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