If found something, return 0.
1450 */
1451int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1452 u64 *start_ret, u64 *end_ret, unsigned long bits,
1453 struct extent_state **cached_state)
1454{
1455 struct extent_state *state;
1456 struct rb_node *n;
1457 int ret = 1;
1458
1459 spin_lock(&tree->lock);
1460 if (cached_state && *cached_state) {
1461 state = *cached_state;
1462 if (state->end == start - 1 && state->tree) {
1463 n = rb_next(&state->rb_node);
1464 while (n) {
1465 state = rb_entry(n, struct extent_state,
1466 rb_node);
1467 if (state->state & bits)
1468 goto got_it;
1469 n = rb_next(n);
1470 }
1471 free_extent_state(*cached_state);
1472 *cached_state = NULL;
1473 goto out;
1474 }
1475 free_extent_state(*cached_state);
1476 *cached_state = NULL;
1477 }
1478
1479 state = find_first_extent_bit_state(tree, start, bits);
1480got_it:
1481 if (state) {
1482 cache_state(state, cached_state);
1483 *start_ret = state->start;
1484 *end_ret = state->end;
1485 ret = 0;
1486 }
1487out:
1488 spin_unlock(&tree->lock);
1489 return ret;
1490}
1491
1492/*
1493 * find a contiguous range of bytes in the file marked as delalloc, not
1494 * more than 'max_bytes'. start and end are used to return the range,
1495 *
1496 * 1 is returned if we find something, 0 if nothing was in the tree
1497 */
1498static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1499 u64 *start, u64 *end, u64 max_bytes,
1500 struct extent_state **cached_state)
1501{
1502 struct rb_node *node;
1503 struct extent_state *state;
1504 u64 cur_start = *start;
1505 u64 found = 0;
1506 u64 total_bytes = 0;
1507
1508 spin_lock(&tree->lock);
1509
1510 /*
1511 * this search will find all the extents that end after
1512 * our range starts.
1513 */
1514 node = tree_search(tree, cur_start);
1515 if (!node) {
1516 if (!found)
1517 *end = (u64)-1;
1518 goto out;
1519 }
1520
1521 while (1) {
1522 state = rb_entry(node, struct extent_state, rb_node);
1523 if (found && (state->start != cur_start ||
1524 (state->state & EXTENT_BOUNDARY))) {
1525 goto out;
1526 }
1527 if (!(state->state & EXTENT_DELALLOC)) {
1528 if (!found)
1529 *end = state->end;
1530 goto out;
1531 }
1532 if (!found) {
1533 *start = state->start;
1534 *cached_state = state;
1535 atomic_inc(&state->refs);
1536 }
1537 found++;
1538 *end = state->end;
1539 cur_start = state->end + 1;
1540 node = rb_next(node);
1541 total_bytes += state->end - state->start + 1;
1542 if (total_bytes >= max_bytes)
1543 break;
1544 if (!node)
1545 break;
1546 }
1547out:
1548 spin_unlock(&tree->lock);
1549 return found;
1550}
1551
1552static noinline void __unlock_for_delalloc(struct inode *inode,
1553 struct page *locked_page,
1554 u64 start, u64 end)
1555{
1556 int ret;
1557 struct page *pages[16];
1558 unsigned long index = start >> PAGE_CACHE_SHIFT;
1559 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1560 unsigned long nr_pages = end_index - index + 1;
1561 int i;
1562
1563 if (index == locked_page->index && end_index == index)
1564 return;
1565
1566 while (nr_pages > 0) {
1567 ret = find_get_pages_contig(inode->i_mapping, index,
1568 min_t(unsigned long, nr_pages,
1569 ARRAY_SIZE(pages)), pages);
1570 for (i = 0; i < ret; i++) {
1571 if (pages[i] != locked_page)
1572 unlock_page(pages[i]);
1573 page_cache_release(pages[i]);
1574 }
1575 nr_pages -= ret;
1576 index += ret;
1577 cond_resched();
1578 }
1579}
1580
1581static noinline int lock_delalloc_pages(struct inode *inode,
1582 struct page *locked_page,
1583 u64 delalloc_start,
1584 u64 delalloc_end)
1585{
1586 unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1587 unsigned long start_index = index;
1588 unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1589 unsigned long pages_locked = 0;
1590 struct page *pages[16];
1591 unsigned long nrpages;
1592 int ret;
1593 int i;
1594
1595 /* the caller is responsible for locking the start index */
1596 if (index == locked_page->index && index == end_index)
1597 return 0;
1598
1599 /* skip the page at the start index */
1600 nrpages = end_index - index + 1;
1601 while (nrpages > 0) {
1602 ret = find_get_pages_contig(inode->i_mapping, index,
1603 min_t(unsigned long,
1604 nrpages, ARRAY_SIZE(pages)), pages);
1605 if (ret == 0) {
1606 ret = -EAGAIN;
1607 goto done;
1608 }
1609 /* now we have an array of pages, lock them all */
1610 for (i = 0; i < ret; i++) {
1611 /*
1612 * the caller is taking responsibility for
1613 * locked_page
1614 */
1615 if (pages[i] != locked_page) {
1616 lock_page(pages[i]);
1617 if (!PageDirty(pages[i]) ||
1618 pages[i]->mapping != inode->i_mapping) {
1619 ret = -EAGAIN;
1620 unlock_page(pages[i]);
1621 page_cache_release(pages[i]);
1622 goto done;
1623 }
1624 }
1625 page_cache_release(pages[i]);
1626 pages_locked++;
1627 }
1628 nrpages -= ret;
1629 index += ret;
1630 cond_resched();
1631 }
1632 ret = 0;
1633done:
1634 if (ret && pages_locked) {
1635 __unlock_for_delalloc(inode, locked_page,
1636 delalloc_start,
1637 ((u64)(start_index + pages_locked - 1)) <<
1638 PAGE_CACHE_SHIFT);
1639 }
1640 return ret;
1641}
1642
1643/*
1644 * find a contiguous range of bytes in the file marked as delalloc, not
1645 * more than 'max_bytes'. start and end are used to return the range,
1646 *
1647 * 1 is returned if we find something, 0 if nothing was in the tree
1648 */
1649STATIC u64 find_lock_delalloc_range(struct inode *inode,
1650 struct extent_io_tree *tree,
1651 struct page *locked_page, u64 *start,
1652 u64 *end, u64 max_bytes)
1653{
1654 u64 delalloc_start;
1655 u64 delalloc_end;
1656 u64 found;
1657 struct extent_state *cached_state = NULL;
1658 int ret;
1659 int loops = 0;
1660
1661again:
1662 /* step one, find a bunch of delalloc bytes starting at start */
1663 delalloc_start = *start;
1664 delalloc_end = 0;
1665 found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1666 max_bytes, &cached_state);
1667 if (!found || delalloc_end <= *start) {
1668 *start = delalloc_start;
1669 *end = delalloc_end;
1670 free_extent_state(cached_state);
1671 return 0;
1672 }
1673
1674 /*
1675 * start comes from the offset of locked_page. We have to lock
1676 * pages in order, so we can't process delalloc bytes before
1677 * locked_page
1678 */
1679 if (delalloc_start < *start)
1680 delalloc_start = *start;
1681
1682 /*
1683 * make sure to limit the number of pages we try to lock down
1684 */
1685 if (delalloc_end + 1 - delalloc_start > max_bytes)
1686 delalloc_end = delalloc_start + max_bytes - 1;
1687
1688 /* step two, lock all the pages after the page that has start */
1689 ret = lock_delalloc_pages(inode, locked_page,
1690 delalloc_start, delalloc_end);
1691 if (ret == -EAGAIN) {
1692 /* some of the pages are gone, lets avoid looping by
1693 * shortening the size of the delalloc range we're searching
1694 */
1695 free_extent_state(cached_state);
1696 if (!loops) {
1697 max_bytes = PAGE_CACHE_SIZE;
1698 loops = 1;
1699 goto again;
1700 } else {
1701 found = 0;
1702 goto out_failed;
1703 }
1704 }
1705 BUG_ON(ret); /* Only valid values are 0 and -EAGAIN */
1706
1707 /* step three, lock the state bits for the whole range */
1708 lock_extent_bits(tree, delalloc_start, delalloc_end, 0, &cached_state);
1709
1710 /* then test to make sure it is all still delalloc */
1711 ret = test_range_bit(tree, delalloc_start, delalloc_end,
1712 EXTENT_DELALLOC, 1, cached_state);
1713 if (!ret) {
1714 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1715 &cached_state, GFP_NOFS);
1716 __unlock_for_delalloc(inode, locked_page,
1717 delalloc_start, delalloc_end);
1718 cond_resched();
1719 goto again;
1720 }
1721 free_extent_state(cached_state);
1722 *start = delalloc_start;
1723 *end = delalloc_end;
1724out_failed:
1725 return found;
1726}
1727
1728int extent_clear_unlock_delalloc(struct inode *inode, u64 start, u64 end,
1729 struct page *locked_page,
1730 unsigned long clear_bits,
1731 unsigned long page_ops)
1732{
1733 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1734 int ret;
1735 struct page *pages[16];
1736 unsigned long index = start >> PAGE_CACHE_SHIFT;
1737 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1738 unsigned long nr_pages = end_index - index + 1;
1739 int i;
1740
1741 clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1742 if (page_ops == 0)
1743 return 0;
1744
1745 while (nr_pages > 0) {
1746 ret = find_get_pages_contig(inode->i_mapping, index,
1747 min_t(unsigned long,
1748 nr_pages, ARRAY_SIZE(pages)), pages);
1749 for (i = 0; i < ret; i++) {
1750
1751 if (page_ops & PAGE_SET_PRIVATE2)
1752 SetPagePrivate2(pages[i]);
1753
1754 if (pages[i] == locked_page) {
1755 page_cache_release(pages[i]);
1756 continue;
1757 }
1758 if (page_ops & PAGE_CLEAR_DIRTY)
1759 clear_page_dirty_for_io(pages[i]);
1760 if (page_ops & PAGE_SET_WRITEBACK)
1761 set_page_writeback(pages[i]);
1762 if (page_ops & PAGE_END_WRITEBACK)
1763 end_page_writeback(pages[i]);
1764 if (page_ops & PAGE_UNLOCK)
1765 unlock_page(pages[i]);
1766 page_cache_release(pages[i]);
1767 }
1768 nr_pages -= ret;
1769 index += ret;
1770 cond_resched();
1771 }
1772 return 0;
1773}
1774
1775/*
1776 * count the number of bytes in the tree that have a given bit(s)
1777 * set. This can be fairly slow, except for EXTENT_DIRTY which is
1778 * cached. The total number found is returned.
1779 */
1780u64 count_range_bits(struct extent_io_tree *tree,
1781 u64 *start, u64 search_end, u64 max_bytes,
1782 unsigned long bits, int contig)
1783{
1784 struct rb_node *node;
1785 struct extent_state *state;
1786 u64 cur_start = *start;
1787 u64 total_bytes = 0;
1788 u64 last = 0;
1789 int found = 0;
1790
1791 if (WARN_ON(search_end <= cur_start))
1792 return 0;
1793
1794 spin_lock(&tree->lock);
1795 if (cur_start == 0 && bits == EXTENT_DIRTY) {
1796 total_bytes = tree->dirty_bytes;
1797 goto out;
1798 }
1799 /*
1800 * this search will find all the extents that end after
1801 * our range starts.
1802 */
1803 node = tree_search(tree, cur_start);
1804 if (!node)
1805 goto out;
1806
1807 while (1) {
1808 state = rb_entry(node, struct extent_state, rb_node);
1809 if (state->start > search_end)
1810 break;
1811 if (contig && found && state->start > last + 1)
1812 break;
1813 if (state->end >= cur_start && (state->state & bits) == bits) {
1814 total_bytes += min(search_end, state->end) + 1 -
1815 max(cur_start, state->start);
1816 if (total_bytes >= max_bytes)
1817 break;
1818 if (!found) {
1819 *start = max(cur_start, state->start);
1820 found = 1;
1821 }
1822 last = state->end;
1823 } else if (contig && found) {
1824 break;
1825 }
1826 node = rb_next(node);
1827 if (!node)
1828 break;
1829 }
1830out:
1831 spin_unlock(&tree->lock);
1832 return total_bytes;
1833}
1834
1835/*
1836 * set the private field for a given byte offset in the tree. If there isn't
1837 * an extent_state there already, this does nothing.
1838 */
1839static int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1840{
1841 struct rb_node *node;
1842 struct extent_state *state;
1843 int ret = 0;
1844
1845 spin_lock(&tree->lock);
1846 /*
1847 * this search will find all the extents that end after
1848 * our range starts.
1849 */
1850 node = tree_search(tree, start);
1851 if (!node) {
1852 ret = -ENOENT;
1853 goto out;
1854 }
1855 state = rb_entry(node, struct extent_state, rb_node);
1856 if (state->start != start) {
1857 ret = -ENOENT;
1858 goto out;
1859 }
1860 state->private = private;
1861out:
1862 spin_unlock(&tree->lock);
1863 return ret;
1864}
1865
1866int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1867{
1868 struct rb_node *node;
1869 struct extent_state *state;
1870 int ret = 0;
1871
1872 spin_lock(&tree->lock);
1873 /*
1874 * this search will find all the extents that end after
1875 * our range starts.
1876 */
1877 node = tree_search(tree, start);
1878 if (!node) {
1879 ret = -ENOENT;
1880 goto out;
1881 }
1882 state = rb_entry(node, struct extent_state, rb_node);
1883 if (state->start != start) {
1884 ret = -ENOENT;
1885 goto out;
1886 }
1887 *private = state->private;
1888out:
1889 spin_unlock(&tree->lock);
1890 return ret;
1891}
1892
1893/*
1894 * searches a range in the state tree for a given mask.
1895 * If 'filled' == 1, this returns 1 only if every extent in the tree
1896 * has the bits set. Otherwise, 1 is returned if any bit in the
1897 * range is found set.
1898 */
1899int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1900 unsigned long bits, int filled, struct extent_state *cached)
1901{
1902 struct extent_state *state = NULL;
1903 struct rb_node *node;
1904 int bitset = 0;
1905
1906 spin_lock(&tree->lock);
1907 if (cached && cached->tree && cached->start <= start &&
1908 cached->end > start)
1909 node = &cached->rb_node;
1910 else
1911 node = tree_search(tree, start);
1912 while (node && start <= end) {
1913 state = rb_entry(node, struct extent_state, rb_node);
1914
1915 if (filled && state->start > start) {
1916 bitset = 0;
1917 break;
1918 }
1919
1920 if (state->start > end)
1921 break;
1922
1923 if (state->state & bits) {
1924 bitset = 1;
1925 if (!filled)
1926 break;
1927 } else if (filled) {
1928 bitset = 0;
1929 break;
1930 }
1931
1932 if (state->end == (u64)-1)
1933 break;
1934
1935 start = state->end + 1;
1936 if (start > end)
1937 break;
1938 node = rb_next(node);
1939 if (!node) {
1940 if (filled)
1941 bitset = 0;
1942 break;
1943 }
1944 }
1945 spin_unlock(&tree->lock);
1946 return bitset;
1947}
1948
1949/*
1950 * helper function to set a given page up to date if all the
1951 * extents in the tree for that page are up to date
1952 */
1953static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
1954{
1955 u64 start = page_offset(page);
1956 u64 end = start + PAGE_CACHE_SIZE - 1;
1957 if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1958 SetPageUptodate(page);
1959}
1960
1961/*
1962 * When IO fails, either with EIO or csum verification fails, we
1963 * try other mirrors that might have a good copy of the data. This
1964 * io_failure_record is used to record state as we go through all the
1965 * mirrors. If another mirror has good data, the page is set up to date
1966 * and things continue. If a good mirror can't be found, the original
1967 * bio end_io callback is called to indicate things have failed.
