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/sched/mm.h>
10#include <linux/spinlock.h>
11#include <linux/blkdev.h>
12#include <linux/swap.h>
13#include <linux/writeback.h>
14#include <linux/pagevec.h>
15#include <linux/prefetch.h>
16#include <linux/fsverity.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 "bio.h"
23#include "locking.h"
24#include "backref.h"
25#include "disk-io.h"
26#include "subpage.h"
27#include "zoned.h"
28#include "block-group.h"
29#include "compression.h"
30#include "fs.h"
31#include "accessors.h"
32#include "file-item.h"
33#include "file.h"
34#include "dev-replace.h"
35#include "super.h"
36#include "transaction.h"
37
38static struct kmem_cache *extent_buffer_cache;
39
40#ifdef CONFIG_BTRFS_DEBUG
41static inline void btrfs_leak_debug_add_eb(struct extent_buffer *eb)
42{
43 struct btrfs_fs_info *fs_info = eb->fs_info;
44 unsigned long flags;
45
46 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
47 list_add(&eb->leak_list, &fs_info->allocated_ebs);
48 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
49}
50
51static inline void btrfs_leak_debug_del_eb(struct extent_buffer *eb)
52{
53 struct btrfs_fs_info *fs_info = eb->fs_info;
54 unsigned long flags;
55
56 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
57 list_del(&eb->leak_list);
58 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
59}
60
61void btrfs_extent_buffer_leak_debug_check(struct btrfs_fs_info *fs_info)
62{
63 struct extent_buffer *eb;
64 unsigned long flags;
65
66 /*
67 * If we didn't get into open_ctree our allocated_ebs will not be
68 * initialized, so just skip this.
69 */
70 if (!fs_info->allocated_ebs.next)
71 return;
72
73 WARN_ON(!list_empty(&fs_info->allocated_ebs));
74 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
75 while (!list_empty(&fs_info->allocated_ebs)) {
76 eb = list_first_entry(&fs_info->allocated_ebs,
77 struct extent_buffer, leak_list);
78 pr_err(
79 "BTRFS: buffer leak start %llu len %u refs %d bflags %lu owner %llu\n",
80 eb->start, eb->len, atomic_read(&eb->refs), eb->bflags,
81 btrfs_header_owner(eb));
82 list_del(&eb->leak_list);
83 WARN_ON_ONCE(1);
84 kmem_cache_free(extent_buffer_cache, eb);
85 }
86 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
87}
88#else
89#define btrfs_leak_debug_add_eb(eb) do {} while (0)
90#define btrfs_leak_debug_del_eb(eb) do {} while (0)
91#endif
92
93/*
94 * Structure to record info about the bio being assembled, and other info like
95 * how many bytes are there before stripe/ordered extent boundary.
96 */
97struct btrfs_bio_ctrl {
98 struct btrfs_bio *bbio;
99 enum btrfs_compression_type compress_type;
100 u32 len_to_oe_boundary;
101 blk_opf_t opf;
102 btrfs_bio_end_io_t end_io_func;
103 struct writeback_control *wbc;
104
105 /*
106 * The sectors of the page which are going to be submitted by
107 * extent_writepage_io().
108 * This is to avoid touching ranges covered by compression/inline.
109 */
110 unsigned long submit_bitmap;
111};
112
113static void submit_one_bio(struct btrfs_bio_ctrl *bio_ctrl)
114{
115 struct btrfs_bio *bbio = bio_ctrl->bbio;
116
117 if (!bbio)
118 return;
119
120 /* Caller should ensure the bio has at least some range added */
121 ASSERT(bbio->bio.bi_iter.bi_size);
122
123 if (btrfs_op(&bbio->bio) == BTRFS_MAP_READ &&
124 bio_ctrl->compress_type != BTRFS_COMPRESS_NONE)
125 btrfs_submit_compressed_read(bbio);
126 else
127 btrfs_submit_bbio(bbio, 0);
128
129 /* The bbio is owned by the end_io handler now */
130 bio_ctrl->bbio = NULL;
131}
132
133/*
134 * Submit or fail the current bio in the bio_ctrl structure.
135 */
136static void submit_write_bio(struct btrfs_bio_ctrl *bio_ctrl, int ret)
137{
138 struct btrfs_bio *bbio = bio_ctrl->bbio;
139
140 if (!bbio)
141 return;
142
143 if (ret) {
144 ASSERT(ret < 0);
145 btrfs_bio_end_io(bbio, errno_to_blk_status(ret));
146 /* The bio is owned by the end_io handler now */
147 bio_ctrl->bbio = NULL;
148 } else {
149 submit_one_bio(bio_ctrl);
150 }
151}
152
153int __init extent_buffer_init_cachep(void)
154{
155 extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
156 sizeof(struct extent_buffer), 0, 0,
157 NULL);
158 if (!extent_buffer_cache)
159 return -ENOMEM;
160
161 return 0;
162}
163
164void __cold extent_buffer_free_cachep(void)
165{
166 /*
167 * Make sure all delayed rcu free are flushed before we
168 * destroy caches.
169 */
170 rcu_barrier();
171 kmem_cache_destroy(extent_buffer_cache);
172}
173
174static void process_one_folio(struct btrfs_fs_info *fs_info,
175 struct folio *folio, const struct folio *locked_folio,
176 unsigned long page_ops, u64 start, u64 end)
177{
178 u32 len;
179
180 ASSERT(end + 1 - start != 0 && end + 1 - start < U32_MAX);
181 len = end + 1 - start;
182
183 if (page_ops & PAGE_SET_ORDERED)
184 btrfs_folio_clamp_set_ordered(fs_info, folio, start, len);
185 if (page_ops & PAGE_START_WRITEBACK) {
186 btrfs_folio_clamp_clear_dirty(fs_info, folio, start, len);
187 btrfs_folio_clamp_set_writeback(fs_info, folio, start, len);
188 }
189 if (page_ops & PAGE_END_WRITEBACK)
190 btrfs_folio_clamp_clear_writeback(fs_info, folio, start, len);
191
192 if (folio != locked_folio && (page_ops & PAGE_UNLOCK))
193 btrfs_folio_end_lock(fs_info, folio, start, len);
194}
195
196static void __process_folios_contig(struct address_space *mapping,
197 const struct folio *locked_folio, u64 start,
198 u64 end, unsigned long page_ops)
199{
200 struct btrfs_fs_info *fs_info = inode_to_fs_info(mapping->host);
201 pgoff_t start_index = start >> PAGE_SHIFT;
202 pgoff_t end_index = end >> PAGE_SHIFT;
203 pgoff_t index = start_index;
204 struct folio_batch fbatch;
205 int i;
206
207 folio_batch_init(&fbatch);
208 while (index <= end_index) {
209 int found_folios;
210
211 found_folios = filemap_get_folios_contig(mapping, &index,
212 end_index, &fbatch);
213 for (i = 0; i < found_folios; i++) {
214 struct folio *folio = fbatch.folios[i];
215
216 process_one_folio(fs_info, folio, locked_folio,
217 page_ops, start, end);
218 }
219 folio_batch_release(&fbatch);
220 cond_resched();
221 }
222}
223
224static noinline void __unlock_for_delalloc(const struct inode *inode,
225 const struct folio *locked_folio,
226 u64 start, u64 end)
227{
228 unsigned long index = start >> PAGE_SHIFT;
229 unsigned long end_index = end >> PAGE_SHIFT;
230
231 ASSERT(locked_folio);
232 if (index == locked_folio->index && end_index == index)
233 return;
234
235 __process_folios_contig(inode->i_mapping, locked_folio, start, end,
236 PAGE_UNLOCK);
237}
238
239static noinline int lock_delalloc_folios(struct inode *inode,
240 const struct folio *locked_folio,
241 u64 start, u64 end)
242{
243 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
244 struct address_space *mapping = inode->i_mapping;
245 pgoff_t start_index = start >> PAGE_SHIFT;
246 pgoff_t end_index = end >> PAGE_SHIFT;
247 pgoff_t index = start_index;
248 u64 processed_end = start;
249 struct folio_batch fbatch;
250
251 if (index == locked_folio->index && index == end_index)
252 return 0;
253
254 folio_batch_init(&fbatch);
255 while (index <= end_index) {
256 unsigned int found_folios, i;
257
258 found_folios = filemap_get_folios_contig(mapping, &index,
259 end_index, &fbatch);
260 if (found_folios == 0)
261 goto out;
262
263 for (i = 0; i < found_folios; i++) {
264 struct folio *folio = fbatch.folios[i];
265 u64 range_start;
266 u32 range_len;
267
268 if (folio == locked_folio)
269 continue;
270
271 folio_lock(folio);
272 if (!folio_test_dirty(folio) || folio->mapping != mapping) {
273 folio_unlock(folio);
274 goto out;
275 }
276 range_start = max_t(u64, folio_pos(folio), start);
277 range_len = min_t(u64, folio_pos(folio) + folio_size(folio),
278 end + 1) - range_start;
279 btrfs_folio_set_lock(fs_info, folio, range_start, range_len);
280
281 processed_end = range_start + range_len - 1;
282 }
283 folio_batch_release(&fbatch);
284 cond_resched();
285 }
286
287 return 0;
288out:
289 folio_batch_release(&fbatch);
290 if (processed_end > start)
291 __unlock_for_delalloc(inode, locked_folio, start,
292 processed_end);
293 return -EAGAIN;
294}
295
296/*
297 * Find and lock a contiguous range of bytes in the file marked as delalloc, no
298 * more than @max_bytes.
299 *
300 * @start: The original start bytenr to search.
301 * Will store the extent range start bytenr.
302 * @end: The original end bytenr of the search range
303 * Will store the extent range end bytenr.
304 *
305 * Return true if we find a delalloc range which starts inside the original
306 * range, and @start/@end will store the delalloc range start/end.
307 *
308 * Return false if we can't find any delalloc range which starts inside the
309 * original range, and @start/@end will be the non-delalloc range start/end.
310 */
311EXPORT_FOR_TESTS
312noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
313 struct folio *locked_folio,
314 u64 *start, u64 *end)
315{
316 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
317 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
318 const u64 orig_start = *start;
319 const u64 orig_end = *end;
320 /* The sanity tests may not set a valid fs_info. */
321 u64 max_bytes = fs_info ? fs_info->max_extent_size : BTRFS_MAX_EXTENT_SIZE;
322 u64 delalloc_start;
323 u64 delalloc_end;
324 bool found;
325 struct extent_state *cached_state = NULL;
326 int ret;
327 int loops = 0;
328
329 /* Caller should pass a valid @end to indicate the search range end */
330 ASSERT(orig_end > orig_start);
331
332 /* The range should at least cover part of the folio */
333 ASSERT(!(orig_start >= folio_pos(locked_folio) + folio_size(locked_folio) ||
334 orig_end <= folio_pos(locked_folio)));
335again:
336 /* step one, find a bunch of delalloc bytes starting at start */
337 delalloc_start = *start;
338 delalloc_end = 0;
339 found = btrfs_find_delalloc_range(tree, &delalloc_start, &delalloc_end,
340 max_bytes, &cached_state);
341 if (!found || delalloc_end <= *start || delalloc_start > orig_end) {
342 *start = delalloc_start;
343
344 /* @delalloc_end can be -1, never go beyond @orig_end */
345 *end = min(delalloc_end, orig_end);
346 free_extent_state(cached_state);
347 return false;
348 }
349
350 /*
351 * start comes from the offset of locked_folio. We have to lock
352 * folios in order, so we can't process delalloc bytes before
353 * locked_folio
354 */
355 if (delalloc_start < *start)
356 delalloc_start = *start;
357
358 /*
359 * make sure to limit the number of folios we try to lock down
360 */
361 if (delalloc_end + 1 - delalloc_start > max_bytes)
362 delalloc_end = delalloc_start + max_bytes - 1;
363
364 /* step two, lock all the folioss after the folios that has start */
365 ret = lock_delalloc_folios(inode, locked_folio, delalloc_start,
366 delalloc_end);
367 ASSERT(!ret || ret == -EAGAIN);
368 if (ret == -EAGAIN) {
369 /* some of the folios are gone, lets avoid looping by
370 * shortening the size of the delalloc range we're searching
371 */
372 free_extent_state(cached_state);
373 cached_state = NULL;
374 if (!loops) {
375 max_bytes = PAGE_SIZE;
376 loops = 1;
377 goto again;
378 } else {
379 found = false;
380 goto out_failed;
381 }
382 }
383
384 /* step three, lock the state bits for the whole range */
385 lock_extent(tree, delalloc_start, delalloc_end, &cached_state);
386
387 /* then test to make sure it is all still delalloc */
388 ret = test_range_bit(tree, delalloc_start, delalloc_end,
389 EXTENT_DELALLOC, cached_state);
390
391 unlock_extent(tree, delalloc_start, delalloc_end, &cached_state);
392 if (!ret) {
393 __unlock_for_delalloc(inode, locked_folio, delalloc_start,
394 delalloc_end);
395 cond_resched();
396 goto again;
397 }
398 *start = delalloc_start;
399 *end = delalloc_end;
400out_failed:
401 return found;
402}
403
404void extent_clear_unlock_delalloc(struct btrfs_inode *inode, u64 start, u64 end,
405 const struct folio *locked_folio,
406 struct extent_state **cached,
407 u32 clear_bits, unsigned long page_ops)
408{
409 clear_extent_bit(&inode->io_tree, start, end, clear_bits, cached);
410
411 __process_folios_contig(inode->vfs_inode.i_mapping, locked_folio, start,
412 end, page_ops);
413}
414
415static bool btrfs_verify_folio(struct folio *folio, u64 start, u32 len)
416{
417 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
418
419 if (!fsverity_active(folio->mapping->host) ||
420 btrfs_folio_test_uptodate(fs_info, folio, start, len) ||
421 start >= i_size_read(folio->mapping->host))
422 return true;
423 return fsverity_verify_folio(folio);
424}
425
426static void end_folio_read(struct folio *folio, bool uptodate, u64 start, u32 len)
427{
428 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
429
430 ASSERT(folio_pos(folio) <= start &&
431 start + len <= folio_pos(folio) + PAGE_SIZE);
432
433 if (uptodate && btrfs_verify_folio(folio, start, len))
434 btrfs_folio_set_uptodate(fs_info, folio, start, len);
435 else
436 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
437
438 if (!btrfs_is_subpage(fs_info, folio->mapping))
439 folio_unlock(folio);
440 else
441 btrfs_folio_end_lock(fs_info, folio, start, len);
442}
443
444/*
445 * After a write IO is done, we need to:
446 *
447 * - clear the uptodate bits on error
448 * - clear the writeback bits in the extent tree for the range
449 * - filio_end_writeback() if there is no more pending io for the folio
450 *
451 * Scheduling is not allowed, so the extent state tree is expected
452 * to have one and only one object corresponding to this IO.
453 */
454static void end_bbio_data_write(struct btrfs_bio *bbio)
455{
456 struct btrfs_fs_info *fs_info = bbio->fs_info;
457 struct bio *bio = &bbio->bio;
458 int error = blk_status_to_errno(bio->bi_status);
459 struct folio_iter fi;
460 const u32 sectorsize = fs_info->sectorsize;
461
462 ASSERT(!bio_flagged(bio, BIO_CLONED));
463 bio_for_each_folio_all(fi, bio) {
464 struct folio *folio = fi.folio;
465 u64 start = folio_pos(folio) + fi.offset;
466 u32 len = fi.length;
467
468 /* Only order 0 (single page) folios are allowed for data. */
469 ASSERT(folio_order(folio) == 0);
470
471 /* Our read/write should always be sector aligned. */
472 if (!IS_ALIGNED(fi.offset, sectorsize))
473 btrfs_err(fs_info,
474 "partial page write in btrfs with offset %zu and length %zu",
475 fi.offset, fi.length);
476 else if (!IS_ALIGNED(fi.length, sectorsize))
477 btrfs_info(fs_info,
478 "incomplete page write with offset %zu and length %zu",
479 fi.offset, fi.length);
480
481 btrfs_finish_ordered_extent(bbio->ordered, folio, start, len,
482 !error);
483 if (error)
484 mapping_set_error(folio->mapping, error);
485 btrfs_folio_clear_writeback(fs_info, folio, start, len);
486 }
487
488 bio_put(bio);
489}
490
491static void begin_folio_read(struct btrfs_fs_info *fs_info, struct folio *folio)
492{
493 ASSERT(folio_test_locked(folio));
494 if (!btrfs_is_subpage(fs_info, folio->mapping))
495 return;
496
497 ASSERT(folio_test_private(folio));
498 btrfs_folio_set_lock(fs_info, folio, folio_pos(folio), PAGE_SIZE);
499}
500
501/*
502 * After a data read IO is done, we need to:
503 *
504 * - clear the uptodate bits on error
505 * - set the uptodate bits if things worked
506 * - set the folio up to date if all extents in the tree are uptodate
507 * - clear the lock bit in the extent tree
508 * - unlock the folio if there are no other extents locked for it
509 *
510 * Scheduling is not allowed, so the extent state tree is expected
511 * to have one and only one object corresponding to this IO.
512 */
513static void end_bbio_data_read(struct btrfs_bio *bbio)
514{
515 struct btrfs_fs_info *fs_info = bbio->fs_info;
516 struct bio *bio = &bbio->bio;
517 struct folio_iter fi;
518 const u32 sectorsize = fs_info->sectorsize;
519
520 ASSERT(!bio_flagged(bio, BIO_CLONED));
521 bio_for_each_folio_all(fi, &bbio->bio) {
522 bool uptodate = !bio->bi_status;
523 struct folio *folio = fi.folio;
524 struct inode *inode = folio->mapping->host;
525 u64 start;
526 u64 end;
527 u32 len;
528
529 /* For now only order 0 folios are supported for data. */
530 ASSERT(folio_order(folio) == 0);
531 btrfs_debug(fs_info,
532 "%s: bi_sector=%llu, err=%d, mirror=%u",
533 __func__, bio->bi_iter.bi_sector, bio->bi_status,
534 bbio->mirror_num);
535
536 /*
537 * We always issue full-sector reads, but if some block in a
538 * folio fails to read, blk_update_request() will advance
539 * bv_offset and adjust bv_len to compensate. Print a warning
540 * for unaligned offsets, and an error if they don't add up to
541 * a full sector.
542 */
543 if (!IS_ALIGNED(fi.offset, sectorsize))
544 btrfs_err(fs_info,
545 "partial page read in btrfs with offset %zu and length %zu",
546 fi.offset, fi.length);
547 else if (!IS_ALIGNED(fi.offset + fi.length, sectorsize))
548 btrfs_info(fs_info,
549 "incomplete page read with offset %zu and length %zu",
550 fi.offset, fi.length);
551
552 start = folio_pos(folio) + fi.offset;
553 end = start + fi.length - 1;
554 len = fi.length;
555
556 if (likely(uptodate)) {
557 loff_t i_size = i_size_read(inode);
558 pgoff_t end_index = i_size >> folio_shift(folio);
559
560 /*
561 * Zero out the remaining part if this range straddles
562 * i_size.
563 *
564 * Here we should only zero the range inside the folio,
565 * not touch anything else.
566 *
567 * NOTE: i_size is exclusive while end is inclusive.
