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/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 "check-integrity.h"
25#include "locking.h"
26#include "rcu-string.h"
27#include "backref.h"
28#include "disk-io.h"
29#include "subpage.h"
30#include "zoned.h"
31#include "block-group.h"
32#include "compression.h"
33#include "fs.h"
34#include "accessors.h"
35#include "file-item.h"
36#include "file.h"
37#include "dev-replace.h"
38#include "super.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 bio *bio;
100 int mirror_num;
101 enum btrfs_compression_type compress_type;
102 u32 len_to_stripe_boundary;
103 u32 len_to_oe_boundary;
104 btrfs_bio_end_io_t end_io_func;
105
106 /*
107 * This is for metadata read, to provide the extra needed verification
108 * info. This has to be provided for submit_one_bio(), as
109 * submit_one_bio() can submit a bio if it ends at stripe boundary. If
110 * no such parent_check is provided, the metadata can hit false alert at
111 * endio time.
112 */
113 struct btrfs_tree_parent_check *parent_check;
114
115 /*
116 * Tell writepage not to lock the state bits for this range, it still
117 * does the unlocking.
118 */
119 bool extent_locked;
120
121 /* Tell the submit_bio code to use REQ_SYNC */
122 bool sync_io;
123};
124
125static void submit_one_bio(struct btrfs_bio_ctrl *bio_ctrl)
126{
127 struct bio *bio;
128 struct bio_vec *bv;
129 struct btrfs_inode *inode;
130 int mirror_num;
131
132 if (!bio_ctrl->bio)
133 return;
134
135 bio = bio_ctrl->bio;
136 bv = bio_first_bvec_all(bio);
137 inode = BTRFS_I(bv->bv_page->mapping->host);
138 mirror_num = bio_ctrl->mirror_num;
139
140 /* Caller should ensure the bio has at least some range added */
141 ASSERT(bio->bi_iter.bi_size);
142
143 btrfs_bio(bio)->file_offset = page_offset(bv->bv_page) + bv->bv_offset;
144
145 if (!is_data_inode(&inode->vfs_inode)) {
146 if (btrfs_op(bio) != BTRFS_MAP_WRITE) {
147 /*
148 * For metadata read, we should have the parent_check,
149 * and copy it to bbio for metadata verification.
150 */
151 ASSERT(bio_ctrl->parent_check);
152 memcpy(&btrfs_bio(bio)->parent_check,
153 bio_ctrl->parent_check,
154 sizeof(struct btrfs_tree_parent_check));
155 }
156 btrfs_submit_metadata_bio(inode, bio, mirror_num);
157 } else if (btrfs_op(bio) == BTRFS_MAP_WRITE) {
158 btrfs_submit_data_write_bio(inode, bio, mirror_num);
159 } else {
160 btrfs_submit_data_read_bio(inode, bio, mirror_num,
161 bio_ctrl->compress_type);
162 }
163
164 /* The bio is owned by the end_io handler now */
165 bio_ctrl->bio = NULL;
166}
167
168/*
169 * Submit or fail the current bio in the bio_ctrl structure.
170 */
171static void submit_write_bio(struct btrfs_bio_ctrl *bio_ctrl, int ret)
172{
173 struct bio *bio = bio_ctrl->bio;
174
175 if (!bio)
176 return;
177
178 if (ret) {
179 ASSERT(ret < 0);
180 btrfs_bio_end_io(btrfs_bio(bio), errno_to_blk_status(ret));
181 /* The bio is owned by the end_io handler now */
182 bio_ctrl->bio = NULL;
183 } else {
184 submit_one_bio(bio_ctrl);
185 }
186}
187
188int __init extent_buffer_init_cachep(void)
189{
190 extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
191 sizeof(struct extent_buffer), 0,
192 SLAB_MEM_SPREAD, NULL);
193 if (!extent_buffer_cache)
194 return -ENOMEM;
195
196 return 0;
197}
198
199void __cold extent_buffer_free_cachep(void)
200{
201 /*
202 * Make sure all delayed rcu free are flushed before we
203 * destroy caches.
204 */
205 rcu_barrier();
206 kmem_cache_destroy(extent_buffer_cache);
207}
208
209void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
210{
211 unsigned long index = start >> PAGE_SHIFT;
212 unsigned long end_index = end >> PAGE_SHIFT;
213 struct page *page;
214
215 while (index <= end_index) {
216 page = find_get_page(inode->i_mapping, index);
217 BUG_ON(!page); /* Pages should be in the extent_io_tree */
218 clear_page_dirty_for_io(page);
219 put_page(page);
220 index++;
221 }
222}
223
224void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end)
225{
226 struct address_space *mapping = inode->i_mapping;
227 unsigned long index = start >> PAGE_SHIFT;
228 unsigned long end_index = end >> PAGE_SHIFT;
229 struct folio *folio;
230
231 while (index <= end_index) {
232 folio = filemap_get_folio(mapping, index);
233 filemap_dirty_folio(mapping, folio);
234 folio_account_redirty(folio);
235 index += folio_nr_pages(folio);
236 folio_put(folio);
237 }
238}
239
240/*
241 * Process one page for __process_pages_contig().
242 *
243 * Return >0 if we hit @page == @locked_page.
244 * Return 0 if we updated the page status.
245 * Return -EGAIN if the we need to try again.
246 * (For PAGE_LOCK case but got dirty page or page not belong to mapping)
247 */
248static int process_one_page(struct btrfs_fs_info *fs_info,
249 struct address_space *mapping,
250 struct page *page, struct page *locked_page,
251 unsigned long page_ops, u64 start, u64 end)
252{
253 u32 len;
254
255 ASSERT(end + 1 - start != 0 && end + 1 - start < U32_MAX);
256 len = end + 1 - start;
257
258 if (page_ops & PAGE_SET_ORDERED)
259 btrfs_page_clamp_set_ordered(fs_info, page, start, len);
260 if (page_ops & PAGE_SET_ERROR)
261 btrfs_page_clamp_set_error(fs_info, page, start, len);
262 if (page_ops & PAGE_START_WRITEBACK) {
263 btrfs_page_clamp_clear_dirty(fs_info, page, start, len);
264 btrfs_page_clamp_set_writeback(fs_info, page, start, len);
265 }
266 if (page_ops & PAGE_END_WRITEBACK)
267 btrfs_page_clamp_clear_writeback(fs_info, page, start, len);
268
269 if (page == locked_page)
270 return 1;
271
272 if (page_ops & PAGE_LOCK) {
273 int ret;
274
275 ret = btrfs_page_start_writer_lock(fs_info, page, start, len);
276 if (ret)
277 return ret;
278 if (!PageDirty(page) || page->mapping != mapping) {
279 btrfs_page_end_writer_lock(fs_info, page, start, len);
280 return -EAGAIN;
281 }
282 }
283 if (page_ops & PAGE_UNLOCK)
284 btrfs_page_end_writer_lock(fs_info, page, start, len);
285 return 0;
286}
287
288static int __process_pages_contig(struct address_space *mapping,
289 struct page *locked_page,
290 u64 start, u64 end, unsigned long page_ops,
291 u64 *processed_end)
292{
293 struct btrfs_fs_info *fs_info = btrfs_sb(mapping->host->i_sb);
294 pgoff_t start_index = start >> PAGE_SHIFT;
295 pgoff_t end_index = end >> PAGE_SHIFT;
296 pgoff_t index = start_index;
297 unsigned long pages_processed = 0;
298 struct folio_batch fbatch;
299 int err = 0;
300 int i;
301
302 if (page_ops & PAGE_LOCK) {
303 ASSERT(page_ops == PAGE_LOCK);
304 ASSERT(processed_end && *processed_end == start);
305 }
306
307 if ((page_ops & PAGE_SET_ERROR) && start_index <= end_index)
308 mapping_set_error(mapping, -EIO);
309
310 folio_batch_init(&fbatch);
311 while (index <= end_index) {
312 int found_folios;
313
314 found_folios = filemap_get_folios_contig(mapping, &index,
315 end_index, &fbatch);
316
317 if (found_folios == 0) {
318 /*
319 * Only if we're going to lock these pages, we can find
320 * nothing at @index.
321 */
322 ASSERT(page_ops & PAGE_LOCK);
323 err = -EAGAIN;
324 goto out;
325 }
326
327 for (i = 0; i < found_folios; i++) {
328 int process_ret;
329 struct folio *folio = fbatch.folios[i];
330 process_ret = process_one_page(fs_info, mapping,
331 &folio->page, locked_page, page_ops,
332 start, end);
333 if (process_ret < 0) {
334 err = -EAGAIN;
335 folio_batch_release(&fbatch);
336 goto out;
337 }
338 pages_processed += folio_nr_pages(folio);
339 }
340 folio_batch_release(&fbatch);
341 cond_resched();
342 }
343out:
344 if (err && processed_end) {
345 /*
346 * Update @processed_end. I know this is awful since it has
347 * two different return value patterns (inclusive vs exclusive).
348 *
349 * But the exclusive pattern is necessary if @start is 0, or we
350 * underflow and check against processed_end won't work as
351 * expected.
352 */
353 if (pages_processed)
354 *processed_end = min(end,
355 ((u64)(start_index + pages_processed) << PAGE_SHIFT) - 1);
356 else
357 *processed_end = start;
358 }
359 return err;
360}
361
362static noinline void __unlock_for_delalloc(struct inode *inode,
363 struct page *locked_page,
364 u64 start, u64 end)
365{
366 unsigned long index = start >> PAGE_SHIFT;
367 unsigned long end_index = end >> PAGE_SHIFT;
368
369 ASSERT(locked_page);
370 if (index == locked_page->index && end_index == index)
371 return;
372
373 __process_pages_contig(inode->i_mapping, locked_page, start, end,
374 PAGE_UNLOCK, NULL);
375}
376
377static noinline int lock_delalloc_pages(struct inode *inode,
378 struct page *locked_page,
379 u64 delalloc_start,
380 u64 delalloc_end)
381{
382 unsigned long index = delalloc_start >> PAGE_SHIFT;
383 unsigned long end_index = delalloc_end >> PAGE_SHIFT;
384 u64 processed_end = delalloc_start;
385 int ret;
386
387 ASSERT(locked_page);
388 if (index == locked_page->index && index == end_index)
389 return 0;
390
391 ret = __process_pages_contig(inode->i_mapping, locked_page, delalloc_start,
392 delalloc_end, PAGE_LOCK, &processed_end);
393 if (ret == -EAGAIN && processed_end > delalloc_start)
394 __unlock_for_delalloc(inode, locked_page, delalloc_start,
395 processed_end);
396 return ret;
397}
398
399/*
400 * Find and lock a contiguous range of bytes in the file marked as delalloc, no
401 * more than @max_bytes.
402 *
403 * @start: The original start bytenr to search.
404 * Will store the extent range start bytenr.
405 * @end: The original end bytenr of the search range
406 * Will store the extent range end bytenr.
407 *
408 * Return true if we find a delalloc range which starts inside the original
409 * range, and @start/@end will store the delalloc range start/end.
410 *
411 * Return false if we can't find any delalloc range which starts inside the
412 * original range, and @start/@end will be the non-delalloc range start/end.
413 */
414EXPORT_FOR_TESTS
415noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
416 struct page *locked_page, u64 *start,
417 u64 *end)
418{
419 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
420 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
421 const u64 orig_start = *start;
422 const u64 orig_end = *end;
423 /* The sanity tests may not set a valid fs_info. */
424 u64 max_bytes = fs_info ? fs_info->max_extent_size : BTRFS_MAX_EXTENT_SIZE;
425 u64 delalloc_start;
426 u64 delalloc_end;
427 bool found;
428 struct extent_state *cached_state = NULL;
429 int ret;
430 int loops = 0;
431
432 /* Caller should pass a valid @end to indicate the search range end */
433 ASSERT(orig_end > orig_start);
434
435 /* The range should at least cover part of the page */
436 ASSERT(!(orig_start >= page_offset(locked_page) + PAGE_SIZE ||
437 orig_end <= page_offset(locked_page)));
438again:
439 /* step one, find a bunch of delalloc bytes starting at start */
440 delalloc_start = *start;
441 delalloc_end = 0;
442 found = btrfs_find_delalloc_range(tree, &delalloc_start, &delalloc_end,
443 max_bytes, &cached_state);
444 if (!found || delalloc_end <= *start || delalloc_start > orig_end) {
445 *start = delalloc_start;
446
447 /* @delalloc_end can be -1, never go beyond @orig_end */
448 *end = min(delalloc_end, orig_end);
449 free_extent_state(cached_state);
450 return false;
451 }
452
453 /*
454 * start comes from the offset of locked_page. We have to lock
455 * pages in order, so we can't process delalloc bytes before
456 * locked_page
457 */
458 if (delalloc_start < *start)
459 delalloc_start = *start;
460
461 /*
462 * make sure to limit the number of pages we try to lock down
463 */
464 if (delalloc_end + 1 - delalloc_start > max_bytes)
465 delalloc_end = delalloc_start + max_bytes - 1;
466
467 /* step two, lock all the pages after the page that has start */
468 ret = lock_delalloc_pages(inode, locked_page,
469 delalloc_start, delalloc_end);
470 ASSERT(!ret || ret == -EAGAIN);
471 if (ret == -EAGAIN) {
472 /* some of the pages are gone, lets avoid looping by
473 * shortening the size of the delalloc range we're searching
474 */
475 free_extent_state(cached_state);
476 cached_state = NULL;
477 if (!loops) {
478 max_bytes = PAGE_SIZE;
479 loops = 1;
480 goto again;
481 } else {
482 found = false;
483 goto out_failed;
484 }
485 }
486
487 /* step three, lock the state bits for the whole range */
488 lock_extent(tree, delalloc_start, delalloc_end, &cached_state);
489
490 /* then test to make sure it is all still delalloc */
491 ret = test_range_bit(tree, delalloc_start, delalloc_end,
492 EXTENT_DELALLOC, 1, cached_state);
493 if (!ret) {
494 unlock_extent(tree, delalloc_start, delalloc_end,
495 &cached_state);
496 __unlock_for_delalloc(inode, locked_page,
497 delalloc_start, delalloc_end);
498 cond_resched();
499 goto again;
500 }
501 free_extent_state(cached_state);
502 *start = delalloc_start;
503 *end = delalloc_end;
504out_failed:
505 return found;
506}
507
508void extent_clear_unlock_delalloc(struct btrfs_inode *inode, u64 start, u64 end,
509 struct page *locked_page,
510 u32 clear_bits, unsigned long page_ops)
511{
512 clear_extent_bit(&inode->io_tree, start, end, clear_bits, NULL);
513
514 __process_pages_contig(inode->vfs_inode.i_mapping, locked_page,
515 start, end, page_ops, NULL);
516}
517
518static int insert_failrec(struct btrfs_inode *inode,
519 struct io_failure_record *failrec)
520{
521 struct rb_node *exist;
522
523 spin_lock(&inode->io_failure_lock);
524 exist = rb_simple_insert(&inode->io_failure_tree, failrec->bytenr,
525 &failrec->rb_node);
526 spin_unlock(&inode->io_failure_lock);
527
528 return (exist == NULL) ? 0 : -EEXIST;
529}
530
531static struct io_failure_record *get_failrec(struct btrfs_inode *inode, u64 start)
532{
533 struct rb_node *node;
534 struct io_failure_record *failrec = ERR_PTR(-ENOENT);
535
536 spin_lock(&inode->io_failure_lock);
537 node = rb_simple_search(&inode->io_failure_tree, start);
538 if (node)
539 failrec = rb_entry(node, struct io_failure_record, rb_node);
540 spin_unlock(&inode->io_failure_lock);
541 return failrec;
542}
543
544static void free_io_failure(struct btrfs_inode *inode,
545 struct io_failure_record *rec)
546{
547 spin_lock(&inode->io_failure_lock);
548 rb_erase(&rec->rb_node, &inode->io_failure_tree);
549 spin_unlock(&inode->io_failure_lock);
550
551 kfree(rec);
552}
553
554static int next_mirror(const struct io_failure_record *failrec, int cur_mirror)
555{
556 if (cur_mirror == failrec->num_copies)
557 return cur_mirror + 1 - failrec->num_copies;
558 return cur_mirror + 1;
559}
560
561static int prev_mirror(const struct io_failure_record *failrec, int cur_mirror)
562{
563 if (cur_mirror == 1)
564 return failrec->num_copies;
565 return cur_mirror - 1;
566}
567
568/*
569 * each time an IO finishes, we do a fast check in the IO failure tree
570 * to see if we need to process or clean up an io_failure_record
571 */
572int btrfs_clean_io_failure(struct btrfs_inode *inode, u64 start,
573 struct page *page, unsigned int pg_offset)
574{
575 struct btrfs_fs_info *fs_info = inode->root->fs_info;
576 struct extent_io_tree *io_tree = &inode->io_tree;
577 u64 ino = btrfs_ino(inode);
578 u64 locked_start, locked_end;
579 struct io_failure_record *failrec;
580 int mirror;
581 int ret;
582
583 failrec = get_failrec(inode, start);
584 if (IS_ERR(failrec))
585 return 0;
586
587 BUG_ON(!failrec->this_mirror);
588
589 if (sb_rdonly(fs_info->sb))
590 goto out;
591
592 ret = find_first_extent_bit(io_tree, failrec->bytenr, &locked_start,
593 &locked_end, EXTENT_LOCKED, NULL);
594 if (ret || locked_start > failrec->bytenr ||
595 locked_end < failrec->bytenr + failrec->len - 1)
596 goto out;
597
598 mirror = failrec->this_mirror;
599 do {
600 mirror = prev_mirror(failrec, mirror);
601 btrfs_repair_io_failure(fs_info, ino, start, failrec->len,
602 failrec->logical, page, pg_offset, mirror);
603 } while (mirror != failrec->failed_mirror);
604
605out:
606 free_io_failure(inode, failrec);
607 return 0;
608}
609
610/*
611 * Can be called when
612 * - hold extent lock
613 * - under ordered extent
614 * - the inode is freeing
615 */
616void btrfs_free_io_failure_record(struct btrfs_inode *inode, u64 start, u64 end)
617{
618 struct io_failure_record *failrec;
619 struct rb_node *node, *next;
620
621 if (RB_EMPTY_ROOT(&inode->io_failure_tree))
622 return;
623
624 spin_lock(&inode->io_failure_lock);
625 node = rb_simple_search_first(&inode->io_failure_tree, start);
626 while (node) {
627 failrec = rb_entry(node, struct io_failure_record, rb_node);
628 if (failrec->bytenr > end)
629 break;
630
631 next = rb_next(node);
632 rb_erase(&failrec->rb_node, &inode->io_failure_tree);
633 kfree(failrec);
634
635 node = next;
636 }
637 spin_unlock(&inode->io_failure_lock);
638}
639
640static struct io_failure_record *btrfs_get_io_failure_record(struct inode *inode,
641 struct btrfs_bio *bbio,
642 unsigned int bio_offset)
643{
644 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
645 u64 start = bbio->file_offset + bio_offset;
646 struct io_failure_record *failrec;
647 const u32 sectorsize = fs_info->sectorsize;
648 int ret;
649
650 failrec = get_failrec(BTRFS_I(inode), start);
651 if (!IS_ERR(failrec)) {
652 btrfs_debug(fs_info,
653 "Get IO Failure Record: (found) logical=%llu, start=%llu, len=%llu",
654 failrec->logical, failrec->bytenr, failrec->len);
655 /*
656 * when data can be on disk more than twice, add to failrec here
657 * (e.g. with a list for failed_mirror) to make
658 * clean_io_failure() clean all those errors at once.
659 */
660 ASSERT(failrec->this_mirror == bbio->mirror_num);
661 ASSERT(failrec->len == fs_info->sectorsize);
662 return failrec;
663 }
664
665 failrec = kzalloc(sizeof(*failrec), GFP_NOFS);
666 if (!failrec)
667 return ERR_PTR(-ENOMEM);
668
669 RB_CLEAR_NODE(&failrec->rb_node);
670 failrec->bytenr = start;
671 failrec->len = sectorsize;
672 failrec->failed_mirror = bbio->mirror_num;
673 failrec->this_mirror = bbio->mirror_num;
674 failrec->logical = (bbio->iter.bi_sector << SECTOR_SHIFT) + bio_offset;
675
676 btrfs_debug(fs_info,
677 "new io failure record logical %llu start %llu",
678 failrec->logical, start);
679
680 failrec->num_copies = btrfs_num_copies(fs_info, failrec->logical, sectorsize);
681 if (failrec->num_copies == 1) {
682 /*
683 * We only have a single copy of the data, so don't bother with
684 * all the retry and error correction code that follows. No
685 * matter what the error is, it is very likely to persist.
686 */
687 btrfs_debug(fs_info,
688 "cannot repair logical %llu num_copies %d",
689 failrec->logical, failrec->num_copies);
690 kfree(failrec);
691 return ERR_PTR(-EIO);
692 }
693
694 /* Set the bits in the private failure tree */
695 ret = insert_failrec(BTRFS_I(inode), failrec);
696 if (ret) {
697 kfree(failrec);
698 return ERR_PTR(ret);
699 }
700
701 return failrec;
702}
703
704int btrfs_repair_one_sector(struct btrfs_inode *inode, struct btrfs_bio *failed_bbio,
705 u32 bio_offset, struct page *page, unsigned int pgoff,
706 bool submit_buffered)
707{
708 u64 start = failed_bbio->file_offset + bio_offset;
709 struct io_failure_record *failrec;
710 struct btrfs_fs_info *fs_info = inode->root->fs_info;
711 struct bio *failed_bio = &failed_bbio->bio;
712 const int icsum = bio_offset >> fs_info->sectorsize_bits;
713 struct bio *repair_bio;
714 struct btrfs_bio *repair_bbio;
715
716 btrfs_debug(fs_info,
717 "repair read error: read error at %llu", start);
718
719 BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
720
721 failrec = btrfs_get_io_failure_record(&inode->vfs_inode, failed_bbio, bio_offset);
722 if (IS_ERR(failrec))
723 return PTR_ERR(failrec);
724
725 /*
726 * There are two premises:
727 * a) deliver good data to the caller
728 * b) correct the bad sectors on disk
729 *
730 * Since we're only doing repair for one sector, we only need to get
731 * a good copy of the failed sector and if we succeed, we have setup
732 * everything for btrfs_repair_io_failure to do the rest for us.
733 */
734 failrec->this_mirror = next_mirror(failrec, failrec->this_mirror);
735 if (failrec->this_mirror == failrec->failed_mirror) {
736 btrfs_debug(fs_info,
737 "failed to repair num_copies %d this_mirror %d failed_mirror %d",
738 failrec->num_copies, failrec->this_mirror, failrec->failed_mirror);
739 free_io_failure(inode, failrec);
740 return -EIO;
741 }
742
743 repair_bio = btrfs_bio_alloc(1, REQ_OP_READ, failed_bbio->end_io,
744 failed_bbio->private);
745 repair_bbio = btrfs_bio(repair_bio);
746 repair_bbio->file_offset = start;
747 repair_bio->bi_iter.bi_sector = failrec->logical >> 9;
748
749 if (failed_bbio->csum) {
750 const u32 csum_size = fs_info->csum_size;
751
752 repair_bbio->csum = repair_bbio->csum_inline;
753 memcpy(repair_bbio->csum,
754 failed_bbio->csum + csum_size * icsum, csum_size);
755 }
756
757 bio_add_page(repair_bio, page, failrec->len, pgoff);
758 repair_bbio->iter = repair_bio->bi_iter;
759
760 btrfs_debug(fs_info,
761 "repair read error: submitting new read to mirror %d",
762 failrec->this_mirror);
763
764 /*
765 * At this point we have a bio, so any errors from bio submission will
766 * be handled by the endio on the repair_bio, so we can't return an
767 * error here.
768 */
769 if (submit_buffered)
770 btrfs_submit_data_read_bio(inode, repair_bio,
771 failrec->this_mirror, 0);
772 else
773 btrfs_submit_dio_repair_bio(inode, repair_bio, failrec->this_mirror);
774
775 return BLK_STS_OK;
776}
777
778static void end_page_read(struct page *page, bool uptodate, u64 start, u32 len)
779{
780 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
781
782 ASSERT(page_offset(page) <= start &&
783 start + len <= page_offset(page) + PAGE_SIZE);
784
785 if (uptodate) {
786 if (fsverity_active(page->mapping->host) &&
787 !PageError(page) &&
788 !PageUptodate(page) &&
789 start < i_size_read(page->mapping->host) &&
790 !fsverity_verify_page(page)) {
791 btrfs_page_set_error(fs_info, page, start, len);
792 } else {
793 btrfs_page_set_uptodate(fs_info, page, start, len);
794 }
795 } else {
796 btrfs_page_clear_uptodate(fs_info, page, start, len);
797 btrfs_page_set_error(fs_info, page, start, len);
798 }
799
800 if (!btrfs_is_subpage(fs_info, page))
801 unlock_page(page);
802 else
803 btrfs_subpage_end_reader(fs_info, page, start, len);
804}
805
806static void end_sector_io(struct page *page, u64 offset, bool uptodate)
807{
808 struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
809 const u32 sectorsize = inode->root->fs_info->sectorsize;
810
811 end_page_read(page, uptodate, offset, sectorsize);
812 unlock_extent(&inode->io_tree, offset, offset + sectorsize - 1, NULL);
813}
814
815static void submit_data_read_repair(struct inode *inode,
816 struct btrfs_bio *failed_bbio,
817 u32 bio_offset, const struct bio_vec *bvec,
818 unsigned int error_bitmap)
819{
820 const unsigned int pgoff = bvec->bv_offset;
821 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
822 struct page *page = bvec->bv_page;
823 const u64 start = page_offset(bvec->bv_page) + bvec->bv_offset;
824 const u64 end = start + bvec->bv_len - 1;
825 const u32 sectorsize = fs_info->sectorsize;
826 const int nr_bits = (end + 1 - start) >> fs_info->sectorsize_bits;
827 int i;
828
829 BUG_ON(bio_op(&failed_bbio->bio) == REQ_OP_WRITE);
830
831 /* This repair is only for data */
832 ASSERT(is_data_inode(inode));
833
834 /* We're here because we had some read errors or csum mismatch */
835 ASSERT(error_bitmap);
836
837 /*
838 * We only get called on buffered IO, thus page must be mapped and bio
839 * must not be cloned.
840 */
841 ASSERT(page->mapping && !bio_flagged(&failed_bbio->bio, BIO_CLONED));
842
843 /* Iterate through all the sectors in the range */
844 for (i = 0; i < nr_bits; i++) {
845 const unsigned int offset = i * sectorsize;
846 bool uptodate = false;
847 int ret;
848
849 if (!(error_bitmap & (1U << i))) {
850 /*
851 * This sector has no error, just end the page read
852 * and unlock the range.
853 */
854 uptodate = true;
855 goto next;
856 }
857
858 ret = btrfs_repair_one_sector(BTRFS_I(inode), failed_bbio,
859 bio_offset + offset, page, pgoff + offset,
860 true);
861 if (!ret) {
862 /*
863 * We have submitted the read repair, the page release
864 * will be handled by the endio function of the
865 * submitted repair bio.
866 * Thus we don't need to do any thing here.
867 */
868 continue;
869 }
870 /*
871 * Continue on failed repair, otherwise the remaining sectors
872 * will not be properly unlocked.
873 */
874next:
875 end_sector_io(page, start + offset, uptodate);
876 }
877}
878
879/* lots and lots of room for performance fixes in the end_bio funcs */
880
881void end_extent_writepage(struct page *page, int err, u64 start, u64 end)
882{
883 struct btrfs_inode *inode;
884 const bool uptodate = (err == 0);
885 int ret = 0;
886
887 ASSERT(page && page->mapping);
888 inode = BTRFS_I(page->mapping->host);
889 btrfs_writepage_endio_finish_ordered(inode, page, start, end, uptodate);
890
891 if (!uptodate) {
892 const struct btrfs_fs_info *fs_info = inode->root->fs_info;
893 u32 len;
894
895 ASSERT(end + 1 - start <= U32_MAX);
896 len = end + 1 - start;
897
898 btrfs_page_clear_uptodate(fs_info, page, start, len);
899 btrfs_page_set_error(fs_info, page, start, len);
900 ret = err < 0 ? err : -EIO;
901 mapping_set_error(page->mapping, ret);
902 }
903}
904
905/*
906 * after a writepage IO is done, we need to:
907 * clear the uptodate bits on error
908 * clear the writeback bits in the extent tree for this IO
909 * end_page_writeback if the page has no more pending IO
910 *
911 * Scheduling is not allowed, so the extent state tree is expected
912 * to have one and only one object corresponding to this IO.
913 */
914static void end_bio_extent_writepage(struct btrfs_bio *bbio)
915{
916 struct bio *bio = &bbio->bio;
917 int error = blk_status_to_errno(bio->bi_status);
918 struct bio_vec *bvec;
919 u64 start;
920 u64 end;
921 struct bvec_iter_all iter_all;
922 bool first_bvec = true;
923
924 ASSERT(!bio_flagged(bio, BIO_CLONED));
925 bio_for_each_segment_all(bvec, bio, iter_all) {
926 struct page *page = bvec->bv_page;
927 struct inode *inode = page->mapping->host;
928 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
929 const u32 sectorsize = fs_info->sectorsize;
930
931 /* Our read/write should always be sector aligned. */
932 if (!IS_ALIGNED(bvec->bv_offset, sectorsize))
933 btrfs_err(fs_info,
934 "partial page write in btrfs with offset %u and length %u",
935 bvec->bv_offset, bvec->bv_len);
936 else if (!IS_ALIGNED(bvec->bv_len, sectorsize))
937 btrfs_info(fs_info,
938 "incomplete page write with offset %u and length %u",
939 bvec->bv_offset, bvec->bv_len);
940
941 start = page_offset(page) + bvec->bv_offset;
942 end = start + bvec->bv_len - 1;
943
944 if (first_bvec) {
945 btrfs_record_physical_zoned(inode, start, bio);
946 first_bvec = false;
947 }
948
949 end_extent_writepage(page, error, start, end);
950
951 btrfs_page_clear_writeback(fs_info, page, start, bvec->bv_len);
952 }
953
954 bio_put(bio);
955}
956
957/*
958 * Record previously processed extent range
959 *
960 * For endio_readpage_release_extent() to handle a full extent range, reducing
961 * the extent io operations.
962 */
963struct processed_extent {
964 struct btrfs_inode *inode;
965 /* Start of the range in @inode */
966 u64 start;
967 /* End of the range in @inode */
968 u64 end;
969 bool uptodate;
970};
971
972/*
973 * Try to release processed extent range
974 *
975 * May not release the extent range right now if the current range is
976 * contiguous to processed extent.
977 *
978 * Will release processed extent when any of @inode, @uptodate, the range is
979 * no longer contiguous to the processed range.
980 *
981 * Passing @inode == NULL will force processed extent to be released.
982 */
983static void endio_readpage_release_extent(struct processed_extent *processed,
984 struct btrfs_inode *inode, u64 start, u64 end,
985 bool uptodate)
986{
987 struct extent_state *cached = NULL;
988 struct extent_io_tree *tree;
989
990 /* The first extent, initialize @processed */
991 if (!processed->inode)
992 goto update;
993
994 /*
995 * Contiguous to processed extent, just uptodate the end.
996 *
997 * Several things to notice:
998 *
999 * - bio can be merged as long as on-disk bytenr is contiguous
1000 * This means we can have page belonging to other inodes, thus need to
1001 * check if the inode still matches.
1002 * - bvec can contain range beyond current page for multi-page bvec
1003 * Thus we need to do processed->end + 1 >= start check
1004 */
1005 if (processed->inode == inode && processed->uptodate == uptodate &&
1006 processed->end + 1 >= start && end >= processed->end) {
1007 processed->end = end;
1008 return;
1009 }
1010
1011 tree = &processed->inode->io_tree;
1012 /*
1013 * Now we don't have range contiguous to the processed range, release
1014 * the processed range now.
1015 */
1016 unlock_extent(tree, processed->start, processed->end, &cached);
1017
1018update:
1019 /* Update processed to current range */
1020 processed->inode = inode;
1021 processed->start = start;
1022 processed->end = end;
1023 processed->uptodate = uptodate;
1024}
1025
1026static void begin_page_read(struct btrfs_fs_info *fs_info, struct page *page)
1027{
1028 ASSERT(PageLocked(page));
1029 if (!btrfs_is_subpage(fs_info, page))
1030 return;
1031
1032 ASSERT(PagePrivate(page));
1033 btrfs_subpage_start_reader(fs_info, page, page_offset(page), PAGE_SIZE);
1034}
1035
1036/*
1037 * Find extent buffer for a givne bytenr.
1038 *
1039 * This is for end_bio_extent_readpage(), thus we can't do any unsafe locking
1040 * in endio context.
1041 */
1042static struct extent_buffer *find_extent_buffer_readpage(
1043 struct btrfs_fs_info *fs_info, struct page *page, u64 bytenr)
1044{
1045 struct extent_buffer *eb;
1046
1047 /*
1048 * For regular sectorsize, we can use page->private to grab extent
1049 * buffer
1050 */
1051 if (fs_info->nodesize >= PAGE_SIZE) {
1052 ASSERT(PagePrivate(page) && page->private);
1053 return (struct extent_buffer *)page->private;
1054 }
1055
1056 /* For subpage case, we need to lookup buffer radix tree */
1057 rcu_read_lock();
1058 eb = radix_tree_lookup(&fs_info->buffer_radix,
1059 bytenr >> fs_info->sectorsize_bits);
1060 rcu_read_unlock();
1061 ASSERT(eb);
1062 return eb;
1063}
1064
1065/*
1066 * after a readpage IO is done, we need to:
1067 * clear the uptodate bits on error
1068 * set the uptodate bits if things worked
1069 * set the page up to date if all extents in the tree are uptodate
1070 * clear the lock bit in the extent tree
1071 * unlock the page if there are no other extents locked for it
1072 *
1073 * Scheduling is not allowed, so the extent state tree is expected
1074 * to have one and only one object corresponding to this IO.
1075 */
1076static void end_bio_extent_readpage(struct btrfs_bio *bbio)
1077{
1078 struct bio *bio = &bbio->bio;
1079 struct bio_vec *bvec;
1080 struct processed_extent processed = { 0 };
1081 /*
1082 * The offset to the beginning of a bio, since one bio can never be
1083 * larger than UINT_MAX, u32 here is enough.
1084 */
1085 u32 bio_offset = 0;
1086 int mirror;
1087 struct bvec_iter_all iter_all;
1088
1089 ASSERT(!bio_flagged(bio, BIO_CLONED));
1090 bio_for_each_segment_all(bvec, bio, iter_all) {
1091 bool uptodate = !bio->bi_status;
1092 struct page *page = bvec->bv_page;
1093 struct inode *inode = page->mapping->host;
1094 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1095 const u32 sectorsize = fs_info->sectorsize;
1096 unsigned int error_bitmap = (unsigned int)-1;
1097 bool repair = false;
1098 u64 start;
1099 u64 end;
1100 u32 len;
1101
1102 btrfs_debug(fs_info,
1103 "end_bio_extent_readpage: bi_sector=%llu, err=%d, mirror=%u",
1104 bio->bi_iter.bi_sector, bio->bi_status,
1105 bbio->mirror_num);
1106
1107 /*
1108 * We always issue full-sector reads, but if some block in a
1109 * page fails to read, blk_update_request() will advance
1110 * bv_offset and adjust bv_len to compensate. Print a warning
1111 * for unaligned offsets, and an error if they don't add up to
1112 * a full sector.
1113 */
1114 if (!IS_ALIGNED(bvec->bv_offset, sectorsize))
1115 btrfs_err(fs_info,
1116 "partial page read in btrfs with offset %u and length %u",
1117 bvec->bv_offset, bvec->bv_len);
1118 else if (!IS_ALIGNED(bvec->bv_offset + bvec->bv_len,
1119 sectorsize))
1120 btrfs_info(fs_info,
1121 "incomplete page read with offset %u and length %u",
1122 bvec->bv_offset, bvec->bv_len);
1123
1124 start = page_offset(page) + bvec->bv_offset;
1125 end = start + bvec->bv_len - 1;
1126 len = bvec->bv_len;
1127
1128 mirror = bbio->mirror_num;
1129 if (likely(uptodate)) {
1130 if (is_data_inode(inode)) {
1131 error_bitmap = btrfs_verify_data_csum(bbio,
1132 bio_offset, page, start, end);
1133 if (error_bitmap)
1134 uptodate = false;
1135 } else {
1136 if (btrfs_validate_metadata_buffer(bbio,
1137 page, start, end, mirror))
1138 uptodate = false;
1139 }
1140 }
1141
1142 if (likely(uptodate)) {
1143 loff_t i_size = i_size_read(inode);
1144 pgoff_t end_index = i_size >> PAGE_SHIFT;
1145
1146 btrfs_clean_io_failure(BTRFS_I(inode), start, page, 0);
1147
1148 /*
1149 * Zero out the remaining part if this range straddles
1150 * i_size.
1151 *
1152 * Here we should only zero the range inside the bvec,
1153 * not touch anything else.
1154 *
1155 * NOTE: i_size is exclusive while end is inclusive.
1156 */
1157 if (page->index == end_index && i_size <= end) {
1158 u32 zero_start = max(offset_in_page(i_size),
1159 offset_in_page(start));
1160
1161 zero_user_segment(page, zero_start,
1162 offset_in_page(end) + 1);
1163 }
1164 } else if (is_data_inode(inode)) {
1165 /*
1166 * Only try to repair bios that actually made it to a
1167 * device. If the bio failed to be submitted mirror
1168 * is 0 and we need to fail it without retrying.
1169 *
1170 * This also includes the high level bios for compressed
1171 * extents - these never make it to a device and repair
1172 * is already handled on the lower compressed bio.
1173 */
1174 if (mirror > 0)
1175 repair = true;
1176 } else {
1177 struct extent_buffer *eb;
1178
1179 eb = find_extent_buffer_readpage(fs_info, page, start);
1180 set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
1181 eb->read_mirror = mirror;
1182 atomic_dec(&eb->io_pages);
1183 }
1184
1185 if (repair) {
1186 /*
1187 * submit_data_read_repair() will handle all the good
1188 * and bad sectors, we just continue to the next bvec.
1189 */
1190 submit_data_read_repair(inode, bbio, bio_offset, bvec,
1191 error_bitmap);
1192 } else {
1193 /* Update page status and unlock */
1194 end_page_read(page, uptodate, start, len);
1195 endio_readpage_release_extent(&processed, BTRFS_I(inode),
1196 start, end, PageUptodate(page));
1197 }
1198
1199 ASSERT(bio_offset + len > bio_offset);
1200 bio_offset += len;
1201
1202 }
1203 /* Release the last extent */
1204 endio_readpage_release_extent(&processed, NULL, 0, 0, false);
1205 btrfs_bio_free_csum(bbio);
1206 bio_put(bio);
1207}
1208
1209/*
1210 * Populate every free slot in a provided array with pages.
1211 *
1212 * @nr_pages: number of pages to allocate
1213 * @page_array: the array to fill with pages; any existing non-null entries in
1214 * the array will be skipped
1215 *
1216 * Return: 0 if all pages were able to be allocated;
1217 * -ENOMEM otherwise, and the caller is responsible for freeing all
1218 * non-null page pointers in the array.
1219 */
1220int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array)
1221{
1222 unsigned int allocated;
1223
1224 for (allocated = 0; allocated < nr_pages;) {
1225 unsigned int last = allocated;
1226
1227 allocated = alloc_pages_bulk_array(GFP_NOFS, nr_pages, page_array);
1228
1229 if (allocated == nr_pages)
1230 return 0;
1231
1232 /*
1233 * During this iteration, no page could be allocated, even
1234 * though alloc_pages_bulk_array() falls back to alloc_page()
1235 * if it could not bulk-allocate. So we must be out of memory.
1236 */
1237 if (allocated == last)
1238 return -ENOMEM;
1239
1240 memalloc_retry_wait(GFP_NOFS);
1241 }
1242 return 0;
1243}
1244
1245/*
1246 * Attempt to add a page to bio.
1247 *
1248 * @bio_ctrl: record both the bio, and its bio_flags
1249 * @page: page to add to the bio
1250 * @disk_bytenr: offset of the new bio or to check whether we are adding
1251 * a contiguous page to the previous one
1252 * @size: portion of page that we want to write
1253 * @pg_offset: starting offset in the page
1254 * @compress_type: compression type of the current bio to see if we can merge them
1255 *
1256 * Attempt to add a page to bio considering stripe alignment etc.
1257 *
1258 * Return >= 0 for the number of bytes added to the bio.
1259 * Can return 0 if the current bio is already at stripe/zone boundary.
1260 * Return <0 for error.
1261 */
1262static int btrfs_bio_add_page(struct btrfs_bio_ctrl *bio_ctrl,
1263 struct page *page,
1264 u64 disk_bytenr, unsigned int size,
1265 unsigned int pg_offset,
1266 enum btrfs_compression_type compress_type)
1267{
1268 struct bio *bio = bio_ctrl->bio;
1269 u32 bio_size = bio->bi_iter.bi_size;
1270 u32 real_size;
1271 const sector_t sector = disk_bytenr >> SECTOR_SHIFT;
1272 bool contig = false;
1273 int ret;
1274
1275 ASSERT(bio);
1276 /* The limit should be calculated when bio_ctrl->bio is allocated */
1277 ASSERT(bio_ctrl->len_to_oe_boundary && bio_ctrl->len_to_stripe_boundary);
1278 if (bio_ctrl->compress_type != compress_type)
1279 return 0;
1280
1281
1282 if (bio->bi_iter.bi_size == 0) {
1283 /* We can always add a page into an empty bio. */
1284 contig = true;
1285 } else if (bio_ctrl->compress_type == BTRFS_COMPRESS_NONE) {
1286 struct bio_vec *bvec = bio_last_bvec_all(bio);
1287
1288 /*
1289 * The contig check requires the following conditions to be met:
1290 * 1) The pages are belonging to the same inode
1291 * This is implied by the call chain.
1292 *
1293 * 2) The range has adjacent logical bytenr
1294 *
1295 * 3) The range has adjacent file offset
1296 * This is required for the usage of btrfs_bio->file_offset.
1297 */
1298 if (bio_end_sector(bio) == sector &&
1299 page_offset(bvec->bv_page) + bvec->bv_offset +
1300 bvec->bv_len == page_offset(page) + pg_offset)
1301 contig = true;
1302 } else {
1303 /*
1304 * For compression, all IO should have its logical bytenr
1305 * set to the starting bytenr of the compressed extent.
1306 */
1307 contig = bio->bi_iter.bi_sector == sector;
1308 }
1309
1310 if (!contig)
1311 return 0;
1312
1313 real_size = min(bio_ctrl->len_to_oe_boundary,
1314 bio_ctrl->len_to_stripe_boundary) - bio_size;
1315 real_size = min(real_size, size);
1316
1317 /*
1318 * If real_size is 0, never call bio_add_*_page(), as even size is 0,
1319 * bio will still execute its endio function on the page!
1320 */
1321 if (real_size == 0)
1322 return 0;
1323
1324 if (bio_op(bio) == REQ_OP_ZONE_APPEND)
1325 ret = bio_add_zone_append_page(bio, page, real_size, pg_offset);
1326 else
1327 ret = bio_add_page(bio, page, real_size, pg_offset);
1328
1329 return ret;
1330}
1331
1332static int calc_bio_boundaries(struct btrfs_bio_ctrl *bio_ctrl,
1333 struct btrfs_inode *inode, u64 file_offset)
1334{
1335 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1336 struct btrfs_io_geometry geom;
1337 struct btrfs_ordered_extent *ordered;
1338 struct extent_map *em;
1339 u64 logical = (bio_ctrl->bio->bi_iter.bi_sector << SECTOR_SHIFT);
1340 int ret;
1341
1342 /*
1343 * Pages for compressed extent are never submitted to disk directly,
1344 * thus it has no real boundary, just set them to U32_MAX.