1968 */
1969struct io_failure_record {
1970 struct page *page;
1971 u64 start;
1972 u64 len;
1973 u64 logical;
1974 unsigned long bio_flags;
1975 int this_mirror;
1976 int failed_mirror;
1977 int in_validation;
1978};
1979
1980static int free_io_failure(struct inode *inode, struct io_failure_record *rec,
1981 int did_repair)
1982{
1983 int ret;
1984 int err = 0;
1985 struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1986
1987 set_state_private(failure_tree, rec->start, 0);
1988 ret = clear_extent_bits(failure_tree, rec->start,
1989 rec->start + rec->len - 1,
1990 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1991 if (ret)
1992 err = ret;
1993
1994 ret = clear_extent_bits(&BTRFS_I(inode)->io_tree, rec->start,
1995 rec->start + rec->len - 1,
1996 EXTENT_DAMAGED, GFP_NOFS);
1997 if (ret && !err)
1998 err = ret;
1999
2000 kfree(rec);
2001 return err;
2002}
2003
2004/*
2005 * this bypasses the standard btrfs submit functions deliberately, as
2006 * the standard behavior is to write all copies in a raid setup. here we only
2007 * want to write the one bad copy. so we do the mapping for ourselves and issue
2008 * submit_bio directly.
2009 * to avoid any synchronization issues, wait for the data after writing, which
2010 * actually prevents the read that triggered the error from finishing.
2011 * currently, there can be no more than two copies of every data bit. thus,
2012 * exactly one rewrite is required.
2013 */
2014int repair_io_failure(struct btrfs_fs_info *fs_info, u64 start,
2015 u64 length, u64 logical, struct page *page,
2016 int mirror_num)
2017{
2018 struct bio *bio;
2019 struct btrfs_device *dev;
2020 u64 map_length = 0;
2021 u64 sector;
2022 struct btrfs_bio *bbio = NULL;
2023 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
2024 int ret;
2025
2026 ASSERT(!(fs_info->sb->s_flags & MS_RDONLY));
2027 BUG_ON(!mirror_num);
2028
2029 /* we can't repair anything in raid56 yet */
2030 if (btrfs_is_parity_mirror(map_tree, logical, length, mirror_num))
2031 return 0;
2032
2033 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2034 if (!bio)
2035 return -EIO;
2036 bio->bi_iter.bi_size = 0;
2037 map_length = length;
2038
2039 ret = btrfs_map_block(fs_info, WRITE, logical,
2040 &map_length, &bbio, mirror_num);
2041 if (ret) {
2042 bio_put(bio);
2043 return -EIO;
2044 }
2045 BUG_ON(mirror_num != bbio->mirror_num);
2046 sector = bbio->stripes[mirror_num-1].physical >> 9;
2047 bio->bi_iter.bi_sector = sector;
2048 dev = bbio->stripes[mirror_num-1].dev;
2049 kfree(bbio);
2050 if (!dev || !dev->bdev || !dev->writeable) {
2051 bio_put(bio);
2052 return -EIO;
2053 }
2054 bio->bi_bdev = dev->bdev;
2055 bio_add_page(bio, page, length, start - page_offset(page));
2056
2057 if (btrfsic_submit_bio_wait(WRITE_SYNC, bio)) {
2058 /* try to remap that extent elsewhere? */
2059 bio_put(bio);
2060 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2061 return -EIO;
2062 }
2063
2064 printk_ratelimited_in_rcu(KERN_INFO
2065 "BTRFS: read error corrected: ino %lu off %llu "
2066 "(dev %s sector %llu)\n", page->mapping->host->i_ino,
2067 start, rcu_str_deref(dev->name), sector);
2068
2069 bio_put(bio);
2070 return 0;
2071}
2072
2073int repair_eb_io_failure(struct btrfs_root *root, struct extent_buffer *eb,
2074 int mirror_num)
2075{
2076 u64 start = eb->start;
2077 unsigned long i, num_pages = num_extent_pages(eb->start, eb->len);
2078 int ret = 0;
2079
2080 if (root->fs_info->sb->s_flags & MS_RDONLY)
2081 return -EROFS;
2082
2083 for (i = 0; i < num_pages; i++) {
2084 struct page *p = extent_buffer_page(eb, i);
2085 ret = repair_io_failure(root->fs_info, start, PAGE_CACHE_SIZE,
2086 start, p, mirror_num);
2087 if (ret)
2088 break;
2089 start += PAGE_CACHE_SIZE;
2090 }
2091
2092 return ret;
2093}
2094
2095/*
2096 * each time an IO finishes, we do a fast check in the IO failure tree
2097 * to see if we need to process or clean up an io_failure_record
2098 */
2099static int clean_io_failure(u64 start, struct page *page)
2100{
2101 u64 private;
2102 u64 private_failure;
2103 struct io_failure_record *failrec;
2104 struct inode *inode = page->mapping->host;
2105 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2106 struct extent_state *state;
2107 int num_copies;
2108 int did_repair = 0;
2109 int ret;
2110
2111 private = 0;
2112 ret = count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
2113 (u64)-1, 1, EXTENT_DIRTY, 0);
2114 if (!ret)
2115 return 0;
2116
2117 ret = get_state_private(&BTRFS_I(inode)->io_failure_tree, start,
2118 &private_failure);
2119 if (ret)
2120 return 0;
2121
2122 failrec = (struct io_failure_record *)(unsigned long) private_failure;
2123 BUG_ON(!failrec->this_mirror);
2124
2125 if (failrec->in_validation) {
2126 /* there was no real error, just free the record */
2127 pr_debug("clean_io_failure: freeing dummy error at %llu\n",
2128 failrec->start);
2129 did_repair = 1;
2130 goto out;
2131 }
2132 if (fs_info->sb->s_flags & MS_RDONLY)
2133 goto out;
2134
2135 spin_lock(&BTRFS_I(inode)->io_tree.lock);
2136 state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
2137 failrec->start,
2138 EXTENT_LOCKED);
2139 spin_unlock(&BTRFS_I(inode)->io_tree.lock);
2140
2141 if (state && state->start <= failrec->start &&
2142 state->end >= failrec->start + failrec->len - 1) {
2143 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2144 failrec->len);
2145 if (num_copies > 1) {
2146 ret = repair_io_failure(fs_info, start, failrec->len,
2147 failrec->logical, page,
2148 failrec->failed_mirror);
2149 did_repair = !ret;
2150 }
2151 ret = 0;
2152 }
2153
2154out:
2155 if (!ret)
2156 ret = free_io_failure(inode, failrec, did_repair);
2157
2158 return ret;
2159}
2160
2161/*
2162 * this is a generic handler for readpage errors (default
2163 * readpage_io_failed_hook). if other copies exist, read those and write back
2164 * good data to the failed position. does not investigate in remapping the
2165 * failed extent elsewhere, hoping the device will be smart enough to do this as
2166 * needed
2167 */
2168
2169static int bio_readpage_error(struct bio *failed_bio, u64 phy_offset,
2170 struct page *page, u64 start, u64 end,
2171 int failed_mirror)
2172{
2173 struct io_failure_record *failrec = NULL;
2174 u64 private;
2175 struct extent_map *em;
2176 struct inode *inode = page->mapping->host;
2177 struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2178 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2179 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2180 struct bio *bio;
2181 struct btrfs_io_bio *btrfs_failed_bio;
2182 struct btrfs_io_bio *btrfs_bio;
2183 int num_copies;
2184 int ret;
2185 int read_mode;
2186 u64 logical;
2187
2188 BUG_ON(failed_bio->bi_rw & REQ_WRITE);
2189
2190 ret = get_state_private(failure_tree, start, &private);
2191 if (ret) {
2192 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2193 if (!failrec)
2194 return -ENOMEM;
2195 failrec->start = start;
2196 failrec->len = end - start + 1;
2197 failrec->this_mirror = 0;
2198 failrec->bio_flags = 0;
2199 failrec->in_validation = 0;
2200
2201 read_lock(&em_tree->lock);
2202 em = lookup_extent_mapping(em_tree, start, failrec->len);
2203 if (!em) {
2204 read_unlock(&em_tree->lock);
2205 kfree(failrec);
2206 return -EIO;
2207 }
2208
2209 if (em->start > start || em->start + em->len <= start) {
2210 free_extent_map(em);
2211 em = NULL;
2212 }
2213 read_unlock(&em_tree->lock);
2214
2215 if (!em) {
2216 kfree(failrec);
2217 return -EIO;
2218 }
2219 logical = start - em->start;
2220 logical = em->block_start + logical;
2221 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2222 logical = em->block_start;
2223 failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2224 extent_set_compress_type(&failrec->bio_flags,
2225 em->compress_type);
2226 }
2227 pr_debug("bio_readpage_error: (new) logical=%llu, start=%llu, "
2228 "len=%llu\n", logical, start, failrec->len);
2229 failrec->logical = logical;
2230 free_extent_map(em);
2231
2232 /* set the bits in the private failure tree */
2233 ret = set_extent_bits(failure_tree, start, end,
2234 EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
2235 if (ret >= 0)
2236 ret = set_state_private(failure_tree, start,
2237 (u64)(unsigned long)failrec);
2238 /* set the bits in the inode's tree */
2239 if (ret >= 0)
2240 ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED,
2241 GFP_NOFS);
2242 if (ret < 0) {
2243 kfree(failrec);
2244 return ret;
2245 }
2246 } else {
2247 failrec = (struct io_failure_record *)(unsigned long)private;
2248 pr_debug("bio_readpage_error: (found) logical=%llu, "
2249 "start=%llu, len=%llu, validation=%d\n",
2250 failrec->logical, failrec->start, failrec->len,
2251 failrec->in_validation);
2252 /*
2253 * when data can be on disk more than twice, add to failrec here
2254 * (e.g. with a list for failed_mirror) to make
2255 * clean_io_failure() clean all those errors at once.
2256 */
2257 }
2258 num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
2259 failrec->logical, failrec->len);
2260 if (num_copies == 1) {
2261 /*
2262 * we only have a single copy of the data, so don't bother with
2263 * all the retry and error correction code that follows. no
2264 * matter what the error is, it is very likely to persist.
2265 */
2266 pr_debug("bio_readpage_error: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
2267 num_copies, failrec->this_mirror, failed_mirror);
2268 free_io_failure(inode, failrec, 0);
2269 return -EIO;
2270 }
2271
2272 /*
2273 * there are two premises:
2274 * a) deliver good data to the caller
2275 * b) correct the bad sectors on disk
2276 */
2277 if (failed_bio->bi_vcnt > 1) {
2278 /*
2279 * to fulfill b), we need to know the exact failing sectors, as
2280 * we don't want to rewrite any more than the failed ones. thus,
2281 * we need separate read requests for the failed bio
2282 *
2283 * if the following BUG_ON triggers, our validation request got
2284 * merged. we need separate requests for our algorithm to work.
2285 */
2286 BUG_ON(failrec->in_validation);
2287 failrec->in_validation = 1;
2288 failrec->this_mirror = failed_mirror;
2289 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
2290 } else {
2291 /*
2292 * we're ready to fulfill a) and b) alongside. get a good copy
2293 * of the failed sector and if we succeed, we have setup
2294 * everything for repair_io_failure to do the rest for us.
2295 */
2296 if (failrec->in_validation) {
2297 BUG_ON(failrec->this_mirror != failed_mirror);
2298 failrec->in_validation = 0;
2299 failrec->this_mirror = 0;
2300 }
2301 failrec->failed_mirror = failed_mirror;
2302 failrec->this_mirror++;
2303 if (failrec->this_mirror == failed_mirror)
2304 failrec->this_mirror++;
2305 read_mode = READ_SYNC;
2306 }
2307
2308 if (failrec->this_mirror > num_copies) {
2309 pr_debug("bio_readpage_error: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
2310 num_copies, failrec->this_mirror, failed_mirror);
2311 free_io_failure(inode, failrec, 0);
2312 return -EIO;
2313 }
2314
2315 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
2316 if (!bio) {
2317 free_io_failure(inode, failrec, 0);
2318 return -EIO;
2319 }
2320 bio->bi_end_io = failed_bio->bi_end_io;
2321 bio->bi_iter.bi_sector = failrec->logical >> 9;
2322 bio->bi_bdev = BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev;
2323 bio->bi_iter.bi_size = 0;
2324
2325 btrfs_failed_bio = btrfs_io_bio(failed_bio);
2326 if (btrfs_failed_bio->csum) {
2327 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2328 u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2329
2330 btrfs_bio = btrfs_io_bio(bio);
2331 btrfs_bio->csum = btrfs_bio->csum_inline;
2332 phy_offset >>= inode->i_sb->s_blocksize_bits;
2333 phy_offset *= csum_size;
2334 memcpy(btrfs_bio->csum, btrfs_failed_bio->csum + phy_offset,
2335 csum_size);
2336 }
2337
2338 bio_add_page(bio, page, failrec->len, start - page_offset(page));
2339
2340 pr_debug("bio_readpage_error: submitting new read[%#x] to "
2341 "this_mirror=%d, num_copies=%d, in_validation=%d\n", read_mode,
2342 failrec->this_mirror, num_copies, failrec->in_validation);
2343
2344 ret = tree->ops->submit_bio_hook(inode, read_mode, bio,
2345 failrec->this_mirror,
2346 failrec->bio_flags, 0);
2347 return ret;
2348}
2349
2350/* lots and lots of room for performance fixes in the end_bio funcs */
2351
2352int end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2353{
2354 int uptodate = (err == 0);
2355 struct extent_io_tree *tree;
2356 int ret;
2357
2358 tree = &BTRFS_I(page->mapping->host)->io_tree;
2359
2360 if (tree->ops && tree->ops->writepage_end_io_hook) {
2361 ret = tree->ops->writepage_end_io_hook(page, start,
2362 end, NULL, uptodate);
2363 if (ret)
2364 uptodate = 0;
2365 }
2366
2367 if (!uptodate) {
2368 ClearPageUptodate(page);
2369 SetPageError(page);
2370 }
2371 return 0;
2372}
2373
2374/*
2375 * after a writepage IO is done, we need to:
2376 * clear the uptodate bits on error
2377 * clear the writeback bits in the extent tree for this IO
2378 * end_page_writeback if the page has no more pending IO
2379 *
2380 * Scheduling is not allowed, so the extent state tree is expected
2381 * to have one and only one object corresponding to this IO.
2382 */
2383static void end_bio_extent_writepage(struct bio *bio, int err)
2384{
2385 struct bio_vec *bvec;
2386 u64 start;
2387 u64 end;
2388 int i;
2389
2390 bio_for_each_segment_all(bvec, bio, i) {
2391 struct page *page = bvec->bv_page;
2392
2393 /* We always issue full-page reads, but if some block
2394 * in a page fails to read, blk_update_request() will
2395 * advance bv_offset and adjust bv_len to compensate.
