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
106static void submit_one_bio(struct btrfs_bio_ctrl *bio_ctrl)
107{
108 struct btrfs_bio *bbio = bio_ctrl->bbio;
109
110 if (!bbio)
111 return;
112
113 /* Caller should ensure the bio has at least some range added */
114 ASSERT(bbio->bio.bi_iter.bi_size);
115
116 if (btrfs_op(&bbio->bio) == BTRFS_MAP_READ &&
117 bio_ctrl->compress_type != BTRFS_COMPRESS_NONE)
118 btrfs_submit_compressed_read(bbio);
119 else
120 btrfs_submit_bio(bbio, 0);
121
122 /* The bbio is owned by the end_io handler now */
123 bio_ctrl->bbio = NULL;
124}
125
126/*
127 * Submit or fail the current bio in the bio_ctrl structure.
128 */
129static void submit_write_bio(struct btrfs_bio_ctrl *bio_ctrl, int ret)
130{
131 struct btrfs_bio *bbio = bio_ctrl->bbio;
132
133 if (!bbio)
134 return;
135
136 if (ret) {
137 ASSERT(ret < 0);
138 btrfs_bio_end_io(bbio, errno_to_blk_status(ret));
139 /* The bio is owned by the end_io handler now */
140 bio_ctrl->bbio = NULL;
141 } else {
142 submit_one_bio(bio_ctrl);
143 }
144}
145
146int __init extent_buffer_init_cachep(void)
147{
148 extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
149 sizeof(struct extent_buffer), 0, 0,
150 NULL);
151 if (!extent_buffer_cache)
152 return -ENOMEM;
153
154 return 0;
155}
156
157void __cold extent_buffer_free_cachep(void)
158{
159 /*
160 * Make sure all delayed rcu free are flushed before we
161 * destroy caches.
162 */
163 rcu_barrier();
164 kmem_cache_destroy(extent_buffer_cache);
165}
166
167void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
168{
169 unsigned long index = start >> PAGE_SHIFT;
170 unsigned long end_index = end >> PAGE_SHIFT;
171 struct page *page;
172
173 while (index <= end_index) {
174 page = find_get_page(inode->i_mapping, index);
175 BUG_ON(!page); /* Pages should be in the extent_io_tree */
176 clear_page_dirty_for_io(page);
177 put_page(page);
178 index++;
179 }
180}
181
182static void process_one_page(struct btrfs_fs_info *fs_info,
183 struct page *page, struct page *locked_page,
184 unsigned long page_ops, u64 start, u64 end)
185{
186 struct folio *folio = page_folio(page);
187 u32 len;
188
189 ASSERT(end + 1 - start != 0 && end + 1 - start < U32_MAX);
190 len = end + 1 - start;
191
192 if (page_ops & PAGE_SET_ORDERED)
193 btrfs_folio_clamp_set_ordered(fs_info, folio, start, len);
194 if (page_ops & PAGE_START_WRITEBACK) {
195 btrfs_folio_clamp_clear_dirty(fs_info, folio, start, len);
196 btrfs_folio_clamp_set_writeback(fs_info, folio, start, len);
197 }
198 if (page_ops & PAGE_END_WRITEBACK)
199 btrfs_folio_clamp_clear_writeback(fs_info, folio, start, len);
200
201 if (page != locked_page && (page_ops & PAGE_UNLOCK))
202 btrfs_folio_end_writer_lock(fs_info, folio, start, len);
203}
204
205static void __process_pages_contig(struct address_space *mapping,
206 struct page *locked_page, u64 start, u64 end,
207 unsigned long page_ops)
208{
209 struct btrfs_fs_info *fs_info = inode_to_fs_info(mapping->host);
210 pgoff_t start_index = start >> PAGE_SHIFT;
211 pgoff_t end_index = end >> PAGE_SHIFT;
212 pgoff_t index = start_index;
213 struct folio_batch fbatch;
214 int i;
215
216 folio_batch_init(&fbatch);
217 while (index <= end_index) {
218 int found_folios;
219
220 found_folios = filemap_get_folios_contig(mapping, &index,
221 end_index, &fbatch);
222 for (i = 0; i < found_folios; i++) {
223 struct folio *folio = fbatch.folios[i];
224
225 process_one_page(fs_info, &folio->page, locked_page,
226 page_ops, start, end);
227 }
228 folio_batch_release(&fbatch);
229 cond_resched();
230 }
231}
232
233static noinline void __unlock_for_delalloc(struct inode *inode,
234 struct page *locked_page,
235 u64 start, u64 end)
236{
237 unsigned long index = start >> PAGE_SHIFT;
238 unsigned long end_index = end >> PAGE_SHIFT;
239
240 ASSERT(locked_page);
241 if (index == locked_page->index && end_index == index)
242 return;
243
244 __process_pages_contig(inode->i_mapping, locked_page, start, end,
245 PAGE_UNLOCK);
246}
247
248static noinline int lock_delalloc_pages(struct inode *inode,
249 struct page *locked_page,
250 u64 start,
251 u64 end)
252{
253 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
254 struct address_space *mapping = inode->i_mapping;
255 pgoff_t start_index = start >> PAGE_SHIFT;
256 pgoff_t end_index = end >> PAGE_SHIFT;
257 pgoff_t index = start_index;
258 u64 processed_end = start;
259 struct folio_batch fbatch;
260
261 if (index == locked_page->index && index == end_index)
262 return 0;
263
264 folio_batch_init(&fbatch);
265 while (index <= end_index) {
266 unsigned int found_folios, i;
267
268 found_folios = filemap_get_folios_contig(mapping, &index,
269 end_index, &fbatch);
270 if (found_folios == 0)
271 goto out;
272
273 for (i = 0; i < found_folios; i++) {
274 struct folio *folio = fbatch.folios[i];
275 struct page *page = folio_page(folio, 0);
276 u32 len = end + 1 - start;
277
278 if (page == locked_page)
279 continue;
280
281 if (btrfs_folio_start_writer_lock(fs_info, folio, start,
282 len))
283 goto out;
284
285 if (!PageDirty(page) || page->mapping != mapping) {
286 btrfs_folio_end_writer_lock(fs_info, folio, start,
287 len);
288 goto out;
289 }
290
291 processed_end = page_offset(page) + PAGE_SIZE - 1;
292 }
293 folio_batch_release(&fbatch);
294 cond_resched();
295 }
296
297 return 0;
298out:
299 folio_batch_release(&fbatch);
300 if (processed_end > start)
301 __unlock_for_delalloc(inode, locked_page, start, processed_end);
302 return -EAGAIN;
303}
304
305/*
306 * Find and lock a contiguous range of bytes in the file marked as delalloc, no
307 * more than @max_bytes.
308 *
309 * @start: The original start bytenr to search.
310 * Will store the extent range start bytenr.
311 * @end: The original end bytenr of the search range
312 * Will store the extent range end bytenr.
313 *
314 * Return true if we find a delalloc range which starts inside the original
315 * range, and @start/@end will store the delalloc range start/end.
316 *
317 * Return false if we can't find any delalloc range which starts inside the
318 * original range, and @start/@end will be the non-delalloc range start/end.
319 */
320EXPORT_FOR_TESTS
321noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
322 struct page *locked_page, u64 *start,
323 u64 *end)
324{
325 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
326 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
327 const u64 orig_start = *start;
328 const u64 orig_end = *end;
329 /* The sanity tests may not set a valid fs_info. */
330 u64 max_bytes = fs_info ? fs_info->max_extent_size : BTRFS_MAX_EXTENT_SIZE;
331 u64 delalloc_start;
332 u64 delalloc_end;
333 bool found;
334 struct extent_state *cached_state = NULL;
335 int ret;
336 int loops = 0;
337
338 /* Caller should pass a valid @end to indicate the search range end */
339 ASSERT(orig_end > orig_start);
340
341 /* The range should at least cover part of the page */
342 ASSERT(!(orig_start >= page_offset(locked_page) + PAGE_SIZE ||
343 orig_end <= page_offset(locked_page)));
344again:
345 /* step one, find a bunch of delalloc bytes starting at start */
346 delalloc_start = *start;
347 delalloc_end = 0;
348 found = btrfs_find_delalloc_range(tree, &delalloc_start, &delalloc_end,
349 max_bytes, &cached_state);
350 if (!found || delalloc_end <= *start || delalloc_start > orig_end) {
351 *start = delalloc_start;
352
353 /* @delalloc_end can be -1, never go beyond @orig_end */
354 *end = min(delalloc_end, orig_end);
355 free_extent_state(cached_state);
356 return false;
357 }
358
359 /*
360 * start comes from the offset of locked_page. We have to lock
361 * pages in order, so we can't process delalloc bytes before
362 * locked_page
363 */
364 if (delalloc_start < *start)
365 delalloc_start = *start;
366
367 /*
368 * make sure to limit the number of pages we try to lock down
369 */
370 if (delalloc_end + 1 - delalloc_start > max_bytes)
371 delalloc_end = delalloc_start + max_bytes - 1;
372
373 /* step two, lock all the pages after the page that has start */
374 ret = lock_delalloc_pages(inode, locked_page,
375 delalloc_start, delalloc_end);
376 ASSERT(!ret || ret == -EAGAIN);
377 if (ret == -EAGAIN) {
378 /* some of the pages are gone, lets avoid looping by
379 * shortening the size of the delalloc range we're searching
380 */
381 free_extent_state(cached_state);
382 cached_state = NULL;
383 if (!loops) {
384 max_bytes = PAGE_SIZE;
385 loops = 1;
386 goto again;
387 } else {
388 found = false;
389 goto out_failed;
390 }
391 }
392
393 /* step three, lock the state bits for the whole range */
394 lock_extent(tree, delalloc_start, delalloc_end, &cached_state);
395
396 /* then test to make sure it is all still delalloc */
397 ret = test_range_bit(tree, delalloc_start, delalloc_end,
398 EXTENT_DELALLOC, cached_state);
399 if (!ret) {
400 unlock_extent(tree, delalloc_start, delalloc_end,
401 &cached_state);
402 __unlock_for_delalloc(inode, locked_page,
403 delalloc_start, delalloc_end);
404 cond_resched();
405 goto again;
406 }
407 free_extent_state(cached_state);
408 *start = delalloc_start;
409 *end = delalloc_end;
410out_failed:
411 return found;
412}
413
414void extent_clear_unlock_delalloc(struct btrfs_inode *inode, u64 start, u64 end,
415 struct page *locked_page,
416 u32 clear_bits, unsigned long page_ops)
417{
418 clear_extent_bit(&inode->io_tree, start, end, clear_bits, NULL);
419
420 __process_pages_contig(inode->vfs_inode.i_mapping, locked_page,
421 start, end, page_ops);
422}
423
424static bool btrfs_verify_page(struct page *page, u64 start)
425{
426 if (!fsverity_active(page->mapping->host) ||
427 PageUptodate(page) ||
428 start >= i_size_read(page->mapping->host))
429 return true;
430 return fsverity_verify_page(page);
431}
432
433static void end_page_read(struct page *page, bool uptodate, u64 start, u32 len)
434{
435 struct btrfs_fs_info *fs_info = page_to_fs_info(page);
436 struct folio *folio = page_folio(page);
437
438 ASSERT(page_offset(page) <= start &&
439 start + len <= page_offset(page) + PAGE_SIZE);
440
441 if (uptodate && btrfs_verify_page(page, start))
442 btrfs_folio_set_uptodate(fs_info, folio, start, len);
443 else
444 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
445
446 if (!btrfs_is_subpage(fs_info, page->mapping))
447 unlock_page(page);
448 else
449 btrfs_subpage_end_reader(fs_info, folio, start, len);
450}
451
452/*
453 * After a write IO is done, we need to:
454 *
455 * - clear the uptodate bits on error
456 * - clear the writeback bits in the extent tree for the range
457 * - filio_end_writeback() if there is no more pending io for the folio
458 *
459 * Scheduling is not allowed, so the extent state tree is expected
460 * to have one and only one object corresponding to this IO.
461 */
462static void end_bbio_data_write(struct btrfs_bio *bbio)
463{
464 struct btrfs_fs_info *fs_info = bbio->fs_info;
465 struct bio *bio = &bbio->bio;
466 int error = blk_status_to_errno(bio->bi_status);
467 struct folio_iter fi;
468 const u32 sectorsize = fs_info->sectorsize;
469
470 ASSERT(!bio_flagged(bio, BIO_CLONED));
471 bio_for_each_folio_all(fi, bio) {
472 struct folio *folio = fi.folio;
473 u64 start = folio_pos(folio) + fi.offset;
474 u32 len = fi.length;
475
476 /* Only order 0 (single page) folios are allowed for data. */
477 ASSERT(folio_order(folio) == 0);
478
479 /* Our read/write should always be sector aligned. */
480 if (!IS_ALIGNED(fi.offset, sectorsize))
481 btrfs_err(fs_info,
482 "partial page write in btrfs with offset %zu and length %zu",
483 fi.offset, fi.length);
484 else if (!IS_ALIGNED(fi.length, sectorsize))
485 btrfs_info(fs_info,
486 "incomplete page write with offset %zu and length %zu",
487 fi.offset, fi.length);
488
489 btrfs_finish_ordered_extent(bbio->ordered,
490 folio_page(folio, 0), start, len, !error);
491 if (error)
492 mapping_set_error(folio->mapping, error);
493 btrfs_folio_clear_writeback(fs_info, folio, start, len);
494 }
495
496 bio_put(bio);
497}
498
499/*
500 * Record previously processed extent range
501 *
502 * For endio_readpage_release_extent() to handle a full extent range, reducing
503 * the extent io operations.
504 */
505struct processed_extent {
506 struct btrfs_inode *inode;
507 /* Start of the range in @inode */
508 u64 start;
509 /* End of the range in @inode */
510 u64 end;
511 bool uptodate;
512};
513
514/*
515 * Try to release processed extent range
516 *
517 * May not release the extent range right now if the current range is
518 * contiguous to processed extent.
519 *
520 * Will release processed extent when any of @inode, @uptodate, the range is
521 * no longer contiguous to the processed range.
522 *
523 * Passing @inode == NULL will force processed extent to be released.
524 */
525static void endio_readpage_release_extent(struct processed_extent *processed,
526 struct btrfs_inode *inode, u64 start, u64 end,
527 bool uptodate)
528{
529 struct extent_state *cached = NULL;
530 struct extent_io_tree *tree;
531
532 /* The first extent, initialize @processed */
533 if (!processed->inode)
534 goto update;
535
536 /*
537 * Contiguous to processed extent, just uptodate the end.
538 *
539 * Several things to notice:
540 *
541 * - bio can be merged as long as on-disk bytenr is contiguous
542 * This means we can have page belonging to other inodes, thus need to
543 * check if the inode still matches.
544 * - bvec can contain range beyond current page for multi-page bvec
545 * Thus we need to do processed->end + 1 >= start check
546 */
547 if (processed->inode == inode && processed->uptodate == uptodate &&
548 processed->end + 1 >= start && end >= processed->end) {
549 processed->end = end;
550 return;
551 }
552
553 tree = &processed->inode->io_tree;
554 /*
555 * Now we don't have range contiguous to the processed range, release
556 * the processed range now.
557 */
558 unlock_extent(tree, processed->start, processed->end, &cached);
559
560update:
561 /* Update processed to current range */
562 processed->inode = inode;
563 processed->start = start;
564 processed->end = end;
565 processed->uptodate = uptodate;
566}
567
568static void begin_page_read(struct btrfs_fs_info *fs_info, struct page *page)
569{
570 struct folio *folio = page_folio(page);
571
572 ASSERT(folio_test_locked(folio));
573 if (!btrfs_is_subpage(fs_info, folio->mapping))
574 return;
575
576 ASSERT(folio_test_private(folio));
577 btrfs_subpage_start_reader(fs_info, folio, page_offset(page), PAGE_SIZE);
578}
579
580/*
581 * After a data read IO is done, we need to:
582 *
583 * - clear the uptodate bits on error
584 * - set the uptodate bits if things worked
585 * - set the folio up to date if all extents in the tree are uptodate
586 * - clear the lock bit in the extent tree
587 * - unlock the folio if there are no other extents locked for it
588 *
589 * Scheduling is not allowed, so the extent state tree is expected
590 * to have one and only one object corresponding to this IO.
591 */
592static void end_bbio_data_read(struct btrfs_bio *bbio)
593{
594 struct btrfs_fs_info *fs_info = bbio->fs_info;
595 struct bio *bio = &bbio->bio;
596 struct processed_extent processed = { 0 };
597 struct folio_iter fi;
598 const u32 sectorsize = fs_info->sectorsize;
599
600 ASSERT(!bio_flagged(bio, BIO_CLONED));
601 bio_for_each_folio_all(fi, &bbio->bio) {
602 bool uptodate = !bio->bi_status;
603 struct folio *folio = fi.folio;
604 struct inode *inode = folio->mapping->host;
605 u64 start;
606 u64 end;
607 u32 len;
608
609 /* For now only order 0 folios are supported for data. */
610 ASSERT(folio_order(folio) == 0);
611 btrfs_debug(fs_info,
612 "%s: bi_sector=%llu, err=%d, mirror=%u",
613 __func__, bio->bi_iter.bi_sector, bio->bi_status,
614 bbio->mirror_num);
615
616 /*
617 * We always issue full-sector reads, but if some block in a
618 * folio fails to read, blk_update_request() will advance
619 * bv_offset and adjust bv_len to compensate. Print a warning
620 * for unaligned offsets, and an error if they don't add up to
621 * a full sector.
622 */
623 if (!IS_ALIGNED(fi.offset, sectorsize))
624 btrfs_err(fs_info,
625 "partial page read in btrfs with offset %zu and length %zu",
626 fi.offset, fi.length);
627 else if (!IS_ALIGNED(fi.offset + fi.length, sectorsize))
628 btrfs_info(fs_info,
629 "incomplete page read with offset %zu and length %zu",
630 fi.offset, fi.length);
631
632 start = folio_pos(folio) + fi.offset;
633 end = start + fi.length - 1;
634 len = fi.length;
635
636 if (likely(uptodate)) {
637 loff_t i_size = i_size_read(inode);
638 pgoff_t end_index = i_size >> folio_shift(folio);
639
640 /*
641 * Zero out the remaining part if this range straddles
642 * i_size.
643 *
644 * Here we should only zero the range inside the folio,
645 * not touch anything else.
646 *
647 * NOTE: i_size is exclusive while end is inclusive.
648 */
649 if (folio_index(folio) == end_index && i_size <= end) {
650 u32 zero_start = max(offset_in_folio(folio, i_size),
651 offset_in_folio(folio, start));
652 u32 zero_len = offset_in_folio(folio, end) + 1 -
653 zero_start;
654
655 folio_zero_range(folio, zero_start, zero_len);
656 }
657 }
658
659 /* Update page status and unlock. */
660 end_page_read(folio_page(folio, 0), uptodate, start, len);
661 endio_readpage_release_extent(&processed, BTRFS_I(inode),
662 start, end, uptodate);
663 }
664 /* Release the last extent */
665 endio_readpage_release_extent(&processed, NULL, 0, 0, false);
666 bio_put(bio);
667}
668
669/*
670 * Populate every free slot in a provided array with pages.
671 *
672 * @nr_pages: number of pages to allocate
673 * @page_array: the array to fill with pages; any existing non-null entries in
674 * the array will be skipped
675 * @extra_gfp: the extra GFP flags for the allocation.
676 *
677 * Return: 0 if all pages were able to be allocated;
678 * -ENOMEM otherwise, the partially allocated pages would be freed and
679 * the array slots zeroed
680 */
681int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array,
682 gfp_t extra_gfp)
683{
684 const gfp_t gfp = GFP_NOFS | extra_gfp;
685 unsigned int allocated;
686
687 for (allocated = 0; allocated < nr_pages;) {
688 unsigned int last = allocated;
689
690 allocated = alloc_pages_bulk_array(gfp, nr_pages, page_array);
691 if (unlikely(allocated == last)) {
692 /* No progress, fail and do cleanup. */
693 for (int i = 0; i < allocated; i++) {
694 __free_page(page_array[i]);
695 page_array[i] = NULL;
696 }
697 return -ENOMEM;
698 }
699 }
700 return 0;
701}
702
703/*
704 * Populate needed folios for the extent buffer.
705 *
706 * For now, the folios populated are always in order 0 (aka, single page).
707 */
708static int alloc_eb_folio_array(struct extent_buffer *eb, gfp_t extra_gfp)
709{
710 struct page *page_array[INLINE_EXTENT_BUFFER_PAGES] = { 0 };
711 int num_pages = num_extent_pages(eb);
712 int ret;
713
714 ret = btrfs_alloc_page_array(num_pages, page_array, extra_gfp);
715 if (ret < 0)
716 return ret;
717
718 for (int i = 0; i < num_pages; i++)
719 eb->folios[i] = page_folio(page_array[i]);
720 eb->folio_size = PAGE_SIZE;
721 eb->folio_shift = PAGE_SHIFT;
722 return 0;
723}
724
725static bool btrfs_bio_is_contig(struct btrfs_bio_ctrl *bio_ctrl,
726 struct page *page, u64 disk_bytenr,
727 unsigned int pg_offset)
728{
729 struct bio *bio = &bio_ctrl->bbio->bio;
730 struct bio_vec *bvec = bio_last_bvec_all(bio);
731 const sector_t sector = disk_bytenr >> SECTOR_SHIFT;
732
733 if (bio_ctrl->compress_type != BTRFS_COMPRESS_NONE) {
734 /*
735 * For compression, all IO should have its logical bytenr set
736 * to the starting bytenr of the compressed extent.
737 */
738 return bio->bi_iter.bi_sector == sector;
739 }
740
741 /*
742 * The contig check requires the following conditions to be met:
743 *
744 * 1) The pages are belonging to the same inode
745 * This is implied by the call chain.
746 *
747 * 2) The range has adjacent logical bytenr
748 *
749 * 3) The range has adjacent file offset
750 * This is required for the usage of btrfs_bio->file_offset.