1345 *
1346 * The split happens for real compressed bio, which happens in
1347 * btrfs_submit_compressed_read/write().
1348 */
1349 if (bio_ctrl->compress_type != BTRFS_COMPRESS_NONE) {
1350 bio_ctrl->len_to_oe_boundary = U32_MAX;
1351 bio_ctrl->len_to_stripe_boundary = U32_MAX;
1352 return 0;
1353 }
1354 em = btrfs_get_chunk_map(fs_info, logical, fs_info->sectorsize);
1355 if (IS_ERR(em))
1356 return PTR_ERR(em);
1357 ret = btrfs_get_io_geometry(fs_info, em, btrfs_op(bio_ctrl->bio),
1358 logical, &geom);
1359 free_extent_map(em);
1360 if (ret < 0) {
1361 return ret;
1362 }
1363 if (geom.len > U32_MAX)
1364 bio_ctrl->len_to_stripe_boundary = U32_MAX;
1365 else
1366 bio_ctrl->len_to_stripe_boundary = (u32)geom.len;
1367
1368 if (bio_op(bio_ctrl->bio) != REQ_OP_ZONE_APPEND) {
1369 bio_ctrl->len_to_oe_boundary = U32_MAX;
1370 return 0;
1371 }
1372
1373 /* Ordered extent not yet created, so we're good */
1374 ordered = btrfs_lookup_ordered_extent(inode, file_offset);
1375 if (!ordered) {
1376 bio_ctrl->len_to_oe_boundary = U32_MAX;
1377 return 0;
1378 }
1379
1380 bio_ctrl->len_to_oe_boundary = min_t(u32, U32_MAX,
1381 ordered->disk_bytenr + ordered->disk_num_bytes - logical);
1382 btrfs_put_ordered_extent(ordered);
1383 return 0;
1384}
1385
1386static int alloc_new_bio(struct btrfs_inode *inode,
1387 struct btrfs_bio_ctrl *bio_ctrl,
1388 struct writeback_control *wbc,
1389 blk_opf_t opf,
1390 u64 disk_bytenr, u32 offset, u64 file_offset,
1391 enum btrfs_compression_type compress_type)
1392{
1393 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1394 struct bio *bio;
1395 int ret;
1396
1397 ASSERT(bio_ctrl->end_io_func);
1398
1399 bio = btrfs_bio_alloc(BIO_MAX_VECS, opf, bio_ctrl->end_io_func, NULL);
1400 /*
1401 * For compressed page range, its disk_bytenr is always @disk_bytenr
1402 * passed in, no matter if we have added any range into previous bio.
1403 */
1404 if (compress_type != BTRFS_COMPRESS_NONE)
1405 bio->bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
1406 else
1407 bio->bi_iter.bi_sector = (disk_bytenr + offset) >> SECTOR_SHIFT;
1408 bio_ctrl->bio = bio;
1409 bio_ctrl->compress_type = compress_type;
1410 ret = calc_bio_boundaries(bio_ctrl, inode, file_offset);
1411 if (ret < 0)
1412 goto error;
1413
1414 if (wbc) {
1415 /*
1416 * For Zone append we need the correct block_device that we are
1417 * going to write to set in the bio to be able to respect the
1418 * hardware limitation. Look it up here:
1419 */
1420 if (bio_op(bio) == REQ_OP_ZONE_APPEND) {
1421 struct btrfs_device *dev;
1422
1423 dev = btrfs_zoned_get_device(fs_info, disk_bytenr,
1424 fs_info->sectorsize);
1425 if (IS_ERR(dev)) {
1426 ret = PTR_ERR(dev);
1427 goto error;
1428 }
1429
1430 bio_set_dev(bio, dev->bdev);
1431 } else {
1432 /*
1433 * Otherwise pick the last added device to support
1434 * cgroup writeback. For multi-device file systems this
1435 * means blk-cgroup policies have to always be set on the
1436 * last added/replaced device. This is a bit odd but has
1437 * been like that for a long time.
1438 */
1439 bio_set_dev(bio, fs_info->fs_devices->latest_dev->bdev);
1440 }
1441 wbc_init_bio(wbc, bio);
1442 } else {
1443 ASSERT(bio_op(bio) != REQ_OP_ZONE_APPEND);
1444 }
1445 return 0;
1446error:
1447 bio_ctrl->bio = NULL;
1448 btrfs_bio_end_io(btrfs_bio(bio), errno_to_blk_status(ret));
1449 return ret;
1450}
1451
1452/*
1453 * @opf: bio REQ_OP_* and REQ_* flags as one value
1454 * @wbc: optional writeback control for io accounting
1455 * @disk_bytenr: logical bytenr where the write will be
1456 * @page: page to add to the bio
1457 * @size: portion of page that we want to write to
1458 * @pg_offset: offset of the new bio or to check whether we are adding
1459 * a contiguous page to the previous one
1460 * @compress_type: compress type for current bio
1461 *
1462 * The will either add the page into the existing @bio_ctrl->bio, or allocate a
1463 * new one in @bio_ctrl->bio.
1464 * The mirror number for this IO should already be initizlied in
1465 * @bio_ctrl->mirror_num.
1466 */
1467static int submit_extent_page(blk_opf_t opf,
1468 struct writeback_control *wbc,
1469 struct btrfs_bio_ctrl *bio_ctrl,
1470 u64 disk_bytenr, struct page *page,
1471 size_t size, unsigned long pg_offset,
1472 enum btrfs_compression_type compress_type,
1473 bool force_bio_submit)
1474{
1475 int ret = 0;
1476 struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
1477 unsigned int cur = pg_offset;
1478
1479 ASSERT(bio_ctrl);
1480
1481 ASSERT(pg_offset < PAGE_SIZE && size <= PAGE_SIZE &&
1482 pg_offset + size <= PAGE_SIZE);
1483
1484 ASSERT(bio_ctrl->end_io_func);
1485
1486 if (force_bio_submit)
1487 submit_one_bio(bio_ctrl);
1488
1489 while (cur < pg_offset + size) {
1490 u32 offset = cur - pg_offset;
1491 int added;
1492
1493 /* Allocate new bio if needed */
1494 if (!bio_ctrl->bio) {
1495 ret = alloc_new_bio(inode, bio_ctrl, wbc, opf,
1496 disk_bytenr, offset,
1497 page_offset(page) + cur,
1498 compress_type);
1499 if (ret < 0)
1500 return ret;
1501 }
1502 /*
1503 * We must go through btrfs_bio_add_page() to ensure each
1504 * page range won't cross various boundaries.
1505 */
1506 if (compress_type != BTRFS_COMPRESS_NONE)
1507 added = btrfs_bio_add_page(bio_ctrl, page, disk_bytenr,
1508 size - offset, pg_offset + offset,
1509 compress_type);
1510 else
1511 added = btrfs_bio_add_page(bio_ctrl, page,
1512 disk_bytenr + offset, size - offset,
1513 pg_offset + offset, compress_type);
1514
1515 /* Metadata page range should never be split */
1516 if (!is_data_inode(&inode->vfs_inode))
1517 ASSERT(added == 0 || added == size - offset);
1518
1519 /* At least we added some page, update the account */
1520 if (wbc && added)
1521 wbc_account_cgroup_owner(wbc, page, added);
1522
1523 /* We have reached boundary, submit right now */
1524 if (added < size - offset) {
1525 /* The bio should contain some page(s) */
1526 ASSERT(bio_ctrl->bio->bi_iter.bi_size);
1527 submit_one_bio(bio_ctrl);
1528 }
1529 cur += added;
1530 }
1531 return 0;
1532}
1533
1534static int attach_extent_buffer_page(struct extent_buffer *eb,
1535 struct page *page,
1536 struct btrfs_subpage *prealloc)
1537{
1538 struct btrfs_fs_info *fs_info = eb->fs_info;
1539 int ret = 0;
1540
1541 /*
1542 * If the page is mapped to btree inode, we should hold the private
1543 * lock to prevent race.
1544 * For cloned or dummy extent buffers, their pages are not mapped and
1545 * will not race with any other ebs.
1546 */
1547 if (page->mapping)
1548 lockdep_assert_held(&page->mapping->private_lock);
1549
1550 if (fs_info->nodesize >= PAGE_SIZE) {
1551 if (!PagePrivate(page))
1552 attach_page_private(page, eb);
1553 else
1554 WARN_ON(page->private != (unsigned long)eb);
1555 return 0;
1556 }
1557
1558 /* Already mapped, just free prealloc */
1559 if (PagePrivate(page)) {
1560 btrfs_free_subpage(prealloc);
1561 return 0;
1562 }
1563
1564 if (prealloc)
1565 /* Has preallocated memory for subpage */
1566 attach_page_private(page, prealloc);
1567 else
1568 /* Do new allocation to attach subpage */
1569 ret = btrfs_attach_subpage(fs_info, page,
1570 BTRFS_SUBPAGE_METADATA);
1571 return ret;
1572}
1573
1574int set_page_extent_mapped(struct page *page)
1575{
1576 struct btrfs_fs_info *fs_info;
1577
1578 ASSERT(page->mapping);
1579
1580 if (PagePrivate(page))
1581 return 0;
1582
1583 fs_info = btrfs_sb(page->mapping->host->i_sb);
1584
1585 if (btrfs_is_subpage(fs_info, page))
1586 return btrfs_attach_subpage(fs_info, page, BTRFS_SUBPAGE_DATA);
1587
1588 attach_page_private(page, (void *)EXTENT_PAGE_PRIVATE);
1589 return 0;
1590}
1591
1592void clear_page_extent_mapped(struct page *page)
1593{
1594 struct btrfs_fs_info *fs_info;
1595
1596 ASSERT(page->mapping);
1597
1598 if (!PagePrivate(page))
1599 return;
1600
1601 fs_info = btrfs_sb(page->mapping->host->i_sb);
1602 if (btrfs_is_subpage(fs_info, page))
1603 return btrfs_detach_subpage(fs_info, page);
1604
1605 detach_page_private(page);
1606}
1607
1608static struct extent_map *
1609__get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
1610 u64 start, u64 len, struct extent_map **em_cached)
1611{
1612 struct extent_map *em;
1613
1614 if (em_cached && *em_cached) {
1615 em = *em_cached;
1616 if (extent_map_in_tree(em) && start >= em->start &&
1617 start < extent_map_end(em)) {
1618 refcount_inc(&em->refs);
1619 return em;
1620 }
1621
1622 free_extent_map(em);
1623 *em_cached = NULL;
1624 }
1625
1626 em = btrfs_get_extent(BTRFS_I(inode), page, pg_offset, start, len);
1627 if (em_cached && !IS_ERR(em)) {
1628 BUG_ON(*em_cached);
1629 refcount_inc(&em->refs);
1630 *em_cached = em;
1631 }
1632 return em;
1633}
1634/*
1635 * basic readpage implementation. Locked extent state structs are inserted
1636 * into the tree that are removed when the IO is done (by the end_io
1637 * handlers)
1638 * XXX JDM: This needs looking at to ensure proper page locking
1639 * return 0 on success, otherwise return error
1640 */
1641static int btrfs_do_readpage(struct page *page, struct extent_map **em_cached,
1642 struct btrfs_bio_ctrl *bio_ctrl,
1643 blk_opf_t read_flags, u64 *prev_em_start)
1644{
1645 struct inode *inode = page->mapping->host;
1646 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1647 u64 start = page_offset(page);
1648 const u64 end = start + PAGE_SIZE - 1;
1649 u64 cur = start;
1650 u64 extent_offset;
1651 u64 last_byte = i_size_read(inode);
1652 u64 block_start;
1653 struct extent_map *em;
1654 int ret = 0;
1655 size_t pg_offset = 0;
1656 size_t iosize;
1657 size_t blocksize = inode->i_sb->s_blocksize;
1658 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1659
1660 ret = set_page_extent_mapped(page);
1661 if (ret < 0) {
1662 unlock_extent(tree, start, end, NULL);
1663 btrfs_page_set_error(fs_info, page, start, PAGE_SIZE);
1664 unlock_page(page);
1665 goto out;
1666 }
1667
1668 if (page->index == last_byte >> PAGE_SHIFT) {
1669 size_t zero_offset = offset_in_page(last_byte);
1670
1671 if (zero_offset) {
1672 iosize = PAGE_SIZE - zero_offset;
1673 memzero_page(page, zero_offset, iosize);
1674 }
1675 }
1676 bio_ctrl->end_io_func = end_bio_extent_readpage;
1677 begin_page_read(fs_info, page);
1678 while (cur <= end) {
1679 unsigned long this_bio_flag = 0;
1680 bool force_bio_submit = false;
1681 u64 disk_bytenr;
1682
1683 ASSERT(IS_ALIGNED(cur, fs_info->sectorsize));
1684 if (cur >= last_byte) {
1685 iosize = PAGE_SIZE - pg_offset;
1686 memzero_page(page, pg_offset, iosize);
1687 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1688 end_page_read(page, true, cur, iosize);
1689 break;
1690 }
1691 em = __get_extent_map(inode, page, pg_offset, cur,
1692 end - cur + 1, em_cached);
1693 if (IS_ERR(em)) {
1694 unlock_extent(tree, cur, end, NULL);
1695 end_page_read(page, false, cur, end + 1 - cur);
1696 ret = PTR_ERR(em);
1697 break;
1698 }
1699 extent_offset = cur - em->start;
1700 BUG_ON(extent_map_end(em) <= cur);
1701 BUG_ON(end < cur);
1702
1703 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
1704 this_bio_flag = em->compress_type;
1705
1706 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1707 iosize = ALIGN(iosize, blocksize);
1708 if (this_bio_flag != BTRFS_COMPRESS_NONE)
1709 disk_bytenr = em->block_start;
1710 else
1711 disk_bytenr = em->block_start + extent_offset;
1712 block_start = em->block_start;
1713 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
1714 block_start = EXTENT_MAP_HOLE;
1715
1716 /*
1717 * If we have a file range that points to a compressed extent
1718 * and it's followed by a consecutive file range that points
1719 * to the same compressed extent (possibly with a different
1720 * offset and/or length, so it either points to the whole extent
1721 * or only part of it), we must make sure we do not submit a
1722 * single bio to populate the pages for the 2 ranges because
1723 * this makes the compressed extent read zero out the pages
1724 * belonging to the 2nd range. Imagine the following scenario:
1725 *
1726 * File layout
1727 * [0 - 8K] [8K - 24K]
1728 * | |
1729 * | |
1730 * points to extent X, points to extent X,
1731 * offset 4K, length of 8K offset 0, length 16K
1732 *
1733 * [extent X, compressed length = 4K uncompressed length = 16K]
1734 *
1735 * If the bio to read the compressed extent covers both ranges,
1736 * it will decompress extent X into the pages belonging to the
1737 * first range and then it will stop, zeroing out the remaining
1738 * pages that belong to the other range that points to extent X.
1739 * So here we make sure we submit 2 bios, one for the first
1740 * range and another one for the third range. Both will target
1741 * the same physical extent from disk, but we can't currently
1742 * make the compressed bio endio callback populate the pages
1743 * for both ranges because each compressed bio is tightly
1744 * coupled with a single extent map, and each range can have
1745 * an extent map with a different offset value relative to the
1746 * uncompressed data of our extent and different lengths. This
1747 * is a corner case so we prioritize correctness over
1748 * non-optimal behavior (submitting 2 bios for the same extent).
1749 */
1750 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) &&
1751 prev_em_start && *prev_em_start != (u64)-1 &&
1752 *prev_em_start != em->start)
1753 force_bio_submit = true;
1754
1755 if (prev_em_start)
1756 *prev_em_start = em->start;
1757
1758 free_extent_map(em);
1759 em = NULL;
1760
1761 /* we've found a hole, just zero and go on */
1762 if (block_start == EXTENT_MAP_HOLE) {
1763 memzero_page(page, pg_offset, iosize);
1764
1765 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1766 end_page_read(page, true, cur, iosize);
1767 cur = cur + iosize;
1768 pg_offset += iosize;
1769 continue;
1770 }
1771 /* the get_extent function already copied into the page */
1772 if (block_start == EXTENT_MAP_INLINE) {
1773 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1774 end_page_read(page, true, cur, iosize);
1775 cur = cur + iosize;
1776 pg_offset += iosize;
1777 continue;
1778 }
1779
1780 ret = submit_extent_page(REQ_OP_READ | read_flags, NULL,
1781 bio_ctrl, disk_bytenr, page, iosize,
1782 pg_offset, this_bio_flag,
1783 force_bio_submit);
1784 if (ret) {
1785 /*
1786 * We have to unlock the remaining range, or the page
1787 * will never be unlocked.
1788 */
1789 unlock_extent(tree, cur, end, NULL);
1790 end_page_read(page, false, cur, end + 1 - cur);
1791 goto out;
1792 }
1793 cur = cur + iosize;
1794 pg_offset += iosize;
1795 }
1796out:
1797 return ret;
1798}
1799
1800int btrfs_read_folio(struct file *file, struct folio *folio)
1801{
1802 struct page *page = &folio->page;
1803 struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
1804 u64 start = page_offset(page);
1805 u64 end = start + PAGE_SIZE - 1;
1806 struct btrfs_bio_ctrl bio_ctrl = { 0 };
1807 int ret;
1808
1809 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
1810
1811 ret = btrfs_do_readpage(page, NULL, &bio_ctrl, 0, NULL);
1812 /*
1813 * If btrfs_do_readpage() failed we will want to submit the assembled
1814 * bio to do the cleanup.
1815 */
1816 submit_one_bio(&bio_ctrl);
1817 return ret;
1818}
1819
1820static inline void contiguous_readpages(struct page *pages[], int nr_pages,
1821 u64 start, u64 end,
1822 struct extent_map **em_cached,
1823 struct btrfs_bio_ctrl *bio_ctrl,
1824 u64 *prev_em_start)
1825{
1826 struct btrfs_inode *inode = BTRFS_I(pages[0]->mapping->host);
1827 int index;
1828
1829 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
1830
1831 for (index = 0; index < nr_pages; index++) {
1832 btrfs_do_readpage(pages[index], em_cached, bio_ctrl,
1833 REQ_RAHEAD, prev_em_start);
1834 put_page(pages[index]);
1835 }
1836}
1837
1838/*
1839 * helper for __extent_writepage, doing all of the delayed allocation setup.
1840 *
1841 * This returns 1 if btrfs_run_delalloc_range function did all the work required
1842 * to write the page (copy into inline extent). In this case the IO has
1843 * been started and the page is already unlocked.
1844 *
1845 * This returns 0 if all went well (page still locked)
1846 * This returns < 0 if there were errors (page still locked)
1847 */
1848static noinline_for_stack int writepage_delalloc(struct btrfs_inode *inode,
1849 struct page *page, struct writeback_control *wbc)
1850{
1851 const u64 page_end = page_offset(page) + PAGE_SIZE - 1;
1852 u64 delalloc_start = page_offset(page);
1853 u64 delalloc_to_write = 0;
1854 /* How many pages are started by btrfs_run_delalloc_range() */
1855 unsigned long nr_written = 0;
1856 int ret;
1857 int page_started = 0;
1858
1859 while (delalloc_start < page_end) {
1860 u64 delalloc_end = page_end;
1861 bool found;
1862
1863 found = find_lock_delalloc_range(&inode->vfs_inode, page,
1864 &delalloc_start,
1865 &delalloc_end);
1866 if (!found) {
1867 delalloc_start = delalloc_end + 1;
1868 continue;
1869 }
1870 ret = btrfs_run_delalloc_range(inode, page, delalloc_start,
1871 delalloc_end, &page_started, &nr_written, wbc);
1872 if (ret) {
1873 btrfs_page_set_error(inode->root->fs_info, page,
1874 page_offset(page), PAGE_SIZE);
1875 return ret;
1876 }
1877 /*
1878 * delalloc_end is already one less than the total length, so
1879 * we don't subtract one from PAGE_SIZE
1880 */
1881 delalloc_to_write += (delalloc_end - delalloc_start +
1882 PAGE_SIZE) >> PAGE_SHIFT;
1883 delalloc_start = delalloc_end + 1;
1884 }
1885 if (wbc->nr_to_write < delalloc_to_write) {
1886 int thresh = 8192;
1887
1888 if (delalloc_to_write < thresh * 2)
1889 thresh = delalloc_to_write;
1890 wbc->nr_to_write = min_t(u64, delalloc_to_write,
1891 thresh);
1892 }
1893
1894 /* Did btrfs_run_dealloc_range() already unlock and start the IO? */
1895 if (page_started) {
1896 /*
1897 * We've unlocked the page, so we can't update the mapping's
1898 * writeback index, just update nr_to_write.
1899 */
1900 wbc->nr_to_write -= nr_written;
1901 return 1;
1902 }
1903
1904 return 0;
1905}
1906
1907/*
1908 * Find the first byte we need to write.
1909 *
1910 * For subpage, one page can contain several sectors, and
1911 * __extent_writepage_io() will just grab all extent maps in the page
1912 * range and try to submit all non-inline/non-compressed extents.
1913 *
1914 * This is a big problem for subpage, we shouldn't re-submit already written
1915 * data at all.
1916 * This function will lookup subpage dirty bit to find which range we really
1917 * need to submit.
1918 *
1919 * Return the next dirty range in [@start, @end).
1920 * If no dirty range is found, @start will be page_offset(page) + PAGE_SIZE.
1921 */
1922static void find_next_dirty_byte(struct btrfs_fs_info *fs_info,
1923 struct page *page, u64 *start, u64 *end)
1924{
1925 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
1926 struct btrfs_subpage_info *spi = fs_info->subpage_info;
1927 u64 orig_start = *start;
1928 /* Declare as unsigned long so we can use bitmap ops */
1929 unsigned long flags;
1930 int range_start_bit;
1931 int range_end_bit;
1932
1933 /*
1934 * For regular sector size == page size case, since one page only
1935 * contains one sector, we return the page offset directly.
1936 */
1937 if (!btrfs_is_subpage(fs_info, page)) {
1938 *start = page_offset(page);
1939 *end = page_offset(page) + PAGE_SIZE;
1940 return;
1941 }
1942
1943 range_start_bit = spi->dirty_offset +
1944 (offset_in_page(orig_start) >> fs_info->sectorsize_bits);
1945
1946 /* We should have the page locked, but just in case */
1947 spin_lock_irqsave(&subpage->lock, flags);
1948 bitmap_next_set_region(subpage->bitmaps, &range_start_bit, &range_end_bit,
1949 spi->dirty_offset + spi->bitmap_nr_bits);
1950 spin_unlock_irqrestore(&subpage->lock, flags);
1951
1952 range_start_bit -= spi->dirty_offset;
1953 range_end_bit -= spi->dirty_offset;
1954
1955 *start = page_offset(page) + range_start_bit * fs_info->sectorsize;
1956 *end = page_offset(page) + range_end_bit * fs_info->sectorsize;
1957}
1958
1959/*
1960 * helper for __extent_writepage. This calls the writepage start hooks,
1961 * and does the loop to map the page into extents and bios.
1962 *
1963 * We return 1 if the IO is started and the page is unlocked,
1964 * 0 if all went well (page still locked)
1965 * < 0 if there were errors (page still locked)
1966 */
1967static noinline_for_stack int __extent_writepage_io(struct btrfs_inode *inode,
1968 struct page *page,
1969 struct writeback_control *wbc,
1970 struct btrfs_bio_ctrl *bio_ctrl,
1971 loff_t i_size,
1972 int *nr_ret)
1973{
1974 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1975 u64 cur = page_offset(page);
1976 u64 end = cur + PAGE_SIZE - 1;
1977 u64 extent_offset;
1978 u64 block_start;
1979 struct extent_map *em;
1980 int saved_ret = 0;
1981 int ret = 0;
1982 int nr = 0;
1983 enum req_op op = REQ_OP_WRITE;
1984 const blk_opf_t write_flags = wbc_to_write_flags(wbc);
1985 bool has_error = false;
1986 bool compressed;
1987
1988 ret = btrfs_writepage_cow_fixup(page);
1989 if (ret) {
1990 /* Fixup worker will requeue */
1991 redirty_page_for_writepage(wbc, page);
1992 unlock_page(page);
1993 return 1;
1994 }
1995
1996 /*
1997 * we don't want to touch the inode after unlocking the page,
1998 * so we update the mapping writeback index now
1999 */
2000 wbc->nr_to_write--;
2001
2002 bio_ctrl->end_io_func = end_bio_extent_writepage;
2003 while (cur <= end) {
2004 u64 disk_bytenr;
2005 u64 em_end;
2006 u64 dirty_range_start = cur;
2007 u64 dirty_range_end;
2008 u32 iosize;
2009
2010 if (cur >= i_size) {
2011 btrfs_writepage_endio_finish_ordered(inode, page, cur,
2012 end, true);
2013 /*
2014 * This range is beyond i_size, thus we don't need to
2015 * bother writing back.
2016 * But we still need to clear the dirty subpage bit, or
2017 * the next time the page gets dirtied, we will try to
2018 * writeback the sectors with subpage dirty bits,
2019 * causing writeback without ordered extent.
2020 */
2021 btrfs_page_clear_dirty(fs_info, page, cur, end + 1 - cur);
2022 break;
2023 }
2024
2025 find_next_dirty_byte(fs_info, page, &dirty_range_start,
2026 &dirty_range_end);
2027 if (cur < dirty_range_start) {
2028 cur = dirty_range_start;
2029 continue;
2030 }
2031
2032 em = btrfs_get_extent(inode, NULL, 0, cur, end - cur + 1);
2033 if (IS_ERR(em)) {
2034 btrfs_page_set_error(fs_info, page, cur, end - cur + 1);
2035 ret = PTR_ERR_OR_ZERO(em);
2036 has_error = true;
2037 if (!saved_ret)
2038 saved_ret = ret;
2039 break;
2040 }
2041
2042 extent_offset = cur - em->start;
2043 em_end = extent_map_end(em);
2044 ASSERT(cur <= em_end);
2045 ASSERT(cur < end);
2046 ASSERT(IS_ALIGNED(em->start, fs_info->sectorsize));
2047 ASSERT(IS_ALIGNED(em->len, fs_info->sectorsize));
2048 block_start = em->block_start;
2049 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2050 disk_bytenr = em->block_start + extent_offset;
2051
2052 /*
2053 * Note that em_end from extent_map_end() and dirty_range_end from
2054 * find_next_dirty_byte() are all exclusive
2055 */
2056 iosize = min(min(em_end, end + 1), dirty_range_end) - cur;
2057
2058 if (btrfs_use_zone_append(inode, em->block_start))
2059 op = REQ_OP_ZONE_APPEND;
2060
2061 free_extent_map(em);
2062 em = NULL;
2063
2064 /*
2065 * compressed and inline extents are written through other
2066 * paths in the FS
2067 */
2068 if (compressed || block_start == EXTENT_MAP_HOLE ||
2069 block_start == EXTENT_MAP_INLINE) {
2070 if (compressed)
2071 nr++;
2072 else
2073 btrfs_writepage_endio_finish_ordered(inode,
2074 page, cur, cur + iosize - 1, true);
2075 btrfs_page_clear_dirty(fs_info, page, cur, iosize);
2076 cur += iosize;
2077 continue;
2078 }
2079
2080 btrfs_set_range_writeback(inode, cur, cur + iosize - 1);
2081 if (!PageWriteback(page)) {
2082 btrfs_err(inode->root->fs_info,
2083 "page %lu not writeback, cur %llu end %llu",
2084 page->index, cur, end);
2085 }
2086
2087 /*
2088 * Although the PageDirty bit is cleared before entering this
2089 * function, subpage dirty bit is not cleared.
2090 * So clear subpage dirty bit here so next time we won't submit
2091 * page for range already written to disk.
2092 */
2093 btrfs_page_clear_dirty(fs_info, page, cur, iosize);
2094
2095 ret = submit_extent_page(op | write_flags, wbc,
2096 bio_ctrl, disk_bytenr,
2097 page, iosize,
2098 cur - page_offset(page),
2099 0, false);
2100 if (ret) {
2101 has_error = true;
2102 if (!saved_ret)
2103 saved_ret = ret;
2104
2105 btrfs_page_set_error(fs_info, page, cur, iosize);
2106 if (PageWriteback(page))
2107 btrfs_page_clear_writeback(fs_info, page, cur,
2108 iosize);
2109 }
2110
2111 cur += iosize;
2112 nr++;
2113 }
2114 /*
2115 * If we finish without problem, we should not only clear page dirty,
2116 * but also empty subpage dirty bits
2117 */
2118 if (!has_error)
2119 btrfs_page_assert_not_dirty(fs_info, page);
2120 else
2121 ret = saved_ret;
2122 *nr_ret = nr;
2123 return ret;
2124}
2125
2126/*
2127 * the writepage semantics are similar to regular writepage. extent
2128 * records are inserted to lock ranges in the tree, and as dirty areas
2129 * are found, they are marked writeback. Then the lock bits are removed
2130 * and the end_io handler clears the writeback ranges
2131 *
2132 * Return 0 if everything goes well.
2133 * Return <0 for error.
2134 */
2135static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2136 struct btrfs_bio_ctrl *bio_ctrl)
2137{
2138 struct folio *folio = page_folio(page);
2139 struct inode *inode = page->mapping->host;
2140 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2141 const u64 page_start = page_offset(page);
2142 const u64 page_end = page_start + PAGE_SIZE - 1;
2143 int ret;
2144 int nr = 0;
2145 size_t pg_offset;
2146 loff_t i_size = i_size_read(inode);
2147 unsigned long end_index = i_size >> PAGE_SHIFT;
2148
2149 trace___extent_writepage(page, inode, wbc);
2150
2151 WARN_ON(!PageLocked(page));
2152
2153 btrfs_page_clear_error(btrfs_sb(inode->i_sb), page,
2154 page_offset(page), PAGE_SIZE);
2155
2156 pg_offset = offset_in_page(i_size);
2157 if (page->index > end_index ||
2158 (page->index == end_index && !pg_offset)) {
2159 folio_invalidate(folio, 0, folio_size(folio));
2160 folio_unlock(folio);
2161 return 0;
2162 }
2163
2164 if (page->index == end_index)
2165 memzero_page(page, pg_offset, PAGE_SIZE - pg_offset);
2166
2167 ret = set_page_extent_mapped(page);
2168 if (ret < 0) {
2169 SetPageError(page);
2170 goto done;
2171 }
2172
2173 if (!bio_ctrl->extent_locked) {
2174 ret = writepage_delalloc(BTRFS_I(inode), page, wbc);
2175 if (ret == 1)
2176 return 0;
2177 if (ret)
2178 goto done;
2179 }
2180
2181 ret = __extent_writepage_io(BTRFS_I(inode), page, wbc, bio_ctrl, i_size,
2182 &nr);
2183 if (ret == 1)
2184 return 0;
2185
2186done:
2187 if (nr == 0) {
2188 /* make sure the mapping tag for page dirty gets cleared */
2189 set_page_writeback(page);
2190 end_page_writeback(page);
2191 }
2192 /*
2193 * Here we used to have a check for PageError() and then set @ret and
2194 * call end_extent_writepage().
2195 *
2196 * But in fact setting @ret here will cause different error paths
2197 * between subpage and regular sectorsize.
2198 *
2199 * For regular page size, we never submit current page, but only add
2200 * current page to current bio.
2201 * The bio submission can only happen in next page.
2202 * Thus if we hit the PageError() branch, @ret is already set to
2203 * non-zero value and will not get updated for regular sectorsize.
2204 *
2205 * But for subpage case, it's possible we submit part of current page,
2206 * thus can get PageError() set by submitted bio of the same page,
2207 * while our @ret is still 0.
2208 *
2209 * So here we unify the behavior and don't set @ret.
2210 * Error can still be properly passed to higher layer as page will
2211 * be set error, here we just don't handle the IO failure.
2212 *
2213 * NOTE: This is just a hotfix for subpage.
2214 * The root fix will be properly ending ordered extent when we hit
2215 * an error during writeback.
2216 *
2217 * But that needs a bigger refactoring, as we not only need to grab the
2218 * submitted OE, but also need to know exactly at which bytenr we hit
2219 * the error.
2220 * Currently the full page based __extent_writepage_io() is not
2221 * capable of that.
2222 */
2223 if (PageError(page))
2224 end_extent_writepage(page, ret, page_start, page_end);
2225 if (bio_ctrl->extent_locked) {
2226 /*
2227 * If bio_ctrl->extent_locked, it's from extent_write_locked_range(),
2228 * the page can either be locked by lock_page() or
2229 * process_one_page().
2230 * Let btrfs_page_unlock_writer() handle both cases.
2231 */
2232 ASSERT(wbc);
2233 btrfs_page_unlock_writer(fs_info, page, wbc->range_start,
2234 wbc->range_end + 1 - wbc->range_start);
2235 } else {
2236 unlock_page(page);
2237 }
2238 ASSERT(ret <= 0);
2239 return ret;
2240}
2241
2242void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
2243{
2244 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
2245 TASK_UNINTERRUPTIBLE);
2246}
2247
2248static void end_extent_buffer_writeback(struct extent_buffer *eb)
2249{
2250 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
2251 smp_mb__after_atomic();
2252 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
2253}
2254
2255/*
2256 * Lock extent buffer status and pages for writeback.
2257 *
2258 * May try to flush write bio if we can't get the lock.
2259 *
2260 * Return 0 if the extent buffer doesn't need to be submitted.
2261 * (E.g. the extent buffer is not dirty)
2262 * Return >0 is the extent buffer is submitted to bio.
2263 * Return <0 if something went wrong, no page is locked.
2264 */
2265static noinline_for_stack int lock_extent_buffer_for_io(struct extent_buffer *eb,
2266 struct btrfs_bio_ctrl *bio_ctrl)
2267{
2268 struct btrfs_fs_info *fs_info = eb->fs_info;
2269 int i, num_pages;
2270 int flush = 0;
2271 int ret = 0;
2272
2273 if (!btrfs_try_tree_write_lock(eb)) {
2274 submit_write_bio(bio_ctrl, 0);
2275 flush = 1;
2276 btrfs_tree_lock(eb);
2277 }
2278
2279 if (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
2280 btrfs_tree_unlock(eb);
2281 if (!bio_ctrl->sync_io)
2282 return 0;
2283 if (!flush) {
2284 submit_write_bio(bio_ctrl, 0);
2285 flush = 1;
2286 }
2287 while (1) {
2288 wait_on_extent_buffer_writeback(eb);
2289 btrfs_tree_lock(eb);
2290 if (!test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags))
2291 break;
2292 btrfs_tree_unlock(eb);
2293 }
2294 }
2295
2296 /*
2297 * We need to do this to prevent races in people who check if the eb is
2298 * under IO since we can end up having no IO bits set for a short period
2299 * of time.
2300 */
2301 spin_lock(&eb->refs_lock);
2302 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
2303 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
2304 spin_unlock(&eb->refs_lock);
2305 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
2306 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
2307 -eb->len,
2308 fs_info->dirty_metadata_batch);
2309 ret = 1;
2310 } else {
2311 spin_unlock(&eb->refs_lock);
2312 }
2313
2314 btrfs_tree_unlock(eb);
2315
2316 /*
2317 * Either we don't need to submit any tree block, or we're submitting
2318 * subpage eb.
2319 * Subpage metadata doesn't use page locking at all, so we can skip
2320 * the page locking.
2321 */
2322 if (!ret || fs_info->nodesize < PAGE_SIZE)
2323 return ret;
2324
2325 num_pages = num_extent_pages(eb);
2326 for (i = 0; i < num_pages; i++) {
2327 struct page *p = eb->pages[i];
2328
2329 if (!trylock_page(p)) {
2330 if (!flush) {
2331 submit_write_bio(bio_ctrl, 0);
2332 flush = 1;
2333 }
2334 lock_page(p);
2335 }
2336 }
2337
2338 return ret;
2339}
2340
2341static void set_btree_ioerr(struct page *page, struct extent_buffer *eb)
2342{
2343 struct btrfs_fs_info *fs_info = eb->fs_info;
2344
2345 btrfs_page_set_error(fs_info, page, eb->start, eb->len);
2346 if (test_and_set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
2347 return;
2348
2349 /*
2350 * A read may stumble upon this buffer later, make sure that it gets an
2351 * error and knows there was an error.
2352 */
2353 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
2354
2355 /*
2356 * We need to set the mapping with the io error as well because a write
2357 * error will flip the file system readonly, and then syncfs() will
2358 * return a 0 because we are readonly if we don't modify the err seq for
2359 * the superblock.
2360 */
2361 mapping_set_error(page->mapping, -EIO);
2362
2363 /*
2364 * If we error out, we should add back the dirty_metadata_bytes
2365 * to make it consistent.
2366 */
2367 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
2368 eb->len, fs_info->dirty_metadata_batch);
2369
2370 /*
2371 * If writeback for a btree extent that doesn't belong to a log tree
2372 * failed, increment the counter transaction->eb_write_errors.
2373 * We do this because while the transaction is running and before it's
2374 * committing (when we call filemap_fdata[write|wait]_range against
2375 * the btree inode), we might have
2376 * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
2377 * returns an error or an error happens during writeback, when we're
2378 * committing the transaction we wouldn't know about it, since the pages
2379 * can be no longer dirty nor marked anymore for writeback (if a
2380 * subsequent modification to the extent buffer didn't happen before the
2381 * transaction commit), which makes filemap_fdata[write|wait]_range not
2382 * able to find the pages tagged with SetPageError at transaction
2383 * commit time. So if this happens we must abort the transaction,
2384 * otherwise we commit a super block with btree roots that point to
2385 * btree nodes/leafs whose content on disk is invalid - either garbage
2386 * or the content of some node/leaf from a past generation that got
2387 * cowed or deleted and is no longer valid.
2388 *
2389 * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
2390 * not be enough - we need to distinguish between log tree extents vs
2391 * non-log tree extents, and the next filemap_fdatawait_range() call
2392 * will catch and clear such errors in the mapping - and that call might
2393 * be from a log sync and not from a transaction commit. Also, checking
2394 * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
2395 * not done and would not be reliable - the eb might have been released
2396 * from memory and reading it back again means that flag would not be
2397 * set (since it's a runtime flag, not persisted on disk).
2398 *
2399 * Using the flags below in the btree inode also makes us achieve the
2400 * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
2401 * writeback for all dirty pages and before filemap_fdatawait_range()
2402 * is called, the writeback for all dirty pages had already finished
2403 * with errors - because we were not using AS_EIO/AS_ENOSPC,
2404 * filemap_fdatawait_range() would return success, as it could not know
2405 * that writeback errors happened (the pages were no longer tagged for
2406 * writeback).
2407 */
2408 switch (eb->log_index) {
2409 case -1:
2410 set_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags);
2411 break;
2412 case 0:
2413 set_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
2414 break;
2415 case 1:
2416 set_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
2417 break;
2418 default:
2419 BUG(); /* unexpected, logic error */
2420 }
2421}
2422
2423/*
2424 * The endio specific version which won't touch any unsafe spinlock in endio
2425 * context.
2426 */
2427static struct extent_buffer *find_extent_buffer_nolock(
2428 struct btrfs_fs_info *fs_info, u64 start)
2429{
2430 struct extent_buffer *eb;
2431
2432 rcu_read_lock();
2433 eb = radix_tree_lookup(&fs_info->buffer_radix,
2434 start >> fs_info->sectorsize_bits);
2435 if (eb && atomic_inc_not_zero(&eb->refs)) {
2436 rcu_read_unlock();
2437 return eb;
2438 }
2439 rcu_read_unlock();
2440 return NULL;
2441}
2442
2443/*
2444 * The endio function for subpage extent buffer write.
2445 *
2446 * Unlike end_bio_extent_buffer_writepage(), we only call end_page_writeback()
2447 * after all extent buffers in the page has finished their writeback.
2448 */
2449static void end_bio_subpage_eb_writepage(struct btrfs_bio *bbio)
2450{
2451 struct bio *bio = &bbio->bio;
2452 struct btrfs_fs_info *fs_info;
2453 struct bio_vec *bvec;
2454 struct bvec_iter_all iter_all;
2455
2456 fs_info = btrfs_sb(bio_first_page_all(bio)->mapping->host->i_sb);
2457 ASSERT(fs_info->nodesize < PAGE_SIZE);
2458
2459 ASSERT(!bio_flagged(bio, BIO_CLONED));
2460 bio_for_each_segment_all(bvec, bio, iter_all) {
2461 struct page *page = bvec->bv_page;
2462 u64 bvec_start = page_offset(page) + bvec->bv_offset;
2463 u64 bvec_end = bvec_start + bvec->bv_len - 1;
2464 u64 cur_bytenr = bvec_start;
2465
2466 ASSERT(IS_ALIGNED(bvec->bv_len, fs_info->nodesize));
2467
2468 /* Iterate through all extent buffers in the range */
2469 while (cur_bytenr <= bvec_end) {
2470 struct extent_buffer *eb;
2471 int done;
2472
2473 /*
2474 * Here we can't use find_extent_buffer(), as it may
2475 * try to lock eb->refs_lock, which is not safe in endio
2476 * context.
2477 */
2478 eb = find_extent_buffer_nolock(fs_info, cur_bytenr);
2479 ASSERT(eb);
2480
2481 cur_bytenr = eb->start + eb->len;
2482
2483 ASSERT(test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags));
2484 done = atomic_dec_and_test(&eb->io_pages);
2485 ASSERT(done);
2486
2487 if (bio->bi_status ||
2488 test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
2489 ClearPageUptodate(page);
2490 set_btree_ioerr(page, eb);
2491 }
2492
2493 btrfs_subpage_clear_writeback(fs_info, page, eb->start,
2494 eb->len);
2495 end_extent_buffer_writeback(eb);
2496 /*
2497 * free_extent_buffer() will grab spinlock which is not
2498 * safe in endio context. Thus here we manually dec
2499 * the ref.
2500 */
2501 atomic_dec(&eb->refs);
2502 }
2503 }
2504 bio_put(bio);
2505}
2506
2507static void end_bio_extent_buffer_writepage(struct btrfs_bio *bbio)
2508{
2509 struct bio *bio = &bbio->bio;
2510 struct bio_vec *bvec;
2511 struct extent_buffer *eb;
2512 int done;
2513 struct bvec_iter_all iter_all;
2514
2515 ASSERT(!bio_flagged(bio, BIO_CLONED));
2516 bio_for_each_segment_all(bvec, bio, iter_all) {
2517 struct page *page = bvec->bv_page;
2518
2519 eb = (struct extent_buffer *)page->private;
2520 BUG_ON(!eb);
2521 done = atomic_dec_and_test(&eb->io_pages);
2522
2523 if (bio->bi_status ||
2524 test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)) {
2525 ClearPageUptodate(page);
2526 set_btree_ioerr(page, eb);
2527 }
2528
2529 end_page_writeback(page);
2530
2531 if (!done)
2532 continue;
2533
2534 end_extent_buffer_writeback(eb);
2535 }
2536
2537 bio_put(bio);
2538}
2539
2540static void prepare_eb_write(struct extent_buffer *eb)
2541{
2542 u32 nritems;
2543 unsigned long start;
2544 unsigned long end;
2545
2546 clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
2547 atomic_set(&eb->io_pages, num_extent_pages(eb));
2548
2549 /* Set btree blocks beyond nritems with 0 to avoid stale content */
2550 nritems = btrfs_header_nritems(eb);
2551 if (btrfs_header_level(eb) > 0) {
2552 end = btrfs_node_key_ptr_offset(eb, nritems);
2553 memzero_extent_buffer(eb, end, eb->len - end);
2554 } else {
2555 /*
2556 * Leaf:
2557 * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
2558 */
2559 start = btrfs_item_nr_offset(eb, nritems);
2560 end = btrfs_item_nr_offset(eb, 0);
2561 if (nritems == 0)
2562 end += BTRFS_LEAF_DATA_SIZE(eb->fs_info);
2563 else
2564 end += btrfs_item_offset(eb, nritems - 1);
2565 memzero_extent_buffer(eb, start, end - start);
2566 }
2567}
2568
2569/*
2570 * Unlike the work in write_one_eb(), we rely completely on extent locking.