2396 * Print a warning for nonzero offsets, and an error
2397 * if they don't add up to a full page. */
2398 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2399 if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2400 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2401 "partial page write in btrfs with offset %u and length %u",
2402 bvec->bv_offset, bvec->bv_len);
2403 else
2404 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2405 "incomplete page write in btrfs with offset %u and "
2406 "length %u",
2407 bvec->bv_offset, bvec->bv_len);
2408 }
2409
2410 start = page_offset(page);
2411 end = start + bvec->bv_offset + bvec->bv_len - 1;
2412
2413 if (end_extent_writepage(page, err, start, end))
2414 continue;
2415
2416 end_page_writeback(page);
2417 }
2418
2419 bio_put(bio);
2420}
2421
2422static void
2423endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2424 int uptodate)
2425{
2426 struct extent_state *cached = NULL;
2427 u64 end = start + len - 1;
2428
2429 if (uptodate && tree->track_uptodate)
2430 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2431 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
2432}
2433
2434/*
2435 * after a readpage IO is done, we need to:
2436 * clear the uptodate bits on error
2437 * set the uptodate bits if things worked
2438 * set the page up to date if all extents in the tree are uptodate
2439 * clear the lock bit in the extent tree
2440 * unlock the page if there are no other extents locked for it
2441 *
2442 * Scheduling is not allowed, so the extent state tree is expected
2443 * to have one and only one object corresponding to this IO.
2444 */
2445static void end_bio_extent_readpage(struct bio *bio, int err)
2446{
2447 struct bio_vec *bvec;
2448 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
2449 struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2450 struct extent_io_tree *tree;
2451 u64 offset = 0;
2452 u64 start;
2453 u64 end;
2454 u64 len;
2455 u64 extent_start = 0;
2456 u64 extent_len = 0;
2457 int mirror;
2458 int ret;
2459 int i;
2460
2461 if (err)
2462 uptodate = 0;
2463
2464 bio_for_each_segment_all(bvec, bio, i) {
2465 struct page *page = bvec->bv_page;
2466 struct inode *inode = page->mapping->host;
2467
2468 pr_debug("end_bio_extent_readpage: bi_sector=%llu, err=%d, "
2469 "mirror=%lu\n", (u64)bio->bi_iter.bi_sector, err,
2470 io_bio->mirror_num);
2471 tree = &BTRFS_I(inode)->io_tree;
2472
2473 /* We always issue full-page reads, but if some block
2474 * in a page fails to read, blk_update_request() will
2475 * advance bv_offset and adjust bv_len to compensate.
2476 * Print a warning for nonzero offsets, and an error
2477 * if they don't add up to a full page. */
2478 if (bvec->bv_offset || bvec->bv_len != PAGE_CACHE_SIZE) {
2479 if (bvec->bv_offset + bvec->bv_len != PAGE_CACHE_SIZE)
2480 btrfs_err(BTRFS_I(page->mapping->host)->root->fs_info,
2481 "partial page read in btrfs with offset %u and length %u",
2482 bvec->bv_offset, bvec->bv_len);
2483 else
2484 btrfs_info(BTRFS_I(page->mapping->host)->root->fs_info,
2485 "incomplete page read in btrfs with offset %u and "
2486 "length %u",
2487 bvec->bv_offset, bvec->bv_len);
2488 }
2489
2490 start = page_offset(page);
2491 end = start + bvec->bv_offset + bvec->bv_len - 1;
2492 len = bvec->bv_len;
2493
2494 mirror = io_bio->mirror_num;
2495 if (likely(uptodate && tree->ops &&
2496 tree->ops->readpage_end_io_hook)) {
2497 ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2498 page, start, end,
2499 mirror);
2500 if (ret)
2501 uptodate = 0;
2502 else
2503 clean_io_failure(start, page);
2504 }
2505
2506 if (likely(uptodate))
2507 goto readpage_ok;
2508
2509 if (tree->ops && tree->ops->readpage_io_failed_hook) {
2510 ret = tree->ops->readpage_io_failed_hook(page, mirror);
2511 if (!ret && !err &&
2512 test_bit(BIO_UPTODATE, &bio->bi_flags))
2513 uptodate = 1;
2514 } else {
2515 /*
2516 * The generic bio_readpage_error handles errors the
2517 * following way: If possible, new read requests are
2518 * created and submitted and will end up in
2519 * end_bio_extent_readpage as well (if we're lucky, not
2520 * in the !uptodate case). In that case it returns 0 and
2521 * we just go on with the next page in our bio. If it
2522 * can't handle the error it will return -EIO and we
2523 * remain responsible for that page.
2524 */
2525 ret = bio_readpage_error(bio, offset, page, start, end,
2526 mirror);
2527 if (ret == 0) {
2528 uptodate =
2529 test_bit(BIO_UPTODATE, &bio->bi_flags);
2530 if (err)
2531 uptodate = 0;
2532 continue;
2533 }
2534 }
2535readpage_ok:
2536 if (likely(uptodate)) {
2537 loff_t i_size = i_size_read(inode);
2538 pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2539 unsigned offset;
2540
2541 /* Zero out the end if this page straddles i_size */
2542 offset = i_size & (PAGE_CACHE_SIZE-1);
2543 if (page->index == end_index && offset)
2544 zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2545 SetPageUptodate(page);
2546 } else {
2547 ClearPageUptodate(page);
2548 SetPageError(page);
2549 }
2550 unlock_page(page);
2551 offset += len;
2552
2553 if (unlikely(!uptodate)) {
2554 if (extent_len) {
2555 endio_readpage_release_extent(tree,
2556 extent_start,
2557 extent_len, 1);
2558 extent_start = 0;
2559 extent_len = 0;
2560 }
2561 endio_readpage_release_extent(tree, start,
2562 end - start + 1, 0);
2563 } else if (!extent_len) {
2564 extent_start = start;
2565 extent_len = end + 1 - start;
2566 } else if (extent_start + extent_len == start) {
2567 extent_len += end + 1 - start;
2568 } else {
2569 endio_readpage_release_extent(tree, extent_start,
2570 extent_len, uptodate);
2571 extent_start = start;
2572 extent_len = end + 1 - start;
2573 }
2574 }
2575
2576 if (extent_len)
2577 endio_readpage_release_extent(tree, extent_start, extent_len,
2578 uptodate);
2579 if (io_bio->end_io)
2580 io_bio->end_io(io_bio, err);
2581 bio_put(bio);
2582}
2583
2584/*
2585 * this allocates from the btrfs_bioset. We're returning a bio right now
2586 * but you can call btrfs_io_bio for the appropriate container_of magic
2587 */
2588struct bio *
2589btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
2590 gfp_t gfp_flags)
2591{
2592 struct btrfs_io_bio *btrfs_bio;
2593 struct bio *bio;
2594
2595 bio = bio_alloc_bioset(gfp_flags, nr_vecs, btrfs_bioset);
2596
2597 if (bio == NULL && (current->flags & PF_MEMALLOC)) {
2598 while (!bio && (nr_vecs /= 2)) {
2599 bio = bio_alloc_bioset(gfp_flags,
2600 nr_vecs, btrfs_bioset);
2601 }
2602 }
2603
2604 if (bio) {
2605 bio->bi_bdev = bdev;
2606 bio->bi_iter.bi_sector = first_sector;
2607 btrfs_bio = btrfs_io_bio(bio);
2608 btrfs_bio->csum = NULL;
2609 btrfs_bio->csum_allocated = NULL;
2610 btrfs_bio->end_io = NULL;
2611 }
2612 return bio;
2613}
2614
2615struct bio *btrfs_bio_clone(struct bio *bio, gfp_t gfp_mask)
2616{
2617 return bio_clone_bioset(bio, gfp_mask, btrfs_bioset);
2618}
2619
2620
2621/* this also allocates from the btrfs_bioset */
2622struct bio *btrfs_io_bio_alloc(gfp_t gfp_mask, unsigned int nr_iovecs)
2623{
2624 struct btrfs_io_bio *btrfs_bio;
2625 struct bio *bio;
2626
2627 bio = bio_alloc_bioset(gfp_mask, nr_iovecs, btrfs_bioset);
2628 if (bio) {
2629 btrfs_bio = btrfs_io_bio(bio);
2630 btrfs_bio->csum = NULL;
2631 btrfs_bio->csum_allocated = NULL;
2632 btrfs_bio->end_io = NULL;
2633 }
2634 return bio;
2635}
2636
2637
2638static int __must_check submit_one_bio(int rw, struct bio *bio,
2639 int mirror_num, unsigned long bio_flags)
2640{
2641 int ret = 0;
2642 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
2643 struct page *page = bvec->bv_page;
2644 struct extent_io_tree *tree = bio->bi_private;
2645 u64 start;
2646
2647 start = page_offset(page) + bvec->bv_offset;
2648
2649 bio->bi_private = NULL;
2650
2651 bio_get(bio);
2652
2653 if (tree->ops && tree->ops->submit_bio_hook)
2654 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
2655 mirror_num, bio_flags, start);
2656 else
2657 btrfsic_submit_bio(rw, bio);
2658
2659 if (bio_flagged(bio, BIO_EOPNOTSUPP))
2660 ret = -EOPNOTSUPP;
2661 bio_put(bio);
2662 return ret;
2663}
2664
2665static int merge_bio(int rw, struct extent_io_tree *tree, struct page *page,
2666 unsigned long offset, size_t size, struct bio *bio,
2667 unsigned long bio_flags)
2668{
2669 int ret = 0;
2670 if (tree->ops && tree->ops->merge_bio_hook)
2671 ret = tree->ops->merge_bio_hook(rw, page, offset, size, bio,
2672 bio_flags);
2673 BUG_ON(ret < 0);
2674 return ret;
2675
2676}
2677
2678static int submit_extent_page(int rw, struct extent_io_tree *tree,
2679 struct page *page, sector_t sector,
2680 size_t size, unsigned long offset,
2681 struct block_device *bdev,
2682 struct bio **bio_ret,
2683 unsigned long max_pages,
2684 bio_end_io_t end_io_func,
2685 int mirror_num,
2686 unsigned long prev_bio_flags,
2687 unsigned long bio_flags)
2688{
2689 int ret = 0;
2690 struct bio *bio;
2691 int nr;
2692 int contig = 0;
2693 int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
2694 int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
2695 size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
2696
2697 if (bio_ret && *bio_ret) {
2698 bio = *bio_ret;
2699 if (old_compressed)
2700 contig = bio->bi_iter.bi_sector == sector;
2701 else
2702 contig = bio_end_sector(bio) == sector;
2703
2704 if (prev_bio_flags != bio_flags || !contig ||
2705 merge_bio(rw, tree, page, offset, page_size, bio, bio_flags) ||
2706 bio_add_page(bio, page, page_size, offset) < page_size) {
2707 ret = submit_one_bio(rw, bio, mirror_num,
2708 prev_bio_flags);
2709 if (ret < 0)
2710 return ret;
2711 bio = NULL;
2712 } else {
2713 return 0;
2714 }
2715 }
2716 if (this_compressed)
2717 nr = BIO_MAX_PAGES;
2718 else
2719 nr = bio_get_nr_vecs(bdev);
2720
2721 bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
2722 if (!bio)
2723 return -ENOMEM;
2724
2725 bio_add_page(bio, page, page_size, offset);
2726 bio->bi_end_io = end_io_func;
2727 bio->bi_private = tree;
2728
2729 if (bio_ret)
2730 *bio_ret = bio;
2731 else
2732 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
2733
2734 return ret;
2735}
2736
2737static void attach_extent_buffer_page(struct extent_buffer *eb,
2738 struct page *page)
2739{
2740 if (!PagePrivate(page)) {
2741 SetPagePrivate(page);
2742 page_cache_get(page);
2743 set_page_private(page, (unsigned long)eb);
2744 } else {
2745 WARN_ON(page->private != (unsigned long)eb);
2746 }
2747}
2748
2749void set_page_extent_mapped(struct page *page)
2750{
2751 if (!PagePrivate(page)) {
2752 SetPagePrivate(page);
2753 page_cache_get(page);
2754 set_page_private(page, EXTENT_PAGE_PRIVATE);
2755 }
2756}
2757
2758static struct extent_map *
2759__get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
2760 u64 start, u64 len, get_extent_t *get_extent,
2761 struct extent_map **em_cached)
2762{
2763 struct extent_map *em;
2764
2765 if (em_cached && *em_cached) {
2766 em = *em_cached;
2767 if (extent_map_in_tree(em) && start >= em->start &&
2768 start < extent_map_end(em)) {
2769 atomic_inc(&em->refs);
2770 return em;
2771 }
2772
2773 free_extent_map(em);
2774 *em_cached = NULL;
2775 }
2776
2777 em = get_extent(inode, page, pg_offset, start, len, 0);
2778 if (em_cached && !IS_ERR_OR_NULL(em)) {
2779 BUG_ON(*em_cached);
2780 atomic_inc(&em->refs);
2781 *em_cached = em;
2782 }
2783 return em;
2784}
2785/*
2786 * basic readpage implementation. Locked extent state structs are inserted
2787 * into the tree that are removed when the IO is done (by the end_io
2788 * handlers)
2789 * XXX JDM: This needs looking at to ensure proper page locking
2790 */
2791static int __do_readpage(struct extent_io_tree *tree,
2792 struct page *page,
2793 get_extent_t *get_extent,
2794 struct extent_map **em_cached,
2795 struct bio **bio, int mirror_num,
2796 unsigned long *bio_flags, int rw)
2797{
2798 struct inode *inode = page->mapping->host;
2799 u64 start = page_offset(page);
2800 u64 page_end = start + PAGE_CACHE_SIZE - 1;
2801 u64 end;
2802 u64 cur = start;
2803 u64 extent_offset;
2804 u64 last_byte = i_size_read(inode);
2805 u64 block_start;
2806 u64 cur_end;
2807 sector_t sector;
2808 struct extent_map *em;
2809 struct block_device *bdev;
2810 int ret;
