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
38 static struct kmem_cache *extent_buffer_cache;
39
40 #ifdef CONFIG_BTRFS_DEBUG
41 static 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
51 static 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
61 void 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 */
97 struct 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 static 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 */
129 static 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
146 int __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
157 void __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
167 static void process_one_page(struct btrfs_fs_info *fs_info,
168 struct page *page, const struct page *locked_page,
169 unsigned long page_ops, u64 start, u64 end)
170 {
171 struct folio *folio = page_folio(page);
172 u32 len;
173
174 ASSERT(end + 1 - start != 0 && end + 1 - start < U32_MAX);
175 len = end + 1 - start;
176
177 if (page_ops & PAGE_SET_ORDERED)
178 btrfs_folio_clamp_set_ordered(fs_info, folio, start, len);
179 if (page_ops & PAGE_START_WRITEBACK) {
180 btrfs_folio_clamp_clear_dirty(fs_info, folio, start, len);
181 btrfs_folio_clamp_set_writeback(fs_info, folio, start, len);
182 }
183 if (page_ops & PAGE_END_WRITEBACK)
184 btrfs_folio_clamp_clear_writeback(fs_info, folio, start, len);
185
186 if (page != locked_page && (page_ops & PAGE_UNLOCK))
187 btrfs_folio_end_writer_lock(fs_info, folio, start, len);
188 }
189
190 static void __process_pages_contig(struct address_space *mapping,
191 const struct page *locked_page, u64 start, u64 end,
192 unsigned long page_ops)
193 {
194 struct btrfs_fs_info *fs_info = inode_to_fs_info(mapping->host);
195 pgoff_t start_index = start >> PAGE_SHIFT;
196 pgoff_t end_index = end >> PAGE_SHIFT;
197 pgoff_t index = start_index;
198 struct folio_batch fbatch;
199 int i;
200
201 folio_batch_init(&fbatch);
202 while (index <= end_index) {
203 int found_folios;
204
205 found_folios = filemap_get_folios_contig(mapping, &index,
206 end_index, &fbatch);
207 for (i = 0; i < found_folios; i++) {
208 struct folio *folio = fbatch.folios[i];
209
210 process_one_page(fs_info, &folio->page, locked_page,
211 page_ops, start, end);
212 }
213 folio_batch_release(&fbatch);
214 cond_resched();
215 }
216 }
217
218 static noinline void __unlock_for_delalloc(const struct inode *inode,
219 const struct page *locked_page,
220 u64 start, u64 end)
221 {
222 unsigned long index = start >> PAGE_SHIFT;
223 unsigned long end_index = end >> PAGE_SHIFT;
224
225 ASSERT(locked_page);
226 if (index == locked_page->index && end_index == index)
227 return;
228
229 __process_pages_contig(inode->i_mapping, locked_page, start, end,
230 PAGE_UNLOCK);
231 }
232
233 static noinline int lock_delalloc_pages(struct inode *inode,
234 const struct page *locked_page,
235 u64 start,
236 u64 end)
237 {
238 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
239 struct address_space *mapping = inode->i_mapping;
240 pgoff_t start_index = start >> PAGE_SHIFT;
241 pgoff_t end_index = end >> PAGE_SHIFT;
242 pgoff_t index = start_index;
243 u64 processed_end = start;
244 struct folio_batch fbatch;
245
246 if (index == locked_page->index && index == end_index)
247 return 0;
248
249 folio_batch_init(&fbatch);
250 while (index <= end_index) {
251 unsigned int found_folios, i;
252
253 found_folios = filemap_get_folios_contig(mapping, &index,
254 end_index, &fbatch);
255 if (found_folios == 0)
256 goto out;
257
258 for (i = 0; i < found_folios; i++) {
259 struct folio *folio = fbatch.folios[i];
260 struct page *page = folio_page(folio, 0);
261 u32 len = end + 1 - start;
262
263 if (page == locked_page)
264 continue;
265
266 if (btrfs_folio_start_writer_lock(fs_info, folio, start,
267 len))
268 goto out;
269
270 if (!PageDirty(page) || page->mapping != mapping) {
271 btrfs_folio_end_writer_lock(fs_info, folio, start,
272 len);
273 goto out;
274 }
275
276 processed_end = page_offset(page) + PAGE_SIZE - 1;
277 }
278 folio_batch_release(&fbatch);
279 cond_resched();
280 }
281
282 return 0;
283 out:
284 folio_batch_release(&fbatch);
285 if (processed_end > start)
286 __unlock_for_delalloc(inode, locked_page, start, processed_end);
287 return -EAGAIN;
288 }
289
290 /*
291 * Find and lock a contiguous range of bytes in the file marked as delalloc, no
292 * more than @max_bytes.
293 *
294 * @start: The original start bytenr to search.
295 * Will store the extent range start bytenr.
296 * @end: The original end bytenr of the search range
297 * Will store the extent range end bytenr.
298 *
299 * Return true if we find a delalloc range which starts inside the original
300 * range, and @start/@end will store the delalloc range start/end.
301 *
302 * Return false if we can't find any delalloc range which starts inside the
303 * original range, and @start/@end will be the non-delalloc range start/end.
304 */
305 EXPORT_FOR_TESTS
306 noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
307 struct page *locked_page, u64 *start,
308 u64 *end)
309 {
310 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
311 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
312 const u64 orig_start = *start;
313 const u64 orig_end = *end;
314 /* The sanity tests may not set a valid fs_info. */
315 u64 max_bytes = fs_info ? fs_info->max_extent_size : BTRFS_MAX_EXTENT_SIZE;
316 u64 delalloc_start;
317 u64 delalloc_end;
318 bool found;
319 struct extent_state *cached_state = NULL;
320 int ret;
321 int loops = 0;
322
323 /* Caller should pass a valid @end to indicate the search range end */
324 ASSERT(orig_end > orig_start);
325
326 /* The range should at least cover part of the page */
327 ASSERT(!(orig_start >= page_offset(locked_page) + PAGE_SIZE ||
328 orig_end <= page_offset(locked_page)));
329 again:
330 /* step one, find a bunch of delalloc bytes starting at start */
331 delalloc_start = *start;
332 delalloc_end = 0;
333 found = btrfs_find_delalloc_range(tree, &delalloc_start, &delalloc_end,
334 max_bytes, &cached_state);
335 if (!found || delalloc_end <= *start || delalloc_start > orig_end) {
336 *start = delalloc_start;
337
338 /* @delalloc_end can be -1, never go beyond @orig_end */
339 *end = min(delalloc_end, orig_end);
340 free_extent_state(cached_state);
341 return false;
342 }
343
344 /*
345 * start comes from the offset of locked_page. We have to lock
346 * pages in order, so we can't process delalloc bytes before
347 * locked_page
348 */
349 if (delalloc_start < *start)
350 delalloc_start = *start;
351
352 /*
353 * make sure to limit the number of pages we try to lock down
354 */
355 if (delalloc_end + 1 - delalloc_start > max_bytes)
356 delalloc_end = delalloc_start + max_bytes - 1;
357
358 /* step two, lock all the pages after the page that has start */
359 ret = lock_delalloc_pages(inode, locked_page,
360 delalloc_start, delalloc_end);
361 ASSERT(!ret || ret == -EAGAIN);
362 if (ret == -EAGAIN) {
363 /* some of the pages are gone, lets avoid looping by
364 * shortening the size of the delalloc range we're searching
365 */
366 free_extent_state(cached_state);
367 cached_state = NULL;
368 if (!loops) {
369 max_bytes = PAGE_SIZE;
370 loops = 1;
371 goto again;
372 } else {
373 found = false;
374 goto out_failed;
375 }
376 }
377
378 /* step three, lock the state bits for the whole range */
379 lock_extent(tree, delalloc_start, delalloc_end, &cached_state);
380
381 /* then test to make sure it is all still delalloc */
382 ret = test_range_bit(tree, delalloc_start, delalloc_end,
383 EXTENT_DELALLOC, cached_state);
384
385 unlock_extent(tree, delalloc_start, delalloc_end, &cached_state);
386 if (!ret) {
387 __unlock_for_delalloc(inode, locked_page,
388 delalloc_start, delalloc_end);
389 cond_resched();
390 goto again;
391 }
392 *start = delalloc_start;
393 *end = delalloc_end;
394 out_failed:
395 return found;
396 }
397
398 void extent_clear_unlock_delalloc(struct btrfs_inode *inode, u64 start, u64 end,
399 const struct page *locked_page,
400 struct extent_state **cached,
401 u32 clear_bits, unsigned long page_ops)
402 {
403 clear_extent_bit(&inode->io_tree, start, end, clear_bits, cached);
404
405 __process_pages_contig(inode->vfs_inode.i_mapping, locked_page,
406 start, end, page_ops);
407 }
408
409 static bool btrfs_verify_page(struct page *page, u64 start)
410 {
411 if (!fsverity_active(page->mapping->host) ||
412 PageUptodate(page) ||
413 start >= i_size_read(page->mapping->host))
414 return true;
415 return fsverity_verify_page(page);
416 }
417
418 static void end_page_read(struct page *page, bool uptodate, u64 start, u32 len)
419 {
420 struct btrfs_fs_info *fs_info = page_to_fs_info(page);
421 struct folio *folio = page_folio(page);
422
423 ASSERT(page_offset(page) <= start &&
424 start + len <= page_offset(page) + PAGE_SIZE);
425
426 if (uptodate && btrfs_verify_page(page, start))
427 btrfs_folio_set_uptodate(fs_info, folio, start, len);
428 else
429 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
430
431 if (!btrfs_is_subpage(fs_info, page->mapping))
432 unlock_page(page);
433 else
434 btrfs_subpage_end_reader(fs_info, folio, start, len);
435 }
436
437 /*
438 * After a write IO is done, we need to:
439 *
440 * - clear the uptodate bits on error
441 * - clear the writeback bits in the extent tree for the range
442 * - filio_end_writeback() if there is no more pending io for the folio
443 *
444 * Scheduling is not allowed, so the extent state tree is expected
445 * to have one and only one object corresponding to this IO.