568 */
569 if (folio_index(folio) == end_index && i_size <= end) {
570 u32 zero_start = max(offset_in_folio(folio, i_size),
571 offset_in_folio(folio, start));
572 u32 zero_len = offset_in_folio(folio, end) + 1 -
573 zero_start;
574
575 folio_zero_range(folio, zero_start, zero_len);
576 }
577 }
578
579 /* Update page status and unlock. */
580 end_folio_read(folio, uptodate, start, len);
581 }
582 bio_put(bio);
583}
584
585/*
586 * Populate every free slot in a provided array with folios using GFP_NOFS.
587 *
588 * @nr_folios: number of folios to allocate
589 * @folio_array: the array to fill with folios; any existing non-NULL entries in
590 * the array will be skipped
591 *
592 * Return: 0 if all folios were able to be allocated;
593 * -ENOMEM otherwise, the partially allocated folios would be freed and
594 * the array slots zeroed
595 */
596int btrfs_alloc_folio_array(unsigned int nr_folios, struct folio **folio_array)
597{
598 for (int i = 0; i < nr_folios; i++) {
599 if (folio_array[i])
600 continue;
601 folio_array[i] = folio_alloc(GFP_NOFS, 0);
602 if (!folio_array[i])
603 goto error;
604 }
605 return 0;
606error:
607 for (int i = 0; i < nr_folios; i++) {
608 if (folio_array[i])
609 folio_put(folio_array[i]);
610 }
611 return -ENOMEM;
612}
613
614/*
615 * Populate every free slot in a provided array with pages, using GFP_NOFS.
616 *
617 * @nr_pages: number of pages to allocate
618 * @page_array: the array to fill with pages; any existing non-null entries in
619 * the array will be skipped
620 * @nofail: whether using __GFP_NOFAIL flag
621 *
622 * Return: 0 if all pages were able to be allocated;
623 * -ENOMEM otherwise, the partially allocated pages would be freed and
624 * the array slots zeroed
625 */
626int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array,
627 bool nofail)
628{
629 const gfp_t gfp = nofail ? (GFP_NOFS | __GFP_NOFAIL) : GFP_NOFS;
630 unsigned int allocated;
631
632 for (allocated = 0; allocated < nr_pages;) {
633 unsigned int last = allocated;
634
635 allocated = alloc_pages_bulk_array(gfp, nr_pages, page_array);
636 if (unlikely(allocated == last)) {
637 /* No progress, fail and do cleanup. */
638 for (int i = 0; i < allocated; i++) {
639 __free_page(page_array[i]);
640 page_array[i] = NULL;
641 }
642 return -ENOMEM;
643 }
644 }
645 return 0;
646}
647
648/*
649 * Populate needed folios for the extent buffer.
650 *
651 * For now, the folios populated are always in order 0 (aka, single page).
652 */
653static int alloc_eb_folio_array(struct extent_buffer *eb, bool nofail)
654{
655 struct page *page_array[INLINE_EXTENT_BUFFER_PAGES] = { 0 };
656 int num_pages = num_extent_pages(eb);
657 int ret;
658
659 ret = btrfs_alloc_page_array(num_pages, page_array, nofail);
660 if (ret < 0)
661 return ret;
662
663 for (int i = 0; i < num_pages; i++)
664 eb->folios[i] = page_folio(page_array[i]);
665 eb->folio_size = PAGE_SIZE;
666 eb->folio_shift = PAGE_SHIFT;
667 return 0;
668}
669
670static bool btrfs_bio_is_contig(struct btrfs_bio_ctrl *bio_ctrl,
671 struct folio *folio, u64 disk_bytenr,
672 unsigned int pg_offset)
673{
674 struct bio *bio = &bio_ctrl->bbio->bio;
675 struct bio_vec *bvec = bio_last_bvec_all(bio);
676 const sector_t sector = disk_bytenr >> SECTOR_SHIFT;
677 struct folio *bv_folio = page_folio(bvec->bv_page);
678
679 if (bio_ctrl->compress_type != BTRFS_COMPRESS_NONE) {
680 /*
681 * For compression, all IO should have its logical bytenr set
682 * to the starting bytenr of the compressed extent.
683 */
684 return bio->bi_iter.bi_sector == sector;
685 }
686
687 /*
688 * The contig check requires the following conditions to be met:
689 *
690 * 1) The folios are belonging to the same inode
691 * This is implied by the call chain.
692 *
693 * 2) The range has adjacent logical bytenr
694 *
695 * 3) The range has adjacent file offset
696 * This is required for the usage of btrfs_bio->file_offset.
697 */
698 return bio_end_sector(bio) == sector &&
699 folio_pos(bv_folio) + bvec->bv_offset + bvec->bv_len ==
700 folio_pos(folio) + pg_offset;
701}
702
703static void alloc_new_bio(struct btrfs_inode *inode,
704 struct btrfs_bio_ctrl *bio_ctrl,
705 u64 disk_bytenr, u64 file_offset)
706{
707 struct btrfs_fs_info *fs_info = inode->root->fs_info;
708 struct btrfs_bio *bbio;
709
710 bbio = btrfs_bio_alloc(BIO_MAX_VECS, bio_ctrl->opf, fs_info,
711 bio_ctrl->end_io_func, NULL);
712 bbio->bio.bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
713 bbio->inode = inode;
714 bbio->file_offset = file_offset;
715 bio_ctrl->bbio = bbio;
716 bio_ctrl->len_to_oe_boundary = U32_MAX;
717
718 /* Limit data write bios to the ordered boundary. */
719 if (bio_ctrl->wbc) {
720 struct btrfs_ordered_extent *ordered;
721
722 ordered = btrfs_lookup_ordered_extent(inode, file_offset);
723 if (ordered) {
724 bio_ctrl->len_to_oe_boundary = min_t(u32, U32_MAX,
725 ordered->file_offset +
726 ordered->disk_num_bytes - file_offset);
727 bbio->ordered = ordered;
728 }
729
730 /*
731 * Pick the last added device to support cgroup writeback. For
732 * multi-device file systems this means blk-cgroup policies have
733 * to always be set on the last added/replaced device.
734 * This is a bit odd but has been like that for a long time.
735 */
736 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
737 wbc_init_bio(bio_ctrl->wbc, &bbio->bio);
738 }
739}
740
741/*
742 * @disk_bytenr: logical bytenr where the write will be
743 * @page: page to add to the bio
744 * @size: portion of page that we want to write to
745 * @pg_offset: offset of the new bio or to check whether we are adding
746 * a contiguous page to the previous one
747 *
748 * The will either add the page into the existing @bio_ctrl->bbio, or allocate a
749 * new one in @bio_ctrl->bbio.
750 * The mirror number for this IO should already be initizlied in
751 * @bio_ctrl->mirror_num.
752 */
753static void submit_extent_folio(struct btrfs_bio_ctrl *bio_ctrl,
754 u64 disk_bytenr, struct folio *folio,
755 size_t size, unsigned long pg_offset)
756{
757 struct btrfs_inode *inode = folio_to_inode(folio);
758
759 ASSERT(pg_offset + size <= PAGE_SIZE);
760 ASSERT(bio_ctrl->end_io_func);
761
762 if (bio_ctrl->bbio &&
763 !btrfs_bio_is_contig(bio_ctrl, folio, disk_bytenr, pg_offset))
764 submit_one_bio(bio_ctrl);
765
766 do {
767 u32 len = size;
768
769 /* Allocate new bio if needed */
770 if (!bio_ctrl->bbio) {
771 alloc_new_bio(inode, bio_ctrl, disk_bytenr,
772 folio_pos(folio) + pg_offset);
773 }
774
775 /* Cap to the current ordered extent boundary if there is one. */
776 if (len > bio_ctrl->len_to_oe_boundary) {
777 ASSERT(bio_ctrl->compress_type == BTRFS_COMPRESS_NONE);
778 ASSERT(is_data_inode(inode));
779 len = bio_ctrl->len_to_oe_boundary;
780 }
781
782 if (!bio_add_folio(&bio_ctrl->bbio->bio, folio, len, pg_offset)) {
783 /* bio full: move on to a new one */
784 submit_one_bio(bio_ctrl);
785 continue;
786 }
787
788 if (bio_ctrl->wbc)
789 wbc_account_cgroup_owner(bio_ctrl->wbc, folio,
790 len);
791
792 size -= len;
793 pg_offset += len;
794 disk_bytenr += len;
795
796 /*
797 * len_to_oe_boundary defaults to U32_MAX, which isn't folio or
798 * sector aligned. alloc_new_bio() then sets it to the end of
799 * our ordered extent for writes into zoned devices.
800 *
801 * When len_to_oe_boundary is tracking an ordered extent, we
802 * trust the ordered extent code to align things properly, and
803 * the check above to cap our write to the ordered extent
804 * boundary is correct.
805 *
806 * When len_to_oe_boundary is U32_MAX, the cap above would
807 * result in a 4095 byte IO for the last folio right before
808 * we hit the bio limit of UINT_MAX. bio_add_folio() has all
809 * the checks required to make sure we don't overflow the bio,
810 * and we should just ignore len_to_oe_boundary completely
811 * unless we're using it to track an ordered extent.
812 *
813 * It's pretty hard to make a bio sized U32_MAX, but it can
814 * happen when the page cache is able to feed us contiguous
815 * folios for large extents.
816 */
817 if (bio_ctrl->len_to_oe_boundary != U32_MAX)
818 bio_ctrl->len_to_oe_boundary -= len;
819
820 /* Ordered extent boundary: move on to a new bio. */
821 if (bio_ctrl->len_to_oe_boundary == 0)
822 submit_one_bio(bio_ctrl);
823 } while (size);
824}
825
826static int attach_extent_buffer_folio(struct extent_buffer *eb,
827 struct folio *folio,
828 struct btrfs_subpage *prealloc)
829{
830 struct btrfs_fs_info *fs_info = eb->fs_info;
831 int ret = 0;
832
833 /*
834 * If the page is mapped to btree inode, we should hold the private
835 * lock to prevent race.
836 * For cloned or dummy extent buffers, their pages are not mapped and
837 * will not race with any other ebs.
838 */
839 if (folio->mapping)
840 lockdep_assert_held(&folio->mapping->i_private_lock);
841
842 if (fs_info->nodesize >= PAGE_SIZE) {
843 if (!folio_test_private(folio))
844 folio_attach_private(folio, eb);
845 else
846 WARN_ON(folio_get_private(folio) != eb);
847 return 0;
848 }
849
850 /* Already mapped, just free prealloc */
851 if (folio_test_private(folio)) {
852 btrfs_free_subpage(prealloc);
853 return 0;
854 }
855
856 if (prealloc)
857 /* Has preallocated memory for subpage */
858 folio_attach_private(folio, prealloc);
859 else
860 /* Do new allocation to attach subpage */
861 ret = btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_METADATA);
862 return ret;
863}
864
865int set_page_extent_mapped(struct page *page)
866{
867 return set_folio_extent_mapped(page_folio(page));
868}
869
870int set_folio_extent_mapped(struct folio *folio)
871{
872 struct btrfs_fs_info *fs_info;
873
874 ASSERT(folio->mapping);
875
876 if (folio_test_private(folio))
877 return 0;
878
879 fs_info = folio_to_fs_info(folio);
880
881 if (btrfs_is_subpage(fs_info, folio->mapping))
882 return btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_DATA);
883
884 folio_attach_private(folio, (void *)EXTENT_FOLIO_PRIVATE);
885 return 0;
886}
887
888void clear_folio_extent_mapped(struct folio *folio)
889{
890 struct btrfs_fs_info *fs_info;
891
892 ASSERT(folio->mapping);
893
894 if (!folio_test_private(folio))
895 return;
896
897 fs_info = folio_to_fs_info(folio);
898 if (btrfs_is_subpage(fs_info, folio->mapping))
899 return btrfs_detach_subpage(fs_info, folio);
900
901 folio_detach_private(folio);
902}
903
904static struct extent_map *get_extent_map(struct btrfs_inode *inode,
905 struct folio *folio, u64 start,
906 u64 len, struct extent_map **em_cached)
907{
908 struct extent_map *em;
909
910 ASSERT(em_cached);
911
912 if (*em_cached) {
913 em = *em_cached;
914 if (extent_map_in_tree(em) && start >= em->start &&
915 start < extent_map_end(em)) {
916 refcount_inc(&em->refs);
917 return em;
918 }
919
920 free_extent_map(em);
921 *em_cached = NULL;
922 }
923
924 em = btrfs_get_extent(inode, folio, start, len);
925 if (!IS_ERR(em)) {
926 BUG_ON(*em_cached);
927 refcount_inc(&em->refs);
928 *em_cached = em;
929 }
930
931 return em;
932}
933/*
934 * basic readpage implementation. Locked extent state structs are inserted
935 * into the tree that are removed when the IO is done (by the end_io
936 * handlers)
937 * XXX JDM: This needs looking at to ensure proper page locking
938 * return 0 on success, otherwise return error
939 */
940static int btrfs_do_readpage(struct folio *folio, struct extent_map **em_cached,
941 struct btrfs_bio_ctrl *bio_ctrl, u64 *prev_em_start)
942{
943 struct inode *inode = folio->mapping->host;
944 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
945 u64 start = folio_pos(folio);
946 const u64 end = start + PAGE_SIZE - 1;
947 u64 cur = start;
948 u64 extent_offset;
949 u64 last_byte = i_size_read(inode);
950 u64 block_start;
951 struct extent_map *em;
952 int ret = 0;
953 size_t pg_offset = 0;
954 size_t iosize;
955 size_t blocksize = fs_info->sectorsize;
956
957 ret = set_folio_extent_mapped(folio);
958 if (ret < 0) {
959 folio_unlock(folio);
960 return ret;
961 }
962
963 if (folio->index == last_byte >> folio_shift(folio)) {
964 size_t zero_offset = offset_in_folio(folio, last_byte);
965
966 if (zero_offset) {
967 iosize = folio_size(folio) - zero_offset;
968 folio_zero_range(folio, zero_offset, iosize);
969 }
970 }
971 bio_ctrl->end_io_func = end_bbio_data_read;
972 begin_folio_read(fs_info, folio);
973 while (cur <= end) {
974 enum btrfs_compression_type compress_type = BTRFS_COMPRESS_NONE;
975 bool force_bio_submit = false;
976 u64 disk_bytenr;
977
978 ASSERT(IS_ALIGNED(cur, fs_info->sectorsize));
979 if (cur >= last_byte) {
980 iosize = folio_size(folio) - pg_offset;
981 folio_zero_range(folio, pg_offset, iosize);
982 end_folio_read(folio, true, cur, iosize);
983 break;
984 }
985 em = get_extent_map(BTRFS_I(inode), folio, cur, end - cur + 1, em_cached);
986 if (IS_ERR(em)) {
987 end_folio_read(folio, false, cur, end + 1 - cur);
988 return PTR_ERR(em);
989 }
990 extent_offset = cur - em->start;
991 BUG_ON(extent_map_end(em) <= cur);
992 BUG_ON(end < cur);
993
994 compress_type = extent_map_compression(em);
995
996 iosize = min(extent_map_end(em) - cur, end - cur + 1);
997 iosize = ALIGN(iosize, blocksize);
998 if (compress_type != BTRFS_COMPRESS_NONE)
999 disk_bytenr = em->disk_bytenr;
1000 else
1001 disk_bytenr = extent_map_block_start(em) + extent_offset;
1002 block_start = extent_map_block_start(em);
1003 if (em->flags & EXTENT_FLAG_PREALLOC)
1004 block_start = EXTENT_MAP_HOLE;
1005
1006 /*
1007 * If we have a file range that points to a compressed extent
1008 * and it's followed by a consecutive file range that points
1009 * to the same compressed extent (possibly with a different
1010 * offset and/or length, so it either points to the whole extent
1011 * or only part of it), we must make sure we do not submit a
1012 * single bio to populate the folios for the 2 ranges because
1013 * this makes the compressed extent read zero out the folios
1014 * belonging to the 2nd range. Imagine the following scenario:
1015 *
1016 * File layout
1017 * [0 - 8K] [8K - 24K]
1018 * | |
1019 * | |
1020 * points to extent X, points to extent X,
1021 * offset 4K, length of 8K offset 0, length 16K
1022 *
1023 * [extent X, compressed length = 4K uncompressed length = 16K]
1024 *
1025 * If the bio to read the compressed extent covers both ranges,
1026 * it will decompress extent X into the folios belonging to the
1027 * first range and then it will stop, zeroing out the remaining
1028 * folios that belong to the other range that points to extent X.
1029 * So here we make sure we submit 2 bios, one for the first
1030 * range and another one for the third range. Both will target
1031 * the same physical extent from disk, but we can't currently
1032 * make the compressed bio endio callback populate the folios
1033 * for both ranges because each compressed bio is tightly
1034 * coupled with a single extent map, and each range can have
1035 * an extent map with a different offset value relative to the
1036 * uncompressed data of our extent and different lengths. This
1037 * is a corner case so we prioritize correctness over
1038 * non-optimal behavior (submitting 2 bios for the same extent).
1039 */
1040 if (compress_type != BTRFS_COMPRESS_NONE &&
1041 prev_em_start && *prev_em_start != (u64)-1 &&
1042 *prev_em_start != em->start)
1043 force_bio_submit = true;
1044
1045 if (prev_em_start)
1046 *prev_em_start = em->start;
1047
1048 free_extent_map(em);
1049 em = NULL;
1050
1051 /* we've found a hole, just zero and go on */
1052 if (block_start == EXTENT_MAP_HOLE) {
1053 folio_zero_range(folio, pg_offset, iosize);
1054
1055 end_folio_read(folio, true, cur, iosize);
1056 cur = cur + iosize;
1057 pg_offset += iosize;
1058 continue;
1059 }
1060 /* the get_extent function already copied into the folio */
1061 if (block_start == EXTENT_MAP_INLINE) {
1062 end_folio_read(folio, true, cur, iosize);
1063 cur = cur + iosize;
1064 pg_offset += iosize;
1065 continue;
1066 }
1067
1068 if (bio_ctrl->compress_type != compress_type) {
1069 submit_one_bio(bio_ctrl);
1070 bio_ctrl->compress_type = compress_type;
1071 }
1072
1073 if (force_bio_submit)
1074 submit_one_bio(bio_ctrl);
1075 submit_extent_folio(bio_ctrl, disk_bytenr, folio, iosize,
1076 pg_offset);
1077 cur = cur + iosize;
1078 pg_offset += iosize;
1079 }
1080
1081 return 0;
1082}
1083
1084int btrfs_read_folio(struct file *file, struct folio *folio)
1085{
1086 struct btrfs_inode *inode = folio_to_inode(folio);
1087 const u64 start = folio_pos(folio);
1088 const u64 end = start + folio_size(folio) - 1;
1089 struct extent_state *cached_state = NULL;
1090 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ };
1091 struct extent_map *em_cached = NULL;
1092 int ret;
1093
1094 btrfs_lock_and_flush_ordered_range(inode, start, end, &cached_state);
1095 ret = btrfs_do_readpage(folio, &em_cached, &bio_ctrl, NULL);
1096 unlock_extent(&inode->io_tree, start, end, &cached_state);
1097
1098 free_extent_map(em_cached);
1099
1100 /*
1101 * If btrfs_do_readpage() failed we will want to submit the assembled
1102 * bio to do the cleanup.
