Loading...
1#include <linux/bitops.h>
2#include <linux/slab.h>
3#include <linux/bio.h>
4#include <linux/mm.h>
5#include <linux/pagemap.h>
6#include <linux/page-flags.h>
7#include <linux/module.h>
8#include <linux/spinlock.h>
9#include <linux/blkdev.h>
10#include <linux/swap.h>
11#include <linux/writeback.h>
12#include <linux/pagevec.h>
13#include <linux/prefetch.h>
14#include <linux/cleancache.h>
15#include "extent_io.h"
16#include "extent_map.h"
17#include "compat.h"
18#include "ctree.h"
19#include "btrfs_inode.h"
20
21static struct kmem_cache *extent_state_cache;
22static struct kmem_cache *extent_buffer_cache;
23
24static LIST_HEAD(buffers);
25static LIST_HEAD(states);
26
27#define LEAK_DEBUG 0
28#if LEAK_DEBUG
29static DEFINE_SPINLOCK(leak_lock);
30#endif
31
32#define BUFFER_LRU_MAX 64
33
34struct tree_entry {
35 u64 start;
36 u64 end;
37 struct rb_node rb_node;
38};
39
40struct extent_page_data {
41 struct bio *bio;
42 struct extent_io_tree *tree;
43 get_extent_t *get_extent;
44
45 /* tells writepage not to lock the state bits for this range
46 * it still does the unlocking
47 */
48 unsigned int extent_locked:1;
49
50 /* tells the submit_bio code to use a WRITE_SYNC */
51 unsigned int sync_io:1;
52};
53
54int __init extent_io_init(void)
55{
56 extent_state_cache = kmem_cache_create("extent_state",
57 sizeof(struct extent_state), 0,
58 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
59 if (!extent_state_cache)
60 return -ENOMEM;
61
62 extent_buffer_cache = kmem_cache_create("extent_buffers",
63 sizeof(struct extent_buffer), 0,
64 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
65 if (!extent_buffer_cache)
66 goto free_state_cache;
67 return 0;
68
69free_state_cache:
70 kmem_cache_destroy(extent_state_cache);
71 return -ENOMEM;
72}
73
74void extent_io_exit(void)
75{
76 struct extent_state *state;
77 struct extent_buffer *eb;
78
79 while (!list_empty(&states)) {
80 state = list_entry(states.next, struct extent_state, leak_list);
81 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
82 "state %lu in tree %p refs %d\n",
83 (unsigned long long)state->start,
84 (unsigned long long)state->end,
85 state->state, state->tree, atomic_read(&state->refs));
86 list_del(&state->leak_list);
87 kmem_cache_free(extent_state_cache, state);
88
89 }
90
91 while (!list_empty(&buffers)) {
92 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
93 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
94 "refs %d\n", (unsigned long long)eb->start,
95 eb->len, atomic_read(&eb->refs));
96 list_del(&eb->leak_list);
97 kmem_cache_free(extent_buffer_cache, eb);
98 }
99 if (extent_state_cache)
100 kmem_cache_destroy(extent_state_cache);
101 if (extent_buffer_cache)
102 kmem_cache_destroy(extent_buffer_cache);
103}
104
105void extent_io_tree_init(struct extent_io_tree *tree,
106 struct address_space *mapping)
107{
108 tree->state = RB_ROOT;
109 INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
110 tree->ops = NULL;
111 tree->dirty_bytes = 0;
112 spin_lock_init(&tree->lock);
113 spin_lock_init(&tree->buffer_lock);
114 tree->mapping = mapping;
115}
116
117static struct extent_state *alloc_extent_state(gfp_t mask)
118{
119 struct extent_state *state;
120#if LEAK_DEBUG
121 unsigned long flags;
122#endif
123
124 state = kmem_cache_alloc(extent_state_cache, mask);
125 if (!state)
126 return state;
127 state->state = 0;
128 state->private = 0;
129 state->tree = NULL;
130#if LEAK_DEBUG
131 spin_lock_irqsave(&leak_lock, flags);
132 list_add(&state->leak_list, &states);
133 spin_unlock_irqrestore(&leak_lock, flags);
134#endif
135 atomic_set(&state->refs, 1);
136 init_waitqueue_head(&state->wq);
137 return state;
138}
139
140void free_extent_state(struct extent_state *state)
141{
142 if (!state)
143 return;
144 if (atomic_dec_and_test(&state->refs)) {
145#if LEAK_DEBUG
146 unsigned long flags;
147#endif
148 WARN_ON(state->tree);
149#if LEAK_DEBUG
150 spin_lock_irqsave(&leak_lock, flags);
151 list_del(&state->leak_list);
152 spin_unlock_irqrestore(&leak_lock, flags);
153#endif
154 kmem_cache_free(extent_state_cache, state);
155 }
156}
157
158static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
159 struct rb_node *node)
160{
161 struct rb_node **p = &root->rb_node;
162 struct rb_node *parent = NULL;
163 struct tree_entry *entry;
164
165 while (*p) {
166 parent = *p;
167 entry = rb_entry(parent, struct tree_entry, rb_node);
168
169 if (offset < entry->start)
170 p = &(*p)->rb_left;
171 else if (offset > entry->end)
172 p = &(*p)->rb_right;
173 else
174 return parent;
175 }
176
177 entry = rb_entry(node, struct tree_entry, rb_node);
178 rb_link_node(node, parent, p);
179 rb_insert_color(node, root);
180 return NULL;
181}
182
183static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
184 struct rb_node **prev_ret,
185 struct rb_node **next_ret)
186{
187 struct rb_root *root = &tree->state;
188 struct rb_node *n = root->rb_node;
189 struct rb_node *prev = NULL;
190 struct rb_node *orig_prev = NULL;
191 struct tree_entry *entry;
192 struct tree_entry *prev_entry = NULL;
193
194 while (n) {
195 entry = rb_entry(n, struct tree_entry, rb_node);
196 prev = n;
197 prev_entry = entry;
198
199 if (offset < entry->start)
200 n = n->rb_left;
201 else if (offset > entry->end)
202 n = n->rb_right;
203 else
204 return n;
205 }
206
207 if (prev_ret) {
208 orig_prev = prev;
209 while (prev && offset > prev_entry->end) {
210 prev = rb_next(prev);
211 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
212 }
213 *prev_ret = prev;
214 prev = orig_prev;
215 }
216
217 if (next_ret) {
218 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
219 while (prev && offset < prev_entry->start) {
220 prev = rb_prev(prev);
221 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
222 }
223 *next_ret = prev;
224 }
225 return NULL;
226}
227
228static inline struct rb_node *tree_search(struct extent_io_tree *tree,
229 u64 offset)
230{
231 struct rb_node *prev = NULL;
232 struct rb_node *ret;
233
234 ret = __etree_search(tree, offset, &prev, NULL);
235 if (!ret)
236 return prev;
237 return ret;
238}
239
240static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
241 struct extent_state *other)
242{
243 if (tree->ops && tree->ops->merge_extent_hook)
244 tree->ops->merge_extent_hook(tree->mapping->host, new,
245 other);
246}
247
248/*
249 * utility function to look for merge candidates inside a given range.
250 * Any extents with matching state are merged together into a single
251 * extent in the tree. Extents with EXTENT_IO in their state field
252 * are not merged because the end_io handlers need to be able to do
253 * operations on them without sleeping (or doing allocations/splits).
254 *
255 * This should be called with the tree lock held.
256 */
257static void merge_state(struct extent_io_tree *tree,
258 struct extent_state *state)
259{
260 struct extent_state *other;
261 struct rb_node *other_node;
262
263 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
264 return;
265
266 other_node = rb_prev(&state->rb_node);
267 if (other_node) {
268 other = rb_entry(other_node, struct extent_state, rb_node);
269 if (other->end == state->start - 1 &&
270 other->state == state->state) {
271 merge_cb(tree, state, other);
272 state->start = other->start;
273 other->tree = NULL;
274 rb_erase(&other->rb_node, &tree->state);
275 free_extent_state(other);
276 }
277 }
278 other_node = rb_next(&state->rb_node);
279 if (other_node) {
280 other = rb_entry(other_node, struct extent_state, rb_node);
281 if (other->start == state->end + 1 &&
282 other->state == state->state) {
283 merge_cb(tree, state, other);
284 state->end = other->end;
285 other->tree = NULL;
286 rb_erase(&other->rb_node, &tree->state);
287 free_extent_state(other);
288 }
289 }
290}
291
292static void set_state_cb(struct extent_io_tree *tree,
293 struct extent_state *state, int *bits)
294{
295 if (tree->ops && tree->ops->set_bit_hook)
296 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
297}
298
299static void clear_state_cb(struct extent_io_tree *tree,
300 struct extent_state *state, int *bits)
301{
302 if (tree->ops && tree->ops->clear_bit_hook)
303 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
304}
305
306static void set_state_bits(struct extent_io_tree *tree,
307 struct extent_state *state, int *bits);
308
309/*
310 * insert an extent_state struct into the tree. 'bits' are set on the
311 * struct before it is inserted.
312 *
313 * This may return -EEXIST if the extent is already there, in which case the
314 * state struct is freed.
315 *
316 * The tree lock is not taken internally. This is a utility function and
317 * probably isn't what you want to call (see set/clear_extent_bit).
318 */
319static int insert_state(struct extent_io_tree *tree,
320 struct extent_state *state, u64 start, u64 end,
321 int *bits)
322{
323 struct rb_node *node;
324
325 if (end < start) {
326 printk(KERN_ERR "btrfs end < start %llu %llu\n",
327 (unsigned long long)end,
328 (unsigned long long)start);
329 WARN_ON(1);
330 }
331 state->start = start;
332 state->end = end;
333
334 set_state_bits(tree, state, bits);
335
336 node = tree_insert(&tree->state, end, &state->rb_node);
337 if (node) {
338 struct extent_state *found;
339 found = rb_entry(node, struct extent_state, rb_node);
340 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
341 "%llu %llu\n", (unsigned long long)found->start,
342 (unsigned long long)found->end,
343 (unsigned long long)start, (unsigned long long)end);
344 return -EEXIST;
345 }
346 state->tree = tree;
347 merge_state(tree, state);
348 return 0;
349}
350
351static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
352 u64 split)
353{
354 if (tree->ops && tree->ops->split_extent_hook)
355 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
356}
357
358/*
359 * split a given extent state struct in two, inserting the preallocated
360 * struct 'prealloc' as the newly created second half. 'split' indicates an
361 * offset inside 'orig' where it should be split.
362 *
363 * Before calling,
364 * the tree has 'orig' at [orig->start, orig->end]. After calling, there
365 * are two extent state structs in the tree:
366 * prealloc: [orig->start, split - 1]
367 * orig: [ split, orig->end ]
368 *
369 * The tree locks are not taken by this function. They need to be held
370 * by the caller.
371 */
372static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
373 struct extent_state *prealloc, u64 split)
374{
375 struct rb_node *node;
376
377 split_cb(tree, orig, split);
378
379 prealloc->start = orig->start;
380 prealloc->end = split - 1;
381 prealloc->state = orig->state;
382 orig->start = split;
383
384 node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
385 if (node) {
386 free_extent_state(prealloc);
387 return -EEXIST;
388 }
389 prealloc->tree = tree;
390 return 0;
391}
392
393/*
394 * utility function to clear some bits in an extent state struct.
395 * it will optionally wake up any one waiting on this state (wake == 1), or
396 * forcibly remove the state from the tree (delete == 1).
397 *
398 * If no bits are set on the state struct after clearing things, the
399 * struct is freed and removed from the tree
400 */
401static int clear_state_bit(struct extent_io_tree *tree,
402 struct extent_state *state,
403 int *bits, int wake)
404{
405 int bits_to_clear = *bits & ~EXTENT_CTLBITS;
406 int ret = state->state & bits_to_clear;
407
408 if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
409 u64 range = state->end - state->start + 1;
410 WARN_ON(range > tree->dirty_bytes);
411 tree->dirty_bytes -= range;
412 }
413 clear_state_cb(tree, state, bits);
414 state->state &= ~bits_to_clear;
415 if (wake)
416 wake_up(&state->wq);
417 if (state->state == 0) {
418 if (state->tree) {
419 rb_erase(&state->rb_node, &tree->state);
420 state->tree = NULL;
421 free_extent_state(state);
422 } else {
423 WARN_ON(1);
424 }
425 } else {
426 merge_state(tree, state);
427 }
428 return ret;
429}
430
431static struct extent_state *
432alloc_extent_state_atomic(struct extent_state *prealloc)
433{
434 if (!prealloc)
435 prealloc = alloc_extent_state(GFP_ATOMIC);
436
437 return prealloc;
438}
439
440/*
441 * clear some bits on a range in the tree. This may require splitting
442 * or inserting elements in the tree, so the gfp mask is used to
443 * indicate which allocations or sleeping are allowed.
444 *
445 * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
446 * the given range from the tree regardless of state (ie for truncate).
447 *
448 * the range [start, end] is inclusive.
449 *
450 * This takes the tree lock, and returns < 0 on error, > 0 if any of the
451 * bits were already set, or zero if none of the bits were already set.
452 */
453int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
454 int bits, int wake, int delete,
455 struct extent_state **cached_state,
456 gfp_t mask)
457{
458 struct extent_state *state;
459 struct extent_state *cached;
460 struct extent_state *prealloc = NULL;
461 struct rb_node *next_node;
462 struct rb_node *node;
463 u64 last_end;
464 int err;
465 int set = 0;
466 int clear = 0;
467
468 if (delete)
469 bits |= ~EXTENT_CTLBITS;
470 bits |= EXTENT_FIRST_DELALLOC;
471
472 if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
473 clear = 1;
474again:
475 if (!prealloc && (mask & __GFP_WAIT)) {
476 prealloc = alloc_extent_state(mask);
477 if (!prealloc)
478 return -ENOMEM;
479 }
480
481 spin_lock(&tree->lock);
482 if (cached_state) {
483 cached = *cached_state;
484
485 if (clear) {
486 *cached_state = NULL;
487 cached_state = NULL;
488 }
489
490 if (cached && cached->tree && cached->start <= start &&
491 cached->end > start) {
492 if (clear)
493 atomic_dec(&cached->refs);
494 state = cached;
495 goto hit_next;
496 }
497 if (clear)
498 free_extent_state(cached);
499 }
500 /*
501 * this search will find the extents that end after
502 * our range starts
503 */
504 node = tree_search(tree, start);
505 if (!node)
506 goto out;
507 state = rb_entry(node, struct extent_state, rb_node);
508hit_next:
509 if (state->start > end)
510 goto out;
511 WARN_ON(state->end < start);
512 last_end = state->end;
513
514 /*
515 * | ---- desired range ---- |
516 * | state | or
517 * | ------------- state -------------- |
518 *
519 * We need to split the extent we found, and may flip
520 * bits on second half.
521 *
522 * If the extent we found extends past our range, we
523 * just split and search again. It'll get split again
524 * the next time though.
525 *
526 * If the extent we found is inside our range, we clear
527 * the desired bit on it.
528 */
529
530 if (state->start < start) {
531 prealloc = alloc_extent_state_atomic(prealloc);
532 BUG_ON(!prealloc);
533 err = split_state(tree, state, prealloc, start);
534 BUG_ON(err == -EEXIST);
535 prealloc = NULL;
536 if (err)
537 goto out;
538 if (state->end <= end) {
539 set |= clear_state_bit(tree, state, &bits, wake);
540 if (last_end == (u64)-1)
541 goto out;
542 start = last_end + 1;
543 }
544 goto search_again;
545 }
546 /*
547 * | ---- desired range ---- |
548 * | state |
549 * We need to split the extent, and clear the bit
550 * on the first half
551 */
552 if (state->start <= end && state->end > end) {
553 prealloc = alloc_extent_state_atomic(prealloc);
554 BUG_ON(!prealloc);
555 err = split_state(tree, state, prealloc, end + 1);
556 BUG_ON(err == -EEXIST);
557 if (wake)
558 wake_up(&state->wq);
559
560 set |= clear_state_bit(tree, prealloc, &bits, wake);
561
562 prealloc = NULL;
563 goto out;
564 }
565
566 if (state->end < end && prealloc && !need_resched())
567 next_node = rb_next(&state->rb_node);
568 else
569 next_node = NULL;
570
571 set |= clear_state_bit(tree, state, &bits, wake);
572 if (last_end == (u64)-1)
573 goto out;
574 start = last_end + 1;
575 if (start <= end && next_node) {
576 state = rb_entry(next_node, struct extent_state,
577 rb_node);
578 if (state->start == start)
579 goto hit_next;
580 }
581 goto search_again;
582
583out:
584 spin_unlock(&tree->lock);
585 if (prealloc)
586 free_extent_state(prealloc);
587
588 return set;
589
590search_again:
591 if (start > end)
592 goto out;
593 spin_unlock(&tree->lock);
594 if (mask & __GFP_WAIT)
595 cond_resched();
596 goto again;
597}
598
599static int wait_on_state(struct extent_io_tree *tree,
600 struct extent_state *state)
601 __releases(tree->lock)
602 __acquires(tree->lock)
603{
604 DEFINE_WAIT(wait);
605 prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
606 spin_unlock(&tree->lock);
607 schedule();
608 spin_lock(&tree->lock);
609 finish_wait(&state->wq, &wait);
610 return 0;
611}
612
613/*
614 * waits for one or more bits to clear on a range in the state tree.
615 * The range [start, end] is inclusive.
616 * The tree lock is taken by this function
617 */
618int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
619{
620 struct extent_state *state;
621 struct rb_node *node;
622
623 spin_lock(&tree->lock);
624again:
625 while (1) {
626 /*
627 * this search will find all the extents that end after
628 * our range starts
629 */
630 node = tree_search(tree, start);
631 if (!node)
632 break;
633
634 state = rb_entry(node, struct extent_state, rb_node);
635
636 if (state->start > end)
637 goto out;
638
639 if (state->state & bits) {
640 start = state->start;
641 atomic_inc(&state->refs);
642 wait_on_state(tree, state);
643 free_extent_state(state);
644 goto again;
645 }
646 start = state->end + 1;
647
648 if (start > end)
649 break;
650
651 cond_resched_lock(&tree->lock);
652 }
653out:
654 spin_unlock(&tree->lock);
655 return 0;
656}
657
658static void set_state_bits(struct extent_io_tree *tree,
659 struct extent_state *state,
660 int *bits)
661{
662 int bits_to_set = *bits & ~EXTENT_CTLBITS;
663
664 set_state_cb(tree, state, bits);
665 if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
666 u64 range = state->end - state->start + 1;
667 tree->dirty_bytes += range;
668 }
669 state->state |= bits_to_set;
670}
671
672static void cache_state(struct extent_state *state,
673 struct extent_state **cached_ptr)
674{
675 if (cached_ptr && !(*cached_ptr)) {
676 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
677 *cached_ptr = state;
678 atomic_inc(&state->refs);
679 }
680 }
681}
682
683static void uncache_state(struct extent_state **cached_ptr)
684{
685 if (cached_ptr && (*cached_ptr)) {
686 struct extent_state *state = *cached_ptr;
687 *cached_ptr = NULL;
688 free_extent_state(state);
689 }
690}
691
692/*
693 * set some bits on a range in the tree. This may require allocations or
694 * sleeping, so the gfp mask is used to indicate what is allowed.
695 *
696 * If any of the exclusive bits are set, this will fail with -EEXIST if some
697 * part of the range already has the desired bits set. The start of the
698 * existing range is returned in failed_start in this case.
699 *
700 * [start, end] is inclusive This takes the tree lock.
701 */
702
703int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
704 int bits, int exclusive_bits, u64 *failed_start,
705 struct extent_state **cached_state, gfp_t mask)
706{
707 struct extent_state *state;
708 struct extent_state *prealloc = NULL;
709 struct rb_node *node;
710 int err = 0;
711 u64 last_start;
712 u64 last_end;
713
714 bits |= EXTENT_FIRST_DELALLOC;
715again:
716 if (!prealloc && (mask & __GFP_WAIT)) {
717 prealloc = alloc_extent_state(mask);
718 BUG_ON(!prealloc);
719 }
720
721 spin_lock(&tree->lock);
722 if (cached_state && *cached_state) {
723 state = *cached_state;
724 if (state->start <= start && state->end > start &&
725 state->tree) {
726 node = &state->rb_node;
727 goto hit_next;
728 }
729 }
730 /*
731 * this search will find all the extents that end after
732 * our range starts.
733 */
734 node = tree_search(tree, start);
735 if (!node) {
736 prealloc = alloc_extent_state_atomic(prealloc);
737 BUG_ON(!prealloc);
738 err = insert_state(tree, prealloc, start, end, &bits);
739 prealloc = NULL;
740 BUG_ON(err == -EEXIST);
741 goto out;
742 }
743 state = rb_entry(node, struct extent_state, rb_node);
744hit_next:
745 last_start = state->start;
746 last_end = state->end;
747
748 /*
749 * | ---- desired range ---- |
750 * | state |
751 *
752 * Just lock what we found and keep going
753 */
754 if (state->start == start && state->end <= end) {
755 struct rb_node *next_node;
756 if (state->state & exclusive_bits) {
757 *failed_start = state->start;
758 err = -EEXIST;
759 goto out;
760 }
761
762 set_state_bits(tree, state, &bits);
763
764 cache_state(state, cached_state);
765 merge_state(tree, state);
766 if (last_end == (u64)-1)
767 goto out;
768
769 start = last_end + 1;
770 next_node = rb_next(&state->rb_node);
771 if (next_node && start < end && prealloc && !need_resched()) {
772 state = rb_entry(next_node, struct extent_state,
773 rb_node);
774 if (state->start == start)
775 goto hit_next;
776 }
777 goto search_again;
778 }
779
780 /*
781 * | ---- desired range ---- |
782 * | state |
783 * or
784 * | ------------- state -------------- |
785 *
786 * We need to split the extent we found, and may flip bits on
787 * second half.
788 *
789 * If the extent we found extends past our
790 * range, we just split and search again. It'll get split
791 * again the next time though.
792 *
793 * If the extent we found is inside our range, we set the
794 * desired bit on it.
795 */
796 if (state->start < start) {
797 if (state->state & exclusive_bits) {
798 *failed_start = start;
799 err = -EEXIST;
800 goto out;
801 }
802
803 prealloc = alloc_extent_state_atomic(prealloc);
804 BUG_ON(!prealloc);
805 err = split_state(tree, state, prealloc, start);
806 BUG_ON(err == -EEXIST);
807 prealloc = NULL;
808 if (err)
809 goto out;
810 if (state->end <= end) {
811 set_state_bits(tree, state, &bits);
812 cache_state(state, cached_state);
813 merge_state(tree, state);
814 if (last_end == (u64)-1)
815 goto out;
816 start = last_end + 1;
817 }
818 goto search_again;
819 }
820 /*
821 * | ---- desired range ---- |
822 * | state | or | state |
823 *
824 * There's a hole, we need to insert something in it and
825 * ignore the extent we found.
826 */
827 if (state->start > start) {
828 u64 this_end;
829 if (end < last_start)
830 this_end = end;
831 else
832 this_end = last_start - 1;
833
834 prealloc = alloc_extent_state_atomic(prealloc);
835 BUG_ON(!prealloc);
836
837 /*
838 * Avoid to free 'prealloc' if it can be merged with
839 * the later extent.
840 */
841 err = insert_state(tree, prealloc, start, this_end,
842 &bits);
843 BUG_ON(err == -EEXIST);
844 if (err) {
845 free_extent_state(prealloc);
846 prealloc = NULL;
847 goto out;
848 }
849 cache_state(prealloc, cached_state);
850 prealloc = NULL;
851 start = this_end + 1;
852 goto search_again;
853 }
854 /*
855 * | ---- desired range ---- |
856 * | state |
857 * We need to split the extent, and set the bit
858 * on the first half
859 */
860 if (state->start <= end && state->end > end) {
861 if (state->state & exclusive_bits) {
862 *failed_start = start;
863 err = -EEXIST;
864 goto out;
865 }
866
867 prealloc = alloc_extent_state_atomic(prealloc);
868 BUG_ON(!prealloc);
869 err = split_state(tree, state, prealloc, end + 1);
870 BUG_ON(err == -EEXIST);
871
872 set_state_bits(tree, prealloc, &bits);
873 cache_state(prealloc, cached_state);
874 merge_state(tree, prealloc);
875 prealloc = NULL;
876 goto out;
877 }
878
879 goto search_again;
880
881out:
882 spin_unlock(&tree->lock);
883 if (prealloc)
884 free_extent_state(prealloc);
885
886 return err;
887
888search_again:
889 if (start > end)
890 goto out;
891 spin_unlock(&tree->lock);
892 if (mask & __GFP_WAIT)
893 cond_resched();
894 goto again;
895}
896
897/* wrappers around set/clear extent bit */
898int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
899 gfp_t mask)
900{
901 return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
902 NULL, mask);
903}
904
905int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
906 int bits, gfp_t mask)
907{
908 return set_extent_bit(tree, start, end, bits, 0, NULL,
909 NULL, mask);
910}
911
912int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
913 int bits, gfp_t mask)
914{
915 return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
916}
917
918int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
919 struct extent_state **cached_state, gfp_t mask)
920{
921 return set_extent_bit(tree, start, end,
922 EXTENT_DELALLOC | EXTENT_DIRTY | EXTENT_UPTODATE,
923 0, NULL, cached_state, mask);
924}
925
926int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
927 gfp_t mask)
928{
929 return clear_extent_bit(tree, start, end,
930 EXTENT_DIRTY | EXTENT_DELALLOC |
931 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
932}
933
934int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
935 gfp_t mask)
936{
937 return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
938 NULL, mask);
939}
940
941int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
942 struct extent_state **cached_state, gfp_t mask)
943{
944 return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
945 NULL, cached_state, mask);
946}
947
948static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
949 u64 end, struct extent_state **cached_state,
950 gfp_t mask)
951{
952 return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
953 cached_state, mask);
954}
955
956/*
957 * either insert or lock state struct between start and end use mask to tell
958 * us if waiting is desired.
959 */
960int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
961 int bits, struct extent_state **cached_state, gfp_t mask)
962{
963 int err;
964 u64 failed_start;
965 while (1) {
966 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
967 EXTENT_LOCKED, &failed_start,
968 cached_state, mask);
969 if (err == -EEXIST && (mask & __GFP_WAIT)) {
970 wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
971 start = failed_start;
972 } else {
973 break;
974 }
975 WARN_ON(start > end);
976 }
977 return err;
978}
979
980int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
981{
982 return lock_extent_bits(tree, start, end, 0, NULL, mask);
983}
984
985int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
986 gfp_t mask)
987{
988 int err;
989 u64 failed_start;
990
991 err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
992 &failed_start, NULL, mask);
993 if (err == -EEXIST) {
994 if (failed_start > start)
995 clear_extent_bit(tree, start, failed_start - 1,
996 EXTENT_LOCKED, 1, 0, NULL, mask);
997 return 0;
998 }
999 return 1;
1000}
1001
1002int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1003 struct extent_state **cached, gfp_t mask)
1004{
1005 return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1006 mask);
1007}
1008
1009int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1010{
1011 return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1012 mask);
1013}
1014
1015/*
1016 * helper function to set both pages and extents in the tree writeback
1017 */
1018static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1019{
1020 unsigned long index = start >> PAGE_CACHE_SHIFT;
1021 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1022 struct page *page;
1023
1024 while (index <= end_index) {
1025 page = find_get_page(tree->mapping, index);
1026 BUG_ON(!page);
1027 set_page_writeback(page);
1028 page_cache_release(page);
1029 index++;
1030 }
1031 return 0;
1032}
1033
1034/* find the first state struct with 'bits' set after 'start', and
1035 * return it. tree->lock must be held. NULL will returned if
1036 * nothing was found after 'start'
1037 */
1038struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1039 u64 start, int bits)
1040{
1041 struct rb_node *node;
1042 struct extent_state *state;
1043
1044 /*
1045 * this search will find all the extents that end after
1046 * our range starts.
1047 */
1048 node = tree_search(tree, start);
1049 if (!node)
1050 goto out;
1051
1052 while (1) {
1053 state = rb_entry(node, struct extent_state, rb_node);
1054 if (state->end >= start && (state->state & bits))
1055 return state;
1056
1057 node = rb_next(node);
1058 if (!node)
1059 break;
1060 }
1061out:
1062 return NULL;
1063}
1064
1065/*
1066 * find the first offset in the io tree with 'bits' set. zero is
1067 * returned if we find something, and *start_ret and *end_ret are
1068 * set to reflect the state struct that was found.
1069 *
1070 * If nothing was found, 1 is returned, < 0 on error
1071 */
1072int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1073 u64 *start_ret, u64 *end_ret, int bits)
1074{
1075 struct extent_state *state;
1076 int ret = 1;
1077
1078 spin_lock(&tree->lock);
1079 state = find_first_extent_bit_state(tree, start, bits);
1080 if (state) {
1081 *start_ret = state->start;
1082 *end_ret = state->end;
1083 ret = 0;
1084 }
1085 spin_unlock(&tree->lock);
1086 return ret;
1087}
1088
1089/*
1090 * find a contiguous range of bytes in the file marked as delalloc, not
1091 * more than 'max_bytes'. start and end are used to return the range,
1092 *
1093 * 1 is returned if we find something, 0 if nothing was in the tree
1094 */
1095static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1096 u64 *start, u64 *end, u64 max_bytes,
1097 struct extent_state **cached_state)
1098{
1099 struct rb_node *node;
1100 struct extent_state *state;
1101 u64 cur_start = *start;
1102 u64 found = 0;
1103 u64 total_bytes = 0;
1104
1105 spin_lock(&tree->lock);
1106
1107 /*
1108 * this search will find all the extents that end after
1109 * our range starts.
1110 */
1111 node = tree_search(tree, cur_start);
1112 if (!node) {
1113 if (!found)
1114 *end = (u64)-1;
1115 goto out;
1116 }
1117
1118 while (1) {
1119 state = rb_entry(node, struct extent_state, rb_node);
1120 if (found && (state->start != cur_start ||
1121 (state->state & EXTENT_BOUNDARY))) {
1122 goto out;
1123 }
1124 if (!(state->state & EXTENT_DELALLOC)) {
1125 if (!found)
1126 *end = state->end;
1127 goto out;
1128 }
1129 if (!found) {
1130 *start = state->start;
1131 *cached_state = state;
1132 atomic_inc(&state->refs);
1133 }
1134 found++;
1135 *end = state->end;
1136 cur_start = state->end + 1;
1137 node = rb_next(node);
1138 if (!node)
1139 break;
1140 total_bytes += state->end - state->start + 1;
1141 if (total_bytes >= max_bytes)
1142 break;
1143 }
1144out:
1145 spin_unlock(&tree->lock);
1146 return found;
1147}
1148
1149static noinline int __unlock_for_delalloc(struct inode *inode,
1150 struct page *locked_page,
1151 u64 start, u64 end)
1152{
1153 int ret;
1154 struct page *pages[16];
1155 unsigned long index = start >> PAGE_CACHE_SHIFT;
1156 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1157 unsigned long nr_pages = end_index - index + 1;
1158 int i;
1159
1160 if (index == locked_page->index && end_index == index)
1161 return 0;
1162
1163 while (nr_pages > 0) {
1164 ret = find_get_pages_contig(inode->i_mapping, index,
1165 min_t(unsigned long, nr_pages,
1166 ARRAY_SIZE(pages)), pages);
1167 for (i = 0; i < ret; i++) {
1168 if (pages[i] != locked_page)
1169 unlock_page(pages[i]);
1170 page_cache_release(pages[i]);
1171 }
1172 nr_pages -= ret;
1173 index += ret;
1174 cond_resched();
1175 }
1176 return 0;
1177}
1178
1179static noinline int lock_delalloc_pages(struct inode *inode,
1180 struct page *locked_page,
1181 u64 delalloc_start,
1182 u64 delalloc_end)
1183{
1184 unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1185 unsigned long start_index = index;
1186 unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1187 unsigned long pages_locked = 0;
1188 struct page *pages[16];
1189 unsigned long nrpages;
1190 int ret;
1191 int i;
1192
1193 /* the caller is responsible for locking the start index */
1194 if (index == locked_page->index && index == end_index)
1195 return 0;
1196
1197 /* skip the page at the start index */
1198 nrpages = end_index - index + 1;
1199 while (nrpages > 0) {
1200 ret = find_get_pages_contig(inode->i_mapping, index,
1201 min_t(unsigned long,
1202 nrpages, ARRAY_SIZE(pages)), pages);
1203 if (ret == 0) {
1204 ret = -EAGAIN;
1205 goto done;
1206 }
1207 /* now we have an array of pages, lock them all */
1208 for (i = 0; i < ret; i++) {
1209 /*
1210 * the caller is taking responsibility for
1211 * locked_page
1212 */
1213 if (pages[i] != locked_page) {
1214 lock_page(pages[i]);
1215 if (!PageDirty(pages[i]) ||
1216 pages[i]->mapping != inode->i_mapping) {
1217 ret = -EAGAIN;
1218 unlock_page(pages[i]);
1219 page_cache_release(pages[i]);
1220 goto done;
1221 }
1222 }
1223 page_cache_release(pages[i]);
1224 pages_locked++;
1225 }
1226 nrpages -= ret;
1227 index += ret;
1228 cond_resched();
1229 }
1230 ret = 0;
1231done:
1232 if (ret && pages_locked) {
1233 __unlock_for_delalloc(inode, locked_page,
1234 delalloc_start,
1235 ((u64)(start_index + pages_locked - 1)) <<
1236 PAGE_CACHE_SHIFT);
1237 }
1238 return ret;
1239}
1240
1241/*
1242 * find a contiguous range of bytes in the file marked as delalloc, not
1243 * more than 'max_bytes'. start and end are used to return the range,
1244 *
1245 * 1 is returned if we find something, 0 if nothing was in the tree
1246 */
1247static noinline u64 find_lock_delalloc_range(struct inode *inode,
1248 struct extent_io_tree *tree,
1249 struct page *locked_page,
1250 u64 *start, u64 *end,
1251 u64 max_bytes)
1252{
1253 u64 delalloc_start;
1254 u64 delalloc_end;
1255 u64 found;
1256 struct extent_state *cached_state = NULL;
1257 int ret;
1258 int loops = 0;
1259
1260again:
1261 /* step one, find a bunch of delalloc bytes starting at start */
1262 delalloc_start = *start;
1263 delalloc_end = 0;
1264 found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1265 max_bytes, &cached_state);
1266 if (!found || delalloc_end <= *start) {
1267 *start = delalloc_start;
1268 *end = delalloc_end;
1269 free_extent_state(cached_state);
1270 return found;
1271 }
1272
1273 /*
1274 * start comes from the offset of locked_page. We have to lock
1275 * pages in order, so we can't process delalloc bytes before
1276 * locked_page
1277 */
1278 if (delalloc_start < *start)
1279 delalloc_start = *start;
1280
1281 /*
1282 * make sure to limit the number of pages we try to lock down
1283 * if we're looping.
1284 */
1285 if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1286 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1287
1288 /* step two, lock all the pages after the page that has start */
1289 ret = lock_delalloc_pages(inode, locked_page,
1290 delalloc_start, delalloc_end);
1291 if (ret == -EAGAIN) {
1292 /* some of the pages are gone, lets avoid looping by
1293 * shortening the size of the delalloc range we're searching
1294 */
1295 free_extent_state(cached_state);
1296 if (!loops) {
1297 unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1298 max_bytes = PAGE_CACHE_SIZE - offset;
1299 loops = 1;
1300 goto again;
1301 } else {
1302 found = 0;
1303 goto out_failed;
1304 }
1305 }
1306 BUG_ON(ret);
1307
1308 /* step three, lock the state bits for the whole range */
1309 lock_extent_bits(tree, delalloc_start, delalloc_end,
1310 0, &cached_state, GFP_NOFS);
1311
1312 /* then test to make sure it is all still delalloc */
1313 ret = test_range_bit(tree, delalloc_start, delalloc_end,
1314 EXTENT_DELALLOC, 1, cached_state);
1315 if (!ret) {
1316 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1317 &cached_state, GFP_NOFS);
1318 __unlock_for_delalloc(inode, locked_page,
1319 delalloc_start, delalloc_end);
1320 cond_resched();
1321 goto again;
1322 }
1323 free_extent_state(cached_state);
1324 *start = delalloc_start;
1325 *end = delalloc_end;
1326out_failed:
1327 return found;
1328}
1329
1330int extent_clear_unlock_delalloc(struct inode *inode,
1331 struct extent_io_tree *tree,
1332 u64 start, u64 end, struct page *locked_page,
1333 unsigned long op)
1334{
1335 int ret;
1336 struct page *pages[16];
1337 unsigned long index = start >> PAGE_CACHE_SHIFT;
1338 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1339 unsigned long nr_pages = end_index - index + 1;
1340 int i;
1341 int clear_bits = 0;
1342
1343 if (op & EXTENT_CLEAR_UNLOCK)
1344 clear_bits |= EXTENT_LOCKED;
1345 if (op & EXTENT_CLEAR_DIRTY)
1346 clear_bits |= EXTENT_DIRTY;
1347
1348 if (op & EXTENT_CLEAR_DELALLOC)
1349 clear_bits |= EXTENT_DELALLOC;
1350
1351 clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1352 if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1353 EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1354 EXTENT_SET_PRIVATE2)))
1355 return 0;
1356
1357 while (nr_pages > 0) {
1358 ret = find_get_pages_contig(inode->i_mapping, index,
1359 min_t(unsigned long,
1360 nr_pages, ARRAY_SIZE(pages)), pages);
1361 for (i = 0; i < ret; i++) {
1362
1363 if (op & EXTENT_SET_PRIVATE2)
1364 SetPagePrivate2(pages[i]);
1365
1366 if (pages[i] == locked_page) {
1367 page_cache_release(pages[i]);
1368 continue;
1369 }
1370 if (op & EXTENT_CLEAR_DIRTY)
1371 clear_page_dirty_for_io(pages[i]);
1372 if (op & EXTENT_SET_WRITEBACK)
1373 set_page_writeback(pages[i]);
1374 if (op & EXTENT_END_WRITEBACK)
1375 end_page_writeback(pages[i]);
1376 if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1377 unlock_page(pages[i]);
1378 page_cache_release(pages[i]);
1379 }
1380 nr_pages -= ret;
1381 index += ret;
1382 cond_resched();
1383 }
1384 return 0;
1385}
1386
1387/*
1388 * count the number of bytes in the tree that have a given bit(s)
1389 * set. This can be fairly slow, except for EXTENT_DIRTY which is
1390 * cached. The total number found is returned.
1391 */
1392u64 count_range_bits(struct extent_io_tree *tree,
1393 u64 *start, u64 search_end, u64 max_bytes,
1394 unsigned long bits, int contig)
1395{
1396 struct rb_node *node;
1397 struct extent_state *state;
1398 u64 cur_start = *start;
1399 u64 total_bytes = 0;
1400 u64 last = 0;
1401 int found = 0;
1402
1403 if (search_end <= cur_start) {
1404 WARN_ON(1);
1405 return 0;
1406 }
1407
1408 spin_lock(&tree->lock);
1409 if (cur_start == 0 && bits == EXTENT_DIRTY) {
1410 total_bytes = tree->dirty_bytes;
1411 goto out;
1412 }
1413 /*
1414 * this search will find all the extents that end after
1415 * our range starts.
1416 */
1417 node = tree_search(tree, cur_start);
1418 if (!node)
1419 goto out;
1420
1421 while (1) {
1422 state = rb_entry(node, struct extent_state, rb_node);
1423 if (state->start > search_end)
1424 break;
1425 if (contig && found && state->start > last + 1)
1426 break;
1427 if (state->end >= cur_start && (state->state & bits) == bits) {
1428 total_bytes += min(search_end, state->end) + 1 -
1429 max(cur_start, state->start);
1430 if (total_bytes >= max_bytes)
1431 break;
1432 if (!found) {
1433 *start = max(cur_start, state->start);
1434 found = 1;
1435 }
1436 last = state->end;
1437 } else if (contig && found) {
1438 break;
1439 }
1440 node = rb_next(node);
1441 if (!node)
1442 break;
1443 }
1444out:
1445 spin_unlock(&tree->lock);
1446 return total_bytes;
1447}
1448
1449/*
1450 * set the private field for a given byte offset in the tree. If there isn't
1451 * an extent_state there already, this does nothing.
1452 */
1453int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1454{
1455 struct rb_node *node;
1456 struct extent_state *state;
1457 int ret = 0;
1458
1459 spin_lock(&tree->lock);
1460 /*
1461 * this search will find all the extents that end after
1462 * our range starts.
1463 */
1464 node = tree_search(tree, start);
1465 if (!node) {
1466 ret = -ENOENT;
1467 goto out;
1468 }
1469 state = rb_entry(node, struct extent_state, rb_node);
1470 if (state->start != start) {
1471 ret = -ENOENT;
1472 goto out;
1473 }
1474 state->private = private;
1475out:
1476 spin_unlock(&tree->lock);
1477 return ret;
1478}
1479
1480int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1481{
1482 struct rb_node *node;
1483 struct extent_state *state;
1484 int ret = 0;
1485
1486 spin_lock(&tree->lock);
1487 /*
1488 * this search will find all the extents that end after
1489 * our range starts.
1490 */
1491 node = tree_search(tree, start);
1492 if (!node) {
1493 ret = -ENOENT;
1494 goto out;
1495 }
1496 state = rb_entry(node, struct extent_state, rb_node);
1497 if (state->start != start) {
1498 ret = -ENOENT;
1499 goto out;
1500 }
1501 *private = state->private;
1502out:
1503 spin_unlock(&tree->lock);
1504 return ret;
1505}
1506
1507/*
1508 * searches a range in the state tree for a given mask.
1509 * If 'filled' == 1, this returns 1 only if every extent in the tree
1510 * has the bits set. Otherwise, 1 is returned if any bit in the
1511 * range is found set.
1512 */
1513int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1514 int bits, int filled, struct extent_state *cached)
1515{
1516 struct extent_state *state = NULL;
1517 struct rb_node *node;
1518 int bitset = 0;
1519
1520 spin_lock(&tree->lock);
1521 if (cached && cached->tree && cached->start <= start &&
1522 cached->end > start)
1523 node = &cached->rb_node;
1524 else
1525 node = tree_search(tree, start);
1526 while (node && start <= end) {
1527 state = rb_entry(node, struct extent_state, rb_node);
1528
1529 if (filled && state->start > start) {
1530 bitset = 0;
1531 break;
1532 }
1533
1534 if (state->start > end)
1535 break;
1536
1537 if (state->state & bits) {
1538 bitset = 1;
1539 if (!filled)
1540 break;
1541 } else if (filled) {
1542 bitset = 0;
1543 break;
1544 }
1545
1546 if (state->end == (u64)-1)
1547 break;
1548
1549 start = state->end + 1;
1550 if (start > end)
1551 break;
1552 node = rb_next(node);
1553 if (!node) {
1554 if (filled)
1555 bitset = 0;
1556 break;
1557 }
1558 }
1559 spin_unlock(&tree->lock);
1560 return bitset;
1561}
1562
1563/*
1564 * helper function to set a given page up to date if all the
1565 * extents in the tree for that page are up to date
1566 */
1567static int check_page_uptodate(struct extent_io_tree *tree,
1568 struct page *page)
1569{
1570 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1571 u64 end = start + PAGE_CACHE_SIZE - 1;
1572 if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1573 SetPageUptodate(page);
1574 return 0;
1575}
1576
1577/*
1578 * helper function to unlock a page if all the extents in the tree
1579 * for that page are unlocked
1580 */
1581static int check_page_locked(struct extent_io_tree *tree,
1582 struct page *page)
1583{
1584 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1585 u64 end = start + PAGE_CACHE_SIZE - 1;
1586 if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1587 unlock_page(page);
1588 return 0;
1589}
1590
1591/*
1592 * helper function to end page writeback if all the extents
1593 * in the tree for that page are done with writeback
1594 */
1595static int check_page_writeback(struct extent_io_tree *tree,
1596 struct page *page)
1597{
1598 end_page_writeback(page);
1599 return 0;
1600}
1601
1602/* lots and lots of room for performance fixes in the end_bio funcs */
1603
1604/*
1605 * after a writepage IO is done, we need to:
1606 * clear the uptodate bits on error
1607 * clear the writeback bits in the extent tree for this IO
1608 * end_page_writeback if the page has no more pending IO
1609 *
1610 * Scheduling is not allowed, so the extent state tree is expected
1611 * to have one and only one object corresponding to this IO.