446 */
447 static void end_bbio_data_write(struct btrfs_bio *bbio)
448 {
449 struct btrfs_fs_info *fs_info = bbio->fs_info;
450 struct bio *bio = &bbio->bio;
451 int error = blk_status_to_errno(bio->bi_status);
452 struct folio_iter fi;
453 const u32 sectorsize = fs_info->sectorsize;
454
455 ASSERT(!bio_flagged(bio, BIO_CLONED));
456 bio_for_each_folio_all(fi, bio) {
457 struct folio *folio = fi.folio;
458 u64 start = folio_pos(folio) + fi.offset;
459 u32 len = fi.length;
460
461 /* Only order 0 (single page) folios are allowed for data. */
462 ASSERT(folio_order(folio) == 0);
463
464 /* Our read/write should always be sector aligned. */
465 if (!IS_ALIGNED(fi.offset, sectorsize))
466 btrfs_err(fs_info,
467 "partial page write in btrfs with offset %zu and length %zu",
468 fi.offset, fi.length);
469 else if (!IS_ALIGNED(fi.length, sectorsize))
470 btrfs_info(fs_info,
471 "incomplete page write with offset %zu and length %zu",
472 fi.offset, fi.length);
473
474 btrfs_finish_ordered_extent(bbio->ordered,
475 folio_page(folio, 0), start, len, !error);
476 if (error)
477 mapping_set_error(folio->mapping, error);
478 btrfs_folio_clear_writeback(fs_info, folio, start, len);
479 }
480
481 bio_put(bio);
482 }
483
484 /*
485 * Record previously processed extent range
486 *
487 * For endio_readpage_release_extent() to handle a full extent range, reducing
488 * the extent io operations.
489 */
490 struct processed_extent {
491 struct btrfs_inode *inode;
492 /* Start of the range in @inode */
493 u64 start;
494 /* End of the range in @inode */
495 u64 end;
496 bool uptodate;
497 };
498
499 /*
500 * Try to release processed extent range
501 *
502 * May not release the extent range right now if the current range is
503 * contiguous to processed extent.
504 *
505 * Will release processed extent when any of @inode, @uptodate, the range is
506 * no longer contiguous to the processed range.
507 *
508 * Passing @inode == NULL will force processed extent to be released.
509 */
510 static void endio_readpage_release_extent(struct processed_extent *processed,
511 struct btrfs_inode *inode, u64 start, u64 end,
512 bool uptodate)
513 {
514 struct extent_state *cached = NULL;
515 struct extent_io_tree *tree;
516
517 /* The first extent, initialize @processed */
518 if (!processed->inode)
519 goto update;
520
521 /*
522 * Contiguous to processed extent, just uptodate the end.
523 *
524 * Several things to notice:
525 *
526 * - bio can be merged as long as on-disk bytenr is contiguous
527 * This means we can have page belonging to other inodes, thus need to
528 * check if the inode still matches.
529 * - bvec can contain range beyond current page for multi-page bvec
530 * Thus we need to do processed->end + 1 >= start check
531 */
532 if (processed->inode == inode && processed->uptodate == uptodate &&
533 processed->end + 1 >= start && end >= processed->end) {
534 processed->end = end;
535 return;
536 }
537
538 tree = &processed->inode->io_tree;
539 /*
540 * Now we don't have range contiguous to the processed range, release
541 * the processed range now.
542 */
543 unlock_extent(tree, processed->start, processed->end, &cached);
544
545 update:
546 /* Update processed to current range */
547 processed->inode = inode;
548 processed->start = start;
549 processed->end = end;
550 processed->uptodate = uptodate;
551 }
552
553 static void begin_page_read(struct btrfs_fs_info *fs_info, struct page *page)
554 {
555 struct folio *folio = page_folio(page);
556
557 ASSERT(folio_test_locked(folio));
558 if (!btrfs_is_subpage(fs_info, folio->mapping))
559 return;
560
561 ASSERT(folio_test_private(folio));
562 btrfs_subpage_start_reader(fs_info, folio, page_offset(page), PAGE_SIZE);
563 }
564
565 /*
566 * After a data read IO is done, we need to:
567 *
568 * - clear the uptodate bits on error
569 * - set the uptodate bits if things worked
570 * - set the folio up to date if all extents in the tree are uptodate
571 * - clear the lock bit in the extent tree
572 * - unlock the folio if there are no other extents locked for it
573 *
574 * Scheduling is not allowed, so the extent state tree is expected
575 * to have one and only one object corresponding to this IO.
576 */
577 static void end_bbio_data_read(struct btrfs_bio *bbio)
578 {
579 struct btrfs_fs_info *fs_info = bbio->fs_info;
580 struct bio *bio = &bbio->bio;
581 struct processed_extent processed = { 0 };
582 struct folio_iter fi;
583 const u32 sectorsize = fs_info->sectorsize;
584
585 ASSERT(!bio_flagged(bio, BIO_CLONED));
586 bio_for_each_folio_all(fi, &bbio->bio) {
587 bool uptodate = !bio->bi_status;
588 struct folio *folio = fi.folio;
589 struct inode *inode = folio->mapping->host;
590 u64 start;
591 u64 end;
592 u32 len;
593
594 /* For now only order 0 folios are supported for data. */
595 ASSERT(folio_order(folio) == 0);
596 btrfs_debug(fs_info,
597 "%s: bi_sector=%llu, err=%d, mirror=%u",
598 __func__, bio->bi_iter.bi_sector, bio->bi_status,
599 bbio->mirror_num);
600
601 /*
602 * We always issue full-sector reads, but if some block in a
603 * folio fails to read, blk_update_request() will advance
604 * bv_offset and adjust bv_len to compensate. Print a warning
605 * for unaligned offsets, and an error if they don't add up to
606 * a full sector.
607 */
608 if (!IS_ALIGNED(fi.offset, sectorsize))
609 btrfs_err(fs_info,
610 "partial page read in btrfs with offset %zu and length %zu",
611 fi.offset, fi.length);
612 else if (!IS_ALIGNED(fi.offset + fi.length, sectorsize))
613 btrfs_info(fs_info,
614 "incomplete page read with offset %zu and length %zu",
615 fi.offset, fi.length);
616
617 start = folio_pos(folio) + fi.offset;
618 end = start + fi.length - 1;
619 len = fi.length;
620
621 if (likely(uptodate)) {
622 loff_t i_size = i_size_read(inode);
623 pgoff_t end_index = i_size >> folio_shift(folio);
624
625 /*
626 * Zero out the remaining part if this range straddles
627 * i_size.
628 *
629 * Here we should only zero the range inside the folio,
630 * not touch anything else.
631 *
632 * NOTE: i_size is exclusive while end is inclusive.
633 */
634 if (folio_index(folio) == end_index && i_size <= end) {
635 u32 zero_start = max(offset_in_folio(folio, i_size),
636 offset_in_folio(folio, start));
637 u32 zero_len = offset_in_folio(folio, end) + 1 -
638 zero_start;
639
640 folio_zero_range(folio, zero_start, zero_len);
641 }
642 }
643
644 /* Update page status and unlock. */
645 end_page_read(folio_page(folio, 0), uptodate, start, len);
646 endio_readpage_release_extent(&processed, BTRFS_I(inode),
647 start, end, uptodate);
648 }
649 /* Release the last extent */
650 endio_readpage_release_extent(&processed, NULL, 0, 0, false);
651 bio_put(bio);
652 }
653
654 /*
655 * Populate every free slot in a provided array with folios using GFP_NOFS.
656 *
657 * @nr_folios: number of folios to allocate
658 * @folio_array: the array to fill with folios; any existing non-NULL entries in
659 * the array will be skipped
660 *
661 * Return: 0 if all folios were able to be allocated;
662 * -ENOMEM otherwise, the partially allocated folios would be freed and
663 * the array slots zeroed
664 */
665 int btrfs_alloc_folio_array(unsigned int nr_folios, struct folio **folio_array)
666 {
667 for (int i = 0; i < nr_folios; i++) {
668 if (folio_array[i])
669 continue;
670 folio_array[i] = folio_alloc(GFP_NOFS, 0);
671 if (!folio_array[i])
672 goto error;
673 }
674 return 0;
675 error:
676 for (int i = 0; i < nr_folios; i++) {
677 if (folio_array[i])
678 folio_put(folio_array[i]);
679 }
680 return -ENOMEM;
681 }
682
683 /*
684 * Populate every free slot in a provided array with pages, using GFP_NOFS.
685 *
686 * @nr_pages: number of pages to allocate
687 * @page_array: the array to fill with pages; any existing non-null entries in
688 * the array will be skipped
689 * @nofail: whether using __GFP_NOFAIL flag
690 *
691 * Return: 0 if all pages were able to be allocated;
692 * -ENOMEM otherwise, the partially allocated pages would be freed and
693 * the array slots zeroed
694 */
695 int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array,
696 bool nofail)
697 {
698 const gfp_t gfp = nofail ? (GFP_NOFS | __GFP_NOFAIL) : GFP_NOFS;
699 unsigned int allocated;
700
701 for (allocated = 0; allocated < nr_pages;) {
702 unsigned int last = allocated;
703
704 allocated = alloc_pages_bulk_array(gfp, nr_pages, page_array);
705 if (unlikely(allocated == last)) {
706 /* No progress, fail and do cleanup. */
707 for (int i = 0; i < allocated; i++) {
708 __free_page(page_array[i]);
709 page_array[i] = NULL;
710 }
711 return -ENOMEM;
712 }
713 }
714 return 0;
715 }
716
717 /*
718 * Populate needed folios for the extent buffer.
719 *
720 * For now, the folios populated are always in order 0 (aka, single page).
721 */
722 static int alloc_eb_folio_array(struct extent_buffer *eb, bool nofail)
723 {
724 struct page *page_array[INLINE_EXTENT_BUFFER_PAGES] = { 0 };
725 int num_pages = num_extent_pages(eb);
726 int ret;
727
728 ret = btrfs_alloc_page_array(num_pages, page_array, nofail);
729 if (ret < 0)
730 return ret;
731
732 for (int i = 0; i < num_pages; i++)
733 eb->folios[i] = page_folio(page_array[i]);
734 eb->folio_size = PAGE_SIZE;
735 eb->folio_shift = PAGE_SHIFT;
736 return 0;
737 }
738
739 static bool btrfs_bio_is_contig(struct btrfs_bio_ctrl *bio_ctrl,
740 struct page *page, u64 disk_bytenr,
741 unsigned int pg_offset)
742 {
743 struct bio *bio = &bio_ctrl->bbio->bio;
744 struct bio_vec *bvec = bio_last_bvec_all(bio);
745 const sector_t sector = disk_bytenr >> SECTOR_SHIFT;
746
747 if (bio_ctrl->compress_type != BTRFS_COMPRESS_NONE) {
748 /*
749 * For compression, all IO should have its logical bytenr set
750 * to the starting bytenr of the compressed extent.