1103 */
1104 submit_one_bio(&bio_ctrl);
1105 return ret;
1106}
1107
1108static void set_delalloc_bitmap(struct folio *folio, unsigned long *delalloc_bitmap,
1109 u64 start, u32 len)
1110{
1111 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
1112 const u64 folio_start = folio_pos(folio);
1113 unsigned int start_bit;
1114 unsigned int nbits;
1115
1116 ASSERT(start >= folio_start && start + len <= folio_start + PAGE_SIZE);
1117 start_bit = (start - folio_start) >> fs_info->sectorsize_bits;
1118 nbits = len >> fs_info->sectorsize_bits;
1119 ASSERT(bitmap_test_range_all_zero(delalloc_bitmap, start_bit, nbits));
1120 bitmap_set(delalloc_bitmap, start_bit, nbits);
1121}
1122
1123static bool find_next_delalloc_bitmap(struct folio *folio,
1124 unsigned long *delalloc_bitmap, u64 start,
1125 u64 *found_start, u32 *found_len)
1126{
1127 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
1128 const u64 folio_start = folio_pos(folio);
1129 const unsigned int bitmap_size = fs_info->sectors_per_page;
1130 unsigned int start_bit;
1131 unsigned int first_zero;
1132 unsigned int first_set;
1133
1134 ASSERT(start >= folio_start && start < folio_start + PAGE_SIZE);
1135
1136 start_bit = (start - folio_start) >> fs_info->sectorsize_bits;
1137 first_set = find_next_bit(delalloc_bitmap, bitmap_size, start_bit);
1138 if (first_set >= bitmap_size)
1139 return false;
1140
1141 *found_start = folio_start + (first_set << fs_info->sectorsize_bits);
1142 first_zero = find_next_zero_bit(delalloc_bitmap, bitmap_size, first_set);
1143 *found_len = (first_zero - first_set) << fs_info->sectorsize_bits;
1144 return true;
1145}
1146
1147/*
1148 * Do all of the delayed allocation setup.
1149 *
1150 * Return >0 if all the dirty blocks are submitted async (compression) or inlined.
1151 * The @folio should no longer be touched (treat it as already unlocked).
1152 *
1153 * Return 0 if there is still dirty block that needs to be submitted through
1154 * extent_writepage_io().
1155 * bio_ctrl->submit_bitmap will indicate which blocks of the folio should be
1156 * submitted, and @folio is still kept locked.
1157 *
1158 * Return <0 if there is any error hit.
1159 * Any allocated ordered extent range covering this folio will be marked
1160 * finished (IOERR), and @folio is still kept locked.
1161 */
1162static noinline_for_stack int writepage_delalloc(struct btrfs_inode *inode,
1163 struct folio *folio,
1164 struct btrfs_bio_ctrl *bio_ctrl)
1165{
1166 struct btrfs_fs_info *fs_info = inode_to_fs_info(&inode->vfs_inode);
1167 struct writeback_control *wbc = bio_ctrl->wbc;
1168 const bool is_subpage = btrfs_is_subpage(fs_info, folio->mapping);
1169 const u64 page_start = folio_pos(folio);
1170 const u64 page_end = page_start + folio_size(folio) - 1;
1171 unsigned long delalloc_bitmap = 0;
1172 /*
1173 * Save the last found delalloc end. As the delalloc end can go beyond
1174 * page boundary, thus we cannot rely on subpage bitmap to locate the
1175 * last delalloc end.
1176 */
1177 u64 last_delalloc_end = 0;
1178 /*
1179 * The range end (exclusive) of the last successfully finished delalloc
1180 * range.
1181 * Any range covered by ordered extent must either be manually marked
1182 * finished (error handling), or has IO submitted (and finish the
1183 * ordered extent normally).
1184 *
1185 * This records the end of ordered extent cleanup if we hit an error.
1186 */
1187 u64 last_finished_delalloc_end = page_start;
1188 u64 delalloc_start = page_start;
1189 u64 delalloc_end = page_end;
1190 u64 delalloc_to_write = 0;
1191 int ret = 0;
1192 int bit;
1193
1194 /* Save the dirty bitmap as our submission bitmap will be a subset of it. */
1195 if (btrfs_is_subpage(fs_info, inode->vfs_inode.i_mapping)) {
1196 ASSERT(fs_info->sectors_per_page > 1);
1197 btrfs_get_subpage_dirty_bitmap(fs_info, folio, &bio_ctrl->submit_bitmap);
1198 } else {
1199 bio_ctrl->submit_bitmap = 1;
1200 }
1201
1202 for_each_set_bit(bit, &bio_ctrl->submit_bitmap, fs_info->sectors_per_page) {
1203 u64 start = page_start + (bit << fs_info->sectorsize_bits);
1204
1205 btrfs_folio_set_lock(fs_info, folio, start, fs_info->sectorsize);
1206 }
1207
1208 /* Lock all (subpage) delalloc ranges inside the folio first. */
1209 while (delalloc_start < page_end) {
1210 delalloc_end = page_end;
1211 if (!find_lock_delalloc_range(&inode->vfs_inode, folio,
1212 &delalloc_start, &delalloc_end)) {
1213 delalloc_start = delalloc_end + 1;
1214 continue;
1215 }
1216 set_delalloc_bitmap(folio, &delalloc_bitmap, delalloc_start,
1217 min(delalloc_end, page_end) + 1 - delalloc_start);
1218 last_delalloc_end = delalloc_end;
1219 delalloc_start = delalloc_end + 1;
1220 }
1221 delalloc_start = page_start;
1222
1223 if (!last_delalloc_end)
1224 goto out;
1225
1226 /* Run the delalloc ranges for the above locked ranges. */
1227 while (delalloc_start < page_end) {
1228 u64 found_start;
1229 u32 found_len;
1230 bool found;
1231
1232 if (!is_subpage) {
1233 /*
1234 * For non-subpage case, the found delalloc range must
1235 * cover this folio and there must be only one locked
1236 * delalloc range.
1237 */
1238 found_start = page_start;
1239 found_len = last_delalloc_end + 1 - found_start;
1240 found = true;
1241 } else {
1242 found = find_next_delalloc_bitmap(folio, &delalloc_bitmap,
1243 delalloc_start, &found_start, &found_len);
1244 }
1245 if (!found)
1246 break;
1247 /*
1248 * The subpage range covers the last sector, the delalloc range may
1249 * end beyond the folio boundary, use the saved delalloc_end
1250 * instead.
1251 */
1252 if (found_start + found_len >= page_end)
1253 found_len = last_delalloc_end + 1 - found_start;
1254
1255 if (ret >= 0) {
1256 /*
1257 * Some delalloc range may be created by previous folios.
1258 * Thus we still need to clean up this range during error
1259 * handling.
1260 */
1261 last_finished_delalloc_end = found_start;
1262 /* No errors hit so far, run the current delalloc range. */
1263 ret = btrfs_run_delalloc_range(inode, folio,
1264 found_start,
1265 found_start + found_len - 1,
1266 wbc);
1267 if (ret >= 0)
1268 last_finished_delalloc_end = found_start + found_len;
1269 } else {
1270 /*
1271 * We've hit an error during previous delalloc range,
1272 * have to cleanup the remaining locked ranges.
1273 */
1274 unlock_extent(&inode->io_tree, found_start,
1275 found_start + found_len - 1, NULL);
1276 __unlock_for_delalloc(&inode->vfs_inode, folio,
1277 found_start,
1278 found_start + found_len - 1);
1279 }
1280
1281 /*
1282 * We have some ranges that's going to be submitted asynchronously
1283 * (compression or inline). These range have their own control
1284 * on when to unlock the pages. We should not touch them
1285 * anymore, so clear the range from the submission bitmap.
1286 */
1287 if (ret > 0) {
1288 unsigned int start_bit = (found_start - page_start) >>
1289 fs_info->sectorsize_bits;
1290 unsigned int end_bit = (min(page_end + 1, found_start + found_len) -
1291 page_start) >> fs_info->sectorsize_bits;
1292 bitmap_clear(&bio_ctrl->submit_bitmap, start_bit, end_bit - start_bit);
1293 }
1294 /*
1295 * Above btrfs_run_delalloc_range() may have unlocked the folio,
1296 * thus for the last range, we cannot touch the folio anymore.
1297 */
1298 if (found_start + found_len >= last_delalloc_end + 1)
1299 break;
1300
1301 delalloc_start = found_start + found_len;
1302 }
1303 /*
1304 * It's possible we had some ordered extents created before we hit
1305 * an error, cleanup non-async successfully created delalloc ranges.
1306 */
1307 if (unlikely(ret < 0)) {
1308 unsigned int bitmap_size = min(
1309 (last_finished_delalloc_end - page_start) >>
1310 fs_info->sectorsize_bits,
1311 fs_info->sectors_per_page);
1312
1313 for_each_set_bit(bit, &bio_ctrl->submit_bitmap, bitmap_size)
1314 btrfs_mark_ordered_io_finished(inode, folio,
1315 page_start + (bit << fs_info->sectorsize_bits),
1316 fs_info->sectorsize, false);
1317 return ret;
1318 }
1319out:
1320 if (last_delalloc_end)
1321 delalloc_end = last_delalloc_end;
1322 else
1323 delalloc_end = page_end;
1324 /*
1325 * delalloc_end is already one less than the total length, so
1326 * we don't subtract one from PAGE_SIZE
1327 */
1328 delalloc_to_write +=
1329 DIV_ROUND_UP(delalloc_end + 1 - page_start, PAGE_SIZE);
1330
1331 /*
1332 * If all ranges are submitted asynchronously, we just need to account
1333 * for them here.
1334 */
1335 if (bitmap_empty(&bio_ctrl->submit_bitmap, fs_info->sectors_per_page)) {
1336 wbc->nr_to_write -= delalloc_to_write;
1337 return 1;
1338 }
1339
1340 if (wbc->nr_to_write < delalloc_to_write) {
1341 int thresh = 8192;
1342
1343 if (delalloc_to_write < thresh * 2)
1344 thresh = delalloc_to_write;
1345 wbc->nr_to_write = min_t(u64, delalloc_to_write,
1346 thresh);
1347 }
1348
1349 return 0;
1350}
1351
1352/*
1353 * Return 0 if we have submitted or queued the sector for submission.
1354 * Return <0 for critical errors.
1355 *
1356 * Caller should make sure filepos < i_size and handle filepos >= i_size case.
1357 */
1358static int submit_one_sector(struct btrfs_inode *inode,
1359 struct folio *folio,
1360 u64 filepos, struct btrfs_bio_ctrl *bio_ctrl,
1361 loff_t i_size)
1362{
1363 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1364 struct extent_map *em;
1365 u64 block_start;
1366 u64 disk_bytenr;
1367 u64 extent_offset;
1368 u64 em_end;
1369 const u32 sectorsize = fs_info->sectorsize;
1370
1371 ASSERT(IS_ALIGNED(filepos, sectorsize));
1372
1373 /* @filepos >= i_size case should be handled by the caller. */
1374 ASSERT(filepos < i_size);
1375
1376 em = btrfs_get_extent(inode, NULL, filepos, sectorsize);
1377 if (IS_ERR(em))
1378 return PTR_ERR_OR_ZERO(em);
1379
1380 extent_offset = filepos - em->start;
1381 em_end = extent_map_end(em);
1382 ASSERT(filepos <= em_end);
1383 ASSERT(IS_ALIGNED(em->start, sectorsize));
1384 ASSERT(IS_ALIGNED(em->len, sectorsize));
1385
1386 block_start = extent_map_block_start(em);
1387 disk_bytenr = extent_map_block_start(em) + extent_offset;
1388
1389 ASSERT(!extent_map_is_compressed(em));
1390 ASSERT(block_start != EXTENT_MAP_HOLE);
1391 ASSERT(block_start != EXTENT_MAP_INLINE);
1392
1393 free_extent_map(em);
1394 em = NULL;
1395
1396 /*
1397 * Although the PageDirty bit is cleared before entering this
1398 * function, subpage dirty bit is not cleared.
1399 * So clear subpage dirty bit here so next time we won't submit
1400 * a folio for a range already written to disk.
1401 */
1402 btrfs_folio_clear_dirty(fs_info, folio, filepos, sectorsize);
1403 btrfs_folio_set_writeback(fs_info, folio, filepos, sectorsize);
1404 /*
1405 * Above call should set the whole folio with writeback flag, even
1406 * just for a single subpage sector.
1407 * As long as the folio is properly locked and the range is correct,
1408 * we should always get the folio with writeback flag.
1409 */
1410 ASSERT(folio_test_writeback(folio));
1411
1412 submit_extent_folio(bio_ctrl, disk_bytenr, folio,
1413 sectorsize, filepos - folio_pos(folio));
1414 return 0;
1415}
1416
1417/*
1418 * Helper for extent_writepage(). This calls the writepage start hooks,
1419 * and does the loop to map the page into extents and bios.
1420 *
1421 * We return 1 if the IO is started and the page is unlocked,
1422 * 0 if all went well (page still locked)
1423 * < 0 if there were errors (page still locked)
1424 */
1425static noinline_for_stack int extent_writepage_io(struct btrfs_inode *inode,
1426 struct folio *folio,
1427 u64 start, u32 len,
1428 struct btrfs_bio_ctrl *bio_ctrl,
1429 loff_t i_size)
1430{
1431 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1432 unsigned long range_bitmap = 0;
1433 bool submitted_io = false;
1434 bool error = false;
1435 const u64 folio_start = folio_pos(folio);
1436 u64 cur;
1437 int bit;
1438 int ret = 0;
1439
1440 ASSERT(start >= folio_start &&
1441 start + len <= folio_start + folio_size(folio));
1442
1443 ret = btrfs_writepage_cow_fixup(folio);
1444 if (ret) {
1445 /* Fixup worker will requeue */
1446 folio_redirty_for_writepage(bio_ctrl->wbc, folio);
1447 folio_unlock(folio);
1448 return 1;
1449 }
1450
1451 for (cur = start; cur < start + len; cur += fs_info->sectorsize)
1452 set_bit((cur - folio_start) >> fs_info->sectorsize_bits, &range_bitmap);
1453 bitmap_and(&bio_ctrl->submit_bitmap, &bio_ctrl->submit_bitmap, &range_bitmap,
1454 fs_info->sectors_per_page);
1455
1456 bio_ctrl->end_io_func = end_bbio_data_write;
1457
1458 for_each_set_bit(bit, &bio_ctrl->submit_bitmap, fs_info->sectors_per_page) {
1459 cur = folio_pos(folio) + (bit << fs_info->sectorsize_bits);
1460
1461 if (cur >= i_size) {
1462 btrfs_mark_ordered_io_finished(inode, folio, cur,
1463 start + len - cur, true);
1464 /*
1465 * This range is beyond i_size, thus we don't need to
1466 * bother writing back.
1467 * But we still need to clear the dirty subpage bit, or
1468 * the next time the folio gets dirtied, we will try to
1469 * writeback the sectors with subpage dirty bits,
1470 * causing writeback without ordered extent.
1471 */
1472 btrfs_folio_clear_dirty(fs_info, folio, cur,
1473 start + len - cur);
1474 break;
1475 }
1476 ret = submit_one_sector(inode, folio, cur, bio_ctrl, i_size);
1477 if (unlikely(ret < 0)) {
1478 /*
1479 * bio_ctrl may contain a bio crossing several folios.
1480 * Submit it immediately so that the bio has a chance
1481 * to finish normally, other than marked as error.
1482 */
1483 submit_one_bio(bio_ctrl);
1484 /*
1485 * Failed to grab the extent map which should be very rare.
1486 * Since there is no bio submitted to finish the ordered
1487 * extent, we have to manually finish this sector.
1488 */
1489 btrfs_mark_ordered_io_finished(inode, folio, cur,
1490 fs_info->sectorsize, false);
1491 error = true;
1492 continue;
1493 }
1494 submitted_io = true;
1495 }
1496
1497 /*
1498 * If we didn't submitted any sector (>= i_size), folio dirty get
1499 * cleared but PAGECACHE_TAG_DIRTY is not cleared (only cleared
1500 * by folio_start_writeback() if the folio is not dirty).
1501 *
1502 * Here we set writeback and clear for the range. If the full folio
1503 * is no longer dirty then we clear the PAGECACHE_TAG_DIRTY tag.
1504 *
1505 * If we hit any error, the corresponding sector will still be dirty
1506 * thus no need to clear PAGECACHE_TAG_DIRTY.
1507 */
1508 if (!submitted_io && !error) {
1509 btrfs_folio_set_writeback(fs_info, folio, start, len);
1510 btrfs_folio_clear_writeback(fs_info, folio, start, len);
1511 }
1512 return ret;
1513}
1514
1515/*
1516 * the writepage semantics are similar to regular writepage. extent
1517 * records are inserted to lock ranges in the tree, and as dirty areas
1518 * are found, they are marked writeback. Then the lock bits are removed
1519 * and the end_io handler clears the writeback ranges
1520 *
1521 * Return 0 if everything goes well.
1522 * Return <0 for error.
1523 */
1524static int extent_writepage(struct folio *folio, struct btrfs_bio_ctrl *bio_ctrl)
1525{
1526 struct btrfs_inode *inode = BTRFS_I(folio->mapping->host);
1527 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1528 int ret;
1529 size_t pg_offset;
1530 loff_t i_size = i_size_read(&inode->vfs_inode);
1531 unsigned long end_index = i_size >> PAGE_SHIFT;
1532
1533 trace_extent_writepage(folio, &inode->vfs_inode, bio_ctrl->wbc);
1534
1535 WARN_ON(!folio_test_locked(folio));
1536
1537 pg_offset = offset_in_folio(folio, i_size);
1538 if (folio->index > end_index ||
1539 (folio->index == end_index && !pg_offset)) {
1540 folio_invalidate(folio, 0, folio_size(folio));
1541 folio_unlock(folio);
1542 return 0;
1543 }
1544
1545 if (folio->index == end_index)
1546 folio_zero_range(folio, pg_offset, folio_size(folio) - pg_offset);
1547
1548 /*
1549 * Default to unlock the whole folio.
1550 * The proper bitmap can only be initialized until writepage_delalloc().
1551 */
1552 bio_ctrl->submit_bitmap = (unsigned long)-1;
1553 ret = set_folio_extent_mapped(folio);
1554 if (ret < 0)
1555 goto done;
1556
1557 ret = writepage_delalloc(inode, folio, bio_ctrl);
1558 if (ret == 1)
1559 return 0;
1560 if (ret)
1561 goto done;
1562
1563 ret = extent_writepage_io(inode, folio, folio_pos(folio),
1564 PAGE_SIZE, bio_ctrl, i_size);
1565 if (ret == 1)
1566 return 0;
1567
1568 bio_ctrl->wbc->nr_to_write--;
1569
1570done:
1571 if (ret < 0)
1572 mapping_set_error(folio->mapping, ret);
1573 /*
1574 * Only unlock ranges that are submitted. As there can be some async
1575 * submitted ranges inside the folio.
1576 */
1577 btrfs_folio_end_lock_bitmap(fs_info, folio, bio_ctrl->submit_bitmap);
1578 ASSERT(ret <= 0);
1579 return ret;
1580}
1581
1582void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
1583{
1584 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
1585 TASK_UNINTERRUPTIBLE);
1586}
1587
1588/*
1589 * Lock extent buffer status and pages for writeback.
1590 *
1591 * Return %false if the extent buffer doesn't need to be submitted (e.g. the
1592 * extent buffer is not dirty)
1593 * Return %true is the extent buffer is submitted to bio.