1612 */
1613static void end_bio_extent_writepage(struct bio *bio, int err)
1614{
1615 int uptodate = err == 0;
1616 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1617 struct extent_io_tree *tree;
1618 u64 start;
1619 u64 end;
1620 int whole_page;
1621 int ret;
1622
1623 do {
1624 struct page *page = bvec->bv_page;
1625 tree = &BTRFS_I(page->mapping->host)->io_tree;
1626
1627 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1628 bvec->bv_offset;
1629 end = start + bvec->bv_len - 1;
1630
1631 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1632 whole_page = 1;
1633 else
1634 whole_page = 0;
1635
1636 if (--bvec >= bio->bi_io_vec)
1637 prefetchw(&bvec->bv_page->flags);
1638 if (tree->ops && tree->ops->writepage_end_io_hook) {
1639 ret = tree->ops->writepage_end_io_hook(page, start,
1640 end, NULL, uptodate);
1641 if (ret)
1642 uptodate = 0;
1643 }
1644
1645 if (!uptodate && tree->ops &&
1646 tree->ops->writepage_io_failed_hook) {
1647 ret = tree->ops->writepage_io_failed_hook(bio, page,
1648 start, end, NULL);
1649 if (ret == 0) {
1650 uptodate = (err == 0);
1651 continue;
1652 }
1653 }
1654
1655 if (!uptodate) {
1656 clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
1657 ClearPageUptodate(page);
1658 SetPageError(page);
1659 }
1660
1661 if (whole_page)
1662 end_page_writeback(page);
1663 else
1664 check_page_writeback(tree, page);
1665 } while (bvec >= bio->bi_io_vec);
1666
1667 bio_put(bio);
1668}
1669
1670/*
1671 * after a readpage IO is done, we need to:
1672 * clear the uptodate bits on error
1673 * set the uptodate bits if things worked
1674 * set the page up to date if all extents in the tree are uptodate
1675 * clear the lock bit in the extent tree
1676 * unlock the page if there are no other extents locked for it
1677 *
1678 * Scheduling is not allowed, so the extent state tree is expected
1679 * to have one and only one object corresponding to this IO.
1680 */
1681static void end_bio_extent_readpage(struct bio *bio, int err)
1682{
1683 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1684 struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
1685 struct bio_vec *bvec = bio->bi_io_vec;
1686 struct extent_io_tree *tree;
1687 u64 start;
1688 u64 end;
1689 int whole_page;
1690 int ret;
1691
1692 if (err)
1693 uptodate = 0;
1694
1695 do {
1696 struct page *page = bvec->bv_page;
1697 struct extent_state *cached = NULL;
1698 struct extent_state *state;
1699
1700 tree = &BTRFS_I(page->mapping->host)->io_tree;
1701
1702 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1703 bvec->bv_offset;
1704 end = start + bvec->bv_len - 1;
1705
1706 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1707 whole_page = 1;
1708 else
1709 whole_page = 0;
1710
1711 if (++bvec <= bvec_end)
1712 prefetchw(&bvec->bv_page->flags);
1713
1714 spin_lock(&tree->lock);
1715 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
1716 if (state && state->start == start) {
1717 /*
1718 * take a reference on the state, unlock will drop
1719 * the ref
1720 */
1721 cache_state(state, &cached);
1722 }
1723 spin_unlock(&tree->lock);
1724
1725 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1726 ret = tree->ops->readpage_end_io_hook(page, start, end,
1727 state);
1728 if (ret)
1729 uptodate = 0;
1730 }
1731 if (!uptodate && tree->ops &&
1732 tree->ops->readpage_io_failed_hook) {
1733 ret = tree->ops->readpage_io_failed_hook(bio, page,
1734 start, end, NULL);
1735 if (ret == 0) {
1736 uptodate =
1737 test_bit(BIO_UPTODATE, &bio->bi_flags);
1738 if (err)
1739 uptodate = 0;
1740 uncache_state(&cached);
1741 continue;
1742 }
1743 }
1744
1745 if (uptodate) {
1746 set_extent_uptodate(tree, start, end, &cached,
1747 GFP_ATOMIC);
1748 }
1749 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
1750
1751 if (whole_page) {
1752 if (uptodate) {
1753 SetPageUptodate(page);
1754 } else {
1755 ClearPageUptodate(page);
1756 SetPageError(page);
1757 }
1758 unlock_page(page);
1759 } else {
1760 if (uptodate) {
1761 check_page_uptodate(tree, page);
1762 } else {
1763 ClearPageUptodate(page);
1764 SetPageError(page);
1765 }
1766 check_page_locked(tree, page);
1767 }
1768 } while (bvec <= bvec_end);
1769
1770 bio_put(bio);
1771}
1772
1773struct bio *
1774btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1775 gfp_t gfp_flags)
1776{
1777 struct bio *bio;
1778
1779 bio = bio_alloc(gfp_flags, nr_vecs);
1780
1781 if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1782 while (!bio && (nr_vecs /= 2))
1783 bio = bio_alloc(gfp_flags, nr_vecs);
1784 }
1785
1786 if (bio) {
1787 bio->bi_size = 0;
1788 bio->bi_bdev = bdev;
1789 bio->bi_sector = first_sector;
1790 }
1791 return bio;
1792}
1793
1794static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1795 unsigned long bio_flags)
1796{
1797 int ret = 0;
1798 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1799 struct page *page = bvec->bv_page;
1800 struct extent_io_tree *tree = bio->bi_private;
1801 u64 start;
1802
1803 start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1804
1805 bio->bi_private = NULL;
1806
1807 bio_get(bio);
1808
1809 if (tree->ops && tree->ops->submit_bio_hook)
1810 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1811 mirror_num, bio_flags, start);
1812 else
1813 submit_bio(rw, bio);
1814 if (bio_flagged(bio, BIO_EOPNOTSUPP))
1815 ret = -EOPNOTSUPP;
1816 bio_put(bio);
1817 return ret;
1818}
1819
1820static int submit_extent_page(int rw, struct extent_io_tree *tree,
1821 struct page *page, sector_t sector,
1822 size_t size, unsigned long offset,
1823 struct block_device *bdev,
1824 struct bio **bio_ret,
1825 unsigned long max_pages,
1826 bio_end_io_t end_io_func,
1827 int mirror_num,
1828 unsigned long prev_bio_flags,
1829 unsigned long bio_flags)
1830{
1831 int ret = 0;
1832 struct bio *bio;
1833 int nr;
1834 int contig = 0;
1835 int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1836 int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1837 size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1838
1839 if (bio_ret && *bio_ret) {
1840 bio = *bio_ret;
1841 if (old_compressed)
1842 contig = bio->bi_sector == sector;
1843 else
1844 contig = bio->bi_sector + (bio->bi_size >> 9) ==
1845 sector;
1846
1847 if (prev_bio_flags != bio_flags || !contig ||
1848 (tree->ops && tree->ops->merge_bio_hook &&
1849 tree->ops->merge_bio_hook(page, offset, page_size, bio,
1850 bio_flags)) ||
1851 bio_add_page(bio, page, page_size, offset) < page_size) {
1852 ret = submit_one_bio(rw, bio, mirror_num,
1853 prev_bio_flags);
1854 bio = NULL;
1855 } else {
1856 return 0;
1857 }
1858 }
1859 if (this_compressed)
1860 nr = BIO_MAX_PAGES;
1861 else
1862 nr = bio_get_nr_vecs(bdev);
1863
1864 bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1865 if (!bio)
1866 return -ENOMEM;
1867
1868 bio_add_page(bio, page, page_size, offset);
1869 bio->bi_end_io = end_io_func;
1870 bio->bi_private = tree;
1871
1872 if (bio_ret)
1873 *bio_ret = bio;
1874 else
1875 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1876
1877 return ret;
1878}
1879
1880void set_page_extent_mapped(struct page *page)
1881{
1882 if (!PagePrivate(page)) {
1883 SetPagePrivate(page);
1884 page_cache_get(page);
1885 set_page_private(page, EXTENT_PAGE_PRIVATE);
1886 }
1887}
1888
1889static void set_page_extent_head(struct page *page, unsigned long len)
1890{
1891 WARN_ON(!PagePrivate(page));
1892 set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1893}
1894
1895/*
1896 * basic readpage implementation. Locked extent state structs are inserted
1897 * into the tree that are removed when the IO is done (by the end_io
1898 * handlers)
1899 */
1900static int __extent_read_full_page(struct extent_io_tree *tree,
1901 struct page *page,
1902 get_extent_t *get_extent,
1903 struct bio **bio, int mirror_num,
1904 unsigned long *bio_flags)
1905{
1906 struct inode *inode = page->mapping->host;
1907 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1908 u64 page_end = start + PAGE_CACHE_SIZE - 1;
1909 u64 end;
1910 u64 cur = start;
1911 u64 extent_offset;
1912 u64 last_byte = i_size_read(inode);
1913 u64 block_start;
1914 u64 cur_end;
1915 sector_t sector;
1916 struct extent_map *em;
1917 struct block_device *bdev;
1918 struct btrfs_ordered_extent *ordered;
1919 int ret;
1920 int nr = 0;
1921 size_t pg_offset = 0;
1922 size_t iosize;
1923 size_t disk_io_size;
1924 size_t blocksize = inode->i_sb->s_blocksize;
1925 unsigned long this_bio_flag = 0;
1926
1927 set_page_extent_mapped(page);
1928
1929 if (!PageUptodate(page)) {
1930 if (cleancache_get_page(page) == 0) {
1931 BUG_ON(blocksize != PAGE_SIZE);
1932 goto out;
1933 }
1934 }
1935
1936 end = page_end;
1937 while (1) {
1938 lock_extent(tree, start, end, GFP_NOFS);
1939 ordered = btrfs_lookup_ordered_extent(inode, start);
1940 if (!ordered)
1941 break;
1942 unlock_extent(tree, start, end, GFP_NOFS);
1943 btrfs_start_ordered_extent(inode, ordered, 1);
1944 btrfs_put_ordered_extent(ordered);
1945 }
1946
1947 if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
1948 char *userpage;
1949 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
1950
1951 if (zero_offset) {
1952 iosize = PAGE_CACHE_SIZE - zero_offset;
1953 userpage = kmap_atomic(page, KM_USER0);
1954 memset(userpage + zero_offset, 0, iosize);
1955 flush_dcache_page(page);
1956 kunmap_atomic(userpage, KM_USER0);
1957 }
1958 }
1959 while (cur <= end) {
1960 if (cur >= last_byte) {
1961 char *userpage;
1962 struct extent_state *cached = NULL;
1963
1964 iosize = PAGE_CACHE_SIZE - pg_offset;
1965 userpage = kmap_atomic(page, KM_USER0);
1966 memset(userpage + pg_offset, 0, iosize);
1967 flush_dcache_page(page);
1968 kunmap_atomic(userpage, KM_USER0);
1969 set_extent_uptodate(tree, cur, cur + iosize - 1,
1970 &cached, GFP_NOFS);
1971 unlock_extent_cached(tree, cur, cur + iosize - 1,
1972 &cached, GFP_NOFS);
1973 break;
1974 }
1975 em = get_extent(inode, page, pg_offset, cur,
1976 end - cur + 1, 0);
1977 if (IS_ERR_OR_NULL(em)) {
1978 SetPageError(page);
1979 unlock_extent(tree, cur, end, GFP_NOFS);
1980 break;
1981 }
1982 extent_offset = cur - em->start;
1983 BUG_ON(extent_map_end(em) <= cur);
1984 BUG_ON(end < cur);
1985
1986 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
1987 this_bio_flag = EXTENT_BIO_COMPRESSED;
1988 extent_set_compress_type(&this_bio_flag,
1989 em->compress_type);
1990 }
1991
1992 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1993 cur_end = min(extent_map_end(em) - 1, end);
1994 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1995 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
1996 disk_io_size = em->block_len;
1997 sector = em->block_start >> 9;
1998 } else {
1999 sector = (em->block_start + extent_offset) >> 9;
2000 disk_io_size = iosize;
2001 }
2002 bdev = em->bdev;
2003 block_start = em->block_start;
2004 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2005 block_start = EXTENT_MAP_HOLE;
2006 free_extent_map(em);
2007 em = NULL;
2008
2009 /* we've found a hole, just zero and go on */
2010 if (block_start == EXTENT_MAP_HOLE) {
2011 char *userpage;
2012 struct extent_state *cached = NULL;
2013
2014 userpage = kmap_atomic(page, KM_USER0);
2015 memset(userpage + pg_offset, 0, iosize);
2016 flush_dcache_page(page);
2017 kunmap_atomic(userpage, KM_USER0);
2018
2019 set_extent_uptodate(tree, cur, cur + iosize - 1,
2020 &cached, GFP_NOFS);
2021 unlock_extent_cached(tree, cur, cur + iosize - 1,
2022 &cached, GFP_NOFS);
2023 cur = cur + iosize;
2024 pg_offset += iosize;
2025 continue;
2026 }
2027 /* the get_extent function already copied into the page */
2028 if (test_range_bit(tree, cur, cur_end,
2029 EXTENT_UPTODATE, 1, NULL)) {
2030 check_page_uptodate(tree, page);
2031 unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2032 cur = cur + iosize;
2033 pg_offset += iosize;
2034 continue;
2035 }
2036 /* we have an inline extent but it didn't get marked up
2037 * to date. Error out
2038 */
2039 if (block_start == EXTENT_MAP_INLINE) {
2040 SetPageError(page);
2041 unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2042 cur = cur + iosize;
2043 pg_offset += iosize;
2044 continue;
2045 }
2046
2047 ret = 0;
2048 if (tree->ops && tree->ops->readpage_io_hook) {
2049 ret = tree->ops->readpage_io_hook(page, cur,
2050 cur + iosize - 1);
2051 }
2052 if (!ret) {
2053 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2054 pnr -= page->index;
2055 ret = submit_extent_page(READ, tree, page,
2056 sector, disk_io_size, pg_offset,
2057 bdev, bio, pnr,
2058 end_bio_extent_readpage, mirror_num,
2059 *bio_flags,
2060 this_bio_flag);
2061 nr++;
2062 *bio_flags = this_bio_flag;
2063 }
2064 if (ret)
2065 SetPageError(page);
2066 cur = cur + iosize;
2067 pg_offset += iosize;
2068 }
2069out:
2070 if (!nr) {
2071 if (!PageError(page))
2072 SetPageUptodate(page);
2073 unlock_page(page);
2074 }
2075 return 0;
2076}
2077
2078int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2079 get_extent_t *get_extent)
2080{
2081 struct bio *bio = NULL;
2082 unsigned long bio_flags = 0;
2083 int ret;
2084
2085 ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2086 &bio_flags);
2087 if (bio)
2088 ret = submit_one_bio(READ, bio, 0, bio_flags);
2089 return ret;
2090}
2091
2092static noinline void update_nr_written(struct page *page,
2093 struct writeback_control *wbc,
2094 unsigned long nr_written)
2095{
2096 wbc->nr_to_write -= nr_written;
2097 if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2098 wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2099 page->mapping->writeback_index = page->index + nr_written;
2100}
2101
2102/*
2103 * the writepage semantics are similar to regular writepage. extent
2104 * records are inserted to lock ranges in the tree, and as dirty areas
2105 * are found, they are marked writeback. Then the lock bits are removed
2106 * and the end_io handler clears the writeback ranges
2107 */
2108static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2109 void *data)
2110{
2111 struct inode *inode = page->mapping->host;
2112 struct extent_page_data *epd = data;
2113 struct extent_io_tree *tree = epd->tree;
2114 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2115 u64 delalloc_start;
2116 u64 page_end = start + PAGE_CACHE_SIZE - 1;
2117 u64 end;
2118 u64 cur = start;
2119 u64 extent_offset;
2120 u64 last_byte = i_size_read(inode);
2121 u64 block_start;
2122 u64 iosize;
2123 sector_t sector;
2124 struct extent_state *cached_state = NULL;
2125 struct extent_map *em;
2126 struct block_device *bdev;
2127 int ret;
2128 int nr = 0;
2129 size_t pg_offset = 0;
2130 size_t blocksize;
2131 loff_t i_size = i_size_read(inode);
2132 unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2133 u64 nr_delalloc;
2134 u64 delalloc_end;
2135 int page_started;
2136 int compressed;
2137 int write_flags;
2138 unsigned long nr_written = 0;
2139
2140 if (wbc->sync_mode == WB_SYNC_ALL)
2141 write_flags = WRITE_SYNC;
2142 else
2143 write_flags = WRITE;
2144
2145 trace___extent_writepage(page, inode, wbc);
2146
2147 WARN_ON(!PageLocked(page));
2148 pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2149 if (page->index > end_index ||
2150 (page->index == end_index && !pg_offset)) {
2151 page->mapping->a_ops->invalidatepage(page, 0);
2152 unlock_page(page);
2153 return 0;
2154 }
2155
2156 if (page->index == end_index) {
2157 char *userpage;
2158
2159 userpage = kmap_atomic(page, KM_USER0);
2160 memset(userpage + pg_offset, 0,
2161 PAGE_CACHE_SIZE - pg_offset);
2162 kunmap_atomic(userpage, KM_USER0);
2163 flush_dcache_page(page);
2164 }
2165 pg_offset = 0;
2166
2167 set_page_extent_mapped(page);
2168
2169 delalloc_start = start;
2170 delalloc_end = 0;
2171 page_started = 0;
2172 if (!epd->extent_locked) {
2173 u64 delalloc_to_write = 0;
2174 /*
2175 * make sure the wbc mapping index is at least updated
2176 * to this page.
2177 */
2178 update_nr_written(page, wbc, 0);
2179
2180 while (delalloc_end < page_end) {
2181 nr_delalloc = find_lock_delalloc_range(inode, tree,
2182 page,
2183 &delalloc_start,
2184 &delalloc_end,
2185 128 * 1024 * 1024);
2186 if (nr_delalloc == 0) {
2187 delalloc_start = delalloc_end + 1;
2188 continue;
2189 }
2190 tree->ops->fill_delalloc(inode, page, delalloc_start,
2191 delalloc_end, &page_started,
2192 &nr_written);
2193 /*
2194 * delalloc_end is already one less than the total
2195 * length, so we don't subtract one from
2196 * PAGE_CACHE_SIZE
2197 */
2198 delalloc_to_write += (delalloc_end - delalloc_start +
2199 PAGE_CACHE_SIZE) >>
2200 PAGE_CACHE_SHIFT;
2201 delalloc_start = delalloc_end + 1;
2202 }
2203 if (wbc->nr_to_write < delalloc_to_write) {
2204 int thresh = 8192;
2205
2206 if (delalloc_to_write < thresh * 2)
2207 thresh = delalloc_to_write;
2208 wbc->nr_to_write = min_t(u64, delalloc_to_write,
2209 thresh);
2210 }
2211
2212 /* did the fill delalloc function already unlock and start
2213 * the IO?
2214 */
2215 if (page_started) {
2216 ret = 0;
2217 /*
2218 * we've unlocked the page, so we can't update
2219 * the mapping's writeback index, just update
2220 * nr_to_write.
2221 */
2222 wbc->nr_to_write -= nr_written;
2223 goto done_unlocked;
2224 }
2225 }
2226 if (tree->ops && tree->ops->writepage_start_hook) {
2227 ret = tree->ops->writepage_start_hook(page, start,
2228 page_end);
2229 if (ret == -EAGAIN) {
2230 redirty_page_for_writepage(wbc, page);
2231 update_nr_written(page, wbc, nr_written);
2232 unlock_page(page);
2233 ret = 0;
2234 goto done_unlocked;
2235 }
2236 }
2237
2238 /*
2239 * we don't want to touch the inode after unlocking the page,
2240 * so we update the mapping writeback index now
2241 */
2242 update_nr_written(page, wbc, nr_written + 1);
2243
2244 end = page_end;
2245 if (last_byte <= start) {
2246 if (tree->ops && tree->ops->writepage_end_io_hook)
2247 tree->ops->writepage_end_io_hook(page, start,
2248 page_end, NULL, 1);
2249 goto done;
2250 }
2251
2252 blocksize = inode->i_sb->s_blocksize;
2253
2254 while (cur <= end) {
2255 if (cur >= last_byte) {
2256 if (tree->ops && tree->ops->writepage_end_io_hook)
2257 tree->ops->writepage_end_io_hook(page, cur,
2258 page_end, NULL, 1);
2259 break;
2260 }
2261 em = epd->get_extent(inode, page, pg_offset, cur,
2262 end - cur + 1, 1);
2263 if (IS_ERR_OR_NULL(em)) {
2264 SetPageError(page);
2265 break;
2266 }
2267
2268 extent_offset = cur - em->start;
2269 BUG_ON(extent_map_end(em) <= cur);
2270 BUG_ON(end < cur);
2271 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2272 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2273 sector = (em->block_start + extent_offset) >> 9;
2274 bdev = em->bdev;
2275 block_start = em->block_start;
2276 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2277 free_extent_map(em);
2278 em = NULL;
2279
2280 /*
2281 * compressed and inline extents are written through other
2282 * paths in the FS
2283 */
2284 if (compressed || block_start == EXTENT_MAP_HOLE ||
2285 block_start == EXTENT_MAP_INLINE) {
2286 /*
2287 * end_io notification does not happen here for
2288 * compressed extents
2289 */
2290 if (!compressed && tree->ops &&
2291 tree->ops->writepage_end_io_hook)
2292 tree->ops->writepage_end_io_hook(page, cur,
2293 cur + iosize - 1,
2294 NULL, 1);
2295 else if (compressed) {
2296 /* we don't want to end_page_writeback on
2297 * a compressed extent. this happens
2298 * elsewhere
2299 */
2300 nr++;
2301 }
2302
2303 cur += iosize;
2304 pg_offset += iosize;
2305 continue;
2306 }
2307 /* leave this out until we have a page_mkwrite call */
2308 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2309 EXTENT_DIRTY, 0, NULL)) {
2310 cur = cur + iosize;
2311 pg_offset += iosize;
2312 continue;
2313 }
2314
2315 if (tree->ops && tree->ops->writepage_io_hook) {
2316 ret = tree->ops->writepage_io_hook(page, cur,
2317 cur + iosize - 1);
2318 } else {
2319 ret = 0;
2320 }
2321 if (ret) {
2322 SetPageError(page);
2323 } else {
2324 unsigned long max_nr = end_index + 1;
2325
2326 set_range_writeback(tree, cur, cur + iosize - 1);
2327 if (!PageWriteback(page)) {
2328 printk(KERN_ERR "btrfs warning page %lu not "
2329 "writeback, cur %llu end %llu\n",
2330 page->index, (unsigned long long)cur,
2331 (unsigned long long)end);
2332 }
2333
2334 ret = submit_extent_page(write_flags, tree, page,
2335 sector, iosize, pg_offset,
2336 bdev, &epd->bio, max_nr,
2337 end_bio_extent_writepage,
2338 0, 0, 0);
2339 if (ret)
2340 SetPageError(page);
2341 }
2342 cur = cur + iosize;
2343 pg_offset += iosize;
2344 nr++;
2345 }
2346done:
2347 if (nr == 0) {
2348 /* make sure the mapping tag for page dirty gets cleared */
2349 set_page_writeback(page);
2350 end_page_writeback(page);
2351 }
2352 unlock_page(page);
2353
2354done_unlocked:
2355
2356 /* drop our reference on any cached states */
2357 free_extent_state(cached_state);
2358 return 0;
2359}
2360
2361/**
2362 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2363 * @mapping: address space structure to write
2364 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2365 * @writepage: function called for each page
2366 * @data: data passed to writepage function
2367 *
2368 * If a page is already under I/O, write_cache_pages() skips it, even
2369 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2370 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2371 * and msync() need to guarantee that all the data which was dirty at the time
2372 * the call was made get new I/O started against them. If wbc->sync_mode is
2373 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2374 * existing IO to complete.
2375 */
2376static int extent_write_cache_pages(struct extent_io_tree *tree,
2377 struct address_space *mapping,
2378 struct writeback_control *wbc,
2379 writepage_t writepage, void *data,
2380 void (*flush_fn)(void *))
2381{
2382 int ret = 0;
2383 int done = 0;
2384 int nr_to_write_done = 0;
2385 struct pagevec pvec;
2386 int nr_pages;
2387 pgoff_t index;
2388 pgoff_t end; /* Inclusive */
2389 int scanned = 0;
2390 int tag;
2391
2392 pagevec_init(&pvec, 0);
2393 if (wbc->range_cyclic) {
2394 index = mapping->writeback_index; /* Start from prev offset */
2395 end = -1;
2396 } else {
2397 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2398 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2399 scanned = 1;
2400 }
2401 if (wbc->sync_mode == WB_SYNC_ALL)
2402 tag = PAGECACHE_TAG_TOWRITE;
2403 else
2404 tag = PAGECACHE_TAG_DIRTY;
2405retry:
2406 if (wbc->sync_mode == WB_SYNC_ALL)
2407 tag_pages_for_writeback(mapping, index, end);
2408 while (!done && !nr_to_write_done && (index <= end) &&
2409 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2410 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2411 unsigned i;
2412
2413 scanned = 1;
2414 for (i = 0; i < nr_pages; i++) {
2415 struct page *page = pvec.pages[i];
2416
2417 /*
2418 * At this point we hold neither mapping->tree_lock nor
2419 * lock on the page itself: the page may be truncated or
2420 * invalidated (changing page->mapping to NULL), or even
2421 * swizzled back from swapper_space to tmpfs file
2422 * mapping
2423 */
2424 if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2425 tree->ops->write_cache_pages_lock_hook(page);
2426 else
2427 lock_page(page);
2428
2429 if (unlikely(page->mapping != mapping)) {
2430 unlock_page(page);
2431 continue;
2432 }
2433
2434 if (!wbc->range_cyclic && page->index > end) {
2435 done = 1;
2436 unlock_page(page);
2437 continue;
2438 }
2439
2440 if (wbc->sync_mode != WB_SYNC_NONE) {
2441 if (PageWriteback(page))
2442 flush_fn(data);
2443 wait_on_page_writeback(page);
2444 }
2445
2446 if (PageWriteback(page) ||
2447 !clear_page_dirty_for_io(page)) {
2448 unlock_page(page);
2449 continue;
2450 }
2451
2452 ret = (*writepage)(page, wbc, data);
2453
2454 if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2455 unlock_page(page);
2456 ret = 0;
2457 }
2458 if (ret)
2459 done = 1;
2460
2461 /*
2462 * the filesystem may choose to bump up nr_to_write.
2463 * We have to make sure to honor the new nr_to_write
2464 * at any time
2465 */
2466 nr_to_write_done = wbc->nr_to_write <= 0;
2467 }
2468 pagevec_release(&pvec);
2469 cond_resched();
2470 }
2471 if (!scanned && !done) {
2472 /*
2473 * We hit the last page and there is more work to be done: wrap
2474 * back to the start of the file
2475 */
2476 scanned = 1;
2477 index = 0;
2478 goto retry;
2479 }
2480 return ret;
2481}
2482
2483static void flush_epd_write_bio(struct extent_page_data *epd)
2484{
2485 if (epd->bio) {
2486 if (epd->sync_io)
2487 submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2488 else
2489 submit_one_bio(WRITE, epd->bio, 0, 0);
2490 epd->bio = NULL;
2491 }
2492}
2493
2494static noinline void flush_write_bio(void *data)
2495{
2496 struct extent_page_data *epd = data;
2497 flush_epd_write_bio(epd);
2498}
2499
2500int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2501 get_extent_t *get_extent,
2502 struct writeback_control *wbc)
2503{
2504 int ret;
2505 struct extent_page_data epd = {
2506 .bio = NULL,
2507 .tree = tree,
2508 .get_extent = get_extent,
2509 .extent_locked = 0,
2510 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2511 };
2512
2513 ret = __extent_writepage(page, wbc, &epd);
2514
2515 flush_epd_write_bio(&epd);
2516 return ret;
2517}
2518
2519int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2520 u64 start, u64 end, get_extent_t *get_extent,
2521 int mode)
2522{
2523 int ret = 0;
2524 struct address_space *mapping = inode->i_mapping;
2525 struct page *page;
2526 unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2527 PAGE_CACHE_SHIFT;
2528
2529 struct extent_page_data epd = {
2530 .bio = NULL,
2531 .tree = tree,
2532 .get_extent = get_extent,
2533 .extent_locked = 1,
2534 .sync_io = mode == WB_SYNC_ALL,
2535 };
2536 struct writeback_control wbc_writepages = {
2537 .sync_mode = mode,
2538 .nr_to_write = nr_pages * 2,
2539 .range_start = start,
2540 .range_end = end + 1,
2541 };
2542
2543 while (start <= end) {
2544 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2545 if (clear_page_dirty_for_io(page))
2546 ret = __extent_writepage(page, &wbc_writepages, &epd);
2547 else {
2548 if (tree->ops && tree->ops->writepage_end_io_hook)
2549 tree->ops->writepage_end_io_hook(page, start,
2550 start + PAGE_CACHE_SIZE - 1,
2551 NULL, 1);
2552 unlock_page(page);
2553 }
2554 page_cache_release(page);
2555 start += PAGE_CACHE_SIZE;
2556 }
2557
2558 flush_epd_write_bio(&epd);
2559 return ret;
2560}
2561
2562int extent_writepages(struct extent_io_tree *tree,
2563 struct address_space *mapping,
2564 get_extent_t *get_extent,
2565 struct writeback_control *wbc)
2566{
2567 int ret = 0;
2568 struct extent_page_data epd = {
2569 .bio = NULL,
2570 .tree = tree,
2571 .get_extent = get_extent,
2572 .extent_locked = 0,
2573 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2574 };
2575
2576 ret = extent_write_cache_pages(tree, mapping, wbc,
2577 __extent_writepage, &epd,
2578 flush_write_bio);
2579 flush_epd_write_bio(&epd);
2580 return ret;
2581}
2582
2583int extent_readpages(struct extent_io_tree *tree,
2584 struct address_space *mapping,
2585 struct list_head *pages, unsigned nr_pages,
2586 get_extent_t get_extent)
2587{
2588 struct bio *bio = NULL;
2589 unsigned page_idx;
2590 unsigned long bio_flags = 0;
2591
2592 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2593 struct page *page = list_entry(pages->prev, struct page, lru);
2594
2595 prefetchw(&page->flags);
2596 list_del(&page->lru);
2597 if (!add_to_page_cache_lru(page, mapping,
2598 page->index, GFP_NOFS)) {
2599 __extent_read_full_page(tree, page, get_extent,
2600 &bio, 0, &bio_flags);
2601 }
2602 page_cache_release(page);
2603 }
2604 BUG_ON(!list_empty(pages));
2605 if (bio)
2606 submit_one_bio(READ, bio, 0, bio_flags);
2607 return 0;
2608}
2609
2610/*
2611 * basic invalidatepage code, this waits on any locked or writeback
2612 * ranges corresponding to the page, and then deletes any extent state
2613 * records from the tree
2614 */
2615int extent_invalidatepage(struct extent_io_tree *tree,
2616 struct page *page, unsigned long offset)
2617{
2618 struct extent_state *cached_state = NULL;
2619 u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2620 u64 end = start + PAGE_CACHE_SIZE - 1;
2621 size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2622
2623 start += (offset + blocksize - 1) & ~(blocksize - 1);
2624 if (start > end)
2625 return 0;
2626
2627 lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
2628 wait_on_page_writeback(page);
2629 clear_extent_bit(tree, start, end,
2630 EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2631 EXTENT_DO_ACCOUNTING,
2632 1, 1, &cached_state, GFP_NOFS);
2633 return 0;
2634}
2635
2636/*
2637 * a helper for releasepage, this tests for areas of the page that
2638 * are locked or under IO and drops the related state bits if it is safe
2639 * to drop the page.
2640 */
2641int try_release_extent_state(struct extent_map_tree *map,
2642 struct extent_io_tree *tree, struct page *page,
2643 gfp_t mask)
2644{
2645 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2646 u64 end = start + PAGE_CACHE_SIZE - 1;
2647 int ret = 1;
2648
2649 if (test_range_bit(tree, start, end,
2650 EXTENT_IOBITS, 0, NULL))
2651 ret = 0;
2652 else {
2653 if ((mask & GFP_NOFS) == GFP_NOFS)
2654 mask = GFP_NOFS;
2655 /*
2656 * at this point we can safely clear everything except the
2657 * locked bit and the nodatasum bit
2658 */
2659 ret = clear_extent_bit(tree, start, end,
2660 ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2661 0, 0, NULL, mask);
2662
2663 /* if clear_extent_bit failed for enomem reasons,
2664 * we can't allow the release to continue.
2665 */
2666 if (ret < 0)
2667 ret = 0;
2668 else
2669 ret = 1;
2670 }
2671 return ret;
2672}
2673
2674/*
2675 * a helper for releasepage. As long as there are no locked extents
2676 * in the range corresponding to the page, both state records and extent
2677 * map records are removed
2678 */
2679int try_release_extent_mapping(struct extent_map_tree *map,
2680 struct extent_io_tree *tree, struct page *page,
2681 gfp_t mask)
2682{
2683 struct extent_map *em;
2684 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2685 u64 end = start + PAGE_CACHE_SIZE - 1;
2686
2687 if ((mask & __GFP_WAIT) &&
2688 page->mapping->host->i_size > 16 * 1024 * 1024) {
2689 u64 len;
2690 while (start <= end) {
2691 len = end - start + 1;
2692 write_lock(&map->lock);
2693 em = lookup_extent_mapping(map, start, len);
2694 if (IS_ERR_OR_NULL(em)) {
2695 write_unlock(&map->lock);
2696 break;
2697 }
2698 if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2699 em->start != start) {
2700 write_unlock(&map->lock);
2701 free_extent_map(em);
2702 break;
2703 }
2704 if (!test_range_bit(tree, em->start,
2705 extent_map_end(em) - 1,
2706 EXTENT_LOCKED | EXTENT_WRITEBACK,
2707 0, NULL)) {
2708 remove_extent_mapping(map, em);
2709 /* once for the rb tree */
2710 free_extent_map(em);
2711 }
2712 start = extent_map_end(em);
2713 write_unlock(&map->lock);
2714
2715 /* once for us */
2716 free_extent_map(em);
2717 }
2718 }
2719 return try_release_extent_state(map, tree, page, mask);
2720}
2721
2722/*
2723 * helper function for fiemap, which doesn't want to see any holes.
2724 * This maps until we find something past 'last'
2725 */
2726static struct extent_map *get_extent_skip_holes(struct inode *inode,
2727 u64 offset,
2728 u64 last,
2729 get_extent_t *get_extent)
2730{
2731 u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
2732 struct extent_map *em;
2733 u64 len;
2734
2735 if (offset >= last)
2736 return NULL;
2737
2738 while(1) {
2739 len = last - offset;
2740 if (len == 0)
2741 break;
2742 len = (len + sectorsize - 1) & ~(sectorsize - 1);
2743 em = get_extent(inode, NULL, 0, offset, len, 0);
2744 if (IS_ERR_OR_NULL(em))
2745 return em;
2746
2747 /* if this isn't a hole return it */
2748 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
2749 em->block_start != EXTENT_MAP_HOLE) {
2750 return em;
2751 }
2752
2753 /* this is a hole, advance to the next extent */
2754 offset = extent_map_end(em);
2755 free_extent_map(em);
2756 if (offset >= last)
2757 break;
2758 }
2759 return NULL;
2760}
2761
2762int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2763 __u64 start, __u64 len, get_extent_t *get_extent)
2764{
2765 int ret = 0;
2766 u64 off = start;
2767 u64 max = start + len;
2768 u32 flags = 0;
2769 u32 found_type;
2770 u64 last;
2771 u64 last_for_get_extent = 0;
2772 u64 disko = 0;
2773 u64 isize = i_size_read(inode);
2774 struct btrfs_key found_key;
2775 struct extent_map *em = NULL;
2776 struct extent_state *cached_state = NULL;
2777 struct btrfs_path *path;
2778 struct btrfs_file_extent_item *item;
2779 int end = 0;
2780 u64 em_start = 0;
2781 u64 em_len = 0;
2782 u64 em_end = 0;
2783 unsigned long emflags;
2784
2785 if (len == 0)
2786 return -EINVAL;
2787
2788 path = btrfs_alloc_path();
2789 if (!path)
2790 return -ENOMEM;
2791 path->leave_spinning = 1;
2792
2793 /*
2794 * lookup the last file extent. We're not using i_size here
2795 * because there might be preallocation past i_size
2796 */
2797 ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
2798 path, btrfs_ino(inode), -1, 0);
2799 if (ret < 0) {
2800 btrfs_free_path(path);
2801 return ret;
2802 }
2803 WARN_ON(!ret);
2804 path->slots[0]--;
2805 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2806 struct btrfs_file_extent_item);
2807 btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
2808 found_type = btrfs_key_type(&found_key);
2809
2810 /* No extents, but there might be delalloc bits */
2811 if (found_key.objectid != btrfs_ino(inode) ||
2812 found_type != BTRFS_EXTENT_DATA_KEY) {
2813 /* have to trust i_size as the end */
2814 last = (u64)-1;
2815 last_for_get_extent = isize;
2816 } else {
2817 /*
2818 * remember the start of the last extent. There are a
2819 * bunch of different factors that go into the length of the
2820 * extent, so its much less complex to remember where it started
2821 */
2822 last = found_key.offset;
2823 last_for_get_extent = last + 1;
2824 }
2825 btrfs_free_path(path);
2826
2827 /*
2828 * we might have some extents allocated but more delalloc past those
2829 * extents. so, we trust isize unless the start of the last extent is
2830 * beyond isize
2831 */
2832 if (last < isize) {
2833 last = (u64)-1;
2834 last_for_get_extent = isize;
2835 }
2836
2837 lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
2838 &cached_state, GFP_NOFS);
2839
2840 em = get_extent_skip_holes(inode, off, last_for_get_extent,
2841 get_extent);
2842 if (!em)
2843 goto out;
2844 if (IS_ERR(em)) {
2845 ret = PTR_ERR(em);
2846 goto out;
2847 }
2848
2849 while (!end) {
2850 u64 offset_in_extent;
2851
2852 /* break if the extent we found is outside the range */
2853 if (em->start >= max || extent_map_end(em) < off)
2854 break;
2855
2856 /*
2857 * get_extent may return an extent that starts before our
2858 * requested range. We have to make sure the ranges
2859 * we return to fiemap always move forward and don't
2860 * overlap, so adjust the offsets here
2861 */
2862 em_start = max(em->start, off);
2863
2864 /*
2865 * record the offset from the start of the extent
2866 * for adjusting the disk offset below
2867 */
2868 offset_in_extent = em_start - em->start;
2869 em_end = extent_map_end(em);
2870 em_len = em_end - em_start;
2871 emflags = em->flags;
2872 disko = 0;
2873 flags = 0;
2874
2875 /*
2876 * bump off for our next call to get_extent
2877 */
2878 off = extent_map_end(em);
2879 if (off >= max)
2880 end = 1;
2881
2882 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
2883 end = 1;
2884 flags |= FIEMAP_EXTENT_LAST;
2885 } else if (em->block_start == EXTENT_MAP_INLINE) {
2886 flags |= (FIEMAP_EXTENT_DATA_INLINE |
2887 FIEMAP_EXTENT_NOT_ALIGNED);
2888 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
2889 flags |= (FIEMAP_EXTENT_DELALLOC |
2890 FIEMAP_EXTENT_UNKNOWN);
2891 } else {
2892 disko = em->block_start + offset_in_extent;
2893 }
2894 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2895 flags |= FIEMAP_EXTENT_ENCODED;
2896
2897 free_extent_map(em);
2898 em = NULL;
2899 if ((em_start >= last) || em_len == (u64)-1 ||
2900 (last == (u64)-1 && isize <= em_end)) {
2901 flags |= FIEMAP_EXTENT_LAST;
2902 end = 1;
2903 }
2904
2905 /* now scan forward to see if this is really the last extent. */
2906 em = get_extent_skip_holes(inode, off, last_for_get_extent,
2907 get_extent);
2908 if (IS_ERR(em)) {
2909 ret = PTR_ERR(em);
2910 goto out;
2911 }
2912 if (!em) {
2913 flags |= FIEMAP_EXTENT_LAST;
2914 end = 1;
2915 }
2916 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
2917 em_len, flags);
2918 if (ret)
2919 goto out_free;
2920 }
2921out_free:
2922 free_extent_map(em);
2923out:
2924 unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
2925 &cached_state, GFP_NOFS);
2926 return ret;
2927}
2928
2929static inline struct page *extent_buffer_page(struct extent_buffer *eb,
2930 unsigned long i)
2931{
2932 struct page *p;
2933 struct address_space *mapping;
2934
2935 if (i == 0)
2936 return eb->first_page;
2937 i += eb->start >> PAGE_CACHE_SHIFT;
2938 mapping = eb->first_page->mapping;
2939 if (!mapping)
2940 return NULL;
2941
2942 /*
2943 * extent_buffer_page is only called after pinning the page
2944 * by increasing the reference count. So we know the page must
2945 * be in the radix tree.
2946 */
2947 rcu_read_lock();
2948 p = radix_tree_lookup(&mapping->page_tree, i);
2949 rcu_read_unlock();
2950
2951 return p;
2952}
2953
2954static inline unsigned long num_extent_pages(u64 start, u64 len)
2955{
2956 return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
2957 (start >> PAGE_CACHE_SHIFT);
2958}
2959
2960static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
2961 u64 start,
2962 unsigned long len,
2963 gfp_t mask)
2964{
2965 struct extent_buffer *eb = NULL;
2966#if LEAK_DEBUG
2967 unsigned long flags;
2968#endif
2969
2970 eb = kmem_cache_zalloc(extent_buffer_cache, mask);
2971 if (eb == NULL)
2972 return NULL;
2973 eb->start = start;
2974 eb->len = len;
2975 rwlock_init(&eb->lock);
2976 atomic_set(&eb->write_locks, 0);
2977 atomic_set(&eb->read_locks, 0);
2978 atomic_set(&eb->blocking_readers, 0);
2979 atomic_set(&eb->blocking_writers, 0);
2980 atomic_set(&eb->spinning_readers, 0);
2981 atomic_set(&eb->spinning_writers, 0);
2982 init_waitqueue_head(&eb->write_lock_wq);
2983 init_waitqueue_head(&eb->read_lock_wq);
2984
2985#if LEAK_DEBUG
2986 spin_lock_irqsave(&leak_lock, flags);
2987 list_add(&eb->leak_list, &buffers);
2988 spin_unlock_irqrestore(&leak_lock, flags);
2989#endif
2990 atomic_set(&eb->refs, 1);
2991
2992 return eb;
2993}
2994
2995static void __free_extent_buffer(struct extent_buffer *eb)
2996{
2997#if LEAK_DEBUG
2998 unsigned long flags;
2999 spin_lock_irqsave(&leak_lock, flags);
3000 list_del(&eb->leak_list);
3001 spin_unlock_irqrestore(&leak_lock, flags);
3002#endif
3003 kmem_cache_free(extent_buffer_cache, eb);
3004}
3005
3006/*
3007 * Helper for releasing extent buffer page.