2571 * Page locking is only utilized at minimum to keep the VMM code happy.
2572 */
2573static int write_one_subpage_eb(struct extent_buffer *eb,
2574 struct writeback_control *wbc,
2575 struct btrfs_bio_ctrl *bio_ctrl)
2576{
2577 struct btrfs_fs_info *fs_info = eb->fs_info;
2578 struct page *page = eb->pages[0];
2579 blk_opf_t write_flags = wbc_to_write_flags(wbc);
2580 bool no_dirty_ebs = false;
2581 int ret;
2582
2583 prepare_eb_write(eb);
2584
2585 /* clear_page_dirty_for_io() in subpage helper needs page locked */
2586 lock_page(page);
2587 btrfs_subpage_set_writeback(fs_info, page, eb->start, eb->len);
2588
2589 /* Check if this is the last dirty bit to update nr_written */
2590 no_dirty_ebs = btrfs_subpage_clear_and_test_dirty(fs_info, page,
2591 eb->start, eb->len);
2592 if (no_dirty_ebs)
2593 clear_page_dirty_for_io(page);
2594
2595 bio_ctrl->end_io_func = end_bio_subpage_eb_writepage;
2596
2597 ret = submit_extent_page(REQ_OP_WRITE | write_flags, wbc,
2598 bio_ctrl, eb->start, page, eb->len,
2599 eb->start - page_offset(page), 0, false);
2600 if (ret) {
2601 btrfs_subpage_clear_writeback(fs_info, page, eb->start, eb->len);
2602 set_btree_ioerr(page, eb);
2603 unlock_page(page);
2604
2605 if (atomic_dec_and_test(&eb->io_pages))
2606 end_extent_buffer_writeback(eb);
2607 return -EIO;
2608 }
2609 unlock_page(page);
2610 /*
2611 * Submission finished without problem, if no range of the page is
2612 * dirty anymore, we have submitted a page. Update nr_written in wbc.
2613 */
2614 if (no_dirty_ebs)
2615 wbc->nr_to_write--;
2616 return ret;
2617}
2618
2619static noinline_for_stack int write_one_eb(struct extent_buffer *eb,
2620 struct writeback_control *wbc,
2621 struct btrfs_bio_ctrl *bio_ctrl)
2622{
2623 u64 disk_bytenr = eb->start;
2624 int i, num_pages;
2625 blk_opf_t write_flags = wbc_to_write_flags(wbc);
2626 int ret = 0;
2627
2628 prepare_eb_write(eb);
2629
2630 bio_ctrl->end_io_func = end_bio_extent_buffer_writepage;
2631
2632 num_pages = num_extent_pages(eb);
2633 for (i = 0; i < num_pages; i++) {
2634 struct page *p = eb->pages[i];
2635
2636 clear_page_dirty_for_io(p);
2637 set_page_writeback(p);
2638 ret = submit_extent_page(REQ_OP_WRITE | write_flags, wbc,
2639 bio_ctrl, disk_bytenr, p,
2640 PAGE_SIZE, 0, 0, false);
2641 if (ret) {
2642 set_btree_ioerr(p, eb);
2643 if (PageWriteback(p))
2644 end_page_writeback(p);
2645 if (atomic_sub_and_test(num_pages - i, &eb->io_pages))
2646 end_extent_buffer_writeback(eb);
2647 ret = -EIO;
2648 break;
2649 }
2650 disk_bytenr += PAGE_SIZE;
2651 wbc->nr_to_write--;
2652 unlock_page(p);
2653 }
2654
2655 if (unlikely(ret)) {
2656 for (; i < num_pages; i++) {
2657 struct page *p = eb->pages[i];
2658 clear_page_dirty_for_io(p);
2659 unlock_page(p);
2660 }
2661 }
2662
2663 return ret;
2664}
2665
2666/*
2667 * Submit one subpage btree page.
2668 *
2669 * The main difference to submit_eb_page() is:
2670 * - Page locking
2671 * For subpage, we don't rely on page locking at all.
2672 *
2673 * - Flush write bio
2674 * We only flush bio if we may be unable to fit current extent buffers into
2675 * current bio.
2676 *
2677 * Return >=0 for the number of submitted extent buffers.
2678 * Return <0 for fatal error.
2679 */
2680static int submit_eb_subpage(struct page *page,
2681 struct writeback_control *wbc,
2682 struct btrfs_bio_ctrl *bio_ctrl)
2683{
2684 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
2685 int submitted = 0;
2686 u64 page_start = page_offset(page);
2687 int bit_start = 0;
2688 int sectors_per_node = fs_info->nodesize >> fs_info->sectorsize_bits;
2689 int ret;
2690
2691 /* Lock and write each dirty extent buffers in the range */
2692 while (bit_start < fs_info->subpage_info->bitmap_nr_bits) {
2693 struct btrfs_subpage *subpage = (struct btrfs_subpage *)page->private;
2694 struct extent_buffer *eb;
2695 unsigned long flags;
2696 u64 start;
2697
2698 /*
2699 * Take private lock to ensure the subpage won't be detached
2700 * in the meantime.
2701 */
2702 spin_lock(&page->mapping->private_lock);
2703 if (!PagePrivate(page)) {
2704 spin_unlock(&page->mapping->private_lock);
2705 break;
2706 }
2707 spin_lock_irqsave(&subpage->lock, flags);
2708 if (!test_bit(bit_start + fs_info->subpage_info->dirty_offset,
2709 subpage->bitmaps)) {
2710 spin_unlock_irqrestore(&subpage->lock, flags);
2711 spin_unlock(&page->mapping->private_lock);
2712 bit_start++;
2713 continue;
2714 }
2715
2716 start = page_start + bit_start * fs_info->sectorsize;
2717 bit_start += sectors_per_node;
2718
2719 /*
2720 * Here we just want to grab the eb without touching extra
2721 * spin locks, so call find_extent_buffer_nolock().
2722 */
2723 eb = find_extent_buffer_nolock(fs_info, start);
2724 spin_unlock_irqrestore(&subpage->lock, flags);
2725 spin_unlock(&page->mapping->private_lock);
2726
2727 /*
2728 * The eb has already reached 0 refs thus find_extent_buffer()
2729 * doesn't return it. We don't need to write back such eb
2730 * anyway.
2731 */
2732 if (!eb)
2733 continue;
2734
2735 ret = lock_extent_buffer_for_io(eb, bio_ctrl);
2736 if (ret == 0) {
2737 free_extent_buffer(eb);
2738 continue;
2739 }
2740 if (ret < 0) {
2741 free_extent_buffer(eb);
2742 goto cleanup;
2743 }
2744 ret = write_one_subpage_eb(eb, wbc, bio_ctrl);
2745 free_extent_buffer(eb);
2746 if (ret < 0)
2747 goto cleanup;
2748 submitted++;
2749 }
2750 return submitted;
2751
2752cleanup:
2753 /* We hit error, end bio for the submitted extent buffers */
2754 submit_write_bio(bio_ctrl, ret);
2755 return ret;
2756}
2757
2758/*
2759 * Submit all page(s) of one extent buffer.
2760 *
2761 * @page: the page of one extent buffer
2762 * @eb_context: to determine if we need to submit this page, if current page
2763 * belongs to this eb, we don't need to submit
2764 *
2765 * The caller should pass each page in their bytenr order, and here we use
2766 * @eb_context to determine if we have submitted pages of one extent buffer.
2767 *
2768 * If we have, we just skip until we hit a new page that doesn't belong to
2769 * current @eb_context.
2770 *
2771 * If not, we submit all the page(s) of the extent buffer.
2772 *
2773 * Return >0 if we have submitted the extent buffer successfully.
2774 * Return 0 if we don't need to submit the page, as it's already submitted by
2775 * previous call.
2776 * Return <0 for fatal error.
2777 */
2778static int submit_eb_page(struct page *page, struct writeback_control *wbc,
2779 struct btrfs_bio_ctrl *bio_ctrl,
2780 struct extent_buffer **eb_context)
2781{
2782 struct address_space *mapping = page->mapping;
2783 struct btrfs_block_group *cache = NULL;
2784 struct extent_buffer *eb;
2785 int ret;
2786
2787 if (!PagePrivate(page))
2788 return 0;
2789
2790 if (btrfs_sb(page->mapping->host->i_sb)->nodesize < PAGE_SIZE)
2791 return submit_eb_subpage(page, wbc, bio_ctrl);
2792
2793 spin_lock(&mapping->private_lock);
2794 if (!PagePrivate(page)) {
2795 spin_unlock(&mapping->private_lock);
2796 return 0;
2797 }
2798
2799 eb = (struct extent_buffer *)page->private;
2800
2801 /*
2802 * Shouldn't happen and normally this would be a BUG_ON but no point
2803 * crashing the machine for something we can survive anyway.
2804 */
2805 if (WARN_ON(!eb)) {
2806 spin_unlock(&mapping->private_lock);
2807 return 0;
2808 }
2809
2810 if (eb == *eb_context) {
2811 spin_unlock(&mapping->private_lock);
2812 return 0;
2813 }
2814 ret = atomic_inc_not_zero(&eb->refs);
2815 spin_unlock(&mapping->private_lock);
2816 if (!ret)
2817 return 0;
2818
2819 if (!btrfs_check_meta_write_pointer(eb->fs_info, eb, &cache)) {
2820 /*
2821 * If for_sync, this hole will be filled with
2822 * trasnsaction commit.
2823 */
2824 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
2825 ret = -EAGAIN;
2826 else
2827 ret = 0;
2828 free_extent_buffer(eb);
2829 return ret;
2830 }
2831
2832 *eb_context = eb;
2833
2834 ret = lock_extent_buffer_for_io(eb, bio_ctrl);
2835 if (ret <= 0) {
2836 btrfs_revert_meta_write_pointer(cache, eb);
2837 if (cache)
2838 btrfs_put_block_group(cache);
2839 free_extent_buffer(eb);
2840 return ret;
2841 }
2842 if (cache) {
2843 /*
2844 * Implies write in zoned mode. Mark the last eb in a block group.
2845 */
2846 btrfs_schedule_zone_finish_bg(cache, eb);
2847 btrfs_put_block_group(cache);
2848 }
2849 ret = write_one_eb(eb, wbc, bio_ctrl);
2850 free_extent_buffer(eb);
2851 if (ret < 0)
2852 return ret;
2853 return 1;
2854}
2855
2856int btree_write_cache_pages(struct address_space *mapping,
2857 struct writeback_control *wbc)
2858{
2859 struct extent_buffer *eb_context = NULL;
2860 struct btrfs_bio_ctrl bio_ctrl = {
2861 .extent_locked = 0,
2862 .sync_io = (wbc->sync_mode == WB_SYNC_ALL),
2863 };
2864 struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
2865 int ret = 0;
2866 int done = 0;
2867 int nr_to_write_done = 0;
2868 struct pagevec pvec;
2869 int nr_pages;
2870 pgoff_t index;
2871 pgoff_t end; /* Inclusive */
2872 int scanned = 0;
2873 xa_mark_t tag;
2874
2875 pagevec_init(&pvec);
2876 if (wbc->range_cyclic) {
2877 index = mapping->writeback_index; /* Start from prev offset */
2878 end = -1;
2879 /*
2880 * Start from the beginning does not need to cycle over the
2881 * range, mark it as scanned.
2882 */
2883 scanned = (index == 0);
2884 } else {
2885 index = wbc->range_start >> PAGE_SHIFT;
2886 end = wbc->range_end >> PAGE_SHIFT;
2887 scanned = 1;
2888 }
2889 if (wbc->sync_mode == WB_SYNC_ALL)
2890 tag = PAGECACHE_TAG_TOWRITE;
2891 else
2892 tag = PAGECACHE_TAG_DIRTY;
2893 btrfs_zoned_meta_io_lock(fs_info);
2894retry:
2895 if (wbc->sync_mode == WB_SYNC_ALL)
2896 tag_pages_for_writeback(mapping, index, end);
2897 while (!done && !nr_to_write_done && (index <= end) &&
2898 (nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2899 tag))) {
2900 unsigned i;
2901
2902 for (i = 0; i < nr_pages; i++) {
2903 struct page *page = pvec.pages[i];
2904
2905 ret = submit_eb_page(page, wbc, &bio_ctrl, &eb_context);
2906 if (ret == 0)
2907 continue;
2908 if (ret < 0) {
2909 done = 1;
2910 break;
2911 }
2912
2913 /*
2914 * the filesystem may choose to bump up nr_to_write.
2915 * We have to make sure to honor the new nr_to_write
2916 * at any time
2917 */
2918 nr_to_write_done = wbc->nr_to_write <= 0;
2919 }
2920 pagevec_release(&pvec);
2921 cond_resched();
2922 }
2923 if (!scanned && !done) {
2924 /*
2925 * We hit the last page and there is more work to be done: wrap
2926 * back to the start of the file
2927 */
2928 scanned = 1;
2929 index = 0;
2930 goto retry;
2931 }
2932 /*
2933 * If something went wrong, don't allow any metadata write bio to be
2934 * submitted.
2935 *
2936 * This would prevent use-after-free if we had dirty pages not
2937 * cleaned up, which can still happen by fuzzed images.
2938 *
2939 * - Bad extent tree
2940 * Allowing existing tree block to be allocated for other trees.
2941 *
2942 * - Log tree operations
2943 * Exiting tree blocks get allocated to log tree, bumps its
2944 * generation, then get cleaned in tree re-balance.
2945 * Such tree block will not be written back, since it's clean,
2946 * thus no WRITTEN flag set.
2947 * And after log writes back, this tree block is not traced by
2948 * any dirty extent_io_tree.
2949 *
2950 * - Offending tree block gets re-dirtied from its original owner
2951 * Since it has bumped generation, no WRITTEN flag, it can be
2952 * reused without COWing. This tree block will not be traced
2953 * by btrfs_transaction::dirty_pages.
2954 *
2955 * Now such dirty tree block will not be cleaned by any dirty
2956 * extent io tree. Thus we don't want to submit such wild eb
2957 * if the fs already has error.
2958 *
2959 * We can get ret > 0 from submit_extent_page() indicating how many ebs
2960 * were submitted. Reset it to 0 to avoid false alerts for the caller.
2961 */
2962 if (ret > 0)
2963 ret = 0;
2964 if (!ret && BTRFS_FS_ERROR(fs_info))
2965 ret = -EROFS;
2966 submit_write_bio(&bio_ctrl, ret);
2967
2968 btrfs_zoned_meta_io_unlock(fs_info);
2969 return ret;
2970}
2971
2972/*
2973 * Walk the list of dirty pages of the given address space and write all of them.
2974 *
2975 * @mapping: address space structure to write
2976 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2977 * @bio_ctrl: holds context for the write, namely the bio
2978 *
2979 * If a page is already under I/O, write_cache_pages() skips it, even
2980 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2981 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2982 * and msync() need to guarantee that all the data which was dirty at the time
2983 * the call was made get new I/O started against them. If wbc->sync_mode is
2984 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2985 * existing IO to complete.
2986 */
2987static int extent_write_cache_pages(struct address_space *mapping,
2988 struct writeback_control *wbc,
2989 struct btrfs_bio_ctrl *bio_ctrl)
2990{
2991 struct inode *inode = mapping->host;
2992 int ret = 0;
2993 int done = 0;
2994 int nr_to_write_done = 0;
2995 struct pagevec pvec;
2996 int nr_pages;
2997 pgoff_t index;
2998 pgoff_t end; /* Inclusive */
2999 pgoff_t done_index;
3000 int range_whole = 0;
3001 int scanned = 0;
3002 xa_mark_t tag;
3003
3004 /*
3005 * We have to hold onto the inode so that ordered extents can do their
3006 * work when the IO finishes. The alternative to this is failing to add
3007 * an ordered extent if the igrab() fails there and that is a huge pain
3008 * to deal with, so instead just hold onto the inode throughout the
3009 * writepages operation. If it fails here we are freeing up the inode
3010 * anyway and we'd rather not waste our time writing out stuff that is
3011 * going to be truncated anyway.
3012 */
3013 if (!igrab(inode))
3014 return 0;
3015
3016 pagevec_init(&pvec);
3017 if (wbc->range_cyclic) {
3018 index = mapping->writeback_index; /* Start from prev offset */
3019 end = -1;
3020 /*
3021 * Start from the beginning does not need to cycle over the
3022 * range, mark it as scanned.
3023 */
3024 scanned = (index == 0);
3025 } else {
3026 index = wbc->range_start >> PAGE_SHIFT;
3027 end = wbc->range_end >> PAGE_SHIFT;
3028 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
3029 range_whole = 1;
3030 scanned = 1;
3031 }
3032
3033 /*
3034 * We do the tagged writepage as long as the snapshot flush bit is set
3035 * and we are the first one who do the filemap_flush() on this inode.
3036 *
3037 * The nr_to_write == LONG_MAX is needed to make sure other flushers do
3038 * not race in and drop the bit.
3039 */
3040 if (range_whole && wbc->nr_to_write == LONG_MAX &&
3041 test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
3042 &BTRFS_I(inode)->runtime_flags))
3043 wbc->tagged_writepages = 1;
3044
3045 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
3046 tag = PAGECACHE_TAG_TOWRITE;
3047 else
3048 tag = PAGECACHE_TAG_DIRTY;
3049retry:
3050 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
3051 tag_pages_for_writeback(mapping, index, end);
3052 done_index = index;
3053 while (!done && !nr_to_write_done && (index <= end) &&
3054 (nr_pages = pagevec_lookup_range_tag(&pvec, mapping,
3055 &index, end, tag))) {
3056 unsigned i;
3057
3058 for (i = 0; i < nr_pages; i++) {
3059 struct page *page = pvec.pages[i];
3060
3061 done_index = page->index + 1;
3062 /*
3063 * At this point we hold neither the i_pages lock nor
3064 * the page lock: the page may be truncated or
3065 * invalidated (changing page->mapping to NULL),
3066 * or even swizzled back from swapper_space to
3067 * tmpfs file mapping
3068 */
3069 if (!trylock_page(page)) {
3070 submit_write_bio(bio_ctrl, 0);
3071 lock_page(page);
3072 }
3073
3074 if (unlikely(page->mapping != mapping)) {
3075 unlock_page(page);
3076 continue;
3077 }
3078
3079 if (wbc->sync_mode != WB_SYNC_NONE) {
3080 if (PageWriteback(page))
3081 submit_write_bio(bio_ctrl, 0);
3082 wait_on_page_writeback(page);
3083 }
3084
3085 if (PageWriteback(page) ||
3086 !clear_page_dirty_for_io(page)) {
3087 unlock_page(page);
3088 continue;
3089 }
3090
3091 ret = __extent_writepage(page, wbc, bio_ctrl);
3092 if (ret < 0) {
3093 done = 1;
3094 break;
3095 }
3096
3097 /*
3098 * the filesystem may choose to bump up nr_to_write.
3099 * We have to make sure to honor the new nr_to_write
3100 * at any time
3101 */
3102 nr_to_write_done = wbc->nr_to_write <= 0;
3103 }
3104 pagevec_release(&pvec);
3105 cond_resched();
3106 }
3107 if (!scanned && !done) {
3108 /*
3109 * We hit the last page and there is more work to be done: wrap
3110 * back to the start of the file
3111 */
3112 scanned = 1;
3113 index = 0;
3114
3115 /*
3116 * If we're looping we could run into a page that is locked by a
3117 * writer and that writer could be waiting on writeback for a
3118 * page in our current bio, and thus deadlock, so flush the
3119 * write bio here.
3120 */
3121 submit_write_bio(bio_ctrl, 0);
3122 goto retry;
3123 }
3124
3125 if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
3126 mapping->writeback_index = done_index;
3127
3128 btrfs_add_delayed_iput(BTRFS_I(inode));
3129 return ret;
3130}
3131
3132/*
3133 * Submit the pages in the range to bio for call sites which delalloc range has
3134 * already been ran (aka, ordered extent inserted) and all pages are still
3135 * locked.
3136 */
3137int extent_write_locked_range(struct inode *inode, u64 start, u64 end)
3138{
3139 bool found_error = false;
3140 int first_error = 0;
3141 int ret = 0;
3142 struct address_space *mapping = inode->i_mapping;
3143 struct page *page;
3144 u64 cur = start;
3145 unsigned long nr_pages;
3146 const u32 sectorsize = btrfs_sb(inode->i_sb)->sectorsize;
3147 struct btrfs_bio_ctrl bio_ctrl = {
3148 .extent_locked = 1,
3149 .sync_io = 1,
3150 };
3151 struct writeback_control wbc_writepages = {
3152 .sync_mode = WB_SYNC_ALL,
3153 .range_start = start,
3154 .range_end = end + 1,
3155 /* We're called from an async helper function */
3156 .punt_to_cgroup = 1,
3157 .no_cgroup_owner = 1,
3158 };
3159
3160 ASSERT(IS_ALIGNED(start, sectorsize) && IS_ALIGNED(end + 1, sectorsize));
3161 nr_pages = (round_up(end, PAGE_SIZE) - round_down(start, PAGE_SIZE)) >>
3162 PAGE_SHIFT;
3163 wbc_writepages.nr_to_write = nr_pages * 2;
3164
3165 wbc_attach_fdatawrite_inode(&wbc_writepages, inode);
3166 while (cur <= end) {
3167 u64 cur_end = min(round_down(cur, PAGE_SIZE) + PAGE_SIZE - 1, end);
3168
3169 page = find_get_page(mapping, cur >> PAGE_SHIFT);
3170 /*
3171 * All pages in the range are locked since
3172 * btrfs_run_delalloc_range(), thus there is no way to clear
3173 * the page dirty flag.
3174 */
3175 ASSERT(PageLocked(page));
3176 ASSERT(PageDirty(page));
3177 clear_page_dirty_for_io(page);
3178 ret = __extent_writepage(page, &wbc_writepages, &bio_ctrl);
3179 ASSERT(ret <= 0);
3180 if (ret < 0) {
3181 found_error = true;
3182 first_error = ret;
3183 }
3184 put_page(page);
3185 cur = cur_end + 1;
3186 }
3187
3188 submit_write_bio(&bio_ctrl, found_error ? ret : 0);
3189
3190 wbc_detach_inode(&wbc_writepages);
3191 if (found_error)
3192 return first_error;
3193 return ret;
3194}
3195
3196int extent_writepages(struct address_space *mapping,
3197 struct writeback_control *wbc)
3198{
3199 struct inode *inode = mapping->host;
3200 int ret = 0;
3201 struct btrfs_bio_ctrl bio_ctrl = {
3202 .extent_locked = 0,
3203 .sync_io = (wbc->sync_mode == WB_SYNC_ALL),
3204 };
3205
3206 /*
3207 * Allow only a single thread to do the reloc work in zoned mode to
3208 * protect the write pointer updates.
3209 */
3210 btrfs_zoned_data_reloc_lock(BTRFS_I(inode));
3211 ret = extent_write_cache_pages(mapping, wbc, &bio_ctrl);
3212 submit_write_bio(&bio_ctrl, ret);
3213 btrfs_zoned_data_reloc_unlock(BTRFS_I(inode));
3214 return ret;
3215}
3216
3217void extent_readahead(struct readahead_control *rac)
3218{
3219 struct btrfs_bio_ctrl bio_ctrl = { 0 };
3220 struct page *pagepool[16];
3221 struct extent_map *em_cached = NULL;
3222 u64 prev_em_start = (u64)-1;
3223 int nr;
3224
3225 while ((nr = readahead_page_batch(rac, pagepool))) {
3226 u64 contig_start = readahead_pos(rac);
3227 u64 contig_end = contig_start + readahead_batch_length(rac) - 1;
3228
3229 contiguous_readpages(pagepool, nr, contig_start, contig_end,
3230 &em_cached, &bio_ctrl, &prev_em_start);
3231 }
3232
3233 if (em_cached)
3234 free_extent_map(em_cached);
3235 submit_one_bio(&bio_ctrl);
3236}
3237
3238/*
3239 * basic invalidate_folio code, this waits on any locked or writeback
3240 * ranges corresponding to the folio, and then deletes any extent state
3241 * records from the tree
3242 */
3243int extent_invalidate_folio(struct extent_io_tree *tree,
3244 struct folio *folio, size_t offset)
3245{
3246 struct extent_state *cached_state = NULL;
3247 u64 start = folio_pos(folio);
3248 u64 end = start + folio_size(folio) - 1;
3249 size_t blocksize = folio->mapping->host->i_sb->s_blocksize;
3250
3251 /* This function is only called for the btree inode */
3252 ASSERT(tree->owner == IO_TREE_BTREE_INODE_IO);
3253
3254 start += ALIGN(offset, blocksize);
3255 if (start > end)
3256 return 0;
3257
3258 lock_extent(tree, start, end, &cached_state);
3259 folio_wait_writeback(folio);
3260
3261 /*
3262 * Currently for btree io tree, only EXTENT_LOCKED is utilized,
3263 * so here we only need to unlock the extent range to free any
3264 * existing extent state.
3265 */
3266 unlock_extent(tree, start, end, &cached_state);
3267 return 0;
3268}
3269
3270/*
3271 * a helper for release_folio, this tests for areas of the page that
3272 * are locked or under IO and drops the related state bits if it is safe
3273 * to drop the page.
3274 */
3275static int try_release_extent_state(struct extent_io_tree *tree,
3276 struct page *page, gfp_t mask)
3277{
3278 u64 start = page_offset(page);
3279 u64 end = start + PAGE_SIZE - 1;
3280 int ret = 1;
3281
3282 if (test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL)) {
3283 ret = 0;
3284 } else {
3285 u32 clear_bits = ~(EXTENT_LOCKED | EXTENT_NODATASUM |
3286 EXTENT_DELALLOC_NEW | EXTENT_CTLBITS);
3287
3288 /*
3289 * At this point we can safely clear everything except the
3290 * locked bit, the nodatasum bit and the delalloc new bit.
3291 * The delalloc new bit will be cleared by ordered extent
3292 * completion.
3293 */
3294 ret = __clear_extent_bit(tree, start, end, clear_bits, NULL,
3295 mask, NULL);
3296
3297 /* if clear_extent_bit failed for enomem reasons,
3298 * we can't allow the release to continue.
3299 */
3300 if (ret < 0)
3301 ret = 0;
3302 else
3303 ret = 1;
3304 }
3305 return ret;
3306}
3307
3308/*
3309 * a helper for release_folio. As long as there are no locked extents
3310 * in the range corresponding to the page, both state records and extent
3311 * map records are removed
3312 */
3313int try_release_extent_mapping(struct page *page, gfp_t mask)
3314{
3315 struct extent_map *em;
3316 u64 start = page_offset(page);
3317 u64 end = start + PAGE_SIZE - 1;
3318 struct btrfs_inode *btrfs_inode = BTRFS_I(page->mapping->host);
3319 struct extent_io_tree *tree = &btrfs_inode->io_tree;
3320 struct extent_map_tree *map = &btrfs_inode->extent_tree;
3321
3322 if (gfpflags_allow_blocking(mask) &&
3323 page->mapping->host->i_size > SZ_16M) {
3324 u64 len;
3325 while (start <= end) {
3326 struct btrfs_fs_info *fs_info;
3327 u64 cur_gen;
3328
3329 len = end - start + 1;
3330 write_lock(&map->lock);
3331 em = lookup_extent_mapping(map, start, len);
3332 if (!em) {
3333 write_unlock(&map->lock);
3334 break;
3335 }
3336 if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
3337 em->start != start) {
3338 write_unlock(&map->lock);
3339 free_extent_map(em);
3340 break;
3341 }
3342 if (test_range_bit(tree, em->start,
3343 extent_map_end(em) - 1,
3344 EXTENT_LOCKED, 0, NULL))
3345 goto next;
3346 /*
3347 * If it's not in the list of modified extents, used
3348 * by a fast fsync, we can remove it. If it's being
3349 * logged we can safely remove it since fsync took an
3350 * extra reference on the em.
3351 */
3352 if (list_empty(&em->list) ||
3353 test_bit(EXTENT_FLAG_LOGGING, &em->flags))
3354 goto remove_em;
3355 /*
3356 * If it's in the list of modified extents, remove it
3357 * only if its generation is older then the current one,
3358 * in which case we don't need it for a fast fsync.
3359 * Otherwise don't remove it, we could be racing with an
3360 * ongoing fast fsync that could miss the new extent.
3361 */
3362 fs_info = btrfs_inode->root->fs_info;
3363 spin_lock(&fs_info->trans_lock);
3364 cur_gen = fs_info->generation;
3365 spin_unlock(&fs_info->trans_lock);
3366 if (em->generation >= cur_gen)
3367 goto next;
3368remove_em:
3369 /*
3370 * We only remove extent maps that are not in the list of
3371 * modified extents or that are in the list but with a
3372 * generation lower then the current generation, so there
3373 * is no need to set the full fsync flag on the inode (it
3374 * hurts the fsync performance for workloads with a data
3375 * size that exceeds or is close to the system's memory).
3376 */
3377 remove_extent_mapping(map, em);
3378 /* once for the rb tree */
3379 free_extent_map(em);
3380next:
3381 start = extent_map_end(em);
3382 write_unlock(&map->lock);
3383
3384 /* once for us */
3385 free_extent_map(em);
3386
3387 cond_resched(); /* Allow large-extent preemption. */
3388 }
3389 }
3390 return try_release_extent_state(tree, page, mask);
3391}
3392
3393/*
3394 * To cache previous fiemap extent
3395 *
3396 * Will be used for merging fiemap extent
3397 */
3398struct fiemap_cache {
3399 u64 offset;
3400 u64 phys;
3401 u64 len;
3402 u32 flags;
3403 bool cached;
3404};
3405
3406/*
3407 * Helper to submit fiemap extent.
3408 *
3409 * Will try to merge current fiemap extent specified by @offset, @phys,
3410 * @len and @flags with cached one.
3411 * And only when we fails to merge, cached one will be submitted as
3412 * fiemap extent.
3413 *
3414 * Return value is the same as fiemap_fill_next_extent().
3415 */
3416static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
3417 struct fiemap_cache *cache,
3418 u64 offset, u64 phys, u64 len, u32 flags)
3419{
3420 int ret = 0;
3421
3422 /* Set at the end of extent_fiemap(). */
3423 ASSERT((flags & FIEMAP_EXTENT_LAST) == 0);
3424
3425 if (!cache->cached)
3426 goto assign;
3427
3428 /*
3429 * Sanity check, extent_fiemap() should have ensured that new
3430 * fiemap extent won't overlap with cached one.
3431 * Not recoverable.
3432 *
3433 * NOTE: Physical address can overlap, due to compression
3434 */
3435 if (cache->offset + cache->len > offset) {
3436 WARN_ON(1);
3437 return -EINVAL;
3438 }
3439
3440 /*
3441 * Only merges fiemap extents if
3442 * 1) Their logical addresses are continuous
3443 *
3444 * 2) Their physical addresses are continuous
3445 * So truly compressed (physical size smaller than logical size)
3446 * extents won't get merged with each other
3447 *
3448 * 3) Share same flags
3449 */
3450 if (cache->offset + cache->len == offset &&
3451 cache->phys + cache->len == phys &&
3452 cache->flags == flags) {
3453 cache->len += len;
3454 return 0;
3455 }
3456
3457 /* Not mergeable, need to submit cached one */
3458 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
3459 cache->len, cache->flags);
3460 cache->cached = false;
3461 if (ret)
3462 return ret;
3463assign:
3464 cache->cached = true;
3465 cache->offset = offset;
3466 cache->phys = phys;
3467 cache->len = len;
3468 cache->flags = flags;
3469
3470 return 0;
3471}
3472
3473/*
3474 * Emit last fiemap cache
3475 *
3476 * The last fiemap cache may still be cached in the following case:
3477 * 0 4k 8k
3478 * |<- Fiemap range ->|
3479 * |<------------ First extent ----------->|
3480 *
3481 * In this case, the first extent range will be cached but not emitted.
3482 * So we must emit it before ending extent_fiemap().
3483 */
3484static int emit_last_fiemap_cache(struct fiemap_extent_info *fieinfo,
3485 struct fiemap_cache *cache)
3486{
3487 int ret;
3488
3489 if (!cache->cached)
3490 return 0;
3491
3492 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
3493 cache->len, cache->flags);
3494 cache->cached = false;
3495 if (ret > 0)
3496 ret = 0;
3497 return ret;
3498}
3499
3500static int fiemap_next_leaf_item(struct btrfs_inode *inode, struct btrfs_path *path)
3501{
3502 struct extent_buffer *clone;
3503 struct btrfs_key key;
3504 int slot;
3505 int ret;
3506
3507 path->slots[0]++;
3508 if (path->slots[0] < btrfs_header_nritems(path->nodes[0]))
3509 return 0;
3510
3511 ret = btrfs_next_leaf(inode->root, path);
3512 if (ret != 0)
3513 return ret;
3514
3515 /*
3516 * Don't bother with cloning if there are no more file extent items for
3517 * our inode.
3518 */
3519 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3520 if (key.objectid != btrfs_ino(inode) || key.type != BTRFS_EXTENT_DATA_KEY)
3521 return 1;
3522
3523 /* See the comment at fiemap_search_slot() about why we clone. */
3524 clone = btrfs_clone_extent_buffer(path->nodes[0]);
3525 if (!clone)
3526 return -ENOMEM;
3527
3528 slot = path->slots[0];
3529 btrfs_release_path(path);
3530 path->nodes[0] = clone;
3531 path->slots[0] = slot;
3532
3533 return 0;
3534}
3535
3536/*
3537 * Search for the first file extent item that starts at a given file offset or
3538 * the one that starts immediately before that offset.
3539 * Returns: 0 on success, < 0 on error, 1 if not found.
3540 */
3541static int fiemap_search_slot(struct btrfs_inode *inode, struct btrfs_path *path,
3542 u64 file_offset)
3543{
3544 const u64 ino = btrfs_ino(inode);
3545 struct btrfs_root *root = inode->root;
3546 struct extent_buffer *clone;
3547 struct btrfs_key key;
3548 int slot;
3549 int ret;
3550
3551 key.objectid = ino;
3552 key.type = BTRFS_EXTENT_DATA_KEY;
3553 key.offset = file_offset;
3554
3555 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3556 if (ret < 0)
3557 return ret;
3558
3559 if (ret > 0 && path->slots[0] > 0) {
3560 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
3561 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
3562 path->slots[0]--;
3563 }
3564
3565 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
3566 ret = btrfs_next_leaf(root, path);
3567 if (ret != 0)
3568 return ret;
3569
3570 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3571 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3572 return 1;
3573 }
3574
3575 /*
3576 * We clone the leaf and use it during fiemap. This is because while
3577 * using the leaf we do expensive things like checking if an extent is
3578 * shared, which can take a long time. In order to prevent blocking
3579 * other tasks for too long, we use a clone of the leaf. We have locked
3580 * the file range in the inode's io tree, so we know none of our file
3581 * extent items can change. This way we avoid blocking other tasks that
3582 * want to insert items for other inodes in the same leaf or b+tree
3583 * rebalance operations (triggered for example when someone is trying
3584 * to push items into this leaf when trying to insert an item in a
3585 * neighbour leaf).
3586 * We also need the private clone because holding a read lock on an
3587 * extent buffer of the subvolume's b+tree will make lockdep unhappy
3588 * when we call fiemap_fill_next_extent(), because that may cause a page
3589 * fault when filling the user space buffer with fiemap data.
3590 */
3591 clone = btrfs_clone_extent_buffer(path->nodes[0]);
3592 if (!clone)
3593 return -ENOMEM;
3594
3595 slot = path->slots[0];
3596 btrfs_release_path(path);
3597 path->nodes[0] = clone;
3598 path->slots[0] = slot;
3599
3600 return 0;
3601}
3602
3603/*
3604 * Process a range which is a hole or a prealloc extent in the inode's subvolume
3605 * btree. If @disk_bytenr is 0, we are dealing with a hole, otherwise a prealloc
3606 * extent. The end offset (@end) is inclusive.
3607 */
3608static int fiemap_process_hole(struct btrfs_inode *inode,
3609 struct fiemap_extent_info *fieinfo,
3610 struct fiemap_cache *cache,
3611 struct extent_state **delalloc_cached_state,
3612 struct btrfs_backref_share_check_ctx *backref_ctx,
3613 u64 disk_bytenr, u64 extent_offset,
3614 u64 extent_gen,
3615 u64 start, u64 end)
3616{
3617 const u64 i_size = i_size_read(&inode->vfs_inode);
3618 u64 cur_offset = start;
3619 u64 last_delalloc_end = 0;
3620 u32 prealloc_flags = FIEMAP_EXTENT_UNWRITTEN;
3621 bool checked_extent_shared = false;
3622 int ret;
3623
3624 /*
3625 * There can be no delalloc past i_size, so don't waste time looking for
3626 * it beyond i_size.
3627 */
3628 while (cur_offset < end && cur_offset < i_size) {
3629 u64 delalloc_start;
3630 u64 delalloc_end;
3631 u64 prealloc_start;
3632 u64 prealloc_len = 0;
3633 bool delalloc;
3634
3635 delalloc = btrfs_find_delalloc_in_range(inode, cur_offset, end,
3636 delalloc_cached_state,
3637 &delalloc_start,
3638 &delalloc_end);
3639 if (!delalloc)
3640 break;
3641
3642 /*
3643 * If this is a prealloc extent we have to report every section
3644 * of it that has no delalloc.
3645 */
3646 if (disk_bytenr != 0) {
3647 if (last_delalloc_end == 0) {
3648 prealloc_start = start;
3649 prealloc_len = delalloc_start - start;
3650 } else {
3651 prealloc_start = last_delalloc_end + 1;
3652 prealloc_len = delalloc_start - prealloc_start;
3653 }
3654 }
3655
3656 if (prealloc_len > 0) {
3657 if (!checked_extent_shared && fieinfo->fi_extents_max) {
3658 ret = btrfs_is_data_extent_shared(inode,
3659 disk_bytenr,
3660 extent_gen,
3661 backref_ctx);
3662 if (ret < 0)
3663 return ret;
3664 else if (ret > 0)
3665 prealloc_flags |= FIEMAP_EXTENT_SHARED;
3666
3667 checked_extent_shared = true;
3668 }
3669 ret = emit_fiemap_extent(fieinfo, cache, prealloc_start,
3670 disk_bytenr + extent_offset,
3671 prealloc_len, prealloc_flags);
3672 if (ret)
3673 return ret;
3674 extent_offset += prealloc_len;
3675 }
3676
3677 ret = emit_fiemap_extent(fieinfo, cache, delalloc_start, 0,
3678 delalloc_end + 1 - delalloc_start,
3679 FIEMAP_EXTENT_DELALLOC |
3680 FIEMAP_EXTENT_UNKNOWN);
3681 if (ret)
3682 return ret;
3683
3684 last_delalloc_end = delalloc_end;
3685 cur_offset = delalloc_end + 1;
3686 extent_offset += cur_offset - delalloc_start;
3687 cond_resched();
3688 }
3689
3690 /*
3691 * Either we found no delalloc for the whole prealloc extent or we have
3692 * a prealloc extent that spans i_size or starts at or after i_size.
3693 */
3694 if (disk_bytenr != 0 && last_delalloc_end < end) {
3695 u64 prealloc_start;
3696 u64 prealloc_len;
3697
3698 if (last_delalloc_end == 0) {
3699 prealloc_start = start;
3700 prealloc_len = end + 1 - start;
3701 } else {
3702 prealloc_start = last_delalloc_end + 1;
3703 prealloc_len = end + 1 - prealloc_start;
3704 }
3705
3706 if (!checked_extent_shared && fieinfo->fi_extents_max) {
3707 ret = btrfs_is_data_extent_shared(inode,
3708 disk_bytenr,
3709 extent_gen,
3710 backref_ctx);
3711 if (ret < 0)
3712 return ret;
3713 else if (ret > 0)
3714 prealloc_flags |= FIEMAP_EXTENT_SHARED;
3715 }
3716 ret = emit_fiemap_extent(fieinfo, cache, prealloc_start,
3717 disk_bytenr + extent_offset,
3718 prealloc_len, prealloc_flags);
3719 if (ret)
3720 return ret;
3721 }
3722
3723 return 0;
3724}
3725
3726static int fiemap_find_last_extent_offset(struct btrfs_inode *inode,
3727 struct btrfs_path *path,
3728 u64 *last_extent_end_ret)
3729{
3730 const u64 ino = btrfs_ino(inode);
3731 struct btrfs_root *root = inode->root;
3732 struct extent_buffer *leaf;
3733 struct btrfs_file_extent_item *ei;
3734 struct btrfs_key key;
3735 u64 disk_bytenr;
3736 int ret;
3737
3738 /*
3739 * Lookup the last file extent. We're not using i_size here because
3740 * there might be preallocation past i_size.
3741 */
3742 ret = btrfs_lookup_file_extent(NULL, root, path, ino, (u64)-1, 0);
3743 /* There can't be a file extent item at offset (u64)-1 */
3744 ASSERT(ret != 0);
3745 if (ret < 0)
3746 return ret;
3747
3748 /*
3749 * For a non-existing key, btrfs_search_slot() always leaves us at a
3750 * slot > 0, except if the btree is empty, which is impossible because
3751 * at least it has the inode item for this inode and all the items for
3752 * the root inode 256.
3753 */
3754 ASSERT(path->slots[0] > 0);
3755 path->slots[0]--;
3756 leaf = path->nodes[0];
3757 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3758 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY) {
3759 /* No file extent items in the subvolume tree. */
3760 *last_extent_end_ret = 0;
3761 return 0;
3762 }
3763
3764 /*
3765 * For an inline extent, the disk_bytenr is where inline data starts at,
3766 * so first check if we have an inline extent item before checking if we
3767 * have an implicit hole (disk_bytenr == 0).
3768 */
3769 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
3770 if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_INLINE) {
3771 *last_extent_end_ret = btrfs_file_extent_end(path);
3772 return 0;
3773 }
3774
3775 /*
3776 * Find the last file extent item that is not a hole (when NO_HOLES is
3777 * not enabled). This should take at most 2 iterations in the worst
3778 * case: we have one hole file extent item at slot 0 of a leaf and
3779 * another hole file extent item as the last item in the previous leaf.
3780 * This is because we merge file extent items that represent holes.
3781 */
3782 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
3783 while (disk_bytenr == 0) {
3784 ret = btrfs_previous_item(root, path, ino, BTRFS_EXTENT_DATA_KEY);
3785 if (ret < 0) {
3786 return ret;
3787 } else if (ret > 0) {
3788 /* No file extent items that are not holes. */
3789 *last_extent_end_ret = 0;
3790 return 0;
3791 }
3792 leaf = path->nodes[0];
3793 ei = btrfs_item_ptr(leaf, path->slots[0],
3794 struct btrfs_file_extent_item);
3795 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
3796 }
3797
3798 *last_extent_end_ret = btrfs_file_extent_end(path);
3799 return 0;
3800}
3801
3802int extent_fiemap(struct btrfs_inode *inode, struct fiemap_extent_info *fieinfo,
3803 u64 start, u64 len)
3804{
3805 const u64 ino = btrfs_ino(inode);
3806 struct extent_state *cached_state = NULL;
3807 struct extent_state *delalloc_cached_state = NULL;
3808 struct btrfs_path *path;
3809 struct fiemap_cache cache = { 0 };
3810 struct btrfs_backref_share_check_ctx *backref_ctx;
3811 u64 last_extent_end;
3812 u64 prev_extent_end;
3813 u64 lockstart;
3814 u64 lockend;
3815 bool stopped = false;
3816 int ret;
3817
3818 backref_ctx = btrfs_alloc_backref_share_check_ctx();
3819 path = btrfs_alloc_path();
3820 if (!backref_ctx || !path) {
3821 ret = -ENOMEM;
3822 goto out;
3823 }
3824
3825 lockstart = round_down(start, inode->root->fs_info->sectorsize);
3826 lockend = round_up(start + len, inode->root->fs_info->sectorsize);
3827 prev_extent_end = lockstart;
3828
3829 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3830 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3831
3832 ret = fiemap_find_last_extent_offset(inode, path, &last_extent_end);
3833 if (ret < 0)
3834 goto out_unlock;
3835 btrfs_release_path(path);
3836
3837 path->reada = READA_FORWARD;
3838 ret = fiemap_search_slot(inode, path, lockstart);
3839 if (ret < 0) {
3840 goto out_unlock;
3841 } else if (ret > 0) {
3842 /*
3843 * No file extent item found, but we may have delalloc between
3844 * the current offset and i_size. So check for that.
