Loading...
1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/bitops.h>
4#include <linux/slab.h>
5#include <linux/bio.h>
6#include <linux/mm.h>
7#include <linux/pagemap.h>
8#include <linux/page-flags.h>
9#include <linux/spinlock.h>
10#include <linux/blkdev.h>
11#include <linux/swap.h>
12#include <linux/writeback.h>
13#include <linux/pagevec.h>
14#include <linux/prefetch.h>
15#include <linux/cleancache.h>
16#include "extent_io.h"
17#include "extent-io-tree.h"
18#include "extent_map.h"
19#include "ctree.h"
20#include "btrfs_inode.h"
21#include "volumes.h"
22#include "check-integrity.h"
23#include "locking.h"
24#include "rcu-string.h"
25#include "backref.h"
26#include "disk-io.h"
27
28static struct kmem_cache *extent_state_cache;
29static struct kmem_cache *extent_buffer_cache;
30static struct bio_set btrfs_bioset;
31
32static inline bool extent_state_in_tree(const struct extent_state *state)
33{
34 return !RB_EMPTY_NODE(&state->rb_node);
35}
36
37#ifdef CONFIG_BTRFS_DEBUG
38static LIST_HEAD(states);
39static DEFINE_SPINLOCK(leak_lock);
40
41static inline void btrfs_leak_debug_add(spinlock_t *lock,
42 struct list_head *new,
43 struct list_head *head)
44{
45 unsigned long flags;
46
47 spin_lock_irqsave(lock, flags);
48 list_add(new, head);
49 spin_unlock_irqrestore(lock, flags);
50}
51
52static inline void btrfs_leak_debug_del(spinlock_t *lock,
53 struct list_head *entry)
54{
55 unsigned long flags;
56
57 spin_lock_irqsave(lock, flags);
58 list_del(entry);
59 spin_unlock_irqrestore(lock, flags);
60}
61
62void btrfs_extent_buffer_leak_debug_check(struct btrfs_fs_info *fs_info)
63{
64 struct extent_buffer *eb;
65 unsigned long flags;
66
67 /*
68 * If we didn't get into open_ctree our allocated_ebs will not be
69 * initialized, so just skip this.
70 */
71 if (!fs_info->allocated_ebs.next)
72 return;
73
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 %lu 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 kmem_cache_free(extent_buffer_cache, eb);
84 }
85 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
86}
87
88static inline void btrfs_extent_state_leak_debug_check(void)
89{
90 struct extent_state *state;
91
92 while (!list_empty(&states)) {
93 state = list_entry(states.next, struct extent_state, leak_list);
94 pr_err("BTRFS: state leak: start %llu end %llu state %u in tree %d refs %d\n",
95 state->start, state->end, state->state,
96 extent_state_in_tree(state),
97 refcount_read(&state->refs));
98 list_del(&state->leak_list);
99 kmem_cache_free(extent_state_cache, state);
100 }
101}
102
103#define btrfs_debug_check_extent_io_range(tree, start, end) \
104 __btrfs_debug_check_extent_io_range(__func__, (tree), (start), (end))
105static inline void __btrfs_debug_check_extent_io_range(const char *caller,
106 struct extent_io_tree *tree, u64 start, u64 end)
107{
108 struct inode *inode = tree->private_data;
109 u64 isize;
110
111 if (!inode || !is_data_inode(inode))
112 return;
113
114 isize = i_size_read(inode);
115 if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
116 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
117 "%s: ino %llu isize %llu odd range [%llu,%llu]",
118 caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
119 }
120}
121#else
122#define btrfs_leak_debug_add(lock, new, head) do {} while (0)
123#define btrfs_leak_debug_del(lock, entry) do {} while (0)
124#define btrfs_extent_state_leak_debug_check() do {} while (0)
125#define btrfs_debug_check_extent_io_range(c, s, e) do {} while (0)
126#endif
127
128struct tree_entry {
129 u64 start;
130 u64 end;
131 struct rb_node rb_node;
132};
133
134struct extent_page_data {
135 struct bio *bio;
136 /* tells writepage not to lock the state bits for this range
137 * it still does the unlocking
138 */
139 unsigned int extent_locked:1;
140
141 /* tells the submit_bio code to use REQ_SYNC */
142 unsigned int sync_io:1;
143};
144
145static int add_extent_changeset(struct extent_state *state, unsigned bits,
146 struct extent_changeset *changeset,
147 int set)
148{
149 int ret;
150
151 if (!changeset)
152 return 0;
153 if (set && (state->state & bits) == bits)
154 return 0;
155 if (!set && (state->state & bits) == 0)
156 return 0;
157 changeset->bytes_changed += state->end - state->start + 1;
158 ret = ulist_add(&changeset->range_changed, state->start, state->end,
159 GFP_ATOMIC);
160 return ret;
161}
162
163static int __must_check submit_one_bio(struct bio *bio, int mirror_num,
164 unsigned long bio_flags)
165{
166 blk_status_t ret = 0;
167 struct extent_io_tree *tree = bio->bi_private;
168
169 bio->bi_private = NULL;
170
171 if (tree->ops)
172 ret = tree->ops->submit_bio_hook(tree->private_data, bio,
173 mirror_num, bio_flags);
174 else
175 btrfsic_submit_bio(bio);
176
177 return blk_status_to_errno(ret);
178}
179
180/* Cleanup unsubmitted bios */
181static void end_write_bio(struct extent_page_data *epd, int ret)
182{
183 if (epd->bio) {
184 epd->bio->bi_status = errno_to_blk_status(ret);
185 bio_endio(epd->bio);
186 epd->bio = NULL;
187 }
188}
189
190/*
191 * Submit bio from extent page data via submit_one_bio
192 *
193 * Return 0 if everything is OK.
194 * Return <0 for error.
195 */
196static int __must_check flush_write_bio(struct extent_page_data *epd)
197{
198 int ret = 0;
199
200 if (epd->bio) {
201 ret = submit_one_bio(epd->bio, 0, 0);
202 /*
203 * Clean up of epd->bio is handled by its endio function.
204 * And endio is either triggered by successful bio execution
205 * or the error handler of submit bio hook.
206 * So at this point, no matter what happened, we don't need
207 * to clean up epd->bio.
208 */
209 epd->bio = NULL;
210 }
211 return ret;
212}
213
214int __init extent_state_cache_init(void)
215{
216 extent_state_cache = kmem_cache_create("btrfs_extent_state",
217 sizeof(struct extent_state), 0,
218 SLAB_MEM_SPREAD, NULL);
219 if (!extent_state_cache)
220 return -ENOMEM;
221 return 0;
222}
223
224int __init extent_io_init(void)
225{
226 extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
227 sizeof(struct extent_buffer), 0,
228 SLAB_MEM_SPREAD, NULL);
229 if (!extent_buffer_cache)
230 return -ENOMEM;
231
232 if (bioset_init(&btrfs_bioset, BIO_POOL_SIZE,
233 offsetof(struct btrfs_io_bio, bio),
234 BIOSET_NEED_BVECS))
235 goto free_buffer_cache;
236
237 if (bioset_integrity_create(&btrfs_bioset, BIO_POOL_SIZE))
238 goto free_bioset;
239
240 return 0;
241
242free_bioset:
243 bioset_exit(&btrfs_bioset);
244
245free_buffer_cache:
246 kmem_cache_destroy(extent_buffer_cache);
247 extent_buffer_cache = NULL;
248 return -ENOMEM;
249}
250
251void __cold extent_state_cache_exit(void)
252{
253 btrfs_extent_state_leak_debug_check();
254 kmem_cache_destroy(extent_state_cache);
255}
256
257void __cold extent_io_exit(void)
258{
259 /*
260 * Make sure all delayed rcu free are flushed before we
261 * destroy caches.
262 */
263 rcu_barrier();
264 kmem_cache_destroy(extent_buffer_cache);
265 bioset_exit(&btrfs_bioset);
266}
267
268/*
269 * For the file_extent_tree, we want to hold the inode lock when we lookup and
270 * update the disk_i_size, but lockdep will complain because our io_tree we hold
271 * the tree lock and get the inode lock when setting delalloc. These two things
272 * are unrelated, so make a class for the file_extent_tree so we don't get the
273 * two locking patterns mixed up.
274 */
275static struct lock_class_key file_extent_tree_class;
276
277void extent_io_tree_init(struct btrfs_fs_info *fs_info,
278 struct extent_io_tree *tree, unsigned int owner,
279 void *private_data)
280{
281 tree->fs_info = fs_info;
282 tree->state = RB_ROOT;
283 tree->ops = NULL;
284 tree->dirty_bytes = 0;
285 spin_lock_init(&tree->lock);
286 tree->private_data = private_data;
287 tree->owner = owner;
288 if (owner == IO_TREE_INODE_FILE_EXTENT)
289 lockdep_set_class(&tree->lock, &file_extent_tree_class);
290}
291
292void extent_io_tree_release(struct extent_io_tree *tree)
293{
294 spin_lock(&tree->lock);
295 /*
296 * Do a single barrier for the waitqueue_active check here, the state
297 * of the waitqueue should not change once extent_io_tree_release is
298 * called.
299 */
300 smp_mb();
301 while (!RB_EMPTY_ROOT(&tree->state)) {
302 struct rb_node *node;
303 struct extent_state *state;
304
305 node = rb_first(&tree->state);
306 state = rb_entry(node, struct extent_state, rb_node);
307 rb_erase(&state->rb_node, &tree->state);
308 RB_CLEAR_NODE(&state->rb_node);
309 /*
310 * btree io trees aren't supposed to have tasks waiting for
311 * changes in the flags of extent states ever.
312 */
313 ASSERT(!waitqueue_active(&state->wq));
314 free_extent_state(state);
315
316 cond_resched_lock(&tree->lock);
317 }
318 spin_unlock(&tree->lock);
319}
320
321static struct extent_state *alloc_extent_state(gfp_t mask)
322{
323 struct extent_state *state;
324
325 /*
326 * The given mask might be not appropriate for the slab allocator,
327 * drop the unsupported bits
328 */
329 mask &= ~(__GFP_DMA32|__GFP_HIGHMEM);
330 state = kmem_cache_alloc(extent_state_cache, mask);
331 if (!state)
332 return state;
333 state->state = 0;
334 state->failrec = NULL;
335 RB_CLEAR_NODE(&state->rb_node);
336 btrfs_leak_debug_add(&leak_lock, &state->leak_list, &states);
337 refcount_set(&state->refs, 1);
338 init_waitqueue_head(&state->wq);
339 trace_alloc_extent_state(state, mask, _RET_IP_);
340 return state;
341}
342
343void free_extent_state(struct extent_state *state)
344{
345 if (!state)
346 return;
347 if (refcount_dec_and_test(&state->refs)) {
348 WARN_ON(extent_state_in_tree(state));
349 btrfs_leak_debug_del(&leak_lock, &state->leak_list);
350 trace_free_extent_state(state, _RET_IP_);
351 kmem_cache_free(extent_state_cache, state);
352 }
353}
354
355static struct rb_node *tree_insert(struct rb_root *root,
356 struct rb_node *search_start,
357 u64 offset,
358 struct rb_node *node,
359 struct rb_node ***p_in,
360 struct rb_node **parent_in)
361{
362 struct rb_node **p;
363 struct rb_node *parent = NULL;
364 struct tree_entry *entry;
365
366 if (p_in && parent_in) {
367 p = *p_in;
368 parent = *parent_in;
369 goto do_insert;
370 }
371
372 p = search_start ? &search_start : &root->rb_node;
373 while (*p) {
374 parent = *p;
375 entry = rb_entry(parent, struct tree_entry, rb_node);
376
377 if (offset < entry->start)
378 p = &(*p)->rb_left;
379 else if (offset > entry->end)
380 p = &(*p)->rb_right;
381 else
382 return parent;
383 }
384
385do_insert:
386 rb_link_node(node, parent, p);
387 rb_insert_color(node, root);
388 return NULL;
389}
390
391/**
392 * __etree_search - searche @tree for an entry that contains @offset. Such
393 * entry would have entry->start <= offset && entry->end >= offset.
394 *
395 * @tree - the tree to search
396 * @offset - offset that should fall within an entry in @tree
397 * @next_ret - pointer to the first entry whose range ends after @offset
398 * @prev - pointer to the first entry whose range begins before @offset
399 * @p_ret - pointer where new node should be anchored (used when inserting an
400 * entry in the tree)
401 * @parent_ret - points to entry which would have been the parent of the entry,
402 * containing @offset
403 *
404 * This function returns a pointer to the entry that contains @offset byte
405 * address. If no such entry exists, then NULL is returned and the other
406 * pointer arguments to the function are filled, otherwise the found entry is
407 * returned and other pointers are left untouched.
408 */
409static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
410 struct rb_node **next_ret,
411 struct rb_node **prev_ret,
412 struct rb_node ***p_ret,
413 struct rb_node **parent_ret)
414{
415 struct rb_root *root = &tree->state;
416 struct rb_node **n = &root->rb_node;
417 struct rb_node *prev = NULL;
418 struct rb_node *orig_prev = NULL;
419 struct tree_entry *entry;
420 struct tree_entry *prev_entry = NULL;
421
422 while (*n) {
423 prev = *n;
424 entry = rb_entry(prev, struct tree_entry, rb_node);
425 prev_entry = entry;
426
427 if (offset < entry->start)
428 n = &(*n)->rb_left;
429 else if (offset > entry->end)
430 n = &(*n)->rb_right;
431 else
432 return *n;
433 }
434
435 if (p_ret)
436 *p_ret = n;
437 if (parent_ret)
438 *parent_ret = prev;
439
440 if (next_ret) {
441 orig_prev = prev;
442 while (prev && offset > prev_entry->end) {
443 prev = rb_next(prev);
444 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
445 }
446 *next_ret = prev;
447 prev = orig_prev;
448 }
449
450 if (prev_ret) {
451 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
452 while (prev && offset < prev_entry->start) {
453 prev = rb_prev(prev);
454 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
455 }
456 *prev_ret = prev;
457 }
458 return NULL;
459}
460
461static inline struct rb_node *
462tree_search_for_insert(struct extent_io_tree *tree,
463 u64 offset,
464 struct rb_node ***p_ret,
465 struct rb_node **parent_ret)
466{
467 struct rb_node *next= NULL;
468 struct rb_node *ret;
469
470 ret = __etree_search(tree, offset, &next, NULL, p_ret, parent_ret);
471 if (!ret)
472 return next;
473 return ret;
474}
475
476static inline struct rb_node *tree_search(struct extent_io_tree *tree,
477 u64 offset)
478{
479 return tree_search_for_insert(tree, offset, NULL, NULL);
480}
481
482/*
483 * utility function to look for merge candidates inside a given range.
484 * Any extents with matching state are merged together into a single
485 * extent in the tree. Extents with EXTENT_IO in their state field
486 * are not merged because the end_io handlers need to be able to do
487 * operations on them without sleeping (or doing allocations/splits).
488 *
489 * This should be called with the tree lock held.
490 */
491static void merge_state(struct extent_io_tree *tree,
492 struct extent_state *state)
493{
494 struct extent_state *other;
495 struct rb_node *other_node;
496
497 if (state->state & (EXTENT_LOCKED | EXTENT_BOUNDARY))
498 return;
499
500 other_node = rb_prev(&state->rb_node);
501 if (other_node) {
502 other = rb_entry(other_node, struct extent_state, rb_node);
503 if (other->end == state->start - 1 &&
504 other->state == state->state) {
505 if (tree->private_data &&
506 is_data_inode(tree->private_data))
507 btrfs_merge_delalloc_extent(tree->private_data,
508 state, other);
509 state->start = other->start;
510 rb_erase(&other->rb_node, &tree->state);
511 RB_CLEAR_NODE(&other->rb_node);
512 free_extent_state(other);
513 }
514 }
515 other_node = rb_next(&state->rb_node);
516 if (other_node) {
517 other = rb_entry(other_node, struct extent_state, rb_node);
518 if (other->start == state->end + 1 &&
519 other->state == state->state) {
520 if (tree->private_data &&
521 is_data_inode(tree->private_data))
522 btrfs_merge_delalloc_extent(tree->private_data,
523 state, other);
524 state->end = other->end;
525 rb_erase(&other->rb_node, &tree->state);
526 RB_CLEAR_NODE(&other->rb_node);
527 free_extent_state(other);
528 }
529 }
530}
531
532static void set_state_bits(struct extent_io_tree *tree,
533 struct extent_state *state, unsigned *bits,
534 struct extent_changeset *changeset);
535
536/*
537 * insert an extent_state struct into the tree. 'bits' are set on the
538 * struct before it is inserted.
539 *
540 * This may return -EEXIST if the extent is already there, in which case the
541 * state struct is freed.
542 *
543 * The tree lock is not taken internally. This is a utility function and
544 * probably isn't what you want to call (see set/clear_extent_bit).
545 */
546static int insert_state(struct extent_io_tree *tree,
547 struct extent_state *state, u64 start, u64 end,
548 struct rb_node ***p,
549 struct rb_node **parent,
550 unsigned *bits, struct extent_changeset *changeset)
551{
552 struct rb_node *node;
553
554 if (end < start) {
555 btrfs_err(tree->fs_info,
556 "insert state: end < start %llu %llu", end, start);
557 WARN_ON(1);
558 }
559 state->start = start;
560 state->end = end;
561
562 set_state_bits(tree, state, bits, changeset);
563
564 node = tree_insert(&tree->state, NULL, end, &state->rb_node, p, parent);
565 if (node) {
566 struct extent_state *found;
567 found = rb_entry(node, struct extent_state, rb_node);
568 btrfs_err(tree->fs_info,
569 "found node %llu %llu on insert of %llu %llu",
570 found->start, found->end, start, end);
571 return -EEXIST;
572 }
573 merge_state(tree, state);
574 return 0;
575}
576
577/*
578 * split a given extent state struct in two, inserting the preallocated
579 * struct 'prealloc' as the newly created second half. 'split' indicates an
580 * offset inside 'orig' where it should be split.
581 *
582 * Before calling,
583 * the tree has 'orig' at [orig->start, orig->end]. After calling, there
584 * are two extent state structs in the tree:
585 * prealloc: [orig->start, split - 1]
586 * orig: [ split, orig->end ]
587 *
588 * The tree locks are not taken by this function. They need to be held
589 * by the caller.
590 */
591static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
592 struct extent_state *prealloc, u64 split)
593{
594 struct rb_node *node;
595
596 if (tree->private_data && is_data_inode(tree->private_data))
597 btrfs_split_delalloc_extent(tree->private_data, orig, split);
598
599 prealloc->start = orig->start;
600 prealloc->end = split - 1;
601 prealloc->state = orig->state;
602 orig->start = split;
603
604 node = tree_insert(&tree->state, &orig->rb_node, prealloc->end,
605 &prealloc->rb_node, NULL, NULL);
606 if (node) {
607 free_extent_state(prealloc);
608 return -EEXIST;
609 }
610 return 0;
611}
612
613static struct extent_state *next_state(struct extent_state *state)
614{
615 struct rb_node *next = rb_next(&state->rb_node);
616 if (next)
617 return rb_entry(next, struct extent_state, rb_node);
618 else
619 return NULL;
620}
621
622/*
623 * utility function to clear some bits in an extent state struct.
624 * it will optionally wake up anyone waiting on this state (wake == 1).
625 *
626 * If no bits are set on the state struct after clearing things, the
627 * struct is freed and removed from the tree
628 */
629static struct extent_state *clear_state_bit(struct extent_io_tree *tree,
630 struct extent_state *state,
631 unsigned *bits, int wake,
632 struct extent_changeset *changeset)
633{
634 struct extent_state *next;
635 unsigned bits_to_clear = *bits & ~EXTENT_CTLBITS;
636 int ret;
637
638 if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
639 u64 range = state->end - state->start + 1;
640 WARN_ON(range > tree->dirty_bytes);
641 tree->dirty_bytes -= range;
642 }
643
644 if (tree->private_data && is_data_inode(tree->private_data))
645 btrfs_clear_delalloc_extent(tree->private_data, state, bits);
646
647 ret = add_extent_changeset(state, bits_to_clear, changeset, 0);
648 BUG_ON(ret < 0);
649 state->state &= ~bits_to_clear;
650 if (wake)
651 wake_up(&state->wq);
652 if (state->state == 0) {
653 next = next_state(state);
654 if (extent_state_in_tree(state)) {
655 rb_erase(&state->rb_node, &tree->state);
656 RB_CLEAR_NODE(&state->rb_node);
657 free_extent_state(state);
658 } else {
659 WARN_ON(1);
660 }
661 } else {
662 merge_state(tree, state);
663 next = next_state(state);
664 }
665 return next;
666}
667
668static struct extent_state *
669alloc_extent_state_atomic(struct extent_state *prealloc)
670{
671 if (!prealloc)
672 prealloc = alloc_extent_state(GFP_ATOMIC);
673
674 return prealloc;
675}
676
677static void extent_io_tree_panic(struct extent_io_tree *tree, int err)
678{
679 struct inode *inode = tree->private_data;
680
681 btrfs_panic(btrfs_sb(inode->i_sb), err,
682 "locking error: extent tree was modified by another thread while locked");
683}
684
685/*
686 * clear some bits on a range in the tree. This may require splitting
687 * or inserting elements in the tree, so the gfp mask is used to
688 * indicate which allocations or sleeping are allowed.
689 *
690 * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
691 * the given range from the tree regardless of state (ie for truncate).
692 *
693 * the range [start, end] is inclusive.
694 *
695 * This takes the tree lock, and returns 0 on success and < 0 on error.
696 */
697int __clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
698 unsigned bits, int wake, int delete,
699 struct extent_state **cached_state,
700 gfp_t mask, struct extent_changeset *changeset)
701{
702 struct extent_state *state;
703 struct extent_state *cached;
704 struct extent_state *prealloc = NULL;
705 struct rb_node *node;
706 u64 last_end;
707 int err;
708 int clear = 0;
709
710 btrfs_debug_check_extent_io_range(tree, start, end);
711 trace_btrfs_clear_extent_bit(tree, start, end - start + 1, bits);
712
713 if (bits & EXTENT_DELALLOC)
714 bits |= EXTENT_NORESERVE;
715
716 if (delete)
717 bits |= ~EXTENT_CTLBITS;
718
719 if (bits & (EXTENT_LOCKED | EXTENT_BOUNDARY))
720 clear = 1;
721again:
722 if (!prealloc && gfpflags_allow_blocking(mask)) {
723 /*
724 * Don't care for allocation failure here because we might end
725 * up not needing the pre-allocated extent state at all, which
726 * is the case if we only have in the tree extent states that
727 * cover our input range and don't cover too any other range.
728 * If we end up needing a new extent state we allocate it later.
729 */
730 prealloc = alloc_extent_state(mask);
731 }
732
733 spin_lock(&tree->lock);
734 if (cached_state) {
735 cached = *cached_state;
736
737 if (clear) {
738 *cached_state = NULL;
739 cached_state = NULL;
740 }
741
742 if (cached && extent_state_in_tree(cached) &&
743 cached->start <= start && cached->end > start) {
744 if (clear)
745 refcount_dec(&cached->refs);
746 state = cached;
747 goto hit_next;
748 }
749 if (clear)
750 free_extent_state(cached);
751 }
752 /*
753 * this search will find the extents that end after
754 * our range starts
755 */
756 node = tree_search(tree, start);
757 if (!node)
758 goto out;
759 state = rb_entry(node, struct extent_state, rb_node);
760hit_next:
761 if (state->start > end)
762 goto out;
763 WARN_ON(state->end < start);
764 last_end = state->end;
765
766 /* the state doesn't have the wanted bits, go ahead */
767 if (!(state->state & bits)) {
768 state = next_state(state);
769 goto next;
770 }
771
772 /*
773 * | ---- desired range ---- |
774 * | state | or
775 * | ------------- state -------------- |
776 *
777 * We need to split the extent we found, and may flip
778 * bits on second half.
779 *
780 * If the extent we found extends past our range, we
781 * just split and search again. It'll get split again
782 * the next time though.
783 *
784 * If the extent we found is inside our range, we clear
785 * the desired bit on it.
786 */
787
788 if (state->start < start) {
789 prealloc = alloc_extent_state_atomic(prealloc);
790 BUG_ON(!prealloc);
791 err = split_state(tree, state, prealloc, start);
792 if (err)
793 extent_io_tree_panic(tree, err);
794
795 prealloc = NULL;
796 if (err)
797 goto out;
798 if (state->end <= end) {
799 state = clear_state_bit(tree, state, &bits, wake,
800 changeset);
801 goto next;
802 }
803 goto search_again;
804 }
805 /*
806 * | ---- desired range ---- |
807 * | state |
808 * We need to split the extent, and clear the bit
809 * on the first half
810 */
811 if (state->start <= end && state->end > end) {
812 prealloc = alloc_extent_state_atomic(prealloc);
813 BUG_ON(!prealloc);
814 err = split_state(tree, state, prealloc, end + 1);
815 if (err)
816 extent_io_tree_panic(tree, err);
817
818 if (wake)
819 wake_up(&state->wq);
820
821 clear_state_bit(tree, prealloc, &bits, wake, changeset);
822
823 prealloc = NULL;
824 goto out;
825 }
826
827 state = clear_state_bit(tree, state, &bits, wake, changeset);
828next:
829 if (last_end == (u64)-1)
830 goto out;
831 start = last_end + 1;
832 if (start <= end && state && !need_resched())
833 goto hit_next;
834
835search_again:
836 if (start > end)
837 goto out;
838 spin_unlock(&tree->lock);
839 if (gfpflags_allow_blocking(mask))
840 cond_resched();
841 goto again;
842
843out:
844 spin_unlock(&tree->lock);
845 if (prealloc)
846 free_extent_state(prealloc);
847
848 return 0;
849
850}
851
852static void wait_on_state(struct extent_io_tree *tree,
853 struct extent_state *state)
854 __releases(tree->lock)
855 __acquires(tree->lock)
856{
857 DEFINE_WAIT(wait);
858 prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
859 spin_unlock(&tree->lock);
860 schedule();
861 spin_lock(&tree->lock);
862 finish_wait(&state->wq, &wait);
863}
864
865/*
866 * waits for one or more bits to clear on a range in the state tree.
867 * The range [start, end] is inclusive.
868 * The tree lock is taken by this function
869 */
870static void wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
871 unsigned long bits)
872{
873 struct extent_state *state;
874 struct rb_node *node;
875
876 btrfs_debug_check_extent_io_range(tree, start, end);
877
878 spin_lock(&tree->lock);
879again:
880 while (1) {
881 /*
882 * this search will find all the extents that end after
883 * our range starts
884 */
885 node = tree_search(tree, start);
886process_node:
887 if (!node)
888 break;
889
890 state = rb_entry(node, struct extent_state, rb_node);
891
892 if (state->start > end)
893 goto out;
894
895 if (state->state & bits) {
896 start = state->start;
897 refcount_inc(&state->refs);
898 wait_on_state(tree, state);
899 free_extent_state(state);
900 goto again;
901 }
902 start = state->end + 1;
903
904 if (start > end)
905 break;
906
907 if (!cond_resched_lock(&tree->lock)) {
908 node = rb_next(node);
909 goto process_node;
910 }
911 }
912out:
913 spin_unlock(&tree->lock);
914}
915
916static void set_state_bits(struct extent_io_tree *tree,
917 struct extent_state *state,
918 unsigned *bits, struct extent_changeset *changeset)
919{
920 unsigned bits_to_set = *bits & ~EXTENT_CTLBITS;
921 int ret;
922
923 if (tree->private_data && is_data_inode(tree->private_data))
924 btrfs_set_delalloc_extent(tree->private_data, state, bits);
925
926 if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
927 u64 range = state->end - state->start + 1;
928 tree->dirty_bytes += range;
929 }
930 ret = add_extent_changeset(state, bits_to_set, changeset, 1);
931 BUG_ON(ret < 0);
932 state->state |= bits_to_set;
933}
934
935static void cache_state_if_flags(struct extent_state *state,
936 struct extent_state **cached_ptr,
937 unsigned flags)
938{
939 if (cached_ptr && !(*cached_ptr)) {
940 if (!flags || (state->state & flags)) {
941 *cached_ptr = state;
942 refcount_inc(&state->refs);
943 }
944 }
945}
946
947static void cache_state(struct extent_state *state,
948 struct extent_state **cached_ptr)
949{
950 return cache_state_if_flags(state, cached_ptr,
951 EXTENT_LOCKED | EXTENT_BOUNDARY);
952}
953
954/*
955 * set some bits on a range in the tree. This may require allocations or
956 * sleeping, so the gfp mask is used to indicate what is allowed.
957 *
958 * If any of the exclusive bits are set, this will fail with -EEXIST if some
959 * part of the range already has the desired bits set. The start of the
960 * existing range is returned in failed_start in this case.
961 *
962 * [start, end] is inclusive This takes the tree lock.
963 */
964
965static int __must_check
966__set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
967 unsigned bits, unsigned exclusive_bits,
968 u64 *failed_start, struct extent_state **cached_state,
969 gfp_t mask, struct extent_changeset *changeset)
970{
971 struct extent_state *state;
972 struct extent_state *prealloc = NULL;
973 struct rb_node *node;
974 struct rb_node **p;
975 struct rb_node *parent;
976 int err = 0;
977 u64 last_start;
978 u64 last_end;
979
980 btrfs_debug_check_extent_io_range(tree, start, end);
981 trace_btrfs_set_extent_bit(tree, start, end - start + 1, bits);
982
983again:
984 if (!prealloc && gfpflags_allow_blocking(mask)) {
985 /*
986 * Don't care for allocation failure here because we might end
987 * up not needing the pre-allocated extent state at all, which
988 * is the case if we only have in the tree extent states that
989 * cover our input range and don't cover too any other range.
990 * If we end up needing a new extent state we allocate it later.
991 */
992 prealloc = alloc_extent_state(mask);
993 }
994
995 spin_lock(&tree->lock);
996 if (cached_state && *cached_state) {
997 state = *cached_state;
998 if (state->start <= start && state->end > start &&
999 extent_state_in_tree(state)) {
1000 node = &state->rb_node;
1001 goto hit_next;
1002 }
1003 }
1004 /*
1005 * this search will find all the extents that end after
1006 * our range starts.
1007 */
1008 node = tree_search_for_insert(tree, start, &p, &parent);
1009 if (!node) {
1010 prealloc = alloc_extent_state_atomic(prealloc);
1011 BUG_ON(!prealloc);
1012 err = insert_state(tree, prealloc, start, end,
1013 &p, &parent, &bits, changeset);
1014 if (err)
1015 extent_io_tree_panic(tree, err);
1016
1017 cache_state(prealloc, cached_state);
1018 prealloc = NULL;
1019 goto out;
1020 }
1021 state = rb_entry(node, struct extent_state, rb_node);
1022hit_next:
1023 last_start = state->start;
1024 last_end = state->end;
1025
1026 /*
1027 * | ---- desired range ---- |
1028 * | state |
1029 *
1030 * Just lock what we found and keep going
1031 */
1032 if (state->start == start && state->end <= end) {
1033 if (state->state & exclusive_bits) {
1034 *failed_start = state->start;
1035 err = -EEXIST;
1036 goto out;
1037 }
1038
1039 set_state_bits(tree, state, &bits, changeset);
1040 cache_state(state, cached_state);
1041 merge_state(tree, state);
1042 if (last_end == (u64)-1)
1043 goto out;
1044 start = last_end + 1;
1045 state = next_state(state);
1046 if (start < end && state && state->start == start &&
1047 !need_resched())
1048 goto hit_next;
1049 goto search_again;
1050 }
1051
1052 /*
1053 * | ---- desired range ---- |
1054 * | state |
1055 * or
1056 * | ------------- state -------------- |
1057 *
1058 * We need to split the extent we found, and may flip bits on
1059 * second half.
1060 *
1061 * If the extent we found extends past our
1062 * range, we just split and search again. It'll get split
1063 * again the next time though.
1064 *
1065 * If the extent we found is inside our range, we set the
1066 * desired bit on it.
1067 */
1068 if (state->start < start) {
1069 if (state->state & exclusive_bits) {
1070 *failed_start = start;
1071 err = -EEXIST;
1072 goto out;
1073 }
1074
1075 /*
1076 * If this extent already has all the bits we want set, then
1077 * skip it, not necessary to split it or do anything with it.
1078 */
1079 if ((state->state & bits) == bits) {
1080 start = state->end + 1;
1081 cache_state(state, cached_state);
1082 goto search_again;
1083 }
1084
1085 prealloc = alloc_extent_state_atomic(prealloc);
1086 BUG_ON(!prealloc);
1087 err = split_state(tree, state, prealloc, start);
1088 if (err)
1089 extent_io_tree_panic(tree, err);
1090
1091 prealloc = NULL;
1092 if (err)
1093 goto out;
1094 if (state->end <= end) {
1095 set_state_bits(tree, state, &bits, changeset);
1096 cache_state(state, cached_state);
1097 merge_state(tree, state);
1098 if (last_end == (u64)-1)
1099 goto out;
1100 start = last_end + 1;
1101 state = next_state(state);
1102 if (start < end && state && state->start == start &&
1103 !need_resched())
1104 goto hit_next;
1105 }
1106 goto search_again;
1107 }
1108 /*
1109 * | ---- desired range ---- |
1110 * | state | or | state |
1111 *
1112 * There's a hole, we need to insert something in it and
1113 * ignore the extent we found.
1114 */
1115 if (state->start > start) {
1116 u64 this_end;
1117 if (end < last_start)
1118 this_end = end;
1119 else
1120 this_end = last_start - 1;
1121
1122 prealloc = alloc_extent_state_atomic(prealloc);
1123 BUG_ON(!prealloc);
1124
1125 /*
1126 * Avoid to free 'prealloc' if it can be merged with
1127 * the later extent.
1128 */
1129 err = insert_state(tree, prealloc, start, this_end,
1130 NULL, NULL, &bits, changeset);
1131 if (err)
1132 extent_io_tree_panic(tree, err);
1133
1134 cache_state(prealloc, cached_state);
1135 prealloc = NULL;
1136 start = this_end + 1;
1137 goto search_again;
1138 }
1139 /*
1140 * | ---- desired range ---- |
1141 * | state |
1142 * We need to split the extent, and set the bit
1143 * on the first half
1144 */
1145 if (state->start <= end && state->end > end) {
1146 if (state->state & exclusive_bits) {
1147 *failed_start = start;
1148 err = -EEXIST;
1149 goto out;
1150 }
1151
1152 prealloc = alloc_extent_state_atomic(prealloc);
1153 BUG_ON(!prealloc);
1154 err = split_state(tree, state, prealloc, end + 1);
1155 if (err)
1156 extent_io_tree_panic(tree, err);
1157
1158 set_state_bits(tree, prealloc, &bits, changeset);
1159 cache_state(prealloc, cached_state);
1160 merge_state(tree, prealloc);
1161 prealloc = NULL;
1162 goto out;
1163 }
1164
1165search_again:
1166 if (start > end)
1167 goto out;
1168 spin_unlock(&tree->lock);
1169 if (gfpflags_allow_blocking(mask))
1170 cond_resched();
1171 goto again;
1172
1173out:
1174 spin_unlock(&tree->lock);
1175 if (prealloc)
1176 free_extent_state(prealloc);
1177
1178 return err;
1179
1180}
1181
1182int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1183 unsigned bits, u64 * failed_start,
1184 struct extent_state **cached_state, gfp_t mask)
1185{
1186 return __set_extent_bit(tree, start, end, bits, 0, failed_start,
1187 cached_state, mask, NULL);
1188}
1189
1190
1191/**
1192 * convert_extent_bit - convert all bits in a given range from one bit to
1193 * another
1194 * @tree: the io tree to search
1195 * @start: the start offset in bytes
1196 * @end: the end offset in bytes (inclusive)
1197 * @bits: the bits to set in this range
1198 * @clear_bits: the bits to clear in this range
1199 * @cached_state: state that we're going to cache
1200 *
1201 * This will go through and set bits for the given range. If any states exist
1202 * already in this range they are set with the given bit and cleared of the
1203 * clear_bits. This is only meant to be used by things that are mergeable, ie
1204 * converting from say DELALLOC to DIRTY. This is not meant to be used with
1205 * boundary bits like LOCK.
1206 *
1207 * All allocations are done with GFP_NOFS.
1208 */
1209int convert_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1210 unsigned bits, unsigned clear_bits,
1211 struct extent_state **cached_state)
1212{
1213 struct extent_state *state;
1214 struct extent_state *prealloc = NULL;
1215 struct rb_node *node;
1216 struct rb_node **p;
1217 struct rb_node *parent;
1218 int err = 0;
1219 u64 last_start;
1220 u64 last_end;
1221 bool first_iteration = true;
1222
1223 btrfs_debug_check_extent_io_range(tree, start, end);
1224 trace_btrfs_convert_extent_bit(tree, start, end - start + 1, bits,
1225 clear_bits);
1226
1227again:
1228 if (!prealloc) {
1229 /*
1230 * Best effort, don't worry if extent state allocation fails
1231 * here for the first iteration. We might have a cached state
1232 * that matches exactly the target range, in which case no
1233 * extent state allocations are needed. We'll only know this
1234 * after locking the tree.
1235 */
1236 prealloc = alloc_extent_state(GFP_NOFS);
1237 if (!prealloc && !first_iteration)
1238 return -ENOMEM;
1239 }
1240
1241 spin_lock(&tree->lock);
1242 if (cached_state && *cached_state) {
1243 state = *cached_state;
1244 if (state->start <= start && state->end > start &&
1245 extent_state_in_tree(state)) {
1246 node = &state->rb_node;
1247 goto hit_next;
1248 }
1249 }
1250
1251 /*
1252 * this search will find all the extents that end after
1253 * our range starts.
1254 */
1255 node = tree_search_for_insert(tree, start, &p, &parent);
1256 if (!node) {
1257 prealloc = alloc_extent_state_atomic(prealloc);
1258 if (!prealloc) {
1259 err = -ENOMEM;
1260 goto out;
1261 }
1262 err = insert_state(tree, prealloc, start, end,
1263 &p, &parent, &bits, NULL);
1264 if (err)
1265 extent_io_tree_panic(tree, err);
1266 cache_state(prealloc, cached_state);
1267 prealloc = NULL;
1268 goto out;
1269 }
1270 state = rb_entry(node, struct extent_state, rb_node);
1271hit_next:
1272 last_start = state->start;
1273 last_end = state->end;
1274
1275 /*
1276 * | ---- desired range ---- |
1277 * | state |
1278 *
1279 * Just lock what we found and keep going
1280 */
1281 if (state->start == start && state->end <= end) {
1282 set_state_bits(tree, state, &bits, NULL);
1283 cache_state(state, cached_state);
1284 state = clear_state_bit(tree, state, &clear_bits, 0, NULL);
1285 if (last_end == (u64)-1)
1286 goto out;
1287 start = last_end + 1;
1288 if (start < end && state && state->start == start &&
1289 !need_resched())
1290 goto hit_next;
1291 goto search_again;
1292 }
1293
1294 /*
1295 * | ---- desired range ---- |
1296 * | state |
1297 * or
1298 * | ------------- state -------------- |
1299 *
1300 * We need to split the extent we found, and may flip bits on
1301 * second half.
1302 *
1303 * If the extent we found extends past our
1304 * range, we just split and search again. It'll get split
1305 * again the next time though.
1306 *
1307 * If the extent we found is inside our range, we set the
1308 * desired bit on it.
1309 */
1310 if (state->start < start) {
1311 prealloc = alloc_extent_state_atomic(prealloc);
1312 if (!prealloc) {
1313 err = -ENOMEM;
1314 goto out;
1315 }
1316 err = split_state(tree, state, prealloc, start);
1317 if (err)
1318 extent_io_tree_panic(tree, err);
1319 prealloc = NULL;
1320 if (err)
1321 goto out;
1322 if (state->end <= end) {
1323 set_state_bits(tree, state, &bits, NULL);
1324 cache_state(state, cached_state);
1325 state = clear_state_bit(tree, state, &clear_bits, 0,
1326 NULL);
1327 if (last_end == (u64)-1)
1328 goto out;
1329 start = last_end + 1;
1330 if (start < end && state && state->start == start &&
1331 !need_resched())
1332 goto hit_next;
1333 }
1334 goto search_again;
1335 }
1336 /*
1337 * | ---- desired range ---- |
1338 * | state | or | state |
1339 *
1340 * There's a hole, we need to insert something in it and
1341 * ignore the extent we found.
1342 */
1343 if (state->start > start) {
1344 u64 this_end;
1345 if (end < last_start)
1346 this_end = end;
1347 else
1348 this_end = last_start - 1;
1349
1350 prealloc = alloc_extent_state_atomic(prealloc);
1351 if (!prealloc) {
1352 err = -ENOMEM;
1353 goto out;
1354 }
1355
1356 /*
1357 * Avoid to free 'prealloc' if it can be merged with
1358 * the later extent.
1359 */
1360 err = insert_state(tree, prealloc, start, this_end,
1361 NULL, NULL, &bits, NULL);
1362 if (err)
1363 extent_io_tree_panic(tree, err);
1364 cache_state(prealloc, cached_state);
1365 prealloc = NULL;
1366 start = this_end + 1;
1367 goto search_again;
1368 }
1369 /*
1370 * | ---- desired range ---- |
1371 * | state |
1372 * We need to split the extent, and set the bit
1373 * on the first half
1374 */
1375 if (state->start <= end && state->end > end) {
1376 prealloc = alloc_extent_state_atomic(prealloc);
1377 if (!prealloc) {
1378 err = -ENOMEM;
1379 goto out;
1380 }
1381
1382 err = split_state(tree, state, prealloc, end + 1);
1383 if (err)
1384 extent_io_tree_panic(tree, err);
1385
1386 set_state_bits(tree, prealloc, &bits, NULL);
1387 cache_state(prealloc, cached_state);
1388 clear_state_bit(tree, prealloc, &clear_bits, 0, NULL);
1389 prealloc = NULL;
1390 goto out;
1391 }
1392
1393search_again:
1394 if (start > end)
1395 goto out;
1396 spin_unlock(&tree->lock);
1397 cond_resched();
1398 first_iteration = false;
1399 goto again;
1400
1401out:
1402 spin_unlock(&tree->lock);
1403 if (prealloc)
1404 free_extent_state(prealloc);
1405
1406 return err;
1407}
1408
1409/* wrappers around set/clear extent bit */
1410int set_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1411 unsigned bits, struct extent_changeset *changeset)
1412{
1413 /*
1414 * We don't support EXTENT_LOCKED yet, as current changeset will
1415 * record any bits changed, so for EXTENT_LOCKED case, it will
1416 * either fail with -EEXIST or changeset will record the whole
1417 * range.
1418 */
1419 BUG_ON(bits & EXTENT_LOCKED);
1420
1421 return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL, GFP_NOFS,
1422 changeset);
1423}
1424
1425int set_extent_bits_nowait(struct extent_io_tree *tree, u64 start, u64 end,
1426 unsigned bits)
1427{
1428 return __set_extent_bit(tree, start, end, bits, 0, NULL, NULL,
1429 GFP_NOWAIT, NULL);
1430}
1431
1432int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
1433 unsigned bits, int wake, int delete,
1434 struct extent_state **cached)
1435{
1436 return __clear_extent_bit(tree, start, end, bits, wake, delete,
1437 cached, GFP_NOFS, NULL);
1438}
1439
1440int clear_record_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1441 unsigned bits, struct extent_changeset *changeset)
1442{
1443 /*
1444 * Don't support EXTENT_LOCKED case, same reason as
1445 * set_record_extent_bits().
1446 */
1447 BUG_ON(bits & EXTENT_LOCKED);
1448
1449 return __clear_extent_bit(tree, start, end, bits, 0, 0, NULL, GFP_NOFS,
1450 changeset);
1451}
1452
1453/*
1454 * either insert or lock state struct between start and end use mask to tell
1455 * us if waiting is desired.
1456 */
1457int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
1458 struct extent_state **cached_state)
1459{
1460 int err;
1461 u64 failed_start;
1462
1463 while (1) {
1464 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED,
1465 EXTENT_LOCKED, &failed_start,
1466 cached_state, GFP_NOFS, NULL);
1467 if (err == -EEXIST) {
1468 wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
1469 start = failed_start;
1470 } else
1471 break;
1472 WARN_ON(start > end);
1473 }
1474 return err;
1475}
1476
1477int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end)
1478{
1479 int err;
1480 u64 failed_start;
1481
1482 err = __set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
1483 &failed_start, NULL, GFP_NOFS, NULL);
1484 if (err == -EEXIST) {
1485 if (failed_start > start)
1486 clear_extent_bit(tree, start, failed_start - 1,
1487 EXTENT_LOCKED, 1, 0, NULL);
1488 return 0;
1489 }
1490 return 1;
1491}
1492
1493void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
1494{
1495 unsigned long index = start >> PAGE_SHIFT;
1496 unsigned long end_index = end >> PAGE_SHIFT;
1497 struct page *page;
1498
1499 while (index <= end_index) {
1500 page = find_get_page(inode->i_mapping, index);
1501 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1502 clear_page_dirty_for_io(page);
1503 put_page(page);
1504 index++;
1505 }
1506}
1507
1508void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
1509{
1510 unsigned long index = start >> PAGE_SHIFT;
1511 unsigned long end_index = end >> PAGE_SHIFT;
1512 struct page *page;
1513
1514 while (index <= end_index) {
1515 page = find_get_page(inode->i_mapping, index);
1516 BUG_ON(!page); /* Pages should be in the extent_io_tree */
1517 __set_page_dirty_nobuffers(page);
1518 account_page_redirty(page);
1519 put_page(page);
1520 index++;
1521 }
1522}
1523
1524/* find the first state struct with 'bits' set after 'start', and
1525 * return it. tree->lock must be held. NULL will returned if
1526 * nothing was found after 'start'
1527 */
1528static struct extent_state *
1529find_first_extent_bit_state(struct extent_io_tree *tree,
1530 u64 start, unsigned bits)
1531{
1532 struct rb_node *node;
1533 struct extent_state *state;
1534
1535 /*
1536 * this search will find all the extents that end after
1537 * our range starts.