751 */
752 return bio_end_sector(bio) == sector &&
753 page_offset(bvec->bv_page) + bvec->bv_offset + bvec->bv_len ==
754 page_offset(page) + pg_offset;
755}
756
757static void alloc_new_bio(struct btrfs_inode *inode,
758 struct btrfs_bio_ctrl *bio_ctrl,
759 u64 disk_bytenr, u64 file_offset)
760{
761 struct btrfs_fs_info *fs_info = inode->root->fs_info;
762 struct btrfs_bio *bbio;
763
764 bbio = btrfs_bio_alloc(BIO_MAX_VECS, bio_ctrl->opf, fs_info,
765 bio_ctrl->end_io_func, NULL);
766 bbio->bio.bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
767 bbio->inode = inode;
768 bbio->file_offset = file_offset;
769 bio_ctrl->bbio = bbio;
770 bio_ctrl->len_to_oe_boundary = U32_MAX;
771
772 /* Limit data write bios to the ordered boundary. */
773 if (bio_ctrl->wbc) {
774 struct btrfs_ordered_extent *ordered;
775
776 ordered = btrfs_lookup_ordered_extent(inode, file_offset);
777 if (ordered) {
778 bio_ctrl->len_to_oe_boundary = min_t(u32, U32_MAX,
779 ordered->file_offset +
780 ordered->disk_num_bytes - file_offset);
781 bbio->ordered = ordered;
782 }
783
784 /*
785 * Pick the last added device to support cgroup writeback. For
786 * multi-device file systems this means blk-cgroup policies have
787 * to always be set on the last added/replaced device.
788 * This is a bit odd but has been like that for a long time.
789 */
790 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
791 wbc_init_bio(bio_ctrl->wbc, &bbio->bio);
792 }
793}
794
795/*
796 * @disk_bytenr: logical bytenr where the write will be
797 * @page: page to add to the bio
798 * @size: portion of page that we want to write to
799 * @pg_offset: offset of the new bio or to check whether we are adding
800 * a contiguous page to the previous one
801 *
802 * The will either add the page into the existing @bio_ctrl->bbio, or allocate a
803 * new one in @bio_ctrl->bbio.
804 * The mirror number for this IO should already be initizlied in
805 * @bio_ctrl->mirror_num.
806 */
807static void submit_extent_page(struct btrfs_bio_ctrl *bio_ctrl,
808 u64 disk_bytenr, struct page *page,
809 size_t size, unsigned long pg_offset)
810{
811 struct btrfs_inode *inode = page_to_inode(page);
812
813 ASSERT(pg_offset + size <= PAGE_SIZE);
814 ASSERT(bio_ctrl->end_io_func);
815
816 if (bio_ctrl->bbio &&
817 !btrfs_bio_is_contig(bio_ctrl, page, disk_bytenr, pg_offset))
818 submit_one_bio(bio_ctrl);
819
820 do {
821 u32 len = size;
822
823 /* Allocate new bio if needed */
824 if (!bio_ctrl->bbio) {
825 alloc_new_bio(inode, bio_ctrl, disk_bytenr,
826 page_offset(page) + pg_offset);
827 }
828
829 /* Cap to the current ordered extent boundary if there is one. */
830 if (len > bio_ctrl->len_to_oe_boundary) {
831 ASSERT(bio_ctrl->compress_type == BTRFS_COMPRESS_NONE);
832 ASSERT(is_data_inode(&inode->vfs_inode));
833 len = bio_ctrl->len_to_oe_boundary;
834 }
835
836 if (bio_add_page(&bio_ctrl->bbio->bio, page, len, pg_offset) != len) {
837 /* bio full: move on to a new one */
838 submit_one_bio(bio_ctrl);
839 continue;
840 }
841
842 if (bio_ctrl->wbc)
843 wbc_account_cgroup_owner(bio_ctrl->wbc, page, len);
844
845 size -= len;
846 pg_offset += len;
847 disk_bytenr += len;
848
849 /*
850 * len_to_oe_boundary defaults to U32_MAX, which isn't page or
851 * sector aligned. alloc_new_bio() then sets it to the end of
852 * our ordered extent for writes into zoned devices.
853 *
854 * When len_to_oe_boundary is tracking an ordered extent, we
855 * trust the ordered extent code to align things properly, and
856 * the check above to cap our write to the ordered extent
857 * boundary is correct.
858 *
859 * When len_to_oe_boundary is U32_MAX, the cap above would
860 * result in a 4095 byte IO for the last page right before
861 * we hit the bio limit of UINT_MAX. bio_add_page() has all
862 * the checks required to make sure we don't overflow the bio,
863 * and we should just ignore len_to_oe_boundary completely
864 * unless we're using it to track an ordered extent.
865 *
866 * It's pretty hard to make a bio sized U32_MAX, but it can
867 * happen when the page cache is able to feed us contiguous
868 * pages for large extents.
869 */
870 if (bio_ctrl->len_to_oe_boundary != U32_MAX)
871 bio_ctrl->len_to_oe_boundary -= len;
872
873 /* Ordered extent boundary: move on to a new bio. */
874 if (bio_ctrl->len_to_oe_boundary == 0)
875 submit_one_bio(bio_ctrl);
876 } while (size);
877}
878
879static int attach_extent_buffer_folio(struct extent_buffer *eb,
880 struct folio *folio,
881 struct btrfs_subpage *prealloc)
882{
883 struct btrfs_fs_info *fs_info = eb->fs_info;
884 int ret = 0;
885
886 /*
887 * If the page is mapped to btree inode, we should hold the private
888 * lock to prevent race.
889 * For cloned or dummy extent buffers, their pages are not mapped and
890 * will not race with any other ebs.
891 */
892 if (folio->mapping)
893 lockdep_assert_held(&folio->mapping->i_private_lock);
894
895 if (fs_info->nodesize >= PAGE_SIZE) {
896 if (!folio_test_private(folio))
897 folio_attach_private(folio, eb);
898 else
899 WARN_ON(folio_get_private(folio) != eb);
900 return 0;
901 }
902
903 /* Already mapped, just free prealloc */
904 if (folio_test_private(folio)) {
905 btrfs_free_subpage(prealloc);
906 return 0;
907 }
908
909 if (prealloc)
910 /* Has preallocated memory for subpage */
911 folio_attach_private(folio, prealloc);
912 else
913 /* Do new allocation to attach subpage */
914 ret = btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_METADATA);
915 return ret;
916}
917
918int set_page_extent_mapped(struct page *page)
919{
920 return set_folio_extent_mapped(page_folio(page));
921}
922
923int set_folio_extent_mapped(struct folio *folio)
924{
925 struct btrfs_fs_info *fs_info;
926
927 ASSERT(folio->mapping);
928
929 if (folio_test_private(folio))
930 return 0;
931
932 fs_info = folio_to_fs_info(folio);
933
934 if (btrfs_is_subpage(fs_info, folio->mapping))
935 return btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_DATA);
936
937 folio_attach_private(folio, (void *)EXTENT_FOLIO_PRIVATE);
938 return 0;
939}
940
941void clear_page_extent_mapped(struct page *page)
942{
943 struct folio *folio = page_folio(page);
944 struct btrfs_fs_info *fs_info;
945
946 ASSERT(page->mapping);
947
948 if (!folio_test_private(folio))
949 return;
950
951 fs_info = page_to_fs_info(page);
952 if (btrfs_is_subpage(fs_info, page->mapping))
953 return btrfs_detach_subpage(fs_info, folio);
954
955 folio_detach_private(folio);
956}
957
958static struct extent_map *__get_extent_map(struct inode *inode, struct page *page,
959 u64 start, u64 len, struct extent_map **em_cached)
960{
961 struct extent_map *em;
962
963 ASSERT(em_cached);
964
965 if (*em_cached) {
966 em = *em_cached;
967 if (extent_map_in_tree(em) && start >= em->start &&
968 start < extent_map_end(em)) {
969 refcount_inc(&em->refs);
970 return em;
971 }
972
973 free_extent_map(em);
974 *em_cached = NULL;
975 }
976
977 em = btrfs_get_extent(BTRFS_I(inode), page, start, len);
978 if (!IS_ERR(em)) {
979 BUG_ON(*em_cached);
980 refcount_inc(&em->refs);
981 *em_cached = em;
982 }
983 return em;
984}
985/*
986 * basic readpage implementation. Locked extent state structs are inserted
987 * into the tree that are removed when the IO is done (by the end_io
988 * handlers)
989 * XXX JDM: This needs looking at to ensure proper page locking
990 * return 0 on success, otherwise return error
991 */
992static int btrfs_do_readpage(struct page *page, struct extent_map **em_cached,
993 struct btrfs_bio_ctrl *bio_ctrl, u64 *prev_em_start)
994{
995 struct inode *inode = page->mapping->host;
996 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
997 u64 start = page_offset(page);
998 const u64 end = start + PAGE_SIZE - 1;
999 u64 cur = start;
1000 u64 extent_offset;
1001 u64 last_byte = i_size_read(inode);
1002 u64 block_start;
1003 struct extent_map *em;
1004 int ret = 0;
1005 size_t pg_offset = 0;
1006 size_t iosize;
1007 size_t blocksize = fs_info->sectorsize;
1008 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1009
1010 ret = set_page_extent_mapped(page);
1011 if (ret < 0) {
1012 unlock_extent(tree, start, end, NULL);
1013 unlock_page(page);
1014 return ret;
1015 }
1016
1017 if (page->index == last_byte >> PAGE_SHIFT) {
1018 size_t zero_offset = offset_in_page(last_byte);
1019
1020 if (zero_offset) {
1021 iosize = PAGE_SIZE - zero_offset;
1022 memzero_page(page, zero_offset, iosize);
1023 }
1024 }
1025 bio_ctrl->end_io_func = end_bbio_data_read;
1026 begin_page_read(fs_info, page);
1027 while (cur <= end) {
1028 enum btrfs_compression_type compress_type = BTRFS_COMPRESS_NONE;
1029 bool force_bio_submit = false;
1030 u64 disk_bytenr;
1031
1032 ASSERT(IS_ALIGNED(cur, fs_info->sectorsize));
1033 if (cur >= last_byte) {
1034 iosize = PAGE_SIZE - pg_offset;
1035 memzero_page(page, pg_offset, iosize);
1036 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1037 end_page_read(page, true, cur, iosize);
1038 break;
1039 }
1040 em = __get_extent_map(inode, page, cur, end - cur + 1, em_cached);
1041 if (IS_ERR(em)) {
1042 unlock_extent(tree, cur, end, NULL);
1043 end_page_read(page, false, cur, end + 1 - cur);
1044 return PTR_ERR(em);
1045 }
1046 extent_offset = cur - em->start;
1047 BUG_ON(extent_map_end(em) <= cur);
1048 BUG_ON(end < cur);
1049
1050 compress_type = extent_map_compression(em);
1051
1052 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1053 iosize = ALIGN(iosize, blocksize);
1054 if (compress_type != BTRFS_COMPRESS_NONE)
1055 disk_bytenr = em->block_start;
1056 else
1057 disk_bytenr = em->block_start + extent_offset;
1058 block_start = em->block_start;
1059 if (em->flags & EXTENT_FLAG_PREALLOC)
1060 block_start = EXTENT_MAP_HOLE;
1061
1062 /*
1063 * If we have a file range that points to a compressed extent
1064 * and it's followed by a consecutive file range that points
1065 * to the same compressed extent (possibly with a different
1066 * offset and/or length, so it either points to the whole extent
1067 * or only part of it), we must make sure we do not submit a
1068 * single bio to populate the pages for the 2 ranges because
1069 * this makes the compressed extent read zero out the pages
1070 * belonging to the 2nd range. Imagine the following scenario:
1071 *
1072 * File layout
1073 * [0 - 8K] [8K - 24K]
1074 * | |
1075 * | |
1076 * points to extent X, points to extent X,
1077 * offset 4K, length of 8K offset 0, length 16K
1078 *
1079 * [extent X, compressed length = 4K uncompressed length = 16K]
1080 *
1081 * If the bio to read the compressed extent covers both ranges,
1082 * it will decompress extent X into the pages belonging to the
1083 * first range and then it will stop, zeroing out the remaining
1084 * pages that belong to the other range that points to extent X.
1085 * So here we make sure we submit 2 bios, one for the first
1086 * range and another one for the third range. Both will target
1087 * the same physical extent from disk, but we can't currently
1088 * make the compressed bio endio callback populate the pages
1089 * for both ranges because each compressed bio is tightly
1090 * coupled with a single extent map, and each range can have
1091 * an extent map with a different offset value relative to the
1092 * uncompressed data of our extent and different lengths. This
1093 * is a corner case so we prioritize correctness over
1094 * non-optimal behavior (submitting 2 bios for the same extent).
1095 */
1096 if (compress_type != BTRFS_COMPRESS_NONE &&
1097 prev_em_start && *prev_em_start != (u64)-1 &&
1098 *prev_em_start != em->start)
1099 force_bio_submit = true;
1100
1101 if (prev_em_start)
1102 *prev_em_start = em->start;
1103
1104 free_extent_map(em);
1105 em = NULL;
1106
1107 /* we've found a hole, just zero and go on */
1108 if (block_start == EXTENT_MAP_HOLE) {
1109 memzero_page(page, pg_offset, iosize);
1110
1111 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1112 end_page_read(page, true, cur, iosize);
1113 cur = cur + iosize;
1114 pg_offset += iosize;
1115 continue;
1116 }
1117 /* the get_extent function already copied into the page */
1118 if (block_start == EXTENT_MAP_INLINE) {
1119 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1120 end_page_read(page, true, cur, iosize);
1121 cur = cur + iosize;
1122 pg_offset += iosize;
1123 continue;
1124 }
1125
1126 if (bio_ctrl->compress_type != compress_type) {
1127 submit_one_bio(bio_ctrl);
1128 bio_ctrl->compress_type = compress_type;
1129 }
1130
1131 if (force_bio_submit)
1132 submit_one_bio(bio_ctrl);
1133 submit_extent_page(bio_ctrl, disk_bytenr, page, iosize,
1134 pg_offset);
1135 cur = cur + iosize;
1136 pg_offset += iosize;
1137 }
1138
1139 return 0;
1140}
1141
1142int btrfs_read_folio(struct file *file, struct folio *folio)
1143{
1144 struct page *page = &folio->page;
1145 struct btrfs_inode *inode = page_to_inode(page);
1146 u64 start = page_offset(page);
1147 u64 end = start + PAGE_SIZE - 1;
1148 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ };
1149 struct extent_map *em_cached = NULL;
1150 int ret;
1151
1152 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
1153
1154 ret = btrfs_do_readpage(page, &em_cached, &bio_ctrl, NULL);
1155 free_extent_map(em_cached);
1156
1157 /*
1158 * If btrfs_do_readpage() failed we will want to submit the assembled
1159 * bio to do the cleanup.
1160 */
1161 submit_one_bio(&bio_ctrl);
1162 return ret;
1163}
1164
1165static inline void contiguous_readpages(struct page *pages[], int nr_pages,
1166 u64 start, u64 end,
1167 struct extent_map **em_cached,
1168 struct btrfs_bio_ctrl *bio_ctrl,
1169 u64 *prev_em_start)
1170{
1171 struct btrfs_inode *inode = page_to_inode(pages[0]);
1172 int index;
1173
1174 ASSERT(em_cached);
1175
1176 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
1177
1178 for (index = 0; index < nr_pages; index++) {
1179 btrfs_do_readpage(pages[index], em_cached, bio_ctrl,
1180 prev_em_start);
1181 put_page(pages[index]);
1182 }
1183}
1184
1185/*
1186 * helper for __extent_writepage, doing all of the delayed allocation setup.
1187 *
1188 * This returns 1 if btrfs_run_delalloc_range function did all the work required
1189 * to write the page (copy into inline extent). In this case the IO has
1190 * been started and the page is already unlocked.
1191 *
1192 * This returns 0 if all went well (page still locked)
1193 * This returns < 0 if there were errors (page still locked)
1194 */
1195static noinline_for_stack int writepage_delalloc(struct btrfs_inode *inode,
1196 struct page *page, struct writeback_control *wbc)
1197{
1198 const u64 page_start = page_offset(page);
1199 const u64 page_end = page_start + PAGE_SIZE - 1;
1200 u64 delalloc_start = page_start;
1201 u64 delalloc_end = page_end;
1202 u64 delalloc_to_write = 0;
1203 int ret = 0;
1204
1205 while (delalloc_start < page_end) {
1206 delalloc_end = page_end;
1207 if (!find_lock_delalloc_range(&inode->vfs_inode, page,
1208 &delalloc_start, &delalloc_end)) {
1209 delalloc_start = delalloc_end + 1;
1210 continue;
1211 }
1212
1213 ret = btrfs_run_delalloc_range(inode, page, delalloc_start,
1214 delalloc_end, wbc);
1215 if (ret < 0)
1216 return ret;
1217
1218 delalloc_start = delalloc_end + 1;
1219 }
1220
1221 /*
1222 * delalloc_end is already one less than the total length, so
1223 * we don't subtract one from PAGE_SIZE
1224 */
1225 delalloc_to_write +=
1226 DIV_ROUND_UP(delalloc_end + 1 - page_start, PAGE_SIZE);
1227
1228 /*
1229 * If btrfs_run_dealloc_range() already started I/O and unlocked
1230 * the pages, we just need to account for them here.
1231 */
1232 if (ret == 1) {
1233 wbc->nr_to_write -= delalloc_to_write;
1234 return 1;
1235 }
1236
1237 if (wbc->nr_to_write < delalloc_to_write) {
1238 int thresh = 8192;
1239
1240 if (delalloc_to_write < thresh * 2)
1241 thresh = delalloc_to_write;
1242 wbc->nr_to_write = min_t(u64, delalloc_to_write,
1243 thresh);
1244 }
1245
1246 return 0;
1247}
1248
1249/*
1250 * Find the first byte we need to write.
1251 *
1252 * For subpage, one page can contain several sectors, and
1253 * __extent_writepage_io() will just grab all extent maps in the page
1254 * range and try to submit all non-inline/non-compressed extents.
1255 *
1256 * This is a big problem for subpage, we shouldn't re-submit already written
1257 * data at all.
1258 * This function will lookup subpage dirty bit to find which range we really
1259 * need to submit.
1260 *
1261 * Return the next dirty range in [@start, @end).
1262 * If no dirty range is found, @start will be page_offset(page) + PAGE_SIZE.
1263 */
1264static void find_next_dirty_byte(struct btrfs_fs_info *fs_info,
1265 struct page *page, u64 *start, u64 *end)
1266{
1267 struct folio *folio = page_folio(page);
1268 struct btrfs_subpage *subpage = folio_get_private(folio);
1269 struct btrfs_subpage_info *spi = fs_info->subpage_info;
1270 u64 orig_start = *start;
1271 /* Declare as unsigned long so we can use bitmap ops */
1272 unsigned long flags;
1273 int range_start_bit;
1274 int range_end_bit;
1275
1276 /*
1277 * For regular sector size == page size case, since one page only
1278 * contains one sector, we return the page offset directly.
1279 */
1280 if (!btrfs_is_subpage(fs_info, page->mapping)) {
1281 *start = page_offset(page);
1282 *end = page_offset(page) + PAGE_SIZE;
1283 return;
1284 }
1285
1286 range_start_bit = spi->dirty_offset +
1287 (offset_in_page(orig_start) >> fs_info->sectorsize_bits);
1288
1289 /* We should have the page locked, but just in case */
1290 spin_lock_irqsave(&subpage->lock, flags);
1291 bitmap_next_set_region(subpage->bitmaps, &range_start_bit, &range_end_bit,
1292 spi->dirty_offset + spi->bitmap_nr_bits);
1293 spin_unlock_irqrestore(&subpage->lock, flags);
1294
1295 range_start_bit -= spi->dirty_offset;
1296 range_end_bit -= spi->dirty_offset;
1297
1298 *start = page_offset(page) + range_start_bit * fs_info->sectorsize;
1299 *end = page_offset(page) + range_end_bit * fs_info->sectorsize;
1300}
1301
1302/*
1303 * helper for __extent_writepage. This calls the writepage start hooks,
1304 * and does the loop to map the page into extents and bios.
1305 *
1306 * We return 1 if the IO is started and the page is unlocked,
1307 * 0 if all went well (page still locked)
1308 * < 0 if there were errors (page still locked)
1309 */
1310static noinline_for_stack int __extent_writepage_io(struct btrfs_inode *inode,
1311 struct page *page,
1312 struct btrfs_bio_ctrl *bio_ctrl,
1313 loff_t i_size,
1314 int *nr_ret)
1315{
1316 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1317 u64 cur = page_offset(page);
1318 u64 end = cur + PAGE_SIZE - 1;
1319 u64 extent_offset;
1320 u64 block_start;
1321 struct extent_map *em;
1322 int ret = 0;
1323 int nr = 0;
1324
1325 ret = btrfs_writepage_cow_fixup(page);
1326 if (ret) {
1327 /* Fixup worker will requeue */
1328 redirty_page_for_writepage(bio_ctrl->wbc, page);
1329 unlock_page(page);
1330 return 1;
1331 }
1332
1333 bio_ctrl->end_io_func = end_bbio_data_write;
1334 while (cur <= end) {
1335 u32 len = end - cur + 1;
1336 u64 disk_bytenr;
1337 u64 em_end;
1338 u64 dirty_range_start = cur;
1339 u64 dirty_range_end;
1340 u32 iosize;
1341
1342 if (cur >= i_size) {
1343 btrfs_mark_ordered_io_finished(inode, page, cur, len,
1344 true);
1345 /*
1346 * This range is beyond i_size, thus we don't need to
1347 * bother writing back.
1348 * But we still need to clear the dirty subpage bit, or
1349 * the next time the page gets dirtied, we will try to
1350 * writeback the sectors with subpage dirty bits,
1351 * causing writeback without ordered extent.