3845 */
3846 ret = 0;
3847 goto check_eof_delalloc;
3848 }
3849
3850 while (prev_extent_end < lockend) {
3851 struct extent_buffer *leaf = path->nodes[0];
3852 struct btrfs_file_extent_item *ei;
3853 struct btrfs_key key;
3854 u64 extent_end;
3855 u64 extent_len;
3856 u64 extent_offset = 0;
3857 u64 extent_gen;
3858 u64 disk_bytenr = 0;
3859 u64 flags = 0;
3860 int extent_type;
3861 u8 compression;
3862
3863 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3864 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3865 break;
3866
3867 extent_end = btrfs_file_extent_end(path);
3868
3869 /*
3870 * The first iteration can leave us at an extent item that ends
3871 * before our range's start. Move to the next item.
3872 */
3873 if (extent_end <= lockstart)
3874 goto next_item;
3875
3876 backref_ctx->curr_leaf_bytenr = leaf->start;
3877
3878 /* We have in implicit hole (NO_HOLES feature enabled). */
3879 if (prev_extent_end < key.offset) {
3880 const u64 range_end = min(key.offset, lockend) - 1;
3881
3882 ret = fiemap_process_hole(inode, fieinfo, &cache,
3883 &delalloc_cached_state,
3884 backref_ctx, 0, 0, 0,
3885 prev_extent_end, range_end);
3886 if (ret < 0) {
3887 goto out_unlock;
3888 } else if (ret > 0) {
3889 /* fiemap_fill_next_extent() told us to stop. */
3890 stopped = true;
3891 break;
3892 }
3893
3894 /* We've reached the end of the fiemap range, stop. */
3895 if (key.offset >= lockend) {
3896 stopped = true;
3897 break;
3898 }
3899 }
3900
3901 extent_len = extent_end - key.offset;
3902 ei = btrfs_item_ptr(leaf, path->slots[0],
3903 struct btrfs_file_extent_item);
3904 compression = btrfs_file_extent_compression(leaf, ei);
3905 extent_type = btrfs_file_extent_type(leaf, ei);
3906 extent_gen = btrfs_file_extent_generation(leaf, ei);
3907
3908 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
3909 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
3910 if (compression == BTRFS_COMPRESS_NONE)
3911 extent_offset = btrfs_file_extent_offset(leaf, ei);
3912 }
3913
3914 if (compression != BTRFS_COMPRESS_NONE)
3915 flags |= FIEMAP_EXTENT_ENCODED;
3916
3917 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
3918 flags |= FIEMAP_EXTENT_DATA_INLINE;
3919 flags |= FIEMAP_EXTENT_NOT_ALIGNED;
3920 ret = emit_fiemap_extent(fieinfo, &cache, key.offset, 0,
3921 extent_len, flags);
3922 } else if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
3923 ret = fiemap_process_hole(inode, fieinfo, &cache,
3924 &delalloc_cached_state,
3925 backref_ctx,
3926 disk_bytenr, extent_offset,
3927 extent_gen, key.offset,
3928 extent_end - 1);
3929 } else if (disk_bytenr == 0) {
3930 /* We have an explicit hole. */
3931 ret = fiemap_process_hole(inode, fieinfo, &cache,
3932 &delalloc_cached_state,
3933 backref_ctx, 0, 0, 0,
3934 key.offset, extent_end - 1);
3935 } else {
3936 /* We have a regular extent. */
3937 if (fieinfo->fi_extents_max) {
3938 ret = btrfs_is_data_extent_shared(inode,
3939 disk_bytenr,
3940 extent_gen,
3941 backref_ctx);
3942 if (ret < 0)
3943 goto out_unlock;
3944 else if (ret > 0)
3945 flags |= FIEMAP_EXTENT_SHARED;
3946 }
3947
3948 ret = emit_fiemap_extent(fieinfo, &cache, key.offset,
3949 disk_bytenr + extent_offset,
3950 extent_len, flags);
3951 }
3952
3953 if (ret < 0) {
3954 goto out_unlock;
3955 } else if (ret > 0) {
3956 /* fiemap_fill_next_extent() told us to stop. */
3957 stopped = true;
3958 break;
3959 }
3960
3961 prev_extent_end = extent_end;
3962next_item:
3963 if (fatal_signal_pending(current)) {
3964 ret = -EINTR;
3965 goto out_unlock;
3966 }
3967
3968 ret = fiemap_next_leaf_item(inode, path);
3969 if (ret < 0) {
3970 goto out_unlock;
3971 } else if (ret > 0) {
3972 /* No more file extent items for this inode. */
3973 break;
3974 }
3975 cond_resched();
3976 }
3977
3978check_eof_delalloc:
3979 /*
3980 * Release (and free) the path before emitting any final entries to
3981 * fiemap_fill_next_extent() to keep lockdep happy. This is because
3982 * once we find no more file extent items exist, we may have a
3983 * non-cloned leaf, and fiemap_fill_next_extent() can trigger page
3984 * faults when copying data to the user space buffer.
3985 */
3986 btrfs_free_path(path);
3987 path = NULL;
3988
3989 if (!stopped && prev_extent_end < lockend) {
3990 ret = fiemap_process_hole(inode, fieinfo, &cache,
3991 &delalloc_cached_state, backref_ctx,
3992 0, 0, 0, prev_extent_end, lockend - 1);
3993 if (ret < 0)
3994 goto out_unlock;
3995 prev_extent_end = lockend;
3996 }
3997
3998 if (cache.cached && cache.offset + cache.len >= last_extent_end) {
3999 const u64 i_size = i_size_read(&inode->vfs_inode);
4000
4001 if (prev_extent_end < i_size) {
4002 u64 delalloc_start;
4003 u64 delalloc_end;
4004 bool delalloc;
4005
4006 delalloc = btrfs_find_delalloc_in_range(inode,
4007 prev_extent_end,
4008 i_size - 1,
4009 &delalloc_cached_state,
4010 &delalloc_start,
4011 &delalloc_end);
4012 if (!delalloc)
4013 cache.flags |= FIEMAP_EXTENT_LAST;
4014 } else {
4015 cache.flags |= FIEMAP_EXTENT_LAST;
4016 }
4017 }
4018
4019 ret = emit_last_fiemap_cache(fieinfo, &cache);
4020
4021out_unlock:
4022 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
4023 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
4024out:
4025 free_extent_state(delalloc_cached_state);
4026 btrfs_free_backref_share_ctx(backref_ctx);
4027 btrfs_free_path(path);
4028 return ret;
4029}
4030
4031static void __free_extent_buffer(struct extent_buffer *eb)
4032{
4033 kmem_cache_free(extent_buffer_cache, eb);
4034}
4035
4036int extent_buffer_under_io(const struct extent_buffer *eb)
4037{
4038 return (atomic_read(&eb->io_pages) ||
4039 test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
4040 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4041}
4042
4043static bool page_range_has_eb(struct btrfs_fs_info *fs_info, struct page *page)
4044{
4045 struct btrfs_subpage *subpage;
4046
4047 lockdep_assert_held(&page->mapping->private_lock);
4048
4049 if (PagePrivate(page)) {
4050 subpage = (struct btrfs_subpage *)page->private;
4051 if (atomic_read(&subpage->eb_refs))
4052 return true;
4053 /*
4054 * Even there is no eb refs here, we may still have
4055 * end_page_read() call relying on page::private.
4056 */
4057 if (atomic_read(&subpage->readers))
4058 return true;
4059 }
4060 return false;
4061}
4062
4063static void detach_extent_buffer_page(struct extent_buffer *eb, struct page *page)
4064{
4065 struct btrfs_fs_info *fs_info = eb->fs_info;
4066 const bool mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
4067
4068 /*
4069 * For mapped eb, we're going to change the page private, which should
4070 * be done under the private_lock.
4071 */
4072 if (mapped)
4073 spin_lock(&page->mapping->private_lock);
4074
4075 if (!PagePrivate(page)) {
4076 if (mapped)
4077 spin_unlock(&page->mapping->private_lock);
4078 return;
4079 }
4080
4081 if (fs_info->nodesize >= PAGE_SIZE) {
4082 /*
4083 * We do this since we'll remove the pages after we've
4084 * removed the eb from the radix tree, so we could race
4085 * and have this page now attached to the new eb. So
4086 * only clear page_private if it's still connected to
4087 * this eb.
4088 */
4089 if (PagePrivate(page) &&
4090 page->private == (unsigned long)eb) {
4091 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
4092 BUG_ON(PageDirty(page));
4093 BUG_ON(PageWriteback(page));
4094 /*
4095 * We need to make sure we haven't be attached
4096 * to a new eb.
4097 */
4098 detach_page_private(page);
4099 }
4100 if (mapped)
4101 spin_unlock(&page->mapping->private_lock);
4102 return;
4103 }
4104
4105 /*
4106 * For subpage, we can have dummy eb with page private. In this case,
4107 * we can directly detach the private as such page is only attached to
4108 * one dummy eb, no sharing.
4109 */
4110 if (!mapped) {
4111 btrfs_detach_subpage(fs_info, page);
4112 return;
4113 }
4114
4115 btrfs_page_dec_eb_refs(fs_info, page);
4116
4117 /*
4118 * We can only detach the page private if there are no other ebs in the
4119 * page range and no unfinished IO.
4120 */
4121 if (!page_range_has_eb(fs_info, page))
4122 btrfs_detach_subpage(fs_info, page);
4123
4124 spin_unlock(&page->mapping->private_lock);
4125}
4126
4127/* Release all pages attached to the extent buffer */
4128static void btrfs_release_extent_buffer_pages(struct extent_buffer *eb)
4129{
4130 int i;
4131 int num_pages;
4132
4133 ASSERT(!extent_buffer_under_io(eb));
4134
4135 num_pages = num_extent_pages(eb);
4136 for (i = 0; i < num_pages; i++) {
4137 struct page *page = eb->pages[i];
4138
4139 if (!page)
4140 continue;
4141
4142 detach_extent_buffer_page(eb, page);
4143
4144 /* One for when we allocated the page */
4145 put_page(page);
4146 }
4147}
4148
4149/*
4150 * Helper for releasing the extent buffer.
4151 */
4152static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
4153{
4154 btrfs_release_extent_buffer_pages(eb);
4155 btrfs_leak_debug_del_eb(eb);
4156 __free_extent_buffer(eb);
4157}
4158
4159static struct extent_buffer *
4160__alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
4161 unsigned long len)
4162{
4163 struct extent_buffer *eb = NULL;
4164
4165 eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
4166 eb->start = start;
4167 eb->len = len;
4168 eb->fs_info = fs_info;
4169 init_rwsem(&eb->lock);
4170
4171 btrfs_leak_debug_add_eb(eb);
4172 INIT_LIST_HEAD(&eb->release_list);
4173
4174 spin_lock_init(&eb->refs_lock);
4175 atomic_set(&eb->refs, 1);
4176 atomic_set(&eb->io_pages, 0);
4177
4178 ASSERT(len <= BTRFS_MAX_METADATA_BLOCKSIZE);
4179
4180 return eb;
4181}
4182
4183struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
4184{
4185 int i;
4186 struct extent_buffer *new;
4187 int num_pages = num_extent_pages(src);
4188 int ret;
4189
4190 new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
4191 if (new == NULL)
4192 return NULL;
4193
4194 /*
4195 * Set UNMAPPED before calling btrfs_release_extent_buffer(), as
4196 * btrfs_release_extent_buffer() have different behavior for
4197 * UNMAPPED subpage extent buffer.
4198 */
4199 set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
4200
4201 ret = btrfs_alloc_page_array(num_pages, new->pages);
4202 if (ret) {
4203 btrfs_release_extent_buffer(new);
4204 return NULL;
4205 }
4206
4207 for (i = 0; i < num_pages; i++) {
4208 int ret;
4209 struct page *p = new->pages[i];
4210
4211 ret = attach_extent_buffer_page(new, p, NULL);
4212 if (ret < 0) {
4213 btrfs_release_extent_buffer(new);
4214 return NULL;
4215 }
4216 WARN_ON(PageDirty(p));
4217 copy_page(page_address(p), page_address(src->pages[i]));
4218 }
4219 set_extent_buffer_uptodate(new);
4220
4221 return new;
4222}
4223
4224struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4225 u64 start, unsigned long len)
4226{
4227 struct extent_buffer *eb;
4228 int num_pages;
4229 int i;
4230 int ret;
4231
4232 eb = __alloc_extent_buffer(fs_info, start, len);
4233 if (!eb)
4234 return NULL;
4235
4236 num_pages = num_extent_pages(eb);
4237 ret = btrfs_alloc_page_array(num_pages, eb->pages);
4238 if (ret)
4239 goto err;
4240
4241 for (i = 0; i < num_pages; i++) {
4242 struct page *p = eb->pages[i];
4243
4244 ret = attach_extent_buffer_page(eb, p, NULL);
4245 if (ret < 0)
4246 goto err;
4247 }
4248
4249 set_extent_buffer_uptodate(eb);
4250 btrfs_set_header_nritems(eb, 0);
4251 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
4252
4253 return eb;
4254err:
4255 for (i = 0; i < num_pages; i++) {
4256 if (eb->pages[i]) {
4257 detach_extent_buffer_page(eb, eb->pages[i]);
4258 __free_page(eb->pages[i]);
4259 }
4260 }
4261 __free_extent_buffer(eb);
4262 return NULL;
4263}
4264
4265struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
4266 u64 start)
4267{
4268 return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
4269}
4270
4271static void check_buffer_tree_ref(struct extent_buffer *eb)
4272{
4273 int refs;
4274 /*
4275 * The TREE_REF bit is first set when the extent_buffer is added
4276 * to the radix tree. It is also reset, if unset, when a new reference
4277 * is created by find_extent_buffer.
4278 *
4279 * It is only cleared in two cases: freeing the last non-tree
4280 * reference to the extent_buffer when its STALE bit is set or
4281 * calling release_folio when the tree reference is the only reference.
4282 *
4283 * In both cases, care is taken to ensure that the extent_buffer's
4284 * pages are not under io. However, release_folio can be concurrently
4285 * called with creating new references, which is prone to race
4286 * conditions between the calls to check_buffer_tree_ref in those
4287 * codepaths and clearing TREE_REF in try_release_extent_buffer.
4288 *
4289 * The actual lifetime of the extent_buffer in the radix tree is
4290 * adequately protected by the refcount, but the TREE_REF bit and
4291 * its corresponding reference are not. To protect against this
4292 * class of races, we call check_buffer_tree_ref from the codepaths
4293 * which trigger io after they set eb->io_pages. Note that once io is
4294 * initiated, TREE_REF can no longer be cleared, so that is the
4295 * moment at which any such race is best fixed.
4296 */
4297 refs = atomic_read(&eb->refs);
4298 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4299 return;
4300
4301 spin_lock(&eb->refs_lock);
4302 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4303 atomic_inc(&eb->refs);
4304 spin_unlock(&eb->refs_lock);
4305}
4306
4307static void mark_extent_buffer_accessed(struct extent_buffer *eb,
4308 struct page *accessed)
4309{
4310 int num_pages, i;
4311
4312 check_buffer_tree_ref(eb);
4313
4314 num_pages = num_extent_pages(eb);
4315 for (i = 0; i < num_pages; i++) {
4316 struct page *p = eb->pages[i];
4317
4318 if (p != accessed)
4319 mark_page_accessed(p);
4320 }
4321}
4322
4323struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
4324 u64 start)
4325{
4326 struct extent_buffer *eb;
4327
4328 eb = find_extent_buffer_nolock(fs_info, start);
4329 if (!eb)
4330 return NULL;
4331 /*
4332 * Lock our eb's refs_lock to avoid races with free_extent_buffer().
4333 * When we get our eb it might be flagged with EXTENT_BUFFER_STALE and
4334 * another task running free_extent_buffer() might have seen that flag
4335 * set, eb->refs == 2, that the buffer isn't under IO (dirty and
4336 * writeback flags not set) and it's still in the tree (flag
4337 * EXTENT_BUFFER_TREE_REF set), therefore being in the process of
4338 * decrementing the extent buffer's reference count twice. So here we
4339 * could race and increment the eb's reference count, clear its stale
4340 * flag, mark it as dirty and drop our reference before the other task
4341 * finishes executing free_extent_buffer, which would later result in
4342 * an attempt to free an extent buffer that is dirty.
4343 */
4344 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
4345 spin_lock(&eb->refs_lock);
4346 spin_unlock(&eb->refs_lock);
4347 }
4348 mark_extent_buffer_accessed(eb, NULL);
4349 return eb;
4350}
4351
4352#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4353struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
4354 u64 start)
4355{
4356 struct extent_buffer *eb, *exists = NULL;
4357 int ret;
4358
4359 eb = find_extent_buffer(fs_info, start);
4360 if (eb)
4361 return eb;
4362 eb = alloc_dummy_extent_buffer(fs_info, start);
4363 if (!eb)
4364 return ERR_PTR(-ENOMEM);
4365 eb->fs_info = fs_info;
4366again:
4367 ret = radix_tree_preload(GFP_NOFS);
4368 if (ret) {
4369 exists = ERR_PTR(ret);
4370 goto free_eb;
4371 }
4372 spin_lock(&fs_info->buffer_lock);
4373 ret = radix_tree_insert(&fs_info->buffer_radix,
4374 start >> fs_info->sectorsize_bits, eb);
4375 spin_unlock(&fs_info->buffer_lock);
4376 radix_tree_preload_end();
4377 if (ret == -EEXIST) {
4378 exists = find_extent_buffer(fs_info, start);
4379 if (exists)
4380 goto free_eb;
4381 else
4382 goto again;
4383 }
4384 check_buffer_tree_ref(eb);
4385 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4386
4387 return eb;
4388free_eb:
4389 btrfs_release_extent_buffer(eb);
4390 return exists;
4391}
4392#endif
4393
4394static struct extent_buffer *grab_extent_buffer(
4395 struct btrfs_fs_info *fs_info, struct page *page)
4396{
4397 struct extent_buffer *exists;
4398
4399 /*
4400 * For subpage case, we completely rely on radix tree to ensure we
4401 * don't try to insert two ebs for the same bytenr. So here we always
4402 * return NULL and just continue.
4403 */
4404 if (fs_info->nodesize < PAGE_SIZE)
4405 return NULL;
4406
4407 /* Page not yet attached to an extent buffer */
4408 if (!PagePrivate(page))
4409 return NULL;
4410
4411 /*
4412 * We could have already allocated an eb for this page and attached one
4413 * so lets see if we can get a ref on the existing eb, and if we can we
4414 * know it's good and we can just return that one, else we know we can
4415 * just overwrite page->private.
4416 */
4417 exists = (struct extent_buffer *)page->private;
4418 if (atomic_inc_not_zero(&exists->refs))
4419 return exists;
4420
4421 WARN_ON(PageDirty(page));
4422 detach_page_private(page);
4423 return NULL;
4424}
4425
4426static int check_eb_alignment(struct btrfs_fs_info *fs_info, u64 start)
4427{
4428 if (!IS_ALIGNED(start, fs_info->sectorsize)) {
4429 btrfs_err(fs_info, "bad tree block start %llu", start);
4430 return -EINVAL;
4431 }
4432
4433 if (fs_info->nodesize < PAGE_SIZE &&
4434 offset_in_page(start) + fs_info->nodesize > PAGE_SIZE) {
4435 btrfs_err(fs_info,
4436 "tree block crosses page boundary, start %llu nodesize %u",
4437 start, fs_info->nodesize);
4438 return -EINVAL;
4439 }
4440 if (fs_info->nodesize >= PAGE_SIZE &&
4441 !PAGE_ALIGNED(start)) {
4442 btrfs_err(fs_info,
4443 "tree block is not page aligned, start %llu nodesize %u",
4444 start, fs_info->nodesize);
4445 return -EINVAL;
4446 }
4447 return 0;
4448}
4449
4450struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
4451 u64 start, u64 owner_root, int level)
4452{
4453 unsigned long len = fs_info->nodesize;
4454 int num_pages;
4455 int i;
4456 unsigned long index = start >> PAGE_SHIFT;
4457 struct extent_buffer *eb;
4458 struct extent_buffer *exists = NULL;
4459 struct page *p;
4460 struct address_space *mapping = fs_info->btree_inode->i_mapping;
4461 u64 lockdep_owner = owner_root;
4462 int uptodate = 1;
4463 int ret;
4464
4465 if (check_eb_alignment(fs_info, start))
4466 return ERR_PTR(-EINVAL);
4467
4468#if BITS_PER_LONG == 32
4469 if (start >= MAX_LFS_FILESIZE) {
4470 btrfs_err_rl(fs_info,
4471 "extent buffer %llu is beyond 32bit page cache limit", start);
4472 btrfs_err_32bit_limit(fs_info);
4473 return ERR_PTR(-EOVERFLOW);
4474 }
4475 if (start >= BTRFS_32BIT_EARLY_WARN_THRESHOLD)
4476 btrfs_warn_32bit_limit(fs_info);
4477#endif
4478
4479 eb = find_extent_buffer(fs_info, start);
4480 if (eb)
4481 return eb;
4482
4483 eb = __alloc_extent_buffer(fs_info, start, len);
4484 if (!eb)
4485 return ERR_PTR(-ENOMEM);
4486
4487 /*
4488 * The reloc trees are just snapshots, so we need them to appear to be
4489 * just like any other fs tree WRT lockdep.
4490 */
4491 if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID)
4492 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
4493
4494 btrfs_set_buffer_lockdep_class(lockdep_owner, eb, level);
4495
4496 num_pages = num_extent_pages(eb);
4497 for (i = 0; i < num_pages; i++, index++) {
4498 struct btrfs_subpage *prealloc = NULL;
4499
4500 p = find_or_create_page(mapping, index, GFP_NOFS|__GFP_NOFAIL);
4501 if (!p) {
4502 exists = ERR_PTR(-ENOMEM);
4503 goto free_eb;
4504 }
4505
4506 /*
4507 * Preallocate page->private for subpage case, so that we won't
4508 * allocate memory with private_lock hold. The memory will be
4509 * freed by attach_extent_buffer_page() or freed manually if
4510 * we exit earlier.
4511 *
4512 * Although we have ensured one subpage eb can only have one
4513 * page, but it may change in the future for 16K page size
4514 * support, so we still preallocate the memory in the loop.
4515 */
4516 if (fs_info->nodesize < PAGE_SIZE) {
4517 prealloc = btrfs_alloc_subpage(fs_info, BTRFS_SUBPAGE_METADATA);
4518 if (IS_ERR(prealloc)) {
4519 ret = PTR_ERR(prealloc);
4520 unlock_page(p);
4521 put_page(p);
4522 exists = ERR_PTR(ret);
4523 goto free_eb;
4524 }
4525 }
4526
4527 spin_lock(&mapping->private_lock);
4528 exists = grab_extent_buffer(fs_info, p);
4529 if (exists) {
4530 spin_unlock(&mapping->private_lock);
4531 unlock_page(p);
4532 put_page(p);
4533 mark_extent_buffer_accessed(exists, p);
4534 btrfs_free_subpage(prealloc);
4535 goto free_eb;
4536 }
4537 /* Should not fail, as we have preallocated the memory */
4538 ret = attach_extent_buffer_page(eb, p, prealloc);
4539 ASSERT(!ret);
4540 /*
4541 * To inform we have extra eb under allocation, so that
4542 * detach_extent_buffer_page() won't release the page private
4543 * when the eb hasn't yet been inserted into radix tree.
4544 *
4545 * The ref will be decreased when the eb released the page, in
4546 * detach_extent_buffer_page().
4547 * Thus needs no special handling in error path.
4548 */
4549 btrfs_page_inc_eb_refs(fs_info, p);
4550 spin_unlock(&mapping->private_lock);
4551
4552 WARN_ON(btrfs_page_test_dirty(fs_info, p, eb->start, eb->len));
4553 eb->pages[i] = p;
4554 if (!PageUptodate(p))
4555 uptodate = 0;
4556
4557 /*
4558 * We can't unlock the pages just yet since the extent buffer
4559 * hasn't been properly inserted in the radix tree, this
4560 * opens a race with btree_release_folio which can free a page
4561 * while we are still filling in all pages for the buffer and
4562 * we could crash.
4563 */
4564 }
4565 if (uptodate)
4566 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4567again:
4568 ret = radix_tree_preload(GFP_NOFS);
4569 if (ret) {
4570 exists = ERR_PTR(ret);
4571 goto free_eb;
4572 }
4573
4574 spin_lock(&fs_info->buffer_lock);
4575 ret = radix_tree_insert(&fs_info->buffer_radix,
4576 start >> fs_info->sectorsize_bits, eb);
4577 spin_unlock(&fs_info->buffer_lock);
4578 radix_tree_preload_end();
4579 if (ret == -EEXIST) {
4580 exists = find_extent_buffer(fs_info, start);
4581 if (exists)
4582 goto free_eb;
4583 else
4584 goto again;
4585 }
4586 /* add one reference for the tree */
4587 check_buffer_tree_ref(eb);
4588 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
4589
4590 /*
4591 * Now it's safe to unlock the pages because any calls to
4592 * btree_release_folio will correctly detect that a page belongs to a
4593 * live buffer and won't free them prematurely.
4594 */
4595 for (i = 0; i < num_pages; i++)
4596 unlock_page(eb->pages[i]);
4597 return eb;
4598
4599free_eb:
4600 WARN_ON(!atomic_dec_and_test(&eb->refs));
4601 for (i = 0; i < num_pages; i++) {
4602 if (eb->pages[i])
4603 unlock_page(eb->pages[i]);
4604 }
4605
4606 btrfs_release_extent_buffer(eb);
4607 return exists;
4608}
4609
4610static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
4611{
4612 struct extent_buffer *eb =
4613 container_of(head, struct extent_buffer, rcu_head);
4614
4615 __free_extent_buffer(eb);
4616}
4617
4618static int release_extent_buffer(struct extent_buffer *eb)
4619 __releases(&eb->refs_lock)
4620{
4621 lockdep_assert_held(&eb->refs_lock);
4622
4623 WARN_ON(atomic_read(&eb->refs) == 0);
4624 if (atomic_dec_and_test(&eb->refs)) {
4625 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
4626 struct btrfs_fs_info *fs_info = eb->fs_info;
4627
4628 spin_unlock(&eb->refs_lock);
4629
4630 spin_lock(&fs_info->buffer_lock);
4631 radix_tree_delete(&fs_info->buffer_radix,
4632 eb->start >> fs_info->sectorsize_bits);
4633 spin_unlock(&fs_info->buffer_lock);
4634 } else {
4635 spin_unlock(&eb->refs_lock);
4636 }
4637
4638 btrfs_leak_debug_del_eb(eb);
4639 /* Should be safe to release our pages at this point */
4640 btrfs_release_extent_buffer_pages(eb);
4641#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4642 if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
4643 __free_extent_buffer(eb);
4644 return 1;
4645 }
4646#endif
4647 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
4648 return 1;
4649 }
4650 spin_unlock(&eb->refs_lock);
4651
4652 return 0;
4653}
4654
4655void free_extent_buffer(struct extent_buffer *eb)
4656{
4657 int refs;
4658 if (!eb)
4659 return;
4660
4661 refs = atomic_read(&eb->refs);
4662 while (1) {
4663 if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
4664 || (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
4665 refs == 1))
4666 break;
4667 if (atomic_try_cmpxchg(&eb->refs, &refs, refs - 1))
4668 return;
4669 }
4670
4671 spin_lock(&eb->refs_lock);
4672 if (atomic_read(&eb->refs) == 2 &&
4673 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
4674 !extent_buffer_under_io(eb) &&
4675 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4676 atomic_dec(&eb->refs);
4677
4678 /*
4679 * I know this is terrible, but it's temporary until we stop tracking
4680 * the uptodate bits and such for the extent buffers.
4681 */
4682 release_extent_buffer(eb);
4683}
4684
4685void free_extent_buffer_stale(struct extent_buffer *eb)
4686{
4687 if (!eb)
4688 return;
4689
4690 spin_lock(&eb->refs_lock);
4691 set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
4692
4693 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
4694 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
4695 atomic_dec(&eb->refs);
4696 release_extent_buffer(eb);
4697}
4698
4699static void btree_clear_page_dirty(struct page *page)
4700{
4701 ASSERT(PageDirty(page));
4702 ASSERT(PageLocked(page));
4703 clear_page_dirty_for_io(page);
4704 xa_lock_irq(&page->mapping->i_pages);
4705 if (!PageDirty(page))
4706 __xa_clear_mark(&page->mapping->i_pages,
4707 page_index(page), PAGECACHE_TAG_DIRTY);
4708 xa_unlock_irq(&page->mapping->i_pages);
4709}
4710
4711static void clear_subpage_extent_buffer_dirty(const struct extent_buffer *eb)
4712{
4713 struct btrfs_fs_info *fs_info = eb->fs_info;
4714 struct page *page = eb->pages[0];
4715 bool last;
4716
4717 /* btree_clear_page_dirty() needs page locked */
4718 lock_page(page);
4719 last = btrfs_subpage_clear_and_test_dirty(fs_info, page, eb->start,
4720 eb->len);
4721 if (last)
4722 btree_clear_page_dirty(page);
4723 unlock_page(page);
4724 WARN_ON(atomic_read(&eb->refs) == 0);
4725}
4726
4727void clear_extent_buffer_dirty(const struct extent_buffer *eb)
4728{
4729 int i;
4730 int num_pages;
4731 struct page *page;
4732
4733 if (eb->fs_info->nodesize < PAGE_SIZE)
4734 return clear_subpage_extent_buffer_dirty(eb);
4735
4736 num_pages = num_extent_pages(eb);
4737
4738 for (i = 0; i < num_pages; i++) {
4739 page = eb->pages[i];
4740 if (!PageDirty(page))
4741 continue;
4742 lock_page(page);
4743 btree_clear_page_dirty(page);
4744 ClearPageError(page);
4745 unlock_page(page);
4746 }
4747 WARN_ON(atomic_read(&eb->refs) == 0);
4748}
4749
4750bool set_extent_buffer_dirty(struct extent_buffer *eb)
4751{
4752 int i;
4753 int num_pages;
4754 bool was_dirty;
4755
4756 check_buffer_tree_ref(eb);
4757
4758 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4759
4760 num_pages = num_extent_pages(eb);
4761 WARN_ON(atomic_read(&eb->refs) == 0);
4762 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4763
4764 if (!was_dirty) {
4765 bool subpage = eb->fs_info->nodesize < PAGE_SIZE;
4766
4767 /*
4768 * For subpage case, we can have other extent buffers in the
4769 * same page, and in clear_subpage_extent_buffer_dirty() we
4770 * have to clear page dirty without subpage lock held.
4771 * This can cause race where our page gets dirty cleared after
4772 * we just set it.
4773 *
4774 * Thankfully, clear_subpage_extent_buffer_dirty() has locked
4775 * its page for other reasons, we can use page lock to prevent
4776 * the above race.
4777 */
4778 if (subpage)
4779 lock_page(eb->pages[0]);
4780 for (i = 0; i < num_pages; i++)
4781 btrfs_page_set_dirty(eb->fs_info, eb->pages[i],
4782 eb->start, eb->len);
4783 if (subpage)
4784 unlock_page(eb->pages[0]);
4785 }
4786#ifdef CONFIG_BTRFS_DEBUG
4787 for (i = 0; i < num_pages; i++)
4788 ASSERT(PageDirty(eb->pages[i]));
4789#endif
4790
4791 return was_dirty;
4792}
4793
4794void clear_extent_buffer_uptodate(struct extent_buffer *eb)
4795{
4796 struct btrfs_fs_info *fs_info = eb->fs_info;
4797 struct page *page;
4798 int num_pages;
4799 int i;
4800
4801 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4802 num_pages = num_extent_pages(eb);
4803 for (i = 0; i < num_pages; i++) {
4804 page = eb->pages[i];
4805 if (!page)
4806 continue;
4807
4808 /*
4809 * This is special handling for metadata subpage, as regular
4810 * btrfs_is_subpage() can not handle cloned/dummy metadata.
4811 */
4812 if (fs_info->nodesize >= PAGE_SIZE)
4813 ClearPageUptodate(page);
4814 else
4815 btrfs_subpage_clear_uptodate(fs_info, page, eb->start,
4816 eb->len);
4817 }
4818}
4819
4820void set_extent_buffer_uptodate(struct extent_buffer *eb)
4821{
4822 struct btrfs_fs_info *fs_info = eb->fs_info;
4823 struct page *page;
4824 int num_pages;
4825 int i;
4826
4827 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4828 num_pages = num_extent_pages(eb);
4829 for (i = 0; i < num_pages; i++) {
4830 page = eb->pages[i];
4831
4832 /*
4833 * This is special handling for metadata subpage, as regular
4834 * btrfs_is_subpage() can not handle cloned/dummy metadata.
4835 */
4836 if (fs_info->nodesize >= PAGE_SIZE)
4837 SetPageUptodate(page);
4838 else
4839 btrfs_subpage_set_uptodate(fs_info, page, eb->start,
4840 eb->len);
4841 }
4842}
4843
4844static int read_extent_buffer_subpage(struct extent_buffer *eb, int wait,
4845 int mirror_num,
4846 struct btrfs_tree_parent_check *check)
4847{
4848 struct btrfs_fs_info *fs_info = eb->fs_info;
4849 struct extent_io_tree *io_tree;
4850 struct page *page = eb->pages[0];
4851 struct extent_state *cached_state = NULL;
4852 struct btrfs_bio_ctrl bio_ctrl = {
4853 .mirror_num = mirror_num,
4854 .parent_check = check,
4855 };
4856 int ret = 0;
4857
4858 ASSERT(!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags));
4859 ASSERT(PagePrivate(page));
4860 ASSERT(check);
4861 io_tree = &BTRFS_I(fs_info->btree_inode)->io_tree;
4862
4863 if (wait == WAIT_NONE) {
4864 if (!try_lock_extent(io_tree, eb->start, eb->start + eb->len - 1,
4865 &cached_state))
4866 return -EAGAIN;
4867 } else {
4868 ret = lock_extent(io_tree, eb->start, eb->start + eb->len - 1,
4869 &cached_state);
4870 if (ret < 0)
4871 return ret;
4872 }
4873
4874 ret = 0;
4875 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags) ||
4876 PageUptodate(page) ||
4877 btrfs_subpage_test_uptodate(fs_info, page, eb->start, eb->len)) {
4878 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4879 unlock_extent(io_tree, eb->start, eb->start + eb->len - 1,
4880 &cached_state);
4881 return ret;
4882 }
4883
4884 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
4885 eb->read_mirror = 0;
4886 atomic_set(&eb->io_pages, 1);
4887 check_buffer_tree_ref(eb);
4888 bio_ctrl.end_io_func = end_bio_extent_readpage;
4889
4890 btrfs_subpage_clear_error(fs_info, page, eb->start, eb->len);
4891
4892 btrfs_subpage_start_reader(fs_info, page, eb->start, eb->len);
4893 ret = submit_extent_page(REQ_OP_READ, NULL, &bio_ctrl,
4894 eb->start, page, eb->len,
4895 eb->start - page_offset(page), 0, true);
4896 if (ret) {
4897 /*
4898 * In the endio function, if we hit something wrong we will
4899 * increase the io_pages, so here we need to decrease it for
4900 * error path.
4901 */
4902 atomic_dec(&eb->io_pages);
4903 }
4904 submit_one_bio(&bio_ctrl);
4905 if (ret || wait != WAIT_COMPLETE) {
4906 free_extent_state(cached_state);
4907 return ret;
4908 }
4909
4910 wait_extent_bit(io_tree, eb->start, eb->start + eb->len - 1,
4911 EXTENT_LOCKED, &cached_state);
4912 if (!test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4913 ret = -EIO;
4914 return ret;
4915}
4916
4917int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num,
4918 struct btrfs_tree_parent_check *check)
4919{
4920 int i;
4921 struct page *page;
4922 int err;
4923 int ret = 0;
4924 int locked_pages = 0;
4925 int all_uptodate = 1;
4926 int num_pages;
4927 unsigned long num_reads = 0;
4928 struct btrfs_bio_ctrl bio_ctrl = {
4929 .mirror_num = mirror_num,
4930 .parent_check = check,
4931 };
4932
4933 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4934 return 0;
4935
4936 /*
4937 * We could have had EXTENT_BUFFER_UPTODATE cleared by the write
4938 * operation, which could potentially still be in flight. In this case
4939 * we simply want to return an error.
4940 */
4941 if (unlikely(test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)))
4942 return -EIO;
4943
4944 if (eb->fs_info->nodesize < PAGE_SIZE)
4945 return read_extent_buffer_subpage(eb, wait, mirror_num, check);
4946
4947 num_pages = num_extent_pages(eb);
4948 for (i = 0; i < num_pages; i++) {
4949 page = eb->pages[i];
4950 if (wait == WAIT_NONE) {
4951 /*
4952 * WAIT_NONE is only utilized by readahead. If we can't
4953 * acquire the lock atomically it means either the eb
4954 * is being read out or under modification.
4955 * Either way the eb will be or has been cached,
4956 * readahead can exit safely.
4957 */
4958 if (!trylock_page(page))
4959 goto unlock_exit;
4960 } else {
4961 lock_page(page);
4962 }
4963 locked_pages++;
4964 }
4965 /*
4966 * We need to firstly lock all pages to make sure that
4967 * the uptodate bit of our pages won't be affected by
4968 * clear_extent_buffer_uptodate().
4969 */
4970 for (i = 0; i < num_pages; i++) {
4971 page = eb->pages[i];
4972 if (!PageUptodate(page)) {
4973 num_reads++;
4974 all_uptodate = 0;
4975 }
4976 }
4977
4978 if (all_uptodate) {
4979 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4980 goto unlock_exit;
4981 }
4982
4983 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
4984 eb->read_mirror = 0;
4985 atomic_set(&eb->io_pages, num_reads);
4986 /*
4987 * It is possible for release_folio to clear the TREE_REF bit before we
4988 * set io_pages. See check_buffer_tree_ref for a more detailed comment.
4989 */
4990 check_buffer_tree_ref(eb);
4991 bio_ctrl.end_io_func = end_bio_extent_readpage;
4992 for (i = 0; i < num_pages; i++) {
4993 page = eb->pages[i];
4994
4995 if (!PageUptodate(page)) {
4996 if (ret) {
4997 atomic_dec(&eb->io_pages);
4998 unlock_page(page);
4999 continue;
5000 }
5001
5002 ClearPageError(page);
5003 err = submit_extent_page(REQ_OP_READ, NULL,
5004 &bio_ctrl, page_offset(page), page,
5005 PAGE_SIZE, 0, 0, false);
5006 if (err) {
5007 /*
5008 * We failed to submit the bio so it's the
5009 * caller's responsibility to perform cleanup
5010 * i.e unlock page/set error bit.
5011 */
5012 ret = err;
5013 SetPageError(page);
5014 unlock_page(page);
5015 atomic_dec(&eb->io_pages);
5016 }
5017 } else {
5018 unlock_page(page);
5019 }
5020 }
5021
5022 submit_one_bio(&bio_ctrl);
5023
5024 if (ret || wait != WAIT_COMPLETE)
5025 return ret;
5026
5027 for (i = 0; i < num_pages; i++) {
5028 page = eb->pages[i];
5029 wait_on_page_locked(page);
5030 if (!PageUptodate(page))
5031 ret = -EIO;
5032 }
5033
5034 return ret;
5035
5036unlock_exit:
5037 while (locked_pages > 0) {
5038 locked_pages--;
5039 page = eb->pages[locked_pages];
5040 unlock_page(page);
5041 }
5042 return ret;
5043}
5044
5045static bool report_eb_range(const struct extent_buffer *eb, unsigned long start,
5046 unsigned long len)
5047{
5048 btrfs_warn(eb->fs_info,
5049 "access to eb bytenr %llu len %lu out of range start %lu len %lu",
5050 eb->start, eb->len, start, len);
5051 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
5052
5053 return true;
5054}
5055
5056/*
5057 * Check if the [start, start + len) range is valid before reading/writing
5058 * the eb.
5059 * NOTE: @start and @len are offset inside the eb, not logical address.
5060 *
5061 * Caller should not touch the dst/src memory if this function returns error.
5062 */
5063static inline int check_eb_range(const struct extent_buffer *eb,
5064 unsigned long start, unsigned long len)
5065{
5066 unsigned long offset;
5067
5068 /* start, start + len should not go beyond eb->len nor overflow */
5069 if (unlikely(check_add_overflow(start, len, &offset) || offset > eb->len))
5070 return report_eb_range(eb, start, len);
5071
5072 return false;
5073}
5074
5075void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
5076 unsigned long start, unsigned long len)
5077{
5078 size_t cur;
5079 size_t offset;
5080 struct page *page;
5081 char *kaddr;
5082 char *dst = (char *)dstv;
5083 unsigned long i = get_eb_page_index(start);
5084
5085 if (check_eb_range(eb, start, len))
5086 return;
5087
5088 offset = get_eb_offset_in_page(eb, start);
5089
5090 while (len > 0) {
5091 page = eb->pages[i];
5092
5093 cur = min(len, (PAGE_SIZE - offset));
5094 kaddr = page_address(page);
5095 memcpy(dst, kaddr + offset, cur);
5096
5097 dst += cur;
5098 len -= cur;
5099 offset = 0;
5100 i++;
5101 }
5102}
5103
5104int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
5105 void __user *dstv,
5106 unsigned long start, unsigned long len)
5107{
5108 size_t cur;
5109 size_t offset;
5110 struct page *page;
5111 char *kaddr;
5112 char __user *dst = (char __user *)dstv;
5113 unsigned long i = get_eb_page_index(start);
5114 int ret = 0;
5115
5116 WARN_ON(start > eb->len);
5117 WARN_ON(start + len > eb->start + eb->len);
5118
5119 offset = get_eb_offset_in_page(eb, start);
5120
5121 while (len > 0) {
5122 page = eb->pages[i];
5123
5124 cur = min(len, (PAGE_SIZE - offset));
5125 kaddr = page_address(page);
5126 if (copy_to_user_nofault(dst, kaddr + offset, cur)) {
5127 ret = -EFAULT;
5128 break;
5129 }
5130
5131 dst += cur;
5132 len -= cur;
5133 offset = 0;
5134 i++;
5135 }
5136
5137 return ret;
5138}
5139
5140int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
5141 unsigned long start, unsigned long len)
5142{
5143 size_t cur;
5144 size_t offset;
5145 struct page *page;
5146 char *kaddr;
5147 char *ptr = (char *)ptrv;
5148 unsigned long i = get_eb_page_index(start);
5149 int ret = 0;
5150
5151 if (check_eb_range(eb, start, len))
5152 return -EINVAL;
5153
5154 offset = get_eb_offset_in_page(eb, start);
5155
5156 while (len > 0) {
5157 page = eb->pages[i];
5158
5159 cur = min(len, (PAGE_SIZE - offset));
5160
5161 kaddr = page_address(page);
5162 ret = memcmp(ptr, kaddr + offset, cur);
5163 if (ret)
5164 break;
5165
5166 ptr += cur;
5167 len -= cur;
5168 offset = 0;
5169 i++;
5170 }
5171 return ret;
5172}
5173
5174/*
5175 * Check that the extent buffer is uptodate.
5176 *
5177 * For regular sector size == PAGE_SIZE case, check if @page is uptodate.
5178 * For subpage case, check if the range covered by the eb has EXTENT_UPTODATE.