1538 */
1539 node = tree_search(tree, start);
1540 if (!node)
1541 goto out;
1542
1543 while (1) {
1544 state = rb_entry(node, struct extent_state, rb_node);
1545 if (state->end >= start && (state->state & bits))
1546 return state;
1547
1548 node = rb_next(node);
1549 if (!node)
1550 break;
1551 }
1552out:
1553 return NULL;
1554}
1555
1556/*
1557 * find the first offset in the io tree with 'bits' set. zero is
1558 * returned if we find something, and *start_ret and *end_ret are
1559 * set to reflect the state struct that was found.
1560 *
1561 * If nothing was found, 1 is returned. If found something, return 0.
1562 */
1563int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1564 u64 *start_ret, u64 *end_ret, unsigned bits,
1565 struct extent_state **cached_state)
1566{
1567 struct extent_state *state;
1568 int ret = 1;
1569
1570 spin_lock(&tree->lock);
1571 if (cached_state && *cached_state) {
1572 state = *cached_state;
1573 if (state->end == start - 1 && extent_state_in_tree(state)) {
1574 while ((state = next_state(state)) != NULL) {
1575 if (state->state & bits)
1576 goto got_it;
1577 }
1578 free_extent_state(*cached_state);
1579 *cached_state = NULL;
1580 goto out;
1581 }
1582 free_extent_state(*cached_state);
1583 *cached_state = NULL;
1584 }
1585
1586 state = find_first_extent_bit_state(tree, start, bits);
1587got_it:
1588 if (state) {
1589 cache_state_if_flags(state, cached_state, 0);
1590 *start_ret = state->start;
1591 *end_ret = state->end;
1592 ret = 0;
1593 }
1594out:
1595 spin_unlock(&tree->lock);
1596 return ret;
1597}
1598
1599/**
1600 * find_contiguous_extent_bit: find a contiguous area of bits
1601 * @tree - io tree to check
1602 * @start - offset to start the search from
1603 * @start_ret - the first offset we found with the bits set
1604 * @end_ret - the final contiguous range of the bits that were set
1605 * @bits - bits to look for
1606 *
1607 * set_extent_bit and clear_extent_bit can temporarily split contiguous ranges
1608 * to set bits appropriately, and then merge them again. During this time it
1609 * will drop the tree->lock, so use this helper if you want to find the actual
1610 * contiguous area for given bits. We will search to the first bit we find, and
1611 * then walk down the tree until we find a non-contiguous area. The area
1612 * returned will be the full contiguous area with the bits set.
1613 */
1614int find_contiguous_extent_bit(struct extent_io_tree *tree, u64 start,
1615 u64 *start_ret, u64 *end_ret, unsigned bits)
1616{
1617 struct extent_state *state;
1618 int ret = 1;
1619
1620 spin_lock(&tree->lock);
1621 state = find_first_extent_bit_state(tree, start, bits);
1622 if (state) {
1623 *start_ret = state->start;
1624 *end_ret = state->end;
1625 while ((state = next_state(state)) != NULL) {
1626 if (state->start > (*end_ret + 1))
1627 break;
1628 *end_ret = state->end;
1629 }
1630 ret = 0;
1631 }
1632 spin_unlock(&tree->lock);
1633 return ret;
1634}
1635
1636/**
1637 * find_first_clear_extent_bit - find the first range that has @bits not set.
1638 * This range could start before @start.
1639 *
1640 * @tree - the tree to search
1641 * @start - the offset at/after which the found extent should start
1642 * @start_ret - records the beginning of the range
1643 * @end_ret - records the end of the range (inclusive)
1644 * @bits - the set of bits which must be unset
1645 *
1646 * Since unallocated range is also considered one which doesn't have the bits
1647 * set it's possible that @end_ret contains -1, this happens in case the range
1648 * spans (last_range_end, end of device]. In this case it's up to the caller to
1649 * trim @end_ret to the appropriate size.
1650 */
1651void find_first_clear_extent_bit(struct extent_io_tree *tree, u64 start,
1652 u64 *start_ret, u64 *end_ret, unsigned bits)
1653{
1654 struct extent_state *state;
1655 struct rb_node *node, *prev = NULL, *next;
1656
1657 spin_lock(&tree->lock);
1658
1659 /* Find first extent with bits cleared */
1660 while (1) {
1661 node = __etree_search(tree, start, &next, &prev, NULL, NULL);
1662 if (!node && !next && !prev) {
1663 /*
1664 * Tree is completely empty, send full range and let
1665 * caller deal with it
1666 */
1667 *start_ret = 0;
1668 *end_ret = -1;
1669 goto out;
1670 } else if (!node && !next) {
1671 /*
1672 * We are past the last allocated chunk, set start at
1673 * the end of the last extent.
1674 */
1675 state = rb_entry(prev, struct extent_state, rb_node);
1676 *start_ret = state->end + 1;
1677 *end_ret = -1;
1678 goto out;
1679 } else if (!node) {
1680 node = next;
1681 }
1682 /*
1683 * At this point 'node' either contains 'start' or start is
1684 * before 'node'
1685 */
1686 state = rb_entry(node, struct extent_state, rb_node);
1687
1688 if (in_range(start, state->start, state->end - state->start + 1)) {
1689 if (state->state & bits) {
1690 /*
1691 * |--range with bits sets--|
1692 * |
1693 * start
1694 */
1695 start = state->end + 1;
1696 } else {
1697 /*
1698 * 'start' falls within a range that doesn't
1699 * have the bits set, so take its start as
1700 * the beginning of the desired range
1701 *
1702 * |--range with bits cleared----|
1703 * |
1704 * start
1705 */
1706 *start_ret = state->start;
1707 break;
1708 }
1709 } else {
1710 /*
1711 * |---prev range---|---hole/unset---|---node range---|
1712 * |
1713 * start
1714 *
1715 * or
1716 *
1717 * |---hole/unset--||--first node--|
1718 * 0 |
1719 * start
1720 */
1721 if (prev) {
1722 state = rb_entry(prev, struct extent_state,
1723 rb_node);
1724 *start_ret = state->end + 1;
1725 } else {
1726 *start_ret = 0;
1727 }
1728 break;
1729 }
1730 }
1731
1732 /*
1733 * Find the longest stretch from start until an entry which has the
1734 * bits set
1735 */
1736 while (1) {
1737 state = rb_entry(node, struct extent_state, rb_node);
1738 if (state->end >= start && !(state->state & bits)) {
1739 *end_ret = state->end;
1740 } else {
1741 *end_ret = state->start - 1;
1742 break;
1743 }
1744
1745 node = rb_next(node);
1746 if (!node)
1747 break;
1748 }
1749out:
1750 spin_unlock(&tree->lock);
1751}
1752
1753/*
1754 * find a contiguous range of bytes in the file marked as delalloc, not
1755 * more than 'max_bytes'. start and end are used to return the range,
1756 *
1757 * true is returned if we find something, false if nothing was in the tree
1758 */
1759bool btrfs_find_delalloc_range(struct extent_io_tree *tree, u64 *start,
1760 u64 *end, u64 max_bytes,
1761 struct extent_state **cached_state)
1762{
1763 struct rb_node *node;
1764 struct extent_state *state;
1765 u64 cur_start = *start;
1766 bool found = false;
1767 u64 total_bytes = 0;
1768
1769 spin_lock(&tree->lock);
1770
1771 /*
1772 * this search will find all the extents that end after
1773 * our range starts.
1774 */
1775 node = tree_search(tree, cur_start);
1776 if (!node) {
1777 *end = (u64)-1;
1778 goto out;
1779 }
1780
1781 while (1) {
1782 state = rb_entry(node, struct extent_state, rb_node);
1783 if (found && (state->start != cur_start ||
1784 (state->state & EXTENT_BOUNDARY))) {
1785 goto out;
1786 }
1787 if (!(state->state & EXTENT_DELALLOC)) {
1788 if (!found)
1789 *end = state->end;
1790 goto out;
1791 }
1792 if (!found) {
1793 *start = state->start;
1794 *cached_state = state;
1795 refcount_inc(&state->refs);
1796 }
1797 found = true;
1798 *end = state->end;
1799 cur_start = state->end + 1;
1800 node = rb_next(node);
1801 total_bytes += state->end - state->start + 1;
1802 if (total_bytes >= max_bytes)
1803 break;
1804 if (!node)
1805 break;
1806 }
1807out:
1808 spin_unlock(&tree->lock);
1809 return found;
1810}
1811
1812static int __process_pages_contig(struct address_space *mapping,
1813 struct page *locked_page,
1814 pgoff_t start_index, pgoff_t end_index,
1815 unsigned long page_ops, pgoff_t *index_ret);
1816
1817static noinline void __unlock_for_delalloc(struct inode *inode,
1818 struct page *locked_page,
1819 u64 start, u64 end)
1820{
1821 unsigned long index = start >> PAGE_SHIFT;
1822 unsigned long end_index = end >> PAGE_SHIFT;
1823
1824 ASSERT(locked_page);
1825 if (index == locked_page->index && end_index == index)
1826 return;
1827
1828 __process_pages_contig(inode->i_mapping, locked_page, index, end_index,
1829 PAGE_UNLOCK, NULL);
1830}
1831
1832static noinline int lock_delalloc_pages(struct inode *inode,
1833 struct page *locked_page,
1834 u64 delalloc_start,
1835 u64 delalloc_end)
1836{
1837 unsigned long index = delalloc_start >> PAGE_SHIFT;
1838 unsigned long index_ret = index;
1839 unsigned long end_index = delalloc_end >> PAGE_SHIFT;
1840 int ret;
1841
1842 ASSERT(locked_page);
1843 if (index == locked_page->index && index == end_index)
1844 return 0;
1845
1846 ret = __process_pages_contig(inode->i_mapping, locked_page, index,
1847 end_index, PAGE_LOCK, &index_ret);
1848 if (ret == -EAGAIN)
1849 __unlock_for_delalloc(inode, locked_page, delalloc_start,
1850 (u64)index_ret << PAGE_SHIFT);
1851 return ret;
1852}
1853
1854/*
1855 * Find and lock a contiguous range of bytes in the file marked as delalloc, no
1856 * more than @max_bytes. @Start and @end are used to return the range,
1857 *
1858 * Return: true if we find something
1859 * false if nothing was in the tree
1860 */
1861EXPORT_FOR_TESTS
1862noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
1863 struct page *locked_page, u64 *start,
1864 u64 *end)
1865{
1866 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1867 u64 max_bytes = BTRFS_MAX_EXTENT_SIZE;
1868 u64 delalloc_start;
1869 u64 delalloc_end;
1870 bool found;
1871 struct extent_state *cached_state = NULL;
1872 int ret;
1873 int loops = 0;
1874
1875again:
1876 /* step one, find a bunch of delalloc bytes starting at start */
1877 delalloc_start = *start;
1878 delalloc_end = 0;
1879 found = btrfs_find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1880 max_bytes, &cached_state);
1881 if (!found || delalloc_end <= *start) {
1882 *start = delalloc_start;
1883 *end = delalloc_end;
1884 free_extent_state(cached_state);
1885 return false;
1886 }
1887
1888 /*
1889 * start comes from the offset of locked_page. We have to lock
1890 * pages in order, so we can't process delalloc bytes before
1891 * locked_page
1892 */
1893 if (delalloc_start < *start)
1894 delalloc_start = *start;
1895
1896 /*
1897 * make sure to limit the number of pages we try to lock down
1898 */
1899 if (delalloc_end + 1 - delalloc_start > max_bytes)
1900 delalloc_end = delalloc_start + max_bytes - 1;
1901
1902 /* step two, lock all the pages after the page that has start */
1903 ret = lock_delalloc_pages(inode, locked_page,
1904 delalloc_start, delalloc_end);
1905 ASSERT(!ret || ret == -EAGAIN);
1906 if (ret == -EAGAIN) {
1907 /* some of the pages are gone, lets avoid looping by
1908 * shortening the size of the delalloc range we're searching
1909 */
1910 free_extent_state(cached_state);
1911 cached_state = NULL;
1912 if (!loops) {
1913 max_bytes = PAGE_SIZE;
1914 loops = 1;
1915 goto again;
1916 } else {
1917 found = false;
1918 goto out_failed;
1919 }
1920 }
1921
1922 /* step three, lock the state bits for the whole range */
1923 lock_extent_bits(tree, delalloc_start, delalloc_end, &cached_state);
1924
1925 /* then test to make sure it is all still delalloc */
1926 ret = test_range_bit(tree, delalloc_start, delalloc_end,
1927 EXTENT_DELALLOC, 1, cached_state);
1928 if (!ret) {
1929 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1930 &cached_state);
1931 __unlock_for_delalloc(inode, locked_page,
1932 delalloc_start, delalloc_end);
1933 cond_resched();
1934 goto again;
1935 }
1936 free_extent_state(cached_state);
1937 *start = delalloc_start;
1938 *end = delalloc_end;
1939out_failed:
1940 return found;
1941}
1942
1943static int __process_pages_contig(struct address_space *mapping,
1944 struct page *locked_page,
1945 pgoff_t start_index, pgoff_t end_index,
1946 unsigned long page_ops, pgoff_t *index_ret)
1947{
1948 unsigned long nr_pages = end_index - start_index + 1;
1949 unsigned long pages_locked = 0;
1950 pgoff_t index = start_index;
1951 struct page *pages[16];
1952 unsigned ret;
1953 int err = 0;
1954 int i;
1955
1956 if (page_ops & PAGE_LOCK) {
1957 ASSERT(page_ops == PAGE_LOCK);
1958 ASSERT(index_ret && *index_ret == start_index);
1959 }
1960
1961 if ((page_ops & PAGE_SET_ERROR) && nr_pages > 0)
1962 mapping_set_error(mapping, -EIO);
1963
1964 while (nr_pages > 0) {
1965 ret = find_get_pages_contig(mapping, index,
1966 min_t(unsigned long,
1967 nr_pages, ARRAY_SIZE(pages)), pages);
1968 if (ret == 0) {
1969 /*
1970 * Only if we're going to lock these pages,
1971 * can we find nothing at @index.
1972 */
1973 ASSERT(page_ops & PAGE_LOCK);
1974 err = -EAGAIN;
1975 goto out;
1976 }
1977
1978 for (i = 0; i < ret; i++) {
1979 if (page_ops & PAGE_SET_PRIVATE2)
1980 SetPagePrivate2(pages[i]);
1981
1982 if (locked_page && pages[i] == locked_page) {
1983 put_page(pages[i]);
1984 pages_locked++;
1985 continue;
1986 }
1987 if (page_ops & PAGE_CLEAR_DIRTY)
1988 clear_page_dirty_for_io(pages[i]);
1989 if (page_ops & PAGE_SET_WRITEBACK)
1990 set_page_writeback(pages[i]);
1991 if (page_ops & PAGE_SET_ERROR)
1992 SetPageError(pages[i]);
1993 if (page_ops & PAGE_END_WRITEBACK)
1994 end_page_writeback(pages[i]);
1995 if (page_ops & PAGE_UNLOCK)
1996 unlock_page(pages[i]);
1997 if (page_ops & PAGE_LOCK) {
1998 lock_page(pages[i]);
1999 if (!PageDirty(pages[i]) ||
2000 pages[i]->mapping != mapping) {
2001 unlock_page(pages[i]);
2002 for (; i < ret; i++)
2003 put_page(pages[i]);
2004 err = -EAGAIN;
2005 goto out;
2006 }
2007 }
2008 put_page(pages[i]);
2009 pages_locked++;
2010 }
2011 nr_pages -= ret;
2012 index += ret;
2013 cond_resched();
2014 }
2015out:
2016 if (err && index_ret)
2017 *index_ret = start_index + pages_locked - 1;
2018 return err;
2019}
2020
2021void extent_clear_unlock_delalloc(struct btrfs_inode *inode, u64 start, u64 end,
2022 struct page *locked_page,
2023 unsigned clear_bits,
2024 unsigned long page_ops)
2025{
2026 clear_extent_bit(&inode->io_tree, start, end, clear_bits, 1, 0, NULL);
2027
2028 __process_pages_contig(inode->vfs_inode.i_mapping, locked_page,
2029 start >> PAGE_SHIFT, end >> PAGE_SHIFT,
2030 page_ops, NULL);
2031}
2032
2033/*
2034 * count the number of bytes in the tree that have a given bit(s)
2035 * set. This can be fairly slow, except for EXTENT_DIRTY which is
2036 * cached. The total number found is returned.
2037 */
2038u64 count_range_bits(struct extent_io_tree *tree,
2039 u64 *start, u64 search_end, u64 max_bytes,
2040 unsigned bits, int contig)
2041{
2042 struct rb_node *node;
2043 struct extent_state *state;
2044 u64 cur_start = *start;
2045 u64 total_bytes = 0;
2046 u64 last = 0;
2047 int found = 0;
2048
2049 if (WARN_ON(search_end <= cur_start))
2050 return 0;
2051
2052 spin_lock(&tree->lock);
2053 if (cur_start == 0 && bits == EXTENT_DIRTY) {
2054 total_bytes = tree->dirty_bytes;
2055 goto out;
2056 }
2057 /*
2058 * this search will find all the extents that end after
2059 * our range starts.
2060 */
2061 node = tree_search(tree, cur_start);
2062 if (!node)
2063 goto out;
2064
2065 while (1) {
2066 state = rb_entry(node, struct extent_state, rb_node);
2067 if (state->start > search_end)
2068 break;
2069 if (contig && found && state->start > last + 1)
2070 break;
2071 if (state->end >= cur_start && (state->state & bits) == bits) {
2072 total_bytes += min(search_end, state->end) + 1 -
2073 max(cur_start, state->start);
2074 if (total_bytes >= max_bytes)
2075 break;
2076 if (!found) {
2077 *start = max(cur_start, state->start);
2078 found = 1;
2079 }
2080 last = state->end;
2081 } else if (contig && found) {
2082 break;
2083 }
2084 node = rb_next(node);
2085 if (!node)
2086 break;
2087 }
2088out:
2089 spin_unlock(&tree->lock);
2090 return total_bytes;
2091}
2092
2093/*
2094 * set the private field for a given byte offset in the tree. If there isn't
2095 * an extent_state there already, this does nothing.
2096 */
2097int set_state_failrec(struct extent_io_tree *tree, u64 start,
2098 struct io_failure_record *failrec)
2099{
2100 struct rb_node *node;
2101 struct extent_state *state;
2102 int ret = 0;
2103
2104 spin_lock(&tree->lock);
2105 /*
2106 * this search will find all the extents that end after
2107 * our range starts.
2108 */
2109 node = tree_search(tree, start);
2110 if (!node) {
2111 ret = -ENOENT;
2112 goto out;
2113 }
2114 state = rb_entry(node, struct extent_state, rb_node);
2115 if (state->start != start) {
2116 ret = -ENOENT;
2117 goto out;
2118 }
2119 state->failrec = failrec;
2120out:
2121 spin_unlock(&tree->lock);
2122 return ret;
2123}
2124
2125struct io_failure_record *get_state_failrec(struct extent_io_tree *tree, u64 start)
2126{
2127 struct rb_node *node;
2128 struct extent_state *state;
2129 struct io_failure_record *failrec;
2130
2131 spin_lock(&tree->lock);
2132 /*
2133 * this search will find all the extents that end after
2134 * our range starts.
2135 */
2136 node = tree_search(tree, start);
2137 if (!node) {
2138 failrec = ERR_PTR(-ENOENT);
2139 goto out;
2140 }
2141 state = rb_entry(node, struct extent_state, rb_node);
2142 if (state->start != start) {
2143 failrec = ERR_PTR(-ENOENT);
2144 goto out;
2145 }
2146
2147 failrec = state->failrec;
2148out:
2149 spin_unlock(&tree->lock);
2150 return failrec;
2151}
2152
2153/*
2154 * searches a range in the state tree for a given mask.
2155 * If 'filled' == 1, this returns 1 only if every extent in the tree
2156 * has the bits set. Otherwise, 1 is returned if any bit in the
2157 * range is found set.
2158 */
2159int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
2160 unsigned bits, int filled, struct extent_state *cached)
2161{
2162 struct extent_state *state = NULL;
2163 struct rb_node *node;
2164 int bitset = 0;
2165
2166 spin_lock(&tree->lock);
2167 if (cached && extent_state_in_tree(cached) && cached->start <= start &&
2168 cached->end > start)
2169 node = &cached->rb_node;
2170 else
2171 node = tree_search(tree, start);
2172 while (node && start <= end) {
2173 state = rb_entry(node, struct extent_state, rb_node);
2174
2175 if (filled && state->start > start) {
2176 bitset = 0;
2177 break;
2178 }
2179
2180 if (state->start > end)
2181 break;
2182
2183 if (state->state & bits) {
2184 bitset = 1;
2185 if (!filled)
2186 break;
2187 } else if (filled) {
2188 bitset = 0;
2189 break;
2190 }
2191
2192 if (state->end == (u64)-1)
2193 break;
2194
2195 start = state->end + 1;
2196 if (start > end)
2197 break;
2198 node = rb_next(node);
2199 if (!node) {
2200 if (filled)
2201 bitset = 0;
2202 break;
2203 }
2204 }
2205 spin_unlock(&tree->lock);
2206 return bitset;
2207}
2208
2209/*
2210 * helper function to set a given page up to date if all the
2211 * extents in the tree for that page are up to date
2212 */
2213static void check_page_uptodate(struct extent_io_tree *tree, struct page *page)
2214{
2215 u64 start = page_offset(page);
2216 u64 end = start + PAGE_SIZE - 1;
2217 if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
2218 SetPageUptodate(page);
2219}
2220
2221int free_io_failure(struct extent_io_tree *failure_tree,
2222 struct extent_io_tree *io_tree,
2223 struct io_failure_record *rec)
2224{
2225 int ret;
2226 int err = 0;
2227
2228 set_state_failrec(failure_tree, rec->start, NULL);
2229 ret = clear_extent_bits(failure_tree, rec->start,
2230 rec->start + rec->len - 1,
2231 EXTENT_LOCKED | EXTENT_DIRTY);
2232 if (ret)
2233 err = ret;
2234
2235 ret = clear_extent_bits(io_tree, rec->start,
2236 rec->start + rec->len - 1,
2237 EXTENT_DAMAGED);
2238 if (ret && !err)
2239 err = ret;
2240
2241 kfree(rec);
2242 return err;
2243}
2244
2245/*
2246 * this bypasses the standard btrfs submit functions deliberately, as
2247 * the standard behavior is to write all copies in a raid setup. here we only
2248 * want to write the one bad copy. so we do the mapping for ourselves and issue
2249 * submit_bio directly.
2250 * to avoid any synchronization issues, wait for the data after writing, which
2251 * actually prevents the read that triggered the error from finishing.
2252 * currently, there can be no more than two copies of every data bit. thus,
2253 * exactly one rewrite is required.
2254 */
2255int repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
2256 u64 length, u64 logical, struct page *page,
2257 unsigned int pg_offset, int mirror_num)
2258{
2259 struct bio *bio;
2260 struct btrfs_device *dev;
2261 u64 map_length = 0;
2262 u64 sector;
2263 struct btrfs_bio *bbio = NULL;
2264 int ret;
2265
2266 ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
2267 BUG_ON(!mirror_num);
2268
2269 bio = btrfs_io_bio_alloc(1);
2270 bio->bi_iter.bi_size = 0;
2271 map_length = length;
2272
2273 /*
2274 * Avoid races with device replace and make sure our bbio has devices
2275 * associated to its stripes that don't go away while we are doing the
2276 * read repair operation.
2277 */
2278 btrfs_bio_counter_inc_blocked(fs_info);
2279 if (btrfs_is_parity_mirror(fs_info, logical, length)) {
2280 /*
2281 * Note that we don't use BTRFS_MAP_WRITE because it's supposed
2282 * to update all raid stripes, but here we just want to correct
2283 * bad stripe, thus BTRFS_MAP_READ is abused to only get the bad
2284 * stripe's dev and sector.
2285 */
2286 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, logical,
2287 &map_length, &bbio, 0);
2288 if (ret) {
2289 btrfs_bio_counter_dec(fs_info);
2290 bio_put(bio);
2291 return -EIO;
2292 }
2293 ASSERT(bbio->mirror_num == 1);
2294 } else {
2295 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, logical,
2296 &map_length, &bbio, mirror_num);
2297 if (ret) {
2298 btrfs_bio_counter_dec(fs_info);
2299 bio_put(bio);
2300 return -EIO;
2301 }
2302 BUG_ON(mirror_num != bbio->mirror_num);
2303 }
2304
2305 sector = bbio->stripes[bbio->mirror_num - 1].physical >> 9;
2306 bio->bi_iter.bi_sector = sector;
2307 dev = bbio->stripes[bbio->mirror_num - 1].dev;
2308 btrfs_put_bbio(bbio);
2309 if (!dev || !dev->bdev ||
2310 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2311 btrfs_bio_counter_dec(fs_info);
2312 bio_put(bio);
2313 return -EIO;
2314 }
2315 bio_set_dev(bio, dev->bdev);
2316 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
2317 bio_add_page(bio, page, length, pg_offset);
2318
2319 if (btrfsic_submit_bio_wait(bio)) {
2320 /* try to remap that extent elsewhere? */
2321 btrfs_bio_counter_dec(fs_info);
2322 bio_put(bio);
2323 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
2324 return -EIO;
2325 }
2326
2327 btrfs_info_rl_in_rcu(fs_info,
2328 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
2329 ino, start,
2330 rcu_str_deref(dev->name), sector);
2331 btrfs_bio_counter_dec(fs_info);
2332 bio_put(bio);
2333 return 0;
2334}
2335
2336int btrfs_repair_eb_io_failure(const struct extent_buffer *eb, int mirror_num)
2337{
2338 struct btrfs_fs_info *fs_info = eb->fs_info;
2339 u64 start = eb->start;
2340 int i, num_pages = num_extent_pages(eb);
2341 int ret = 0;
2342
2343 if (sb_rdonly(fs_info->sb))
2344 return -EROFS;
2345
2346 for (i = 0; i < num_pages; i++) {
2347 struct page *p = eb->pages[i];
2348
2349 ret = repair_io_failure(fs_info, 0, start, PAGE_SIZE, start, p,
2350 start - page_offset(p), mirror_num);
2351 if (ret)
2352 break;
2353 start += PAGE_SIZE;
2354 }
2355
2356 return ret;
2357}
2358
2359/*
2360 * each time an IO finishes, we do a fast check in the IO failure tree
2361 * to see if we need to process or clean up an io_failure_record
2362 */
2363int clean_io_failure(struct btrfs_fs_info *fs_info,
2364 struct extent_io_tree *failure_tree,
2365 struct extent_io_tree *io_tree, u64 start,
2366 struct page *page, u64 ino, unsigned int pg_offset)
2367{
2368 u64 private;
2369 struct io_failure_record *failrec;
2370 struct extent_state *state;
2371 int num_copies;
2372 int ret;
2373
2374 private = 0;
2375 ret = count_range_bits(failure_tree, &private, (u64)-1, 1,
2376 EXTENT_DIRTY, 0);
2377 if (!ret)
2378 return 0;
2379
2380 failrec = get_state_failrec(failure_tree, start);
2381 if (IS_ERR(failrec))
2382 return 0;
2383
2384 BUG_ON(!failrec->this_mirror);
2385
2386 if (failrec->in_validation) {
2387 /* there was no real error, just free the record */
2388 btrfs_debug(fs_info,
2389 "clean_io_failure: freeing dummy error at %llu",
2390 failrec->start);
2391 goto out;
2392 }
2393 if (sb_rdonly(fs_info->sb))
2394 goto out;
2395
2396 spin_lock(&io_tree->lock);
2397 state = find_first_extent_bit_state(io_tree,
2398 failrec->start,
2399 EXTENT_LOCKED);
2400 spin_unlock(&io_tree->lock);
2401
2402 if (state && state->start <= failrec->start &&
2403 state->end >= failrec->start + failrec->len - 1) {
2404 num_copies = btrfs_num_copies(fs_info, failrec->logical,
2405 failrec->len);
2406 if (num_copies > 1) {
2407 repair_io_failure(fs_info, ino, start, failrec->len,
2408 failrec->logical, page, pg_offset,
2409 failrec->failed_mirror);
2410 }
2411 }
2412
2413out:
2414 free_io_failure(failure_tree, io_tree, failrec);
2415
2416 return 0;
2417}
2418
2419/*
2420 * Can be called when
2421 * - hold extent lock
2422 * - under ordered extent
2423 * - the inode is freeing
2424 */
2425void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
2426{
2427 struct extent_io_tree *failure_tree = &inode->io_failure_tree;
2428 struct io_failure_record *failrec;
2429 struct extent_state *state, *next;
2430
2431 if (RB_EMPTY_ROOT(&failure_tree->state))
2432 return;
2433
2434 spin_lock(&failure_tree->lock);
2435 state = find_first_extent_bit_state(failure_tree, start, EXTENT_DIRTY);
2436 while (state) {
2437 if (state->start > end)
2438 break;
2439
2440 ASSERT(state->end <= end);
2441
2442 next = next_state(state);
2443
2444 failrec = state->failrec;
2445 free_extent_state(state);
2446 kfree(failrec);
2447
2448 state = next;
2449 }
2450 spin_unlock(&failure_tree->lock);
2451}
2452
2453static struct io_failure_record *btrfs_get_io_failure_record(struct inode *inode,
2454 u64 start, u64 end)
2455{
2456 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2457 struct io_failure_record *failrec;
2458 struct extent_map *em;
2459 struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2460 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2461 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2462 int ret;
2463 u64 logical;
2464
2465 failrec = get_state_failrec(failure_tree, start);
2466 if (!IS_ERR(failrec)) {
2467 btrfs_debug(fs_info,
2468 "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu, validation=%d",
2469 failrec->logical, failrec->start, failrec->len,
2470 failrec->in_validation);
2471 /*
2472 * when data can be on disk more than twice, add to failrec here
2473 * (e.g. with a list for failed_mirror) to make
2474 * clean_io_failure() clean all those errors at once.
2475 */
2476
2477 return failrec;
2478 }
2479
2480 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
2481 if (!failrec)
2482 return ERR_PTR(-ENOMEM);
2483
2484 failrec->start = start;
2485 failrec->len = end - start + 1;
2486 failrec->this_mirror = 0;
2487 failrec->bio_flags = 0;
2488 failrec->in_validation = 0;
2489
2490 read_lock(&em_tree->lock);
2491 em = lookup_extent_mapping(em_tree, start, failrec->len);
2492 if (!em) {
2493 read_unlock(&em_tree->lock);
2494 kfree(failrec);
2495 return ERR_PTR(-EIO);
2496 }
2497
2498 if (em->start > start || em->start + em->len <= start) {
2499 free_extent_map(em);
2500 em = NULL;
2501 }
2502 read_unlock(&em_tree->lock);
2503 if (!em) {
2504 kfree(failrec);
2505 return ERR_PTR(-EIO);
2506 }
2507
2508 logical = start - em->start;
2509 logical = em->block_start + logical;
2510 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
2511 logical = em->block_start;
2512 failrec->bio_flags = EXTENT_BIO_COMPRESSED;
2513 extent_set_compress_type(&failrec->bio_flags, em->compress_type);
2514 }
2515
2516 btrfs_debug(fs_info,
2517 "Get IO Failure Record: (new) logical=%llu, start=%llu, len=%llu",
2518 logical, start, failrec->len);
2519
2520 failrec->logical = logical;
2521 free_extent_map(em);
2522
2523 /* Set the bits in the private failure tree */
2524 ret = set_extent_bits(failure_tree, start, end,
2525 EXTENT_LOCKED | EXTENT_DIRTY);
2526 if (ret >= 0) {
2527 ret = set_state_failrec(failure_tree, start, failrec);
2528 /* Set the bits in the inode's tree */
2529 ret = set_extent_bits(tree, start, end, EXTENT_DAMAGED);
2530 } else if (ret < 0) {
2531 kfree(failrec);
2532 return ERR_PTR(ret);
2533 }
2534
2535 return failrec;
2536}
2537
2538static bool btrfs_check_repairable(struct inode *inode, bool needs_validation,
2539 struct io_failure_record *failrec,
2540 int failed_mirror)
2541{
2542 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2543 int num_copies;
2544
2545 num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
2546 if (num_copies == 1) {
2547 /*
2548 * we only have a single copy of the data, so don't bother with
2549 * all the retry and error correction code that follows. no
2550 * matter what the error is, it is very likely to persist.
2551 */
2552 btrfs_debug(fs_info,
2553 "Check Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
2554 num_copies, failrec->this_mirror, failed_mirror);
2555 return false;
2556 }
2557
2558 /*
2559 * there are two premises:
2560 * a) deliver good data to the caller
2561 * b) correct the bad sectors on disk
2562 */
2563 if (needs_validation) {
2564 /*
2565 * to fulfill b), we need to know the exact failing sectors, as
2566 * we don't want to rewrite any more than the failed ones. thus,
2567 * we need separate read requests for the failed bio
2568 *
2569 * if the following BUG_ON triggers, our validation request got
2570 * merged. we need separate requests for our algorithm to work.
2571 */
2572 BUG_ON(failrec->in_validation);
2573 failrec->in_validation = 1;
2574 failrec->this_mirror = failed_mirror;
2575 } else {
2576 /*
2577 * we're ready to fulfill a) and b) alongside. get a good copy
2578 * of the failed sector and if we succeed, we have setup
2579 * everything for repair_io_failure to do the rest for us.
2580 */
2581 if (failrec->in_validation) {
2582 BUG_ON(failrec->this_mirror != failed_mirror);
2583 failrec->in_validation = 0;
2584 failrec->this_mirror = 0;
2585 }
2586 failrec->failed_mirror = failed_mirror;
2587 failrec->this_mirror++;
2588 if (failrec->this_mirror == failed_mirror)
2589 failrec->this_mirror++;
2590 }
2591
2592 if (failrec->this_mirror > num_copies) {
2593 btrfs_debug(fs_info,
2594 "Check Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
2595 num_copies, failrec->this_mirror, failed_mirror);
2596 return false;
2597 }
2598
2599 return true;
2600}
2601
2602static bool btrfs_io_needs_validation(struct inode *inode, struct bio *bio)
2603{
2604 u64 len = 0;
2605 const u32 blocksize = inode->i_sb->s_blocksize;
2606
2607 /*
2608 * If bi_status is BLK_STS_OK, then this was a checksum error, not an
2609 * I/O error. In this case, we already know exactly which sector was
2610 * bad, so we don't need to validate.
2611 */
2612 if (bio->bi_status == BLK_STS_OK)
2613 return false;
2614
2615 /*
2616 * We need to validate each sector individually if the failed I/O was
2617 * for multiple sectors.
2618 *
2619 * There are a few possible bios that can end up here:
2620 * 1. A buffered read bio, which is not cloned.
2621 * 2. A direct I/O read bio, which is cloned.
2622 * 3. A (buffered or direct) repair bio, which is not cloned.
2623 *
2624 * For cloned bios (case 2), we can get the size from
2625 * btrfs_io_bio->iter; for non-cloned bios (cases 1 and 3), we can get
2626 * it from the bvecs.
2627 */
2628 if (bio_flagged(bio, BIO_CLONED)) {
2629 if (btrfs_io_bio(bio)->iter.bi_size > blocksize)
2630 return true;
2631 } else {
2632 struct bio_vec *bvec;
2633 int i;
2634
2635 bio_for_each_bvec_all(bvec, bio, i) {
2636 len += bvec->bv_len;
2637 if (len > blocksize)
2638 return true;
2639 }
2640 }
2641 return false;
2642}
2643
2644blk_status_t btrfs_submit_read_repair(struct inode *inode,
2645 struct bio *failed_bio, u64 phy_offset,
2646 struct page *page, unsigned int pgoff,
2647 u64 start, u64 end, int failed_mirror,
2648 submit_bio_hook_t *submit_bio_hook)
2649{
2650 struct io_failure_record *failrec;
2651 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2652 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
2653 struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
2654 struct btrfs_io_bio *failed_io_bio = btrfs_io_bio(failed_bio);
2655 const int icsum = phy_offset >> inode->i_sb->s_blocksize_bits;
2656 bool need_validation;
2657 struct bio *repair_bio;
2658 struct btrfs_io_bio *repair_io_bio;
2659 blk_status_t status;
2660
2661 btrfs_debug(fs_info,
2662 "repair read error: read error at %llu", start);
2663
2664 BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
2665
2666 failrec = btrfs_get_io_failure_record(inode, start, end);
2667 if (IS_ERR(failrec))
2668 return errno_to_blk_status(PTR_ERR(failrec));
2669
2670 need_validation = btrfs_io_needs_validation(inode, failed_bio);
2671
2672 if (!btrfs_check_repairable(inode, need_validation, failrec,
2673 failed_mirror)) {
2674 free_io_failure(failure_tree, tree, failrec);
2675 return BLK_STS_IOERR;
2676 }
2677
2678 repair_bio = btrfs_io_bio_alloc(1);
2679 repair_io_bio = btrfs_io_bio(repair_bio);
2680 repair_bio->bi_opf = REQ_OP_READ;
2681 if (need_validation)
2682 repair_bio->bi_opf |= REQ_FAILFAST_DEV;
2683 repair_bio->bi_end_io = failed_bio->bi_end_io;
2684 repair_bio->bi_iter.bi_sector = failrec->logical >> 9;
2685 repair_bio->bi_private = failed_bio->bi_private;
2686
2687 if (failed_io_bio->csum) {
2688 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
2689
2690 repair_io_bio->csum = repair_io_bio->csum_inline;
2691 memcpy(repair_io_bio->csum,
2692 failed_io_bio->csum + csum_size * icsum, csum_size);
2693 }
2694
2695 bio_add_page(repair_bio, page, failrec->len, pgoff);
2696 repair_io_bio->logical = failrec->start;
2697 repair_io_bio->iter = repair_bio->bi_iter;
2698
2699 btrfs_debug(btrfs_sb(inode->i_sb),
2700"repair read error: submitting new read to mirror %d, in_validation=%d",
2701 failrec->this_mirror, failrec->in_validation);
2702
2703 status = submit_bio_hook(inode, repair_bio, failrec->this_mirror,
2704 failrec->bio_flags);
2705 if (status) {
2706 free_io_failure(failure_tree, tree, failrec);
2707 bio_put(repair_bio);
2708 }
2709 return status;
2710}
2711
2712/* lots and lots of room for performance fixes in the end_bio funcs */
2713
2714void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
2715{
2716 int uptodate = (err == 0);
2717 int ret = 0;
2718
2719 btrfs_writepage_endio_finish_ordered(page, start, end, uptodate);
2720
2721 if (!uptodate) {
2722 ClearPageUptodate(page);
2723 SetPageError(page);
2724 ret = err < 0 ? err : -EIO;
2725 mapping_set_error(page->mapping, ret);
2726 }
2727}
2728
2729/*
2730 * after a writepage IO is done, we need to:
2731 * clear the uptodate bits on error
2732 * clear the writeback bits in the extent tree for this IO
2733 * end_page_writeback if the page has no more pending IO
2734 *
2735 * Scheduling is not allowed, so the extent state tree is expected
2736 * to have one and only one object corresponding to this IO.
2737 */
2738static void end_bio_extent_writepage(struct bio *bio)
2739{
2740 int error = blk_status_to_errno(bio->bi_status);
2741 struct bio_vec *bvec;
2742 u64 start;
2743 u64 end;
2744 struct bvec_iter_all iter_all;
2745
2746 ASSERT(!bio_flagged(bio, BIO_CLONED));
2747 bio_for_each_segment_all(bvec, bio, iter_all) {
2748 struct page *page = bvec->bv_page;
2749 struct inode *inode = page->mapping->host;
2750 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2751
2752 /* We always issue full-page reads, but if some block
2753 * in a page fails to read, blk_update_request() will
2754 * advance bv_offset and adjust bv_len to compensate.
2755 * Print a warning for nonzero offsets, and an error
2756 * if they don't add up to a full page. */
2757 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2758 if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2759 btrfs_err(fs_info,
2760 "partial page write in btrfs with offset %u and length %u",
2761 bvec->bv_offset, bvec->bv_len);
2762 else
2763 btrfs_info(fs_info,
2764 "incomplete page write in btrfs with offset %u and length %u",
2765 bvec->bv_offset, bvec->bv_len);
2766 }
2767
2768 start = page_offset(page);
2769 end = start + bvec->bv_offset + bvec->bv_len - 1;
2770
2771 end_extent_writepage(page, error, start, end);
2772 end_page_writeback(page);
2773 }
2774
2775 bio_put(bio);
2776}
2777
2778static void
2779endio_readpage_release_extent(struct extent_io_tree *tree, u64 start, u64 len,
2780 int uptodate)
2781{
2782 struct extent_state *cached = NULL;
2783 u64 end = start + len - 1;
2784
2785 if (uptodate && tree->track_uptodate)
2786 set_extent_uptodate(tree, start, end, &cached, GFP_ATOMIC);
2787 unlock_extent_cached_atomic(tree, start, end, &cached);
2788}
2789
2790/*
2791 * after a readpage IO is done, we need to:
2792 * clear the uptodate bits on error
2793 * set the uptodate bits if things worked
2794 * set the page up to date if all extents in the tree are uptodate
2795 * clear the lock bit in the extent tree
2796 * unlock the page if there are no other extents locked for it
2797 *
2798 * Scheduling is not allowed, so the extent state tree is expected
2799 * to have one and only one object corresponding to this IO.
2800 */
2801static void end_bio_extent_readpage(struct bio *bio)
2802{
2803 struct bio_vec *bvec;
2804 int uptodate = !bio->bi_status;
2805 struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
2806 struct extent_io_tree *tree, *failure_tree;
2807 u64 offset = 0;
2808 u64 start;
2809 u64 end;
2810 u64 len;
2811 u64 extent_start = 0;
2812 u64 extent_len = 0;
2813 int mirror;
2814 int ret;
2815 struct bvec_iter_all iter_all;
2816
2817 ASSERT(!bio_flagged(bio, BIO_CLONED));
2818 bio_for_each_segment_all(bvec, bio, iter_all) {
2819 struct page *page = bvec->bv_page;
2820 struct inode *inode = page->mapping->host;
2821 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2822 bool data_inode = btrfs_ino(BTRFS_I(inode))
2823 != BTRFS_BTREE_INODE_OBJECTID;
2824
2825 btrfs_debug(fs_info,
2826 "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
2827 (u64)bio->bi_iter.bi_sector, bio->bi_status,
2828 io_bio->mirror_num);
2829 tree = &BTRFS_I(inode)->io_tree;
2830 failure_tree = &BTRFS_I(inode)->io_failure_tree;
2831
2832 /* We always issue full-page reads, but if some block
2833 * in a page fails to read, blk_update_request() will
2834 * advance bv_offset and adjust bv_len to compensate.
2835 * Print a warning for nonzero offsets, and an error
2836 * if they don't add up to a full page. */
2837 if (bvec->bv_offset || bvec->bv_len != PAGE_SIZE) {
2838 if (bvec->bv_offset + bvec->bv_len != PAGE_SIZE)
2839 btrfs_err(fs_info,
2840 "partial page read in btrfs with offset %u and length %u",
2841 bvec->bv_offset, bvec->bv_len);
2842 else
2843 btrfs_info(fs_info,
2844 "incomplete page read in btrfs with offset %u and length %u",
2845 bvec->bv_offset, bvec->bv_len);
2846 }
2847
2848 start = page_offset(page);
2849 end = start + bvec->bv_offset + bvec->bv_len - 1;
2850 len = bvec->bv_len;
2851
2852 mirror = io_bio->mirror_num;
2853 if (likely(uptodate)) {
2854 ret = tree->ops->readpage_end_io_hook(io_bio, offset,
2855 page, start, end,
2856 mirror);
2857 if (ret)
2858 uptodate = 0;
2859 else
2860 clean_io_failure(BTRFS_I(inode)->root->fs_info,
2861 failure_tree, tree, start,
2862 page,
2863 btrfs_ino(BTRFS_I(inode)), 0);
2864 }
2865
2866 if (likely(uptodate))
2867 goto readpage_ok;
2868
2869 if (data_inode) {
2870
2871 /*
2872 * The generic bio_readpage_error handles errors the
2873 * following way: If possible, new read requests are
2874 * created and submitted and will end up in
2875 * end_bio_extent_readpage as well (if we're lucky,
2876 * not in the !uptodate case). In that case it returns
2877 * 0 and we just go on with the next page in our bio.
2878 * If it can't handle the error it will return -EIO and
2879 * we remain responsible for that page.
2880 */
2881 if (!btrfs_submit_read_repair(inode, bio, offset, page,
2882 start - page_offset(page),
2883 start, end, mirror,
2884 tree->ops->submit_bio_hook)) {
2885 uptodate = !bio->bi_status;
2886 offset += len;
2887 continue;
2888 }
2889 } else {
2890 struct extent_buffer *eb;
2891
2892 eb = (struct extent_buffer *)page->private;
2893 set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
2894 eb->read_mirror = mirror;
2895 atomic_dec(&eb->io_pages);
2896 if (test_and_clear_bit(EXTENT_BUFFER_READAHEAD,
2897 &eb->bflags))
2898 btree_readahead_hook(eb, -EIO);
2899 }
2900readpage_ok:
2901 if (likely(uptodate)) {
2902 loff_t i_size = i_size_read(inode);
2903 pgoff_t end_index = i_size >> PAGE_SHIFT;
2904 unsigned off;
2905
2906 /* Zero out the end if this page straddles i_size */
2907 off = offset_in_page(i_size);
2908 if (page->index == end_index && off)
2909 zero_user_segment(page, off, PAGE_SIZE);
2910 SetPageUptodate(page);
2911 } else {
2912 ClearPageUptodate(page);
2913 SetPageError(page);
2914 }
2915 unlock_page(page);
2916 offset += len;
2917
2918 if (unlikely(!uptodate)) {
2919 if (extent_len) {
2920 endio_readpage_release_extent(tree,
2921 extent_start,
2922 extent_len, 1);
2923 extent_start = 0;
2924 extent_len = 0;
2925 }
2926 endio_readpage_release_extent(tree, start,
2927 end - start + 1, 0);
2928 } else if (!extent_len) {
2929 extent_start = start;
2930 extent_len = end + 1 - start;
2931 } else if (extent_start + extent_len == start) {
2932 extent_len += end + 1 - start;
2933 } else {
2934 endio_readpage_release_extent(tree, extent_start,
2935 extent_len, uptodate);
2936 extent_start = start;
2937 extent_len = end + 1 - start;
2938 }
2939 }
2940
2941 if (extent_len)
2942 endio_readpage_release_extent(tree, extent_start, extent_len,
2943 uptodate);
2944 btrfs_io_bio_free_csum(io_bio);
2945 bio_put(bio);
2946}
2947
2948/*
2949 * Initialize the members up to but not including 'bio'. Use after allocating a
2950 * new bio by bio_alloc_bioset as it does not initialize the bytes outside of
2951 * 'bio' because use of __GFP_ZERO is not supported.