2811 int nr = 0;
2812 int parent_locked = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2813 size_t pg_offset = 0;
2814 size_t iosize;
2815 size_t disk_io_size;
2816 size_t blocksize = inode->i_sb->s_blocksize;
2817 unsigned long this_bio_flag = *bio_flags & EXTENT_BIO_PARENT_LOCKED;
2818
2819 set_page_extent_mapped(page);
2820
2821 end = page_end;
2822 if (!PageUptodate(page)) {
2823 if (cleancache_get_page(page) == 0) {
2824 BUG_ON(blocksize != PAGE_SIZE);
2825 unlock_extent(tree, start, end);
2826 goto out;
2827 }
2828 }
2829
2830 if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
2831 char *userpage;
2832 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
2833
2834 if (zero_offset) {
2835 iosize = PAGE_CACHE_SIZE - zero_offset;
2836 userpage = kmap_atomic(page);
2837 memset(userpage + zero_offset, 0, iosize);
2838 flush_dcache_page(page);
2839 kunmap_atomic(userpage);
2840 }
2841 }
2842 while (cur <= end) {
2843 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2844
2845 if (cur >= last_byte) {
2846 char *userpage;
2847 struct extent_state *cached = NULL;
2848
2849 iosize = PAGE_CACHE_SIZE - pg_offset;
2850 userpage = kmap_atomic(page);
2851 memset(userpage + pg_offset, 0, iosize);
2852 flush_dcache_page(page);
2853 kunmap_atomic(userpage);
2854 set_extent_uptodate(tree, cur, cur + iosize - 1,
2855 &cached, GFP_NOFS);
2856 if (!parent_locked)
2857 unlock_extent_cached(tree, cur,
2858 cur + iosize - 1,
2859 &cached, GFP_NOFS);
2860 break;
2861 }
2862 em = __get_extent_map(inode, page, pg_offset, cur,
2863 end - cur + 1, get_extent, em_cached);
2864 if (IS_ERR_OR_NULL(em)) {
2865 SetPageError(page);
2866 if (!parent_locked)
2867 unlock_extent(tree, cur, end);
2868 break;
2869 }
2870 extent_offset = cur - em->start;
2871 BUG_ON(extent_map_end(em) <= cur);
2872 BUG_ON(end < cur);
2873
2874 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2875 this_bio_flag |= EXTENT_BIO_COMPRESSED;
2876 extent_set_compress_type(&this_bio_flag,
2877 em->compress_type);
2878 }
2879
2880 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2881 cur_end = min(extent_map_end(em) - 1, end);
2882 iosize = ALIGN(iosize, blocksize);
2883 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
2884 disk_io_size = em->block_len;
2885 sector = em->block_start >> 9;
2886 } else {
2887 sector = (em->block_start + extent_offset) >> 9;
2888 disk_io_size = iosize;
2889 }
2890 bdev = em->bdev;
2891 block_start = em->block_start;
2892 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2893 block_start = EXTENT_MAP_HOLE;
2894 free_extent_map(em);
2895 em = NULL;
2896
2897 /* we've found a hole, just zero and go on */
2898 if (block_start == EXTENT_MAP_HOLE) {
2899 char *userpage;
2900 struct extent_state *cached = NULL;
2901
2902 userpage = kmap_atomic(page);
2903 memset(userpage + pg_offset, 0, iosize);
2904 flush_dcache_page(page);
2905 kunmap_atomic(userpage);
2906
2907 set_extent_uptodate(tree, cur, cur + iosize - 1,
2908 &cached, GFP_NOFS);
2909 unlock_extent_cached(tree, cur, cur + iosize - 1,
2910 &cached, GFP_NOFS);
2911 cur = cur + iosize;
2912 pg_offset += iosize;
2913 continue;
2914 }
2915 /* the get_extent function already copied into the page */
2916 if (test_range_bit(tree, cur, cur_end,
2917 EXTENT_UPTODATE, 1, NULL)) {
2918 check_page_uptodate(tree, page);
2919 if (!parent_locked)
2920 unlock_extent(tree, cur, cur + iosize - 1);
2921 cur = cur + iosize;
2922 pg_offset += iosize;
2923 continue;
2924 }
2925 /* we have an inline extent but it didn't get marked up
2926 * to date. Error out
2927 */
2928 if (block_start == EXTENT_MAP_INLINE) {
2929 SetPageError(page);
2930 if (!parent_locked)
2931 unlock_extent(tree, cur, cur + iosize - 1);
2932 cur = cur + iosize;
2933 pg_offset += iosize;
2934 continue;
2935 }
2936
2937 pnr -= page->index;
2938 ret = submit_extent_page(rw, tree, page,
2939 sector, disk_io_size, pg_offset,
2940 bdev, bio, pnr,
2941 end_bio_extent_readpage, mirror_num,
2942 *bio_flags,
2943 this_bio_flag);
2944 if (!ret) {
2945 nr++;
2946 *bio_flags = this_bio_flag;
2947 } else {
2948 SetPageError(page);
2949 if (!parent_locked)
2950 unlock_extent(tree, cur, cur + iosize - 1);
2951 }
2952 cur = cur + iosize;
2953 pg_offset += iosize;
2954 }
2955out:
2956 if (!nr) {
2957 if (!PageError(page))
2958 SetPageUptodate(page);
2959 unlock_page(page);
2960 }
2961 return 0;
2962}
2963
2964static inline void __do_contiguous_readpages(struct extent_io_tree *tree,
2965 struct page *pages[], int nr_pages,
2966 u64 start, u64 end,
2967 get_extent_t *get_extent,
2968 struct extent_map **em_cached,
2969 struct bio **bio, int mirror_num,
2970 unsigned long *bio_flags, int rw)
2971{
2972 struct inode *inode;
2973 struct btrfs_ordered_extent *ordered;
2974 int index;
2975
2976 inode = pages[0]->mapping->host;
2977 while (1) {
2978 lock_extent(tree, start, end);
2979 ordered = btrfs_lookup_ordered_range(inode, start,
2980 end - start + 1);
2981 if (!ordered)
2982 break;
2983 unlock_extent(tree, start, end);
2984 btrfs_start_ordered_extent(inode, ordered, 1);
2985 btrfs_put_ordered_extent(ordered);
2986 }
2987
2988 for (index = 0; index < nr_pages; index++) {
2989 __do_readpage(tree, pages[index], get_extent, em_cached, bio,
2990 mirror_num, bio_flags, rw);
2991 page_cache_release(pages[index]);
2992 }
2993}
2994
2995static void __extent_readpages(struct extent_io_tree *tree,
2996 struct page *pages[],
2997 int nr_pages, get_extent_t *get_extent,
2998 struct extent_map **em_cached,
2999 struct bio **bio, int mirror_num,
3000 unsigned long *bio_flags, int rw)
3001{
3002 u64 start = 0;
3003 u64 end = 0;
3004 u64 page_start;
3005 int index;
3006 int first_index = 0;
3007
3008 for (index = 0; index < nr_pages; index++) {
3009 page_start = page_offset(pages[index]);
3010 if (!end) {
3011 start = page_start;
3012 end = start + PAGE_CACHE_SIZE - 1;
3013 first_index = index;
3014 } else if (end + 1 == page_start) {
3015 end += PAGE_CACHE_SIZE;
3016 } else {
3017 __do_contiguous_readpages(tree, &pages[first_index],
3018 index - first_index, start,
3019 end, get_extent, em_cached,
3020 bio, mirror_num, bio_flags,
3021 rw);
3022 start = page_start;
3023 end = start + PAGE_CACHE_SIZE - 1;
3024 first_index = index;
3025 }
3026 }
3027
3028 if (end)
3029 __do_contiguous_readpages(tree, &pages[first_index],
3030 index - first_index, start,
3031 end, get_extent, em_cached, bio,
3032 mirror_num, bio_flags, rw);
3033}
3034
3035static int __extent_read_full_page(struct extent_io_tree *tree,
3036 struct page *page,
3037 get_extent_t *get_extent,
3038 struct bio **bio, int mirror_num,
3039 unsigned long *bio_flags, int rw)
3040{
3041 struct inode *inode = page->mapping->host;
3042 struct btrfs_ordered_extent *ordered;
3043 u64 start = page_offset(page);
3044 u64 end = start + PAGE_CACHE_SIZE - 1;
3045 int ret;
3046
3047 while (1) {
3048 lock_extent(tree, start, end);
3049 ordered = btrfs_lookup_ordered_extent(inode, start);
3050 if (!ordered)
3051 break;
3052 unlock_extent(tree, start, end);
3053 btrfs_start_ordered_extent(inode, ordered, 1);
3054 btrfs_put_ordered_extent(ordered);
3055 }
3056
3057 ret = __do_readpage(tree, page, get_extent, NULL, bio, mirror_num,
3058 bio_flags, rw);
3059 return ret;
3060}
3061
3062int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
3063 get_extent_t *get_extent, int mirror_num)
3064{
3065 struct bio *bio = NULL;
3066 unsigned long bio_flags = 0;
3067 int ret;
3068
3069 ret = __extent_read_full_page(tree, page, get_extent, &bio, mirror_num,
3070 &bio_flags, READ);
3071 if (bio)
3072 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3073 return ret;
3074}
3075
3076int extent_read_full_page_nolock(struct extent_io_tree *tree, struct page *page,
3077 get_extent_t *get_extent, int mirror_num)
3078{
3079 struct bio *bio = NULL;
3080 unsigned long bio_flags = EXTENT_BIO_PARENT_LOCKED;
3081 int ret;
3082
3083 ret = __do_readpage(tree, page, get_extent, NULL, &bio, mirror_num,
3084 &bio_flags, READ);
3085 if (bio)
3086 ret = submit_one_bio(READ, bio, mirror_num, bio_flags);
3087 return ret;
3088}
3089
3090static noinline void update_nr_written(struct page *page,
3091 struct writeback_control *wbc,
3092 unsigned long nr_written)
3093{
3094 wbc->nr_to_write -= nr_written;
3095 if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
3096 wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
3097 page->mapping->writeback_index = page->index + nr_written;
3098}
3099
3100/*
3101 * the writepage semantics are similar to regular writepage. extent
3102 * records are inserted to lock ranges in the tree, and as dirty areas
3103 * are found, they are marked writeback. Then the lock bits are removed
3104 * and the end_io handler clears the writeback ranges
3105 */
3106static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3107 void *data)
3108{
3109 struct inode *inode = page->mapping->host;
3110 struct extent_page_data *epd = data;
3111 struct extent_io_tree *tree = epd->tree;
3112 u64 start = page_offset(page);
3113 u64 delalloc_start;
3114 u64 page_end = start + PAGE_CACHE_SIZE - 1;
3115 u64 end;
3116 u64 cur = start;
3117 u64 extent_offset;
3118 u64 last_byte = i_size_read(inode);
3119 u64 block_start;
3120 u64 iosize;
3121 sector_t sector;
3122 struct extent_state *cached_state = NULL;
3123 struct extent_map *em;
3124 struct block_device *bdev;
3125 int ret;
3126 int nr = 0;
3127 size_t pg_offset = 0;
3128 size_t blocksize;
3129 loff_t i_size = i_size_read(inode);
3130 unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
3131 u64 nr_delalloc;
3132 u64 delalloc_end;
3133 int page_started;
3134 int compressed;
3135 int write_flags;
3136 unsigned long nr_written = 0;
3137 bool fill_delalloc = true;
3138
3139 if (wbc->sync_mode == WB_SYNC_ALL)
3140 write_flags = WRITE_SYNC;
3141 else
3142 write_flags = WRITE;
3143
3144 trace___extent_writepage(page, inode, wbc);
3145
3146 WARN_ON(!PageLocked(page));
3147
3148 ClearPageError(page);
3149
3150 pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
3151 if (page->index > end_index ||
3152 (page->index == end_index && !pg_offset)) {
3153 page->mapping->a_ops->invalidatepage(page, 0, PAGE_CACHE_SIZE);
3154 unlock_page(page);
3155 return 0;
3156 }
3157
3158 if (page->index == end_index) {
3159 char *userpage;
3160
3161 userpage = kmap_atomic(page);
3162 memset(userpage + pg_offset, 0,
3163 PAGE_CACHE_SIZE - pg_offset);
3164 kunmap_atomic(userpage);
3165 flush_dcache_page(page);
3166 }
3167 pg_offset = 0;
3168
3169 set_page_extent_mapped(page);
3170
3171 if (!tree->ops || !tree->ops->fill_delalloc)
3172 fill_delalloc = false;
3173
3174 delalloc_start = start;
3175 delalloc_end = 0;
3176 page_started = 0;
3177 if (!epd->extent_locked && fill_delalloc) {
3178 u64 delalloc_to_write = 0;
3179 /*
3180 * make sure the wbc mapping index is at least updated
3181 * to this page.
3182 */
3183 update_nr_written(page, wbc, 0);
3184
3185 while (delalloc_end < page_end) {
3186 nr_delalloc = find_lock_delalloc_range(inode, tree,
3187 page,
3188 &delalloc_start,
3189 &delalloc_end,
3190 128 * 1024 * 1024);
3191 if (nr_delalloc == 0) {
3192 delalloc_start = delalloc_end + 1;
3193 continue;
3194 }
3195 ret = tree->ops->fill_delalloc(inode, page,
3196 delalloc_start,
3197 delalloc_end,
3198 &page_started,
3199 &nr_written);
3200 /* File system has been set read-only */
3201 if (ret) {
3202 SetPageError(page);
3203 goto done;
3204 }
3205 /*
3206 * delalloc_end is already one less than the total
3207 * length, so we don't subtract one from
3208 * PAGE_CACHE_SIZE
3209 */
3210 delalloc_to_write += (delalloc_end - delalloc_start +
3211 PAGE_CACHE_SIZE) >>
3212 PAGE_CACHE_SHIFT;
3213 delalloc_start = delalloc_end + 1;
3214 }
3215 if (wbc->nr_to_write < delalloc_to_write) {
3216 int thresh = 8192;
3217
3218 if (delalloc_to_write < thresh * 2)
3219 thresh = delalloc_to_write;
3220 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3221 thresh);
3222 }
3223
3224 /* did the fill delalloc function already unlock and start
3225 * the IO?
3226 */
3227 if (page_started) {
3228 ret = 0;
3229 /*
3230 * we've unlocked the page, so we can't update
3231 * the mapping's writeback index, just update
3232 * nr_to_write.