5179 */
5180static void assert_eb_page_uptodate(const struct extent_buffer *eb,
5181 struct page *page)
5182{
5183 struct btrfs_fs_info *fs_info = eb->fs_info;
5184
5185 /*
5186 * If we are using the commit root we could potentially clear a page
5187 * Uptodate while we're using the extent buffer that we've previously
5188 * looked up. We don't want to complain in this case, as the page was
5189 * valid before, we just didn't write it out. Instead we want to catch
5190 * the case where we didn't actually read the block properly, which
5191 * would have !PageUptodate && !PageError, as we clear PageError before
5192 * reading.
5193 */
5194 if (fs_info->nodesize < PAGE_SIZE) {
5195 bool uptodate, error;
5196
5197 uptodate = btrfs_subpage_test_uptodate(fs_info, page,
5198 eb->start, eb->len);
5199 error = btrfs_subpage_test_error(fs_info, page, eb->start, eb->len);
5200 WARN_ON(!uptodate && !error);
5201 } else {
5202 WARN_ON(!PageUptodate(page) && !PageError(page));
5203 }
5204}
5205
5206void write_extent_buffer_chunk_tree_uuid(const struct extent_buffer *eb,
5207 const void *srcv)
5208{
5209 char *kaddr;
5210
5211 assert_eb_page_uptodate(eb, eb->pages[0]);
5212 kaddr = page_address(eb->pages[0]) +
5213 get_eb_offset_in_page(eb, offsetof(struct btrfs_header,
5214 chunk_tree_uuid));
5215 memcpy(kaddr, srcv, BTRFS_FSID_SIZE);
5216}
5217
5218void write_extent_buffer_fsid(const struct extent_buffer *eb, const void *srcv)
5219{
5220 char *kaddr;
5221
5222 assert_eb_page_uptodate(eb, eb->pages[0]);
5223 kaddr = page_address(eb->pages[0]) +
5224 get_eb_offset_in_page(eb, offsetof(struct btrfs_header, fsid));
5225 memcpy(kaddr, srcv, BTRFS_FSID_SIZE);
5226}
5227
5228void write_extent_buffer(const struct extent_buffer *eb, const void *srcv,
5229 unsigned long start, unsigned long len)
5230{
5231 size_t cur;
5232 size_t offset;
5233 struct page *page;
5234 char *kaddr;
5235 char *src = (char *)srcv;
5236 unsigned long i = get_eb_page_index(start);
5237
5238 WARN_ON(test_bit(EXTENT_BUFFER_NO_CHECK, &eb->bflags));
5239
5240 if (check_eb_range(eb, start, len))
5241 return;
5242
5243 offset = get_eb_offset_in_page(eb, start);
5244
5245 while (len > 0) {
5246 page = eb->pages[i];
5247 assert_eb_page_uptodate(eb, page);
5248
5249 cur = min(len, PAGE_SIZE - offset);
5250 kaddr = page_address(page);
5251 memcpy(kaddr + offset, src, cur);
5252
5253 src += cur;
5254 len -= cur;
5255 offset = 0;
5256 i++;
5257 }
5258}
5259
5260void memzero_extent_buffer(const struct extent_buffer *eb, unsigned long start,
5261 unsigned long len)
5262{
5263 size_t cur;
5264 size_t offset;
5265 struct page *page;
5266 char *kaddr;
5267 unsigned long i = get_eb_page_index(start);
5268
5269 if (check_eb_range(eb, start, len))
5270 return;
5271
5272 offset = get_eb_offset_in_page(eb, start);
5273
5274 while (len > 0) {
5275 page = eb->pages[i];
5276 assert_eb_page_uptodate(eb, page);
5277
5278 cur = min(len, PAGE_SIZE - offset);
5279 kaddr = page_address(page);
5280 memset(kaddr + offset, 0, cur);
5281
5282 len -= cur;
5283 offset = 0;
5284 i++;
5285 }
5286}
5287
5288void copy_extent_buffer_full(const struct extent_buffer *dst,
5289 const struct extent_buffer *src)
5290{
5291 int i;
5292 int num_pages;
5293
5294 ASSERT(dst->len == src->len);
5295
5296 if (dst->fs_info->nodesize >= PAGE_SIZE) {
5297 num_pages = num_extent_pages(dst);
5298 for (i = 0; i < num_pages; i++)
5299 copy_page(page_address(dst->pages[i]),
5300 page_address(src->pages[i]));
5301 } else {
5302 size_t src_offset = get_eb_offset_in_page(src, 0);
5303 size_t dst_offset = get_eb_offset_in_page(dst, 0);
5304
5305 ASSERT(src->fs_info->nodesize < PAGE_SIZE);
5306 memcpy(page_address(dst->pages[0]) + dst_offset,
5307 page_address(src->pages[0]) + src_offset,
5308 src->len);
5309 }
5310}
5311
5312void copy_extent_buffer(const struct extent_buffer *dst,
5313 const struct extent_buffer *src,
5314 unsigned long dst_offset, unsigned long src_offset,
5315 unsigned long len)
5316{
5317 u64 dst_len = dst->len;
5318 size_t cur;
5319 size_t offset;
5320 struct page *page;
5321 char *kaddr;
5322 unsigned long i = get_eb_page_index(dst_offset);
5323
5324 if (check_eb_range(dst, dst_offset, len) ||
5325 check_eb_range(src, src_offset, len))
5326 return;
5327
5328 WARN_ON(src->len != dst_len);
5329
5330 offset = get_eb_offset_in_page(dst, dst_offset);
5331
5332 while (len > 0) {
5333 page = dst->pages[i];
5334 assert_eb_page_uptodate(dst, page);
5335
5336 cur = min(len, (unsigned long)(PAGE_SIZE - offset));
5337
5338 kaddr = page_address(page);
5339 read_extent_buffer(src, kaddr + offset, src_offset, cur);
5340
5341 src_offset += cur;
5342 len -= cur;
5343 offset = 0;
5344 i++;
5345 }
5346}
5347
5348/*
5349 * eb_bitmap_offset() - calculate the page and offset of the byte containing the
5350 * given bit number
5351 * @eb: the extent buffer
5352 * @start: offset of the bitmap item in the extent buffer
5353 * @nr: bit number
5354 * @page_index: return index of the page in the extent buffer that contains the
5355 * given bit number
5356 * @page_offset: return offset into the page given by page_index
5357 *
5358 * This helper hides the ugliness of finding the byte in an extent buffer which
5359 * contains a given bit.
5360 */
5361static inline void eb_bitmap_offset(const struct extent_buffer *eb,
5362 unsigned long start, unsigned long nr,
5363 unsigned long *page_index,
5364 size_t *page_offset)
5365{
5366 size_t byte_offset = BIT_BYTE(nr);
5367 size_t offset;
5368
5369 /*
5370 * The byte we want is the offset of the extent buffer + the offset of
5371 * the bitmap item in the extent buffer + the offset of the byte in the
5372 * bitmap item.
5373 */
5374 offset = start + offset_in_page(eb->start) + byte_offset;
5375
5376 *page_index = offset >> PAGE_SHIFT;
5377 *page_offset = offset_in_page(offset);
5378}
5379
5380/*
5381 * Determine whether a bit in a bitmap item is set.
5382 *
5383 * @eb: the extent buffer
5384 * @start: offset of the bitmap item in the extent buffer
5385 * @nr: bit number to test
5386 */
5387int extent_buffer_test_bit(const struct extent_buffer *eb, unsigned long start,
5388 unsigned long nr)
5389{
5390 u8 *kaddr;
5391 struct page *page;
5392 unsigned long i;
5393 size_t offset;
5394
5395 eb_bitmap_offset(eb, start, nr, &i, &offset);
5396 page = eb->pages[i];
5397 assert_eb_page_uptodate(eb, page);
5398 kaddr = page_address(page);
5399 return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
5400}
5401
5402/*
5403 * Set an area of a bitmap to 1.
5404 *
5405 * @eb: the extent buffer
5406 * @start: offset of the bitmap item in the extent buffer
5407 * @pos: bit number of the first bit
5408 * @len: number of bits to set
5409 */
5410void extent_buffer_bitmap_set(const struct extent_buffer *eb, unsigned long start,
5411 unsigned long pos, unsigned long len)
5412{
5413 u8 *kaddr;
5414 struct page *page;
5415 unsigned long i;
5416 size_t offset;
5417 const unsigned int size = pos + len;
5418 int bits_to_set = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5419 u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(pos);
5420
5421 eb_bitmap_offset(eb, start, pos, &i, &offset);
5422 page = eb->pages[i];
5423 assert_eb_page_uptodate(eb, page);
5424 kaddr = page_address(page);
5425
5426 while (len >= bits_to_set) {
5427 kaddr[offset] |= mask_to_set;
5428 len -= bits_to_set;
5429 bits_to_set = BITS_PER_BYTE;
5430 mask_to_set = ~0;
5431 if (++offset >= PAGE_SIZE && len > 0) {
5432 offset = 0;
5433 page = eb->pages[++i];
5434 assert_eb_page_uptodate(eb, page);
5435 kaddr = page_address(page);
5436 }
5437 }
5438 if (len) {
5439 mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
5440 kaddr[offset] |= mask_to_set;
5441 }
5442}
5443
5444
5445/*
5446 * Clear an area of a bitmap.
5447 *
5448 * @eb: the extent buffer
5449 * @start: offset of the bitmap item in the extent buffer
5450 * @pos: bit number of the first bit
5451 * @len: number of bits to clear
5452 */
5453void extent_buffer_bitmap_clear(const struct extent_buffer *eb,
5454 unsigned long start, unsigned long pos,
5455 unsigned long len)
5456{
5457 u8 *kaddr;
5458 struct page *page;
5459 unsigned long i;
5460 size_t offset;
5461 const unsigned int size = pos + len;
5462 int bits_to_clear = BITS_PER_BYTE - (pos % BITS_PER_BYTE);
5463 u8 mask_to_clear = BITMAP_FIRST_BYTE_MASK(pos);
5464
5465 eb_bitmap_offset(eb, start, pos, &i, &offset);
5466 page = eb->pages[i];
5467 assert_eb_page_uptodate(eb, page);
5468 kaddr = page_address(page);
5469
5470 while (len >= bits_to_clear) {
5471 kaddr[offset] &= ~mask_to_clear;
5472 len -= bits_to_clear;
5473 bits_to_clear = BITS_PER_BYTE;
5474 mask_to_clear = ~0;
5475 if (++offset >= PAGE_SIZE && len > 0) {
5476 offset = 0;
5477 page = eb->pages[++i];
5478 assert_eb_page_uptodate(eb, page);
5479 kaddr = page_address(page);
5480 }
5481 }
5482 if (len) {
5483 mask_to_clear &= BITMAP_LAST_BYTE_MASK(size);
5484 kaddr[offset] &= ~mask_to_clear;
5485 }
5486}
5487
5488static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
5489{
5490 unsigned long distance = (src > dst) ? src - dst : dst - src;
5491 return distance < len;
5492}
5493
5494static void copy_pages(struct page *dst_page, struct page *src_page,
5495 unsigned long dst_off, unsigned long src_off,
5496 unsigned long len)
5497{
5498 char *dst_kaddr = page_address(dst_page);
5499 char *src_kaddr;
5500 int must_memmove = 0;
5501
5502 if (dst_page != src_page) {
5503 src_kaddr = page_address(src_page);
5504 } else {
5505 src_kaddr = dst_kaddr;
5506 if (areas_overlap(src_off, dst_off, len))
5507 must_memmove = 1;
5508 }
5509
5510 if (must_memmove)
5511 memmove(dst_kaddr + dst_off, src_kaddr + src_off, len);
5512 else
5513 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
5514}
5515
5516void memcpy_extent_buffer(const struct extent_buffer *dst,
5517 unsigned long dst_offset, unsigned long src_offset,
5518 unsigned long len)
5519{
5520 size_t cur;
5521 size_t dst_off_in_page;
5522 size_t src_off_in_page;
5523 unsigned long dst_i;
5524 unsigned long src_i;
5525
5526 if (check_eb_range(dst, dst_offset, len) ||
5527 check_eb_range(dst, src_offset, len))
5528 return;
5529
5530 while (len > 0) {
5531 dst_off_in_page = get_eb_offset_in_page(dst, dst_offset);
5532 src_off_in_page = get_eb_offset_in_page(dst, src_offset);
5533
5534 dst_i = get_eb_page_index(dst_offset);
5535 src_i = get_eb_page_index(src_offset);
5536
5537 cur = min(len, (unsigned long)(PAGE_SIZE -
5538 src_off_in_page));
5539 cur = min_t(unsigned long, cur,
5540 (unsigned long)(PAGE_SIZE - dst_off_in_page));
5541
5542 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5543 dst_off_in_page, src_off_in_page, cur);
5544
5545 src_offset += cur;
5546 dst_offset += cur;
5547 len -= cur;
5548 }
5549}
5550
5551void memmove_extent_buffer(const struct extent_buffer *dst,
5552 unsigned long dst_offset, unsigned long src_offset,
5553 unsigned long len)
5554{
5555 size_t cur;
5556 size_t dst_off_in_page;
5557 size_t src_off_in_page;
5558 unsigned long dst_end = dst_offset + len - 1;
5559 unsigned long src_end = src_offset + len - 1;
5560 unsigned long dst_i;
5561 unsigned long src_i;
5562
5563 if (check_eb_range(dst, dst_offset, len) ||
5564 check_eb_range(dst, src_offset, len))
5565 return;
5566 if (dst_offset < src_offset) {
5567 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
5568 return;
5569 }
5570 while (len > 0) {
5571 dst_i = get_eb_page_index(dst_end);
5572 src_i = get_eb_page_index(src_end);
5573
5574 dst_off_in_page = get_eb_offset_in_page(dst, dst_end);
5575 src_off_in_page = get_eb_offset_in_page(dst, src_end);
5576
5577 cur = min_t(unsigned long, len, src_off_in_page + 1);
5578 cur = min(cur, dst_off_in_page + 1);
5579 copy_pages(dst->pages[dst_i], dst->pages[src_i],
5580 dst_off_in_page - cur + 1,
5581 src_off_in_page - cur + 1, cur);
5582
5583 dst_end -= cur;
5584 src_end -= cur;
5585 len -= cur;
5586 }
5587}
5588
5589#define GANG_LOOKUP_SIZE 16
5590static struct extent_buffer *get_next_extent_buffer(
5591 struct btrfs_fs_info *fs_info, struct page *page, u64 bytenr)
5592{
5593 struct extent_buffer *gang[GANG_LOOKUP_SIZE];
5594 struct extent_buffer *found = NULL;
5595 u64 page_start = page_offset(page);
5596 u64 cur = page_start;
5597
5598 ASSERT(in_range(bytenr, page_start, PAGE_SIZE));
5599 lockdep_assert_held(&fs_info->buffer_lock);
5600
5601 while (cur < page_start + PAGE_SIZE) {
5602 int ret;
5603 int i;
5604
5605 ret = radix_tree_gang_lookup(&fs_info->buffer_radix,
5606 (void **)gang, cur >> fs_info->sectorsize_bits,
5607 min_t(unsigned int, GANG_LOOKUP_SIZE,
5608 PAGE_SIZE / fs_info->nodesize));
5609 if (ret == 0)
5610 goto out;
5611 for (i = 0; i < ret; i++) {
5612 /* Already beyond page end */
5613 if (gang[i]->start >= page_start + PAGE_SIZE)
5614 goto out;
5615 /* Found one */
5616 if (gang[i]->start >= bytenr) {
5617 found = gang[i];
5618 goto out;
5619 }
5620 }
5621 cur = gang[ret - 1]->start + gang[ret - 1]->len;
5622 }
5623out:
5624 return found;
5625}
5626
5627static int try_release_subpage_extent_buffer(struct page *page)
5628{
5629 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
5630 u64 cur = page_offset(page);
5631 const u64 end = page_offset(page) + PAGE_SIZE;
5632 int ret;
5633
5634 while (cur < end) {
5635 struct extent_buffer *eb = NULL;
5636
5637 /*
5638 * Unlike try_release_extent_buffer() which uses page->private
5639 * to grab buffer, for subpage case we rely on radix tree, thus
5640 * we need to ensure radix tree consistency.
5641 *
5642 * We also want an atomic snapshot of the radix tree, thus go
5643 * with spinlock rather than RCU.
5644 */
5645 spin_lock(&fs_info->buffer_lock);
5646 eb = get_next_extent_buffer(fs_info, page, cur);
5647 if (!eb) {
5648 /* No more eb in the page range after or at cur */
5649 spin_unlock(&fs_info->buffer_lock);
5650 break;
5651 }
5652 cur = eb->start + eb->len;
5653
5654 /*
5655 * The same as try_release_extent_buffer(), to ensure the eb
5656 * won't disappear out from under us.
5657 */
5658 spin_lock(&eb->refs_lock);
5659 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5660 spin_unlock(&eb->refs_lock);
5661 spin_unlock(&fs_info->buffer_lock);
5662 break;
5663 }
5664 spin_unlock(&fs_info->buffer_lock);
5665
5666 /*
5667 * If tree ref isn't set then we know the ref on this eb is a
5668 * real ref, so just return, this eb will likely be freed soon
5669 * anyway.
5670 */
5671 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5672 spin_unlock(&eb->refs_lock);
5673 break;
5674 }
5675
5676 /*
5677 * Here we don't care about the return value, we will always
5678 * check the page private at the end. And
5679 * release_extent_buffer() will release the refs_lock.
5680 */
5681 release_extent_buffer(eb);
5682 }
5683 /*
5684 * Finally to check if we have cleared page private, as if we have
5685 * released all ebs in the page, the page private should be cleared now.
5686 */
5687 spin_lock(&page->mapping->private_lock);
5688 if (!PagePrivate(page))
5689 ret = 1;
5690 else
5691 ret = 0;
5692 spin_unlock(&page->mapping->private_lock);
5693 return ret;
5694
5695}
5696
5697int try_release_extent_buffer(struct page *page)
5698{
5699 struct extent_buffer *eb;
5700
5701 if (btrfs_sb(page->mapping->host->i_sb)->nodesize < PAGE_SIZE)
5702 return try_release_subpage_extent_buffer(page);
5703
5704 /*
5705 * We need to make sure nobody is changing page->private, as we rely on
5706 * page->private as the pointer to extent buffer.
5707 */
5708 spin_lock(&page->mapping->private_lock);
5709 if (!PagePrivate(page)) {
5710 spin_unlock(&page->mapping->private_lock);
5711 return 1;
5712 }
5713
5714 eb = (struct extent_buffer *)page->private;
5715 BUG_ON(!eb);
5716
5717 /*
5718 * This is a little awful but should be ok, we need to make sure that
5719 * the eb doesn't disappear out from under us while we're looking at
5720 * this page.
5721 */
5722 spin_lock(&eb->refs_lock);
5723 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
5724 spin_unlock(&eb->refs_lock);
5725 spin_unlock(&page->mapping->private_lock);
5726 return 0;
5727 }
5728 spin_unlock(&page->mapping->private_lock);
5729
5730 /*
5731 * If tree ref isn't set then we know the ref on this eb is a real ref,
5732 * so just return, this page will likely be freed soon anyway.
5733 */
5734 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
5735 spin_unlock(&eb->refs_lock);
5736 return 0;
5737 }
5738
5739 return release_extent_buffer(eb);
5740}
5741
5742/*
5743 * btrfs_readahead_tree_block - attempt to readahead a child block
5744 * @fs_info: the fs_info
5745 * @bytenr: bytenr to read
5746 * @owner_root: objectid of the root that owns this eb
5747 * @gen: generation for the uptodate check, can be 0
5748 * @level: level for the eb
5749 *
5750 * Attempt to readahead a tree block at @bytenr. If @gen is 0 then we do a
5751 * normal uptodate check of the eb, without checking the generation. If we have
5752 * to read the block we will not block on anything.
5753 */
5754void btrfs_readahead_tree_block(struct btrfs_fs_info *fs_info,
5755 u64 bytenr, u64 owner_root, u64 gen, int level)
5756{
5757 struct btrfs_tree_parent_check check = {
5758 .has_first_key = 0,
5759 .level = level,
5760 .transid = gen
5761 };
5762 struct extent_buffer *eb;
5763 int ret;
5764
5765 eb = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
5766 if (IS_ERR(eb))
5767 return;
5768
5769 if (btrfs_buffer_uptodate(eb, gen, 1)) {
5770 free_extent_buffer(eb);
5771 return;
5772 }
5773
5774 ret = read_extent_buffer_pages(eb, WAIT_NONE, 0, &check);
5775 if (ret < 0)
5776 free_extent_buffer_stale(eb);
5777 else
5778 free_extent_buffer(eb);
5779}
5780
5781/*
5782 * btrfs_readahead_node_child - readahead a node's child block
5783 * @node: parent node we're reading from
5784 * @slot: slot in the parent node for the child we want to read
5785 *
5786 * A helper for btrfs_readahead_tree_block, we simply read the bytenr pointed at
5787 * the slot in the node provided.
5788 */
5789void btrfs_readahead_node_child(struct extent_buffer *node, int slot)
5790{
5791 btrfs_readahead_tree_block(node->fs_info,
5792 btrfs_node_blockptr(node, slot),
5793 btrfs_header_owner(node),
5794 btrfs_node_ptr_generation(node, slot),
5795 btrfs_header_level(node) - 1);
5796}
1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/bitops.h>
4#include <linux/slab.h>
5#include <linux/bio.h>
6#include <linux/mm.h>
7#include <linux/pagemap.h>
8#include <linux/page-flags.h>
9#include <linux/sched/mm.h>
10#include <linux/spinlock.h>
11#include <linux/blkdev.h>
12#include <linux/swap.h>
13#include <linux/writeback.h>
14#include <linux/pagevec.h>
15#include <linux/prefetch.h>
16#include <linux/fsverity.h>
17#include "extent_io.h"
18#include "extent-io-tree.h"
19#include "extent_map.h"
20#include "ctree.h"
21#include "btrfs_inode.h"
22#include "bio.h"
23#include "locking.h"
24#include "backref.h"
25#include "disk-io.h"
26#include "subpage.h"
27#include "zoned.h"
28#include "block-group.h"
29#include "compression.h"
30#include "fs.h"
31#include "accessors.h"
32#include "file-item.h"
33#include "file.h"
34#include "dev-replace.h"
35#include "super.h"
36#include "transaction.h"
37
38static struct kmem_cache *extent_buffer_cache;
39
40#ifdef CONFIG_BTRFS_DEBUG
41static inline void btrfs_leak_debug_add_eb(struct extent_buffer *eb)
42{
43 struct btrfs_fs_info *fs_info = eb->fs_info;
44 unsigned long flags;
45
46 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
47 list_add(&eb->leak_list, &fs_info->allocated_ebs);
48 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
49}
50
51static inline void btrfs_leak_debug_del_eb(struct extent_buffer *eb)
52{
53 struct btrfs_fs_info *fs_info = eb->fs_info;
54 unsigned long flags;
55
56 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
57 list_del(&eb->leak_list);
58 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
59}
60
61void btrfs_extent_buffer_leak_debug_check(struct btrfs_fs_info *fs_info)
62{
63 struct extent_buffer *eb;
64 unsigned long flags;
65
66 /*
67 * If we didn't get into open_ctree our allocated_ebs will not be
68 * initialized, so just skip this.
69 */
70 if (!fs_info->allocated_ebs.next)
71 return;
72
73 WARN_ON(!list_empty(&fs_info->allocated_ebs));
74 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
75 while (!list_empty(&fs_info->allocated_ebs)) {
76 eb = list_first_entry(&fs_info->allocated_ebs,
77 struct extent_buffer, leak_list);
78 pr_err(
79 "BTRFS: buffer leak start %llu len %u refs %d bflags %lu owner %llu\n",
80 eb->start, eb->len, atomic_read(&eb->refs), eb->bflags,
81 btrfs_header_owner(eb));
82 list_del(&eb->leak_list);
83 WARN_ON_ONCE(1);
84 kmem_cache_free(extent_buffer_cache, eb);
85 }
86 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
87}
88#else
89#define btrfs_leak_debug_add_eb(eb) do {} while (0)
90#define btrfs_leak_debug_del_eb(eb) do {} while (0)
91#endif
92
93/*
94 * Structure to record info about the bio being assembled, and other info like
95 * how many bytes are there before stripe/ordered extent boundary.
96 */
97struct btrfs_bio_ctrl {
98 struct btrfs_bio *bbio;
99 enum btrfs_compression_type compress_type;
100 u32 len_to_oe_boundary;
101 blk_opf_t opf;
102 btrfs_bio_end_io_t end_io_func;
103 struct writeback_control *wbc;
104
105 /*
106 * The sectors of the page which are going to be submitted by
107 * extent_writepage_io().
108 * This is to avoid touching ranges covered by compression/inline.
109 */
110 unsigned long submit_bitmap;
111};
112
113static void submit_one_bio(struct btrfs_bio_ctrl *bio_ctrl)
114{
115 struct btrfs_bio *bbio = bio_ctrl->bbio;
116
117 if (!bbio)
118 return;
119
120 /* Caller should ensure the bio has at least some range added */
121 ASSERT(bbio->bio.bi_iter.bi_size);
122
123 if (btrfs_op(&bbio->bio) == BTRFS_MAP_READ &&
124 bio_ctrl->compress_type != BTRFS_COMPRESS_NONE)
125 btrfs_submit_compressed_read(bbio);
126 else
127 btrfs_submit_bbio(bbio, 0);
128
129 /* The bbio is owned by the end_io handler now */
130 bio_ctrl->bbio = NULL;
131}
132
133/*
134 * Submit or fail the current bio in the bio_ctrl structure.
135 */
136static void submit_write_bio(struct btrfs_bio_ctrl *bio_ctrl, int ret)
137{
138 struct btrfs_bio *bbio = bio_ctrl->bbio;
139
140 if (!bbio)
141 return;
142
143 if (ret) {
144 ASSERT(ret < 0);
145 btrfs_bio_end_io(bbio, errno_to_blk_status(ret));
146 /* The bio is owned by the end_io handler now */
147 bio_ctrl->bbio = NULL;
148 } else {
149 submit_one_bio(bio_ctrl);
150 }
151}
152
153int __init extent_buffer_init_cachep(void)
154{
155 extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
156 sizeof(struct extent_buffer), 0, 0,
157 NULL);
158 if (!extent_buffer_cache)
159 return -ENOMEM;
160
161 return 0;
162}
163
164void __cold extent_buffer_free_cachep(void)
165{
166 /*
167 * Make sure all delayed rcu free are flushed before we
168 * destroy caches.
169 */
170 rcu_barrier();
171 kmem_cache_destroy(extent_buffer_cache);
172}
173
174static void process_one_folio(struct btrfs_fs_info *fs_info,
175 struct folio *folio, const struct folio *locked_folio,
176 unsigned long page_ops, u64 start, u64 end)
177{
178 u32 len;
179
180 ASSERT(end + 1 - start != 0 && end + 1 - start < U32_MAX);
181 len = end + 1 - start;
182
183 if (page_ops & PAGE_SET_ORDERED)
184 btrfs_folio_clamp_set_ordered(fs_info, folio, start, len);
185 if (page_ops & PAGE_START_WRITEBACK) {
186 btrfs_folio_clamp_clear_dirty(fs_info, folio, start, len);
187 btrfs_folio_clamp_set_writeback(fs_info, folio, start, len);
188 }
189 if (page_ops & PAGE_END_WRITEBACK)
190 btrfs_folio_clamp_clear_writeback(fs_info, folio, start, len);
191
192 if (folio != locked_folio && (page_ops & PAGE_UNLOCK))
193 btrfs_folio_end_lock(fs_info, folio, start, len);
194}
195
196static void __process_folios_contig(struct address_space *mapping,
197 const struct folio *locked_folio, u64 start,
198 u64 end, unsigned long page_ops)
199{
200 struct btrfs_fs_info *fs_info = inode_to_fs_info(mapping->host);
201 pgoff_t start_index = start >> PAGE_SHIFT;
202 pgoff_t end_index = end >> PAGE_SHIFT;
203 pgoff_t index = start_index;
204 struct folio_batch fbatch;
205 int i;
206
207 folio_batch_init(&fbatch);
208 while (index <= end_index) {
209 int found_folios;
210
211 found_folios = filemap_get_folios_contig(mapping, &index,
212 end_index, &fbatch);
213 for (i = 0; i < found_folios; i++) {
214 struct folio *folio = fbatch.folios[i];
215
216 process_one_folio(fs_info, folio, locked_folio,
217 page_ops, start, end);
218 }
219 folio_batch_release(&fbatch);
220 cond_resched();
221 }
222}
223
224static noinline void __unlock_for_delalloc(const struct inode *inode,
225 const struct folio *locked_folio,
226 u64 start, u64 end)
227{
228 unsigned long index = start >> PAGE_SHIFT;
229 unsigned long end_index = end >> PAGE_SHIFT;
230
231 ASSERT(locked_folio);
232 if (index == locked_folio->index && end_index == index)
233 return;
234
235 __process_folios_contig(inode->i_mapping, locked_folio, start, end,
236 PAGE_UNLOCK);
237}
238
239static noinline int lock_delalloc_folios(struct inode *inode,
240 const struct folio *locked_folio,
241 u64 start, u64 end)
242{
243 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
244 struct address_space *mapping = inode->i_mapping;
245 pgoff_t start_index = start >> PAGE_SHIFT;
246 pgoff_t end_index = end >> PAGE_SHIFT;
247 pgoff_t index = start_index;
248 u64 processed_end = start;
249 struct folio_batch fbatch;
250
251 if (index == locked_folio->index && index == end_index)
252 return 0;
253
254 folio_batch_init(&fbatch);
255 while (index <= end_index) {
256 unsigned int found_folios, i;
257
258 found_folios = filemap_get_folios_contig(mapping, &index,
259 end_index, &fbatch);
260 if (found_folios == 0)
261 goto out;
262
263 for (i = 0; i < found_folios; i++) {
264 struct folio *folio = fbatch.folios[i];
265 u64 range_start;
266 u32 range_len;
267
268 if (folio == locked_folio)
269 continue;
270
271 folio_lock(folio);
272 if (!folio_test_dirty(folio) || folio->mapping != mapping) {
273 folio_unlock(folio);
274 goto out;
275 }
276 range_start = max_t(u64, folio_pos(folio), start);
277 range_len = min_t(u64, folio_pos(folio) + folio_size(folio),
278 end + 1) - range_start;
279 btrfs_folio_set_lock(fs_info, folio, range_start, range_len);
280
281 processed_end = range_start + range_len - 1;
282 }
283 folio_batch_release(&fbatch);
284 cond_resched();
285 }
286
287 return 0;
288out:
289 folio_batch_release(&fbatch);
290 if (processed_end > start)
291 __unlock_for_delalloc(inode, locked_folio, start,
292 processed_end);
293 return -EAGAIN;
294}
295
296/*
297 * Find and lock a contiguous range of bytes in the file marked as delalloc, no
298 * more than @max_bytes.
299 *
300 * @start: The original start bytenr to search.
301 * Will store the extent range start bytenr.
302 * @end: The original end bytenr of the search range
303 * Will store the extent range end bytenr.
304 *
305 * Return true if we find a delalloc range which starts inside the original
306 * range, and @start/@end will store the delalloc range start/end.
307 *
308 * Return false if we can't find any delalloc range which starts inside the
309 * original range, and @start/@end will be the non-delalloc range start/end.
310 */
311EXPORT_FOR_TESTS
312noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
313 struct folio *locked_folio,
314 u64 *start, u64 *end)
315{
316 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
317 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
318 const u64 orig_start = *start;
319 const u64 orig_end = *end;
320 /* The sanity tests may not set a valid fs_info. */
321 u64 max_bytes = fs_info ? fs_info->max_extent_size : BTRFS_MAX_EXTENT_SIZE;
322 u64 delalloc_start;
323 u64 delalloc_end;
324 bool found;
325 struct extent_state *cached_state = NULL;
326 int ret;
327 int loops = 0;
328
329 /* Caller should pass a valid @end to indicate the search range end */
330 ASSERT(orig_end > orig_start);
331
332 /* The range should at least cover part of the folio */
333 ASSERT(!(orig_start >= folio_pos(locked_folio) + folio_size(locked_folio) ||
334 orig_end <= folio_pos(locked_folio)));
335again:
336 /* step one, find a bunch of delalloc bytes starting at start */
337 delalloc_start = *start;
338 delalloc_end = 0;
339 found = btrfs_find_delalloc_range(tree, &delalloc_start, &delalloc_end,
340 max_bytes, &cached_state);
341 if (!found || delalloc_end <= *start || delalloc_start > orig_end) {
342 *start = delalloc_start;
343
344 /* @delalloc_end can be -1, never go beyond @orig_end */
345 *end = min(delalloc_end, orig_end);
346 free_extent_state(cached_state);
347 return false;
348 }
349
350 /*
351 * start comes from the offset of locked_folio. We have to lock
352 * folios in order, so we can't process delalloc bytes before
353 * locked_folio
354 */
355 if (delalloc_start < *start)
356 delalloc_start = *start;
357
358 /*
359 * make sure to limit the number of folios we try to lock down
360 */
361 if (delalloc_end + 1 - delalloc_start > max_bytes)
362 delalloc_end = delalloc_start + max_bytes - 1;
363
364 /* step two, lock all the folioss after the folios that has start */
365 ret = lock_delalloc_folios(inode, locked_folio, delalloc_start,
366 delalloc_end);
367 ASSERT(!ret || ret == -EAGAIN);
368 if (ret == -EAGAIN) {
369 /* some of the folios are gone, lets avoid looping by
370 * shortening the size of the delalloc range we're searching
371 */
372 free_extent_state(cached_state);
373 cached_state = NULL;
374 if (!loops) {
375 max_bytes = PAGE_SIZE;
376 loops = 1;
377 goto again;
378 } else {
379 found = false;
380 goto out_failed;
381 }
382 }
383
384 /* step three, lock the state bits for the whole range */
385 lock_extent(tree, delalloc_start, delalloc_end, &cached_state);
386
387 /* then test to make sure it is all still delalloc */
388 ret = test_range_bit(tree, delalloc_start, delalloc_end,
389 EXTENT_DELALLOC, cached_state);
390
391 unlock_extent(tree, delalloc_start, delalloc_end, &cached_state);
392 if (!ret) {
393 __unlock_for_delalloc(inode, locked_folio, delalloc_start,
394 delalloc_end);
395 cond_resched();
396 goto again;
397 }
398 *start = delalloc_start;
399 *end = delalloc_end;
400out_failed:
401 return found;
402}
403
404void extent_clear_unlock_delalloc(struct btrfs_inode *inode, u64 start, u64 end,
405 const struct folio *locked_folio,
406 struct extent_state **cached,
407 u32 clear_bits, unsigned long page_ops)
408{
409 clear_extent_bit(&inode->io_tree, start, end, clear_bits, cached);
410
411 __process_folios_contig(inode->vfs_inode.i_mapping, locked_folio, start,
412 end, page_ops);
413}
414
415static bool btrfs_verify_folio(struct folio *folio, u64 start, u32 len)
416{
417 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
418
419 if (!fsverity_active(folio->mapping->host) ||
420 btrfs_folio_test_uptodate(fs_info, folio, start, len) ||
421 start >= i_size_read(folio->mapping->host))
422 return true;
423 return fsverity_verify_folio(folio);
424}
425
426static void end_folio_read(struct folio *folio, bool uptodate, u64 start, u32 len)
427{
428 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
429
430 ASSERT(folio_pos(folio) <= start &&
431 start + len <= folio_pos(folio) + PAGE_SIZE);
432
433 if (uptodate && btrfs_verify_folio(folio, start, len))
434 btrfs_folio_set_uptodate(fs_info, folio, start, len);
435 else
436 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
437
438 if (!btrfs_is_subpage(fs_info, folio->mapping))
439 folio_unlock(folio);
440 else
441 btrfs_folio_end_lock(fs_info, folio, start, len);
442}
443
444/*
445 * After a write IO is done, we need to:
446 *
447 * - clear the uptodate bits on error
448 * - clear the writeback bits in the extent tree for the range
449 * - filio_end_writeback() if there is no more pending io for the folio
450 *
451 * Scheduling is not allowed, so the extent state tree is expected
452 * to have one and only one object corresponding to this IO.
453 */
454static void end_bbio_data_write(struct btrfs_bio *bbio)
455{
456 struct btrfs_fs_info *fs_info = bbio->fs_info;
457 struct bio *bio = &bbio->bio;
458 int error = blk_status_to_errno(bio->bi_status);
459 struct folio_iter fi;
460 const u32 sectorsize = fs_info->sectorsize;
461
462 ASSERT(!bio_flagged(bio, BIO_CLONED));
463 bio_for_each_folio_all(fi, bio) {
464 struct folio *folio = fi.folio;
465 u64 start = folio_pos(folio) + fi.offset;
466 u32 len = fi.length;
467
468 /* Only order 0 (single page) folios are allowed for data. */
469 ASSERT(folio_order(folio) == 0);
470
471 /* Our read/write should always be sector aligned. */
472 if (!IS_ALIGNED(fi.offset, sectorsize))
473 btrfs_err(fs_info,
474 "partial page write in btrfs with offset %zu and length %zu",
475 fi.offset, fi.length);
476 else if (!IS_ALIGNED(fi.length, sectorsize))
477 btrfs_info(fs_info,
478 "incomplete page write with offset %zu and length %zu",
479 fi.offset, fi.length);
480
481 btrfs_finish_ordered_extent(bbio->ordered, folio, start, len,
482 !error);
483 if (error)
484 mapping_set_error(folio->mapping, error);
485 btrfs_folio_clear_writeback(fs_info, folio, start, len);
486 }
487
488 bio_put(bio);
489}
490
491static void begin_folio_read(struct btrfs_fs_info *fs_info, struct folio *folio)
492{
493 ASSERT(folio_test_locked(folio));
494 if (!btrfs_is_subpage(fs_info, folio->mapping))
495 return;
496
497 ASSERT(folio_test_private(folio));
498 btrfs_folio_set_lock(fs_info, folio, folio_pos(folio), PAGE_SIZE);
499}
500
501/*
502 * After a data read IO is done, we need to:
503 *
504 * - clear the uptodate bits on error
505 * - set the uptodate bits if things worked
506 * - set the folio up to date if all extents in the tree are uptodate
507 * - clear the lock bit in the extent tree
508 * - unlock the folio if there are no other extents locked for it
509 *
510 * Scheduling is not allowed, so the extent state tree is expected
511 * to have one and only one object corresponding to this IO.
512 */
513static void end_bbio_data_read(struct btrfs_bio *bbio)
514{
515 struct btrfs_fs_info *fs_info = bbio->fs_info;
516 struct bio *bio = &bbio->bio;
517 struct folio_iter fi;
518 const u32 sectorsize = fs_info->sectorsize;
519
520 ASSERT(!bio_flagged(bio, BIO_CLONED));
521 bio_for_each_folio_all(fi, &bbio->bio) {
522 bool uptodate = !bio->bi_status;
523 struct folio *folio = fi.folio;
524 struct inode *inode = folio->mapping->host;
525 u64 start;
526 u64 end;
527 u32 len;
528
529 /* For now only order 0 folios are supported for data. */
530 ASSERT(folio_order(folio) == 0);
531 btrfs_debug(fs_info,
532 "%s: bi_sector=%llu, err=%d, mirror=%u",
533 __func__, bio->bi_iter.bi_sector, bio->bi_status,
534 bbio->mirror_num);
535
536 /*
537 * We always issue full-sector reads, but if some block in a
538 * folio fails to read, blk_update_request() will advance
539 * bv_offset and adjust bv_len to compensate. Print a warning
540 * for unaligned offsets, and an error if they don't add up to
541 * a full sector.
542 */
543 if (!IS_ALIGNED(fi.offset, sectorsize))
544 btrfs_err(fs_info,
545 "partial page read in btrfs with offset %zu and length %zu",
546 fi.offset, fi.length);
547 else if (!IS_ALIGNED(fi.offset + fi.length, sectorsize))
548 btrfs_info(fs_info,
549 "incomplete page read with offset %zu and length %zu",
550 fi.offset, fi.length);
551
552 start = folio_pos(folio) + fi.offset;
553 end = start + fi.length - 1;
554 len = fi.length;
555
556 if (likely(uptodate)) {
557 loff_t i_size = i_size_read(inode);
558 pgoff_t end_index = i_size >> folio_shift(folio);
559
560 /*
561 * Zero out the remaining part if this range straddles
562 * i_size.
563 *
564 * Here we should only zero the range inside the folio,
565 * not touch anything else.
566 *
567 * NOTE: i_size is exclusive while end is inclusive.
568 */
569 if (folio_index(folio) == end_index && i_size <= end) {
570 u32 zero_start = max(offset_in_folio(folio, i_size),
571 offset_in_folio(folio, start));
572 u32 zero_len = offset_in_folio(folio, end) + 1 -
573 zero_start;
574
575 folio_zero_range(folio, zero_start, zero_len);
576 }
577 }
578
579 /* Update page status and unlock. */
580 end_folio_read(folio, uptodate, start, len);
581 }
582 bio_put(bio);
583}
584
585/*
586 * Populate every free slot in a provided array with folios using GFP_NOFS.
587 *
588 * @nr_folios: number of folios to allocate
589 * @folio_array: the array to fill with folios; any existing non-NULL entries in
590 * the array will be skipped
591 *
592 * Return: 0 if all folios were able to be allocated;
593 * -ENOMEM otherwise, the partially allocated folios would be freed and
594 * the array slots zeroed
595 */
596int btrfs_alloc_folio_array(unsigned int nr_folios, struct folio **folio_array)
597{
598 for (int i = 0; i < nr_folios; i++) {
599 if (folio_array[i])
600 continue;
601 folio_array[i] = folio_alloc(GFP_NOFS, 0);
602 if (!folio_array[i])
603 goto error;
604 }
605 return 0;
606error:
607 for (int i = 0; i < nr_folios; i++) {
608 if (folio_array[i])
609 folio_put(folio_array[i]);
610 }
611 return -ENOMEM;
612}
613
614/*
615 * Populate every free slot in a provided array with pages, using GFP_NOFS.
616 *
617 * @nr_pages: number of pages to allocate
618 * @page_array: the array to fill with pages; any existing non-null entries in
619 * the array will be skipped
620 * @nofail: whether using __GFP_NOFAIL flag
621 *
622 * Return: 0 if all pages were able to be allocated;
623 * -ENOMEM otherwise, the partially allocated pages would be freed and
624 * the array slots zeroed
625 */
626int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array,
627 bool nofail)
628{
629 const gfp_t gfp = nofail ? (GFP_NOFS | __GFP_NOFAIL) : GFP_NOFS;
630 unsigned int allocated;
631
632 for (allocated = 0; allocated < nr_pages;) {
633 unsigned int last = allocated;
634
635 allocated = alloc_pages_bulk_array(gfp, nr_pages, page_array);
636 if (unlikely(allocated == last)) {
637 /* No progress, fail and do cleanup. */
638 for (int i = 0; i < allocated; i++) {
639 __free_page(page_array[i]);
640 page_array[i] = NULL;
641 }
642 return -ENOMEM;
643 }
644 }
645 return 0;
646}
647
648/*
649 * Populate needed folios for the extent buffer.
650 *
651 * For now, the folios populated are always in order 0 (aka, single page).