1352 */
1353 btrfs_folio_clear_dirty(fs_info, page_folio(page), cur, len);
1354 break;
1355 }
1356
1357 find_next_dirty_byte(fs_info, page, &dirty_range_start,
1358 &dirty_range_end);
1359 if (cur < dirty_range_start) {
1360 cur = dirty_range_start;
1361 continue;
1362 }
1363
1364 em = btrfs_get_extent(inode, NULL, cur, len);
1365 if (IS_ERR(em)) {
1366 ret = PTR_ERR_OR_ZERO(em);
1367 goto out_error;
1368 }
1369
1370 extent_offset = cur - em->start;
1371 em_end = extent_map_end(em);
1372 ASSERT(cur <= em_end);
1373 ASSERT(cur < end);
1374 ASSERT(IS_ALIGNED(em->start, fs_info->sectorsize));
1375 ASSERT(IS_ALIGNED(em->len, fs_info->sectorsize));
1376
1377 block_start = em->block_start;
1378 disk_bytenr = em->block_start + extent_offset;
1379
1380 ASSERT(!extent_map_is_compressed(em));
1381 ASSERT(block_start != EXTENT_MAP_HOLE);
1382 ASSERT(block_start != EXTENT_MAP_INLINE);
1383
1384 /*
1385 * Note that em_end from extent_map_end() and dirty_range_end from
1386 * find_next_dirty_byte() are all exclusive
1387 */
1388 iosize = min(min(em_end, end + 1), dirty_range_end) - cur;
1389 free_extent_map(em);
1390 em = NULL;
1391
1392 btrfs_set_range_writeback(inode, cur, cur + iosize - 1);
1393 if (!PageWriteback(page)) {
1394 btrfs_err(inode->root->fs_info,
1395 "page %lu not writeback, cur %llu end %llu",
1396 page->index, cur, end);
1397 }
1398
1399 /*
1400 * Although the PageDirty bit is cleared before entering this
1401 * function, subpage dirty bit is not cleared.
1402 * So clear subpage dirty bit here so next time we won't submit
1403 * page for range already written to disk.
1404 */
1405 btrfs_folio_clear_dirty(fs_info, page_folio(page), cur, iosize);
1406
1407 submit_extent_page(bio_ctrl, disk_bytenr, page, iosize,
1408 cur - page_offset(page));
1409 cur += iosize;
1410 nr++;
1411 }
1412
1413 btrfs_folio_assert_not_dirty(fs_info, page_folio(page));
1414 *nr_ret = nr;
1415 return 0;
1416
1417out_error:
1418 /*
1419 * If we finish without problem, we should not only clear page dirty,
1420 * but also empty subpage dirty bits
1421 */
1422 *nr_ret = nr;
1423 return ret;
1424}
1425
1426/*
1427 * the writepage semantics are similar to regular writepage. extent
1428 * records are inserted to lock ranges in the tree, and as dirty areas
1429 * are found, they are marked writeback. Then the lock bits are removed
1430 * and the end_io handler clears the writeback ranges
1431 *
1432 * Return 0 if everything goes well.
1433 * Return <0 for error.
1434 */
1435static int __extent_writepage(struct page *page, struct btrfs_bio_ctrl *bio_ctrl)
1436{
1437 struct folio *folio = page_folio(page);
1438 struct inode *inode = page->mapping->host;
1439 const u64 page_start = page_offset(page);
1440 int ret;
1441 int nr = 0;
1442 size_t pg_offset;
1443 loff_t i_size = i_size_read(inode);
1444 unsigned long end_index = i_size >> PAGE_SHIFT;
1445
1446 trace___extent_writepage(page, inode, bio_ctrl->wbc);
1447
1448 WARN_ON(!PageLocked(page));
1449
1450 pg_offset = offset_in_page(i_size);
1451 if (page->index > end_index ||
1452 (page->index == end_index && !pg_offset)) {
1453 folio_invalidate(folio, 0, folio_size(folio));
1454 folio_unlock(folio);
1455 return 0;
1456 }
1457
1458 if (page->index == end_index)
1459 memzero_page(page, pg_offset, PAGE_SIZE - pg_offset);
1460
1461 ret = set_page_extent_mapped(page);
1462 if (ret < 0)
1463 goto done;
1464
1465 ret = writepage_delalloc(BTRFS_I(inode), page, bio_ctrl->wbc);
1466 if (ret == 1)
1467 return 0;
1468 if (ret)
1469 goto done;
1470
1471 ret = __extent_writepage_io(BTRFS_I(inode), page, bio_ctrl, i_size, &nr);
1472 if (ret == 1)
1473 return 0;
1474
1475 bio_ctrl->wbc->nr_to_write--;
1476
1477done:
1478 if (nr == 0) {
1479 /* make sure the mapping tag for page dirty gets cleared */
1480 set_page_writeback(page);
1481 end_page_writeback(page);
1482 }
1483 if (ret) {
1484 btrfs_mark_ordered_io_finished(BTRFS_I(inode), page, page_start,
1485 PAGE_SIZE, !ret);
1486 mapping_set_error(page->mapping, ret);
1487 }
1488 unlock_page(page);
1489 ASSERT(ret <= 0);
1490 return ret;
1491}
1492
1493void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
1494{
1495 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
1496 TASK_UNINTERRUPTIBLE);
1497}
1498
1499/*
1500 * Lock extent buffer status and pages for writeback.
1501 *
1502 * Return %false if the extent buffer doesn't need to be submitted (e.g. the
1503 * extent buffer is not dirty)
1504 * Return %true is the extent buffer is submitted to bio.
1505 */
1506static noinline_for_stack bool lock_extent_buffer_for_io(struct extent_buffer *eb,
1507 struct writeback_control *wbc)
1508{
1509 struct btrfs_fs_info *fs_info = eb->fs_info;
1510 bool ret = false;
1511
1512 btrfs_tree_lock(eb);
1513 while (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
1514 btrfs_tree_unlock(eb);
1515 if (wbc->sync_mode != WB_SYNC_ALL)
1516 return false;
1517 wait_on_extent_buffer_writeback(eb);
1518 btrfs_tree_lock(eb);
1519 }
1520
1521 /*
1522 * We need to do this to prevent races in people who check if the eb is
1523 * under IO since we can end up having no IO bits set for a short period
1524 * of time.
1525 */
1526 spin_lock(&eb->refs_lock);
1527 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
1528 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1529 spin_unlock(&eb->refs_lock);
1530 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
1531 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
1532 -eb->len,
1533 fs_info->dirty_metadata_batch);
1534 ret = true;
1535 } else {
1536 spin_unlock(&eb->refs_lock);
1537 }
1538 btrfs_tree_unlock(eb);
1539 return ret;
1540}
1541
1542static void set_btree_ioerr(struct extent_buffer *eb)
1543{
1544 struct btrfs_fs_info *fs_info = eb->fs_info;
1545
1546 set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1547
1548 /*
1549 * A read may stumble upon this buffer later, make sure that it gets an
1550 * error and knows there was an error.
1551 */
1552 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
1553
1554 /*
1555 * We need to set the mapping with the io error as well because a write
1556 * error will flip the file system readonly, and then syncfs() will
1557 * return a 0 because we are readonly if we don't modify the err seq for
1558 * the superblock.
1559 */
1560 mapping_set_error(eb->fs_info->btree_inode->i_mapping, -EIO);
1561
1562 /*
1563 * If writeback for a btree extent that doesn't belong to a log tree
1564 * failed, increment the counter transaction->eb_write_errors.
1565 * We do this because while the transaction is running and before it's
1566 * committing (when we call filemap_fdata[write|wait]_range against
1567 * the btree inode), we might have
1568 * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
1569 * returns an error or an error happens during writeback, when we're
1570 * committing the transaction we wouldn't know about it, since the pages
1571 * can be no longer dirty nor marked anymore for writeback (if a
1572 * subsequent modification to the extent buffer didn't happen before the
1573 * transaction commit), which makes filemap_fdata[write|wait]_range not
1574 * able to find the pages tagged with SetPageError at transaction
1575 * commit time. So if this happens we must abort the transaction,
1576 * otherwise we commit a super block with btree roots that point to
1577 * btree nodes/leafs whose content on disk is invalid - either garbage
1578 * or the content of some node/leaf from a past generation that got
1579 * cowed or deleted and is no longer valid.
1580 *
1581 * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
1582 * not be enough - we need to distinguish between log tree extents vs
1583 * non-log tree extents, and the next filemap_fdatawait_range() call
1584 * will catch and clear such errors in the mapping - and that call might
1585 * be from a log sync and not from a transaction commit. Also, checking
1586 * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
1587 * not done and would not be reliable - the eb might have been released
1588 * from memory and reading it back again means that flag would not be
1589 * set (since it's a runtime flag, not persisted on disk).
1590 *
1591 * Using the flags below in the btree inode also makes us achieve the
1592 * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
1593 * writeback for all dirty pages and before filemap_fdatawait_range()
1594 * is called, the writeback for all dirty pages had already finished
1595 * with errors - because we were not using AS_EIO/AS_ENOSPC,
1596 * filemap_fdatawait_range() would return success, as it could not know
1597 * that writeback errors happened (the pages were no longer tagged for
1598 * writeback).
1599 */
1600 switch (eb->log_index) {
1601 case -1:
1602 set_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags);
1603 break;
1604 case 0:
1605 set_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
1606 break;
1607 case 1:
1608 set_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
1609 break;
1610 default:
1611 BUG(); /* unexpected, logic error */
1612 }
1613}
1614
1615/*
1616 * The endio specific version which won't touch any unsafe spinlock in endio
1617 * context.
1618 */
1619static struct extent_buffer *find_extent_buffer_nolock(
1620 struct btrfs_fs_info *fs_info, u64 start)
1621{
1622 struct extent_buffer *eb;
1623
1624 rcu_read_lock();
1625 eb = radix_tree_lookup(&fs_info->buffer_radix,
1626 start >> fs_info->sectorsize_bits);
1627 if (eb && atomic_inc_not_zero(&eb->refs)) {
1628 rcu_read_unlock();
1629 return eb;
1630 }
1631 rcu_read_unlock();
1632 return NULL;
1633}
1634
1635static void end_bbio_meta_write(struct btrfs_bio *bbio)
1636{
1637 struct extent_buffer *eb = bbio->private;
1638 struct btrfs_fs_info *fs_info = eb->fs_info;
1639 bool uptodate = !bbio->bio.bi_status;
1640 struct folio_iter fi;
1641 u32 bio_offset = 0;
1642
1643 if (!uptodate)
1644 set_btree_ioerr(eb);
1645
1646 bio_for_each_folio_all(fi, &bbio->bio) {
1647 u64 start = eb->start + bio_offset;
1648 struct folio *folio = fi.folio;
1649 u32 len = fi.length;
1650
1651 btrfs_folio_clear_writeback(fs_info, folio, start, len);
1652 bio_offset += len;
1653 }
1654
1655 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1656 smp_mb__after_atomic();
1657 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
1658
1659 bio_put(&bbio->bio);
1660}
1661
1662static void prepare_eb_write(struct extent_buffer *eb)
1663{
1664 u32 nritems;
1665 unsigned long start;
1666 unsigned long end;
1667
1668 clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1669
1670 /* Set btree blocks beyond nritems with 0 to avoid stale content */
1671 nritems = btrfs_header_nritems(eb);
1672 if (btrfs_header_level(eb) > 0) {
1673 end = btrfs_node_key_ptr_offset(eb, nritems);
1674 memzero_extent_buffer(eb, end, eb->len - end);
1675 } else {
1676 /*
1677 * Leaf:
1678 * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
1679 */
1680 start = btrfs_item_nr_offset(eb, nritems);
1681 end = btrfs_item_nr_offset(eb, 0);
1682 if (nritems == 0)
1683 end += BTRFS_LEAF_DATA_SIZE(eb->fs_info);
1684 else
1685 end += btrfs_item_offset(eb, nritems - 1);
1686 memzero_extent_buffer(eb, start, end - start);
1687 }
1688}
1689
1690static noinline_for_stack void write_one_eb(struct extent_buffer *eb,
1691 struct writeback_control *wbc)
1692{
1693 struct btrfs_fs_info *fs_info = eb->fs_info;
1694 struct btrfs_bio *bbio;
1695
1696 prepare_eb_write(eb);
1697
1698 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
1699 REQ_OP_WRITE | REQ_META | wbc_to_write_flags(wbc),
1700 eb->fs_info, end_bbio_meta_write, eb);
1701 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
1702 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
1703 wbc_init_bio(wbc, &bbio->bio);
1704 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
1705 bbio->file_offset = eb->start;
1706 if (fs_info->nodesize < PAGE_SIZE) {
1707 struct folio *folio = eb->folios[0];
1708 bool ret;
1709
1710 folio_lock(folio);
1711 btrfs_subpage_set_writeback(fs_info, folio, eb->start, eb->len);
1712 if (btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start,
1713 eb->len)) {
1714 folio_clear_dirty_for_io(folio);
1715 wbc->nr_to_write--;
1716 }
1717 ret = bio_add_folio(&bbio->bio, folio, eb->len,
1718 eb->start - folio_pos(folio));
1719 ASSERT(ret);
1720 wbc_account_cgroup_owner(wbc, folio_page(folio, 0), eb->len);
1721 folio_unlock(folio);
1722 } else {
1723 int num_folios = num_extent_folios(eb);
1724
1725 for (int i = 0; i < num_folios; i++) {
1726 struct folio *folio = eb->folios[i];
1727 bool ret;
1728
1729 folio_lock(folio);
1730 folio_clear_dirty_for_io(folio);
1731 folio_start_writeback(folio);
1732 ret = bio_add_folio(&bbio->bio, folio, eb->folio_size, 0);
1733 ASSERT(ret);
1734 wbc_account_cgroup_owner(wbc, folio_page(folio, 0),
1735 eb->folio_size);
1736 wbc->nr_to_write -= folio_nr_pages(folio);
1737 folio_unlock(folio);
1738 }
1739 }
1740 btrfs_submit_bio(bbio, 0);
1741}
1742
1743/*
1744 * Submit one subpage btree page.
1745 *
1746 * The main difference to submit_eb_page() is:
1747 * - Page locking
1748 * For subpage, we don't rely on page locking at all.
1749 *
1750 * - Flush write bio
1751 * We only flush bio if we may be unable to fit current extent buffers into
1752 * current bio.
1753 *
1754 * Return >=0 for the number of submitted extent buffers.
1755 * Return <0 for fatal error.
1756 */
1757static int submit_eb_subpage(struct page *page, struct writeback_control *wbc)
1758{
1759 struct btrfs_fs_info *fs_info = page_to_fs_info(page);
1760 struct folio *folio = page_folio(page);
1761 int submitted = 0;
1762 u64 page_start = page_offset(page);
1763 int bit_start = 0;
1764 int sectors_per_node = fs_info->nodesize >> fs_info->sectorsize_bits;
1765
1766 /* Lock and write each dirty extent buffers in the range */
1767 while (bit_start < fs_info->subpage_info->bitmap_nr_bits) {
1768 struct btrfs_subpage *subpage = folio_get_private(folio);
1769 struct extent_buffer *eb;
1770 unsigned long flags;
1771 u64 start;
1772
1773 /*
1774 * Take private lock to ensure the subpage won't be detached
1775 * in the meantime.
1776 */
1777 spin_lock(&page->mapping->i_private_lock);
1778 if (!folio_test_private(folio)) {
1779 spin_unlock(&page->mapping->i_private_lock);
1780 break;
1781 }
1782 spin_lock_irqsave(&subpage->lock, flags);
1783 if (!test_bit(bit_start + fs_info->subpage_info->dirty_offset,
1784 subpage->bitmaps)) {
1785 spin_unlock_irqrestore(&subpage->lock, flags);
1786 spin_unlock(&page->mapping->i_private_lock);
1787 bit_start++;
1788 continue;
1789 }
1790
1791 start = page_start + bit_start * fs_info->sectorsize;
1792 bit_start += sectors_per_node;
1793
1794 /*
1795 * Here we just want to grab the eb without touching extra
1796 * spin locks, so call find_extent_buffer_nolock().
1797 */
1798 eb = find_extent_buffer_nolock(fs_info, start);
1799 spin_unlock_irqrestore(&subpage->lock, flags);
1800 spin_unlock(&page->mapping->i_private_lock);
1801
1802 /*
1803 * The eb has already reached 0 refs thus find_extent_buffer()
1804 * doesn't return it. We don't need to write back such eb
1805 * anyway.
1806 */
1807 if (!eb)
1808 continue;
1809
1810 if (lock_extent_buffer_for_io(eb, wbc)) {
1811 write_one_eb(eb, wbc);
1812 submitted++;
1813 }
1814 free_extent_buffer(eb);
1815 }
1816 return submitted;
1817}
1818
1819/*
1820 * Submit all page(s) of one extent buffer.
1821 *
1822 * @page: the page of one extent buffer
1823 * @eb_context: to determine if we need to submit this page, if current page
1824 * belongs to this eb, we don't need to submit
1825 *
1826 * The caller should pass each page in their bytenr order, and here we use
1827 * @eb_context to determine if we have submitted pages of one extent buffer.
1828 *
1829 * If we have, we just skip until we hit a new page that doesn't belong to
1830 * current @eb_context.
1831 *
1832 * If not, we submit all the page(s) of the extent buffer.
1833 *
1834 * Return >0 if we have submitted the extent buffer successfully.
1835 * Return 0 if we don't need to submit the page, as it's already submitted by
1836 * previous call.
1837 * Return <0 for fatal error.
1838 */
1839static int submit_eb_page(struct page *page, struct btrfs_eb_write_context *ctx)
1840{
1841 struct writeback_control *wbc = ctx->wbc;
1842 struct address_space *mapping = page->mapping;
1843 struct folio *folio = page_folio(page);
1844 struct extent_buffer *eb;
1845 int ret;
1846
1847 if (!folio_test_private(folio))
1848 return 0;
1849
1850 if (page_to_fs_info(page)->nodesize < PAGE_SIZE)
1851 return submit_eb_subpage(page, wbc);
1852
1853 spin_lock(&mapping->i_private_lock);
1854 if (!folio_test_private(folio)) {
1855 spin_unlock(&mapping->i_private_lock);
1856 return 0;
1857 }
1858
1859 eb = folio_get_private(folio);
1860
1861 /*
1862 * Shouldn't happen and normally this would be a BUG_ON but no point
1863 * crashing the machine for something we can survive anyway.
1864 */
1865 if (WARN_ON(!eb)) {
1866 spin_unlock(&mapping->i_private_lock);
1867 return 0;
1868 }
1869
1870 if (eb == ctx->eb) {
1871 spin_unlock(&mapping->i_private_lock);
1872 return 0;
1873 }
1874 ret = atomic_inc_not_zero(&eb->refs);
1875 spin_unlock(&mapping->i_private_lock);
1876 if (!ret)
1877 return 0;
1878
1879 ctx->eb = eb;
1880
1881 ret = btrfs_check_meta_write_pointer(eb->fs_info, ctx);
1882 if (ret) {
1883 if (ret == -EBUSY)
1884 ret = 0;
1885 free_extent_buffer(eb);
1886 return ret;
1887 }
1888
1889 if (!lock_extent_buffer_for_io(eb, wbc)) {
1890 free_extent_buffer(eb);
1891 return 0;
1892 }
1893 /* Implies write in zoned mode. */
1894 if (ctx->zoned_bg) {
1895 /* Mark the last eb in the block group. */
1896 btrfs_schedule_zone_finish_bg(ctx->zoned_bg, eb);
1897 ctx->zoned_bg->meta_write_pointer += eb->len;
1898 }
1899 write_one_eb(eb, wbc);
1900 free_extent_buffer(eb);
1901 return 1;
1902}
1903
1904int btree_write_cache_pages(struct address_space *mapping,
1905 struct writeback_control *wbc)
1906{
1907 struct btrfs_eb_write_context ctx = { .wbc = wbc };
1908 struct btrfs_fs_info *fs_info = inode_to_fs_info(mapping->host);
1909 int ret = 0;
1910 int done = 0;
1911 int nr_to_write_done = 0;
1912 struct folio_batch fbatch;
1913 unsigned int nr_folios;
1914 pgoff_t index;
1915 pgoff_t end; /* Inclusive */
1916 int scanned = 0;
1917 xa_mark_t tag;
1918
1919 folio_batch_init(&fbatch);
1920 if (wbc->range_cyclic) {
1921 index = mapping->writeback_index; /* Start from prev offset */
1922 end = -1;
1923 /*
1924 * Start from the beginning does not need to cycle over the
1925 * range, mark it as scanned.
1926 */
1927 scanned = (index == 0);
1928 } else {
1929 index = wbc->range_start >> PAGE_SHIFT;
1930 end = wbc->range_end >> PAGE_SHIFT;
1931 scanned = 1;
1932 }
1933 if (wbc->sync_mode == WB_SYNC_ALL)
1934 tag = PAGECACHE_TAG_TOWRITE;
1935 else
1936 tag = PAGECACHE_TAG_DIRTY;
1937 btrfs_zoned_meta_io_lock(fs_info);
1938retry:
1939 if (wbc->sync_mode == WB_SYNC_ALL)
1940 tag_pages_for_writeback(mapping, index, end);
1941 while (!done && !nr_to_write_done && (index <= end) &&
1942 (nr_folios = filemap_get_folios_tag(mapping, &index, end,
1943 tag, &fbatch))) {
1944 unsigned i;
1945
1946 for (i = 0; i < nr_folios; i++) {
1947 struct folio *folio = fbatch.folios[i];
1948
1949 ret = submit_eb_page(&folio->page, &ctx);
1950 if (ret == 0)
1951 continue;
1952 if (ret < 0) {
1953 done = 1;
1954 break;
1955 }
1956
1957 /*
1958 * the filesystem may choose to bump up nr_to_write.
1959 * We have to make sure to honor the new nr_to_write
1960 * at any time
1961 */
1962 nr_to_write_done = wbc->nr_to_write <= 0;
1963 }
1964 folio_batch_release(&fbatch);
1965 cond_resched();
1966 }
1967 if (!scanned && !done) {
1968 /*
1969 * We hit the last page and there is more work to be done: wrap
1970 * back to the start of the file
1971 */
1972 scanned = 1;
1973 index = 0;
1974 goto retry;
1975 }
1976 /*
1977 * If something went wrong, don't allow any metadata write bio to be
1978 * submitted.
1979 *
1980 * This would prevent use-after-free if we had dirty pages not
1981 * cleaned up, which can still happen by fuzzed images.
1982 *
1983 * - Bad extent tree
1984 * Allowing existing tree block to be allocated for other trees.