751 */
752 return bio->bi_iter.bi_sector == sector;
753 }
754
755 /*
756 * The contig check requires the following conditions to be met:
757 *
758 * 1) The pages are belonging to the same inode
759 * This is implied by the call chain.
760 *
761 * 2) The range has adjacent logical bytenr
762 *
763 * 3) The range has adjacent file offset
764 * This is required for the usage of btrfs_bio->file_offset.
765 */
766 return bio_end_sector(bio) == sector &&
767 page_offset(bvec->bv_page) + bvec->bv_offset + bvec->bv_len ==
768 page_offset(page) + pg_offset;
769 }
770
771 static void alloc_new_bio(struct btrfs_inode *inode,
772 struct btrfs_bio_ctrl *bio_ctrl,
773 u64 disk_bytenr, u64 file_offset)
774 {
775 struct btrfs_fs_info *fs_info = inode->root->fs_info;
776 struct btrfs_bio *bbio;
777
778 bbio = btrfs_bio_alloc(BIO_MAX_VECS, bio_ctrl->opf, fs_info,
779 bio_ctrl->end_io_func, NULL);
780 bbio->bio.bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
781 bbio->inode = inode;
782 bbio->file_offset = file_offset;
783 bio_ctrl->bbio = bbio;
784 bio_ctrl->len_to_oe_boundary = U32_MAX;
785
786 /* Limit data write bios to the ordered boundary. */
787 if (bio_ctrl->wbc) {
788 struct btrfs_ordered_extent *ordered;
789
790 ordered = btrfs_lookup_ordered_extent(inode, file_offset);
791 if (ordered) {
792 bio_ctrl->len_to_oe_boundary = min_t(u32, U32_MAX,
793 ordered->file_offset +
794 ordered->disk_num_bytes - file_offset);
795 bbio->ordered = ordered;
796 }
797
798 /*
799 * Pick the last added device to support cgroup writeback. For
800 * multi-device file systems this means blk-cgroup policies have
801 * to always be set on the last added/replaced device.
802 * This is a bit odd but has been like that for a long time.
803 */
804 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
805 wbc_init_bio(bio_ctrl->wbc, &bbio->bio);
806 }
807 }
808
809 /*
810 * @disk_bytenr: logical bytenr where the write will be
811 * @page: page to add to the bio
812 * @size: portion of page that we want to write to
813 * @pg_offset: offset of the new bio or to check whether we are adding
814 * a contiguous page to the previous one
815 *
816 * The will either add the page into the existing @bio_ctrl->bbio, or allocate a
817 * new one in @bio_ctrl->bbio.
818 * The mirror number for this IO should already be initizlied in
819 * @bio_ctrl->mirror_num.
820 */
821 static void submit_extent_page(struct btrfs_bio_ctrl *bio_ctrl,
822 u64 disk_bytenr, struct page *page,
823 size_t size, unsigned long pg_offset)
824 {
825 struct btrfs_inode *inode = page_to_inode(page);
826
827 ASSERT(pg_offset + size <= PAGE_SIZE);
828 ASSERT(bio_ctrl->end_io_func);
829
830 if (bio_ctrl->bbio &&
831 !btrfs_bio_is_contig(bio_ctrl, page, disk_bytenr, pg_offset))
832 submit_one_bio(bio_ctrl);
833
834 do {
835 u32 len = size;
836
837 /* Allocate new bio if needed */
838 if (!bio_ctrl->bbio) {
839 alloc_new_bio(inode, bio_ctrl, disk_bytenr,
840 page_offset(page) + pg_offset);
841 }
842
843 /* Cap to the current ordered extent boundary if there is one. */
844 if (len > bio_ctrl->len_to_oe_boundary) {
845 ASSERT(bio_ctrl->compress_type == BTRFS_COMPRESS_NONE);
846 ASSERT(is_data_inode(inode));
847 len = bio_ctrl->len_to_oe_boundary;
848 }
849
850 if (bio_add_page(&bio_ctrl->bbio->bio, page, len, pg_offset) != len) {
851 /* bio full: move on to a new one */
852 submit_one_bio(bio_ctrl);
853 continue;
854 }
855
856 if (bio_ctrl->wbc)
857 wbc_account_cgroup_owner(bio_ctrl->wbc, page, len);
858
859 size -= len;
860 pg_offset += len;
861 disk_bytenr += len;
862
863 /*
864 * len_to_oe_boundary defaults to U32_MAX, which isn't page or
865 * sector aligned. alloc_new_bio() then sets it to the end of
866 * our ordered extent for writes into zoned devices.
867 *
868 * When len_to_oe_boundary is tracking an ordered extent, we
869 * trust the ordered extent code to align things properly, and
870 * the check above to cap our write to the ordered extent
871 * boundary is correct.
872 *
873 * When len_to_oe_boundary is U32_MAX, the cap above would
874 * result in a 4095 byte IO for the last page right before
875 * we hit the bio limit of UINT_MAX. bio_add_page() has all
876 * the checks required to make sure we don't overflow the bio,
877 * and we should just ignore len_to_oe_boundary completely
878 * unless we're using it to track an ordered extent.
879 *
880 * It's pretty hard to make a bio sized U32_MAX, but it can
881 * happen when the page cache is able to feed us contiguous
882 * pages for large extents.
883 */
884 if (bio_ctrl->len_to_oe_boundary != U32_MAX)
885 bio_ctrl->len_to_oe_boundary -= len;
886
887 /* Ordered extent boundary: move on to a new bio. */
888 if (bio_ctrl->len_to_oe_boundary == 0)
889 submit_one_bio(bio_ctrl);
890 } while (size);
891 }
892
893 static int attach_extent_buffer_folio(struct extent_buffer *eb,
894 struct folio *folio,
895 struct btrfs_subpage *prealloc)
896 {
897 struct btrfs_fs_info *fs_info = eb->fs_info;
898 int ret = 0;
899
900 /*
901 * If the page is mapped to btree inode, we should hold the private
902 * lock to prevent race.
903 * For cloned or dummy extent buffers, their pages are not mapped and
904 * will not race with any other ebs.
905 */
906 if (folio->mapping)
907 lockdep_assert_held(&folio->mapping->i_private_lock);
908
909 if (fs_info->nodesize >= PAGE_SIZE) {
910 if (!folio_test_private(folio))
911 folio_attach_private(folio, eb);
912 else
913 WARN_ON(folio_get_private(folio) != eb);
914 return 0;
915 }
916
917 /* Already mapped, just free prealloc */
918 if (folio_test_private(folio)) {
919 btrfs_free_subpage(prealloc);
920 return 0;
921 }
922
923 if (prealloc)
924 /* Has preallocated memory for subpage */
925 folio_attach_private(folio, prealloc);
926 else
927 /* Do new allocation to attach subpage */
928 ret = btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_METADATA);
929 return ret;
930 }
931
932 int set_page_extent_mapped(struct page *page)
933 {
934 return set_folio_extent_mapped(page_folio(page));
935 }
936
937 int set_folio_extent_mapped(struct folio *folio)
938 {
939 struct btrfs_fs_info *fs_info;
940
941 ASSERT(folio->mapping);
942
943 if (folio_test_private(folio))
944 return 0;
945
946 fs_info = folio_to_fs_info(folio);
947
948 if (btrfs_is_subpage(fs_info, folio->mapping))
949 return btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_DATA);
950
951 folio_attach_private(folio, (void *)EXTENT_FOLIO_PRIVATE);
952 return 0;
953 }
954
955 void clear_page_extent_mapped(struct page *page)
956 {
957 struct folio *folio = page_folio(page);
958 struct btrfs_fs_info *fs_info;
959
960 ASSERT(page->mapping);
961
962 if (!folio_test_private(folio))
963 return;
964
965 fs_info = page_to_fs_info(page);
966 if (btrfs_is_subpage(fs_info, page->mapping))
967 return btrfs_detach_subpage(fs_info, folio);
968
969 folio_detach_private(folio);
970 }
971
972 static struct extent_map *__get_extent_map(struct inode *inode, struct page *page,
973 u64 start, u64 len, struct extent_map **em_cached)
974 {
975 struct extent_map *em;
976
977 ASSERT(em_cached);
978
979 if (*em_cached) {
980 em = *em_cached;
981 if (extent_map_in_tree(em) && start >= em->start &&
982 start < extent_map_end(em)) {
983 refcount_inc(&em->refs);
984 return em;
985 }
986
987 free_extent_map(em);
988 *em_cached = NULL;
989 }
990
991 em = btrfs_get_extent(BTRFS_I(inode), page, start, len);
992 if (!IS_ERR(em)) {
993 BUG_ON(*em_cached);
994 refcount_inc(&em->refs);
995 *em_cached = em;
996 }
997 return em;
998 }
999 /*
1000 * basic readpage implementation. Locked extent state structs are inserted
1001 * into the tree that are removed when the IO is done (by the end_io
1002 * handlers)
1003 * XXX JDM: This needs looking at to ensure proper page locking
1004 * return 0 on success, otherwise return error
1005 */
1006 static int btrfs_do_readpage(struct page *page, struct extent_map **em_cached,
1007 struct btrfs_bio_ctrl *bio_ctrl, u64 *prev_em_start)
1008 {
1009 struct inode *inode = page->mapping->host;
1010 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1011 u64 start = page_offset(page);
1012 const u64 end = start + PAGE_SIZE - 1;
1013 u64 cur = start;
1014 u64 extent_offset;
1015 u64 last_byte = i_size_read(inode);
1016 u64 block_start;
1017 struct extent_map *em;
1018 int ret = 0;
1019 size_t pg_offset = 0;
1020 size_t iosize;
1021 size_t blocksize = fs_info->sectorsize;
1022 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1023
1024 ret = set_page_extent_mapped(page);
1025 if (ret < 0) {
1026 unlock_extent(tree, start, end, NULL);
1027 unlock_page(page);
1028 return ret;
1029 }
1030
1031 if (page->index == last_byte >> PAGE_SHIFT) {
1032 size_t zero_offset = offset_in_page(last_byte);
1033
1034 if (zero_offset) {
1035 iosize = PAGE_SIZE - zero_offset;
1036 memzero_page(page, zero_offset, iosize);
1037 }
1038 }
1039 bio_ctrl->end_io_func = end_bbio_data_read;
1040 begin_page_read(fs_info, page);
1041 while (cur <= end) {
1042 enum btrfs_compression_type compress_type = BTRFS_COMPRESS_NONE;
1043 bool force_bio_submit = false;
1044 u64 disk_bytenr;
1045
1046 ASSERT(IS_ALIGNED(cur, fs_info->sectorsize));
1047 if (cur >= last_byte) {
1048 iosize = PAGE_SIZE - pg_offset;
1049 memzero_page(page, pg_offset, iosize);
1050 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1051 end_page_read(page, true, cur, iosize);
1052 break;
1053 }
1054 em = __get_extent_map(inode, page, cur, end - cur + 1, em_cached);
1055 if (IS_ERR(em)) {
1056 unlock_extent(tree, cur, end, NULL);
1057 end_page_read(page, false, cur, end + 1 - cur);
1058 return PTR_ERR(em);
1059 }
1060 extent_offset = cur - em->start;
1061 BUG_ON(extent_map_end(em) <= cur);
1062 BUG_ON(end < cur);
1063
1064 compress_type = extent_map_compression(em);
1065
1066 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1067 iosize = ALIGN(iosize, blocksize);
1068 if (compress_type != BTRFS_COMPRESS_NONE)
1069 disk_bytenr = em->disk_bytenr;
1070 else
1071 disk_bytenr = extent_map_block_start(em) + extent_offset;
1072 block_start = extent_map_block_start(em);
1073 if (em->flags & EXTENT_FLAG_PREALLOC)
1074 block_start = EXTENT_MAP_HOLE;
1075
1076 /*
1077 * If we have a file range that points to a compressed extent
1078 * and it's followed by a consecutive file range that points
1079 * to the same compressed extent (possibly with a different
1080 * offset and/or length, so it either points to the whole extent
1081 * or only part of it), we must make sure we do not submit a
1082 * single bio to populate the pages for the 2 ranges because
1083 * this makes the compressed extent read zero out the pages
1084 * belonging to the 2nd range. Imagine the following scenario:
1085 *
1086 * File layout
1087 * [0 - 8K] [8K - 24K]
1088 * | |
1089 * | |
1090 * points to extent X, points to extent X,
1091 * offset 4K, length of 8K offset 0, length 16K
1092 *
1093 * [extent X, compressed length = 4K uncompressed length = 16K]
1094 *
1095 * If the bio to read the compressed extent covers both ranges,
1096 * it will decompress extent X into the pages belonging to the
1097 * first range and then it will stop, zeroing out the remaining
1098 * pages that belong to the other range that points to extent X.