652 */
653static int alloc_eb_folio_array(struct extent_buffer *eb, bool nofail)
654{
655 struct page *page_array[INLINE_EXTENT_BUFFER_PAGES] = { 0 };
656 int num_pages = num_extent_pages(eb);
657 int ret;
658
659 ret = btrfs_alloc_page_array(num_pages, page_array, nofail);
660 if (ret < 0)
661 return ret;
662
663 for (int i = 0; i < num_pages; i++)
664 eb->folios[i] = page_folio(page_array[i]);
665 eb->folio_size = PAGE_SIZE;
666 eb->folio_shift = PAGE_SHIFT;
667 return 0;
668}
669
670static bool btrfs_bio_is_contig(struct btrfs_bio_ctrl *bio_ctrl,
671 struct folio *folio, u64 disk_bytenr,
672 unsigned int pg_offset)
673{
674 struct bio *bio = &bio_ctrl->bbio->bio;
675 struct bio_vec *bvec = bio_last_bvec_all(bio);
676 const sector_t sector = disk_bytenr >> SECTOR_SHIFT;
677 struct folio *bv_folio = page_folio(bvec->bv_page);
678
679 if (bio_ctrl->compress_type != BTRFS_COMPRESS_NONE) {
680 /*
681 * For compression, all IO should have its logical bytenr set
682 * to the starting bytenr of the compressed extent.
683 */
684 return bio->bi_iter.bi_sector == sector;
685 }
686
687 /*
688 * The contig check requires the following conditions to be met:
689 *
690 * 1) The folios are belonging to the same inode
691 * This is implied by the call chain.
692 *
693 * 2) The range has adjacent logical bytenr
694 *
695 * 3) The range has adjacent file offset
696 * This is required for the usage of btrfs_bio->file_offset.
697 */
698 return bio_end_sector(bio) == sector &&
699 folio_pos(bv_folio) + bvec->bv_offset + bvec->bv_len ==
700 folio_pos(folio) + pg_offset;
701}
702
703static void alloc_new_bio(struct btrfs_inode *inode,
704 struct btrfs_bio_ctrl *bio_ctrl,
705 u64 disk_bytenr, u64 file_offset)
706{
707 struct btrfs_fs_info *fs_info = inode->root->fs_info;
708 struct btrfs_bio *bbio;
709
710 bbio = btrfs_bio_alloc(BIO_MAX_VECS, bio_ctrl->opf, fs_info,
711 bio_ctrl->end_io_func, NULL);
712 bbio->bio.bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
713 bbio->inode = inode;
714 bbio->file_offset = file_offset;
715 bio_ctrl->bbio = bbio;
716 bio_ctrl->len_to_oe_boundary = U32_MAX;
717
718 /* Limit data write bios to the ordered boundary. */
719 if (bio_ctrl->wbc) {
720 struct btrfs_ordered_extent *ordered;
721
722 ordered = btrfs_lookup_ordered_extent(inode, file_offset);
723 if (ordered) {
724 bio_ctrl->len_to_oe_boundary = min_t(u32, U32_MAX,
725 ordered->file_offset +
726 ordered->disk_num_bytes - file_offset);
727 bbio->ordered = ordered;
728 }
729
730 /*
731 * Pick the last added device to support cgroup writeback. For
732 * multi-device file systems this means blk-cgroup policies have
733 * to always be set on the last added/replaced device.
734 * This is a bit odd but has been like that for a long time.
735 */
736 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
737 wbc_init_bio(bio_ctrl->wbc, &bbio->bio);
738 }
739}
740
741/*
742 * @disk_bytenr: logical bytenr where the write will be
743 * @page: page to add to the bio
744 * @size: portion of page that we want to write to
745 * @pg_offset: offset of the new bio or to check whether we are adding
746 * a contiguous page to the previous one
747 *
748 * The will either add the page into the existing @bio_ctrl->bbio, or allocate a
749 * new one in @bio_ctrl->bbio.
750 * The mirror number for this IO should already be initizlied in
751 * @bio_ctrl->mirror_num.
752 */
753static void submit_extent_folio(struct btrfs_bio_ctrl *bio_ctrl,
754 u64 disk_bytenr, struct folio *folio,
755 size_t size, unsigned long pg_offset)
756{
757 struct btrfs_inode *inode = folio_to_inode(folio);
758
759 ASSERT(pg_offset + size <= PAGE_SIZE);
760 ASSERT(bio_ctrl->end_io_func);
761
762 if (bio_ctrl->bbio &&
763 !btrfs_bio_is_contig(bio_ctrl, folio, disk_bytenr, pg_offset))
764 submit_one_bio(bio_ctrl);
765
766 do {
767 u32 len = size;
768
769 /* Allocate new bio if needed */
770 if (!bio_ctrl->bbio) {
771 alloc_new_bio(inode, bio_ctrl, disk_bytenr,
772 folio_pos(folio) + pg_offset);
773 }
774
775 /* Cap to the current ordered extent boundary if there is one. */
776 if (len > bio_ctrl->len_to_oe_boundary) {
777 ASSERT(bio_ctrl->compress_type == BTRFS_COMPRESS_NONE);
778 ASSERT(is_data_inode(inode));
779 len = bio_ctrl->len_to_oe_boundary;
780 }
781
782 if (!bio_add_folio(&bio_ctrl->bbio->bio, folio, len, pg_offset)) {
783 /* bio full: move on to a new one */
784 submit_one_bio(bio_ctrl);
785 continue;
786 }
787
788 if (bio_ctrl->wbc)
789 wbc_account_cgroup_owner(bio_ctrl->wbc, folio,
790 len);
791
792 size -= len;
793 pg_offset += len;
794 disk_bytenr += len;
795
796 /*
797 * len_to_oe_boundary defaults to U32_MAX, which isn't folio or
798 * sector aligned. alloc_new_bio() then sets it to the end of
799 * our ordered extent for writes into zoned devices.
800 *
801 * When len_to_oe_boundary is tracking an ordered extent, we
802 * trust the ordered extent code to align things properly, and
803 * the check above to cap our write to the ordered extent
804 * boundary is correct.
805 *
806 * When len_to_oe_boundary is U32_MAX, the cap above would
807 * result in a 4095 byte IO for the last folio right before
808 * we hit the bio limit of UINT_MAX. bio_add_folio() has all
809 * the checks required to make sure we don't overflow the bio,
810 * and we should just ignore len_to_oe_boundary completely
811 * unless we're using it to track an ordered extent.
812 *
813 * It's pretty hard to make a bio sized U32_MAX, but it can
814 * happen when the page cache is able to feed us contiguous
815 * folios for large extents.
816 */
817 if (bio_ctrl->len_to_oe_boundary != U32_MAX)
818 bio_ctrl->len_to_oe_boundary -= len;
819
820 /* Ordered extent boundary: move on to a new bio. */
821 if (bio_ctrl->len_to_oe_boundary == 0)
822 submit_one_bio(bio_ctrl);
823 } while (size);
824}
825
826static int attach_extent_buffer_folio(struct extent_buffer *eb,
827 struct folio *folio,
828 struct btrfs_subpage *prealloc)
829{
830 struct btrfs_fs_info *fs_info = eb->fs_info;
831 int ret = 0;
832
833 /*
834 * If the page is mapped to btree inode, we should hold the private
835 * lock to prevent race.
836 * For cloned or dummy extent buffers, their pages are not mapped and
837 * will not race with any other ebs.
838 */
839 if (folio->mapping)
840 lockdep_assert_held(&folio->mapping->i_private_lock);
841
842 if (fs_info->nodesize >= PAGE_SIZE) {
843 if (!folio_test_private(folio))
844 folio_attach_private(folio, eb);
845 else
846 WARN_ON(folio_get_private(folio) != eb);
847 return 0;
848 }
849
850 /* Already mapped, just free prealloc */
851 if (folio_test_private(folio)) {
852 btrfs_free_subpage(prealloc);
853 return 0;
854 }
855
856 if (prealloc)
857 /* Has preallocated memory for subpage */
858 folio_attach_private(folio, prealloc);
859 else
860 /* Do new allocation to attach subpage */
861 ret = btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_METADATA);
862 return ret;
863}
864
865int set_page_extent_mapped(struct page *page)
866{
867 return set_folio_extent_mapped(page_folio(page));
868}
869
870int set_folio_extent_mapped(struct folio *folio)
871{
872 struct btrfs_fs_info *fs_info;
873
874 ASSERT(folio->mapping);
875
876 if (folio_test_private(folio))
877 return 0;
878
879 fs_info = folio_to_fs_info(folio);
880
881 if (btrfs_is_subpage(fs_info, folio->mapping))
882 return btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_DATA);
883
884 folio_attach_private(folio, (void *)EXTENT_FOLIO_PRIVATE);
885 return 0;
886}
887
888void clear_folio_extent_mapped(struct folio *folio)
889{
890 struct btrfs_fs_info *fs_info;
891
892 ASSERT(folio->mapping);
893
894 if (!folio_test_private(folio))
895 return;
896
897 fs_info = folio_to_fs_info(folio);
898 if (btrfs_is_subpage(fs_info, folio->mapping))
899 return btrfs_detach_subpage(fs_info, folio);
900
901 folio_detach_private(folio);
902}
903
904static struct extent_map *get_extent_map(struct btrfs_inode *inode,
905 struct folio *folio, u64 start,
906 u64 len, struct extent_map **em_cached)
907{
908 struct extent_map *em;
909
910 ASSERT(em_cached);
911
912 if (*em_cached) {
913 em = *em_cached;
914 if (extent_map_in_tree(em) && start >= em->start &&
915 start < extent_map_end(em)) {
916 refcount_inc(&em->refs);
917 return em;
918 }
919
920 free_extent_map(em);
921 *em_cached = NULL;
922 }
923
924 em = btrfs_get_extent(inode, folio, start, len);
925 if (!IS_ERR(em)) {
926 BUG_ON(*em_cached);
927 refcount_inc(&em->refs);
928 *em_cached = em;
929 }
930
931 return em;
932}
933/*
934 * basic readpage implementation. Locked extent state structs are inserted
935 * into the tree that are removed when the IO is done (by the end_io
936 * handlers)
937 * XXX JDM: This needs looking at to ensure proper page locking
938 * return 0 on success, otherwise return error
939 */
940static int btrfs_do_readpage(struct folio *folio, struct extent_map **em_cached,
941 struct btrfs_bio_ctrl *bio_ctrl, u64 *prev_em_start)
942{
943 struct inode *inode = folio->mapping->host;
944 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
945 u64 start = folio_pos(folio);
946 const u64 end = start + PAGE_SIZE - 1;
947 u64 cur = start;
948 u64 extent_offset;
949 u64 last_byte = i_size_read(inode);
950 u64 block_start;
951 struct extent_map *em;
952 int ret = 0;
953 size_t pg_offset = 0;
954 size_t iosize;
955 size_t blocksize = fs_info->sectorsize;
956
957 ret = set_folio_extent_mapped(folio);
958 if (ret < 0) {
959 folio_unlock(folio);
960 return ret;
961 }
962
963 if (folio->index == last_byte >> folio_shift(folio)) {
964 size_t zero_offset = offset_in_folio(folio, last_byte);
965
966 if (zero_offset) {
967 iosize = folio_size(folio) - zero_offset;
968 folio_zero_range(folio, zero_offset, iosize);
969 }
970 }
971 bio_ctrl->end_io_func = end_bbio_data_read;
972 begin_folio_read(fs_info, folio);
973 while (cur <= end) {
974 enum btrfs_compression_type compress_type = BTRFS_COMPRESS_NONE;
975 bool force_bio_submit = false;
976 u64 disk_bytenr;
977
978 ASSERT(IS_ALIGNED(cur, fs_info->sectorsize));
979 if (cur >= last_byte) {
980 iosize = folio_size(folio) - pg_offset;
981 folio_zero_range(folio, pg_offset, iosize);
982 end_folio_read(folio, true, cur, iosize);
983 break;
984 }
985 em = get_extent_map(BTRFS_I(inode), folio, cur, end - cur + 1, em_cached);
986 if (IS_ERR(em)) {
987 end_folio_read(folio, false, cur, end + 1 - cur);
988 return PTR_ERR(em);
989 }
990 extent_offset = cur - em->start;
991 BUG_ON(extent_map_end(em) <= cur);
992 BUG_ON(end < cur);
993
994 compress_type = extent_map_compression(em);
995
996 iosize = min(extent_map_end(em) - cur, end - cur + 1);
997 iosize = ALIGN(iosize, blocksize);
998 if (compress_type != BTRFS_COMPRESS_NONE)
999 disk_bytenr = em->disk_bytenr;
1000 else
1001 disk_bytenr = extent_map_block_start(em) + extent_offset;
1002 block_start = extent_map_block_start(em);
1003 if (em->flags & EXTENT_FLAG_PREALLOC)
1004 block_start = EXTENT_MAP_HOLE;
1005
1006 /*
1007 * If we have a file range that points to a compressed extent
1008 * and it's followed by a consecutive file range that points
1009 * to the same compressed extent (possibly with a different
1010 * offset and/or length, so it either points to the whole extent
1011 * or only part of it), we must make sure we do not submit a
1012 * single bio to populate the folios for the 2 ranges because
1013 * this makes the compressed extent read zero out the folios
1014 * belonging to the 2nd range. Imagine the following scenario:
1015 *
1016 * File layout
1017 * [0 - 8K] [8K - 24K]
1018 * | |
1019 * | |
1020 * points to extent X, points to extent X,
1021 * offset 4K, length of 8K offset 0, length 16K
1022 *
1023 * [extent X, compressed length = 4K uncompressed length = 16K]
1024 *
1025 * If the bio to read the compressed extent covers both ranges,
1026 * it will decompress extent X into the folios belonging to the
1027 * first range and then it will stop, zeroing out the remaining
1028 * folios that belong to the other range that points to extent X.
1029 * So here we make sure we submit 2 bios, one for the first
1030 * range and another one for the third range. Both will target
1031 * the same physical extent from disk, but we can't currently
1032 * make the compressed bio endio callback populate the folios
1033 * for both ranges because each compressed bio is tightly
1034 * coupled with a single extent map, and each range can have
1035 * an extent map with a different offset value relative to the
1036 * uncompressed data of our extent and different lengths. This
1037 * is a corner case so we prioritize correctness over
1038 * non-optimal behavior (submitting 2 bios for the same extent).
1039 */
1040 if (compress_type != BTRFS_COMPRESS_NONE &&
1041 prev_em_start && *prev_em_start != (u64)-1 &&
1042 *prev_em_start != em->start)
1043 force_bio_submit = true;
1044
1045 if (prev_em_start)
1046 *prev_em_start = em->start;
1047
1048 free_extent_map(em);
1049 em = NULL;
1050
1051 /* we've found a hole, just zero and go on */
1052 if (block_start == EXTENT_MAP_HOLE) {
1053 folio_zero_range(folio, pg_offset, iosize);
1054
1055 end_folio_read(folio, true, cur, iosize);
1056 cur = cur + iosize;
1057 pg_offset += iosize;
1058 continue;
1059 }
1060 /* the get_extent function already copied into the folio */
1061 if (block_start == EXTENT_MAP_INLINE) {
1062 end_folio_read(folio, true, cur, iosize);
1063 cur = cur + iosize;
1064 pg_offset += iosize;
1065 continue;
1066 }
1067
1068 if (bio_ctrl->compress_type != compress_type) {
1069 submit_one_bio(bio_ctrl);
1070 bio_ctrl->compress_type = compress_type;
1071 }
1072
1073 if (force_bio_submit)
1074 submit_one_bio(bio_ctrl);
1075 submit_extent_folio(bio_ctrl, disk_bytenr, folio, iosize,
1076 pg_offset);
1077 cur = cur + iosize;
1078 pg_offset += iosize;
1079 }
1080
1081 return 0;
1082}
1083
1084int btrfs_read_folio(struct file *file, struct folio *folio)
1085{
1086 struct btrfs_inode *inode = folio_to_inode(folio);
1087 const u64 start = folio_pos(folio);
1088 const u64 end = start + folio_size(folio) - 1;
1089 struct extent_state *cached_state = NULL;
1090 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ };
1091 struct extent_map *em_cached = NULL;
1092 int ret;
1093
1094 btrfs_lock_and_flush_ordered_range(inode, start, end, &cached_state);
1095 ret = btrfs_do_readpage(folio, &em_cached, &bio_ctrl, NULL);
1096 unlock_extent(&inode->io_tree, start, end, &cached_state);
1097
1098 free_extent_map(em_cached);
1099
1100 /*
1101 * If btrfs_do_readpage() failed we will want to submit the assembled
1102 * bio to do the cleanup.
1103 */
1104 submit_one_bio(&bio_ctrl);
1105 return ret;
1106}
1107
1108static void set_delalloc_bitmap(struct folio *folio, unsigned long *delalloc_bitmap,
1109 u64 start, u32 len)
1110{
1111 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
1112 const u64 folio_start = folio_pos(folio);
1113 unsigned int start_bit;
1114 unsigned int nbits;
1115
1116 ASSERT(start >= folio_start && start + len <= folio_start + PAGE_SIZE);
1117 start_bit = (start - folio_start) >> fs_info->sectorsize_bits;
1118 nbits = len >> fs_info->sectorsize_bits;
1119 ASSERT(bitmap_test_range_all_zero(delalloc_bitmap, start_bit, nbits));
1120 bitmap_set(delalloc_bitmap, start_bit, nbits);
1121}
1122
1123static bool find_next_delalloc_bitmap(struct folio *folio,
1124 unsigned long *delalloc_bitmap, u64 start,
1125 u64 *found_start, u32 *found_len)
1126{
1127 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
1128 const u64 folio_start = folio_pos(folio);
1129 const unsigned int bitmap_size = fs_info->sectors_per_page;
1130 unsigned int start_bit;
1131 unsigned int first_zero;
1132 unsigned int first_set;
1133
1134 ASSERT(start >= folio_start && start < folio_start + PAGE_SIZE);
1135
1136 start_bit = (start - folio_start) >> fs_info->sectorsize_bits;
1137 first_set = find_next_bit(delalloc_bitmap, bitmap_size, start_bit);
1138 if (first_set >= bitmap_size)
1139 return false;
1140
1141 *found_start = folio_start + (first_set << fs_info->sectorsize_bits);
1142 first_zero = find_next_zero_bit(delalloc_bitmap, bitmap_size, first_set);
1143 *found_len = (first_zero - first_set) << fs_info->sectorsize_bits;
1144 return true;
1145}
1146
1147/*
1148 * Do all of the delayed allocation setup.
1149 *
1150 * Return >0 if all the dirty blocks are submitted async (compression) or inlined.
1151 * The @folio should no longer be touched (treat it as already unlocked).
1152 *
1153 * Return 0 if there is still dirty block that needs to be submitted through
1154 * extent_writepage_io().
1155 * bio_ctrl->submit_bitmap will indicate which blocks of the folio should be
1156 * submitted, and @folio is still kept locked.
1157 *
1158 * Return <0 if there is any error hit.
1159 * Any allocated ordered extent range covering this folio will be marked
1160 * finished (IOERR), and @folio is still kept locked.
1161 */
1162static noinline_for_stack int writepage_delalloc(struct btrfs_inode *inode,
1163 struct folio *folio,
1164 struct btrfs_bio_ctrl *bio_ctrl)
1165{
1166 struct btrfs_fs_info *fs_info = inode_to_fs_info(&inode->vfs_inode);
1167 struct writeback_control *wbc = bio_ctrl->wbc;
1168 const bool is_subpage = btrfs_is_subpage(fs_info, folio->mapping);
1169 const u64 page_start = folio_pos(folio);
1170 const u64 page_end = page_start + folio_size(folio) - 1;
1171 unsigned long delalloc_bitmap = 0;
1172 /*
1173 * Save the last found delalloc end. As the delalloc end can go beyond
1174 * page boundary, thus we cannot rely on subpage bitmap to locate the
1175 * last delalloc end.
1176 */
1177 u64 last_delalloc_end = 0;
1178 /*
1179 * The range end (exclusive) of the last successfully finished delalloc
1180 * range.
1181 * Any range covered by ordered extent must either be manually marked
1182 * finished (error handling), or has IO submitted (and finish the
1183 * ordered extent normally).
1184 *
1185 * This records the end of ordered extent cleanup if we hit an error.
1186 */
1187 u64 last_finished_delalloc_end = page_start;
1188 u64 delalloc_start = page_start;
1189 u64 delalloc_end = page_end;
1190 u64 delalloc_to_write = 0;
1191 int ret = 0;
1192 int bit;
1193
1194 /* Save the dirty bitmap as our submission bitmap will be a subset of it. */
1195 if (btrfs_is_subpage(fs_info, inode->vfs_inode.i_mapping)) {
1196 ASSERT(fs_info->sectors_per_page > 1);
1197 btrfs_get_subpage_dirty_bitmap(fs_info, folio, &bio_ctrl->submit_bitmap);
1198 } else {
1199 bio_ctrl->submit_bitmap = 1;
1200 }
1201
1202 for_each_set_bit(bit, &bio_ctrl->submit_bitmap, fs_info->sectors_per_page) {
1203 u64 start = page_start + (bit << fs_info->sectorsize_bits);
1204
1205 btrfs_folio_set_lock(fs_info, folio, start, fs_info->sectorsize);
1206 }
1207
1208 /* Lock all (subpage) delalloc ranges inside the folio first. */
1209 while (delalloc_start < page_end) {
1210 delalloc_end = page_end;
1211 if (!find_lock_delalloc_range(&inode->vfs_inode, folio,
1212 &delalloc_start, &delalloc_end)) {
1213 delalloc_start = delalloc_end + 1;
1214 continue;
1215 }
1216 set_delalloc_bitmap(folio, &delalloc_bitmap, delalloc_start,
1217 min(delalloc_end, page_end) + 1 - delalloc_start);
1218 last_delalloc_end = delalloc_end;
1219 delalloc_start = delalloc_end + 1;
1220 }
1221 delalloc_start = page_start;
1222
1223 if (!last_delalloc_end)
1224 goto out;
1225
1226 /* Run the delalloc ranges for the above locked ranges. */
1227 while (delalloc_start < page_end) {
1228 u64 found_start;
1229 u32 found_len;
1230 bool found;
1231
1232 if (!is_subpage) {
1233 /*
1234 * For non-subpage case, the found delalloc range must
1235 * cover this folio and there must be only one locked
1236 * delalloc range.
1237 */
1238 found_start = page_start;
1239 found_len = last_delalloc_end + 1 - found_start;
1240 found = true;
1241 } else {
1242 found = find_next_delalloc_bitmap(folio, &delalloc_bitmap,
1243 delalloc_start, &found_start, &found_len);
1244 }
1245 if (!found)
1246 break;
1247 /*
1248 * The subpage range covers the last sector, the delalloc range may
1249 * end beyond the folio boundary, use the saved delalloc_end
1250 * instead.
1251 */
1252 if (found_start + found_len >= page_end)
1253 found_len = last_delalloc_end + 1 - found_start;
1254
1255 if (ret >= 0) {
1256 /*
1257 * Some delalloc range may be created by previous folios.
1258 * Thus we still need to clean up this range during error
1259 * handling.
1260 */
1261 last_finished_delalloc_end = found_start;
1262 /* No errors hit so far, run the current delalloc range. */
1263 ret = btrfs_run_delalloc_range(inode, folio,
1264 found_start,
1265 found_start + found_len - 1,
1266 wbc);
1267 if (ret >= 0)
1268 last_finished_delalloc_end = found_start + found_len;
1269 } else {
1270 /*
1271 * We've hit an error during previous delalloc range,
1272 * have to cleanup the remaining locked ranges.
1273 */
1274 unlock_extent(&inode->io_tree, found_start,
1275 found_start + found_len - 1, NULL);
1276 __unlock_for_delalloc(&inode->vfs_inode, folio,
1277 found_start,
1278 found_start + found_len - 1);
1279 }
1280
1281 /*
1282 * We have some ranges that's going to be submitted asynchronously
1283 * (compression or inline). These range have their own control
1284 * on when to unlock the pages. We should not touch them
1285 * anymore, so clear the range from the submission bitmap.
1286 */
1287 if (ret > 0) {
1288 unsigned int start_bit = (found_start - page_start) >>
1289 fs_info->sectorsize_bits;
1290 unsigned int end_bit = (min(page_end + 1, found_start + found_len) -
1291 page_start) >> fs_info->sectorsize_bits;
1292 bitmap_clear(&bio_ctrl->submit_bitmap, start_bit, end_bit - start_bit);
1293 }
1294 /*
1295 * Above btrfs_run_delalloc_range() may have unlocked the folio,
1296 * thus for the last range, we cannot touch the folio anymore.
1297 */
1298 if (found_start + found_len >= last_delalloc_end + 1)
1299 break;
1300
1301 delalloc_start = found_start + found_len;
1302 }
1303 /*
1304 * It's possible we had some ordered extents created before we hit
1305 * an error, cleanup non-async successfully created delalloc ranges.
1306 */
1307 if (unlikely(ret < 0)) {
1308 unsigned int bitmap_size = min(
1309 (last_finished_delalloc_end - page_start) >>
1310 fs_info->sectorsize_bits,
1311 fs_info->sectors_per_page);
1312
1313 for_each_set_bit(bit, &bio_ctrl->submit_bitmap, bitmap_size)
1314 btrfs_mark_ordered_io_finished(inode, folio,
1315 page_start + (bit << fs_info->sectorsize_bits),
1316 fs_info->sectorsize, false);
1317 return ret;
1318 }
1319out:
1320 if (last_delalloc_end)
1321 delalloc_end = last_delalloc_end;
1322 else
1323 delalloc_end = page_end;
1324 /*
1325 * delalloc_end is already one less than the total length, so
1326 * we don't subtract one from PAGE_SIZE
1327 */
1328 delalloc_to_write +=
1329 DIV_ROUND_UP(delalloc_end + 1 - page_start, PAGE_SIZE);
1330
1331 /*
1332 * If all ranges are submitted asynchronously, we just need to account
1333 * for them here.
1334 */
1335 if (bitmap_empty(&bio_ctrl->submit_bitmap, fs_info->sectors_per_page)) {
1336 wbc->nr_to_write -= delalloc_to_write;
1337 return 1;
1338 }
1339
1340 if (wbc->nr_to_write < delalloc_to_write) {
1341 int thresh = 8192;
1342
1343 if (delalloc_to_write < thresh * 2)
1344 thresh = delalloc_to_write;
1345 wbc->nr_to_write = min_t(u64, delalloc_to_write,
1346 thresh);
1347 }
1348
1349 return 0;
1350}
1351
1352/*
1353 * Return 0 if we have submitted or queued the sector for submission.
1354 * Return <0 for critical errors.
1355 *
1356 * Caller should make sure filepos < i_size and handle filepos >= i_size case.
1357 */
1358static int submit_one_sector(struct btrfs_inode *inode,
1359 struct folio *folio,
1360 u64 filepos, struct btrfs_bio_ctrl *bio_ctrl,
1361 loff_t i_size)
1362{
1363 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1364 struct extent_map *em;
1365 u64 block_start;
1366 u64 disk_bytenr;
1367 u64 extent_offset;
1368 u64 em_end;
1369 const u32 sectorsize = fs_info->sectorsize;
1370
1371 ASSERT(IS_ALIGNED(filepos, sectorsize));
1372
1373 /* @filepos >= i_size case should be handled by the caller. */
1374 ASSERT(filepos < i_size);
1375
1376 em = btrfs_get_extent(inode, NULL, filepos, sectorsize);
1377 if (IS_ERR(em))
1378 return PTR_ERR_OR_ZERO(em);
1379
1380 extent_offset = filepos - em->start;
1381 em_end = extent_map_end(em);
1382 ASSERT(filepos <= em_end);
1383 ASSERT(IS_ALIGNED(em->start, sectorsize));
1384 ASSERT(IS_ALIGNED(em->len, sectorsize));
1385
1386 block_start = extent_map_block_start(em);
1387 disk_bytenr = extent_map_block_start(em) + extent_offset;
1388
1389 ASSERT(!extent_map_is_compressed(em));
1390 ASSERT(block_start != EXTENT_MAP_HOLE);
1391 ASSERT(block_start != EXTENT_MAP_INLINE);
1392
1393 free_extent_map(em);
1394 em = NULL;
1395
1396 /*
1397 * Although the PageDirty bit is cleared before entering this
1398 * function, subpage dirty bit is not cleared.
1399 * So clear subpage dirty bit here so next time we won't submit
1400 * a folio for a range already written to disk.
1401 */
1402 btrfs_folio_clear_dirty(fs_info, folio, filepos, sectorsize);
1403 btrfs_folio_set_writeback(fs_info, folio, filepos, sectorsize);
1404 /*
1405 * Above call should set the whole folio with writeback flag, even
1406 * just for a single subpage sector.
1407 * As long as the folio is properly locked and the range is correct,
1408 * we should always get the folio with writeback flag.
1409 */
1410 ASSERT(folio_test_writeback(folio));
1411
1412 submit_extent_folio(bio_ctrl, disk_bytenr, folio,
1413 sectorsize, filepos - folio_pos(folio));
1414 return 0;
1415}
1416
1417/*
1418 * Helper for extent_writepage(). This calls the writepage start hooks,
1419 * and does the loop to map the page into extents and bios.
1420 *
1421 * We return 1 if the IO is started and the page is unlocked,
1422 * 0 if all went well (page still locked)
1423 * < 0 if there were errors (page still locked)
1424 */
1425static noinline_for_stack int extent_writepage_io(struct btrfs_inode *inode,
1426 struct folio *folio,
1427 u64 start, u32 len,
1428 struct btrfs_bio_ctrl *bio_ctrl,
1429 loff_t i_size)
1430{
1431 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1432 unsigned long range_bitmap = 0;
1433 bool submitted_io = false;
1434 bool error = false;
1435 const u64 folio_start = folio_pos(folio);
1436 u64 cur;
1437 int bit;
1438 int ret = 0;
1439
1440 ASSERT(start >= folio_start &&
1441 start + len <= folio_start + folio_size(folio));
1442
1443 ret = btrfs_writepage_cow_fixup(folio);
1444 if (ret) {
1445 /* Fixup worker will requeue */
1446 folio_redirty_for_writepage(bio_ctrl->wbc, folio);
1447 folio_unlock(folio);
1448 return 1;
1449 }
1450
1451 for (cur = start; cur < start + len; cur += fs_info->sectorsize)
1452 set_bit((cur - folio_start) >> fs_info->sectorsize_bits, &range_bitmap);
1453 bitmap_and(&bio_ctrl->submit_bitmap, &bio_ctrl->submit_bitmap, &range_bitmap,
1454 fs_info->sectors_per_page);
1455
1456 bio_ctrl->end_io_func = end_bbio_data_write;
1457
1458 for_each_set_bit(bit, &bio_ctrl->submit_bitmap, fs_info->sectors_per_page) {
1459 cur = folio_pos(folio) + (bit << fs_info->sectorsize_bits);
1460
1461 if (cur >= i_size) {
1462 btrfs_mark_ordered_io_finished(inode, folio, cur,
1463 start + len - cur, true);
1464 /*
1465 * This range is beyond i_size, thus we don't need to
1466 * bother writing back.
1467 * But we still need to clear the dirty subpage bit, or
1468 * the next time the folio gets dirtied, we will try to
1469 * writeback the sectors with subpage dirty bits,
1470 * causing writeback without ordered extent.
1471 */
1472 btrfs_folio_clear_dirty(fs_info, folio, cur,
1473 start + len - cur);
1474 break;
1475 }
1476 ret = submit_one_sector(inode, folio, cur, bio_ctrl, i_size);
1477 if (unlikely(ret < 0)) {
1478 /*
1479 * bio_ctrl may contain a bio crossing several folios.
1480 * Submit it immediately so that the bio has a chance
1481 * to finish normally, other than marked as error.
1482 */
1483 submit_one_bio(bio_ctrl);
1484 /*
1485 * Failed to grab the extent map which should be very rare.
1486 * Since there is no bio submitted to finish the ordered
1487 * extent, we have to manually finish this sector.
1488 */
1489 btrfs_mark_ordered_io_finished(inode, folio, cur,
1490 fs_info->sectorsize, false);
1491 error = true;
1492 continue;
1493 }
1494 submitted_io = true;
1495 }
1496
1497 /*
1498 * If we didn't submitted any sector (>= i_size), folio dirty get
1499 * cleared but PAGECACHE_TAG_DIRTY is not cleared (only cleared
1500 * by folio_start_writeback() if the folio is not dirty).
1501 *
1502 * Here we set writeback and clear for the range. If the full folio
1503 * is no longer dirty then we clear the PAGECACHE_TAG_DIRTY tag.
1504 *
1505 * If we hit any error, the corresponding sector will still be dirty
1506 * thus no need to clear PAGECACHE_TAG_DIRTY.
1507 */
1508 if (!submitted_io && !error) {
1509 btrfs_folio_set_writeback(fs_info, folio, start, len);
1510 btrfs_folio_clear_writeback(fs_info, folio, start, len);
1511 }
1512 return ret;
1513}
1514
1515/*
1516 * the writepage semantics are similar to regular writepage. extent
1517 * records are inserted to lock ranges in the tree, and as dirty areas
1518 * are found, they are marked writeback. Then the lock bits are removed
1519 * and the end_io handler clears the writeback ranges
1520 *
1521 * Return 0 if everything goes well.
1522 * Return <0 for error.
1523 */
1524static int extent_writepage(struct folio *folio, struct btrfs_bio_ctrl *bio_ctrl)
1525{
1526 struct btrfs_inode *inode = BTRFS_I(folio->mapping->host);
1527 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1528 int ret;
1529 size_t pg_offset;
1530 loff_t i_size = i_size_read(&inode->vfs_inode);
1531 unsigned long end_index = i_size >> PAGE_SHIFT;
1532
1533 trace_extent_writepage(folio, &inode->vfs_inode, bio_ctrl->wbc);
1534
1535 WARN_ON(!folio_test_locked(folio));
1536
1537 pg_offset = offset_in_folio(folio, i_size);
1538 if (folio->index > end_index ||
1539 (folio->index == end_index && !pg_offset)) {
1540 folio_invalidate(folio, 0, folio_size(folio));
1541 folio_unlock(folio);
1542 return 0;
1543 }
1544
1545 if (folio->index == end_index)
1546 folio_zero_range(folio, pg_offset, folio_size(folio) - pg_offset);
1547
1548 /*
1549 * Default to unlock the whole folio.
1550 * The proper bitmap can only be initialized until writepage_delalloc().
1551 */
1552 bio_ctrl->submit_bitmap = (unsigned long)-1;
1553 ret = set_folio_extent_mapped(folio);
1554 if (ret < 0)
1555 goto done;
1556
1557 ret = writepage_delalloc(inode, folio, bio_ctrl);
1558 if (ret == 1)
1559 return 0;
1560 if (ret)
1561 goto done;
1562
1563 ret = extent_writepage_io(inode, folio, folio_pos(folio),
1564 PAGE_SIZE, bio_ctrl, i_size);
1565 if (ret == 1)
1566 return 0;
1567
1568 bio_ctrl->wbc->nr_to_write--;
1569
1570done:
1571 if (ret < 0)
1572 mapping_set_error(folio->mapping, ret);
1573 /*
1574 * Only unlock ranges that are submitted. As there can be some async
1575 * submitted ranges inside the folio.
1576 */
1577 btrfs_folio_end_lock_bitmap(fs_info, folio, bio_ctrl->submit_bitmap);
1578 ASSERT(ret <= 0);
1579 return ret;
1580}
1581
1582void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
1583{
1584 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
1585 TASK_UNINTERRUPTIBLE);
1586}
1587
1588/*
1589 * Lock extent buffer status and pages for writeback.
1590 *
1591 * Return %false if the extent buffer doesn't need to be submitted (e.g. the
1592 * extent buffer is not dirty)
1593 * Return %true is the extent buffer is submitted to bio.
1594 */
1595static noinline_for_stack bool lock_extent_buffer_for_io(struct extent_buffer *eb,
1596 struct writeback_control *wbc)
1597{
1598 struct btrfs_fs_info *fs_info = eb->fs_info;
1599 bool ret = false;
1600
1601 btrfs_tree_lock(eb);
1602 while (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
1603 btrfs_tree_unlock(eb);
1604 if (wbc->sync_mode != WB_SYNC_ALL)
1605 return false;
1606 wait_on_extent_buffer_writeback(eb);
1607 btrfs_tree_lock(eb);
1608 }
1609
1610 /*
1611 * We need to do this to prevent races in people who check if the eb is
1612 * under IO since we can end up having no IO bits set for a short period
1613 * of time.
1614 */
1615 spin_lock(&eb->refs_lock);
1616 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
1617 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1618 spin_unlock(&eb->refs_lock);
1619 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
1620 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
1621 -eb->len,
1622 fs_info->dirty_metadata_batch);
1623 ret = true;
1624 } else {
1625 spin_unlock(&eb->refs_lock);
1626 }
1627 btrfs_tree_unlock(eb);
1628 return ret;
1629}
1630
1631static void set_btree_ioerr(struct extent_buffer *eb)
1632{
1633 struct btrfs_fs_info *fs_info = eb->fs_info;
1634
1635 set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1636
1637 /*
1638 * A read may stumble upon this buffer later, make sure that it gets an
1639 * error and knows there was an error.
1640 */
1641 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
1642
1643 /*
1644 * We need to set the mapping with the io error as well because a write
1645 * error will flip the file system readonly, and then syncfs() will
1646 * return a 0 because we are readonly if we don't modify the err seq for
1647 * the superblock.
1648 */
1649 mapping_set_error(eb->fs_info->btree_inode->i_mapping, -EIO);
1650
1651 /*
1652 * If writeback for a btree extent that doesn't belong to a log tree
1653 * failed, increment the counter transaction->eb_write_errors.
1654 * We do this because while the transaction is running and before it's
1655 * committing (when we call filemap_fdata[write|wait]_range against
1656 * the btree inode), we might have
1657 * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
1658 * returns an error or an error happens during writeback, when we're
1659 * committing the transaction we wouldn't know about it, since the pages
1660 * can be no longer dirty nor marked anymore for writeback (if a
1661 * subsequent modification to the extent buffer didn't happen before the
1662 * transaction commit), which makes filemap_fdata[write|wait]_range not
1663 * able to find the pages which contain errors at transaction
1664 * commit time. So if this happens we must abort the transaction,
1665 * otherwise we commit a super block with btree roots that point to
1666 * btree nodes/leafs whose content on disk is invalid - either garbage
1667 * or the content of some node/leaf from a past generation that got
1668 * cowed or deleted and is no longer valid.
1669 *
1670 * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
1671 * not be enough - we need to distinguish between log tree extents vs
1672 * non-log tree extents, and the next filemap_fdatawait_range() call
1673 * will catch and clear such errors in the mapping - and that call might
1674 * be from a log sync and not from a transaction commit. Also, checking
1675 * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
1676 * not done and would not be reliable - the eb might have been released
1677 * from memory and reading it back again means that flag would not be
1678 * set (since it's a runtime flag, not persisted on disk).
1679 *
1680 * Using the flags below in the btree inode also makes us achieve the
1681 * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
1682 * writeback for all dirty pages and before filemap_fdatawait_range()
1683 * is called, the writeback for all dirty pages had already finished
1684 * with errors - because we were not using AS_EIO/AS_ENOSPC,
1685 * filemap_fdatawait_range() would return success, as it could not know
1686 * that writeback errors happened (the pages were no longer tagged for
1687 * writeback).
1688 */
1689 switch (eb->log_index) {
1690 case -1:
1691 set_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags);
1692 break;
1693 case 0:
1694 set_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
1695 break;
1696 case 1:
1697 set_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
1698 break;
1699 default:
1700 BUG(); /* unexpected, logic error */
1701 }
1702}
1703
1704/*
1705 * The endio specific version which won't touch any unsafe spinlock in endio
1706 * context.
1707 */
1708static struct extent_buffer *find_extent_buffer_nolock(
1709 const struct btrfs_fs_info *fs_info, u64 start)
1710{
1711 struct extent_buffer *eb;
1712
1713 rcu_read_lock();
1714 eb = radix_tree_lookup(&fs_info->buffer_radix,
1715 start >> fs_info->sectorsize_bits);
1716 if (eb && atomic_inc_not_zero(&eb->refs)) {
1717 rcu_read_unlock();
1718 return eb;
1719 }
1720 rcu_read_unlock();
1721 return NULL;
1722}
1723
1724static void end_bbio_meta_write(struct btrfs_bio *bbio)
1725{
1726 struct extent_buffer *eb = bbio->private;
1727 struct btrfs_fs_info *fs_info = eb->fs_info;
1728 bool uptodate = !bbio->bio.bi_status;
1729 struct folio_iter fi;
1730 u32 bio_offset = 0;
1731
1732 if (!uptodate)
1733 set_btree_ioerr(eb);
1734
1735 bio_for_each_folio_all(fi, &bbio->bio) {
1736 u64 start = eb->start + bio_offset;
1737 struct folio *folio = fi.folio;
1738 u32 len = fi.length;
1739
1740 btrfs_folio_clear_writeback(fs_info, folio, start, len);
1741 bio_offset += len;
1742 }
1743
1744 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1745 smp_mb__after_atomic();
1746 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
1747
1748 bio_put(&bbio->bio);
1749}
1750
1751static void prepare_eb_write(struct extent_buffer *eb)
1752{
1753 u32 nritems;
1754 unsigned long start;
1755 unsigned long end;
1756
1757 clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1758
1759 /* Set btree blocks beyond nritems with 0 to avoid stale content */
1760 nritems = btrfs_header_nritems(eb);
1761 if (btrfs_header_level(eb) > 0) {
1762 end = btrfs_node_key_ptr_offset(eb, nritems);
1763 memzero_extent_buffer(eb, end, eb->len - end);
1764 } else {
1765 /*
1766 * Leaf:
1767 * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
1768 */
1769 start = btrfs_item_nr_offset(eb, nritems);
1770 end = btrfs_item_nr_offset(eb, 0);
1771 if (nritems == 0)
1772 end += BTRFS_LEAF_DATA_SIZE(eb->fs_info);
1773 else
1774 end += btrfs_item_offset(eb, nritems - 1);
1775 memzero_extent_buffer(eb, start, end - start);
1776 }
1777}
1778
1779static noinline_for_stack void write_one_eb(struct extent_buffer *eb,
1780 struct writeback_control *wbc)
1781{
1782 struct btrfs_fs_info *fs_info = eb->fs_info;
1783 struct btrfs_bio *bbio;
1784
1785 prepare_eb_write(eb);
1786
1787 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
1788 REQ_OP_WRITE | REQ_META | wbc_to_write_flags(wbc),
1789 eb->fs_info, end_bbio_meta_write, eb);
1790 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
1791 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
1792 wbc_init_bio(wbc, &bbio->bio);
1793 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
1794 bbio->file_offset = eb->start;
1795 if (fs_info->nodesize < PAGE_SIZE) {
1796 struct folio *folio = eb->folios[0];
1797 bool ret;
1798
1799 folio_lock(folio);
1800 btrfs_subpage_set_writeback(fs_info, folio, eb->start, eb->len);
1801 if (btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start,
1802 eb->len)) {
1803 folio_clear_dirty_for_io(folio);
1804 wbc->nr_to_write--;
1805 }
1806 ret = bio_add_folio(&bbio->bio, folio, eb->len,
1807 eb->start - folio_pos(folio));
1808 ASSERT(ret);
1809 wbc_account_cgroup_owner(wbc, folio, eb->len);
1810 folio_unlock(folio);
1811 } else {
1812 int num_folios = num_extent_folios(eb);
1813
1814 for (int i = 0; i < num_folios; i++) {
1815 struct folio *folio = eb->folios[i];
1816 bool ret;
1817
1818 folio_lock(folio);
1819 folio_clear_dirty_for_io(folio);
1820 folio_start_writeback(folio);
1821 ret = bio_add_folio(&bbio->bio, folio, eb->folio_size, 0);
1822 ASSERT(ret);
1823 wbc_account_cgroup_owner(wbc, folio, eb->folio_size);
1824 wbc->nr_to_write -= folio_nr_pages(folio);
1825 folio_unlock(folio);
1826 }
1827 }
1828 btrfs_submit_bbio(bbio, 0);
1829}
1830
1831/*
1832 * Submit one subpage btree page.
1833 *
1834 * The main difference to submit_eb_page() is:
1835 * - Page locking
1836 * For subpage, we don't rely on page locking at all.
1837 *
1838 * - Flush write bio
1839 * We only flush bio if we may be unable to fit current extent buffers into
1840 * current bio.