1985 *
1986 * - Log tree operations
1987 * Exiting tree blocks get allocated to log tree, bumps its
1988 * generation, then get cleaned in tree re-balance.
1989 * Such tree block will not be written back, since it's clean,
1990 * thus no WRITTEN flag set.
1991 * And after log writes back, this tree block is not traced by
1992 * any dirty extent_io_tree.
1993 *
1994 * - Offending tree block gets re-dirtied from its original owner
1995 * Since it has bumped generation, no WRITTEN flag, it can be
1996 * reused without COWing. This tree block will not be traced
1997 * by btrfs_transaction::dirty_pages.
1998 *
1999 * Now such dirty tree block will not be cleaned by any dirty
2000 * extent io tree. Thus we don't want to submit such wild eb
2001 * if the fs already has error.
2002 *
2003 * We can get ret > 0 from submit_extent_page() indicating how many ebs
2004 * were submitted. Reset it to 0 to avoid false alerts for the caller.
2005 */
2006 if (ret > 0)
2007 ret = 0;
2008 if (!ret && BTRFS_FS_ERROR(fs_info))
2009 ret = -EROFS;
2010
2011 if (ctx.zoned_bg)
2012 btrfs_put_block_group(ctx.zoned_bg);
2013 btrfs_zoned_meta_io_unlock(fs_info);
2014 return ret;
2015}
2016
2017/*
2018 * Walk the list of dirty pages of the given address space and write all of them.
2019 *
2020 * @mapping: address space structure to write
2021 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2022 * @bio_ctrl: holds context for the write, namely the bio
2023 *
2024 * If a page is already under I/O, write_cache_pages() skips it, even
2025 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2026 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2027 * and msync() need to guarantee that all the data which was dirty at the time
2028 * the call was made get new I/O started against them. If wbc->sync_mode is
2029 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2030 * existing IO to complete.
2031 */
2032static int extent_write_cache_pages(struct address_space *mapping,
2033 struct btrfs_bio_ctrl *bio_ctrl)
2034{
2035 struct writeback_control *wbc = bio_ctrl->wbc;
2036 struct inode *inode = mapping->host;
2037 int ret = 0;
2038 int done = 0;
2039 int nr_to_write_done = 0;
2040 struct folio_batch fbatch;
2041 unsigned int nr_folios;
2042 pgoff_t index;
2043 pgoff_t end; /* Inclusive */
2044 pgoff_t done_index;
2045 int range_whole = 0;
2046 int scanned = 0;
2047 xa_mark_t tag;
2048
2049 /*
2050 * We have to hold onto the inode so that ordered extents can do their
2051 * work when the IO finishes. The alternative to this is failing to add
2052 * an ordered extent if the igrab() fails there and that is a huge pain
2053 * to deal with, so instead just hold onto the inode throughout the
2054 * writepages operation. If it fails here we are freeing up the inode
2055 * anyway and we'd rather not waste our time writing out stuff that is
2056 * going to be truncated anyway.
2057 */
2058 if (!igrab(inode))
2059 return 0;
2060
2061 folio_batch_init(&fbatch);
2062 if (wbc->range_cyclic) {
2063 index = mapping->writeback_index; /* Start from prev offset */
2064 end = -1;
2065 /*
2066 * Start from the beginning does not need to cycle over the
2067 * range, mark it as scanned.
2068 */
2069 scanned = (index == 0);
2070 } else {
2071 index = wbc->range_start >> PAGE_SHIFT;
2072 end = wbc->range_end >> PAGE_SHIFT;
2073 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2074 range_whole = 1;
2075 scanned = 1;
2076 }
2077
2078 /*
2079 * We do the tagged writepage as long as the snapshot flush bit is set
2080 * and we are the first one who do the filemap_flush() on this inode.
2081 *
2082 * The nr_to_write == LONG_MAX is needed to make sure other flushers do
2083 * not race in and drop the bit.
2084 */
2085 if (range_whole && wbc->nr_to_write == LONG_MAX &&
2086 test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
2087 &BTRFS_I(inode)->runtime_flags))
2088 wbc->tagged_writepages = 1;
2089
2090 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2091 tag = PAGECACHE_TAG_TOWRITE;
2092 else
2093 tag = PAGECACHE_TAG_DIRTY;
2094retry:
2095 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2096 tag_pages_for_writeback(mapping, index, end);
2097 done_index = index;
2098 while (!done && !nr_to_write_done && (index <= end) &&
2099 (nr_folios = filemap_get_folios_tag(mapping, &index,
2100 end, tag, &fbatch))) {
2101 unsigned i;
2102
2103 for (i = 0; i < nr_folios; i++) {
2104 struct folio *folio = fbatch.folios[i];
2105
2106 done_index = folio_next_index(folio);
2107 /*
2108 * At this point we hold neither the i_pages lock nor
2109 * the page lock: the page may be truncated or
2110 * invalidated (changing page->mapping to NULL),
2111 * or even swizzled back from swapper_space to
2112 * tmpfs file mapping
2113 */
2114 if (!folio_trylock(folio)) {
2115 submit_write_bio(bio_ctrl, 0);
2116 folio_lock(folio);
2117 }
2118
2119 if (unlikely(folio->mapping != mapping)) {
2120 folio_unlock(folio);
2121 continue;
2122 }
2123
2124 if (!folio_test_dirty(folio)) {
2125 /* Someone wrote it for us. */
2126 folio_unlock(folio);
2127 continue;
2128 }
2129
2130 if (wbc->sync_mode != WB_SYNC_NONE) {
2131 if (folio_test_writeback(folio))
2132 submit_write_bio(bio_ctrl, 0);
2133 folio_wait_writeback(folio);
2134 }
2135
2136 if (folio_test_writeback(folio) ||
2137 !folio_clear_dirty_for_io(folio)) {
2138 folio_unlock(folio);
2139 continue;
2140 }
2141
2142 ret = __extent_writepage(&folio->page, bio_ctrl);
2143 if (ret < 0) {
2144 done = 1;
2145 break;
2146 }
2147
2148 /*
2149 * The filesystem may choose to bump up nr_to_write.
2150 * We have to make sure to honor the new nr_to_write
2151 * at any time.
2152 */
2153 nr_to_write_done = (wbc->sync_mode == WB_SYNC_NONE &&
2154 wbc->nr_to_write <= 0);
2155 }
2156 folio_batch_release(&fbatch);
2157 cond_resched();
2158 }
2159 if (!scanned && !done) {
2160 /*
2161 * We hit the last page and there is more work to be done: wrap
2162 * back to the start of the file
2163 */
2164 scanned = 1;
2165 index = 0;
2166
2167 /*
2168 * If we're looping we could run into a page that is locked by a
2169 * writer and that writer could be waiting on writeback for a
2170 * page in our current bio, and thus deadlock, so flush the
2171 * write bio here.
2172 */
2173 submit_write_bio(bio_ctrl, 0);
2174 goto retry;
2175 }
2176
2177 if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
2178 mapping->writeback_index = done_index;
2179
2180 btrfs_add_delayed_iput(BTRFS_I(inode));
2181 return ret;
2182}
2183
2184/*
2185 * Submit the pages in the range to bio for call sites which delalloc range has
2186 * already been ran (aka, ordered extent inserted) and all pages are still
2187 * locked.
2188 */
2189void extent_write_locked_range(struct inode *inode, struct page *locked_page,
2190 u64 start, u64 end, struct writeback_control *wbc,
2191 bool pages_dirty)
2192{
2193 bool found_error = false;
2194 int ret = 0;
2195 struct address_space *mapping = inode->i_mapping;
2196 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
2197 const u32 sectorsize = fs_info->sectorsize;
2198 loff_t i_size = i_size_read(inode);
2199 u64 cur = start;
2200 struct btrfs_bio_ctrl bio_ctrl = {
2201 .wbc = wbc,
2202 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2203 };
2204
2205 if (wbc->no_cgroup_owner)
2206 bio_ctrl.opf |= REQ_BTRFS_CGROUP_PUNT;
2207
2208 ASSERT(IS_ALIGNED(start, sectorsize) && IS_ALIGNED(end + 1, sectorsize));
2209
2210 while (cur <= end) {
2211 u64 cur_end = min(round_down(cur, PAGE_SIZE) + PAGE_SIZE - 1, end);
2212 u32 cur_len = cur_end + 1 - cur;
2213 struct page *page;
2214 int nr = 0;
2215
2216 page = find_get_page(mapping, cur >> PAGE_SHIFT);
2217 ASSERT(PageLocked(page));
2218 if (pages_dirty && page != locked_page) {
2219 ASSERT(PageDirty(page));
2220 clear_page_dirty_for_io(page);
2221 }
2222
2223 ret = __extent_writepage_io(BTRFS_I(inode), page, &bio_ctrl,
2224 i_size, &nr);
2225 if (ret == 1)
2226 goto next_page;
2227
2228 /* Make sure the mapping tag for page dirty gets cleared. */
2229 if (nr == 0) {
2230 set_page_writeback(page);
2231 end_page_writeback(page);
2232 }
2233 if (ret) {
2234 btrfs_mark_ordered_io_finished(BTRFS_I(inode), page,
2235 cur, cur_len, !ret);
2236 mapping_set_error(page->mapping, ret);
2237 }
2238 btrfs_folio_unlock_writer(fs_info, page_folio(page), cur, cur_len);
2239 if (ret < 0)
2240 found_error = true;
2241next_page:
2242 put_page(page);
2243 cur = cur_end + 1;
2244 }
2245
2246 submit_write_bio(&bio_ctrl, found_error ? ret : 0);
2247}
2248
2249int extent_writepages(struct address_space *mapping,
2250 struct writeback_control *wbc)
2251{
2252 struct inode *inode = mapping->host;
2253 int ret = 0;
2254 struct btrfs_bio_ctrl bio_ctrl = {
2255 .wbc = wbc,
2256 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2257 };
2258
2259 /*
2260 * Allow only a single thread to do the reloc work in zoned mode to
2261 * protect the write pointer updates.
2262 */
2263 btrfs_zoned_data_reloc_lock(BTRFS_I(inode));
2264 ret = extent_write_cache_pages(mapping, &bio_ctrl);
2265 submit_write_bio(&bio_ctrl, ret);
2266 btrfs_zoned_data_reloc_unlock(BTRFS_I(inode));
2267 return ret;
2268}
2269
2270void extent_readahead(struct readahead_control *rac)
2271{
2272 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ | REQ_RAHEAD };
2273 struct page *pagepool[16];
2274 struct extent_map *em_cached = NULL;
2275 u64 prev_em_start = (u64)-1;
2276 int nr;
2277
2278 while ((nr = readahead_page_batch(rac, pagepool))) {
2279 u64 contig_start = readahead_pos(rac);
2280 u64 contig_end = contig_start + readahead_batch_length(rac) - 1;
2281
2282 contiguous_readpages(pagepool, nr, contig_start, contig_end,
2283 &em_cached, &bio_ctrl, &prev_em_start);
2284 }
2285
2286 if (em_cached)
2287 free_extent_map(em_cached);
2288 submit_one_bio(&bio_ctrl);
2289}
2290
2291/*
2292 * basic invalidate_folio code, this waits on any locked or writeback
2293 * ranges corresponding to the folio, and then deletes any extent state
2294 * records from the tree
2295 */
2296int extent_invalidate_folio(struct extent_io_tree *tree,
2297 struct folio *folio, size_t offset)
2298{
2299 struct extent_state *cached_state = NULL;
2300 u64 start = folio_pos(folio);
2301 u64 end = start + folio_size(folio) - 1;
2302 size_t blocksize = folio_to_fs_info(folio)->sectorsize;
2303
2304 /* This function is only called for the btree inode */
2305 ASSERT(tree->owner == IO_TREE_BTREE_INODE_IO);
2306
2307 start += ALIGN(offset, blocksize);
2308 if (start > end)
2309 return 0;
2310
2311 lock_extent(tree, start, end, &cached_state);
2312 folio_wait_writeback(folio);
2313
2314 /*
2315 * Currently for btree io tree, only EXTENT_LOCKED is utilized,
2316 * so here we only need to unlock the extent range to free any
2317 * existing extent state.
2318 */
2319 unlock_extent(tree, start, end, &cached_state);
2320 return 0;
2321}
2322
2323/*
2324 * a helper for release_folio, this tests for areas of the page that
2325 * are locked or under IO and drops the related state bits if it is safe
2326 * to drop the page.
2327 */
2328static int try_release_extent_state(struct extent_io_tree *tree,
2329 struct page *page, gfp_t mask)
2330{
2331 u64 start = page_offset(page);
2332 u64 end = start + PAGE_SIZE - 1;
2333 int ret = 1;
2334
2335 if (test_range_bit_exists(tree, start, end, EXTENT_LOCKED)) {
2336 ret = 0;
2337 } else {
2338 u32 clear_bits = ~(EXTENT_LOCKED | EXTENT_NODATASUM |
2339 EXTENT_DELALLOC_NEW | EXTENT_CTLBITS |
2340 EXTENT_QGROUP_RESERVED);
2341
2342 /*
2343 * At this point we can safely clear everything except the
2344 * locked bit, the nodatasum bit and the delalloc new bit.
2345 * The delalloc new bit will be cleared by ordered extent
2346 * completion.
2347 */
2348 ret = __clear_extent_bit(tree, start, end, clear_bits, NULL, NULL);
2349
2350 /* if clear_extent_bit failed for enomem reasons,
2351 * we can't allow the release to continue.
2352 */
2353 if (ret < 0)
2354 ret = 0;
2355 else
2356 ret = 1;
2357 }
2358 return ret;
2359}
2360
2361/*
2362 * a helper for release_folio. As long as there are no locked extents
2363 * in the range corresponding to the page, both state records and extent
2364 * map records are removed
2365 */
2366int try_release_extent_mapping(struct page *page, gfp_t mask)
2367{
2368 struct extent_map *em;
2369 u64 start = page_offset(page);
2370 u64 end = start + PAGE_SIZE - 1;
2371 struct btrfs_inode *btrfs_inode = page_to_inode(page);
2372 struct extent_io_tree *tree = &btrfs_inode->io_tree;
2373 struct extent_map_tree *map = &btrfs_inode->extent_tree;
2374
2375 if (gfpflags_allow_blocking(mask) &&
2376 page->mapping->host->i_size > SZ_16M) {
2377 u64 len;
2378 while (start <= end) {
2379 struct btrfs_fs_info *fs_info;
2380 u64 cur_gen;
2381
2382 len = end - start + 1;
2383 write_lock(&map->lock);
2384 em = lookup_extent_mapping(map, start, len);
2385 if (!em) {
2386 write_unlock(&map->lock);
2387 break;
2388 }
2389 if ((em->flags & EXTENT_FLAG_PINNED) ||
2390 em->start != start) {
2391 write_unlock(&map->lock);
2392 free_extent_map(em);
2393 break;
2394 }
2395 if (test_range_bit_exists(tree, em->start,
2396 extent_map_end(em) - 1,
2397 EXTENT_LOCKED))
2398 goto next;
2399 /*
2400 * If it's not in the list of modified extents, used
2401 * by a fast fsync, we can remove it. If it's being
2402 * logged we can safely remove it since fsync took an
2403 * extra reference on the em.
2404 */
2405 if (list_empty(&em->list) ||
2406 (em->flags & EXTENT_FLAG_LOGGING))
2407 goto remove_em;
2408 /*
2409 * If it's in the list of modified extents, remove it
2410 * only if its generation is older then the current one,
2411 * in which case we don't need it for a fast fsync.
2412 * Otherwise don't remove it, we could be racing with an
2413 * ongoing fast fsync that could miss the new extent.
2414 */
2415 fs_info = btrfs_inode->root->fs_info;
2416 spin_lock(&fs_info->trans_lock);
2417 cur_gen = fs_info->generation;
2418 spin_unlock(&fs_info->trans_lock);
2419 if (em->generation >= cur_gen)
2420 goto next;
2421remove_em:
2422 /*
2423 * We only remove extent maps that are not in the list of
2424 * modified extents or that are in the list but with a
2425 * generation lower then the current generation, so there
2426 * is no need to set the full fsync flag on the inode (it
2427 * hurts the fsync performance for workloads with a data
2428 * size that exceeds or is close to the system's memory).
2429 */
2430 remove_extent_mapping(map, em);
2431 /* once for the rb tree */
2432 free_extent_map(em);
2433next:
2434 start = extent_map_end(em);
2435 write_unlock(&map->lock);
2436
2437 /* once for us */
2438 free_extent_map(em);
2439
2440 cond_resched(); /* Allow large-extent preemption. */
2441 }
2442 }
2443 return try_release_extent_state(tree, page, mask);
2444}
2445
2446struct btrfs_fiemap_entry {
2447 u64 offset;
2448 u64 phys;
2449 u64 len;
2450 u32 flags;
2451};
2452
2453/*
2454 * Indicate the caller of emit_fiemap_extent() that it needs to unlock the file
2455 * range from the inode's io tree, unlock the subvolume tree search path, flush
2456 * the fiemap cache and relock the file range and research the subvolume tree.
2457 * The value here is something negative that can't be confused with a valid
2458 * errno value and different from 1 because that's also a return value from
2459 * fiemap_fill_next_extent() and also it's often used to mean some btree search
2460 * did not find a key, so make it some distinct negative value.
2461 */
2462#define BTRFS_FIEMAP_FLUSH_CACHE (-(MAX_ERRNO + 1))
2463
2464/*
2465 * Used to:
2466 *
2467 * - Cache the next entry to be emitted to the fiemap buffer, so that we can
2468 * merge extents that are contiguous and can be grouped as a single one;
2469 *
2470 * - Store extents ready to be written to the fiemap buffer in an intermediary
2471 * buffer. This intermediary buffer is to ensure that in case the fiemap
2472 * buffer is memory mapped to the fiemap target file, we don't deadlock
2473 * during btrfs_page_mkwrite(). This is because during fiemap we are locking
2474 * an extent range in order to prevent races with delalloc flushing and
2475 * ordered extent completion, which is needed in order to reliably detect
2476 * delalloc in holes and prealloc extents. And this can lead to a deadlock
2477 * if the fiemap buffer is memory mapped to the file we are running fiemap
2478 * against (a silly, useless in practice scenario, but possible) because
2479 * btrfs_page_mkwrite() will try to lock the same extent range.
2480 */
2481struct fiemap_cache {
2482 /* An array of ready fiemap entries. */
2483 struct btrfs_fiemap_entry *entries;
2484 /* Number of entries in the entries array. */
2485 int entries_size;
2486 /* Index of the next entry in the entries array to write to. */
2487 int entries_pos;
2488 /*
2489 * Once the entries array is full, this indicates what's the offset for
2490 * the next file extent item we must search for in the inode's subvolume
2491 * tree after unlocking the extent range in the inode's io tree and
2492 * releasing the search path.
2493 */
2494 u64 next_search_offset;
2495 /*
2496 * This matches struct fiemap_extent_info::fi_mapped_extents, we use it
2497 * to count ourselves emitted extents and stop instead of relying on
2498 * fiemap_fill_next_extent() because we buffer ready fiemap entries at
2499 * the @entries array, and we want to stop as soon as we hit the max
2500 * amount of extents to map, not just to save time but also to make the
2501 * logic at extent_fiemap() simpler.
2502 */
2503 unsigned int extents_mapped;
2504 /* Fields for the cached extent (unsubmitted, not ready, extent). */
2505 u64 offset;
2506 u64 phys;
2507 u64 len;
2508 u32 flags;
2509 bool cached;
2510};
2511
2512static int flush_fiemap_cache(struct fiemap_extent_info *fieinfo,
2513 struct fiemap_cache *cache)
2514{
2515 for (int i = 0; i < cache->entries_pos; i++) {
2516 struct btrfs_fiemap_entry *entry = &cache->entries[i];
2517 int ret;
2518
2519 ret = fiemap_fill_next_extent(fieinfo, entry->offset,
2520 entry->phys, entry->len,
2521 entry->flags);
2522 /*
2523 * Ignore 1 (reached max entries) because we keep track of that
2524 * ourselves in emit_fiemap_extent().
2525 */
2526 if (ret < 0)
2527 return ret;
2528 }
2529 cache->entries_pos = 0;
2530
2531 return 0;
2532}
2533
2534/*
2535 * Helper to submit fiemap extent.
2536 *
2537 * Will try to merge current fiemap extent specified by @offset, @phys,
2538 * @len and @flags with cached one.
2539 * And only when we fails to merge, cached one will be submitted as
2540 * fiemap extent.
2541 *
2542 * Return value is the same as fiemap_fill_next_extent().
2543 */
2544static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
2545 struct fiemap_cache *cache,
2546 u64 offset, u64 phys, u64 len, u32 flags)
2547{
2548 struct btrfs_fiemap_entry *entry;
2549 u64 cache_end;
2550
2551 /* Set at the end of extent_fiemap(). */
2552 ASSERT((flags & FIEMAP_EXTENT_LAST) == 0);
2553
2554 if (!cache->cached)
2555 goto assign;
2556
2557 /*
2558 * When iterating the extents of the inode, at extent_fiemap(), we may
2559 * find an extent that starts at an offset behind the end offset of the
2560 * previous extent we processed. This happens if fiemap is called
2561 * without FIEMAP_FLAG_SYNC and there are ordered extents completing
2562 * after we had to unlock the file range, release the search path, emit
2563 * the fiemap extents stored in the buffer (cache->entries array) and
2564 * the lock the remainder of the range and re-search the btree.
2565 *
2566 * For example we are in leaf X processing its last item, which is the
2567 * file extent item for file range [512K, 1M[, and after
2568 * btrfs_next_leaf() releases the path, there's an ordered extent that
2569 * completes for the file range [768K, 2M[, and that results in trimming
2570 * the file extent item so that it now corresponds to the file range
2571 * [512K, 768K[ and a new file extent item is inserted for the file
2572 * range [768K, 2M[, which may end up as the last item of leaf X or as
2573 * the first item of the next leaf - in either case btrfs_next_leaf()
2574 * will leave us with a path pointing to the new extent item, for the
2575 * file range [768K, 2M[, since that's the first key that follows the
2576 * last one we processed. So in order not to report overlapping extents
2577 * to user space, we trim the length of the previously cached extent and
2578 * emit it.