1099 * So here we make sure we submit 2 bios, one for the first
1100 * range and another one for the third range. Both will target
1101 * the same physical extent from disk, but we can't currently
1102 * make the compressed bio endio callback populate the pages
1103 * for both ranges because each compressed bio is tightly
1104 * coupled with a single extent map, and each range can have
1105 * an extent map with a different offset value relative to the
1106 * uncompressed data of our extent and different lengths. This
1107 * is a corner case so we prioritize correctness over
1108 * non-optimal behavior (submitting 2 bios for the same extent).
1109 */
1110 if (compress_type != BTRFS_COMPRESS_NONE &&
1111 prev_em_start && *prev_em_start != (u64)-1 &&
1112 *prev_em_start != em->start)
1113 force_bio_submit = true;
1114
1115 if (prev_em_start)
1116 *prev_em_start = em->start;
1117
1118 free_extent_map(em);
1119 em = NULL;
1120
1121 /* we've found a hole, just zero and go on */
1122 if (block_start == EXTENT_MAP_HOLE) {
1123 memzero_page(page, pg_offset, iosize);
1124
1125 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1126 end_page_read(page, true, cur, iosize);
1127 cur = cur + iosize;
1128 pg_offset += iosize;
1129 continue;
1130 }
1131 /* the get_extent function already copied into the page */
1132 if (block_start == EXTENT_MAP_INLINE) {
1133 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1134 end_page_read(page, true, cur, iosize);
1135 cur = cur + iosize;
1136 pg_offset += iosize;
1137 continue;
1138 }
1139
1140 if (bio_ctrl->compress_type != compress_type) {
1141 submit_one_bio(bio_ctrl);
1142 bio_ctrl->compress_type = compress_type;
1143 }
1144
1145 if (force_bio_submit)
1146 submit_one_bio(bio_ctrl);
1147 submit_extent_page(bio_ctrl, disk_bytenr, page, iosize,
1148 pg_offset);
1149 cur = cur + iosize;
1150 pg_offset += iosize;
1151 }
1152
1153 return 0;
1154 }
1155
1156 int btrfs_read_folio(struct file *file, struct folio *folio)
1157 {
1158 struct page *page = &folio->page;
1159 struct btrfs_inode *inode = page_to_inode(page);
1160 u64 start = page_offset(page);
1161 u64 end = start + PAGE_SIZE - 1;
1162 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ };
1163 struct extent_map *em_cached = NULL;
1164 int ret;
1165
1166 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
1167
1168 ret = btrfs_do_readpage(page, &em_cached, &bio_ctrl, NULL);
1169 free_extent_map(em_cached);
1170
1171 /*
1172 * If btrfs_do_readpage() failed we will want to submit the assembled
1173 * bio to do the cleanup.
1174 */
1175 submit_one_bio(&bio_ctrl);
1176 return ret;
1177 }
1178
1179 static inline void contiguous_readpages(struct page *pages[], int nr_pages,
1180 u64 start, u64 end,
1181 struct extent_map **em_cached,
1182 struct btrfs_bio_ctrl *bio_ctrl,
1183 u64 *prev_em_start)
1184 {
1185 struct btrfs_inode *inode = page_to_inode(pages[0]);
1186 int index;
1187
1188 ASSERT(em_cached);
1189
1190 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
1191
1192 for (index = 0; index < nr_pages; index++) {
1193 btrfs_do_readpage(pages[index], em_cached, bio_ctrl,
1194 prev_em_start);
1195 put_page(pages[index]);
1196 }
1197 }
1198
1199 /*
1200 * helper for __extent_writepage, doing all of the delayed allocation setup.
1201 *
1202 * This returns 1 if btrfs_run_delalloc_range function did all the work required
1203 * to write the page (copy into inline extent). In this case the IO has
1204 * been started and the page is already unlocked.
1205 *
1206 * This returns 0 if all went well (page still locked)
1207 * This returns < 0 if there were errors (page still locked)
1208 */
1209 static noinline_for_stack int writepage_delalloc(struct btrfs_inode *inode,
1210 struct page *page, struct writeback_control *wbc)
1211 {
1212 struct btrfs_fs_info *fs_info = inode_to_fs_info(&inode->vfs_inode);
1213 struct folio *folio = page_folio(page);
1214 const bool is_subpage = btrfs_is_subpage(fs_info, page->mapping);
1215 const u64 page_start = page_offset(page);
1216 const u64 page_end = page_start + PAGE_SIZE - 1;
1217 /*
1218 * Save the last found delalloc end. As the delalloc end can go beyond
1219 * page boundary, thus we cannot rely on subpage bitmap to locate the
1220 * last delalloc end.
1221 */
1222 u64 last_delalloc_end = 0;
1223 u64 delalloc_start = page_start;
1224 u64 delalloc_end = page_end;
1225 u64 delalloc_to_write = 0;
1226 int ret = 0;
1227
1228 /* Lock all (subpage) delalloc ranges inside the page first. */
1229 while (delalloc_start < page_end) {
1230 delalloc_end = page_end;
1231 if (!find_lock_delalloc_range(&inode->vfs_inode, page,
1232 &delalloc_start, &delalloc_end)) {
1233 delalloc_start = delalloc_end + 1;
1234 continue;
1235 }
1236 btrfs_folio_set_writer_lock(fs_info, folio, delalloc_start,
1237 min(delalloc_end, page_end) + 1 -
1238 delalloc_start);
1239 last_delalloc_end = delalloc_end;
1240 delalloc_start = delalloc_end + 1;
1241 }
1242 delalloc_start = page_start;
1243
1244 if (!last_delalloc_end)
1245 goto out;
1246
1247 /* Run the delalloc ranges for the above locked ranges. */
1248 while (delalloc_start < page_end) {
1249 u64 found_start;
1250 u32 found_len;
1251 bool found;
1252
1253 if (!is_subpage) {
1254 /*
1255 * For non-subpage case, the found delalloc range must
1256 * cover this page and there must be only one locked
1257 * delalloc range.
1258 */
1259 found_start = page_start;
1260 found_len = last_delalloc_end + 1 - found_start;
1261 found = true;
1262 } else {
1263 found = btrfs_subpage_find_writer_locked(fs_info, folio,
1264 delalloc_start, &found_start, &found_len);
1265 }
1266 if (!found)
1267 break;
1268 /*
1269 * The subpage range covers the last sector, the delalloc range may
1270 * end beyond the page boundary, use the saved delalloc_end
1271 * instead.
1272 */
1273 if (found_start + found_len >= page_end)
1274 found_len = last_delalloc_end + 1 - found_start;
1275
1276 if (ret >= 0) {
1277 /* No errors hit so far, run the current delalloc range. */
1278 ret = btrfs_run_delalloc_range(inode, page, found_start,
1279 found_start + found_len - 1,
1280 wbc);
1281 } else {
1282 /*
1283 * We've hit an error during previous delalloc range,
1284 * have to cleanup the remaining locked ranges.
1285 */
1286 unlock_extent(&inode->io_tree, found_start,
1287 found_start + found_len - 1, NULL);
1288 __unlock_for_delalloc(&inode->vfs_inode, page, found_start,
1289 found_start + found_len - 1);
1290 }
1291
1292 /*
1293 * We can hit btrfs_run_delalloc_range() with >0 return value.
1294 *
1295 * This happens when either the IO is already done and page
1296 * unlocked (inline) or the IO submission and page unlock would
1297 * be handled as async (compression).
1298 *
1299 * Inline is only possible for regular sectorsize for now.
1300 *
1301 * Compression is possible for both subpage and regular cases,
1302 * but even for subpage compression only happens for page aligned
1303 * range, thus the found delalloc range must go beyond current
1304 * page.
1305 */
1306 if (ret > 0)
1307 ASSERT(!is_subpage || found_start + found_len >= page_end);
1308
1309 /*
1310 * Above btrfs_run_delalloc_range() may have unlocked the page,
1311 * thus for the last range, we cannot touch the page anymore.
