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