1841 *
1842 * Return >=0 for the number of submitted extent buffers.
1843 * Return <0 for fatal error.
1844 */
1845static int submit_eb_subpage(struct folio *folio, struct writeback_control *wbc)
1846{
1847 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
1848 int submitted = 0;
1849 u64 folio_start = folio_pos(folio);
1850 int bit_start = 0;
1851 int sectors_per_node = fs_info->nodesize >> fs_info->sectorsize_bits;
1852
1853 /* Lock and write each dirty extent buffers in the range */
1854 while (bit_start < fs_info->sectors_per_page) {
1855 struct btrfs_subpage *subpage = folio_get_private(folio);
1856 struct extent_buffer *eb;
1857 unsigned long flags;
1858 u64 start;
1859
1860 /*
1861 * Take private lock to ensure the subpage won't be detached
1862 * in the meantime.
1863 */
1864 spin_lock(&folio->mapping->i_private_lock);
1865 if (!folio_test_private(folio)) {
1866 spin_unlock(&folio->mapping->i_private_lock);
1867 break;
1868 }
1869 spin_lock_irqsave(&subpage->lock, flags);
1870 if (!test_bit(bit_start + btrfs_bitmap_nr_dirty * fs_info->sectors_per_page,
1871 subpage->bitmaps)) {
1872 spin_unlock_irqrestore(&subpage->lock, flags);
1873 spin_unlock(&folio->mapping->i_private_lock);
1874 bit_start++;
1875 continue;
1876 }
1877
1878 start = folio_start + bit_start * fs_info->sectorsize;
1879 bit_start += sectors_per_node;
1880
1881 /*
1882 * Here we just want to grab the eb without touching extra
1883 * spin locks, so call find_extent_buffer_nolock().
1884 */
1885 eb = find_extent_buffer_nolock(fs_info, start);
1886 spin_unlock_irqrestore(&subpage->lock, flags);
1887 spin_unlock(&folio->mapping->i_private_lock);
1888
1889 /*
1890 * The eb has already reached 0 refs thus find_extent_buffer()
1891 * doesn't return it. We don't need to write back such eb
1892 * anyway.
1893 */
1894 if (!eb)
1895 continue;
1896
1897 if (lock_extent_buffer_for_io(eb, wbc)) {
1898 write_one_eb(eb, wbc);
1899 submitted++;
1900 }
1901 free_extent_buffer(eb);
1902 }
1903 return submitted;
1904}
1905
1906/*
1907 * Submit all page(s) of one extent buffer.
1908 *
1909 * @page: the page of one extent buffer
1910 * @eb_context: to determine if we need to submit this page, if current page
1911 * belongs to this eb, we don't need to submit
1912 *
1913 * The caller should pass each page in their bytenr order, and here we use
1914 * @eb_context to determine if we have submitted pages of one extent buffer.
1915 *
1916 * If we have, we just skip until we hit a new page that doesn't belong to
1917 * current @eb_context.
1918 *
1919 * If not, we submit all the page(s) of the extent buffer.
1920 *
1921 * Return >0 if we have submitted the extent buffer successfully.
1922 * Return 0 if we don't need to submit the page, as it's already submitted by
1923 * previous call.
1924 * Return <0 for fatal error.
1925 */
1926static int submit_eb_page(struct folio *folio, struct btrfs_eb_write_context *ctx)
1927{
1928 struct writeback_control *wbc = ctx->wbc;
1929 struct address_space *mapping = folio->mapping;
1930 struct extent_buffer *eb;
1931 int ret;
1932
1933 if (!folio_test_private(folio))
1934 return 0;
1935
1936 if (folio_to_fs_info(folio)->nodesize < PAGE_SIZE)
1937 return submit_eb_subpage(folio, wbc);
1938
1939 spin_lock(&mapping->i_private_lock);
1940 if (!folio_test_private(folio)) {
1941 spin_unlock(&mapping->i_private_lock);
1942 return 0;
1943 }
1944
1945 eb = folio_get_private(folio);
1946
1947 /*
1948 * Shouldn't happen and normally this would be a BUG_ON but no point
1949 * crashing the machine for something we can survive anyway.
1950 */
1951 if (WARN_ON(!eb)) {
1952 spin_unlock(&mapping->i_private_lock);
1953 return 0;
1954 }
1955
1956 if (eb == ctx->eb) {
1957 spin_unlock(&mapping->i_private_lock);
1958 return 0;
1959 }
1960 ret = atomic_inc_not_zero(&eb->refs);
1961 spin_unlock(&mapping->i_private_lock);
1962 if (!ret)
1963 return 0;
1964
1965 ctx->eb = eb;
1966
1967 ret = btrfs_check_meta_write_pointer(eb->fs_info, ctx);
1968 if (ret) {
1969 if (ret == -EBUSY)
1970 ret = 0;
1971 free_extent_buffer(eb);
1972 return ret;
1973 }
1974
1975 if (!lock_extent_buffer_for_io(eb, wbc)) {
1976 free_extent_buffer(eb);
1977 return 0;
1978 }
1979 /* Implies write in zoned mode. */
1980 if (ctx->zoned_bg) {
1981 /* Mark the last eb in the block group. */
1982 btrfs_schedule_zone_finish_bg(ctx->zoned_bg, eb);
1983 ctx->zoned_bg->meta_write_pointer += eb->len;
1984 }
1985 write_one_eb(eb, wbc);
1986 free_extent_buffer(eb);
1987 return 1;
1988}
1989
1990int btree_write_cache_pages(struct address_space *mapping,
1991 struct writeback_control *wbc)
1992{
1993 struct btrfs_eb_write_context ctx = { .wbc = wbc };
1994 struct btrfs_fs_info *fs_info = inode_to_fs_info(mapping->host);
1995 int ret = 0;
1996 int done = 0;
1997 int nr_to_write_done = 0;
1998 struct folio_batch fbatch;
1999 unsigned int nr_folios;
2000 pgoff_t index;
2001 pgoff_t end; /* Inclusive */
2002 int scanned = 0;
2003 xa_mark_t tag;
2004
2005 folio_batch_init(&fbatch);
2006 if (wbc->range_cyclic) {
2007 index = mapping->writeback_index; /* Start from prev offset */
2008 end = -1;
2009 /*
2010 * Start from the beginning does not need to cycle over the
2011 * range, mark it as scanned.
2012 */
2013 scanned = (index == 0);
2014 } else {
2015 index = wbc->range_start >> PAGE_SHIFT;
2016 end = wbc->range_end >> PAGE_SHIFT;
2017 scanned = 1;
2018 }
2019 if (wbc->sync_mode == WB_SYNC_ALL)
2020 tag = PAGECACHE_TAG_TOWRITE;
2021 else
2022 tag = PAGECACHE_TAG_DIRTY;
2023 btrfs_zoned_meta_io_lock(fs_info);
2024retry:
2025 if (wbc->sync_mode == WB_SYNC_ALL)
2026 tag_pages_for_writeback(mapping, index, end);
2027 while (!done && !nr_to_write_done && (index <= end) &&
2028 (nr_folios = filemap_get_folios_tag(mapping, &index, end,
2029 tag, &fbatch))) {
2030 unsigned i;
2031
2032 for (i = 0; i < nr_folios; i++) {
2033 struct folio *folio = fbatch.folios[i];
2034
2035 ret = submit_eb_page(folio, &ctx);
2036 if (ret == 0)
2037 continue;
2038 if (ret < 0) {
2039 done = 1;
2040 break;
2041 }
2042
2043 /*
2044 * the filesystem may choose to bump up nr_to_write.
2045 * We have to make sure to honor the new nr_to_write
2046 * at any time
2047 */
2048 nr_to_write_done = wbc->nr_to_write <= 0;
2049 }
2050 folio_batch_release(&fbatch);
2051 cond_resched();
2052 }
2053 if (!scanned && !done) {
2054 /*
2055 * We hit the last page and there is more work to be done: wrap
2056 * back to the start of the file
2057 */
2058 scanned = 1;
2059 index = 0;
2060 goto retry;
2061 }
2062 /*
2063 * If something went wrong, don't allow any metadata write bio to be
2064 * submitted.
2065 *
2066 * This would prevent use-after-free if we had dirty pages not
2067 * cleaned up, which can still happen by fuzzed images.
2068 *
2069 * - Bad extent tree
2070 * Allowing existing tree block to be allocated for other trees.
2071 *
2072 * - Log tree operations
2073 * Exiting tree blocks get allocated to log tree, bumps its
2074 * generation, then get cleaned in tree re-balance.
2075 * Such tree block will not be written back, since it's clean,
2076 * thus no WRITTEN flag set.
2077 * And after log writes back, this tree block is not traced by
2078 * any dirty extent_io_tree.
2079 *
2080 * - Offending tree block gets re-dirtied from its original owner
2081 * Since it has bumped generation, no WRITTEN flag, it can be
2082 * reused without COWing. This tree block will not be traced
2083 * by btrfs_transaction::dirty_pages.
2084 *
2085 * Now such dirty tree block will not be cleaned by any dirty
2086 * extent io tree. Thus we don't want to submit such wild eb
2087 * if the fs already has error.
2088 *
2089 * We can get ret > 0 from submit_extent_folio() indicating how many ebs
2090 * were submitted. Reset it to 0 to avoid false alerts for the caller.
2091 */
2092 if (ret > 0)
2093 ret = 0;
2094 if (!ret && BTRFS_FS_ERROR(fs_info))
2095 ret = -EROFS;
2096
2097 if (ctx.zoned_bg)
2098 btrfs_put_block_group(ctx.zoned_bg);
2099 btrfs_zoned_meta_io_unlock(fs_info);
2100 return ret;
2101}
2102
2103/*
2104 * Walk the list of dirty pages of the given address space and write all of them.
2105 *
2106 * @mapping: address space structure to write
2107 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2108 * @bio_ctrl: holds context for the write, namely the bio
2109 *
2110 * If a page is already under I/O, write_cache_pages() skips it, even
2111 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2112 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2113 * and msync() need to guarantee that all the data which was dirty at the time
2114 * the call was made get new I/O started against them. If wbc->sync_mode is
2115 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2116 * existing IO to complete.
2117 */
2118static int extent_write_cache_pages(struct address_space *mapping,
2119 struct btrfs_bio_ctrl *bio_ctrl)
2120{
2121 struct writeback_control *wbc = bio_ctrl->wbc;
2122 struct inode *inode = mapping->host;
2123 int ret = 0;
2124 int done = 0;
2125 int nr_to_write_done = 0;
2126 struct folio_batch fbatch;
2127 unsigned int nr_folios;
2128 pgoff_t index;
2129 pgoff_t end; /* Inclusive */
2130 pgoff_t done_index;
2131 int range_whole = 0;
2132 int scanned = 0;
2133 xa_mark_t tag;
2134
2135 /*
2136 * We have to hold onto the inode so that ordered extents can do their
2137 * work when the IO finishes. The alternative to this is failing to add
2138 * an ordered extent if the igrab() fails there and that is a huge pain
2139 * to deal with, so instead just hold onto the inode throughout the
2140 * writepages operation. If it fails here we are freeing up the inode
2141 * anyway and we'd rather not waste our time writing out stuff that is
2142 * going to be truncated anyway.
2143 */
2144 if (!igrab(inode))
2145 return 0;
2146
2147 folio_batch_init(&fbatch);
2148 if (wbc->range_cyclic) {
2149 index = mapping->writeback_index; /* Start from prev offset */
2150 end = -1;
2151 /*
2152 * Start from the beginning does not need to cycle over the
2153 * range, mark it as scanned.
2154 */
2155 scanned = (index == 0);
2156 } else {
2157 index = wbc->range_start >> PAGE_SHIFT;
2158 end = wbc->range_end >> PAGE_SHIFT;
2159 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2160 range_whole = 1;
2161 scanned = 1;
2162 }
2163
2164 /*
2165 * We do the tagged writepage as long as the snapshot flush bit is set
2166 * and we are the first one who do the filemap_flush() on this inode.
2167 *
2168 * The nr_to_write == LONG_MAX is needed to make sure other flushers do
2169 * not race in and drop the bit.
2170 */
2171 if (range_whole && wbc->nr_to_write == LONG_MAX &&
2172 test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
2173 &BTRFS_I(inode)->runtime_flags))
2174 wbc->tagged_writepages = 1;
2175
2176 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2177 tag = PAGECACHE_TAG_TOWRITE;
2178 else
2179 tag = PAGECACHE_TAG_DIRTY;
2180retry:
2181 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2182 tag_pages_for_writeback(mapping, index, end);
2183 done_index = index;
2184 while (!done && !nr_to_write_done && (index <= end) &&
2185 (nr_folios = filemap_get_folios_tag(mapping, &index,
2186 end, tag, &fbatch))) {
2187 unsigned i;
2188
2189 for (i = 0; i < nr_folios; i++) {
2190 struct folio *folio = fbatch.folios[i];
2191
2192 done_index = folio_next_index(folio);
2193 /*
2194 * At this point we hold neither the i_pages lock nor
2195 * the page lock: the page may be truncated or
2196 * invalidated (changing page->mapping to NULL),
2197 * or even swizzled back from swapper_space to
2198 * tmpfs file mapping
2199 */
2200 if (!folio_trylock(folio)) {
2201 submit_write_bio(bio_ctrl, 0);
2202 folio_lock(folio);
2203 }
2204
2205 if (unlikely(folio->mapping != mapping)) {
2206 folio_unlock(folio);
2207 continue;
2208 }
2209
2210 if (!folio_test_dirty(folio)) {
2211 /* Someone wrote it for us. */
2212 folio_unlock(folio);
2213 continue;
2214 }
2215
2216 /*
2217 * For subpage case, compression can lead to mixed
2218 * writeback and dirty flags, e.g:
2219 * 0 32K 64K 96K 128K
2220 * | |//////||/////| |//|
2221 *
2222 * In above case, [32K, 96K) is asynchronously submitted
2223 * for compression, and [124K, 128K) needs to be written back.
2224 *
2225 * If we didn't wait wrtiteback for page 64K, [128K, 128K)
2226 * won't be submitted as the page still has writeback flag
2227 * and will be skipped in the next check.
2228 *
2229 * This mixed writeback and dirty case is only possible for
2230 * subpage case.
2231 *
2232 * TODO: Remove this check after migrating compression to
2233 * regular submission.
2234 */
2235 if (wbc->sync_mode != WB_SYNC_NONE ||
2236 btrfs_is_subpage(inode_to_fs_info(inode), mapping)) {
2237 if (folio_test_writeback(folio))
2238 submit_write_bio(bio_ctrl, 0);
2239 folio_wait_writeback(folio);
2240 }
2241
2242 if (folio_test_writeback(folio) ||
2243 !folio_clear_dirty_for_io(folio)) {
2244 folio_unlock(folio);
2245 continue;
2246 }
2247
2248 ret = extent_writepage(folio, bio_ctrl);
2249 if (ret < 0) {
2250 done = 1;
2251 break;
2252 }
2253
2254 /*
2255 * The filesystem may choose to bump up nr_to_write.
2256 * We have to make sure to honor the new nr_to_write
2257 * at any time.
2258 */
2259 nr_to_write_done = (wbc->sync_mode == WB_SYNC_NONE &&
2260 wbc->nr_to_write <= 0);
2261 }
2262 folio_batch_release(&fbatch);
2263 cond_resched();
2264 }
2265 if (!scanned && !done) {
2266 /*
2267 * We hit the last page and there is more work to be done: wrap
2268 * back to the start of the file
2269 */
2270 scanned = 1;
2271 index = 0;
2272
2273 /*
2274 * If we're looping we could run into a page that is locked by a
2275 * writer and that writer could be waiting on writeback for a
2276 * page in our current bio, and thus deadlock, so flush the
2277 * write bio here.
2278 */
2279 submit_write_bio(bio_ctrl, 0);
2280 goto retry;
2281 }
2282
2283 if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
2284 mapping->writeback_index = done_index;
2285
2286 btrfs_add_delayed_iput(BTRFS_I(inode));
2287 return ret;
2288}
2289
2290/*
2291 * Submit the pages in the range to bio for call sites which delalloc range has
2292 * already been ran (aka, ordered extent inserted) and all pages are still
2293 * locked.
2294 */
2295void extent_write_locked_range(struct inode *inode, const struct folio *locked_folio,
2296 u64 start, u64 end, struct writeback_control *wbc,
2297 bool pages_dirty)
2298{
2299 bool found_error = false;
2300 int ret = 0;
2301 struct address_space *mapping = inode->i_mapping;
2302 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
2303 const u32 sectorsize = fs_info->sectorsize;
2304 loff_t i_size = i_size_read(inode);
2305 u64 cur = start;
2306 struct btrfs_bio_ctrl bio_ctrl = {
2307 .wbc = wbc,
2308 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2309 };
2310
2311 if (wbc->no_cgroup_owner)
2312 bio_ctrl.opf |= REQ_BTRFS_CGROUP_PUNT;
2313
2314 ASSERT(IS_ALIGNED(start, sectorsize) && IS_ALIGNED(end + 1, sectorsize));
2315
2316 while (cur <= end) {
2317 u64 cur_end = min(round_down(cur, PAGE_SIZE) + PAGE_SIZE - 1, end);
2318 u32 cur_len = cur_end + 1 - cur;
2319 struct folio *folio;
2320
2321 folio = filemap_get_folio(mapping, cur >> PAGE_SHIFT);
2322
2323 /*
2324 * This shouldn't happen, the pages are pinned and locked, this
2325 * code is just in case, but shouldn't actually be run.
2326 */
2327 if (IS_ERR(folio)) {
2328 btrfs_mark_ordered_io_finished(BTRFS_I(inode), NULL,
2329 cur, cur_len, false);
2330 mapping_set_error(mapping, PTR_ERR(folio));
2331 cur = cur_end + 1;
2332 continue;
2333 }
2334
2335 ASSERT(folio_test_locked(folio));
2336 if (pages_dirty && folio != locked_folio)
2337 ASSERT(folio_test_dirty(folio));
2338
2339 /*
2340 * Set the submission bitmap to submit all sectors.
2341 * extent_writepage_io() will do the truncation correctly.
2342 */
2343 bio_ctrl.submit_bitmap = (unsigned long)-1;
2344 ret = extent_writepage_io(BTRFS_I(inode), folio, cur, cur_len,
2345 &bio_ctrl, i_size);
2346 if (ret == 1)
2347 goto next_page;
2348
2349 if (ret)
2350 mapping_set_error(mapping, ret);
2351 btrfs_folio_end_lock(fs_info, folio, cur, cur_len);
2352 if (ret < 0)
2353 found_error = true;
2354next_page:
2355 folio_put(folio);
2356 cur = cur_end + 1;
2357 }
2358
2359 submit_write_bio(&bio_ctrl, found_error ? ret : 0);
2360}
2361
2362int btrfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
2363{
2364 struct inode *inode = mapping->host;
2365 int ret = 0;
2366 struct btrfs_bio_ctrl bio_ctrl = {
2367 .wbc = wbc,
2368 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2369 };
2370
2371 /*
2372 * Allow only a single thread to do the reloc work in zoned mode to
2373 * protect the write pointer updates.
2374 */
2375 btrfs_zoned_data_reloc_lock(BTRFS_I(inode));
2376 ret = extent_write_cache_pages(mapping, &bio_ctrl);
2377 submit_write_bio(&bio_ctrl, ret);
2378 btrfs_zoned_data_reloc_unlock(BTRFS_I(inode));
2379 return ret;
2380}
2381
2382void btrfs_readahead(struct readahead_control *rac)
2383{
2384 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ | REQ_RAHEAD };
2385 struct folio *folio;
2386 struct btrfs_inode *inode = BTRFS_I(rac->mapping->host);
2387 const u64 start = readahead_pos(rac);
2388 const u64 end = start + readahead_length(rac) - 1;
2389 struct extent_state *cached_state = NULL;
2390 struct extent_map *em_cached = NULL;
2391 u64 prev_em_start = (u64)-1;
2392
2393 btrfs_lock_and_flush_ordered_range(inode, start, end, &cached_state);
2394
2395 while ((folio = readahead_folio(rac)) != NULL)
2396 btrfs_do_readpage(folio, &em_cached, &bio_ctrl, &prev_em_start);
2397
2398 unlock_extent(&inode->io_tree, start, end, &cached_state);
2399
2400 if (em_cached)
2401 free_extent_map(em_cached);
2402 submit_one_bio(&bio_ctrl);
2403}
2404
2405/*
2406 * basic invalidate_folio code, this waits on any locked or writeback
2407 * ranges corresponding to the folio, and then deletes any extent state
2408 * records from the tree
2409 */
2410int extent_invalidate_folio(struct extent_io_tree *tree,
2411 struct folio *folio, size_t offset)
2412{
2413 struct extent_state *cached_state = NULL;
2414 u64 start = folio_pos(folio);
2415 u64 end = start + folio_size(folio) - 1;
2416 size_t blocksize = folio_to_fs_info(folio)->sectorsize;
2417
2418 /* This function is only called for the btree inode */
2419 ASSERT(tree->owner == IO_TREE_BTREE_INODE_IO);
2420
2421 start += ALIGN(offset, blocksize);
2422 if (start > end)
2423 return 0;
2424
2425 lock_extent(tree, start, end, &cached_state);
2426 folio_wait_writeback(folio);
2427
2428 /*
2429 * Currently for btree io tree, only EXTENT_LOCKED is utilized,
2430 * so here we only need to unlock the extent range to free any
2431 * existing extent state.
2432 */
2433 unlock_extent(tree, start, end, &cached_state);
2434 return 0;
2435}
2436
2437/*
2438 * a helper for release_folio, this tests for areas of the page that
2439 * are locked or under IO and drops the related state bits if it is safe
2440 * to drop the page.
2441 */
2442static bool try_release_extent_state(struct extent_io_tree *tree,
2443 struct folio *folio)
2444{
2445 u64 start = folio_pos(folio);
2446 u64 end = start + PAGE_SIZE - 1;
2447 bool ret;
2448
2449 if (test_range_bit_exists(tree, start, end, EXTENT_LOCKED)) {
2450 ret = false;
2451 } else {
2452 u32 clear_bits = ~(EXTENT_LOCKED | EXTENT_NODATASUM |
2453 EXTENT_DELALLOC_NEW | EXTENT_CTLBITS |
2454 EXTENT_QGROUP_RESERVED);
2455 int ret2;
2456
2457 /*
2458 * At this point we can safely clear everything except the
2459 * locked bit, the nodatasum bit and the delalloc new bit.
2460 * The delalloc new bit will be cleared by ordered extent
2461 * completion.
2462 */
2463 ret2 = __clear_extent_bit(tree, start, end, clear_bits, NULL, NULL);
2464
2465 /* if clear_extent_bit failed for enomem reasons,
2466 * we can't allow the release to continue.
2467 */
2468 if (ret2 < 0)
2469 ret = false;
2470 else
2471 ret = true;
2472 }
2473 return ret;
2474}
2475
2476/*
2477 * a helper for release_folio. As long as there are no locked extents
2478 * in the range corresponding to the page, both state records and extent
2479 * map records are removed
2480 */
2481bool try_release_extent_mapping(struct folio *folio, gfp_t mask)
2482{
2483 u64 start = folio_pos(folio);
2484 u64 end = start + PAGE_SIZE - 1;
2485 struct btrfs_inode *inode = folio_to_inode(folio);
2486 struct extent_io_tree *io_tree = &inode->io_tree;
2487
2488 while (start <= end) {
2489 const u64 cur_gen = btrfs_get_fs_generation(inode->root->fs_info);
2490 const u64 len = end - start + 1;
2491 struct extent_map_tree *extent_tree = &inode->extent_tree;
2492 struct extent_map *em;
2493
2494 write_lock(&extent_tree->lock);
2495 em = lookup_extent_mapping(extent_tree, start, len);
2496 if (!em) {
2497 write_unlock(&extent_tree->lock);
2498 break;
2499 }
2500 if ((em->flags & EXTENT_FLAG_PINNED) || em->start != start) {
2501 write_unlock(&extent_tree->lock);
2502 free_extent_map(em);
2503 break;
2504 }
2505 if (test_range_bit_exists(io_tree, em->start,
2506 extent_map_end(em) - 1, EXTENT_LOCKED))
2507 goto next;
2508 /*
2509 * If it's not in the list of modified extents, used by a fast
2510 * fsync, we can remove it. If it's being logged we can safely
2511 * remove it since fsync took an extra reference on the em.
2512 */
2513 if (list_empty(&em->list) || (em->flags & EXTENT_FLAG_LOGGING))
2514 goto remove_em;
2515 /*
2516 * If it's in the list of modified extents, remove it only if
2517 * its generation is older then the current one, in which case
2518 * we don't need it for a fast fsync. Otherwise don't remove it,
2519 * we could be racing with an ongoing fast fsync that could miss
2520 * the new extent.
2521 */
2522 if (em->generation >= cur_gen)
2523 goto next;
2524remove_em:
2525 /*
2526 * We only remove extent maps that are not in the list of
2527 * modified extents or that are in the list but with a
2528 * generation lower then the current generation, so there is no
2529 * need to set the full fsync flag on the inode (it hurts the
2530 * fsync performance for workloads with a data size that exceeds
2531 * or is close to the system's memory).
2532 */
2533 remove_extent_mapping(inode, em);
2534 /* Once for the inode's extent map tree. */
2535 free_extent_map(em);
2536next:
2537 start = extent_map_end(em);
2538 write_unlock(&extent_tree->lock);
2539
2540 /* Once for us, for the lookup_extent_mapping() reference. */
2541 free_extent_map(em);
2542
2543 if (need_resched()) {
2544 /*
2545 * If we need to resched but we can't block just exit
2546 * and leave any remaining extent maps.
2547 */
2548 if (!gfpflags_allow_blocking(mask))
2549 break;
2550
2551 cond_resched();
2552 }
2553 }
2554 return try_release_extent_state(io_tree, folio);
2555}
2556
2557static void __free_extent_buffer(struct extent_buffer *eb)
2558{
2559 kmem_cache_free(extent_buffer_cache, eb);
2560}
2561
2562static int extent_buffer_under_io(const struct extent_buffer *eb)
2563{
2564 return (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
2565 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
2566}
2567
2568static bool folio_range_has_eb(struct folio *folio)
2569{
2570 struct btrfs_subpage *subpage;
2571
2572 lockdep_assert_held(&folio->mapping->i_private_lock);
2573
2574 if (folio_test_private(folio)) {
2575 subpage = folio_get_private(folio);
2576 if (atomic_read(&subpage->eb_refs))
2577 return true;
2578 }
2579 return false;
2580}
2581
2582static void detach_extent_buffer_folio(const struct extent_buffer *eb, struct folio *folio)
2583{
2584 struct btrfs_fs_info *fs_info = eb->fs_info;
2585 const bool mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
2586
2587 /*
2588 * For mapped eb, we're going to change the folio private, which should
2589 * be done under the i_private_lock.
2590 */
2591 if (mapped)
2592 spin_lock(&folio->mapping->i_private_lock);
2593
2594 if (!folio_test_private(folio)) {
2595 if (mapped)
2596 spin_unlock(&folio->mapping->i_private_lock);
2597 return;
2598 }
2599
2600 if (fs_info->nodesize >= PAGE_SIZE) {
2601 /*
2602 * We do this since we'll remove the pages after we've
2603 * removed the eb from the radix tree, so we could race
2604 * and have this page now attached to the new eb. So
2605 * only clear folio if it's still connected to
2606 * this eb.
2607 */
2608 if (folio_test_private(folio) && folio_get_private(folio) == eb) {
2609 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
2610 BUG_ON(folio_test_dirty(folio));
2611 BUG_ON(folio_test_writeback(folio));
2612 /* We need to make sure we haven't be attached to a new eb. */
2613 folio_detach_private(folio);
2614 }
2615 if (mapped)
2616 spin_unlock(&folio->mapping->i_private_lock);
2617 return;
2618 }
2619
2620 /*
2621 * For subpage, we can have dummy eb with folio private attached. In
2622 * this case, we can directly detach the private as such folio is only
2623 * attached to one dummy eb, no sharing.
2624 */
2625 if (!mapped) {
2626 btrfs_detach_subpage(fs_info, folio);
2627 return;
2628 }
2629
2630 btrfs_folio_dec_eb_refs(fs_info, folio);
2631
2632 /*
2633 * We can only detach the folio private if there are no other ebs in the
2634 * page range and no unfinished IO.
2635 */
2636 if (!folio_range_has_eb(folio))
2637 btrfs_detach_subpage(fs_info, folio);
2638
2639 spin_unlock(&folio->mapping->i_private_lock);
2640}
2641
2642/* Release all pages attached to the extent buffer */
2643static void btrfs_release_extent_buffer_pages(const struct extent_buffer *eb)
2644{
2645 ASSERT(!extent_buffer_under_io(eb));
2646
2647 for (int i = 0; i < INLINE_EXTENT_BUFFER_PAGES; i++) {
2648 struct folio *folio = eb->folios[i];
2649
2650 if (!folio)
2651 continue;
2652
2653 detach_extent_buffer_folio(eb, folio);
2654
2655 /* One for when we allocated the folio. */
2656 folio_put(folio);
2657 }
2658}
2659
2660/*
2661 * Helper for releasing the extent buffer.
2662 */
2663static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
2664{
2665 btrfs_release_extent_buffer_pages(eb);
2666 btrfs_leak_debug_del_eb(eb);
2667 __free_extent_buffer(eb);
2668}
2669
2670static struct extent_buffer *
2671__alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
2672 unsigned long len)
2673{
2674 struct extent_buffer *eb = NULL;
2675
2676 eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
2677 eb->start = start;
2678 eb->len = len;
2679 eb->fs_info = fs_info;
2680 init_rwsem(&eb->lock);
2681
2682 btrfs_leak_debug_add_eb(eb);
2683
2684 spin_lock_init(&eb->refs_lock);
2685 atomic_set(&eb->refs, 1);
2686
2687 ASSERT(len <= BTRFS_MAX_METADATA_BLOCKSIZE);
2688
2689 return eb;
2690}
2691
2692struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
2693{
2694 struct extent_buffer *new;
2695 int num_folios = num_extent_folios(src);
2696 int ret;
2697
2698 new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
2699 if (new == NULL)
2700 return NULL;
2701
2702 /*
2703 * Set UNMAPPED before calling btrfs_release_extent_buffer(), as
2704 * btrfs_release_extent_buffer() have different behavior for
2705 * UNMAPPED subpage extent buffer.
2706 */
2707 set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
2708
2709 ret = alloc_eb_folio_array(new, false);
2710 if (ret) {
2711 btrfs_release_extent_buffer(new);
2712 return NULL;
2713 }
2714
2715 for (int i = 0; i < num_folios; i++) {
2716 struct folio *folio = new->folios[i];
2717
2718 ret = attach_extent_buffer_folio(new, folio, NULL);
2719 if (ret < 0) {
2720 btrfs_release_extent_buffer(new);
2721 return NULL;
2722 }
2723 WARN_ON(folio_test_dirty(folio));
2724 }
2725 copy_extent_buffer_full(new, src);
2726 set_extent_buffer_uptodate(new);
2727
2728 return new;
2729}
2730
2731struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
2732 u64 start, unsigned long len)
2733{
2734 struct extent_buffer *eb;
2735 int num_folios = 0;
2736 int ret;
2737
2738 eb = __alloc_extent_buffer(fs_info, start, len);
2739 if (!eb)
2740 return NULL;
2741
2742 ret = alloc_eb_folio_array(eb, false);
2743 if (ret)
2744 goto err;
2745
2746 num_folios = num_extent_folios(eb);
2747 for (int i = 0; i < num_folios; i++) {
2748 ret = attach_extent_buffer_folio(eb, eb->folios[i], NULL);
2749 if (ret < 0)
2750 goto err;
2751 }
2752
2753 set_extent_buffer_uptodate(eb);
2754 btrfs_set_header_nritems(eb, 0);
2755 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
2756
2757 return eb;
2758err:
2759 for (int i = 0; i < num_folios; i++) {
2760 if (eb->folios[i]) {
2761 detach_extent_buffer_folio(eb, eb->folios[i]);
2762 folio_put(eb->folios[i]);
2763 }
2764 }
2765 __free_extent_buffer(eb);
2766 return NULL;
2767}
2768
2769struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
2770 u64 start)
2771{
2772 return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
2773}
2774
2775static void check_buffer_tree_ref(struct extent_buffer *eb)
2776{
2777 int refs;
2778 /*
2779 * The TREE_REF bit is first set when the extent_buffer is added
2780 * to the radix tree. It is also reset, if unset, when a new reference
2781 * is created by find_extent_buffer.
2782 *
2783 * It is only cleared in two cases: freeing the last non-tree
2784 * reference to the extent_buffer when its STALE bit is set or
2785 * calling release_folio when the tree reference is the only reference.
2786 *
2787 * In both cases, care is taken to ensure that the extent_buffer's
2788 * pages are not under io. However, release_folio can be concurrently
2789 * called with creating new references, which is prone to race
2790 * conditions between the calls to check_buffer_tree_ref in those
2791 * codepaths and clearing TREE_REF in try_release_extent_buffer.
2792 *
2793 * The actual lifetime of the extent_buffer in the radix tree is
2794 * adequately protected by the refcount, but the TREE_REF bit and
2795 * its corresponding reference are not. To protect against this
2796 * class of races, we call check_buffer_tree_ref from the codepaths
2797 * which trigger io. Note that once io is initiated, TREE_REF can no
2798 * longer be cleared, so that is the moment at which any such race is
2799 * best fixed.
2800 */
2801 refs = atomic_read(&eb->refs);
2802 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
2803 return;
2804
2805 spin_lock(&eb->refs_lock);
2806 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
2807 atomic_inc(&eb->refs);
2808 spin_unlock(&eb->refs_lock);
2809}
2810
2811static void mark_extent_buffer_accessed(struct extent_buffer *eb)
2812{
2813 int num_folios= num_extent_folios(eb);
2814
2815 check_buffer_tree_ref(eb);
2816
2817 for (int i = 0; i < num_folios; i++)
2818 folio_mark_accessed(eb->folios[i]);
2819}
2820
2821struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
2822 u64 start)
2823{
2824 struct extent_buffer *eb;
2825
2826 eb = find_extent_buffer_nolock(fs_info, start);
2827 if (!eb)
2828 return NULL;
2829 /*
2830 * Lock our eb's refs_lock to avoid races with free_extent_buffer().
2831 * When we get our eb it might be flagged with EXTENT_BUFFER_STALE and
2832 * another task running free_extent_buffer() might have seen that flag
2833 * set, eb->refs == 2, that the buffer isn't under IO (dirty and
2834 * writeback flags not set) and it's still in the tree (flag
2835 * EXTENT_BUFFER_TREE_REF set), therefore being in the process of
2836 * decrementing the extent buffer's reference count twice. So here we
2837 * could race and increment the eb's reference count, clear its stale
2838 * flag, mark it as dirty and drop our reference before the other task
2839 * finishes executing free_extent_buffer, which would later result in
2840 * an attempt to free an extent buffer that is dirty.
2841 */
2842 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
2843 spin_lock(&eb->refs_lock);
2844 spin_unlock(&eb->refs_lock);
2845 }
2846 mark_extent_buffer_accessed(eb);
2847 return eb;
2848}
2849
2850#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
2851struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
2852 u64 start)
2853{
2854 struct extent_buffer *eb, *exists = NULL;
2855 int ret;
2856
2857 eb = find_extent_buffer(fs_info, start);
2858 if (eb)
2859 return eb;
2860 eb = alloc_dummy_extent_buffer(fs_info, start);
2861 if (!eb)
2862 return ERR_PTR(-ENOMEM);
2863 eb->fs_info = fs_info;
2864again:
2865 ret = radix_tree_preload(GFP_NOFS);
2866 if (ret) {
2867 exists = ERR_PTR(ret);
2868 goto free_eb;
2869 }
2870 spin_lock(&fs_info->buffer_lock);
2871 ret = radix_tree_insert(&fs_info->buffer_radix,
2872 start >> fs_info->sectorsize_bits, eb);
2873 spin_unlock(&fs_info->buffer_lock);
2874 radix_tree_preload_end();
2875 if (ret == -EEXIST) {
2876 exists = find_extent_buffer(fs_info, start);
2877 if (exists)
2878 goto free_eb;
2879 else
2880 goto again;
2881 }
2882 check_buffer_tree_ref(eb);
2883 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
2884
2885 return eb;
2886free_eb:
2887 btrfs_release_extent_buffer(eb);
2888 return exists;
2889}
2890#endif
2891
2892static struct extent_buffer *grab_extent_buffer(
2893 struct btrfs_fs_info *fs_info, struct page *page)
2894{
2895 struct folio *folio = page_folio(page);
2896 struct extent_buffer *exists;
2897
2898 lockdep_assert_held(&page->mapping->i_private_lock);
2899
2900 /*
2901 * For subpage case, we completely rely on radix tree to ensure we
2902 * don't try to insert two ebs for the same bytenr. So here we always
2903 * return NULL and just continue.
2904 */
2905 if (fs_info->nodesize < PAGE_SIZE)
2906 return NULL;
2907
2908 /* Page not yet attached to an extent buffer */
2909 if (!folio_test_private(folio))
2910 return NULL;
2911
2912 /*
2913 * We could have already allocated an eb for this page and attached one
2914 * so lets see if we can get a ref on the existing eb, and if we can we
2915 * know it's good and we can just return that one, else we know we can
2916 * just overwrite folio private.
2917 */
2918 exists = folio_get_private(folio);
2919 if (atomic_inc_not_zero(&exists->refs))
2920 return exists;
2921
2922 WARN_ON(PageDirty(page));
2923 folio_detach_private(folio);
2924 return NULL;
2925}
2926
2927static int check_eb_alignment(struct btrfs_fs_info *fs_info, u64 start)
2928{
2929 if (!IS_ALIGNED(start, fs_info->sectorsize)) {
2930 btrfs_err(fs_info, "bad tree block start %llu", start);
2931 return -EINVAL;
2932 }
2933
2934 if (fs_info->nodesize < PAGE_SIZE &&
2935 offset_in_page(start) + fs_info->nodesize > PAGE_SIZE) {
2936 btrfs_err(fs_info,
2937 "tree block crosses page boundary, start %llu nodesize %u",
2938 start, fs_info->nodesize);
2939 return -EINVAL;
2940 }
2941 if (fs_info->nodesize >= PAGE_SIZE &&
2942 !PAGE_ALIGNED(start)) {
2943 btrfs_err(fs_info,
2944 "tree block is not page aligned, start %llu nodesize %u",
2945 start, fs_info->nodesize);
2946 return -EINVAL;
2947 }
2948 if (!IS_ALIGNED(start, fs_info->nodesize) &&
2949 !test_and_set_bit(BTRFS_FS_UNALIGNED_TREE_BLOCK, &fs_info->flags)) {
2950 btrfs_warn(fs_info,
2951"tree block not nodesize aligned, start %llu nodesize %u, can be resolved by a full metadata balance",
2952 start, fs_info->nodesize);
2953 }
2954 return 0;
2955}
2956
2957
2958/*
2959 * Return 0 if eb->folios[i] is attached to btree inode successfully.
2960 * Return >0 if there is already another extent buffer for the range,
2961 * and @found_eb_ret would be updated.
2962 * Return -EAGAIN if the filemap has an existing folio but with different size
2963 * than @eb.
2964 * The caller needs to free the existing folios and retry using the same order.
2965 */
2966static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
2967 struct btrfs_subpage *prealloc,
2968 struct extent_buffer **found_eb_ret)
2969{
2970
2971 struct btrfs_fs_info *fs_info = eb->fs_info;
2972 struct address_space *mapping = fs_info->btree_inode->i_mapping;
2973 const unsigned long index = eb->start >> PAGE_SHIFT;
2974 struct folio *existing_folio = NULL;
2975 int ret;
2976
2977 ASSERT(found_eb_ret);
2978
2979 /* Caller should ensure the folio exists. */
2980 ASSERT(eb->folios[i]);
2981
2982retry:
2983 ret = filemap_add_folio(mapping, eb->folios[i], index + i,
2984 GFP_NOFS | __GFP_NOFAIL);
2985 if (!ret)
2986 goto finish;
2987
2988 existing_folio = filemap_lock_folio(mapping, index + i);
2989 /* The page cache only exists for a very short time, just retry. */
2990 if (IS_ERR(existing_folio)) {
2991 existing_folio = NULL;
2992 goto retry;
2993 }
2994
2995 /* For now, we should only have single-page folios for btree inode. */
2996 ASSERT(folio_nr_pages(existing_folio) == 1);
2997
2998 if (folio_size(existing_folio) != eb->folio_size) {
2999 folio_unlock(existing_folio);
3000 folio_put(existing_folio);
3001 return -EAGAIN;
3002 }
3003
3004finish:
3005 spin_lock(&mapping->i_private_lock);
3006 if (existing_folio && fs_info->nodesize < PAGE_SIZE) {
3007 /* We're going to reuse the existing page, can drop our folio now. */
3008 __free_page(folio_page(eb->folios[i], 0));
3009 eb->folios[i] = existing_folio;
3010 } else if (existing_folio) {
3011 struct extent_buffer *existing_eb;
3012
3013 existing_eb = grab_extent_buffer(fs_info,
3014 folio_page(existing_folio, 0));
3015 if (existing_eb) {
3016 /* The extent buffer still exists, we can use it directly. */
3017 *found_eb_ret = existing_eb;
3018 spin_unlock(&mapping->i_private_lock);
3019 folio_unlock(existing_folio);
3020 folio_put(existing_folio);
3021 return 1;
3022 }
3023 /* The extent buffer no longer exists, we can reuse the folio. */
3024 __free_page(folio_page(eb->folios[i], 0));
3025 eb->folios[i] = existing_folio;
3026 }
3027 eb->folio_size = folio_size(eb->folios[i]);
3028 eb->folio_shift = folio_shift(eb->folios[i]);
3029 /* Should not fail, as we have preallocated the memory. */
3030 ret = attach_extent_buffer_folio(eb, eb->folios[i], prealloc);
3031 ASSERT(!ret);
3032 /*
3033 * To inform we have an extra eb under allocation, so that
3034 * detach_extent_buffer_page() won't release the folio private when the
3035 * eb hasn't been inserted into radix tree yet.
3036 *
3037 * The ref will be decreased when the eb releases the page, in
3038 * detach_extent_buffer_page(). Thus needs no special handling in the
3039 * error path.
3040 */
3041 btrfs_folio_inc_eb_refs(fs_info, eb->folios[i]);
3042 spin_unlock(&mapping->i_private_lock);
3043 return 0;
3044}
3045
3046struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
3047 u64 start, u64 owner_root, int level)
3048{
3049 unsigned long len = fs_info->nodesize;
3050 int num_folios;
3051 int attached = 0;
3052 struct extent_buffer *eb;
3053 struct extent_buffer *existing_eb = NULL;
3054 struct btrfs_subpage *prealloc = NULL;
3055 u64 lockdep_owner = owner_root;
3056 bool page_contig = true;
3057 int uptodate = 1;
3058 int ret;
3059
3060 if (check_eb_alignment(fs_info, start))
3061 return ERR_PTR(-EINVAL);
3062
3063#if BITS_PER_LONG == 32
3064 if (start >= MAX_LFS_FILESIZE) {
3065 btrfs_err_rl(fs_info,
3066 "extent buffer %llu is beyond 32bit page cache limit", start);
3067 btrfs_err_32bit_limit(fs_info);
3068 return ERR_PTR(-EOVERFLOW);
3069 }
3070 if (start >= BTRFS_32BIT_EARLY_WARN_THRESHOLD)
3071 btrfs_warn_32bit_limit(fs_info);
3072#endif
3073
3074 eb = find_extent_buffer(fs_info, start);
3075 if (eb)
3076 return eb;
3077
3078 eb = __alloc_extent_buffer(fs_info, start, len);
3079 if (!eb)
3080 return ERR_PTR(-ENOMEM);
3081
3082 /*
3083 * The reloc trees are just snapshots, so we need them to appear to be
3084 * just like any other fs tree WRT lockdep.