3233 */
3234 wbc->nr_to_write -= nr_written;
3235 goto done_unlocked;
3236 }
3237 }
3238 if (tree->ops && tree->ops->writepage_start_hook) {
3239 ret = tree->ops->writepage_start_hook(page, start,
3240 page_end);
3241 if (ret) {
3242 /* Fixup worker will requeue */
3243 if (ret == -EBUSY)
3244 wbc->pages_skipped++;
3245 else
3246 redirty_page_for_writepage(wbc, page);
3247 update_nr_written(page, wbc, nr_written);
3248 unlock_page(page);
3249 ret = 0;
3250 goto done_unlocked;
3251 }
3252 }
3253
3254 /*
3255 * we don't want to touch the inode after unlocking the page,
3256 * so we update the mapping writeback index now
3257 */
3258 update_nr_written(page, wbc, nr_written + 1);
3259
3260 end = page_end;
3261 if (last_byte <= start) {
3262 if (tree->ops && tree->ops->writepage_end_io_hook)
3263 tree->ops->writepage_end_io_hook(page, start,
3264 page_end, NULL, 1);
3265 goto done;
3266 }
3267
3268 blocksize = inode->i_sb->s_blocksize;
3269
3270 while (cur <= end) {
3271 if (cur >= last_byte) {
3272 if (tree->ops && tree->ops->writepage_end_io_hook)
3273 tree->ops->writepage_end_io_hook(page, cur,
3274 page_end, NULL, 1);
3275 break;
3276 }
3277 em = epd->get_extent(inode, page, pg_offset, cur,
3278 end - cur + 1, 1);
3279 if (IS_ERR_OR_NULL(em)) {
3280 SetPageError(page);
3281 break;
3282 }
3283
3284 extent_offset = cur - em->start;
3285 BUG_ON(extent_map_end(em) <= cur);
3286 BUG_ON(end < cur);
3287 iosize = min(extent_map_end(em) - cur, end - cur + 1);
3288 iosize = ALIGN(iosize, blocksize);
3289 sector = (em->block_start + extent_offset) >> 9;
3290 bdev = em->bdev;
3291 block_start = em->block_start;
3292 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3293 free_extent_map(em);
3294 em = NULL;
3295
3296 /*
3297 * compressed and inline extents are written through other
3298 * paths in the FS
3299 */
3300 if (compressed || block_start == EXTENT_MAP_HOLE ||
3301 block_start == EXTENT_MAP_INLINE) {
3302 /*
3303 * end_io notification does not happen here for
3304 * compressed extents
3305 */
3306 if (!compressed && tree->ops &&
3307 tree->ops->writepage_end_io_hook)
3308 tree->ops->writepage_end_io_hook(page, cur,
3309 cur + iosize - 1,
3310 NULL, 1);
3311 else if (compressed) {
3312 /* we don't want to end_page_writeback on
3313 * a compressed extent. this happens
3314 * elsewhere
3315 */
3316 nr++;
3317 }
3318
3319 cur += iosize;
3320 pg_offset += iosize;
3321 continue;
3322 }
3323 /* leave this out until we have a page_mkwrite call */
3324 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
3325 EXTENT_DIRTY, 0, NULL)) {
3326 cur = cur + iosize;
3327 pg_offset += iosize;
3328 continue;
3329 }
3330
3331 if (tree->ops && tree->ops->writepage_io_hook) {
3332 ret = tree->ops->writepage_io_hook(page, cur,
3333 cur + iosize - 1);
3334 } else {
3335 ret = 0;
3336 }
3337 if (ret) {
3338 SetPageError(page);
3339 } else {
3340 unsigned long max_nr = end_index + 1;
3341
3342 set_range_writeback(tree, cur, cur + iosize - 1);
3343 if (!PageWriteback(page)) {
3344 btrfs_err(BTRFS_I(inode)->root->fs_info,
3345 "page %lu not writeback, cur %llu end %llu",
3346 page->index, cur, end);
3347 }
3348
3349 ret = submit_extent_page(write_flags, tree, page,
3350 sector, iosize, pg_offset,
3351 bdev, &epd->bio, max_nr,
3352 end_bio_extent_writepage,
3353 0, 0, 0);
3354 if (ret)
3355 SetPageError(page);
3356 }
3357 cur = cur + iosize;
3358 pg_offset += iosize;
3359 nr++;
3360 }
3361done:
3362 if (nr == 0) {
3363 /* make sure the mapping tag for page dirty gets cleared */
3364 set_page_writeback(page);
3365 end_page_writeback(page);
3366 }
3367 unlock_page(page);
3368
3369done_unlocked:
3370
3371 /* drop our reference on any cached states */
3372 free_extent_state(cached_state);
3373 return 0;
3374}
3375
3376static int eb_wait(void *word)
3377{
3378 io_schedule();
3379 return 0;
3380}
3381
3382void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3383{
3384 wait_on_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK, eb_wait,
3385 TASK_UNINTERRUPTIBLE);
3386}
3387
3388static int lock_extent_buffer_for_io(struct extent_buffer *eb,
3389 struct btrfs_fs_info *fs_info,
3390 struct extent_page_data *epd)
3391{
3392 unsigned long i, num_pages;
3393 int flush = 0;
3394 int ret = 0;
3395
3396 if (!btrfs_try_tree_write_lock(eb)) {
3397 flush = 1;
3398 flush_write_bio(epd);
3399 btrfs_tree_lock(eb);
3400 }
3401
3402 if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3403 btrfs_tree_unlock(eb);
3404 if (!epd->sync_io)
3405 return 0;
3406 if (!flush) {
3407 flush_write_bio(epd);
3408 flush = 1;
3409 }
3410 while (1) {
3411 wait_on_extent_buffer_writeback(eb);
3412 btrfs_tree_lock(eb);
3413 if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3414 break;
3415 btrfs_tree_unlock(eb);
3416 }
3417 }
3418
3419 /*
3420 * We need to do this to prevent races in people who check if the eb is
3421 * under IO since we can end up having no IO bits set for a short period
3422 * of time.
3423 */
3424 spin_lock(&eb->refs_lock);
3425 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3426 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3427 spin_unlock(&eb->refs_lock);
3428 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3429 __percpu_counter_add(&fs_info->dirty_metadata_bytes,
3430 -eb->len,
3431 fs_info->dirty_metadata_batch);
3432 ret = 1;
3433 } else {
3434 spin_unlock(&eb->refs_lock);
3435 }
3436
3437 btrfs_tree_unlock(eb);
3438
3439 if (!ret)
3440 return ret;
3441
3442 num_pages = num_extent_pages(eb->start, eb->len);
3443 for (i = 0; i < num_pages; i++) {
3444 struct page *p = extent_buffer_page(eb, i);
3445
3446 if (!trylock_page(p)) {
3447 if (!flush) {
3448 flush_write_bio(epd);
3449 flush = 1;
3450 }
3451 lock_page(p);
3452 }
3453 }
3454
3455 return ret;
3456}
3457
3458static void end_extent_buffer_writeback(struct extent_buffer *eb)
3459{
3460 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3461 smp_mb__after_clear_bit();
3462 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3463}
3464
3465static void end_bio_extent_buffer_writepage(struct bio *bio, int err)
3466{
3467 struct bio_vec *bvec;
3468 struct extent_buffer *eb;
3469 int i, done;
3470
3471 bio_for_each_segment_all(bvec, bio, i) {
3472 struct page *page = bvec->bv_page;
3473
3474 eb = (struct extent_buffer *)page->private;
3475 BUG_ON(!eb);
3476 done = atomic_dec_and_test(&eb->io_pages);
3477
3478 if (err || test_bit(EXTENT_BUFFER_IOERR, &eb->bflags)) {
3479 set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3480 ClearPageUptodate(page);
3481 SetPageError(page);
3482 }
3483
3484 end_page_writeback(page);
3485
3486 if (!done)
3487 continue;
3488
3489 end_extent_buffer_writeback(eb);
3490 }
3491
3492 bio_put(bio);
3493}
3494
3495static int write_one_eb(struct extent_buffer *eb,
3496 struct btrfs_fs_info *fs_info,
3497 struct writeback_control *wbc,
3498 struct extent_page_data *epd)
3499{
3500 struct block_device *bdev = fs_info->fs_devices->latest_bdev;
3501 struct extent_io_tree *tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
3502 u64 offset = eb->start;
3503 unsigned long i, num_pages;
3504 unsigned long bio_flags = 0;
3505 int rw = (epd->sync_io ? WRITE_SYNC : WRITE) | REQ_META;
3506 int ret = 0;
3507
3508 clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3509 num_pages = num_extent_pages(eb->start, eb->len);
3510 atomic_set(&eb->io_pages, num_pages);
3511 if (btrfs_header_owner(eb) == BTRFS_TREE_LOG_OBJECTID)
3512 bio_flags = EXTENT_BIO_TREE_LOG;
3513
3514 for (i = 0; i < num_pages; i++) {
3515 struct page *p = extent_buffer_page(eb, i);
3516
3517 clear_page_dirty_for_io(p);
3518 set_page_writeback(p);
3519 ret = submit_extent_page(rw, tree, p, offset >> 9,
3520 PAGE_CACHE_SIZE, 0, bdev, &epd->bio,
3521 -1, end_bio_extent_buffer_writepage,
3522 0, epd->bio_flags, bio_flags);
3523 epd->bio_flags = bio_flags;
3524 if (ret) {
3525 set_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
3526 SetPageError(p);
3527 if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3528 end_extent_buffer_writeback(eb);
3529 ret = -EIO;
3530 break;
3531 }
3532 offset += PAGE_CACHE_SIZE;
3533 update_nr_written(p, wbc, 1);
3534 unlock_page(p);
3535 }
3536
3537 if (unlikely(ret)) {
3538 for (; i < num_pages; i++) {
3539 struct page *p = extent_buffer_page(eb, i);
3540 unlock_page(p);
3541 }
3542 }
3543
3544 return ret;
3545}
3546
3547int btree_write_cache_pages(struct address_space *mapping,
3548 struct writeback_control *wbc)
3549{
3550 struct extent_io_tree *tree = &BTRFS_I(mapping->host)->io_tree;
3551 struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3552 struct extent_buffer *eb, *prev_eb = NULL;
3553 struct extent_page_data epd = {
3554 .bio = NULL,
3555 .tree = tree,
3556 .extent_locked = 0,
3557 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3558 .bio_flags = 0,
3559 };
3560 int ret = 0;
3561 int done = 0;
3562 int nr_to_write_done = 0;
3563 struct pagevec pvec;
3564 int nr_pages;
3565 pgoff_t index;
3566 pgoff_t end; /* Inclusive */
3567 int scanned = 0;
3568 int tag;
3569
3570 pagevec_init(&pvec, 0);
3571 if (wbc->range_cyclic) {
3572 index = mapping->writeback_index; /* Start from prev offset */
3573 end = -1;
3574 } else {
3575 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3576 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3577 scanned = 1;
3578 }
3579 if (wbc->sync_mode == WB_SYNC_ALL)
3580 tag = PAGECACHE_TAG_TOWRITE;
3581 else
3582 tag = PAGECACHE_TAG_DIRTY;
3583retry:
3584 if (wbc->sync_mode == WB_SYNC_ALL)
3585 tag_pages_for_writeback(mapping, index, end);
3586 while (!done && !nr_to_write_done && (index <= end) &&
3587 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3588 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3589 unsigned i;
3590
3591 scanned = 1;
3592 for (i = 0; i < nr_pages; i++) {
3593 struct page *page = pvec.pages[i];
3594
3595 if (!PagePrivate(page))
3596 continue;
3597
3598 if (!wbc->range_cyclic && page->index > end) {
3599 done = 1;
3600 break;
3601 }
3602
3603 spin_lock(&mapping->private_lock);
3604 if (!PagePrivate(page)) {
3605 spin_unlock(&mapping->private_lock);
3606 continue;
3607 }
3608
3609 eb = (struct extent_buffer *)page->private;
3610
3611 /*
3612 * Shouldn't happen and normally this would be a BUG_ON
3613 * but no sense in crashing the users box for something
3614 * we can survive anyway.
3615 */
3616 if (WARN_ON(!eb)) {
3617 spin_unlock(&mapping->private_lock);
3618 continue;
3619 }
3620
3621 if (eb == prev_eb) {
3622 spin_unlock(&mapping->private_lock);
3623 continue;
3624 }
3625
3626 ret = atomic_inc_not_zero(&eb->refs);
3627 spin_unlock(&mapping->private_lock);
3628 if (!ret)
3629 continue;
3630
3631 prev_eb = eb;
3632 ret = lock_extent_buffer_for_io(eb, fs_info, &epd);
3633 if (!ret) {
3634 free_extent_buffer(eb);
3635 continue;
3636 }
3637
3638 ret = write_one_eb(eb, fs_info, wbc, &epd);
3639 if (ret) {
3640 done = 1;
3641 free_extent_buffer(eb);
3642 break;
3643 }
3644 free_extent_buffer(eb);
3645
3646 /*
3647 * the filesystem may choose to bump up nr_to_write.
3648 * We have to make sure to honor the new nr_to_write
3649 * at any time
3650 */
3651 nr_to_write_done = wbc->nr_to_write <= 0;
3652 }
3653 pagevec_release(&pvec);
3654 cond_resched();
3655 }
3656 if (!scanned && !done) {
3657 /*
3658 * We hit the last page and there is more work to be done: wrap
3659 * back to the start of the file
3660 */
3661 scanned = 1;
3662 index = 0;
3663 goto retry;
3664 }
3665 flush_write_bio(&epd);
3666 return ret;
3667}
3668
3669/**
3670 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
3671 * @mapping: address space structure to write
3672 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
3673 * @writepage: function called for each page
3674 * @data: data passed to writepage function
3675 *
3676 * If a page is already under I/O, write_cache_pages() skips it, even
3677 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
3678 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
3679 * and msync() need to guarantee that all the data which was dirty at the time
3680 * the call was made get new I/O started against them. If wbc->sync_mode is
3681 * WB_SYNC_ALL then we were called for data integrity and we must wait for
3682 * existing IO to complete.
3683 */
3684static int extent_write_cache_pages(struct extent_io_tree *tree,
3685 struct address_space *mapping,
3686 struct writeback_control *wbc,
3687 writepage_t writepage, void *data,
3688 void (*flush_fn)(void *))
3689{
3690 struct inode *inode = mapping->host;
3691 int ret = 0;
3692 int done = 0;
3693 int nr_to_write_done = 0;
3694 struct pagevec pvec;
3695 int nr_pages;
3696 pgoff_t index;
3697 pgoff_t end; /* Inclusive */
3698 int scanned = 0;
3699 int tag;
3700
3701 /*
3702 * We have to hold onto the inode so that ordered extents can do their
3703 * work when the IO finishes. The alternative to this is failing to add
3704 * an ordered extent if the igrab() fails there and that is a huge pain
3705 * to deal with, so instead just hold onto the inode throughout the
3706 * writepages operation. If it fails here we are freeing up the inode
3707 * anyway and we'd rather not waste our time writing out stuff that is
3708 * going to be truncated anyway.
3709 */
3710 if (!igrab(inode))
3711 return 0;
3712
3713 pagevec_init(&pvec, 0);
3714 if (wbc->range_cyclic) {
3715 index = mapping->writeback_index; /* Start from prev offset */
3716 end = -1;
3717 } else {
3718 index = wbc->range_start >> PAGE_CACHE_SHIFT;
3719 end = wbc->range_end >> PAGE_CACHE_SHIFT;
3720 scanned = 1;
3721 }
3722 if (wbc->sync_mode == WB_SYNC_ALL)
3723 tag = PAGECACHE_TAG_TOWRITE;
3724 else
3725 tag = PAGECACHE_TAG_DIRTY;
3726retry:
3727 if (wbc->sync_mode == WB_SYNC_ALL)
3728 tag_pages_for_writeback(mapping, index, end);
3729 while (!done && !nr_to_write_done && (index <= end) &&
3730 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
3731 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
3732 unsigned i;
3733
3734 scanned = 1;
3735 for (i = 0; i < nr_pages; i++) {
3736 struct page *page = pvec.pages[i];
3737
3738 /*
3739 * At this point we hold neither mapping->tree_lock nor
3740 * lock on the page itself: the page may be truncated or
3741 * invalidated (changing page->mapping to NULL), or even
3742 * swizzled back from swapper_space to tmpfs file
3743 * mapping
3744 */
3745 if (!trylock_page(page)) {
3746 flush_fn(data);
3747 lock_page(page);
3748 }
3749
3750 if (unlikely(page->mapping != mapping)) {
3751 unlock_page(page);
3752 continue;
3753 }
3754
3755 if (!wbc->range_cyclic && page->index > end) {
3756 done = 1;
3757 unlock_page(page);
3758 continue;
3759 }
3760
3761 if (wbc->sync_mode != WB_SYNC_NONE) {
3762 if (PageWriteback(page))
3763 flush_fn(data);
3764 wait_on_page_writeback(page);
3765 }
3766
3767 if (PageWriteback(page) ||
3768 !clear_page_dirty_for_io(page)) {
3769 unlock_page(page);
3770 continue;
3771 }
3772
3773 ret = (*writepage)(page, wbc, data);
3774
3775 if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
3776 unlock_page(page);
3777 ret = 0;
3778 }
3779 if (ret)
3780 done = 1;
3781
3782 /*
3783 * the filesystem may choose to bump up nr_to_write.