2952 */
2953static inline void btrfs_io_bio_init(struct btrfs_io_bio *btrfs_bio)
2954{
2955 memset(btrfs_bio, 0, offsetof(struct btrfs_io_bio, bio));
2956}
2957
2958/*
2959 * The following helpers allocate a bio. As it's backed by a bioset, it'll
2960 * never fail. We're returning a bio right now but you can call btrfs_io_bio
2961 * for the appropriate container_of magic
2962 */
2963struct bio *btrfs_bio_alloc(u64 first_byte)
2964{
2965 struct bio *bio;
2966
2967 bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, &btrfs_bioset);
2968 bio->bi_iter.bi_sector = first_byte >> 9;
2969 btrfs_io_bio_init(btrfs_io_bio(bio));
2970 return bio;
2971}
2972
2973struct bio *btrfs_bio_clone(struct bio *bio)
2974{
2975 struct btrfs_io_bio *btrfs_bio;
2976 struct bio *new;
2977
2978 /* Bio allocation backed by a bioset does not fail */
2979 new = bio_clone_fast(bio, GFP_NOFS, &btrfs_bioset);
2980 btrfs_bio = btrfs_io_bio(new);
2981 btrfs_io_bio_init(btrfs_bio);
2982 btrfs_bio->iter = bio->bi_iter;
2983 return new;
2984}
2985
2986struct bio *btrfs_io_bio_alloc(unsigned int nr_iovecs)
2987{
2988 struct bio *bio;
2989
2990 /* Bio allocation backed by a bioset does not fail */
2991 bio = bio_alloc_bioset(GFP_NOFS, nr_iovecs, &btrfs_bioset);
2992 btrfs_io_bio_init(btrfs_io_bio(bio));
2993 return bio;
2994}
2995
2996struct bio *btrfs_bio_clone_partial(struct bio *orig, int offset, int size)
2997{
2998 struct bio *bio;
2999 struct btrfs_io_bio *btrfs_bio;
3000
3001 /* this will never fail when it's backed by a bioset */
3002 bio = bio_clone_fast(orig, GFP_NOFS, &btrfs_bioset);
3003 ASSERT(bio);
3004
3005 btrfs_bio = btrfs_io_bio(bio);
3006 btrfs_io_bio_init(btrfs_bio);
3007
3008 bio_trim(bio, offset >> 9, size >> 9);
3009 btrfs_bio->iter = bio->bi_iter;
3010 return bio;
3011}
3012
3013/*
3014 * @opf: bio REQ_OP_* and REQ_* flags as one value
3015 * @wbc: optional writeback control for io accounting
3016 * @page: page to add to the bio
3017 * @pg_offset: offset of the new bio or to check whether we are adding
3018 * a contiguous page to the previous one
3019 * @size: portion of page that we want to write
3020 * @offset: starting offset in the page
3021 * @bio_ret: must be valid pointer, newly allocated bio will be stored there
3022 * @end_io_func: end_io callback for new bio
3023 * @mirror_num: desired mirror to read/write
3024 * @prev_bio_flags: flags of previous bio to see if we can merge the current one
3025 * @bio_flags: flags of the current bio to see if we can merge them
3026 */
3027static int submit_extent_page(unsigned int opf,
3028 struct writeback_control *wbc,
3029 struct page *page, u64 offset,
3030 size_t size, unsigned long pg_offset,
3031 struct bio **bio_ret,
3032 bio_end_io_t end_io_func,
3033 int mirror_num,
3034 unsigned long prev_bio_flags,
3035 unsigned long bio_flags,
3036 bool force_bio_submit)
3037{
3038 int ret = 0;
3039 struct bio *bio;
3040 size_t page_size = min_t(size_t, size, PAGE_SIZE);
3041 sector_t sector = offset >> 9;
3042 struct extent_io_tree *tree = &BTRFS_I(page->mapping->host)->io_tree;
3043
3044 ASSERT(bio_ret);
3045
3046 if (*bio_ret) {
3047 bool contig;
3048 bool can_merge = true;
3049
3050 bio = *bio_ret;
3051 if (prev_bio_flags & EXTENT_BIO_COMPRESSED)
3052 contig = bio->bi_iter.bi_sector == sector;
3053 else
3054 contig = bio_end_sector(bio) == sector;
3055
3056 ASSERT(tree->ops);
3057 if (btrfs_bio_fits_in_stripe(page, page_size, bio, bio_flags))
3058 can_merge = false;
3059
3060 if (prev_bio_flags != bio_flags || !contig || !can_merge ||
3061 force_bio_submit ||
3062 bio_add_page(bio, page, page_size, pg_offset) < page_size) {
3063 ret = submit_one_bio(bio, mirror_num, prev_bio_flags);
3064 if (ret < 0) {
3065 *bio_ret = NULL;
3066 return ret;
3067 }
3068 bio = NULL;
3069 } else {
3070 if (wbc)
3071 wbc_account_cgroup_owner(wbc, page, page_size);
3072 return 0;
3073 }
3074 }
3075
3076 bio = btrfs_bio_alloc(offset);
3077 bio_add_page(bio, page, page_size, pg_offset);
3078 bio->bi_end_io = end_io_func;
3079 bio->bi_private = tree;
3080 bio->bi_write_hint = page->mapping->host->i_write_hint;
3081 bio->bi_opf = opf;
3082 if (wbc) {
3083 struct block_device *bdev;
3084
3085 bdev = BTRFS_I(page->mapping->host)->root->fs_info->fs_devices->latest_bdev;
3086 bio_set_dev(bio, bdev);
3087 wbc_init_bio(wbc, bio);
3088 wbc_account_cgroup_owner(wbc, page, page_size);
3089 }
3090
3091 *bio_ret = bio;
3092
3093 return ret;
3094}
3095
3096static void attach_extent_buffer_page(struct extent_buffer *eb,
3097 struct page *page)
3098{
3099 if (!PagePrivate(page))
3100 attach_page_private(page, eb);
3101 else
3102 WARN_ON(page->private != (unsigned long)eb);
3103}
3104
3105void set_page_extent_mapped(struct page *page)
3106{
3107 if (!PagePrivate(page))
3108 attach_page_private(page, (void *)EXTENT_PAGE_PRIVATE);
3109}
3110
3111static struct extent_map *
3112__get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
3113 u64 start, u64 len, get_extent_t *get_extent,
3114 struct extent_map **em_cached)
3115{
3116 struct extent_map *em;
3117
3118 if (em_cached && *em_cached) {
3119 em = *em_cached;
3120 if (extent_map_in_tree(em) && start >= em->start &&
3121 start < extent_map_end(em)) {
3122 refcount_inc(&em->refs);
3123 return em;
3124 }
3125
3126 free_extent_map(em);
3127 *em_cached = NULL;
3128 }
3129
3130 em = get_extent(BTRFS_I(inode), page, pg_offset, start, len);
3131 if (em_cached && !IS_ERR_OR_NULL(em)) {
3132 BUG_ON(*em_cached);
3133 refcount_inc(&em->refs);
3134 *em_cached = em;
3135 }
3136 return em;
3137}
3138/*
3139 * basic readpage implementation. Locked extent state structs are inserted
3140 * into the tree that are removed when the IO is done (by the end_io
3141 * handlers)
3142 * XXX JDM: This needs looking at to ensure proper page locking
3143 * return 0 on success, otherwise return error
3144 */
3145static int __do_readpage(struct page *page,
3146 get_extent_t *get_extent,
3147 struct extent_map **em_cached,
3148 struct bio **bio, int mirror_num,
3149 unsigned long *bio_flags, unsigned int read_flags,
3150 u64 *prev_em_start)
3151{
3152 struct inode *inode = page->mapping->host;
3153 u64 start = page_offset(page);
3154 const u64 end = start + PAGE_SIZE - 1;
3155 u64 cur = start;
3156 u64 extent_offset;
3157 u64 last_byte = i_size_read(inode);
3158 u64 block_start;
3159 u64 cur_end;
3160 struct extent_map *em;
3161 int ret = 0;
3162 int nr = 0;
3163 size_t pg_offset = 0;
3164 size_t iosize;
3165 size_t disk_io_size;
3166 size_t blocksize = inode->i_sb->s_blocksize;
3167 unsigned long this_bio_flag = 0;
3168 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
3169
3170 set_page_extent_mapped(page);
3171
3172 if (!PageUptodate(page)) {
3173 if (cleancache_get_page(page) == 0) {
3174 BUG_ON(blocksize != PAGE_SIZE);
3175 unlock_extent(tree, start, end);
3176 goto out;
3177 }
3178 }
3179
3180 if (page->index == last_byte >> PAGE_SHIFT) {
3181 char *userpage;
3182 size_t zero_offset = offset_in_page(last_byte);
3183
3184 if (zero_offset) {
3185 iosize = PAGE_SIZE - zero_offset;
3186 userpage = kmap_atomic(page);
3187 memset(userpage + zero_offset, 0, iosize);
3188 flush_dcache_page(page);
3189 kunmap_atomic(userpage);
3190 }
3191 }
3192 while (cur <= end) {
3193 bool force_bio_submit = false;
3194 u64 offset;
3195
3196 if (cur >= last_byte) {
3197 char *userpage;
3198 struct extent_state *cached = NULL;
3199
3200 iosize = PAGE_SIZE - pg_offset;
3201 userpage = kmap_atomic(page);
3202 memset(userpage + pg_offset, 0, iosize);
3203 flush_dcache_page(page);
3204 kunmap_atomic(userpage);
3205 set_extent_uptodate(tree, cur, cur + iosize - 1,
3206 &cached, GFP_NOFS);
3207 unlock_extent_cached(tree, cur,
3208 cur + iosize - 1, &cached);
3209 break;
3210 }
3211 em = __get_extent_map(inode, page, pg_offset, cur,
3212 end - cur + 1, get_extent, em_cached);
3213 if (IS_ERR_OR_NULL(em)) {
3214 SetPageError(page);
3215 unlock_extent(tree, cur, end);
3216 break;
3217 }
3218 extent_offset = cur - em->start;
3219 BUG_ON(extent_map_end(em) <= cur);
3220 BUG_ON(end < cur);
3221
3222 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
3223 this_bio_flag |= EXTENT_BIO_COMPRESSED;
3224 extent_set_compress_type(&this_bio_flag,
3225 em->compress_type);
3226 }
3227
3228 iosize = min(extent_map_end(em) - cur, end - cur + 1);
3229 cur_end = min(extent_map_end(em) - 1, end);
3230 iosize = ALIGN(iosize, blocksize);
3231 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
3232 disk_io_size = em->block_len;
3233 offset = em->block_start;
3234 } else {
3235 offset = em->block_start + extent_offset;
3236 disk_io_size = iosize;
3237 }
3238 block_start = em->block_start;
3239 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3240 block_start = EXTENT_MAP_HOLE;
3241
3242 /*
3243 * If we have a file range that points to a compressed extent
3244 * and it's followed by a consecutive file range that points to
3245 * to the same compressed extent (possibly with a different
3246 * offset and/or length, so it either points to the whole extent
3247 * or only part of it), we must make sure we do not submit a
3248 * single bio to populate the pages for the 2 ranges because
3249 * this makes the compressed extent read zero out the pages
3250 * belonging to the 2nd range. Imagine the following scenario:
3251 *
3252 * File layout
3253 * [0 - 8K] [8K - 24K]
3254 * | |
3255 * | |
3256 * points to extent X, points to extent X,
3257 * offset 4K, length of 8K offset 0, length 16K
3258 *
3259 * [extent X, compressed length = 4K uncompressed length = 16K]
3260 *
3261 * If the bio to read the compressed extent covers both ranges,
3262 * it will decompress extent X into the pages belonging to the
3263 * first range and then it will stop, zeroing out the remaining
3264 * pages that belong to the other range that points to extent X.
3265 * So here we make sure we submit 2 bios, one for the first
3266 * range and another one for the third range. Both will target
3267 * the same physical extent from disk, but we can't currently
3268 * make the compressed bio endio callback populate the pages
3269 * for both ranges because each compressed bio is tightly
3270 * coupled with a single extent map, and each range can have
3271 * an extent map with a different offset value relative to the
3272 * uncompressed data of our extent and different lengths. This
3273 * is a corner case so we prioritize correctness over
3274 * non-optimal behavior (submitting 2 bios for the same extent).
3275 */
3276 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
3277 prev_em_start && *prev_em_start != (u64)-1 &&
3278 *prev_em_start != em->start)
3279 force_bio_submit = true;
3280
3281 if (prev_em_start)
3282 *prev_em_start = em->start;
3283
3284 free_extent_map(em);
3285 em = NULL;
3286
3287 /* we've found a hole, just zero and go on */
3288 if (block_start == EXTENT_MAP_HOLE) {
3289 char *userpage;
3290 struct extent_state *cached = NULL;
3291
3292 userpage = kmap_atomic(page);
3293 memset(userpage + pg_offset, 0, iosize);
3294 flush_dcache_page(page);
3295 kunmap_atomic(userpage);
3296
3297 set_extent_uptodate(tree, cur, cur + iosize - 1,
3298 &cached, GFP_NOFS);
3299 unlock_extent_cached(tree, cur,
3300 cur + iosize - 1, &cached);
3301 cur = cur + iosize;
3302 pg_offset += iosize;
3303 continue;
3304 }
3305 /* the get_extent function already copied into the page */
3306 if (test_range_bit(tree, cur, cur_end,
3307 EXTENT_UPTODATE, 1, NULL)) {
3308 check_page_uptodate(tree, page);
3309 unlock_extent(tree, cur, cur + iosize - 1);
3310 cur = cur + iosize;
3311 pg_offset += iosize;
3312 continue;
3313 }
3314 /* we have an inline extent but it didn't get marked up
3315 * to date. Error out
3316 */
3317 if (block_start == EXTENT_MAP_INLINE) {
3318 SetPageError(page);
3319 unlock_extent(tree, cur, cur + iosize - 1);
3320 cur = cur + iosize;
3321 pg_offset += iosize;
3322 continue;
3323 }
3324
3325 ret = submit_extent_page(REQ_OP_READ | read_flags, NULL,
3326 page, offset, disk_io_size,
3327 pg_offset, bio,
3328 end_bio_extent_readpage, mirror_num,
3329 *bio_flags,
3330 this_bio_flag,
3331 force_bio_submit);
3332 if (!ret) {
3333 nr++;
3334 *bio_flags = this_bio_flag;
3335 } else {
3336 SetPageError(page);
3337 unlock_extent(tree, cur, cur + iosize - 1);
3338 goto out;
3339 }
3340 cur = cur + iosize;
3341 pg_offset += iosize;
3342 }
3343out:
3344 if (!nr) {
3345 if (!PageError(page))
3346 SetPageUptodate(page);
3347 unlock_page(page);
3348 }
3349 return ret;
3350}
3351
3352static inline void contiguous_readpages(struct page *pages[], int nr_pages,
3353 u64 start, u64 end,
3354 struct extent_map **em_cached,
3355 struct bio **bio,
3356 unsigned long *bio_flags,
3357 u64 *prev_em_start)
3358{
3359 struct btrfs_inode *inode = BTRFS_I(pages[0]->mapping->host);
3360 int index;
3361
3362 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
3363
3364 for (index = 0; index < nr_pages; index++) {
3365 __do_readpage(pages[index], btrfs_get_extent, em_cached,
3366 bio, 0, bio_flags, REQ_RAHEAD, prev_em_start);
3367 put_page(pages[index]);
3368 }
3369}
3370
3371static int __extent_read_full_page(struct page *page,
3372 get_extent_t *get_extent,
3373 struct bio **bio, int mirror_num,
3374 unsigned long *bio_flags,
3375 unsigned int read_flags)
3376{
3377 struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
3378 u64 start = page_offset(page);
3379 u64 end = start + PAGE_SIZE - 1;
3380 int ret;
3381
3382 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
3383
3384 ret = __do_readpage(page, get_extent, NULL, bio, mirror_num,
3385 bio_flags, read_flags, NULL);
3386 return ret;
3387}
3388
3389int extent_read_full_page(struct page *page, get_extent_t *get_extent,
3390 int mirror_num)
3391{
3392 struct bio *bio = NULL;
3393 unsigned long bio_flags = 0;
3394 int ret;
3395
3396 ret = __extent_read_full_page(page, get_extent, &bio, mirror_num,
3397 &bio_flags, 0);
3398 if (bio)
3399 ret = submit_one_bio(bio, mirror_num, bio_flags);
3400 return ret;
3401}
3402
3403static void update_nr_written(struct writeback_control *wbc,
3404 unsigned long nr_written)
3405{
3406 wbc->nr_to_write -= nr_written;
3407}
3408
3409/*
3410 * helper for __extent_writepage, doing all of the delayed allocation setup.
3411 *
3412 * This returns 1 if btrfs_run_delalloc_range function did all the work required
3413 * to write the page (copy into inline extent). In this case the IO has
3414 * been started and the page is already unlocked.
3415 *
3416 * This returns 0 if all went well (page still locked)
3417 * This returns < 0 if there were errors (page still locked)
3418 */
3419static noinline_for_stack int writepage_delalloc(struct btrfs_inode *inode,
3420 struct page *page, struct writeback_control *wbc,
3421 u64 delalloc_start, unsigned long *nr_written)
3422{
3423 u64 page_end = delalloc_start + PAGE_SIZE - 1;
3424 bool found;
3425 u64 delalloc_to_write = 0;
3426 u64 delalloc_end = 0;
3427 int ret;
3428 int page_started = 0;
3429
3430
3431 while (delalloc_end < page_end) {
3432 found = find_lock_delalloc_range(&inode->vfs_inode, page,
3433 &delalloc_start,
3434 &delalloc_end);
3435 if (!found) {
3436 delalloc_start = delalloc_end + 1;
3437 continue;
3438 }
3439 ret = btrfs_run_delalloc_range(inode, page, delalloc_start,
3440 delalloc_end, &page_started, nr_written, wbc);
3441 if (ret) {
3442 SetPageError(page);
3443 /*
3444 * btrfs_run_delalloc_range should return < 0 for error
3445 * but just in case, we use > 0 here meaning the IO is
3446 * started, so we don't want to return > 0 unless
3447 * things are going well.
3448 */
3449 return ret < 0 ? ret : -EIO;
3450 }
3451 /*
3452 * delalloc_end is already one less than the total length, so
3453 * we don't subtract one from PAGE_SIZE
3454 */
3455 delalloc_to_write += (delalloc_end - delalloc_start +
3456 PAGE_SIZE) >> PAGE_SHIFT;
3457 delalloc_start = delalloc_end + 1;
3458 }
3459 if (wbc->nr_to_write < delalloc_to_write) {
3460 int thresh = 8192;
3461
3462 if (delalloc_to_write < thresh * 2)
3463 thresh = delalloc_to_write;
3464 wbc->nr_to_write = min_t(u64, delalloc_to_write,
3465 thresh);
3466 }
3467
3468 /* did the fill delalloc function already unlock and start
3469 * the IO?
3470 */
3471 if (page_started) {
3472 /*
3473 * we've unlocked the page, so we can't update
3474 * the mapping's writeback index, just update
3475 * nr_to_write.
3476 */
3477 wbc->nr_to_write -= *nr_written;
3478 return 1;
3479 }
3480
3481 return 0;
3482}
3483
3484/*
3485 * helper for __extent_writepage. This calls the writepage start hooks,
3486 * and does the loop to map the page into extents and bios.
3487 *
3488 * We return 1 if the IO is started and the page is unlocked,
3489 * 0 if all went well (page still locked)
3490 * < 0 if there were errors (page still locked)
3491 */
3492static noinline_for_stack int __extent_writepage_io(struct btrfs_inode *inode,
3493 struct page *page,
3494 struct writeback_control *wbc,
3495 struct extent_page_data *epd,
3496 loff_t i_size,
3497 unsigned long nr_written,
3498 int *nr_ret)
3499{
3500 struct extent_io_tree *tree = &inode->io_tree;
3501 u64 start = page_offset(page);
3502 u64 page_end = start + PAGE_SIZE - 1;
3503 u64 end;
3504 u64 cur = start;
3505 u64 extent_offset;
3506 u64 block_start;
3507 u64 iosize;
3508 struct extent_map *em;
3509 size_t pg_offset = 0;
3510 size_t blocksize;
3511 int ret = 0;
3512 int nr = 0;
3513 const unsigned int write_flags = wbc_to_write_flags(wbc);
3514 bool compressed;
3515
3516 ret = btrfs_writepage_cow_fixup(page, start, page_end);
3517 if (ret) {
3518 /* Fixup worker will requeue */
3519 redirty_page_for_writepage(wbc, page);
3520 update_nr_written(wbc, nr_written);
3521 unlock_page(page);
3522 return 1;
3523 }
3524
3525 /*
3526 * we don't want to touch the inode after unlocking the page,
3527 * so we update the mapping writeback index now
3528 */
3529 update_nr_written(wbc, nr_written + 1);
3530
3531 end = page_end;
3532 blocksize = inode->vfs_inode.i_sb->s_blocksize;
3533
3534 while (cur <= end) {
3535 u64 em_end;
3536 u64 offset;
3537
3538 if (cur >= i_size) {
3539 btrfs_writepage_endio_finish_ordered(page, cur,
3540 page_end, 1);
3541 break;
3542 }
3543 em = btrfs_get_extent(inode, NULL, 0, cur, end - cur + 1);
3544 if (IS_ERR_OR_NULL(em)) {
3545 SetPageError(page);
3546 ret = PTR_ERR_OR_ZERO(em);
3547 break;
3548 }
3549
3550 extent_offset = cur - em->start;
3551 em_end = extent_map_end(em);
3552 BUG_ON(em_end <= cur);
3553 BUG_ON(end < cur);
3554 iosize = min(em_end - cur, end - cur + 1);
3555 iosize = ALIGN(iosize, blocksize);
3556 offset = em->block_start + extent_offset;
3557 block_start = em->block_start;
3558 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
3559 free_extent_map(em);
3560 em = NULL;
3561
3562 /*
3563 * compressed and inline extents are written through other
3564 * paths in the FS
3565 */
3566 if (compressed || block_start == EXTENT_MAP_HOLE ||
3567 block_start == EXTENT_MAP_INLINE) {
3568 if (compressed)
3569 nr++;
3570 else
3571 btrfs_writepage_endio_finish_ordered(page, cur,
3572 cur + iosize - 1, 1);
3573 cur += iosize;
3574 pg_offset += iosize;
3575 continue;
3576 }
3577
3578 btrfs_set_range_writeback(tree, cur, cur + iosize - 1);
3579 if (!PageWriteback(page)) {
3580 btrfs_err(inode->root->fs_info,
3581 "page %lu not writeback, cur %llu end %llu",
3582 page->index, cur, end);
3583 }
3584
3585 ret = submit_extent_page(REQ_OP_WRITE | write_flags, wbc,
3586 page, offset, iosize, pg_offset,
3587 &epd->bio,
3588 end_bio_extent_writepage,
3589 0, 0, 0, false);
3590 if (ret) {
3591 SetPageError(page);
3592 if (PageWriteback(page))
3593 end_page_writeback(page);
3594 }
3595
3596 cur = cur + iosize;
3597 pg_offset += iosize;
3598 nr++;
3599 }
3600 *nr_ret = nr;
3601 return ret;
3602}
3603
3604/*
3605 * the writepage semantics are similar to regular writepage. extent
3606 * records are inserted to lock ranges in the tree, and as dirty areas
3607 * are found, they are marked writeback. Then the lock bits are removed
3608 * and the end_io handler clears the writeback ranges
3609 *
3610 * Return 0 if everything goes well.
3611 * Return <0 for error.
3612 */
3613static int __extent_writepage(struct page *page, struct writeback_control *wbc,
3614 struct extent_page_data *epd)
3615{
3616 struct inode *inode = page->mapping->host;
3617 u64 start = page_offset(page);
3618 u64 page_end = start + PAGE_SIZE - 1;
3619 int ret;
3620 int nr = 0;
3621 size_t pg_offset;
3622 loff_t i_size = i_size_read(inode);
3623 unsigned long end_index = i_size >> PAGE_SHIFT;
3624 unsigned long nr_written = 0;
3625
3626 trace___extent_writepage(page, inode, wbc);
3627
3628 WARN_ON(!PageLocked(page));
3629
3630 ClearPageError(page);
3631
3632 pg_offset = offset_in_page(i_size);
3633 if (page->index > end_index ||
3634 (page->index == end_index && !pg_offset)) {
3635 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
3636 unlock_page(page);
3637 return 0;
3638 }
3639
3640 if (page->index == end_index) {
3641 char *userpage;
3642
3643 userpage = kmap_atomic(page);
3644 memset(userpage + pg_offset, 0,
3645 PAGE_SIZE - pg_offset);
3646 kunmap_atomic(userpage);
3647 flush_dcache_page(page);
3648 }
3649
3650 set_page_extent_mapped(page);
3651
3652 if (!epd->extent_locked) {
3653 ret = writepage_delalloc(BTRFS_I(inode), page, wbc, start,
3654 &nr_written);
3655 if (ret == 1)
3656 return 0;
3657 if (ret)
3658 goto done;
3659 }
3660
3661 ret = __extent_writepage_io(BTRFS_I(inode), page, wbc, epd, i_size,
3662 nr_written, &nr);
3663 if (ret == 1)
3664 return 0;
3665
3666done:
3667 if (nr == 0) {
3668 /* make sure the mapping tag for page dirty gets cleared */
3669 set_page_writeback(page);
3670 end_page_writeback(page);
3671 }
3672 if (PageError(page)) {
3673 ret = ret < 0 ? ret : -EIO;
3674 end_extent_writepage(page, ret, start, page_end);
3675 }
3676 unlock_page(page);
3677 ASSERT(ret <= 0);
3678 return ret;
3679}
3680
3681void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
3682{
3683 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
3684 TASK_UNINTERRUPTIBLE);
3685}
3686
3687static void end_extent_buffer_writeback(struct extent_buffer *eb)
3688{
3689 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3690 smp_mb__after_atomic();
3691 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
3692}
3693
3694/*
3695 * Lock eb pages and flush the bio if we can't the locks
3696 *
3697 * Return 0 if nothing went wrong
3698 * Return >0 is same as 0, except bio is not submitted
3699 * Return <0 if something went wrong, no page is locked
3700 */
3701static noinline_for_stack int lock_extent_buffer_for_io(struct extent_buffer *eb,
3702 struct extent_page_data *epd)
3703{
3704 struct btrfs_fs_info *fs_info = eb->fs_info;
3705 int i, num_pages, failed_page_nr;
3706 int flush = 0;
3707 int ret = 0;
3708
3709 if (!btrfs_try_tree_write_lock(eb)) {
3710 ret = flush_write_bio(epd);
3711 if (ret < 0)
3712 return ret;
3713 flush = 1;
3714 btrfs_tree_lock(eb);
3715 }
3716
3717 if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
3718 btrfs_tree_unlock(eb);
3719 if (!epd->sync_io)
3720 return 0;
3721 if (!flush) {
3722 ret = flush_write_bio(epd);
3723 if (ret < 0)
3724 return ret;
3725 flush = 1;
3726 }
3727 while (1) {
3728 wait_on_extent_buffer_writeback(eb);
3729 btrfs_tree_lock(eb);
3730 if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
3731 break;
3732 btrfs_tree_unlock(eb);
3733 }
3734 }
3735
3736 /*
3737 * We need to do this to prevent races in people who check if the eb is
3738 * under IO since we can end up having no IO bits set for a short period
3739 * of time.
3740 */
3741 spin_lock(&eb->refs_lock);
3742 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3743 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
3744 spin_unlock(&eb->refs_lock);
3745 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3746 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
3747 -eb->len,
3748 fs_info->dirty_metadata_batch);
3749 ret = 1;
3750 } else {
3751 spin_unlock(&eb->refs_lock);
3752 }
3753
3754 btrfs_tree_unlock(eb);
3755
3756 if (!ret)
3757 return ret;
3758
3759 num_pages = num_extent_pages(eb);
3760 for (i = 0; i < num_pages; i++) {
3761 struct page *p = eb->pages[i];
3762
3763 if (!trylock_page(p)) {
3764 if (!flush) {
3765 int err;
3766
3767 err = flush_write_bio(epd);
3768 if (err < 0) {
3769 ret = err;
3770 failed_page_nr = i;
3771 goto err_unlock;
3772 }
3773 flush = 1;
3774 }
3775 lock_page(p);
3776 }
3777 }
3778
3779 return ret;
3780err_unlock:
3781 /* Unlock already locked pages */
3782 for (i = 0; i < failed_page_nr; i++)
3783 unlock_page(eb->pages[i]);
3784 /*
3785 * Clear EXTENT_BUFFER_WRITEBACK and wake up anyone waiting on it.
3786 * Also set back EXTENT_BUFFER_DIRTY so future attempts to this eb can
3787 * be made and undo everything done before.
3788 */
3789 btrfs_tree_lock(eb);
3790 spin_lock(&eb->refs_lock);
3791 set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3792 end_extent_buffer_writeback(eb);
3793 spin_unlock(&eb->refs_lock);
3794 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes, eb->len,
3795 fs_info->dirty_metadata_batch);
3796 btrfs_clear_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
3797 btrfs_tree_unlock(eb);
3798 return ret;
3799}
3800
3801static void set_btree_ioerr(struct page *page)
3802{
3803 struct extent_buffer *eb = (struct extent_buffer *)page->private;
3804 struct btrfs_fs_info *fs_info;
3805
3806 SetPageError(page);
3807 if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3808 return;
3809
3810 /*
3811 * If we error out, we should add back the dirty_metadata_bytes
3812 * to make it consistent.
3813 */
3814 fs_info = eb->fs_info;
3815 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
3816 eb->len, fs_info->dirty_metadata_batch);
3817
3818 /*
3819 * If writeback for a btree extent that doesn't belong to a log tree
3820 * failed, increment the counter transaction->eb_write_errors.
3821 * We do this because while the transaction is running and before it's
3822 * committing (when we call filemap_fdata[write|wait]_range against
3823 * the btree inode), we might have
3824 * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
3825 * returns an error or an error happens during writeback, when we're
3826 * committing the transaction we wouldn't know about it, since the pages
3827 * can be no longer dirty nor marked anymore for writeback (if a
3828 * subsequent modification to the extent buffer didn't happen before the
3829 * transaction commit), which makes filemap_fdata[write|wait]_range not
3830 * able to find the pages tagged with SetPageError at transaction
3831 * commit time. So if this happens we must abort the transaction,
3832 * otherwise we commit a super block with btree roots that point to
3833 * btree nodes/leafs whose content on disk is invalid - either garbage
3834 * or the content of some node/leaf from a past generation that got
3835 * cowed or deleted and is no longer valid.
3836 *
3837 * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
3838 * not be enough - we need to distinguish between log tree extents vs
3839 * non-log tree extents, and the next filemap_fdatawait_range() call
3840 * will catch and clear such errors in the mapping - and that call might
3841 * be from a log sync and not from a transaction commit. Also, checking
3842 * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
3843 * not done and would not be reliable - the eb might have been released
3844 * from memory and reading it back again means that flag would not be
3845 * set (since it's a runtime flag, not persisted on disk).
3846 *
3847 * Using the flags below in the btree inode also makes us achieve the
3848 * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
3849 * writeback for all dirty pages and before filemap_fdatawait_range()
3850 * is called, the writeback for all dirty pages had already finished
3851 * with errors - because we were not using AS_EIO/AS_ENOSPC,
3852 * filemap_fdatawait_range() would return success, as it could not know
3853 * that writeback errors happened (the pages were no longer tagged for
3854 * writeback).
3855 */
3856 switch (eb->log_index) {
3857 case -1:
3858 set_bit(BTRFS_FS_BTREE_ERR, &eb->fs_info->flags);
3859 break;
3860 case 0:
3861 set_bit(BTRFS_FS_LOG1_ERR, &eb->fs_info->flags);
3862 break;
3863 case 1:
3864 set_bit(BTRFS_FS_LOG2_ERR, &eb->fs_info->flags);
3865 break;
3866 default:
3867 BUG(); /* unexpected, logic error */
3868 }
3869}
3870
3871static void end_bio_extent_buffer_writepage(struct bio *bio)
3872{
3873 struct bio_vec *bvec;
3874 struct extent_buffer *eb;
3875 int done;
3876 struct bvec_iter_all iter_all;
3877
3878 ASSERT(!bio_flagged(bio, BIO_CLONED));
3879 bio_for_each_segment_all(bvec, bio, iter_all) {
3880 struct page *page = bvec->bv_page;
3881
3882 eb = (struct extent_buffer *)page->private;
3883 BUG_ON(!eb);
3884 done = atomic_dec_and_test(&eb->io_pages);
3885
3886 if (bio->bi_status ||
3887 test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
3888 ClearPageUptodate(page);
3889 set_btree_ioerr(page);
3890 }
3891
3892 end_page_writeback(page);
3893
3894 if (!done)
3895 continue;
3896
3897 end_extent_buffer_writeback(eb);
3898 }
3899
3900 bio_put(bio);
3901}
3902
3903static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
3904 struct writeback_control *wbc,
3905 struct extent_page_data *epd)
3906{
3907 u64 offset = eb->start;
3908 u32 nritems;
3909 int i, num_pages;
3910 unsigned long start, end;
3911 unsigned int write_flags = wbc_to_write_flags(wbc) | REQ_META;
3912 int ret = 0;
3913
3914 clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
3915 num_pages = num_extent_pages(eb);
3916 atomic_set(&eb->io_pages, num_pages);
3917
3918 /* set btree blocks beyond nritems with 0 to avoid stale content. */
3919 nritems = btrfs_header_nritems(eb);
3920 if (btrfs_header_level(eb) > 0) {
3921 end = btrfs_node_key_ptr_offset(nritems);
3922
3923 memzero_extent_buffer(eb, end, eb->len - end);
3924 } else {
3925 /*
3926 * leaf:
3927 * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
3928 */
3929 start = btrfs_item_nr_offset(nritems);
3930 end = BTRFS_LEAF_DATA_OFFSET + leaf_data_end(eb);
3931 memzero_extent_buffer(eb, start, end - start);
3932 }
3933
3934 for (i = 0; i < num_pages; i++) {
3935 struct page *p = eb->pages[i];
3936
3937 clear_page_dirty_for_io(p);
3938 set_page_writeback(p);
3939 ret = submit_extent_page(REQ_OP_WRITE | write_flags, wbc,
3940 p, offset, PAGE_SIZE, 0,
3941 &epd->bio,
3942 end_bio_extent_buffer_writepage,
3943 0, 0, 0, false);
3944 if (ret) {
3945 set_btree_ioerr(p);
3946 if (PageWriteback(p))
3947 end_page_writeback(p);
3948 if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
3949 end_extent_buffer_writeback(eb);
3950 ret = -EIO;
3951 break;
3952 }
3953 offset += PAGE_SIZE;
3954 update_nr_written(wbc, 1);
3955 unlock_page(p);
3956 }
3957
3958 if (unlikely(ret)) {
3959 for (; i < num_pages; i++) {
3960 struct page *p = eb->pages[i];
3961 clear_page_dirty_for_io(p);
3962 unlock_page(p);
3963 }
3964 }
3965
3966 return ret;
3967}
3968
3969int btree_write_cache_pages(struct address_space *mapping,
3970 struct writeback_control *wbc)
3971{
3972 struct extent_buffer *eb, *prev_eb = NULL;
3973 struct extent_page_data epd = {
3974 .bio = NULL,
3975 .extent_locked = 0,
3976 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
3977 };
3978 struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
3979 int ret = 0;
3980 int done = 0;
3981 int nr_to_write_done = 0;
3982 struct pagevec pvec;
3983 int nr_pages;
3984 pgoff_t index;
3985 pgoff_t end; /* Inclusive */
3986 int scanned = 0;
3987 xa_mark_t tag;
3988
3989 pagevec_init(&pvec);
3990 if (wbc->range_cyclic) {
3991 index = mapping->writeback_index; /* Start from prev offset */
3992 end = -1;
3993 /*
3994 * Start from the beginning does not need to cycle over the
3995 * range, mark it as scanned.
3996 */
3997 scanned = (index == 0);
3998 } else {
3999 index = wbc->range_start >> PAGE_SHIFT;
4000 end = wbc->range_end >> PAGE_SHIFT;
4001 scanned = 1;
4002 }
4003 if (wbc->sync_mode == WB_SYNC_ALL)
4004 tag = PAGECACHE_TAG_TOWRITE;
4005 else
4006 tag = PAGECACHE_TAG_DIRTY;
4007retry:
4008 if (wbc->sync_mode == WB_SYNC_ALL)
4009 tag_pages_for_writeback(mapping, index, end);
4010 while (!done && !nr_to_write_done && (index <= end) &&
4011 (nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
4012 tag))) {
4013 unsigned i;
4014
4015 for (i = 0; i < nr_pages; i++) {
4016 struct page *page = pvec.pages[i];
4017
4018 if (!PagePrivate(page))
4019 continue;
4020
4021 spin_lock(&mapping->private_lock);
4022 if (!PagePrivate(page)) {
4023 spin_unlock(&mapping->private_lock);
4024 continue;
4025 }
4026
4027 eb = (struct extent_buffer *)page->private;
4028
4029 /*
4030 * Shouldn't happen and normally this would be a BUG_ON
4031 * but no sense in crashing the users box for something
4032 * we can survive anyway.
4033 */
4034 if (WARN_ON(!eb)) {
4035 spin_unlock(&mapping->private_lock);
4036 continue;
4037 }
4038
4039 if (eb == prev_eb) {
4040 spin_unlock(&mapping->private_lock);
4041 continue;
4042 }
4043
4044 ret = atomic_inc_not_zero(&eb->refs);
4045 spin_unlock(&mapping->private_lock);
4046 if (!ret)
4047 continue;
4048
4049 prev_eb = eb;
4050 ret = lock_extent_buffer_for_io(eb, &epd);
4051 if (!ret) {
4052 free_extent_buffer(eb);
4053 continue;
4054 } else if (ret < 0) {
4055 done = 1;
4056 free_extent_buffer(eb);
4057 break;
4058 }
4059
4060 ret = write_one_eb(eb, wbc, &epd);
4061 if (ret) {
4062 done = 1;
4063 free_extent_buffer(eb);
4064 break;
4065 }
4066 free_extent_buffer(eb);
4067
4068 /*
4069 * the filesystem may choose to bump up nr_to_write.
4070 * We have to make sure to honor the new nr_to_write
4071 * at any time
4072 */
4073 nr_to_write_done = wbc->nr_to_write <= 0;
4074 }
4075 pagevec_release(&pvec);
4076 cond_resched();
4077 }
4078 if (!scanned && !done) {
4079 /*
4080 * We hit the last page and there is more work to be done: wrap
4081 * back to the start of the file
4082 */
4083 scanned = 1;
4084 index = 0;
4085 goto retry;
4086 }
4087 ASSERT(ret <= 0);
4088 if (ret < 0) {
4089 end_write_bio(&epd, ret);
4090 return ret;
4091 }
4092 /*
4093 * If something went wrong, don't allow any metadata write bio to be
4094 * submitted.
4095 *
4096 * This would prevent use-after-free if we had dirty pages not
4097 * cleaned up, which can still happen by fuzzed images.
4098 *
4099 * - Bad extent tree
4100 * Allowing existing tree block to be allocated for other trees.
4101 *
4102 * - Log tree operations
4103 * Exiting tree blocks get allocated to log tree, bumps its
4104 * generation, then get cleaned in tree re-balance.
4105 * Such tree block will not be written back, since it's clean,
4106 * thus no WRITTEN flag set.
4107 * And after log writes back, this tree block is not traced by
4108 * any dirty extent_io_tree.
4109 *
4110 * - Offending tree block gets re-dirtied from its original owner
4111 * Since it has bumped generation, no WRITTEN flag, it can be
4112 * reused without COWing. This tree block will not be traced
4113 * by btrfs_transaction::dirty_pages.
4114 *
4115 * Now such dirty tree block will not be cleaned by any dirty
4116 * extent io tree. Thus we don't want to submit such wild eb
4117 * if the fs already has error.
4118 */
4119 if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
4120 ret = flush_write_bio(&epd);
4121 } else {
4122 ret = -EROFS;
4123 end_write_bio(&epd, ret);
4124 }
4125 return ret;
4126}
4127
4128/**
4129 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
4130 * @mapping: address space structure to write
4131 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
4132 * @data: data passed to __extent_writepage function
4133 *
4134 * If a page is already under I/O, write_cache_pages() skips it, even
4135 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
4136 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
4137 * and msync() need to guarantee that all the data which was dirty at the time
4138 * the call was made get new I/O started against them. If wbc->sync_mode is
4139 * WB_SYNC_ALL then we were called for data integrity and we must wait for
4140 * existing IO to complete.
4141 */
4142static int extent_write_cache_pages(struct address_space *mapping,
4143 struct writeback_control *wbc,
4144 struct extent_page_data *epd)
4145{
4146 struct inode *inode = mapping->host;
4147 int ret = 0;
4148 int done = 0;
4149 int nr_to_write_done = 0;
4150 struct pagevec pvec;
4151 int nr_pages;
4152 pgoff_t index;
4153 pgoff_t end; /* Inclusive */
4154 pgoff_t done_index;
4155 int range_whole = 0;
4156 int scanned = 0;
4157 xa_mark_t tag;
4158
4159 /*
4160 * We have to hold onto the inode so that ordered extents can do their
4161 * work when the IO finishes. The alternative to this is failing to add
4162 * an ordered extent if the igrab() fails there and that is a huge pain
4163 * to deal with, so instead just hold onto the inode throughout the
4164 * writepages operation. If it fails here we are freeing up the inode
4165 * anyway and we'd rather not waste our time writing out stuff that is
4166 * going to be truncated anyway.
4167 */
4168 if (!igrab(inode))
4169 return 0;
4170
4171 pagevec_init(&pvec);
4172 if (wbc->range_cyclic) {
4173 index = mapping->writeback_index; /* Start from prev offset */
4174 end = -1;
4175 /*
4176 * Start from the beginning does not need to cycle over the
4177 * range, mark it as scanned.
4178 */
4179 scanned = (index == 0);
4180 } else {
4181 index = wbc->range_start >> PAGE_SHIFT;
4182 end = wbc->range_end >> PAGE_SHIFT;
4183 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
4184 range_whole = 1;
4185 scanned = 1;
4186 }
4187
4188 /*
4189 * We do the tagged writepage as long as the snapshot flush bit is set
4190 * and we are the first one who do the filemap_flush() on this inode.
4191 *
4192 * The nr_to_write == LONG_MAX is needed to make sure other flushers do
4193 * not race in and drop the bit.
4194 */
4195 if (range_whole && wbc->nr_to_write == LONG_MAX &&
4196 test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
4197 &BTRFS_I(inode)->runtime_flags))
4198 wbc->tagged_writepages = 1;
4199
4200 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
4201 tag = PAGECACHE_TAG_TOWRITE;
4202 else
4203 tag = PAGECACHE_TAG_DIRTY;
4204retry:
4205 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
4206 tag_pages_for_writeback(mapping, index, end);
4207 done_index = index;
4208 while (!done && !nr_to_write_done && (index <= end) &&
4209 (nr_pages = pagevec_lookup_range_tag(&pvec, mapping,
4210 &index, end, tag))) {
4211 unsigned i;
4212
4213 for (i = 0; i < nr_pages; i++) {
4214 struct page *page = pvec.pages[i];
4215
4216 done_index = page->index + 1;
4217 /*
4218 * At this point we hold neither the i_pages lock nor
4219 * the page lock: the page may be truncated or
4220 * invalidated (changing page->mapping to NULL),
4221 * or even swizzled back from swapper_space to
4222 * tmpfs file mapping
4223 */
4224 if (!trylock_page(page)) {
4225 ret = flush_write_bio(epd);
4226 BUG_ON(ret < 0);
4227 lock_page(page);
4228 }
4229
4230 if (unlikely(page->mapping != mapping)) {
4231 unlock_page(page);
4232 continue;
4233 }
4234
4235 if (wbc->sync_mode != WB_SYNC_NONE) {
4236 if (PageWriteback(page)) {
4237 ret = flush_write_bio(epd);
4238 BUG_ON(ret < 0);
4239 }
4240 wait_on_page_writeback(page);
4241 }
4242
4243 if (PageWriteback(page) ||
4244 !clear_page_dirty_for_io(page)) {
4245 unlock_page(page);
4246 continue;
4247 }
4248
4249 ret = __extent_writepage(page, wbc, epd);
4250 if (ret < 0) {
4251 done = 1;
4252 break;
4253 }
4254
4255 /*
4256 * the filesystem may choose to bump up nr_to_write.
4257 * We have to make sure to honor the new nr_to_write
4258 * at any time
4259 */
4260 nr_to_write_done = wbc->nr_to_write <= 0;
4261 }
4262 pagevec_release(&pvec);
4263 cond_resched();
4264 }
4265 if (!scanned && !done) {
4266 /*
4267 * We hit the last page and there is more work to be done: wrap
4268 * back to the start of the file
4269 */
4270 scanned = 1;
4271 index = 0;
4272
4273 /*
4274 * If we're looping we could run into a page that is locked by a
4275 * writer and that writer could be waiting on writeback for a
4276 * page in our current bio, and thus deadlock, so flush the
4277 * write bio here.