3008 */
3009static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3010 unsigned long start_idx)
3011{
3012 unsigned long index;
3013 struct page *page;
3014
3015 if (!eb->first_page)
3016 return;
3017
3018 index = num_extent_pages(eb->start, eb->len);
3019 if (start_idx >= index)
3020 return;
3021
3022 do {
3023 index--;
3024 page = extent_buffer_page(eb, index);
3025 if (page)
3026 page_cache_release(page);
3027 } while (index != start_idx);
3028}
3029
3030/*
3031 * Helper for releasing the extent buffer.
3032 */
3033static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3034{
3035 btrfs_release_extent_buffer_page(eb, 0);
3036 __free_extent_buffer(eb);
3037}
3038
3039struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3040 u64 start, unsigned long len,
3041 struct page *page0)
3042{
3043 unsigned long num_pages = num_extent_pages(start, len);
3044 unsigned long i;
3045 unsigned long index = start >> PAGE_CACHE_SHIFT;
3046 struct extent_buffer *eb;
3047 struct extent_buffer *exists = NULL;
3048 struct page *p;
3049 struct address_space *mapping = tree->mapping;
3050 int uptodate = 1;
3051 int ret;
3052
3053 rcu_read_lock();
3054 eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3055 if (eb && atomic_inc_not_zero(&eb->refs)) {
3056 rcu_read_unlock();
3057 mark_page_accessed(eb->first_page);
3058 return eb;
3059 }
3060 rcu_read_unlock();
3061
3062 eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
3063 if (!eb)
3064 return NULL;
3065
3066 if (page0) {
3067 eb->first_page = page0;
3068 i = 1;
3069 index++;
3070 page_cache_get(page0);
3071 mark_page_accessed(page0);
3072 set_page_extent_mapped(page0);
3073 set_page_extent_head(page0, len);
3074 uptodate = PageUptodate(page0);
3075 } else {
3076 i = 0;
3077 }
3078 for (; i < num_pages; i++, index++) {
3079 p = find_or_create_page(mapping, index, GFP_NOFS);
3080 if (!p) {
3081 WARN_ON(1);
3082 goto free_eb;
3083 }
3084 set_page_extent_mapped(p);
3085 mark_page_accessed(p);
3086 if (i == 0) {
3087 eb->first_page = p;
3088 set_page_extent_head(p, len);
3089 } else {
3090 set_page_private(p, EXTENT_PAGE_PRIVATE);
3091 }
3092 if (!PageUptodate(p))
3093 uptodate = 0;
3094
3095 /*
3096 * see below about how we avoid a nasty race with release page
3097 * and why we unlock later
3098 */
3099 if (i != 0)
3100 unlock_page(p);
3101 }
3102 if (uptodate)
3103 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3104
3105 ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
3106 if (ret)
3107 goto free_eb;
3108
3109 spin_lock(&tree->buffer_lock);
3110 ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
3111 if (ret == -EEXIST) {
3112 exists = radix_tree_lookup(&tree->buffer,
3113 start >> PAGE_CACHE_SHIFT);
3114 /* add one reference for the caller */
3115 atomic_inc(&exists->refs);
3116 spin_unlock(&tree->buffer_lock);
3117 radix_tree_preload_end();
3118 goto free_eb;
3119 }
3120 /* add one reference for the tree */
3121 atomic_inc(&eb->refs);
3122 spin_unlock(&tree->buffer_lock);
3123 radix_tree_preload_end();
3124
3125 /*
3126 * there is a race where release page may have
3127 * tried to find this extent buffer in the radix
3128 * but failed. It will tell the VM it is safe to
3129 * reclaim the, and it will clear the page private bit.
3130 * We must make sure to set the page private bit properly
3131 * after the extent buffer is in the radix tree so
3132 * it doesn't get lost
3133 */
3134 set_page_extent_mapped(eb->first_page);
3135 set_page_extent_head(eb->first_page, eb->len);
3136 if (!page0)
3137 unlock_page(eb->first_page);
3138 return eb;
3139
3140free_eb:
3141 if (eb->first_page && !page0)
3142 unlock_page(eb->first_page);
3143
3144 if (!atomic_dec_and_test(&eb->refs))
3145 return exists;
3146 btrfs_release_extent_buffer(eb);
3147 return exists;
3148}
3149
3150struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3151 u64 start, unsigned long len)
3152{
3153 struct extent_buffer *eb;
3154
3155 rcu_read_lock();
3156 eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3157 if (eb && atomic_inc_not_zero(&eb->refs)) {
3158 rcu_read_unlock();
3159 mark_page_accessed(eb->first_page);
3160 return eb;
3161 }
3162 rcu_read_unlock();
3163
3164 return NULL;
3165}
3166
3167void free_extent_buffer(struct extent_buffer *eb)
3168{
3169 if (!eb)
3170 return;
3171
3172 if (!atomic_dec_and_test(&eb->refs))
3173 return;
3174
3175 WARN_ON(1);
3176}
3177
3178int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3179 struct extent_buffer *eb)
3180{
3181 unsigned long i;
3182 unsigned long num_pages;
3183 struct page *page;
3184
3185 num_pages = num_extent_pages(eb->start, eb->len);
3186
3187 for (i = 0; i < num_pages; i++) {
3188 page = extent_buffer_page(eb, i);
3189 if (!PageDirty(page))
3190 continue;
3191
3192 lock_page(page);
3193 WARN_ON(!PagePrivate(page));
3194
3195 set_page_extent_mapped(page);
3196 if (i == 0)
3197 set_page_extent_head(page, eb->len);
3198
3199 clear_page_dirty_for_io(page);
3200 spin_lock_irq(&page->mapping->tree_lock);
3201 if (!PageDirty(page)) {
3202 radix_tree_tag_clear(&page->mapping->page_tree,
3203 page_index(page),
3204 PAGECACHE_TAG_DIRTY);
3205 }
3206 spin_unlock_irq(&page->mapping->tree_lock);
3207 unlock_page(page);
3208 }
3209 return 0;
3210}
3211
3212int set_extent_buffer_dirty(struct extent_io_tree *tree,
3213 struct extent_buffer *eb)
3214{
3215 unsigned long i;
3216 unsigned long num_pages;
3217 int was_dirty = 0;
3218
3219 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3220 num_pages = num_extent_pages(eb->start, eb->len);
3221 for (i = 0; i < num_pages; i++)
3222 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3223 return was_dirty;
3224}
3225
3226static int __eb_straddles_pages(u64 start, u64 len)
3227{
3228 if (len < PAGE_CACHE_SIZE)
3229 return 1;
3230 if (start & (PAGE_CACHE_SIZE - 1))
3231 return 1;
3232 if ((start + len) & (PAGE_CACHE_SIZE - 1))
3233 return 1;
3234 return 0;
3235}
3236
3237static int eb_straddles_pages(struct extent_buffer *eb)
3238{
3239 return __eb_straddles_pages(eb->start, eb->len);
3240}
3241
3242int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3243 struct extent_buffer *eb,
3244 struct extent_state **cached_state)
3245{
3246 unsigned long i;
3247 struct page *page;
3248 unsigned long num_pages;
3249
3250 num_pages = num_extent_pages(eb->start, eb->len);
3251 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3252
3253 if (eb_straddles_pages(eb)) {
3254 clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3255 cached_state, GFP_NOFS);
3256 }
3257 for (i = 0; i < num_pages; i++) {
3258 page = extent_buffer_page(eb, i);
3259 if (page)
3260 ClearPageUptodate(page);
3261 }
3262 return 0;
3263}
3264
3265int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3266 struct extent_buffer *eb)
3267{
3268 unsigned long i;
3269 struct page *page;
3270 unsigned long num_pages;
3271
3272 num_pages = num_extent_pages(eb->start, eb->len);
3273
3274 if (eb_straddles_pages(eb)) {
3275 set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3276 NULL, GFP_NOFS);
3277 }
3278 for (i = 0; i < num_pages; i++) {
3279 page = extent_buffer_page(eb, i);
3280 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3281 ((i == num_pages - 1) &&
3282 ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3283 check_page_uptodate(tree, page);
3284 continue;
3285 }
3286 SetPageUptodate(page);
3287 }
3288 return 0;
3289}
3290
3291int extent_range_uptodate(struct extent_io_tree *tree,
3292 u64 start, u64 end)
3293{
3294 struct page *page;
3295 int ret;
3296 int pg_uptodate = 1;
3297 int uptodate;
3298 unsigned long index;
3299
3300 if (__eb_straddles_pages(start, end - start + 1)) {
3301 ret = test_range_bit(tree, start, end,
3302 EXTENT_UPTODATE, 1, NULL);
3303 if (ret)
3304 return 1;
3305 }
3306 while (start <= end) {
3307 index = start >> PAGE_CACHE_SHIFT;
3308 page = find_get_page(tree->mapping, index);
3309 uptodate = PageUptodate(page);
3310 page_cache_release(page);
3311 if (!uptodate) {
3312 pg_uptodate = 0;
3313 break;
3314 }
3315 start += PAGE_CACHE_SIZE;
3316 }
3317 return pg_uptodate;
3318}
3319
3320int extent_buffer_uptodate(struct extent_io_tree *tree,
3321 struct extent_buffer *eb,
3322 struct extent_state *cached_state)
3323{
3324 int ret = 0;
3325 unsigned long num_pages;
3326 unsigned long i;
3327 struct page *page;
3328 int pg_uptodate = 1;
3329
3330 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3331 return 1;
3332
3333 if (eb_straddles_pages(eb)) {
3334 ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3335 EXTENT_UPTODATE, 1, cached_state);
3336 if (ret)
3337 return ret;
3338 }
3339
3340 num_pages = num_extent_pages(eb->start, eb->len);
3341 for (i = 0; i < num_pages; i++) {
3342 page = extent_buffer_page(eb, i);
3343 if (!PageUptodate(page)) {
3344 pg_uptodate = 0;
3345 break;
3346 }
3347 }
3348 return pg_uptodate;
3349}
3350
3351int read_extent_buffer_pages(struct extent_io_tree *tree,
3352 struct extent_buffer *eb,
3353 u64 start, int wait,
3354 get_extent_t *get_extent, int mirror_num)
3355{
3356 unsigned long i;
3357 unsigned long start_i;
3358 struct page *page;
3359 int err;
3360 int ret = 0;
3361 int locked_pages = 0;
3362 int all_uptodate = 1;
3363 int inc_all_pages = 0;
3364 unsigned long num_pages;
3365 struct bio *bio = NULL;
3366 unsigned long bio_flags = 0;
3367
3368 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3369 return 0;
3370
3371 if (eb_straddles_pages(eb)) {
3372 if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3373 EXTENT_UPTODATE, 1, NULL)) {
3374 return 0;
3375 }
3376 }
3377
3378 if (start) {
3379 WARN_ON(start < eb->start);
3380 start_i = (start >> PAGE_CACHE_SHIFT) -
3381 (eb->start >> PAGE_CACHE_SHIFT);
3382 } else {
3383 start_i = 0;
3384 }
3385
3386 num_pages = num_extent_pages(eb->start, eb->len);
3387 for (i = start_i; i < num_pages; i++) {
3388 page = extent_buffer_page(eb, i);
3389 if (!wait) {
3390 if (!trylock_page(page))
3391 goto unlock_exit;
3392 } else {
3393 lock_page(page);
3394 }
3395 locked_pages++;
3396 if (!PageUptodate(page))
3397 all_uptodate = 0;
3398 }
3399 if (all_uptodate) {
3400 if (start_i == 0)
3401 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3402 goto unlock_exit;
3403 }
3404
3405 for (i = start_i; i < num_pages; i++) {
3406 page = extent_buffer_page(eb, i);
3407
3408 WARN_ON(!PagePrivate(page));
3409
3410 set_page_extent_mapped(page);
3411 if (i == 0)
3412 set_page_extent_head(page, eb->len);
3413
3414 if (inc_all_pages)
3415 page_cache_get(page);
3416 if (!PageUptodate(page)) {
3417 if (start_i == 0)
3418 inc_all_pages = 1;
3419 ClearPageError(page);
3420 err = __extent_read_full_page(tree, page,
3421 get_extent, &bio,
3422 mirror_num, &bio_flags);
3423 if (err)
3424 ret = err;
3425 } else {
3426 unlock_page(page);
3427 }
3428 }
3429
3430 if (bio)
3431 submit_one_bio(READ, bio, mirror_num, bio_flags);
3432
3433 if (ret || !wait)
3434 return ret;
3435
3436 for (i = start_i; i < num_pages; i++) {
3437 page = extent_buffer_page(eb, i);
3438 wait_on_page_locked(page);
3439 if (!PageUptodate(page))
3440 ret = -EIO;
3441 }
3442
3443 if (!ret)
3444 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3445 return ret;
3446
3447unlock_exit:
3448 i = start_i;
3449 while (locked_pages > 0) {
3450 page = extent_buffer_page(eb, i);
3451 i++;
3452 unlock_page(page);
3453 locked_pages--;
3454 }
3455 return ret;
3456}
3457
3458void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3459 unsigned long start,
3460 unsigned long len)
3461{
3462 size_t cur;
3463 size_t offset;
3464 struct page *page;
3465 char *kaddr;
3466 char *dst = (char *)dstv;
3467 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3468 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3469
3470 WARN_ON(start > eb->len);
3471 WARN_ON(start + len > eb->start + eb->len);
3472
3473 offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3474
3475 while (len > 0) {
3476 page = extent_buffer_page(eb, i);
3477
3478 cur = min(len, (PAGE_CACHE_SIZE - offset));
3479 kaddr = page_address(page);
3480 memcpy(dst, kaddr + offset, cur);
3481
3482 dst += cur;
3483 len -= cur;
3484 offset = 0;
3485 i++;
3486 }
3487}
3488
3489int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3490 unsigned long min_len, char **map,
3491 unsigned long *map_start,
3492 unsigned long *map_len)
3493{
3494 size_t offset = start & (PAGE_CACHE_SIZE - 1);
3495 char *kaddr;
3496 struct page *p;
3497 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3498 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3499 unsigned long end_i = (start_offset + start + min_len - 1) >>
3500 PAGE_CACHE_SHIFT;
3501
3502 if (i != end_i)
3503 return -EINVAL;
3504
3505 if (i == 0) {
3506 offset = start_offset;
3507 *map_start = 0;
3508 } else {
3509 offset = 0;
3510 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3511 }
3512
3513 if (start + min_len > eb->len) {
3514 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
3515 "wanted %lu %lu\n", (unsigned long long)eb->start,
3516 eb->len, start, min_len);
3517 WARN_ON(1);
3518 return -EINVAL;
3519 }
3520
3521 p = extent_buffer_page(eb, i);
3522 kaddr = page_address(p);
3523 *map = kaddr + offset;
3524 *map_len = PAGE_CACHE_SIZE - offset;
3525 return 0;
3526}
3527
3528int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3529 unsigned long start,
3530 unsigned long len)
3531{
3532 size_t cur;
3533 size_t offset;
3534 struct page *page;
3535 char *kaddr;
3536 char *ptr = (char *)ptrv;
3537 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3538 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3539 int ret = 0;
3540
3541 WARN_ON(start > eb->len);
3542 WARN_ON(start + len > eb->start + eb->len);
3543
3544 offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3545
3546 while (len > 0) {
3547 page = extent_buffer_page(eb, i);
3548
3549 cur = min(len, (PAGE_CACHE_SIZE - offset));
3550
3551 kaddr = page_address(page);
3552 ret = memcmp(ptr, kaddr + offset, cur);
3553 if (ret)
3554 break;
3555
3556 ptr += cur;
3557 len -= cur;
3558 offset = 0;
3559 i++;
3560 }
3561 return ret;
3562}
3563
3564void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3565 unsigned long start, unsigned long len)
3566{
3567 size_t cur;
3568 size_t offset;
3569 struct page *page;
3570 char *kaddr;
3571 char *src = (char *)srcv;
3572 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3573 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3574
3575 WARN_ON(start > eb->len);
3576 WARN_ON(start + len > eb->start + eb->len);
3577
3578 offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3579
3580 while (len > 0) {
3581 page = extent_buffer_page(eb, i);
3582 WARN_ON(!PageUptodate(page));
3583
3584 cur = min(len, PAGE_CACHE_SIZE - offset);
3585 kaddr = page_address(page);
3586 memcpy(kaddr + offset, src, cur);
3587
3588 src += cur;
3589 len -= cur;
3590 offset = 0;
3591 i++;
3592 }
3593}
3594
3595void memset_extent_buffer(struct extent_buffer *eb, char c,
3596 unsigned long start, unsigned long len)
3597{
3598 size_t cur;
3599 size_t offset;
3600 struct page *page;
3601 char *kaddr;
3602 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3603 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3604
3605 WARN_ON(start > eb->len);
3606 WARN_ON(start + len > eb->start + eb->len);
3607
3608 offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3609
3610 while (len > 0) {
3611 page = extent_buffer_page(eb, i);
3612 WARN_ON(!PageUptodate(page));
3613
3614 cur = min(len, PAGE_CACHE_SIZE - offset);
3615 kaddr = page_address(page);
3616 memset(kaddr + offset, c, cur);
3617
3618 len -= cur;
3619 offset = 0;
3620 i++;
3621 }
3622}
3623
3624void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3625 unsigned long dst_offset, unsigned long src_offset,
3626 unsigned long len)
3627{
3628 u64 dst_len = dst->len;
3629 size_t cur;
3630 size_t offset;
3631 struct page *page;
3632 char *kaddr;
3633 size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3634 unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3635
3636 WARN_ON(src->len != dst_len);
3637
3638 offset = (start_offset + dst_offset) &
3639 ((unsigned long)PAGE_CACHE_SIZE - 1);
3640
3641 while (len > 0) {
3642 page = extent_buffer_page(dst, i);
3643 WARN_ON(!PageUptodate(page));
3644
3645 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3646
3647 kaddr = page_address(page);
3648 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3649
3650 src_offset += cur;
3651 len -= cur;
3652 offset = 0;
3653 i++;
3654 }
3655}
3656
3657static void move_pages(struct page *dst_page, struct page *src_page,
3658 unsigned long dst_off, unsigned long src_off,
3659 unsigned long len)
3660{
3661 char *dst_kaddr = page_address(dst_page);
3662 if (dst_page == src_page) {
3663 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3664 } else {
3665 char *src_kaddr = page_address(src_page);
3666 char *p = dst_kaddr + dst_off + len;
3667 char *s = src_kaddr + src_off + len;
3668
3669 while (len--)
3670 *--p = *--s;
3671 }
3672}
3673
3674static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
3675{
3676 unsigned long distance = (src > dst) ? src - dst : dst - src;
3677 return distance < len;
3678}
3679
3680static void copy_pages(struct page *dst_page, struct page *src_page,
3681 unsigned long dst_off, unsigned long src_off,
3682 unsigned long len)
3683{
3684 char *dst_kaddr = page_address(dst_page);
3685 char *src_kaddr;
3686
3687 if (dst_page != src_page) {
3688 src_kaddr = page_address(src_page);
3689 } else {
3690 src_kaddr = dst_kaddr;
3691 BUG_ON(areas_overlap(src_off, dst_off, len));
3692 }
3693
3694 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3695}
3696
3697void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3698 unsigned long src_offset, unsigned long len)
3699{
3700 size_t cur;
3701 size_t dst_off_in_page;
3702 size_t src_off_in_page;
3703 size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3704 unsigned long dst_i;
3705 unsigned long src_i;
3706
3707 if (src_offset + len > dst->len) {
3708 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3709 "len %lu dst len %lu\n", src_offset, len, dst->len);
3710 BUG_ON(1);
3711 }
3712 if (dst_offset + len > dst->len) {
3713 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3714 "len %lu dst len %lu\n", dst_offset, len, dst->len);
3715 BUG_ON(1);
3716 }
3717
3718 while (len > 0) {
3719 dst_off_in_page = (start_offset + dst_offset) &
3720 ((unsigned long)PAGE_CACHE_SIZE - 1);
3721 src_off_in_page = (start_offset + src_offset) &
3722 ((unsigned long)PAGE_CACHE_SIZE - 1);
3723
3724 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3725 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3726
3727 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3728 src_off_in_page));
3729 cur = min_t(unsigned long, cur,
3730 (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3731
3732 copy_pages(extent_buffer_page(dst, dst_i),
3733 extent_buffer_page(dst, src_i),
3734 dst_off_in_page, src_off_in_page, cur);
3735
3736 src_offset += cur;
3737 dst_offset += cur;
3738 len -= cur;
3739 }
3740}
3741
3742void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3743 unsigned long src_offset, unsigned long len)
3744{
3745 size_t cur;
3746 size_t dst_off_in_page;
3747 size_t src_off_in_page;
3748 unsigned long dst_end = dst_offset + len - 1;
3749 unsigned long src_end = src_offset + len - 1;
3750 size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3751 unsigned long dst_i;
3752 unsigned long src_i;
3753
3754 if (src_offset + len > dst->len) {
3755 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3756 "len %lu len %lu\n", src_offset, len, dst->len);
3757 BUG_ON(1);
3758 }
3759 if (dst_offset + len > dst->len) {
3760 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3761 "len %lu len %lu\n", dst_offset, len, dst->len);
3762 BUG_ON(1);
3763 }
3764 if (!areas_overlap(src_offset, dst_offset, len)) {
3765 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3766 return;
3767 }
3768 while (len > 0) {
3769 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3770 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3771
3772 dst_off_in_page = (start_offset + dst_end) &
3773 ((unsigned long)PAGE_CACHE_SIZE - 1);
3774 src_off_in_page = (start_offset + src_end) &
3775 ((unsigned long)PAGE_CACHE_SIZE - 1);
3776
3777 cur = min_t(unsigned long, len, src_off_in_page + 1);
3778 cur = min(cur, dst_off_in_page + 1);
3779 move_pages(extent_buffer_page(dst, dst_i),
3780 extent_buffer_page(dst, src_i),
3781 dst_off_in_page - cur + 1,
3782 src_off_in_page - cur + 1, cur);
3783
3784 dst_end -= cur;
3785 src_end -= cur;
3786 len -= cur;
3787 }
3788}
3789
3790static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3791{
3792 struct extent_buffer *eb =
3793 container_of(head, struct extent_buffer, rcu_head);
3794
3795 btrfs_release_extent_buffer(eb);
3796}
3797
3798int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
3799{
3800 u64 start = page_offset(page);
3801 struct extent_buffer *eb;
3802 int ret = 1;
3803
3804 spin_lock(&tree->buffer_lock);
3805 eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3806 if (!eb) {
3807 spin_unlock(&tree->buffer_lock);
3808 return ret;
3809 }
3810
3811 if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3812 ret = 0;
3813 goto out;
3814 }
3815
3816 /*
3817 * set @eb->refs to 0 if it is already 1, and then release the @eb.
3818 * Or go back.
3819 */
3820 if (atomic_cmpxchg(&eb->refs, 1, 0) != 1) {
3821 ret = 0;
3822 goto out;
3823 }
3824
3825 radix_tree_delete(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3826out:
3827 spin_unlock(&tree->buffer_lock);
3828
3829 /* at this point we can safely release the extent buffer */
3830 if (atomic_read(&eb->refs) == 0)
3831 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3832 return ret;
3833}
1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/bitops.h>
4#include <linux/slab.h>
5#include <linux/bio.h>
6#include <linux/mm.h>
7#include <linux/pagemap.h>
8#include <linux/page-flags.h>
9#include <linux/spinlock.h>
10#include <linux/blkdev.h>
11#include <linux/swap.h>
12#include <linux/writeback.h>
13#include <linux/pagevec.h>
14#include <linux/prefetch.h>
15#include <linux/cleancache.h>
16#include "misc.h"
17#include "extent_io.h"
18#include "extent-io-tree.h"
19#include "extent_map.h"
20#include "ctree.h"
21#include "btrfs_inode.h"
22#include "volumes.h"
23#include "check-integrity.h"
24#include "locking.h"
25#include "rcu-string.h"
26#include "backref.h"
27#include "disk-io.h"
28#include "subpage.h"
29#include "zoned.h"
30#include "block-group.h"
31
32static struct kmem_cache *extent_state_cache;
33static struct kmem_cache *extent_buffer_cache;
34static struct bio_set btrfs_bioset;
35
36static inline bool extent_state_in_tree(const struct extent_state *state)
37{
38 return !RB_EMPTY_NODE(&state->rb_node);
39}
40
41#ifdef CONFIG_BTRFS_DEBUG
42static LIST_HEAD(states);
43static DEFINE_SPINLOCK(leak_lock);
44
45static inline void btrfs_leak_debug_add(spinlock_t *lock,
46 struct list_head *new,
47 struct list_head *head)
48{
49 unsigned long flags;
50
51 spin_lock_irqsave(lock, flags);
52 list_add(new, head);
53 spin_unlock_irqrestore(lock, flags);
54}
55
56static inline void btrfs_leak_debug_del(spinlock_t *lock,
57 struct list_head *entry)
58{
59 unsigned long flags;
60
61 spin_lock_irqsave(lock, flags);
62 list_del(entry);
63 spin_unlock_irqrestore(lock, flags);
64}
65
66void btrfs_extent_buffer_leak_debug_check(struct btrfs_fs_info *fs_info)
67{
68 struct extent_buffer *eb;
69 unsigned long flags;
70
71 /*
72 * If we didn't get into open_ctree our allocated_ebs will not be
73 * initialized, so just skip this.
74 */
75 if (!fs_info->allocated_ebs.next)
76 return;
77
78 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
79 while (!list_empty(&fs_info->allocated_ebs)) {
80 eb = list_first_entry(&fs_info->allocated_ebs,
81 struct extent_buffer, leak_list);
82 pr_err(
83 "BTRFS: buffer leak start %llu len %lu refs %d bflags %lu owner %llu\n",
84 eb->start, eb->len, atomic_read(&eb->refs), eb->bflags,
85 btrfs_header_owner(eb));
86 list_del(&eb->leak_list);
87 kmem_cache_free(extent_buffer_cache, eb);
88 }
89 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
90}
91
92static inline void btrfs_extent_state_leak_debug_check(void)
93{
94 struct extent_state *state;
95
96 while (!list_empty(&states)) {
97 state = list_entry(states.next, struct extent_state, leak_list);
98 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
99 state->start, state->end, state->state,
100 extent_state_in_tree(state),
101 refcount_read(&state->refs));
102 list_del(&state->leak_list);
103 kmem_cache_free(extent_state_cache, state);
104 }
105}
106
107#define btrfs_debug_check_extent_io_range(tree, start, end) \
108 __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
109static inline void __btrfs_debug_check_extent_io_range(const char *caller,
110 struct extent_io_tree *tree, u64 start, u64 end)
111{
112 struct inode *inode = tree->private_data;
113 u64 isize;
114
115 if (!inode || !is_data_inode(inode))
116 return;
117
118 isize = i_size_read(inode);
119 if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
120 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
121 "%s: ino %llu isize %llu odd range [%llu,%llu]",
122 caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
123 }
124}
125#else
126#define btrfs_leak_debug_add(lock, new, head) do {} while (0)
127#define btrfs_leak_debug_del(lock, entry) do {} while (0)
128#define btrfs_extent_state_leak_debug_check() do {} while (0)
129#define btrfs_debug_check_extent_io_range(c, s, e) do {} while (0)
130#endif
131
132struct tree_entry {
133 u64 start;
134 u64 end;
135 struct rb_node rb_node;
136};
137
138struct extent_page_data {
139 struct btrfs_bio_ctrl bio_ctrl;
140 /* tells writepage not to lock the state bits for this range
141 * it still does the unlocking
142 */
143 unsigned int extent_locked:1;
144
145 /* tells the submit_bio code to use REQ_SYNC */
146 unsigned int sync_io:1;
147};
148
149static int add_extent_changeset(struct extent_state *state, u32 bits,
150 struct extent_changeset *changeset,
151 int set)
152{
153 int ret;
154
155 if (!changeset)
156 return 0;
157 if (set && (state->state & bits) == bits)
158 return 0;
159 if (!set && (state->state & bits) == 0)
160 return 0;
161 changeset->bytes_changed += state->end - state->start + 1;
162 ret = ulist_add(&changeset->range_changed, state->start, state->end,
163 GFP_ATOMIC);
164 return ret;
165}
166
167int __must_check submit_one_bio(struct bio *bio, int mirror_num,
168 unsigned long bio_flags)
169{
170 blk_status_t ret = 0;
171 struct extent_io_tree *tree = bio->bi_private;
172
173 bio->bi_private = NULL;
174
175 if (is_data_inode(tree->private_data))
176 ret = btrfs_submit_data_bio(tree->private_data, bio, mirror_num,
177 bio_flags);
178 else
179 ret = btrfs_submit_metadata_bio(tree->private_data, bio,
180 mirror_num, bio_flags);
181
182 return blk_status_to_errno(ret);
183}
184
185/* Cleanup unsubmitted bios */
186static void end_write_bio(struct extent_page_data *epd, int ret)
187{
188 struct bio *bio = epd->bio_ctrl.bio;
189
190 if (bio) {
191 bio->bi_status = errno_to_blk_status(ret);
192 bio_endio(bio);
193 epd->bio_ctrl.bio = NULL;
194 }
195}
196
197/*
198 * Submit bio from extent page data via submit_one_bio
199 *
200 * Return 0 if everything is OK.
201 * Return <0 for error.
202 */
203static int __must_check flush_write_bio(struct extent_page_data *epd)
204{
205 int ret = 0;
206 struct bio *bio = epd->bio_ctrl.bio;
207
208 if (bio) {
209 ret = submit_one_bio(bio, 0, 0);
210 /*
211 * Clean up of epd->bio is handled by its endio function.
212 * And endio is either triggered by successful bio execution
213 * or the error handler of submit bio hook.
214 * So at this point, no matter what happened, we don't need
215 * to clean up epd->bio.
216 */
217 epd->bio_ctrl.bio = NULL;
218 }
219 return ret;
220}
221
222int __init extent_state_cache_init(void)
223{
224 extent_state_cache = kmem_cache_create("btrfs_extent_state",
225 sizeof(struct extent_state), 0,
226 SLAB_MEM_SPREAD, NULL);
227 if (!extent_state_cache)
228 return -ENOMEM;
229 return 0;
230}
231
232int __init extent_io_init(void)
233{
234 extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
235 sizeof(struct extent_buffer), 0,
236 SLAB_MEM_SPREAD, NULL);
237 if (!extent_buffer_cache)
238 return -ENOMEM;
239
240 if (bioset_init(&btrfs_bioset, BIO_POOL_SIZE,
241 offsetof(struct btrfs_io_bio, bio),
242 BIOSET_NEED_BVECS))
243 goto free_buffer_cache;
244
245 if (bioset_integrity_create(&btrfs_bioset, BIO_POOL_SIZE))
246 goto free_bioset;
247
248 return 0;
249
250free_bioset:
251 bioset_exit(&btrfs_bioset);
252
253free_buffer_cache:
254 kmem_cache_destroy(extent_buffer_cache);
255 extent_buffer_cache = NULL;
256 return -ENOMEM;
257}
258
259void __cold extent_state_cache_exit(void)
260{
261 btrfs_extent_state_leak_debug_check();
262 kmem_cache_destroy(extent_state_cache);
263}
264
265void __cold extent_io_exit(void)
266{
267 /*
268 * Make sure all delayed rcu free are flushed before we
269 * destroy caches.
270 */
271 rcu_barrier();
272 kmem_cache_destroy(extent_buffer_cache);
273 bioset_exit(&btrfs_bioset);
274}
275
276/*
277 * For the file_extent_tree, we want to hold the inode lock when we lookup and
278 * update the disk_i_size, but lockdep will complain because our io_tree we hold
279 * the tree lock and get the inode lock when setting delalloc. These two things
280 * are unrelated, so make a class for the file_extent_tree so we don't get the
281 * two locking patterns mixed up.
282 */
283static struct lock_class_key file_extent_tree_class;
284
285void extent_io_tree_init(struct btrfs_fs_info *fs_info,
286 struct extent_io_tree *tree, unsigned int owner,
287 void *private_data)
288{
289 tree->fs_info = fs_info;
290 tree->state = RB_ROOT;
291 tree->dirty_bytes = 0;
292 spin_lock_init(&tree->lock);
293 tree->private_data = private_data;
294 tree->owner = owner;
295 if (owner == IO_TREE_INODE_FILE_EXTENT)
296 lockdep_set_class(&tree->lock, &file_extent_tree_class);
297}
298
299void extent_io_tree_release(struct extent_io_tree *tree)
300{
301 spin_lock(&tree->lock);
302 /*
303 * Do a single barrier for the waitqueue_active check here, the state
304 * of the waitqueue should not change once extent_io_tree_release is
305 * called.
306 */
307 smp_mb();
308 while (!RB_EMPTY_ROOT(&tree->state)) {
309 struct rb_node *node;
310 struct extent_state *state;
311
312 node = rb_first(&tree->state);
313 state = rb_entry(node, struct extent_state, rb_node);
314 rb_erase(&state->rb_node, &tree->state);
315 RB_CLEAR_NODE(&state->rb_node);
316 /*
317 * btree io trees aren't supposed to have tasks waiting for
318 * changes in the flags of extent states ever.
319 */
320 ASSERT(!waitqueue_active(&state->wq));
321 free_extent_state(state);
322
323 cond_resched_lock(&tree->lock);
324 }
325 spin_unlock(&tree->lock);
326}
327
328static struct extent_state *alloc_extent_state(gfp_t mask)
329{
330 struct extent_state *state;
331
332 /*
333 * The given mask might be not appropriate for the slab allocator,
334 * drop the unsupported bits
335 */
336 mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
337 state = kmem_cache_alloc(extent_state_cache, mask);
338 if (!state)
339 return state;
340 state->state = 0;
341 state->failrec = NULL;
342 RB_CLEAR_NODE(&state->rb_node);
343 btrfs_leak_debug_add(&leak_lock, &state->leak_list, &states);
344 refcount_set(&state->refs, 1);
345 init_waitqueue_head(&state->wq);
346 trace_alloc_extent_state(state, mask, _RET_IP_);
347 return state;
348}
349
350void free_extent_state(struct extent_state *state)
351{
352 if (!state)
353 return;
354 if (refcount_dec_and_test(&state->refs)) {
355 WARN_ON(extent_state_in_tree(state));
356 btrfs_leak_debug_del(&leak_lock, &state->leak_list);
357 trace_free_extent_state(state, _RET_IP_);
358 kmem_cache_free(extent_state_cache, state);
359 }
360}
361
362static struct rb_node *tree_insert(struct rb_root *root,
363 struct rb_node *search_start,
364 u64 offset,
365 struct rb_node *node,
366 struct rb_node ***p_in,
367 struct rb_node **parent_in)
368{
369 struct rb_node **p;
370 struct rb_node *parent = NULL;
371 struct tree_entry *entry;
372
373 if (p_in && parent_in) {
374 p = *p_in;
375 parent = *parent_in;
376 goto do_insert;
377 }
378
379 p = search_start ? &search_start : &root->rb_node;
380 while (*p) {
381 parent = *p;
382 entry = rb_entry(parent, struct tree_entry, rb_node);
383
384 if (offset < entry->start)
385 p = &(*p)->rb_left;
386 else if (offset > entry->end)
387 p = &(*p)->rb_right;
388 else
389 return parent;
390 }
391
392do_insert:
393 rb_link_node(node, parent, p);
394 rb_insert_color(node, root);
395 return NULL;
396}
397
398/**
399 * Search @tree for an entry that contains @offset. Such entry would have
400 * entry->start <= offset && entry->end >= offset.
401 *
402 * @tree: the tree to search
403 * @offset: offset that should fall within an entry in @tree
404 * @next_ret: pointer to the first entry whose range ends after @offset
405 * @prev_ret: pointer to the first entry whose range begins before @offset
406 * @p_ret: pointer where new node should be anchored (used when inserting an
407 * entry in the tree)
408 * @parent_ret: points to entry which would have been the parent of the entry,
409 * containing @offset
410 *
411 * This function returns a pointer to the entry that contains @offset byte
412 * address. If no such entry exists, then NULL is returned and the other
413 * pointer arguments to the function are filled, otherwise the found entry is
414 * returned and other pointers are left untouched.
415 */
416static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
417 struct rb_node **next_ret,
418 struct rb_node **prev_ret,
419 struct rb_node ***p_ret,
420 struct rb_node **parent_ret)
421{
422 struct rb_root *root = &tree->state;
423 struct rb_node **n = &root->rb_node;
424 struct rb_node *prev = NULL;
425 struct rb_node *orig_prev = NULL;
426 struct tree_entry *entry;
427 struct tree_entry *prev_entry = NULL;
428
429 while (*n) {
430 prev = *n;
431 entry = rb_entry(prev, struct tree_entry, rb_node);
432 prev_entry = entry;
433
434 if (offset < entry->start)
435 n = &(*n)->rb_left;
436 else if (offset > entry->end)
437 n = &(*n)->rb_right;
438 else
439 return *n;
440 }
441
442 if (p_ret)
443 *p_ret = n;
444 if (parent_ret)
445 *parent_ret = prev;
446
447 if (next_ret) {
448 orig_prev = prev;
449 while (prev && offset > prev_entry->end) {
450 prev = rb_next(prev);
451 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
452 }
453 *next_ret = prev;
454 prev = orig_prev;
455 }
456
457 if (prev_ret) {
458 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
459 while (prev && offset < prev_entry->start) {
460 prev = rb_prev(prev);
461 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
462 }
463 *prev_ret = prev;
464 }
465 return NULL;
466}
467
468static inline struct rb_node *
469tree_search_for_insert(struct extent_io_tree *tree,
470 u64 offset,
471 struct rb_node ***p_ret,
472 struct rb_node **parent_ret)
473{
474 struct rb_node *next= NULL;
475 struct rb_node *ret;
476
477 ret = __etree_search(tree, offset, &next, NULL, p_ret, parent_ret);
478 if (!ret)
479 return next;
480 return ret;
481}
482
483static inline struct rb_node *tree_search(struct extent_io_tree *tree,
484 u64 offset)
485{
486 return tree_search_for_insert(tree, offset, NULL, NULL);
487}
488
489/*
490 * utility function to look for merge candidates inside a given range.
491 * Any extents with matching state are merged together into a single
492 * extent in the tree. Extents with EXTENT_IO in their state field
493 * are not merged because the end_io handlers need to be able to do
494 * operations on them without sleeping (or doing allocations/splits).
495 *
496 * This should be called with the tree lock held.
497 */
498static void merge_state(struct extent_io_tree *tree,
499 struct extent_state *state)
500{
501 struct extent_state *other;
502 struct rb_node *other_node;
503
504 if (state->state & (EXTENT_LOCKED | EXTENT_BOUNDARY))
505 return;
506
507 other_node = rb_prev(&state->rb_node);
508 if (other_node) {
509 other = rb_entry(other_node, struct extent_state, rb_node);
510 if (other->end == state->start - 1 &&
511 other->state == state->state) {
512 if (tree->private_data &&
513 is_data_inode(tree->private_data))
514 btrfs_merge_delalloc_extent(tree->private_data,
515 state, other);
516 state->start = other->start;
517 rb_erase(&other->rb_node, &tree->state);
518 RB_CLEAR_NODE(&other->rb_node);
519 free_extent_state(other);
520 }
521 }
522 other_node = rb_next(&state->rb_node);
523 if (other_node) {
524 other = rb_entry(other_node, struct extent_state, rb_node);
525 if (other->start == state->end + 1 &&
526 other->state == state->state) {
527 if (tree->private_data &&
528 is_data_inode(tree->private_data))
529 btrfs_merge_delalloc_extent(tree->private_data,
530 state, other);
531 state->end = other->end;
532 rb_erase(&other->rb_node, &tree->state);
533 RB_CLEAR_NODE(&other->rb_node);
534 free_extent_state(other);
535 }
536 }
537}
538
539static void set_state_bits(struct extent_io_tree *tree,
540 struct extent_state *state, u32 *bits,
541 struct extent_changeset *changeset);
542
543/*
544 * insert an extent_state struct into the tree. 'bits' are set on the
545 * struct before it is inserted.
546 *
547 * This may return -EEXIST if the extent is already there, in which case the
548 * state struct is freed.
549 *
550 * The tree lock is not taken internally. This is a utility function and
551 * probably isn't what you want to call (see set/clear_extent_bit).
552 */
553static int insert_state(struct extent_io_tree *tree,
554 struct extent_state *state, u64 start, u64 end,
555 struct rb_node ***p,
556 struct rb_node **parent,
557 u32 *bits, struct extent_changeset *changeset)
558{
559 struct rb_node *node;
560
561 if (end < start) {
562 btrfs_err(tree->fs_info,
563 "insert state: end < start %llu %llu", end, start);
564 WARN_ON(1);
565 }
566 state->start = start;
567 state->end = end;
568
569 set_state_bits(tree, state, bits, changeset);
570
571 node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
572 if (node) {
573 struct extent_state *found;
574 found = rb_entry(node, struct extent_state, rb_node);
575 btrfs_err(tree->fs_info,
576 "found node %llu %llu on insert of %llu %llu",
577 found->start, found->end, start, end);
578 return -EEXIST;
579 }
580 merge_state(tree, state);
581 return 0;
582}
583
584/*
585 * split a given extent state struct in two, inserting the preallocated
586 * struct 'prealloc' as the newly created second half. 'split' indicates an
587 * offset inside 'orig' where it should be split.
588 *
589 * Before calling,
590 * the tree has 'orig' at [orig->start, orig->end]. After calling, there
591 * are two extent state structs in the tree:
592 * prealloc: [orig->start, split - 1]
593 * orig: [ split, orig->end ]
594 *
595 * The tree locks are not taken by this function. They need to be held
596 * by the caller.
597 */
598static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
599 struct extent_state *prealloc, u64 split)
600{
601 struct rb_node *node;
602
603 if (tree->private_data && is_data_inode(tree->private_data))
604 btrfs_split_delalloc_extent(tree->private_data, orig, split);
605
606 prealloc->start = orig->start;
607 prealloc->end = split - 1;
608 prealloc->state = orig->state;
609 orig->start = split;
610
611 node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
612 &prealloc->rb_node, NULL, NULL);
613 if (node) {
614 free_extent_state(prealloc);
615 return -EEXIST;
616 }
617 return 0;
618}
619
620static struct extent_state *next_state(struct extent_state *state)
621{
622 struct rb_node *next = rb_next(&state->rb_node);
623 if (next)
624 return rb_entry(next, struct extent_state, rb_node);
625 else
626 return NULL;
627}
628
629/*
630 * utility function to clear some bits in an extent state struct.
631 * it will optionally wake up anyone waiting on this state (wake == 1).