3784 * We have to make sure to honor the new nr_to_write
3785 * at any time
3786 */
3787 nr_to_write_done = wbc->nr_to_write <= 0;
3788 }
3789 pagevec_release(&pvec);
3790 cond_resched();
3791 }
3792 if (!scanned && !done) {
3793 /*
3794 * We hit the last page and there is more work to be done: wrap
3795 * back to the start of the file
3796 */
3797 scanned = 1;
3798 index = 0;
3799 goto retry;
3800 }
3801 btrfs_add_delayed_iput(inode);
3802 return ret;
3803}
3804
3805static void flush_epd_write_bio(struct extent_page_data *epd)
3806{
3807 if (epd->bio) {
3808 int rw = WRITE;
3809 int ret;
3810
3811 if (epd->sync_io)
3812 rw = WRITE_SYNC;
3813
3814 ret = submit_one_bio(rw, epd->bio, 0, epd->bio_flags);
3815 BUG_ON(ret < 0); /* -ENOMEM */
3816 epd->bio = NULL;
3817 }
3818}
3819
3820static noinline void flush_write_bio(void *data)
3821{
3822 struct extent_page_data *epd = data;
3823 flush_epd_write_bio(epd);
3824}
3825
3826int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
3827 get_extent_t *get_extent,
3828 struct writeback_control *wbc)
3829{
3830 int ret;
3831 struct extent_page_data epd = {
3832 .bio = NULL,
3833 .tree = tree,
3834 .get_extent = get_extent,
3835 .extent_locked = 0,
3836 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3837 .bio_flags = 0,
3838 };
3839
3840 ret = __extent_writepage(page, wbc, &epd);
3841
3842 flush_epd_write_bio(&epd);
3843 return ret;
3844}
3845
3846int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
3847 u64 start, u64 end, get_extent_t *get_extent,
3848 int mode)
3849{
3850 int ret = 0;
3851 struct address_space *mapping = inode->i_mapping;
3852 struct page *page;
3853 unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
3854 PAGE_CACHE_SHIFT;
3855
3856 struct extent_page_data epd = {
3857 .bio = NULL,
3858 .tree = tree,
3859 .get_extent = get_extent,
3860 .extent_locked = 1,
3861 .sync_io = mode == WB_SYNC_ALL,
3862 .bio_flags = 0,
3863 };
3864 struct writeback_control wbc_writepages = {
3865 .sync_mode = mode,
3866 .nr_to_write = nr_pages * 2,
3867 .range_start = start,
3868 .range_end = end + 1,
3869 };
3870
3871 while (start <= end) {
3872 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
3873 if (clear_page_dirty_for_io(page))
3874 ret = __extent_writepage(page, &wbc_writepages, &epd);
3875 else {
3876 if (tree->ops && tree->ops->writepage_end_io_hook)
3877 tree->ops->writepage_end_io_hook(page, start,
3878 start + PAGE_CACHE_SIZE - 1,
3879 NULL, 1);
3880 unlock_page(page);
3881 }
3882 page_cache_release(page);
3883 start += PAGE_CACHE_SIZE;
3884 }
3885
3886 flush_epd_write_bio(&epd);
3887 return ret;
3888}
3889
3890int extent_writepages(struct extent_io_tree *tree,
3891 struct address_space *mapping,
3892 get_extent_t *get_extent,
3893 struct writeback_control *wbc)
3894{
3895 int ret = 0;
3896 struct extent_page_data epd = {
3897 .bio = NULL,
3898 .tree = tree,
3899 .get_extent = get_extent,
3900 .extent_locked = 0,
3901 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3902 .bio_flags = 0,
3903 };
3904
3905 ret = extent_write_cache_pages(tree, mapping, wbc,
3906 __extent_writepage, &epd,
3907 flush_write_bio);
3908 flush_epd_write_bio(&epd);
3909 return ret;
3910}
3911
3912int extent_readpages(struct extent_io_tree *tree,
3913 struct address_space *mapping,
3914 struct list_head *pages, unsigned nr_pages,
3915 get_extent_t get_extent)
3916{
3917 struct bio *bio = NULL;
3918 unsigned page_idx;
3919 unsigned long bio_flags = 0;
3920 struct page *pagepool[16];
3921 struct page *page;
3922 struct extent_map *em_cached = NULL;
3923 int nr = 0;
3924
3925 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
3926 page = list_entry(pages->prev, struct page, lru);
3927
3928 prefetchw(&page->flags);
3929 list_del(&page->lru);
3930 if (add_to_page_cache_lru(page, mapping,
3931 page->index, GFP_NOFS)) {
3932 page_cache_release(page);
3933 continue;
3934 }
3935
3936 pagepool[nr++] = page;
3937 if (nr < ARRAY_SIZE(pagepool))
3938 continue;
3939 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
3940 &bio, 0, &bio_flags, READ);
3941 nr = 0;
3942 }
3943 if (nr)
3944 __extent_readpages(tree, pagepool, nr, get_extent, &em_cached,
3945 &bio, 0, &bio_flags, READ);
3946
3947 if (em_cached)
3948 free_extent_map(em_cached);
3949
3950 BUG_ON(!list_empty(pages));
3951 if (bio)
3952 return submit_one_bio(READ, bio, 0, bio_flags);
3953 return 0;
3954}
3955
3956/*
3957 * basic invalidatepage code, this waits on any locked or writeback
3958 * ranges corresponding to the page, and then deletes any extent state
3959 * records from the tree
3960 */
3961int extent_invalidatepage(struct extent_io_tree *tree,
3962 struct page *page, unsigned long offset)
3963{
3964 struct extent_state *cached_state = NULL;
3965 u64 start = page_offset(page);
3966 u64 end = start + PAGE_CACHE_SIZE - 1;
3967 size_t blocksize = page->mapping->host->i_sb->s_blocksize;
3968
3969 start += ALIGN(offset, blocksize);
3970 if (start > end)
3971 return 0;
3972
3973 lock_extent_bits(tree, start, end, 0, &cached_state);
3974 wait_on_page_writeback(page);
3975 clear_extent_bit(tree, start, end,
3976 EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
3977 EXTENT_DO_ACCOUNTING,
3978 1, 1, &cached_state, GFP_NOFS);
3979 return 0;
3980}
3981
3982/*
3983 * a helper for releasepage, this tests for areas of the page that
3984 * are locked or under IO and drops the related state bits if it is safe
3985 * to drop the page.
3986 */
3987static int try_release_extent_state(struct extent_map_tree *map,
3988 struct extent_io_tree *tree,
3989 struct page *page, gfp_t mask)
3990{
3991 u64 start = page_offset(page);
3992 u64 end = start + PAGE_CACHE_SIZE - 1;
3993 int ret = 1;
3994
3995 if (test_range_bit(tree, start, end,
3996 EXTENT_IOBITS, 0, NULL))
3997 ret = 0;
3998 else {
3999 if ((mask & GFP_NOFS) == GFP_NOFS)
4000 mask = GFP_NOFS;
4001 /*
4002 * at this point we can safely clear everything except the
4003 * locked bit and the nodatasum bit
4004 */
4005 ret = clear_extent_bit(tree, start, end,
4006 ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4007 0, 0, NULL, mask);
4008
4009 /* if clear_extent_bit failed for enomem reasons,
4010 * we can't allow the release to continue.
4011 */
4012 if (ret < 0)
4013 ret = 0;
4014 else
4015 ret = 1;
4016 }
4017 return ret;
4018}
4019
4020/*
4021 * a helper for releasepage. As long as there are no locked extents
4022 * in the range corresponding to the page, both state records and extent
4023 * map records are removed
4024 */
4025int try_release_extent_mapping(struct extent_map_tree *map,
4026 struct extent_io_tree *tree, struct page *page,
4027 gfp_t mask)
4028{
4029 struct extent_map *em;
4030 u64 start = page_offset(page);
4031 u64 end = start + PAGE_CACHE_SIZE - 1;
4032
4033 if ((mask & __GFP_WAIT) &&
4034 page->mapping->host->i_size > 16 * 1024 * 1024) {
4035 u64 len;
4036 while (start <= end) {
4037 len = end - start + 1;
4038 write_lock(&map->lock);
4039 em = lookup_extent_mapping(map, start, len);
4040 if (!em) {
4041 write_unlock(&map->lock);
4042 break;
4043 }
4044 if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4045 em->start != start) {
4046 write_unlock(&map->lock);
4047 free_extent_map(em);
4048 break;
4049 }
4050 if (!test_range_bit(tree, em->start,
4051 extent_map_end(em) - 1,
4052 EXTENT_LOCKED | EXTENT_WRITEBACK,
4053 0, NULL)) {
4054 remove_extent_mapping(map, em);
4055 /* once for the rb tree */
4056 free_extent_map(em);
4057 }
4058 start = extent_map_end(em);
4059 write_unlock(&map->lock);
4060
4061 /* once for us */
4062 free_extent_map(em);
4063 }
4064 }
4065 return try_release_extent_state(map, tree, page, mask);
4066}
4067
4068/*
4069 * helper function for fiemap, which doesn't want to see any holes.
4070 * This maps until we find something past 'last'
4071 */
4072static struct extent_map *get_extent_skip_holes(struct inode *inode,
4073 u64 offset,
4074 u64 last,
4075 get_extent_t *get_extent)
4076{
4077 u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
4078 struct extent_map *em;
4079 u64 len;
4080
4081 if (offset >= last)
4082 return NULL;
4083
4084 while (1) {
4085 len = last - offset;
4086 if (len == 0)
4087 break;
4088 len = ALIGN(len, sectorsize);
4089 em = get_extent(inode, NULL, 0, offset, len, 0);
4090 if (IS_ERR_OR_NULL(em))
4091 return em;
4092
4093 /* if this isn't a hole return it */
4094 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
4095 em->block_start != EXTENT_MAP_HOLE) {
4096 return em;
4097 }
4098
4099 /* this is a hole, advance to the next extent */
4100 offset = extent_map_end(em);
4101 free_extent_map(em);
4102 if (offset >= last)
4103 break;
4104 }
4105 return NULL;
4106}
4107
4108static noinline int count_ext_ref(u64 inum, u64 offset, u64 root_id, void *ctx)
4109{
4110 unsigned long cnt = *((unsigned long *)ctx);
4111
4112 cnt++;
4113 *((unsigned long *)ctx) = cnt;
4114
4115 /* Now we're sure that the extent is shared. */
4116 if (cnt > 1)
4117 return 1;
4118 return 0;
4119}
4120
4121int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4122 __u64 start, __u64 len, get_extent_t *get_extent)
4123{
4124 int ret = 0;
4125 u64 off = start;
4126 u64 max = start + len;
4127 u32 flags = 0;
4128 u32 found_type;
4129 u64 last;
4130 u64 last_for_get_extent = 0;
4131 u64 disko = 0;
4132 u64 isize = i_size_read(inode);
4133 struct btrfs_key found_key;
4134 struct extent_map *em = NULL;
4135 struct extent_state *cached_state = NULL;
4136 struct btrfs_path *path;
4137 int end = 0;
4138 u64 em_start = 0;
4139 u64 em_len = 0;
4140 u64 em_end = 0;
4141
4142 if (len == 0)
4143 return -EINVAL;
4144
4145 path = btrfs_alloc_path();
4146 if (!path)
4147 return -ENOMEM;
4148 path->leave_spinning = 1;
4149
4150 start = ALIGN(start, BTRFS_I(inode)->root->sectorsize);
4151 len = ALIGN(len, BTRFS_I(inode)->root->sectorsize);
4152
4153 /*
4154 * lookup the last file extent. We're not using i_size here
4155 * because there might be preallocation past i_size
4156 */
4157 ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
4158 path, btrfs_ino(inode), -1, 0);
4159 if (ret < 0) {
4160 btrfs_free_path(path);
4161 return ret;
4162 }
4163 WARN_ON(!ret);
4164 path->slots[0]--;
4165 btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4166 found_type = btrfs_key_type(&found_key);
4167
4168 /* No extents, but there might be delalloc bits */
4169 if (found_key.objectid != btrfs_ino(inode) ||
4170 found_type != BTRFS_EXTENT_DATA_KEY) {
4171 /* have to trust i_size as the end */
4172 last = (u64)-1;
4173 last_for_get_extent = isize;
4174 } else {
4175 /*
4176 * remember the start of the last extent. There are a
4177 * bunch of different factors that go into the length of the
4178 * extent, so its much less complex to remember where it started
4179 */
4180 last = found_key.offset;
4181 last_for_get_extent = last + 1;
4182 }
4183 btrfs_release_path(path);
4184
4185 /*
4186 * we might have some extents allocated but more delalloc past those
4187 * extents. so, we trust isize unless the start of the last extent is
4188 * beyond isize
4189 */
4190 if (last < isize) {
4191 last = (u64)-1;
4192 last_for_get_extent = isize;
4193 }
4194
4195 lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1, 0,
4196 &cached_state);
4197
4198 em = get_extent_skip_holes(inode, start, last_for_get_extent,
4199 get_extent);
4200 if (!em)
4201 goto out;
4202 if (IS_ERR(em)) {
4203 ret = PTR_ERR(em);
4204 goto out;
4205 }
4206
4207 while (!end) {
4208 u64 offset_in_extent = 0;
4209
4210 /* break if the extent we found is outside the range */
4211 if (em->start >= max || extent_map_end(em) < off)
4212 break;
4213
4214 /*
4215 * get_extent may return an extent that starts before our
4216 * requested range. We have to make sure the ranges
4217 * we return to fiemap always move forward and don't
4218 * overlap, so adjust the offsets here
4219 */
4220 em_start = max(em->start, off);
4221
4222 /*
4223 * record the offset from the start of the extent
4224 * for adjusting the disk offset below. Only do this if the
4225 * extent isn't compressed since our in ram offset may be past
4226 * what we have actually allocated on disk.
4227 */
4228 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4229 offset_in_extent = em_start - em->start;
4230 em_end = extent_map_end(em);
4231 em_len = em_end - em_start;
4232 disko = 0;
4233 flags = 0;
4234
4235 /*
4236 * bump off for our next call to get_extent
4237 */
4238 off = extent_map_end(em);
4239 if (off >= max)
4240 end = 1;
4241
4242 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4243 end = 1;
4244 flags |= FIEMAP_EXTENT_LAST;
4245 } else if (em->block_start == EXTENT_MAP_INLINE) {
4246 flags |= (FIEMAP_EXTENT_DATA_INLINE |
4247 FIEMAP_EXTENT_NOT_ALIGNED);
4248 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4249 flags |= (FIEMAP_EXTENT_DELALLOC |
4250 FIEMAP_EXTENT_UNKNOWN);
4251 } else {
4252 unsigned long ref_cnt = 0;
4253
4254 disko = em->block_start + offset_in_extent;
4255
4256 /*
4257 * As btrfs supports shared space, this information
4258 * can be exported to userspace tools via
4259 * flag FIEMAP_EXTENT_SHARED.