1594 */
1595static noinline_for_stack bool lock_extent_buffer_for_io(struct extent_buffer *eb,
1596 struct writeback_control *wbc)
1597{
1598 struct btrfs_fs_info *fs_info = eb->fs_info;
1599 bool ret = false;
1600
1601 btrfs_tree_lock(eb);
1602 while (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
1603 btrfs_tree_unlock(eb);
1604 if (wbc->sync_mode != WB_SYNC_ALL)
1605 return false;
1606 wait_on_extent_buffer_writeback(eb);
1607 btrfs_tree_lock(eb);
1608 }
1609
1610 /*
1611 * We need to do this to prevent races in people who check if the eb is
1612 * under IO since we can end up having no IO bits set for a short period
1613 * of time.
1614 */
1615 spin_lock(&eb->refs_lock);
1616 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
1617 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1618 spin_unlock(&eb->refs_lock);
1619 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
1620 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
1621 -eb->len,
1622 fs_info->dirty_metadata_batch);
1623 ret = true;
1624 } else {
1625 spin_unlock(&eb->refs_lock);
1626 }
1627 btrfs_tree_unlock(eb);
1628 return ret;
1629}
1630
1631static void set_btree_ioerr(struct extent_buffer *eb)
1632{
1633 struct btrfs_fs_info *fs_info = eb->fs_info;
1634
1635 set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1636
1637 /*
1638 * A read may stumble upon this buffer later, make sure that it gets an
1639 * error and knows there was an error.
1640 */
1641 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
1642
1643 /*
1644 * We need to set the mapping with the io error as well because a write
1645 * error will flip the file system readonly, and then syncfs() will
1646 * return a 0 because we are readonly if we don't modify the err seq for
1647 * the superblock.
1648 */
1649 mapping_set_error(eb->fs_info->btree_inode->i_mapping, -EIO);
1650
1651 /*
1652 * If writeback for a btree extent that doesn't belong to a log tree
1653 * failed, increment the counter transaction->eb_write_errors.
1654 * We do this because while the transaction is running and before it's
1655 * committing (when we call filemap_fdata[write|wait]_range against
1656 * the btree inode), we might have
1657 * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
1658 * returns an error or an error happens during writeback, when we're
1659 * committing the transaction we wouldn't know about it, since the pages
1660 * can be no longer dirty nor marked anymore for writeback (if a
1661 * subsequent modification to the extent buffer didn't happen before the
1662 * transaction commit), which makes filemap_fdata[write|wait]_range not
1663 * able to find the pages which contain errors at transaction
1664 * commit time. So if this happens we must abort the transaction,
1665 * otherwise we commit a super block with btree roots that point to
1666 * btree nodes/leafs whose content on disk is invalid - either garbage
1667 * or the content of some node/leaf from a past generation that got
1668 * cowed or deleted and is no longer valid.
1669 *
1670 * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
1671 * not be enough - we need to distinguish between log tree extents vs
1672 * non-log tree extents, and the next filemap_fdatawait_range() call
1673 * will catch and clear such errors in the mapping - and that call might
1674 * be from a log sync and not from a transaction commit. Also, checking
1675 * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
1676 * not done and would not be reliable - the eb might have been released
1677 * from memory and reading it back again means that flag would not be
1678 * set (since it's a runtime flag, not persisted on disk).
1679 *
1680 * Using the flags below in the btree inode also makes us achieve the
1681 * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
1682 * writeback for all dirty pages and before filemap_fdatawait_range()
1683 * is called, the writeback for all dirty pages had already finished
1684 * with errors - because we were not using AS_EIO/AS_ENOSPC,
1685 * filemap_fdatawait_range() would return success, as it could not know
1686 * that writeback errors happened (the pages were no longer tagged for
1687 * writeback).
1688 */
1689 switch (eb->log_index) {
1690 case -1:
1691 set_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags);
1692 break;
1693 case 0:
1694 set_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
1695 break;
1696 case 1:
1697 set_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
1698 break;
1699 default:
1700 BUG(); /* unexpected, logic error */
1701 }
1702}
1703
1704/*
1705 * The endio specific version which won't touch any unsafe spinlock in endio
1706 * context.
1707 */
1708static struct extent_buffer *find_extent_buffer_nolock(
1709 const struct btrfs_fs_info *fs_info, u64 start)
1710{
1711 struct extent_buffer *eb;
1712
1713 rcu_read_lock();
1714 eb = radix_tree_lookup(&fs_info->buffer_radix,
1715 start >> fs_info->sectorsize_bits);
1716 if (eb && atomic_inc_not_zero(&eb->refs)) {
1717 rcu_read_unlock();
1718 return eb;
1719 }
1720 rcu_read_unlock();
1721 return NULL;
1722}
1723
1724static void end_bbio_meta_write(struct btrfs_bio *bbio)
1725{
1726 struct extent_buffer *eb = bbio->private;
1727 struct btrfs_fs_info *fs_info = eb->fs_info;
1728 bool uptodate = !bbio->bio.bi_status;
1729 struct folio_iter fi;
1730 u32 bio_offset = 0;
1731
1732 if (!uptodate)
1733 set_btree_ioerr(eb);
1734
1735 bio_for_each_folio_all(fi, &bbio->bio) {
1736 u64 start = eb->start + bio_offset;
1737 struct folio *folio = fi.folio;
1738 u32 len = fi.length;
1739
1740 btrfs_folio_clear_writeback(fs_info, folio, start, len);
1741 bio_offset += len;
1742 }
1743
1744 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1745 smp_mb__after_atomic();
1746 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
1747
1748 bio_put(&bbio->bio);
1749}
1750
1751static void prepare_eb_write(struct extent_buffer *eb)
1752{
1753 u32 nritems;
1754 unsigned long start;
1755 unsigned long end;
1756
1757 clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1758
1759 /* Set btree blocks beyond nritems with 0 to avoid stale content */
1760 nritems = btrfs_header_nritems(eb);
1761 if (btrfs_header_level(eb) > 0) {
1762 end = btrfs_node_key_ptr_offset(eb, nritems);
1763 memzero_extent_buffer(eb, end, eb->len - end);
1764 } else {
1765 /*
1766 * Leaf:
1767 * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
1768 */
1769 start = btrfs_item_nr_offset(eb, nritems);
1770 end = btrfs_item_nr_offset(eb, 0);
1771 if (nritems == 0)
1772 end += BTRFS_LEAF_DATA_SIZE(eb->fs_info);
1773 else
1774 end += btrfs_item_offset(eb, nritems - 1);
1775 memzero_extent_buffer(eb, start, end - start);
1776 }
1777}
1778
1779static noinline_for_stack void write_one_eb(struct extent_buffer *eb,
1780 struct writeback_control *wbc)
1781{
1782 struct btrfs_fs_info *fs_info = eb->fs_info;
1783 struct btrfs_bio *bbio;
1784
1785 prepare_eb_write(eb);
1786
1787 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
1788 REQ_OP_WRITE | REQ_META | wbc_to_write_flags(wbc),
1789 eb->fs_info, end_bbio_meta_write, eb);
1790 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
1791 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
1792 wbc_init_bio(wbc, &bbio->bio);
1793 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
1794 bbio->file_offset = eb->start;
1795 if (fs_info->nodesize < PAGE_SIZE) {
1796 struct folio *folio = eb->folios[0];
1797 bool ret;
1798
1799 folio_lock(folio);
1800 btrfs_subpage_set_writeback(fs_info, folio, eb->start, eb->len);
1801 if (btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start,
1802 eb->len)) {
1803 folio_clear_dirty_for_io(folio);
1804 wbc->nr_to_write--;
1805 }
1806 ret = bio_add_folio(&bbio->bio, folio, eb->len,
1807 eb->start - folio_pos(folio));
1808 ASSERT(ret);
1809 wbc_account_cgroup_owner(wbc, folio, eb->len);
1810 folio_unlock(folio);
1811 } else {
1812 int num_folios = num_extent_folios(eb);
1813
1814 for (int i = 0; i < num_folios; i++) {
1815 struct folio *folio = eb->folios[i];
1816 bool ret;
1817
1818 folio_lock(folio);
1819 folio_clear_dirty_for_io(folio);
1820 folio_start_writeback(folio);
1821 ret = bio_add_folio(&bbio->bio, folio, eb->folio_size, 0);
1822 ASSERT(ret);
1823 wbc_account_cgroup_owner(wbc, folio, eb->folio_size);
1824 wbc->nr_to_write -= folio_nr_pages(folio);
1825 folio_unlock(folio);
1826 }
1827 }
1828 btrfs_submit_bbio(bbio, 0);
1829}
1830
1831/*
1832 * Submit one subpage btree page.
1833 *
1834 * The main difference to submit_eb_page() is:
1835 * - Page locking
1836 * For subpage, we don't rely on page locking at all.
1837 *
1838 * - Flush write bio
1839 * We only flush bio if we may be unable to fit current extent buffers into
1840 * current bio.
1841 *
1842 * Return >=0 for the number of submitted extent buffers.
1843 * Return <0 for fatal error.
1844 */
1845static int submit_eb_subpage(struct folio *folio, struct writeback_control *wbc)
1846{
1847 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
1848 int submitted = 0;
1849 u64 folio_start = folio_pos(folio);
1850 int bit_start = 0;
1851 int sectors_per_node = fs_info->nodesize >> fs_info->sectorsize_bits;
1852
1853 /* Lock and write each dirty extent buffers in the range */
1854 while (bit_start < fs_info->sectors_per_page) {
1855 struct btrfs_subpage *subpage = folio_get_private(folio);
1856 struct extent_buffer *eb;
1857 unsigned long flags;
1858 u64 start;
1859
1860 /*
1861 * Take private lock to ensure the subpage won't be detached
1862 * in the meantime.
1863 */
1864 spin_lock(&folio->mapping->i_private_lock);
1865 if (!folio_test_private(folio)) {
1866 spin_unlock(&folio->mapping->i_private_lock);
1867 break;
1868 }
1869 spin_lock_irqsave(&subpage->lock, flags);
1870 if (!test_bit(bit_start + btrfs_bitmap_nr_dirty * fs_info->sectors_per_page,
1871 subpage->bitmaps)) {
1872 spin_unlock_irqrestore(&subpage->lock, flags);
1873 spin_unlock(&folio->mapping->i_private_lock);
1874 bit_start++;
1875 continue;
1876 }
1877
1878 start = folio_start + bit_start * fs_info->sectorsize;
1879 bit_start += sectors_per_node;
1880
1881 /*
1882 * Here we just want to grab the eb without touching extra
1883 * spin locks, so call find_extent_buffer_nolock().
1884 */
1885 eb = find_extent_buffer_nolock(fs_info, start);
1886 spin_unlock_irqrestore(&subpage->lock, flags);
1887 spin_unlock(&folio->mapping->i_private_lock);
1888
1889 /*
1890 * The eb has already reached 0 refs thus find_extent_buffer()
1891 * doesn't return it. We don't need to write back such eb
1892 * anyway.
1893 */
1894 if (!eb)
1895 continue;
1896
1897 if (lock_extent_buffer_for_io(eb, wbc)) {
1898 write_one_eb(eb, wbc);
1899 submitted++;
1900 }
1901 free_extent_buffer(eb);
1902 }
1903 return submitted;
1904}
1905
1906/*
1907 * Submit all page(s) of one extent buffer.
1908 *
1909 * @page: the page of one extent buffer
1910 * @eb_context: to determine if we need to submit this page, if current page
1911 * belongs to this eb, we don't need to submit
1912 *
1913 * The caller should pass each page in their bytenr order, and here we use
1914 * @eb_context to determine if we have submitted pages of one extent buffer.
1915 *
1916 * If we have, we just skip until we hit a new page that doesn't belong to
1917 * current @eb_context.
1918 *
1919 * If not, we submit all the page(s) of the extent buffer.
1920 *
1921 * Return >0 if we have submitted the extent buffer successfully.
1922 * Return 0 if we don't need to submit the page, as it's already submitted by
1923 * previous call.
1924 * Return <0 for fatal error.
1925 */
1926static int submit_eb_page(struct folio *folio, struct btrfs_eb_write_context *ctx)
1927{
1928 struct writeback_control *wbc = ctx->wbc;
1929 struct address_space *mapping = folio->mapping;
1930 struct extent_buffer *eb;
1931 int ret;
1932
1933 if (!folio_test_private(folio))
1934 return 0;
1935
1936 if (folio_to_fs_info(folio)->nodesize < PAGE_SIZE)
1937 return submit_eb_subpage(folio, wbc);
1938
1939 spin_lock(&mapping->i_private_lock);
1940 if (!folio_test_private(folio)) {
1941 spin_unlock(&mapping->i_private_lock);
1942 return 0;
1943 }
1944
1945 eb = folio_get_private(folio);
1946
1947 /*
1948 * Shouldn't happen and normally this would be a BUG_ON but no point
1949 * crashing the machine for something we can survive anyway.
1950 */
1951 if (WARN_ON(!eb)) {
1952 spin_unlock(&mapping->i_private_lock);
1953 return 0;
1954 }
1955
1956 if (eb == ctx->eb) {
1957 spin_unlock(&mapping->i_private_lock);
1958 return 0;
1959 }
1960 ret = atomic_inc_not_zero(&eb->refs);
1961 spin_unlock(&mapping->i_private_lock);
1962 if (!ret)
1963 return 0;
1964
1965 ctx->eb = eb;
1966
1967 ret = btrfs_check_meta_write_pointer(eb->fs_info, ctx);
1968 if (ret) {
1969 if (ret == -EBUSY)
1970 ret = 0;
1971 free_extent_buffer(eb);
1972 return ret;
1973 }
1974
1975 if (!lock_extent_buffer_for_io(eb, wbc)) {
1976 free_extent_buffer(eb);
1977 return 0;
1978 }
1979 /* Implies write in zoned mode. */
1980 if (ctx->zoned_bg) {
1981 /* Mark the last eb in the block group. */
1982 btrfs_schedule_zone_finish_bg(ctx->zoned_bg, eb);
1983 ctx->zoned_bg->meta_write_pointer += eb->len;
1984 }
1985 write_one_eb(eb, wbc);
1986 free_extent_buffer(eb);
1987 return 1;
1988}
1989
1990int btree_write_cache_pages(struct address_space *mapping,
1991 struct writeback_control *wbc)
1992{
1993 struct btrfs_eb_write_context ctx = { .wbc = wbc };
1994 struct btrfs_fs_info *fs_info = inode_to_fs_info(mapping->host);
1995 int ret = 0;
1996 int done = 0;
1997 int nr_to_write_done = 0;
1998 struct folio_batch fbatch;
1999 unsigned int nr_folios;
2000 pgoff_t index;
2001 pgoff_t end; /* Inclusive */
2002 int scanned = 0;
2003 xa_mark_t tag;
2004
2005 folio_batch_init(&fbatch);
2006 if (wbc->range_cyclic) {
2007 index = mapping->writeback_index; /* Start from prev offset */
2008 end = -1;
2009 /*
2010 * Start from the beginning does not need to cycle over the
2011 * range, mark it as scanned.
2012 */
2013 scanned = (index == 0);
2014 } else {
2015 index = wbc->range_start >> PAGE_SHIFT;
2016 end = wbc->range_end >> PAGE_SHIFT;
2017 scanned = 1;
2018 }
2019 if (wbc->sync_mode == WB_SYNC_ALL)
2020 tag = PAGECACHE_TAG_TOWRITE;
2021 else
2022 tag = PAGECACHE_TAG_DIRTY;
2023 btrfs_zoned_meta_io_lock(fs_info);
2024retry:
2025 if (wbc->sync_mode == WB_SYNC_ALL)
2026 tag_pages_for_writeback(mapping, index, end);
2027 while (!done && !nr_to_write_done && (index <= end) &&
2028 (nr_folios = filemap_get_folios_tag(mapping, &index, end,
2029 tag, &fbatch))) {
2030 unsigned i;
2031
2032 for (i = 0; i < nr_folios; i++) {
2033 struct folio *folio = fbatch.folios[i];
2034
2035 ret = submit_eb_page(folio, &ctx);
2036 if (ret == 0)
2037 continue;
2038 if (ret < 0) {
2039 done = 1;
2040 break;
2041 }
2042
2043 /*
2044 * the filesystem may choose to bump up nr_to_write.
2045 * We have to make sure to honor the new nr_to_write
2046 * at any time
2047 */
2048 nr_to_write_done = wbc->nr_to_write <= 0;
2049 }
2050 folio_batch_release(&fbatch);
2051 cond_resched();
2052 }
2053 if (!scanned && !done) {
2054 /*
2055 * We hit the last page and there is more work to be done: wrap
2056 * back to the start of the file
2057 */
2058 scanned = 1;
2059 index = 0;
2060 goto retry;
2061 }
2062 /*
2063 * If something went wrong, don't allow any metadata write bio to be
2064 * submitted.
2065 *
2066 * This would prevent use-after-free if we had dirty pages not
2067 * cleaned up, which can still happen by fuzzed images.
2068 *
2069 * - Bad extent tree
2070 * Allowing existing tree block to be allocated for other trees.
2071 *
2072 * - Log tree operations
2073 * Exiting tree blocks get allocated to log tree, bumps its
2074 * generation, then get cleaned in tree re-balance.
2075 * Such tree block will not be written back, since it's clean,
2076 * thus no WRITTEN flag set.
2077 * And after log writes back, this tree block is not traced by
2078 * any dirty extent_io_tree.
2079 *
2080 * - Offending tree block gets re-dirtied from its original owner
2081 * Since it has bumped generation, no WRITTEN flag, it can be
2082 * reused without COWing. This tree block will not be traced
2083 * by btrfs_transaction::dirty_pages.
2084 *
2085 * Now such dirty tree block will not be cleaned by any dirty
2086 * extent io tree. Thus we don't want to submit such wild eb
2087 * if the fs already has error.
2088 *
2089 * We can get ret > 0 from submit_extent_folio() indicating how many ebs
2090 * were submitted. Reset it to 0 to avoid false alerts for the caller.
2091 */
2092 if (ret > 0)
2093 ret = 0;
2094 if (!ret && BTRFS_FS_ERROR(fs_info))
2095 ret = -EROFS;
2096
2097 if (ctx.zoned_bg)
2098 btrfs_put_block_group(ctx.zoned_bg);
2099 btrfs_zoned_meta_io_unlock(fs_info);
2100 return ret;
2101}
2102
2103/*
2104 * Walk the list of dirty pages of the given address space and write all of them.
2105 *
2106 * @mapping: address space structure to write
2107 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2108 * @bio_ctrl: holds context for the write, namely the bio
2109 *
2110 * If a page is already under I/O, write_cache_pages() skips it, even
2111 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2112 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2113 * and msync() need to guarantee that all the data which was dirty at the time
2114 * the call was made get new I/O started against them. If wbc->sync_mode is
2115 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2116 * existing IO to complete.
2117 */
2118static int extent_write_cache_pages(struct address_space *mapping,
2119 struct btrfs_bio_ctrl *bio_ctrl)
2120{
2121 struct writeback_control *wbc = bio_ctrl->wbc;
2122 struct inode *inode = mapping->host;
2123 int ret = 0;
2124 int done = 0;
2125 int nr_to_write_done = 0;
2126 struct folio_batch fbatch;
2127 unsigned int nr_folios;
2128 pgoff_t index;
2129 pgoff_t end; /* Inclusive */
2130 pgoff_t done_index;
2131 int range_whole = 0;
2132 int scanned = 0;
2133 xa_mark_t tag;
2134
2135 /*
2136 * We have to hold onto the inode so that ordered extents can do their
2137 * work when the IO finishes. The alternative to this is failing to add
2138 * an ordered extent if the igrab() fails there and that is a huge pain
2139 * to deal with, so instead just hold onto the inode throughout the
2140 * writepages operation. If it fails here we are freeing up the inode
2141 * anyway and we'd rather not waste our time writing out stuff that is
2142 * going to be truncated anyway.