1312 */
1313 if (found_start + found_len >= last_delalloc_end + 1)
1314 break;
1315
1316 delalloc_start = found_start + found_len;
1317 }
1318 if (ret < 0)
1319 return ret;
1320 out:
1321 if (last_delalloc_end)
1322 delalloc_end = last_delalloc_end;
1323 else
1324 delalloc_end = page_end;
1325 /*
1326 * delalloc_end is already one less than the total length, so
1327 * we don't subtract one from PAGE_SIZE
1328 */
1329 delalloc_to_write +=
1330 DIV_ROUND_UP(delalloc_end + 1 - page_start, PAGE_SIZE);
1331
1332 /*
1333 * If btrfs_run_dealloc_range() already started I/O and unlocked
1334 * the pages, we just need to account for them here.
1335 */
1336 if (ret == 1) {
1337 wbc->nr_to_write -= delalloc_to_write;
1338 return 1;
1339 }
1340
1341 if (wbc->nr_to_write < delalloc_to_write) {
1342 int thresh = 8192;
1343
1344 if (delalloc_to_write < thresh * 2)
1345 thresh = delalloc_to_write;
1346 wbc->nr_to_write = min_t(u64, delalloc_to_write,
1347 thresh);
1348 }
1349
1350 return 0;
1351 }
1352
1353 /*
1354 * Find the first byte we need to write.
1355 *
1356 * For subpage, one page can contain several sectors, and
1357 * __extent_writepage_io() will just grab all extent maps in the page
1358 * range and try to submit all non-inline/non-compressed extents.
1359 *
1360 * This is a big problem for subpage, we shouldn't re-submit already written
1361 * data at all.
1362 * This function will lookup subpage dirty bit to find which range we really
1363 * need to submit.
1364 *
1365 * Return the next dirty range in [@start, @end).
1366 * If no dirty range is found, @start will be page_offset(page) + PAGE_SIZE.
1367 */
1368 static void find_next_dirty_byte(const struct btrfs_fs_info *fs_info,
1369 struct page *page, u64 *start, u64 *end)
1370 {
1371 struct folio *folio = page_folio(page);
1372 struct btrfs_subpage *subpage = folio_get_private(folio);
1373 struct btrfs_subpage_info *spi = fs_info->subpage_info;
1374 u64 orig_start = *start;
1375 /* Declare as unsigned long so we can use bitmap ops */
1376 unsigned long flags;
1377 int range_start_bit;
1378 int range_end_bit;
1379
1380 /*
1381 * For regular sector size == page size case, since one page only
1382 * contains one sector, we return the page offset directly.
1383 */
1384 if (!btrfs_is_subpage(fs_info, page->mapping)) {
1385 *start = page_offset(page);
1386 *end = page_offset(page) + PAGE_SIZE;
1387 return;
1388 }
1389
1390 range_start_bit = spi->dirty_offset +
1391 (offset_in_page(orig_start) >> fs_info->sectorsize_bits);
1392
1393 /* We should have the page locked, but just in case */
1394 spin_lock_irqsave(&subpage->lock, flags);
1395 bitmap_next_set_region(subpage->bitmaps, &range_start_bit, &range_end_bit,
1396 spi->dirty_offset + spi->bitmap_nr_bits);
1397 spin_unlock_irqrestore(&subpage->lock, flags);
1398
1399 range_start_bit -= spi->dirty_offset;
1400 range_end_bit -= spi->dirty_offset;
1401
1402 *start = page_offset(page) + range_start_bit * fs_info->sectorsize;
1403 *end = page_offset(page) + range_end_bit * fs_info->sectorsize;
1404 }
1405
1406 /*
1407 * helper for __extent_writepage. This calls the writepage start hooks,
1408 * and does the loop to map the page into extents and bios.
1409 *
1410 * We return 1 if the IO is started and the page is unlocked,
1411 * 0 if all went well (page still locked)
1412 * < 0 if there were errors (page still locked)
1413 */
1414 static noinline_for_stack int __extent_writepage_io(struct btrfs_inode *inode,
1415 struct page *page, u64 start, u32 len,
1416 struct btrfs_bio_ctrl *bio_ctrl,
1417 loff_t i_size,
1418 int *nr_ret)
1419 {
1420 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1421 u64 cur = start;
1422 u64 end = start + len - 1;
1423 u64 extent_offset;
1424 u64 block_start;
1425 struct extent_map *em;
1426 int ret = 0;
1427 int nr = 0;
1428
1429 ASSERT(start >= page_offset(page) &&
1430 start + len <= page_offset(page) + PAGE_SIZE);
1431
1432 ret = btrfs_writepage_cow_fixup(page);
1433 if (ret) {
1434 /* Fixup worker will requeue */
1435 redirty_page_for_writepage(bio_ctrl->wbc, page);
1436 unlock_page(page);
1437 return 1;
1438 }
1439
1440 bio_ctrl->end_io_func = end_bbio_data_write;
1441 while (cur <= end) {
1442 u32 len = end - cur + 1;
1443 u64 disk_bytenr;
1444 u64 em_end;
1445 u64 dirty_range_start = cur;
1446 u64 dirty_range_end;
1447 u32 iosize;
1448
1449 if (cur >= i_size) {
1450 btrfs_mark_ordered_io_finished(inode, page, cur, len,
1451 true);
1452 /*
1453 * This range is beyond i_size, thus we don't need to
1454 * bother writing back.
1455 * But we still need to clear the dirty subpage bit, or
1456 * the next time the page gets dirtied, we will try to
1457 * writeback the sectors with subpage dirty bits,
1458 * causing writeback without ordered extent.
1459 */
1460 btrfs_folio_clear_dirty(fs_info, page_folio(page), cur, len);
1461 break;
1462 }
1463
1464 find_next_dirty_byte(fs_info, page, &dirty_range_start,
1465 &dirty_range_end);
1466 if (cur < dirty_range_start) {
1467 cur = dirty_range_start;
1468 continue;
1469 }
1470
1471 em = btrfs_get_extent(inode, NULL, cur, len);
1472 if (IS_ERR(em)) {
1473 ret = PTR_ERR_OR_ZERO(em);
1474 goto out_error;
1475 }
1476
1477 extent_offset = cur - em->start;
1478 em_end = extent_map_end(em);
1479 ASSERT(cur <= em_end);
1480 ASSERT(cur < end);
1481 ASSERT(IS_ALIGNED(em->start, fs_info->sectorsize));
1482 ASSERT(IS_ALIGNED(em->len, fs_info->sectorsize));
1483
1484 block_start = extent_map_block_start(em);
1485 disk_bytenr = extent_map_block_start(em) + extent_offset;
1486
1487 ASSERT(!extent_map_is_compressed(em));
1488 ASSERT(block_start != EXTENT_MAP_HOLE);
1489 ASSERT(block_start != EXTENT_MAP_INLINE);
1490
1491 /*
1492 * Note that em_end from extent_map_end() and dirty_range_end from
1493 * find_next_dirty_byte() are all exclusive
1494 */
1495 iosize = min(min(em_end, end + 1), dirty_range_end) - cur;
1496 free_extent_map(em);
1497 em = NULL;
1498
1499 btrfs_set_range_writeback(inode, cur, cur + iosize - 1);
1500 if (!PageWriteback(page)) {
1501 btrfs_err(inode->root->fs_info,
1502 "page %lu not writeback, cur %llu end %llu",
1503 page->index, cur, end);
1504 }
1505
1506 /*
1507 * Although the PageDirty bit is cleared before entering this
1508 * function, subpage dirty bit is not cleared.
1509 * So clear subpage dirty bit here so next time we won't submit
1510 * page for range already written to disk.
1511 */
1512 btrfs_folio_clear_dirty(fs_info, page_folio(page), cur, iosize);
1513
1514 submit_extent_page(bio_ctrl, disk_bytenr, page, iosize,
1515 cur - page_offset(page));
1516 cur += iosize;
1517 nr++;
1518 }
1519
1520 btrfs_folio_assert_not_dirty(fs_info, page_folio(page), start, len);
1521 *nr_ret = nr;
1522 return 0;
1523
1524 out_error:
1525 /*
1526 * If we finish without problem, we should not only clear page dirty,
1527 * but also empty subpage dirty bits
1528 */
1529 *nr_ret = nr;
1530 return ret;
1531 }
1532
1533 /*
1534 * the writepage semantics are similar to regular writepage. extent
1535 * records are inserted to lock ranges in the tree, and as dirty areas
1536 * are found, they are marked writeback. Then the lock bits are removed
1537 * and the end_io handler clears the writeback ranges
1538 *
1539 * Return 0 if everything goes well.
1540 * Return <0 for error.
1541 */
1542 static int __extent_writepage(struct page *page, struct btrfs_bio_ctrl *bio_ctrl)
1543 {
1544 struct folio *folio = page_folio(page);
1545 struct inode *inode = page->mapping->host;
1546 const u64 page_start = page_offset(page);
1547 int ret;
1548 int nr = 0;
1549 size_t pg_offset;
1550 loff_t i_size = i_size_read(inode);
1551 unsigned long end_index = i_size >> PAGE_SHIFT;
1552
1553 trace___extent_writepage(page, inode, bio_ctrl->wbc);
1554
1555 WARN_ON(!PageLocked(page));
1556
1557 pg_offset = offset_in_page(i_size);
1558 if (page->index > end_index ||
1559 (page->index == end_index && !pg_offset)) {
1560 folio_invalidate(folio, 0, folio_size(folio));
1561 folio_unlock(folio);
1562 return 0;
1563 }
1564
1565 if (page->index == end_index)
1566 memzero_page(page, pg_offset, PAGE_SIZE - pg_offset);
1567
1568 ret = set_page_extent_mapped(page);
1569 if (ret < 0)
1570 goto done;
1571
1572 ret = writepage_delalloc(BTRFS_I(inode), page, bio_ctrl->wbc);
1573 if (ret == 1)
1574 return 0;
1575 if (ret)
1576 goto done;
1577
1578 ret = __extent_writepage_io(BTRFS_I(inode), page, page_offset(page),
1579 PAGE_SIZE, bio_ctrl, i_size, &nr);
1580 if (ret == 1)
1581 return 0;
1582
1583 bio_ctrl->wbc->nr_to_write--;
1584
1585 done:
1586 if (nr == 0) {
1587 /* make sure the mapping tag for page dirty gets cleared */
1588 set_page_writeback(page);
1589 end_page_writeback(page);
1590 }
1591 if (ret) {
1592 btrfs_mark_ordered_io_finished(BTRFS_I(inode), page, page_start,
1593 PAGE_SIZE, !ret);
1594 mapping_set_error(page->mapping, ret);
1595 }
1596
1597 btrfs_folio_end_all_writers(inode_to_fs_info(inode), folio);
1598 ASSERT(ret <= 0);
1599 return ret;
1600 }
1601
1602 void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
1603 {
1604 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
1605 TASK_UNINTERRUPTIBLE);
1606 }
1607
1608 /*
1609 * Lock extent buffer status and pages for writeback.