2579 *
2580 * Upon calling btrfs_next_leaf() we may also find an extent with an
2581 * offset smaller than or equals to cache->offset, and this happens
2582 * when we had a hole or prealloc extent with several delalloc ranges in
2583 * it, but after btrfs_next_leaf() released the path, delalloc was
2584 * flushed and the resulting ordered extents were completed, so we can
2585 * now have found a file extent item for an offset that is smaller than
2586 * or equals to what we have in cache->offset. We deal with this as
2587 * described below.
2588 */
2589 cache_end = cache->offset + cache->len;
2590 if (cache_end > offset) {
2591 if (offset == cache->offset) {
2592 /*
2593 * We cached a dealloc range (found in the io tree) for
2594 * a hole or prealloc extent and we have now found a
2595 * file extent item for the same offset. What we have
2596 * now is more recent and up to date, so discard what
2597 * we had in the cache and use what we have just found.
2598 */
2599 goto assign;
2600 } else if (offset > cache->offset) {
2601 /*
2602 * The extent range we previously found ends after the
2603 * offset of the file extent item we found and that
2604 * offset falls somewhere in the middle of that previous
2605 * extent range. So adjust the range we previously found
2606 * to end at the offset of the file extent item we have
2607 * just found, since this extent is more up to date.
2608 * Emit that adjusted range and cache the file extent
2609 * item we have just found. This corresponds to the case
2610 * where a previously found file extent item was split
2611 * due to an ordered extent completing.
2612 */
2613 cache->len = offset - cache->offset;
2614 goto emit;
2615 } else {
2616 const u64 range_end = offset + len;
2617
2618 /*
2619 * The offset of the file extent item we have just found
2620 * is behind the cached offset. This means we were
2621 * processing a hole or prealloc extent for which we
2622 * have found delalloc ranges (in the io tree), so what
2623 * we have in the cache is the last delalloc range we
2624 * found while the file extent item we found can be
2625 * either for a whole delalloc range we previously
2626 * emmitted or only a part of that range.
2627 *
2628 * We have two cases here:
2629 *
2630 * 1) The file extent item's range ends at or behind the
2631 * cached extent's end. In this case just ignore the
2632 * current file extent item because we don't want to
2633 * overlap with previous ranges that may have been
2634 * emmitted already;
2635 *
2636 * 2) The file extent item starts behind the currently
2637 * cached extent but its end offset goes beyond the
2638 * end offset of the cached extent. We don't want to
2639 * overlap with a previous range that may have been
2640 * emmitted already, so we emit the currently cached
2641 * extent and then partially store the current file
2642 * extent item's range in the cache, for the subrange
2643 * going the cached extent's end to the end of the
2644 * file extent item.
2645 */
2646 if (range_end <= cache_end)
2647 return 0;
2648
2649 if (!(flags & (FIEMAP_EXTENT_ENCODED | FIEMAP_EXTENT_DELALLOC)))
2650 phys += cache_end - offset;
2651
2652 offset = cache_end;
2653 len = range_end - cache_end;
2654 goto emit;
2655 }
2656 }
2657
2658 /*
2659 * Only merges fiemap extents if
2660 * 1) Their logical addresses are continuous
2661 *
2662 * 2) Their physical addresses are continuous
2663 * So truly compressed (physical size smaller than logical size)
2664 * extents won't get merged with each other
2665 *
2666 * 3) Share same flags
2667 */
2668 if (cache->offset + cache->len == offset &&
2669 cache->phys + cache->len == phys &&
2670 cache->flags == flags) {
2671 cache->len += len;
2672 return 0;
2673 }
2674
2675emit:
2676 /* Not mergeable, need to submit cached one */
2677
2678 if (cache->entries_pos == cache->entries_size) {
2679 /*
2680 * We will need to research for the end offset of the last
2681 * stored extent and not from the current offset, because after
2682 * unlocking the range and releasing the path, if there's a hole
2683 * between that end offset and this current offset, a new extent
2684 * may have been inserted due to a new write, so we don't want
2685 * to miss it.
2686 */
2687 entry = &cache->entries[cache->entries_size - 1];
2688 cache->next_search_offset = entry->offset + entry->len;
2689 cache->cached = false;
2690
2691 return BTRFS_FIEMAP_FLUSH_CACHE;
2692 }
2693
2694 entry = &cache->entries[cache->entries_pos];
2695 entry->offset = cache->offset;
2696 entry->phys = cache->phys;
2697 entry->len = cache->len;
2698 entry->flags = cache->flags;
2699 cache->entries_pos++;
2700 cache->extents_mapped++;
2701
2702 if (cache->extents_mapped == fieinfo->fi_extents_max) {
2703 cache->cached = false;
2704 return 1;
2705 }
2706assign:
2707 cache->cached = true;
2708 cache->offset = offset;
2709 cache->phys = phys;
2710 cache->len = len;
2711 cache->flags = flags;
2712
2713 return 0;
2714}
2715
2716/*
2717 * Emit last fiemap cache
2718 *
2719 * The last fiemap cache may still be cached in the following case:
2720 * 0 4k 8k
2721 * |<- Fiemap range ->|
2722 * |<------------ First extent ----------->|
2723 *
2724 * In this case, the first extent range will be cached but not emitted.
2725 * So we must emit it before ending extent_fiemap().
2726 */
2727static int emit_last_fiemap_cache(struct fiemap_extent_info *fieinfo,
2728 struct fiemap_cache *cache)
2729{
2730 int ret;
2731
2732 if (!cache->cached)
2733 return 0;
2734
2735 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
2736 cache->len, cache->flags);
2737 cache->cached = false;
2738 if (ret > 0)
2739 ret = 0;
2740 return ret;
2741}
2742
2743static int fiemap_next_leaf_item(struct btrfs_inode *inode, struct btrfs_path *path)
2744{
2745 struct extent_buffer *clone = path->nodes[0];
2746 struct btrfs_key key;
2747 int slot;
2748 int ret;
2749
2750 path->slots[0]++;
2751 if (path->slots[0] < btrfs_header_nritems(path->nodes[0]))
2752 return 0;
2753
2754 /*
2755 * Add a temporary extra ref to an already cloned extent buffer to
2756 * prevent btrfs_next_leaf() freeing it, we want to reuse it to avoid
2757 * the cost of allocating a new one.
2758 */
2759 ASSERT(test_bit(EXTENT_BUFFER_UNMAPPED, &clone->bflags));
2760 atomic_inc(&clone->refs);
2761
2762 ret = btrfs_next_leaf(inode->root, path);
2763 if (ret != 0)
2764 goto out;
2765
2766 /*
2767 * Don't bother with cloning if there are no more file extent items for
2768 * our inode.
2769 */
2770 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2771 if (key.objectid != btrfs_ino(inode) || key.type != BTRFS_EXTENT_DATA_KEY) {
2772 ret = 1;
2773 goto out;
2774 }
2775
2776 /*
2777 * Important to preserve the start field, for the optimizations when
2778 * checking if extents are shared (see extent_fiemap()).
2779 *
2780 * We must set ->start before calling copy_extent_buffer_full(). If we
2781 * are on sub-pagesize blocksize, we use ->start to determine the offset
2782 * into the folio where our eb exists, and if we update ->start after
2783 * the fact then any subsequent reads of the eb may read from a
2784 * different offset in the folio than where we originally copied into.
2785 */
2786 clone->start = path->nodes[0]->start;
2787 /* See the comment at fiemap_search_slot() about why we clone. */
2788 copy_extent_buffer_full(clone, path->nodes[0]);
2789
2790 slot = path->slots[0];
2791 btrfs_release_path(path);
2792 path->nodes[0] = clone;
2793 path->slots[0] = slot;
2794out:
2795 if (ret)
2796 free_extent_buffer(clone);
2797
2798 return ret;
2799}
2800
2801/*
2802 * Search for the first file extent item that starts at a given file offset or
2803 * the one that starts immediately before that offset.
2804 * Returns: 0 on success, < 0 on error, 1 if not found.
2805 */
2806static int fiemap_search_slot(struct btrfs_inode *inode, struct btrfs_path *path,
2807 u64 file_offset)
2808{
2809 const u64 ino = btrfs_ino(inode);
2810 struct btrfs_root *root = inode->root;
2811 struct extent_buffer *clone;
2812 struct btrfs_key key;
2813 int slot;
2814 int ret;
2815
2816 key.objectid = ino;
2817 key.type = BTRFS_EXTENT_DATA_KEY;
2818 key.offset = file_offset;
2819
2820 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2821 if (ret < 0)
2822 return ret;
2823
2824 if (ret > 0 && path->slots[0] > 0) {
2825 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
2826 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
2827 path->slots[0]--;
2828 }
2829
2830 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2831 ret = btrfs_next_leaf(root, path);
2832 if (ret != 0)
2833 return ret;
2834
2835 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2836 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
2837 return 1;
2838 }
2839
2840 /*
2841 * We clone the leaf and use it during fiemap. This is because while
2842 * using the leaf we do expensive things like checking if an extent is
2843 * shared, which can take a long time. In order to prevent blocking
2844 * other tasks for too long, we use a clone of the leaf. We have locked
2845 * the file range in the inode's io tree, so we know none of our file
2846 * extent items can change. This way we avoid blocking other tasks that
2847 * want to insert items for other inodes in the same leaf or b+tree
2848 * rebalance operations (triggered for example when someone is trying
2849 * to push items into this leaf when trying to insert an item in a
2850 * neighbour leaf).
2851 * We also need the private clone because holding a read lock on an
2852 * extent buffer of the subvolume's b+tree will make lockdep unhappy
2853 * when we check if extents are shared, as backref walking may need to
2854 * lock the same leaf we are processing.
2855 */
2856 clone = btrfs_clone_extent_buffer(path->nodes[0]);
2857 if (!clone)
2858 return -ENOMEM;
2859
2860 slot = path->slots[0];
2861 btrfs_release_path(path);
2862 path->nodes[0] = clone;
2863 path->slots[0] = slot;
2864
2865 return 0;
2866}
2867
2868/*
2869 * Process a range which is a hole or a prealloc extent in the inode's subvolume
2870 * btree. If @disk_bytenr is 0, we are dealing with a hole, otherwise a prealloc
2871 * extent. The end offset (@end) is inclusive.
2872 */
2873static int fiemap_process_hole(struct btrfs_inode *inode,
2874 struct fiemap_extent_info *fieinfo,
2875 struct fiemap_cache *cache,
2876 struct extent_state **delalloc_cached_state,
2877 struct btrfs_backref_share_check_ctx *backref_ctx,
2878 u64 disk_bytenr, u64 extent_offset,
2879 u64 extent_gen,
2880 u64 start, u64 end)
2881{
2882 const u64 i_size = i_size_read(&inode->vfs_inode);
2883 u64 cur_offset = start;
2884 u64 last_delalloc_end = 0;
2885 u32 prealloc_flags = FIEMAP_EXTENT_UNWRITTEN;
2886 bool checked_extent_shared = false;
2887 int ret;
2888
2889 /*
2890 * There can be no delalloc past i_size, so don't waste time looking for
2891 * it beyond i_size.
2892 */
2893 while (cur_offset < end && cur_offset < i_size) {
2894 u64 delalloc_start;
2895 u64 delalloc_end;
2896 u64 prealloc_start;
2897 u64 prealloc_len = 0;
2898 bool delalloc;
2899
2900 delalloc = btrfs_find_delalloc_in_range(inode, cur_offset, end,
2901 delalloc_cached_state,
2902 &delalloc_start,
2903 &delalloc_end);
2904 if (!delalloc)
2905 break;
2906
2907 /*
2908 * If this is a prealloc extent we have to report every section
2909 * of it that has no delalloc.
2910 */
2911 if (disk_bytenr != 0) {
2912 if (last_delalloc_end == 0) {
2913 prealloc_start = start;
2914 prealloc_len = delalloc_start - start;
2915 } else {
2916 prealloc_start = last_delalloc_end + 1;
2917 prealloc_len = delalloc_start - prealloc_start;
2918 }
2919 }
2920
2921 if (prealloc_len > 0) {
2922 if (!checked_extent_shared && fieinfo->fi_extents_max) {
2923 ret = btrfs_is_data_extent_shared(inode,
2924 disk_bytenr,
2925 extent_gen,
2926 backref_ctx);
2927 if (ret < 0)
2928 return ret;
2929 else if (ret > 0)
2930 prealloc_flags |= FIEMAP_EXTENT_SHARED;
2931
2932 checked_extent_shared = true;
2933 }
2934 ret = emit_fiemap_extent(fieinfo, cache, prealloc_start,
2935 disk_bytenr + extent_offset,
2936 prealloc_len, prealloc_flags);
2937 if (ret)
2938 return ret;
2939 extent_offset += prealloc_len;
2940 }
2941
2942 ret = emit_fiemap_extent(fieinfo, cache, delalloc_start, 0,
2943 delalloc_end + 1 - delalloc_start,
2944 FIEMAP_EXTENT_DELALLOC |
2945 FIEMAP_EXTENT_UNKNOWN);
2946 if (ret)
2947 return ret;
2948
2949 last_delalloc_end = delalloc_end;
2950 cur_offset = delalloc_end + 1;
2951 extent_offset += cur_offset - delalloc_start;
2952 cond_resched();
2953 }
2954
2955 /*
2956 * Either we found no delalloc for the whole prealloc extent or we have
2957 * a prealloc extent that spans i_size or starts at or after i_size.
2958 */
2959 if (disk_bytenr != 0 && last_delalloc_end < end) {
2960 u64 prealloc_start;
2961 u64 prealloc_len;
2962
2963 if (last_delalloc_end == 0) {
2964 prealloc_start = start;
2965 prealloc_len = end + 1 - start;
2966 } else {
2967 prealloc_start = last_delalloc_end + 1;
2968 prealloc_len = end + 1 - prealloc_start;
2969 }
2970
2971 if (!checked_extent_shared && fieinfo->fi_extents_max) {
2972 ret = btrfs_is_data_extent_shared(inode,
2973 disk_bytenr,
2974 extent_gen,
2975 backref_ctx);
2976 if (ret < 0)
2977 return ret;
2978 else if (ret > 0)
2979 prealloc_flags |= FIEMAP_EXTENT_SHARED;
2980 }
2981 ret = emit_fiemap_extent(fieinfo, cache, prealloc_start,
2982 disk_bytenr + extent_offset,
2983 prealloc_len, prealloc_flags);
2984 if (ret)
2985 return ret;
2986 }
2987
2988 return 0;
2989}
2990
2991static int fiemap_find_last_extent_offset(struct btrfs_inode *inode,
2992 struct btrfs_path *path,
2993 u64 *last_extent_end_ret)
2994{
2995 const u64 ino = btrfs_ino(inode);
2996 struct btrfs_root *root = inode->root;
2997 struct extent_buffer *leaf;
2998 struct btrfs_file_extent_item *ei;
2999 struct btrfs_key key;
3000 u64 disk_bytenr;
3001 int ret;
3002
3003 /*
3004 * Lookup the last file extent. We're not using i_size here because
3005 * there might be preallocation past i_size.
3006 */
3007 ret = btrfs_lookup_file_extent(NULL, root, path, ino, (u64)-1, 0);
3008 /* There can't be a file extent item at offset (u64)-1 */
3009 ASSERT(ret != 0);
3010 if (ret < 0)
3011 return ret;
3012
3013 /*
3014 * For a non-existing key, btrfs_search_slot() always leaves us at a
3015 * slot > 0, except if the btree is empty, which is impossible because
3016 * at least it has the inode item for this inode and all the items for
3017 * the root inode 256.
3018 */
3019 ASSERT(path->slots[0] > 0);
3020 path->slots[0]--;
3021 leaf = path->nodes[0];
3022 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3023 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY) {
3024 /* No file extent items in the subvolume tree. */
3025 *last_extent_end_ret = 0;
3026 return 0;
3027 }
3028
3029 /*
3030 * For an inline extent, the disk_bytenr is where inline data starts at,
3031 * so first check if we have an inline extent item before checking if we
3032 * have an implicit hole (disk_bytenr == 0).
3033 */
3034 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
3035 if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_INLINE) {
3036 *last_extent_end_ret = btrfs_file_extent_end(path);
3037 return 0;
3038 }
3039
3040 /*
3041 * Find the last file extent item that is not a hole (when NO_HOLES is
3042 * not enabled). This should take at most 2 iterations in the worst
3043 * case: we have one hole file extent item at slot 0 of a leaf and
3044 * another hole file extent item as the last item in the previous leaf.
3045 * This is because we merge file extent items that represent holes.
3046 */
3047 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
3048 while (disk_bytenr == 0) {
3049 ret = btrfs_previous_item(root, path, ino, BTRFS_EXTENT_DATA_KEY);
3050 if (ret < 0) {
3051 return ret;
3052 } else if (ret > 0) {
3053 /* No file extent items that are not holes. */
3054 *last_extent_end_ret = 0;
3055 return 0;
3056 }
3057 leaf = path->nodes[0];
3058 ei = btrfs_item_ptr(leaf, path->slots[0],
3059 struct btrfs_file_extent_item);
3060 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
3061 }
3062
3063 *last_extent_end_ret = btrfs_file_extent_end(path);
3064 return 0;
3065}
3066
3067int extent_fiemap(struct btrfs_inode *inode, struct fiemap_extent_info *fieinfo,
3068 u64 start, u64 len)
3069{
3070 const u64 ino = btrfs_ino(inode);
3071 struct extent_state *cached_state = NULL;
3072 struct extent_state *delalloc_cached_state = NULL;
3073 struct btrfs_path *path;
3074 struct fiemap_cache cache = { 0 };
3075 struct btrfs_backref_share_check_ctx *backref_ctx;
3076 u64 last_extent_end;
3077 u64 prev_extent_end;
3078 u64 range_start;
3079 u64 range_end;
3080 const u64 sectorsize = inode->root->fs_info->sectorsize;
3081 bool stopped = false;
3082 int ret;
3083
3084 cache.entries_size = PAGE_SIZE / sizeof(struct btrfs_fiemap_entry);
3085 cache.entries = kmalloc_array(cache.entries_size,
3086 sizeof(struct btrfs_fiemap_entry),
3087 GFP_KERNEL);
3088 backref_ctx = btrfs_alloc_backref_share_check_ctx();
3089 path = btrfs_alloc_path();
3090 if (!cache.entries || !backref_ctx || !path) {
3091 ret = -ENOMEM;
3092 goto out;
3093 }
3094
3095restart:
3096 range_start = round_down(start, sectorsize);
3097 range_end = round_up(start + len, sectorsize);
3098 prev_extent_end = range_start;
3099
3100 lock_extent(&inode->io_tree, range_start, range_end, &cached_state);
3101
3102 ret = fiemap_find_last_extent_offset(inode, path, &last_extent_end);
3103 if (ret < 0)
3104 goto out_unlock;
3105 btrfs_release_path(path);
3106
3107 path->reada = READA_FORWARD;
3108 ret = fiemap_search_slot(inode, path, range_start);
3109 if (ret < 0) {
3110 goto out_unlock;
3111 } else if (ret > 0) {
3112 /*
3113 * No file extent item found, but we may have delalloc between
3114 * the current offset and i_size. So check for that.
3115 */
3116 ret = 0;
3117 goto check_eof_delalloc;
3118 }
3119
3120 while (prev_extent_end < range_end) {
3121 struct extent_buffer *leaf = path->nodes[0];
3122 struct btrfs_file_extent_item *ei;
3123 struct btrfs_key key;
3124 u64 extent_end;
3125 u64 extent_len;
3126 u64 extent_offset = 0;
3127 u64 extent_gen;
3128 u64 disk_bytenr = 0;
3129 u64 flags = 0;
3130 int extent_type;
3131 u8 compression;
3132
3133 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3134 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3135 break;
3136
3137 extent_end = btrfs_file_extent_end(path);
3138
3139 /*
3140 * The first iteration can leave us at an extent item that ends
3141 * before our range's start. Move to the next item.