2143 */
2144 if (!igrab(inode))
2145 return 0;
2146
2147 folio_batch_init(&fbatch);
2148 if (wbc->range_cyclic) {
2149 index = mapping->writeback_index; /* Start from prev offset */
2150 end = -1;
2151 /*
2152 * Start from the beginning does not need to cycle over the
2153 * range, mark it as scanned.
2154 */
2155 scanned = (index == 0);
2156 } else {
2157 index = wbc->range_start >> PAGE_SHIFT;
2158 end = wbc->range_end >> PAGE_SHIFT;
2159 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2160 range_whole = 1;
2161 scanned = 1;
2162 }
2163
2164 /*
2165 * We do the tagged writepage as long as the snapshot flush bit is set
2166 * and we are the first one who do the filemap_flush() on this inode.
2167 *
2168 * The nr_to_write == LONG_MAX is needed to make sure other flushers do
2169 * not race in and drop the bit.
2170 */
2171 if (range_whole && wbc->nr_to_write == LONG_MAX &&
2172 test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
2173 &BTRFS_I(inode)->runtime_flags))
2174 wbc->tagged_writepages = 1;
2175
2176 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2177 tag = PAGECACHE_TAG_TOWRITE;
2178 else
2179 tag = PAGECACHE_TAG_DIRTY;
2180retry:
2181 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2182 tag_pages_for_writeback(mapping, index, end);
2183 done_index = index;
2184 while (!done && !nr_to_write_done && (index <= end) &&
2185 (nr_folios = filemap_get_folios_tag(mapping, &index,
2186 end, tag, &fbatch))) {
2187 unsigned i;
2188
2189 for (i = 0; i < nr_folios; i++) {
2190 struct folio *folio = fbatch.folios[i];
2191
2192 done_index = folio_next_index(folio);
2193 /*
2194 * At this point we hold neither the i_pages lock nor
2195 * the page lock: the page may be truncated or
2196 * invalidated (changing page->mapping to NULL),
2197 * or even swizzled back from swapper_space to
2198 * tmpfs file mapping
2199 */
2200 if (!folio_trylock(folio)) {
2201 submit_write_bio(bio_ctrl, 0);
2202 folio_lock(folio);
2203 }
2204
2205 if (unlikely(folio->mapping != mapping)) {
2206 folio_unlock(folio);
2207 continue;
2208 }
2209
2210 if (!folio_test_dirty(folio)) {
2211 /* Someone wrote it for us. */
2212 folio_unlock(folio);
2213 continue;
2214 }
2215
2216 /*
2217 * For subpage case, compression can lead to mixed
2218 * writeback and dirty flags, e.g:
2219 * 0 32K 64K 96K 128K
2220 * | |//////||/////| |//|
2221 *
2222 * In above case, [32K, 96K) is asynchronously submitted
2223 * for compression, and [124K, 128K) needs to be written back.
2224 *
2225 * If we didn't wait wrtiteback for page 64K, [128K, 128K)
2226 * won't be submitted as the page still has writeback flag
2227 * and will be skipped in the next check.
2228 *
2229 * This mixed writeback and dirty case is only possible for
2230 * subpage case.
2231 *
2232 * TODO: Remove this check after migrating compression to
2233 * regular submission.
2234 */
2235 if (wbc->sync_mode != WB_SYNC_NONE ||
2236 btrfs_is_subpage(inode_to_fs_info(inode), mapping)) {
2237 if (folio_test_writeback(folio))
2238 submit_write_bio(bio_ctrl, 0);
2239 folio_wait_writeback(folio);
2240 }
2241
2242 if (folio_test_writeback(folio) ||
2243 !folio_clear_dirty_for_io(folio)) {
2244 folio_unlock(folio);
2245 continue;
2246 }
2247
2248 ret = extent_writepage(folio, bio_ctrl);
2249 if (ret < 0) {
2250 done = 1;
2251 break;
2252 }
2253
2254 /*
2255 * The filesystem may choose to bump up nr_to_write.
2256 * We have to make sure to honor the new nr_to_write
2257 * at any time.
2258 */
2259 nr_to_write_done = (wbc->sync_mode == WB_SYNC_NONE &&
2260 wbc->nr_to_write <= 0);
2261 }
2262 folio_batch_release(&fbatch);
2263 cond_resched();
2264 }
2265 if (!scanned && !done) {
2266 /*
2267 * We hit the last page and there is more work to be done: wrap
2268 * back to the start of the file
2269 */
2270 scanned = 1;
2271 index = 0;
2272
2273 /*
2274 * If we're looping we could run into a page that is locked by a
2275 * writer and that writer could be waiting on writeback for a
2276 * page in our current bio, and thus deadlock, so flush the
2277 * write bio here.
2278 */
2279 submit_write_bio(bio_ctrl, 0);
2280 goto retry;
2281 }
2282
2283 if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
2284 mapping->writeback_index = done_index;
2285
2286 btrfs_add_delayed_iput(BTRFS_I(inode));
2287 return ret;
2288}
2289
2290/*
2291 * Submit the pages in the range to bio for call sites which delalloc range has
2292 * already been ran (aka, ordered extent inserted) and all pages are still
2293 * locked.
2294 */
2295void extent_write_locked_range(struct inode *inode, const struct folio *locked_folio,
2296 u64 start, u64 end, struct writeback_control *wbc,
2297 bool pages_dirty)
2298{
2299 bool found_error = false;
2300 int ret = 0;
2301 struct address_space *mapping = inode->i_mapping;
2302 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
2303 const u32 sectorsize = fs_info->sectorsize;
2304 loff_t i_size = i_size_read(inode);
2305 u64 cur = start;
2306 struct btrfs_bio_ctrl bio_ctrl = {
2307 .wbc = wbc,
2308 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2309 };
2310
2311 if (wbc->no_cgroup_owner)
2312 bio_ctrl.opf |= REQ_BTRFS_CGROUP_PUNT;
2313
2314 ASSERT(IS_ALIGNED(start, sectorsize) && IS_ALIGNED(end + 1, sectorsize));
2315
2316 while (cur <= end) {
2317 u64 cur_end = min(round_down(cur, PAGE_SIZE) + PAGE_SIZE - 1, end);
2318 u32 cur_len = cur_end + 1 - cur;
2319 struct folio *folio;
2320
2321 folio = filemap_get_folio(mapping, cur >> PAGE_SHIFT);
2322
2323 /*
2324 * This shouldn't happen, the pages are pinned and locked, this
2325 * code is just in case, but shouldn't actually be run.
2326 */
2327 if (IS_ERR(folio)) {
2328 btrfs_mark_ordered_io_finished(BTRFS_I(inode), NULL,
2329 cur, cur_len, false);
2330 mapping_set_error(mapping, PTR_ERR(folio));
2331 cur = cur_end + 1;
2332 continue;
2333 }
2334
2335 ASSERT(folio_test_locked(folio));
2336 if (pages_dirty && folio != locked_folio)
2337 ASSERT(folio_test_dirty(folio));
2338
2339 /*
2340 * Set the submission bitmap to submit all sectors.
2341 * extent_writepage_io() will do the truncation correctly.
2342 */
2343 bio_ctrl.submit_bitmap = (unsigned long)-1;
2344 ret = extent_writepage_io(BTRFS_I(inode), folio, cur, cur_len,
2345 &bio_ctrl, i_size);
2346 if (ret == 1)
2347 goto next_page;
2348
2349 if (ret)
2350 mapping_set_error(mapping, ret);
2351 btrfs_folio_end_lock(fs_info, folio, cur, cur_len);
2352 if (ret < 0)
2353 found_error = true;
2354next_page:
2355 folio_put(folio);
2356 cur = cur_end + 1;
2357 }
2358
2359 submit_write_bio(&bio_ctrl, found_error ? ret : 0);
2360}
2361
2362int btrfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
2363{
2364 struct inode *inode = mapping->host;
2365 int ret = 0;
2366 struct btrfs_bio_ctrl bio_ctrl = {
2367 .wbc = wbc,
2368 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2369 };
2370
2371 /*
2372 * Allow only a single thread to do the reloc work in zoned mode to
2373 * protect the write pointer updates.
2374 */
2375 btrfs_zoned_data_reloc_lock(BTRFS_I(inode));
2376 ret = extent_write_cache_pages(mapping, &bio_ctrl);
2377 submit_write_bio(&bio_ctrl, ret);
2378 btrfs_zoned_data_reloc_unlock(BTRFS_I(inode));
2379 return ret;
2380}
2381
2382void btrfs_readahead(struct readahead_control *rac)
2383{
2384 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ | REQ_RAHEAD };
2385 struct folio *folio;
2386 struct btrfs_inode *inode = BTRFS_I(rac->mapping->host);
2387 const u64 start = readahead_pos(rac);
2388 const u64 end = start + readahead_length(rac) - 1;
2389 struct extent_state *cached_state = NULL;
2390 struct extent_map *em_cached = NULL;
2391 u64 prev_em_start = (u64)-1;
2392
2393 btrfs_lock_and_flush_ordered_range(inode, start, end, &cached_state);
2394
2395 while ((folio = readahead_folio(rac)) != NULL)
2396 btrfs_do_readpage(folio, &em_cached, &bio_ctrl, &prev_em_start);
2397
2398 unlock_extent(&inode->io_tree, start, end, &cached_state);
2399
2400 if (em_cached)
2401 free_extent_map(em_cached);
2402 submit_one_bio(&bio_ctrl);
2403}
2404
2405/*
2406 * basic invalidate_folio code, this waits on any locked or writeback
2407 * ranges corresponding to the folio, and then deletes any extent state
2408 * records from the tree
2409 */
2410int extent_invalidate_folio(struct extent_io_tree *tree,
2411 struct folio *folio, size_t offset)
2412{
2413 struct extent_state *cached_state = NULL;
2414 u64 start = folio_pos(folio);
2415 u64 end = start + folio_size(folio) - 1;
2416 size_t blocksize = folio_to_fs_info(folio)->sectorsize;
2417
2418 /* This function is only called for the btree inode */
2419 ASSERT(tree->owner == IO_TREE_BTREE_INODE_IO);
2420
2421 start += ALIGN(offset, blocksize);
2422 if (start > end)
2423 return 0;
2424
2425 lock_extent(tree, start, end, &cached_state);
2426 folio_wait_writeback(folio);
2427
2428 /*
2429 * Currently for btree io tree, only EXTENT_LOCKED is utilized,
2430 * so here we only need to unlock the extent range to free any
2431 * existing extent state.
2432 */
2433 unlock_extent(tree, start, end, &cached_state);
2434 return 0;
2435}
2436
2437/*
2438 * a helper for release_folio, this tests for areas of the page that
2439 * are locked or under IO and drops the related state bits if it is safe
2440 * to drop the page.
2441 */
2442static bool try_release_extent_state(struct extent_io_tree *tree,
2443 struct folio *folio)
2444{
2445 u64 start = folio_pos(folio);
2446 u64 end = start + PAGE_SIZE - 1;
2447 bool ret;
2448
2449 if (test_range_bit_exists(tree, start, end, EXTENT_LOCKED)) {
2450 ret = false;
2451 } else {
2452 u32 clear_bits = ~(EXTENT_LOCKED | EXTENT_NODATASUM |
2453 EXTENT_DELALLOC_NEW | EXTENT_CTLBITS |
2454 EXTENT_QGROUP_RESERVED);
2455 int ret2;
2456
2457 /*
2458 * At this point we can safely clear everything except the
2459 * locked bit, the nodatasum bit and the delalloc new bit.
2460 * The delalloc new bit will be cleared by ordered extent
2461 * completion.
2462 */
2463 ret2 = __clear_extent_bit(tree, start, end, clear_bits, NULL, NULL);
2464
2465 /* if clear_extent_bit failed for enomem reasons,
2466 * we can't allow the release to continue.
2467 */
2468 if (ret2 < 0)
2469 ret = false;
2470 else
2471 ret = true;
2472 }
2473 return ret;
2474}
2475
2476/*
2477 * a helper for release_folio. As long as there are no locked extents
2478 * in the range corresponding to the page, both state records and extent
2479 * map records are removed
2480 */
2481bool try_release_extent_mapping(struct folio *folio, gfp_t mask)
2482{
2483 u64 start = folio_pos(folio);
2484 u64 end = start + PAGE_SIZE - 1;
2485 struct btrfs_inode *inode = folio_to_inode(folio);
2486 struct extent_io_tree *io_tree = &inode->io_tree;
2487
2488 while (start <= end) {
2489 const u64 cur_gen = btrfs_get_fs_generation(inode->root->fs_info);
2490 const u64 len = end - start + 1;
2491 struct extent_map_tree *extent_tree = &inode->extent_tree;
2492 struct extent_map *em;
2493
2494 write_lock(&extent_tree->lock);
2495 em = lookup_extent_mapping(extent_tree, start, len);
2496 if (!em) {
2497 write_unlock(&extent_tree->lock);
2498 break;
2499 }
2500 if ((em->flags & EXTENT_FLAG_PINNED) || em->start != start) {
2501 write_unlock(&extent_tree->lock);
2502 free_extent_map(em);
2503 break;
2504 }
2505 if (test_range_bit_exists(io_tree, em->start,
2506 extent_map_end(em) - 1, EXTENT_LOCKED))
2507 goto next;
2508 /*
2509 * If it's not in the list of modified extents, used by a fast
2510 * fsync, we can remove it. If it's being logged we can safely
2511 * remove it since fsync took an extra reference on the em.
2512 */
2513 if (list_empty(&em->list) || (em->flags & EXTENT_FLAG_LOGGING))
2514 goto remove_em;
2515 /*
2516 * If it's in the list of modified extents, remove it only if
2517 * its generation is older then the current one, in which case
2518 * we don't need it for a fast fsync. Otherwise don't remove it,
2519 * we could be racing with an ongoing fast fsync that could miss
2520 * the new extent.
2521 */
2522 if (em->generation >= cur_gen)
2523 goto next;
2524remove_em:
2525 /*
2526 * We only remove extent maps that are not in the list of
2527 * modified extents or that are in the list but with a
2528 * generation lower then the current generation, so there is no
2529 * need to set the full fsync flag on the inode (it hurts the
2530 * fsync performance for workloads with a data size that exceeds
2531 * or is close to the system's memory).
2532 */
2533 remove_extent_mapping(inode, em);
2534 /* Once for the inode's extent map tree. */
2535 free_extent_map(em);
2536next:
2537 start = extent_map_end(em);
2538 write_unlock(&extent_tree->lock);
2539
2540 /* Once for us, for the lookup_extent_mapping() reference. */
2541 free_extent_map(em);
2542
2543 if (need_resched()) {
2544 /*
2545 * If we need to resched but we can't block just exit
2546 * and leave any remaining extent maps.
2547 */
2548 if (!gfpflags_allow_blocking(mask))
2549 break;
2550
2551 cond_resched();
2552 }
2553 }
2554 return try_release_extent_state(io_tree, folio);
2555}
2556
2557static void __free_extent_buffer(struct extent_buffer *eb)
2558{
2559 kmem_cache_free(extent_buffer_cache, eb);
2560}
2561
2562static int extent_buffer_under_io(const struct extent_buffer *eb)
2563{
2564 return (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
2565 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
2566}
2567
2568static bool folio_range_has_eb(struct folio *folio)
2569{
2570 struct btrfs_subpage *subpage;
2571
2572 lockdep_assert_held(&folio->mapping->i_private_lock);
2573
2574 if (folio_test_private(folio)) {
2575 subpage = folio_get_private(folio);
2576 if (atomic_read(&subpage->eb_refs))
2577 return true;
2578 }
2579 return false;
2580}
2581
2582static void detach_extent_buffer_folio(const struct extent_buffer *eb, struct folio *folio)
2583{
2584 struct btrfs_fs_info *fs_info = eb->fs_info;
2585 const bool mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
2586
2587 /*
2588 * For mapped eb, we're going to change the folio private, which should
2589 * be done under the i_private_lock.
2590 */
2591 if (mapped)
2592 spin_lock(&folio->mapping->i_private_lock);
2593
2594 if (!folio_test_private(folio)) {
2595 if (mapped)
2596 spin_unlock(&folio->mapping->i_private_lock);
2597 return;
2598 }
2599
2600 if (fs_info->nodesize >= PAGE_SIZE) {
2601 /*
2602 * We do this since we'll remove the pages after we've
2603 * removed the eb from the radix tree, so we could race
2604 * and have this page now attached to the new eb. So
2605 * only clear folio if it's still connected to
2606 * this eb.
2607 */
2608 if (folio_test_private(folio) && folio_get_private(folio) == eb) {
2609 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
2610 BUG_ON(folio_test_dirty(folio));
2611 BUG_ON(folio_test_writeback(folio));
2612 /* We need to make sure we haven't be attached to a new eb. */
2613 folio_detach_private(folio);
2614 }
2615 if (mapped)
2616 spin_unlock(&folio->mapping->i_private_lock);
2617 return;
2618 }
2619
2620 /*
2621 * For subpage, we can have dummy eb with folio private attached. In
2622 * this case, we can directly detach the private as such folio is only
2623 * attached to one dummy eb, no sharing.
2624 */
2625 if (!mapped) {
2626 btrfs_detach_subpage(fs_info, folio);
2627 return;
2628 }
2629
2630 btrfs_folio_dec_eb_refs(fs_info, folio);
2631
2632 /*
2633 * We can only detach the folio private if there are no other ebs in the
2634 * page range and no unfinished IO.
2635 */
2636 if (!folio_range_has_eb(folio))
2637 btrfs_detach_subpage(fs_info, folio);
2638
2639 spin_unlock(&folio->mapping->i_private_lock);
2640}
2641
2642/* Release all pages attached to the extent buffer */
2643static void btrfs_release_extent_buffer_pages(const struct extent_buffer *eb)
2644{
2645 ASSERT(!extent_buffer_under_io(eb));
2646
2647 for (int i = 0; i < INLINE_EXTENT_BUFFER_PAGES; i++) {
2648 struct folio *folio = eb->folios[i];
2649
2650 if (!folio)
2651 continue;
2652
2653 detach_extent_buffer_folio(eb, folio);
2654
2655 /* One for when we allocated the folio. */
2656 folio_put(folio);
2657 }
2658}
2659
2660/*
2661 * Helper for releasing the extent buffer.
2662 */
2663static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
2664{
2665 btrfs_release_extent_buffer_pages(eb);
2666 btrfs_leak_debug_del_eb(eb);
2667 __free_extent_buffer(eb);
2668}
2669
2670static struct extent_buffer *
2671__alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
2672 unsigned long len)
2673{
2674 struct extent_buffer *eb = NULL;
2675
2676 eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
2677 eb->start = start;
2678 eb->len = len;
2679 eb->fs_info = fs_info;
2680 init_rwsem(&eb->lock);
2681
2682 btrfs_leak_debug_add_eb(eb);
2683
2684 spin_lock_init(&eb->refs_lock);
2685 atomic_set(&eb->refs, 1);
2686
2687 ASSERT(len <= BTRFS_MAX_METADATA_BLOCKSIZE);
2688
2689 return eb;
2690}
2691
2692struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
2693{
2694 struct extent_buffer *new;
2695 int num_folios = num_extent_folios(src);
2696 int ret;
2697
2698 new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
2699 if (new == NULL)
2700 return NULL;
2701
2702 /*
2703 * Set UNMAPPED before calling btrfs_release_extent_buffer(), as
2704 * btrfs_release_extent_buffer() have different behavior for
2705 * UNMAPPED subpage extent buffer.