632 *
633 * If no bits are set on the state struct after clearing things, the
634 * struct is freed and removed from the tree
635 */
636static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
637 struct extent_state *state,
638 u32 *bits, int wake,
639 struct extent_changeset *changeset)
640{
641 struct extent_state *next;
642 u32 bits_to_clear = *bits & ~EXTENT_CTLBITS;
643 int ret;
644
645 if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
646 u64 range = state->end - state->start + 1;
647 WARN_ON(range > tree->dirty_bytes);
648 tree->dirty_bytes -= range;
649 }
650
651 if (tree->private_data && is_data_inode(tree->private_data))
652 btrfs_clear_delalloc_extent(tree->private_data, state, bits);
653
654 ret = add_extent_changeset(state, bits_to_clear, changeset, 0);
655 BUG_ON(ret < 0);
656 state->state &= ~bits_to_clear;
657 if (wake)
658 wake_up(&state->wq);
659 if (state->state == 0) {
660 next = next_state(state);
661 if (extent_state_in_tree(state)) {
662 rb_erase(&state->rb_node, &tree->state);
663 RB_CLEAR_NODE(&state->rb_node);
664 free_extent_state(state);
665 } else {
666 WARN_ON(1);
667 }
668 } else {
669 merge_state(tree, state);
670 next = next_state(state);
671 }
672 return next;
673}
674
675static struct extent_state *
676alloc_extent_state_atomic(struct extent_state *prealloc)
677{
678 if (!prealloc)
679 prealloc = alloc_extent_state(GFP_ATOMIC);
680
681 return prealloc;
682}
683
684static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
685{
686 btrfs_panic(tree->fs_info, err,
687 "locking error: extent tree was modified by another thread while locked");
688}
689
690/*
691 * clear some bits on a range in the tree. This may require splitting
692 * or inserting elements in the tree, so the gfp mask is used to
693 * indicate which allocations or sleeping are allowed.
694 *
695 * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
696 * the given range from the tree regardless of state (ie for truncate).
697 *
698 * the range [start, end] is inclusive.
699 *
700 * This takes the tree lock, and returns 0 on success and < 0 on error.
701 */
702int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
703 u32 bits, int wake, int delete,
704 struct extent_state **cached_state,
705 gfp_t mask, struct extent_changeset *changeset)
706{
707 struct extent_state *state;
708 struct extent_state *cached;
709 struct extent_state *prealloc = NULL;
710 struct rb_node *node;
711 u64 last_end;
712 int err;
713 int clear = 0;
714
715 btrfs_debug_check_extent_io_range(tree, start, end);
716 trace_btrfs_clear_extent_bit(tree, start, end - start + 1, bits);
717
718 if (bits & EXTENT_DELALLOC)
719 bits |= EXTENT_NORESERVE;
720
721 if (delete)
722 bits |= ~EXTENT_CTLBITS;
723
724 if (bits & (EXTENT_LOCKED | EXTENT_BOUNDARY))
725 clear = 1;
726again:
727 if (!prealloc && gfpflags_allow_blocking(mask)) {
728 /*
729 * Don't care for allocation failure here because we might end
730 * up not needing the pre-allocated extent state at all, which
731 * is the case if we only have in the tree extent states that
732 * cover our input range and don't cover too any other range.
733 * If we end up needing a new extent state we allocate it later.
734 */
735 prealloc = alloc_extent_state(mask);
736 }
737
738 spin_lock(&tree->lock);
739 if (cached_state) {
740 cached = *cached_state;
741
742 if (clear) {
743 *cached_state = NULL;
744 cached_state = NULL;
745 }
746
747 if (cached && extent_state_in_tree(cached) &&
748 cached->start <= start && cached->end > start) {
749 if (clear)
750 refcount_dec(&cached->refs);
751 state = cached;
752 goto hit_next;
753 }
754 if (clear)
755 free_extent_state(cached);
756 }
757 /*
758 * this search will find the extents that end after
759 * our range starts
760 */
761 node = tree_search(tree, start);
762 if (!node)
763 goto out;
764 state = rb_entry(node, struct extent_state, rb_node);
765hit_next:
766 if (state->start > end)
767 goto out;
768 WARN_ON(state->end < start);
769 last_end = state->end;
770
771 /* the state doesn't have the wanted bits, go ahead */
772 if (!(state->state & bits)) {
773 state = next_state(state);
774 goto next;
775 }
776
777 /*
778 * | ---- desired range ---- |
779 * | state | or
780 * | ------------- state -------------- |
781 *
782 * We need to split the extent we found, and may flip
783 * bits on second half.
784 *
785 * If the extent we found extends past our range, we
786 * just split and search again. It'll get split again
787 * the next time though.
788 *
789 * If the extent we found is inside our range, we clear
790 * the desired bit on it.
791 */
792
793 if (state->start < start) {
794 prealloc = alloc_extent_state_atomic(prealloc);
795 BUG_ON(!prealloc);
796 err = split_state(tree, state, prealloc, start);
797 if (err)
798 extent_io_tree_panic(tree, err);
799
800 prealloc = NULL;
801 if (err)
802 goto out;
803 if (state->end <= end) {
804 state = clear_state_bit(tree, state, &bits, wake,
805 changeset);
806 goto next;
807 }
808 goto search_again;
809 }
810 /*
811 * | ---- desired range ---- |
812 * | state |
813 * We need to split the extent, and clear the bit
814 * on the first half
815 */
816 if (state->start <= end && state->end > end) {
817 prealloc = alloc_extent_state_atomic(prealloc);
818 BUG_ON(!prealloc);
819 err = split_state(tree, state, prealloc, end + 1);
820 if (err)
821 extent_io_tree_panic(tree, err);
822
823 if (wake)
824 wake_up(&state->wq);
825
826 clear_state_bit(tree, prealloc, &bits, wake, changeset);
827
828 prealloc = NULL;
829 goto out;
830 }
831
832 state = clear_state_bit(tree, state, &bits, wake, changeset);
833next:
834 if (last_end == (u64)-1)
835 goto out;
836 start = last_end + 1;
837 if (start <= end && state && !need_resched())
838 goto hit_next;
839
840search_again:
841 if (start > end)
842 goto out;
843 spin_unlock(&tree->lock);
844 if (gfpflags_allow_blocking(mask))
845 cond_resched();
846 goto again;
847
848out:
849 spin_unlock(&tree->lock);
850 if (prealloc)
851 free_extent_state(prealloc);
852
853 return 0;
854
855}
856
857static void wait_on_state(struct extent_io_tree *tree,
858 struct extent_state *state)
859 __releases(tree->lock)
860 __acquires(tree->lock)
861{
862 DEFINE_WAIT(wait);
863 prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
864 spin_unlock(&tree->lock);
865 schedule();
866 spin_lock(&tree->lock);
867 finish_wait(&state->wq, &wait);
868}
869
870/*
871 * waits for one or more bits to clear on a range in the state tree.
872 * The range [start, end] is inclusive.
873 * The tree lock is taken by this function
874 */
875static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
876 u32 bits)
877{
878 struct extent_state *state;
879 struct rb_node *node;
880
881 btrfs_debug_check_extent_io_range(tree, start, end);
882
883 spin_lock(&tree->lock);
884again:
885 while (1) {
886 /*
887 * this search will find all the extents that end after
888 * our range starts
889 */
890 node = tree_search(tree, start);
891process_node:
892 if (!node)
893 break;
894
895 state = rb_entry(node, struct extent_state, rb_node);
896
897 if (state->start > end)
898 goto out;
899
900 if (state->state & bits) {
901 start = state->start;
902 refcount_inc(&state->refs);
903 wait_on_state(tree, state);
904 free_extent_state(state);
905 goto again;
906 }
907 start = state->end + 1;
908
909 if (start > end)
910 break;
911
912 if (!cond_resched_lock(&tree->lock)) {
913 node = rb_next(node);
914 goto process_node;
915 }
916 }
917out:
918 spin_unlock(&tree->lock);
919}
920
921static void set_state_bits(struct extent_io_tree *tree,
922 struct extent_state *state,
923 u32 *bits, struct extent_changeset *changeset)
924{
925 u32 bits_to_set = *bits & ~EXTENT_CTLBITS;
926 int ret;
927
928 if (tree->private_data && is_data_inode(tree->private_data))
929 btrfs_set_delalloc_extent(tree->private_data, state, bits);
930
931 if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
932 u64 range = state->end - state->start + 1;
933 tree->dirty_bytes += range;
934 }
935 ret = add_extent_changeset(state, bits_to_set, changeset, 1);
936 BUG_ON(ret < 0);
937 state->state |= bits_to_set;
938}
939
940static void cache_state_if_flags(struct extent_state *state,
941 struct extent_state **cached_ptr,
942 unsigned flags)
943{
944 if (cached_ptr && !(*cached_ptr)) {
945 if (!flags || (state->state & flags)) {
946 *cached_ptr = state;
947 refcount_inc(&state->refs);
948 }
949 }
950}
951
952static void cache_state(struct extent_state *state,
953 struct extent_state **cached_ptr)
954{
955 return cache_state_if_flags(state, cached_ptr,
956 EXTENT_LOCKED | EXTENT_BOUNDARY);
957}
958
959/*
960 * set some bits on a range in the tree. This may require allocations or
961 * sleeping, so the gfp mask is used to indicate what is allowed.
962 *
963 * If any of the exclusive bits are set, this will fail with -EEXIST if some
964 * part of the range already has the desired bits set. The start of the
965 * existing range is returned in failed_start in this case.
966 *
967 * [start, end] is inclusive This takes the tree lock.
968 */
969int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, u32 bits,
970 u32 exclusive_bits, u64 *failed_start,
971 struct extent_state **cached_state, gfp_t mask,
972 struct extent_changeset *changeset)
973{
974 struct extent_state *state;
975 struct extent_state *prealloc = NULL;
976 struct rb_node *node;
977 struct rb_node **p;
978 struct rb_node *parent;
979 int err = 0;
980 u64 last_start;
981 u64 last_end;
982
983 btrfs_debug_check_extent_io_range(tree, start, end);
984 trace_btrfs_set_extent_bit(tree, start, end - start + 1, bits);
985
986 if (exclusive_bits)
987 ASSERT(failed_start);
988 else
989 ASSERT(failed_start == NULL);
990again:
991 if (!prealloc && gfpflags_allow_blocking(mask)) {
992 /*
993 * Don't care for allocation failure here because we might end
994 * up not needing the pre-allocated extent state at all, which
995 * is the case if we only have in the tree extent states that
996 * cover our input range and don't cover too any other range.
997 * If we end up needing a new extent state we allocate it later.
998 */
999 prealloc = alloc_extent_state(mask);
1000 }
1001
1002 spin_lock(&tree->lock);
1003 if (cached_state && *cached_state) {
1004 state = *cached_state;
1005 if (state->start <= start && state->end > start &&
1006 extent_state_in_tree(state)) {
1007 node = &state->rb_node;
1008 goto hit_next;
1009 }
1010 }
1011 /*
1012 * this search will find all the extents that end after
1013 * our range starts.
1014 */
1015 node = tree_search_for_insert(tree, start, &p, &parent);
1016 if (!node) {
1017 prealloc = alloc_extent_state_atomic(prealloc);
1018 BUG_ON(!prealloc);
1019 err = insert_state(tree, prealloc, start, end,
1020 &p, &parent, &bits, changeset);
1021 if (err)
1022 extent_io_tree_panic(tree, err);
1023
1024 cache_state(prealloc, cached_state);
1025 prealloc = NULL;
1026 goto out;
1027 }
1028 state = rb_entry(node, struct extent_state, rb_node);
1029hit_next:
1030 last_start = state->start;
1031 last_end = state->end;
1032
1033 /*
1034 * | ---- desired range ---- |
1035 * | state |
1036 *
1037 * Just lock what we found and keep going
1038 */
1039 if (state->start == start && state->end <= end) {
1040 if (state->state & exclusive_bits) {
1041 *failed_start = state->start;
1042 err = -EEXIST;
1043 goto out;
1044 }
1045
1046 set_state_bits(tree, state, &bits, changeset);
1047 cache_state(state, cached_state);
1048 merge_state(tree, state);
1049 if (last_end == (u64)-1)
1050 goto out;
1051 start = last_end + 1;
1052 state = next_state(state);
1053 if (start < end && state && state->start == start &&
1054 !need_resched())
1055 goto hit_next;
1056 goto search_again;
1057 }
1058
1059 /*
1060 * | ---- desired range ---- |
1061 * | state |
1062 * or
1063 * | ------------- state -------------- |
1064 *
1065 * We need to split the extent we found, and may flip bits on
1066 * second half.
1067 *
1068 * If the extent we found extends past our
1069 * range, we just split and search again. It'll get split
1070 * again the next time though.
1071 *
1072 * If the extent we found is inside our range, we set the
1073 * desired bit on it.
1074 */
1075 if (state->start < start) {
1076 if (state->state & exclusive_bits) {
1077 *failed_start = start;
1078 err = -EEXIST;
1079 goto out;
1080 }
1081
1082 /*
1083 * If this extent already has all the bits we want set, then
1084 * skip it, not necessary to split it or do anything with it.
1085 */
1086 if ((state->state & bits) == bits) {
1087 start = state->end + 1;
1088 cache_state(state, cached_state);
1089 goto search_again;
1090 }
1091
1092 prealloc = alloc_extent_state_atomic(prealloc);
1093 BUG_ON(!prealloc);
1094 err = split_state(tree, state, prealloc, start);
1095 if (err)
1096 extent_io_tree_panic(tree, err);
1097
1098 prealloc = NULL;
1099 if (err)
1100 goto out;
1101 if (state->end <= end) {
1102 set_state_bits(tree, state, &bits, changeset);
1103 cache_state(state, cached_state);
1104 merge_state(tree, state);
1105 if (last_end == (u64)-1)
1106 goto out;
1107 start = last_end + 1;
1108 state = next_state(state);
1109 if (start < end && state && state->start == start &&
1110 !need_resched())
1111 goto hit_next;
1112 }
1113 goto search_again;
1114 }
1115 /*
1116 * | ---- desired range ---- |
1117 * | state | or | state |
1118 *
1119 * There's a hole, we need to insert something in it and
1120 * ignore the extent we found.
1121 */
1122 if (state->start > start) {
1123 u64 this_end;
1124 if (end < last_start)
1125 this_end = end;
1126 else
1127 this_end = last_start - 1;
1128
1129 prealloc = alloc_extent_state_atomic(prealloc);
1130 BUG_ON(!prealloc);
1131
1132 /*
1133 * Avoid to free 'prealloc' if it can be merged with
1134 * the later extent.
1135 */
1136 err = insert_state(tree, prealloc, start, this_end,
1137 NULL, NULL, &bits, changeset);
1138 if (err)
1139 extent_io_tree_panic(tree, err);
1140
1141 cache_state(prealloc, cached_state);
1142 prealloc = NULL;
1143 start = this_end + 1;
1144 goto search_again;
1145 }
1146 /*
1147 * | ---- desired range ---- |
1148 * | state |
1149 * We need to split the extent, and set the bit
1150 * on the first half
1151 */
1152 if (state->start <= end && state->end > end) {
1153 if (state->state & exclusive_bits) {
1154 *failed_start = start;
1155 err = -EEXIST;
1156 goto out;
1157 }
1158
1159 prealloc = alloc_extent_state_atomic(prealloc);
1160 BUG_ON(!prealloc);
1161 err = split_state(tree, state, prealloc, end + 1);
1162 if (err)
1163 extent_io_tree_panic(tree, err);
1164
1165 set_state_bits(tree, prealloc, &bits, changeset);
1166 cache_state(prealloc, cached_state);
1167 merge_state(tree, prealloc);
1168 prealloc = NULL;
1169 goto out;
1170 }
1171
1172search_again:
1173 if (start > end)
1174 goto out;
1175 spin_unlock(&tree->lock);
1176 if (gfpflags_allow_blocking(mask))
1177 cond_resched();
1178 goto again;
1179
1180out:
1181 spin_unlock(&tree->lock);
1182 if (prealloc)
1183 free_extent_state(prealloc);
1184
1185 return err;
1186
1187}
1188
1189/**
1190 * convert_extent_bit - convert all bits in a given range from one bit to
1191 * another
1192 * @tree: the io tree to search
1193 * @start: the start offset in bytes
1194 * @end: the end offset in bytes (inclusive)
1195 * @bits: the bits to set in this range
1196 * @clear_bits: the bits to clear in this range
1197 * @cached_state: state that we're going to cache
1198 *
1199 * This will go through and set bits for the given range. If any states exist
1200 * already in this range they are set with the given bit and cleared of the
1201 * clear_bits. This is only meant to be used by things that are mergeable, ie
1202 * converting from say DELALLOC to DIRTY. This is not meant to be used with
1203 * boundary bits like LOCK.
1204 *
1205 * All allocations are done with GFP_NOFS.
1206 */
1207int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1208 u32 bits, u32 clear_bits,
1209 struct extent_state **cached_state)
1210{
1211 struct extent_state *state;
1212 struct extent_state *prealloc = NULL;
1213 struct rb_node *node;
1214 struct rb_node **p;
1215 struct rb_node *parent;
1216 int err = 0;
1217 u64 last_start;
1218 u64 last_end;
1219 bool first_iteration = true;
1220
1221 btrfs_debug_check_extent_io_range(tree, start, end);
1222 trace_btrfs_convert_extent_bit(tree, start, end - start + 1, bits,
1223 clear_bits);
1224
1225again:
1226 if (!prealloc) {
1227 /*
1228 * Best effort, don't worry if extent state allocation fails
1229 * here for the first iteration. We might have a cached state
1230 * that matches exactly the target range, in which case no
1231 * extent state allocations are needed. We'll only know this
1232 * after locking the tree.
1233 */
1234 prealloc = alloc_extent_state(GFP_NOFS);
1235 if (!prealloc && !first_iteration)
1236 return -ENOMEM;
1237 }
1238
1239 spin_lock(&tree->lock);
1240 if (cached_state && *cached_state) {
1241 state = *cached_state;
1242 if (state->start <= start && state->end > start &&
1243 extent_state_in_tree(state)) {
1244 node = &state->rb_node;
1245 goto hit_next;
1246 }
1247 }
1248
1249 /*
1250 * this search will find all the extents that end after
1251 * our range starts.
1252 */
1253 node = tree_search_for_insert(tree, start, &p, &parent);
1254 if (!node) {
1255 prealloc = alloc_extent_state_atomic(prealloc);
1256 if (!prealloc) {
1257 err = -ENOMEM;
1258 goto out;
1259 }
1260 err = insert_state(tree, prealloc, start, end,
1261 &p, &parent, &bits, NULL);
1262 if (err)
1263 extent_io_tree_panic(tree, err);
1264 cache_state(prealloc, cached_state);
1265 prealloc = NULL;
1266 goto out;
1267 }
1268 state = rb_entry(node, struct extent_state, rb_node);
1269hit_next:
1270 last_start = state->start;
1271 last_end = state->end;
1272
1273 /*
1274 * | ---- desired range ---- |
1275 * | state |
1276 *
1277 * Just lock what we found and keep going
1278 */
1279 if (state->start == start && state->end <= end) {
1280 set_state_bits(tree, state, &bits, NULL);
1281 cache_state(state, cached_state);
1282 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1283 if (last_end == (u64)-1)
1284 goto out;
1285 start = last_end + 1;
1286 if (start < end && state && state->start == start &&
1287 !need_resched())
1288 goto hit_next;
1289 goto search_again;
1290 }
1291
1292 /*
1293 * | ---- desired range ---- |
1294 * | state |
1295 * or
1296 * | ------------- state -------------- |
1297 *
1298 * We need to split the extent we found, and may flip bits on
1299 * second half.
1300 *
1301 * If the extent we found extends past our
1302 * range, we just split and search again. It'll get split
1303 * again the next time though.
1304 *
1305 * If the extent we found is inside our range, we set the
1306 * desired bit on it.
1307 */
1308 if (state->start < start) {
1309 prealloc = alloc_extent_state_atomic(prealloc);
1310 if (!prealloc) {
1311 err = -ENOMEM;
1312 goto out;
1313 }
1314 err = split_state(tree, state, prealloc, start);
1315 if (err)
1316 extent_io_tree_panic(tree, err);
1317 prealloc = NULL;
1318 if (err)
1319 goto out;
1320 if (state->end <= end) {
1321 set_state_bits(tree, state, &bits, NULL);
1322 cache_state(state, cached_state);
1323 state = clear_state_bit(tree, state, &clear_bits, 0,
1324 NULL);
1325 if (last_end == (u64)-1)
1326 goto out;
1327 start = last_end + 1;
1328 if (start < end && state && state->start == start &&
1329 !need_resched())
1330 goto hit_next;
1331 }
1332 goto search_again;
1333 }
1334 /*
1335 * | ---- desired range ---- |
1336 * | state | or | state |
1337 *
1338 * There's a hole, we need to insert something in it and
1339 * ignore the extent we found.
1340 */
1341 if (state->start > start) {
1342 u64 this_end;
1343 if (end < last_start)
1344 this_end = end;
1345 else
1346 this_end = last_start - 1;
1347
1348 prealloc = alloc_extent_state_atomic(prealloc);
1349 if (!prealloc) {
1350 err = -ENOMEM;
1351 goto out;
1352 }
1353
1354 /*
1355 * Avoid to free 'prealloc' if it can be merged with
1356 * the later extent.
1357 */
1358 err = insert_state(tree, prealloc, start, this_end,
1359 NULL, NULL, &bits, NULL);
1360 if (err)
1361 extent_io_tree_panic(tree, err);
1362 cache_state(prealloc, cached_state);
1363 prealloc = NULL;
1364 start = this_end + 1;
1365 goto search_again;
1366 }
1367 /*
1368 * | ---- desired range ---- |
1369 * | state |
1370 * We need to split the extent, and set the bit
1371 * on the first half
1372 */
1373 if (state->start <= end && state->end > end) {
1374 prealloc = alloc_extent_state_atomic(prealloc);
1375 if (!prealloc) {
1376 err = -ENOMEM;
1377 goto out;
1378 }
1379
1380 err = split_state(tree, state, prealloc, end + 1);
1381 if (err)
1382 extent_io_tree_panic(tree, err);
1383
1384 set_state_bits(tree, prealloc, &bits, NULL);
1385 cache_state(prealloc, cached_state);
1386 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1387 prealloc = NULL;
1388 goto out;
1389 }
1390
1391search_again:
1392 if (start > end)
1393 goto out;
1394 spin_unlock(&tree->lock);
1395 cond_resched();
1396 first_iteration = false;
1397 goto again;
1398
1399out:
1400 spin_unlock(&tree->lock);
1401 if (prealloc)
1402 free_extent_state(prealloc);
1403
1404 return err;
1405}
1406
1407/* wrappers around set/clear extent bit */
1408int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1409 u32 bits, struct extent_changeset *changeset)
1410{
1411 /*
1412 * We don't support EXTENT_LOCKED yet, as current changeset will
1413 * record any bits changed, so for EXTENT_LOCKED case, it will
1414 * either fail with -EEXIST or changeset will record the whole
1415 * range.
1416 */
1417 BUG_ON(bits & EXTENT_LOCKED);
1418
1419 return set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1420 changeset);
1421}
1422
1423int set_extent_bits_nowait(struct extent_io_tree *tree, u64 start, u64 end,
1424 u32 bits)
1425{
1426 return set_extent_bit(tree, start, end, bits, 0, NULL, NULL,
1427 GFP_NOWAIT, NULL);
1428}
1429
1430int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1431 u32 bits, int wake, int delete,
1432 struct extent_state **cached)
1433{
1434 return __clear_extent_bit(tree, start, end, bits, wake, delete,
1435 cached, GFP_NOFS, NULL);
1436}
1437
1438int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1439 u32 bits, struct extent_changeset *changeset)
1440{
1441 /*
1442 * Don't support EXTENT_LOCKED case, same reason as
1443 * set_record_extent_bits().
1444 */
1445 BUG_ON(bits & EXTENT_LOCKED);
1446
1447 return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1448 changeset);
1449}
1450
1451/*
1452 * either insert or lock state struct between start and end use mask to tell
1453 * us if waiting is desired.
1454 */
1455int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1456 struct extent_state **cached_state)
1457{
1458 int err;
1459 u64 failed_start;
1460
1461 while (1) {
1462 err = set_extent_bit(tree, start, end, EXTENT_LOCKED,
1463 EXTENT_LOCKED, &failed_start,
1464 cached_state, GFP_NOFS, NULL);
1465 if (err == -EEXIST) {
1466 wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1467 start = failed_start;
1468 } else
1469 break;
1470 WARN_ON(start > end);
1471 }
1472 return err;
1473}
1474
1475int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1476{
1477 int err;
1478 u64 failed_start;
1479
1480 err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1481 &failed_start, NULL, GFP_NOFS, NULL);
1482 if (err == -EEXIST) {
1483 if (failed_start > start)
1484 clear_extent_bit(tree, start, failed_start - 1,
1485 EXTENT_LOCKED, 1, 0, NULL);
1486 return 0;
1487 }
1488 return 1;
1489}
1490
1491void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1492{
1493 unsigned long index = start >> PAGE_SHIFT;
1494 unsigned long end_index = end >> PAGE_SHIFT;
1495 struct page *page;
1496
1497 while (index <= end_index) {
1498 page = find_get_page(inode->i_mapping, index);
1499 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1500 clear_page_dirty_for_io(page);
1501 put_page(page);
1502 index++;
1503 }
1504}
1505
1506void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1507{
1508 unsigned long index = start >> PAGE_SHIFT;
1509 unsigned long end_index = end >> PAGE_SHIFT;
1510 struct page *page;
1511
1512 while (index <= end_index) {
1513 page = find_get_page(inode->i_mapping, index);
1514 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1515 __set_page_dirty_nobuffers(page);
1516 account_page_redirty(page);
1517 put_page(page);
1518 index++;
1519 }
1520}
1521
1522/* find the first state struct with 'bits' set after 'start', and
1523 * return it. tree->lock must be held. NULL will returned if
1524 * nothing was found after 'start'
1525 */
1526static struct extent_state *
1527find_first_extent_bit_state(struct extent_io_tree *tree, u64 start, u32 bits)
1528{
1529 struct rb_node *node;
1530 struct extent_state *state;
1531
1532 /*
1533 * this search will find all the extents that end after
1534 * our range starts.
1535 */
1536 node = tree_search(tree, start);
1537 if (!node)
1538 goto out;
1539
1540 while (1) {
1541 state = rb_entry(node, struct extent_state, rb_node);
1542 if (state->end >= start && (state->state & bits))
1543 return state;
1544
1545 node = rb_next(node);
1546 if (!node)
1547 break;
1548 }
1549out:
1550 return NULL;
1551}
1552
1553/*
1554 * Find the first offset in the io tree with one or more @bits set.
1555 *
1556 * Note: If there are multiple bits set in @bits, any of them will match.
1557 *
1558 * Return 0 if we find something, and update @start_ret and @end_ret.
1559 * Return 1 if we found nothing.
1560 */
1561int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1562 u64 *start_ret, u64 *end_ret, u32 bits,
1563 struct extent_state **cached_state)
1564{
1565 struct extent_state *state;
1566 int ret = 1;
1567
1568 spin_lock(&tree->lock);
1569 if (cached_state && *cached_state) {
1570 state = *cached_state;
1571 if (state->end == start - 1 && extent_state_in_tree(state)) {
1572 while ((state = next_state(state)) != NULL) {
1573 if (state->state & bits)
1574 goto got_it;
1575 }
1576 free_extent_state(*cached_state);
1577 *cached_state = NULL;
1578 goto out;
1579 }
1580 free_extent_state(*cached_state);
1581 *cached_state = NULL;
1582 }
1583
1584 state = find_first_extent_bit_state(tree, start, bits);
1585got_it:
1586 if (state) {
1587 cache_state_if_flags(state, cached_state, 0);
1588 *start_ret = state->start;
1589 *end_ret = state->end;
1590 ret = 0;
1591 }
1592out:
1593 spin_unlock(&tree->lock);
1594 return ret;
1595}
1596
1597/**
1598 * Find a contiguous area of bits
1599 *
1600 * @tree: io tree to check
1601 * @start: offset to start the search from
1602 * @start_ret: the first offset we found with the bits set
1603 * @end_ret: the final contiguous range of the bits that were set
1604 * @bits: bits to look for
1605 *
1606 * set_extent_bit and clear_extent_bit can temporarily split contiguous ranges
1607 * to set bits appropriately, and then merge them again. During this time it
1608 * will drop the tree->lock, so use this helper if you want to find the actual
1609 * contiguous area for given bits. We will search to the first bit we find, and
1610 * then walk down the tree until we find a non-contiguous area. The area
1611 * returned will be the full contiguous area with the bits set.
1612 */
1613int find_contiguous_extent_bit(struct extent_io_tree *tree, u64 start,
1614 u64 *start_ret, u64 *end_ret, u32 bits)
1615{
1616 struct extent_state *state;
1617 int ret = 1;
1618
1619 spin_lock(&tree->lock);
1620 state = find_first_extent_bit_state(tree, start, bits);
1621 if (state) {
1622 *start_ret = state->start;
1623 *end_ret = state->end;
1624 while ((state = next_state(state)) != NULL) {
1625 if (state->start > (*end_ret + 1))
1626 break;
1627 *end_ret = state->end;
1628 }
1629 ret = 0;
1630 }
1631 spin_unlock(&tree->lock);
1632 return ret;
1633}
1634
1635/**
1636 * Find the first range that has @bits not set. This range could start before
1637 * @start.
1638 *
1639 * @tree: the tree to search
1640 * @start: offset at/after which the found extent should start
1641 * @start_ret: records the beginning of the range
1642 * @end_ret: records the end of the range (inclusive)
1643 * @bits: the set of bits which must be unset
1644 *
1645 * Since unallocated range is also considered one which doesn't have the bits
1646 * set it's possible that @end_ret contains -1, this happens in case the range
1647 * spans (last_range_end, end of device]. In this case it's up to the caller to
1648 * trim @end_ret to the appropriate size.
1649 */
1650void find_first_clear_extent_bit(struct extent_io_tree *tree, u64 start,
1651 u64 *start_ret, u64 *end_ret, u32 bits)
1652{
1653 struct extent_state *state;
1654 struct rb_node *node, *prev = NULL, *next;
1655
1656 spin_lock(&tree->lock);
1657
1658 /* Find first extent with bits cleared */
1659 while (1) {
1660 node = __etree_search(tree, start, &next, &prev, NULL, NULL);
1661 if (!node && !next && !prev) {
1662 /*
1663 * Tree is completely empty, send full range and let
1664 * caller deal with it
1665 */
1666 *start_ret = 0;
1667 *end_ret = -1;
1668 goto out;
1669 } else if (!node && !next) {
1670 /*
1671 * We are past the last allocated chunk, set start at
1672 * the end of the last extent.
1673 */
1674 state = rb_entry(prev, struct extent_state, rb_node);
1675 *start_ret = state->end + 1;
1676 *end_ret = -1;
1677 goto out;
1678 } else if (!node) {
1679 node = next;
1680 }
1681 /*
1682 * At this point 'node' either contains 'start' or start is
1683 * before 'node'
1684 */
1685 state = rb_entry(node, struct extent_state, rb_node);
1686
1687 if (in_range(start, state->start, state->end - state->start + 1)) {
1688 if (state->state & bits) {
1689 /*
1690 * |--range with bits sets--|
1691 * |
1692 * start
1693 */
1694 start = state->end + 1;
1695 } else {
1696 /*
1697 * 'start' falls within a range that doesn't
1698 * have the bits set, so take its start as
1699 * the beginning of the desired range
1700 *
1701 * |--range with bits cleared----|
1702 * |
1703 * start
1704 */
1705 *start_ret = state->start;
1706 break;
1707 }
1708 } else {
1709 /*
1710 * |---prev range---|---hole/unset---|---node range---|
1711 * |
1712 * start
1713 *
1714 * or
1715 *
1716 * |---hole/unset--||--first node--|
1717 * 0 |
1718 * start
1719 */
1720 if (prev) {
1721 state = rb_entry(prev, struct extent_state,
1722 rb_node);
1723 *start_ret = state->end + 1;
1724 } else {
1725 *start_ret = 0;
1726 }
1727 break;
1728 }
1729 }
1730
1731 /*
1732 * Find the longest stretch from start until an entry which has the
1733 * bits set
1734 */
1735 while (1) {
1736 state = rb_entry(node, struct extent_state, rb_node);
1737 if (state->end >= start && !(state->state & bits)) {
1738 *end_ret = state->end;
1739 } else {
1740 *end_ret = state->start - 1;
1741 break;
1742 }
1743
1744 node = rb_next(node);
1745 if (!node)
1746 break;
1747 }
1748out:
1749 spin_unlock(&tree->lock);
1750}
1751
1752/*
1753 * find a contiguous range of bytes in the file marked as delalloc, not
1754 * more than 'max_bytes'. start and end are used to return the range,
1755 *
1756 * true is returned if we find something, false if nothing was in the tree
1757 */
1758bool btrfs_find_delalloc_range(struct extent_io_tree *tree, u64 *start,
1759 u64 *end, u64 max_bytes,
1760 struct extent_state **cached_state)
1761{
1762 struct rb_node *node;
1763 struct extent_state *state;
1764 u64 cur_start = *start;
1765 bool found = false;
1766 u64 total_bytes = 0;
1767
1768 spin_lock(&tree->lock);
1769
1770 /*
1771 * this search will find all the extents that end after
1772 * our range starts.
1773 */
1774 node = tree_search(tree, cur_start);
1775 if (!node) {
1776 *end = (u64)-1;
1777 goto out;
1778 }
1779
1780 while (1) {
1781 state = rb_entry(node, struct extent_state, rb_node);
1782 if (found && (state->start != cur_start ||
1783 (state->state & EXTENT_BOUNDARY))) {
1784 goto out;
1785 }
1786 if (!(state->state & EXTENT_DELALLOC)) {
1787 if (!found)
1788 *end = state->end;
1789 goto out;
1790 }
1791 if (!found) {
1792 *start = state->start;
1793 *cached_state = state;
1794 refcount_inc(&state->refs);
1795 }
1796 found = true;
1797 *end = state->end;
1798 cur_start = state->end + 1;
1799 node = rb_next(node);
1800 total_bytes += state->end - state->start + 1;
1801 if (total_bytes >= max_bytes)
1802 break;
1803 if (!node)
1804 break;
1805 }
1806out:
1807 spin_unlock(&tree->lock);
1808 return found;
1809}
1810
1811/*
1812 * Process one page for __process_pages_contig().
1813 *
1814 * Return >0 if we hit @page == @locked_page.
1815 * Return 0 if we updated the page status.
1816 * Return -EGAIN if the we need to try again.
1817 * (For PAGE_LOCK case but got dirty page or page not belong to mapping)
1818 */
1819static int process_one_page(struct btrfs_fs_info *fs_info,
1820 struct address_space *mapping,
1821 struct page *page, struct page *locked_page,
1822 unsigned long page_ops, u64 start, u64 end)
1823{
1824 u32 len;
1825
1826 ASSERT(end + 1 - start != 0 && end + 1 - start < U32_MAX);
1827 len = end + 1 - start;
1828
1829 if (page_ops & PAGE_SET_ORDERED)
1830 btrfs_page_clamp_set_ordered(fs_info, page, start, len);
1831 if (page_ops & PAGE_SET_ERROR)
1832 btrfs_page_clamp_set_error(fs_info, page, start, len);
1833 if (page_ops & PAGE_START_WRITEBACK) {
1834 btrfs_page_clamp_clear_dirty(fs_info, page, start, len);
1835 btrfs_page_clamp_set_writeback(fs_info, page, start, len);
1836 }
1837 if (page_ops & PAGE_END_WRITEBACK)
1838 btrfs_page_clamp_clear_writeback(fs_info, page, start, len);
1839
1840 if (page == locked_page)
1841 return 1;
1842
1843 if (page_ops & PAGE_LOCK) {
1844 int ret;
1845
1846 ret = btrfs_page_start_writer_lock(fs_info, page, start, len);
1847 if (ret)
1848 return ret;
1849 if (!PageDirty(page) || page->mapping != mapping) {
1850 btrfs_page_end_writer_lock(fs_info, page, start, len);
1851 return -EAGAIN;
1852 }
1853 }
1854 if (page_ops & PAGE_UNLOCK)
1855 btrfs_page_end_writer_lock(fs_info, page, start, len);
1856 return 0;
1857}
1858
1859static int __process_pages_contig(struct address_space *mapping,
1860 struct page *locked_page,
1861 u64 start, u64 end, unsigned long page_ops,
1862 u64 *processed_end)
1863{
1864 struct btrfs_fs_info *fs_info = btrfs_sb(mapping->host->i_sb);
1865 pgoff_t start_index = start >> PAGE_SHIFT;
1866 pgoff_t end_index = end >> PAGE_SHIFT;
1867 pgoff_t index = start_index;
1868 unsigned long nr_pages = end_index - start_index + 1;
1869 unsigned long pages_processed = 0;
1870 struct page *pages[16];
1871 int err = 0;
1872 int i;
1873
1874 if (page_ops & PAGE_LOCK) {
1875 ASSERT(page_ops == PAGE_LOCK);
1876 ASSERT(processed_end && *processed_end == start);
1877 }
1878
1879 if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1880 mapping_set_error(mapping, -EIO);
1881
1882 while (nr_pages > 0) {
1883 int found_pages;
1884
1885 found_pages = find_get_pages_contig(mapping, index,
1886 min_t(unsigned long,
1887 nr_pages, ARRAY_SIZE(pages)), pages);
1888 if (found_pages == 0) {
1889 /*
1890 * Only if we're going to lock these pages, we can find
1891 * nothing at @index.
1892 */
1893 ASSERT(page_ops & PAGE_LOCK);
1894 err = -EAGAIN;
1895 goto out;
1896 }
1897
1898 for (i = 0; i < found_pages; i++) {
1899 int process_ret;
1900
1901 process_ret = process_one_page(fs_info, mapping,
1902 pages[i], locked_page, page_ops,
1903 start, end);
1904 if (process_ret < 0) {
1905 for (; i < found_pages; i++)
1906 put_page(pages[i]);
1907 err = -EAGAIN;
1908 goto out;
1909 }
1910 put_page(pages[i]);
1911 pages_processed++;
1912 }
1913 nr_pages -= found_pages;
1914 index += found_pages;
1915 cond_resched();
1916 }
1917out:
1918 if (err && processed_end) {
1919 /*
1920 * Update @processed_end. I know this is awful since it has
1921 * two different return value patterns (inclusive vs exclusive).
1922 *
1923 * But the exclusive pattern is necessary if @start is 0, or we
1924 * underflow and check against processed_end won't work as
1925 * expected.
1926 */
1927 if (pages_processed)
1928 *processed_end = min(end,
1929 ((u64)(start_index + pages_processed) << PAGE_SHIFT) - 1);
1930 else
1931 *processed_end = start;
1932 }
1933 return err;
1934}
1935
1936static noinline void __unlock_for_delalloc(struct inode *inode,
1937 struct page *locked_page,
1938 u64 start, u64 end)
1939{
1940 unsigned long index = start >> PAGE_SHIFT;
1941 unsigned long end_index = end >> PAGE_SHIFT;
1942
1943 ASSERT(locked_page);
1944 if (index == locked_page->index && end_index == index)
1945 return;
1946
1947 __process_pages_contig(inode->i_mapping, locked_page, start, end,
1948 PAGE_UNLOCK, NULL);
1949}
1950
1951static noinline int lock_delalloc_pages(struct inode *inode,
1952 struct page *locked_page,
1953 u64 delalloc_start,
1954 u64 delalloc_end)
1955{
1956 unsigned long index = delalloc_start >> PAGE_SHIFT;
1957 unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1958 u64 processed_end = delalloc_start;
1959 int ret;
1960
1961 ASSERT(locked_page);
1962 if (index == locked_page->index && index == end_index)
1963 return 0;
1964
1965 ret = __process_pages_contig(inode->i_mapping, locked_page, delalloc_start,
1966 delalloc_end, PAGE_LOCK, &processed_end);
1967 if (ret == -EAGAIN && processed_end > delalloc_start)
1968 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1969 processed_end);
1970 return ret;
1971}
1972
1973/*
1974 * Find and lock a contiguous range of bytes in the file marked as delalloc, no
1975 * more than @max_bytes. @Start and @end are used to return the range,
1976 *
1977 * Return: true if we find something
1978 * false if nothing was in the tree
1979 */
1980EXPORT_FOR_TESTS
1981noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
1982 struct page *locked_page, u64 *start,
1983 u64 *end)
1984{
1985 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1986 u64 max_bytes = BTRFS_MAX_EXTENT_SIZE;
1987 u64 delalloc_start;
1988 u64 delalloc_end;
1989 bool found;
1990 struct extent_state *cached_state = NULL;
1991 int ret;
1992 int loops = 0;
1993
1994again:
1995 /* step one, find a bunch of delalloc bytes starting at start */
1996 delalloc_start = *start;
1997 delalloc_end = 0;
1998 found = btrfs_find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1999 max_bytes, &cached_state);
2000 if (!found || delalloc_end <= *start) {
2001 *start = delalloc_start;
2002 *end = delalloc_end;
2003 free_extent_state(cached_state);
2004 return false;
2005 }
2006
2007 /*
2008 * start comes from the offset of locked_page. We have to lock
2009 * pages in order, so we can't process delalloc bytes before
2010 * locked_page
2011 */
2012 if (delalloc_start < *start)
2013 delalloc_start = *start;
2014
2015 /*
2016 * make sure to limit the number of pages we try to lock down
2017 */
2018 if (delalloc_end + 1 - delalloc_start > max_bytes)
2019 delalloc_end = delalloc_start + max_bytes - 1;
2020
2021 /* step two, lock all the pages after the page that has start */
2022 ret = lock_delalloc_pages(inode, locked_page,
2023 delalloc_start, delalloc_end);
2024 ASSERT(!ret || ret == -EAGAIN);
2025 if (ret == -EAGAIN) {
2026 /* some of the pages are gone, lets avoid looping by
2027 * shortening the size of the delalloc range we're searching
2028 */
2029 free_extent_state(cached_state);
2030 cached_state = NULL;
2031 if (!loops) {
2032 max_bytes = PAGE_SIZE;
2033 loops = 1;
2034 goto again;
2035 } else {
2036 found = false;
2037 goto out_failed;
2038 }
2039 }
2040
2041 /* step three, lock the state bits for the whole range */
2042 lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
2043
2044 /* then test to make sure it is all still delalloc */
2045 ret = test_range_bit(tree, delalloc_start, delalloc_end,
2046 EXTENT_DELALLOC, 1, cached_state);
2047 if (!ret) {
2048 unlock_extent_cached(tree, delalloc_start, delalloc_end,
2049 &cached_state);
2050 __unlock_for_delalloc(inode, locked_page,
2051 delalloc_start, delalloc_end);
2052 cond_resched();
2053 goto again;
2054 }
2055 free_extent_state(cached_state);
2056 *start = delalloc_start;
2057 *end = delalloc_end;
2058out_failed:
2059 return found;
2060}
2061
2062void extent_clear_unlock_delalloc(struct btrfs_inode *inode, u64 start, u64 end,
2063 struct page *locked_page,
2064 u32 clear_bits, unsigned long page_ops)
2065{
2066 clear_extent_bit(&inode->io_tree, start, end, clear_bits, 1, 0, NULL);
2067
2068 __process_pages_contig(inode->vfs_inode.i_mapping, locked_page,
2069 start, end, page_ops, NULL);
2070}
2071
2072/*
2073 * count the number of bytes in the tree that have a given bit(s)
2074 * set. This can be fairly slow, except for EXTENT_DIRTY which is
2075 * cached. The total number found is returned.