3142 */
3143 if (extent_end <= range_start)
3144 goto next_item;
3145
3146 backref_ctx->curr_leaf_bytenr = leaf->start;
3147
3148 /* We have in implicit hole (NO_HOLES feature enabled). */
3149 if (prev_extent_end < key.offset) {
3150 const u64 hole_end = min(key.offset, range_end) - 1;
3151
3152 ret = fiemap_process_hole(inode, fieinfo, &cache,
3153 &delalloc_cached_state,
3154 backref_ctx, 0, 0, 0,
3155 prev_extent_end, hole_end);
3156 if (ret < 0) {
3157 goto out_unlock;
3158 } else if (ret > 0) {
3159 /* fiemap_fill_next_extent() told us to stop. */
3160 stopped = true;
3161 break;
3162 }
3163
3164 /* We've reached the end of the fiemap range, stop. */
3165 if (key.offset >= range_end) {
3166 stopped = true;
3167 break;
3168 }
3169 }
3170
3171 extent_len = extent_end - key.offset;
3172 ei = btrfs_item_ptr(leaf, path->slots[0],
3173 struct btrfs_file_extent_item);
3174 compression = btrfs_file_extent_compression(leaf, ei);
3175 extent_type = btrfs_file_extent_type(leaf, ei);
3176 extent_gen = btrfs_file_extent_generation(leaf, ei);
3177
3178 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
3179 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
3180 if (compression == BTRFS_COMPRESS_NONE)
3181 extent_offset = btrfs_file_extent_offset(leaf, ei);
3182 }
3183
3184 if (compression != BTRFS_COMPRESS_NONE)
3185 flags |= FIEMAP_EXTENT_ENCODED;
3186
3187 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
3188 flags |= FIEMAP_EXTENT_DATA_INLINE;
3189 flags |= FIEMAP_EXTENT_NOT_ALIGNED;
3190 ret = emit_fiemap_extent(fieinfo, &cache, key.offset, 0,
3191 extent_len, flags);
3192 } else if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
3193 ret = fiemap_process_hole(inode, fieinfo, &cache,
3194 &delalloc_cached_state,
3195 backref_ctx,
3196 disk_bytenr, extent_offset,
3197 extent_gen, key.offset,
3198 extent_end - 1);
3199 } else if (disk_bytenr == 0) {
3200 /* We have an explicit hole. */
3201 ret = fiemap_process_hole(inode, fieinfo, &cache,
3202 &delalloc_cached_state,
3203 backref_ctx, 0, 0, 0,
3204 key.offset, extent_end - 1);
3205 } else {
3206 /* We have a regular extent. */
3207 if (fieinfo->fi_extents_max) {
3208 ret = btrfs_is_data_extent_shared(inode,
3209 disk_bytenr,
3210 extent_gen,
3211 backref_ctx);
3212 if (ret < 0)
3213 goto out_unlock;
3214 else if (ret > 0)
3215 flags |= FIEMAP_EXTENT_SHARED;
3216 }
3217
3218 ret = emit_fiemap_extent(fieinfo, &cache, key.offset,
3219 disk_bytenr + extent_offset,
3220 extent_len, flags);
3221 }
3222
3223 if (ret < 0) {
3224 goto out_unlock;
3225 } else if (ret > 0) {
3226 /* emit_fiemap_extent() told us to stop. */
3227 stopped = true;
3228 break;
3229 }
3230
3231 prev_extent_end = extent_end;
3232next_item:
3233 if (fatal_signal_pending(current)) {
3234 ret = -EINTR;
3235 goto out_unlock;
3236 }
3237
3238 ret = fiemap_next_leaf_item(inode, path);
3239 if (ret < 0) {
3240 goto out_unlock;
3241 } else if (ret > 0) {
3242 /* No more file extent items for this inode. */
3243 break;
3244 }
3245 cond_resched();
3246 }
3247
3248check_eof_delalloc:
3249 if (!stopped && prev_extent_end < range_end) {
3250 ret = fiemap_process_hole(inode, fieinfo, &cache,
3251 &delalloc_cached_state, backref_ctx,
3252 0, 0, 0, prev_extent_end, range_end - 1);
3253 if (ret < 0)
3254 goto out_unlock;
3255 prev_extent_end = range_end;
3256 }
3257
3258 if (cache.cached && cache.offset + cache.len >= last_extent_end) {
3259 const u64 i_size = i_size_read(&inode->vfs_inode);
3260
3261 if (prev_extent_end < i_size) {
3262 u64 delalloc_start;
3263 u64 delalloc_end;
3264 bool delalloc;
3265
3266 delalloc = btrfs_find_delalloc_in_range(inode,
3267 prev_extent_end,
3268 i_size - 1,
3269 &delalloc_cached_state,
3270 &delalloc_start,
3271 &delalloc_end);
3272 if (!delalloc)
3273 cache.flags |= FIEMAP_EXTENT_LAST;
3274 } else {
3275 cache.flags |= FIEMAP_EXTENT_LAST;
3276 }
3277 }
3278
3279out_unlock:
3280 unlock_extent(&inode->io_tree, range_start, range_end, &cached_state);
3281
3282 if (ret == BTRFS_FIEMAP_FLUSH_CACHE) {
3283 btrfs_release_path(path);
3284 ret = flush_fiemap_cache(fieinfo, &cache);
3285 if (ret)
3286 goto out;
3287 len -= cache.next_search_offset - start;
3288 start = cache.next_search_offset;
3289 goto restart;
3290 } else if (ret < 0) {
3291 goto out;
3292 }
3293
3294 /*
3295 * Must free the path before emitting to the fiemap buffer because we
3296 * may have a non-cloned leaf and if the fiemap buffer is memory mapped
3297 * to a file, a write into it (through btrfs_page_mkwrite()) may trigger
3298 * waiting for an ordered extent that in order to complete needs to
3299 * modify that leaf, therefore leading to a deadlock.
3300 */
3301 btrfs_free_path(path);
3302 path = NULL;
3303
3304 ret = flush_fiemap_cache(fieinfo, &cache);
3305 if (ret)
3306 goto out;
3307
3308 ret = emit_last_fiemap_cache(fieinfo, &cache);
3309out:
3310 free_extent_state(delalloc_cached_state);
3311 kfree(cache.entries);
3312 btrfs_free_backref_share_ctx(backref_ctx);
3313 btrfs_free_path(path);
3314 return ret;
3315}
3316
3317static void __free_extent_buffer(struct extent_buffer *eb)
3318{
3319 kmem_cache_free(extent_buffer_cache, eb);
3320}
3321
3322static int extent_buffer_under_io(const struct extent_buffer *eb)
3323{
3324 return (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
3325 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3326}
3327
3328static bool folio_range_has_eb(struct btrfs_fs_info *fs_info, struct folio *folio)
3329{
3330 struct btrfs_subpage *subpage;
3331
3332 lockdep_assert_held(&folio->mapping->i_private_lock);
3333
3334 if (folio_test_private(folio)) {
3335 subpage = folio_get_private(folio);
3336 if (atomic_read(&subpage->eb_refs))
3337 return true;
3338 /*
3339 * Even there is no eb refs here, we may still have
3340 * end_page_read() call relying on page::private.
3341 */
3342 if (atomic_read(&subpage->readers))
3343 return true;
3344 }
3345 return false;
3346}
3347
3348static void detach_extent_buffer_folio(struct extent_buffer *eb, struct folio *folio)
3349{
3350 struct btrfs_fs_info *fs_info = eb->fs_info;
3351 const bool mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3352
3353 /*
3354 * For mapped eb, we're going to change the folio private, which should
3355 * be done under the i_private_lock.
3356 */
3357 if (mapped)
3358 spin_lock(&folio->mapping->i_private_lock);
3359
3360 if (!folio_test_private(folio)) {
3361 if (mapped)
3362 spin_unlock(&folio->mapping->i_private_lock);
3363 return;
3364 }
3365
3366 if (fs_info->nodesize >= PAGE_SIZE) {
3367 /*
3368 * We do this since we'll remove the pages after we've
3369 * removed the eb from the radix tree, so we could race
3370 * and have this page now attached to the new eb. So
3371 * only clear folio if it's still connected to
3372 * this eb.
3373 */
3374 if (folio_test_private(folio) && folio_get_private(folio) == eb) {
3375 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3376 BUG_ON(folio_test_dirty(folio));
3377 BUG_ON(folio_test_writeback(folio));
3378 /* We need to make sure we haven't be attached to a new eb. */
3379 folio_detach_private(folio);
3380 }
3381 if (mapped)
3382 spin_unlock(&folio->mapping->i_private_lock);
3383 return;
3384 }
3385
3386 /*
3387 * For subpage, we can have dummy eb with folio private attached. In
3388 * this case, we can directly detach the private as such folio is only
3389 * attached to one dummy eb, no sharing.
3390 */
3391 if (!mapped) {
3392 btrfs_detach_subpage(fs_info, folio);
3393 return;
3394 }
3395
3396 btrfs_folio_dec_eb_refs(fs_info, folio);
3397
3398 /*
3399 * We can only detach the folio private if there are no other ebs in the
3400 * page range and no unfinished IO.
3401 */
3402 if (!folio_range_has_eb(fs_info, folio))
3403 btrfs_detach_subpage(fs_info, folio);
3404
3405 spin_unlock(&folio->mapping->i_private_lock);
3406}
3407
3408/* Release all pages attached to the extent buffer */
3409static void btrfs_release_extent_buffer_pages(struct extent_buffer *eb)
3410{
3411 ASSERT(!extent_buffer_under_io(eb));
3412
3413 for (int i = 0; i < INLINE_EXTENT_BUFFER_PAGES; i++) {
3414 struct folio *folio = eb->folios[i];
3415
3416 if (!folio)
3417 continue;
3418
3419 detach_extent_buffer_folio(eb, folio);
3420
3421 /* One for when we allocated the folio. */
3422 folio_put(folio);
3423 }
3424}
3425
3426/*
3427 * Helper for releasing the extent buffer.
3428 */
3429static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3430{
3431 btrfs_release_extent_buffer_pages(eb);
3432 btrfs_leak_debug_del_eb(eb);
3433 __free_extent_buffer(eb);
3434}
3435
3436static struct extent_buffer *
3437__alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
3438 unsigned long len)
3439{
3440 struct extent_buffer *eb = NULL;
3441
3442 eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
3443 eb->start = start;
3444 eb->len = len;
3445 eb->fs_info = fs_info;
3446 init_rwsem(&eb->lock);
3447
3448 btrfs_leak_debug_add_eb(eb);
3449
3450 spin_lock_init(&eb->refs_lock);
3451 atomic_set(&eb->refs, 1);
3452
3453 ASSERT(len <= BTRFS_MAX_METADATA_BLOCKSIZE);
3454
3455 return eb;
3456}
3457
3458struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
3459{
3460 struct extent_buffer *new;
3461 int num_folios = num_extent_folios(src);
3462 int ret;
3463
3464 new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
3465 if (new == NULL)
3466 return NULL;
3467
3468 /*
3469 * Set UNMAPPED before calling btrfs_release_extent_buffer(), as
3470 * btrfs_release_extent_buffer() have different behavior for
3471 * UNMAPPED subpage extent buffer.
3472 */
3473 set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
3474
3475 ret = alloc_eb_folio_array(new, 0);
3476 if (ret) {
3477 btrfs_release_extent_buffer(new);
3478 return NULL;
3479 }
3480
3481 for (int i = 0; i < num_folios; i++) {
3482 struct folio *folio = new->folios[i];
3483 int ret;
3484
3485 ret = attach_extent_buffer_folio(new, folio, NULL);
3486 if (ret < 0) {
3487 btrfs_release_extent_buffer(new);
3488 return NULL;
3489 }
3490 WARN_ON(folio_test_dirty(folio));
3491 }
3492 copy_extent_buffer_full(new, src);
3493 set_extent_buffer_uptodate(new);
3494
3495 return new;
3496}
3497
3498struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
3499 u64 start, unsigned long len)
3500{
3501 struct extent_buffer *eb;
3502 int num_folios = 0;
3503 int ret;
3504
3505 eb = __alloc_extent_buffer(fs_info, start, len);
3506 if (!eb)
3507 return NULL;
3508
3509 ret = alloc_eb_folio_array(eb, 0);
3510 if (ret)
3511 goto err;
3512
3513 num_folios = num_extent_folios(eb);
3514 for (int i = 0; i < num_folios; i++) {
3515 ret = attach_extent_buffer_folio(eb, eb->folios[i], NULL);
3516 if (ret < 0)
3517 goto err;
3518 }
3519
3520 set_extent_buffer_uptodate(eb);
3521 btrfs_set_header_nritems(eb, 0);
3522 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3523
3524 return eb;
3525err:
3526 for (int i = 0; i < num_folios; i++) {
3527 if (eb->folios[i]) {
3528 detach_extent_buffer_folio(eb, eb->folios[i]);
3529 __folio_put(eb->folios[i]);
3530 }
3531 }
3532 __free_extent_buffer(eb);
3533 return NULL;
3534}
3535
3536struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
3537 u64 start)
3538{
3539 return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
3540}
3541
3542static void check_buffer_tree_ref(struct extent_buffer *eb)
3543{
3544 int refs;
3545 /*
3546 * The TREE_REF bit is first set when the extent_buffer is added
3547 * to the radix tree. It is also reset, if unset, when a new reference
3548 * is created by find_extent_buffer.
3549 *
3550 * It is only cleared in two cases: freeing the last non-tree
3551 * reference to the extent_buffer when its STALE bit is set or
3552 * calling release_folio when the tree reference is the only reference.
3553 *
3554 * In both cases, care is taken to ensure that the extent_buffer's
3555 * pages are not under io. However, release_folio can be concurrently
3556 * called with creating new references, which is prone to race
3557 * conditions between the calls to check_buffer_tree_ref in those
3558 * codepaths and clearing TREE_REF in try_release_extent_buffer.
3559 *
3560 * The actual lifetime of the extent_buffer in the radix tree is
3561 * adequately protected by the refcount, but the TREE_REF bit and
3562 * its corresponding reference are not. To protect against this
3563 * class of races, we call check_buffer_tree_ref from the codepaths
3564 * which trigger io. Note that once io is initiated, TREE_REF can no
3565 * longer be cleared, so that is the moment at which any such race is
3566 * best fixed.
3567 */
3568 refs = atomic_read(&eb->refs);
3569 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3570 return;
3571
3572 spin_lock(&eb->refs_lock);
3573 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3574 atomic_inc(&eb->refs);
3575 spin_unlock(&eb->refs_lock);
3576}
3577
3578static void mark_extent_buffer_accessed(struct extent_buffer *eb)
3579{
3580 int num_folios= num_extent_folios(eb);
3581
3582 check_buffer_tree_ref(eb);
3583
3584 for (int i = 0; i < num_folios; i++)
3585 folio_mark_accessed(eb->folios[i]);
3586}
3587
3588struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
3589 u64 start)
3590{
3591 struct extent_buffer *eb;
3592
3593 eb = find_extent_buffer_nolock(fs_info, start);
3594 if (!eb)
3595 return NULL;
3596 /*
3597 * Lock our eb's refs_lock to avoid races with free_extent_buffer().
3598 * When we get our eb it might be flagged with EXTENT_BUFFER_STALE and
3599 * another task running free_extent_buffer() might have seen that flag
3600 * set, eb->refs == 2, that the buffer isn't under IO (dirty and
3601 * writeback flags not set) and it's still in the tree (flag
3602 * EXTENT_BUFFER_TREE_REF set), therefore being in the process of
3603 * decrementing the extent buffer's reference count twice. So here we
3604 * could race and increment the eb's reference count, clear its stale
3605 * flag, mark it as dirty and drop our reference before the other task
3606 * finishes executing free_extent_buffer, which would later result in
3607 * an attempt to free an extent buffer that is dirty.
3608 */
3609 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
3610 spin_lock(&eb->refs_lock);
3611 spin_unlock(&eb->refs_lock);
3612 }
3613 mark_extent_buffer_accessed(eb);
3614 return eb;
3615}
3616
3617#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3618struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
3619 u64 start)
3620{
3621 struct extent_buffer *eb, *exists = NULL;
3622 int ret;
3623
3624 eb = find_extent_buffer(fs_info, start);
3625 if (eb)
3626 return eb;
3627 eb = alloc_dummy_extent_buffer(fs_info, start);
3628 if (!eb)
3629 return ERR_PTR(-ENOMEM);
3630 eb->fs_info = fs_info;
3631again:
3632 ret = radix_tree_preload(GFP_NOFS);
3633 if (ret) {
3634 exists = ERR_PTR(ret);
3635 goto free_eb;
3636 }
3637 spin_lock(&fs_info->buffer_lock);
3638 ret = radix_tree_insert(&fs_info->buffer_radix,
3639 start >> fs_info->sectorsize_bits, eb);
3640 spin_unlock(&fs_info->buffer_lock);
3641 radix_tree_preload_end();
3642 if (ret == -EEXIST) {
3643 exists = find_extent_buffer(fs_info, start);
3644 if (exists)
3645 goto free_eb;
3646 else
3647 goto again;
3648 }
3649 check_buffer_tree_ref(eb);
3650 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
3651
3652 return eb;
3653free_eb:
3654 btrfs_release_extent_buffer(eb);
3655 return exists;
3656}
3657#endif
3658
3659static struct extent_buffer *grab_extent_buffer(
3660 struct btrfs_fs_info *fs_info, struct page *page)
3661{
3662 struct folio *folio = page_folio(page);
3663 struct extent_buffer *exists;
3664
3665 /*
3666 * For subpage case, we completely rely on radix tree to ensure we
3667 * don't try to insert two ebs for the same bytenr. So here we always
3668 * return NULL and just continue.
3669 */
3670 if (fs_info->nodesize < PAGE_SIZE)
3671 return NULL;
3672
3673 /* Page not yet attached to an extent buffer */
3674 if (!folio_test_private(folio))
3675 return NULL;
3676
3677 /*
3678 * We could have already allocated an eb for this page and attached one
3679 * so lets see if we can get a ref on the existing eb, and if we can we
3680 * know it's good and we can just return that one, else we know we can
3681 * just overwrite folio private.
3682 */
3683 exists = folio_get_private(folio);
3684 if (atomic_inc_not_zero(&exists->refs))
3685 return exists;
3686
3687 WARN_ON(PageDirty(page));
3688 folio_detach_private(folio);
3689 return NULL;
3690}
3691
3692static int check_eb_alignment(struct btrfs_fs_info *fs_info, u64 start)
3693{
3694 if (!IS_ALIGNED(start, fs_info->sectorsize)) {
3695 btrfs_err(fs_info, "bad tree block start %llu", start);
3696 return -EINVAL;
3697 }
3698
3699 if (fs_info->nodesize < PAGE_SIZE &&
3700 offset_in_page(start) + fs_info->nodesize > PAGE_SIZE) {
3701 btrfs_err(fs_info,
3702 "tree block crosses page boundary, start %llu nodesize %u",
3703 start, fs_info->nodesize);
3704 return -EINVAL;
3705 }
3706 if (fs_info->nodesize >= PAGE_SIZE &&
3707 !PAGE_ALIGNED(start)) {
3708 btrfs_err(fs_info,
3709 "tree block is not page aligned, start %llu nodesize %u",
3710 start, fs_info->nodesize);
3711 return -EINVAL;
3712 }
3713 if (!IS_ALIGNED(start, fs_info->nodesize) &&
3714 !test_and_set_bit(BTRFS_FS_UNALIGNED_TREE_BLOCK, &fs_info->flags)) {
3715 btrfs_warn(fs_info,
3716"tree block not nodesize aligned, start %llu nodesize %u, can be resolved by a full metadata balance",
3717 start, fs_info->nodesize);
3718 }
3719 return 0;
3720}
3721
3722
3723/*
3724 * Return 0 if eb->folios[i] is attached to btree inode successfully.
3725 * Return >0 if there is already another extent buffer for the range,
3726 * and @found_eb_ret would be updated.
3727 * Return -EAGAIN if the filemap has an existing folio but with different size
3728 * than @eb.
3729 * The caller needs to free the existing folios and retry using the same order.
3730 */
3731static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
3732 struct extent_buffer **found_eb_ret)
3733{
3734
3735 struct btrfs_fs_info *fs_info = eb->fs_info;
3736 struct address_space *mapping = fs_info->btree_inode->i_mapping;
3737 const unsigned long index = eb->start >> PAGE_SHIFT;
3738 struct folio *existing_folio;
3739 int ret;
3740
3741 ASSERT(found_eb_ret);
3742
3743 /* Caller should ensure the folio exists. */
3744 ASSERT(eb->folios[i]);
3745
3746retry:
3747 ret = filemap_add_folio(mapping, eb->folios[i], index + i,
3748 GFP_NOFS | __GFP_NOFAIL);
3749 if (!ret)
3750 return 0;
3751
3752 existing_folio = filemap_lock_folio(mapping, index + i);
3753 /* The page cache only exists for a very short time, just retry. */
3754 if (IS_ERR(existing_folio))
3755 goto retry;
3756
3757 /* For now, we should only have single-page folios for btree inode. */
3758 ASSERT(folio_nr_pages(existing_folio) == 1);
3759
3760 if (folio_size(existing_folio) != eb->folio_size) {
3761 folio_unlock(existing_folio);
3762 folio_put(existing_folio);
3763 return -EAGAIN;
3764 }
3765
3766 if (fs_info->nodesize < PAGE_SIZE) {
3767 /*
3768 * We're going to reuse the existing page, can drop our page
3769 * and subpage structure now.
3770 */
3771 __free_page(folio_page(eb->folios[i], 0));
3772 eb->folios[i] = existing_folio;
3773 } else {
3774 struct extent_buffer *existing_eb;
3775
3776 existing_eb = grab_extent_buffer(fs_info,
3777 folio_page(existing_folio, 0));
3778 if (existing_eb) {
3779 /* The extent buffer still exists, we can use it directly. */
3780 *found_eb_ret = existing_eb;
3781 folio_unlock(existing_folio);
3782 folio_put(existing_folio);
3783 return 1;
3784 }
3785 /* The extent buffer no longer exists, we can reuse the folio. */
3786 __free_page(folio_page(eb->folios[i], 0));
3787 eb->folios[i] = existing_folio;
3788 }
3789 return 0;
3790}
3791
3792struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
3793 u64 start, u64 owner_root, int level)
3794{
3795 unsigned long len = fs_info->nodesize;
3796 int num_folios;
3797 int attached = 0;
3798 struct extent_buffer *eb;
3799 struct extent_buffer *existing_eb = NULL;
3800 struct address_space *mapping = fs_info->btree_inode->i_mapping;
3801 struct btrfs_subpage *prealloc = NULL;
3802 u64 lockdep_owner = owner_root;
3803 bool page_contig = true;
3804 int uptodate = 1;
3805 int ret;
3806
3807 if (check_eb_alignment(fs_info, start))
3808 return ERR_PTR(-EINVAL);
3809
3810#if BITS_PER_LONG == 32
3811 if (start >= MAX_LFS_FILESIZE) {
3812 btrfs_err_rl(fs_info,
3813 "extent buffer %llu is beyond 32bit page cache limit", start);
3814 btrfs_err_32bit_limit(fs_info);
3815 return ERR_PTR(-EOVERFLOW);
3816 }
3817 if (start >= BTRFS_32BIT_EARLY_WARN_THRESHOLD)
3818 btrfs_warn_32bit_limit(fs_info);
3819#endif
3820
3821 eb = find_extent_buffer(fs_info, start);
3822 if (eb)
3823 return eb;
3824
3825 eb = __alloc_extent_buffer(fs_info, start, len);
3826 if (!eb)
3827 return ERR_PTR(-ENOMEM);
3828
3829 /*
3830 * The reloc trees are just snapshots, so we need them to appear to be
3831 * just like any other fs tree WRT lockdep.
3832 */
3833 if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID)
3834 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
3835
3836 btrfs_set_buffer_lockdep_class(lockdep_owner, eb, level);
3837
3838 /*
3839 * Preallocate folio private for subpage case, so that we won't
3840 * allocate memory with i_private_lock nor page lock hold.
3841 *
3842 * The memory will be freed by attach_extent_buffer_page() or freed
3843 * manually if we exit earlier.