1610 *
1611 * Return %false if the extent buffer doesn't need to be submitted (e.g. the
1612 * extent buffer is not dirty)
1613 * Return %true is the extent buffer is submitted to bio.
1614 */
1615 static noinline_for_stack bool lock_extent_buffer_for_io(struct extent_buffer *eb,
1616 struct writeback_control *wbc)
1617 {
1618 struct btrfs_fs_info *fs_info = eb->fs_info;
1619 bool ret = false;
1620
1621 btrfs_tree_lock(eb);
1622 while (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
1623 btrfs_tree_unlock(eb);
1624 if (wbc->sync_mode != WB_SYNC_ALL)
1625 return false;
1626 wait_on_extent_buffer_writeback(eb);
1627 btrfs_tree_lock(eb);
1628 }
1629
1630 /*
1631 * We need to do this to prevent races in people who check if the eb is
1632 * under IO since we can end up having no IO bits set for a short period
1633 * of time.
1634 */
1635 spin_lock(&eb->refs_lock);
1636 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
1637 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1638 spin_unlock(&eb->refs_lock);
1639 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
1640 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
1641 -eb->len,
1642 fs_info->dirty_metadata_batch);
1643 ret = true;
1644 } else {
1645 spin_unlock(&eb->refs_lock);
1646 }
1647 btrfs_tree_unlock(eb);
1648 return ret;
1649 }
1650
1651 static void set_btree_ioerr(struct extent_buffer *eb)
1652 {
1653 struct btrfs_fs_info *fs_info = eb->fs_info;
1654
1655 set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1656
1657 /*
1658 * A read may stumble upon this buffer later, make sure that it gets an
1659 * error and knows there was an error.
1660 */
1661 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
1662
1663 /*
1664 * We need to set the mapping with the io error as well because a write
1665 * error will flip the file system readonly, and then syncfs() will
1666 * return a 0 because we are readonly if we don't modify the err seq for
1667 * the superblock.
1668 */
1669 mapping_set_error(eb->fs_info->btree_inode->i_mapping, -EIO);
1670
1671 /*
1672 * If writeback for a btree extent that doesn't belong to a log tree
1673 * failed, increment the counter transaction->eb_write_errors.
1674 * We do this because while the transaction is running and before it's
1675 * committing (when we call filemap_fdata[write|wait]_range against
1676 * the btree inode), we might have
1677 * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
1678 * returns an error or an error happens during writeback, when we're
1679 * committing the transaction we wouldn't know about it, since the pages
1680 * can be no longer dirty nor marked anymore for writeback (if a
1681 * subsequent modification to the extent buffer didn't happen before the
1682 * transaction commit), which makes filemap_fdata[write|wait]_range not
1683 * able to find the pages which contain errors at transaction
1684 * commit time. So if this happens we must abort the transaction,
1685 * otherwise we commit a super block with btree roots that point to
1686 * btree nodes/leafs whose content on disk is invalid - either garbage
1687 * or the content of some node/leaf from a past generation that got
1688 * cowed or deleted and is no longer valid.
1689 *
1690 * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
1691 * not be enough - we need to distinguish between log tree extents vs
1692 * non-log tree extents, and the next filemap_fdatawait_range() call
1693 * will catch and clear such errors in the mapping - and that call might
1694 * be from a log sync and not from a transaction commit. Also, checking
1695 * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
1696 * not done and would not be reliable - the eb might have been released
1697 * from memory and reading it back again means that flag would not be
1698 * set (since it's a runtime flag, not persisted on disk).
1699 *
1700 * Using the flags below in the btree inode also makes us achieve the
1701 * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
1702 * writeback for all dirty pages and before filemap_fdatawait_range()
1703 * is called, the writeback for all dirty pages had already finished
1704 * with errors - because we were not using AS_EIO/AS_ENOSPC,
1705 * filemap_fdatawait_range() would return success, as it could not know
1706 * that writeback errors happened (the pages were no longer tagged for
1707 * writeback).
1708 */
1709 switch (eb->log_index) {
1710 case -1:
1711 set_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags);
1712 break;
1713 case 0:
1714 set_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
1715 break;
1716 case 1:
1717 set_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
1718 break;
1719 default:
1720 BUG(); /* unexpected, logic error */
1721 }
1722 }
1723
1724 /*
1725 * The endio specific version which won't touch any unsafe spinlock in endio
1726 * context.
1727 */
1728 static struct extent_buffer *find_extent_buffer_nolock(
1729 const struct btrfs_fs_info *fs_info, u64 start)
1730 {
1731 struct extent_buffer *eb;
1732
1733 rcu_read_lock();
1734 eb = radix_tree_lookup(&fs_info->buffer_radix,
1735 start >> fs_info->sectorsize_bits);
1736 if (eb && atomic_inc_not_zero(&eb->refs)) {
1737 rcu_read_unlock();
1738 return eb;
1739 }
1740 rcu_read_unlock();
1741 return NULL;
1742 }
1743
1744 static void end_bbio_meta_write(struct btrfs_bio *bbio)
1745 {
1746 struct extent_buffer *eb = bbio->private;
1747 struct btrfs_fs_info *fs_info = eb->fs_info;
1748 bool uptodate = !bbio->bio.bi_status;
1749 struct folio_iter fi;
1750 u32 bio_offset = 0;
1751
1752 if (!uptodate)
1753 set_btree_ioerr(eb);
1754
1755 bio_for_each_folio_all(fi, &bbio->bio) {
1756 u64 start = eb->start + bio_offset;
1757 struct folio *folio = fi.folio;
1758 u32 len = fi.length;
1759
1760 btrfs_folio_clear_writeback(fs_info, folio, start, len);
1761 bio_offset += len;
1762 }
1763
1764 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1765 smp_mb__after_atomic();
1766 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
1767
1768 bio_put(&bbio->bio);
1769 }
1770
1771 static void prepare_eb_write(struct extent_buffer *eb)
1772 {
1773 u32 nritems;
1774 unsigned long start;
1775 unsigned long end;
1776
1777 clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1778
1779 /* Set btree blocks beyond nritems with 0 to avoid stale content */
1780 nritems = btrfs_header_nritems(eb);
1781 if (btrfs_header_level(eb) > 0) {
1782 end = btrfs_node_key_ptr_offset(eb, nritems);
1783 memzero_extent_buffer(eb, end, eb->len - end);
1784 } else {
1785 /*
1786 * Leaf:
1787 * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
1788 */
1789 start = btrfs_item_nr_offset(eb, nritems);
1790 end = btrfs_item_nr_offset(eb, 0);
1791 if (nritems == 0)
1792 end += BTRFS_LEAF_DATA_SIZE(eb->fs_info);
1793 else
1794 end += btrfs_item_offset(eb, nritems - 1);
1795 memzero_extent_buffer(eb, start, end - start);
1796 }
1797 }
1798
1799 static noinline_for_stack void write_one_eb(struct extent_buffer *eb,
1800 struct writeback_control *wbc)
1801 {
1802 struct btrfs_fs_info *fs_info = eb->fs_info;
1803 struct btrfs_bio *bbio;
1804
1805 prepare_eb_write(eb);
1806
1807 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
1808 REQ_OP_WRITE | REQ_META | wbc_to_write_flags(wbc),
1809 eb->fs_info, end_bbio_meta_write, eb);
1810 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
1811 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
1812 wbc_init_bio(wbc, &bbio->bio);
1813 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
1814 bbio->file_offset = eb->start;
1815 if (fs_info->nodesize < PAGE_SIZE) {
1816 struct folio *folio = eb->folios[0];
1817 bool ret;
1818
1819 folio_lock(folio);
1820 btrfs_subpage_set_writeback(fs_info, folio, eb->start, eb->len);
1821 if (btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start,
1822 eb->len)) {
1823 folio_clear_dirty_for_io(folio);
1824 wbc->nr_to_write--;
1825 }
1826 ret = bio_add_folio(&bbio->bio, folio, eb->len,
1827 eb->start - folio_pos(folio));
1828 ASSERT(ret);
1829 wbc_account_cgroup_owner(wbc, folio_page(folio, 0), eb->len);
1830 folio_unlock(folio);
1831 } else {
1832 int num_folios = num_extent_folios(eb);
1833
1834 for (int i = 0; i < num_folios; i++) {
1835 struct folio *folio = eb->folios[i];
1836 bool ret;
1837
1838 folio_lock(folio);
1839 folio_clear_dirty_for_io(folio);
1840 folio_start_writeback(folio);
1841 ret = bio_add_folio(&bbio->bio, folio, eb->folio_size, 0);
1842 ASSERT(ret);
1843 wbc_account_cgroup_owner(wbc, folio_page(folio, 0),
1844 eb->folio_size);
1845 wbc->nr_to_write -= folio_nr_pages(folio);
1846 folio_unlock(folio);
1847 }
1848 }
1849 btrfs_submit_bio(bbio, 0);
1850 }
1851
1852 /*
1853 * Submit one subpage btree page.
1854 *
1855 * The main difference to submit_eb_page() is:
1856 * - Page locking
1857 * For subpage, we don't rely on page locking at all.
1858 *
1859 * - Flush write bio
1860 * We only flush bio if we may be unable to fit current extent buffers into
1861 * current bio.
1862 *
1863 * Return >=0 for the number of submitted extent buffers.
1864 * Return <0 for fatal error.
1865 */
1866 static int submit_eb_subpage(struct page *page, struct writeback_control *wbc)
1867 {
1868 struct btrfs_fs_info *fs_info = page_to_fs_info(page);
1869 struct folio *folio = page_folio(page);
1870 int submitted = 0;
1871 u64 page_start = page_offset(page);
1872 int bit_start = 0;
1873 int sectors_per_node = fs_info->nodesize >> fs_info->sectorsize_bits;
1874
1875 /* Lock and write each dirty extent buffers in the range */
1876 while (bit_start < fs_info->subpage_info->bitmap_nr_bits) {
1877 struct btrfs_subpage *subpage = folio_get_private(folio);
1878 struct extent_buffer *eb;
1879 unsigned long flags;
1880 u64 start;
1881
1882 /*
1883 * Take private lock to ensure the subpage won't be detached
1884 * in the meantime.