2076 */
2077u64 count_range_bits(struct extent_io_tree *tree,
2078 u64 *start, u64 search_end, u64 max_bytes,
2079 u32 bits, int contig)
2080{
2081 struct rb_node *node;
2082 struct extent_state *state;
2083 u64 cur_start = *start;
2084 u64 total_bytes = 0;
2085 u64 last = 0;
2086 int found = 0;
2087
2088 if (WARN_ON(search_end <= cur_start))
2089 return 0;
2090
2091 spin_lock(&tree->lock);
2092 if (cur_start == 0 && bits == EXTENT_DIRTY) {
2093 total_bytes = tree->dirty_bytes;
2094 goto out;
2095 }
2096 /*
2097 * this search will find all the extents that end after
2098 * our range starts.
2099 */
2100 node = tree_search(tree, cur_start);
2101 if (!node)
2102 goto out;
2103
2104 while (1) {
2105 state = rb_entry(node, struct extent_state, rb_node);
2106 if (state->start > search_end)
2107 break;
2108 if (contig && found && state->start > last + 1)
2109 break;
2110 if (state->end >= cur_start && (state->state & bits) == bits) {
2111 total_bytes += min(search_end, state->end) + 1 -
2112 max(cur_start, state->start);
2113 if (total_bytes >= max_bytes)
2114 break;
2115 if (!found) {
2116 *start = max(cur_start, state->start);
2117 found = 1;
2118 }
2119 last = state->end;
2120 } else if (contig && found) {
2121 break;
2122 }
2123 node = rb_next(node);
2124 if (!node)
2125 break;
2126 }
2127out:
2128 spin_unlock(&tree->lock);
2129 return total_bytes;
2130}
2131
2132/*
2133 * set the private field for a given byte offset in the tree. If there isn't
2134 * an extent_state there already, this does nothing.
2135 */
2136int set_state_failrec(struct extent_io_tree *tree, u64 start,
2137 struct io_failure_record *failrec)
2138{
2139 struct rb_node *node;
2140 struct extent_state *state;
2141 int ret = 0;
2142
2143 spin_lock(&tree->lock);
2144 /*
2145 * this search will find all the extents that end after
2146 * our range starts.
2147 */
2148 node = tree_search(tree, start);
2149 if (!node) {
2150 ret = -ENOENT;
2151 goto out;
2152 }
2153 state = rb_entry(node, struct extent_state, rb_node);
2154 if (state->start != start) {
2155 ret = -ENOENT;
2156 goto out;
2157 }
2158 state->failrec = failrec;
2159out:
2160 spin_unlock(&tree->lock);
2161 return ret;
2162}
2163
2164struct io_failure_record *get_state_failrec(struct extent_io_tree *tree, u64 start)
2165{
2166 struct rb_node *node;
2167 struct extent_state *state;
2168 struct io_failure_record *failrec;
2169
2170 spin_lock(&tree->lock);
2171 /*
2172 * this search will find all the extents that end after
2173 * our range starts.
2174 */
2175 node = tree_search(tree, start);
2176 if (!node) {
2177 failrec = ERR_PTR(-ENOENT);
2178 goto out;
2179 }
2180 state = rb_entry(node, struct extent_state, rb_node);
2181 if (state->start != start) {
2182 failrec = ERR_PTR(-ENOENT);
2183 goto out;
2184 }
2185
2186 failrec = state->failrec;
2187out:
2188 spin_unlock(&tree->lock);
2189 return failrec;
2190}
2191
2192/*
2193 * searches a range in the state tree for a given mask.
2194 * If 'filled' == 1, this returns 1 only if every extent in the tree
2195 * has the bits set. Otherwise, 1 is returned if any bit in the
2196 * range is found set.
2197 */
2198int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
2199 u32 bits, int filled, struct extent_state *cached)
2200{
2201 struct extent_state *state = NULL;
2202 struct rb_node *node;
2203 int bitset = 0;
2204
2205 spin_lock(&tree->lock);
2206 if (cached && extent_state_in_tree(cached) && cached->start <= start &&
2207 cached->end > start)
2208 node = &cached->rb_node;
2209 else
2210 node = tree_search(tree, start);
2211 while (node && start <= end) {
2212 state = rb_entry(node, struct extent_state, rb_node);
2213
2214 if (filled && state->start > start) {
2215 bitset = 0;
2216 break;
2217 }
2218
2219 if (state->start > end)
2220 break;
2221
2222 if (state->state & bits) {
2223 bitset = 1;
2224 if (!filled)
2225 break;
2226 } else if (filled) {
2227 bitset = 0;
2228 break;
2229 }
2230
2231 if (state->end == (u64)-1)
2232 break;
2233
2234 start = state->end + 1;
2235 if (start > end)
2236 break;
2237 node = rb_next(node);
2238 if (!node) {
2239 if (filled)
2240 bitset = 0;
2241 break;
2242 }
2243 }
2244 spin_unlock(&tree->lock);
2245 return bitset;
2246}
2247
2248/*
2249 * helper function to set a given page up to date if all the
2250 * extents in the tree for that page are up to date
2251 */
2252static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
2253{
2254 u64 start = page_offset(page);
2255 u64 end = start + PAGE_SIZE - 1;
2256 if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
2257 SetPageUptodate(page);
2258}
2259
2260int free_io_failure(struct extent_io_tree *failure_tree,
2261 struct extent_io_tree *io_tree,
2262 struct io_failure_record *rec)
2263{
2264 int ret;
2265 int err = 0;
2266
2267 set_state_failrec(failure_tree, rec->start, NULL);
2268 ret = clear_extent_bits(failure_tree, rec->start,
2269 rec->start + rec->len - 1,
2270 EXTENT_LOCKED | EXTENT_DIRTY);
2271 if (ret)
2272 err = ret;
2273
2274 ret = clear_extent_bits(io_tree, rec->start,
2275 rec->start + rec->len - 1,
2276 EXTENT_DAMAGED);
2277 if (ret && !err)
2278 err = ret;
2279
2280 kfree(rec);
2281 return err;
2282}
2283
2284/*
2285 * this bypasses the standard btrfs submit functions deliberately, as
2286 * the standard behavior is to write all copies in a raid setup. here we only
2287 * want to write the one bad copy. so we do the mapping for ourselves and issue
2288 * submit_bio directly.
2289 * to avoid any synchronization issues, wait for the data after writing, which
2290 * actually prevents the read that triggered the error from finishing.
2291 * currently, there can be no more than two copies of every data bit. thus,
2292 * exactly one rewrite is required.
2293 */
2294int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
2295 u64 length, u64 logical, struct page *page,
2296 unsigned int pg_offset, int mirror_num)
2297{
2298 struct bio *bio;
2299 struct btrfs_device *dev;
2300 u64 map_length = 0;
2301 u64 sector;
2302 struct btrfs_bio *bbio = NULL;
2303 int ret;
2304
2305 ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
2306 BUG_ON(!mirror_num);
2307
2308 if (btrfs_is_zoned(fs_info))
2309 return btrfs_repair_one_zone(fs_info, logical);
2310
2311 bio = btrfs_io_bio_alloc(1);
2312 bio->bi_iter.bi_size = 0;
2313 map_length = length;
2314
2315 /*
2316 * Avoid races with device replace and make sure our bbio has devices
2317 * associated to its stripes that don't go away while we are doing the
2318 * read repair operation.
2319 */
2320 btrfs_bio_counter_inc_blocked(fs_info);
2321 if (btrfs_is_parity_mirror(fs_info, logical, length)) {
2322 /*
2323 * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2324 * to update all raid stripes, but here we just want to correct
2325 * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2326 * stripe's dev and sector.
2327 */
2328 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2329 &map_length, &bbio, 0);
2330 if (ret) {
2331 btrfs_bio_counter_dec(fs_info);
2332 bio_put(bio);
2333 return -EIO;
2334 }
2335 ASSERT(bbio->mirror_num == 1);
2336 } else {
2337 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2338 &map_length, &bbio, mirror_num);
2339 if (ret) {
2340 btrfs_bio_counter_dec(fs_info);
2341 bio_put(bio);
2342 return -EIO;
2343 }
2344 BUG_ON(mirror_num != bbio->mirror_num);
2345 }
2346
2347 sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2348 bio->bi_iter.bi_sector = sector;
2349 dev = bbio->stripes[bbio->mirror_num - 1].dev;
2350 btrfs_put_bbio(bbio);
2351 if (!dev || !dev->bdev ||
2352 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2353 btrfs_bio_counter_dec(fs_info);
2354 bio_put(bio);
2355 return -EIO;
2356 }
2357 bio_set_dev(bio, dev->bdev);
2358 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2359 bio_add_page(bio, page, length, pg_offset);
2360
2361 if (btrfsic_submit_bio_wait(bio)) {
2362 /* try to remap that extent elsewhere? */
2363 btrfs_bio_counter_dec(fs_info);
2364 bio_put(bio);
2365 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2366 return -EIO;
2367 }
2368
2369 btrfs_info_rl_in_rcu(fs_info,
2370 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2371 ino, start,
2372 rcu_str_deref(dev->name), sector);
2373 btrfs_bio_counter_dec(fs_info);
2374 bio_put(bio);
2375 return 0;
2376}
2377
2378int btrfs_repair_eb_io_failure(const struct extent_buffer *eb, int mirror_num)
2379{
2380 struct btrfs_fs_info *fs_info = eb->fs_info;
2381 u64 start = eb->start;
2382 int i, num_pages = num_extent_pages(eb);
2383 int ret = 0;
2384
2385 if (sb_rdonly(fs_info->sb))
2386 return -EROFS;
2387
2388 for (i = 0; i < num_pages; i++) {
2389 struct page *p = eb->pages[i];
2390
2391 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2392 start - page_offset(p), mirror_num);
2393 if (ret)
2394 break;
2395 start += PAGE_SIZE;
2396 }
2397
2398 return ret;
2399}
2400
2401/*
2402 * each time an IO finishes, we do a fast check in the IO failure tree
2403 * to see if we need to process or clean up an io_failure_record
2404 */
2405int clean_io_failure(struct btrfs_fs_info *fs_info,
2406 struct extent_io_tree *failure_tree,
2407 struct extent_io_tree *io_tree, u64 start,
2408 struct page *page, u64 ino, unsigned int pg_offset)
2409{
2410 u64 private;
2411 struct io_failure_record *failrec;
2412 struct extent_state *state;
2413 int num_copies;
2414 int ret;
2415
2416 private = 0;
2417 ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2418 EXTENT_DIRTY, 0);
2419 if (!ret)
2420 return 0;
2421
2422 failrec = get_state_failrec(failure_tree, start);
2423 if (IS_ERR(failrec))
2424 return 0;
2425
2426 BUG_ON(!failrec->this_mirror);
2427
2428 if (sb_rdonly(fs_info->sb))
2429 goto out;
2430
2431 spin_lock(&io_tree->lock);
2432 state = find_first_extent_bit_state(io_tree,
2433 failrec->start,
2434 EXTENT_LOCKED);
2435 spin_unlock(&io_tree->lock);
2436
2437 if (state && state->start <= failrec->start &&
2438 state->end >= failrec->start + failrec->len - 1) {
2439 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2440 failrec->len);
2441 if (num_copies > 1) {
2442 repair_io_failure(fs_info, ino, start, failrec->len,
2443 failrec->logical, page, pg_offset,
2444 failrec->failed_mirror);
2445 }
2446 }
2447
2448out:
2449 free_io_failure(failure_tree, io_tree, failrec);
2450
2451 return 0;
2452}
2453
2454/*
2455 * Can be called when
2456 * - hold extent lock
2457 * - under ordered extent
2458 * - the inode is freeing
2459 */
2460void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2461{
2462 struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2463 struct io_failure_record *failrec;
2464 struct extent_state *state, *next;
2465
2466 if (RB_EMPTY_ROOT(&failure_tree->state))
2467 return;
2468
2469 spin_lock(&failure_tree->lock);
2470 state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2471 while (state) {
2472 if (state->start > end)
2473 break;
2474
2475 ASSERT(state->end <= end);
2476
2477 next = next_state(state);
2478
2479 failrec = state->failrec;
2480 free_extent_state(state);
2481 kfree(failrec);
2482
2483 state = next;
2484 }
2485 spin_unlock(&failure_tree->lock);
2486}
2487
2488static struct io_failure_record *btrfs_get_io_failure_record(struct inode *inode,
2489 u64 start)
2490{
2491 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2492 struct io_failure_record *failrec;
2493 struct extent_map *em;
2494 struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2495 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2496 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2497 const u32 sectorsize = fs_info->sectorsize;
2498 int ret;
2499 u64 logical;
2500
2501 failrec = get_state_failrec(failure_tree, start);
2502 if (!IS_ERR(failrec)) {
2503 btrfs_debug(fs_info,
2504 "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu",
2505 failrec->logical, failrec->start, failrec->len);
2506 /*
2507 * when data can be on disk more than twice, add to failrec here
2508 * (e.g. with a list for failed_mirror) to make
2509 * clean_io_failure() clean all those errors at once.
2510 */
2511
2512 return failrec;
2513 }
2514
2515 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2516 if (!failrec)
2517 return ERR_PTR(-ENOMEM);
2518
2519 failrec->start = start;
2520 failrec->len = sectorsize;
2521 failrec->this_mirror = 0;
2522 failrec->bio_flags = 0;
2523
2524 read_lock(&em_tree->lock);
2525 em = lookup_extent_mapping(em_tree, start, failrec->len);
2526 if (!em) {
2527 read_unlock(&em_tree->lock);
2528 kfree(failrec);
2529 return ERR_PTR(-EIO);
2530 }
2531
2532 if (em->start > start || em->start + em->len <= start) {
2533 free_extent_map(em);
2534 em = NULL;
2535 }
2536 read_unlock(&em_tree->lock);
2537 if (!em) {
2538 kfree(failrec);
2539 return ERR_PTR(-EIO);
2540 }
2541
2542 logical = start - em->start;
2543 logical = em->block_start + logical;
2544 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2545 logical = em->block_start;
2546 failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2547 extent_set_compress_type(&failrec->bio_flags, em->compress_type);
2548 }
2549
2550 btrfs_debug(fs_info,
2551 "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2552 logical, start, failrec->len);
2553
2554 failrec->logical = logical;
2555 free_extent_map(em);
2556
2557 /* Set the bits in the private failure tree */
2558 ret = set_extent_bits(failure_tree, start, start + sectorsize - 1,
2559 EXTENT_LOCKED | EXTENT_DIRTY);
2560 if (ret >= 0) {
2561 ret = set_state_failrec(failure_tree, start, failrec);
2562 /* Set the bits in the inode's tree */
2563 ret = set_extent_bits(tree, start, start + sectorsize - 1,
2564 EXTENT_DAMAGED);
2565 } else if (ret < 0) {
2566 kfree(failrec);
2567 return ERR_PTR(ret);
2568 }
2569
2570 return failrec;
2571}
2572
2573static bool btrfs_check_repairable(struct inode *inode,
2574 struct io_failure_record *failrec,
2575 int failed_mirror)
2576{
2577 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2578 int num_copies;
2579
2580 num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2581 if (num_copies == 1) {
2582 /*
2583 * we only have a single copy of the data, so don't bother with
2584 * all the retry and error correction code that follows. no
2585 * matter what the error is, it is very likely to persist.
2586 */
2587 btrfs_debug(fs_info,
2588 "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2589 num_copies, failrec->this_mirror, failed_mirror);
2590 return false;
2591 }
2592
2593 /* The failure record should only contain one sector */
2594 ASSERT(failrec->len == fs_info->sectorsize);
2595
2596 /*
2597 * There are two premises:
2598 * a) deliver good data to the caller
2599 * b) correct the bad sectors on disk
2600 *
2601 * Since we're only doing repair for one sector, we only need to get
2602 * a good copy of the failed sector and if we succeed, we have setup
2603 * everything for repair_io_failure to do the rest for us.
2604 */
2605 failrec->failed_mirror = failed_mirror;
2606 failrec->this_mirror++;
2607 if (failrec->this_mirror == failed_mirror)
2608 failrec->this_mirror++;
2609
2610 if (failrec->this_mirror > num_copies) {
2611 btrfs_debug(fs_info,
2612 "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2613 num_copies, failrec->this_mirror, failed_mirror);
2614 return false;
2615 }
2616
2617 return true;
2618}
2619
2620int btrfs_repair_one_sector(struct inode *inode,
2621 struct bio *failed_bio, u32 bio_offset,
2622 struct page *page, unsigned int pgoff,
2623 u64 start, int failed_mirror,
2624 submit_bio_hook_t *submit_bio_hook)
2625{
2626 struct io_failure_record *failrec;
2627 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2628 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2629 struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2630 struct btrfs_io_bio *failed_io_bio = btrfs_io_bio(failed_bio);
2631 const int icsum = bio_offset >> fs_info->sectorsize_bits;
2632 struct bio *repair_bio;
2633 struct btrfs_io_bio *repair_io_bio;
2634 blk_status_t status;
2635
2636 btrfs_debug(fs_info,
2637 "repair read error: read error at %llu", start);
2638
2639 BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2640
2641 failrec = btrfs_get_io_failure_record(inode, start);
2642 if (IS_ERR(failrec))
2643 return PTR_ERR(failrec);
2644
2645
2646 if (!btrfs_check_repairable(inode, failrec, failed_mirror)) {
2647 free_io_failure(failure_tree, tree, failrec);
2648 return -EIO;
2649 }
2650
2651 repair_bio = btrfs_io_bio_alloc(1);
2652 repair_io_bio = btrfs_io_bio(repair_bio);
2653 repair_bio->bi_opf = REQ_OP_READ;
2654 repair_bio->bi_end_io = failed_bio->bi_end_io;
2655 repair_bio->bi_iter.bi_sector = failrec->logical >> 9;
2656 repair_bio->bi_private = failed_bio->bi_private;
2657
2658 if (failed_io_bio->csum) {
2659 const u32 csum_size = fs_info->csum_size;
2660
2661 repair_io_bio->csum = repair_io_bio->csum_inline;
2662 memcpy(repair_io_bio->csum,
2663 failed_io_bio->csum + csum_size * icsum, csum_size);
2664 }
2665
2666 bio_add_page(repair_bio, page, failrec->len, pgoff);
2667 repair_io_bio->logical = failrec->start;
2668 repair_io_bio->iter = repair_bio->bi_iter;
2669
2670 btrfs_debug(btrfs_sb(inode->i_sb),
2671 "repair read error: submitting new read to mirror %d",
2672 failrec->this_mirror);
2673
2674 status = submit_bio_hook(inode, repair_bio, failrec->this_mirror,
2675 failrec->bio_flags);
2676 if (status) {
2677 free_io_failure(failure_tree, tree, failrec);
2678 bio_put(repair_bio);
2679 }
2680 return blk_status_to_errno(status);
2681}
2682
2683static void end_page_read(struct page *page, bool uptodate, u64 start, u32 len)
2684{
2685 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
2686
2687 ASSERT(page_offset(page) <= start &&
2688 start + len <= page_offset(page) + PAGE_SIZE);
2689
2690 if (uptodate) {
2691 btrfs_page_set_uptodate(fs_info, page, start, len);
2692 } else {
2693 btrfs_page_clear_uptodate(fs_info, page, start, len);
2694 btrfs_page_set_error(fs_info, page, start, len);
2695 }
2696
2697 if (fs_info->sectorsize == PAGE_SIZE)
2698 unlock_page(page);
2699 else
2700 btrfs_subpage_end_reader(fs_info, page, start, len);
2701}
2702
2703static blk_status_t submit_read_repair(struct inode *inode,
2704 struct bio *failed_bio, u32 bio_offset,
2705 struct page *page, unsigned int pgoff,
2706 u64 start, u64 end, int failed_mirror,
2707 unsigned int error_bitmap,
2708 submit_bio_hook_t *submit_bio_hook)
2709{
2710 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2711 const u32 sectorsize = fs_info->sectorsize;
2712 const int nr_bits = (end + 1 - start) >> fs_info->sectorsize_bits;
2713 int error = 0;
2714 int i;
2715
2716 BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2717
2718 /* We're here because we had some read errors or csum mismatch */
2719 ASSERT(error_bitmap);
2720
2721 /*
2722 * We only get called on buffered IO, thus page must be mapped and bio
2723 * must not be cloned.
2724 */
2725 ASSERT(page->mapping && !bio_flagged(failed_bio, BIO_CLONED));
2726
2727 /* Iterate through all the sectors in the range */
2728 for (i = 0; i < nr_bits; i++) {
2729 const unsigned int offset = i * sectorsize;
2730 struct extent_state *cached = NULL;
2731 bool uptodate = false;
2732 int ret;
2733
2734 if (!(error_bitmap & (1U << i))) {
2735 /*
2736 * This sector has no error, just end the page read
2737 * and unlock the range.
2738 */
2739 uptodate = true;
2740 goto next;
2741 }
2742
2743 ret = btrfs_repair_one_sector(inode, failed_bio,
2744 bio_offset + offset,
2745 page, pgoff + offset, start + offset,
2746 failed_mirror, submit_bio_hook);
2747 if (!ret) {
2748 /*
2749 * We have submitted the read repair, the page release
2750 * will be handled by the endio function of the
2751 * submitted repair bio.
2752 * Thus we don't need to do any thing here.
2753 */
2754 continue;
2755 }
2756 /*
2757 * Repair failed, just record the error but still continue.
2758 * Or the remaining sectors will not be properly unlocked.
2759 */
2760 if (!error)
2761 error = ret;
2762next:
2763 end_page_read(page, uptodate, start + offset, sectorsize);
2764 if (uptodate)
2765 set_extent_uptodate(&BTRFS_I(inode)->io_tree,
2766 start + offset,
2767 start + offset + sectorsize - 1,
2768 &cached, GFP_ATOMIC);
2769 unlock_extent_cached_atomic(&BTRFS_I(inode)->io_tree,
2770 start + offset,
2771 start + offset + sectorsize - 1,
2772 &cached);
2773 }
2774 return errno_to_blk_status(error);
2775}
2776
2777/* lots and lots of room for performance fixes in the end_bio funcs */
2778
2779void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2780{
2781 struct btrfs_inode *inode;
2782 int uptodate = (err == 0);
2783 int ret = 0;
2784
2785 ASSERT(page && page->mapping);
2786 inode = BTRFS_I(page->mapping->host);
2787 btrfs_writepage_endio_finish_ordered(inode, page, start, end, uptodate);
2788
2789 if (!uptodate) {
2790 ClearPageUptodate(page);
2791 SetPageError(page);
2792 ret = err < 0 ? err : -EIO;
2793 mapping_set_error(page->mapping, ret);
2794 }
2795}
2796
2797/*
2798 * after a writepage IO is done, we need to:
2799 * clear the uptodate bits on error
2800 * clear the writeback bits in the extent tree for this IO
2801 * end_page_writeback if the page has no more pending IO
2802 *
2803 * Scheduling is not allowed, so the extent state tree is expected
2804 * to have one and only one object corresponding to this IO.
2805 */
2806static void end_bio_extent_writepage(struct bio *bio)
2807{
2808 int error = blk_status_to_errno(bio->bi_status);
2809 struct bio_vec *bvec;
2810 u64 start;
2811 u64 end;
2812 struct bvec_iter_all iter_all;
2813 bool first_bvec = true;
2814
2815 ASSERT(!bio_flagged(bio, BIO_CLONED));
2816 bio_for_each_segment_all(bvec, bio, iter_all) {
2817 struct page *page = bvec->bv_page;
2818 struct inode *inode = page->mapping->host;
2819 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2820 const u32 sectorsize = fs_info->sectorsize;
2821
2822 /* Our read/write should always be sector aligned. */
2823 if (!IS_ALIGNED(bvec->bv_offset, sectorsize))
2824 btrfs_err(fs_info,
2825 "partial page write in btrfs with offset %u and length %u",
2826 bvec->bv_offset, bvec->bv_len);
2827 else if (!IS_ALIGNED(bvec->bv_len, sectorsize))
2828 btrfs_info(fs_info,
2829 "incomplete page write with offset %u and length %u",
2830 bvec->bv_offset, bvec->bv_len);
2831
2832 start = page_offset(page) + bvec->bv_offset;
2833 end = start + bvec->bv_len - 1;
2834
2835 if (first_bvec) {
2836 btrfs_record_physical_zoned(inode, start, bio);
2837 first_bvec = false;
2838 }
2839
2840 end_extent_writepage(page, error, start, end);
2841
2842 btrfs_page_clear_writeback(fs_info, page, start, bvec->bv_len);
2843 }
2844
2845 bio_put(bio);
2846}
2847
2848/*
2849 * Record previously processed extent range
2850 *
2851 * For endio_readpage_release_extent() to handle a full extent range, reducing
2852 * the extent io operations.
2853 */
2854struct processed_extent {
2855 struct btrfs_inode *inode;
2856 /* Start of the range in @inode */
2857 u64 start;
2858 /* End of the range in @inode */
2859 u64 end;
2860 bool uptodate;
2861};
2862
2863/*
2864 * Try to release processed extent range
2865 *
2866 * May not release the extent range right now if the current range is
2867 * contiguous to processed extent.
2868 *
2869 * Will release processed extent when any of @inode, @uptodate, the range is
2870 * no longer contiguous to the processed range.
2871 *
2872 * Passing @inode == NULL will force processed extent to be released.
2873 */
2874static void endio_readpage_release_extent(struct processed_extent *processed,
2875 struct btrfs_inode *inode, u64 start, u64 end,
2876 bool uptodate)
2877{
2878 struct extent_state *cached = NULL;
2879 struct extent_io_tree *tree;
2880
2881 /* The first extent, initialize @processed */
2882 if (!processed->inode)
2883 goto update;
2884
2885 /*
2886 * Contiguous to processed extent, just uptodate the end.
2887 *
2888 * Several things to notice:
2889 *
2890 * - bio can be merged as long as on-disk bytenr is contiguous
2891 * This means we can have page belonging to other inodes, thus need to
2892 * check if the inode still matches.
2893 * - bvec can contain range beyond current page for multi-page bvec
2894 * Thus we need to do processed->end + 1 >= start check
2895 */
2896 if (processed->inode == inode && processed->uptodate == uptodate &&
2897 processed->end + 1 >= start && end >= processed->end) {
2898 processed->end = end;
2899 return;
2900 }
2901
2902 tree = &processed->inode->io_tree;
2903 /*
2904 * Now we don't have range contiguous to the processed range, release
2905 * the processed range now.
2906 */
2907 if (processed->uptodate && tree->track_uptodate)
2908 set_extent_uptodate(tree, processed->start, processed->end,
2909 &cached, GFP_ATOMIC);
2910 unlock_extent_cached_atomic(tree, processed->start, processed->end,
2911 &cached);
2912
2913update:
2914 /* Update processed to current range */
2915 processed->inode = inode;
2916 processed->start = start;
2917 processed->end = end;
2918 processed->uptodate = uptodate;
2919}
2920
2921static void begin_page_read(struct btrfs_fs_info *fs_info, struct page *page)
2922{
2923 ASSERT(PageLocked(page));
2924 if (fs_info->sectorsize == PAGE_SIZE)
2925 return;
2926
2927 ASSERT(PagePrivate(page));
2928 btrfs_subpage_start_reader(fs_info, page, page_offset(page), PAGE_SIZE);
2929}
2930
2931/*
2932 * Find extent buffer for a givne bytenr.
2933 *
2934 * This is for end_bio_extent_readpage(), thus we can't do any unsafe locking
2935 * in endio context.
2936 */
2937static struct extent_buffer *find_extent_buffer_readpage(
2938 struct btrfs_fs_info *fs_info, struct page *page, u64 bytenr)
2939{
2940 struct extent_buffer *eb;
2941
2942 /*
2943 * For regular sectorsize, we can use page->private to grab extent
2944 * buffer
2945 */
2946 if (fs_info->sectorsize == PAGE_SIZE) {
2947 ASSERT(PagePrivate(page) && page->private);
2948 return (struct extent_buffer *)page->private;
2949 }
2950
2951 /* For subpage case, we need to lookup buffer radix tree */
2952 rcu_read_lock();
2953 eb = radix_tree_lookup(&fs_info->buffer_radix,
2954 bytenr >> fs_info->sectorsize_bits);
2955 rcu_read_unlock();
2956 ASSERT(eb);
2957 return eb;
2958}
2959
2960/*
2961 * after a readpage IO is done, we need to:
2962 * clear the uptodate bits on error
2963 * set the uptodate bits if things worked
2964 * set the page up to date if all extents in the tree are uptodate
2965 * clear the lock bit in the extent tree
2966 * unlock the page if there are no other extents locked for it
2967 *
2968 * Scheduling is not allowed, so the extent state tree is expected
2969 * to have one and only one object corresponding to this IO.
2970 */
2971static void end_bio_extent_readpage(struct bio *bio)
2972{
2973 struct bio_vec *bvec;
2974 struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2975 struct extent_io_tree *tree, *failure_tree;
2976 struct processed_extent processed = { 0 };
2977 /*
2978 * The offset to the beginning of a bio, since one bio can never be
2979 * larger than UINT_MAX, u32 here is enough.
2980 */
2981 u32 bio_offset = 0;
2982 int mirror;
2983 int ret;
2984 struct bvec_iter_all iter_all;
2985
2986 ASSERT(!bio_flagged(bio, BIO_CLONED));
2987 bio_for_each_segment_all(bvec, bio, iter_all) {
2988 bool uptodate = !bio->bi_status;
2989 struct page *page = bvec->bv_page;
2990 struct inode *inode = page->mapping->host;
2991 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2992 const u32 sectorsize = fs_info->sectorsize;
2993 unsigned int error_bitmap = (unsigned int)-1;
2994 u64 start;
2995 u64 end;
2996 u32 len;
2997
2998 btrfs_debug(fs_info,
2999 "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
3000 bio->bi_iter.bi_sector, bio->bi_status,
3001 io_bio->mirror_num);
3002 tree = &BTRFS_I(inode)->io_tree;
3003 failure_tree = &BTRFS_I(inode)->io_failure_tree;
3004
3005 /*
3006 * We always issue full-sector reads, but if some block in a
3007 * page fails to read, blk_update_request() will advance
3008 * bv_offset and adjust bv_len to compensate. Print a warning
3009 * for unaligned offsets, and an error if they don't add up to
3010 * a full sector.
3011 */
3012 if (!IS_ALIGNED(bvec->bv_offset, sectorsize))
3013 btrfs_err(fs_info,
3014 "partial page read in btrfs with offset %u and length %u",
3015 bvec->bv_offset, bvec->bv_len);
3016 else if (!IS_ALIGNED(bvec->bv_offset + bvec->bv_len,
3017 sectorsize))
3018 btrfs_info(fs_info,
3019 "incomplete page read with offset %u and length %u",
3020 bvec->bv_offset, bvec->bv_len);
3021
3022 start = page_offset(page) + bvec->bv_offset;
3023 end = start + bvec->bv_len - 1;
3024 len = bvec->bv_len;
3025
3026 mirror = io_bio->mirror_num;
3027 if (likely(uptodate)) {
3028 if (is_data_inode(inode)) {
3029 error_bitmap = btrfs_verify_data_csum(io_bio,
3030 bio_offset, page, start, end);
3031 ret = error_bitmap;
3032 } else {
3033 ret = btrfs_validate_metadata_buffer(io_bio,
3034 page, start, end, mirror);
3035 }
3036 if (ret)
3037 uptodate = false;
3038 else
3039 clean_io_failure(BTRFS_I(inode)->root->fs_info,
3040 failure_tree, tree, start,
3041 page,
3042 btrfs_ino(BTRFS_I(inode)), 0);
3043 }
3044
3045 if (likely(uptodate))
3046 goto readpage_ok;
3047
3048 if (is_data_inode(inode)) {
3049 /*
3050 * btrfs_submit_read_repair() will handle all the good
3051 * and bad sectors, we just continue to the next bvec.
3052 */
3053 submit_read_repair(inode, bio, bio_offset, page,
3054 start - page_offset(page), start,
3055 end, mirror, error_bitmap,
3056 btrfs_submit_data_bio);
3057
3058 ASSERT(bio_offset + len > bio_offset);
3059 bio_offset += len;
3060 continue;
3061 } else {
3062 struct extent_buffer *eb;
3063
3064 eb = find_extent_buffer_readpage(fs_info, page, start);
3065 set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
3066 eb->read_mirror = mirror;
3067 atomic_dec(&eb->io_pages);
3068 if (test_and_clear_bit(EXTENT_BUFFER_READAHEAD,
3069 &eb->bflags))
3070 btree_readahead_hook(eb, -EIO);
3071 }
3072readpage_ok:
3073 if (likely(uptodate)) {
3074 loff_t i_size = i_size_read(inode);
3075 pgoff_t end_index = i_size >> PAGE_SHIFT;
3076
3077 /*
3078 * Zero out the remaining part if this range straddles
3079 * i_size.
3080 *
3081 * Here we should only zero the range inside the bvec,
3082 * not touch anything else.
3083 *
3084 * NOTE: i_size is exclusive while end is inclusive.
3085 */
3086 if (page->index == end_index && i_size <= end) {
3087 u32 zero_start = max(offset_in_page(i_size),
3088 offset_in_page(start));
3089
3090 zero_user_segment(page, zero_start,
3091 offset_in_page(end) + 1);
3092 }
3093 }
3094 ASSERT(bio_offset + len > bio_offset);
3095 bio_offset += len;
3096
3097 /* Update page status and unlock */
3098 end_page_read(page, uptodate, start, len);
3099 endio_readpage_release_extent(&processed, BTRFS_I(inode),
3100 start, end, uptodate);
3101 }
3102 /* Release the last extent */
3103 endio_readpage_release_extent(&processed, NULL, 0, 0, false);
3104 btrfs_io_bio_free_csum(io_bio);
3105 bio_put(bio);
3106}
3107
3108/*
3109 * Initialize the members up to but not including 'bio'. Use after allocating a
3110 * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
3111 * 'bio' because use of __GFP_ZERO is not supported.
3112 */
3113static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
3114{
3115 memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
3116}
3117
3118/*
3119 * The following helpers allocate a bio. As it's backed by a bioset, it'll
3120 * never fail. We're returning a bio right now but you can call btrfs_io_bio
3121 * for the appropriate container_of magic
3122 */
3123struct bio *btrfs_bio_alloc(u64 first_byte)
3124{
3125 struct bio *bio;
3126
3127 bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_VECS, &btrfs_bioset);
3128 bio->bi_iter.bi_sector = first_byte >> 9;
3129 btrfs_io_bio_init(btrfs_io_bio(bio));
3130 return bio;
3131}
3132
3133struct bio *btrfs_bio_clone(struct bio *bio)
3134{
3135 struct btrfs_io_bio *btrfs_bio;
3136 struct bio *new;
3137
3138 /* Bio allocation backed by a bioset does not fail */
3139 new = bio_clone_fast(bio, GFP_NOFS, &btrfs_bioset);
3140 btrfs_bio = btrfs_io_bio(new);
3141 btrfs_io_bio_init(btrfs_bio);
3142 btrfs_bio->iter = bio->bi_iter;
3143 return new;
3144}
3145
3146struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
3147{
3148 struct bio *bio;
3149
3150 /* Bio allocation backed by a bioset does not fail */
3151 bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, &btrfs_bioset);
3152 btrfs_io_bio_init(btrfs_io_bio(bio));
3153 return bio;
3154}
3155
3156struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
3157{
3158 struct bio *bio;
3159 struct btrfs_io_bio *btrfs_bio;
3160
3161 /* this will never fail when it's backed by a bioset */
3162 bio = bio_clone_fast(orig, GFP_NOFS, &btrfs_bioset);
3163 ASSERT(bio);
3164
3165 btrfs_bio = btrfs_io_bio(bio);
3166 btrfs_io_bio_init(btrfs_bio);
3167
3168 bio_trim(bio, offset >> 9, size >> 9);
3169 btrfs_bio->iter = bio->bi_iter;
3170 return bio;
3171}
3172
3173/**
3174 * Attempt to add a page to bio
3175 *
3176 * @bio: destination bio
3177 * @page: page to add to the bio
3178 * @disk_bytenr: offset of the new bio or to check whether we are adding
3179 * a contiguous page to the previous one
3180 * @pg_offset: starting offset in the page
3181 * @size: portion of page that we want to write
3182 * @prev_bio_flags: flags of previous bio to see if we can merge the current one
3183 * @bio_flags: flags of the current bio to see if we can merge them
3184 * @return: true if page was added, false otherwise
3185 *
3186 * Attempt to add a page to bio considering stripe alignment etc.
3187 *
3188 * Return true if successfully page added. Otherwise, return false.
3189 */
3190static bool btrfs_bio_add_page(struct btrfs_bio_ctrl *bio_ctrl,
3191 struct page *page,
3192 u64 disk_bytenr, unsigned int size,
3193 unsigned int pg_offset,
3194 unsigned long bio_flags)
3195{
3196 struct bio *bio = bio_ctrl->bio;
3197 u32 bio_size = bio->bi_iter.bi_size;
3198 const sector_t sector = disk_bytenr >> SECTOR_SHIFT;
3199 bool contig;
3200 int ret;
3201
3202 ASSERT(bio);
3203 /* The limit should be calculated when bio_ctrl->bio is allocated */
3204 ASSERT(bio_ctrl->len_to_oe_boundary && bio_ctrl->len_to_stripe_boundary);
3205 if (bio_ctrl->bio_flags != bio_flags)
3206 return false;
3207
3208 if (bio_ctrl->bio_flags & EXTENT_BIO_COMPRESSED)
3209 contig = bio->bi_iter.bi_sector == sector;
3210 else
3211 contig = bio_end_sector(bio) == sector;
3212 if (!contig)
3213 return false;
3214
3215 if (bio_size + size > bio_ctrl->len_to_oe_boundary ||
3216 bio_size + size > bio_ctrl->len_to_stripe_boundary)
3217 return false;
3218
3219 if (bio_op(bio) == REQ_OP_ZONE_APPEND)
3220 ret = bio_add_zone_append_page(bio, page, size, pg_offset);
3221 else
3222 ret = bio_add_page(bio, page, size, pg_offset);
3223
3224 return ret == size;
3225}
3226
3227static int calc_bio_boundaries(struct btrfs_bio_ctrl *bio_ctrl,
3228 struct btrfs_inode *inode)
3229{
3230 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3231 struct btrfs_io_geometry geom;
3232 struct btrfs_ordered_extent *ordered;
3233 struct extent_map *em;
3234 u64 logical = (bio_ctrl->bio->bi_iter.bi_sector << SECTOR_SHIFT);
3235 int ret;
3236
3237 /*
3238 * Pages for compressed extent are never submitted to disk directly,
3239 * thus it has no real boundary, just set them to U32_MAX.
3240 *
3241 * The split happens for real compressed bio, which happens in
3242 * btrfs_submit_compressed_read/write().
3243 */
3244 if (bio_ctrl->bio_flags & EXTENT_BIO_COMPRESSED) {
3245 bio_ctrl->len_to_oe_boundary = U32_MAX;
3246 bio_ctrl->len_to_stripe_boundary = U32_MAX;
3247 return 0;
3248 }
3249 em = btrfs_get_chunk_map(fs_info, logical, fs_info->sectorsize);
3250 if (IS_ERR(em))
3251 return PTR_ERR(em);
3252 ret = btrfs_get_io_geometry(fs_info, em, btrfs_op(bio_ctrl->bio),
3253 logical, &geom);
3254 free_extent_map(em);
3255 if (ret < 0) {
3256 return ret;
3257 }
3258 if (geom.len > U32_MAX)
3259 bio_ctrl->len_to_stripe_boundary = U32_MAX;
3260 else
3261 bio_ctrl->len_to_stripe_boundary = (u32)geom.len;
3262
3263 if (!btrfs_is_zoned(fs_info) ||
3264 bio_op(bio_ctrl->bio) != REQ_OP_ZONE_APPEND) {
3265 bio_ctrl->len_to_oe_boundary = U32_MAX;
3266 return 0;
3267 }
3268
3269 ASSERT(fs_info->max_zone_append_size > 0);
3270 /* Ordered extent not yet created, so we're good */
3271 ordered = btrfs_lookup_ordered_extent(inode, logical);
3272 if (!ordered) {
3273 bio_ctrl->len_to_oe_boundary = U32_MAX;
3274 return 0;
3275 }
3276
3277 bio_ctrl->len_to_oe_boundary = min_t(u32, U32_MAX,
3278 ordered->disk_bytenr + ordered->disk_num_bytes - logical);
3279 btrfs_put_ordered_extent(ordered);
3280 return 0;
3281}
3282
3283/*
3284 * @opf: bio REQ_OP_* and REQ_* flags as one value
3285 * @wbc: optional writeback control for io accounting
3286 * @page: page to add to the bio
3287 * @disk_bytenr: logical bytenr where the write will be
3288 * @size: portion of page that we want to write to
3289 * @pg_offset: offset of the new bio or to check whether we are adding
3290 * a contiguous page to the previous one
3291 * @bio_ret: must be valid pointer, newly allocated bio will be stored there
3292 * @end_io_func: end_io callback for new bio
3293 * @mirror_num: desired mirror to read/write
3294 * @prev_bio_flags: flags of previous bio to see if we can merge the current one
3295 * @bio_flags: flags of the current bio to see if we can merge them
3296 */
3297static int submit_extent_page(unsigned int opf,
3298 struct writeback_control *wbc,
3299 struct btrfs_bio_ctrl *bio_ctrl,
3300 struct page *page, u64 disk_bytenr,
3301 size_t size, unsigned long pg_offset,
3302 bio_end_io_t end_io_func,
3303 int mirror_num,
3304 unsigned long bio_flags,
3305 bool force_bio_submit)
3306{
3307 int ret = 0;
3308 struct bio *bio;
3309 size_t io_size = min_t(size_t, size, PAGE_SIZE);
3310 struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
3311 struct extent_io_tree *tree = &inode->io_tree;
3312 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3313
3314 ASSERT(bio_ctrl);
3315
3316 ASSERT(pg_offset < PAGE_SIZE && size <= PAGE_SIZE &&
3317 pg_offset + size <= PAGE_SIZE);
3318 if (bio_ctrl->bio) {
3319 bio = bio_ctrl->bio;
3320 if (force_bio_submit ||
3321 !btrfs_bio_add_page(bio_ctrl, page, disk_bytenr, io_size,
3322 pg_offset, bio_flags)) {
3323 ret = submit_one_bio(bio, mirror_num, bio_ctrl->bio_flags);
3324 bio_ctrl->bio = NULL;
3325 if (ret < 0)
3326 return ret;
3327 } else {
3328 if (wbc)
3329 wbc_account_cgroup_owner(wbc, page, io_size);
3330 return 0;
3331 }
3332 }
3333
3334 bio = btrfs_bio_alloc(disk_bytenr);
3335 bio_add_page(bio, page, io_size, pg_offset);
3336 bio->bi_end_io = end_io_func;
3337 bio->bi_private = tree;
3338 bio->bi_write_hint = page->mapping->host->i_write_hint;
3339 bio->bi_opf = opf;
3340 if (wbc) {
3341 struct block_device *bdev;
3342
3343 bdev = fs_info->fs_devices->latest_bdev;
3344 bio_set_dev(bio, bdev);
3345 wbc_init_bio(wbc, bio);
3346 wbc_account_cgroup_owner(wbc, page, io_size);
3347 }
3348 if (btrfs_is_zoned(fs_info) && bio_op(bio) == REQ_OP_ZONE_APPEND) {
3349 struct btrfs_device *device;
3350
3351 device = btrfs_zoned_get_device(fs_info, disk_bytenr, io_size);
3352 if (IS_ERR(device))
3353 return PTR_ERR(device);
3354
3355 btrfs_io_bio(bio)->device = device;
3356 }
3357
3358 bio_ctrl->bio = bio;
3359 bio_ctrl->bio_flags = bio_flags;
3360 ret = calc_bio_boundaries(bio_ctrl, inode);
3361
3362 return ret;
3363}
3364
3365static int attach_extent_buffer_page(struct extent_buffer *eb,
3366 struct page *page,
3367 struct btrfs_subpage *prealloc)
3368{
3369 struct btrfs_fs_info *fs_info = eb->fs_info;
3370 int ret = 0;
3371
3372 /*
3373 * If the page is mapped to btree inode, we should hold the private
3374 * lock to prevent race.
