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