1885 */
1886 spin_lock(&page->mapping->i_private_lock);
1887 if (!folio_test_private(folio)) {
1888 spin_unlock(&page->mapping->i_private_lock);
1889 break;
1890 }
1891 spin_lock_irqsave(&subpage->lock, flags);
1892 if (!test_bit(bit_start + fs_info->subpage_info->dirty_offset,
1893 subpage->bitmaps)) {
1894 spin_unlock_irqrestore(&subpage->lock, flags);
1895 spin_unlock(&page->mapping->i_private_lock);
1896 bit_start++;
1897 continue;
1898 }
1899
1900 start = page_start + bit_start * fs_info->sectorsize;
1901 bit_start += sectors_per_node;
1902
1903 /*
1904 * Here we just want to grab the eb without touching extra
1905 * spin locks, so call find_extent_buffer_nolock().
1906 */
1907 eb = find_extent_buffer_nolock(fs_info, start);
1908 spin_unlock_irqrestore(&subpage->lock, flags);
1909 spin_unlock(&page->mapping->i_private_lock);
1910
1911 /*
1912 * The eb has already reached 0 refs thus find_extent_buffer()
1913 * doesn't return it. We don't need to write back such eb
1914 * anyway.
1915 */
1916 if (!eb)
1917 continue;
1918
1919 if (lock_extent_buffer_for_io(eb, wbc)) {
1920 write_one_eb(eb, wbc);
1921 submitted++;
1922 }
1923 free_extent_buffer(eb);
1924 }
1925 return submitted;
1926 }
1927
1928 /*
1929 * Submit all page(s) of one extent buffer.
1930 *
1931 * @page: the page of one extent buffer
1932 * @eb_context: to determine if we need to submit this page, if current page
1933 * belongs to this eb, we don't need to submit
1934 *
1935 * The caller should pass each page in their bytenr order, and here we use
1936 * @eb_context to determine if we have submitted pages of one extent buffer.
1937 *
1938 * If we have, we just skip until we hit a new page that doesn't belong to
1939 * current @eb_context.
1940 *
1941 * If not, we submit all the page(s) of the extent buffer.
1942 *
1943 * Return >0 if we have submitted the extent buffer successfully.
1944 * Return 0 if we don't need to submit the page, as it's already submitted by
1945 * previous call.
1946 * Return <0 for fatal error.
1947 */
1948 static int submit_eb_page(struct page *page, struct btrfs_eb_write_context *ctx)
1949 {
1950 struct writeback_control *wbc = ctx->wbc;
1951 struct address_space *mapping = page->mapping;
1952 struct folio *folio = page_folio(page);
1953 struct extent_buffer *eb;
1954 int ret;
1955
1956 if (!folio_test_private(folio))
1957 return 0;
1958
1959 if (page_to_fs_info(page)->nodesize < PAGE_SIZE)
1960 return submit_eb_subpage(page, wbc);
1961
1962 spin_lock(&mapping->i_private_lock);
1963 if (!folio_test_private(folio)) {
1964 spin_unlock(&mapping->i_private_lock);
1965 return 0;
1966 }
1967
1968 eb = folio_get_private(folio);
1969
1970 /*
1971 * Shouldn't happen and normally this would be a BUG_ON but no point
1972 * crashing the machine for something we can survive anyway.
1973 */
1974 if (WARN_ON(!eb)) {
1975 spin_unlock(&mapping->i_private_lock);
1976 return 0;
1977 }
1978
1979 if (eb == ctx->eb) {
1980 spin_unlock(&mapping->i_private_lock);
1981 return 0;
1982 }
1983 ret = atomic_inc_not_zero(&eb->refs);
1984 spin_unlock(&mapping->i_private_lock);
1985 if (!ret)
1986 return 0;
1987
1988 ctx->eb = eb;
1989
1990 ret = btrfs_check_meta_write_pointer(eb->fs_info, ctx);
1991 if (ret) {
1992 if (ret == -EBUSY)
1993 ret = 0;
1994 free_extent_buffer(eb);
1995 return ret;
1996 }
1997
1998 if (!lock_extent_buffer_for_io(eb, wbc)) {
1999 free_extent_buffer(eb);
2000 return 0;
2001 }
2002 /* Implies write in zoned mode. */
2003 if (ctx->zoned_bg) {
2004 /* Mark the last eb in the block group. */
2005 btrfs_schedule_zone_finish_bg(ctx->zoned_bg, eb);
2006 ctx->zoned_bg->meta_write_pointer += eb->len;
2007 }
2008 write_one_eb(eb, wbc);
2009 free_extent_buffer(eb);
2010 return 1;
2011 }
2012
2013 int btree_write_cache_pages(struct address_space *mapping,
2014 struct writeback_control *wbc)
2015 {
2016 struct btrfs_eb_write_context ctx = { .wbc = wbc };
2017 struct btrfs_fs_info *fs_info = inode_to_fs_info(mapping->host);
2018 int ret = 0;
2019 int done = 0;
2020 int nr_to_write_done = 0;
2021 struct folio_batch fbatch;
2022 unsigned int nr_folios;
2023 pgoff_t index;
2024 pgoff_t end; /* Inclusive */
2025 int scanned = 0;
2026 xa_mark_t tag;
2027
2028 folio_batch_init(&fbatch);
2029 if (wbc->range_cyclic) {
2030 index = mapping->writeback_index; /* Start from prev offset */
2031 end = -1;
2032 /*
2033 * Start from the beginning does not need to cycle over the
2034 * range, mark it as scanned.
2035 */
2036 scanned = (index == 0);
2037 } else {
2038 index = wbc->range_start >> PAGE_SHIFT;
2039 end = wbc->range_end >> PAGE_SHIFT;
2040 scanned = 1;
2041 }
2042 if (wbc->sync_mode == WB_SYNC_ALL)
2043 tag = PAGECACHE_TAG_TOWRITE;
2044 else
2045 tag = PAGECACHE_TAG_DIRTY;
2046 btrfs_zoned_meta_io_lock(fs_info);
2047 retry:
2048 if (wbc->sync_mode == WB_SYNC_ALL)
2049 tag_pages_for_writeback(mapping, index, end);
2050 while (!done && !nr_to_write_done && (index <= end) &&
2051 (nr_folios = filemap_get_folios_tag(mapping, &index, end,
2052 tag, &fbatch))) {
2053 unsigned i;
2054
2055 for (i = 0; i < nr_folios; i++) {
2056 struct folio *folio = fbatch.folios[i];
2057
2058 ret = submit_eb_page(&folio->page, &ctx);
2059 if (ret == 0)
2060 continue;
2061 if (ret < 0) {
2062 done = 1;
2063 break;
2064 }
2065
2066 /*
2067 * the filesystem may choose to bump up nr_to_write.
2068 * We have to make sure to honor the new nr_to_write
2069 * at any time
2070 */
2071 nr_to_write_done = wbc->nr_to_write <= 0;
2072 }
2073 folio_batch_release(&fbatch);
2074 cond_resched();
2075 }
2076 if (!scanned && !done) {
2077 /*
2078 * We hit the last page and there is more work to be done: wrap
2079 * back to the start of the file
2080 */
2081 scanned = 1;
2082 index = 0;
2083 goto retry;
2084 }
2085 /*
2086 * If something went wrong, don't allow any metadata write bio to be
2087 * submitted.
2088 *
2089 * This would prevent use-after-free if we had dirty pages not
2090 * cleaned up, which can still happen by fuzzed images.
2091 *
2092 * - Bad extent tree
2093 * Allowing existing tree block to be allocated for other trees.
2094 *
2095 * - Log tree operations
2096 * Exiting tree blocks get allocated to log tree, bumps its
2097 * generation, then get cleaned in tree re-balance.
2098 * Such tree block will not be written back, since it's clean,
2099 * thus no WRITTEN flag set.
2100 * And after log writes back, this tree block is not traced by
2101 * any dirty extent_io_tree.
2102 *
2103 * - Offending tree block gets re-dirtied from its original owner
2104 * Since it has bumped generation, no WRITTEN flag, it can be
2105 * reused without COWing. This tree block will not be traced
2106 * by btrfs_transaction::dirty_pages.
2107 *
2108 * Now such dirty tree block will not be cleaned by any dirty
2109 * extent io tree. Thus we don't want to submit such wild eb
2110 * if the fs already has error.
2111 *
2112 * We can get ret > 0 from submit_extent_page() indicating how many ebs
2113 * were submitted. Reset it to 0 to avoid false alerts for the caller.
2114 */
2115 if (ret > 0)
2116 ret = 0;
2117 if (!ret && BTRFS_FS_ERROR(fs_info))
2118 ret = -EROFS;
2119
2120 if (ctx.zoned_bg)
2121 btrfs_put_block_group(ctx.zoned_bg);
2122 btrfs_zoned_meta_io_unlock(fs_info);
2123 return ret;
2124 }
2125
2126 /*
2127 * Walk the list of dirty pages of the given address space and write all of them.
2128 *
2129 * @mapping: address space structure to write
2130 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2131 * @bio_ctrl: holds context for the write, namely the bio
2132 *
2133 * If a page is already under I/O, write_cache_pages() skips it, even
2134 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2135 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2136 * and msync() need to guarantee that all the data which was dirty at the time
2137 * the call was made get new I/O started against them. If wbc->sync_mode is
2138 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2139 * existing IO to complete.
2140 */
2141 static int extent_write_cache_pages(struct address_space *mapping,
2142 struct btrfs_bio_ctrl *bio_ctrl)
2143 {
2144 struct writeback_control *wbc = bio_ctrl->wbc;
2145 struct inode *inode = mapping->host;
2146 int ret = 0;
2147 int done = 0;
2148 int nr_to_write_done = 0;
2149 struct folio_batch fbatch;
2150 unsigned int nr_folios;
2151 pgoff_t index;
2152 pgoff_t end; /* Inclusive */
2153 pgoff_t done_index;
2154 int range_whole = 0;
2155 int scanned = 0;
2156 xa_mark_t tag;
2157
2158 /*
2159 * We have to hold onto the inode so that ordered extents can do their
2160 * work when the IO finishes. The alternative to this is failing to add
2161 * an ordered extent if the igrab() fails there and that is a huge pain
2162 * to deal with, so instead just hold onto the inode throughout the
2163 * writepages operation. If it fails here we are freeing up the inode
2164 * anyway and we'd rather not waste our time writing out stuff that is
2165 * going to be truncated anyway.