3375 * For cloned or dummy extent buffers, their pages are not mapped and
3376 * will not race with any other ebs.
3377 */
3378 if (page->mapping)
3379 lockdep_assert_held(&page->mapping->private_lock);
3380
3381 if (fs_info->sectorsize == PAGE_SIZE) {
3382 if (!PagePrivate(page))
3383 attach_page_private(page, eb);
3384 else
3385 WARN_ON(page->private != (unsigned long)eb);
3386 return 0;
3387 }
3388
3389 /* Already mapped, just free prealloc */
3390 if (PagePrivate(page)) {
3391 btrfs_free_subpage(prealloc);
3392 return 0;
3393 }
3394
3395 if (prealloc)
3396 /* Has preallocated memory for subpage */
3397 attach_page_private(page, prealloc);
3398 else
3399 /* Do new allocation to attach subpage */
3400 ret = btrfs_attach_subpage(fs_info, page,
3401 BTRFS_SUBPAGE_METADATA);
3402 return ret;
3403}
3404
3405int set_page_extent_mapped(struct page *page)
3406{
3407 struct btrfs_fs_info *fs_info;
3408
3409 ASSERT(page->mapping);
3410
3411 if (PagePrivate(page))
3412 return 0;
3413
3414 fs_info = btrfs_sb(page->mapping->host->i_sb);
3415
3416 if (fs_info->sectorsize < PAGE_SIZE)
3417 return btrfs_attach_subpage(fs_info, page, BTRFS_SUBPAGE_DATA);
3418
3419 attach_page_private(page, (void *)EXTENT_PAGE_PRIVATE);
3420 return 0;
3421}
3422
3423void clear_page_extent_mapped(struct page *page)
3424{
3425 struct btrfs_fs_info *fs_info;
3426
3427 ASSERT(page->mapping);
3428
3429 if (!PagePrivate(page))
3430 return;
3431
3432 fs_info = btrfs_sb(page->mapping->host->i_sb);
3433 if (fs_info->sectorsize < PAGE_SIZE)
3434 return btrfs_detach_subpage(fs_info, page);
3435
3436 detach_page_private(page);
3437}
3438
3439static struct extent_map *
3440__get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
3441 u64 start, u64 len, struct extent_map **em_cached)
3442{
3443 struct extent_map *em;
3444
3445 if (em_cached && *em_cached) {
3446 em = *em_cached;
3447 if (extent_map_in_tree(em) && start >= em->start &&
3448 start < extent_map_end(em)) {
3449 refcount_inc(&em->refs);
3450 return em;
3451 }
3452
3453 free_extent_map(em);
3454 *em_cached = NULL;
3455 }
3456
3457 em = btrfs_get_extent(BTRFS_I(inode), page, pg_offset, start, len);
3458 if (em_cached && !IS_ERR_OR_NULL(em)) {
3459 BUG_ON(*em_cached);
3460 refcount_inc(&em->refs);
3461 *em_cached = em;
3462 }
3463 return em;
3464}
3465/*
3466 * basic readpage implementation. Locked extent state structs are inserted
3467 * into the tree that are removed when the IO is done (by the end_io
3468 * handlers)
3469 * XXX JDM: This needs looking at to ensure proper page locking
3470 * return 0 on success, otherwise return error
3471 */
3472int btrfs_do_readpage(struct page *page, struct extent_map **em_cached,
3473 struct btrfs_bio_ctrl *bio_ctrl,
3474 unsigned int read_flags, u64 *prev_em_start)
3475{
3476 struct inode *inode = page->mapping->host;
3477 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3478 u64 start = page_offset(page);
3479 const u64 end = start + PAGE_SIZE - 1;
3480 u64 cur = start;
3481 u64 extent_offset;
3482 u64 last_byte = i_size_read(inode);
3483 u64 block_start;
3484 u64 cur_end;
3485 struct extent_map *em;
3486 int ret = 0;
3487 int nr = 0;
3488 size_t pg_offset = 0;
3489 size_t iosize;
3490 size_t blocksize = inode->i_sb->s_blocksize;
3491 unsigned long this_bio_flag = 0;
3492 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
3493
3494 ret = set_page_extent_mapped(page);
3495 if (ret < 0) {
3496 unlock_extent(tree, start, end);
3497 btrfs_page_set_error(fs_info, page, start, PAGE_SIZE);
3498 unlock_page(page);
3499 goto out;
3500 }
3501
3502 if (!PageUptodate(page)) {
3503 if (cleancache_get_page(page) == 0) {
3504 BUG_ON(blocksize != PAGE_SIZE);
3505 unlock_extent(tree, start, end);
3506 unlock_page(page);
3507 goto out;
3508 }
3509 }
3510
3511 if (page->index == last_byte >> PAGE_SHIFT) {
3512 size_t zero_offset = offset_in_page(last_byte);
3513
3514 if (zero_offset) {
3515 iosize = PAGE_SIZE - zero_offset;
3516 memzero_page(page, zero_offset, iosize);
3517 flush_dcache_page(page);
3518 }
3519 }
3520 begin_page_read(fs_info, page);
3521 while (cur <= end) {
3522 bool force_bio_submit = false;
3523 u64 disk_bytenr;
3524
3525 if (cur >= last_byte) {
3526 struct extent_state *cached = NULL;
3527
3528 iosize = PAGE_SIZE - pg_offset;
3529 memzero_page(page, pg_offset, iosize);
3530 flush_dcache_page(page);
3531 set_extent_uptodate(tree, cur, cur + iosize - 1,
3532 &cached, GFP_NOFS);
3533 unlock_extent_cached(tree, cur,
3534 cur + iosize - 1, &cached);
3535 end_page_read(page, true, cur, iosize);
3536 break;
3537 }
3538 em = __get_extent_map(inode, page, pg_offset, cur,
3539 end - cur + 1, em_cached);
3540 if (IS_ERR_OR_NULL(em)) {
3541 unlock_extent(tree, cur, end);
3542 end_page_read(page, false, cur, end + 1 - cur);
3543 break;
3544 }
3545 extent_offset = cur - em->start;
3546 BUG_ON(extent_map_end(em) <= cur);
3547 BUG_ON(end < cur);
3548
3549 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
3550 this_bio_flag |= EXTENT_BIO_COMPRESSED;
3551 extent_set_compress_type(&this_bio_flag,
3552 em->compress_type);
3553 }
3554
3555 iosize = min(extent_map_end(em) - cur, end - cur + 1);
3556 cur_end = min(extent_map_end(em) - 1, end);
3557 iosize = ALIGN(iosize, blocksize);
3558 if (this_bio_flag & EXTENT_BIO_COMPRESSED)
3559 disk_bytenr = em->block_start;
3560 else
3561 disk_bytenr = em->block_start + extent_offset;
3562 block_start = em->block_start;
3563 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3564 block_start = EXTENT_MAP_HOLE;
3565
3566 /*
3567 * If we have a file range that points to a compressed extent
3568 * and it's followed by a consecutive file range that points
3569 * to the same compressed extent (possibly with a different
3570 * offset and/or length, so it either points to the whole extent
3571 * or only part of it), we must make sure we do not submit a
3572 * single bio to populate the pages for the 2 ranges because
3573 * this makes the compressed extent read zero out the pages
3574 * belonging to the 2nd range. Imagine the following scenario:
3575 *
3576 * File layout
3577 * [0 - 8K] [8K - 24K]
3578 * | |
3579 * | |
3580 * points to extent X, points to extent X,
3581 * offset 4K, length of 8K offset 0, length 16K
3582 *
3583 * [extent X, compressed length = 4K uncompressed length = 16K]
3584 *
3585 * If the bio to read the compressed extent covers both ranges,
3586 * it will decompress extent X into the pages belonging to the
3587 * first range and then it will stop, zeroing out the remaining
3588 * pages that belong to the other range that points to extent X.
3589 * So here we make sure we submit 2 bios, one for the first
3590 * range and another one for the third range. Both will target
3591 * the same physical extent from disk, but we can't currently
3592 * make the compressed bio endio callback populate the pages
3593 * for both ranges because each compressed bio is tightly
3594 * coupled with a single extent map, and each range can have
3595 * an extent map with a different offset value relative to the
3596 * uncompressed data of our extent and different lengths. This
3597 * is a corner case so we prioritize correctness over
3598 * non-optimal behavior (submitting 2 bios for the same extent).
3599 */
3600 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3601 prev_em_start && *prev_em_start != (u64)-1 &&
3602 *prev_em_start != em->start)
3603 force_bio_submit = true;
3604
3605 if (prev_em_start)
3606 *prev_em_start = em->start;
3607
3608 free_extent_map(em);
3609 em = NULL;
3610
3611 /* we've found a hole, just zero and go on */
3612 if (block_start == EXTENT_MAP_HOLE) {
3613 struct extent_state *cached = NULL;
3614
3615 memzero_page(page, pg_offset, iosize);
3616 flush_dcache_page(page);
3617
3618 set_extent_uptodate(tree, cur, cur + iosize - 1,
3619 &cached, GFP_NOFS);
3620 unlock_extent_cached(tree, cur,
3621 cur + iosize - 1, &cached);
3622 end_page_read(page, true, cur, iosize);
3623 cur = cur + iosize;
3624 pg_offset += iosize;
3625 continue;
3626 }
3627 /* the get_extent function already copied into the page */
3628 if (test_range_bit(tree, cur, cur_end,
3629 EXTENT_UPTODATE, 1, NULL)) {
3630 check_page_uptodate(tree, page);
3631 unlock_extent(tree, cur, cur + iosize - 1);
3632 end_page_read(page, true, cur, iosize);
3633 cur = cur + iosize;
3634 pg_offset += iosize;
3635 continue;
3636 }
3637 /* we have an inline extent but it didn't get marked up
3638 * to date. Error out
3639 */
3640 if (block_start == EXTENT_MAP_INLINE) {
3641 unlock_extent(tree, cur, cur + iosize - 1);
3642 end_page_read(page, false, cur, iosize);
3643 cur = cur + iosize;
3644 pg_offset += iosize;
3645 continue;
3646 }
3647
3648 ret = submit_extent_page(REQ_OP_READ | read_flags, NULL,
3649 bio_ctrl, page, disk_bytenr, iosize,
3650 pg_offset,
3651 end_bio_extent_readpage, 0,
3652 this_bio_flag,
3653 force_bio_submit);
3654 if (!ret) {
3655 nr++;
3656 } else {
3657 unlock_extent(tree, cur, cur + iosize - 1);
3658 end_page_read(page, false, cur, iosize);
3659 goto out;
3660 }
3661 cur = cur + iosize;
3662 pg_offset += iosize;
3663 }
3664out:
3665 return ret;
3666}
3667
3668static inline void contiguous_readpages(struct page *pages[], int nr_pages,
3669 u64 start, u64 end,
3670 struct extent_map **em_cached,
3671 struct btrfs_bio_ctrl *bio_ctrl,
3672 u64 *prev_em_start)
3673{
3674 struct btrfs_inode *inode = BTRFS_I(pages[0]->mapping->host);
3675 int index;
3676
3677 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
3678
3679 for (index = 0; index < nr_pages; index++) {
3680 btrfs_do_readpage(pages[index], em_cached, bio_ctrl,
3681 REQ_RAHEAD, prev_em_start);
3682 put_page(pages[index]);
3683 }
3684}
3685
3686static void update_nr_written(struct writeback_control *wbc,
3687 unsigned long nr_written)
3688{
3689 wbc->nr_to_write -= nr_written;
3690}
3691
3692/*
3693 * helper for __extent_writepage, doing all of the delayed allocation setup.
3694 *
3695 * This returns 1 if btrfs_run_delalloc_range function did all the work required
3696 * to write the page (copy into inline extent). In this case the IO has
3697 * been started and the page is already unlocked.
3698 *
3699 * This returns 0 if all went well (page still locked)
3700 * This returns < 0 if there were errors (page still locked)
3701 */
3702static noinline_for_stack int writepage_delalloc(struct btrfs_inode *inode,
3703 struct page *page, struct writeback_control *wbc,
3704 u64 delalloc_start, unsigned long *nr_written)
3705{
3706 u64 page_end = delalloc_start + PAGE_SIZE - 1;
3707 bool found;
3708 u64 delalloc_to_write = 0;
3709 u64 delalloc_end = 0;
3710 int ret;
3711 int page_started = 0;
3712
3713
3714 while (delalloc_end < page_end) {
3715 found = find_lock_delalloc_range(&inode->vfs_inode, page,
3716 &delalloc_start,
3717 &delalloc_end);
3718 if (!found) {
3719 delalloc_start = delalloc_end + 1;
3720 continue;
3721 }
3722 ret = btrfs_run_delalloc_range(inode, page, delalloc_start,
3723 delalloc_end, &page_started, nr_written, wbc);
3724 if (ret) {
3725 SetPageError(page);
3726 /*
3727 * btrfs_run_delalloc_range should return < 0 for error
3728 * but just in case, we use > 0 here meaning the IO is
3729 * started, so we don't want to return > 0 unless
3730 * things are going well.
3731 */
3732 return ret < 0 ? ret : -EIO;
3733 }
3734 /*
3735 * delalloc_end is already one less than the total length, so
3736 * we don't subtract one from PAGE_SIZE
3737 */
3738 delalloc_to_write += (delalloc_end - delalloc_start +
3739 PAGE_SIZE) >> PAGE_SHIFT;
3740 delalloc_start = delalloc_end + 1;
3741 }
3742 if (wbc->nr_to_write < delalloc_to_write) {
3743 int thresh = 8192;
3744
3745 if (delalloc_to_write < thresh * 2)
3746 thresh = delalloc_to_write;
3747 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3748 thresh);
3749 }
3750
3751 /* did the fill delalloc function already unlock and start
3752 * the IO?
3753 */
3754 if (page_started) {
3755 /*
3756 * we've unlocked the page, so we can't update
3757 * the mapping's writeback index, just update
3758 * nr_to_write.
3759 */
3760 wbc->nr_to_write -= *nr_written;
3761 return 1;
3762 }
3763
3764 return 0;
3765}
3766
3767/*
3768 * Find the first byte we need to write.
3769 *
3770 * For subpage, one page can contain several sectors, and
3771 * __extent_writepage_io() will just grab all extent maps in the page
3772 * range and try to submit all non-inline/non-compressed extents.
3773 *
3774 * This is a big problem for subpage, we shouldn't re-submit already written
3775 * data at all.
3776 * This function will lookup subpage dirty bit to find which range we really
3777 * need to submit.
3778 *
3779 * Return the next dirty range in [@start, @end).
3780 * If no dirty range is found, @start will be page_offset(page) + PAGE_SIZE.
3781 */
3782static void find_next_dirty_byte(struct btrfs_fs_info *fs_info,
3783 struct page *page, u64 *start, u64 *end)
3784{
3785 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
3786 u64 orig_start = *start;
3787 /* Declare as unsigned long so we can use bitmap ops */
3788 unsigned long dirty_bitmap;
3789 unsigned long flags;
3790 int nbits = (orig_start - page_offset(page)) >> fs_info->sectorsize_bits;
3791 int range_start_bit = nbits;
3792 int range_end_bit;
3793
3794 /*
3795 * For regular sector size == page size case, since one page only
3796 * contains one sector, we return the page offset directly.
3797 */
3798 if (fs_info->sectorsize == PAGE_SIZE) {
3799 *start = page_offset(page);
3800 *end = page_offset(page) + PAGE_SIZE;
3801 return;
3802 }
3803
3804 /* We should have the page locked, but just in case */
3805 spin_lock_irqsave(&subpage->lock, flags);
3806 dirty_bitmap = subpage->dirty_bitmap;
3807 spin_unlock_irqrestore(&subpage->lock, flags);
3808
3809 bitmap_next_set_region(&dirty_bitmap, &range_start_bit, &range_end_bit,
3810 BTRFS_SUBPAGE_BITMAP_SIZE);
3811 *start = page_offset(page) + range_start_bit * fs_info->sectorsize;
3812 *end = page_offset(page) + range_end_bit * fs_info->sectorsize;
3813}
3814
3815/*
3816 * helper for __extent_writepage. This calls the writepage start hooks,
3817 * and does the loop to map the page into extents and bios.
3818 *
3819 * We return 1 if the IO is started and the page is unlocked,
3820 * 0 if all went well (page still locked)
3821 * < 0 if there were errors (page still locked)
3822 */
3823static noinline_for_stack int __extent_writepage_io(struct btrfs_inode *inode,
3824 struct page *page,
3825 struct writeback_control *wbc,
3826 struct extent_page_data *epd,
3827 loff_t i_size,
3828 unsigned long nr_written,
3829 int *nr_ret)
3830{
3831 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3832 u64 start = page_offset(page);
3833 u64 end = start + PAGE_SIZE - 1;
3834 u64 cur = start;
3835 u64 extent_offset;
3836 u64 block_start;
3837 struct extent_map *em;
3838 int ret = 0;
3839 int nr = 0;
3840 u32 opf = REQ_OP_WRITE;
3841 const unsigned int write_flags = wbc_to_write_flags(wbc);
3842 bool compressed;
3843
3844 ret = btrfs_writepage_cow_fixup(page, start, end);
3845 if (ret) {
3846 /* Fixup worker will requeue */
3847 redirty_page_for_writepage(wbc, page);
3848 update_nr_written(wbc, nr_written);
3849 unlock_page(page);
3850 return 1;
3851 }
3852
3853 /*
3854 * we don't want to touch the inode after unlocking the page,
3855 * so we update the mapping writeback index now
3856 */
3857 update_nr_written(wbc, nr_written + 1);
3858
3859 while (cur <= end) {
3860 u64 disk_bytenr;
3861 u64 em_end;
3862 u64 dirty_range_start = cur;
3863 u64 dirty_range_end;
3864 u32 iosize;
3865
3866 if (cur >= i_size) {
3867 btrfs_writepage_endio_finish_ordered(inode, page, cur,
3868 end, 1);
3869 break;
3870 }
3871
3872 find_next_dirty_byte(fs_info, page, &dirty_range_start,
3873 &dirty_range_end);
3874 if (cur < dirty_range_start) {
3875 cur = dirty_range_start;
3876 continue;
3877 }
3878
3879 em = btrfs_get_extent(inode, NULL, 0, cur, end - cur + 1);
3880 if (IS_ERR_OR_NULL(em)) {
3881 btrfs_page_set_error(fs_info, page, cur, end - cur + 1);
3882 ret = PTR_ERR_OR_ZERO(em);
3883 break;
3884 }
3885
3886 extent_offset = cur - em->start;
3887 em_end = extent_map_end(em);
3888 ASSERT(cur <= em_end);
3889 ASSERT(cur < end);
3890 ASSERT(IS_ALIGNED(em->start, fs_info->sectorsize));
3891 ASSERT(IS_ALIGNED(em->len, fs_info->sectorsize));
3892 block_start = em->block_start;
3893 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3894 disk_bytenr = em->block_start + extent_offset;
3895
3896 /*
3897 * Note that em_end from extent_map_end() and dirty_range_end from
3898 * find_next_dirty_byte() are all exclusive
3899 */
3900 iosize = min(min(em_end, end + 1), dirty_range_end) - cur;
3901
3902 if (btrfs_use_zone_append(inode, em->block_start))
3903 opf = REQ_OP_ZONE_APPEND;
3904
3905 free_extent_map(em);
3906 em = NULL;
3907
3908 /*
3909 * compressed and inline extents are written through other
3910 * paths in the FS
3911 */
3912 if (compressed || block_start == EXTENT_MAP_HOLE ||
3913 block_start == EXTENT_MAP_INLINE) {
3914 if (compressed)
3915 nr++;
3916 else
3917 btrfs_writepage_endio_finish_ordered(inode,
3918 page, cur, cur + iosize - 1, 1);
3919 cur += iosize;
3920 continue;
3921 }
3922
3923 btrfs_set_range_writeback(inode, cur, cur + iosize - 1);
3924 if (!PageWriteback(page)) {
3925 btrfs_err(inode->root->fs_info,
3926 "page %lu not writeback, cur %llu end %llu",
3927 page->index, cur, end);
3928 }
3929
3930 /*
3931 * Although the PageDirty bit is cleared before entering this
3932 * function, subpage dirty bit is not cleared.
3933 * So clear subpage dirty bit here so next time we won't submit
3934 * page for range already written to disk.
3935 */
3936 btrfs_page_clear_dirty(fs_info, page, cur, iosize);
3937
3938 ret = submit_extent_page(opf | write_flags, wbc,
3939 &epd->bio_ctrl, page,
3940 disk_bytenr, iosize,
3941 cur - page_offset(page),
3942 end_bio_extent_writepage,
3943 0, 0, false);
3944 if (ret) {
3945 btrfs_page_set_error(fs_info, page, cur, iosize);
3946 if (PageWriteback(page))
3947 btrfs_page_clear_writeback(fs_info, page, cur,
3948 iosize);
3949 }
3950
3951 cur += iosize;
3952 nr++;
3953 }
3954 *nr_ret = nr;
3955 return ret;
3956}
3957
3958/*
3959 * the writepage semantics are similar to regular writepage. extent
3960 * records are inserted to lock ranges in the tree, and as dirty areas
3961 * are found, they are marked writeback. Then the lock bits are removed
3962 * and the end_io handler clears the writeback ranges
3963 *
3964 * Return 0 if everything goes well.
3965 * Return <0 for error.
3966 */
3967static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3968 struct extent_page_data *epd)
3969{
3970 struct inode *inode = page->mapping->host;
3971 u64 start = page_offset(page);
3972 u64 page_end = start + PAGE_SIZE - 1;
3973 int ret;
3974 int nr = 0;
3975 size_t pg_offset;
3976 loff_t i_size = i_size_read(inode);
3977 unsigned long end_index = i_size >> PAGE_SHIFT;
3978 unsigned long nr_written = 0;
3979
3980 trace___extent_writepage(page, inode, wbc);
3981
3982 WARN_ON(!PageLocked(page));
3983
3984 ClearPageError(page);
3985
3986 pg_offset = offset_in_page(i_size);
3987 if (page->index > end_index ||
3988 (page->index == end_index && !pg_offset)) {
3989 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3990 unlock_page(page);
3991 return 0;
3992 }
3993
3994 if (page->index == end_index) {
3995 memzero_page(page, pg_offset, PAGE_SIZE - pg_offset);
3996 flush_dcache_page(page);
3997 }
3998
3999 ret = set_page_extent_mapped(page);
4000 if (ret < 0) {
4001 SetPageError(page);
4002 goto done;
4003 }
4004
4005 if (!epd->extent_locked) {
4006 ret = writepage_delalloc(BTRFS_I(inode), page, wbc, start,
4007 &nr_written);
4008 if (ret == 1)
4009 return 0;
4010 if (ret)
4011 goto done;
4012 }
4013
4014 ret = __extent_writepage_io(BTRFS_I(inode), page, wbc, epd, i_size,
4015 nr_written, &nr);
4016 if (ret == 1)
4017 return 0;
4018
4019done:
4020 if (nr == 0) {
4021 /* make sure the mapping tag for page dirty gets cleared */
4022 set_page_writeback(page);
4023 end_page_writeback(page);
4024 }
4025 if (PageError(page)) {
4026 ret = ret < 0 ? ret : -EIO;
4027 end_extent_writepage(page, ret, start, page_end);
4028 }
4029 unlock_page(page);
4030 ASSERT(ret <= 0);
4031 return ret;
4032}
4033
4034void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
4035{
4036 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
4037 TASK_UNINTERRUPTIBLE);
4038}
4039
4040static void end_extent_buffer_writeback(struct extent_buffer *eb)
4041{
4042 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
4043 smp_mb__after_atomic();
4044 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
4045}
4046
4047/*
4048 * Lock extent buffer status and pages for writeback.
4049 *
4050 * May try to flush write bio if we can't get the lock.
4051 *
4052 * Return 0 if the extent buffer doesn't need to be submitted.
4053 * (E.g. the extent buffer is not dirty)
4054 * Return >0 is the extent buffer is submitted to bio.
4055 * Return <0 if something went wrong, no page is locked.
4056 */
4057static noinline_for_stack int lock_extent_buffer_for_io(struct extent_buffer *eb,
4058 struct extent_page_data *epd)
4059{
4060 struct btrfs_fs_info *fs_info = eb->fs_info;
4061 int i, num_pages, failed_page_nr;
4062 int flush = 0;
4063 int ret = 0;
4064
4065 if (!btrfs_try_tree_write_lock(eb)) {
4066 ret = flush_write_bio(epd);
4067 if (ret < 0)
4068 return ret;
4069 flush = 1;
4070 btrfs_tree_lock(eb);
4071 }
4072
4073 if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
4074 btrfs_tree_unlock(eb);
4075 if (!epd->sync_io)
4076 return 0;
4077 if (!flush) {
4078 ret = flush_write_bio(epd);
4079 if (ret < 0)
4080 return ret;
4081 flush = 1;
4082 }
4083 while (1) {
4084 wait_on_extent_buffer_writeback(eb);
4085 btrfs_tree_lock(eb);
4086 if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
4087 break;
4088 btrfs_tree_unlock(eb);
4089 }
4090 }
4091
4092 /*
4093 * We need to do this to prevent races in people who check if the eb is
4094 * under IO since we can end up having no IO bits set for a short period
4095 * of time.
4096 */
4097 spin_lock(&eb->refs_lock);
4098 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
4099 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
4100 spin_unlock(&eb->refs_lock);
4101 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
4102 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
4103 -eb->len,
4104 fs_info->dirty_metadata_batch);
4105 ret = 1;
4106 } else {
4107 spin_unlock(&eb->refs_lock);
4108 }
4109
4110 btrfs_tree_unlock(eb);
4111
4112 /*
4113 * Either we don't need to submit any tree block, or we're submitting
4114 * subpage eb.
4115 * Subpage metadata doesn't use page locking at all, so we can skip
4116 * the page locking.
4117 */
4118 if (!ret || fs_info->sectorsize < PAGE_SIZE)
4119 return ret;
4120
4121 num_pages = num_extent_pages(eb);
4122 for (i = 0; i < num_pages; i++) {
4123 struct page *p = eb->pages[i];
4124
4125 if (!trylock_page(p)) {
4126 if (!flush) {
4127 int err;
4128
4129 err = flush_write_bio(epd);
4130 if (err < 0) {
4131 ret = err;
4132 failed_page_nr = i;
4133 goto err_unlock;
4134 }
4135 flush = 1;
4136 }
4137 lock_page(p);
4138 }
4139 }
4140
4141 return ret;
4142err_unlock:
4143 /* Unlock already locked pages */
4144 for (i = 0; i < failed_page_nr; i++)
4145 unlock_page(eb->pages[i]);
4146 /*
4147 * Clear EXTENT_BUFFER_WRITEBACK and wake up anyone waiting on it.
4148 * Also set back EXTENT_BUFFER_DIRTY so future attempts to this eb can
4149 * be made and undo everything done before.
4150 */
4151 btrfs_tree_lock(eb);
4152 spin_lock(&eb->refs_lock);
4153 set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4154 end_extent_buffer_writeback(eb);
4155 spin_unlock(&eb->refs_lock);
4156 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes, eb->len,
4157 fs_info->dirty_metadata_batch);
4158 btrfs_clear_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
4159 btrfs_tree_unlock(eb);
4160 return ret;
4161}
4162
4163static void set_btree_ioerr(struct page *page, struct extent_buffer *eb)
4164{
4165 struct btrfs_fs_info *fs_info = eb->fs_info;
4166
4167 btrfs_page_set_error(fs_info, page, eb->start, eb->len);
4168 if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
4169 return;
4170
4171 /*
4172 * If we error out, we should add back the dirty_metadata_bytes
4173 * to make it consistent.
4174 */
4175 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
4176 eb->len, fs_info->dirty_metadata_batch);
4177
4178 /*
4179 * If writeback for a btree extent that doesn't belong to a log tree
4180 * failed, increment the counter transaction->eb_write_errors.
4181 * We do this because while the transaction is running and before it's
4182 * committing (when we call filemap_fdata[write|wait]_range against
4183 * the btree inode), we might have
4184 * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
4185 * returns an error or an error happens during writeback, when we're
4186 * committing the transaction we wouldn't know about it, since the pages
4187 * can be no longer dirty nor marked anymore for writeback (if a
4188 * subsequent modification to the extent buffer didn't happen before the
4189 * transaction commit), which makes filemap_fdata[write|wait]_range not
4190 * able to find the pages tagged with SetPageError at transaction
4191 * commit time. So if this happens we must abort the transaction,
4192 * otherwise we commit a super block with btree roots that point to
4193 * btree nodes/leafs whose content on disk is invalid - either garbage
4194 * or the content of some node/leaf from a past generation that got
4195 * cowed or deleted and is no longer valid.
4196 *
4197 * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
4198 * not be enough - we need to distinguish between log tree extents vs
4199 * non-log tree extents, and the next filemap_fdatawait_range() call
4200 * will catch and clear such errors in the mapping - and that call might
4201 * be from a log sync and not from a transaction commit. Also, checking
4202 * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
4203 * not done and would not be reliable - the eb might have been released
4204 * from memory and reading it back again means that flag would not be
4205 * set (since it's a runtime flag, not persisted on disk).
4206 *
4207 * Using the flags below in the btree inode also makes us achieve the
4208 * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
4209 * writeback for all dirty pages and before filemap_fdatawait_range()
4210 * is called, the writeback for all dirty pages had already finished
4211 * with errors - because we were not using AS_EIO/AS_ENOSPC,
4212 * filemap_fdatawait_range() would return success, as it could not know
4213 * that writeback errors happened (the pages were no longer tagged for
4214 * writeback).
4215 */
4216 switch (eb->log_index) {
4217 case -1:
4218 set_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags);
4219 break;
4220 case 0:
4221 set_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
4222 break;
4223 case 1:
4224 set_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
4225 break;
4226 default:
4227 BUG(); /* unexpected, logic error */
4228 }
4229}
4230
4231/*
4232 * The endio specific version which won't touch any unsafe spinlock in endio
4233 * context.
4234 */
4235static struct extent_buffer *find_extent_buffer_nolock(
4236 struct btrfs_fs_info *fs_info, u64 start)
4237{
4238 struct extent_buffer *eb;
4239
4240 rcu_read_lock();
4241 eb = radix_tree_lookup(&fs_info->buffer_radix,
4242 start >> fs_info->sectorsize_bits);
4243 if (eb && atomic_inc_not_zero(&eb->refs)) {
4244 rcu_read_unlock();
4245 return eb;
4246 }
4247 rcu_read_unlock();
4248 return NULL;
4249}
4250
4251/*
4252 * The endio function for subpage extent buffer write.
4253 *
4254 * Unlike end_bio_extent_buffer_writepage(), we only call end_page_writeback()
4255 * after all extent buffers in the page has finished their writeback.
4256 */
4257static void end_bio_subpage_eb_writepage(struct bio *bio)
4258{
4259 struct btrfs_fs_info *fs_info;
4260 struct bio_vec *bvec;
4261 struct bvec_iter_all iter_all;
4262
4263 fs_info = btrfs_sb(bio_first_page_all(bio)->mapping->host->i_sb);
4264 ASSERT(fs_info->sectorsize < PAGE_SIZE);
4265
4266 ASSERT(!bio_flagged(bio, BIO_CLONED));
4267 bio_for_each_segment_all(bvec, bio, iter_all) {
4268 struct page *page = bvec->bv_page;
4269 u64 bvec_start = page_offset(page) + bvec->bv_offset;
4270 u64 bvec_end = bvec_start + bvec->bv_len - 1;
4271 u64 cur_bytenr = bvec_start;
4272
4273 ASSERT(IS_ALIGNED(bvec->bv_len, fs_info->nodesize));
4274
4275 /* Iterate through all extent buffers in the range */
4276 while (cur_bytenr <= bvec_end) {
4277 struct extent_buffer *eb;
4278 int done;
4279
4280 /*
4281 * Here we can't use find_extent_buffer(), as it may
4282 * try to lock eb->refs_lock, which is not safe in endio
4283 * context.
4284 */
4285 eb = find_extent_buffer_nolock(fs_info, cur_bytenr);
4286 ASSERT(eb);
4287
4288 cur_bytenr = eb->start + eb->len;
4289
4290 ASSERT(test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags));
4291 done = atomic_dec_and_test(&eb->io_pages);
4292 ASSERT(done);
4293
4294 if (bio->bi_status ||
4295 test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
4296 ClearPageUptodate(page);
4297 set_btree_ioerr(page, eb);
4298 }
4299
4300 btrfs_subpage_clear_writeback(fs_info, page, eb->start,
4301 eb->len);
4302 end_extent_buffer_writeback(eb);
4303 /*
4304 * free_extent_buffer() will grab spinlock which is not
4305 * safe in endio context. Thus here we manually dec
4306 * the ref.
4307 */
4308 atomic_dec(&eb->refs);
4309 }
4310 }
4311 bio_put(bio);
4312}
4313
4314static void end_bio_extent_buffer_writepage(struct bio *bio)
4315{
4316 struct bio_vec *bvec;
4317 struct extent_buffer *eb;
4318 int done;
4319 struct bvec_iter_all iter_all;
4320
4321 ASSERT(!bio_flagged(bio, BIO_CLONED));
4322 bio_for_each_segment_all(bvec, bio, iter_all) {
4323 struct page *page = bvec->bv_page;
4324
4325 eb = (struct extent_buffer *)page->private;
4326 BUG_ON(!eb);
4327 done = atomic_dec_and_test(&eb->io_pages);
4328
4329 if (bio->bi_status ||
4330 test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
4331 ClearPageUptodate(page);
4332 set_btree_ioerr(page, eb);
4333 }
4334
4335 end_page_writeback(page);
4336
4337 if (!done)
4338 continue;
4339
4340 end_extent_buffer_writeback(eb);
4341 }
4342
4343 bio_put(bio);
4344}
4345
4346static void prepare_eb_write(struct extent_buffer *eb)
4347{
4348 u32 nritems;
4349 unsigned long start;
4350 unsigned long end;
4351
4352 clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
4353 atomic_set(&eb->io_pages, num_extent_pages(eb));
4354
4355 /* Set btree blocks beyond nritems with 0 to avoid stale content */
4356 nritems = btrfs_header_nritems(eb);
4357 if (btrfs_header_level(eb) > 0) {
4358 end = btrfs_node_key_ptr_offset(nritems);
4359 memzero_extent_buffer(eb, end, eb->len - end);
4360 } else {
4361 /*
4362 * Leaf:
4363 * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
4364 */
4365 start = btrfs_item_nr_offset(nritems);
4366 end = BTRFS_LEAF_DATA_OFFSET + leaf_data_end(eb);
4367 memzero_extent_buffer(eb, start, end - start);
4368 }
4369}
4370
4371/*
4372 * Unlike the work in write_one_eb(), we rely completely on extent locking.
4373 * Page locking is only utilized at minimum to keep the VMM code happy.
4374 */
4375static int write_one_subpage_eb(struct extent_buffer *eb,
4376 struct writeback_control *wbc,
4377 struct extent_page_data *epd)
4378{
4379 struct btrfs_fs_info *fs_info = eb->fs_info;
4380 struct page *page = eb->pages[0];
4381 unsigned int write_flags = wbc_to_write_flags(wbc) | REQ_META;
4382 bool no_dirty_ebs = false;
4383 int ret;
4384
4385 prepare_eb_write(eb);
4386
4387 /* clear_page_dirty_for_io() in subpage helper needs page locked */
4388 lock_page(page);
4389 btrfs_subpage_set_writeback(fs_info, page, eb->start, eb->len);
4390
4391 /* Check if this is the last dirty bit to update nr_written */
4392 no_dirty_ebs = btrfs_subpage_clear_and_test_dirty(fs_info, page,
4393 eb->start, eb->len);
4394 if (no_dirty_ebs)
4395 clear_page_dirty_for_io(page);
4396
4397 ret = submit_extent_page(REQ_OP_WRITE | write_flags, wbc,
4398 &epd->bio_ctrl, page, eb->start, eb->len,
4399 eb->start - page_offset(page),
4400 end_bio_subpage_eb_writepage, 0, 0, false);
4401 if (ret) {
4402 btrfs_subpage_clear_writeback(fs_info, page, eb->start, eb->len);
4403 set_btree_ioerr(page, eb);
4404 unlock_page(page);
4405
4406 if (atomic_dec_and_test(&eb->io_pages))
4407 end_extent_buffer_writeback(eb);
4408 return -EIO;
4409 }
4410 unlock_page(page);
4411 /*
4412 * Submission finished without problem, if no range of the page is
4413 * dirty anymore, we have submitted a page. Update nr_written in wbc.
4414 */
4415 if (no_dirty_ebs)
4416 update_nr_written(wbc, 1);
4417 return ret;
4418}
4419
4420static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
4421 struct writeback_control *wbc,
4422 struct extent_page_data *epd)
4423{
4424 u64 disk_bytenr = eb->start;
4425 int i, num_pages;
4426 unsigned int write_flags = wbc_to_write_flags(wbc) | REQ_META;
4427 int ret = 0;
4428
4429 prepare_eb_write(eb);
4430
4431 num_pages = num_extent_pages(eb);
4432 for (i = 0; i < num_pages; i++) {
4433 struct page *p = eb->pages[i];
4434
4435 clear_page_dirty_for_io(p);
4436 set_page_writeback(p);
4437 ret = submit_extent_page(REQ_OP_WRITE | write_flags, wbc,
4438 &epd->bio_ctrl, p, disk_bytenr,
4439 PAGE_SIZE, 0,
4440 end_bio_extent_buffer_writepage,
4441 0, 0, false);
4442 if (ret) {
4443 set_btree_ioerr(p, eb);
4444 if (PageWriteback(p))
4445 end_page_writeback(p);
4446 if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
4447 end_extent_buffer_writeback(eb);
4448 ret = -EIO;
4449 break;
4450 }
4451 disk_bytenr += PAGE_SIZE;
4452 update_nr_written(wbc, 1);
4453 unlock_page(p);
4454 }
4455
4456 if (unlikely(ret)) {
4457 for (; i < num_pages; i++) {
4458 struct page *p = eb->pages[i];
4459 clear_page_dirty_for_io(p);
4460 unlock_page(p);
4461 }
4462 }
4463
4464 return ret;
4465}
4466
4467/*
4468 * Submit one subpage btree page.
4469 *
4470 * The main difference to submit_eb_page() is:
4471 * - Page locking
4472 * For subpage, we don't rely on page locking at all.
4473 *
4474 * - Flush write bio
4475 * We only flush bio if we may be unable to fit current extent buffers into
4476 * current bio.
4477 *
4478 * Return >=0 for the number of submitted extent buffers.
4479 * Return <0 for fatal error.
4480 */
4481static int submit_eb_subpage(struct page *page,
4482 struct writeback_control *wbc,
4483 struct extent_page_data *epd)
4484{
4485 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
4486 int submitted = 0;
4487 u64 page_start = page_offset(page);
4488 int bit_start = 0;
4489 const int nbits = BTRFS_SUBPAGE_BITMAP_SIZE;
4490 int sectors_per_node = fs_info->nodesize >> fs_info->sectorsize_bits;
4491 int ret;
4492
4493 /* Lock and write each dirty extent buffers in the range */
4494 while (bit_start < nbits) {
4495 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
4496 struct extent_buffer *eb;
4497 unsigned long flags;
4498 u64 start;
4499
4500 /*
4501 * Take private lock to ensure the subpage won't be detached
4502 * in the meantime.
4503 */
4504 spin_lock(&page->mapping->private_lock);
4505 if (!PagePrivate(page)) {
4506 spin_unlock(&page->mapping->private_lock);
4507 break;
4508 }
4509 spin_lock_irqsave(&subpage->lock, flags);
4510 if (!((1 << bit_start) & subpage->dirty_bitmap)) {
4511 spin_unlock_irqrestore(&subpage->lock, flags);
4512 spin_unlock(&page->mapping->private_lock);
4513 bit_start++;
4514 continue;
4515 }
4516
4517 start = page_start + bit_start * fs_info->sectorsize;
4518 bit_start += sectors_per_node;
4519
4520 /*
4521 * Here we just want to grab the eb without touching extra
4522 * spin locks, so call find_extent_buffer_nolock().
4523 */
4524 eb = find_extent_buffer_nolock(fs_info, start);
4525 spin_unlock_irqrestore(&subpage->lock, flags);
4526 spin_unlock(&page->mapping->private_lock);
4527
4528 /*
4529 * The eb has already reached 0 refs thus find_extent_buffer()
4530 * doesn't return it. We don't need to write back such eb
4531 * anyway.
4532 */
4533 if (!eb)
4534 continue;
4535
4536 ret = lock_extent_buffer_for_io(eb, epd);
4537 if (ret == 0) {
4538 free_extent_buffer(eb);
4539 continue;
4540 }
4541 if (ret < 0) {
4542 free_extent_buffer(eb);
4543 goto cleanup;
4544 }
4545 ret = write_one_subpage_eb(eb, wbc, epd);
4546 free_extent_buffer(eb);
4547 if (ret < 0)
4548 goto cleanup;
4549 submitted++;
4550 }
4551 return submitted;
4552
4553cleanup:
4554 /* We hit error, end bio for the submitted extent buffers */
4555 end_write_bio(epd, ret);
4556 return ret;
4557}
4558
4559/*
4560 * Submit all page(s) of one extent buffer.
4561 *
4562 * @page: the page of one extent buffer
4563 * @eb_context: to determine if we need to submit this page, if current page
4564 * belongs to this eb, we don't need to submit
4565 *
4566 * The caller should pass each page in their bytenr order, and here we use
4567 * @eb_context to determine if we have submitted pages of one extent buffer.
4568 *
4569 * If we have, we just skip until we hit a new page that doesn't belong to
4570 * current @eb_context.
4571 *
4572 * If not, we submit all the page(s) of the extent buffer.
4573 *
4574 * Return >0 if we have submitted the extent buffer successfully.
4575 * Return 0 if we don't need to submit the page, as it's already submitted by
4576 * previous call.
4577 * Return <0 for fatal error.
4578 */
4579static int submit_eb_page(struct page *page, struct writeback_control *wbc,
4580 struct extent_page_data *epd,
4581 struct extent_buffer **eb_context)
4582{
4583 struct address_space *mapping = page->mapping;
4584 struct btrfs_block_group *cache = NULL;
4585 struct extent_buffer *eb;
4586 int ret;
4587
4588 if (!PagePrivate(page))
4589 return 0;
4590
4591 if (btrfs_sb(page->mapping->host->i_sb)->sectorsize < PAGE_SIZE)
4592 return submit_eb_subpage(page, wbc, epd);
4593
4594 spin_lock(&mapping->private_lock);
4595 if (!PagePrivate(page)) {
4596 spin_unlock(&mapping->private_lock);
4597 return 0;
4598 }
4599
4600 eb = (struct extent_buffer *)page->private;
4601
4602 /*
4603 * Shouldn't happen and normally this would be a BUG_ON but no point
4604 * crashing the machine for something we can survive anyway.