2706 */
2707 set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
2708
2709 ret = alloc_eb_folio_array(new, false);
2710 if (ret) {
2711 btrfs_release_extent_buffer(new);
2712 return NULL;
2713 }
2714
2715 for (int i = 0; i < num_folios; i++) {
2716 struct folio *folio = new->folios[i];
2717
2718 ret = attach_extent_buffer_folio(new, folio, NULL);
2719 if (ret < 0) {
2720 btrfs_release_extent_buffer(new);
2721 return NULL;
2722 }
2723 WARN_ON(folio_test_dirty(folio));
2724 }
2725 copy_extent_buffer_full(new, src);
2726 set_extent_buffer_uptodate(new);
2727
2728 return new;
2729}
2730
2731struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
2732 u64 start, unsigned long len)
2733{
2734 struct extent_buffer *eb;
2735 int num_folios = 0;
2736 int ret;
2737
2738 eb = __alloc_extent_buffer(fs_info, start, len);
2739 if (!eb)
2740 return NULL;
2741
2742 ret = alloc_eb_folio_array(eb, false);
2743 if (ret)
2744 goto err;
2745
2746 num_folios = num_extent_folios(eb);
2747 for (int i = 0; i < num_folios; i++) {
2748 ret = attach_extent_buffer_folio(eb, eb->folios[i], NULL);
2749 if (ret < 0)
2750 goto err;
2751 }
2752
2753 set_extent_buffer_uptodate(eb);
2754 btrfs_set_header_nritems(eb, 0);
2755 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
2756
2757 return eb;
2758err:
2759 for (int i = 0; i < num_folios; i++) {
2760 if (eb->folios[i]) {
2761 detach_extent_buffer_folio(eb, eb->folios[i]);
2762 folio_put(eb->folios[i]);
2763 }
2764 }
2765 __free_extent_buffer(eb);
2766 return NULL;
2767}
2768
2769struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
2770 u64 start)
2771{
2772 return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
2773}
2774
2775static void check_buffer_tree_ref(struct extent_buffer *eb)
2776{
2777 int refs;
2778 /*
2779 * The TREE_REF bit is first set when the extent_buffer is added
2780 * to the radix tree. It is also reset, if unset, when a new reference
2781 * is created by find_extent_buffer.
2782 *
2783 * It is only cleared in two cases: freeing the last non-tree
2784 * reference to the extent_buffer when its STALE bit is set or
2785 * calling release_folio when the tree reference is the only reference.
2786 *
2787 * In both cases, care is taken to ensure that the extent_buffer's
2788 * pages are not under io. However, release_folio can be concurrently
2789 * called with creating new references, which is prone to race
2790 * conditions between the calls to check_buffer_tree_ref in those
2791 * codepaths and clearing TREE_REF in try_release_extent_buffer.
2792 *
2793 * The actual lifetime of the extent_buffer in the radix tree is
2794 * adequately protected by the refcount, but the TREE_REF bit and
2795 * its corresponding reference are not. To protect against this
2796 * class of races, we call check_buffer_tree_ref from the codepaths
2797 * which trigger io. Note that once io is initiated, TREE_REF can no
2798 * longer be cleared, so that is the moment at which any such race is
2799 * best fixed.
2800 */
2801 refs = atomic_read(&eb->refs);
2802 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
2803 return;
2804
2805 spin_lock(&eb->refs_lock);
2806 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
2807 atomic_inc(&eb->refs);
2808 spin_unlock(&eb->refs_lock);
2809}
2810
2811static void mark_extent_buffer_accessed(struct extent_buffer *eb)
2812{
2813 int num_folios= num_extent_folios(eb);
2814
2815 check_buffer_tree_ref(eb);
2816
2817 for (int i = 0; i < num_folios; i++)
2818 folio_mark_accessed(eb->folios[i]);
2819}
2820
2821struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
2822 u64 start)
2823{
2824 struct extent_buffer *eb;
2825
2826 eb = find_extent_buffer_nolock(fs_info, start);
2827 if (!eb)
2828 return NULL;
2829 /*
2830 * Lock our eb's refs_lock to avoid races with free_extent_buffer().
2831 * When we get our eb it might be flagged with EXTENT_BUFFER_STALE and
2832 * another task running free_extent_buffer() might have seen that flag
2833 * set, eb->refs == 2, that the buffer isn't under IO (dirty and
2834 * writeback flags not set) and it's still in the tree (flag
2835 * EXTENT_BUFFER_TREE_REF set), therefore being in the process of
2836 * decrementing the extent buffer's reference count twice. So here we
2837 * could race and increment the eb's reference count, clear its stale
2838 * flag, mark it as dirty and drop our reference before the other task
2839 * finishes executing free_extent_buffer, which would later result in
2840 * an attempt to free an extent buffer that is dirty.
2841 */
2842 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
2843 spin_lock(&eb->refs_lock);
2844 spin_unlock(&eb->refs_lock);
2845 }
2846 mark_extent_buffer_accessed(eb);
2847 return eb;
2848}
2849
2850#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
2851struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
2852 u64 start)
2853{
2854 struct extent_buffer *eb, *exists = NULL;
2855 int ret;
2856
2857 eb = find_extent_buffer(fs_info, start);
2858 if (eb)
2859 return eb;
2860 eb = alloc_dummy_extent_buffer(fs_info, start);
2861 if (!eb)
2862 return ERR_PTR(-ENOMEM);
2863 eb->fs_info = fs_info;
2864again:
2865 ret = radix_tree_preload(GFP_NOFS);
2866 if (ret) {
2867 exists = ERR_PTR(ret);
2868 goto free_eb;
2869 }
2870 spin_lock(&fs_info->buffer_lock);
2871 ret = radix_tree_insert(&fs_info->buffer_radix,
2872 start >> fs_info->sectorsize_bits, eb);
2873 spin_unlock(&fs_info->buffer_lock);
2874 radix_tree_preload_end();
2875 if (ret == -EEXIST) {
2876 exists = find_extent_buffer(fs_info, start);
2877 if (exists)
2878 goto free_eb;
2879 else
2880 goto again;
2881 }
2882 check_buffer_tree_ref(eb);
2883 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
2884
2885 return eb;
2886free_eb:
2887 btrfs_release_extent_buffer(eb);
2888 return exists;
2889}
2890#endif
2891
2892static struct extent_buffer *grab_extent_buffer(
2893 struct btrfs_fs_info *fs_info, struct page *page)
2894{
2895 struct folio *folio = page_folio(page);
2896 struct extent_buffer *exists;
2897
2898 lockdep_assert_held(&page->mapping->i_private_lock);
2899
2900 /*
2901 * For subpage case, we completely rely on radix tree to ensure we
2902 * don't try to insert two ebs for the same bytenr. So here we always
2903 * return NULL and just continue.
2904 */
2905 if (fs_info->nodesize < PAGE_SIZE)
2906 return NULL;
2907
2908 /* Page not yet attached to an extent buffer */
2909 if (!folio_test_private(folio))
2910 return NULL;
2911
2912 /*
2913 * We could have already allocated an eb for this page and attached one
2914 * so lets see if we can get a ref on the existing eb, and if we can we
2915 * know it's good and we can just return that one, else we know we can
2916 * just overwrite folio private.
2917 */
2918 exists = folio_get_private(folio);
2919 if (atomic_inc_not_zero(&exists->refs))
2920 return exists;
2921
2922 WARN_ON(PageDirty(page));
2923 folio_detach_private(folio);
2924 return NULL;
2925}
2926
2927static int check_eb_alignment(struct btrfs_fs_info *fs_info, u64 start)
2928{
2929 if (!IS_ALIGNED(start, fs_info->sectorsize)) {
2930 btrfs_err(fs_info, "bad tree block start %llu", start);
2931 return -EINVAL;
2932 }
2933
2934 if (fs_info->nodesize < PAGE_SIZE &&
2935 offset_in_page(start) + fs_info->nodesize > PAGE_SIZE) {
2936 btrfs_err(fs_info,
2937 "tree block crosses page boundary, start %llu nodesize %u",
2938 start, fs_info->nodesize);
2939 return -EINVAL;
2940 }
2941 if (fs_info->nodesize >= PAGE_SIZE &&
2942 !PAGE_ALIGNED(start)) {
2943 btrfs_err(fs_info,
2944 "tree block is not page aligned, start %llu nodesize %u",
2945 start, fs_info->nodesize);
2946 return -EINVAL;
2947 }
2948 if (!IS_ALIGNED(start, fs_info->nodesize) &&
2949 !test_and_set_bit(BTRFS_FS_UNALIGNED_TREE_BLOCK, &fs_info->flags)) {
2950 btrfs_warn(fs_info,
2951"tree block not nodesize aligned, start %llu nodesize %u, can be resolved by a full metadata balance",
2952 start, fs_info->nodesize);
2953 }
2954 return 0;
2955}
2956
2957
2958/*
2959 * Return 0 if eb->folios[i] is attached to btree inode successfully.
2960 * Return >0 if there is already another extent buffer for the range,
2961 * and @found_eb_ret would be updated.
2962 * Return -EAGAIN if the filemap has an existing folio but with different size
2963 * than @eb.
2964 * The caller needs to free the existing folios and retry using the same order.
2965 */
2966static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
2967 struct btrfs_subpage *prealloc,
2968 struct extent_buffer **found_eb_ret)
2969{
2970
2971 struct btrfs_fs_info *fs_info = eb->fs_info;
2972 struct address_space *mapping = fs_info->btree_inode->i_mapping;
2973 const unsigned long index = eb->start >> PAGE_SHIFT;
2974 struct folio *existing_folio = NULL;
2975 int ret;
2976
2977 ASSERT(found_eb_ret);
2978
2979 /* Caller should ensure the folio exists. */
2980 ASSERT(eb->folios[i]);
2981
2982retry:
2983 ret = filemap_add_folio(mapping, eb->folios[i], index + i,
2984 GFP_NOFS | __GFP_NOFAIL);
2985 if (!ret)
2986 goto finish;
2987
2988 existing_folio = filemap_lock_folio(mapping, index + i);
2989 /* The page cache only exists for a very short time, just retry. */
2990 if (IS_ERR(existing_folio)) {
2991 existing_folio = NULL;
2992 goto retry;
2993 }
2994
2995 /* For now, we should only have single-page folios for btree inode. */
2996 ASSERT(folio_nr_pages(existing_folio) == 1);
2997
2998 if (folio_size(existing_folio) != eb->folio_size) {
2999 folio_unlock(existing_folio);
3000 folio_put(existing_folio);
3001 return -EAGAIN;
3002 }
3003
3004finish:
3005 spin_lock(&mapping->i_private_lock);
3006 if (existing_folio && fs_info->nodesize < PAGE_SIZE) {
3007 /* We're going to reuse the existing page, can drop our folio now. */
3008 __free_page(folio_page(eb->folios[i], 0));
3009 eb->folios[i] = existing_folio;
3010 } else if (existing_folio) {
3011 struct extent_buffer *existing_eb;
3012
3013 existing_eb = grab_extent_buffer(fs_info,
3014 folio_page(existing_folio, 0));
3015 if (existing_eb) {
3016 /* The extent buffer still exists, we can use it directly. */
3017 *found_eb_ret = existing_eb;
3018 spin_unlock(&mapping->i_private_lock);
3019 folio_unlock(existing_folio);
3020 folio_put(existing_folio);
3021 return 1;
3022 }
3023 /* The extent buffer no longer exists, we can reuse the folio. */
3024 __free_page(folio_page(eb->folios[i], 0));
3025 eb->folios[i] = existing_folio;
3026 }
3027 eb->folio_size = folio_size(eb->folios[i]);
3028 eb->folio_shift = folio_shift(eb->folios[i]);
3029 /* Should not fail, as we have preallocated the memory. */
3030 ret = attach_extent_buffer_folio(eb, eb->folios[i], prealloc);
3031 ASSERT(!ret);
3032 /*
3033 * To inform we have an extra eb under allocation, so that
3034 * detach_extent_buffer_page() won't release the folio private when the
3035 * eb hasn't been inserted into radix tree yet.
3036 *
3037 * The ref will be decreased when the eb releases the page, in
3038 * detach_extent_buffer_page(). Thus needs no special handling in the
3039 * error path.
3040 */
3041 btrfs_folio_inc_eb_refs(fs_info, eb->folios[i]);
3042 spin_unlock(&mapping->i_private_lock);
3043 return 0;
3044}
3045
3046struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
3047 u64 start, u64 owner_root, int level)
3048{
3049 unsigned long len = fs_info->nodesize;
3050 int num_folios;
3051 int attached = 0;
3052 struct extent_buffer *eb;
3053 struct extent_buffer *existing_eb = NULL;
3054 struct btrfs_subpage *prealloc = NULL;
3055 u64 lockdep_owner = owner_root;
3056 bool page_contig = true;
3057 int uptodate = 1;
3058 int ret;
3059
3060 if (check_eb_alignment(fs_info, start))
3061 return ERR_PTR(-EINVAL);
3062
3063#if BITS_PER_LONG == 32
3064 if (start >= MAX_LFS_FILESIZE) {
3065 btrfs_err_rl(fs_info,
3066 "extent buffer %llu is beyond 32bit page cache limit", start);
3067 btrfs_err_32bit_limit(fs_info);
3068 return ERR_PTR(-EOVERFLOW);
3069 }
3070 if (start >= BTRFS_32BIT_EARLY_WARN_THRESHOLD)
3071 btrfs_warn_32bit_limit(fs_info);
3072#endif
3073
3074 eb = find_extent_buffer(fs_info, start);
3075 if (eb)
3076 return eb;
3077
3078 eb = __alloc_extent_buffer(fs_info, start, len);
3079 if (!eb)
3080 return ERR_PTR(-ENOMEM);
3081
3082 /*
3083 * The reloc trees are just snapshots, so we need them to appear to be
3084 * just like any other fs tree WRT lockdep.
3085 */
3086 if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID)
3087 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
3088
3089 btrfs_set_buffer_lockdep_class(lockdep_owner, eb, level);
3090
3091 /*
3092 * Preallocate folio private for subpage case, so that we won't
3093 * allocate memory with i_private_lock nor page lock hold.
3094 *
3095 * The memory will be freed by attach_extent_buffer_page() or freed
3096 * manually if we exit earlier.
3097 */
3098 if (fs_info->nodesize < PAGE_SIZE) {
3099 prealloc = btrfs_alloc_subpage(fs_info, BTRFS_SUBPAGE_METADATA);
3100 if (IS_ERR(prealloc)) {
3101 ret = PTR_ERR(prealloc);
3102 goto out;
3103 }
3104 }
3105
3106reallocate:
3107 /* Allocate all pages first. */
3108 ret = alloc_eb_folio_array(eb, true);
3109 if (ret < 0) {
3110 btrfs_free_subpage(prealloc);
3111 goto out;
3112 }
3113
3114 num_folios = num_extent_folios(eb);
3115 /* Attach all pages to the filemap. */
3116 for (int i = 0; i < num_folios; i++) {
3117 struct folio *folio;
3118
3119 ret = attach_eb_folio_to_filemap(eb, i, prealloc, &existing_eb);
3120 if (ret > 0) {
3121 ASSERT(existing_eb);
3122 goto out;
3123 }
3124
3125 /*
3126 * TODO: Special handling for a corner case where the order of
3127 * folios mismatch between the new eb and filemap.
3128 *
3129 * This happens when:
3130 *
3131 * - the new eb is using higher order folio
3132 *
3133 * - the filemap is still using 0-order folios for the range
3134 * This can happen at the previous eb allocation, and we don't
3135 * have higher order folio for the call.
3136 *
3137 * - the existing eb has already been freed
3138 *
3139 * In this case, we have to free the existing folios first, and
3140 * re-allocate using the same order.
3141 * Thankfully this is not going to happen yet, as we're still
3142 * using 0-order folios.
3143 */
3144 if (unlikely(ret == -EAGAIN)) {
3145 ASSERT(0);
3146 goto reallocate;
3147 }
3148 attached++;
3149
3150 /*
3151 * Only after attach_eb_folio_to_filemap(), eb->folios[] is
3152 * reliable, as we may choose to reuse the existing page cache
3153 * and free the allocated page.
3154 */
3155 folio = eb->folios[i];
3156 WARN_ON(btrfs_folio_test_dirty(fs_info, folio, eb->start, eb->len));
3157
3158 /*
3159 * Check if the current page is physically contiguous with previous eb
3160 * page.
3161 * At this stage, either we allocated a large folio, thus @i
3162 * would only be 0, or we fall back to per-page allocation.
3163 */
3164 if (i && folio_page(eb->folios[i - 1], 0) + 1 != folio_page(folio, 0))
3165 page_contig = false;
3166
3167 if (!btrfs_folio_test_uptodate(fs_info, folio, eb->start, eb->len))
3168 uptodate = 0;
3169
3170 /*
3171 * We can't unlock the pages just yet since the extent buffer
3172 * hasn't been properly inserted in the radix tree, this
3173 * opens a race with btree_release_folio which can free a page
3174 * while we are still filling in all pages for the buffer and
3175 * we could crash.
3176 */
3177 }
3178 if (uptodate)
3179 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3180 /* All pages are physically contiguous, can skip cross page handling. */
3181 if (page_contig)
3182 eb->addr = folio_address(eb->folios[0]) + offset_in_page(eb->start);
3183again:
3184 ret = radix_tree_preload(GFP_NOFS);
3185 if (ret)
3186 goto out;
3187
3188 spin_lock(&fs_info->buffer_lock);
3189 ret = radix_tree_insert(&fs_info->buffer_radix,
3190 start >> fs_info->sectorsize_bits, eb);
3191 spin_unlock(&fs_info->buffer_lock);
3192 radix_tree_preload_end();
3193 if (ret == -EEXIST) {
3194 ret = 0;
3195 existing_eb = find_extent_buffer(fs_info, start);
3196 if (existing_eb)
3197 goto out;
3198 else
3199 goto again;
3200 }
3201 /* add one reference for the tree */
3202 check_buffer_tree_ref(eb);
3203 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
3204
3205 /*
3206 * Now it's safe to unlock the pages because any calls to
3207 * btree_release_folio will correctly detect that a page belongs to a
3208 * live buffer and won't free them prematurely.
3209 */
3210 for (int i = 0; i < num_folios; i++)
3211 unlock_page(folio_page(eb->folios[i], 0));
3212 return eb;
3213
3214out:
3215 WARN_ON(!atomic_dec_and_test(&eb->refs));
3216
3217 /*
3218 * Any attached folios need to be detached before we unlock them. This
3219 * is because when we're inserting our new folios into the mapping, and
3220 * then attaching our eb to that folio. If we fail to insert our folio
3221 * we'll lookup the folio for that index, and grab that EB. We do not
3222 * want that to grab this eb, as we're getting ready to free it. So we
3223 * have to detach it first and then unlock it.
3224 *
3225 * We have to drop our reference and NULL it out here because in the
3226 * subpage case detaching does a btrfs_folio_dec_eb_refs() for our eb.