3844 */
3845 if (fs_info->nodesize < PAGE_SIZE) {
3846 prealloc = btrfs_alloc_subpage(fs_info, BTRFS_SUBPAGE_METADATA);
3847 if (IS_ERR(prealloc)) {
3848 ret = PTR_ERR(prealloc);
3849 goto out;
3850 }
3851 }
3852
3853reallocate:
3854 /* Allocate all pages first. */
3855 ret = alloc_eb_folio_array(eb, __GFP_NOFAIL);
3856 if (ret < 0) {
3857 btrfs_free_subpage(prealloc);
3858 goto out;
3859 }
3860
3861 num_folios = num_extent_folios(eb);
3862 /* Attach all pages to the filemap. */
3863 for (int i = 0; i < num_folios; i++) {
3864 struct folio *folio;
3865
3866 ret = attach_eb_folio_to_filemap(eb, i, &existing_eb);
3867 if (ret > 0) {
3868 ASSERT(existing_eb);
3869 goto out;
3870 }
3871
3872 /*
3873 * TODO: Special handling for a corner case where the order of
3874 * folios mismatch between the new eb and filemap.
3875 *
3876 * This happens when:
3877 *
3878 * - the new eb is using higher order folio
3879 *
3880 * - the filemap is still using 0-order folios for the range
3881 * This can happen at the previous eb allocation, and we don't
3882 * have higher order folio for the call.
3883 *
3884 * - the existing eb has already been freed
3885 *
3886 * In this case, we have to free the existing folios first, and
3887 * re-allocate using the same order.
3888 * Thankfully this is not going to happen yet, as we're still
3889 * using 0-order folios.
3890 */
3891 if (unlikely(ret == -EAGAIN)) {
3892 ASSERT(0);
3893 goto reallocate;
3894 }
3895 attached++;
3896
3897 /*
3898 * Only after attach_eb_folio_to_filemap(), eb->folios[] is
3899 * reliable, as we may choose to reuse the existing page cache
3900 * and free the allocated page.
3901 */
3902 folio = eb->folios[i];
3903 eb->folio_size = folio_size(folio);
3904 eb->folio_shift = folio_shift(folio);
3905 spin_lock(&mapping->i_private_lock);
3906 /* Should not fail, as we have preallocated the memory */
3907 ret = attach_extent_buffer_folio(eb, folio, prealloc);
3908 ASSERT(!ret);
3909 /*
3910 * To inform we have extra eb under allocation, so that
3911 * detach_extent_buffer_page() won't release the folio private
3912 * when the eb hasn't yet been inserted into radix tree.
3913 *
3914 * The ref will be decreased when the eb released the page, in
3915 * detach_extent_buffer_page().
3916 * Thus needs no special handling in error path.
3917 */
3918 btrfs_folio_inc_eb_refs(fs_info, folio);
3919 spin_unlock(&mapping->i_private_lock);
3920
3921 WARN_ON(btrfs_folio_test_dirty(fs_info, folio, eb->start, eb->len));
3922
3923 /*
3924 * Check if the current page is physically contiguous with previous eb
3925 * page.
3926 * At this stage, either we allocated a large folio, thus @i
3927 * would only be 0, or we fall back to per-page allocation.
3928 */
3929 if (i && folio_page(eb->folios[i - 1], 0) + 1 != folio_page(folio, 0))
3930 page_contig = false;
3931
3932 if (!btrfs_folio_test_uptodate(fs_info, folio, eb->start, eb->len))
3933 uptodate = 0;
3934
3935 /*
3936 * We can't unlock the pages just yet since the extent buffer
3937 * hasn't been properly inserted in the radix tree, this
3938 * opens a race with btree_release_folio which can free a page
3939 * while we are still filling in all pages for the buffer and
3940 * we could crash.
3941 */
3942 }
3943 if (uptodate)
3944 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3945 /* All pages are physically contiguous, can skip cross page handling. */
3946 if (page_contig)
3947 eb->addr = folio_address(eb->folios[0]) + offset_in_page(eb->start);
3948again:
3949 ret = radix_tree_preload(GFP_NOFS);
3950 if (ret)
3951 goto out;
3952
3953 spin_lock(&fs_info->buffer_lock);
3954 ret = radix_tree_insert(&fs_info->buffer_radix,
3955 start >> fs_info->sectorsize_bits, eb);
3956 spin_unlock(&fs_info->buffer_lock);
3957 radix_tree_preload_end();
3958 if (ret == -EEXIST) {
3959 ret = 0;
3960 existing_eb = find_extent_buffer(fs_info, start);
3961 if (existing_eb)
3962 goto out;
3963 else
3964 goto again;
3965 }
3966 /* add one reference for the tree */
3967 check_buffer_tree_ref(eb);
3968 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
3969
3970 /*
3971 * Now it's safe to unlock the pages because any calls to
3972 * btree_release_folio will correctly detect that a page belongs to a
3973 * live buffer and won't free them prematurely.
3974 */
3975 for (int i = 0; i < num_folios; i++)
3976 unlock_page(folio_page(eb->folios[i], 0));
3977 return eb;
3978
3979out:
3980 WARN_ON(!atomic_dec_and_test(&eb->refs));
3981
3982 /*
3983 * Any attached folios need to be detached before we unlock them. This
3984 * is because when we're inserting our new folios into the mapping, and
3985 * then attaching our eb to that folio. If we fail to insert our folio
3986 * we'll lookup the folio for that index, and grab that EB. We do not
3987 * want that to grab this eb, as we're getting ready to free it. So we
3988 * have to detach it first and then unlock it.
3989 *
3990 * We have to drop our reference and NULL it out here because in the
3991 * subpage case detaching does a btrfs_folio_dec_eb_refs() for our eb.
3992 * Below when we call btrfs_release_extent_buffer() we will call
3993 * detach_extent_buffer_folio() on our remaining pages in the !subpage
3994 * case. If we left eb->folios[i] populated in the subpage case we'd
3995 * double put our reference and be super sad.
3996 */
3997 for (int i = 0; i < attached; i++) {
3998 ASSERT(eb->folios[i]);
3999 detach_extent_buffer_folio(eb, eb->folios[i]);
4000 unlock_page(folio_page(eb->folios[i], 0));
4001 folio_put(eb->folios[i]);
4002 eb->folios[i] = NULL;
4003 }
4004 /*
4005 * Now all pages of that extent buffer is unmapped, set UNMAPPED flag,
4006 * so it can be cleaned up without utlizing page->mapping.
4007 */
4008 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
4009
4010 btrfs_release_extent_buffer(eb);
4011 if (ret < 0)
4012 return ERR_PTR(ret);
4013 ASSERT(existing_eb);
4014 return existing_eb;
4015}
4016
4017static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4018{
4019 struct extent_buffer *eb =
4020 container_of(head, struct extent_buffer, rcu_head);
4021
4022 __free_extent_buffer(eb);
4023}
4024
4025static int release_extent_buffer(struct extent_buffer *eb)
4026 __releases(&eb->refs_lock)
4027{
4028 lockdep_assert_held(&eb->refs_lock);
4029
4030 WARN_ON(atomic_read(&eb->refs) == 0);
4031 if (atomic_dec_and_test(&eb->refs)) {
4032 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4033 struct btrfs_fs_info *fs_info = eb->fs_info;
4034
4035 spin_unlock(&eb->refs_lock);
4036
4037 spin_lock(&fs_info->buffer_lock);
4038 radix_tree_delete(&fs_info->buffer_radix,
4039 eb->start >> fs_info->sectorsize_bits);
4040 spin_unlock(&fs_info->buffer_lock);
4041 } else {
4042 spin_unlock(&eb->refs_lock);
4043 }
4044
4045 btrfs_leak_debug_del_eb(eb);
4046 /* Should be safe to release our pages at this point */
4047 btrfs_release_extent_buffer_pages(eb);
4048#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4049 if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
4050 __free_extent_buffer(eb);
4051 return 1;
4052 }
4053#endif
4054 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4055 return 1;
4056 }
4057 spin_unlock(&eb->refs_lock);
4058
4059 return 0;
4060}
4061
4062void free_extent_buffer(struct extent_buffer *eb)
4063{
4064 int refs;
4065 if (!eb)
4066 return;
4067
4068 refs = atomic_read(&eb->refs);
4069 while (1) {
4070 if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
4071 || (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
4072 refs == 1))
4073 break;
4074 if (atomic_try_cmpxchg(&eb->refs, &refs, refs - 1))
4075 return;
4076 }
4077
4078 spin_lock(&eb->refs_lock);
4079 if (atomic_read(&eb->refs) == 2 &&
4080 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4081 !extent_buffer_under_io(eb) &&
4082 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4083 atomic_dec(&eb->refs);
4084
4085 /*
4086 * I know this is terrible, but it's temporary until we stop tracking
4087 * the uptodate bits and such for the extent buffers.
4088 */
4089 release_extent_buffer(eb);
4090}
4091
4092void free_extent_buffer_stale(struct extent_buffer *eb)
4093{
4094 if (!eb)
4095 return;
4096
4097 spin_lock(&eb->refs_lock);
4098 set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4099
4100 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4101 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4102 atomic_dec(&eb->refs);
4103 release_extent_buffer(eb);
4104}
4105
4106static void btree_clear_folio_dirty(struct folio *folio)
4107{
4108 ASSERT(folio_test_dirty(folio));
4109 ASSERT(folio_test_locked(folio));
4110 folio_clear_dirty_for_io(folio);
4111 xa_lock_irq(&folio->mapping->i_pages);
4112 if (!folio_test_dirty(folio))
4113 __xa_clear_mark(&folio->mapping->i_pages,
4114 folio_index(folio), PAGECACHE_TAG_DIRTY);
4115 xa_unlock_irq(&folio->mapping->i_pages);
4116}
4117
4118static void clear_subpage_extent_buffer_dirty(const struct extent_buffer *eb)
4119{
4120 struct btrfs_fs_info *fs_info = eb->fs_info;
4121 struct folio *folio = eb->folios[0];
4122 bool last;
4123
4124 /* btree_clear_folio_dirty() needs page locked. */
4125 folio_lock(folio);
4126 last = btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start, eb->len);
4127 if (last)
4128 btree_clear_folio_dirty(folio);
4129 folio_unlock(folio);
4130 WARN_ON(atomic_read(&eb->refs) == 0);
4131}
4132
4133void btrfs_clear_buffer_dirty(struct btrfs_trans_handle *trans,
4134 struct extent_buffer *eb)
4135{
4136 struct btrfs_fs_info *fs_info = eb->fs_info;
4137 int num_folios;
4138
4139 btrfs_assert_tree_write_locked(eb);
4140
4141 if (trans && btrfs_header_generation(eb) != trans->transid)
4142 return;
4143
4144 /*
4145 * Instead of clearing the dirty flag off of the buffer, mark it as
4146 * EXTENT_BUFFER_ZONED_ZEROOUT. This allows us to preserve
4147 * write-ordering in zoned mode, without the need to later re-dirty
4148 * the extent_buffer.
4149 *
4150 * The actual zeroout of the buffer will happen later in
4151 * btree_csum_one_bio.
4152 */
4153 if (btrfs_is_zoned(fs_info) && test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
4154 set_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags);
4155 return;
4156 }
4157
4158 if (!test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags))
4159 return;
4160
4161 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes, -eb->len,
4162 fs_info->dirty_metadata_batch);
4163
4164 if (eb->fs_info->nodesize < PAGE_SIZE)
4165 return clear_subpage_extent_buffer_dirty(eb);
4166
4167 num_folios = num_extent_folios(eb);
4168 for (int i = 0; i < num_folios; i++) {
4169 struct folio *folio = eb->folios[i];
4170
4171 if (!folio_test_dirty(folio))
4172 continue;
4173 folio_lock(folio);
4174 btree_clear_folio_dirty(folio);
4175 folio_unlock(folio);
4176 }
4177 WARN_ON(atomic_read(&eb->refs) == 0);
4178}
4179
4180void set_extent_buffer_dirty(struct extent_buffer *eb)
4181{
4182 int num_folios;
4183 bool was_dirty;
4184
4185 check_buffer_tree_ref(eb);
4186
4187 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4188
4189 num_folios = num_extent_folios(eb);
4190 WARN_ON(atomic_read(&eb->refs) == 0);
4191 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4192 WARN_ON(test_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags));
4193
4194 if (!was_dirty) {
4195 bool subpage = eb->fs_info->nodesize < PAGE_SIZE;
4196
4197 /*
4198 * For subpage case, we can have other extent buffers in the
4199 * same page, and in clear_subpage_extent_buffer_dirty() we
4200 * have to clear page dirty without subpage lock held.
4201 * This can cause race where our page gets dirty cleared after
4202 * we just set it.
4203 *
4204 * Thankfully, clear_subpage_extent_buffer_dirty() has locked
4205 * its page for other reasons, we can use page lock to prevent
4206 * the above race.
4207 */
4208 if (subpage)
4209 lock_page(folio_page(eb->folios[0], 0));
4210 for (int i = 0; i < num_folios; i++)
4211 btrfs_folio_set_dirty(eb->fs_info, eb->folios[i],
4212 eb->start, eb->len);
4213 if (subpage)
4214 unlock_page(folio_page(eb->folios[0], 0));
4215 percpu_counter_add_batch(&eb->fs_info->dirty_metadata_bytes,
4216 eb->len,
4217 eb->fs_info->dirty_metadata_batch);
4218 }
4219#ifdef CONFIG_BTRFS_DEBUG
4220 for (int i = 0; i < num_folios; i++)
4221 ASSERT(folio_test_dirty(eb->folios[i]));
4222#endif
4223}
4224
4225void clear_extent_buffer_uptodate(struct extent_buffer *eb)
4226{
4227 struct btrfs_fs_info *fs_info = eb->fs_info;
4228 int num_folios = num_extent_folios(eb);
4229
4230 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4231 for (int i = 0; i < num_folios; i++) {
4232 struct folio *folio = eb->folios[i];
4233
4234 if (!folio)
4235 continue;
4236
4237 /*
4238 * This is special handling for metadata subpage, as regular
4239 * btrfs_is_subpage() can not handle cloned/dummy metadata.
4240 */
4241 if (fs_info->nodesize >= PAGE_SIZE)
4242 folio_clear_uptodate(folio);
4243 else
4244 btrfs_subpage_clear_uptodate(fs_info, folio,
4245 eb->start, eb->len);
4246 }
4247}
4248
4249void set_extent_buffer_uptodate(struct extent_buffer *eb)
4250{
4251 struct btrfs_fs_info *fs_info = eb->fs_info;
4252 int num_folios = num_extent_folios(eb);
4253
4254 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4255 for (int i = 0; i < num_folios; i++) {
4256 struct folio *folio = eb->folios[i];
4257
4258 /*
4259 * This is special handling for metadata subpage, as regular
4260 * btrfs_is_subpage() can not handle cloned/dummy metadata.
4261 */
4262 if (fs_info->nodesize >= PAGE_SIZE)
4263 folio_mark_uptodate(folio);
4264 else
4265 btrfs_subpage_set_uptodate(fs_info, folio,
4266 eb->start, eb->len);
4267 }
4268}
4269
4270static void end_bbio_meta_read(struct btrfs_bio *bbio)
4271{
4272 struct extent_buffer *eb = bbio->private;
4273 struct btrfs_fs_info *fs_info = eb->fs_info;
4274 bool uptodate = !bbio->bio.bi_status;
4275 struct folio_iter fi;
4276 u32 bio_offset = 0;
4277
4278 eb->read_mirror = bbio->mirror_num;
4279
4280 if (uptodate &&
4281 btrfs_validate_extent_buffer(eb, &bbio->parent_check) < 0)
4282 uptodate = false;
4283
4284 if (uptodate) {
4285 set_extent_buffer_uptodate(eb);
4286 } else {
4287 clear_extent_buffer_uptodate(eb);
4288 set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
4289 }
4290
4291 bio_for_each_folio_all(fi, &bbio->bio) {
4292 struct folio *folio = fi.folio;
4293 u64 start = eb->start + bio_offset;
4294 u32 len = fi.length;
4295
4296 if (uptodate)
4297 btrfs_folio_set_uptodate(fs_info, folio, start, len);
4298 else
4299 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
4300
4301 bio_offset += len;
4302 }
4303
4304 clear_bit(EXTENT_BUFFER_READING, &eb->bflags);
4305 smp_mb__after_atomic();
4306 wake_up_bit(&eb->bflags, EXTENT_BUFFER_READING);
4307 free_extent_buffer(eb);
4308
4309 bio_put(&bbio->bio);
4310}
4311
4312int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num,
4313 struct btrfs_tree_parent_check *check)
4314{
4315 struct btrfs_bio *bbio;
4316 bool ret;
4317
4318 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4319 return 0;
4320
4321 /*
4322 * We could have had EXTENT_BUFFER_UPTODATE cleared by the write
4323 * operation, which could potentially still be in flight. In this case
4324 * we simply want to return an error.
4325 */
4326 if (unlikely(test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)))
4327 return -EIO;
4328
4329 /* Someone else is already reading the buffer, just wait for it. */
4330 if (test_and_set_bit(EXTENT_BUFFER_READING, &eb->bflags))
4331 goto done;
4332
4333 /*
4334 * Between the initial test_bit(EXTENT_BUFFER_UPTODATE) and the above
4335 * test_and_set_bit(EXTENT_BUFFER_READING), someone else could have
4336 * started and finished reading the same eb. In this case, UPTODATE
4337 * will now be set, and we shouldn't read it in again.
4338 */
4339 if (unlikely(test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))) {
4340 clear_bit(EXTENT_BUFFER_READING, &eb->bflags);
4341 smp_mb__after_atomic();
4342 wake_up_bit(&eb->bflags, EXTENT_BUFFER_READING);
4343 return 0;
4344 }
4345
4346 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
4347 eb->read_mirror = 0;
4348 check_buffer_tree_ref(eb);
4349 atomic_inc(&eb->refs);
4350
4351 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
4352 REQ_OP_READ | REQ_META, eb->fs_info,
4353 end_bbio_meta_read, eb);
4354 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
4355 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
4356 bbio->file_offset = eb->start;
4357 memcpy(&bbio->parent_check, check, sizeof(*check));
4358 if (eb->fs_info->nodesize < PAGE_SIZE) {
4359 ret = bio_add_folio(&bbio->bio, eb->folios[0], eb->len,
4360 eb->start - folio_pos(eb->folios[0]));
4361 ASSERT(ret);
4362 } else {
4363 int num_folios = num_extent_folios(eb);
4364
4365 for (int i = 0; i < num_folios; i++) {
4366 struct folio *folio = eb->folios[i];
4367
4368 ret = bio_add_folio(&bbio->bio, folio, eb->folio_size, 0);
4369 ASSERT(ret);
4370 }
4371 }
4372 btrfs_submit_bio(bbio, mirror_num);
4373
4374done:
4375 if (wait == WAIT_COMPLETE) {
4376 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_READING, TASK_UNINTERRUPTIBLE);
4377 if (!test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4378 return -EIO;
4379 }
4380
4381 return 0;
4382}
4383
4384static bool report_eb_range(const struct extent_buffer *eb, unsigned long start,
4385 unsigned long len)
4386{
4387 btrfs_warn(eb->fs_info,
4388 "access to eb bytenr %llu len %u out of range start %lu len %lu",
4389 eb->start, eb->len, start, len);
4390 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4391
4392 return true;
4393}
4394
4395/*
4396 * Check if the [start, start + len) range is valid before reading/writing
4397 * the eb.
4398 * NOTE: @start and @len are offset inside the eb, not logical address.
4399 *
4400 * Caller should not touch the dst/src memory if this function returns error.
4401 */
4402static inline int check_eb_range(const struct extent_buffer *eb,
4403 unsigned long start, unsigned long len)
4404{
4405 unsigned long offset;
4406
4407 /* start, start + len should not go beyond eb->len nor overflow */
4408 if (unlikely(check_add_overflow(start, len, &offset) || offset > eb->len))
4409 return report_eb_range(eb, start, len);
4410
4411 return false;
4412}
4413
4414void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
4415 unsigned long start, unsigned long len)
4416{
4417 const int unit_size = eb->folio_size;
4418 size_t cur;
4419 size_t offset;
4420 char *dst = (char *)dstv;
4421 unsigned long i = get_eb_folio_index(eb, start);
4422
4423 if (check_eb_range(eb, start, len)) {
4424 /*
4425 * Invalid range hit, reset the memory, so callers won't get
4426 * some random garbage for their uninitialized memory.
4427 */
4428 memset(dstv, 0, len);
4429 return;
4430 }
4431
4432 if (eb->addr) {
4433 memcpy(dstv, eb->addr + start, len);
4434 return;
4435 }
4436
4437 offset = get_eb_offset_in_folio(eb, start);
4438
4439 while (len > 0) {
4440 char *kaddr;
4441
4442 cur = min(len, unit_size - offset);
4443 kaddr = folio_address(eb->folios[i]);
4444 memcpy(dst, kaddr + offset, cur);
4445
4446 dst += cur;
4447 len -= cur;
4448 offset = 0;
4449 i++;
4450 }
4451}
4452
4453int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
4454 void __user *dstv,
4455 unsigned long start, unsigned long len)
4456{
4457 const int unit_size = eb->folio_size;
4458 size_t cur;
4459 size_t offset;
4460 char __user *dst = (char __user *)dstv;
4461 unsigned long i = get_eb_folio_index(eb, start);
4462 int ret = 0;
4463
4464 WARN_ON(start > eb->len);
4465 WARN_ON(start + len > eb->start + eb->len);
4466
4467 if (eb->addr) {
4468 if (copy_to_user_nofault(dstv, eb->addr + start, len))
4469 ret = -EFAULT;
4470 return ret;
4471 }
4472
4473 offset = get_eb_offset_in_folio(eb, start);
4474
4475 while (len > 0) {
4476 char *kaddr;
4477
4478 cur = min(len, unit_size - offset);
4479 kaddr = folio_address(eb->folios[i]);
4480 if (copy_to_user_nofault(dst, kaddr + offset, cur)) {
4481 ret = -EFAULT;
4482 break;
4483 }
4484
4485 dst += cur;
4486 len -= cur;
4487 offset = 0;
4488 i++;
4489 }
4490
4491 return ret;
4492}
4493
4494int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
4495 unsigned long start, unsigned long len)
4496{
4497 const int unit_size = eb->folio_size;
4498 size_t cur;
4499 size_t offset;
4500 char *kaddr;
4501 char *ptr = (char *)ptrv;
4502 unsigned long i = get_eb_folio_index(eb, start);
4503 int ret = 0;
4504
4505 if (check_eb_range(eb, start, len))
4506 return -EINVAL;
4507
4508 if (eb->addr)
4509 return memcmp(ptrv, eb->addr + start, len);
4510
4511 offset = get_eb_offset_in_folio(eb, start);
4512
4513 while (len > 0) {
4514 cur = min(len, unit_size - offset);
4515 kaddr = folio_address(eb->folios[i]);
4516 ret = memcmp(ptr, kaddr + offset, cur);
4517 if (ret)
4518 break;
4519
4520 ptr += cur;
4521 len -= cur;
4522 offset = 0;
4523 i++;
4524 }
4525 return ret;
4526}
4527
4528/*
4529 * Check that the extent buffer is uptodate.