3085 */
3086 if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID)
3087 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
3088
3089 btrfs_set_buffer_lockdep_class(lockdep_owner, eb, level);
3090
3091 /*
3092 * Preallocate folio private for subpage case, so that we won't
3093 * allocate memory with i_private_lock nor page lock hold.
3094 *
3095 * The memory will be freed by attach_extent_buffer_page() or freed
3096 * manually if we exit earlier.
3097 */
3098 if (fs_info->nodesize < PAGE_SIZE) {
3099 prealloc = btrfs_alloc_subpage(fs_info, BTRFS_SUBPAGE_METADATA);
3100 if (IS_ERR(prealloc)) {
3101 ret = PTR_ERR(prealloc);
3102 goto out;
3103 }
3104 }
3105
3106reallocate:
3107 /* Allocate all pages first. */
3108 ret = alloc_eb_folio_array(eb, true);
3109 if (ret < 0) {
3110 btrfs_free_subpage(prealloc);
3111 goto out;
3112 }
3113
3114 num_folios = num_extent_folios(eb);
3115 /* Attach all pages to the filemap. */
3116 for (int i = 0; i < num_folios; i++) {
3117 struct folio *folio;
3118
3119 ret = attach_eb_folio_to_filemap(eb, i, prealloc, &existing_eb);
3120 if (ret > 0) {
3121 ASSERT(existing_eb);
3122 goto out;
3123 }
3124
3125 /*
3126 * TODO: Special handling for a corner case where the order of
3127 * folios mismatch between the new eb and filemap.
3128 *
3129 * This happens when:
3130 *
3131 * - the new eb is using higher order folio
3132 *
3133 * - the filemap is still using 0-order folios for the range
3134 * This can happen at the previous eb allocation, and we don't
3135 * have higher order folio for the call.
3136 *
3137 * - the existing eb has already been freed
3138 *
3139 * In this case, we have to free the existing folios first, and
3140 * re-allocate using the same order.
3141 * Thankfully this is not going to happen yet, as we're still
3142 * using 0-order folios.
3143 */
3144 if (unlikely(ret == -EAGAIN)) {
3145 ASSERT(0);
3146 goto reallocate;
3147 }
3148 attached++;
3149
3150 /*
3151 * Only after attach_eb_folio_to_filemap(), eb->folios[] is
3152 * reliable, as we may choose to reuse the existing page cache
3153 * and free the allocated page.
3154 */
3155 folio = eb->folios[i];
3156 WARN_ON(btrfs_folio_test_dirty(fs_info, folio, eb->start, eb->len));
3157
3158 /*
3159 * Check if the current page is physically contiguous with previous eb
3160 * page.
3161 * At this stage, either we allocated a large folio, thus @i
3162 * would only be 0, or we fall back to per-page allocation.
3163 */
3164 if (i && folio_page(eb->folios[i - 1], 0) + 1 != folio_page(folio, 0))
3165 page_contig = false;
3166
3167 if (!btrfs_folio_test_uptodate(fs_info, folio, eb->start, eb->len))
3168 uptodate = 0;
3169
3170 /*
3171 * We can't unlock the pages just yet since the extent buffer
3172 * hasn't been properly inserted in the radix tree, this
3173 * opens a race with btree_release_folio which can free a page
3174 * while we are still filling in all pages for the buffer and
3175 * we could crash.
3176 */
3177 }
3178 if (uptodate)
3179 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3180 /* All pages are physically contiguous, can skip cross page handling. */
3181 if (page_contig)
3182 eb->addr = folio_address(eb->folios[0]) + offset_in_page(eb->start);
3183again:
3184 ret = radix_tree_preload(GFP_NOFS);
3185 if (ret)
3186 goto out;
3187
3188 spin_lock(&fs_info->buffer_lock);
3189 ret = radix_tree_insert(&fs_info->buffer_radix,
3190 start >> fs_info->sectorsize_bits, eb);
3191 spin_unlock(&fs_info->buffer_lock);
3192 radix_tree_preload_end();
3193 if (ret == -EEXIST) {
3194 ret = 0;
3195 existing_eb = find_extent_buffer(fs_info, start);
3196 if (existing_eb)
3197 goto out;
3198 else
3199 goto again;
3200 }
3201 /* add one reference for the tree */
3202 check_buffer_tree_ref(eb);
3203 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
3204
3205 /*
3206 * Now it's safe to unlock the pages because any calls to
3207 * btree_release_folio will correctly detect that a page belongs to a
3208 * live buffer and won't free them prematurely.
3209 */
3210 for (int i = 0; i < num_folios; i++)
3211 unlock_page(folio_page(eb->folios[i], 0));
3212 return eb;
3213
3214out:
3215 WARN_ON(!atomic_dec_and_test(&eb->refs));
3216
3217 /*
3218 * Any attached folios need to be detached before we unlock them. This
3219 * is because when we're inserting our new folios into the mapping, and
3220 * then attaching our eb to that folio. If we fail to insert our folio
3221 * we'll lookup the folio for that index, and grab that EB. We do not
3222 * want that to grab this eb, as we're getting ready to free it. So we
3223 * have to detach it first and then unlock it.
3224 *
3225 * We have to drop our reference and NULL it out here because in the
3226 * subpage case detaching does a btrfs_folio_dec_eb_refs() for our eb.
3227 * Below when we call btrfs_release_extent_buffer() we will call
3228 * detach_extent_buffer_folio() on our remaining pages in the !subpage
3229 * case. If we left eb->folios[i] populated in the subpage case we'd
3230 * double put our reference and be super sad.
3231 */
3232 for (int i = 0; i < attached; i++) {
3233 ASSERT(eb->folios[i]);
3234 detach_extent_buffer_folio(eb, eb->folios[i]);
3235 unlock_page(folio_page(eb->folios[i], 0));
3236 folio_put(eb->folios[i]);
3237 eb->folios[i] = NULL;
3238 }
3239 /*
3240 * Now all pages of that extent buffer is unmapped, set UNMAPPED flag,
3241 * so it can be cleaned up without utilizing page->mapping.
3242 */
3243 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3244
3245 btrfs_release_extent_buffer(eb);
3246 if (ret < 0)
3247 return ERR_PTR(ret);
3248 ASSERT(existing_eb);
3249 return existing_eb;
3250}
3251
3252static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3253{
3254 struct extent_buffer *eb =
3255 container_of(head, struct extent_buffer, rcu_head);
3256
3257 __free_extent_buffer(eb);
3258}
3259
3260static int release_extent_buffer(struct extent_buffer *eb)
3261 __releases(&eb->refs_lock)
3262{
3263 lockdep_assert_held(&eb->refs_lock);
3264
3265 WARN_ON(atomic_read(&eb->refs) == 0);
3266 if (atomic_dec_and_test(&eb->refs)) {
3267 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
3268 struct btrfs_fs_info *fs_info = eb->fs_info;
3269
3270 spin_unlock(&eb->refs_lock);
3271
3272 spin_lock(&fs_info->buffer_lock);
3273 radix_tree_delete(&fs_info->buffer_radix,
3274 eb->start >> fs_info->sectorsize_bits);
3275 spin_unlock(&fs_info->buffer_lock);
3276 } else {
3277 spin_unlock(&eb->refs_lock);
3278 }
3279
3280 btrfs_leak_debug_del_eb(eb);
3281 /* Should be safe to release our pages at this point */
3282 btrfs_release_extent_buffer_pages(eb);
3283#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3284 if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
3285 __free_extent_buffer(eb);
3286 return 1;
3287 }
3288#endif
3289 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3290 return 1;
3291 }
3292 spin_unlock(&eb->refs_lock);
3293
3294 return 0;
3295}
3296
3297void free_extent_buffer(struct extent_buffer *eb)
3298{
3299 int refs;
3300 if (!eb)
3301 return;
3302
3303 refs = atomic_read(&eb->refs);
3304 while (1) {
3305 if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
3306 || (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
3307 refs == 1))
3308 break;
3309 if (atomic_try_cmpxchg(&eb->refs, &refs, refs - 1))
3310 return;
3311 }
3312
3313 spin_lock(&eb->refs_lock);
3314 if (atomic_read(&eb->refs) == 2 &&
3315 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
3316 !extent_buffer_under_io(eb) &&
3317 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3318 atomic_dec(&eb->refs);
3319
3320 /*
3321 * I know this is terrible, but it's temporary until we stop tracking
3322 * the uptodate bits and such for the extent buffers.
3323 */
3324 release_extent_buffer(eb);
3325}
3326
3327void free_extent_buffer_stale(struct extent_buffer *eb)
3328{
3329 if (!eb)
3330 return;
3331
3332 spin_lock(&eb->refs_lock);
3333 set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
3334
3335 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
3336 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3337 atomic_dec(&eb->refs);
3338 release_extent_buffer(eb);
3339}
3340
3341static void btree_clear_folio_dirty(struct folio *folio)
3342{
3343 ASSERT(folio_test_dirty(folio));
3344 ASSERT(folio_test_locked(folio));
3345 folio_clear_dirty_for_io(folio);
3346 xa_lock_irq(&folio->mapping->i_pages);
3347 if (!folio_test_dirty(folio))
3348 __xa_clear_mark(&folio->mapping->i_pages,
3349 folio_index(folio), PAGECACHE_TAG_DIRTY);
3350 xa_unlock_irq(&folio->mapping->i_pages);
3351}
3352
3353static void clear_subpage_extent_buffer_dirty(const struct extent_buffer *eb)
3354{
3355 struct btrfs_fs_info *fs_info = eb->fs_info;
3356 struct folio *folio = eb->folios[0];
3357 bool last;
3358
3359 /* btree_clear_folio_dirty() needs page locked. */
3360 folio_lock(folio);
3361 last = btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start, eb->len);
3362 if (last)
3363 btree_clear_folio_dirty(folio);
3364 folio_unlock(folio);
3365 WARN_ON(atomic_read(&eb->refs) == 0);
3366}
3367
3368void btrfs_clear_buffer_dirty(struct btrfs_trans_handle *trans,
3369 struct extent_buffer *eb)
3370{
3371 struct btrfs_fs_info *fs_info = eb->fs_info;
3372 int num_folios;
3373
3374 btrfs_assert_tree_write_locked(eb);
3375
3376 if (trans && btrfs_header_generation(eb) != trans->transid)
3377 return;
3378
3379 /*
3380 * Instead of clearing the dirty flag off of the buffer, mark it as
3381 * EXTENT_BUFFER_ZONED_ZEROOUT. This allows us to preserve
3382 * write-ordering in zoned mode, without the need to later re-dirty
3383 * the extent_buffer.
3384 *
3385 * The actual zeroout of the buffer will happen later in
3386 * btree_csum_one_bio.
3387 */
3388 if (btrfs_is_zoned(fs_info) && test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3389 set_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags);
3390 return;
3391 }
3392
3393 if (!test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags))
3394 return;
3395
3396 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes, -eb->len,
3397 fs_info->dirty_metadata_batch);
3398
3399 if (eb->fs_info->nodesize < PAGE_SIZE)
3400 return clear_subpage_extent_buffer_dirty(eb);
3401
3402 num_folios = num_extent_folios(eb);
3403 for (int i = 0; i < num_folios; i++) {
3404 struct folio *folio = eb->folios[i];
3405
3406 if (!folio_test_dirty(folio))
3407 continue;
3408 folio_lock(folio);
3409 btree_clear_folio_dirty(folio);
3410 folio_unlock(folio);
3411 }
3412 WARN_ON(atomic_read(&eb->refs) == 0);
3413}
3414
3415void set_extent_buffer_dirty(struct extent_buffer *eb)
3416{
3417 int num_folios;
3418 bool was_dirty;
3419
3420 check_buffer_tree_ref(eb);
3421
3422 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3423
3424 num_folios = num_extent_folios(eb);
3425 WARN_ON(atomic_read(&eb->refs) == 0);
3426 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
3427 WARN_ON(test_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags));
3428
3429 if (!was_dirty) {
3430 bool subpage = eb->fs_info->nodesize < PAGE_SIZE;
3431
3432 /*
3433 * For subpage case, we can have other extent buffers in the
3434 * same page, and in clear_subpage_extent_buffer_dirty() we
3435 * have to clear page dirty without subpage lock held.
3436 * This can cause race where our page gets dirty cleared after
3437 * we just set it.
3438 *
3439 * Thankfully, clear_subpage_extent_buffer_dirty() has locked
3440 * its page for other reasons, we can use page lock to prevent
3441 * the above race.
3442 */
3443 if (subpage)
3444 lock_page(folio_page(eb->folios[0], 0));
3445 for (int i = 0; i < num_folios; i++)
3446 btrfs_folio_set_dirty(eb->fs_info, eb->folios[i],
3447 eb->start, eb->len);
3448 if (subpage)
3449 unlock_page(folio_page(eb->folios[0], 0));
3450 percpu_counter_add_batch(&eb->fs_info->dirty_metadata_bytes,
3451 eb->len,
3452 eb->fs_info->dirty_metadata_batch);
3453 }
3454#ifdef CONFIG_BTRFS_DEBUG
3455 for (int i = 0; i < num_folios; i++)
3456 ASSERT(folio_test_dirty(eb->folios[i]));
3457#endif
3458}
3459
3460void clear_extent_buffer_uptodate(struct extent_buffer *eb)
3461{
3462 struct btrfs_fs_info *fs_info = eb->fs_info;
3463 int num_folios = num_extent_folios(eb);
3464
3465 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3466 for (int i = 0; i < num_folios; i++) {
3467 struct folio *folio = eb->folios[i];
3468
3469 if (!folio)
3470 continue;
3471
3472 /*
3473 * This is special handling for metadata subpage, as regular
3474 * btrfs_is_subpage() can not handle cloned/dummy metadata.
3475 */
3476 if (fs_info->nodesize >= PAGE_SIZE)
3477 folio_clear_uptodate(folio);
3478 else
3479 btrfs_subpage_clear_uptodate(fs_info, folio,
3480 eb->start, eb->len);
3481 }
3482}
3483
3484void set_extent_buffer_uptodate(struct extent_buffer *eb)
3485{
3486 struct btrfs_fs_info *fs_info = eb->fs_info;
3487 int num_folios = num_extent_folios(eb);
3488
3489 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3490 for (int i = 0; i < num_folios; i++) {
3491 struct folio *folio = eb->folios[i];
3492
3493 /*
3494 * This is special handling for metadata subpage, as regular
3495 * btrfs_is_subpage() can not handle cloned/dummy metadata.
3496 */
3497 if (fs_info->nodesize >= PAGE_SIZE)
3498 folio_mark_uptodate(folio);
3499 else
3500 btrfs_subpage_set_uptodate(fs_info, folio,
3501 eb->start, eb->len);
3502 }
3503}
3504
3505static void clear_extent_buffer_reading(struct extent_buffer *eb)
3506{
3507 clear_bit(EXTENT_BUFFER_READING, &eb->bflags);
3508 smp_mb__after_atomic();
3509 wake_up_bit(&eb->bflags, EXTENT_BUFFER_READING);
3510}
3511
3512static void end_bbio_meta_read(struct btrfs_bio *bbio)
3513{
3514 struct extent_buffer *eb = bbio->private;
3515 struct btrfs_fs_info *fs_info = eb->fs_info;
3516 bool uptodate = !bbio->bio.bi_status;
3517 struct folio_iter fi;
3518 u32 bio_offset = 0;
3519
3520 /*
3521 * If the extent buffer is marked UPTODATE before the read operation
3522 * completes, other calls to read_extent_buffer_pages() will return
3523 * early without waiting for the read to finish, causing data races.
3524 */
3525 WARN_ON(test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags));
3526
3527 eb->read_mirror = bbio->mirror_num;
3528
3529 if (uptodate &&
3530 btrfs_validate_extent_buffer(eb, &bbio->parent_check) < 0)
3531 uptodate = false;
3532
3533 if (uptodate) {
3534 set_extent_buffer_uptodate(eb);
3535 } else {
3536 clear_extent_buffer_uptodate(eb);
3537 set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
3538 }
3539
3540 bio_for_each_folio_all(fi, &bbio->bio) {
3541 struct folio *folio = fi.folio;
3542 u64 start = eb->start + bio_offset;
3543 u32 len = fi.length;
3544
3545 if (uptodate)
3546 btrfs_folio_set_uptodate(fs_info, folio, start, len);
3547 else
3548 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
3549
3550 bio_offset += len;
3551 }
3552
3553 clear_extent_buffer_reading(eb);
3554 free_extent_buffer(eb);
3555
3556 bio_put(&bbio->bio);
3557}
3558
3559int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num,
3560 const struct btrfs_tree_parent_check *check)
3561{
3562 struct btrfs_bio *bbio;
3563 bool ret;
3564
3565 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3566 return 0;
3567
3568 /*
3569 * We could have had EXTENT_BUFFER_UPTODATE cleared by the write
3570 * operation, which could potentially still be in flight. In this case
3571 * we simply want to return an error.
3572 */
3573 if (unlikely(test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)))
3574 return -EIO;
3575
3576 /* Someone else is already reading the buffer, just wait for it. */
3577 if (test_and_set_bit(EXTENT_BUFFER_READING, &eb->bflags))
3578 goto done;
3579
3580 /*
3581 * Between the initial test_bit(EXTENT_BUFFER_UPTODATE) and the above
3582 * test_and_set_bit(EXTENT_BUFFER_READING), someone else could have
3583 * started and finished reading the same eb. In this case, UPTODATE
3584 * will now be set, and we shouldn't read it in again.
3585 */
3586 if (unlikely(test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))) {
3587 clear_extent_buffer_reading(eb);
3588 return 0;
3589 }
3590
3591 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
3592 eb->read_mirror = 0;
3593 check_buffer_tree_ref(eb);
3594 atomic_inc(&eb->refs);
3595
3596 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
3597 REQ_OP_READ | REQ_META, eb->fs_info,
3598 end_bbio_meta_read, eb);
3599 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
3600 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
3601 bbio->file_offset = eb->start;
3602 memcpy(&bbio->parent_check, check, sizeof(*check));
3603 if (eb->fs_info->nodesize < PAGE_SIZE) {
3604 ret = bio_add_folio(&bbio->bio, eb->folios[0], eb->len,
3605 eb->start - folio_pos(eb->folios[0]));
3606 ASSERT(ret);
3607 } else {
3608 int num_folios = num_extent_folios(eb);
3609
3610 for (int i = 0; i < num_folios; i++) {
3611 struct folio *folio = eb->folios[i];
3612
3613 ret = bio_add_folio(&bbio->bio, folio, eb->folio_size, 0);
3614 ASSERT(ret);
3615 }
3616 }
3617 btrfs_submit_bbio(bbio, mirror_num);
3618
3619done:
3620 if (wait == WAIT_COMPLETE) {
3621 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_READING, TASK_UNINTERRUPTIBLE);
3622 if (!test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3623 return -EIO;
3624 }
3625
3626 return 0;
3627}
3628
3629static bool report_eb_range(const struct extent_buffer *eb, unsigned long start,
3630 unsigned long len)
3631{
3632 btrfs_warn(eb->fs_info,
3633 "access to eb bytenr %llu len %u out of range start %lu len %lu",
3634 eb->start, eb->len, start, len);
3635 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
3636
3637 return true;
3638}
3639
3640/*
3641 * Check if the [start, start + len) range is valid before reading/writing
3642 * the eb.
3643 * NOTE: @start and @len are offset inside the eb, not logical address.
3644 *
3645 * Caller should not touch the dst/src memory if this function returns error.
3646 */
3647static inline int check_eb_range(const struct extent_buffer *eb,
3648 unsigned long start, unsigned long len)
3649{
3650 unsigned long offset;
3651
3652 /* start, start + len should not go beyond eb->len nor overflow */
3653 if (unlikely(check_add_overflow(start, len, &offset) || offset > eb->len))
3654 return report_eb_range(eb, start, len);
3655
3656 return false;
3657}
3658
3659void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
3660 unsigned long start, unsigned long len)
3661{
3662 const int unit_size = eb->folio_size;
3663 size_t cur;
3664 size_t offset;
3665 char *dst = (char *)dstv;
3666 unsigned long i = get_eb_folio_index(eb, start);
3667
3668 if (check_eb_range(eb, start, len)) {
3669 /*
3670 * Invalid range hit, reset the memory, so callers won't get
3671 * some random garbage for their uninitialized memory.
3672 */
3673 memset(dstv, 0, len);
3674 return;
3675 }
3676
3677 if (eb->addr) {
3678 memcpy(dstv, eb->addr + start, len);
3679 return;
3680 }
3681
3682 offset = get_eb_offset_in_folio(eb, start);
3683
3684 while (len > 0) {
3685 char *kaddr;
3686
3687 cur = min(len, unit_size - offset);
3688 kaddr = folio_address(eb->folios[i]);
3689 memcpy(dst, kaddr + offset, cur);
3690
3691 dst += cur;
3692 len -= cur;
3693 offset = 0;
3694 i++;
3695 }
3696}
3697
3698int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
3699 void __user *dstv,
3700 unsigned long start, unsigned long len)
3701{
3702 const int unit_size = eb->folio_size;
3703 size_t cur;
3704 size_t offset;
3705 char __user *dst = (char __user *)dstv;
3706 unsigned long i = get_eb_folio_index(eb, start);
3707 int ret = 0;
3708
3709 WARN_ON(start > eb->len);
3710 WARN_ON(start + len > eb->start + eb->len);
3711
3712 if (eb->addr) {
3713 if (copy_to_user_nofault(dstv, eb->addr + start, len))
3714 ret = -EFAULT;
3715 return ret;
3716 }
3717
3718 offset = get_eb_offset_in_folio(eb, start);
3719
3720 while (len > 0) {
3721 char *kaddr;
3722
3723 cur = min(len, unit_size - offset);
3724 kaddr = folio_address(eb->folios[i]);
3725 if (copy_to_user_nofault(dst, kaddr + offset, cur)) {
3726 ret = -EFAULT;
3727 break;
3728 }
3729
3730 dst += cur;
3731 len -= cur;
3732 offset = 0;
3733 i++;
3734 }
3735
3736 return ret;
3737}
3738
3739int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
3740 unsigned long start, unsigned long len)
3741{
3742 const int unit_size = eb->folio_size;
3743 size_t cur;
3744 size_t offset;
3745 char *kaddr;
3746 char *ptr = (char *)ptrv;
3747 unsigned long i = get_eb_folio_index(eb, start);
3748 int ret = 0;
3749
3750 if (check_eb_range(eb, start, len))
3751 return -EINVAL;
3752
3753 if (eb->addr)
3754 return memcmp(ptrv, eb->addr + start, len);
3755
3756 offset = get_eb_offset_in_folio(eb, start);
3757
3758 while (len > 0) {
3759 cur = min(len, unit_size - offset);
3760 kaddr = folio_address(eb->folios[i]);
3761 ret = memcmp(ptr, kaddr + offset, cur);
3762 if (ret)
3763 break;
3764
3765 ptr += cur;
3766 len -= cur;
3767 offset = 0;
3768 i++;
3769 }
3770 return ret;
3771}
3772
3773/*
3774 * Check that the extent buffer is uptodate.
3775 *
3776 * For regular sector size == PAGE_SIZE case, check if @page is uptodate.
3777 * For subpage case, check if the range covered by the eb has EXTENT_UPTODATE.
3778 */
3779static void assert_eb_folio_uptodate(const struct extent_buffer *eb, int i)
3780{
3781 struct btrfs_fs_info *fs_info = eb->fs_info;
3782 struct folio *folio = eb->folios[i];
3783
3784 ASSERT(folio);
3785
3786 /*
3787 * If we are using the commit root we could potentially clear a page
3788 * Uptodate while we're using the extent buffer that we've previously
3789 * looked up. We don't want to complain in this case, as the page was
3790 * valid before, we just didn't write it out. Instead we want to catch
3791 * the case where we didn't actually read the block properly, which
3792 * would have !PageUptodate and !EXTENT_BUFFER_WRITE_ERR.
3793 */
3794 if (test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3795 return;
3796
3797 if (fs_info->nodesize < PAGE_SIZE) {
3798 folio = eb->folios[0];
3799 ASSERT(i == 0);
3800 if (WARN_ON(!btrfs_subpage_test_uptodate(fs_info, folio,
3801 eb->start, eb->len)))
3802 btrfs_subpage_dump_bitmap(fs_info, folio, eb->start, eb->len);
3803 } else {
3804 WARN_ON(!folio_test_uptodate(folio));
3805 }
3806}
3807
3808static void __write_extent_buffer(const struct extent_buffer *eb,
3809 const void *srcv, unsigned long start,
3810 unsigned long len, bool use_memmove)
3811{
3812 const int unit_size = eb->folio_size;
3813 size_t cur;
3814 size_t offset;
3815 char *kaddr;
3816 const char *src = (const char *)srcv;
3817 unsigned long i = get_eb_folio_index(eb, start);
3818 /* For unmapped (dummy) ebs, no need to check their uptodate status. */
3819 const bool check_uptodate = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3820
3821 if (check_eb_range(eb, start, len))
3822 return;
3823
3824 if (eb->addr) {
3825 if (use_memmove)
3826 memmove(eb->addr + start, srcv, len);
3827 else
3828 memcpy(eb->addr + start, srcv, len);
3829 return;
3830 }
3831
3832 offset = get_eb_offset_in_folio(eb, start);
3833
3834 while (len > 0) {
3835 if (check_uptodate)
3836 assert_eb_folio_uptodate(eb, i);
3837
3838 cur = min(len, unit_size - offset);
3839 kaddr = folio_address(eb->folios[i]);
3840 if (use_memmove)
3841 memmove(kaddr + offset, src, cur);
3842 else
3843 memcpy(kaddr + offset, src, cur);
3844
3845 src += cur;
3846 len -= cur;
3847 offset = 0;
3848 i++;
3849 }
3850}
3851
3852void write_extent_buffer(const struct extent_buffer *eb, const void *srcv,
3853 unsigned long start, unsigned long len)
3854{
3855 return __write_extent_buffer(eb, srcv, start, len, false);
3856}
3857
3858static void memset_extent_buffer(const struct extent_buffer *eb, int c,
3859 unsigned long start, unsigned long len)
3860{
3861 const int unit_size = eb->folio_size;
3862 unsigned long cur = start;
3863
3864 if (eb->addr) {
3865 memset(eb->addr + start, c, len);
3866 return;
3867 }
3868
3869 while (cur < start + len) {
3870 unsigned long index = get_eb_folio_index(eb, cur);
3871 unsigned int offset = get_eb_offset_in_folio(eb, cur);
3872 unsigned int cur_len = min(start + len - cur, unit_size - offset);
3873
3874 assert_eb_folio_uptodate(eb, index);
3875 memset(folio_address(eb->folios[index]) + offset, c, cur_len);
3876
3877 cur += cur_len;
3878 }
3879}
3880
3881void memzero_extent_buffer(const struct extent_buffer *eb, unsigned long start,
3882 unsigned long len)
3883{
3884 if (check_eb_range(eb, start, len))
3885 return;
3886 return memset_extent_buffer(eb, 0, start, len);
3887}
3888
3889void copy_extent_buffer_full(const struct extent_buffer *dst,
3890 const struct extent_buffer *src)
3891{
3892 const int unit_size = src->folio_size;
3893 unsigned long cur = 0;
3894
3895 ASSERT(dst->len == src->len);
3896
3897 while (cur < src->len) {
3898 unsigned long index = get_eb_folio_index(src, cur);
3899 unsigned long offset = get_eb_offset_in_folio(src, cur);
3900 unsigned long cur_len = min(src->len, unit_size - offset);
3901 void *addr = folio_address(src->folios[index]) + offset;
3902
3903 write_extent_buffer(dst, addr, cur, cur_len);
3904
3905 cur += cur_len;
3906 }
3907}
3908
3909void copy_extent_buffer(const struct extent_buffer *dst,
3910 const struct extent_buffer *src,
3911 unsigned long dst_offset, unsigned long src_offset,
3912 unsigned long len)
3913{
3914 const int unit_size = dst->folio_size;
3915 u64 dst_len = dst->len;
3916 size_t cur;
3917 size_t offset;
3918 char *kaddr;
3919 unsigned long i = get_eb_folio_index(dst, dst_offset);
3920
3921 if (check_eb_range(dst, dst_offset, len) ||
3922 check_eb_range(src, src_offset, len))
3923 return;
3924
3925 WARN_ON(src->len != dst_len);
3926
3927 offset = get_eb_offset_in_folio(dst, dst_offset);
3928
3929 while (len > 0) {
3930 assert_eb_folio_uptodate(dst, i);
3931
3932 cur = min(len, (unsigned long)(unit_size - offset));
3933
3934 kaddr = folio_address(dst->folios[i]);
3935 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3936
3937 src_offset += cur;
3938 len -= cur;
3939 offset = 0;
3940 i++;
3941 }
3942}
3943
3944/*
3945 * Calculate the folio and offset of the byte containing the given bit number.
3946 *
3947 * @eb: the extent buffer
3948 * @start: offset of the bitmap item in the extent buffer
3949 * @nr: bit number
3950 * @folio_index: return index of the folio in the extent buffer that contains
3951 * the given bit number
3952 * @folio_offset: return offset into the folio given by folio_index
3953 *
3954 * This helper hides the ugliness of finding the byte in an extent buffer which
3955 * contains a given bit.
3956 */
3957static inline void eb_bitmap_offset(const struct extent_buffer *eb,
3958 unsigned long start, unsigned long nr,
3959 unsigned long *folio_index,
3960 size_t *folio_offset)
3961{
3962 size_t byte_offset = BIT_BYTE(nr);
3963 size_t offset;
3964
3965 /*
3966 * The byte we want is the offset of the extent buffer + the offset of
3967 * the bitmap item in the extent buffer + the offset of the byte in the
3968 * bitmap item.
3969 */
3970 offset = start + offset_in_eb_folio(eb, eb->start) + byte_offset;
3971
3972 *folio_index = offset >> eb->folio_shift;
3973 *folio_offset = offset_in_eb_folio(eb, offset);
3974}
3975
3976/*
3977 * Determine whether a bit in a bitmap item is set.
3978 *
3979 * @eb: the extent buffer
3980 * @start: offset of the bitmap item in the extent buffer
3981 * @nr: bit number to test
3982 */
3983int extent_buffer_test_bit(const struct extent_buffer *eb, unsigned long start,
3984 unsigned long nr)
3985{
3986 unsigned long i;
3987 size_t offset;
3988 u8 *kaddr;
3989
3990 eb_bitmap_offset(eb, start, nr, &i, &offset);
3991 assert_eb_folio_uptodate(eb, i);
3992 kaddr = folio_address(eb->folios[i]);
3993 return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
3994}
3995
3996static u8 *extent_buffer_get_byte(const struct extent_buffer *eb, unsigned long bytenr)
3997{
3998 unsigned long index = get_eb_folio_index(eb, bytenr);
3999
4000 if (check_eb_range(eb, bytenr, 1))
4001 return NULL;
4002 return folio_address(eb->folios[index]) + get_eb_offset_in_folio(eb, bytenr);
4003}
4004
4005/*
4006 * Set an area of a bitmap to 1.
4007 *
4008 * @eb: the extent buffer
4009 * @start: offset of the bitmap item in the extent buffer
4010 * @pos: bit number of the first bit
4011 * @len: number of bits to set
4012 */
4013void extent_buffer_bitmap_set(const struct extent_buffer *eb, unsigned long start,
4014 unsigned long pos, unsigned long len)
4015{
4016 unsigned int first_byte = start + BIT_BYTE(pos);
4017 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4018 const bool same_byte = (first_byte == last_byte);
4019 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4020 u8 *kaddr;
4021
4022 if (same_byte)
4023 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4024
4025 /* Handle the first byte. */
4026 kaddr = extent_buffer_get_byte(eb, first_byte);
4027 *kaddr |= mask;
4028 if (same_byte)
4029 return;
4030
4031 /* Handle the byte aligned part. */
4032 ASSERT(first_byte + 1 <= last_byte);
4033 memset_extent_buffer(eb, 0xff, first_byte + 1, last_byte - first_byte - 1);
4034
4035 /* Handle the last byte. */
4036 kaddr = extent_buffer_get_byte(eb, last_byte);
4037 *kaddr |= BITMAP_LAST_BYTE_MASK(pos + len);
4038}
4039
4040
4041/*
4042 * Clear an area of a bitmap.
4043 *
4044 * @eb: the extent buffer
4045 * @start: offset of the bitmap item in the extent buffer
4046 * @pos: bit number of the first bit
4047 * @len: number of bits to clear
4048 */
4049void extent_buffer_bitmap_clear(const struct extent_buffer *eb,
4050 unsigned long start, unsigned long pos,
4051 unsigned long len)
4052{
4053 unsigned int first_byte = start + BIT_BYTE(pos);
4054 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4055 const bool same_byte = (first_byte == last_byte);
4056 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4057 u8 *kaddr;
4058
4059 if (same_byte)
4060 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4061
4062 /* Handle the first byte. */
4063 kaddr = extent_buffer_get_byte(eb, first_byte);
4064 *kaddr &= ~mask;
4065 if (same_byte)
4066 return;
4067
4068 /* Handle the byte aligned part. */
4069 ASSERT(first_byte + 1 <= last_byte);
4070 memset_extent_buffer(eb, 0, first_byte + 1, last_byte - first_byte - 1);
4071
4072 /* Handle the last byte. */
4073 kaddr = extent_buffer_get_byte(eb, last_byte);
4074 *kaddr &= ~BITMAP_LAST_BYTE_MASK(pos + len);
4075}
4076
4077static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4078{
4079 unsigned long distance = (src > dst) ? src - dst : dst - src;
4080 return distance < len;
4081}
4082
4083void memcpy_extent_buffer(const struct extent_buffer *dst,
4084 unsigned long dst_offset, unsigned long src_offset,
4085 unsigned long len)
4086{
4087 const int unit_size = dst->folio_size;
4088 unsigned long cur_off = 0;
4089
4090 if (check_eb_range(dst, dst_offset, len) ||
4091 check_eb_range(dst, src_offset, len))
4092 return;
4093
4094 if (dst->addr) {
4095 const bool use_memmove = areas_overlap(src_offset, dst_offset, len);
4096
4097 if (use_memmove)
4098 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4099 else
4100 memcpy(dst->addr + dst_offset, dst->addr + src_offset, len);
4101 return;
4102 }
4103
4104 while (cur_off < len) {
4105 unsigned long cur_src = cur_off + src_offset;
4106 unsigned long folio_index = get_eb_folio_index(dst, cur_src);
4107 unsigned long folio_off = get_eb_offset_in_folio(dst, cur_src);
4108 unsigned long cur_len = min(src_offset + len - cur_src,
4109 unit_size - folio_off);
4110 void *src_addr = folio_address(dst->folios[folio_index]) + folio_off;
4111 const bool use_memmove = areas_overlap(src_offset + cur_off,
4112 dst_offset + cur_off, cur_len);
4113
4114 __write_extent_buffer(dst, src_addr, dst_offset + cur_off, cur_len,
4115 use_memmove);
4116 cur_off += cur_len;
4117 }
4118}
4119
4120void memmove_extent_buffer(const struct extent_buffer *dst,
4121 unsigned long dst_offset, unsigned long src_offset,
4122 unsigned long len)
4123{
4124 unsigned long dst_end = dst_offset + len - 1;
4125 unsigned long src_end = src_offset + len - 1;
4126
4127 if (check_eb_range(dst, dst_offset, len) ||
4128 check_eb_range(dst, src_offset, len))
4129 return;
4130
4131 if (dst_offset < src_offset) {
4132 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4133 return;
4134 }
4135
4136 if (dst->addr) {
4137 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4138 return;
4139 }
4140
4141 while (len > 0) {
4142 unsigned long src_i;
4143 size_t cur;
4144 size_t dst_off_in_folio;
4145 size_t src_off_in_folio;
4146 void *src_addr;
4147 bool use_memmove;
4148
4149 src_i = get_eb_folio_index(dst, src_end);
4150
4151 dst_off_in_folio = get_eb_offset_in_folio(dst, dst_end);
4152 src_off_in_folio = get_eb_offset_in_folio(dst, src_end);
4153
4154 cur = min_t(unsigned long, len, src_off_in_folio + 1);
4155 cur = min(cur, dst_off_in_folio + 1);
4156
4157 src_addr = folio_address(dst->folios[src_i]) + src_off_in_folio -
4158 cur + 1;
4159 use_memmove = areas_overlap(src_end - cur + 1, dst_end - cur + 1,
4160 cur);
4161
4162 __write_extent_buffer(dst, src_addr, dst_end - cur + 1, cur,
4163 use_memmove);
4164
4165 dst_end -= cur;
4166 src_end -= cur;
4167 len -= cur;
4168 }
4169}
4170
4171#define GANG_LOOKUP_SIZE 16
4172static struct extent_buffer *get_next_extent_buffer(
4173 const struct btrfs_fs_info *fs_info, struct folio *folio, u64 bytenr)
4174{
4175 struct extent_buffer *gang[GANG_LOOKUP_SIZE];
4176 struct extent_buffer *found = NULL;
4177 u64 folio_start = folio_pos(folio);
4178 u64 cur = folio_start;
4179
4180 ASSERT(in_range(bytenr, folio_start, PAGE_SIZE));
4181 lockdep_assert_held(&fs_info->buffer_lock);
4182
4183 while (cur < folio_start + PAGE_SIZE) {
4184 int ret;
4185 int i;
4186
4187 ret = radix_tree_gang_lookup(&fs_info->buffer_radix,
4188 (void **)gang, cur >> fs_info->sectorsize_bits,
4189 min_t(unsigned int, GANG_LOOKUP_SIZE,
4190 PAGE_SIZE / fs_info->nodesize));
4191 if (ret == 0)
4192 goto out;
4193 for (i = 0; i < ret; i++) {
4194 /* Already beyond page end */
4195 if (gang[i]->start >= folio_start + PAGE_SIZE)
4196 goto out;
4197 /* Found one */
4198 if (gang[i]->start >= bytenr) {
4199 found = gang[i];
4200 goto out;
4201 }
4202 }
4203 cur = gang[ret - 1]->start + gang[ret - 1]->len;
4204 }
4205out:
4206 return found;
4207}
4208
4209static int try_release_subpage_extent_buffer(struct folio *folio)
4210{
4211 struct btrfs_fs_info *fs_info = folio_to_fs_info(folio);
4212 u64 cur = folio_pos(folio);
4213 const u64 end = cur + PAGE_SIZE;
4214 int ret;
4215
4216 while (cur < end) {
4217 struct extent_buffer *eb = NULL;
4218
4219 /*
4220 * Unlike try_release_extent_buffer() which uses folio private
4221 * to grab buffer, for subpage case we rely on radix tree, thus
4222 * we need to ensure radix tree consistency.
4223 *
4224 * We also want an atomic snapshot of the radix tree, thus go
4225 * with spinlock rather than RCU.
4226 */
4227 spin_lock(&fs_info->buffer_lock);
4228 eb = get_next_extent_buffer(fs_info, folio, cur);
4229 if (!eb) {
4230 /* No more eb in the page range after or at cur */
4231 spin_unlock(&fs_info->buffer_lock);
4232 break;
4233 }
4234 cur = eb->start + eb->len;
4235
4236 /*
4237 * The same as try_release_extent_buffer(), to ensure the eb
4238 * won't disappear out from under us.
4239 */
4240 spin_lock(&eb->refs_lock);
4241 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4242 spin_unlock(&eb->refs_lock);
4243 spin_unlock(&fs_info->buffer_lock);
4244 break;
4245 }
4246 spin_unlock(&fs_info->buffer_lock);
4247
4248 /*
4249 * If tree ref isn't set then we know the ref on this eb is a
4250 * real ref, so just return, this eb will likely be freed soon
4251 * anyway.
4252 */
4253 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4254 spin_unlock(&eb->refs_lock);
4255 break;
4256 }
4257
4258 /*
4259 * Here we don't care about the return value, we will always
4260 * check the folio private at the end. And
4261 * release_extent_buffer() will release the refs_lock.
4262 */
4263 release_extent_buffer(eb);
4264 }
4265 /*
4266 * Finally to check if we have cleared folio private, as if we have
4267 * released all ebs in the page, the folio private should be cleared now.
4268 */
4269 spin_lock(&folio->mapping->i_private_lock);
4270 if (!folio_test_private(folio))
4271 ret = 1;
4272 else
4273 ret = 0;
4274 spin_unlock(&folio->mapping->i_private_lock);
4275 return ret;
4276
4277}
4278
4279int try_release_extent_buffer(struct folio *folio)
4280{
4281 struct extent_buffer *eb;
4282
4283 if (folio_to_fs_info(folio)->nodesize < PAGE_SIZE)
4284 return try_release_subpage_extent_buffer(folio);
4285
4286 /*
4287 * We need to make sure nobody is changing folio private, as we rely on
4288 * folio private as the pointer to extent buffer.
4289 */
4290 spin_lock(&folio->mapping->i_private_lock);
4291 if (!folio_test_private(folio)) {
4292 spin_unlock(&folio->mapping->i_private_lock);
4293 return 1;
4294 }
4295
4296 eb = folio_get_private(folio);
4297 BUG_ON(!eb);
4298
4299 /*
4300 * This is a little awful but should be ok, we need to make sure that
4301 * the eb doesn't disappear out from under us while we're looking at
4302 * this page.
4303 */
4304 spin_lock(&eb->refs_lock);
4305 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4306 spin_unlock(&eb->refs_lock);
4307 spin_unlock(&folio->mapping->i_private_lock);
4308 return 0;
4309 }
4310 spin_unlock(&folio->mapping->i_private_lock);
4311
4312 /*
4313 * If tree ref isn't set then we know the ref on this eb is a real ref,
4314 * so just return, this page will likely be freed soon anyway.
4315 */
4316 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4317 spin_unlock(&eb->refs_lock);
4318 return 0;
4319 }
4320
4321 return release_extent_buffer(eb);
4322}
4323
4324/*
4325 * Attempt to readahead a child block.
4326 *
4327 * @fs_info: the fs_info
4328 * @bytenr: bytenr to read
4329 * @owner_root: objectid of the root that owns this eb
4330 * @gen: generation for the uptodate check, can be 0
4331 * @level: level for the eb
4332 *
4333 * Attempt to readahead a tree block at @bytenr. If @gen is 0 then we do a
4334 * normal uptodate check of the eb, without checking the generation. If we have
4335 * to read the block we will not block on anything.
4336 */
4337void btrfs_readahead_tree_block(struct btrfs_fs_info *fs_info,
4338 u64 bytenr, u64 owner_root, u64 gen, int level)
4339{
4340 struct btrfs_tree_parent_check check = {
4341 .level = level,
4342 .transid = gen
4343 };
4344 struct extent_buffer *eb;
4345 int ret;
4346
4347 eb = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
4348 if (IS_ERR(eb))
4349 return;
4350
4351 if (btrfs_buffer_uptodate(eb, gen, 1)) {
4352 free_extent_buffer(eb);
4353 return;
4354 }
4355
4356 ret = read_extent_buffer_pages(eb, WAIT_NONE, 0, &check);
4357 if (ret < 0)
4358 free_extent_buffer_stale(eb);
4359 else
4360 free_extent_buffer(eb);
4361}
4362
4363/*
4364 * Readahead a node's child block.
4365 *
4366 * @node: parent node we're reading from
4367 * @slot: slot in the parent node for the child we want to read
4368 *
4369 * A helper for btrfs_readahead_tree_block, we simply read the bytenr pointed at
4370 * the slot in the node provided.
4371 */
4372void btrfs_readahead_node_child(struct extent_buffer *node, int slot)
4373{
4374 btrfs_readahead_tree_block(node->fs_info,
4375 btrfs_node_blockptr(node, slot),
4376 btrfs_header_owner(node),
4377 btrfs_node_ptr_generation(node, slot),
4378 btrfs_header_level(node) - 1);
4379}