4605 */
4606 if (WARN_ON(!eb)) {
4607 spin_unlock(&mapping->private_lock);
4608 return 0;
4609 }
4610
4611 if (eb == *eb_context) {
4612 spin_unlock(&mapping->private_lock);
4613 return 0;
4614 }
4615 ret = atomic_inc_not_zero(&eb->refs);
4616 spin_unlock(&mapping->private_lock);
4617 if (!ret)
4618 return 0;
4619
4620 if (!btrfs_check_meta_write_pointer(eb->fs_info, eb, &cache)) {
4621 /*
4622 * If for_sync, this hole will be filled with
4623 * trasnsaction commit.
4624 */
4625 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
4626 ret = -EAGAIN;
4627 else
4628 ret = 0;
4629 free_extent_buffer(eb);
4630 return ret;
4631 }
4632
4633 *eb_context = eb;
4634
4635 ret = lock_extent_buffer_for_io(eb, epd);
4636 if (ret <= 0) {
4637 btrfs_revert_meta_write_pointer(cache, eb);
4638 if (cache)
4639 btrfs_put_block_group(cache);
4640 free_extent_buffer(eb);
4641 return ret;
4642 }
4643 if (cache)
4644 btrfs_put_block_group(cache);
4645 ret = write_one_eb(eb, wbc, epd);
4646 free_extent_buffer(eb);
4647 if (ret < 0)
4648 return ret;
4649 return 1;
4650}
4651
4652int btree_write_cache_pages(struct address_space *mapping,
4653 struct writeback_control *wbc)
4654{
4655 struct extent_buffer *eb_context = NULL;
4656 struct extent_page_data epd = {
4657 .bio_ctrl = { 0 },
4658 .extent_locked = 0,
4659 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4660 };
4661 struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
4662 int ret = 0;
4663 int done = 0;
4664 int nr_to_write_done = 0;
4665 struct pagevec pvec;
4666 int nr_pages;
4667 pgoff_t index;
4668 pgoff_t end; /* Inclusive */
4669 int scanned = 0;
4670 xa_mark_t tag;
4671
4672 pagevec_init(&pvec);
4673 if (wbc->range_cyclic) {
4674 index = mapping->writeback_index; /* Start from prev offset */
4675 end = -1;
4676 /*
4677 * Start from the beginning does not need to cycle over the
4678 * range, mark it as scanned.
4679 */
4680 scanned = (index == 0);
4681 } else {
4682 index = wbc->range_start >> PAGE_SHIFT;
4683 end = wbc->range_end >> PAGE_SHIFT;
4684 scanned = 1;
4685 }
4686 if (wbc->sync_mode == WB_SYNC_ALL)
4687 tag = PAGECACHE_TAG_TOWRITE;
4688 else
4689 tag = PAGECACHE_TAG_DIRTY;
4690 btrfs_zoned_meta_io_lock(fs_info);
4691retry:
4692 if (wbc->sync_mode == WB_SYNC_ALL)
4693 tag_pages_for_writeback(mapping, index, end);
4694 while (!done && !nr_to_write_done && (index <= end) &&
4695 (nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
4696 tag))) {
4697 unsigned i;
4698
4699 for (i = 0; i < nr_pages; i++) {
4700 struct page *page = pvec.pages[i];
4701
4702 ret = submit_eb_page(page, wbc, &epd, &eb_context);
4703 if (ret == 0)
4704 continue;
4705 if (ret < 0) {
4706 done = 1;
4707 break;
4708 }
4709
4710 /*
4711 * the filesystem may choose to bump up nr_to_write.
4712 * We have to make sure to honor the new nr_to_write
4713 * at any time
4714 */
4715 nr_to_write_done = wbc->nr_to_write <= 0;
4716 }
4717 pagevec_release(&pvec);
4718 cond_resched();
4719 }
4720 if (!scanned && !done) {
4721 /*
4722 * We hit the last page and there is more work to be done: wrap
4723 * back to the start of the file
4724 */
4725 scanned = 1;
4726 index = 0;
4727 goto retry;
4728 }
4729 if (ret < 0) {
4730 end_write_bio(&epd, ret);
4731 goto out;
4732 }
4733 /*
4734 * If something went wrong, don't allow any metadata write bio to be
4735 * submitted.
4736 *
4737 * This would prevent use-after-free if we had dirty pages not
4738 * cleaned up, which can still happen by fuzzed images.
4739 *
4740 * - Bad extent tree
4741 * Allowing existing tree block to be allocated for other trees.
4742 *
4743 * - Log tree operations
4744 * Exiting tree blocks get allocated to log tree, bumps its
4745 * generation, then get cleaned in tree re-balance.
4746 * Such tree block will not be written back, since it's clean,
4747 * thus no WRITTEN flag set.
4748 * And after log writes back, this tree block is not traced by
4749 * any dirty extent_io_tree.
4750 *
4751 * - Offending tree block gets re-dirtied from its original owner
4752 * Since it has bumped generation, no WRITTEN flag, it can be
4753 * reused without COWing. This tree block will not be traced
4754 * by btrfs_transaction::dirty_pages.
4755 *
4756 * Now such dirty tree block will not be cleaned by any dirty
4757 * extent io tree. Thus we don't want to submit such wild eb
4758 * if the fs already has error.
4759 */
4760 if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
4761 ret = flush_write_bio(&epd);
4762 } else {
4763 ret = -EROFS;
4764 end_write_bio(&epd, ret);
4765 }
4766out:
4767 btrfs_zoned_meta_io_unlock(fs_info);
4768 return ret;
4769}
4770
4771/**
4772 * Walk the list of dirty pages of the given address space and write all of them.
4773 *
4774 * @mapping: address space structure to write
4775 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
4776 * @epd: holds context for the write, namely the bio
4777 *
4778 * If a page is already under I/O, write_cache_pages() skips it, even
4779 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
4780 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
4781 * and msync() need to guarantee that all the data which was dirty at the time
4782 * the call was made get new I/O started against them. If wbc->sync_mode is
4783 * WB_SYNC_ALL then we were called for data integrity and we must wait for
4784 * existing IO to complete.
4785 */
4786static int extent_write_cache_pages(struct address_space *mapping,
4787 struct writeback_control *wbc,
4788 struct extent_page_data *epd)
4789{
4790 struct inode *inode = mapping->host;
4791 int ret = 0;
4792 int done = 0;
4793 int nr_to_write_done = 0;
4794 struct pagevec pvec;
4795 int nr_pages;
4796 pgoff_t index;
4797 pgoff_t end; /* Inclusive */
4798 pgoff_t done_index;
4799 int range_whole = 0;
4800 int scanned = 0;
4801 xa_mark_t tag;
4802
4803 /*
4804 * We have to hold onto the inode so that ordered extents can do their
4805 * work when the IO finishes. The alternative to this is failing to add
4806 * an ordered extent if the igrab() fails there and that is a huge pain
4807 * to deal with, so instead just hold onto the inode throughout the
4808 * writepages operation. If it fails here we are freeing up the inode
4809 * anyway and we'd rather not waste our time writing out stuff that is
4810 * going to be truncated anyway.
4811 */
4812 if (!igrab(inode))
4813 return 0;
4814
4815 pagevec_init(&pvec);
4816 if (wbc->range_cyclic) {
4817 index = mapping->writeback_index; /* Start from prev offset */
4818 end = -1;
4819 /*
4820 * Start from the beginning does not need to cycle over the
4821 * range, mark it as scanned.
4822 */
4823 scanned = (index == 0);
4824 } else {
4825 index = wbc->range_start >> PAGE_SHIFT;
4826 end = wbc->range_end >> PAGE_SHIFT;
4827 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
4828 range_whole = 1;
4829 scanned = 1;
4830 }
4831
4832 /*
4833 * We do the tagged writepage as long as the snapshot flush bit is set
4834 * and we are the first one who do the filemap_flush() on this inode.
4835 *
4836 * The nr_to_write == LONG_MAX is needed to make sure other flushers do
4837 * not race in and drop the bit.
4838 */
4839 if (range_whole && wbc->nr_to_write == LONG_MAX &&
4840 test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
4841 &BTRFS_I(inode)->runtime_flags))
4842 wbc->tagged_writepages = 1;
4843
4844 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
4845 tag = PAGECACHE_TAG_TOWRITE;
4846 else
4847 tag = PAGECACHE_TAG_DIRTY;
4848retry:
4849 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
4850 tag_pages_for_writeback(mapping, index, end);
4851 done_index = index;
4852 while (!done && !nr_to_write_done && (index <= end) &&
4853 (nr_pages = pagevec_lookup_range_tag(&pvec, mapping,
4854 &index, end, tag))) {
4855 unsigned i;
4856
4857 for (i = 0; i < nr_pages; i++) {
4858 struct page *page = pvec.pages[i];
4859
4860 done_index = page->index + 1;
4861 /*
4862 * At this point we hold neither the i_pages lock nor
4863 * the page lock: the page may be truncated or
4864 * invalidated (changing page->mapping to NULL),
4865 * or even swizzled back from swapper_space to
4866 * tmpfs file mapping
4867 */
4868 if (!trylock_page(page)) {
4869 ret = flush_write_bio(epd);
4870 BUG_ON(ret < 0);
4871 lock_page(page);
4872 }
4873
4874 if (unlikely(page->mapping != mapping)) {
4875 unlock_page(page);
4876 continue;
4877 }
4878
4879 if (wbc->sync_mode != WB_SYNC_NONE) {
4880 if (PageWriteback(page)) {
4881 ret = flush_write_bio(epd);
4882 BUG_ON(ret < 0);
4883 }
4884 wait_on_page_writeback(page);
4885 }
4886
4887 if (PageWriteback(page) ||
4888 !clear_page_dirty_for_io(page)) {
4889 unlock_page(page);
4890 continue;
4891 }
4892
4893 ret = __extent_writepage(page, wbc, epd);
4894 if (ret < 0) {
4895 done = 1;
4896 break;
4897 }
4898
4899 /*
4900 * the filesystem may choose to bump up nr_to_write.
4901 * We have to make sure to honor the new nr_to_write
4902 * at any time
4903 */
4904 nr_to_write_done = wbc->nr_to_write <= 0;
4905 }
4906 pagevec_release(&pvec);
4907 cond_resched();
4908 }
4909 if (!scanned && !done) {
4910 /*
4911 * We hit the last page and there is more work to be done: wrap
4912 * back to the start of the file
4913 */
4914 scanned = 1;
4915 index = 0;
4916
4917 /*
4918 * If we're looping we could run into a page that is locked by a
4919 * writer and that writer could be waiting on writeback for a
4920 * page in our current bio, and thus deadlock, so flush the
4921 * write bio here.
4922 */
4923 ret = flush_write_bio(epd);
4924 if (!ret)
4925 goto retry;
4926 }
4927
4928 if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4929 mapping->writeback_index = done_index;
4930
4931 btrfs_add_delayed_iput(inode);
4932 return ret;
4933}
4934
4935int extent_write_full_page(struct page *page, struct writeback_control *wbc)
4936{
4937 int ret;
4938 struct extent_page_data epd = {
4939 .bio_ctrl = { 0 },
4940 .extent_locked = 0,
4941 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4942 };
4943
4944 ret = __extent_writepage(page, wbc, &epd);
4945 ASSERT(ret <= 0);
4946 if (ret < 0) {
4947 end_write_bio(&epd, ret);
4948 return ret;
4949 }
4950
4951 ret = flush_write_bio(&epd);
4952 ASSERT(ret <= 0);
4953 return ret;
4954}
4955
4956int extent_write_locked_range(struct inode *inode, u64 start, u64 end,
4957 int mode)
4958{
4959 int ret = 0;
4960 struct address_space *mapping = inode->i_mapping;
4961 struct page *page;
4962 unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4963 PAGE_SHIFT;
4964
4965 struct extent_page_data epd = {
4966 .bio_ctrl = { 0 },
4967 .extent_locked = 1,
4968 .sync_io = mode == WB_SYNC_ALL,
4969 };
4970 struct writeback_control wbc_writepages = {
4971 .sync_mode = mode,
4972 .nr_to_write = nr_pages * 2,
4973 .range_start = start,
4974 .range_end = end + 1,
4975 /* We're called from an async helper function */
4976 .punt_to_cgroup = 1,
4977 .no_cgroup_owner = 1,
4978 };
4979
4980 wbc_attach_fdatawrite_inode(&wbc_writepages, inode);
4981 while (start <= end) {
4982 page = find_get_page(mapping, start >> PAGE_SHIFT);
4983 if (clear_page_dirty_for_io(page))
4984 ret = __extent_writepage(page, &wbc_writepages, &epd);
4985 else {
4986 btrfs_writepage_endio_finish_ordered(BTRFS_I(inode),
4987 page, start, start + PAGE_SIZE - 1, 1);
4988 unlock_page(page);
4989 }
4990 put_page(page);
4991 start += PAGE_SIZE;
4992 }
4993
4994 ASSERT(ret <= 0);
4995 if (ret == 0)
4996 ret = flush_write_bio(&epd);
4997 else
4998 end_write_bio(&epd, ret);
4999
5000 wbc_detach_inode(&wbc_writepages);
5001 return ret;
5002}
5003
5004int extent_writepages(struct address_space *mapping,
5005 struct writeback_control *wbc)
5006{
5007 int ret = 0;
5008 struct extent_page_data epd = {
5009 .bio_ctrl = { 0 },
5010 .extent_locked = 0,
5011 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
5012 };
5013
5014 ret = extent_write_cache_pages(mapping, wbc, &epd);
5015 ASSERT(ret <= 0);
5016 if (ret < 0) {
5017 end_write_bio(&epd, ret);
5018 return ret;
5019 }
5020 ret = flush_write_bio(&epd);
5021 return ret;
5022}
5023
5024void extent_readahead(struct readahead_control *rac)
5025{
5026 struct btrfs_bio_ctrl bio_ctrl = { 0 };
5027 struct page *pagepool[16];
5028 struct extent_map *em_cached = NULL;
5029 u64 prev_em_start = (u64)-1;
5030 int nr;
5031
5032 while ((nr = readahead_page_batch(rac, pagepool))) {
5033 u64 contig_start = readahead_pos(rac);
5034 u64 contig_end = contig_start + readahead_batch_length(rac) - 1;
5035
5036 contiguous_readpages(pagepool, nr, contig_start, contig_end,
5037 &em_cached, &bio_ctrl, &prev_em_start);
5038 }
5039
5040 if (em_cached)
5041 free_extent_map(em_cached);
5042
5043 if (bio_ctrl.bio) {
5044 if (submit_one_bio(bio_ctrl.bio, 0, bio_ctrl.bio_flags))
5045 return;
5046 }
5047}
5048
5049/*
5050 * basic invalidatepage code, this waits on any locked or writeback
5051 * ranges corresponding to the page, and then deletes any extent state
5052 * records from the tree
5053 */
5054int extent_invalidatepage(struct extent_io_tree *tree,
5055 struct page *page, unsigned long offset)
5056{
5057 struct extent_state *cached_state = NULL;
5058 u64 start = page_offset(page);
5059 u64 end = start + PAGE_SIZE - 1;
5060 size_t blocksize = page->mapping->host->i_sb->s_blocksize;
5061
5062 /* This function is only called for the btree inode */
5063 ASSERT(tree->owner == IO_TREE_BTREE_INODE_IO);
5064
5065 start += ALIGN(offset, blocksize);
5066 if (start > end)
5067 return 0;
5068
5069 lock_extent_bits(tree, start, end, &cached_state);
5070 wait_on_page_writeback(page);
5071
5072 /*
5073 * Currently for btree io tree, only EXTENT_LOCKED is utilized,
5074 * so here we only need to unlock the extent range to free any
5075 * existing extent state.
5076 */
5077 unlock_extent_cached(tree, start, end, &cached_state);
5078 return 0;
5079}
5080
5081/*
5082 * a helper for releasepage, this tests for areas of the page that
5083 * are locked or under IO and drops the related state bits if it is safe
5084 * to drop the page.
5085 */
5086static int try_release_extent_state(struct extent_io_tree *tree,
5087 struct page *page, gfp_t mask)
5088{
5089 u64 start = page_offset(page);
5090 u64 end = start + PAGE_SIZE - 1;
5091 int ret = 1;
5092
5093 if (test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL)) {
5094 ret = 0;
5095 } else {
5096 /*
5097 * At this point we can safely clear everything except the
5098 * locked bit, the nodatasum bit and the delalloc new bit.
5099 * The delalloc new bit will be cleared by ordered extent
5100 * completion.
5101 */
5102 ret = __clear_extent_bit(tree, start, end,
5103 ~(EXTENT_LOCKED | EXTENT_NODATASUM | EXTENT_DELALLOC_NEW),
5104 0, 0, NULL, mask, NULL);
5105
5106 /* if clear_extent_bit failed for enomem reasons,
5107 * we can't allow the release to continue.
5108 */
5109 if (ret < 0)
5110 ret = 0;
5111 else
5112 ret = 1;
5113 }
5114 return ret;
5115}
5116
5117/*
5118 * a helper for releasepage. As long as there are no locked extents
5119 * in the range corresponding to the page, both state records and extent
5120 * map records are removed
5121 */
5122int try_release_extent_mapping(struct page *page, gfp_t mask)
5123{
5124 struct extent_map *em;
5125 u64 start = page_offset(page);
5126 u64 end = start + PAGE_SIZE - 1;
5127 struct btrfs_inode *btrfs_inode = BTRFS_I(page->mapping->host);
5128 struct extent_io_tree *tree = &btrfs_inode->io_tree;
5129 struct extent_map_tree *map = &btrfs_inode->extent_tree;
5130
5131 if (gfpflags_allow_blocking(mask) &&
5132 page->mapping->host->i_size > SZ_16M) {
5133 u64 len;
5134 while (start <= end) {
5135 struct btrfs_fs_info *fs_info;
5136 u64 cur_gen;
5137
5138 len = end - start + 1;
5139 write_lock(&map->lock);
5140 em = lookup_extent_mapping(map, start, len);
5141 if (!em) {
5142 write_unlock(&map->lock);
5143 break;
5144 }
5145 if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
5146 em->start != start) {
5147 write_unlock(&map->lock);
5148 free_extent_map(em);
5149 break;
5150 }
5151 if (test_range_bit(tree, em->start,
5152 extent_map_end(em) - 1,
5153 EXTENT_LOCKED, 0, NULL))
5154 goto next;
5155 /*
5156 * If it's not in the list of modified extents, used
5157 * by a fast fsync, we can remove it. If it's being
5158 * logged we can safely remove it since fsync took an
5159 * extra reference on the em.
5160 */
5161 if (list_empty(&em->list) ||
5162 test_bit(EXTENT_FLAG_LOGGING, &em->flags))
5163 goto remove_em;
5164 /*
5165 * If it's in the list of modified extents, remove it
5166 * only if its generation is older then the current one,
5167 * in which case we don't need it for a fast fsync.
5168 * Otherwise don't remove it, we could be racing with an
5169 * ongoing fast fsync that could miss the new extent.
5170 */
5171 fs_info = btrfs_inode->root->fs_info;
5172 spin_lock(&fs_info->trans_lock);
5173 cur_gen = fs_info->generation;
5174 spin_unlock(&fs_info->trans_lock);
5175 if (em->generation >= cur_gen)
5176 goto next;
5177remove_em:
5178 /*
5179 * We only remove extent maps that are not in the list of
5180 * modified extents or that are in the list but with a
5181 * generation lower then the current generation, so there
5182 * is no need to set the full fsync flag on the inode (it
5183 * hurts the fsync performance for workloads with a data
5184 * size that exceeds or is close to the system's memory).
5185 */
5186 remove_extent_mapping(map, em);
5187 /* once for the rb tree */
5188 free_extent_map(em);
5189next:
5190 start = extent_map_end(em);
5191 write_unlock(&map->lock);
5192
5193 /* once for us */
5194 free_extent_map(em);
5195
5196 cond_resched(); /* Allow large-extent preemption. */
5197 }
5198 }
5199 return try_release_extent_state(tree, page, mask);
5200}
5201
5202/*
5203 * helper function for fiemap, which doesn't want to see any holes.
5204 * This maps until we find something past 'last'
5205 */
5206static struct extent_map *get_extent_skip_holes(struct btrfs_inode *inode,
5207 u64 offset, u64 last)
5208{
5209 u64 sectorsize = btrfs_inode_sectorsize(inode);
5210 struct extent_map *em;
5211 u64 len;
5212
5213 if (offset >= last)
5214 return NULL;
5215
5216 while (1) {
5217 len = last - offset;
5218 if (len == 0)
5219 break;
5220 len = ALIGN(len, sectorsize);
5221 em = btrfs_get_extent_fiemap(inode, offset, len);
5222 if (IS_ERR_OR_NULL(em))
5223 return em;
5224
5225 /* if this isn't a hole return it */
5226 if (em->block_start != EXTENT_MAP_HOLE)
5227 return em;
5228
5229 /* this is a hole, advance to the next extent */
5230 offset = extent_map_end(em);
5231 free_extent_map(em);
5232 if (offset >= last)
5233 break;
5234 }
5235 return NULL;
5236}
5237
5238/*
5239 * To cache previous fiemap extent
5240 *
5241 * Will be used for merging fiemap extent
5242 */
5243struct fiemap_cache {
5244 u64 offset;
5245 u64 phys;
5246 u64 len;
5247 u32 flags;
5248 bool cached;
5249};
5250
5251/*
5252 * Helper to submit fiemap extent.
5253 *
5254 * Will try to merge current fiemap extent specified by @offset, @phys,
5255 * @len and @flags with cached one.
5256 * And only when we fails to merge, cached one will be submitted as
5257 * fiemap extent.
5258 *
5259 * Return value is the same as fiemap_fill_next_extent().
5260 */
5261static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
5262 struct fiemap_cache *cache,
5263 u64 offset, u64 phys, u64 len, u32 flags)
5264{
5265 int ret = 0;
5266
5267 if (!cache->cached)
5268 goto assign;
5269
5270 /*
5271 * Sanity check, extent_fiemap() should have ensured that new
5272 * fiemap extent won't overlap with cached one.
5273 * Not recoverable.
5274 *
5275 * NOTE: Physical address can overlap, due to compression
5276 */
5277 if (cache->offset + cache->len > offset) {
5278 WARN_ON(1);
5279 return -EINVAL;
5280 }
5281
5282 /*
5283 * Only merges fiemap extents if
5284 * 1) Their logical addresses are continuous
5285 *
5286 * 2) Their physical addresses are continuous
5287 * So truly compressed (physical size smaller than logical size)
5288 * extents won't get merged with each other
5289 *
5290 * 3) Share same flags except FIEMAP_EXTENT_LAST
5291 * So regular extent won't get merged with prealloc extent
5292 */
5293 if (cache->offset + cache->len == offset &&
5294 cache->phys + cache->len == phys &&
5295 (cache->flags & ~FIEMAP_EXTENT_LAST) ==
5296 (flags & ~FIEMAP_EXTENT_LAST)) {
5297 cache->len += len;
5298 cache->flags |= flags;
5299 goto try_submit_last;
5300 }
5301
5302 /* Not mergeable, need to submit cached one */
5303 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
5304 cache->len, cache->flags);
5305 cache->cached = false;
5306 if (ret)
5307 return ret;
5308assign:
5309 cache->cached = true;
5310 cache->offset = offset;
5311 cache->phys = phys;
5312 cache->len = len;
5313 cache->flags = flags;
5314try_submit_last:
5315 if (cache->flags & FIEMAP_EXTENT_LAST) {
5316 ret = fiemap_fill_next_extent(fieinfo, cache->offset,
5317 cache->phys, cache->len, cache->flags);
5318 cache->cached = false;
5319 }
5320 return ret;
5321}
5322
5323/*
5324 * Emit last fiemap cache
5325 *
5326 * The last fiemap cache may still be cached in the following case:
5327 * 0 4k 8k
5328 * |<- Fiemap range ->|
5329 * |<------------ First extent ----------->|
5330 *
5331 * In this case, the first extent range will be cached but not emitted.
5332 * So we must emit it before ending extent_fiemap().
5333 */
5334static int emit_last_fiemap_cache(struct fiemap_extent_info *fieinfo,
5335 struct fiemap_cache *cache)
5336{
5337 int ret;
5338
5339 if (!cache->cached)
5340 return 0;
5341
5342 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
5343 cache->len, cache->flags);
5344 cache->cached = false;
5345 if (ret > 0)
5346 ret = 0;
5347 return ret;
5348}
5349
5350int extent_fiemap(struct btrfs_inode *inode, struct fiemap_extent_info *fieinfo,
5351 u64 start, u64 len)
5352{
5353 int ret = 0;
5354 u64 off;
5355 u64 max = start + len;
5356 u32 flags = 0;
5357 u32 found_type;
5358 u64 last;
5359 u64 last_for_get_extent = 0;
5360 u64 disko = 0;
5361 u64 isize = i_size_read(&inode->vfs_inode);
5362 struct btrfs_key found_key;
5363 struct extent_map *em = NULL;
5364 struct extent_state *cached_state = NULL;
5365 struct btrfs_path *path;
5366 struct btrfs_root *root = inode->root;
5367 struct fiemap_cache cache = { 0 };
5368 struct ulist *roots;
5369 struct ulist *tmp_ulist;
5370 int end = 0;
5371 u64 em_start = 0;
5372 u64 em_len = 0;
5373 u64 em_end = 0;
5374
5375 if (len == 0)
5376 return -EINVAL;
5377
5378 path = btrfs_alloc_path();
5379 if (!path)
5380 return -ENOMEM;
5381
5382 roots = ulist_alloc(GFP_KERNEL);
5383 tmp_ulist = ulist_alloc(GFP_KERNEL);
5384 if (!roots || !tmp_ulist) {
5385 ret = -ENOMEM;
5386 goto out_free_ulist;
5387 }
5388
5389 /*
5390 * We can't initialize that to 'start' as this could miss extents due
5391 * to extent item merging
5392 */
5393 off = 0;
5394 start = round_down(start, btrfs_inode_sectorsize(inode));
5395 len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
5396
5397 /*
5398 * lookup the last file extent. We're not using i_size here
5399 * because there might be preallocation past i_size
5400 */
5401 ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
5402 0);
5403 if (ret < 0) {
5404 goto out_free_ulist;
5405 } else {
5406 WARN_ON(!ret);
5407 if (ret == 1)
5408 ret = 0;
5409 }
5410
5411 path->slots[0]--;
5412 btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
5413 found_type = found_key.type;
5414
5415 /* No extents, but there might be delalloc bits */
5416 if (found_key.objectid != btrfs_ino(inode) ||
5417 found_type != BTRFS_EXTENT_DATA_KEY) {
5418 /* have to trust i_size as the end */
5419 last = (u64)-1;
5420 last_for_get_extent = isize;
5421 } else {
5422 /*
5423 * remember the start of the last extent. There are a
5424 * bunch of different factors that go into the length of the
5425 * extent, so its much less complex to remember where it started
5426 */
5427 last = found_key.offset;
5428 last_for_get_extent = last + 1;
5429 }
5430 btrfs_release_path(path);
5431
5432 /*
5433 * we might have some extents allocated but more delalloc past those
5434 * extents. so, we trust isize unless the start of the last extent is
5435 * beyond isize
5436 */
5437 if (last < isize) {
5438 last = (u64)-1;
5439 last_for_get_extent = isize;
5440 }
5441
5442 lock_extent_bits(&inode->io_tree, start, start + len - 1,
5443 &cached_state);
5444
5445 em = get_extent_skip_holes(inode, start, last_for_get_extent);
5446 if (!em)
5447 goto out;
5448 if (IS_ERR(em)) {
5449 ret = PTR_ERR(em);
5450 goto out;
5451 }
5452
5453 while (!end) {
5454 u64 offset_in_extent = 0;
5455
5456 /* break if the extent we found is outside the range */
5457 if (em->start >= max || extent_map_end(em) < off)
5458 break;
5459
5460 /*
5461 * get_extent may return an extent that starts before our
5462 * requested range. We have to make sure the ranges
5463 * we return to fiemap always move forward and don't
5464 * overlap, so adjust the offsets here
5465 */
5466 em_start = max(em->start, off);
5467
5468 /*
5469 * record the offset from the start of the extent
5470 * for adjusting the disk offset below. Only do this if the
5471 * extent isn't compressed since our in ram offset may be past
5472 * what we have actually allocated on disk.
5473 */
5474 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
5475 offset_in_extent = em_start - em->start;
5476 em_end = extent_map_end(em);
5477 em_len = em_end - em_start;
5478 flags = 0;
5479 if (em->block_start < EXTENT_MAP_LAST_BYTE)
5480 disko = em->block_start + offset_in_extent;
5481 else
5482 disko = 0;
5483
5484 /*
5485 * bump off for our next call to get_extent
5486 */
5487 off = extent_map_end(em);
5488 if (off >= max)
5489 end = 1;
5490
5491 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
5492 end = 1;
5493 flags |= FIEMAP_EXTENT_LAST;
5494 } else if (em->block_start == EXTENT_MAP_INLINE) {
5495 flags |= (FIEMAP_EXTENT_DATA_INLINE |
5496 FIEMAP_EXTENT_NOT_ALIGNED);
5497 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
5498 flags |= (FIEMAP_EXTENT_DELALLOC |
5499 FIEMAP_EXTENT_UNKNOWN);
5500 } else if (fieinfo->fi_extents_max) {
5501 u64 bytenr = em->block_start -
5502 (em->start - em->orig_start);
5503
5504 /*
5505 * As btrfs supports shared space, this information
5506 * can be exported to userspace tools via
5507 * flag FIEMAP_EXTENT_SHARED. If fi_extents_max == 0
5508 * then we're just getting a count and we can skip the
5509 * lookup stuff.
5510 */
5511 ret = btrfs_check_shared(root, btrfs_ino(inode),
5512 bytenr, roots, tmp_ulist);
5513 if (ret < 0)
5514 goto out_free;
5515 if (ret)
5516 flags |= FIEMAP_EXTENT_SHARED;
5517 ret = 0;
5518 }
5519 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
5520 flags |= FIEMAP_EXTENT_ENCODED;
5521 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
5522 flags |= FIEMAP_EXTENT_UNWRITTEN;
5523
5524 free_extent_map(em);
5525 em = NULL;
5526 if ((em_start >= last) || em_len == (u64)-1 ||
5527 (last == (u64)-1 && isize <= em_end)) {
5528 flags |= FIEMAP_EXTENT_LAST;
5529 end = 1;
5530 }
5531
5532 /* now scan forward to see if this is really the last extent. */
5533 em = get_extent_skip_holes(inode, off, last_for_get_extent);
5534 if (IS_ERR(em)) {
5535 ret = PTR_ERR(em);
5536 goto out;
5537 }
5538 if (!em) {
5539 flags |= FIEMAP_EXTENT_LAST;
5540 end = 1;
5541 }
5542 ret = emit_fiemap_extent(fieinfo, &cache, em_start, disko,
5543 em_len, flags);
5544 if (ret) {
5545 if (ret == 1)
5546 ret = 0;
5547 goto out_free;
5548 }
5549 }
5550out_free:
5551 if (!ret)
5552 ret = emit_last_fiemap_cache(fieinfo, &cache);
5553 free_extent_map(em);
5554out:
5555 unlock_extent_cached(&inode->io_tree, start, start + len - 1,
5556 &cached_state);
5557
5558out_free_ulist:
5559 btrfs_free_path(path);
5560 ulist_free(roots);
5561 ulist_free(tmp_ulist);
5562 return ret;
5563}
5564
5565static void __free_extent_buffer(struct extent_buffer *eb)
5566{
5567 kmem_cache_free(extent_buffer_cache, eb);
5568}
5569
5570int extent_buffer_under_io(const struct extent_buffer *eb)
5571{
5572 return (atomic_read(&eb->io_pages) ||
5573 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
5574 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
5575}
5576
5577static bool page_range_has_eb(struct btrfs_fs_info *fs_info, struct page *page)
5578{
5579 struct btrfs_subpage *subpage;
5580
5581 lockdep_assert_held(&page->mapping->private_lock);
5582
5583 if (PagePrivate(page)) {
5584 subpage = (struct btrfs_subpage *)page->private;
5585 if (atomic_read(&subpage->eb_refs))
5586 return true;
5587 /*
5588 * Even there is no eb refs here, we may still have
5589 * end_page_read() call relying on page::private.
5590 */
5591 if (atomic_read(&subpage->readers))
5592 return true;
5593 }
5594 return false;
5595}
5596
5597static void detach_extent_buffer_page(struct extent_buffer *eb, struct page *page)
5598{
5599 struct btrfs_fs_info *fs_info = eb->fs_info;
5600 const bool mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
5601
5602 /*
5603 * For mapped eb, we're going to change the page private, which should
5604 * be done under the private_lock.
5605 */
5606 if (mapped)
5607 spin_lock(&page->mapping->private_lock);
5608
5609 if (!PagePrivate(page)) {
5610 if (mapped)
5611 spin_unlock(&page->mapping->private_lock);
5612 return;
5613 }
5614
5615 if (fs_info->sectorsize == PAGE_SIZE) {
5616 /*
5617 * We do this since we'll remove the pages after we've
5618 * removed the eb from the radix tree, so we could race
5619 * and have this page now attached to the new eb. So
5620 * only clear page_private if it's still connected to
5621 * this eb.
5622 */
5623 if (PagePrivate(page) &&
5624 page->private == (unsigned long)eb) {
5625 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
5626 BUG_ON(PageDirty(page));
5627 BUG_ON(PageWriteback(page));
5628 /*
5629 * We need to make sure we haven't be attached
5630 * to a new eb.
5631 */
5632 detach_page_private(page);
5633 }
5634 if (mapped)
5635 spin_unlock(&page->mapping->private_lock);
5636 return;
5637 }
5638
5639 /*
5640 * For subpage, we can have dummy eb with page private. In this case,
5641 * we can directly detach the private as such page is only attached to
5642 * one dummy eb, no sharing.
5643 */
5644 if (!mapped) {
5645 btrfs_detach_subpage(fs_info, page);
5646 return;
5647 }
5648
5649 btrfs_page_dec_eb_refs(fs_info, page);
5650
5651 /*
5652 * We can only detach the page private if there are no other ebs in the
5653 * page range and no unfinished IO.
5654 */
5655 if (!page_range_has_eb(fs_info, page))
5656 btrfs_detach_subpage(fs_info, page);
5657
5658 spin_unlock(&page->mapping->private_lock);
5659}
5660
5661/* Release all pages attached to the extent buffer */
5662static void btrfs_release_extent_buffer_pages(struct extent_buffer *eb)
5663{
5664 int i;
5665 int num_pages;
5666
5667 ASSERT(!extent_buffer_under_io(eb));
5668
5669 num_pages = num_extent_pages(eb);
5670 for (i = 0; i < num_pages; i++) {
5671 struct page *page = eb->pages[i];
5672
5673 if (!page)
5674 continue;
5675
5676 detach_extent_buffer_page(eb, page);
5677
5678 /* One for when we allocated the page */
5679 put_page(page);
5680 }
5681}
5682
5683/*
5684 * Helper for releasing the extent buffer.
5685 */
5686static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
5687{
5688 btrfs_release_extent_buffer_pages(eb);
5689 btrfs_leak_debug_del(&eb->fs_info->eb_leak_lock, &eb->leak_list);
5690 __free_extent_buffer(eb);
5691}
5692
5693static struct extent_buffer *
5694__alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
5695 unsigned long len)
5696{
5697 struct extent_buffer *eb = NULL;
5698
5699 eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
5700 eb->start = start;
5701 eb->len = len;
5702 eb->fs_info = fs_info;
5703 eb->bflags = 0;
5704 init_rwsem(&eb->lock);
5705
5706 btrfs_leak_debug_add(&fs_info->eb_leak_lock, &eb->leak_list,
5707 &fs_info->allocated_ebs);
5708 INIT_LIST_HEAD(&eb->release_list);
5709
5710 spin_lock_init(&eb->refs_lock);
5711 atomic_set(&eb->refs, 1);
5712 atomic_set(&eb->io_pages, 0);
5713
5714 ASSERT(len <= BTRFS_MAX_METADATA_BLOCKSIZE);
5715
5716 return eb;
5717}
5718
5719struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
5720{
5721 int i;
5722 struct page *p;
5723 struct extent_buffer *new;
5724 int num_pages = num_extent_pages(src);
5725
5726 new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
5727 if (new == NULL)
5728 return NULL;
5729
5730 /*
5731 * Set UNMAPPED before calling btrfs_release_extent_buffer(), as
5732 * btrfs_release_extent_buffer() have different behavior for
5733 * UNMAPPED subpage extent buffer.
5734 */
5735 set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
5736
5737 for (i = 0; i < num_pages; i++) {
5738 int ret;
5739
5740 p = alloc_page(GFP_NOFS);
5741 if (!p) {
5742 btrfs_release_extent_buffer(new);
5743 return NULL;
5744 }
5745 ret = attach_extent_buffer_page(new, p, NULL);
5746 if (ret < 0) {
5747 put_page(p);
5748 btrfs_release_extent_buffer(new);
5749 return NULL;
5750 }
5751 WARN_ON(PageDirty(p));
5752 new->pages[i] = p;
5753 copy_page(page_address(p), page_address(src->pages[i]));
5754 }
5755 set_extent_buffer_uptodate(new);
5756
5757 return new;
5758}
5759
5760struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
5761 u64 start, unsigned long len)
5762{
5763 struct extent_buffer *eb;
5764 int num_pages;
5765 int i;
5766
5767 eb = __alloc_extent_buffer(fs_info, start, len);
5768 if (!eb)
5769 return NULL;
5770
5771 num_pages = num_extent_pages(eb);
5772 for (i = 0; i < num_pages; i++) {
5773 int ret;
5774
5775 eb->pages[i] = alloc_page(GFP_NOFS);
5776 if (!eb->pages[i])
5777 goto err;
5778 ret = attach_extent_buffer_page(eb, eb->pages[i], NULL);
5779 if (ret < 0)
5780 goto err;
5781 }
5782 set_extent_buffer_uptodate(eb);
5783 btrfs_set_header_nritems(eb, 0);
5784 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
5785
5786 return eb;
5787err:
5788 for (; i > 0; i--) {
5789 detach_extent_buffer_page(eb, eb->pages[i - 1]);
5790 __free_page(eb->pages[i - 1]);
5791 }
5792 __free_extent_buffer(eb);
5793 return NULL;
5794}
5795
5796struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
5797 u64 start)
5798{
5799 return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
5800}
5801
5802static void check_buffer_tree_ref(struct extent_buffer *eb)
5803{
5804 int refs;
5805 /*
5806 * The TREE_REF bit is first set when the extent_buffer is added
5807 * to the radix tree. It is also reset, if unset, when a new reference
5808 * is created by find_extent_buffer.
5809 *
5810 * It is only cleared in two cases: freeing the last non-tree
5811 * reference to the extent_buffer when its STALE bit is set or
5812 * calling releasepage when the tree reference is the only reference.
5813 *
5814 * In both cases, care is taken to ensure that the extent_buffer's
5815 * pages are not under io. However, releasepage can be concurrently
5816 * called with creating new references, which is prone to race
5817 * conditions between the calls to check_buffer_tree_ref in those
5818 * codepaths and clearing TREE_REF in try_release_extent_buffer.
5819 *
5820 * The actual lifetime of the extent_buffer in the radix tree is
5821 * adequately protected by the refcount, but the TREE_REF bit and
5822 * its corresponding reference are not. To protect against this
5823 * class of races, we call check_buffer_tree_ref from the codepaths
5824 * which trigger io after they set eb->io_pages. Note that once io is
5825 * initiated, TREE_REF can no longer be cleared, so that is the
5826 * moment at which any such race is best fixed.
5827 */
5828 refs = atomic_read(&eb->refs);
5829 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5830 return;
5831
5832 spin_lock(&eb->refs_lock);
5833 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5834 atomic_inc(&eb->refs);
5835 spin_unlock(&eb->refs_lock);
5836}
5837
5838static void mark_extent_buffer_accessed(struct extent_buffer *eb,
5839 struct page *accessed)
5840{
5841 int num_pages, i;
5842
5843 check_buffer_tree_ref(eb);
5844
5845 num_pages = num_extent_pages(eb);
5846 for (i = 0; i < num_pages; i++) {
5847 struct page *p = eb->pages[i];
5848
5849 if (p != accessed)
5850 mark_page_accessed(p);
5851 }
5852}
5853
5854struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
5855 u64 start)
5856{
5857 struct extent_buffer *eb;
5858
5859 eb = find_extent_buffer_nolock(fs_info, start);
5860 if (!eb)
5861 return NULL;
5862 /*
5863 * Lock our eb's refs_lock to avoid races with free_extent_buffer().
5864 * When we get our eb it might be flagged with EXTENT_BUFFER_STALE and
5865 * another task running free_extent_buffer() might have seen that flag
5866 * set, eb->refs == 2, that the buffer isn't under IO (dirty and
5867 * writeback flags not set) and it's still in the tree (flag
5868 * EXTENT_BUFFER_TREE_REF set), therefore being in the process of
5869 * decrementing the extent buffer's reference count twice. So here we
5870 * could race and increment the eb's reference count, clear its stale
5871 * flag, mark it as dirty and drop our reference before the other task
5872 * finishes executing free_extent_buffer, which would later result in
5873 * an attempt to free an extent buffer that is dirty.
5874 */
5875 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
5876 spin_lock(&eb->refs_lock);
5877 spin_unlock(&eb->refs_lock);
5878 }
5879 mark_extent_buffer_accessed(eb, NULL);
5880 return eb;
5881}
5882
5883#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5884struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
5885 u64 start)
5886{
5887 struct extent_buffer *eb, *exists = NULL;
5888 int ret;
5889
5890 eb = find_extent_buffer(fs_info, start);
5891 if (eb)
5892 return eb;
5893 eb = alloc_dummy_extent_buffer(fs_info, start);
5894 if (!eb)
5895 return ERR_PTR(-ENOMEM);
5896 eb->fs_info = fs_info;
5897again:
5898 ret = radix_tree_preload(GFP_NOFS);
5899 if (ret) {
5900 exists = ERR_PTR(ret);
5901 goto free_eb;
5902 }
5903 spin_lock(&fs_info->buffer_lock);
5904 ret = radix_tree_insert(&fs_info->buffer_radix,
5905 start >> fs_info->sectorsize_bits, eb);
5906 spin_unlock(&fs_info->buffer_lock);
5907 radix_tree_preload_end();
5908 if (ret == -EEXIST) {
5909 exists = find_extent_buffer(fs_info, start);
5910 if (exists)
5911 goto free_eb;
5912 else
5913 goto again;
5914 }
5915 check_buffer_tree_ref(eb);
5916 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5917
5918 return eb;
5919free_eb:
5920 btrfs_release_extent_buffer(eb);
5921 return exists;
5922}
5923#endif
5924
5925static struct extent_buffer *grab_extent_buffer(
5926 struct btrfs_fs_info *fs_info, struct page *page)
5927{
5928 struct extent_buffer *exists;
5929
5930 /*
5931 * For subpage case, we completely rely on radix tree to ensure we
5932 * don't try to insert two ebs for the same bytenr. So here we always
5933 * return NULL and just continue.