4530 *
4531 * For regular sector size == PAGE_SIZE case, check if @page is uptodate.
4532 * For subpage case, check if the range covered by the eb has EXTENT_UPTODATE.
4533 */
4534static void assert_eb_folio_uptodate(const struct extent_buffer *eb, int i)
4535{
4536 struct btrfs_fs_info *fs_info = eb->fs_info;
4537 struct folio *folio = eb->folios[i];
4538
4539 ASSERT(folio);
4540
4541 /*
4542 * If we are using the commit root we could potentially clear a page
4543 * Uptodate while we're using the extent buffer that we've previously
4544 * looked up. We don't want to complain in this case, as the page was
4545 * valid before, we just didn't write it out. Instead we want to catch
4546 * the case where we didn't actually read the block properly, which
4547 * would have !PageUptodate and !EXTENT_BUFFER_WRITE_ERR.
4548 */
4549 if (test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
4550 return;
4551
4552 if (fs_info->nodesize < PAGE_SIZE) {
4553 struct folio *folio = eb->folios[0];
4554
4555 ASSERT(i == 0);
4556 if (WARN_ON(!btrfs_subpage_test_uptodate(fs_info, folio,
4557 eb->start, eb->len)))
4558 btrfs_subpage_dump_bitmap(fs_info, folio, eb->start, eb->len);
4559 } else {
4560 WARN_ON(!folio_test_uptodate(folio));
4561 }
4562}
4563
4564static void __write_extent_buffer(const struct extent_buffer *eb,
4565 const void *srcv, unsigned long start,
4566 unsigned long len, bool use_memmove)
4567{
4568 const int unit_size = eb->folio_size;
4569 size_t cur;
4570 size_t offset;
4571 char *kaddr;
4572 char *src = (char *)srcv;
4573 unsigned long i = get_eb_folio_index(eb, start);
4574 /* For unmapped (dummy) ebs, no need to check their uptodate status. */
4575 const bool check_uptodate = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
4576
4577 if (check_eb_range(eb, start, len))
4578 return;
4579
4580 if (eb->addr) {
4581 if (use_memmove)
4582 memmove(eb->addr + start, srcv, len);
4583 else
4584 memcpy(eb->addr + start, srcv, len);
4585 return;
4586 }
4587
4588 offset = get_eb_offset_in_folio(eb, start);
4589
4590 while (len > 0) {
4591 if (check_uptodate)
4592 assert_eb_folio_uptodate(eb, i);
4593
4594 cur = min(len, unit_size - offset);
4595 kaddr = folio_address(eb->folios[i]);
4596 if (use_memmove)
4597 memmove(kaddr + offset, src, cur);
4598 else
4599 memcpy(kaddr + offset, src, cur);
4600
4601 src += cur;
4602 len -= cur;
4603 offset = 0;
4604 i++;
4605 }
4606}
4607
4608void write_extent_buffer(const struct extent_buffer *eb, const void *srcv,
4609 unsigned long start, unsigned long len)
4610{
4611 return __write_extent_buffer(eb, srcv, start, len, false);
4612}
4613
4614static void memset_extent_buffer(const struct extent_buffer *eb, int c,
4615 unsigned long start, unsigned long len)
4616{
4617 const int unit_size = eb->folio_size;
4618 unsigned long cur = start;
4619
4620 if (eb->addr) {
4621 memset(eb->addr + start, c, len);
4622 return;
4623 }
4624
4625 while (cur < start + len) {
4626 unsigned long index = get_eb_folio_index(eb, cur);
4627 unsigned int offset = get_eb_offset_in_folio(eb, cur);
4628 unsigned int cur_len = min(start + len - cur, unit_size - offset);
4629
4630 assert_eb_folio_uptodate(eb, index);
4631 memset(folio_address(eb->folios[index]) + offset, c, cur_len);
4632
4633 cur += cur_len;
4634 }
4635}
4636
4637void memzero_extent_buffer(const struct extent_buffer *eb, unsigned long start,
4638 unsigned long len)
4639{
4640 if (check_eb_range(eb, start, len))
4641 return;
4642 return memset_extent_buffer(eb, 0, start, len);
4643}
4644
4645void copy_extent_buffer_full(const struct extent_buffer *dst,
4646 const struct extent_buffer *src)
4647{
4648 const int unit_size = src->folio_size;
4649 unsigned long cur = 0;
4650
4651 ASSERT(dst->len == src->len);
4652
4653 while (cur < src->len) {
4654 unsigned long index = get_eb_folio_index(src, cur);
4655 unsigned long offset = get_eb_offset_in_folio(src, cur);
4656 unsigned long cur_len = min(src->len, unit_size - offset);
4657 void *addr = folio_address(src->folios[index]) + offset;
4658
4659 write_extent_buffer(dst, addr, cur, cur_len);
4660
4661 cur += cur_len;
4662 }
4663}
4664
4665void copy_extent_buffer(const struct extent_buffer *dst,
4666 const struct extent_buffer *src,
4667 unsigned long dst_offset, unsigned long src_offset,
4668 unsigned long len)
4669{
4670 const int unit_size = dst->folio_size;
4671 u64 dst_len = dst->len;
4672 size_t cur;
4673 size_t offset;
4674 char *kaddr;
4675 unsigned long i = get_eb_folio_index(dst, dst_offset);
4676
4677 if (check_eb_range(dst, dst_offset, len) ||
4678 check_eb_range(src, src_offset, len))
4679 return;
4680
4681 WARN_ON(src->len != dst_len);
4682
4683 offset = get_eb_offset_in_folio(dst, dst_offset);
4684
4685 while (len > 0) {
4686 assert_eb_folio_uptodate(dst, i);
4687
4688 cur = min(len, (unsigned long)(unit_size - offset));
4689
4690 kaddr = folio_address(dst->folios[i]);
4691 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4692
4693 src_offset += cur;
4694 len -= cur;
4695 offset = 0;
4696 i++;
4697 }
4698}
4699
4700/*
4701 * Calculate the folio and offset of the byte containing the given bit number.
4702 *
4703 * @eb: the extent buffer
4704 * @start: offset of the bitmap item in the extent buffer
4705 * @nr: bit number
4706 * @folio_index: return index of the folio in the extent buffer that contains
4707 * the given bit number
4708 * @folio_offset: return offset into the folio given by folio_index
4709 *
4710 * This helper hides the ugliness of finding the byte in an extent buffer which
4711 * contains a given bit.
4712 */
4713static inline void eb_bitmap_offset(const struct extent_buffer *eb,
4714 unsigned long start, unsigned long nr,
4715 unsigned long *folio_index,
4716 size_t *folio_offset)
4717{
4718 size_t byte_offset = BIT_BYTE(nr);
4719 size_t offset;
4720
4721 /*
4722 * The byte we want is the offset of the extent buffer + the offset of
4723 * the bitmap item in the extent buffer + the offset of the byte in the
4724 * bitmap item.
4725 */
4726 offset = start + offset_in_eb_folio(eb, eb->start) + byte_offset;
4727
4728 *folio_index = offset >> eb->folio_shift;
4729 *folio_offset = offset_in_eb_folio(eb, offset);
4730}
4731
4732/*
4733 * Determine whether a bit in a bitmap item is set.
4734 *
4735 * @eb: the extent buffer
4736 * @start: offset of the bitmap item in the extent buffer
4737 * @nr: bit number to test
4738 */
4739int extent_buffer_test_bit(const struct extent_buffer *eb, unsigned long start,
4740 unsigned long nr)
4741{
4742 unsigned long i;
4743 size_t offset;
4744 u8 *kaddr;
4745
4746 eb_bitmap_offset(eb, start, nr, &i, &offset);
4747 assert_eb_folio_uptodate(eb, i);
4748 kaddr = folio_address(eb->folios[i]);
4749 return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
4750}
4751
4752static u8 *extent_buffer_get_byte(const struct extent_buffer *eb, unsigned long bytenr)
4753{
4754 unsigned long index = get_eb_folio_index(eb, bytenr);
4755
4756 if (check_eb_range(eb, bytenr, 1))
4757 return NULL;
4758 return folio_address(eb->folios[index]) + get_eb_offset_in_folio(eb, bytenr);
4759}
4760
4761/*
4762 * Set an area of a bitmap to 1.
4763 *
4764 * @eb: the extent buffer
4765 * @start: offset of the bitmap item in the extent buffer
4766 * @pos: bit number of the first bit
4767 * @len: number of bits to set
4768 */
4769void extent_buffer_bitmap_set(const struct extent_buffer *eb, unsigned long start,
4770 unsigned long pos, unsigned long len)
4771{
4772 unsigned int first_byte = start + BIT_BYTE(pos);
4773 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4774 const bool same_byte = (first_byte == last_byte);
4775 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4776 u8 *kaddr;
4777
4778 if (same_byte)
4779 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4780
4781 /* Handle the first byte. */
4782 kaddr = extent_buffer_get_byte(eb, first_byte);
4783 *kaddr |= mask;
4784 if (same_byte)
4785 return;
4786
4787 /* Handle the byte aligned part. */
4788 ASSERT(first_byte + 1 <= last_byte);
4789 memset_extent_buffer(eb, 0xff, first_byte + 1, last_byte - first_byte - 1);
4790
4791 /* Handle the last byte. */
4792 kaddr = extent_buffer_get_byte(eb, last_byte);
4793 *kaddr |= BITMAP_LAST_BYTE_MASK(pos + len);
4794}
4795
4796
4797/*
4798 * Clear an area of a bitmap.
4799 *
4800 * @eb: the extent buffer
4801 * @start: offset of the bitmap item in the extent buffer
4802 * @pos: bit number of the first bit
4803 * @len: number of bits to clear
4804 */
4805void extent_buffer_bitmap_clear(const struct extent_buffer *eb,
4806 unsigned long start, unsigned long pos,
4807 unsigned long len)
4808{
4809 unsigned int first_byte = start + BIT_BYTE(pos);
4810 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4811 const bool same_byte = (first_byte == last_byte);
4812 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4813 u8 *kaddr;
4814
4815 if (same_byte)
4816 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4817
4818 /* Handle the first byte. */
4819 kaddr = extent_buffer_get_byte(eb, first_byte);
4820 *kaddr &= ~mask;
4821 if (same_byte)
4822 return;
4823
4824 /* Handle the byte aligned part. */
4825 ASSERT(first_byte + 1 <= last_byte);
4826 memset_extent_buffer(eb, 0, first_byte + 1, last_byte - first_byte - 1);
4827
4828 /* Handle the last byte. */
4829 kaddr = extent_buffer_get_byte(eb, last_byte);
4830 *kaddr &= ~BITMAP_LAST_BYTE_MASK(pos + len);
4831}
4832
4833static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4834{
4835 unsigned long distance = (src > dst) ? src - dst : dst - src;
4836 return distance < len;
4837}
4838
4839void memcpy_extent_buffer(const struct extent_buffer *dst,
4840 unsigned long dst_offset, unsigned long src_offset,
4841 unsigned long len)
4842{
4843 const int unit_size = dst->folio_size;
4844 unsigned long cur_off = 0;
4845
4846 if (check_eb_range(dst, dst_offset, len) ||
4847 check_eb_range(dst, src_offset, len))
4848 return;
4849
4850 if (dst->addr) {
4851 const bool use_memmove = areas_overlap(src_offset, dst_offset, len);
4852
4853 if (use_memmove)
4854 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4855 else
4856 memcpy(dst->addr + dst_offset, dst->addr + src_offset, len);
4857 return;
4858 }
4859
4860 while (cur_off < len) {
4861 unsigned long cur_src = cur_off + src_offset;
4862 unsigned long folio_index = get_eb_folio_index(dst, cur_src);
4863 unsigned long folio_off = get_eb_offset_in_folio(dst, cur_src);
4864 unsigned long cur_len = min(src_offset + len - cur_src,
4865 unit_size - folio_off);
4866 void *src_addr = folio_address(dst->folios[folio_index]) + folio_off;
4867 const bool use_memmove = areas_overlap(src_offset + cur_off,
4868 dst_offset + cur_off, cur_len);
4869
4870 __write_extent_buffer(dst, src_addr, dst_offset + cur_off, cur_len,
4871 use_memmove);
4872 cur_off += cur_len;
4873 }
4874}
4875
4876void memmove_extent_buffer(const struct extent_buffer *dst,
4877 unsigned long dst_offset, unsigned long src_offset,
4878 unsigned long len)
4879{
4880 unsigned long dst_end = dst_offset + len - 1;
4881 unsigned long src_end = src_offset + len - 1;
4882
4883 if (check_eb_range(dst, dst_offset, len) ||
4884 check_eb_range(dst, src_offset, len))
4885 return;
4886
4887 if (dst_offset < src_offset) {
4888 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4889 return;
4890 }
4891
4892 if (dst->addr) {
4893 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4894 return;
4895 }
4896
4897 while (len > 0) {
4898 unsigned long src_i;
4899 size_t cur;
4900 size_t dst_off_in_folio;
4901 size_t src_off_in_folio;
4902 void *src_addr;
4903 bool use_memmove;
4904
4905 src_i = get_eb_folio_index(dst, src_end);
4906
4907 dst_off_in_folio = get_eb_offset_in_folio(dst, dst_end);
4908 src_off_in_folio = get_eb_offset_in_folio(dst, src_end);
4909
4910 cur = min_t(unsigned long, len, src_off_in_folio + 1);
4911 cur = min(cur, dst_off_in_folio + 1);
4912
4913 src_addr = folio_address(dst->folios[src_i]) + src_off_in_folio -
4914 cur + 1;
4915 use_memmove = areas_overlap(src_end - cur + 1, dst_end - cur + 1,
4916 cur);
4917
4918 __write_extent_buffer(dst, src_addr, dst_end - cur + 1, cur,
4919 use_memmove);
4920
4921 dst_end -= cur;
4922 src_end -= cur;
4923 len -= cur;
4924 }
4925}
4926
4927#define GANG_LOOKUP_SIZE 16
4928static struct extent_buffer *get_next_extent_buffer(
4929 struct btrfs_fs_info *fs_info, struct page *page, u64 bytenr)
4930{
4931 struct extent_buffer *gang[GANG_LOOKUP_SIZE];
4932 struct extent_buffer *found = NULL;
4933 u64 page_start = page_offset(page);
4934 u64 cur = page_start;
4935
4936 ASSERT(in_range(bytenr, page_start, PAGE_SIZE));
4937 lockdep_assert_held(&fs_info->buffer_lock);
4938
4939 while (cur < page_start + PAGE_SIZE) {
4940 int ret;
4941 int i;
4942
4943 ret = radix_tree_gang_lookup(&fs_info->buffer_radix,
4944 (void **)gang, cur >> fs_info->sectorsize_bits,
4945 min_t(unsigned int, GANG_LOOKUP_SIZE,
4946 PAGE_SIZE / fs_info->nodesize));
4947 if (ret == 0)
4948 goto out;
4949 for (i = 0; i < ret; i++) {
4950 /* Already beyond page end */
4951 if (gang[i]->start >= page_start + PAGE_SIZE)
4952 goto out;
4953 /* Found one */
4954 if (gang[i]->start >= bytenr) {
4955 found = gang[i];
4956 goto out;
4957 }
4958 }
4959 cur = gang[ret - 1]->start + gang[ret - 1]->len;
4960 }
4961out:
4962 return found;
4963}
4964
4965static int try_release_subpage_extent_buffer(struct page *page)
4966{
4967 struct btrfs_fs_info *fs_info = page_to_fs_info(page);
4968 u64 cur = page_offset(page);
4969 const u64 end = page_offset(page) + PAGE_SIZE;
4970 int ret;
4971
4972 while (cur < end) {
4973 struct extent_buffer *eb = NULL;
4974
4975 /*
4976 * Unlike try_release_extent_buffer() which uses folio private
4977 * to grab buffer, for subpage case we rely on radix tree, thus
4978 * we need to ensure radix tree consistency.
4979 *
4980 * We also want an atomic snapshot of the radix tree, thus go
4981 * with spinlock rather than RCU.
4982 */
4983 spin_lock(&fs_info->buffer_lock);
4984 eb = get_next_extent_buffer(fs_info, page, cur);
4985 if (!eb) {
4986 /* No more eb in the page range after or at cur */
4987 spin_unlock(&fs_info->buffer_lock);
4988 break;
4989 }
4990 cur = eb->start + eb->len;
4991
4992 /*
4993 * The same as try_release_extent_buffer(), to ensure the eb
4994 * won't disappear out from under us.
4995 */
4996 spin_lock(&eb->refs_lock);
4997 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4998 spin_unlock(&eb->refs_lock);
4999 spin_unlock(&fs_info->buffer_lock);
5000 break;
5001 }
5002 spin_unlock(&fs_info->buffer_lock);
5003
5004 /*
5005 * If tree ref isn't set then we know the ref on this eb is a
5006 * real ref, so just return, this eb will likely be freed soon
5007 * anyway.
5008 */
5009 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5010 spin_unlock(&eb->refs_lock);
5011 break;
5012 }
5013
5014 /*
5015 * Here we don't care about the return value, we will always
5016 * check the folio private at the end. And
5017 * release_extent_buffer() will release the refs_lock.
5018 */
5019 release_extent_buffer(eb);
5020 }
5021 /*
5022 * Finally to check if we have cleared folio private, as if we have
5023 * released all ebs in the page, the folio private should be cleared now.
5024 */
5025 spin_lock(&page->mapping->i_private_lock);
5026 if (!folio_test_private(page_folio(page)))
5027 ret = 1;
5028 else
5029 ret = 0;
5030 spin_unlock(&page->mapping->i_private_lock);
5031 return ret;
5032
5033}
5034
5035int try_release_extent_buffer(struct page *page)
5036{
5037 struct folio *folio = page_folio(page);
5038 struct extent_buffer *eb;
5039
5040 if (page_to_fs_info(page)->nodesize < PAGE_SIZE)
5041 return try_release_subpage_extent_buffer(page);
5042
5043 /*
5044 * We need to make sure nobody is changing folio private, as we rely on
5045 * folio private as the pointer to extent buffer.
5046 */
5047 spin_lock(&page->mapping->i_private_lock);
5048 if (!folio_test_private(folio)) {
5049 spin_unlock(&page->mapping->i_private_lock);
5050 return 1;
5051 }
5052
5053 eb = folio_get_private(folio);
5054 BUG_ON(!eb);
5055
5056 /*
5057 * This is a little awful but should be ok, we need to make sure that
5058 * the eb doesn't disappear out from under us while we're looking at
5059 * this page.
5060 */
5061 spin_lock(&eb->refs_lock);
5062 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5063 spin_unlock(&eb->refs_lock);
5064 spin_unlock(&page->mapping->i_private_lock);
5065 return 0;
5066 }
5067 spin_unlock(&page->mapping->i_private_lock);
5068
5069 /*
5070 * If tree ref isn't set then we know the ref on this eb is a real ref,
5071 * so just return, this page will likely be freed soon anyway.
5072 */
5073 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5074 spin_unlock(&eb->refs_lock);
5075 return 0;
5076 }
5077
5078 return release_extent_buffer(eb);
5079}
5080
5081/*
5082 * Attempt to readahead a child block.
5083 *
5084 * @fs_info: the fs_info
5085 * @bytenr: bytenr to read
5086 * @owner_root: objectid of the root that owns this eb
5087 * @gen: generation for the uptodate check, can be 0
5088 * @level: level for the eb
5089 *
5090 * Attempt to readahead a tree block at @bytenr. If @gen is 0 then we do a
5091 * normal uptodate check of the eb, without checking the generation. If we have
5092 * to read the block we will not block on anything.
5093 */
5094void btrfs_readahead_tree_block(struct btrfs_fs_info *fs_info,
5095 u64 bytenr, u64 owner_root, u64 gen, int level)
5096{
5097 struct btrfs_tree_parent_check check = {
5098 .has_first_key = 0,
5099 .level = level,
5100 .transid = gen
5101 };
5102 struct extent_buffer *eb;
5103 int ret;
5104
5105 eb = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
5106 if (IS_ERR(eb))
5107 return;
5108
5109 if (btrfs_buffer_uptodate(eb, gen, 1)) {
5110 free_extent_buffer(eb);
5111 return;
5112 }
5113
5114 ret = read_extent_buffer_pages(eb, WAIT_NONE, 0, &check);
5115 if (ret < 0)
5116 free_extent_buffer_stale(eb);
5117 else
5118 free_extent_buffer(eb);
5119}
5120
5121/*
5122 * Readahead a node's child block.
5123 *
5124 * @node: parent node we're reading from
5125 * @slot: slot in the parent node for the child we want to read
5126 *
5127 * A helper for btrfs_readahead_tree_block, we simply read the bytenr pointed at
5128 * the slot in the node provided.
5129 */
5130void btrfs_readahead_node_child(struct extent_buffer *node, int slot)
5131{
5132 btrfs_readahead_tree_block(node->fs_info,
5133 btrfs_node_blockptr(node, slot),
5134 btrfs_header_owner(node),
5135 btrfs_node_ptr_generation(node, slot),
5136 btrfs_header_level(node) - 1);
5137}