4278 */
4279 ret = flush_write_bio(epd);
4280 if (!ret)
4281 goto retry;
4282 }
4283
4284 if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
4285 mapping->writeback_index = done_index;
4286
4287 btrfs_add_delayed_iput(inode);
4288 return ret;
4289}
4290
4291int extent_write_full_page(struct page *page, struct writeback_control *wbc)
4292{
4293 int ret;
4294 struct extent_page_data epd = {
4295 .bio = NULL,
4296 .extent_locked = 0,
4297 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4298 };
4299
4300 ret = __extent_writepage(page, wbc, &epd);
4301 ASSERT(ret <= 0);
4302 if (ret < 0) {
4303 end_write_bio(&epd, ret);
4304 return ret;
4305 }
4306
4307 ret = flush_write_bio(&epd);
4308 ASSERT(ret <= 0);
4309 return ret;
4310}
4311
4312int extent_write_locked_range(struct inode *inode, u64 start, u64 end,
4313 int mode)
4314{
4315 int ret = 0;
4316 struct address_space *mapping = inode->i_mapping;
4317 struct page *page;
4318 unsigned long nr_pages = (end - start + PAGE_SIZE) >>
4319 PAGE_SHIFT;
4320
4321 struct extent_page_data epd = {
4322 .bio = NULL,
4323 .extent_locked = 1,
4324 .sync_io = mode == WB_SYNC_ALL,
4325 };
4326 struct writeback_control wbc_writepages = {
4327 .sync_mode = mode,
4328 .nr_to_write = nr_pages * 2,
4329 .range_start = start,
4330 .range_end = end + 1,
4331 /* We're called from an async helper function */
4332 .punt_to_cgroup = 1,
4333 .no_cgroup_owner = 1,
4334 };
4335
4336 wbc_attach_fdatawrite_inode(&wbc_writepages, inode);
4337 while (start <= end) {
4338 page = find_get_page(mapping, start >> PAGE_SHIFT);
4339 if (clear_page_dirty_for_io(page))
4340 ret = __extent_writepage(page, &wbc_writepages, &epd);
4341 else {
4342 btrfs_writepage_endio_finish_ordered(page, start,
4343 start + PAGE_SIZE - 1, 1);
4344 unlock_page(page);
4345 }
4346 put_page(page);
4347 start += PAGE_SIZE;
4348 }
4349
4350 ASSERT(ret <= 0);
4351 if (ret == 0)
4352 ret = flush_write_bio(&epd);
4353 else
4354 end_write_bio(&epd, ret);
4355
4356 wbc_detach_inode(&wbc_writepages);
4357 return ret;
4358}
4359
4360int extent_writepages(struct address_space *mapping,
4361 struct writeback_control *wbc)
4362{
4363 int ret = 0;
4364 struct extent_page_data epd = {
4365 .bio = NULL,
4366 .extent_locked = 0,
4367 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
4368 };
4369
4370 ret = extent_write_cache_pages(mapping, wbc, &epd);
4371 ASSERT(ret <= 0);
4372 if (ret < 0) {
4373 end_write_bio(&epd, ret);
4374 return ret;
4375 }
4376 ret = flush_write_bio(&epd);
4377 return ret;
4378}
4379
4380void extent_readahead(struct readahead_control *rac)
4381{
4382 struct bio *bio = NULL;
4383 unsigned long bio_flags = 0;
4384 struct page *pagepool[16];
4385 struct extent_map *em_cached = NULL;
4386 u64 prev_em_start = (u64)-1;
4387 int nr;
4388
4389 while ((nr = readahead_page_batch(rac, pagepool))) {
4390 u64 contig_start = page_offset(pagepool[0]);
4391 u64 contig_end = page_offset(pagepool[nr - 1]) + PAGE_SIZE - 1;
4392
4393 ASSERT(contig_start + nr * PAGE_SIZE - 1 == contig_end);
4394
4395 contiguous_readpages(pagepool, nr, contig_start, contig_end,
4396 &em_cached, &bio, &bio_flags, &prev_em_start);
4397 }
4398
4399 if (em_cached)
4400 free_extent_map(em_cached);
4401
4402 if (bio) {
4403 if (submit_one_bio(bio, 0, bio_flags))
4404 return;
4405 }
4406}
4407
4408/*
4409 * basic invalidatepage code, this waits on any locked or writeback
4410 * ranges corresponding to the page, and then deletes any extent state
4411 * records from the tree
4412 */
4413int extent_invalidatepage(struct extent_io_tree *tree,
4414 struct page *page, unsigned long offset)
4415{
4416 struct extent_state *cached_state = NULL;
4417 u64 start = page_offset(page);
4418 u64 end = start + PAGE_SIZE - 1;
4419 size_t blocksize = page->mapping->host->i_sb->s_blocksize;
4420
4421 start += ALIGN(offset, blocksize);
4422 if (start > end)
4423 return 0;
4424
4425 lock_extent_bits(tree, start, end, &cached_state);
4426 wait_on_page_writeback(page);
4427 clear_extent_bit(tree, start, end, EXTENT_LOCKED | EXTENT_DELALLOC |
4428 EXTENT_DO_ACCOUNTING, 1, 1, &cached_state);
4429 return 0;
4430}
4431
4432/*
4433 * a helper for releasepage, this tests for areas of the page that
4434 * are locked or under IO and drops the related state bits if it is safe
4435 * to drop the page.
4436 */
4437static int try_release_extent_state(struct extent_io_tree *tree,
4438 struct page *page, gfp_t mask)
4439{
4440 u64 start = page_offset(page);
4441 u64 end = start + PAGE_SIZE - 1;
4442 int ret = 1;
4443
4444 if (test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL)) {
4445 ret = 0;
4446 } else {
4447 /*
4448 * at this point we can safely clear everything except the
4449 * locked bit and the nodatasum bit
4450 */
4451 ret = __clear_extent_bit(tree, start, end,
4452 ~(EXTENT_LOCKED | EXTENT_NODATASUM),
4453 0, 0, NULL, mask, NULL);
4454
4455 /* if clear_extent_bit failed for enomem reasons,
4456 * we can't allow the release to continue.
4457 */
4458 if (ret < 0)
4459 ret = 0;
4460 else
4461 ret = 1;
4462 }
4463 return ret;
4464}
4465
4466/*
4467 * a helper for releasepage. As long as there are no locked extents
4468 * in the range corresponding to the page, both state records and extent
4469 * map records are removed
4470 */
4471int try_release_extent_mapping(struct page *page, gfp_t mask)
4472{
4473 struct extent_map *em;
4474 u64 start = page_offset(page);
4475 u64 end = start + PAGE_SIZE - 1;
4476 struct btrfs_inode *btrfs_inode = BTRFS_I(page->mapping->host);
4477 struct extent_io_tree *tree = &btrfs_inode->io_tree;
4478 struct extent_map_tree *map = &btrfs_inode->extent_tree;
4479
4480 if (gfpflags_allow_blocking(mask) &&
4481 page->mapping->host->i_size > SZ_16M) {
4482 u64 len;
4483 while (start <= end) {
4484 struct btrfs_fs_info *fs_info;
4485 u64 cur_gen;
4486
4487 len = end - start + 1;
4488 write_lock(&map->lock);
4489 em = lookup_extent_mapping(map, start, len);
4490 if (!em) {
4491 write_unlock(&map->lock);
4492 break;
4493 }
4494 if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
4495 em->start != start) {
4496 write_unlock(&map->lock);
4497 free_extent_map(em);
4498 break;
4499 }
4500 if (test_range_bit(tree, em->start,
4501 extent_map_end(em) - 1,
4502 EXTENT_LOCKED, 0, NULL))
4503 goto next;
4504 /*
4505 * If it's not in the list of modified extents, used
4506 * by a fast fsync, we can remove it. If it's being
4507 * logged we can safely remove it since fsync took an
4508 * extra reference on the em.
4509 */
4510 if (list_empty(&em->list) ||
4511 test_bit(EXTENT_FLAG_LOGGING, &em->flags))
4512 goto remove_em;
4513 /*
4514 * If it's in the list of modified extents, remove it
4515 * only if its generation is older then the current one,
4516 * in which case we don't need it for a fast fsync.
4517 * Otherwise don't remove it, we could be racing with an
4518 * ongoing fast fsync that could miss the new extent.
4519 */
4520 fs_info = btrfs_inode->root->fs_info;
4521 spin_lock(&fs_info->trans_lock);
4522 cur_gen = fs_info->generation;
4523 spin_unlock(&fs_info->trans_lock);
4524 if (em->generation >= cur_gen)
4525 goto next;
4526remove_em:
4527 /*
4528 * We only remove extent maps that are not in the list of
4529 * modified extents or that are in the list but with a
4530 * generation lower then the current generation, so there
4531 * is no need to set the full fsync flag on the inode (it
4532 * hurts the fsync performance for workloads with a data
4533 * size that exceeds or is close to the system's memory).
4534 */
4535 remove_extent_mapping(map, em);
4536 /* once for the rb tree */
4537 free_extent_map(em);
4538next:
4539 start = extent_map_end(em);
4540 write_unlock(&map->lock);
4541
4542 /* once for us */
4543 free_extent_map(em);
4544
4545 cond_resched(); /* Allow large-extent preemption. */
4546 }
4547 }
4548 return try_release_extent_state(tree, page, mask);
4549}
4550
4551/*
4552 * helper function for fiemap, which doesn't want to see any holes.
4553 * This maps until we find something past 'last'
4554 */
4555static struct extent_map *get_extent_skip_holes(struct inode *inode,
4556 u64 offset, u64 last)
4557{
4558 u64 sectorsize = btrfs_inode_sectorsize(inode);
4559 struct extent_map *em;
4560 u64 len;
4561
4562 if (offset >= last)
4563 return NULL;
4564
4565 while (1) {
4566 len = last - offset;
4567 if (len == 0)
4568 break;
4569 len = ALIGN(len, sectorsize);
4570 em = btrfs_get_extent_fiemap(BTRFS_I(inode), offset, len);
4571 if (IS_ERR_OR_NULL(em))
4572 return em;
4573
4574 /* if this isn't a hole return it */
4575 if (em->block_start != EXTENT_MAP_HOLE)
4576 return em;
4577
4578 /* this is a hole, advance to the next extent */
4579 offset = extent_map_end(em);
4580 free_extent_map(em);
4581 if (offset >= last)
4582 break;
4583 }
4584 return NULL;
4585}
4586
4587/*
4588 * To cache previous fiemap extent
4589 *
4590 * Will be used for merging fiemap extent
4591 */
4592struct fiemap_cache {
4593 u64 offset;
4594 u64 phys;
4595 u64 len;
4596 u32 flags;
4597 bool cached;
4598};
4599
4600/*
4601 * Helper to submit fiemap extent.
4602 *
4603 * Will try to merge current fiemap extent specified by @offset, @phys,
4604 * @len and @flags with cached one.
4605 * And only when we fails to merge, cached one will be submitted as
4606 * fiemap extent.
4607 *
4608 * Return value is the same as fiemap_fill_next_extent().
4609 */
4610static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
4611 struct fiemap_cache *cache,
4612 u64 offset, u64 phys, u64 len, u32 flags)
4613{
4614 int ret = 0;
4615
4616 if (!cache->cached)
4617 goto assign;
4618
4619 /*
4620 * Sanity check, extent_fiemap() should have ensured that new
4621 * fiemap extent won't overlap with cached one.
4622 * Not recoverable.
4623 *
4624 * NOTE: Physical address can overlap, due to compression
4625 */
4626 if (cache->offset + cache->len > offset) {
4627 WARN_ON(1);
4628 return -EINVAL;
4629 }
4630
4631 /*
4632 * Only merges fiemap extents if
4633 * 1) Their logical addresses are continuous
4634 *
4635 * 2) Their physical addresses are continuous
4636 * So truly compressed (physical size smaller than logical size)
4637 * extents won't get merged with each other
4638 *
4639 * 3) Share same flags except FIEMAP_EXTENT_LAST
4640 * So regular extent won't get merged with prealloc extent
4641 */
4642 if (cache->offset + cache->len == offset &&
4643 cache->phys + cache->len == phys &&
4644 (cache->flags & ~FIEMAP_EXTENT_LAST) ==
4645 (flags & ~FIEMAP_EXTENT_LAST)) {
4646 cache->len += len;
4647 cache->flags |= flags;
4648 goto try_submit_last;
4649 }
4650
4651 /* Not mergeable, need to submit cached one */
4652 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4653 cache->len, cache->flags);
4654 cache->cached = false;
4655 if (ret)
4656 return ret;
4657assign:
4658 cache->cached = true;
4659 cache->offset = offset;
4660 cache->phys = phys;
4661 cache->len = len;
4662 cache->flags = flags;
4663try_submit_last:
4664 if (cache->flags & FIEMAP_EXTENT_LAST) {
4665 ret = fiemap_fill_next_extent(fieinfo, cache->offset,
4666 cache->phys, cache->len, cache->flags);
4667 cache->cached = false;
4668 }
4669 return ret;
4670}
4671
4672/*
4673 * Emit last fiemap cache
4674 *
4675 * The last fiemap cache may still be cached in the following case:
4676 * 0 4k 8k
4677 * |<- Fiemap range ->|
4678 * |<------------ First extent ----------->|
4679 *
4680 * In this case, the first extent range will be cached but not emitted.
4681 * So we must emit it before ending extent_fiemap().
4682 */
4683static int emit_last_fiemap_cache(struct fiemap_extent_info *fieinfo,
4684 struct fiemap_cache *cache)
4685{
4686 int ret;
4687
4688 if (!cache->cached)
4689 return 0;
4690
4691 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
4692 cache->len, cache->flags);
4693 cache->cached = false;
4694 if (ret > 0)
4695 ret = 0;
4696 return ret;
4697}
4698
4699int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
4700 u64 start, u64 len)
4701{
4702 int ret = 0;
4703 u64 off = start;
4704 u64 max = start + len;
4705 u32 flags = 0;
4706 u32 found_type;
4707 u64 last;
4708 u64 last_for_get_extent = 0;
4709 u64 disko = 0;
4710 u64 isize = i_size_read(inode);
4711 struct btrfs_key found_key;
4712 struct extent_map *em = NULL;
4713 struct extent_state *cached_state = NULL;
4714 struct btrfs_path *path;
4715 struct btrfs_root *root = BTRFS_I(inode)->root;
4716 struct fiemap_cache cache = { 0 };
4717 struct ulist *roots;
4718 struct ulist *tmp_ulist;
4719 int end = 0;
4720 u64 em_start = 0;
4721 u64 em_len = 0;
4722 u64 em_end = 0;
4723
4724 if (len == 0)
4725 return -EINVAL;
4726
4727 path = btrfs_alloc_path();
4728 if (!path)
4729 return -ENOMEM;
4730 path->leave_spinning = 1;
4731
4732 roots = ulist_alloc(GFP_KERNEL);
4733 tmp_ulist = ulist_alloc(GFP_KERNEL);
4734 if (!roots || !tmp_ulist) {
4735 ret = -ENOMEM;
4736 goto out_free_ulist;
4737 }
4738
4739 start = round_down(start, btrfs_inode_sectorsize(inode));
4740 len = round_up(max, btrfs_inode_sectorsize(inode)) - start;
4741
4742 /*
4743 * lookup the last file extent. We're not using i_size here
4744 * because there might be preallocation past i_size
4745 */
4746 ret = btrfs_lookup_file_extent(NULL, root, path,
4747 btrfs_ino(BTRFS_I(inode)), -1, 0);
4748 if (ret < 0) {
4749 goto out_free_ulist;
4750 } else {
4751 WARN_ON(!ret);
4752 if (ret == 1)
4753 ret = 0;
4754 }
4755
4756 path->slots[0]--;
4757 btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
4758 found_type = found_key.type;
4759
4760 /* No extents, but there might be delalloc bits */
4761 if (found_key.objectid != btrfs_ino(BTRFS_I(inode)) ||
4762 found_type != BTRFS_EXTENT_DATA_KEY) {
4763 /* have to trust i_size as the end */
4764 last = (u64)-1;
4765 last_for_get_extent = isize;
4766 } else {
4767 /*
4768 * remember the start of the last extent. There are a
4769 * bunch of different factors that go into the length of the
4770 * extent, so its much less complex to remember where it started
4771 */
4772 last = found_key.offset;
4773 last_for_get_extent = last + 1;
4774 }
4775 btrfs_release_path(path);
4776
4777 /*
4778 * we might have some extents allocated but more delalloc past those
4779 * extents. so, we trust isize unless the start of the last extent is
4780 * beyond isize
4781 */
4782 if (last < isize) {
4783 last = (u64)-1;
4784 last_for_get_extent = isize;
4785 }
4786
4787 lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4788 &cached_state);
4789
4790 em = get_extent_skip_holes(inode, start, last_for_get_extent);
4791 if (!em)
4792 goto out;
4793 if (IS_ERR(em)) {
4794 ret = PTR_ERR(em);
4795 goto out;
4796 }
4797
4798 while (!end) {
4799 u64 offset_in_extent = 0;
4800
4801 /* break if the extent we found is outside the range */
4802 if (em->start >= max || extent_map_end(em) < off)
4803 break;
4804
4805 /*
4806 * get_extent may return an extent that starts before our
4807 * requested range. We have to make sure the ranges
4808 * we return to fiemap always move forward and don't
4809 * overlap, so adjust the offsets here
4810 */
4811 em_start = max(em->start, off);
4812
4813 /*
4814 * record the offset from the start of the extent
4815 * for adjusting the disk offset below. Only do this if the
4816 * extent isn't compressed since our in ram offset may be past
4817 * what we have actually allocated on disk.
4818 */
4819 if (!test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4820 offset_in_extent = em_start - em->start;
4821 em_end = extent_map_end(em);
4822 em_len = em_end - em_start;
4823 flags = 0;
4824 if (em->block_start < EXTENT_MAP_LAST_BYTE)
4825 disko = em->block_start + offset_in_extent;
4826 else
4827 disko = 0;
4828
4829 /*
4830 * bump off for our next call to get_extent
4831 */
4832 off = extent_map_end(em);
4833 if (off >= max)
4834 end = 1;
4835
4836 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
4837 end = 1;
4838 flags |= FIEMAP_EXTENT_LAST;
4839 } else if (em->block_start == EXTENT_MAP_INLINE) {
4840 flags |= (FIEMAP_EXTENT_DATA_INLINE |
4841 FIEMAP_EXTENT_NOT_ALIGNED);
4842 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
4843 flags |= (FIEMAP_EXTENT_DELALLOC |
4844 FIEMAP_EXTENT_UNKNOWN);
4845 } else if (fieinfo->fi_extents_max) {
4846 u64 bytenr = em->block_start -
4847 (em->start - em->orig_start);
4848
4849 /*
4850 * As btrfs supports shared space, this information
4851 * can be exported to userspace tools via
4852 * flag FIEMAP_EXTENT_SHARED. If fi_extents_max == 0
4853 * then we're just getting a count and we can skip the
4854 * lookup stuff.
4855 */
4856 ret = btrfs_check_shared(root,
4857 btrfs_ino(BTRFS_I(inode)),
4858 bytenr, roots, tmp_ulist);
4859 if (ret < 0)
4860 goto out_free;
4861 if (ret)
4862 flags |= FIEMAP_EXTENT_SHARED;
4863 ret = 0;
4864 }
4865 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
4866 flags |= FIEMAP_EXTENT_ENCODED;
4867 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
4868 flags |= FIEMAP_EXTENT_UNWRITTEN;
4869
4870 free_extent_map(em);
4871 em = NULL;
4872 if ((em_start >= last) || em_len == (u64)-1 ||
4873 (last == (u64)-1 && isize <= em_end)) {
4874 flags |= FIEMAP_EXTENT_LAST;
4875 end = 1;
4876 }
4877
4878 /* now scan forward to see if this is really the last extent. */
4879 em = get_extent_skip_holes(inode, off, last_for_get_extent);
4880 if (IS_ERR(em)) {
4881 ret = PTR_ERR(em);
4882 goto out;
4883 }
4884 if (!em) {
4885 flags |= FIEMAP_EXTENT_LAST;
4886 end = 1;
4887 }
4888 ret = emit_fiemap_extent(fieinfo, &cache, em_start, disko,
4889 em_len, flags);
4890 if (ret) {
4891 if (ret == 1)
4892 ret = 0;
4893 goto out_free;
4894 }
4895 }
4896out_free:
4897 if (!ret)
4898 ret = emit_last_fiemap_cache(fieinfo, &cache);
4899 free_extent_map(em);
4900out:
4901 unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len - 1,
4902 &cached_state);
4903
4904out_free_ulist:
4905 btrfs_free_path(path);
4906 ulist_free(roots);
4907 ulist_free(tmp_ulist);
4908 return ret;
4909}
4910
4911static void __free_extent_buffer(struct extent_buffer *eb)
4912{
4913 kmem_cache_free(extent_buffer_cache, eb);
4914}
4915
4916int extent_buffer_under_io(const struct extent_buffer *eb)
4917{
4918 return (atomic_read(&eb->io_pages) ||
4919 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4920 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4921}
4922
4923/*
4924 * Release all pages attached to the extent buffer.
4925 */
4926static void btrfs_release_extent_buffer_pages(struct extent_buffer *eb)
4927{
4928 int i;
4929 int num_pages;
4930 int mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
4931
4932 BUG_ON(extent_buffer_under_io(eb));
4933
4934 num_pages = num_extent_pages(eb);
4935 for (i = 0; i < num_pages; i++) {
4936 struct page *page = eb->pages[i];
4937
4938 if (!page)
4939 continue;
4940 if (mapped)
4941 spin_lock(&page->mapping->private_lock);
4942 /*
4943 * We do this since we'll remove the pages after we've
4944 * removed the eb from the radix tree, so we could race
4945 * and have this page now attached to the new eb. So
4946 * only clear page_private if it's still connected to
4947 * this eb.
4948 */
4949 if (PagePrivate(page) &&
4950 page->private == (unsigned long)eb) {
4951 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4952 BUG_ON(PageDirty(page));
4953 BUG_ON(PageWriteback(page));
4954 /*
4955 * We need to make sure we haven't be attached
4956 * to a new eb.
4957 */
4958 detach_page_private(page);
4959 }
4960
4961 if (mapped)
4962 spin_unlock(&page->mapping->private_lock);
4963
4964 /* One for when we allocated the page */
4965 put_page(page);
4966 }
4967}
4968
4969/*
4970 * Helper for releasing the extent buffer.
4971 */
4972static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4973{
4974 btrfs_release_extent_buffer_pages(eb);
4975 btrfs_leak_debug_del(&eb->fs_info->eb_leak_lock, &eb->leak_list);
4976 __free_extent_buffer(eb);
4977}
4978
4979static struct extent_buffer *
4980__alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4981 unsigned long len)
4982{
4983 struct extent_buffer *eb = NULL;
4984
4985 eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4986 eb->start = start;
4987 eb->len = len;
4988 eb->fs_info = fs_info;
4989 eb->bflags = 0;
4990 rwlock_init(&eb->lock);
4991 atomic_set(&eb->blocking_readers, 0);
4992 eb->blocking_writers = 0;
4993 eb->lock_nested = false;
4994 init_waitqueue_head(&eb->write_lock_wq);
4995 init_waitqueue_head(&eb->read_lock_wq);
4996
4997 btrfs_leak_debug_add(&fs_info->eb_leak_lock, &eb->leak_list,
4998 &fs_info->allocated_ebs);
4999
5000 spin_lock_init(&eb->refs_lock);
5001 atomic_set(&eb->refs, 1);
5002 atomic_set(&eb->io_pages, 0);
5003
5004 /*
5005 * Sanity checks, currently the maximum is 64k covered by 16x 4k pages
5006 */
5007 BUILD_BUG_ON(BTRFS_MAX_METADATA_BLOCKSIZE
5008 > MAX_INLINE_EXTENT_BUFFER_SIZE);
5009 BUG_ON(len > MAX_INLINE_EXTENT_BUFFER_SIZE);
5010
5011#ifdef CONFIG_BTRFS_DEBUG
5012 eb->spinning_writers = 0;
5013 atomic_set(&eb->spinning_readers, 0);
5014 atomic_set(&eb->read_locks, 0);
5015 eb->write_locks = 0;
5016#endif
5017
5018 return eb;
5019}
5020
5021struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
5022{
5023 int i;
5024 struct page *p;
5025 struct extent_buffer *new;
5026 int num_pages = num_extent_pages(src);
5027
5028 new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
5029 if (new == NULL)
5030 return NULL;
5031
5032 for (i = 0; i < num_pages; i++) {
5033 p = alloc_page(GFP_NOFS);
5034 if (!p) {
5035 btrfs_release_extent_buffer(new);
5036 return NULL;
5037 }
5038 attach_extent_buffer_page(new, p);
5039 WARN_ON(PageDirty(p));
5040 SetPageUptodate(p);
5041 new->pages[i] = p;
5042 copy_page(page_address(p), page_address(src->pages[i]));
5043 }
5044
5045 set_bit(EXTENT_BUFFER_UPTODATE, &new->bflags);
5046 set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
5047
5048 return new;
5049}
5050
5051struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
5052 u64 start, unsigned long len)
5053{
5054 struct extent_buffer *eb;
5055 int num_pages;
5056 int i;
5057
5058 eb = __alloc_extent_buffer(fs_info, start, len);
5059 if (!eb)
5060 return NULL;
5061
5062 num_pages = num_extent_pages(eb);
5063 for (i = 0; i < num_pages; i++) {
5064 eb->pages[i] = alloc_page(GFP_NOFS);
5065 if (!eb->pages[i])
5066 goto err;
5067 }
5068 set_extent_buffer_uptodate(eb);
5069 btrfs_set_header_nritems(eb, 0);
5070 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
5071
5072 return eb;
5073err:
5074 for (; i > 0; i--)
5075 __free_page(eb->pages[i - 1]);
5076 __free_extent_buffer(eb);
5077 return NULL;
5078}
5079
5080struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
5081 u64 start)
5082{
5083 return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
5084}
5085
5086static void check_buffer_tree_ref(struct extent_buffer *eb)
5087{
5088 int refs;
5089 /*
5090 * The TREE_REF bit is first set when the extent_buffer is added
5091 * to the radix tree. It is also reset, if unset, when a new reference
5092 * is created by find_extent_buffer.
5093 *
5094 * It is only cleared in two cases: freeing the last non-tree
5095 * reference to the extent_buffer when its STALE bit is set or
5096 * calling releasepage when the tree reference is the only reference.
5097 *
5098 * In both cases, care is taken to ensure that the extent_buffer's
5099 * pages are not under io. However, releasepage can be concurrently
5100 * called with creating new references, which is prone to race
5101 * conditions between the calls to check_buffer_tree_ref in those
5102 * codepaths and clearing TREE_REF in try_release_extent_buffer.
5103 *
5104 * The actual lifetime of the extent_buffer in the radix tree is
5105 * adequately protected by the refcount, but the TREE_REF bit and
5106 * its corresponding reference are not. To protect against this
5107 * class of races, we call check_buffer_tree_ref from the codepaths
5108 * which trigger io after they set eb->io_pages. Note that once io is
5109 * initiated, TREE_REF can no longer be cleared, so that is the
5110 * moment at which any such race is best fixed.
5111 */
5112 refs = atomic_read(&eb->refs);
5113 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5114 return;
5115
5116 spin_lock(&eb->refs_lock);
5117 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5118 atomic_inc(&eb->refs);
5119 spin_unlock(&eb->refs_lock);
5120}
5121
5122static void mark_extent_buffer_accessed(struct extent_buffer *eb,
5123 struct page *accessed)
5124{
5125 int num_pages, i;
5126
5127 check_buffer_tree_ref(eb);
5128
5129 num_pages = num_extent_pages(eb);
5130 for (i = 0; i < num_pages; i++) {
5131 struct page *p = eb->pages[i];
5132
5133 if (p != accessed)
5134 mark_page_accessed(p);
5135 }
5136}
5137
5138struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
5139 u64 start)
5140{
5141 struct extent_buffer *eb;
5142
5143 rcu_read_lock();
5144 eb = radix_tree_lookup(&fs_info->buffer_radix,
5145 start >> PAGE_SHIFT);
5146 if (eb && atomic_inc_not_zero(&eb->refs)) {
5147 rcu_read_unlock();
5148 /*
5149 * Lock our eb's refs_lock to avoid races with
5150 * free_extent_buffer. When we get our eb it might be flagged
5151 * with EXTENT_BUFFER_STALE and another task running
5152 * free_extent_buffer might have seen that flag set,
5153 * eb->refs == 2, that the buffer isn't under IO (dirty and
5154 * writeback flags not set) and it's still in the tree (flag
5155 * EXTENT_BUFFER_TREE_REF set), therefore being in the process
5156 * of decrementing the extent buffer's reference count twice.
5157 * So here we could race and increment the eb's reference count,
5158 * clear its stale flag, mark it as dirty and drop our reference
5159 * before the other task finishes executing free_extent_buffer,
5160 * which would later result in an attempt to free an extent
5161 * buffer that is dirty.
5162 */
5163 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
5164 spin_lock(&eb->refs_lock);
5165 spin_unlock(&eb->refs_lock);
5166 }
5167 mark_extent_buffer_accessed(eb, NULL);
5168 return eb;
5169 }
5170 rcu_read_unlock();
5171
5172 return NULL;
5173}
5174
5175#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5176struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
5177 u64 start)
5178{
5179 struct extent_buffer *eb, *exists = NULL;
5180 int ret;
5181
5182 eb = find_extent_buffer(fs_info, start);
5183 if (eb)
5184 return eb;
5185 eb = alloc_dummy_extent_buffer(fs_info, start);
5186 if (!eb)
5187 return ERR_PTR(-ENOMEM);
5188 eb->fs_info = fs_info;
5189again:
5190 ret = radix_tree_preload(GFP_NOFS);
5191 if (ret) {
5192 exists = ERR_PTR(ret);
5193 goto free_eb;
5194 }
5195 spin_lock(&fs_info->buffer_lock);
5196 ret = radix_tree_insert(&fs_info->buffer_radix,
5197 start >> PAGE_SHIFT, eb);
5198 spin_unlock(&fs_info->buffer_lock);
5199 radix_tree_preload_end();
5200 if (ret == -EEXIST) {
5201 exists = find_extent_buffer(fs_info, start);
5202 if (exists)
5203 goto free_eb;
5204 else
5205 goto again;
5206 }
5207 check_buffer_tree_ref(eb);
5208 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5209
5210 return eb;
5211free_eb:
5212 btrfs_release_extent_buffer(eb);
5213 return exists;
5214}
5215#endif
5216
5217struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
5218 u64 start)
5219{
5220 unsigned long len = fs_info->nodesize;
5221 int num_pages;
5222 int i;
5223 unsigned long index = start >> PAGE_SHIFT;
5224 struct extent_buffer *eb;
5225 struct extent_buffer *exists = NULL;
5226 struct page *p;
5227 struct address_space *mapping = fs_info->btree_inode->i_mapping;
5228 int uptodate = 1;
5229 int ret;
5230
5231 if (!IS_ALIGNED(start, fs_info->sectorsize)) {
5232 btrfs_err(fs_info, "bad tree block start %llu", start);
5233 return ERR_PTR(-EINVAL);
5234 }
5235
5236 eb = find_extent_buffer(fs_info, start);
5237 if (eb)
5238 return eb;
5239
5240 eb = __alloc_extent_buffer(fs_info, start, len);
5241 if (!eb)
5242 return ERR_PTR(-ENOMEM);
5243
5244 num_pages = num_extent_pages(eb);
5245 for (i = 0; i < num_pages; i++, index++) {
5246 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
5247 if (!p) {
5248 exists = ERR_PTR(-ENOMEM);
5249 goto free_eb;
5250 }
5251
5252 spin_lock(&mapping->private_lock);
5253 if (PagePrivate(p)) {
5254 /*
5255 * We could have already allocated an eb for this page
5256 * and attached one so lets see if we can get a ref on
5257 * the existing eb, and if we can we know it's good and
5258 * we can just return that one, else we know we can just
5259 * overwrite page->private.
5260 */
5261 exists = (struct extent_buffer *)p->private;
5262 if (atomic_inc_not_zero(&exists->refs)) {
5263 spin_unlock(&mapping->private_lock);
5264 unlock_page(p);
5265 put_page(p);
5266 mark_extent_buffer_accessed(exists, p);
5267 goto free_eb;
5268 }
5269 exists = NULL;
5270
5271 /*
5272 * Do this so attach doesn't complain and we need to
5273 * drop the ref the old guy had.
5274 */
5275 ClearPagePrivate(p);
5276 WARN_ON(PageDirty(p));
5277 put_page(p);
5278 }
5279 attach_extent_buffer_page(eb, p);
5280 spin_unlock(&mapping->private_lock);
5281 WARN_ON(PageDirty(p));
5282 eb->pages[i] = p;
5283 if (!PageUptodate(p))
5284 uptodate = 0;
5285
5286 /*
5287 * We can't unlock the pages just yet since the extent buffer
5288 * hasn't been properly inserted in the radix tree, this
5289 * opens a race with btree_releasepage which can free a page
5290 * while we are still filling in all pages for the buffer and
5291 * we could crash.
5292 */
5293 }
5294 if (uptodate)
5295 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5296again:
5297 ret = radix_tree_preload(GFP_NOFS);
5298 if (ret) {
5299 exists = ERR_PTR(ret);
5300 goto free_eb;
5301 }
5302
5303 spin_lock(&fs_info->buffer_lock);
5304 ret = radix_tree_insert(&fs_info->buffer_radix,
5305 start >> PAGE_SHIFT, eb);
5306 spin_unlock(&fs_info->buffer_lock);
5307 radix_tree_preload_end();
5308 if (ret == -EEXIST) {
5309 exists = find_extent_buffer(fs_info, start);
5310 if (exists)
5311 goto free_eb;
5312 else
5313 goto again;
5314 }
5315 /* add one reference for the tree */
5316 check_buffer_tree_ref(eb);
5317 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
5318
5319 /*
5320 * Now it's safe to unlock the pages because any calls to
5321 * btree_releasepage will correctly detect that a page belongs to a
5322 * live buffer and won't free them prematurely.
5323 */
5324 for (i = 0; i < num_pages; i++)
5325 unlock_page(eb->pages[i]);
5326 return eb;
5327
5328free_eb:
5329 WARN_ON(!atomic_dec_and_test(&eb->refs));
5330 for (i = 0; i < num_pages; i++) {
5331 if (eb->pages[i])
5332 unlock_page(eb->pages[i]);
5333 }
5334
5335 btrfs_release_extent_buffer(eb);
5336 return exists;
5337}
5338
5339static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
5340{
5341 struct extent_buffer *eb =
5342 container_of(head, struct extent_buffer, rcu_head);
5343
5344 __free_extent_buffer(eb);
5345}
5346
5347static int release_extent_buffer(struct extent_buffer *eb)
5348 __releases(&eb->refs_lock)
5349{
5350 lockdep_assert_held(&eb->refs_lock);
5351
5352 WARN_ON(atomic_read(&eb->refs) == 0);
5353 if (atomic_dec_and_test(&eb->refs)) {
5354 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
5355 struct btrfs_fs_info *fs_info = eb->fs_info;
5356
5357 spin_unlock(&eb->refs_lock);
5358
5359 spin_lock(&fs_info->buffer_lock);
5360 radix_tree_delete(&fs_info->buffer_radix,
5361 eb->start >> PAGE_SHIFT);
5362 spin_unlock(&fs_info->buffer_lock);
5363 } else {
5364 spin_unlock(&eb->refs_lock);
5365 }
5366
5367 btrfs_leak_debug_del(&eb->fs_info->eb_leak_lock, &eb->leak_list);
5368 /* Should be safe to release our pages at this point */
5369 btrfs_release_extent_buffer_pages(eb);
5370#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
5371 if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
5372 __free_extent_buffer(eb);
5373 return 1;
5374 }
5375#endif
5376 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
5377 return 1;
5378 }
5379 spin_unlock(&eb->refs_lock);
5380
5381 return 0;
5382}
5383
5384void free_extent_buffer(struct extent_buffer *eb)
5385{
5386 int refs;
5387 int old;
5388 if (!eb)
5389 return;
5390
5391 while (1) {
5392 refs = atomic_read(&eb->refs);
5393 if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
5394 || (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
5395 refs == 1))
5396 break;
5397 old = atomic_cmpxchg(&eb->refs, refs, refs - 1);
5398 if (old == refs)
5399 return;
5400 }
5401
5402 spin_lock(&eb->refs_lock);
5403 if (atomic_read(&eb->refs) == 2 &&
5404 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
5405 !extent_buffer_under_io(eb) &&
5406 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5407 atomic_dec(&eb->refs);
5408
5409 /*
5410 * I know this is terrible, but it's temporary until we stop tracking
5411 * the uptodate bits and such for the extent buffers.
5412 */
5413 release_extent_buffer(eb);
5414}
5415
5416void free_extent_buffer_stale(struct extent_buffer *eb)
5417{
5418 if (!eb)
5419 return;
5420
5421 spin_lock(&eb->refs_lock);
5422 set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
5423
5424 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
5425 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
5426 atomic_dec(&eb->refs);
5427 release_extent_buffer(eb);
5428}
5429
5430void clear_extent_buffer_dirty(const struct extent_buffer *eb)
5431{
5432 int i;
5433 int num_pages;
5434 struct page *page;
5435
5436 num_pages = num_extent_pages(eb);
5437
5438 for (i = 0; i < num_pages; i++) {
5439 page = eb->pages[i];
5440 if (!PageDirty(page))
5441 continue;
5442
5443 lock_page(page);
5444 WARN_ON(!PagePrivate(page));
5445
5446 clear_page_dirty_for_io(page);
5447 xa_lock_irq(&page->mapping->i_pages);
5448 if (!PageDirty(page))
5449 __xa_clear_mark(&page->mapping->i_pages,
5450 page_index(page), PAGECACHE_TAG_DIRTY);
5451 xa_unlock_irq(&page->mapping->i_pages);
5452 ClearPageError(page);
5453 unlock_page(page);
5454 }
5455 WARN_ON(atomic_read(&eb->refs) == 0);
5456}
5457
5458bool set_extent_buffer_dirty(struct extent_buffer *eb)
5459{
5460 int i;
5461 int num_pages;
5462 bool was_dirty;
5463
5464 check_buffer_tree_ref(eb);
5465
5466 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
5467
5468 num_pages = num_extent_pages(eb);
5469 WARN_ON(atomic_read(&eb->refs) == 0);
5470 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
5471
5472 if (!was_dirty)
5473 for (i = 0; i < num_pages; i++)
5474 set_page_dirty(eb->pages[i]);
5475
5476#ifdef CONFIG_BTRFS_DEBUG
5477 for (i = 0; i < num_pages; i++)
5478 ASSERT(PageDirty(eb->pages[i]));
5479#endif
5480
5481 return was_dirty;
5482}
5483
5484void clear_extent_buffer_uptodate(struct extent_buffer *eb)
5485{
5486 int i;
5487 struct page *page;
5488 int num_pages;
5489
5490 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5491 num_pages = num_extent_pages(eb);
5492 for (i = 0; i < num_pages; i++) {
5493 page = eb->pages[i];
5494 if (page)
5495 ClearPageUptodate(page);
5496 }
5497}
5498
5499void set_extent_buffer_uptodate(struct extent_buffer *eb)
5500{
5501 int i;
5502 struct page *page;
5503 int num_pages;
5504
5505 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5506 num_pages = num_extent_pages(eb);
5507 for (i = 0; i < num_pages; i++) {
5508 page = eb->pages[i];
5509 SetPageUptodate(page);
5510 }
5511}
5512
5513int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num)
5514{
5515 int i;
5516 struct page *page;
5517 int err;
5518 int ret = 0;
5519 int locked_pages = 0;
5520 int all_uptodate = 1;
5521 int num_pages;
5522 unsigned long num_reads = 0;
5523 struct bio *bio = NULL;
5524 unsigned long bio_flags = 0;
5525
5526 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
5527 return 0;
5528
5529 num_pages = num_extent_pages(eb);
5530 for (i = 0; i < num_pages; i++) {
5531 page = eb->pages[i];
5532 if (wait == WAIT_NONE) {
5533 if (!trylock_page(page))
5534 goto unlock_exit;
5535 } else {
5536 lock_page(page);
5537 }
5538 locked_pages++;
5539 }
5540 /*
5541 * We need to firstly lock all pages to make sure that
5542 * the uptodate bit of our pages won't be affected by
5543 * clear_extent_buffer_uptodate().
5544 */
5545 for (i = 0; i < num_pages; i++) {
5546 page = eb->pages[i];
5547 if (!PageUptodate(page)) {
5548 num_reads++;
5549 all_uptodate = 0;
5550 }
5551 }
5552
5553 if (all_uptodate) {
5554 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
5555 goto unlock_exit;
5556 }
5557
5558 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
5559 eb->read_mirror = 0;
5560 atomic_set(&eb->io_pages, num_reads);
5561 /*
5562 * It is possible for releasepage to clear the TREE_REF bit before we
5563 * set io_pages. See check_buffer_tree_ref for a more detailed comment.
5564 */
5565 check_buffer_tree_ref(eb);
5566 for (i = 0; i < num_pages; i++) {
5567 page = eb->pages[i];
5568
5569 if (!PageUptodate(page)) {
5570 if (ret) {
5571 atomic_dec(&eb->io_pages);
5572 unlock_page(page);
5573 continue;
5574 }
5575
5576 ClearPageError(page);
5577 err = __extent_read_full_page(page,
5578 btree_get_extent, &bio,
5579 mirror_num, &bio_flags,
5580 REQ_META);
5581 if (err) {
5582 ret = err;
5583 /*
5584 * We use &bio in above __extent_read_full_page,
5585 * so we ensure that if it returns error, the
5586 * current page fails to add itself to bio and
5587 * it's been unlocked.
5588 *
5589 * We must dec io_pages by ourselves.
5590 */
5591 atomic_dec(&eb->io_pages);
5592 }
5593 } else {
5594 unlock_page(page);
5595 }
5596 }
5597
5598 if (bio) {
5599 err = submit_one_bio(bio, mirror_num, bio_flags);
5600 if (err)
5601 return err;
5602 }
5603
5604 if (ret || wait != WAIT_COMPLETE)
5605 return ret;
5606
5607 for (i = 0; i < num_pages; i++) {
5608 page = eb->pages[i];
5609 wait_on_page_locked(page);
5610 if (!PageUptodate(page))
5611 ret = -EIO;
5612 }
5613
5614 return ret;
5615
5616unlock_exit:
5617 while (locked_pages > 0) {
5618 locked_pages--;
5619 page = eb->pages[locked_pages];
5620 unlock_page(page);
5621 }
5622 return ret;
5623}
5624
5625void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
5626 unsigned long start, unsigned long len)
5627{
5628 size_t cur;
5629 size_t offset;
5630 struct page *page;
5631 char *kaddr;
5632 char *dst = (char *)dstv;
5633 unsigned long i = start >> PAGE_SHIFT;
5634
5635 if (start + len > eb->len) {
5636 WARN(1, KERN_ERR "btrfs bad mapping eb start %llu len %lu, wanted %lu %lu\n",
5637 eb->start, eb->len, start, len);
5638 memset(dst, 0, len);
5639 return;
5640 }
5641
5642 offset = offset_in_page(start);
5643
5644 while (len > 0) {
5645 page = eb->pages[i];
5646
5647 cur = min(len, (PAGE_SIZE - offset));
5648 kaddr = page_address(page);
5649 memcpy(dst, kaddr + offset, cur);
5650
5651 dst += cur;
5652 len -= cur;
5653 offset = 0;
5654 i++;
5655 }
5656}
5657
5658int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
5659 void __user *dstv,
5660 unsigned long start, unsigned long len)
5661{
5662 size_t cur;
5663 size_t offset;
5664 struct page *page;
5665 char *kaddr;
5666 char __user *dst = (char __user *)dstv;
5667 unsigned long i = start >> PAGE_SHIFT;
5668 int ret = 0;
5669
5670 WARN_ON(start > eb->len);
5671 WARN_ON(start + len > eb->start + eb->len);
5672
5673 offset = offset_in_page(start);
5674
5675 while (len > 0) {
5676 page = eb->pages[i];
5677
5678 cur = min(len, (PAGE_SIZE - offset));
5679 kaddr = page_address(page);
5680 if (copy_to_user_nofault(dst, kaddr + offset, cur)) {
5681 ret = -EFAULT;
5682 break;
5683 }
5684
5685 dst += cur;
5686 len -= cur;
5687 offset = 0;
5688 i++;
5689 }
5690
5691 return ret;
5692}
5693
5694int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
5695 unsigned long start, unsigned long len)
5696{
5697 size_t cur;
5698 size_t offset;
5699 struct page *page;
5700 char *kaddr;
5701 char *ptr = (char *)ptrv;
5702 unsigned long i = start >> PAGE_SHIFT;
5703 int ret = 0;
5704
5705 WARN_ON(start > eb->len);
5706 WARN_ON(start + len > eb->start + eb->len);
5707
5708 offset = offset_in_page(start);
5709
5710 while (len > 0) {
5711 page = eb->pages[i];
5712
5713 cur = min(len, (PAGE_SIZE - offset));
5714
5715 kaddr = page_address(page);
5716 ret = memcmp(ptr, kaddr + offset, cur);
5717 if (ret)
5718 break;
5719
5720 ptr += cur;
5721 len -= cur;
5722 offset = 0;
5723 i++;
5724 }
5725 return ret;
5726}
5727
5728void write_extent_buffer_chunk_tree_uuid(const struct extent_buffer *eb,
5729 const void *srcv)
5730{
5731 char *kaddr;
5732
5733 WARN_ON(!PageUptodate(eb->pages[0]));
5734 kaddr = page_address(eb->pages[0]);
5735 memcpy(kaddr + offsetof(struct btrfs_header, chunk_tree_uuid), srcv,
5736 BTRFS_FSID_SIZE);
5737}
5738
5739void write_extent_buffer_fsid(const struct extent_buffer *eb, const void *srcv)
5740{
5741 char *kaddr;
5742
5743 WARN_ON(!PageUptodate(eb->pages[0]));
5744 kaddr = page_address(eb->pages[0]);
5745 memcpy(kaddr + offsetof(struct btrfs_header, fsid), srcv,
5746 BTRFS_FSID_SIZE);
5747}
5748
5749void write_extent_buffer(const struct extent_buffer *eb, const void *srcv,
5750 unsigned long start, unsigned long len)
5751{
5752 size_t cur;
5753 size_t offset;
5754 struct page *page;
5755 char *kaddr;
5756 char *src = (char *)srcv;
5757 unsigned long i = start >> PAGE_SHIFT;
5758
5759 WARN_ON(start > eb->len);
5760 WARN_ON(start + len > eb->start + eb->len);
5761
5762 offset = offset_in_page(start);
5763
5764 while (len > 0) {
5765 page = eb->pages[i];
5766 WARN_ON(!PageUptodate(page));
5767
5768 cur = min(len, PAGE_SIZE - offset);
5769 kaddr = page_address(page);
5770 memcpy(kaddr + offset, src, cur);
5771
5772 src += cur;
5773 len -= cur;
5774 offset = 0;
5775 i++;
5776 }
5777}
5778
5779void memzero_extent_buffer(const struct extent_buffer *eb, unsigned long start,
5780 unsigned long len)
5781{
5782 size_t cur;
5783 size_t offset;
5784 struct page *page;
5785 char *kaddr;
5786 unsigned long i = start >> PAGE_SHIFT;
5787
5788 WARN_ON(start > eb->len);
5789 WARN_ON(start + len > eb->start + eb->len);
5790
5791 offset = offset_in_page(start);
5792
5793 while (len > 0) {
5794 page = eb->pages[i];
5795 WARN_ON(!PageUptodate(page));
5796
5797 cur = min(len, PAGE_SIZE - offset);
5798 kaddr = page_address(page);
5799 memset(kaddr + offset, 0, cur);
5800
5801 len -= cur;
5802 offset = 0;
5803 i++;
5804 }
5805}
5806
5807void copy_extent_buffer_full(const struct extent_buffer *dst,
5808 const struct extent_buffer *src)
5809{
5810 int i;
5811 int num_pages;
5812
5813 ASSERT(dst->len == src->len);
5814
5815 num_pages = num_extent_pages(dst);
5816 for (i = 0; i < num_pages; i++)
5817 copy_page(page_address(dst->pages[i]),
5818 page_address(src->pages[i]));
5819}
5820
5821void copy_extent_buffer(const struct extent_buffer *dst,
5822 const struct extent_buffer *src,
5823 unsigned long dst_offset, unsigned long src_offset,
5824 unsigned long len)
5825{
5826 u64 dst_len = dst->len;
5827 size_t cur;
5828 size_t offset;
5829 struct page *page;
5830 char *kaddr;
5831 unsigned long i = dst_offset >> PAGE_SHIFT;
5832
5833 WARN_ON(src->len != dst_len);
5834
5835 offset = offset_in_page(dst_offset);
5836
5837 while (len > 0) {
5838 page = dst->pages[i];
5839 WARN_ON(!PageUptodate(page));
5840
5841 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5842
5843 kaddr = page_address(page);
5844 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5845
5846 src_offset += cur;
5847 len -= cur;
5848 offset = 0;
5849 i++;
5850 }
5851}
5852
5853/*
5854 * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5855 * given bit number
5856 * @eb: the extent buffer
5857 * @start: offset of the bitmap item in the extent buffer
5858 * @nr: bit number
5859 * @page_index: return index of the page in the extent buffer that contains the
5860 * given bit number
5861 * @page_offset: return offset into the page given by page_index
5862 *
5863 * This helper hides the ugliness of finding the byte in an extent buffer which
5864 * contains a given bit.