4260 */
4261 ret = iterate_inodes_from_logical(
4262 em->block_start,
4263 BTRFS_I(inode)->root->fs_info,
4264 path, count_ext_ref, &ref_cnt);
4265 if (ret < 0 && ret != -ENOENT)
4266 goto out_free;
4267
4268 if (ref_cnt > 1)
4269 flags |= FIEMAP_EXTENT_SHARED;
4270 }
4271 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4272 flags |= FIEMAP_EXTENT_ENCODED;
4273
4274 free_extent_map(em);
4275 em = NULL;
4276 if ((em_start >= last) || em_len == (u64)-1 ||
4277 (last == (u64)-1 && isize <= em_end)) {
4278 flags |= FIEMAP_EXTENT_LAST;
4279 end = 1;
4280 }
4281
4282 /* now scan forward to see if this is really the last extent. */
4283 em = get_extent_skip_holes(inode, off, last_for_get_extent,
4284 get_extent);
4285 if (IS_ERR(em)) {
4286 ret = PTR_ERR(em);
4287 goto out;
4288 }
4289 if (!em) {
4290 flags |= FIEMAP_EXTENT_LAST;
4291 end = 1;
4292 }
4293 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
4294 em_len, flags);
4295 if (ret)
4296 goto out_free;
4297 }
4298out_free:
4299 free_extent_map(em);
4300out:
4301 btrfs_free_path(path);
4302 unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4303 &cached_state, GFP_NOFS);
4304 return ret;
4305}
4306
4307static void __free_extent_buffer(struct extent_buffer *eb)
4308{
4309 btrfs_leak_debug_del(&eb->leak_list);
4310 kmem_cache_free(extent_buffer_cache, eb);
4311}
4312
4313int extent_buffer_under_io(struct extent_buffer *eb)
4314{
4315 return (atomic_read(&eb->io_pages) ||
4316 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4317 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4318}
4319
4320/*
4321 * Helper for releasing extent buffer page.
4322 */
4323static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
4324 unsigned long start_idx)
4325{
4326 unsigned long index;
4327 unsigned long num_pages;
4328 struct page *page;
4329 int mapped = !test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4330
4331 BUG_ON(extent_buffer_under_io(eb));
4332
4333 num_pages = num_extent_pages(eb->start, eb->len);
4334 index = start_idx + num_pages;
4335 if (start_idx >= index)
4336 return;
4337
4338 do {
4339 index--;
4340 page = extent_buffer_page(eb, index);
4341 if (page && mapped) {
4342 spin_lock(&page->mapping->private_lock);
4343 /*
4344 * We do this since we'll remove the pages after we've
4345 * removed the eb from the radix tree, so we could race
4346 * and have this page now attached to the new eb. So
4347 * only clear page_private if it's still connected to
4348 * this eb.
4349 */
4350 if (PagePrivate(page) &&
4351 page->private == (unsigned long)eb) {
4352 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4353 BUG_ON(PageDirty(page));
4354 BUG_ON(PageWriteback(page));
4355 /*
4356 * We need to make sure we haven't be attached
4357 * to a new eb.
4358 */
4359 ClearPagePrivate(page);
4360 set_page_private(page, 0);
4361 /* One for the page private */
4362 page_cache_release(page);
4363 }
4364 spin_unlock(&page->mapping->private_lock);
4365
4366 }
4367 if (page) {
4368 /* One for when we alloced the page */
4369 page_cache_release(page);
4370 }
4371 } while (index != start_idx);
4372}
4373
4374/*
4375 * Helper for releasing the extent buffer.
4376 */
4377static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4378{
4379 btrfs_release_extent_buffer_page(eb, 0);
4380 __free_extent_buffer(eb);
4381}
4382
4383static struct extent_buffer *
4384__alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4385 unsigned long len, gfp_t mask)
4386{
4387 struct extent_buffer *eb = NULL;
4388
4389 eb = kmem_cache_zalloc(extent_buffer_cache, mask);
4390 if (eb == NULL)
4391 return NULL;
4392 eb->start = start;
4393 eb->len = len;
4394 eb->fs_info = fs_info;
4395 eb->bflags = 0;
4396 rwlock_init(&eb->lock);
4397 atomic_set(&eb->write_locks, 0);
4398 atomic_set(&eb->read_locks, 0);
4399 atomic_set(&eb->blocking_readers, 0);
4400 atomic_set(&eb->blocking_writers, 0);
4401 atomic_set(&eb->spinning_readers, 0);
4402 atomic_set(&eb->spinning_writers, 0);
4403 eb->lock_nested = 0;
4404 init_waitqueue_head(&eb->write_lock_wq);
4405 init_waitqueue_head(&eb->read_lock_wq);
4406
4407 btrfs_leak_debug_add(&eb->leak_list, &buffers);
4408
4409 spin_lock_init(&eb->refs_lock);
4410 atomic_set(&eb->refs, 1);
4411 atomic_set(&eb->io_pages, 0);
4412
4413 /*
4414 * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
4415 */
4416 BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
4417 > MAX_INLINE_EXTENT_BUFFER_SIZE);
4418 BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
4419
4420 return eb;
4421}
4422
4423struct extent_buffer *btrfs_clone_extent_buffer(struct extent_buffer *src)
4424{
4425 unsigned long i;
4426 struct page *p;
4427 struct extent_buffer *new;
4428 unsigned long num_pages = num_extent_pages(src->start, src->len);
4429
4430 new = __alloc_extent_buffer(NULL, src->start, src->len, GFP_NOFS);
4431 if (new == NULL)
4432 return NULL;
4433
4434 for (i = 0; i < num_pages; i++) {
4435 p = alloc_page(GFP_NOFS);
4436 if (!p) {
4437 btrfs_release_extent_buffer(new);
4438 return NULL;
4439 }
4440 attach_extent_buffer_page(new, p);
4441 WARN_ON(PageDirty(p));
4442 SetPageUptodate(p);
4443 new->pages[i] = p;
4444 }
4445
4446 copy_extent_buffer(new, src, 0, 0, src->len);
4447 set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
4448 set_bit(EXTENT_BUFFER_DUMMY, &new->bflags);
4449
4450 return new;
4451}
4452
4453struct extent_buffer *alloc_dummy_extent_buffer(u64 start, unsigned long len)
4454{
4455 struct extent_buffer *eb;
4456 unsigned long num_pages = num_extent_pages(0, len);
4457 unsigned long i;
4458
4459 eb = __alloc_extent_buffer(NULL, start, len, GFP_NOFS);
4460 if (!eb)
4461 return NULL;
4462
4463 for (i = 0; i < num_pages; i++) {
4464 eb->pages[i] = alloc_page(GFP_NOFS);
4465 if (!eb->pages[i])
4466 goto err;
4467 }
4468 set_extent_buffer_uptodate(eb);
4469 btrfs_set_header_nritems(eb, 0);
4470 set_bit(EXTENT_BUFFER_DUMMY, &eb->bflags);
4471
4472 return eb;
4473err:
4474 for (; i > 0; i--)
4475 __free_page(eb->pages[i - 1]);
4476 __free_extent_buffer(eb);
4477 return NULL;
4478}
4479
4480static void check_buffer_tree_ref(struct extent_buffer *eb)
4481{
4482 int refs;
4483 /* the ref bit is tricky. We have to make sure it is set
4484 * if we have the buffer dirty. Otherwise the
4485 * code to free a buffer can end up dropping a dirty
4486 * page
4487 *
4488 * Once the ref bit is set, it won't go away while the
4489 * buffer is dirty or in writeback, and it also won't
4490 * go away while we have the reference count on the
4491 * eb bumped.
4492 *
4493 * We can't just set the ref bit without bumping the
4494 * ref on the eb because free_extent_buffer might
4495 * see the ref bit and try to clear it. If this happens
4496 * free_extent_buffer might end up dropping our original
4497 * ref by mistake and freeing the page before we are able
4498 * to add one more ref.
4499 *
4500 * So bump the ref count first, then set the bit. If someone
4501 * beat us to it, drop the ref we added.
4502 */
4503 refs = atomic_read(&eb->refs);
4504 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4505 return;
4506
4507 spin_lock(&eb->refs_lock);
4508 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4509 atomic_inc(&eb->refs);
4510 spin_unlock(&eb->refs_lock);
4511}
4512
4513static void mark_extent_buffer_accessed(struct extent_buffer *eb)
4514{
4515 unsigned long num_pages, i;
4516
4517 check_buffer_tree_ref(eb);
4518
4519 num_pages = num_extent_pages(eb->start, eb->len);
4520 for (i = 0; i < num_pages; i++) {
4521 struct page *p = extent_buffer_page(eb, i);
4522 mark_page_accessed(p);
4523 }
4524}
4525
4526struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4527 u64 start)
4528{
4529 struct extent_buffer *eb;
4530
4531 rcu_read_lock();
4532 eb = radix_tree_lookup(&fs_info->buffer_radix,
4533 start >> PAGE_CACHE_SHIFT);
4534 if (eb && atomic_inc_not_zero(&eb->refs)) {
4535 rcu_read_unlock();
4536 mark_extent_buffer_accessed(eb);
4537 return eb;
4538 }
4539 rcu_read_unlock();
4540
4541 return NULL;
4542}
4543
4544struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4545 u64 start, unsigned long len)
4546{
4547 unsigned long num_pages = num_extent_pages(start, len);
4548 unsigned long i;
4549 unsigned long index = start >> PAGE_CACHE_SHIFT;
4550 struct extent_buffer *eb;
4551 struct extent_buffer *exists = NULL;
4552 struct page *p;
4553 struct address_space *mapping = fs_info->btree_inode->i_mapping;
4554 int uptodate = 1;
4555 int ret;
4556
4557 eb = find_extent_buffer(fs_info, start);
4558 if (eb)
4559 return eb;
4560
4561 eb = __alloc_extent_buffer(fs_info, start, len, GFP_NOFS);
4562 if (!eb)
4563 return NULL;
4564
4565 for (i = 0; i < num_pages; i++, index++) {
4566 p = find_or_create_page(mapping, index, GFP_NOFS);
4567 if (!p)
4568 goto free_eb;
4569
4570 spin_lock(&mapping->private_lock);
4571 if (PagePrivate(p)) {
4572 /*
4573 * We could have already allocated an eb for this page
4574 * and attached one so lets see if we can get a ref on
4575 * the existing eb, and if we can we know it's good and
4576 * we can just return that one, else we know we can just
4577 * overwrite page->private.
4578 */
4579 exists = (struct extent_buffer *)p->private;
4580 if (atomic_inc_not_zero(&exists->refs)) {
4581 spin_unlock(&mapping->private_lock);
4582 unlock_page(p);
4583 page_cache_release(p);
4584 mark_extent_buffer_accessed(exists);
4585 goto free_eb;
4586 }
4587
4588 /*
4589 * Do this so attach doesn't complain and we need to
4590 * drop the ref the old guy had.
4591 */
4592 ClearPagePrivate(p);
4593 WARN_ON(PageDirty(p));
4594 page_cache_release(p);
4595 }
4596 attach_extent_buffer_page(eb, p);
4597 spin_unlock(&mapping->private_lock);
4598 WARN_ON(PageDirty(p));
4599 mark_page_accessed(p);
4600 eb->pages[i] = p;
4601 if (!PageUptodate(p))
4602 uptodate = 0;
4603
4604 /*
4605 * see below about how we avoid a nasty race with release page
4606 * and why we unlock later
4607 */
4608 }
4609 if (uptodate)
4610 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4611again:
4612 ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
4613 if (ret)
4614 goto free_eb;
4615
4616 spin_lock(&fs_info->buffer_lock);
4617 ret = radix_tree_insert(&fs_info->buffer_radix,
4618 start >> PAGE_CACHE_SHIFT, eb);
4619 spin_unlock(&fs_info->buffer_lock);
4620 radix_tree_preload_end();
4621 if (ret == -EEXIST) {
4622 exists = find_extent_buffer(fs_info, start);
4623 if (exists)
4624 goto free_eb;
4625 else
4626 goto again;
4627 }
4628 /* add one reference for the tree */
4629 check_buffer_tree_ref(eb);
4630 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4631
4632 /*
4633 * there is a race where release page may have
4634 * tried to find this extent buffer in the radix
4635 * but failed. It will tell the VM it is safe to
4636 * reclaim the, and it will clear the page private bit.
4637 * We must make sure to set the page private bit properly
4638 * after the extent buffer is in the radix tree so
4639 * it doesn't get lost
4640 */
4641 SetPageChecked(eb->pages[0]);
4642 for (i = 1; i < num_pages; i++) {
4643 p = extent_buffer_page(eb, i);
4644 ClearPageChecked(p);
4645 unlock_page(p);
4646 }
4647 unlock_page(eb->pages[0]);
4648 return eb;
4649
4650free_eb:
4651 for (i = 0; i < num_pages; i++) {
4652 if (eb->pages[i])
4653 unlock_page(eb->pages[i]);
4654 }
4655
4656 WARN_ON(!atomic_dec_and_test(&eb->refs));
4657 btrfs_release_extent_buffer(eb);
4658 return exists;
4659}
4660
4661static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4662{
4663 struct extent_buffer *eb =
4664 container_of(head, struct extent_buffer, rcu_head);
4665
4666 __free_extent_buffer(eb);
4667}
4668
4669/* Expects to have eb->eb_lock already held */
4670static int release_extent_buffer(struct extent_buffer *eb)
4671{
4672 WARN_ON(atomic_read(&eb->refs) == 0);
4673 if (atomic_dec_and_test(&eb->refs)) {
4674 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4675 struct btrfs_fs_info *fs_info = eb->fs_info;
4676
4677 spin_unlock(&eb->refs_lock);
4678
4679 spin_lock(&fs_info->buffer_lock);
4680 radix_tree_delete(&fs_info->buffer_radix,
4681 eb->start >> PAGE_CACHE_SHIFT);
4682 spin_unlock(&fs_info->buffer_lock);
4683 } else {
4684 spin_unlock(&eb->refs_lock);
4685 }
4686
4687 /* Should be safe to release our pages at this point */
4688 btrfs_release_extent_buffer_page(eb, 0);
4689 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4690 return 1;
4691 }
4692 spin_unlock(&eb->refs_lock);
4693
4694 return 0;
4695}
4696
4697void free_extent_buffer(struct extent_buffer *eb)
4698{
4699 int refs;
4700 int old;
4701 if (!eb)
4702 return;
4703
4704 while (1) {
4705 refs = atomic_read(&eb->refs);
4706 if (refs <= 3)
4707 break;
4708 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
4709 if (old == refs)
4710 return;
4711 }
4712
4713 spin_lock(&eb->refs_lock);
4714 if (atomic_read(&eb->refs) == 2 &&
4715 test_bit(EXTENT_BUFFER_DUMMY, &eb->bflags))
4716 atomic_dec(&eb->refs);
4717
4718 if (atomic_read(&eb->refs) == 2 &&
4719 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4720 !extent_buffer_under_io(eb) &&
4721 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4722 atomic_dec(&eb->refs);
4723
4724 /*
4725 * I know this is terrible, but it's temporary until we stop tracking
4726 * the uptodate bits and such for the extent buffers.