3227 * Below when we call btrfs_release_extent_buffer() we will call
3228 * detach_extent_buffer_folio() on our remaining pages in the !subpage
3229 * case. If we left eb->folios[i] populated in the subpage case we'd
3230 * double put our reference and be super sad.
3231 */
3232 for (int i = 0; i < attached; i++) {
3233 ASSERT(eb->folios[i]);
3234 detach_extent_buffer_folio(eb, eb->folios[i]);
3235 unlock_page(folio_page(eb->folios[i], 0));
3236 folio_put(eb->folios[i]);
3237 eb->folios[i] = NULL;
3238 }
3239 /*
3240 * Now all pages of that extent buffer is unmapped, set UNMAPPED flag,
3241 * so it can be cleaned up without utilizing page->mapping.
3242 */
3243 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3244
3245 btrfs_release_extent_buffer(eb);
3246 if (ret < 0)
3247 return ERR_PTR(ret);
3248 ASSERT(existing_eb);
3249 return existing_eb;
3250}
3251
3252static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3253{
3254 struct extent_buffer *eb =
3255 container_of(head, struct extent_buffer, rcu_head);
3256
3257 __free_extent_buffer(eb);
3258}
3259
3260static int release_extent_buffer(struct extent_buffer *eb)
3261 __releases(&eb->refs_lock)
3262{
3263 lockdep_assert_held(&eb->refs_lock);
3264
3265 WARN_ON(atomic_read(&eb->refs) == 0);
3266 if (atomic_dec_and_test(&eb->refs)) {
3267 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
3268 struct btrfs_fs_info *fs_info = eb->fs_info;
3269
3270 spin_unlock(&eb->refs_lock);
3271
3272 spin_lock(&fs_info->buffer_lock);
3273 radix_tree_delete(&fs_info->buffer_radix,
3274 eb->start >> fs_info->sectorsize_bits);
3275 spin_unlock(&fs_info->buffer_lock);
3276 } else {
3277 spin_unlock(&eb->refs_lock);
3278 }
3279
3280 btrfs_leak_debug_del_eb(eb);
3281 /* Should be safe to release our pages at this point */
3282 btrfs_release_extent_buffer_pages(eb);
3283#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3284 if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
3285 __free_extent_buffer(eb);
3286 return 1;
3287 }
3288#endif
3289 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3290 return 1;
3291 }
3292 spin_unlock(&eb->refs_lock);
3293
3294 return 0;
3295}
3296
3297void free_extent_buffer(struct extent_buffer *eb)
3298{
3299 int refs;
3300 if (!eb)
3301 return;
3302
3303 refs = atomic_read(&eb->refs);
3304 while (1) {
3305 if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
3306 || (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
3307 refs == 1))
3308 break;
3309 if (atomic_try_cmpxchg(&eb->refs, &refs, refs - 1))
3310 return;
3311 }
3312
3313 spin_lock(&eb->refs_lock);
3314 if (atomic_read(&eb->refs) == 2 &&
3315 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
3316 !extent_buffer_under_io(eb) &&
3317 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3318 atomic_dec(&eb->refs);
3319
3320 /*
3321 * I know this is terrible, but it's temporary until we stop tracking
3322 * the uptodate bits and such for the extent buffers.
3323 */
3324 release_extent_buffer(eb);
3325}
3326
3327void free_extent_buffer_stale(struct extent_buffer *eb)
3328{
3329 if (!eb)
3330 return;
3331
3332 spin_lock(&eb->refs_lock);
3333 set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
3334
3335 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
3336 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3337 atomic_dec(&eb->refs);
3338 release_extent_buffer(eb);
3339}
3340
3341static void btree_clear_folio_dirty(struct folio *folio)
3342{
3343 ASSERT(folio_test_dirty(folio));
3344 ASSERT(folio_test_locked(folio));
3345 folio_clear_dirty_for_io(folio);
3346 xa_lock_irq(&folio->mapping->i_pages);
3347 if (!folio_test_dirty(folio))
3348 __xa_clear_mark(&folio->mapping->i_pages,
3349 folio_index(folio), PAGECACHE_TAG_DIRTY);
3350 xa_unlock_irq(&folio->mapping->i_pages);
3351}
3352
3353static void clear_subpage_extent_buffer_dirty(const struct extent_buffer *eb)
3354{
3355 struct btrfs_fs_info *fs_info = eb->fs_info;
3356 struct folio *folio = eb->folios[0];
3357 bool last;
3358
3359 /* btree_clear_folio_dirty() needs page locked. */
3360 folio_lock(folio);
3361 last = btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start, eb->len);
3362 if (last)
3363 btree_clear_folio_dirty(folio);
3364 folio_unlock(folio);
3365 WARN_ON(atomic_read(&eb->refs) == 0);
3366}
3367
3368void btrfs_clear_buffer_dirty(struct btrfs_trans_handle *trans,
3369 struct extent_buffer *eb)
3370{
3371 struct btrfs_fs_info *fs_info = eb->fs_info;
3372 int num_folios;
3373
3374 btrfs_assert_tree_write_locked(eb);
3375
3376 if (trans && btrfs_header_generation(eb) != trans->transid)
3377 return;
3378
3379 /*
3380 * Instead of clearing the dirty flag off of the buffer, mark it as
3381 * EXTENT_BUFFER_ZONED_ZEROOUT. This allows us to preserve
3382 * write-ordering in zoned mode, without the need to later re-dirty
3383 * the extent_buffer.
3384 *
3385 * The actual zeroout of the buffer will happen later in
3386 * btree_csum_one_bio.
3387 */
3388 if (btrfs_is_zoned(fs_info) && test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3389 set_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags);
3390 return;
3391 }
3392
3393 if (!test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags))
3394 return;
3395
3396 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes, -eb->len,
3397 fs_info->dirty_metadata_batch);
3398
3399 if (eb->fs_info->nodesize < PAGE_SIZE)
3400 return clear_subpage_extent_buffer_dirty(eb);
3401
3402 num_folios = num_extent_folios(eb);
3403 for (int i = 0; i < num_folios; i++) {
3404 struct folio *folio = eb->folios[i];
3405
3406 if (!folio_test_dirty(folio))
3407 continue;
3408 folio_lock(folio);
3409 btree_clear_folio_dirty(folio);
3410 folio_unlock(folio);
3411 }
3412 WARN_ON(atomic_read(&eb->refs) == 0);
3413}
3414
3415void set_extent_buffer_dirty(struct extent_buffer *eb)
3416{
3417 int num_folios;
3418 bool was_dirty;
3419
3420 check_buffer_tree_ref(eb);
3421
3422 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3423
3424 num_folios = num_extent_folios(eb);
3425 WARN_ON(atomic_read(&eb->refs) == 0);
3426 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
3427 WARN_ON(test_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags));
3428
3429 if (!was_dirty) {
3430 bool subpage = eb->fs_info->nodesize < PAGE_SIZE;
3431
3432 /*
3433 * For subpage case, we can have other extent buffers in the
3434 * same page, and in clear_subpage_extent_buffer_dirty() we
3435 * have to clear page dirty without subpage lock held.
3436 * This can cause race where our page gets dirty cleared after
3437 * we just set it.
3438 *
3439 * Thankfully, clear_subpage_extent_buffer_dirty() has locked
3440 * its page for other reasons, we can use page lock to prevent
3441 * the above race.
3442 */
3443 if (subpage)
3444 lock_page(folio_page(eb->folios[0], 0));
3445 for (int i = 0; i < num_folios; i++)
3446 btrfs_folio_set_dirty(eb->fs_info, eb->folios[i],
3447 eb->start, eb->len);
3448 if (subpage)
3449 unlock_page(folio_page(eb->folios[0], 0));
3450 percpu_counter_add_batch(&eb->fs_info->dirty_metadata_bytes,
3451 eb->len,
3452 eb->fs_info->dirty_metadata_batch);
3453 }
3454#ifdef CONFIG_BTRFS_DEBUG
3455 for (int i = 0; i < num_folios; i++)
3456 ASSERT(folio_test_dirty(eb->folios[i]));
3457#endif
3458}
3459
3460void clear_extent_buffer_uptodate(struct extent_buffer *eb)
3461{
3462 struct btrfs_fs_info *fs_info = eb->fs_info;
3463 int num_folios = num_extent_folios(eb);
3464
3465 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3466 for (int i = 0; i < num_folios; i++) {
3467 struct folio *folio = eb->folios[i];
3468
3469 if (!folio)
3470 continue;
3471
3472 /*
3473 * This is special handling for metadata subpage, as regular
3474 * btrfs_is_subpage() can not handle cloned/dummy metadata.
3475 */
3476 if (fs_info->nodesize >= PAGE_SIZE)
3477 folio_clear_uptodate(folio);
3478 else
3479 btrfs_subpage_clear_uptodate(fs_info, folio,
3480 eb->start, eb->len);
3481 }
3482}
3483
3484void set_extent_buffer_uptodate(struct extent_buffer *eb)
3485{
3486 struct btrfs_fs_info *fs_info = eb->fs_info;
3487 int num_folios = num_extent_folios(eb);
3488
3489 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3490 for (int i = 0; i < num_folios; i++) {
3491 struct folio *folio = eb->folios[i];
3492
3493 /*
3494 * This is special handling for metadata subpage, as regular
3495 * btrfs_is_subpage() can not handle cloned/dummy metadata.
3496 */
3497 if (fs_info->nodesize >= PAGE_SIZE)
3498 folio_mark_uptodate(folio);
3499 else
3500 btrfs_subpage_set_uptodate(fs_info, folio,
3501 eb->start, eb->len);
3502 }
3503}
3504
3505static void clear_extent_buffer_reading(struct extent_buffer *eb)
3506{
3507 clear_bit(EXTENT_BUFFER_READING, &eb->bflags);
3508 smp_mb__after_atomic();
3509 wake_up_bit(&eb->bflags, EXTENT_BUFFER_READING);
3510}
3511
3512static void end_bbio_meta_read(struct btrfs_bio *bbio)
3513{
3514 struct extent_buffer *eb = bbio->private;
3515 struct btrfs_fs_info *fs_info = eb->fs_info;
3516 bool uptodate = !bbio->bio.bi_status;
3517 struct folio_iter fi;
3518 u32 bio_offset = 0;
3519
3520 /*
3521 * If the extent buffer is marked UPTODATE before the read operation
3522 * completes, other calls to read_extent_buffer_pages() will return
3523 * early without waiting for the read to finish, causing data races.
3524 */
3525 WARN_ON(test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags));
3526
3527 eb->read_mirror = bbio->mirror_num;
3528
3529 if (uptodate &&
3530 btrfs_validate_extent_buffer(eb, &bbio->parent_check) < 0)
3531 uptodate = false;
3532
3533 if (uptodate) {
3534 set_extent_buffer_uptodate(eb);
3535 } else {
3536 clear_extent_buffer_uptodate(eb);
3537 set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
3538 }
3539
3540 bio_for_each_folio_all(fi, &bbio->bio) {
3541 struct folio *folio = fi.folio;
3542 u64 start = eb->start + bio_offset;
3543 u32 len = fi.length;
3544
3545 if (uptodate)
3546 btrfs_folio_set_uptodate(fs_info, folio, start, len);
3547 else
3548 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
3549
3550 bio_offset += len;
3551 }
3552
3553 clear_extent_buffer_reading(eb);
3554 free_extent_buffer(eb);
3555
3556 bio_put(&bbio->bio);
3557}
3558
3559int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num,
3560 const struct btrfs_tree_parent_check *check)
3561{
3562 struct btrfs_bio *bbio;
3563 bool ret;
3564
3565 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3566 return 0;
3567
3568 /*
3569 * We could have had EXTENT_BUFFER_UPTODATE cleared by the write
3570 * operation, which could potentially still be in flight. In this case
3571 * we simply want to return an error.
3572 */
3573 if (unlikely(test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)))
3574 return -EIO;
3575
3576 /* Someone else is already reading the buffer, just wait for it. */
3577 if (test_and_set_bit(EXTENT_BUFFER_READING, &eb->bflags))
3578 goto done;
3579
3580 /*
3581 * Between the initial test_bit(EXTENT_BUFFER_UPTODATE) and the above
3582 * test_and_set_bit(EXTENT_BUFFER_READING), someone else could have
3583 * started and finished reading the same eb. In this case, UPTODATE
3584 * will now be set, and we shouldn't read it in again.
3585 */
3586 if (unlikely(test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))) {
3587 clear_extent_buffer_reading(eb);
3588 return 0;
3589 }
3590
3591 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
3592 eb->read_mirror = 0;
3593 check_buffer_tree_ref(eb);
3594 atomic_inc(&eb->refs);
3595
3596 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
3597 REQ_OP_READ | REQ_META, eb->fs_info,
3598 end_bbio_meta_read, eb);
3599 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
3600 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
3601 bbio->file_offset = eb->start;
3602 memcpy(&bbio->parent_check, check, sizeof(*check));
3603 if (eb->fs_info->nodesize < PAGE_SIZE) {
3604 ret = bio_add_folio(&bbio->bio, eb->folios[0], eb->len,
3605 eb->start - folio_pos(eb->folios[0]));
3606 ASSERT(ret);
3607 } else {
3608 int num_folios = num_extent_folios(eb);
3609
3610 for (int i = 0; i < num_folios; i++) {
3611 struct folio *folio = eb->folios[i];
3612
3613 ret = bio_add_folio(&bbio->bio, folio, eb->folio_size, 0);
3614 ASSERT(ret);
3615 }
3616 }
3617 btrfs_submit_bbio(bbio, mirror_num);
3618
3619done:
3620 if (wait == WAIT_COMPLETE) {
3621 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_READING, TASK_UNINTERRUPTIBLE);
3622 if (!test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3623 return -EIO;
3624 }
3625
3626 return 0;
3627}
3628
3629static bool report_eb_range(const struct extent_buffer *eb, unsigned long start,
3630 unsigned long len)
3631{
3632 btrfs_warn(eb->fs_info,
3633 "access to eb bytenr %llu len %u out of range start %lu len %lu",
3634 eb->start, eb->len, start, len);
3635 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
3636
3637 return true;
3638}
3639
3640/*
3641 * Check if the [start, start + len) range is valid before reading/writing
3642 * the eb.
3643 * NOTE: @start and @len are offset inside the eb, not logical address.
3644 *
3645 * Caller should not touch the dst/src memory if this function returns error.
3646 */
3647static inline int check_eb_range(const struct extent_buffer *eb,
3648 unsigned long start, unsigned long len)
3649{
3650 unsigned long offset;
3651
3652 /* start, start + len should not go beyond eb->len nor overflow */
3653 if (unlikely(check_add_overflow(start, len, &offset) || offset > eb->len))
3654 return report_eb_range(eb, start, len);
3655
3656 return false;
3657}
3658
3659void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
3660 unsigned long start, unsigned long len)
3661{
3662 const int unit_size = eb->folio_size;
3663 size_t cur;
3664 size_t offset;
3665 char *dst = (char *)dstv;
3666 unsigned long i = get_eb_folio_index(eb, start);
3667
3668 if (check_eb_range(eb, start, len)) {
3669 /*
3670 * Invalid range hit, reset the memory, so callers won't get
3671 * some random garbage for their uninitialized memory.
3672 */
3673 memset(dstv, 0, len);
3674 return;
3675 }
3676
3677 if (eb->addr) {
3678 memcpy(dstv, eb->addr + start, len);
3679 return;
3680 }
3681
3682 offset = get_eb_offset_in_folio(eb, start);
3683
3684 while (len > 0) {
3685 char *kaddr;
3686
3687 cur = min(len, unit_size - offset);
3688 kaddr = folio_address(eb->folios[i]);
3689 memcpy(dst, kaddr + offset, cur);
3690
3691 dst += cur;
3692 len -= cur;
3693 offset = 0;
3694 i++;
3695 }
3696}
3697
3698int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
3699 void __user *dstv,
3700 unsigned long start, unsigned long len)
3701{
3702 const int unit_size = eb->folio_size;
3703 size_t cur;
3704 size_t offset;
3705 char __user *dst = (char __user *)dstv;
3706 unsigned long i = get_eb_folio_index(eb, start);
3707 int ret = 0;
3708
3709 WARN_ON(start > eb->len);
3710 WARN_ON(start + len > eb->start + eb->len);
3711
3712 if (eb->addr) {
3713 if (copy_to_user_nofault(dstv, eb->addr + start, len))
3714 ret = -EFAULT;
3715 return ret;
3716 }
3717
3718 offset = get_eb_offset_in_folio(eb, start);
3719
3720 while (len > 0) {
3721 char *kaddr;
3722
3723 cur = min(len, unit_size - offset);
3724 kaddr = folio_address(eb->folios[i]);
3725 if (copy_to_user_nofault(dst, kaddr + offset, cur)) {
3726 ret = -EFAULT;
3727 break;
3728 }
3729
3730 dst += cur;
3731 len -= cur;
3732 offset = 0;
3733 i++;
3734 }
3735
3736 return ret;
3737}
3738
3739int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
3740 unsigned long start, unsigned long len)
3741{
3742 const int unit_size = eb->folio_size;
3743 size_t cur;
3744 size_t offset;
3745 char *kaddr;
3746 char *ptr = (char *)ptrv;
3747 unsigned long i = get_eb_folio_index(eb, start);
3748 int ret = 0;
3749
3750 if (check_eb_range(eb, start, len))
3751 return -EINVAL;
3752
3753 if (eb->addr)
3754 return memcmp(ptrv, eb->addr + start, len);
3755
3756 offset = get_eb_offset_in_folio(eb, start);
3757
3758 while (len > 0) {
3759 cur = min(len, unit_size - offset);
3760 kaddr = folio_address(eb->folios[i]);
3761 ret = memcmp(ptr, kaddr + offset, cur);
3762 if (ret)
3763 break;
3764
3765 ptr += cur;
3766 len -= cur;
3767 offset = 0;
3768 i++;
3769 }
3770 return ret;
3771}
3772
3773/*
3774 * Check that the extent buffer is uptodate.
3775 *
3776 * For regular sector size == PAGE_SIZE case, check if @page is uptodate.
3777 * For subpage case, check if the range covered by the eb has EXTENT_UPTODATE.
3778 */
3779static void assert_eb_folio_uptodate(const struct extent_buffer *eb, int i)
3780{
3781 struct btrfs_fs_info *fs_info = eb->fs_info;
3782 struct folio *folio = eb->folios[i];
3783
3784 ASSERT(folio);
3785
3786 /*
3787 * If we are using the commit root we could potentially clear a page
3788 * Uptodate while we're using the extent buffer that we've previously
3789 * looked up. We don't want to complain in this case, as the page was
3790 * valid before, we just didn't write it out. Instead we want to catch
3791 * the case where we didn't actually read the block properly, which
3792 * would have !PageUptodate and !EXTENT_BUFFER_WRITE_ERR.