5865 */
5866static inline void eb_bitmap_offset(const struct extent_buffer *eb,
5867 unsigned long start, unsigned long nr,
5868 unsigned long *page_index,
5869 size_t *page_offset)
5870{
5871 size_t byte_offset = BIT_BYTE(nr);
5872 size_t offset;
5873
5874 /*
5875 * The byte we want is the offset of the extent buffer + the offset of
5876 * the bitmap item in the extent buffer + the offset of the byte in the
5877 * bitmap item.
5878 */
5879 offset = start + byte_offset;
5880
5881 *page_index = offset >> PAGE_SHIFT;
5882 *page_offset = offset_in_page(offset);
5883}
5884
5885/**
5886 * extent_buffer_test_bit - determine whether a bit in a bitmap item is set
5887 * @eb: the extent buffer
5888 * @start: offset of the bitmap item in the extent buffer
5889 * @nr: bit number to test
5890 */
5891int extent_buffer_test_bit(const struct extent_buffer *eb, unsigned long start,
5892 unsigned long nr)
5893{
5894 u8 *kaddr;
5895 struct page *page;
5896 unsigned long i;
5897 size_t offset;
5898
5899 eb_bitmap_offset(eb, start, nr, &i, &offset);
5900 page = eb->pages[i];
5901 WARN_ON(!PageUptodate(page));
5902 kaddr = page_address(page);
5903 return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5904}
5905
5906/**
5907 * extent_buffer_bitmap_set - set an area of a bitmap
5908 * @eb: the extent buffer
5909 * @start: offset of the bitmap item in the extent buffer
5910 * @pos: bit number of the first bit
5911 * @len: number of bits to set
5912 */
5913void extent_buffer_bitmap_set(const struct extent_buffer *eb, unsigned long start,
5914 unsigned long pos, unsigned long len)
5915{
5916 u8 *kaddr;
5917 struct page *page;
5918 unsigned long i;
5919 size_t offset;
5920 const unsigned int size = pos + len;
5921 int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5922 u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5923
5924 eb_bitmap_offset(eb, start, pos, &i, &offset);
5925 page = eb->pages[i];
5926 WARN_ON(!PageUptodate(page));
5927 kaddr = page_address(page);
5928
5929 while (len >= bits_to_set) {
5930 kaddr[offset] |= mask_to_set;
5931 len -= bits_to_set;
5932 bits_to_set = BITS_PER_BYTE;
5933 mask_to_set = ~0;
5934 if (++offset >= PAGE_SIZE && len > 0) {
5935 offset = 0;
5936 page = eb->pages[++i];
5937 WARN_ON(!PageUptodate(page));
5938 kaddr = page_address(page);
5939 }
5940 }
5941 if (len) {
5942 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5943 kaddr[offset] |= mask_to_set;
5944 }
5945}
5946
5947
5948/**
5949 * extent_buffer_bitmap_clear - clear an area of a bitmap
5950 * @eb: the extent buffer
5951 * @start: offset of the bitmap item in the extent buffer
5952 * @pos: bit number of the first bit
5953 * @len: number of bits to clear
5954 */
5955void extent_buffer_bitmap_clear(const struct extent_buffer *eb,
5956 unsigned long start, unsigned long pos,
5957 unsigned long len)
5958{
5959 u8 *kaddr;
5960 struct page *page;
5961 unsigned long i;
5962 size_t offset;
5963 const unsigned int size = pos + len;
5964 int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5965 u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5966
5967 eb_bitmap_offset(eb, start, pos, &i, &offset);
5968 page = eb->pages[i];
5969 WARN_ON(!PageUptodate(page));
5970 kaddr = page_address(page);
5971
5972 while (len >= bits_to_clear) {
5973 kaddr[offset] &= ~mask_to_clear;
5974 len -= bits_to_clear;
5975 bits_to_clear = BITS_PER_BYTE;
5976 mask_to_clear = ~0;
5977 if (++offset >= PAGE_SIZE && len > 0) {
5978 offset = 0;
5979 page = eb->pages[++i];
5980 WARN_ON(!PageUptodate(page));
5981 kaddr = page_address(page);
5982 }
5983 }
5984 if (len) {
5985 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5986 kaddr[offset] &= ~mask_to_clear;
5987 }
5988}
5989
5990static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5991{
5992 unsigned long distance = (src > dst) ? src - dst : dst - src;
5993 return distance < len;
5994}
5995
5996static void copy_pages(struct page *dst_page, struct page *src_page,
5997 unsigned long dst_off, unsigned long src_off,
5998 unsigned long len)
5999{
6000 char *dst_kaddr = page_address(dst_page);
6001 char *src_kaddr;
6002 int must_memmove = 0;
6003
6004 if (dst_page != src_page) {
6005 src_kaddr = page_address(src_page);
6006 } else {
6007 src_kaddr = dst_kaddr;
6008 if (areas_overlap(src_off, dst_off, len))
6009 must_memmove = 1;
6010 }
6011
6012 if (must_memmove)
6013 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
6014 else
6015 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
6016}
6017
6018void memcpy_extent_buffer(const struct extent_buffer *dst,
6019 unsigned long dst_offset, unsigned long src_offset,
6020 unsigned long len)
6021{
6022 struct btrfs_fs_info *fs_info = dst->fs_info;
6023 size_t cur;
6024 size_t dst_off_in_page;
6025 size_t src_off_in_page;
6026 unsigned long dst_i;
6027 unsigned long src_i;
6028
6029 if (src_offset + len > dst->len) {
6030 btrfs_err(fs_info,
6031 "memmove bogus src_offset %lu move len %lu dst len %lu",
6032 src_offset, len, dst->len);
6033 BUG();
6034 }
6035 if (dst_offset + len > dst->len) {
6036 btrfs_err(fs_info,
6037 "memmove bogus dst_offset %lu move len %lu dst len %lu",
6038 dst_offset, len, dst->len);
6039 BUG();
6040 }
6041
6042 while (len > 0) {
6043 dst_off_in_page = offset_in_page(dst_offset);
6044 src_off_in_page = offset_in_page(src_offset);
6045
6046 dst_i = dst_offset >> PAGE_SHIFT;
6047 src_i = src_offset >> PAGE_SHIFT;
6048
6049 cur = min(len, (unsigned long)(PAGE_SIZE -
6050 src_off_in_page));
6051 cur = min_t(unsigned long, cur,
6052 (unsigned long)(PAGE_SIZE - dst_off_in_page));
6053
6054 copy_pages(dst->pages[dst_i], dst->pages[src_i],
6055 dst_off_in_page, src_off_in_page, cur);
6056
6057 src_offset += cur;
6058 dst_offset += cur;
6059 len -= cur;
6060 }
6061}
6062
6063void memmove_extent_buffer(const struct extent_buffer *dst,
6064 unsigned long dst_offset, unsigned long src_offset,
6065 unsigned long len)
6066{
6067 struct btrfs_fs_info *fs_info = dst->fs_info;
6068 size_t cur;
6069 size_t dst_off_in_page;
6070 size_t src_off_in_page;
6071 unsigned long dst_end = dst_offset + len - 1;
6072 unsigned long src_end = src_offset + len - 1;
6073 unsigned long dst_i;
6074 unsigned long src_i;
6075
6076 if (src_offset + len > dst->len) {
6077 btrfs_err(fs_info,
6078 "memmove bogus src_offset %lu move len %lu len %lu",
6079 src_offset, len, dst->len);
6080 BUG();
6081 }
6082 if (dst_offset + len > dst->len) {
6083 btrfs_err(fs_info,
6084 "memmove bogus dst_offset %lu move len %lu len %lu",
6085 dst_offset, len, dst->len);
6086 BUG();
6087 }
6088 if (dst_offset < src_offset) {
6089 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
6090 return;
6091 }
6092 while (len > 0) {
6093 dst_i = dst_end >> PAGE_SHIFT;
6094 src_i = src_end >> PAGE_SHIFT;
6095
6096 dst_off_in_page = offset_in_page(dst_end);
6097 src_off_in_page = offset_in_page(src_end);
6098
6099 cur = min_t(unsigned long, len, src_off_in_page + 1);
6100 cur = min(cur, dst_off_in_page + 1);
6101 copy_pages(dst->pages[dst_i], dst->pages[src_i],
6102 dst_off_in_page - cur + 1,
6103 src_off_in_page - cur + 1, cur);
6104
6105 dst_end -= cur;
6106 src_end -= cur;
6107 len -= cur;
6108 }
6109}
6110
6111int try_release_extent_buffer(struct page *page)
6112{
6113 struct extent_buffer *eb;
6114
6115 /*
6116 * We need to make sure nobody is attaching this page to an eb right
6117 * now.
6118 */
6119 spin_lock(&page->mapping->private_lock);
6120 if (!PagePrivate(page)) {
6121 spin_unlock(&page->mapping->private_lock);
6122 return 1;
6123 }
6124
6125 eb = (struct extent_buffer *)page->private;
6126 BUG_ON(!eb);
6127
6128 /*
6129 * This is a little awful but should be ok, we need to make sure that
6130 * the eb doesn't disappear out from under us while we're looking at
6131 * this page.
6132 */
6133 spin_lock(&eb->refs_lock);
6134 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
6135 spin_unlock(&eb->refs_lock);
6136 spin_unlock(&page->mapping->private_lock);
6137 return 0;
6138 }
6139 spin_unlock(&page->mapping->private_lock);
6140
6141 /*
6142 * If tree ref isn't set then we know the ref on this eb is a real ref,
6143 * so just return, this page will likely be freed soon anyway.
6144 */
6145 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
6146 spin_unlock(&eb->refs_lock);
6147 return 0;
6148 }
6149
6150 return release_extent_buffer(eb);
6151}
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 "misc.h"
18#include "extent_io.h"
19#include "extent-io-tree.h"
20#include "extent_map.h"
21#include "ctree.h"
22#include "btrfs_inode.h"
23#include "bio.h"
24#include "locking.h"
25#include "rcu-string.h"
26#include "backref.h"
27#include "disk-io.h"
28#include "subpage.h"
29#include "zoned.h"
30#include "block-group.h"
31#include "compression.h"
32#include "fs.h"
33#include "accessors.h"
34#include "file-item.h"
35#include "file.h"
36#include "dev-replace.h"
37#include "super.h"
38#include "transaction.h"
39
40static struct kmem_cache *extent_buffer_cache;
41
42#ifdef CONFIG_BTRFS_DEBUG
43static inline void btrfs_leak_debug_add_eb(struct extent_buffer *eb)
44{
45 struct btrfs_fs_info *fs_info = eb->fs_info;
46 unsigned long flags;
47
48 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
49 list_add(&eb->leak_list, &fs_info->allocated_ebs);
50 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
51}
52
53static inline void btrfs_leak_debug_del_eb(struct extent_buffer *eb)
54{
55 struct btrfs_fs_info *fs_info = eb->fs_info;
56 unsigned long flags;
57
58 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
59 list_del(&eb->leak_list);
60 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
61}
62
63void btrfs_extent_buffer_leak_debug_check(struct btrfs_fs_info *fs_info)
64{
65 struct extent_buffer *eb;
66 unsigned long flags;
67
68 /*
69 * If we didn't get into open_ctree our allocated_ebs will not be
70 * initialized, so just skip this.
71 */
72 if (!fs_info->allocated_ebs.next)
73 return;
74
75 WARN_ON(!list_empty(&fs_info->allocated_ebs));
76 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
77 while (!list_empty(&fs_info->allocated_ebs)) {
78 eb = list_first_entry(&fs_info->allocated_ebs,
79 struct extent_buffer, leak_list);
80 pr_err(
81 "BTRFS: buffer leak start %llu len %lu refs %d bflags %lu owner %llu\n",
82 eb->start, eb->len, atomic_read(&eb->refs), eb->bflags,
83 btrfs_header_owner(eb));
84 list_del(&eb->leak_list);
85 kmem_cache_free(extent_buffer_cache, eb);
86 }
87 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
88}
89#else
90#define btrfs_leak_debug_add_eb(eb) do {} while (0)
91#define btrfs_leak_debug_del_eb(eb) do {} while (0)
92#endif
93
94/*
95 * Structure to record info about the bio being assembled, and other info like
96 * how many bytes are there before stripe/ordered extent boundary.
97 */
98struct btrfs_bio_ctrl {
99 struct btrfs_bio *bbio;
100 enum btrfs_compression_type compress_type;
101 u32 len_to_oe_boundary;
102 blk_opf_t opf;
103 btrfs_bio_end_io_t end_io_func;
104 struct writeback_control *wbc;
105};
106
107static void submit_one_bio(struct btrfs_bio_ctrl *bio_ctrl)
108{
109 struct btrfs_bio *bbio = bio_ctrl->bbio;
110
111 if (!bbio)
112 return;
113
114 /* Caller should ensure the bio has at least some range added */
115 ASSERT(bbio->bio.bi_iter.bi_size);
116
117 if (btrfs_op(&bbio->bio) == BTRFS_MAP_READ &&
118 bio_ctrl->compress_type != BTRFS_COMPRESS_NONE)
119 btrfs_submit_compressed_read(bbio);
120 else
121 btrfs_submit_bio(bbio, 0);
122
123 /* The bbio is owned by the end_io handler now */
124 bio_ctrl->bbio = NULL;
125}
126
127/*
128 * Submit or fail the current bio in the bio_ctrl structure.
129 */
130static void submit_write_bio(struct btrfs_bio_ctrl *bio_ctrl, int ret)
131{
132 struct btrfs_bio *bbio = bio_ctrl->bbio;
133
134 if (!bbio)
135 return;
136
137 if (ret) {
138 ASSERT(ret < 0);
139 btrfs_bio_end_io(bbio, errno_to_blk_status(ret));
140 /* The bio is owned by the end_io handler now */
141 bio_ctrl->bbio = NULL;
142 } else {
143 submit_one_bio(bio_ctrl);
144 }
145}
146
147int __init extent_buffer_init_cachep(void)
148{
149 extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
150 sizeof(struct extent_buffer), 0,
151 SLAB_MEM_SPREAD, NULL);
152 if (!extent_buffer_cache)
153 return -ENOMEM;
154
155 return 0;
156}
157
158void __cold extent_buffer_free_cachep(void)
159{
160 /*
161 * Make sure all delayed rcu free are flushed before we
162 * destroy caches.
163 */
164 rcu_barrier();
165 kmem_cache_destroy(extent_buffer_cache);
166}
167
168void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
169{
170 unsigned long index = start >> PAGE_SHIFT;
171 unsigned long end_index = end >> PAGE_SHIFT;
172 struct page *page;
173
174 while (index <= end_index) {
175 page = find_get_page(inode->i_mapping, index);
176 BUG_ON(!page); /* Pages should be in the extent_io_tree */
177 clear_page_dirty_for_io(page);
178 put_page(page);
179 index++;
180 }
181}
182
183static void process_one_page(struct btrfs_fs_info *fs_info,
184 struct page *page, struct page *locked_page,
185 unsigned long page_ops, u64 start, u64 end)
186{
187 struct folio *folio = page_folio(page);
188 u32 len;
189
190 ASSERT(end + 1 - start != 0 && end + 1 - start < U32_MAX);
191 len = end + 1 - start;
192
193 if (page_ops & PAGE_SET_ORDERED)
194 btrfs_folio_clamp_set_ordered(fs_info, folio, start, len);
195 if (page_ops & PAGE_START_WRITEBACK) {
196 btrfs_folio_clamp_clear_dirty(fs_info, folio, start, len);
197 btrfs_folio_clamp_set_writeback(fs_info, folio, start, len);
198 }
199 if (page_ops & PAGE_END_WRITEBACK)
200 btrfs_folio_clamp_clear_writeback(fs_info, folio, start, len);
201
202 if (page != locked_page && (page_ops & PAGE_UNLOCK))
203 btrfs_folio_end_writer_lock(fs_info, folio, start, len);
204}
205
206static void __process_pages_contig(struct address_space *mapping,
207 struct page *locked_page, u64 start, u64 end,
208 unsigned long page_ops)
209{
210 struct btrfs_fs_info *fs_info = btrfs_sb(mapping->host->i_sb);
211 pgoff_t start_index = start >> PAGE_SHIFT;
212 pgoff_t end_index = end >> PAGE_SHIFT;
213 pgoff_t index = start_index;
214 struct folio_batch fbatch;
215 int i;
216
217 folio_batch_init(&fbatch);
218 while (index <= end_index) {
219 int found_folios;
220
221 found_folios = filemap_get_folios_contig(mapping, &index,
222 end_index, &fbatch);
223 for (i = 0; i < found_folios; i++) {
224 struct folio *folio = fbatch.folios[i];
225
226 process_one_page(fs_info, &folio->page, locked_page,
227 page_ops, start, end);
228 }
229 folio_batch_release(&fbatch);
230 cond_resched();
231 }
232}
233
234static noinline void __unlock_for_delalloc(struct inode *inode,
235 struct page *locked_page,
236 u64 start, u64 end)
237{
238 unsigned long index = start >> PAGE_SHIFT;
239 unsigned long end_index = end >> PAGE_SHIFT;
240
241 ASSERT(locked_page);
242 if (index == locked_page->index && end_index == index)
243 return;
244
245 __process_pages_contig(inode->i_mapping, locked_page, start, end,
246 PAGE_UNLOCK);
247}
248
249static noinline int lock_delalloc_pages(struct inode *inode,
250 struct page *locked_page,
251 u64 start,
252 u64 end)
253{
254 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
255 struct address_space *mapping = inode->i_mapping;
256 pgoff_t start_index = start >> PAGE_SHIFT;
257 pgoff_t end_index = end >> PAGE_SHIFT;
258 pgoff_t index = start_index;
259 u64 processed_end = start;
260 struct folio_batch fbatch;
261
262 if (index == locked_page->index && index == end_index)
263 return 0;
264
265 folio_batch_init(&fbatch);
266 while (index <= end_index) {
267 unsigned int found_folios, i;
268
269 found_folios = filemap_get_folios_contig(mapping, &index,
270 end_index, &fbatch);
271 if (found_folios == 0)
272 goto out;
273
274 for (i = 0; i < found_folios; i++) {
275 struct folio *folio = fbatch.folios[i];
276 struct page *page = folio_page(folio, 0);
277 u32 len = end + 1 - start;
278
279 if (page == locked_page)
280 continue;
281
282 if (btrfs_folio_start_writer_lock(fs_info, folio, start,
283 len))
284 goto out;
285
286 if (!PageDirty(page) || page->mapping != mapping) {
287 btrfs_folio_end_writer_lock(fs_info, folio, start,
288 len);
289 goto out;
290 }
291
292 processed_end = page_offset(page) + PAGE_SIZE - 1;
293 }
294 folio_batch_release(&fbatch);
295 cond_resched();
296 }
297
298 return 0;
299out:
300 folio_batch_release(&fbatch);
301 if (processed_end > start)
302 __unlock_for_delalloc(inode, locked_page, start, processed_end);
303 return -EAGAIN;
304}
305
306/*
307 * Find and lock a contiguous range of bytes in the file marked as delalloc, no
308 * more than @max_bytes.
309 *
310 * @start: The original start bytenr to search.
311 * Will store the extent range start bytenr.
312 * @end: The original end bytenr of the search range
313 * Will store the extent range end bytenr.
314 *
315 * Return true if we find a delalloc range which starts inside the original
316 * range, and @start/@end will store the delalloc range start/end.
317 *
318 * Return false if we can't find any delalloc range which starts inside the
319 * original range, and @start/@end will be the non-delalloc range start/end.
320 */
321EXPORT_FOR_TESTS
322noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
323 struct page *locked_page, u64 *start,
324 u64 *end)
325{
326 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
327 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
328 const u64 orig_start = *start;
329 const u64 orig_end = *end;
330 /* The sanity tests may not set a valid fs_info. */
331 u64 max_bytes = fs_info ? fs_info->max_extent_size : BTRFS_MAX_EXTENT_SIZE;
332 u64 delalloc_start;
333 u64 delalloc_end;
334 bool found;
335 struct extent_state *cached_state = NULL;
336 int ret;
337 int loops = 0;
338
339 /* Caller should pass a valid @end to indicate the search range end */
340 ASSERT(orig_end > orig_start);
341
342 /* The range should at least cover part of the page */
343 ASSERT(!(orig_start >= page_offset(locked_page) + PAGE_SIZE ||
344 orig_end <= page_offset(locked_page)));
345again:
346 /* step one, find a bunch of delalloc bytes starting at start */
347 delalloc_start = *start;
348 delalloc_end = 0;
349 found = btrfs_find_delalloc_range(tree, &delalloc_start, &delalloc_end,
350 max_bytes, &cached_state);
351 if (!found || delalloc_end <= *start || delalloc_start > orig_end) {
352 *start = delalloc_start;
353
354 /* @delalloc_end can be -1, never go beyond @orig_end */
355 *end = min(delalloc_end, orig_end);
356 free_extent_state(cached_state);
357 return false;
358 }
359
360 /*
361 * start comes from the offset of locked_page. We have to lock
362 * pages in order, so we can't process delalloc bytes before
363 * locked_page
364 */
365 if (delalloc_start < *start)
366 delalloc_start = *start;
367
368 /*
369 * make sure to limit the number of pages we try to lock down
370 */
371 if (delalloc_end + 1 - delalloc_start > max_bytes)
372 delalloc_end = delalloc_start + max_bytes - 1;
373
374 /* step two, lock all the pages after the page that has start */
375 ret = lock_delalloc_pages(inode, locked_page,
376 delalloc_start, delalloc_end);
377 ASSERT(!ret || ret == -EAGAIN);
378 if (ret == -EAGAIN) {
379 /* some of the pages are gone, lets avoid looping by
380 * shortening the size of the delalloc range we're searching
381 */
382 free_extent_state(cached_state);
383 cached_state = NULL;
384 if (!loops) {
385 max_bytes = PAGE_SIZE;
386 loops = 1;
387 goto again;
388 } else {
389 found = false;
390 goto out_failed;
391 }
392 }
393
394 /* step three, lock the state bits for the whole range */
395 lock_extent(tree, delalloc_start, delalloc_end, &cached_state);
396
397 /* then test to make sure it is all still delalloc */
398 ret = test_range_bit(tree, delalloc_start, delalloc_end,
399 EXTENT_DELALLOC, cached_state);
400 if (!ret) {
401 unlock_extent(tree, delalloc_start, delalloc_end,
402 &cached_state);
403 __unlock_for_delalloc(inode, locked_page,
404 delalloc_start, delalloc_end);
405 cond_resched();
406 goto again;
407 }
408 free_extent_state(cached_state);
409 *start = delalloc_start;
410 *end = delalloc_end;
411out_failed:
412 return found;
413}
414
415void extent_clear_unlock_delalloc(struct btrfs_inode *inode, u64 start, u64 end,
416 struct page *locked_page,
417 u32 clear_bits, unsigned long page_ops)
418{
419 clear_extent_bit(&inode->io_tree, start, end, clear_bits, NULL);
420
421 __process_pages_contig(inode->vfs_inode.i_mapping, locked_page,
422 start, end, page_ops);
423}
424
425static bool btrfs_verify_page(struct page *page, u64 start)
426{
427 if (!fsverity_active(page->mapping->host) ||
428 PageUptodate(page) ||
429 start >= i_size_read(page->mapping->host))
430 return true;
431 return fsverity_verify_page(page);
432}
433
434static void end_page_read(struct page *page, bool uptodate, u64 start, u32 len)
435{
436 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
437 struct folio *folio = page_folio(page);
438
439 ASSERT(page_offset(page) <= start &&
440 start + len <= page_offset(page) + PAGE_SIZE);
441
442 if (uptodate && btrfs_verify_page(page, start))
443 btrfs_folio_set_uptodate(fs_info, folio, start, len);
444 else
445 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
446
447 if (!btrfs_is_subpage(fs_info, page->mapping))
448 unlock_page(page);
449 else
450 btrfs_subpage_end_reader(fs_info, folio, start, len);
451}
452
453/*
454 * After a write IO is done, we need to:
455 *
456 * - clear the uptodate bits on error
457 * - clear the writeback bits in the extent tree for the range
458 * - filio_end_writeback() if there is no more pending io for the folio
459 *
460 * Scheduling is not allowed, so the extent state tree is expected
461 * to have one and only one object corresponding to this IO.
462 */
463static void end_bbio_data_write(struct btrfs_bio *bbio)
464{
465 struct bio *bio = &bbio->bio;
466 int error = blk_status_to_errno(bio->bi_status);
467 struct folio_iter fi;
468
469 ASSERT(!bio_flagged(bio, BIO_CLONED));
470 bio_for_each_folio_all(fi, bio) {
471 struct folio *folio = fi.folio;
472 struct inode *inode = folio->mapping->host;
473 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
474 const u32 sectorsize = fs_info->sectorsize;
475 u64 start = folio_pos(folio) + fi.offset;
476 u32 len = fi.length;
477
478 /* Only order 0 (single page) folios are allowed for data. */
479 ASSERT(folio_order(folio) == 0);
480
481 /* Our read/write should always be sector aligned. */
482 if (!IS_ALIGNED(fi.offset, sectorsize))
483 btrfs_err(fs_info,
484 "partial page write in btrfs with offset %zu and length %zu",
485 fi.offset, fi.length);
486 else if (!IS_ALIGNED(fi.length, sectorsize))
487 btrfs_info(fs_info,
488 "incomplete page write with offset %zu and length %zu",
489 fi.offset, fi.length);
490
491 btrfs_finish_ordered_extent(bbio->ordered,
492 folio_page(folio, 0), start, len, !error);
493 if (error)
494 mapping_set_error(folio->mapping, error);
495 btrfs_folio_clear_writeback(fs_info, folio, start, len);
496 }
497
498 bio_put(bio);
499}
500
501/*
502 * Record previously processed extent range
503 *
504 * For endio_readpage_release_extent() to handle a full extent range, reducing
505 * the extent io operations.
506 */
507struct processed_extent {
508 struct btrfs_inode *inode;
509 /* Start of the range in @inode */
510 u64 start;
511 /* End of the range in @inode */
512 u64 end;
513 bool uptodate;
514};
515
516/*
517 * Try to release processed extent range
518 *
519 * May not release the extent range right now if the current range is
520 * contiguous to processed extent.
521 *
522 * Will release processed extent when any of @inode, @uptodate, the range is
523 * no longer contiguous to the processed range.
524 *
525 * Passing @inode == NULL will force processed extent to be released.
526 */
527static void endio_readpage_release_extent(struct processed_extent *processed,
528 struct btrfs_inode *inode, u64 start, u64 end,
529 bool uptodate)
530{
531 struct extent_state *cached = NULL;
532 struct extent_io_tree *tree;
533
534 /* The first extent, initialize @processed */
535 if (!processed->inode)
536 goto update;
537
538 /*
539 * Contiguous to processed extent, just uptodate the end.
540 *
541 * Several things to notice:
542 *
543 * - bio can be merged as long as on-disk bytenr is contiguous
544 * This means we can have page belonging to other inodes, thus need to
545 * check if the inode still matches.
546 * - bvec can contain range beyond current page for multi-page bvec
547 * Thus we need to do processed->end + 1 >= start check
548 */
549 if (processed->inode == inode && processed->uptodate == uptodate &&
550 processed->end + 1 >= start && end >= processed->end) {
551 processed->end = end;
552 return;
553 }
554
555 tree = &processed->inode->io_tree;
556 /*
557 * Now we don't have range contiguous to the processed range, release
558 * the processed range now.
559 */
560 unlock_extent(tree, processed->start, processed->end, &cached);
561
562update:
563 /* Update processed to current range */
564 processed->inode = inode;
565 processed->start = start;
566 processed->end = end;
567 processed->uptodate = uptodate;
568}
569
570static void begin_page_read(struct btrfs_fs_info *fs_info, struct page *page)
571{
572 struct folio *folio = page_folio(page);
573
574 ASSERT(folio_test_locked(folio));
575 if (!btrfs_is_subpage(fs_info, folio->mapping))
576 return;
577
578 ASSERT(folio_test_private(folio));
579 btrfs_subpage_start_reader(fs_info, folio, page_offset(page), PAGE_SIZE);
580}
581
582/*
583 * After a data read IO is done, we need to:
584 *
585 * - clear the uptodate bits on error
586 * - set the uptodate bits if things worked
587 * - set the folio up to date if all extents in the tree are uptodate
588 * - clear the lock bit in the extent tree
589 * - unlock the folio if there are no other extents locked for it
590 *
591 * Scheduling is not allowed, so the extent state tree is expected
592 * to have one and only one object corresponding to this IO.
593 */
594static void end_bbio_data_read(struct btrfs_bio *bbio)
595{
596 struct bio *bio = &bbio->bio;
597 struct processed_extent processed = { 0 };
598 struct folio_iter fi;
599 /*
600 * The offset to the beginning of a bio, since one bio can never be
601 * larger than UINT_MAX, u32 here is enough.
602 */
603 u32 bio_offset = 0;
604
605 ASSERT(!bio_flagged(bio, BIO_CLONED));
606 bio_for_each_folio_all(fi, &bbio->bio) {
607 bool uptodate = !bio->bi_status;
608 struct folio *folio = fi.folio;
609 struct inode *inode = folio->mapping->host;
610 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
611 const u32 sectorsize = fs_info->sectorsize;
612 u64 start;
613 u64 end;
614 u32 len;
615
616 /* For now only order 0 folios are supported for data. */
617 ASSERT(folio_order(folio) == 0);
618 btrfs_debug(fs_info,
619 "%s: bi_sector=%llu, err=%d, mirror=%u",
620 __func__, bio->bi_iter.bi_sector, bio->bi_status,
621 bbio->mirror_num);
622
623 /*
624 * We always issue full-sector reads, but if some block in a
625 * folio fails to read, blk_update_request() will advance
626 * bv_offset and adjust bv_len to compensate. Print a warning
627 * for unaligned offsets, and an error if they don't add up to
628 * a full sector.
629 */
630 if (!IS_ALIGNED(fi.offset, sectorsize))
631 btrfs_err(fs_info,
632 "partial page read in btrfs with offset %zu and length %zu",
633 fi.offset, fi.length);
634 else if (!IS_ALIGNED(fi.offset + fi.length, sectorsize))
635 btrfs_info(fs_info,
636 "incomplete page read with offset %zu and length %zu",
637 fi.offset, fi.length);
638
639 start = folio_pos(folio) + fi.offset;
640 end = start + fi.length - 1;
641 len = fi.length;
642
643 if (likely(uptodate)) {
644 loff_t i_size = i_size_read(inode);
645 pgoff_t end_index = i_size >> folio_shift(folio);
646
647 /*
648 * Zero out the remaining part if this range straddles
649 * i_size.
650 *
651 * Here we should only zero the range inside the folio,
652 * not touch anything else.
653 *
654 * NOTE: i_size is exclusive while end is inclusive.
655 */
656 if (folio_index(folio) == end_index && i_size <= end) {
657 u32 zero_start = max(offset_in_folio(folio, i_size),
658 offset_in_folio(folio, start));
659 u32 zero_len = offset_in_folio(folio, end) + 1 -
660 zero_start;
661
662 folio_zero_range(folio, zero_start, zero_len);
663 }
664 }
665
666 /* Update page status and unlock. */
667 end_page_read(folio_page(folio, 0), uptodate, start, len);
668 endio_readpage_release_extent(&processed, BTRFS_I(inode),
669 start, end, uptodate);
670
671 ASSERT(bio_offset + len > bio_offset);
672 bio_offset += len;
673
674 }
675 /* Release the last extent */
676 endio_readpage_release_extent(&processed, NULL, 0, 0, false);
677 bio_put(bio);
678}
679
680/*
681 * Populate every free slot in a provided array with pages.
682 *
683 * @nr_pages: number of pages to allocate
684 * @page_array: the array to fill with pages; any existing non-null entries in
685 * the array will be skipped
686 * @extra_gfp: the extra GFP flags for the allocation.
687 *
688 * Return: 0 if all pages were able to be allocated;
689 * -ENOMEM otherwise, the partially allocated pages would be freed and
690 * the array slots zeroed
691 */
692int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array,
693 gfp_t extra_gfp)
694{
695 unsigned int allocated;
696
697 for (allocated = 0; allocated < nr_pages;) {
698 unsigned int last = allocated;
699
700 allocated = alloc_pages_bulk_array(GFP_NOFS | extra_gfp,
701 nr_pages, page_array);
702
703 if (allocated == nr_pages)
704 return 0;
705
706 /*
707 * During this iteration, no page could be allocated, even
708 * though alloc_pages_bulk_array() falls back to alloc_page()
709 * if it could not bulk-allocate. So we must be out of memory.
710 */
711 if (allocated == last) {
712 for (int i = 0; i < allocated; i++) {
713 __free_page(page_array[i]);
714 page_array[i] = NULL;
715 }
716 return -ENOMEM;
717 }
718
719 memalloc_retry_wait(GFP_NOFS);
720 }
721 return 0;
722}
723
724/*
725 * Populate needed folios for the extent buffer.
726 *
727 * For now, the folios populated are always in order 0 (aka, single page).
728 */
729static int alloc_eb_folio_array(struct extent_buffer *eb, gfp_t extra_gfp)
730{
731 struct page *page_array[INLINE_EXTENT_BUFFER_PAGES] = { 0 };
732 int num_pages = num_extent_pages(eb);
733 int ret;
734
735 ret = btrfs_alloc_page_array(num_pages, page_array, extra_gfp);
736 if (ret < 0)
737 return ret;
738
739 for (int i = 0; i < num_pages; i++)
740 eb->folios[i] = page_folio(page_array[i]);
741 return 0;
742}
743
744static bool btrfs_bio_is_contig(struct btrfs_bio_ctrl *bio_ctrl,
745 struct page *page, u64 disk_bytenr,
746 unsigned int pg_offset)
747{
748 struct bio *bio = &bio_ctrl->bbio->bio;
749 struct bio_vec *bvec = bio_last_bvec_all(bio);
750 const sector_t sector = disk_bytenr >> SECTOR_SHIFT;
751
752 if (bio_ctrl->compress_type != BTRFS_COMPRESS_NONE) {
753 /*
754 * For compression, all IO should have its logical bytenr set
755 * to the starting bytenr of the compressed extent.
756 */
757 return bio->bi_iter.bi_sector == sector;
758 }
759
760 /*
761 * The contig check requires the following conditions to be met:
762 *
763 * 1) The pages are belonging to the same inode
764 * This is implied by the call chain.
765 *
766 * 2) The range has adjacent logical bytenr
767 *
768 * 3) The range has adjacent file offset
769 * This is required for the usage of btrfs_bio->file_offset.
770 */
771 return bio_end_sector(bio) == sector &&
772 page_offset(bvec->bv_page) + bvec->bv_offset + bvec->bv_len ==
773 page_offset(page) + pg_offset;
774}
775
776static void alloc_new_bio(struct btrfs_inode *inode,
777 struct btrfs_bio_ctrl *bio_ctrl,
778 u64 disk_bytenr, u64 file_offset)
779{
780 struct btrfs_fs_info *fs_info = inode->root->fs_info;
781 struct btrfs_bio *bbio;
782
783 bbio = btrfs_bio_alloc(BIO_MAX_VECS, bio_ctrl->opf, fs_info,
784 bio_ctrl->end_io_func, NULL);
785 bbio->bio.bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
786 bbio->inode = inode;
787 bbio->file_offset = file_offset;
788 bio_ctrl->bbio = bbio;
789 bio_ctrl->len_to_oe_boundary = U32_MAX;
790
791 /* Limit data write bios to the ordered boundary. */
792 if (bio_ctrl->wbc) {
793 struct btrfs_ordered_extent *ordered;
794
795 ordered = btrfs_lookup_ordered_extent(inode, file_offset);
796 if (ordered) {
797 bio_ctrl->len_to_oe_boundary = min_t(u32, U32_MAX,
798 ordered->file_offset +
799 ordered->disk_num_bytes - file_offset);
800 bbio->ordered = ordered;
801 }
802
803 /*
804 * Pick the last added device to support cgroup writeback. For
805 * multi-device file systems this means blk-cgroup policies have
806 * to always be set on the last added/replaced device.
807 * This is a bit odd but has been like that for a long time.
808 */
809 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
810 wbc_init_bio(bio_ctrl->wbc, &bbio->bio);
811 }
812}
813
814/*
815 * @disk_bytenr: logical bytenr where the write will be
816 * @page: page to add to the bio
817 * @size: portion of page that we want to write to
818 * @pg_offset: offset of the new bio or to check whether we are adding
819 * a contiguous page to the previous one
820 *
821 * The will either add the page into the existing @bio_ctrl->bbio, or allocate a
822 * new one in @bio_ctrl->bbio.
823 * The mirror number for this IO should already be initizlied in
824 * @bio_ctrl->mirror_num.
825 */
826static void submit_extent_page(struct btrfs_bio_ctrl *bio_ctrl,
827 u64 disk_bytenr, struct page *page,
828 size_t size, unsigned long pg_offset)
829{
830 struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
831
832 ASSERT(pg_offset + size <= PAGE_SIZE);
833 ASSERT(bio_ctrl->end_io_func);
834
835 if (bio_ctrl->bbio &&
836 !btrfs_bio_is_contig(bio_ctrl, page, disk_bytenr, pg_offset))
837 submit_one_bio(bio_ctrl);
838
839 do {
840 u32 len = size;
841
842 /* Allocate new bio if needed */
843 if (!bio_ctrl->bbio) {
844 alloc_new_bio(inode, bio_ctrl, disk_bytenr,
845 page_offset(page) + pg_offset);
846 }
847
848 /* Cap to the current ordered extent boundary if there is one. */
849 if (len > bio_ctrl->len_to_oe_boundary) {
850 ASSERT(bio_ctrl->compress_type == BTRFS_COMPRESS_NONE);
851 ASSERT(is_data_inode(&inode->vfs_inode));
852 len = bio_ctrl->len_to_oe_boundary;
853 }
854
855 if (bio_add_page(&bio_ctrl->bbio->bio, page, len, pg_offset) != len) {
856 /* bio full: move on to a new one */
857 submit_one_bio(bio_ctrl);
858 continue;
859 }
860
861 if (bio_ctrl->wbc)
862 wbc_account_cgroup_owner(bio_ctrl->wbc, page, len);
863
864 size -= len;
865 pg_offset += len;
866 disk_bytenr += len;
867
868 /*
869 * len_to_oe_boundary defaults to U32_MAX, which isn't page or
870 * sector aligned. alloc_new_bio() then sets it to the end of
871 * our ordered extent for writes into zoned devices.
872 *
873 * When len_to_oe_boundary is tracking an ordered extent, we
874 * trust the ordered extent code to align things properly, and
875 * the check above to cap our write to the ordered extent
876 * boundary is correct.
877 *
878 * When len_to_oe_boundary is U32_MAX, the cap above would
879 * result in a 4095 byte IO for the last page right before
880 * we hit the bio limit of UINT_MAX. bio_add_page() has all
881 * the checks required to make sure we don't overflow the bio,
882 * and we should just ignore len_to_oe_boundary completely
883 * unless we're using it to track an ordered extent.
884 *
885 * It's pretty hard to make a bio sized U32_MAX, but it can
886 * happen when the page cache is able to feed us contiguous
887 * pages for large extents.
888 */
889 if (bio_ctrl->len_to_oe_boundary != U32_MAX)
890 bio_ctrl->len_to_oe_boundary -= len;
891
892 /* Ordered extent boundary: move on to a new bio. */
893 if (bio_ctrl->len_to_oe_boundary == 0)
894 submit_one_bio(bio_ctrl);
895 } while (size);
896}
897
898static int attach_extent_buffer_folio(struct extent_buffer *eb,
899 struct folio *folio,
900 struct btrfs_subpage *prealloc)
901{
902 struct btrfs_fs_info *fs_info = eb->fs_info;
903 int ret = 0;
904
905 /*
906 * If the page is mapped to btree inode, we should hold the private
907 * lock to prevent race.
908 * For cloned or dummy extent buffers, their pages are not mapped and
909 * will not race with any other ebs.
910 */
911 if (folio->mapping)
912 lockdep_assert_held(&folio->mapping->i_private_lock);
913
914 if (fs_info->nodesize >= PAGE_SIZE) {
915 if (!folio_test_private(folio))
916 folio_attach_private(folio, eb);
917 else
918 WARN_ON(folio_get_private(folio) != eb);
919 return 0;
920 }
921
922 /* Already mapped, just free prealloc */
923 if (folio_test_private(folio)) {
924 btrfs_free_subpage(prealloc);
925 return 0;
926 }
927
928 if (prealloc)
929 /* Has preallocated memory for subpage */
930 folio_attach_private(folio, prealloc);
931 else
932 /* Do new allocation to attach subpage */
933 ret = btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_METADATA);
934 return ret;
935}
936
937int set_page_extent_mapped(struct page *page)
938{
939 struct folio *folio = page_folio(page);
940 struct btrfs_fs_info *fs_info;
941
942 ASSERT(page->mapping);
943
944 if (folio_test_private(folio))
945 return 0;
946
947 fs_info = btrfs_sb(page->mapping->host->i_sb);
948
949 if (btrfs_is_subpage(fs_info, page->mapping))
950 return btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_DATA);
951
952 folio_attach_private(folio, (void *)EXTENT_FOLIO_PRIVATE);
953 return 0;
954}
955
956void clear_page_extent_mapped(struct page *page)
957{
958 struct folio *folio = page_folio(page);
959 struct btrfs_fs_info *fs_info;
960
961 ASSERT(page->mapping);
962
963 if (!folio_test_private(folio))
964 return;
965
966 fs_info = btrfs_sb(page->mapping->host->i_sb);
967 if (btrfs_is_subpage(fs_info, page->mapping))
968 return btrfs_detach_subpage(fs_info, folio);
969
970 folio_detach_private(folio);
971}
972
973static struct extent_map *
974__get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
975 u64 start, u64 len, struct extent_map **em_cached)
976{
977 struct extent_map *em;
978
979 if (em_cached && *em_cached) {
980 em = *em_cached;
981 if (extent_map_in_tree(em) && start >= em->start &&
982 start < extent_map_end(em)) {
983 refcount_inc(&em->refs);
984 return em;
985 }
986
987 free_extent_map(em);
988 *em_cached = NULL;
989 }
990
991 em = btrfs_get_extent(BTRFS_I(inode), page, pg_offset, start, len);
992 if (em_cached && !IS_ERR(em)) {
993 BUG_ON(*em_cached);
994 refcount_inc(&em->refs);
995 *em_cached = em;
996 }
997 return em;
998}
999/*
1000 * basic readpage implementation. Locked extent state structs are inserted
1001 * into the tree that are removed when the IO is done (by the end_io
1002 * handlers)
1003 * XXX JDM: This needs looking at to ensure proper page locking
1004 * return 0 on success, otherwise return error
1005 */
1006static int btrfs_do_readpage(struct page *page, struct extent_map **em_cached,
1007 struct btrfs_bio_ctrl *bio_ctrl, u64 *prev_em_start)
1008{
1009 struct inode *inode = page->mapping->host;
1010 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1011 u64 start = page_offset(page);
1012 const u64 end = start + PAGE_SIZE - 1;
1013 u64 cur = start;
1014 u64 extent_offset;
1015 u64 last_byte = i_size_read(inode);
1016 u64 block_start;
1017 struct extent_map *em;
1018 int ret = 0;
1019 size_t pg_offset = 0;
1020 size_t iosize;
1021 size_t blocksize = inode->i_sb->s_blocksize;
1022 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1023
1024 ret = set_page_extent_mapped(page);
1025 if (ret < 0) {
1026 unlock_extent(tree, start, end, NULL);
1027 unlock_page(page);
1028 return ret;
1029 }
1030
1031 if (page->index == last_byte >> PAGE_SHIFT) {
1032 size_t zero_offset = offset_in_page(last_byte);
1033
1034 if (zero_offset) {
1035 iosize = PAGE_SIZE - zero_offset;
1036 memzero_page(page, zero_offset, iosize);
1037 }
1038 }
1039 bio_ctrl->end_io_func = end_bbio_data_read;
1040 begin_page_read(fs_info, page);
1041 while (cur <= end) {
1042 enum btrfs_compression_type compress_type = BTRFS_COMPRESS_NONE;
1043 bool force_bio_submit = false;
1044 u64 disk_bytenr;
1045
1046 ASSERT(IS_ALIGNED(cur, fs_info->sectorsize));
1047 if (cur >= last_byte) {
1048 iosize = PAGE_SIZE - pg_offset;
1049 memzero_page(page, pg_offset, iosize);
1050 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1051 end_page_read(page, true, cur, iosize);
1052 break;
1053 }
1054 em = __get_extent_map(inode, page, pg_offset, cur,
1055 end - cur + 1, em_cached);
1056 if (IS_ERR(em)) {
1057 unlock_extent(tree, cur, end, NULL);
1058 end_page_read(page, false, cur, end + 1 - cur);
1059 return PTR_ERR(em);
1060 }
1061 extent_offset = cur - em->start;
1062 BUG_ON(extent_map_end(em) <= cur);
1063 BUG_ON(end < cur);
1064
1065 compress_type = extent_map_compression(em);
1066
1067 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1068 iosize = ALIGN(iosize, blocksize);
1069 if (compress_type != BTRFS_COMPRESS_NONE)
1070 disk_bytenr = em->block_start;
1071 else
1072 disk_bytenr = em->block_start + extent_offset;
1073 block_start = em->block_start;
1074 if (em->flags & EXTENT_FLAG_PREALLOC)
1075 block_start = EXTENT_MAP_HOLE;
1076
1077 /*
1078 * If we have a file range that points to a compressed extent
1079 * and it's followed by a consecutive file range that points
1080 * to the same compressed extent (possibly with a different
1081 * offset and/or length, so it either points to the whole extent
1082 * or only part of it), we must make sure we do not submit a
1083 * single bio to populate the pages for the 2 ranges because
1084 * this makes the compressed extent read zero out the pages
1085 * belonging to the 2nd range. Imagine the following scenario:
1086 *
1087 * File layout
1088 * [0 - 8K] [8K - 24K]
1089 * | |
1090 * | |
1091 * points to extent X, points to extent X,
1092 * offset 4K, length of 8K offset 0, length 16K
1093 *
1094 * [extent X, compressed length = 4K uncompressed length = 16K]
1095 *
1096 * If the bio to read the compressed extent covers both ranges,
1097 * it will decompress extent X into the pages belonging to the
1098 * first range and then it will stop, zeroing out the remaining
1099 * pages that belong to the other range that points to extent X.