4727 */
4728 release_extent_buffer(eb);
4729}
4730
4731void free_extent_buffer_stale(struct extent_buffer *eb)
4732{
4733 if (!eb)
4734 return;
4735
4736 spin_lock(&eb->refs_lock);
4737 set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4738
4739 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4740 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4741 atomic_dec(&eb->refs);
4742 release_extent_buffer(eb);
4743}
4744
4745void clear_extent_buffer_dirty(struct extent_buffer *eb)
4746{
4747 unsigned long i;
4748 unsigned long num_pages;
4749 struct page *page;
4750
4751 num_pages = num_extent_pages(eb->start, eb->len);
4752
4753 for (i = 0; i < num_pages; i++) {
4754 page = extent_buffer_page(eb, i);
4755 if (!PageDirty(page))
4756 continue;
4757
4758 lock_page(page);
4759 WARN_ON(!PagePrivate(page));
4760
4761 clear_page_dirty_for_io(page);
4762 spin_lock_irq(&page->mapping->tree_lock);
4763 if (!PageDirty(page)) {
4764 radix_tree_tag_clear(&page->mapping->page_tree,
4765 page_index(page),
4766 PAGECACHE_TAG_DIRTY);
4767 }
4768 spin_unlock_irq(&page->mapping->tree_lock);
4769 ClearPageError(page);
4770 unlock_page(page);
4771 }
4772 WARN_ON(atomic_read(&eb->refs) == 0);
4773}
4774
4775int set_extent_buffer_dirty(struct extent_buffer *eb)
4776{
4777 unsigned long i;
4778 unsigned long num_pages;
4779 int was_dirty = 0;
4780
4781 check_buffer_tree_ref(eb);
4782
4783 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4784
4785 num_pages = num_extent_pages(eb->start, eb->len);
4786 WARN_ON(atomic_read(&eb->refs) == 0);
4787 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4788
4789 for (i = 0; i < num_pages; i++)
4790 set_page_dirty(extent_buffer_page(eb, i));
4791 return was_dirty;
4792}
4793
4794int clear_extent_buffer_uptodate(struct extent_buffer *eb)
4795{
4796 unsigned long i;
4797 struct page *page;
4798 unsigned long num_pages;
4799
4800 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4801 num_pages = num_extent_pages(eb->start, eb->len);
4802 for (i = 0; i < num_pages; i++) {
4803 page = extent_buffer_page(eb, i);
4804 if (page)
4805 ClearPageUptodate(page);
4806 }
4807 return 0;
4808}
4809
4810int set_extent_buffer_uptodate(struct extent_buffer *eb)
4811{
4812 unsigned long i;
4813 struct page *page;
4814 unsigned long num_pages;
4815
4816 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4817 num_pages = num_extent_pages(eb->start, eb->len);
4818 for (i = 0; i < num_pages; i++) {
4819 page = extent_buffer_page(eb, i);
4820 SetPageUptodate(page);
4821 }
4822 return 0;
4823}
4824
4825int extent_buffer_uptodate(struct extent_buffer *eb)
4826{
4827 return test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4828}
4829
4830int read_extent_buffer_pages(struct extent_io_tree *tree,
4831 struct extent_buffer *eb, u64 start, int wait,
4832 get_extent_t *get_extent, int mirror_num)
4833{
4834 unsigned long i;
4835 unsigned long start_i;
4836 struct page *page;
4837 int err;
4838 int ret = 0;
4839 int locked_pages = 0;
4840 int all_uptodate = 1;
4841 unsigned long num_pages;
4842 unsigned long num_reads = 0;
4843 struct bio *bio = NULL;
4844 unsigned long bio_flags = 0;
4845
4846 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4847 return 0;
4848
4849 if (start) {
4850 WARN_ON(start < eb->start);
4851 start_i = (start >> PAGE_CACHE_SHIFT) -
4852 (eb->start >> PAGE_CACHE_SHIFT);
4853 } else {
4854 start_i = 0;
4855 }
4856
4857 num_pages = num_extent_pages(eb->start, eb->len);
4858 for (i = start_i; i < num_pages; i++) {
4859 page = extent_buffer_page(eb, i);
4860 if (wait == WAIT_NONE) {
4861 if (!trylock_page(page))
4862 goto unlock_exit;
4863 } else {
4864 lock_page(page);
4865 }
4866 locked_pages++;
4867 if (!PageUptodate(page)) {
4868 num_reads++;
4869 all_uptodate = 0;
4870 }
4871 }
4872 if (all_uptodate) {
4873 if (start_i == 0)
4874 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4875 goto unlock_exit;
4876 }
4877
4878 clear_bit(EXTENT_BUFFER_IOERR, &eb->bflags);
4879 eb->read_mirror = 0;
4880 atomic_set(&eb->io_pages, num_reads);
4881 for (i = start_i; i < num_pages; i++) {
4882 page = extent_buffer_page(eb, i);
4883 if (!PageUptodate(page)) {
4884 ClearPageError(page);
4885 err = __extent_read_full_page(tree, page,
4886 get_extent, &bio,
4887 mirror_num, &bio_flags,
4888 READ | REQ_META);
4889 if (err)
4890 ret = err;
4891 } else {
4892 unlock_page(page);
4893 }
4894 }
4895
4896 if (bio) {
4897 err = submit_one_bio(READ | REQ_META, bio, mirror_num,
4898 bio_flags);
4899 if (err)
4900 return err;
4901 }
4902
4903 if (ret || wait != WAIT_COMPLETE)
4904 return ret;
4905
4906 for (i = start_i; i < num_pages; i++) {
4907 page = extent_buffer_page(eb, i);
4908 wait_on_page_locked(page);
4909 if (!PageUptodate(page))
4910 ret = -EIO;
4911 }
4912
4913 return ret;
4914
4915unlock_exit:
4916 i = start_i;
4917 while (locked_pages > 0) {
4918 page = extent_buffer_page(eb, i);
4919 i++;
4920 unlock_page(page);
4921 locked_pages--;
4922 }
4923 return ret;
4924}
4925
4926void read_extent_buffer(struct extent_buffer *eb, void *dstv,
4927 unsigned long start,
4928 unsigned long len)
4929{
4930 size_t cur;
4931 size_t offset;
4932 struct page *page;
4933 char *kaddr;
4934 char *dst = (char *)dstv;
4935 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4936 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4937
4938 WARN_ON(start > eb->len);
4939 WARN_ON(start + len > eb->start + eb->len);
4940
4941 offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
4942
4943 while (len > 0) {
4944 page = extent_buffer_page(eb, i);
4945
4946 cur = min(len, (PAGE_CACHE_SIZE - offset));
4947 kaddr = page_address(page);
4948 memcpy(dst, kaddr + offset, cur);
4949
4950 dst += cur;
4951 len -= cur;
4952 offset = 0;
4953 i++;
4954 }
4955}
4956
4957int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
4958 unsigned long min_len, char **map,
4959 unsigned long *map_start,
4960 unsigned long *map_len)
4961{
4962 size_t offset = start & (PAGE_CACHE_SIZE - 1);
4963 char *kaddr;
4964 struct page *p;
4965 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
4966 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
4967 unsigned long end_i = (start_offset + start + min_len - 1) >>
4968 PAGE_CACHE_SHIFT;
4969
4970 if (i != end_i)
4971 return -EINVAL;
4972
4973 if (i == 0) {
4974 offset = start_offset;
4975 *map_start = 0;
4976 } else {
4977 offset = 0;
4978 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
4979 }
4980
4981 if (start + min_len > eb->len) {
4982 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
4983 "wanted %lu %lu\n",
4984 eb->start, eb->len, start, min_len);
4985 return -EINVAL;
4986 }
4987
4988 p = extent_buffer_page(eb, i);
4989 kaddr = page_address(p);
4990 *map = kaddr + offset;
4991 *map_len = PAGE_CACHE_SIZE - offset;
4992 return 0;
4993}
4994
4995int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
4996 unsigned long start,
4997 unsigned long len)
4998{
4999 size_t cur;
5000 size_t offset;
5001 struct page *page;
5002 char *kaddr;
5003 char *ptr = (char *)ptrv;
5004 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5005 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5006 int ret = 0;
5007
5008 WARN_ON(start > eb->len);
5009 WARN_ON(start + len > eb->start + eb->len);
5010
5011 offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5012
5013 while (len > 0) {
5014 page = extent_buffer_page(eb, i);
5015
5016 cur = min(len, (PAGE_CACHE_SIZE - offset));
5017
5018 kaddr = page_address(page);
5019 ret = memcmp(ptr, kaddr + offset, cur);
5020 if (ret)
5021 break;
5022
5023 ptr += cur;
5024 len -= cur;
5025 offset = 0;
5026 i++;
5027 }
5028 return ret;
5029}
5030
5031void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
5032 unsigned long start, unsigned long len)
5033{
5034 size_t cur;
5035 size_t offset;
5036 struct page *page;
5037 char *kaddr;
5038 char *src = (char *)srcv;
5039 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5040 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5041
5042 WARN_ON(start > eb->len);
5043 WARN_ON(start + len > eb->start + eb->len);
5044
5045 offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5046
5047 while (len > 0) {
5048 page = extent_buffer_page(eb, i);
5049 WARN_ON(!PageUptodate(page));
5050
5051 cur = min(len, PAGE_CACHE_SIZE - offset);
5052 kaddr = page_address(page);
5053 memcpy(kaddr + offset, src, cur);
5054
5055 src += cur;
5056 len -= cur;
5057 offset = 0;
5058 i++;
5059 }
5060}
5061
5062void memset_extent_buffer(struct extent_buffer *eb, char c,
5063 unsigned long start, unsigned long len)
5064{
5065 size_t cur;
5066 size_t offset;
5067 struct page *page;
5068 char *kaddr;
5069 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
5070 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
5071
5072 WARN_ON(start > eb->len);
5073 WARN_ON(start + len > eb->start + eb->len);
5074
5075 offset = (start_offset + start) & (PAGE_CACHE_SIZE - 1);
5076
5077 while (len > 0) {
5078 page = extent_buffer_page(eb, i);
5079 WARN_ON(!PageUptodate(page));
5080
5081 cur = min(len, PAGE_CACHE_SIZE - offset);
5082 kaddr = page_address(page);
5083 memset(kaddr + offset, c, cur);
5084
5085 len -= cur;
5086 offset = 0;
5087 i++;
5088 }
5089}
5090
5091void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
5092 unsigned long dst_offset, unsigned long src_offset,
5093 unsigned long len)
5094{
5095 u64 dst_len = dst->len;
5096 size_t cur;
5097 size_t offset;
5098 struct page *page;
5099 char *kaddr;
5100 size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5101 unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5102
5103 WARN_ON(src->len != dst_len);
5104
5105 offset = (start_offset + dst_offset) &
5106 (PAGE_CACHE_SIZE - 1);
5107
5108 while (len > 0) {
5109 page = extent_buffer_page(dst, i);
5110 WARN_ON(!PageUptodate(page));
5111
5112 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
5113
5114 kaddr = page_address(page);
5115 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5116
5117 src_offset += cur;
5118 len -= cur;
5119 offset = 0;
5120 i++;
5121 }
5122}
5123
5124static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5125{
5126 unsigned long distance = (src > dst) ? src - dst : dst - src;
5127 return distance < len;
5128}
5129
5130static void copy_pages(struct page *dst_page, struct page *src_page,
5131 unsigned long dst_off, unsigned long src_off,
5132 unsigned long len)
5133{
5134 char *dst_kaddr = page_address(dst_page);
5135 char *src_kaddr;
5136 int must_memmove = 0;
5137
5138 if (dst_page != src_page) {
5139 src_kaddr = page_address(src_page);
5140 } else {
5141 src_kaddr = dst_kaddr;
5142 if (areas_overlap(src_off, dst_off, len))
5143 must_memmove = 1;
5144 }
5145
5146 if (must_memmove)
5147 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5148 else
5149 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5150}
5151
5152void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5153 unsigned long src_offset, unsigned long len)
5154{
5155 size_t cur;
5156 size_t dst_off_in_page;
5157 size_t src_off_in_page;
5158 size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5159 unsigned long dst_i;
5160 unsigned long src_i;
5161
5162 if (src_offset + len > dst->len) {
5163 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5164 "len %lu dst len %lu\n", src_offset, len, dst->len);
5165 BUG_ON(1);
5166 }
5167 if (dst_offset + len > dst->len) {
5168 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5169 "len %lu dst len %lu\n", dst_offset, len, dst->len);
5170 BUG_ON(1);
5171 }
5172
5173 while (len > 0) {
5174 dst_off_in_page = (start_offset + dst_offset) &
5175 (PAGE_CACHE_SIZE - 1);
5176 src_off_in_page = (start_offset + src_offset) &
5177 (PAGE_CACHE_SIZE - 1);
5178
5179 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
5180 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
5181
5182 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
5183 src_off_in_page));
5184 cur = min_t(unsigned long, cur,
5185 (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
5186
5187 copy_pages(extent_buffer_page(dst, dst_i),
5188 extent_buffer_page(dst, src_i),
5189 dst_off_in_page, src_off_in_page, cur);
5190
5191 src_offset += cur;
5192 dst_offset += cur;
5193 len -= cur;
5194 }
5195}
5196
5197void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
5198 unsigned long src_offset, unsigned long len)
5199{
5200 size_t cur;
5201 size_t dst_off_in_page;
5202 size_t src_off_in_page;
5203 unsigned long dst_end = dst_offset + len - 1;
5204 unsigned long src_end = src_offset + len - 1;
5205 size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
5206 unsigned long dst_i;
5207 unsigned long src_i;
5208
5209 if (src_offset + len > dst->len) {
5210 printk(KERN_ERR "BTRFS: memmove bogus src_offset %lu move "
5211 "len %lu len %lu\n", src_offset, len, dst->len);
5212 BUG_ON(1);
5213 }
5214 if (dst_offset + len > dst->len) {
5215 printk(KERN_ERR "BTRFS: memmove bogus dst_offset %lu move "
5216 "len %lu len %lu\n", dst_offset, len, dst->len);
5217 BUG_ON(1);
5218 }
5219 if (dst_offset < src_offset) {
5220 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5221 return;
5222 }
5223 while (len > 0) {
5224 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
5225 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
5226
5227 dst_off_in_page = (start_offset + dst_end) &
5228 (PAGE_CACHE_SIZE - 1);
5229 src_off_in_page = (start_offset + src_end) &
5230 (PAGE_CACHE_SIZE - 1);
5231
5232 cur = min_t(unsigned long, len, src_off_in_page + 1);
5233 cur = min(cur, dst_off_in_page + 1);
5234 copy_pages(extent_buffer_page(dst, dst_i),
5235 extent_buffer_page(dst, src_i),
5236 dst_off_in_page - cur + 1,
5237 src_off_in_page - cur + 1, cur);
5238
5239 dst_end -= cur;
5240 src_end -= cur;
5241 len -= cur;
5242 }
5243}
5244
5245int try_release_extent_buffer(struct page *page)
5246{
5247 struct extent_buffer *eb;
5248
5249 /*
5250 * We need to make sure noboody is attaching this page to an eb right
5251 * now.
5252 */
5253 spin_lock(&page->mapping->private_lock);
5254 if (!PagePrivate(page)) {
5255 spin_unlock(&page->mapping->private_lock);
5256 return 1;
5257 }
5258
5259 eb = (struct extent_buffer *)page->private;
5260 BUG_ON(!eb);
5261
5262 /*
5263 * This is a little awful but should be ok, we need to make sure that
5264 * the eb doesn't disappear out from under us while we're looking at
5265 * this page.
5266 */
5267 spin_lock(&eb->refs_lock);
5268 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5269 spin_unlock(&eb->refs_lock);
5270 spin_unlock(&page->mapping->private_lock);
5271 return 0;
5272 }
5273 spin_unlock(&page->mapping->private_lock);
5274
5275 /*
5276 * If tree ref isn't set then we know the ref on this eb is a real ref,
5277 * so just return, this page will likely be freed soon anyway.
5278 */
5279 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5280 spin_unlock(&eb->refs_lock);
5281 return 0;
5282 }
5283
5284 return release_extent_buffer(eb);
5285}