3793 */
3794 if (test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3795 return;
3796
3797 if (fs_info->nodesize < PAGE_SIZE) {
3798 folio = eb->folios[0];
3799 ASSERT(i == 0);
3800 if (WARN_ON(!btrfs_subpage_test_uptodate(fs_info, folio,
3801 eb->start, eb->len)))
3802 btrfs_subpage_dump_bitmap(fs_info, folio, eb->start, eb->len);
3803 } else {
3804 WARN_ON(!folio_test_uptodate(folio));
3805 }
3806}
3807
3808static void __write_extent_buffer(const struct extent_buffer *eb,
3809 const void *srcv, unsigned long start,
3810 unsigned long len, bool use_memmove)
3811{
3812 const int unit_size = eb->folio_size;
3813 size_t cur;
3814 size_t offset;
3815 char *kaddr;
3816 const char *src = (const char *)srcv;
3817 unsigned long i = get_eb_folio_index(eb, start);
3818 /* For unmapped (dummy) ebs, no need to check their uptodate status. */
3819 const bool check_uptodate = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3820
3821 if (check_eb_range(eb, start, len))
3822 return;
3823
3824 if (eb->addr) {
3825 if (use_memmove)
3826 memmove(eb->addr + start, srcv, len);
3827 else
3828 memcpy(eb->addr + start, srcv, len);
3829 return;
3830 }
3831
3832 offset = get_eb_offset_in_folio(eb, start);
3833
3834 while (len > 0) {
3835 if (check_uptodate)
3836 assert_eb_folio_uptodate(eb, i);
3837
3838 cur = min(len, unit_size - offset);
3839 kaddr = folio_address(eb->folios[i]);
3840 if (use_memmove)
3841 memmove(kaddr + offset, src, cur);
3842 else
3843 memcpy(kaddr + offset, src, cur);
3844
3845 src += cur;
3846 len -= cur;
3847 offset = 0;
3848 i++;
3849 }
3850}
3851
3852void write_extent_buffer(const struct extent_buffer *eb, const void *srcv,
3853 unsigned long start, unsigned long len)
3854{
3855 return __write_extent_buffer(eb, srcv, start, len, false);
3856}
3857
3858static void memset_extent_buffer(const struct extent_buffer *eb, int c,
3859 unsigned long start, unsigned long len)
3860{
3861 const int unit_size = eb->folio_size;
3862 unsigned long cur = start;
3863
3864 if (eb->addr) {
3865 memset(eb->addr + start, c, len);
3866 return;
3867 }
3868
3869 while (cur < start + len) {
3870 unsigned long index = get_eb_folio_index(eb, cur);
3871 unsigned int offset = get_eb_offset_in_folio(eb, cur);
3872 unsigned int cur_len = min(start + len - cur, unit_size - offset);
3873
3874 assert_eb_folio_uptodate(eb, index);
3875 memset(folio_address(eb->folios[index]) + offset, c, cur_len);
3876
3877 cur += cur_len;
3878 }
3879}
3880
3881void memzero_extent_buffer(const struct extent_buffer *eb, unsigned long start,
3882 unsigned long len)
3883{
3884 if (check_eb_range(eb, start, len))
3885 return;
3886 return memset_extent_buffer(eb, 0, start, len);
3887}
3888
3889void copy_extent_buffer_full(const struct extent_buffer *dst,
3890 const struct extent_buffer *src)
3891{
3892 const int unit_size = src->folio_size;
3893 unsigned long cur = 0;
3894
3895 ASSERT(dst->len == src->len);
3896
3897 while (cur < src->len) {
3898 unsigned long index = get_eb_folio_index(src, cur);
3899 unsigned long offset = get_eb_offset_in_folio(src, cur);
3900 unsigned long cur_len = min(src->len, unit_size - offset);
3901 void *addr = folio_address(src->folios[index]) + offset;
3902
3903 write_extent_buffer(dst, addr, cur, cur_len);
3904
3905 cur += cur_len;
3906 }
3907}
3908
3909void copy_extent_buffer(const struct extent_buffer *dst,
3910 const struct extent_buffer *src,
3911 unsigned long dst_offset, unsigned long src_offset,
3912 unsigned long len)
3913{
3914 const int unit_size = dst->folio_size;
3915 u64 dst_len = dst->len;
3916 size_t cur;
3917 size_t offset;
3918 char *kaddr;
3919 unsigned long i = get_eb_folio_index(dst, dst_offset);
3920
3921 if (check_eb_range(dst, dst_offset, len) ||
3922 check_eb_range(src, src_offset, len))
3923 return;
3924
3925 WARN_ON(src->len != dst_len);
3926
3927 offset = get_eb_offset_in_folio(dst, dst_offset);
3928
3929 while (len > 0) {
3930 assert_eb_folio_uptodate(dst, i);
3931
3932 cur = min(len, (unsigned long)(unit_size - offset));
3933
3934 kaddr = folio_address(dst->folios[i]);
3935 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3936
3937 src_offset += cur;
3938 len -= cur;
3939 offset = 0;
3940 i++;
3941 }
3942}
3943
3944/*
3945 * Calculate the folio and offset of the byte containing the given bit number.
3946 *
3947 * @eb: the extent buffer
3948 * @start: offset of the bitmap item in the extent buffer
3949 * @nr: bit number
3950 * @folio_index: return index of the folio in the extent buffer that contains
3951 * the given bit number
3952 * @folio_offset: return offset into the folio given by folio_index
3953 *
3954 * This helper hides the ugliness of finding the byte in an extent buffer which
3955 * contains a given bit.
3956 */
3957static inline void eb_bitmap_offset(const struct extent_buffer *eb,
3958 unsigned long start, unsigned long nr,
3959 unsigned long *folio_index,
3960 size_t *folio_offset)
3961{
3962 size_t byte_offset = BIT_BYTE(nr);
3963 size_t offset;
3964
3965 /*
3966 * The byte we want is the offset of the extent buffer + the offset of
3967 * the bitmap item in the extent buffer + the offset of the byte in the
3968 * bitmap item.
3969 */
3970 offset = start + offset_in_eb_folio(eb, eb->start) + byte_offset;
3971
3972 *folio_index = offset >> eb->folio_shift;
3973 *folio_offset = offset_in_eb_folio(eb, offset);
3974}
3975
3976/*
3977 * Determine whether a bit in a bitmap item is set.
3978 *
3979 * @eb: the extent buffer
3980 * @start: offset of the bitmap item in the extent buffer
3981 * @nr: bit number to test
3982 */
3983int extent_buffer_test_bit(const struct extent_buffer *eb, unsigned long start,
3984 unsigned long nr)
3985{
3986 unsigned long i;
3987 size_t offset;
3988 u8 *kaddr;
3989
3990 eb_bitmap_offset(eb, start, nr, &i, &offset);
3991 assert_eb_folio_uptodate(eb, i);
3992 kaddr = folio_address(eb->folios[i]);
3993 return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
3994}
3995
3996static u8 *extent_buffer_get_byte(const struct extent_buffer *eb, unsigned long bytenr)
3997{
3998 unsigned long index = get_eb_folio_index(eb, bytenr);
3999
4000 if (check_eb_range(eb, bytenr, 1))
4001 return NULL;
4002 return folio_address(eb->folios[index]) + get_eb_offset_in_folio(eb, bytenr);
4003}
4004
4005/*
4006 * Set an area of a bitmap to 1.
4007 *
4008 * @eb: the extent buffer
4009 * @start: offset of the bitmap item in the extent buffer
4010 * @pos: bit number of the first bit
4011 * @len: number of bits to set
4012 */
4013void extent_buffer_bitmap_set(const struct extent_buffer *eb, unsigned long start,
4014 unsigned long pos, unsigned long len)
4015{
4016 unsigned int first_byte = start + BIT_BYTE(pos);
4017 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4018 const bool same_byte = (first_byte == last_byte);
4019 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4020 u8 *kaddr;
4021
4022 if (same_byte)
4023 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4024
4025 /* Handle the first byte. */
4026 kaddr = extent_buffer_get_byte(eb, first_byte);
4027 *kaddr |= mask;
4028 if (same_byte)
4029 return;
4030
4031 /* Handle the byte aligned part. */
4032 ASSERT(first_byte + 1 <= last_byte);
4033 memset_extent_buffer(eb, 0xff, first_byte + 1, last_byte - first_byte - 1);
4034
4035 /* Handle the last byte. */
4036 kaddr = extent_buffer_get_byte(eb, last_byte);
4037 *kaddr |= BITMAP_LAST_BYTE_MASK(pos + len);
4038}
4039
4040
4041/*
4042 * Clear an area of a bitmap.
4043 *
4044 * @eb: the extent buffer
4045 * @start: offset of the bitmap item in the extent buffer
4046 * @pos: bit number of the first bit
4047 * @len: number of bits to clear
4048 */
4049void extent_buffer_bitmap_clear(const struct extent_buffer *eb,
4050 unsigned long start, unsigned long pos,
4051 unsigned long len)
4052{
4053 unsigned int first_byte = start + BIT_BYTE(pos);
4054 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4055 const bool same_byte = (first_byte == last_byte);
4056 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4057 u8 *kaddr;
4058
4059 if (same_byte)
4060 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4061
4062 /* Handle the first byte. */
4063 kaddr = extent_buffer_get_byte(eb, first_byte);
4064 *kaddr &= ~mask;
4065 if (same_byte)
4066 return;
4067
4068 /* Handle the byte aligned part. */
4069 ASSERT(first_byte + 1 <= last_byte);
4070 memset_extent_buffer(eb, 0, first_byte + 1, last_byte - first_byte - 1);
4071
4072 /* Handle the last byte. */
4073 kaddr = extent_buffer_get_byte(eb, last_byte);
4074 *kaddr &= ~BITMAP_LAST_BYTE_MASK(pos + len);
4075}
4076
4077static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4078{
4079 unsigned long distance = (src > dst) ? src - dst : dst - src;
4080 return distance < len;
4081}
4082
4083void memcpy_extent_buffer(const struct extent_buffer *dst,
4084 unsigned long dst_offset, unsigned long src_offset,
4085 unsigned long len)
4086{
4087 const int unit_size = dst->folio_size;
4088 unsigned long cur_off = 0;
4089
4090 if (check_eb_range(dst, dst_offset, len) ||
4091 check_eb_range(dst, src_offset, len))
4092 return;
4093
4094 if (dst->addr) {
4095 const bool use_memmove = areas_overlap(src_offset, dst_offset, len);
4096
4097 if (use_memmove)
4098 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4099 else
4100 memcpy(dst->addr + dst_offset, dst->addr + src_offset, len);
4101 return;
4102 }
4103
4104 while (cur_off < len) {
4105 unsigned long cur_src = cur_off + src_offset;
4106 unsigned long folio_index = get_eb_folio_index(dst, cur_src);
4107 unsigned long folio_off = get_eb_offset_in_folio(dst, cur_src);
4108 unsigned long cur_len = min(src_offset + len - cur_src,
4109 unit_size - folio_off);
4110 void *src_addr = folio_address(dst->folios[folio_index]) + folio_off;
4111 const bool use_memmove = areas_overlap(src_offset + cur_off,
4112 dst_offset + cur_off, cur_len);
4113
4114 __write_extent_buffer(dst, src_addr, dst_offset + cur_off, cur_len,
4115 use_memmove);
4116 cur_off += cur_len;
4117 }
4118}
4119
4120void memmove_extent_buffer(const struct extent_buffer *dst,
4121 unsigned long dst_offset, unsigned long src_offset,
4122 unsigned long len)
4123{
4124 unsigned long dst_end = dst_offset + len - 1;
4125 unsigned long src_end = src_offset + len - 1;
4126
4127 if (check_eb_range(dst, dst_offset, len) ||
4128 check_eb_range(dst, src_offset, len))
4129 return;
4130
4131 if (dst_offset < src_offset) {
4132 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4133 return;
4134 }
4135
4136 if (dst->addr) {
4137 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4138 return;
4139 }
4140
4141 while (len > 0) {
4142 unsigned long src_i;
4143 size_t cur;
4144 size_t dst_off_in_folio;
4145 size_t src_off_in_folio;
4146 void *src_addr;
4147 bool use_memmove;
4148
4149 src_i = get_eb_folio_index(dst, src_end);
4150
4151 dst_off_in_folio = get_eb_offset_in_folio(dst, dst_end);
4152 src_off_in_folio = get_eb_offset_in_folio(dst, src_end);
4153
4154 cur = min_t(unsigned long, len, src_off_in_folio + 1);
4155 cur = min(cur, dst_off_in_folio + 1);
4156
4157 src_addr = folio_address(dst->folios[src_i]) + src_off_in_folio -
4158 cur + 1;
4159 use_memmove = areas_overlap(src_end - cur + 1, dst_end - cur + 1,
4160 cur);
4161
4162 __write_extent_buffer(dst, src_addr, dst_end - cur + 1, cur,
4163 use_memmove);
4164
4165 dst_end -= cur;
4166 src_end -= cur;
4167 len -= cur;
4168 }
4169}
4170
4171#define GANG_LOOKUP_SIZE 16
4172static struct extent_buffer *get_next_extent_buffer(
4173 const struct btrfs_fs_info *fs_info, struct folio *folio, u64 bytenr)
4174{
4175 struct extent_buffer *gang[GANG_LOOKUP_SIZE];
4176 struct extent_buffer *found = NULL;
4177 u64 folio_start = folio_pos(folio);
4178 u64 cur = folio_start;
4179
4180 ASSERT(in_range(bytenr, folio_start, PAGE_SIZE));
4181 lockdep_assert_held(&fs_info->buffer_lock);
4182
4183 while (cur < folio_start + PAGE_SIZE) {
4184 int ret;
4185 int i;
4186
4187 ret = radix_tree_gang_lookup(&fs_info->buffer_radix,
4188 (void **)gang, cur >> fs_info->sectorsize_bits,
4189 min_t(unsigned int, GANG_LOOKUP_SIZE,
4190 PAGE_SIZE / fs_info->nodesize));
4191 if (ret == 0)
4192 goto out;
4193 for (i = 0; i < ret; i++) {
4194 /* Already beyond page end */
4195 if (gang[i]->start >= folio_start + PAGE_SIZE)
4196 goto out;
4197 /* Found one */
4198 if (gang[i]->start >= bytenr) {
4199 found = gang[i];
4200 goto out;
4201 }
4202 }
4203 cur = gang[ret - 1]->start + gang[ret - 1]->len;
4204 }
4205out:
4206 return found;
4207}
4208
4209static int try_release_subpage_extent_buffer(struct folio *folio)
4210{
4211 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
4212 u64 cur = folio_pos(folio);
4213 const u64 end = cur + PAGE_SIZE;
4214 int ret;
4215
4216 while (cur < end) {
4217 struct extent_buffer *eb = NULL;
4218
4219 /*
4220 * Unlike try_release_extent_buffer() which uses folio private
4221 * to grab buffer, for subpage case we rely on radix tree, thus
4222 * we need to ensure radix tree consistency.
4223 *
4224 * We also want an atomic snapshot of the radix tree, thus go
4225 * with spinlock rather than RCU.
4226 */
4227 spin_lock(&fs_info->buffer_lock);
4228 eb = get_next_extent_buffer(fs_info, folio, cur);
4229 if (!eb) {
4230 /* No more eb in the page range after or at cur */
4231 spin_unlock(&fs_info->buffer_lock);
4232 break;
4233 }
4234 cur = eb->start + eb->len;
4235
4236 /*
4237 * The same as try_release_extent_buffer(), to ensure the eb
4238 * won't disappear out from under us.
4239 */
4240 spin_lock(&eb->refs_lock);
4241 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4242 spin_unlock(&eb->refs_lock);
4243 spin_unlock(&fs_info->buffer_lock);
4244 break;
4245 }
4246 spin_unlock(&fs_info->buffer_lock);
4247
4248 /*
4249 * If tree ref isn't set then we know the ref on this eb is a
4250 * real ref, so just return, this eb will likely be freed soon
4251 * anyway.
4252 */
4253 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4254 spin_unlock(&eb->refs_lock);
4255 break;
4256 }
4257
4258 /*
4259 * Here we don't care about the return value, we will always
4260 * check the folio private at the end. And
4261 * release_extent_buffer() will release the refs_lock.
4262 */
4263 release_extent_buffer(eb);
4264 }
4265 /*
4266 * Finally to check if we have cleared folio private, as if we have
4267 * released all ebs in the page, the folio private should be cleared now.
4268 */
4269 spin_lock(&folio->mapping->i_private_lock);
4270 if (!folio_test_private(folio))
4271 ret = 1;
4272 else
4273 ret = 0;
4274 spin_unlock(&folio->mapping->i_private_lock);
4275 return ret;
4276
4277}
4278
4279int try_release_extent_buffer(struct folio *folio)
4280{
4281 struct extent_buffer *eb;
4282
4283 if (folio_to_fs_info(folio)->nodesize < PAGE_SIZE)
4284 return try_release_subpage_extent_buffer(folio);
4285
4286 /*
4287 * We need to make sure nobody is changing folio private, as we rely on
4288 * folio private as the pointer to extent buffer.
4289 */
4290 spin_lock(&folio->mapping->i_private_lock);
4291 if (!folio_test_private(folio)) {
4292 spin_unlock(&folio->mapping->i_private_lock);
4293 return 1;
4294 }
4295
4296 eb = folio_get_private(folio);
4297 BUG_ON(!eb);
4298
4299 /*
4300 * This is a little awful but should be ok, we need to make sure that
4301 * the eb doesn't disappear out from under us while we're looking at
4302 * this page.
4303 */
4304 spin_lock(&eb->refs_lock);
4305 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4306 spin_unlock(&eb->refs_lock);
4307 spin_unlock(&folio->mapping->i_private_lock);
4308 return 0;
4309 }
4310 spin_unlock(&folio->mapping->i_private_lock);
4311
4312 /*
4313 * If tree ref isn't set then we know the ref on this eb is a real ref,
4314 * so just return, this page will likely be freed soon anyway.
4315 */
4316 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4317 spin_unlock(&eb->refs_lock);
4318 return 0;
4319 }
4320
4321 return release_extent_buffer(eb);
4322}
4323
4324/*
4325 * Attempt to readahead a child block.
4326 *
4327 * @fs_info: the fs_info
4328 * @bytenr: bytenr to read
4329 * @owner_root: objectid of the root that owns this eb
4330 * @gen: generation for the uptodate check, can be 0
4331 * @level: level for the eb
4332 *
4333 * Attempt to readahead a tree block at @bytenr. If @gen is 0 then we do a
4334 * normal uptodate check of the eb, without checking the generation. If we have
4335 * to read the block we will not block on anything.
4336 */
4337void btrfs_readahead_tree_block(struct btrfs_fs_info *fs_info,
4338 u64 bytenr, u64 owner_root, u64 gen, int level)
4339{
4340 struct btrfs_tree_parent_check check = {
4341 .level = level,
4342 .transid = gen
4343 };
4344 struct extent_buffer *eb;
4345 int ret;
4346
4347 eb = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
4348 if (IS_ERR(eb))
4349 return;
4350
4351 if (btrfs_buffer_uptodate(eb, gen, 1)) {
4352 free_extent_buffer(eb);
4353 return;
4354 }
4355
4356 ret = read_extent_buffer_pages(eb, WAIT_NONE, 0, &check);
4357 if (ret < 0)
4358 free_extent_buffer_stale(eb);
4359 else
4360 free_extent_buffer(eb);
4361}
4362
4363/*
4364 * Readahead a node's child block.
4365 *
4366 * @node: parent node we're reading from
4367 * @slot: slot in the parent node for the child we want to read
4368 *
4369 * A helper for btrfs_readahead_tree_block, we simply read the bytenr pointed at
4370 * the slot in the node provided.
4371 */
4372void btrfs_readahead_node_child(struct extent_buffer *node, int slot)
4373{
4374 btrfs_readahead_tree_block(node->fs_info,
4375 btrfs_node_blockptr(node, slot),
4376 btrfs_header_owner(node),
4377 btrfs_node_ptr_generation(node, slot),
4378 btrfs_header_level(node) - 1);
4379}