1100 * So here we make sure we submit 2 bios, one for the first
1101 * range and another one for the third range. Both will target
1102 * the same physical extent from disk, but we can't currently
1103 * make the compressed bio endio callback populate the pages
1104 * for both ranges because each compressed bio is tightly
1105 * coupled with a single extent map, and each range can have
1106 * an extent map with a different offset value relative to the
1107 * uncompressed data of our extent and different lengths. This
1108 * is a corner case so we prioritize correctness over
1109 * non-optimal behavior (submitting 2 bios for the same extent).
1110 */
1111 if (compress_type != BTRFS_COMPRESS_NONE &&
1112 prev_em_start && *prev_em_start != (u64)-1 &&
1113 *prev_em_start != em->start)
1114 force_bio_submit = true;
1115
1116 if (prev_em_start)
1117 *prev_em_start = em->start;
1118
1119 free_extent_map(em);
1120 em = NULL;
1121
1122 /* we've found a hole, just zero and go on */
1123 if (block_start == EXTENT_MAP_HOLE) {
1124 memzero_page(page, pg_offset, iosize);
1125
1126 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1127 end_page_read(page, true, cur, iosize);
1128 cur = cur + iosize;
1129 pg_offset += iosize;
1130 continue;
1131 }
1132 /* the get_extent function already copied into the page */
1133 if (block_start == EXTENT_MAP_INLINE) {
1134 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1135 end_page_read(page, true, cur, iosize);
1136 cur = cur + iosize;
1137 pg_offset += iosize;
1138 continue;
1139 }
1140
1141 if (bio_ctrl->compress_type != compress_type) {
1142 submit_one_bio(bio_ctrl);
1143 bio_ctrl->compress_type = compress_type;
1144 }
1145
1146 if (force_bio_submit)
1147 submit_one_bio(bio_ctrl);
1148 submit_extent_page(bio_ctrl, disk_bytenr, page, iosize,
1149 pg_offset);
1150 cur = cur + iosize;
1151 pg_offset += iosize;
1152 }
1153
1154 return 0;
1155}
1156
1157int btrfs_read_folio(struct file *file, struct folio *folio)
1158{
1159 struct page *page = &folio->page;
1160 struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
1161 u64 start = page_offset(page);
1162 u64 end = start + PAGE_SIZE - 1;
1163 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ };
1164 int ret;
1165
1166 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
1167
1168 ret = btrfs_do_readpage(page, NULL, &bio_ctrl, NULL);
1169 /*
1170 * If btrfs_do_readpage() failed we will want to submit the assembled
1171 * bio to do the cleanup.
1172 */
1173 submit_one_bio(&bio_ctrl);
1174 return ret;
1175}
1176
1177static inline void contiguous_readpages(struct page *pages[], int nr_pages,
1178 u64 start, u64 end,
1179 struct extent_map **em_cached,
1180 struct btrfs_bio_ctrl *bio_ctrl,
1181 u64 *prev_em_start)
1182{
1183 struct btrfs_inode *inode = BTRFS_I(pages[0]->mapping->host);
1184 int index;
1185
1186 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
1187
1188 for (index = 0; index < nr_pages; index++) {
1189 btrfs_do_readpage(pages[index], em_cached, bio_ctrl,
1190 prev_em_start);
1191 put_page(pages[index]);
1192 }
1193}
1194
1195/*
1196 * helper for __extent_writepage, doing all of the delayed allocation setup.
1197 *
1198 * This returns 1 if btrfs_run_delalloc_range function did all the work required
1199 * to write the page (copy into inline extent). In this case the IO has
1200 * been started and the page is already unlocked.
1201 *
1202 * This returns 0 if all went well (page still locked)
1203 * This returns < 0 if there were errors (page still locked)
1204 */
1205static noinline_for_stack int writepage_delalloc(struct btrfs_inode *inode,
1206 struct page *page, struct writeback_control *wbc)
1207{
1208 const u64 page_start = page_offset(page);
1209 const u64 page_end = page_start + PAGE_SIZE - 1;
1210 u64 delalloc_start = page_start;
1211 u64 delalloc_end = page_end;
1212 u64 delalloc_to_write = 0;
1213 int ret = 0;
1214
1215 while (delalloc_start < page_end) {
1216 delalloc_end = page_end;
1217 if (!find_lock_delalloc_range(&inode->vfs_inode, page,
1218 &delalloc_start, &delalloc_end)) {
1219 delalloc_start = delalloc_end + 1;
1220 continue;
1221 }
1222
1223 ret = btrfs_run_delalloc_range(inode, page, delalloc_start,
1224 delalloc_end, wbc);
1225 if (ret < 0)
1226 return ret;
1227
1228 delalloc_start = delalloc_end + 1;
1229 }
1230
1231 /*
1232 * delalloc_end is already one less than the total length, so
1233 * we don't subtract one from PAGE_SIZE
1234 */
1235 delalloc_to_write +=
1236 DIV_ROUND_UP(delalloc_end + 1 - page_start, PAGE_SIZE);
1237
1238 /*
1239 * If btrfs_run_dealloc_range() already started I/O and unlocked
1240 * the pages, we just need to account for them here.
1241 */
1242 if (ret == 1) {
1243 wbc->nr_to_write -= delalloc_to_write;
1244 return 1;
1245 }
1246
1247 if (wbc->nr_to_write < delalloc_to_write) {
1248 int thresh = 8192;
1249
1250 if (delalloc_to_write < thresh * 2)
1251 thresh = delalloc_to_write;
1252 wbc->nr_to_write = min_t(u64, delalloc_to_write,
1253 thresh);
1254 }
1255
1256 return 0;
1257}
1258
1259/*
1260 * Find the first byte we need to write.
1261 *
1262 * For subpage, one page can contain several sectors, and
1263 * __extent_writepage_io() will just grab all extent maps in the page
1264 * range and try to submit all non-inline/non-compressed extents.
1265 *
1266 * This is a big problem for subpage, we shouldn't re-submit already written
1267 * data at all.
1268 * This function will lookup subpage dirty bit to find which range we really
1269 * need to submit.
1270 *
1271 * Return the next dirty range in [@start, @end).
1272 * If no dirty range is found, @start will be page_offset(page) + PAGE_SIZE.
1273 */
1274static void find_next_dirty_byte(struct btrfs_fs_info *fs_info,
1275 struct page *page, u64 *start, u64 *end)
1276{
1277 struct folio *folio = page_folio(page);
1278 struct btrfs_subpage *subpage = folio_get_private(folio);
1279 struct btrfs_subpage_info *spi = fs_info->subpage_info;
1280 u64 orig_start = *start;
1281 /* Declare as unsigned long so we can use bitmap ops */
1282 unsigned long flags;
1283 int range_start_bit;
1284 int range_end_bit;
1285
1286 /*
1287 * For regular sector size == page size case, since one page only
1288 * contains one sector, we return the page offset directly.
1289 */
1290 if (!btrfs_is_subpage(fs_info, page->mapping)) {
1291 *start = page_offset(page);
1292 *end = page_offset(page) + PAGE_SIZE;
1293 return;
1294 }
1295
1296 range_start_bit = spi->dirty_offset +
1297 (offset_in_page(orig_start) >> fs_info->sectorsize_bits);
1298
1299 /* We should have the page locked, but just in case */
1300 spin_lock_irqsave(&subpage->lock, flags);
1301 bitmap_next_set_region(subpage->bitmaps, &range_start_bit, &range_end_bit,
1302 spi->dirty_offset + spi->bitmap_nr_bits);
1303 spin_unlock_irqrestore(&subpage->lock, flags);
1304
1305 range_start_bit -= spi->dirty_offset;
1306 range_end_bit -= spi->dirty_offset;
1307
1308 *start = page_offset(page) + range_start_bit * fs_info->sectorsize;
1309 *end = page_offset(page) + range_end_bit * fs_info->sectorsize;
1310}
1311
1312/*
1313 * helper for __extent_writepage. This calls the writepage start hooks,
1314 * and does the loop to map the page into extents and bios.
1315 *
1316 * We return 1 if the IO is started and the page is unlocked,
1317 * 0 if all went well (page still locked)
1318 * < 0 if there were errors (page still locked)
1319 */
1320static noinline_for_stack int __extent_writepage_io(struct btrfs_inode *inode,
1321 struct page *page,
1322 struct btrfs_bio_ctrl *bio_ctrl,
1323 loff_t i_size,
1324 int *nr_ret)
1325{
1326 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1327 u64 cur = page_offset(page);
1328 u64 end = cur + PAGE_SIZE - 1;
1329 u64 extent_offset;
1330 u64 block_start;
1331 struct extent_map *em;
1332 int ret = 0;
1333 int nr = 0;
1334
1335 ret = btrfs_writepage_cow_fixup(page);
1336 if (ret) {
1337 /* Fixup worker will requeue */
1338 redirty_page_for_writepage(bio_ctrl->wbc, page);
1339 unlock_page(page);
1340 return 1;
1341 }
1342
1343 bio_ctrl->end_io_func = end_bbio_data_write;
1344 while (cur <= end) {
1345 u32 len = end - cur + 1;
1346 u64 disk_bytenr;
1347 u64 em_end;
1348 u64 dirty_range_start = cur;
1349 u64 dirty_range_end;
1350 u32 iosize;
1351
1352 if (cur >= i_size) {
1353 btrfs_mark_ordered_io_finished(inode, page, cur, len,
1354 true);
1355 /*
1356 * This range is beyond i_size, thus we don't need to
1357 * bother writing back.
1358 * But we still need to clear the dirty subpage bit, or
1359 * the next time the page gets dirtied, we will try to
1360 * writeback the sectors with subpage dirty bits,
1361 * causing writeback without ordered extent.
1362 */
1363 btrfs_folio_clear_dirty(fs_info, page_folio(page), cur, len);
1364 break;
1365 }
1366
1367 find_next_dirty_byte(fs_info, page, &dirty_range_start,
1368 &dirty_range_end);
1369 if (cur < dirty_range_start) {
1370 cur = dirty_range_start;
1371 continue;
1372 }
1373
1374 em = btrfs_get_extent(inode, NULL, 0, cur, len);
1375 if (IS_ERR(em)) {
1376 ret = PTR_ERR_OR_ZERO(em);
1377 goto out_error;
1378 }
1379
1380 extent_offset = cur - em->start;
1381 em_end = extent_map_end(em);
1382 ASSERT(cur <= em_end);
1383 ASSERT(cur < end);
1384 ASSERT(IS_ALIGNED(em->start, fs_info->sectorsize));
1385 ASSERT(IS_ALIGNED(em->len, fs_info->sectorsize));
1386
1387 block_start = em->block_start;
1388 disk_bytenr = em->block_start + extent_offset;
1389
1390 ASSERT(!extent_map_is_compressed(em));
1391 ASSERT(block_start != EXTENT_MAP_HOLE);
1392 ASSERT(block_start != EXTENT_MAP_INLINE);
1393
1394 /*
1395 * Note that em_end from extent_map_end() and dirty_range_end from
1396 * find_next_dirty_byte() are all exclusive
1397 */
1398 iosize = min(min(em_end, end + 1), dirty_range_end) - cur;
1399 free_extent_map(em);
1400 em = NULL;
1401
1402 btrfs_set_range_writeback(inode, cur, cur + iosize - 1);
1403 if (!PageWriteback(page)) {
1404 btrfs_err(inode->root->fs_info,
1405 "page %lu not writeback, cur %llu end %llu",
1406 page->index, cur, end);
1407 }
1408
1409 /*
1410 * Although the PageDirty bit is cleared before entering this
1411 * function, subpage dirty bit is not cleared.
1412 * So clear subpage dirty bit here so next time we won't submit
1413 * page for range already written to disk.
1414 */
1415 btrfs_folio_clear_dirty(fs_info, page_folio(page), cur, iosize);
1416
1417 submit_extent_page(bio_ctrl, disk_bytenr, page, iosize,
1418 cur - page_offset(page));
1419 cur += iosize;
1420 nr++;
1421 }
1422
1423 btrfs_folio_assert_not_dirty(fs_info, page_folio(page));
1424 *nr_ret = nr;
1425 return 0;
1426
1427out_error:
1428 /*
1429 * If we finish without problem, we should not only clear page dirty,
1430 * but also empty subpage dirty bits
1431 */
1432 *nr_ret = nr;
1433 return ret;
1434}
1435
1436/*
1437 * the writepage semantics are similar to regular writepage. extent
1438 * records are inserted to lock ranges in the tree, and as dirty areas
1439 * are found, they are marked writeback. Then the lock bits are removed
1440 * and the end_io handler clears the writeback ranges
1441 *
1442 * Return 0 if everything goes well.
1443 * Return <0 for error.
1444 */
1445static int __extent_writepage(struct page *page, struct btrfs_bio_ctrl *bio_ctrl)
1446{
1447 struct folio *folio = page_folio(page);
1448 struct inode *inode = page->mapping->host;
1449 const u64 page_start = page_offset(page);
1450 int ret;
1451 int nr = 0;
1452 size_t pg_offset;
1453 loff_t i_size = i_size_read(inode);
1454 unsigned long end_index = i_size >> PAGE_SHIFT;
1455
1456 trace___extent_writepage(page, inode, bio_ctrl->wbc);
1457
1458 WARN_ON(!PageLocked(page));
1459
1460 pg_offset = offset_in_page(i_size);
1461 if (page->index > end_index ||
1462 (page->index == end_index && !pg_offset)) {
1463 folio_invalidate(folio, 0, folio_size(folio));
1464 folio_unlock(folio);
1465 return 0;
1466 }
1467
1468 if (page->index == end_index)
1469 memzero_page(page, pg_offset, PAGE_SIZE - pg_offset);
1470
1471 ret = set_page_extent_mapped(page);
1472 if (ret < 0)
1473 goto done;
1474
1475 ret = writepage_delalloc(BTRFS_I(inode), page, bio_ctrl->wbc);
1476 if (ret == 1)
1477 return 0;
1478 if (ret)
1479 goto done;
1480
1481 ret = __extent_writepage_io(BTRFS_I(inode), page, bio_ctrl, i_size, &nr);
1482 if (ret == 1)
1483 return 0;
1484
1485 bio_ctrl->wbc->nr_to_write--;
1486
1487done:
1488 if (nr == 0) {
1489 /* make sure the mapping tag for page dirty gets cleared */
1490 set_page_writeback(page);
1491 end_page_writeback(page);
1492 }
1493 if (ret) {
1494 btrfs_mark_ordered_io_finished(BTRFS_I(inode), page, page_start,
1495 PAGE_SIZE, !ret);
1496 mapping_set_error(page->mapping, ret);
1497 }
1498 unlock_page(page);
1499 ASSERT(ret <= 0);
1500 return ret;
1501}
1502
1503void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
1504{
1505 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
1506 TASK_UNINTERRUPTIBLE);
1507}
1508
1509/*
1510 * Lock extent buffer status and pages for writeback.
1511 *
1512 * Return %false if the extent buffer doesn't need to be submitted (e.g. the
1513 * extent buffer is not dirty)
1514 * Return %true is the extent buffer is submitted to bio.
1515 */
1516static noinline_for_stack bool lock_extent_buffer_for_io(struct extent_buffer *eb,
1517 struct writeback_control *wbc)
1518{
1519 struct btrfs_fs_info *fs_info = eb->fs_info;
1520 bool ret = false;
1521
1522 btrfs_tree_lock(eb);
1523 while (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
1524 btrfs_tree_unlock(eb);
1525 if (wbc->sync_mode != WB_SYNC_ALL)
1526 return false;
1527 wait_on_extent_buffer_writeback(eb);
1528 btrfs_tree_lock(eb);
1529 }
1530
1531 /*
1532 * We need to do this to prevent races in people who check if the eb is
1533 * under IO since we can end up having no IO bits set for a short period
1534 * of time.
1535 */
1536 spin_lock(&eb->refs_lock);
1537 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
1538 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1539 spin_unlock(&eb->refs_lock);
1540 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
1541 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
1542 -eb->len,
1543 fs_info->dirty_metadata_batch);
1544 ret = true;
1545 } else {
1546 spin_unlock(&eb->refs_lock);
1547 }
1548 btrfs_tree_unlock(eb);
1549 return ret;
1550}
1551
1552static void set_btree_ioerr(struct extent_buffer *eb)
1553{
1554 struct btrfs_fs_info *fs_info = eb->fs_info;
1555
1556 set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1557
1558 /*
1559 * A read may stumble upon this buffer later, make sure that it gets an
1560 * error and knows there was an error.
1561 */
1562 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
1563
1564 /*
1565 * We need to set the mapping with the io error as well because a write
1566 * error will flip the file system readonly, and then syncfs() will
1567 * return a 0 because we are readonly if we don't modify the err seq for
1568 * the superblock.
1569 */
1570 mapping_set_error(eb->fs_info->btree_inode->i_mapping, -EIO);
1571
1572 /*
1573 * If writeback for a btree extent that doesn't belong to a log tree
1574 * failed, increment the counter transaction->eb_write_errors.
1575 * We do this because while the transaction is running and before it's
1576 * committing (when we call filemap_fdata[write|wait]_range against
1577 * the btree inode), we might have
1578 * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
1579 * returns an error or an error happens during writeback, when we're
1580 * committing the transaction we wouldn't know about it, since the pages
1581 * can be no longer dirty nor marked anymore for writeback (if a
1582 * subsequent modification to the extent buffer didn't happen before the
1583 * transaction commit), which makes filemap_fdata[write|wait]_range not
1584 * able to find the pages tagged with SetPageError at transaction
1585 * commit time. So if this happens we must abort the transaction,
1586 * otherwise we commit a super block with btree roots that point to
1587 * btree nodes/leafs whose content on disk is invalid - either garbage
1588 * or the content of some node/leaf from a past generation that got
1589 * cowed or deleted and is no longer valid.
1590 *
1591 * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
1592 * not be enough - we need to distinguish between log tree extents vs
1593 * non-log tree extents, and the next filemap_fdatawait_range() call
1594 * will catch and clear such errors in the mapping - and that call might
1595 * be from a log sync and not from a transaction commit. Also, checking
1596 * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
1597 * not done and would not be reliable - the eb might have been released
1598 * from memory and reading it back again means that flag would not be
1599 * set (since it's a runtime flag, not persisted on disk).
1600 *
1601 * Using the flags below in the btree inode also makes us achieve the
1602 * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
1603 * writeback for all dirty pages and before filemap_fdatawait_range()
1604 * is called, the writeback for all dirty pages had already finished
1605 * with errors - because we were not using AS_EIO/AS_ENOSPC,
1606 * filemap_fdatawait_range() would return success, as it could not know
1607 * that writeback errors happened (the pages were no longer tagged for
1608 * writeback).
1609 */
1610 switch (eb->log_index) {
1611 case -1:
1612 set_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags);
1613 break;
1614 case 0:
1615 set_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
1616 break;
1617 case 1:
1618 set_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
1619 break;
1620 default:
1621 BUG(); /* unexpected, logic error */
1622 }
1623}
1624
1625/*
1626 * The endio specific version which won't touch any unsafe spinlock in endio
1627 * context.
1628 */
1629static struct extent_buffer *find_extent_buffer_nolock(
1630 struct btrfs_fs_info *fs_info, u64 start)
1631{
1632 struct extent_buffer *eb;
1633
1634 rcu_read_lock();
1635 eb = radix_tree_lookup(&fs_info->buffer_radix,
1636 start >> fs_info->sectorsize_bits);
1637 if (eb && atomic_inc_not_zero(&eb->refs)) {
1638 rcu_read_unlock();
1639 return eb;
1640 }
1641 rcu_read_unlock();
1642 return NULL;
1643}
1644
1645static void end_bbio_meta_write(struct btrfs_bio *bbio)
1646{
1647 struct extent_buffer *eb = bbio->private;
1648 struct btrfs_fs_info *fs_info = eb->fs_info;
1649 bool uptodate = !bbio->bio.bi_status;
1650 struct folio_iter fi;
1651 u32 bio_offset = 0;
1652
1653 if (!uptodate)
1654 set_btree_ioerr(eb);
1655
1656 bio_for_each_folio_all(fi, &bbio->bio) {
1657 u64 start = eb->start + bio_offset;
1658 struct folio *folio = fi.folio;
1659 u32 len = fi.length;
1660
1661 btrfs_folio_clear_writeback(fs_info, folio, start, len);
1662 bio_offset += len;
1663 }
1664
1665 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1666 smp_mb__after_atomic();
1667 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
1668
1669 bio_put(&bbio->bio);
1670}
1671
1672static void prepare_eb_write(struct extent_buffer *eb)
1673{
1674 u32 nritems;
1675 unsigned long start;
1676 unsigned long end;
1677
1678 clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1679
1680 /* Set btree blocks beyond nritems with 0 to avoid stale content */
1681 nritems = btrfs_header_nritems(eb);
1682 if (btrfs_header_level(eb) > 0) {
1683 end = btrfs_node_key_ptr_offset(eb, nritems);
1684 memzero_extent_buffer(eb, end, eb->len - end);
1685 } else {
1686 /*
1687 * Leaf:
1688 * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
1689 */
1690 start = btrfs_item_nr_offset(eb, nritems);
1691 end = btrfs_item_nr_offset(eb, 0);
1692 if (nritems == 0)
1693 end += BTRFS_LEAF_DATA_SIZE(eb->fs_info);
1694 else
1695 end += btrfs_item_offset(eb, nritems - 1);
1696 memzero_extent_buffer(eb, start, end - start);
1697 }
1698}
1699
1700static noinline_for_stack void write_one_eb(struct extent_buffer *eb,
1701 struct writeback_control *wbc)
1702{
1703 struct btrfs_fs_info *fs_info = eb->fs_info;
1704 struct btrfs_bio *bbio;
1705
1706 prepare_eb_write(eb);
1707
1708 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
1709 REQ_OP_WRITE | REQ_META | wbc_to_write_flags(wbc),
1710 eb->fs_info, end_bbio_meta_write, eb);
1711 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
1712 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
1713 wbc_init_bio(wbc, &bbio->bio);
1714 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
1715 bbio->file_offset = eb->start;
1716 if (fs_info->nodesize < PAGE_SIZE) {
1717 struct folio *folio = eb->folios[0];
1718 bool ret;
1719
1720 folio_lock(folio);
1721 btrfs_subpage_set_writeback(fs_info, folio, eb->start, eb->len);
1722 if (btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start,
1723 eb->len)) {
1724 folio_clear_dirty_for_io(folio);
1725 wbc->nr_to_write--;
1726 }
1727 ret = bio_add_folio(&bbio->bio, folio, eb->len,
1728 eb->start - folio_pos(folio));
1729 ASSERT(ret);
1730 wbc_account_cgroup_owner(wbc, folio_page(folio, 0), eb->len);
1731 folio_unlock(folio);
1732 } else {
1733 int num_folios = num_extent_folios(eb);
1734
1735 for (int i = 0; i < num_folios; i++) {
1736 struct folio *folio = eb->folios[i];
1737 bool ret;
1738
1739 folio_lock(folio);
1740 folio_clear_dirty_for_io(folio);
1741 folio_start_writeback(folio);
1742 ret = bio_add_folio(&bbio->bio, folio, folio_size(folio), 0);
1743 ASSERT(ret);
1744 wbc_account_cgroup_owner(wbc, folio_page(folio, 0),
1745 folio_size(folio));
1746 wbc->nr_to_write -= folio_nr_pages(folio);
1747 folio_unlock(folio);
1748 }
1749 }
1750 btrfs_submit_bio(bbio, 0);
1751}
1752
1753/*
1754 * Submit one subpage btree page.
1755 *
1756 * The main difference to submit_eb_page() is:
1757 * - Page locking
1758 * For subpage, we don't rely on page locking at all.
1759 *
1760 * - Flush write bio
1761 * We only flush bio if we may be unable to fit current extent buffers into
1762 * current bio.
1763 *
1764 * Return >=0 for the number of submitted extent buffers.
1765 * Return <0 for fatal error.
1766 */
1767static int submit_eb_subpage(struct page *page, struct writeback_control *wbc)
1768{
1769 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
1770 struct folio *folio = page_folio(page);
1771 int submitted = 0;
1772 u64 page_start = page_offset(page);
1773 int bit_start = 0;
1774 int sectors_per_node = fs_info->nodesize >> fs_info->sectorsize_bits;
1775
1776 /* Lock and write each dirty extent buffers in the range */
1777 while (bit_start < fs_info->subpage_info->bitmap_nr_bits) {
1778 struct btrfs_subpage *subpage = folio_get_private(folio);
1779 struct extent_buffer *eb;
1780 unsigned long flags;
1781 u64 start;
1782
1783 /*
1784 * Take private lock to ensure the subpage won't be detached
1785 * in the meantime.
1786 */
1787 spin_lock(&page->mapping->i_private_lock);
1788 if (!folio_test_private(folio)) {
1789 spin_unlock(&page->mapping->i_private_lock);
1790 break;
1791 }
1792 spin_lock_irqsave(&subpage->lock, flags);
1793 if (!test_bit(bit_start + fs_info->subpage_info->dirty_offset,
1794 subpage->bitmaps)) {
1795 spin_unlock_irqrestore(&subpage->lock, flags);
1796 spin_unlock(&page->mapping->i_private_lock);
1797 bit_start++;
1798 continue;
1799 }
1800
1801 start = page_start + bit_start * fs_info->sectorsize;
1802 bit_start += sectors_per_node;
1803
1804 /*
1805 * Here we just want to grab the eb without touching extra
1806 * spin locks, so call find_extent_buffer_nolock().
1807 */
1808 eb = find_extent_buffer_nolock(fs_info, start);
1809 spin_unlock_irqrestore(&subpage->lock, flags);
1810 spin_unlock(&page->mapping->i_private_lock);
1811
1812 /*
1813 * The eb has already reached 0 refs thus find_extent_buffer()
1814 * doesn't return it. We don't need to write back such eb
1815 * anyway.
1816 */
1817 if (!eb)
1818 continue;
1819
1820 if (lock_extent_buffer_for_io(eb, wbc)) {
1821 write_one_eb(eb, wbc);
1822 submitted++;
1823 }
1824 free_extent_buffer(eb);
1825 }
1826 return submitted;
1827}
1828
1829/*
1830 * Submit all page(s) of one extent buffer.
1831 *
1832 * @page: the page of one extent buffer
1833 * @eb_context: to determine if we need to submit this page, if current page
1834 * belongs to this eb, we don't need to submit
1835 *
1836 * The caller should pass each page in their bytenr order, and here we use
1837 * @eb_context to determine if we have submitted pages of one extent buffer.
1838 *
1839 * If we have, we just skip until we hit a new page that doesn't belong to
1840 * current @eb_context.
1841 *
1842 * If not, we submit all the page(s) of the extent buffer.
1843 *
1844 * Return >0 if we have submitted the extent buffer successfully.
1845 * Return 0 if we don't need to submit the page, as it's already submitted by
1846 * previous call.
1847 * Return <0 for fatal error.
1848 */
1849static int submit_eb_page(struct page *page, struct btrfs_eb_write_context *ctx)
1850{
1851 struct writeback_control *wbc = ctx->wbc;
1852 struct address_space *mapping = page->mapping;
1853 struct folio *folio = page_folio(page);
1854 struct extent_buffer *eb;
1855 int ret;
1856
1857 if (!folio_test_private(folio))
1858 return 0;
1859
1860 if (btrfs_sb(page->mapping->host->i_sb)->nodesize < PAGE_SIZE)
1861 return submit_eb_subpage(page, wbc);
1862
1863 spin_lock(&mapping->i_private_lock);
1864 if (!folio_test_private(folio)) {
1865 spin_unlock(&mapping->i_private_lock);
1866 return 0;
1867 }
1868
1869 eb = folio_get_private(folio);
1870
1871 /*
1872 * Shouldn't happen and normally this would be a BUG_ON but no point
1873 * crashing the machine for something we can survive anyway.
1874 */
1875 if (WARN_ON(!eb)) {
1876 spin_unlock(&mapping->i_private_lock);
1877 return 0;
1878 }
1879
1880 if (eb == ctx->eb) {
1881 spin_unlock(&mapping->i_private_lock);
1882 return 0;
1883 }
1884 ret = atomic_inc_not_zero(&eb->refs);
1885 spin_unlock(&mapping->i_private_lock);
1886 if (!ret)
1887 return 0;
1888
1889 ctx->eb = eb;
1890
1891 ret = btrfs_check_meta_write_pointer(eb->fs_info, ctx);
1892 if (ret) {
1893 if (ret == -EBUSY)
1894 ret = 0;
1895 free_extent_buffer(eb);
1896 return ret;
1897 }
1898
1899 if (!lock_extent_buffer_for_io(eb, wbc)) {
1900 free_extent_buffer(eb);
1901 return 0;
1902 }
1903 /* Implies write in zoned mode. */
1904 if (ctx->zoned_bg) {
1905 /* Mark the last eb in the block group. */
1906 btrfs_schedule_zone_finish_bg(ctx->zoned_bg, eb);
1907 ctx->zoned_bg->meta_write_pointer += eb->len;
1908 }
1909 write_one_eb(eb, wbc);
1910 free_extent_buffer(eb);
1911 return 1;
1912}
1913
1914int btree_write_cache_pages(struct address_space *mapping,
1915 struct writeback_control *wbc)
1916{
1917 struct btrfs_eb_write_context ctx = { .wbc = wbc };
1918 struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
1919 int ret = 0;
1920 int done = 0;
1921 int nr_to_write_done = 0;
1922 struct folio_batch fbatch;
1923 unsigned int nr_folios;
1924 pgoff_t index;
1925 pgoff_t end; /* Inclusive */
1926 int scanned = 0;
1927 xa_mark_t tag;
1928
1929 folio_batch_init(&fbatch);
1930 if (wbc->range_cyclic) {
1931 index = mapping->writeback_index; /* Start from prev offset */
1932 end = -1;
1933 /*
1934 * Start from the beginning does not need to cycle over the
1935 * range, mark it as scanned.
1936 */
1937 scanned = (index == 0);
1938 } else {
1939 index = wbc->range_start >> PAGE_SHIFT;
1940 end = wbc->range_end >> PAGE_SHIFT;
1941 scanned = 1;
1942 }
1943 if (wbc->sync_mode == WB_SYNC_ALL)
1944 tag = PAGECACHE_TAG_TOWRITE;
1945 else
1946 tag = PAGECACHE_TAG_DIRTY;
1947 btrfs_zoned_meta_io_lock(fs_info);
1948retry:
1949 if (wbc->sync_mode == WB_SYNC_ALL)
1950 tag_pages_for_writeback(mapping, index, end);
1951 while (!done && !nr_to_write_done && (index <= end) &&
1952 (nr_folios = filemap_get_folios_tag(mapping, &index, end,
1953 tag, &fbatch))) {
1954 unsigned i;
1955
1956 for (i = 0; i < nr_folios; i++) {
1957 struct folio *folio = fbatch.folios[i];
1958
1959 ret = submit_eb_page(&folio->page, &ctx);
1960 if (ret == 0)
1961 continue;
1962 if (ret < 0) {
1963 done = 1;
1964 break;
1965 }
1966
1967 /*
1968 * the filesystem may choose to bump up nr_to_write.
1969 * We have to make sure to honor the new nr_to_write
1970 * at any time
1971 */
1972 nr_to_write_done = wbc->nr_to_write <= 0;
1973 }
1974 folio_batch_release(&fbatch);
1975 cond_resched();
1976 }
1977 if (!scanned && !done) {
1978 /*
1979 * We hit the last page and there is more work to be done: wrap
1980 * back to the start of the file
1981 */
1982 scanned = 1;
1983 index = 0;
1984 goto retry;
1985 }
1986 /*
1987 * If something went wrong, don't allow any metadata write bio to be
1988 * submitted.
1989 *
1990 * This would prevent use-after-free if we had dirty pages not
1991 * cleaned up, which can still happen by fuzzed images.
1992 *
1993 * - Bad extent tree
1994 * Allowing existing tree block to be allocated for other trees.
1995 *
1996 * - Log tree operations
1997 * Exiting tree blocks get allocated to log tree, bumps its
1998 * generation, then get cleaned in tree re-balance.
1999 * Such tree block will not be written back, since it's clean,
2000 * thus no WRITTEN flag set.
2001 * And after log writes back, this tree block is not traced by
2002 * any dirty extent_io_tree.
2003 *
2004 * - Offending tree block gets re-dirtied from its original owner
2005 * Since it has bumped generation, no WRITTEN flag, it can be
2006 * reused without COWing. This tree block will not be traced
2007 * by btrfs_transaction::dirty_pages.
2008 *
2009 * Now such dirty tree block will not be cleaned by any dirty
2010 * extent io tree. Thus we don't want to submit such wild eb
2011 * if the fs already has error.
2012 *
2013 * We can get ret > 0 from submit_extent_page() indicating how many ebs
2014 * were submitted. Reset it to 0 to avoid false alerts for the caller.
2015 */
2016 if (ret > 0)
2017 ret = 0;
2018 if (!ret && BTRFS_FS_ERROR(fs_info))
2019 ret = -EROFS;
2020
2021 if (ctx.zoned_bg)
2022 btrfs_put_block_group(ctx.zoned_bg);
2023 btrfs_zoned_meta_io_unlock(fs_info);
2024 return ret;
2025}
2026
2027/*
2028 * Walk the list of dirty pages of the given address space and write all of them.
2029 *
2030 * @mapping: address space structure to write
2031 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2032 * @bio_ctrl: holds context for the write, namely the bio
2033 *
2034 * If a page is already under I/O, write_cache_pages() skips it, even
2035 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2036 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2037 * and msync() need to guarantee that all the data which was dirty at the time
2038 * the call was made get new I/O started against them. If wbc->sync_mode is
2039 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2040 * existing IO to complete.
2041 */
2042static int extent_write_cache_pages(struct address_space *mapping,
2043 struct btrfs_bio_ctrl *bio_ctrl)
2044{
2045 struct writeback_control *wbc = bio_ctrl->wbc;
2046 struct inode *inode = mapping->host;
2047 int ret = 0;
2048 int done = 0;
2049 int nr_to_write_done = 0;
2050 struct folio_batch fbatch;
2051 unsigned int nr_folios;
2052 pgoff_t index;
2053 pgoff_t end; /* Inclusive */
2054 pgoff_t done_index;
2055 int range_whole = 0;
2056 int scanned = 0;
2057 xa_mark_t tag;
2058
2059 /*
2060 * We have to hold onto the inode so that ordered extents can do their
2061 * work when the IO finishes. The alternative to this is failing to add
2062 * an ordered extent if the igrab() fails there and that is a huge pain
2063 * to deal with, so instead just hold onto the inode throughout the
2064 * writepages operation. If it fails here we are freeing up the inode
2065 * anyway and we'd rather not waste our time writing out stuff that is
2066 * going to be truncated anyway.
2067 */
2068 if (!igrab(inode))
2069 return 0;
2070
2071 folio_batch_init(&fbatch);
2072 if (wbc->range_cyclic) {
2073 index = mapping->writeback_index; /* Start from prev offset */
2074 end = -1;
2075 /*
2076 * Start from the beginning does not need to cycle over the
2077 * range, mark it as scanned.
2078 */
2079 scanned = (index == 0);
2080 } else {
2081 index = wbc->range_start >> PAGE_SHIFT;
2082 end = wbc->range_end >> PAGE_SHIFT;
2083 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2084 range_whole = 1;
2085 scanned = 1;
2086 }
2087
2088 /*
2089 * We do the tagged writepage as long as the snapshot flush bit is set
2090 * and we are the first one who do the filemap_flush() on this inode.
2091 *
2092 * The nr_to_write == LONG_MAX is needed to make sure other flushers do
2093 * not race in and drop the bit.
2094 */
2095 if (range_whole && wbc->nr_to_write == LONG_MAX &&
2096 test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
2097 &BTRFS_I(inode)->runtime_flags))
2098 wbc->tagged_writepages = 1;
2099
2100 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2101 tag = PAGECACHE_TAG_TOWRITE;
2102 else
2103 tag = PAGECACHE_TAG_DIRTY;
2104retry:
2105 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2106 tag_pages_for_writeback(mapping, index, end);
2107 done_index = index;
2108 while (!done && !nr_to_write_done && (index <= end) &&
2109 (nr_folios = filemap_get_folios_tag(mapping, &index,
2110 end, tag, &fbatch))) {
2111 unsigned i;
2112
2113 for (i = 0; i < nr_folios; i++) {
2114 struct folio *folio = fbatch.folios[i];
2115
2116 done_index = folio_next_index(folio);
2117 /*
2118 * At this point we hold neither the i_pages lock nor
2119 * the page lock: the page may be truncated or
2120 * invalidated (changing page->mapping to NULL),
2121 * or even swizzled back from swapper_space to
2122 * tmpfs file mapping
2123 */
2124 if (!folio_trylock(folio)) {
2125 submit_write_bio(bio_ctrl, 0);
2126 folio_lock(folio);
2127 }
2128
2129 if (unlikely(folio->mapping != mapping)) {
2130 folio_unlock(folio);
2131 continue;
2132 }
2133
2134 if (!folio_test_dirty(folio)) {
2135 /* Someone wrote it for us. */
2136 folio_unlock(folio);
2137 continue;
2138 }
2139
2140 if (wbc->sync_mode != WB_SYNC_NONE) {
2141 if (folio_test_writeback(folio))
2142 submit_write_bio(bio_ctrl, 0);
2143 folio_wait_writeback(folio);
2144 }
2145
2146 if (folio_test_writeback(folio) ||
2147 !folio_clear_dirty_for_io(folio)) {
2148 folio_unlock(folio);
2149 continue;
2150 }
2151
2152 ret = __extent_writepage(&folio->page, bio_ctrl);
2153 if (ret < 0) {
2154 done = 1;
2155 break;
2156 }
2157
2158 /*
2159 * The filesystem may choose to bump up nr_to_write.
2160 * We have to make sure to honor the new nr_to_write
2161 * at any time.
2162 */
2163 nr_to_write_done = (wbc->sync_mode == WB_SYNC_NONE &&
2164 wbc->nr_to_write <= 0);
2165 }
2166 folio_batch_release(&fbatch);
2167 cond_resched();
2168 }
2169 if (!scanned && !done) {
2170 /*
2171 * We hit the last page and there is more work to be done: wrap
2172 * back to the start of the file
2173 */
2174 scanned = 1;
2175 index = 0;
2176
2177 /*
2178 * If we're looping we could run into a page that is locked by a
2179 * writer and that writer could be waiting on writeback for a
2180 * page in our current bio, and thus deadlock, so flush the
2181 * write bio here.
2182 */
2183 submit_write_bio(bio_ctrl, 0);
2184 goto retry;
2185 }
2186
2187 if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
2188 mapping->writeback_index = done_index;
2189
2190 btrfs_add_delayed_iput(BTRFS_I(inode));
2191 return ret;
2192}
2193
2194/*
2195 * Submit the pages in the range to bio for call sites which delalloc range has
2196 * already been ran (aka, ordered extent inserted) and all pages are still
2197 * locked.
2198 */
2199void extent_write_locked_range(struct inode *inode, struct page *locked_page,
2200 u64 start, u64 end, struct writeback_control *wbc,
2201 bool pages_dirty)
2202{
2203 bool found_error = false;
2204 int ret = 0;
2205 struct address_space *mapping = inode->i_mapping;
2206 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2207 const u32 sectorsize = fs_info->sectorsize;
2208 loff_t i_size = i_size_read(inode);
2209 u64 cur = start;
2210 struct btrfs_bio_ctrl bio_ctrl = {
2211 .wbc = wbc,
2212 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2213 };
2214
2215 if (wbc->no_cgroup_owner)
2216 bio_ctrl.opf |= REQ_BTRFS_CGROUP_PUNT;
2217
2218 ASSERT(IS_ALIGNED(start, sectorsize) && IS_ALIGNED(end + 1, sectorsize));
2219
2220 while (cur <= end) {
2221 u64 cur_end = min(round_down(cur, PAGE_SIZE) + PAGE_SIZE - 1, end);
2222 u32 cur_len = cur_end + 1 - cur;
2223 struct page *page;
2224 int nr = 0;
2225
2226 page = find_get_page(mapping, cur >> PAGE_SHIFT);
2227 ASSERT(PageLocked(page));
2228 if (pages_dirty && page != locked_page) {
2229 ASSERT(PageDirty(page));
2230 clear_page_dirty_for_io(page);
2231 }
2232
2233 ret = __extent_writepage_io(BTRFS_I(inode), page, &bio_ctrl,
2234 i_size, &nr);
2235 if (ret == 1)
2236 goto next_page;
2237
2238 /* Make sure the mapping tag for page dirty gets cleared. */
2239 if (nr == 0) {
2240 set_page_writeback(page);
2241 end_page_writeback(page);
2242 }
2243 if (ret) {
2244 btrfs_mark_ordered_io_finished(BTRFS_I(inode), page,
2245 cur, cur_len, !ret);
2246 mapping_set_error(page->mapping, ret);
2247 }
2248 btrfs_folio_unlock_writer(fs_info, page_folio(page), cur, cur_len);
2249 if (ret < 0)
2250 found_error = true;
2251next_page:
2252 put_page(page);
2253 cur = cur_end + 1;
2254 }
2255
2256 submit_write_bio(&bio_ctrl, found_error ? ret : 0);
2257}
2258
2259int extent_writepages(struct address_space *mapping,
2260 struct writeback_control *wbc)
2261{
2262 struct inode *inode = mapping->host;
2263 int ret = 0;
2264 struct btrfs_bio_ctrl bio_ctrl = {
2265 .wbc = wbc,
2266 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2267 };
2268
2269 /*
2270 * Allow only a single thread to do the reloc work in zoned mode to
2271 * protect the write pointer updates.
2272 */
2273 btrfs_zoned_data_reloc_lock(BTRFS_I(inode));
2274 ret = extent_write_cache_pages(mapping, &bio_ctrl);
2275 submit_write_bio(&bio_ctrl, ret);
2276 btrfs_zoned_data_reloc_unlock(BTRFS_I(inode));
2277 return ret;
2278}
2279
2280void extent_readahead(struct readahead_control *rac)
2281{
2282 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ | REQ_RAHEAD };
2283 struct page *pagepool[16];
2284 struct extent_map *em_cached = NULL;
2285 u64 prev_em_start = (u64)-1;
2286 int nr;
2287
2288 while ((nr = readahead_page_batch(rac, pagepool))) {
2289 u64 contig_start = readahead_pos(rac);
2290 u64 contig_end = contig_start + readahead_batch_length(rac) - 1;
2291
2292 contiguous_readpages(pagepool, nr, contig_start, contig_end,
2293 &em_cached, &bio_ctrl, &prev_em_start);
2294 }
2295
2296 if (em_cached)
2297 free_extent_map(em_cached);
2298 submit_one_bio(&bio_ctrl);
2299}
2300
2301/*
2302 * basic invalidate_folio code, this waits on any locked or writeback
2303 * ranges corresponding to the folio, and then deletes any extent state
2304 * records from the tree
2305 */
2306int extent_invalidate_folio(struct extent_io_tree *tree,
2307 struct folio *folio, size_t offset)
2308{
2309 struct extent_state *cached_state = NULL;
2310 u64 start = folio_pos(folio);
2311 u64 end = start + folio_size(folio) - 1;
2312 size_t blocksize = folio->mapping->host->i_sb->s_blocksize;
2313
2314 /* This function is only called for the btree inode */
2315 ASSERT(tree->owner == IO_TREE_BTREE_INODE_IO);
2316
2317 start += ALIGN(offset, blocksize);
2318 if (start > end)
2319 return 0;
2320
2321 lock_extent(tree, start, end, &cached_state);
2322 folio_wait_writeback(folio);
2323
2324 /*
2325 * Currently for btree io tree, only EXTENT_LOCKED is utilized,
2326 * so here we only need to unlock the extent range to free any
2327 * existing extent state.