5934 */
5935 if (fs_info->sectorsize < PAGE_SIZE)
5936 return NULL;
5937
5938 /* Page not yet attached to an extent buffer */
5939 if (!PagePrivate(page))
5940 return NULL;
5941
5942 /*
5943 * We could have already allocated an eb for this page and attached one
5944 * so lets see if we can get a ref on the existing eb, and if we can we
5945 * know it's good and we can just return that one, else we know we can
5946 * just overwrite page->private.
5947 */
5948 exists = (struct extent_buffer *)page->private;
5949 if (atomic_inc_not_zero(&exists->refs))
5950 return exists;
5951
5952 WARN_ON(PageDirty(page));
5953 detach_page_private(page);
5954 return NULL;
5955}
5956
5957struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
5958 u64 start, u64 owner_root, int level)
5959{
5960 unsigned long len = fs_info->nodesize;
5961 int num_pages;
5962 int i;
5963 unsigned long index = start >> PAGE_SHIFT;
5964 struct extent_buffer *eb;
5965 struct extent_buffer *exists = NULL;
5966 struct page *p;
5967 struct address_space *mapping = fs_info->btree_inode->i_mapping;
5968 int uptodate = 1;
5969 int ret;
5970
5971 if (!IS_ALIGNED(start, fs_info->sectorsize)) {
5972 btrfs_err(fs_info, "bad tree block start %llu", start);
5973 return ERR_PTR(-EINVAL);
5974 }
5975
5976#if BITS_PER_LONG == 32
5977 if (start >= MAX_LFS_FILESIZE) {
5978 btrfs_err_rl(fs_info,
5979 "extent buffer %llu is beyond 32bit page cache limit", start);
5980 btrfs_err_32bit_limit(fs_info);
5981 return ERR_PTR(-EOVERFLOW);
5982 }
5983 if (start >= BTRFS_32BIT_EARLY_WARN_THRESHOLD)
5984 btrfs_warn_32bit_limit(fs_info);
5985#endif
5986
5987 if (fs_info->sectorsize < PAGE_SIZE &&
5988 offset_in_page(start) + len > PAGE_SIZE) {
5989 btrfs_err(fs_info,
5990 "tree block crosses page boundary, start %llu nodesize %lu",
5991 start, len);
5992 return ERR_PTR(-EINVAL);
5993 }
5994
5995 eb = find_extent_buffer(fs_info, start);
5996 if (eb)
5997 return eb;
5998
5999 eb = __alloc_extent_buffer(fs_info, start, len);
6000 if (!eb)
6001 return ERR_PTR(-ENOMEM);
6002 btrfs_set_buffer_lockdep_class(owner_root, eb, level);
6003
6004 num_pages = num_extent_pages(eb);
6005 for (i = 0; i < num_pages; i++, index++) {
6006 struct btrfs_subpage *prealloc = NULL;
6007
6008 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
6009 if (!p) {
6010 exists = ERR_PTR(-ENOMEM);
6011 goto free_eb;
6012 }
6013
6014 /*
6015 * Preallocate page->private for subpage case, so that we won't
6016 * allocate memory with private_lock hold. The memory will be
6017 * freed by attach_extent_buffer_page() or freed manually if
6018 * we exit earlier.
6019 *
6020 * Although we have ensured one subpage eb can only have one
6021 * page, but it may change in the future for 16K page size
6022 * support, so we still preallocate the memory in the loop.
6023 */
6024 ret = btrfs_alloc_subpage(fs_info, &prealloc,
6025 BTRFS_SUBPAGE_METADATA);
6026 if (ret < 0) {
6027 unlock_page(p);
6028 put_page(p);
6029 exists = ERR_PTR(ret);
6030 goto free_eb;
6031 }
6032
6033 spin_lock(&mapping->private_lock);
6034 exists = grab_extent_buffer(fs_info, p);
6035 if (exists) {
6036 spin_unlock(&mapping->private_lock);
6037 unlock_page(p);
6038 put_page(p);
6039 mark_extent_buffer_accessed(exists, p);
6040 btrfs_free_subpage(prealloc);
6041 goto free_eb;
6042 }
6043 /* Should not fail, as we have preallocated the memory */
6044 ret = attach_extent_buffer_page(eb, p, prealloc);
6045 ASSERT(!ret);
6046 /*
6047 * To inform we have extra eb under allocation, so that
6048 * detach_extent_buffer_page() won't release the page private
6049 * when the eb hasn't yet been inserted into radix tree.
6050 *
6051 * The ref will be decreased when the eb released the page, in
6052 * detach_extent_buffer_page().
6053 * Thus needs no special handling in error path.
6054 */
6055 btrfs_page_inc_eb_refs(fs_info, p);
6056 spin_unlock(&mapping->private_lock);
6057
6058 WARN_ON(btrfs_page_test_dirty(fs_info, p, eb->start, eb->len));
6059 eb->pages[i] = p;
6060 if (!PageUptodate(p))
6061 uptodate = 0;
6062
6063 /*
6064 * We can't unlock the pages just yet since the extent buffer
6065 * hasn't been properly inserted in the radix tree, this
6066 * opens a race with btree_releasepage which can free a page
6067 * while we are still filling in all pages for the buffer and
6068 * we could crash.
6069 */
6070 }
6071 if (uptodate)
6072 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
6073again:
6074 ret = radix_tree_preload(GFP_NOFS);
6075 if (ret) {
6076 exists = ERR_PTR(ret);
6077 goto free_eb;
6078 }
6079
6080 spin_lock(&fs_info->buffer_lock);
6081 ret = radix_tree_insert(&fs_info->buffer_radix,
6082 start >> fs_info->sectorsize_bits, eb);
6083 spin_unlock(&fs_info->buffer_lock);
6084 radix_tree_preload_end();
6085 if (ret == -EEXIST) {
6086 exists = find_extent_buffer(fs_info, start);
6087 if (exists)
6088 goto free_eb;
6089 else
6090 goto again;
6091 }
6092 /* add one reference for the tree */
6093 check_buffer_tree_ref(eb);
6094 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
6095
6096 /*
6097 * Now it's safe to unlock the pages because any calls to
6098 * btree_releasepage will correctly detect that a page belongs to a
6099 * live buffer and won't free them prematurely.
6100 */
6101 for (i = 0; i < num_pages; i++)
6102 unlock_page(eb->pages[i]);
6103 return eb;
6104
6105free_eb:
6106 WARN_ON(!atomic_dec_and_test(&eb->refs));
6107 for (i = 0; i < num_pages; i++) {
6108 if (eb->pages[i])
6109 unlock_page(eb->pages[i]);
6110 }
6111
6112 btrfs_release_extent_buffer(eb);
6113 return exists;
6114}
6115
6116static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
6117{
6118 struct extent_buffer *eb =
6119 container_of(head, struct extent_buffer, rcu_head);
6120
6121 __free_extent_buffer(eb);
6122}
6123
6124static int release_extent_buffer(struct extent_buffer *eb)
6125 __releases(&eb->refs_lock)
6126{
6127 lockdep_assert_held(&eb->refs_lock);
6128
6129 WARN_ON(atomic_read(&eb->refs) == 0);
6130 if (atomic_dec_and_test(&eb->refs)) {
6131 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
6132 struct btrfs_fs_info *fs_info = eb->fs_info;
6133
6134 spin_unlock(&eb->refs_lock);
6135
6136 spin_lock(&fs_info->buffer_lock);
6137 radix_tree_delete(&fs_info->buffer_radix,
6138 eb->start >> fs_info->sectorsize_bits);
6139 spin_unlock(&fs_info->buffer_lock);
6140 } else {
6141 spin_unlock(&eb->refs_lock);
6142 }
6143
6144 btrfs_leak_debug_del(&eb->fs_info->eb_leak_lock, &eb->leak_list);
6145 /* Should be safe to release our pages at this point */
6146 btrfs_release_extent_buffer_pages(eb);
6147#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
6148 if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
6149 __free_extent_buffer(eb);
6150 return 1;
6151 }
6152#endif
6153 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
6154 return 1;
6155 }
6156 spin_unlock(&eb->refs_lock);
6157
6158 return 0;
6159}
6160
6161void free_extent_buffer(struct extent_buffer *eb)
6162{
6163 int refs;
6164 int old;
6165 if (!eb)
6166 return;
6167
6168 while (1) {
6169 refs = atomic_read(&eb->refs);
6170 if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
6171 || (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
6172 refs == 1))
6173 break;
6174 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
6175 if (old == refs)
6176 return;
6177 }
6178
6179 spin_lock(&eb->refs_lock);
6180 if (atomic_read(&eb->refs) == 2 &&
6181 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
6182 !extent_buffer_under_io(eb) &&
6183 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
6184 atomic_dec(&eb->refs);
6185
6186 /*
6187 * I know this is terrible, but it's temporary until we stop tracking
6188 * the uptodate bits and such for the extent buffers.
6189 */
6190 release_extent_buffer(eb);
6191}
6192
6193void free_extent_buffer_stale(struct extent_buffer *eb)
6194{
6195 if (!eb)
6196 return;
6197
6198 spin_lock(&eb->refs_lock);
6199 set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
6200
6201 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
6202 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
6203 atomic_dec(&eb->refs);
6204 release_extent_buffer(eb);
6205}
6206
6207static void btree_clear_page_dirty(struct page *page)
6208{
6209 ASSERT(PageDirty(page));
6210 ASSERT(PageLocked(page));
6211 clear_page_dirty_for_io(page);
6212 xa_lock_irq(&page->mapping->i_pages);
6213 if (!PageDirty(page))
6214 __xa_clear_mark(&page->mapping->i_pages,
6215 page_index(page), PAGECACHE_TAG_DIRTY);
6216 xa_unlock_irq(&page->mapping->i_pages);
6217}
6218
6219static void clear_subpage_extent_buffer_dirty(const struct extent_buffer *eb)
6220{
6221 struct btrfs_fs_info *fs_info = eb->fs_info;
6222 struct page *page = eb->pages[0];
6223 bool last;
6224
6225 /* btree_clear_page_dirty() needs page locked */
6226 lock_page(page);
6227 last = btrfs_subpage_clear_and_test_dirty(fs_info, page, eb->start,
6228 eb->len);
6229 if (last)
6230 btree_clear_page_dirty(page);
6231 unlock_page(page);
6232 WARN_ON(atomic_read(&eb->refs) == 0);
6233}
6234
6235void clear_extent_buffer_dirty(const struct extent_buffer *eb)
6236{
6237 int i;
6238 int num_pages;
6239 struct page *page;
6240
6241 if (eb->fs_info->sectorsize < PAGE_SIZE)
6242 return clear_subpage_extent_buffer_dirty(eb);
6243
6244 num_pages = num_extent_pages(eb);
6245
6246 for (i = 0; i < num_pages; i++) {
6247 page = eb->pages[i];
6248 if (!PageDirty(page))
6249 continue;
6250 lock_page(page);
6251 btree_clear_page_dirty(page);
6252 ClearPageError(page);
6253 unlock_page(page);
6254 }
6255 WARN_ON(atomic_read(&eb->refs) == 0);
6256}
6257
6258bool set_extent_buffer_dirty(struct extent_buffer *eb)
6259{
6260 int i;
6261 int num_pages;
6262 bool was_dirty;
6263
6264 check_buffer_tree_ref(eb);
6265
6266 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
6267
6268 num_pages = num_extent_pages(eb);
6269 WARN_ON(atomic_read(&eb->refs) == 0);
6270 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
6271
6272 if (!was_dirty) {
6273 bool subpage = eb->fs_info->sectorsize < PAGE_SIZE;
6274
6275 /*
6276 * For subpage case, we can have other extent buffers in the
6277 * same page, and in clear_subpage_extent_buffer_dirty() we
6278 * have to clear page dirty without subpage lock held.
6279 * This can cause race where our page gets dirty cleared after
6280 * we just set it.
6281 *
6282 * Thankfully, clear_subpage_extent_buffer_dirty() has locked
6283 * its page for other reasons, we can use page lock to prevent
6284 * the above race.
6285 */
6286 if (subpage)
6287 lock_page(eb->pages[0]);
6288 for (i = 0; i < num_pages; i++)
6289 btrfs_page_set_dirty(eb->fs_info, eb->pages[i],
6290 eb->start, eb->len);
6291 if (subpage)
6292 unlock_page(eb->pages[0]);
6293 }
6294#ifdef CONFIG_BTRFS_DEBUG
6295 for (i = 0; i < num_pages; i++)
6296 ASSERT(PageDirty(eb->pages[i]));
6297#endif
6298
6299 return was_dirty;
6300}
6301
6302void clear_extent_buffer_uptodate(struct extent_buffer *eb)
6303{
6304 struct btrfs_fs_info *fs_info = eb->fs_info;
6305 struct page *page;
6306 int num_pages;
6307 int i;
6308
6309 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
6310 num_pages = num_extent_pages(eb);
6311 for (i = 0; i < num_pages; i++) {
6312 page = eb->pages[i];
6313 if (page)
6314 btrfs_page_clear_uptodate(fs_info, page,
6315 eb->start, eb->len);
6316 }
6317}
6318
6319void set_extent_buffer_uptodate(struct extent_buffer *eb)
6320{
6321 struct btrfs_fs_info *fs_info = eb->fs_info;
6322 struct page *page;
6323 int num_pages;
6324 int i;
6325
6326 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
6327 num_pages = num_extent_pages(eb);
6328 for (i = 0; i < num_pages; i++) {
6329 page = eb->pages[i];
6330 btrfs_page_set_uptodate(fs_info, page, eb->start, eb->len);
6331 }
6332}
6333
6334static int read_extent_buffer_subpage(struct extent_buffer *eb, int wait,
6335 int mirror_num)
6336{
6337 struct btrfs_fs_info *fs_info = eb->fs_info;
6338 struct extent_io_tree *io_tree;
6339 struct page *page = eb->pages[0];
6340 struct btrfs_bio_ctrl bio_ctrl = { 0 };
6341 int ret = 0;
6342
6343 ASSERT(!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags));
6344 ASSERT(PagePrivate(page));
6345 io_tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
6346
6347 if (wait == WAIT_NONE) {
6348 if (!try_lock_extent(io_tree, eb->start, eb->start + eb->len - 1))
6349 return -EAGAIN;
6350 } else {
6351 ret = lock_extent(io_tree, eb->start, eb->start + eb->len - 1);
6352 if (ret < 0)
6353 return ret;
6354 }
6355
6356 ret = 0;
6357 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags) ||
6358 PageUptodate(page) ||
6359 btrfs_subpage_test_uptodate(fs_info, page, eb->start, eb->len)) {
6360 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
6361 unlock_extent(io_tree, eb->start, eb->start + eb->len - 1);
6362 return ret;
6363 }
6364
6365 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
6366 eb->read_mirror = 0;
6367 atomic_set(&eb->io_pages, 1);
6368 check_buffer_tree_ref(eb);
6369 btrfs_subpage_clear_error(fs_info, page, eb->start, eb->len);
6370
6371 btrfs_subpage_start_reader(fs_info, page, eb->start, eb->len);
6372 ret = submit_extent_page(REQ_OP_READ | REQ_META, NULL, &bio_ctrl,
6373 page, eb->start, eb->len,
6374 eb->start - page_offset(page),
6375 end_bio_extent_readpage, mirror_num, 0,
6376 true);
6377 if (ret) {
6378 /*
6379 * In the endio function, if we hit something wrong we will
6380 * increase the io_pages, so here we need to decrease it for
6381 * error path.
6382 */
6383 atomic_dec(&eb->io_pages);
6384 }
6385 if (bio_ctrl.bio) {
6386 int tmp;
6387
6388 tmp = submit_one_bio(bio_ctrl.bio, mirror_num, 0);
6389 bio_ctrl.bio = NULL;
6390 if (tmp < 0)
6391 return tmp;
6392 }
6393 if (ret || wait != WAIT_COMPLETE)
6394 return ret;
6395
6396 wait_extent_bit(io_tree, eb->start, eb->start + eb->len - 1, EXTENT_LOCKED);
6397 if (!test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
6398 ret = -EIO;
6399 return ret;
6400}
6401
6402int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num)
6403{
6404 int i;
6405 struct page *page;
6406 int err;
6407 int ret = 0;
6408 int locked_pages = 0;
6409 int all_uptodate = 1;
6410 int num_pages;
6411 unsigned long num_reads = 0;
6412 struct btrfs_bio_ctrl bio_ctrl = { 0 };
6413
6414 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
6415 return 0;
6416
6417 if (eb->fs_info->sectorsize < PAGE_SIZE)
6418 return read_extent_buffer_subpage(eb, wait, mirror_num);
6419
6420 num_pages = num_extent_pages(eb);
6421 for (i = 0; i < num_pages; i++) {
6422 page = eb->pages[i];
6423 if (wait == WAIT_NONE) {
6424 /*
6425 * WAIT_NONE is only utilized by readahead. If we can't
6426 * acquire the lock atomically it means either the eb
6427 * is being read out or under modification.
6428 * Either way the eb will be or has been cached,
6429 * readahead can exit safely.
6430 */
6431 if (!trylock_page(page))
6432 goto unlock_exit;
6433 } else {
6434 lock_page(page);
6435 }
6436 locked_pages++;
6437 }
6438 /*
6439 * We need to firstly lock all pages to make sure that
6440 * the uptodate bit of our pages won't be affected by
6441 * clear_extent_buffer_uptodate().
6442 */
6443 for (i = 0; i < num_pages; i++) {
6444 page = eb->pages[i];
6445 if (!PageUptodate(page)) {
6446 num_reads++;
6447 all_uptodate = 0;
6448 }
6449 }
6450
6451 if (all_uptodate) {
6452 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
6453 goto unlock_exit;
6454 }
6455
6456 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
6457 eb->read_mirror = 0;
6458 atomic_set(&eb->io_pages, num_reads);
6459 /*
6460 * It is possible for releasepage to clear the TREE_REF bit before we
6461 * set io_pages. See check_buffer_tree_ref for a more detailed comment.
6462 */
6463 check_buffer_tree_ref(eb);
6464 for (i = 0; i < num_pages; i++) {
6465 page = eb->pages[i];
6466
6467 if (!PageUptodate(page)) {
6468 if (ret) {
6469 atomic_dec(&eb->io_pages);
6470 unlock_page(page);
6471 continue;
6472 }
6473
6474 ClearPageError(page);
6475 err = submit_extent_page(REQ_OP_READ | REQ_META, NULL,
6476 &bio_ctrl, page, page_offset(page),
6477 PAGE_SIZE, 0, end_bio_extent_readpage,
6478 mirror_num, 0, false);
6479 if (err) {
6480 /*
6481 * We failed to submit the bio so it's the
6482 * caller's responsibility to perform cleanup
6483 * i.e unlock page/set error bit.
6484 */
6485 ret = err;
6486 SetPageError(page);
6487 unlock_page(page);
6488 atomic_dec(&eb->io_pages);
6489 }
6490 } else {
6491 unlock_page(page);
6492 }
6493 }
6494
6495 if (bio_ctrl.bio) {
6496 err = submit_one_bio(bio_ctrl.bio, mirror_num, bio_ctrl.bio_flags);
6497 bio_ctrl.bio = NULL;
6498 if (err)
6499 return err;
6500 }
6501
6502 if (ret || wait != WAIT_COMPLETE)
6503 return ret;
6504
6505 for (i = 0; i < num_pages; i++) {
6506 page = eb->pages[i];
6507 wait_on_page_locked(page);
6508 if (!PageUptodate(page))
6509 ret = -EIO;
6510 }
6511
6512 return ret;
6513
6514unlock_exit:
6515 while (locked_pages > 0) {
6516 locked_pages--;
6517 page = eb->pages[locked_pages];
6518 unlock_page(page);
6519 }
6520 return ret;
6521}
6522
6523static bool report_eb_range(const struct extent_buffer *eb, unsigned long start,
6524 unsigned long len)
6525{
6526 btrfs_warn(eb->fs_info,
6527 "access to eb bytenr %llu len %lu out of range start %lu len %lu",
6528 eb->start, eb->len, start, len);
6529 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
6530
6531 return true;
6532}
6533
6534/*
6535 * Check if the [start, start + len) range is valid before reading/writing
6536 * the eb.
6537 * NOTE: @start and @len are offset inside the eb, not logical address.
6538 *
6539 * Caller should not touch the dst/src memory if this function returns error.
6540 */
6541static inline int check_eb_range(const struct extent_buffer *eb,
6542 unsigned long start, unsigned long len)
6543{
6544 unsigned long offset;
6545
6546 /* start, start + len should not go beyond eb->len nor overflow */
6547 if (unlikely(check_add_overflow(start, len, &offset) || offset > eb->len))
6548 return report_eb_range(eb, start, len);
6549
6550 return false;
6551}
6552
6553void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
6554 unsigned long start, unsigned long len)
6555{
6556 size_t cur;
6557 size_t offset;
6558 struct page *page;
6559 char *kaddr;
6560 char *dst = (char *)dstv;
6561 unsigned long i = get_eb_page_index(start);
6562
6563 if (check_eb_range(eb, start, len))
6564 return;
6565
6566 offset = get_eb_offset_in_page(eb, start);
6567
6568 while (len > 0) {
6569 page = eb->pages[i];
6570
6571 cur = min(len, (PAGE_SIZE - offset));
6572 kaddr = page_address(page);
6573 memcpy(dst, kaddr + offset, cur);
6574
6575 dst += cur;
6576 len -= cur;
6577 offset = 0;
6578 i++;
6579 }
6580}
6581
6582int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
6583 void __user *dstv,
6584 unsigned long start, unsigned long len)
6585{
6586 size_t cur;
6587 size_t offset;
6588 struct page *page;
6589 char *kaddr;
6590 char __user *dst = (char __user *)dstv;
6591 unsigned long i = get_eb_page_index(start);
6592 int ret = 0;
6593
6594 WARN_ON(start > eb->len);
6595 WARN_ON(start + len > eb->start + eb->len);
6596
6597 offset = get_eb_offset_in_page(eb, start);
6598
6599 while (len > 0) {
6600 page = eb->pages[i];
6601
6602 cur = min(len, (PAGE_SIZE - offset));
6603 kaddr = page_address(page);
6604 if (copy_to_user_nofault(dst, kaddr + offset, cur)) {
6605 ret = -EFAULT;
6606 break;
6607 }
6608
6609 dst += cur;
6610 len -= cur;
6611 offset = 0;
6612 i++;
6613 }
6614
6615 return ret;
6616}
6617
6618int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
6619 unsigned long start, unsigned long len)
6620{
6621 size_t cur;
6622 size_t offset;
6623 struct page *page;
6624 char *kaddr;
6625 char *ptr = (char *)ptrv;
6626 unsigned long i = get_eb_page_index(start);
6627 int ret = 0;
6628
6629 if (check_eb_range(eb, start, len))
6630 return -EINVAL;
6631
6632 offset = get_eb_offset_in_page(eb, start);
6633
6634 while (len > 0) {
6635 page = eb->pages[i];
6636
6637 cur = min(len, (PAGE_SIZE - offset));
6638
6639 kaddr = page_address(page);
6640 ret = memcmp(ptr, kaddr + offset, cur);
6641 if (ret)
6642 break;
6643
6644 ptr += cur;
6645 len -= cur;
6646 offset = 0;
6647 i++;
6648 }
6649 return ret;
6650}
6651
6652/*
6653 * Check that the extent buffer is uptodate.
6654 *
6655 * For regular sector size == PAGE_SIZE case, check if @page is uptodate.
6656 * For subpage case, check if the range covered by the eb has EXTENT_UPTODATE.
6657 */
6658static void assert_eb_page_uptodate(const struct extent_buffer *eb,
6659 struct page *page)
6660{
6661 struct btrfs_fs_info *fs_info = eb->fs_info;
6662
6663 if (fs_info->sectorsize < PAGE_SIZE) {
6664 bool uptodate;
6665
6666 uptodate = btrfs_subpage_test_uptodate(fs_info, page,
6667 eb->start, eb->len);
6668 WARN_ON(!uptodate);
6669 } else {
6670 WARN_ON(!PageUptodate(page));
6671 }
6672}
6673
6674void write_extent_buffer_chunk_tree_uuid(const struct extent_buffer *eb,
6675 const void *srcv)
6676{
6677 char *kaddr;
6678
6679 assert_eb_page_uptodate(eb, eb->pages[0]);
6680 kaddr = page_address(eb->pages[0]) +
6681 get_eb_offset_in_page(eb, offsetof(struct btrfs_header,
6682 chunk_tree_uuid));
6683 memcpy(kaddr, srcv, BTRFS_FSID_SIZE);
6684}
6685
6686void write_extent_buffer_fsid(const struct extent_buffer *eb, const void *srcv)
6687{
6688 char *kaddr;
6689
6690 assert_eb_page_uptodate(eb, eb->pages[0]);
6691 kaddr = page_address(eb->pages[0]) +
6692 get_eb_offset_in_page(eb, offsetof(struct btrfs_header, fsid));
6693 memcpy(kaddr, srcv, BTRFS_FSID_SIZE);
6694}
6695
6696void write_extent_buffer(const struct extent_buffer *eb, const void *srcv,
6697 unsigned long start, unsigned long len)
6698{
6699 size_t cur;
6700 size_t offset;
6701 struct page *page;
6702 char *kaddr;
6703 char *src = (char *)srcv;
6704 unsigned long i = get_eb_page_index(start);
6705
6706 WARN_ON(test_bit(EXTENT_BUFFER_NO_CHECK, &eb->bflags));
6707
6708 if (check_eb_range(eb, start, len))
6709 return;
6710
6711 offset = get_eb_offset_in_page(eb, start);
6712
6713 while (len > 0) {
6714 page = eb->pages[i];
6715 assert_eb_page_uptodate(eb, page);
6716
6717 cur = min(len, PAGE_SIZE - offset);
6718 kaddr = page_address(page);
6719 memcpy(kaddr + offset, src, cur);
6720
6721 src += cur;
6722 len -= cur;
6723 offset = 0;
6724 i++;
6725 }
6726}
6727
6728void memzero_extent_buffer(const struct extent_buffer *eb, unsigned long start,
6729 unsigned long len)
6730{
6731 size_t cur;
6732 size_t offset;
6733 struct page *page;
6734 char *kaddr;
6735 unsigned long i = get_eb_page_index(start);
6736
6737 if (check_eb_range(eb, start, len))
6738 return;
6739
6740 offset = get_eb_offset_in_page(eb, start);
6741
6742 while (len > 0) {
6743 page = eb->pages[i];
6744 assert_eb_page_uptodate(eb, page);
6745
6746 cur = min(len, PAGE_SIZE - offset);
6747 kaddr = page_address(page);
6748 memset(kaddr + offset, 0, cur);
6749
6750 len -= cur;
6751 offset = 0;
6752 i++;
6753 }
6754}
6755
6756void copy_extent_buffer_full(const struct extent_buffer *dst,
6757 const struct extent_buffer *src)
6758{
6759 int i;
6760 int num_pages;
6761
6762 ASSERT(dst->len == src->len);
6763
6764 if (dst->fs_info->sectorsize == PAGE_SIZE) {
6765 num_pages = num_extent_pages(dst);
6766 for (i = 0; i < num_pages; i++)
6767 copy_page(page_address(dst->pages[i]),
6768 page_address(src->pages[i]));
6769 } else {
6770 size_t src_offset = get_eb_offset_in_page(src, 0);
6771 size_t dst_offset = get_eb_offset_in_page(dst, 0);
6772
6773 ASSERT(src->fs_info->sectorsize < PAGE_SIZE);
6774 memcpy(page_address(dst->pages[0]) + dst_offset,
6775 page_address(src->pages[0]) + src_offset,
6776 src->len);
6777 }
6778}
6779
6780void copy_extent_buffer(const struct extent_buffer *dst,
6781 const struct extent_buffer *src,
6782 unsigned long dst_offset, unsigned long src_offset,
6783 unsigned long len)
6784{
6785 u64 dst_len = dst->len;
6786 size_t cur;
6787 size_t offset;
6788 struct page *page;
6789 char *kaddr;
6790 unsigned long i = get_eb_page_index(dst_offset);
6791
6792 if (check_eb_range(dst, dst_offset, len) ||
6793 check_eb_range(src, src_offset, len))
6794 return;
6795
6796 WARN_ON(src->len != dst_len);
6797
6798 offset = get_eb_offset_in_page(dst, dst_offset);
6799
6800 while (len > 0) {
6801 page = dst->pages[i];
6802 assert_eb_page_uptodate(dst, page);
6803
6804 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
6805
6806 kaddr = page_address(page);
6807 read_extent_buffer(src, kaddr + offset, src_offset, cur);
6808
6809 src_offset += cur;
6810 len -= cur;
6811 offset = 0;
6812 i++;
6813 }
6814}
6815
6816/*
6817 * eb_bitmap_offset() - calculate the page and offset of the byte containing the
6818 * given bit number
6819 * @eb: the extent buffer
6820 * @start: offset of the bitmap item in the extent buffer
6821 * @nr: bit number
6822 * @page_index: return index of the page in the extent buffer that contains the
6823 * given bit number
6824 * @page_offset: return offset into the page given by page_index
6825 *
6826 * This helper hides the ugliness of finding the byte in an extent buffer which
6827 * contains a given bit.
6828 */
6829static inline void eb_bitmap_offset(const struct extent_buffer *eb,
6830 unsigned long start, unsigned long nr,
6831 unsigned long *page_index,
6832 size_t *page_offset)
6833{
6834 size_t byte_offset = BIT_BYTE(nr);
6835 size_t offset;
6836
6837 /*
6838 * The byte we want is the offset of the extent buffer + the offset of
6839 * the bitmap item in the extent buffer + the offset of the byte in the
6840 * bitmap item.
6841 */
6842 offset = start + offset_in_page(eb->start) + byte_offset;
6843
6844 *page_index = offset >> PAGE_SHIFT;
6845 *page_offset = offset_in_page(offset);
6846}
6847
6848/**
6849 * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
6850 * @eb: the extent buffer
6851 * @start: offset of the bitmap item in the extent buffer
6852 * @nr: bit number to test
6853 */
6854int extent_buffer_test_bit(const struct extent_buffer *eb, unsigned long start,
6855 unsigned long nr)
6856{
6857 u8 *kaddr;
6858 struct page *page;
6859 unsigned long i;
6860 size_t offset;
6861
6862 eb_bitmap_offset(eb, start, nr, &i, &offset);
6863 page = eb->pages[i];
6864 assert_eb_page_uptodate(eb, page);
6865 kaddr = page_address(page);
6866 return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
6867}
6868
6869/**
6870 * extent_buffer_bitmap_set - set an area of a bitmap
6871 * @eb: the extent buffer
6872 * @start: offset of the bitmap item in the extent buffer
6873 * @pos: bit number of the first bit
6874 * @len: number of bits to set
6875 */
6876void extent_buffer_bitmap_set(const struct extent_buffer *eb, unsigned long start,
6877 unsigned long pos, unsigned long len)
6878{
6879 u8 *kaddr;
6880 struct page *page;
6881 unsigned long i;
6882 size_t offset;
6883 const unsigned int size = pos + len;
6884 int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
6885 u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
6886
6887 eb_bitmap_offset(eb, start, pos, &i, &offset);
6888 page = eb->pages[i];
6889 assert_eb_page_uptodate(eb, page);
6890 kaddr = page_address(page);
6891
6892 while (len >= bits_to_set) {
6893 kaddr[offset] |= mask_to_set;
6894 len -= bits_to_set;
6895 bits_to_set = BITS_PER_BYTE;
6896 mask_to_set = ~0;
6897 if (++offset >= PAGE_SIZE && len > 0) {
6898 offset = 0;
6899 page = eb->pages[++i];
6900 assert_eb_page_uptodate(eb, page);
6901 kaddr = page_address(page);
6902 }
6903 }
6904 if (len) {
6905 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
6906 kaddr[offset] |= mask_to_set;
6907 }
6908}
6909
6910
6911/**
6912 * extent_buffer_bitmap_clear - clear an area of a bitmap
6913 * @eb: the extent buffer
6914 * @start: offset of the bitmap item in the extent buffer
6915 * @pos: bit number of the first bit
6916 * @len: number of bits to clear
6917 */
6918void extent_buffer_bitmap_clear(const struct extent_buffer *eb,
6919 unsigned long start, unsigned long pos,
6920 unsigned long len)
6921{
6922 u8 *kaddr;
6923 struct page *page;
6924 unsigned long i;
6925 size_t offset;
6926 const unsigned int size = pos + len;
6927 int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
6928 u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
6929
6930 eb_bitmap_offset(eb, start, pos, &i, &offset);
6931 page = eb->pages[i];
6932 assert_eb_page_uptodate(eb, page);
6933 kaddr = page_address(page);
6934
6935 while (len >= bits_to_clear) {
6936 kaddr[offset] &= ~mask_to_clear;
6937 len -= bits_to_clear;
6938 bits_to_clear = BITS_PER_BYTE;
6939 mask_to_clear = ~0;
6940 if (++offset >= PAGE_SIZE && len > 0) {
6941 offset = 0;
6942 page = eb->pages[++i];
6943 assert_eb_page_uptodate(eb, page);
6944 kaddr = page_address(page);
6945 }
6946 }
6947 if (len) {
6948 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
6949 kaddr[offset] &= ~mask_to_clear;
6950 }
6951}
6952
6953static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
6954{
6955 unsigned long distance = (src > dst) ? src - dst : dst - src;
6956 return distance < len;
6957}
6958
6959static void copy_pages(struct page *dst_page, struct page *src_page,
6960 unsigned long dst_off, unsigned long src_off,
6961 unsigned long len)
6962{
6963 char *dst_kaddr = page_address(dst_page);
6964 char *src_kaddr;
6965 int must_memmove = 0;
6966
6967 if (dst_page != src_page) {
6968 src_kaddr = page_address(src_page);
6969 } else {
6970 src_kaddr = dst_kaddr;
6971 if (areas_overlap(src_off, dst_off, len))
6972 must_memmove = 1;
6973 }
6974
6975 if (must_memmove)
6976 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
6977 else
6978 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
6979}
6980
6981void memcpy_extent_buffer(const struct extent_buffer *dst,
6982 unsigned long dst_offset, unsigned long src_offset,
6983 unsigned long len)
6984{
6985 size_t cur;
6986 size_t dst_off_in_page;
6987 size_t src_off_in_page;
6988 unsigned long dst_i;
6989 unsigned long src_i;
6990
6991 if (check_eb_range(dst, dst_offset, len) ||
6992 check_eb_range(dst, src_offset, len))
6993 return;
6994
6995 while (len > 0) {
6996 dst_off_in_page = get_eb_offset_in_page(dst, dst_offset);
6997 src_off_in_page = get_eb_offset_in_page(dst, src_offset);
6998
6999 dst_i = get_eb_page_index(dst_offset);
7000 src_i = get_eb_page_index(src_offset);
7001
7002 cur = min(len, (unsigned long)(PAGE_SIZE -
7003 src_off_in_page));
7004 cur = min_t(unsigned long, cur,
7005 (unsigned long)(PAGE_SIZE - dst_off_in_page));
7006
7007 copy_pages(dst->pages[dst_i], dst->pages[src_i],
7008 dst_off_in_page, src_off_in_page, cur);
7009
7010 src_offset += cur;
7011 dst_offset += cur;
7012 len -= cur;
7013 }
7014}
7015
7016void memmove_extent_buffer(const struct extent_buffer *dst,
7017 unsigned long dst_offset, unsigned long src_offset,
7018 unsigned long len)
7019{
7020 size_t cur;
7021 size_t dst_off_in_page;
7022 size_t src_off_in_page;
7023 unsigned long dst_end = dst_offset + len - 1;
7024 unsigned long src_end = src_offset + len - 1;
7025 unsigned long dst_i;
7026 unsigned long src_i;
7027
7028 if (check_eb_range(dst, dst_offset, len) ||
7029 check_eb_range(dst, src_offset, len))
7030 return;
7031 if (dst_offset < src_offset) {
7032 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
7033 return;
7034 }
7035 while (len > 0) {
7036 dst_i = get_eb_page_index(dst_end);
7037 src_i = get_eb_page_index(src_end);
7038
7039 dst_off_in_page = get_eb_offset_in_page(dst, dst_end);
7040 src_off_in_page = get_eb_offset_in_page(dst, src_end);
7041
7042 cur = min_t(unsigned long, len, src_off_in_page + 1);
7043 cur = min(cur, dst_off_in_page + 1);
7044 copy_pages(dst->pages[dst_i], dst->pages[src_i],
7045 dst_off_in_page - cur + 1,
7046 src_off_in_page - cur + 1, cur);
7047
7048 dst_end -= cur;
7049 src_end -= cur;
7050 len -= cur;
7051 }
7052}
7053
7054static struct extent_buffer *get_next_extent_buffer(
7055 struct btrfs_fs_info *fs_info, struct page *page, u64 bytenr)
7056{
7057 struct extent_buffer *gang[BTRFS_SUBPAGE_BITMAP_SIZE];
7058 struct extent_buffer *found = NULL;
7059 u64 page_start = page_offset(page);
7060 int ret;
7061 int i;
7062
7063 ASSERT(in_range(bytenr, page_start, PAGE_SIZE));
7064 ASSERT(PAGE_SIZE / fs_info->nodesize <= BTRFS_SUBPAGE_BITMAP_SIZE);
7065 lockdep_assert_held(&fs_info->buffer_lock);
7066
7067 ret = radix_tree_gang_lookup(&fs_info->buffer_radix, (void **)gang,
7068 bytenr >> fs_info->sectorsize_bits,
7069 PAGE_SIZE / fs_info->nodesize);
7070 for (i = 0; i < ret; i++) {
7071 /* Already beyond page end */
7072 if (gang[i]->start >= page_start + PAGE_SIZE)
7073 break;
7074 /* Found one */
7075 if (gang[i]->start >= bytenr) {
7076 found = gang[i];
7077 break;
7078 }
7079 }
7080 return found;
7081}
7082
7083static int try_release_subpage_extent_buffer(struct page *page)
7084{
7085 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
7086 u64 cur = page_offset(page);
7087 const u64 end = page_offset(page) + PAGE_SIZE;
7088 int ret;
7089
7090 while (cur < end) {
7091 struct extent_buffer *eb = NULL;
7092
7093 /*
7094 * Unlike try_release_extent_buffer() which uses page->private
7095 * to grab buffer, for subpage case we rely on radix tree, thus
7096 * we need to ensure radix tree consistency.
7097 *
7098 * We also want an atomic snapshot of the radix tree, thus go
7099 * with spinlock rather than RCU.
7100 */
7101 spin_lock(&fs_info->buffer_lock);
7102 eb = get_next_extent_buffer(fs_info, page, cur);
7103 if (!eb) {
7104 /* No more eb in the page range after or at cur */
7105 spin_unlock(&fs_info->buffer_lock);
7106 break;
7107 }
7108 cur = eb->start + eb->len;
7109
7110 /*
7111 * The same as try_release_extent_buffer(), to ensure the eb
7112 * won't disappear out from under us.
7113 */
7114 spin_lock(&eb->refs_lock);
7115 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
7116 spin_unlock(&eb->refs_lock);
7117 spin_unlock(&fs_info->buffer_lock);
7118 break;
7119 }
7120 spin_unlock(&fs_info->buffer_lock);
7121
7122 /*
7123 * If tree ref isn't set then we know the ref on this eb is a
7124 * real ref, so just return, this eb will likely be freed soon
7125 * anyway.
7126 */
7127 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
7128 spin_unlock(&eb->refs_lock);
7129 break;
7130 }
7131
7132 /*
7133 * Here we don't care about the return value, we will always
7134 * check the page private at the end. And
7135 * release_extent_buffer() will release the refs_lock.
7136 */
7137 release_extent_buffer(eb);
7138 }
7139 /*
7140 * Finally to check if we have cleared page private, as if we have
7141 * released all ebs in the page, the page private should be cleared now.
7142 */
7143 spin_lock(&page->mapping->private_lock);
7144 if (!PagePrivate(page))
7145 ret = 1;
7146 else
7147 ret = 0;
7148 spin_unlock(&page->mapping->private_lock);
7149 return ret;
7150
7151}
7152
7153int try_release_extent_buffer(struct page *page)
7154{
7155 struct extent_buffer *eb;
7156
7157 if (btrfs_sb(page->mapping->host->i_sb)->sectorsize < PAGE_SIZE)
7158 return try_release_subpage_extent_buffer(page);
7159
7160 /*
7161 * We need to make sure nobody is changing page->private, as we rely on
7162 * page->private as the pointer to extent buffer.
7163 */
7164 spin_lock(&page->mapping->private_lock);
7165 if (!PagePrivate(page)) {
7166 spin_unlock(&page->mapping->private_lock);
7167 return 1;
7168 }
7169
7170 eb = (struct extent_buffer *)page->private;
7171 BUG_ON(!eb);
7172
7173 /*
7174 * This is a little awful but should be ok, we need to make sure that
7175 * the eb doesn't disappear out from under us while we're looking at
7176 * this page.
7177 */
7178 spin_lock(&eb->refs_lock);
7179 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
7180 spin_unlock(&eb->refs_lock);
7181 spin_unlock(&page->mapping->private_lock);
7182 return 0;
7183 }
7184 spin_unlock(&page->mapping->private_lock);
7185
7186 /*
7187 * If tree ref isn't set then we know the ref on this eb is a real ref,
7188 * so just return, this page will likely be freed soon anyway.
7189 */
7190 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
7191 spin_unlock(&eb->refs_lock);
7192 return 0;
7193 }
7194
7195 return release_extent_buffer(eb);
7196}
7197
7198/*
7199 * btrfs_readahead_tree_block - attempt to readahead a child block
7200 * @fs_info: the fs_info
7201 * @bytenr: bytenr to read
7202 * @owner_root: objectid of the root that owns this eb
7203 * @gen: generation for the uptodate check, can be 0
7204 * @level: level for the eb
7205 *
7206 * Attempt to readahead a tree block at @bytenr. If @gen is 0 then we do a
7207 * normal uptodate check of the eb, without checking the generation. If we have
7208 * to read the block we will not block on anything.
7209 */
7210void btrfs_readahead_tree_block(struct btrfs_fs_info *fs_info,
7211 u64 bytenr, u64 owner_root, u64 gen, int level)
7212{
7213 struct extent_buffer *eb;
7214 int ret;
7215
7216 eb = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
7217 if (IS_ERR(eb))
7218 return;
7219
7220 if (btrfs_buffer_uptodate(eb, gen, 1)) {
7221 free_extent_buffer(eb);
7222 return;
7223 }
7224
7225 ret = read_extent_buffer_pages(eb, WAIT_NONE, 0);
7226 if (ret < 0)
7227 free_extent_buffer_stale(eb);
7228 else
7229 free_extent_buffer(eb);
7230}
7231
7232/*
7233 * btrfs_readahead_node_child - readahead a node's child block
7234 * @node: parent node we're reading from
7235 * @slot: slot in the parent node for the child we want to read
7236 *
7237 * A helper for btrfs_readahead_tree_block, we simply read the bytenr pointed at
7238 * the slot in the node provided.
7239 */
7240void btrfs_readahead_node_child(struct extent_buffer *node, int slot)
7241{
7242 btrfs_readahead_tree_block(node->fs_info,
7243 btrfs_node_blockptr(node, slot),
7244 btrfs_header_owner(node),
7245 btrfs_node_ptr_generation(node, slot),
7246 btrfs_header_level(node) - 1);
7247}