2328 */
2329 unlock_extent(tree, start, end, &cached_state);
2330 return 0;
2331}
2332
2333/*
2334 * a helper for release_folio, this tests for areas of the page that
2335 * are locked or under IO and drops the related state bits if it is safe
2336 * to drop the page.
2337 */
2338static int try_release_extent_state(struct extent_io_tree *tree,
2339 struct page *page, gfp_t mask)
2340{
2341 u64 start = page_offset(page);
2342 u64 end = start + PAGE_SIZE - 1;
2343 int ret = 1;
2344
2345 if (test_range_bit_exists(tree, start, end, EXTENT_LOCKED)) {
2346 ret = 0;
2347 } else {
2348 u32 clear_bits = ~(EXTENT_LOCKED | EXTENT_NODATASUM |
2349 EXTENT_DELALLOC_NEW | EXTENT_CTLBITS |
2350 EXTENT_QGROUP_RESERVED);
2351
2352 /*
2353 * At this point we can safely clear everything except the
2354 * locked bit, the nodatasum bit and the delalloc new bit.
2355 * The delalloc new bit will be cleared by ordered extent
2356 * completion.
2357 */
2358 ret = __clear_extent_bit(tree, start, end, clear_bits, NULL, NULL);
2359
2360 /* if clear_extent_bit failed for enomem reasons,
2361 * we can't allow the release to continue.
2362 */
2363 if (ret < 0)
2364 ret = 0;
2365 else
2366 ret = 1;
2367 }
2368 return ret;
2369}
2370
2371/*
2372 * a helper for release_folio. As long as there are no locked extents
2373 * in the range corresponding to the page, both state records and extent
2374 * map records are removed
2375 */
2376int try_release_extent_mapping(struct page *page, gfp_t mask)
2377{
2378 struct extent_map *em;
2379 u64 start = page_offset(page);
2380 u64 end = start + PAGE_SIZE - 1;
2381 struct btrfs_inode *btrfs_inode = BTRFS_I(page->mapping->host);
2382 struct extent_io_tree *tree = &btrfs_inode->io_tree;
2383 struct extent_map_tree *map = &btrfs_inode->extent_tree;
2384
2385 if (gfpflags_allow_blocking(mask) &&
2386 page->mapping->host->i_size > SZ_16M) {
2387 u64 len;
2388 while (start <= end) {
2389 struct btrfs_fs_info *fs_info;
2390 u64 cur_gen;
2391
2392 len = end - start + 1;
2393 write_lock(&map->lock);
2394 em = lookup_extent_mapping(map, start, len);
2395 if (!em) {
2396 write_unlock(&map->lock);
2397 break;
2398 }
2399 if ((em->flags & EXTENT_FLAG_PINNED) ||
2400 em->start != start) {
2401 write_unlock(&map->lock);
2402 free_extent_map(em);
2403 break;
2404 }
2405 if (test_range_bit_exists(tree, em->start,
2406 extent_map_end(em) - 1,
2407 EXTENT_LOCKED))
2408 goto next;
2409 /*
2410 * If it's not in the list of modified extents, used
2411 * by a fast fsync, we can remove it. If it's being
2412 * logged we can safely remove it since fsync took an
2413 * extra reference on the em.
2414 */
2415 if (list_empty(&em->list) ||
2416 (em->flags & EXTENT_FLAG_LOGGING))
2417 goto remove_em;
2418 /*
2419 * If it's in the list of modified extents, remove it
2420 * only if its generation is older then the current one,
2421 * in which case we don't need it for a fast fsync.
2422 * Otherwise don't remove it, we could be racing with an
2423 * ongoing fast fsync that could miss the new extent.
2424 */
2425 fs_info = btrfs_inode->root->fs_info;
2426 spin_lock(&fs_info->trans_lock);
2427 cur_gen = fs_info->generation;
2428 spin_unlock(&fs_info->trans_lock);
2429 if (em->generation >= cur_gen)
2430 goto next;
2431remove_em:
2432 /*
2433 * We only remove extent maps that are not in the list of
2434 * modified extents or that are in the list but with a
2435 * generation lower then the current generation, so there
2436 * is no need to set the full fsync flag on the inode (it
2437 * hurts the fsync performance for workloads with a data
2438 * size that exceeds or is close to the system's memory).
2439 */
2440 remove_extent_mapping(map, em);
2441 /* once for the rb tree */
2442 free_extent_map(em);
2443next:
2444 start = extent_map_end(em);
2445 write_unlock(&map->lock);
2446
2447 /* once for us */
2448 free_extent_map(em);
2449
2450 cond_resched(); /* Allow large-extent preemption. */
2451 }
2452 }
2453 return try_release_extent_state(tree, page, mask);
2454}
2455
2456/*
2457 * To cache previous fiemap extent
2458 *
2459 * Will be used for merging fiemap extent
2460 */
2461struct fiemap_cache {
2462 u64 offset;
2463 u64 phys;
2464 u64 len;
2465 u32 flags;
2466 bool cached;
2467};
2468
2469/*
2470 * Helper to submit fiemap extent.
2471 *
2472 * Will try to merge current fiemap extent specified by @offset, @phys,
2473 * @len and @flags with cached one.
2474 * And only when we fails to merge, cached one will be submitted as
2475 * fiemap extent.
2476 *
2477 * Return value is the same as fiemap_fill_next_extent().
2478 */
2479static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
2480 struct fiemap_cache *cache,
2481 u64 offset, u64 phys, u64 len, u32 flags)
2482{
2483 u64 cache_end;
2484 int ret = 0;
2485
2486 /* Set at the end of extent_fiemap(). */
2487 ASSERT((flags & FIEMAP_EXTENT_LAST) == 0);
2488
2489 if (!cache->cached)
2490 goto assign;
2491
2492 /*
2493 * When iterating the extents of the inode, at extent_fiemap(), we may
2494 * find an extent that starts at an offset behind the end offset of the
2495 * previous extent we processed. This happens if fiemap is called
2496 * without FIEMAP_FLAG_SYNC and there are ordered extents completing
2497 * while we call btrfs_next_leaf() (through fiemap_next_leaf_item()).
2498 *
2499 * For example we are in leaf X processing its last item, which is the
2500 * file extent item for file range [512K, 1M[, and after
2501 * btrfs_next_leaf() releases the path, there's an ordered extent that
2502 * completes for the file range [768K, 2M[, and that results in trimming
2503 * the file extent item so that it now corresponds to the file range
2504 * [512K, 768K[ and a new file extent item is inserted for the file
2505 * range [768K, 2M[, which may end up as the last item of leaf X or as
2506 * the first item of the next leaf - in either case btrfs_next_leaf()
2507 * will leave us with a path pointing to the new extent item, for the
2508 * file range [768K, 2M[, since that's the first key that follows the
2509 * last one we processed. So in order not to report overlapping extents
2510 * to user space, we trim the length of the previously cached extent and
2511 * emit it.
2512 *
2513 * Upon calling btrfs_next_leaf() we may also find an extent with an
2514 * offset smaller than or equals to cache->offset, and this happens
2515 * when we had a hole or prealloc extent with several delalloc ranges in
2516 * it, but after btrfs_next_leaf() released the path, delalloc was
2517 * flushed and the resulting ordered extents were completed, so we can
2518 * now have found a file extent item for an offset that is smaller than
2519 * or equals to what we have in cache->offset. We deal with this as
2520 * described below.
2521 */
2522 cache_end = cache->offset + cache->len;
2523 if (cache_end > offset) {
2524 if (offset == cache->offset) {
2525 /*
2526 * We cached a dealloc range (found in the io tree) for
2527 * a hole or prealloc extent and we have now found a
2528 * file extent item for the same offset. What we have
2529 * now is more recent and up to date, so discard what
2530 * we had in the cache and use what we have just found.
2531 */
2532 goto assign;
2533 } else if (offset > cache->offset) {
2534 /*
2535 * The extent range we previously found ends after the
2536 * offset of the file extent item we found and that
2537 * offset falls somewhere in the middle of that previous
2538 * extent range. So adjust the range we previously found
2539 * to end at the offset of the file extent item we have
2540 * just found, since this extent is more up to date.
2541 * Emit that adjusted range and cache the file extent
2542 * item we have just found. This corresponds to the case
2543 * where a previously found file extent item was split
2544 * due to an ordered extent completing.
2545 */
2546 cache->len = offset - cache->offset;
2547 goto emit;
2548 } else {
2549 const u64 range_end = offset + len;
2550
2551 /*
2552 * The offset of the file extent item we have just found
2553 * is behind the cached offset. This means we were
2554 * processing a hole or prealloc extent for which we
2555 * have found delalloc ranges (in the io tree), so what
2556 * we have in the cache is the last delalloc range we
2557 * found while the file extent item we found can be
2558 * either for a whole delalloc range we previously
2559 * emmitted or only a part of that range.
2560 *
2561 * We have two cases here:
2562 *
2563 * 1) The file extent item's range ends at or behind the
2564 * cached extent's end. In this case just ignore the
2565 * current file extent item because we don't want to
2566 * overlap with previous ranges that may have been
2567 * emmitted already;
2568 *
2569 * 2) The file extent item starts behind the currently
2570 * cached extent but its end offset goes beyond the
2571 * end offset of the cached extent. We don't want to
2572 * overlap with a previous range that may have been
2573 * emmitted already, so we emit the currently cached
2574 * extent and then partially store the current file
2575 * extent item's range in the cache, for the subrange
2576 * going the cached extent's end to the end of the
2577 * file extent item.
2578 */
2579 if (range_end <= cache_end)
2580 return 0;
2581
2582 if (!(flags & (FIEMAP_EXTENT_ENCODED | FIEMAP_EXTENT_DELALLOC)))
2583 phys += cache_end - offset;
2584
2585 offset = cache_end;
2586 len = range_end - cache_end;
2587 goto emit;
2588 }
2589 }
2590
2591 /*
2592 * Only merges fiemap extents if
2593 * 1) Their logical addresses are continuous
2594 *
2595 * 2) Their physical addresses are continuous
2596 * So truly compressed (physical size smaller than logical size)
2597 * extents won't get merged with each other
2598 *
2599 * 3) Share same flags
2600 */
2601 if (cache->offset + cache->len == offset &&
2602 cache->phys + cache->len == phys &&
2603 cache->flags == flags) {
2604 cache->len += len;
2605 return 0;
2606 }
2607
2608emit:
2609 /* Not mergeable, need to submit cached one */
2610 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
2611 cache->len, cache->flags);
2612 cache->cached = false;
2613 if (ret)
2614 return ret;
2615assign:
2616 cache->cached = true;
2617 cache->offset = offset;
2618 cache->phys = phys;
2619 cache->len = len;
2620 cache->flags = flags;
2621
2622 return 0;
2623}
2624
2625/*
2626 * Emit last fiemap cache
2627 *
2628 * The last fiemap cache may still be cached in the following case:
2629 * 0 4k 8k
2630 * |<- Fiemap range ->|
2631 * |<------------ First extent ----------->|
2632 *
2633 * In this case, the first extent range will be cached but not emitted.
2634 * So we must emit it before ending extent_fiemap().
2635 */
2636static int emit_last_fiemap_cache(struct fiemap_extent_info *fieinfo,
2637 struct fiemap_cache *cache)
2638{
2639 int ret;
2640
2641 if (!cache->cached)
2642 return 0;
2643
2644 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
2645 cache->len, cache->flags);
2646 cache->cached = false;
2647 if (ret > 0)
2648 ret = 0;
2649 return ret;
2650}
2651
2652static int fiemap_next_leaf_item(struct btrfs_inode *inode, struct btrfs_path *path)
2653{
2654 struct extent_buffer *clone;
2655 struct btrfs_key key;
2656 int slot;
2657 int ret;
2658
2659 path->slots[0]++;
2660 if (path->slots[0] < btrfs_header_nritems(path->nodes[0]))
2661 return 0;
2662
2663 ret = btrfs_next_leaf(inode->root, path);
2664 if (ret != 0)
2665 return ret;
2666
2667 /*
2668 * Don't bother with cloning if there are no more file extent items for
2669 * our inode.
2670 */
2671 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2672 if (key.objectid != btrfs_ino(inode) || key.type != BTRFS_EXTENT_DATA_KEY)
2673 return 1;
2674
2675 /* See the comment at fiemap_search_slot() about why we clone. */
2676 clone = btrfs_clone_extent_buffer(path->nodes[0]);
2677 if (!clone)
2678 return -ENOMEM;
2679
2680 slot = path->slots[0];
2681 btrfs_release_path(path);
2682 path->nodes[0] = clone;
2683 path->slots[0] = slot;
2684
2685 return 0;
2686}
2687
2688/*
2689 * Search for the first file extent item that starts at a given file offset or
2690 * the one that starts immediately before that offset.
2691 * Returns: 0 on success, < 0 on error, 1 if not found.
2692 */
2693static int fiemap_search_slot(struct btrfs_inode *inode, struct btrfs_path *path,
2694 u64 file_offset)
2695{
2696 const u64 ino = btrfs_ino(inode);
2697 struct btrfs_root *root = inode->root;
2698 struct extent_buffer *clone;
2699 struct btrfs_key key;
2700 int slot;
2701 int ret;
2702
2703 key.objectid = ino;
2704 key.type = BTRFS_EXTENT_DATA_KEY;
2705 key.offset = file_offset;
2706
2707 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2708 if (ret < 0)
2709 return ret;
2710
2711 if (ret > 0 && path->slots[0] > 0) {
2712 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
2713 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
2714 path->slots[0]--;
2715 }
2716
2717 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2718 ret = btrfs_next_leaf(root, path);
2719 if (ret != 0)
2720 return ret;
2721
2722 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2723 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
2724 return 1;
2725 }
2726
2727 /*
2728 * We clone the leaf and use it during fiemap. This is because while
2729 * using the leaf we do expensive things like checking if an extent is
2730 * shared, which can take a long time. In order to prevent blocking
2731 * other tasks for too long, we use a clone of the leaf. We have locked
2732 * the file range in the inode's io tree, so we know none of our file
2733 * extent items can change. This way we avoid blocking other tasks that
2734 * want to insert items for other inodes in the same leaf or b+tree
2735 * rebalance operations (triggered for example when someone is trying
2736 * to push items into this leaf when trying to insert an item in a
2737 * neighbour leaf).
2738 * We also need the private clone because holding a read lock on an
2739 * extent buffer of the subvolume's b+tree will make lockdep unhappy
2740 * when we call fiemap_fill_next_extent(), because that may cause a page
2741 * fault when filling the user space buffer with fiemap data.
2742 */
2743 clone = btrfs_clone_extent_buffer(path->nodes[0]);
2744 if (!clone)
2745 return -ENOMEM;
2746
2747 slot = path->slots[0];
2748 btrfs_release_path(path);
2749 path->nodes[0] = clone;
2750 path->slots[0] = slot;
2751
2752 return 0;
2753}
2754
2755/*
2756 * Process a range which is a hole or a prealloc extent in the inode's subvolume
2757 * btree. If @disk_bytenr is 0, we are dealing with a hole, otherwise a prealloc
2758 * extent. The end offset (@end) is inclusive.
2759 */
2760static int fiemap_process_hole(struct btrfs_inode *inode,
2761 struct fiemap_extent_info *fieinfo,
2762 struct fiemap_cache *cache,
2763 struct extent_state **delalloc_cached_state,
2764 struct btrfs_backref_share_check_ctx *backref_ctx,
2765 u64 disk_bytenr, u64 extent_offset,
2766 u64 extent_gen,
2767 u64 start, u64 end)
2768{
2769 const u64 i_size = i_size_read(&inode->vfs_inode);
2770 u64 cur_offset = start;
2771 u64 last_delalloc_end = 0;
2772 u32 prealloc_flags = FIEMAP_EXTENT_UNWRITTEN;
2773 bool checked_extent_shared = false;
2774 int ret;
2775
2776 /*
2777 * There can be no delalloc past i_size, so don't waste time looking for
2778 * it beyond i_size.
2779 */
2780 while (cur_offset < end && cur_offset < i_size) {
2781 struct extent_state *cached_state = NULL;
2782 u64 delalloc_start;
2783 u64 delalloc_end;
2784 u64 prealloc_start;
2785 u64 lockstart;
2786 u64 lockend;
2787 u64 prealloc_len = 0;
2788 bool delalloc;
2789
2790 lockstart = round_down(cur_offset, inode->root->fs_info->sectorsize);
2791 lockend = round_up(end, inode->root->fs_info->sectorsize);
2792
2793 /*
2794 * We are only locking for the delalloc range because that's the
2795 * only thing that can change here. With fiemap we have a lock
2796 * on the inode, so no buffered or direct writes can happen.
2797 *
2798 * However mmaps and normal page writeback will cause this to
2799 * change arbitrarily. We have to lock the extent lock here to
2800 * make sure that nobody messes with the tree while we're doing
2801 * btrfs_find_delalloc_in_range.
2802 */
2803 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
2804 delalloc = btrfs_find_delalloc_in_range(inode, cur_offset, end,
2805 delalloc_cached_state,
2806 &delalloc_start,
2807 &delalloc_end);
2808 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
2809 if (!delalloc)
2810 break;
2811
2812 /*
2813 * If this is a prealloc extent we have to report every section
2814 * of it that has no delalloc.
2815 */
2816 if (disk_bytenr != 0) {
2817 if (last_delalloc_end == 0) {
2818 prealloc_start = start;
2819 prealloc_len = delalloc_start - start;
2820 } else {
2821 prealloc_start = last_delalloc_end + 1;
2822 prealloc_len = delalloc_start - prealloc_start;
2823 }
2824 }
2825
2826 if (prealloc_len > 0) {
2827 if (!checked_extent_shared && fieinfo->fi_extents_max) {
2828 ret = btrfs_is_data_extent_shared(inode,
2829 disk_bytenr,
2830 extent_gen,
2831 backref_ctx);
2832 if (ret < 0)
2833 return ret;
2834 else if (ret > 0)
2835 prealloc_flags |= FIEMAP_EXTENT_SHARED;
2836
2837 checked_extent_shared = true;
2838 }
2839 ret = emit_fiemap_extent(fieinfo, cache, prealloc_start,
2840 disk_bytenr + extent_offset,
2841 prealloc_len, prealloc_flags);
2842 if (ret)
2843 return ret;
2844 extent_offset += prealloc_len;
2845 }
2846
2847 ret = emit_fiemap_extent(fieinfo, cache, delalloc_start, 0,
2848 delalloc_end + 1 - delalloc_start,
2849 FIEMAP_EXTENT_DELALLOC |
2850 FIEMAP_EXTENT_UNKNOWN);
2851 if (ret)
2852 return ret;
2853
2854 last_delalloc_end = delalloc_end;
2855 cur_offset = delalloc_end + 1;
2856 extent_offset += cur_offset - delalloc_start;
2857 cond_resched();
2858 }
2859
2860 /*
2861 * Either we found no delalloc for the whole prealloc extent or we have
2862 * a prealloc extent that spans i_size or starts at or after i_size.
2863 */
2864 if (disk_bytenr != 0 && last_delalloc_end < end) {
2865 u64 prealloc_start;
2866 u64 prealloc_len;
2867
2868 if (last_delalloc_end == 0) {
2869 prealloc_start = start;
2870 prealloc_len = end + 1 - start;
2871 } else {
2872 prealloc_start = last_delalloc_end + 1;
2873 prealloc_len = end + 1 - prealloc_start;
2874 }
2875
2876 if (!checked_extent_shared && fieinfo->fi_extents_max) {
2877 ret = btrfs_is_data_extent_shared(inode,
2878 disk_bytenr,
2879 extent_gen,
2880 backref_ctx);
2881 if (ret < 0)
2882 return ret;
2883 else if (ret > 0)
2884 prealloc_flags |= FIEMAP_EXTENT_SHARED;
2885 }
2886 ret = emit_fiemap_extent(fieinfo, cache, prealloc_start,
2887 disk_bytenr + extent_offset,
2888 prealloc_len, prealloc_flags);
2889 if (ret)
2890 return ret;
2891 }
2892
2893 return 0;
2894}
2895
2896static int fiemap_find_last_extent_offset(struct btrfs_inode *inode,
2897 struct btrfs_path *path,
2898 u64 *last_extent_end_ret)
2899{
2900 const u64 ino = btrfs_ino(inode);
2901 struct btrfs_root *root = inode->root;
2902 struct extent_buffer *leaf;
2903 struct btrfs_file_extent_item *ei;
2904 struct btrfs_key key;
2905 u64 disk_bytenr;
2906 int ret;
2907
2908 /*
2909 * Lookup the last file extent. We're not using i_size here because
2910 * there might be preallocation past i_size.
2911 */
2912 ret = btrfs_lookup_file_extent(NULL, root, path, ino, (u64)-1, 0);
2913 /* There can't be a file extent item at offset (u64)-1 */
2914 ASSERT(ret != 0);
2915 if (ret < 0)
2916 return ret;
2917
2918 /*
2919 * For a non-existing key, btrfs_search_slot() always leaves us at a
2920 * slot > 0, except if the btree is empty, which is impossible because
2921 * at least it has the inode item for this inode and all the items for
2922 * the root inode 256.
2923 */
2924 ASSERT(path->slots[0] > 0);
2925 path->slots[0]--;
2926 leaf = path->nodes[0];
2927 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2928 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY) {
2929 /* No file extent items in the subvolume tree. */
2930 *last_extent_end_ret = 0;
2931 return 0;
2932 }
2933
2934 /*
2935 * For an inline extent, the disk_bytenr is where inline data starts at,
2936 * so first check if we have an inline extent item before checking if we
2937 * have an implicit hole (disk_bytenr == 0).
2938 */
2939 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
2940 if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_INLINE) {
2941 *last_extent_end_ret = btrfs_file_extent_end(path);
2942 return 0;
2943 }
2944
2945 /*
2946 * Find the last file extent item that is not a hole (when NO_HOLES is
2947 * not enabled). This should take at most 2 iterations in the worst
2948 * case: we have one hole file extent item at slot 0 of a leaf and
2949 * another hole file extent item as the last item in the previous leaf.
2950 * This is because we merge file extent items that represent holes.
2951 */
2952 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
2953 while (disk_bytenr == 0) {
2954 ret = btrfs_previous_item(root, path, ino, BTRFS_EXTENT_DATA_KEY);
2955 if (ret < 0) {
2956 return ret;
2957 } else if (ret > 0) {
2958 /* No file extent items that are not holes. */
2959 *last_extent_end_ret = 0;
2960 return 0;
2961 }
2962 leaf = path->nodes[0];
2963 ei = btrfs_item_ptr(leaf, path->slots[0],
2964 struct btrfs_file_extent_item);
2965 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
2966 }
2967
2968 *last_extent_end_ret = btrfs_file_extent_end(path);
2969 return 0;
2970}
2971
2972int extent_fiemap(struct btrfs_inode *inode, struct fiemap_extent_info *fieinfo,
2973 u64 start, u64 len)
2974{
2975 const u64 ino = btrfs_ino(inode);
2976 struct extent_state *delalloc_cached_state = NULL;
2977 struct btrfs_path *path;
2978 struct fiemap_cache cache = { 0 };
2979 struct btrfs_backref_share_check_ctx *backref_ctx;
2980 u64 last_extent_end;
2981 u64 prev_extent_end;
2982 u64 range_start;
2983 u64 range_end;
2984 const u64 sectorsize = inode->root->fs_info->sectorsize;
2985 bool stopped = false;
2986 int ret;
2987
2988 backref_ctx = btrfs_alloc_backref_share_check_ctx();
2989 path = btrfs_alloc_path();
2990 if (!backref_ctx || !path) {
2991 ret = -ENOMEM;
2992 goto out;
2993 }
2994
2995 range_start = round_down(start, sectorsize);
2996 range_end = round_up(start + len, sectorsize);
2997 prev_extent_end = range_start;
2998
2999 ret = fiemap_find_last_extent_offset(inode, path, &last_extent_end);
3000 if (ret < 0)
3001 goto out;
3002 btrfs_release_path(path);
3003
3004 path->reada = READA_FORWARD;
3005 ret = fiemap_search_slot(inode, path, range_start);
3006 if (ret < 0) {
3007 goto out;
3008 } else if (ret > 0) {
3009 /*
3010 * No file extent item found, but we may have delalloc between
3011 * the current offset and i_size. So check for that.
3012 */
3013 ret = 0;
3014 goto check_eof_delalloc;
3015 }
3016
3017 while (prev_extent_end < range_end) {
3018 struct extent_buffer *leaf = path->nodes[0];
3019 struct btrfs_file_extent_item *ei;
3020 struct btrfs_key key;
3021 u64 extent_end;
3022 u64 extent_len;
3023 u64 extent_offset = 0;
3024 u64 extent_gen;
3025 u64 disk_bytenr = 0;
3026 u64 flags = 0;
3027 int extent_type;
3028 u8 compression;
3029
3030 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3031 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3032 break;
3033
3034 extent_end = btrfs_file_extent_end(path);
3035
3036 /*
3037 * The first iteration can leave us at an extent item that ends
3038 * before our range's start. Move to the next item.
3039 */
3040 if (extent_end <= range_start)
3041 goto next_item;
3042
3043 backref_ctx->curr_leaf_bytenr = leaf->start;
3044
3045 /* We have in implicit hole (NO_HOLES feature enabled). */
3046 if (prev_extent_end < key.offset) {
3047 const u64 hole_end = min(key.offset, range_end) - 1;
3048
3049 ret = fiemap_process_hole(inode, fieinfo, &cache,
3050 &delalloc_cached_state,
3051 backref_ctx, 0, 0, 0,
3052 prev_extent_end, hole_end);
3053 if (ret < 0) {
3054 goto out;
3055 } else if (ret > 0) {
3056 /* fiemap_fill_next_extent() told us to stop. */
3057 stopped = true;
3058 break;
3059 }
3060
3061 /* We've reached the end of the fiemap range, stop. */
3062 if (key.offset >= range_end) {
3063 stopped = true;
3064 break;
3065 }
3066 }
3067
3068 extent_len = extent_end - key.offset;
3069 ei = btrfs_item_ptr(leaf, path->slots[0],
3070 struct btrfs_file_extent_item);
3071 compression = btrfs_file_extent_compression(leaf, ei);
3072 extent_type = btrfs_file_extent_type(leaf, ei);
3073 extent_gen = btrfs_file_extent_generation(leaf, ei);
3074
3075 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
3076 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
3077 if (compression == BTRFS_COMPRESS_NONE)
3078 extent_offset = btrfs_file_extent_offset(leaf, ei);
3079 }
3080
3081 if (compression != BTRFS_COMPRESS_NONE)
3082 flags |= FIEMAP_EXTENT_ENCODED;
3083
3084 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
3085 flags |= FIEMAP_EXTENT_DATA_INLINE;
3086 flags |= FIEMAP_EXTENT_NOT_ALIGNED;
3087 ret = emit_fiemap_extent(fieinfo, &cache, key.offset, 0,
3088 extent_len, flags);
3089 } else if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
3090 ret = fiemap_process_hole(inode, fieinfo, &cache,
3091 &delalloc_cached_state,
3092 backref_ctx,
3093 disk_bytenr, extent_offset,
3094 extent_gen, key.offset,
3095 extent_end - 1);
3096 } else if (disk_bytenr == 0) {
3097 /* We have an explicit hole. */
3098 ret = fiemap_process_hole(inode, fieinfo, &cache,
3099 &delalloc_cached_state,
3100 backref_ctx, 0, 0, 0,
3101 key.offset, extent_end - 1);
3102 } else {
3103 /* We have a regular extent. */
3104 if (fieinfo->fi_extents_max) {
3105 ret = btrfs_is_data_extent_shared(inode,
3106 disk_bytenr,
3107 extent_gen,
3108 backref_ctx);
3109 if (ret < 0)
3110 goto out;
3111 else if (ret > 0)
3112 flags |= FIEMAP_EXTENT_SHARED;
3113 }
3114
3115 ret = emit_fiemap_extent(fieinfo, &cache, key.offset,
3116 disk_bytenr + extent_offset,
3117 extent_len, flags);
3118 }
3119
3120 if (ret < 0) {
3121 goto out;
3122 } else if (ret > 0) {
3123 /* fiemap_fill_next_extent() told us to stop. */
3124 stopped = true;
3125 break;
3126 }
3127
3128 prev_extent_end = extent_end;
3129next_item:
3130 if (fatal_signal_pending(current)) {
3131 ret = -EINTR;
3132 goto out;
3133 }
3134
3135 ret = fiemap_next_leaf_item(inode, path);
3136 if (ret < 0) {
3137 goto out;
3138 } else if (ret > 0) {
3139 /* No more file extent items for this inode. */
3140 break;
3141 }
3142 cond_resched();
3143 }
3144
3145check_eof_delalloc:
3146 /*
3147 * Release (and free) the path before emitting any final entries to
3148 * fiemap_fill_next_extent() to keep lockdep happy. This is because
3149 * once we find no more file extent items exist, we may have a
3150 * non-cloned leaf, and fiemap_fill_next_extent() can trigger page
3151 * faults when copying data to the user space buffer.
3152 */
3153 btrfs_free_path(path);
3154 path = NULL;
3155
3156 if (!stopped && prev_extent_end < range_end) {
3157 ret = fiemap_process_hole(inode, fieinfo, &cache,
3158 &delalloc_cached_state, backref_ctx,
3159 0, 0, 0, prev_extent_end, range_end - 1);
3160 if (ret < 0)
3161 goto out;
3162 prev_extent_end = range_end;
3163 }
3164
3165 if (cache.cached && cache.offset + cache.len >= last_extent_end) {
3166 const u64 i_size = i_size_read(&inode->vfs_inode);
3167
3168 if (prev_extent_end < i_size) {
3169 struct extent_state *cached_state = NULL;
3170 u64 delalloc_start;
3171 u64 delalloc_end;
3172 u64 lockstart;
3173 u64 lockend;
3174 bool delalloc;
3175
3176 lockstart = round_down(prev_extent_end, sectorsize);
3177 lockend = round_up(i_size, sectorsize);
3178
3179 /*
3180 * See the comment in fiemap_process_hole as to why
3181 * we're doing the locking here.
3182 */
3183 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3184 delalloc = btrfs_find_delalloc_in_range(inode,
3185 prev_extent_end,
3186 i_size - 1,
3187 &delalloc_cached_state,
3188 &delalloc_start,
3189 &delalloc_end);
3190 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3191 if (!delalloc)
3192 cache.flags |= FIEMAP_EXTENT_LAST;
3193 } else {
3194 cache.flags |= FIEMAP_EXTENT_LAST;
3195 }
3196 }
3197
3198 ret = emit_last_fiemap_cache(fieinfo, &cache);
3199out:
3200 free_extent_state(delalloc_cached_state);
3201 btrfs_free_backref_share_ctx(backref_ctx);
3202 btrfs_free_path(path);
3203 return ret;
3204}
3205
3206static void __free_extent_buffer(struct extent_buffer *eb)
3207{
3208 kmem_cache_free(extent_buffer_cache, eb);
3209}
3210
3211static int extent_buffer_under_io(const struct extent_buffer *eb)
3212{
3213 return (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
3214 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3215}
3216
3217static bool folio_range_has_eb(struct btrfs_fs_info *fs_info, struct folio *folio)
3218{
3219 struct btrfs_subpage *subpage;
3220
3221 lockdep_assert_held(&folio->mapping->i_private_lock);
3222
3223 if (folio_test_private(folio)) {
3224 subpage = folio_get_private(folio);
3225 if (atomic_read(&subpage->eb_refs))
3226 return true;
3227 /*
3228 * Even there is no eb refs here, we may still have
3229 * end_page_read() call relying on page::private.
3230 */
3231 if (atomic_read(&subpage->readers))
3232 return true;
3233 }
3234 return false;
3235}
3236
3237static void detach_extent_buffer_folio(struct extent_buffer *eb, struct folio *folio)
3238{
3239 struct btrfs_fs_info *fs_info = eb->fs_info;
3240 const bool mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3241
3242 /*
3243 * For mapped eb, we're going to change the folio private, which should
3244 * be done under the i_private_lock.
3245 */
3246 if (mapped)
3247 spin_lock(&folio->mapping->i_private_lock);
3248
3249 if (!folio_test_private(folio)) {
3250 if (mapped)
3251 spin_unlock(&folio->mapping->i_private_lock);
3252 return;
3253 }
3254
3255 if (fs_info->nodesize >= PAGE_SIZE) {
3256 /*
3257 * We do this since we'll remove the pages after we've
3258 * removed the eb from the radix tree, so we could race
3259 * and have this page now attached to the new eb. So
3260 * only clear folio if it's still connected to
3261 * this eb.
3262 */
3263 if (folio_test_private(folio) && folio_get_private(folio) == eb) {
3264 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3265 BUG_ON(folio_test_dirty(folio));
3266 BUG_ON(folio_test_writeback(folio));
3267 /* We need to make sure we haven't be attached to a new eb. */
3268 folio_detach_private(folio);
3269 }
3270 if (mapped)
3271 spin_unlock(&folio->mapping->i_private_lock);
3272 return;
3273 }
3274
3275 /*
3276 * For subpage, we can have dummy eb with folio private attached. In
3277 * this case, we can directly detach the private as such folio is only
3278 * attached to one dummy eb, no sharing.
3279 */
3280 if (!mapped) {
3281 btrfs_detach_subpage(fs_info, folio);
3282 return;
3283 }
3284
3285 btrfs_folio_dec_eb_refs(fs_info, folio);
3286
3287 /*
3288 * We can only detach the folio private if there are no other ebs in the
3289 * page range and no unfinished IO.
3290 */
3291 if (!folio_range_has_eb(fs_info, folio))
3292 btrfs_detach_subpage(fs_info, folio);
3293
3294 spin_unlock(&folio->mapping->i_private_lock);
3295}
3296
3297/* Release all pages attached to the extent buffer */
3298static void btrfs_release_extent_buffer_pages(struct extent_buffer *eb)
3299{
3300 ASSERT(!extent_buffer_under_io(eb));
3301
3302 for (int i = 0; i < INLINE_EXTENT_BUFFER_PAGES; i++) {
3303 struct folio *folio = eb->folios[i];
3304
3305 if (!folio)
3306 continue;
3307
3308 detach_extent_buffer_folio(eb, folio);
3309
3310 /* One for when we allocated the folio. */
3311 folio_put(folio);
3312 }
3313}
3314
3315/*
3316 * Helper for releasing the extent buffer.
3317 */
3318static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3319{
3320 btrfs_release_extent_buffer_pages(eb);
3321 btrfs_leak_debug_del_eb(eb);
3322 __free_extent_buffer(eb);
3323}
3324
3325static struct extent_buffer *
3326__alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
3327 unsigned long len)
3328{
3329 struct extent_buffer *eb = NULL;
3330
3331 eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
3332 eb->start = start;
3333 eb->len = len;
3334 eb->fs_info = fs_info;
3335 init_rwsem(&eb->lock);
3336
3337 btrfs_leak_debug_add_eb(eb);
3338
3339 spin_lock_init(&eb->refs_lock);
3340 atomic_set(&eb->refs, 1);
3341
3342 ASSERT(len <= BTRFS_MAX_METADATA_BLOCKSIZE);
3343
3344 return eb;
3345}
3346
3347struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
3348{
3349 struct extent_buffer *new;
3350 int num_folios = num_extent_folios(src);
3351 int ret;
3352
3353 new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
3354 if (new == NULL)
3355 return NULL;
3356
3357 /*
3358 * Set UNMAPPED before calling btrfs_release_extent_buffer(), as
3359 * btrfs_release_extent_buffer() have different behavior for
3360 * UNMAPPED subpage extent buffer.
3361 */
3362 set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
3363
3364 ret = alloc_eb_folio_array(new, 0);
3365 if (ret) {
3366 btrfs_release_extent_buffer(new);
3367 return NULL;
3368 }
3369
3370 for (int i = 0; i < num_folios; i++) {
3371 struct folio *folio = new->folios[i];
3372 int ret;
3373
3374 ret = attach_extent_buffer_folio(new, folio, NULL);
3375 if (ret < 0) {
3376 btrfs_release_extent_buffer(new);
3377 return NULL;
3378 }
3379 WARN_ON(folio_test_dirty(folio));
3380 }
3381 copy_extent_buffer_full(new, src);
3382 set_extent_buffer_uptodate(new);
3383
3384 return new;
3385}
3386
3387struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
3388 u64 start, unsigned long len)
3389{
3390 struct extent_buffer *eb;
3391 int num_folios = 0;
3392 int ret;
3393
3394 eb = __alloc_extent_buffer(fs_info, start, len);
3395 if (!eb)
3396 return NULL;
3397
3398 ret = alloc_eb_folio_array(eb, 0);
3399 if (ret)
3400 goto err;
3401
3402 num_folios = num_extent_folios(eb);
3403 for (int i = 0; i < num_folios; i++) {
3404 ret = attach_extent_buffer_folio(eb, eb->folios[i], NULL);
3405 if (ret < 0)
3406 goto err;
3407 }
3408
3409 set_extent_buffer_uptodate(eb);
3410 btrfs_set_header_nritems(eb, 0);
3411 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3412
3413 return eb;
3414err:
3415 for (int i = 0; i < num_folios; i++) {
3416 if (eb->folios[i]) {
3417 detach_extent_buffer_folio(eb, eb->folios[i]);
3418 __folio_put(eb->folios[i]);
3419 }
3420 }
3421 __free_extent_buffer(eb);
3422 return NULL;
3423}
3424
3425struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
3426 u64 start)
3427{
3428 return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
3429}
3430
3431static void check_buffer_tree_ref(struct extent_buffer *eb)
3432{
3433 int refs;
3434 /*
3435 * The TREE_REF bit is first set when the extent_buffer is added
3436 * to the radix tree. It is also reset, if unset, when a new reference
3437 * is created by find_extent_buffer.
3438 *
3439 * It is only cleared in two cases: freeing the last non-tree
3440 * reference to the extent_buffer when its STALE bit is set or
3441 * calling release_folio when the tree reference is the only reference.
3442 *
3443 * In both cases, care is taken to ensure that the extent_buffer's
3444 * pages are not under io. However, release_folio can be concurrently
3445 * called with creating new references, which is prone to race
3446 * conditions between the calls to check_buffer_tree_ref in those
3447 * codepaths and clearing TREE_REF in try_release_extent_buffer.
3448 *
3449 * The actual lifetime of the extent_buffer in the radix tree is
3450 * adequately protected by the refcount, but the TREE_REF bit and
3451 * its corresponding reference are not. To protect against this
3452 * class of races, we call check_buffer_tree_ref from the codepaths
3453 * which trigger io. Note that once io is initiated, TREE_REF can no
3454 * longer be cleared, so that is the moment at which any such race is
3455 * best fixed.
3456 */
3457 refs = atomic_read(&eb->refs);
3458 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3459 return;
3460
3461 spin_lock(&eb->refs_lock);
3462 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3463 atomic_inc(&eb->refs);
3464 spin_unlock(&eb->refs_lock);
3465}
3466
3467static void mark_extent_buffer_accessed(struct extent_buffer *eb)
3468{
3469 int num_folios= num_extent_folios(eb);
3470
3471 check_buffer_tree_ref(eb);
3472
3473 for (int i = 0; i < num_folios; i++)
3474 folio_mark_accessed(eb->folios[i]);
3475}
3476
3477struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
3478 u64 start)
3479{
3480 struct extent_buffer *eb;
3481
3482 eb = find_extent_buffer_nolock(fs_info, start);
3483 if (!eb)
3484 return NULL;
3485 /*
3486 * Lock our eb's refs_lock to avoid races with free_extent_buffer().
3487 * When we get our eb it might be flagged with EXTENT_BUFFER_STALE and
3488 * another task running free_extent_buffer() might have seen that flag
3489 * set, eb->refs == 2, that the buffer isn't under IO (dirty and
3490 * writeback flags not set) and it's still in the tree (flag
3491 * EXTENT_BUFFER_TREE_REF set), therefore being in the process of
3492 * decrementing the extent buffer's reference count twice. So here we
3493 * could race and increment the eb's reference count, clear its stale
3494 * flag, mark it as dirty and drop our reference before the other task
3495 * finishes executing free_extent_buffer, which would later result in
3496 * an attempt to free an extent buffer that is dirty.
3497 */
3498 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
3499 spin_lock(&eb->refs_lock);
3500 spin_unlock(&eb->refs_lock);
3501 }
3502 mark_extent_buffer_accessed(eb);
3503 return eb;
3504}
3505
3506#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3507struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
3508 u64 start)
3509{
3510 struct extent_buffer *eb, *exists = NULL;
3511 int ret;
3512
3513 eb = find_extent_buffer(fs_info, start);
3514 if (eb)
3515 return eb;
3516 eb = alloc_dummy_extent_buffer(fs_info, start);
3517 if (!eb)
3518 return ERR_PTR(-ENOMEM);
3519 eb->fs_info = fs_info;
3520again:
3521 ret = radix_tree_preload(GFP_NOFS);
3522 if (ret) {
3523 exists = ERR_PTR(ret);
3524 goto free_eb;
3525 }
3526 spin_lock(&fs_info->buffer_lock);
3527 ret = radix_tree_insert(&fs_info->buffer_radix,
3528 start >> fs_info->sectorsize_bits, eb);
3529 spin_unlock(&fs_info->buffer_lock);
3530 radix_tree_preload_end();
3531 if (ret == -EEXIST) {
3532 exists = find_extent_buffer(fs_info, start);
3533 if (exists)
3534 goto free_eb;
3535 else
3536 goto again;
3537 }
3538 check_buffer_tree_ref(eb);
3539 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
3540
3541 return eb;
3542free_eb:
3543 btrfs_release_extent_buffer(eb);
3544 return exists;
3545}
3546#endif
3547
3548static struct extent_buffer *grab_extent_buffer(
3549 struct btrfs_fs_info *fs_info, struct page *page)
3550{
3551 struct folio *folio = page_folio(page);
3552 struct extent_buffer *exists;
3553
3554 /*
3555 * For subpage case, we completely rely on radix tree to ensure we
3556 * don't try to insert two ebs for the same bytenr. So here we always
3557 * return NULL and just continue.
3558 */
3559 if (fs_info->nodesize < PAGE_SIZE)
3560 return NULL;
3561
3562 /* Page not yet attached to an extent buffer */
3563 if (!folio_test_private(folio))
3564 return NULL;
3565
3566 /*
3567 * We could have already allocated an eb for this page and attached one
3568 * so lets see if we can get a ref on the existing eb, and if we can we
3569 * know it's good and we can just return that one, else we know we can
3570 * just overwrite folio private.
3571 */
3572 exists = folio_get_private(folio);
3573 if (atomic_inc_not_zero(&exists->refs))
3574 return exists;
3575
3576 WARN_ON(PageDirty(page));
3577 folio_detach_private(folio);
3578 return NULL;
3579}
3580
3581static int check_eb_alignment(struct btrfs_fs_info *fs_info, u64 start)
3582{
3583 if (!IS_ALIGNED(start, fs_info->sectorsize)) {
3584 btrfs_err(fs_info, "bad tree block start %llu", start);
3585 return -EINVAL;
3586 }
3587
3588 if (fs_info->nodesize < PAGE_SIZE &&
3589 offset_in_page(start) + fs_info->nodesize > PAGE_SIZE) {
3590 btrfs_err(fs_info,
3591 "tree block crosses page boundary, start %llu nodesize %u",
3592 start, fs_info->nodesize);
3593 return -EINVAL;
3594 }
3595 if (fs_info->nodesize >= PAGE_SIZE &&
3596 !PAGE_ALIGNED(start)) {
3597 btrfs_err(fs_info,
3598 "tree block is not page aligned, start %llu nodesize %u",
3599 start, fs_info->nodesize);
3600 return -EINVAL;
3601 }
3602 if (!IS_ALIGNED(start, fs_info->nodesize) &&
3603 !test_and_set_bit(BTRFS_FS_UNALIGNED_TREE_BLOCK, &fs_info->flags)) {
3604 btrfs_warn(fs_info,
3605"tree block not nodesize aligned, start %llu nodesize %u, can be resolved by a full metadata balance",
3606 start, fs_info->nodesize);
3607 }
3608 return 0;
3609}
3610
3611
3612/*
3613 * Return 0 if eb->folios[i] is attached to btree inode successfully.
3614 * Return >0 if there is already another extent buffer for the range,
3615 * and @found_eb_ret would be updated.
3616 * Return -EAGAIN if the filemap has an existing folio but with different size
3617 * than @eb.
3618 * The caller needs to free the existing folios and retry using the same order.
3619 */
3620static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
3621 struct extent_buffer **found_eb_ret)
3622{
3623
3624 struct btrfs_fs_info *fs_info = eb->fs_info;
3625 struct address_space *mapping = fs_info->btree_inode->i_mapping;
3626 const unsigned long index = eb->start >> PAGE_SHIFT;
3627 struct folio *existing_folio;
3628 int ret;
3629
3630 ASSERT(found_eb_ret);
3631
3632 /* Caller should ensure the folio exists. */
3633 ASSERT(eb->folios[i]);
3634
3635retry:
3636 ret = filemap_add_folio(mapping, eb->folios[i], index + i,
3637 GFP_NOFS | __GFP_NOFAIL);
3638 if (!ret)
3639 return 0;
3640
3641 existing_folio = filemap_lock_folio(mapping, index + i);
3642 /* The page cache only exists for a very short time, just retry. */
3643 if (IS_ERR(existing_folio))
3644 goto retry;
3645
3646 /* For now, we should only have single-page folios for btree inode. */
3647 ASSERT(folio_nr_pages(existing_folio) == 1);
3648
3649 if (folio_size(existing_folio) != folio_size(eb->folios[0])) {
3650 folio_unlock(existing_folio);
3651 folio_put(existing_folio);
3652 return -EAGAIN;
3653 }
3654
3655 if (fs_info->nodesize < PAGE_SIZE) {
3656 /*
3657 * We're going to reuse the existing page, can drop our page
3658 * and subpage structure now.
3659 */
3660 __free_page(folio_page(eb->folios[i], 0));
3661 eb->folios[i] = existing_folio;
3662 } else {
3663 struct extent_buffer *existing_eb;
3664
3665 existing_eb = grab_extent_buffer(fs_info,
3666 folio_page(existing_folio, 0));
3667 if (existing_eb) {
3668 /* The extent buffer still exists, we can use it directly. */
3669 *found_eb_ret = existing_eb;
3670 folio_unlock(existing_folio);
3671 folio_put(existing_folio);
3672 return 1;
3673 }
3674 /* The extent buffer no longer exists, we can reuse the folio. */
3675 __free_page(folio_page(eb->folios[i], 0));
3676 eb->folios[i] = existing_folio;
3677 }
3678 return 0;
3679}
3680
3681struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
3682 u64 start, u64 owner_root, int level)
3683{
3684 unsigned long len = fs_info->nodesize;
3685 int num_folios;
3686 int attached = 0;
3687 struct extent_buffer *eb;
3688 struct extent_buffer *existing_eb = NULL;
3689 struct address_space *mapping = fs_info->btree_inode->i_mapping;
3690 struct btrfs_subpage *prealloc = NULL;
3691 u64 lockdep_owner = owner_root;
3692 bool page_contig = true;
3693 int uptodate = 1;
3694 int ret;
3695
3696 if (check_eb_alignment(fs_info, start))
3697 return ERR_PTR(-EINVAL);
3698
3699#if BITS_PER_LONG == 32
3700 if (start >= MAX_LFS_FILESIZE) {
3701 btrfs_err_rl(fs_info,
3702 "extent buffer %llu is beyond 32bit page cache limit", start);
3703 btrfs_err_32bit_limit(fs_info);
3704 return ERR_PTR(-EOVERFLOW);
3705 }
3706 if (start >= BTRFS_32BIT_EARLY_WARN_THRESHOLD)
3707 btrfs_warn_32bit_limit(fs_info);
3708#endif
3709
3710 eb = find_extent_buffer(fs_info, start);
3711 if (eb)
3712 return eb;
3713
3714 eb = __alloc_extent_buffer(fs_info, start, len);
3715 if (!eb)
3716 return ERR_PTR(-ENOMEM);
3717
3718 /*
3719 * The reloc trees are just snapshots, so we need them to appear to be
3720 * just like any other fs tree WRT lockdep.
3721 */
3722 if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID)
3723 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
3724
3725 btrfs_set_buffer_lockdep_class(lockdep_owner, eb, level);
3726
3727 /*
3728 * Preallocate folio private for subpage case, so that we won't
3729 * allocate memory with i_private_lock nor page lock hold.
3730 *
3731 * The memory will be freed by attach_extent_buffer_page() or freed
3732 * manually if we exit earlier.
3733 */
3734 if (fs_info->nodesize < PAGE_SIZE) {
3735 prealloc = btrfs_alloc_subpage(fs_info, BTRFS_SUBPAGE_METADATA);
3736 if (IS_ERR(prealloc)) {
3737 ret = PTR_ERR(prealloc);
3738 goto out;
3739 }
3740 }
3741
3742reallocate:
3743 /* Allocate all pages first. */
3744 ret = alloc_eb_folio_array(eb, __GFP_NOFAIL);
3745 if (ret < 0) {
3746 btrfs_free_subpage(prealloc);
3747 goto out;
3748 }
3749
3750 num_folios = num_extent_folios(eb);
3751 /* Attach all pages to the filemap. */
3752 for (int i = 0; i < num_folios; i++) {
3753 struct folio *folio;
3754
3755 ret = attach_eb_folio_to_filemap(eb, i, &existing_eb);
3756 if (ret > 0) {
3757 ASSERT(existing_eb);
3758 goto out;
3759 }
3760
3761 /*
3762 * TODO: Special handling for a corner case where the order of
3763 * folios mismatch between the new eb and filemap.
3764 *
3765 * This happens when:
3766 *
3767 * - the new eb is using higher order folio
3768 *
3769 * - the filemap is still using 0-order folios for the range
3770 * This can happen at the previous eb allocation, and we don't
3771 * have higher order folio for the call.
3772 *
3773 * - the existing eb has already been freed
3774 *
3775 * In this case, we have to free the existing folios first, and
3776 * re-allocate using the same order.
3777 * Thankfully this is not going to happen yet, as we're still
3778 * using 0-order folios.
3779 */
3780 if (unlikely(ret == -EAGAIN)) {
3781 ASSERT(0);
3782 goto reallocate;
3783 }
3784 attached++;
3785
3786 /*
3787 * Only after attach_eb_folio_to_filemap(), eb->folios[] is
3788 * reliable, as we may choose to reuse the existing page cache
3789 * and free the allocated page.
3790 */
3791 folio = eb->folios[i];
3792 spin_lock(&mapping->i_private_lock);
3793 /* Should not fail, as we have preallocated the memory */
3794 ret = attach_extent_buffer_folio(eb, folio, prealloc);
3795 ASSERT(!ret);
3796 /*
3797 * To inform we have extra eb under allocation, so that
3798 * detach_extent_buffer_page() won't release the folio private
3799 * when the eb hasn't yet been inserted into radix tree.
3800 *
3801 * The ref will be decreased when the eb released the page, in
3802 * detach_extent_buffer_page().
3803 * Thus needs no special handling in error path.
3804 */
3805 btrfs_folio_inc_eb_refs(fs_info, folio);
3806 spin_unlock(&mapping->i_private_lock);
3807
3808 WARN_ON(btrfs_folio_test_dirty(fs_info, folio, eb->start, eb->len));
3809
3810 /*
3811 * Check if the current page is physically contiguous with previous eb
3812 * page.
3813 * At this stage, either we allocated a large folio, thus @i
3814 * would only be 0, or we fall back to per-page allocation.
3815 */
3816 if (i && folio_page(eb->folios[i - 1], 0) + 1 != folio_page(folio, 0))
3817 page_contig = false;
3818
3819 if (!btrfs_folio_test_uptodate(fs_info, folio, eb->start, eb->len))
3820 uptodate = 0;
3821
3822 /*
3823 * We can't unlock the pages just yet since the extent buffer
3824 * hasn't been properly inserted in the radix tree, this
3825 * opens a race with btree_release_folio which can free a page
3826 * while we are still filling in all pages for the buffer and
3827 * we could crash.
3828 */
3829 }
3830 if (uptodate)
3831 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3832 /* All pages are physically contiguous, can skip cross page handling. */
3833 if (page_contig)
3834 eb->addr = folio_address(eb->folios[0]) + offset_in_page(eb->start);
3835again:
3836 ret = radix_tree_preload(GFP_NOFS);
3837 if (ret)
3838 goto out;
3839
3840 spin_lock(&fs_info->buffer_lock);
3841 ret = radix_tree_insert(&fs_info->buffer_radix,
3842 start >> fs_info->sectorsize_bits, eb);
3843 spin_unlock(&fs_info->buffer_lock);
3844 radix_tree_preload_end();
3845 if (ret == -EEXIST) {
3846 ret = 0;
3847 existing_eb = find_extent_buffer(fs_info, start);
3848 if (existing_eb)
3849 goto out;
3850 else
3851 goto again;
3852 }
3853 /* add one reference for the tree */
3854 check_buffer_tree_ref(eb);
3855 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
3856
3857 /*
3858 * Now it's safe to unlock the pages because any calls to
3859 * btree_release_folio will correctly detect that a page belongs to a
3860 * live buffer and won't free them prematurely.
3861 */
3862 for (int i = 0; i < num_folios; i++)
3863 unlock_page(folio_page(eb->folios[i], 0));
3864 return eb;
3865
3866out:
3867 WARN_ON(!atomic_dec_and_test(&eb->refs));
3868
3869 /*
3870 * Any attached folios need to be detached before we unlock them. This
3871 * is because when we're inserting our new folios into the mapping, and
3872 * then attaching our eb to that folio. If we fail to insert our folio
3873 * we'll lookup the folio for that index, and grab that EB. We do not
3874 * want that to grab this eb, as we're getting ready to free it. So we
3875 * have to detach it first and then unlock it.
3876 *
3877 * We have to drop our reference and NULL it out here because in the
3878 * subpage case detaching does a btrfs_folio_dec_eb_refs() for our eb.
3879 * Below when we call btrfs_release_extent_buffer() we will call
3880 * detach_extent_buffer_folio() on our remaining pages in the !subpage
3881 * case. If we left eb->folios[i] populated in the subpage case we'd
3882 * double put our reference and be super sad.
3883 */
3884 for (int i = 0; i < attached; i++) {
3885 ASSERT(eb->folios[i]);
3886 detach_extent_buffer_folio(eb, eb->folios[i]);
3887 unlock_page(folio_page(eb->folios[i], 0));
3888 folio_put(eb->folios[i]);
3889 eb->folios[i] = NULL;
3890 }
3891 /*
3892 * Now all pages of that extent buffer is unmapped, set UNMAPPED flag,
3893 * so it can be cleaned up without utlizing page->mapping.
3894 */
3895 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3896
3897 btrfs_release_extent_buffer(eb);
3898 if (ret < 0)
3899 return ERR_PTR(ret);
3900 ASSERT(existing_eb);
3901 return existing_eb;
3902}
3903
3904static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3905{
3906 struct extent_buffer *eb =
3907 container_of(head, struct extent_buffer, rcu_head);
3908
3909 __free_extent_buffer(eb);
3910}
3911
3912static int release_extent_buffer(struct extent_buffer *eb)
3913 __releases(&eb->refs_lock)
3914{
3915 lockdep_assert_held(&eb->refs_lock);
3916
3917 WARN_ON(atomic_read(&eb->refs) == 0);
3918 if (atomic_dec_and_test(&eb->refs)) {
3919 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
3920 struct btrfs_fs_info *fs_info = eb->fs_info;
3921
3922 spin_unlock(&eb->refs_lock);
3923
3924 spin_lock(&fs_info->buffer_lock);
3925 radix_tree_delete(&fs_info->buffer_radix,
3926 eb->start >> fs_info->sectorsize_bits);
3927 spin_unlock(&fs_info->buffer_lock);
3928 } else {
3929 spin_unlock(&eb->refs_lock);
3930 }
3931
3932 btrfs_leak_debug_del_eb(eb);
3933 /* Should be safe to release our pages at this point */
3934 btrfs_release_extent_buffer_pages(eb);
3935#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3936 if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
3937 __free_extent_buffer(eb);
3938 return 1;
3939 }
3940#endif
3941 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3942 return 1;
3943 }
3944 spin_unlock(&eb->refs_lock);
3945
3946 return 0;
3947}
3948
3949void free_extent_buffer(struct extent_buffer *eb)
3950{
3951 int refs;
3952 if (!eb)
3953 return;
3954
3955 refs = atomic_read(&eb->refs);
3956 while (1) {
3957 if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
3958 || (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
3959 refs == 1))
3960 break;
3961 if (atomic_try_cmpxchg(&eb->refs, &refs, refs - 1))
3962 return;
3963 }
3964
3965 spin_lock(&eb->refs_lock);
3966 if (atomic_read(&eb->refs) == 2 &&
3967 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
3968 !extent_buffer_under_io(eb) &&
3969 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3970 atomic_dec(&eb->refs);
3971
3972 /*
3973 * I know this is terrible, but it's temporary until we stop tracking
3974 * the uptodate bits and such for the extent buffers.
3975 */
3976 release_extent_buffer(eb);
3977}
3978
3979void free_extent_buffer_stale(struct extent_buffer *eb)
3980{
3981 if (!eb)
3982 return;
3983
3984 spin_lock(&eb->refs_lock);
3985 set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
3986
3987 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
3988 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3989 atomic_dec(&eb->refs);
3990 release_extent_buffer(eb);
3991}
3992
3993static void btree_clear_folio_dirty(struct folio *folio)
3994{
3995 ASSERT(folio_test_dirty(folio));
3996 ASSERT(folio_test_locked(folio));
3997 folio_clear_dirty_for_io(folio);
3998 xa_lock_irq(&folio->mapping->i_pages);
3999 if (!folio_test_dirty(folio))
4000 __xa_clear_mark(&folio->mapping->i_pages,
4001 folio_index(folio), PAGECACHE_TAG_DIRTY);
4002 xa_unlock_irq(&folio->mapping->i_pages);
4003}
4004
4005static void clear_subpage_extent_buffer_dirty(const struct extent_buffer *eb)
4006{
4007 struct btrfs_fs_info *fs_info = eb->fs_info;
4008 struct folio *folio = eb->folios[0];
4009 bool last;
4010
4011 /* btree_clear_folio_dirty() needs page locked. */
4012 folio_lock(folio);
4013 last = btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start, eb->len);
4014 if (last)
4015 btree_clear_folio_dirty(folio);
4016 folio_unlock(folio);
4017 WARN_ON(atomic_read(&eb->refs) == 0);
4018}
4019
4020void btrfs_clear_buffer_dirty(struct btrfs_trans_handle *trans,
4021 struct extent_buffer *eb)
4022{
4023 struct btrfs_fs_info *fs_info = eb->fs_info;
4024 int num_folios;
4025
4026 btrfs_assert_tree_write_locked(eb);
4027
4028 if (trans && btrfs_header_generation(eb) != trans->transid)
4029 return;
4030
4031 /*
4032 * Instead of clearing the dirty flag off of the buffer, mark it as
4033 * EXTENT_BUFFER_ZONED_ZEROOUT. This allows us to preserve
4034 * write-ordering in zoned mode, without the need to later re-dirty
4035 * the extent_buffer.
4036 *
4037 * The actual zeroout of the buffer will happen later in
4038 * btree_csum_one_bio.
4039 */
4040 if (btrfs_is_zoned(fs_info)) {
4041 set_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags);
4042 return;
4043 }
4044
4045 if (!test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags))
4046 return;
4047
4048 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes, -eb->len,
4049 fs_info->dirty_metadata_batch);
4050
4051 if (eb->fs_info->nodesize < PAGE_SIZE)
4052 return clear_subpage_extent_buffer_dirty(eb);
4053
4054 num_folios = num_extent_folios(eb);
4055 for (int i = 0; i < num_folios; i++) {
4056 struct folio *folio = eb->folios[i];
4057
4058 if (!folio_test_dirty(folio))
4059 continue;
4060 folio_lock(folio);
4061 btree_clear_folio_dirty(folio);
4062 folio_unlock(folio);
4063 }
4064 WARN_ON(atomic_read(&eb->refs) == 0);
4065}
4066
4067void set_extent_buffer_dirty(struct extent_buffer *eb)
4068{
4069 int num_folios;
4070 bool was_dirty;
4071
4072 check_buffer_tree_ref(eb);
4073
4074 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4075
4076 num_folios = num_extent_folios(eb);
4077 WARN_ON(atomic_read(&eb->refs) == 0);
4078 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4079
4080 if (!was_dirty) {
4081 bool subpage = eb->fs_info->nodesize < PAGE_SIZE;
4082
4083 /*
4084 * For subpage case, we can have other extent buffers in the
4085 * same page, and in clear_subpage_extent_buffer_dirty() we
4086 * have to clear page dirty without subpage lock held.
4087 * This can cause race where our page gets dirty cleared after
4088 * we just set it.
4089 *
4090 * Thankfully, clear_subpage_extent_buffer_dirty() has locked
4091 * its page for other reasons, we can use page lock to prevent
4092 * the above race.
4093 */
4094 if (subpage)
4095 lock_page(folio_page(eb->folios[0], 0));
4096 for (int i = 0; i < num_folios; i++)
4097 btrfs_folio_set_dirty(eb->fs_info, eb->folios[i],
4098 eb->start, eb->len);
4099 if (subpage)
4100 unlock_page(folio_page(eb->folios[0], 0));
4101 percpu_counter_add_batch(&eb->fs_info->dirty_metadata_bytes,
4102 eb->len,
4103 eb->fs_info->dirty_metadata_batch);
4104 }
4105#ifdef CONFIG_BTRFS_DEBUG
4106 for (int i = 0; i < num_folios; i++)
4107 ASSERT(folio_test_dirty(eb->folios[i]));
4108#endif
4109}
4110
4111void clear_extent_buffer_uptodate(struct extent_buffer *eb)
4112{
4113 struct btrfs_fs_info *fs_info = eb->fs_info;
4114 int num_folios = num_extent_folios(eb);
4115
4116 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4117 for (int i = 0; i < num_folios; i++) {
4118 struct folio *folio = eb->folios[i];
4119
4120 if (!folio)
4121 continue;
4122
4123 /*
4124 * This is special handling for metadata subpage, as regular
4125 * btrfs_is_subpage() can not handle cloned/dummy metadata.
4126 */
4127 if (fs_info->nodesize >= PAGE_SIZE)
4128 folio_clear_uptodate(folio);
4129 else
4130 btrfs_subpage_clear_uptodate(fs_info, folio,
4131 eb->start, eb->len);
4132 }
4133}
4134
4135void set_extent_buffer_uptodate(struct extent_buffer *eb)
4136{
4137 struct btrfs_fs_info *fs_info = eb->fs_info;
4138 int num_folios = num_extent_folios(eb);
4139
4140 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4141 for (int i = 0; i < num_folios; i++) {
4142 struct folio *folio = eb->folios[i];
4143
4144 /*
4145 * This is special handling for metadata subpage, as regular
4146 * btrfs_is_subpage() can not handle cloned/dummy metadata.
4147 */
4148 if (fs_info->nodesize >= PAGE_SIZE)
4149 folio_mark_uptodate(folio);
4150 else
4151 btrfs_subpage_set_uptodate(fs_info, folio,
4152 eb->start, eb->len);
4153 }
4154}
4155
4156static void end_bbio_meta_read(struct btrfs_bio *bbio)
4157{
4158 struct extent_buffer *eb = bbio->private;
4159 struct btrfs_fs_info *fs_info = eb->fs_info;
4160 bool uptodate = !bbio->bio.bi_status;
4161 struct folio_iter fi;
4162 u32 bio_offset = 0;
4163
4164 eb->read_mirror = bbio->mirror_num;
4165
4166 if (uptodate &&
4167 btrfs_validate_extent_buffer(eb, &bbio->parent_check) < 0)
4168 uptodate = false;
4169
4170 if (uptodate) {
4171 set_extent_buffer_uptodate(eb);
4172 } else {
4173 clear_extent_buffer_uptodate(eb);
4174 set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
4175 }
4176
4177 bio_for_each_folio_all(fi, &bbio->bio) {
4178 struct folio *folio = fi.folio;
4179 u64 start = eb->start + bio_offset;
4180 u32 len = fi.length;
4181
4182 if (uptodate)
4183 btrfs_folio_set_uptodate(fs_info, folio, start, len);
4184 else
4185 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
4186
4187 bio_offset += len;
4188 }
4189
4190 clear_bit(EXTENT_BUFFER_READING, &eb->bflags);
4191 smp_mb__after_atomic();
4192 wake_up_bit(&eb->bflags, EXTENT_BUFFER_READING);
4193 free_extent_buffer(eb);
4194
4195 bio_put(&bbio->bio);
4196}
4197
4198int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num,
4199 struct btrfs_tree_parent_check *check)
4200{
4201 struct btrfs_bio *bbio;
4202 bool ret;
4203
4204 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4205 return 0;
4206
4207 /*
4208 * We could have had EXTENT_BUFFER_UPTODATE cleared by the write
4209 * operation, which could potentially still be in flight. In this case
4210 * we simply want to return an error.
4211 */
4212 if (unlikely(test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)))
4213 return -EIO;
4214
4215 /* Someone else is already reading the buffer, just wait for it. */
4216 if (test_and_set_bit(EXTENT_BUFFER_READING, &eb->bflags))
4217 goto done;
4218
4219 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
4220 eb->read_mirror = 0;
4221 check_buffer_tree_ref(eb);
4222 atomic_inc(&eb->refs);
4223
4224 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
4225 REQ_OP_READ | REQ_META, eb->fs_info,
4226 end_bbio_meta_read, eb);
4227 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
4228 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
4229 bbio->file_offset = eb->start;
4230 memcpy(&bbio->parent_check, check, sizeof(*check));
4231 if (eb->fs_info->nodesize < PAGE_SIZE) {
4232 ret = bio_add_folio(&bbio->bio, eb->folios[0], eb->len,
4233 eb->start - folio_pos(eb->folios[0]));
4234 ASSERT(ret);
4235 } else {
4236 int num_folios = num_extent_folios(eb);
4237
4238 for (int i = 0; i < num_folios; i++) {
4239 struct folio *folio = eb->folios[i];
4240
4241 ret = bio_add_folio(&bbio->bio, folio, folio_size(folio), 0);
4242 ASSERT(ret);
4243 }
4244 }
4245 btrfs_submit_bio(bbio, mirror_num);
4246
4247done:
4248 if (wait == WAIT_COMPLETE) {
4249 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_READING, TASK_UNINTERRUPTIBLE);
4250 if (!test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4251 return -EIO;
4252 }
4253
4254 return 0;
4255}
4256
4257static bool report_eb_range(const struct extent_buffer *eb, unsigned long start,
4258 unsigned long len)
4259{
4260 btrfs_warn(eb->fs_info,
4261 "access to eb bytenr %llu len %lu out of range start %lu len %lu",
4262 eb->start, eb->len, start, len);
4263 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4264
4265 return true;
4266}
4267
4268/*
4269 * Check if the [start, start + len) range is valid before reading/writing
4270 * the eb.
4271 * NOTE: @start and @len are offset inside the eb, not logical address.
4272 *
4273 * Caller should not touch the dst/src memory if this function returns error.
4274 */
4275static inline int check_eb_range(const struct extent_buffer *eb,
4276 unsigned long start, unsigned long len)
4277{
4278 unsigned long offset;
4279
4280 /* start, start + len should not go beyond eb->len nor overflow */
4281 if (unlikely(check_add_overflow(start, len, &offset) || offset > eb->len))
4282 return report_eb_range(eb, start, len);
4283
4284 return false;
4285}
4286
4287void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
4288 unsigned long start, unsigned long len)
4289{
4290 const int unit_size = folio_size(eb->folios[0]);
4291 size_t cur;
4292 size_t offset;
4293 char *dst = (char *)dstv;
4294 unsigned long i = get_eb_folio_index(eb, start);
4295
4296 if (check_eb_range(eb, start, len)) {
4297 /*
4298 * Invalid range hit, reset the memory, so callers won't get
4299 * some random garbage for their uninitialized memory.
4300 */
4301 memset(dstv, 0, len);
4302 return;
4303 }
4304
4305 if (eb->addr) {
4306 memcpy(dstv, eb->addr + start, len);
4307 return;
4308 }
4309
4310 offset = get_eb_offset_in_folio(eb, start);
4311
4312 while (len > 0) {
4313 char *kaddr;
4314
4315 cur = min(len, unit_size - offset);
4316 kaddr = folio_address(eb->folios[i]);
4317 memcpy(dst, kaddr + offset, cur);
4318
4319 dst += cur;
4320 len -= cur;
4321 offset = 0;
4322 i++;
4323 }
4324}
4325
4326int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
4327 void __user *dstv,
4328 unsigned long start, unsigned long len)
4329{
4330 const int unit_size = folio_size(eb->folios[0]);
4331 size_t cur;
4332 size_t offset;
4333 char __user *dst = (char __user *)dstv;
4334 unsigned long i = get_eb_folio_index(eb, start);
4335 int ret = 0;
4336
4337 WARN_ON(start > eb->len);
4338 WARN_ON(start + len > eb->start + eb->len);
4339
4340 if (eb->addr) {
4341 if (copy_to_user_nofault(dstv, eb->addr + start, len))
4342 ret = -EFAULT;
4343 return ret;
4344 }
4345
4346 offset = get_eb_offset_in_folio(eb, start);
4347
4348 while (len > 0) {
4349 char *kaddr;
4350
4351 cur = min(len, unit_size - offset);
4352 kaddr = folio_address(eb->folios[i]);
4353 if (copy_to_user_nofault(dst, kaddr + offset, cur)) {
4354 ret = -EFAULT;
4355 break;
4356 }
4357
4358 dst += cur;
4359 len -= cur;
4360 offset = 0;
4361 i++;
4362 }
4363
4364 return ret;
4365}
4366
4367int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
4368 unsigned long start, unsigned long len)
4369{
4370 const int unit_size = folio_size(eb->folios[0]);
4371 size_t cur;
4372 size_t offset;
4373 char *kaddr;
4374 char *ptr = (char *)ptrv;
4375 unsigned long i = get_eb_folio_index(eb, start);
4376 int ret = 0;
4377
4378 if (check_eb_range(eb, start, len))
4379 return -EINVAL;
4380
4381 if (eb->addr)
4382 return memcmp(ptrv, eb->addr + start, len);
4383
4384 offset = get_eb_offset_in_folio(eb, start);
4385
4386 while (len > 0) {
4387 cur = min(len, unit_size - offset);
4388 kaddr = folio_address(eb->folios[i]);
4389 ret = memcmp(ptr, kaddr + offset, cur);
4390 if (ret)
4391 break;
4392
4393 ptr += cur;
4394 len -= cur;
4395 offset = 0;
4396 i++;
4397 }
4398 return ret;
4399}
4400
4401/*
4402 * Check that the extent buffer is uptodate.
4403 *
4404 * For regular sector size == PAGE_SIZE case, check if @page is uptodate.
4405 * For subpage case, check if the range covered by the eb has EXTENT_UPTODATE.
4406 */
4407static void assert_eb_folio_uptodate(const struct extent_buffer *eb, int i)
4408{
4409 struct btrfs_fs_info *fs_info = eb->fs_info;
4410 struct folio *folio = eb->folios[i];
4411
4412 ASSERT(folio);
4413
4414 /*
4415 * If we are using the commit root we could potentially clear a page
4416 * Uptodate while we're using the extent buffer that we've previously
4417 * looked up. We don't want to complain in this case, as the page was
4418 * valid before, we just didn't write it out. Instead we want to catch
4419 * the case where we didn't actually read the block properly, which
4420 * would have !PageUptodate and !EXTENT_BUFFER_WRITE_ERR.
4421 */
4422 if (test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
4423 return;
4424
4425 if (fs_info->nodesize < PAGE_SIZE) {
4426 struct folio *folio = eb->folios[0];
4427
4428 ASSERT(i == 0);
4429 if (WARN_ON(!btrfs_subpage_test_uptodate(fs_info, folio,
4430 eb->start, eb->len)))
4431 btrfs_subpage_dump_bitmap(fs_info, folio, eb->start, eb->len);
4432 } else {
4433 WARN_ON(!folio_test_uptodate(folio));
4434 }
4435}
4436
4437static void __write_extent_buffer(const struct extent_buffer *eb,
4438 const void *srcv, unsigned long start,
4439 unsigned long len, bool use_memmove)
4440{
4441 const int unit_size = folio_size(eb->folios[0]);
4442 size_t cur;
4443 size_t offset;
4444 char *kaddr;
4445 char *src = (char *)srcv;
4446 unsigned long i = get_eb_folio_index(eb, start);
4447 /* For unmapped (dummy) ebs, no need to check their uptodate status. */
4448 const bool check_uptodate = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
4449
4450 if (check_eb_range(eb, start, len))
4451 return;
4452
4453 if (eb->addr) {
4454 if (use_memmove)
4455 memmove(eb->addr + start, srcv, len);
4456 else
4457 memcpy(eb->addr + start, srcv, len);
4458 return;
4459 }
4460
4461 offset = get_eb_offset_in_folio(eb, start);
4462
4463 while (len > 0) {
4464 if (check_uptodate)
4465 assert_eb_folio_uptodate(eb, i);
4466
4467 cur = min(len, unit_size - offset);
4468 kaddr = folio_address(eb->folios[i]);
4469 if (use_memmove)
4470 memmove(kaddr + offset, src, cur);
4471 else
4472 memcpy(kaddr + offset, src, cur);
4473
4474 src += cur;
4475 len -= cur;
4476 offset = 0;
4477 i++;
4478 }
4479}
4480
4481void write_extent_buffer(const struct extent_buffer *eb, const void *srcv,
4482 unsigned long start, unsigned long len)
4483{
4484 return __write_extent_buffer(eb, srcv, start, len, false);
4485}
4486
4487static void memset_extent_buffer(const struct extent_buffer *eb, int c,
4488 unsigned long start, unsigned long len)
4489{
4490 const int unit_size = folio_size(eb->folios[0]);
4491 unsigned long cur = start;
4492
4493 if (eb->addr) {
4494 memset(eb->addr + start, c, len);
4495 return;
4496 }
4497
4498 while (cur < start + len) {
4499 unsigned long index = get_eb_folio_index(eb, cur);
4500 unsigned int offset = get_eb_offset_in_folio(eb, cur);
4501 unsigned int cur_len = min(start + len - cur, unit_size - offset);
4502
4503 assert_eb_folio_uptodate(eb, index);
4504 memset(folio_address(eb->folios[index]) + offset, c, cur_len);
4505
4506 cur += cur_len;
4507 }
4508}
4509
4510void memzero_extent_buffer(const struct extent_buffer *eb, unsigned long start,
4511 unsigned long len)
4512{
4513 if (check_eb_range(eb, start, len))
4514 return;
4515 return memset_extent_buffer(eb, 0, start, len);
4516}
4517
4518void copy_extent_buffer_full(const struct extent_buffer *dst,
4519 const struct extent_buffer *src)
4520{
4521 const int unit_size = folio_size(src->folios[0]);
4522 unsigned long cur = 0;
4523
4524 ASSERT(dst->len == src->len);
4525
4526 while (cur < src->len) {
4527 unsigned long index = get_eb_folio_index(src, cur);
4528 unsigned long offset = get_eb_offset_in_folio(src, cur);
4529 unsigned long cur_len = min(src->len, unit_size - offset);
4530 void *addr = folio_address(src->folios[index]) + offset;
4531
4532 write_extent_buffer(dst, addr, cur, cur_len);
4533
4534 cur += cur_len;
4535 }
4536}
4537
4538void copy_extent_buffer(const struct extent_buffer *dst,
4539 const struct extent_buffer *src,
4540 unsigned long dst_offset, unsigned long src_offset,
4541 unsigned long len)
4542{
4543 const int unit_size = folio_size(dst->folios[0]);
4544 u64 dst_len = dst->len;
4545 size_t cur;
4546 size_t offset;
4547 char *kaddr;
4548 unsigned long i = get_eb_folio_index(dst, dst_offset);
4549
4550 if (check_eb_range(dst, dst_offset, len) ||
4551 check_eb_range(src, src_offset, len))
4552 return;
4553
4554 WARN_ON(src->len != dst_len);
4555
4556 offset = get_eb_offset_in_folio(dst, dst_offset);
4557
4558 while (len > 0) {
4559 assert_eb_folio_uptodate(dst, i);
4560
4561 cur = min(len, (unsigned long)(unit_size - offset));
4562
4563 kaddr = folio_address(dst->folios[i]);
4564 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4565
4566 src_offset += cur;
4567 len -= cur;
4568 offset = 0;
4569 i++;
4570 }
4571}
4572
4573/*
4574 * Calculate the folio and offset of the byte containing the given bit number.
4575 *
4576 * @eb: the extent buffer
4577 * @start: offset of the bitmap item in the extent buffer
4578 * @nr: bit number
4579 * @folio_index: return index of the folio in the extent buffer that contains
4580 * the given bit number
4581 * @folio_offset: return offset into the folio given by folio_index
4582 *
4583 * This helper hides the ugliness of finding the byte in an extent buffer which
4584 * contains a given bit.
4585 */
4586static inline void eb_bitmap_offset(const struct extent_buffer *eb,
4587 unsigned long start, unsigned long nr,
4588 unsigned long *folio_index,
4589 size_t *folio_offset)
4590{
4591 size_t byte_offset = BIT_BYTE(nr);
4592 size_t offset;
4593
4594 /*
4595 * The byte we want is the offset of the extent buffer + the offset of
4596 * the bitmap item in the extent buffer + the offset of the byte in the
4597 * bitmap item.
4598 */
4599 offset = start + offset_in_folio(eb->folios[0], eb->start) + byte_offset;
4600
4601 *folio_index = offset >> folio_shift(eb->folios[0]);
4602 *folio_offset = offset_in_folio(eb->folios[0], offset);
4603}
4604
4605/*
4606 * Determine whether a bit in a bitmap item is set.
4607 *
4608 * @eb: the extent buffer
4609 * @start: offset of the bitmap item in the extent buffer
4610 * @nr: bit number to test
4611 */
4612int extent_buffer_test_bit(const struct extent_buffer *eb, unsigned long start,
4613 unsigned long nr)
4614{
4615 unsigned long i;
4616 size_t offset;
4617 u8 *kaddr;
4618
4619 eb_bitmap_offset(eb, start, nr, &i, &offset);
4620 assert_eb_folio_uptodate(eb, i);
4621 kaddr = folio_address(eb->folios[i]);
4622 return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
4623}
4624
4625static u8 *extent_buffer_get_byte(const struct extent_buffer *eb, unsigned long bytenr)
4626{
4627 unsigned long index = get_eb_folio_index(eb, bytenr);
4628
4629 if (check_eb_range(eb, bytenr, 1))
4630 return NULL;
4631 return folio_address(eb->folios[index]) + get_eb_offset_in_folio(eb, bytenr);
4632}
4633
4634/*
4635 * Set an area of a bitmap to 1.
4636 *
4637 * @eb: the extent buffer
4638 * @start: offset of the bitmap item in the extent buffer
4639 * @pos: bit number of the first bit
4640 * @len: number of bits to set
4641 */
4642void extent_buffer_bitmap_set(const struct extent_buffer *eb, unsigned long start,
4643 unsigned long pos, unsigned long len)
4644{
4645 unsigned int first_byte = start + BIT_BYTE(pos);
4646 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4647 const bool same_byte = (first_byte == last_byte);
4648 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4649 u8 *kaddr;
4650
4651 if (same_byte)
4652 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4653
4654 /* Handle the first byte. */
4655 kaddr = extent_buffer_get_byte(eb, first_byte);
4656 *kaddr |= mask;
4657 if (same_byte)
4658 return;
4659
4660 /* Handle the byte aligned part. */
4661 ASSERT(first_byte + 1 <= last_byte);
4662 memset_extent_buffer(eb, 0xff, first_byte + 1, last_byte - first_byte - 1);
4663
4664 /* Handle the last byte. */
4665 kaddr = extent_buffer_get_byte(eb, last_byte);
4666 *kaddr |= BITMAP_LAST_BYTE_MASK(pos + len);
4667}
4668
4669
4670/*
4671 * Clear an area of a bitmap.
4672 *
4673 * @eb: the extent buffer
4674 * @start: offset of the bitmap item in the extent buffer
4675 * @pos: bit number of the first bit
4676 * @len: number of bits to clear
4677 */
4678void extent_buffer_bitmap_clear(const struct extent_buffer *eb,
4679 unsigned long start, unsigned long pos,
4680 unsigned long len)
4681{
4682 unsigned int first_byte = start + BIT_BYTE(pos);
4683 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4684 const bool same_byte = (first_byte == last_byte);
4685 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4686 u8 *kaddr;
4687
4688 if (same_byte)
4689 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4690
4691 /* Handle the first byte. */
4692 kaddr = extent_buffer_get_byte(eb, first_byte);
4693 *kaddr &= ~mask;
4694 if (same_byte)
4695 return;
4696
4697 /* Handle the byte aligned part. */
4698 ASSERT(first_byte + 1 <= last_byte);
4699 memset_extent_buffer(eb, 0, first_byte + 1, last_byte - first_byte - 1);
4700
4701 /* Handle the last byte. */
4702 kaddr = extent_buffer_get_byte(eb, last_byte);
4703 *kaddr &= ~BITMAP_LAST_BYTE_MASK(pos + len);
4704}
4705
4706static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4707{
4708 unsigned long distance = (src > dst) ? src - dst : dst - src;
4709 return distance < len;
4710}
4711
4712void memcpy_extent_buffer(const struct extent_buffer *dst,
4713 unsigned long dst_offset, unsigned long src_offset,
4714 unsigned long len)
4715{
4716 const int unit_size = folio_size(dst->folios[0]);
4717 unsigned long cur_off = 0;
4718
4719 if (check_eb_range(dst, dst_offset, len) ||
4720 check_eb_range(dst, src_offset, len))
4721 return;
4722
4723 if (dst->addr) {
4724 const bool use_memmove = areas_overlap(src_offset, dst_offset, len);
4725
4726 if (use_memmove)
4727 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4728 else
4729 memcpy(dst->addr + dst_offset, dst->addr + src_offset, len);
4730 return;
4731 }
4732
4733 while (cur_off < len) {
4734 unsigned long cur_src = cur_off + src_offset;
4735 unsigned long folio_index = get_eb_folio_index(dst, cur_src);
4736 unsigned long folio_off = get_eb_offset_in_folio(dst, cur_src);
4737 unsigned long cur_len = min(src_offset + len - cur_src,
4738 unit_size - folio_off);
4739 void *src_addr = folio_address(dst->folios[folio_index]) + folio_off;
4740 const bool use_memmove = areas_overlap(src_offset + cur_off,
4741 dst_offset + cur_off, cur_len);
4742
4743 __write_extent_buffer(dst, src_addr, dst_offset + cur_off, cur_len,
4744 use_memmove);
4745 cur_off += cur_len;
4746 }
4747}
4748
4749void memmove_extent_buffer(const struct extent_buffer *dst,
4750 unsigned long dst_offset, unsigned long src_offset,
4751 unsigned long len)
4752{
4753 unsigned long dst_end = dst_offset + len - 1;
4754 unsigned long src_end = src_offset + len - 1;
4755
4756 if (check_eb_range(dst, dst_offset, len) ||
4757 check_eb_range(dst, src_offset, len))
4758 return;
4759
4760 if (dst_offset < src_offset) {
4761 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4762 return;
4763 }
4764
4765 if (dst->addr) {
4766 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4767 return;
4768 }
4769
4770 while (len > 0) {
4771 unsigned long src_i;
4772 size_t cur;
4773 size_t dst_off_in_folio;
4774 size_t src_off_in_folio;
4775 void *src_addr;
4776 bool use_memmove;
4777
4778 src_i = get_eb_folio_index(dst, src_end);
4779
4780 dst_off_in_folio = get_eb_offset_in_folio(dst, dst_end);
4781 src_off_in_folio = get_eb_offset_in_folio(dst, src_end);
4782
4783 cur = min_t(unsigned long, len, src_off_in_folio + 1);
4784 cur = min(cur, dst_off_in_folio + 1);
4785
4786 src_addr = folio_address(dst->folios[src_i]) + src_off_in_folio -
4787 cur + 1;
4788 use_memmove = areas_overlap(src_end - cur + 1, dst_end - cur + 1,
4789 cur);
4790
4791 __write_extent_buffer(dst, src_addr, dst_end - cur + 1, cur,
4792 use_memmove);
4793
4794 dst_end -= cur;
4795 src_end -= cur;
4796 len -= cur;
4797 }
4798}
4799
4800#define GANG_LOOKUP_SIZE 16
4801static struct extent_buffer *get_next_extent_buffer(
4802 struct btrfs_fs_info *fs_info, struct page *page, u64 bytenr)
4803{
4804 struct extent_buffer *gang[GANG_LOOKUP_SIZE];
4805 struct extent_buffer *found = NULL;
4806 u64 page_start = page_offset(page);
4807 u64 cur = page_start;
4808
4809 ASSERT(in_range(bytenr, page_start, PAGE_SIZE));
4810 lockdep_assert_held(&fs_info->buffer_lock);
4811
4812 while (cur < page_start + PAGE_SIZE) {
4813 int ret;
4814 int i;
4815
4816 ret = radix_tree_gang_lookup(&fs_info->buffer_radix,
4817 (void **)gang, cur >> fs_info->sectorsize_bits,
4818 min_t(unsigned int, GANG_LOOKUP_SIZE,
4819 PAGE_SIZE / fs_info->nodesize));
4820 if (ret == 0)
4821 goto out;
4822 for (i = 0; i < ret; i++) {
4823 /* Already beyond page end */
4824 if (gang[i]->start >= page_start + PAGE_SIZE)
4825 goto out;
4826 /* Found one */
4827 if (gang[i]->start >= bytenr) {
4828 found = gang[i];
4829 goto out;
4830 }
4831 }
4832 cur = gang[ret - 1]->start + gang[ret - 1]->len;
4833 }
4834out:
4835 return found;
4836}
4837
4838static int try_release_subpage_extent_buffer(struct page *page)
4839{
4840 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
4841 u64 cur = page_offset(page);
4842 const u64 end = page_offset(page) + PAGE_SIZE;
4843 int ret;
4844
4845 while (cur < end) {
4846 struct extent_buffer *eb = NULL;
4847
4848 /*
4849 * Unlike try_release_extent_buffer() which uses folio private
4850 * to grab buffer, for subpage case we rely on radix tree, thus
4851 * we need to ensure radix tree consistency.
4852 *
4853 * We also want an atomic snapshot of the radix tree, thus go
4854 * with spinlock rather than RCU.
4855 */
4856 spin_lock(&fs_info->buffer_lock);
4857 eb = get_next_extent_buffer(fs_info, page, cur);
4858 if (!eb) {
4859 /* No more eb in the page range after or at cur */
4860 spin_unlock(&fs_info->buffer_lock);
4861 break;
4862 }
4863 cur = eb->start + eb->len;
4864
4865 /*
4866 * The same as try_release_extent_buffer(), to ensure the eb
4867 * won't disappear out from under us.
4868 */
4869 spin_lock(&eb->refs_lock);
4870 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4871 spin_unlock(&eb->refs_lock);
4872 spin_unlock(&fs_info->buffer_lock);
4873 break;
4874 }
4875 spin_unlock(&fs_info->buffer_lock);
4876
4877 /*
4878 * If tree ref isn't set then we know the ref on this eb is a
4879 * real ref, so just return, this eb will likely be freed soon
4880 * anyway.
4881 */
4882 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4883 spin_unlock(&eb->refs_lock);
4884 break;
4885 }
4886
4887 /*
4888 * Here we don't care about the return value, we will always
4889 * check the folio private at the end. And
4890 * release_extent_buffer() will release the refs_lock.
4891 */
4892 release_extent_buffer(eb);
4893 }
4894 /*
4895 * Finally to check if we have cleared folio private, as if we have
4896 * released all ebs in the page, the folio private should be cleared now.
4897 */
4898 spin_lock(&page->mapping->i_private_lock);
4899 if (!folio_test_private(page_folio(page)))
4900 ret = 1;
4901 else
4902 ret = 0;
4903 spin_unlock(&page->mapping->i_private_lock);
4904 return ret;
4905
4906}
4907
4908int try_release_extent_buffer(struct page *page)
4909{
4910 struct folio *folio = page_folio(page);
4911 struct extent_buffer *eb;
4912
4913 if (btrfs_sb(page->mapping->host->i_sb)->nodesize < PAGE_SIZE)
4914 return try_release_subpage_extent_buffer(page);
4915
4916 /*
4917 * We need to make sure nobody is changing folio private, as we rely on
4918 * folio private as the pointer to extent buffer.
4919 */
4920 spin_lock(&page->mapping->i_private_lock);
4921 if (!folio_test_private(folio)) {
4922 spin_unlock(&page->mapping->i_private_lock);
4923 return 1;
4924 }
4925
4926 eb = folio_get_private(folio);
4927 BUG_ON(!eb);
4928
4929 /*
4930 * This is a little awful but should be ok, we need to make sure that
4931 * the eb doesn't disappear out from under us while we're looking at
4932 * this page.
4933 */
4934 spin_lock(&eb->refs_lock);
4935 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4936 spin_unlock(&eb->refs_lock);
4937 spin_unlock(&page->mapping->i_private_lock);
4938 return 0;
4939 }
4940 spin_unlock(&page->mapping->i_private_lock);
4941
4942 /*
4943 * If tree ref isn't set then we know the ref on this eb is a real ref,
4944 * so just return, this page will likely be freed soon anyway.
4945 */
4946 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4947 spin_unlock(&eb->refs_lock);
4948 return 0;
4949 }
4950
4951 return release_extent_buffer(eb);
4952}
4953
4954/*
4955 * Attempt to readahead a child block.
4956 *
4957 * @fs_info: the fs_info
4958 * @bytenr: bytenr to read
4959 * @owner_root: objectid of the root that owns this eb
4960 * @gen: generation for the uptodate check, can be 0
4961 * @level: level for the eb
4962 *
4963 * Attempt to readahead a tree block at @bytenr. If @gen is 0 then we do a
4964 * normal uptodate check of the eb, without checking the generation. If we have
4965 * to read the block we will not block on anything.
4966 */
4967void btrfs_readahead_tree_block(struct btrfs_fs_info *fs_info,
4968 u64 bytenr, u64 owner_root, u64 gen, int level)
4969{
4970 struct btrfs_tree_parent_check check = {
4971 .has_first_key = 0,
4972 .level = level,
4973 .transid = gen
4974 };
4975 struct extent_buffer *eb;
4976 int ret;
4977
4978 eb = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
4979 if (IS_ERR(eb))
4980 return;
4981
4982 if (btrfs_buffer_uptodate(eb, gen, 1)) {
4983 free_extent_buffer(eb);
4984 return;
4985 }
4986
4987 ret = read_extent_buffer_pages(eb, WAIT_NONE, 0, &check);
4988 if (ret < 0)
4989 free_extent_buffer_stale(eb);
4990 else
4991 free_extent_buffer(eb);
4992}
4993
4994/*
4995 * Readahead a node's child block.
4996 *
4997 * @node: parent node we're reading from
4998 * @slot: slot in the parent node for the child we want to read
4999 *
5000 * A helper for btrfs_readahead_tree_block, we simply read the bytenr pointed at
5001 * the slot in the node provided.
5002 */
5003void btrfs_readahead_node_child(struct extent_buffer *node, int slot)
5004{
5005 btrfs_readahead_tree_block(node->fs_info,
5006 btrfs_node_blockptr(node, slot),
5007 btrfs_header_owner(node),
5008 btrfs_node_ptr_generation(node, slot),
5009 btrfs_header_level(node) - 1);
5010}