2166 */
2167 if (!igrab(inode))
2168 return 0;
2169
2170 folio_batch_init(&fbatch);
2171 if (wbc->range_cyclic) {
2172 index = mapping->writeback_index; /* Start from prev offset */
2173 end = -1;
2174 /*
2175 * Start from the beginning does not need to cycle over the
2176 * range, mark it as scanned.
2177 */
2178 scanned = (index == 0);
2179 } else {
2180 index = wbc->range_start >> PAGE_SHIFT;
2181 end = wbc->range_end >> PAGE_SHIFT;
2182 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2183 range_whole = 1;
2184 scanned = 1;
2185 }
2186
2187 /*
2188 * We do the tagged writepage as long as the snapshot flush bit is set
2189 * and we are the first one who do the filemap_flush() on this inode.
2190 *
2191 * The nr_to_write == LONG_MAX is needed to make sure other flushers do
2192 * not race in and drop the bit.
2193 */
2194 if (range_whole && wbc->nr_to_write == LONG_MAX &&
2195 test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
2196 &BTRFS_I(inode)->runtime_flags))
2197 wbc->tagged_writepages = 1;
2198
2199 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2200 tag = PAGECACHE_TAG_TOWRITE;
2201 else
2202 tag = PAGECACHE_TAG_DIRTY;
2203 retry:
2204 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2205 tag_pages_for_writeback(mapping, index, end);
2206 done_index = index;
2207 while (!done && !nr_to_write_done && (index <= end) &&
2208 (nr_folios = filemap_get_folios_tag(mapping, &index,
2209 end, tag, &fbatch))) {
2210 unsigned i;
2211
2212 for (i = 0; i < nr_folios; i++) {
2213 struct folio *folio = fbatch.folios[i];
2214
2215 done_index = folio_next_index(folio);
2216 /*
2217 * At this point we hold neither the i_pages lock nor
2218 * the page lock: the page may be truncated or
2219 * invalidated (changing page->mapping to NULL),
2220 * or even swizzled back from swapper_space to
2221 * tmpfs file mapping
2222 */
2223 if (!folio_trylock(folio)) {
2224 submit_write_bio(bio_ctrl, 0);
2225 folio_lock(folio);
2226 }
2227
2228 if (unlikely(folio->mapping != mapping)) {
2229 folio_unlock(folio);
2230 continue;
2231 }
2232
2233 if (!folio_test_dirty(folio)) {
2234 /* Someone wrote it for us. */
2235 folio_unlock(folio);
2236 continue;
2237 }
2238
2239 if (wbc->sync_mode != WB_SYNC_NONE) {
2240 if (folio_test_writeback(folio))
2241 submit_write_bio(bio_ctrl, 0);
2242 folio_wait_writeback(folio);
2243 }
2244
2245 if (folio_test_writeback(folio) ||
2246 !folio_clear_dirty_for_io(folio)) {
2247 folio_unlock(folio);
2248 continue;
2249 }
2250
2251 ret = __extent_writepage(&folio->page, bio_ctrl);
2252 if (ret < 0) {
2253 done = 1;
2254 break;
2255 }
2256
2257 /*
2258 * The filesystem may choose to bump up nr_to_write.
2259 * We have to make sure to honor the new nr_to_write
2260 * at any time.
2261 */
2262 nr_to_write_done = (wbc->sync_mode == WB_SYNC_NONE &&
2263 wbc->nr_to_write <= 0);
2264 }
2265 folio_batch_release(&fbatch);
2266 cond_resched();
2267 }
2268 if (!scanned && !done) {
2269 /*
2270 * We hit the last page and there is more work to be done: wrap
2271 * back to the start of the file
2272 */
2273 scanned = 1;
2274 index = 0;
2275
2276 /*
2277 * If we're looping we could run into a page that is locked by a
2278 * writer and that writer could be waiting on writeback for a
2279 * page in our current bio, and thus deadlock, so flush the
2280 * write bio here.
2281 */
2282 submit_write_bio(bio_ctrl, 0);
2283 goto retry;
2284 }
2285
2286 if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
2287 mapping->writeback_index = done_index;
2288
2289 btrfs_add_delayed_iput(BTRFS_I(inode));
2290 return ret;
2291 }
2292
2293 /*
2294 * Submit the pages in the range to bio for call sites which delalloc range has
2295 * already been ran (aka, ordered extent inserted) and all pages are still
2296 * locked.
2297 */
2298 void extent_write_locked_range(struct inode *inode, const struct page *locked_page,
2299 u64 start, u64 end, struct writeback_control *wbc,
2300 bool pages_dirty)
2301 {
2302 bool found_error = false;
2303 int ret = 0;
2304 struct address_space *mapping = inode->i_mapping;
2305 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
2306 const u32 sectorsize = fs_info->sectorsize;
2307 loff_t i_size = i_size_read(inode);
2308 u64 cur = start;
2309 struct btrfs_bio_ctrl bio_ctrl = {
2310 .wbc = wbc,
2311 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2312 };
2313
2314 if (wbc->no_cgroup_owner)
2315 bio_ctrl.opf |= REQ_BTRFS_CGROUP_PUNT;
2316
2317 ASSERT(IS_ALIGNED(start, sectorsize) && IS_ALIGNED(end + 1, sectorsize));
2318
2319 while (cur <= end) {
2320 u64 cur_end = min(round_down(cur, PAGE_SIZE) + PAGE_SIZE - 1, end);
2321 u32 cur_len = cur_end + 1 - cur;
2322 struct page *page;
2323 int nr = 0;
2324
2325 page = find_get_page(mapping, cur >> PAGE_SHIFT);
2326 ASSERT(PageLocked(page));
2327 if (pages_dirty && page != locked_page)
2328 ASSERT(PageDirty(page));
2329
2330 ret = __extent_writepage_io(BTRFS_I(inode), page, cur, cur_len,
2331 &bio_ctrl, i_size, &nr);
2332 if (ret == 1)
2333 goto next_page;
2334
2335 /* Make sure the mapping tag for page dirty gets cleared. */
2336 if (nr == 0) {
2337 struct folio *folio;
2338
2339 folio = page_folio(page);
2340 btrfs_folio_set_writeback(fs_info, folio, cur, cur_len);
2341 btrfs_folio_clear_writeback(fs_info, folio, cur, cur_len);
2342 }
2343 if (ret) {
2344 btrfs_mark_ordered_io_finished(BTRFS_I(inode), page,
2345 cur, cur_len, !ret);
2346 mapping_set_error(page->mapping, ret);
2347 }
2348 btrfs_folio_unlock_writer(fs_info, page_folio(page), cur, cur_len);
2349 if (ret < 0)
2350 found_error = true;
2351 next_page:
2352 put_page(page);
2353 cur = cur_end + 1;
2354 }
2355
2356 submit_write_bio(&bio_ctrl, found_error ? ret : 0);
2357 }
2358
2359 int btrfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
2360 {
2361 struct inode *inode = mapping->host;
2362 int ret = 0;
2363 struct btrfs_bio_ctrl bio_ctrl = {
2364 .wbc = wbc,
2365 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2366 };
2367
2368 /*
2369 * Allow only a single thread to do the reloc work in zoned mode to
2370 * protect the write pointer updates.
2371 */
2372 btrfs_zoned_data_reloc_lock(BTRFS_I(inode));
2373 ret = extent_write_cache_pages(mapping, &bio_ctrl);
2374 submit_write_bio(&bio_ctrl, ret);
2375 btrfs_zoned_data_reloc_unlock(BTRFS_I(inode));
2376 return ret;
2377 }
2378
2379 void btrfs_readahead(struct readahead_control *rac)
2380 {
2381 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ | REQ_RAHEAD };
2382 struct page *pagepool[16];
2383 struct extent_map *em_cached = NULL;
2384 u64 prev_em_start = (u64)-1;
2385 int nr;
2386
2387 while ((nr = readahead_page_batch(rac, pagepool))) {
2388 u64 contig_start = readahead_pos(rac);
2389 u64 contig_end = contig_start + readahead_batch_length(rac) - 1;
2390
2391 contiguous_readpages(pagepool, nr, contig_start, contig_end,
2392 &em_cached, &bio_ctrl, &prev_em_start);
2393 }
2394
2395 if (em_cached)
2396 free_extent_map(em_cached);
2397 submit_one_bio(&bio_ctrl);
2398 }
2399
2400 /*
2401 * basic invalidate_folio code, this waits on any locked or writeback
2402 * ranges corresponding to the folio, and then deletes any extent state
2403 * records from the tree
2404 */
2405 int extent_invalidate_folio(struct extent_io_tree *tree,
2406 struct folio *folio, size_t offset)
2407 {
2408 struct extent_state *cached_state = NULL;
2409 u64 start = folio_pos(folio);
2410 u64 end = start + folio_size(folio) - 1;
2411 size_t blocksize = folio_to_fs_info(folio)->sectorsize;
2412
2413 /* This function is only called for the btree inode */
2414 ASSERT(tree->owner == IO_TREE_BTREE_INODE_IO);
2415
2416 start += ALIGN(offset, blocksize);
2417 if (start > end)
2418 return 0;
2419
2420 lock_extent(tree, start, end, &cached_state);
2421 folio_wait_writeback(folio);
2422
2423 /*
2424 * Currently for btree io tree, only EXTENT_LOCKED is utilized,
2425 * so here we only need to unlock the extent range to free any
2426 * existing extent state.
2427 */
2428 unlock_extent(tree, start, end, &cached_state);
2429 return 0;
2430 }
2431
2432 /*
2433 * a helper for release_folio, this tests for areas of the page that
2434 * are locked or under IO and drops the related state bits if it is safe
2435 * to drop the page.
2436 */
2437 static bool try_release_extent_state(struct extent_io_tree *tree,
2438 struct page *page, gfp_t mask)
2439 {
2440 u64 start = page_offset(page);
2441 u64 end = start + PAGE_SIZE - 1;
2442 bool ret;
2443
2444 if (test_range_bit_exists(tree, start, end, EXTENT_LOCKED)) {
2445 ret = false;
2446 } else {
2447 u32 clear_bits = ~(EXTENT_LOCKED | EXTENT_NODATASUM |
2448 EXTENT_DELALLOC_NEW | EXTENT_CTLBITS |
2449 EXTENT_QGROUP_RESERVED);
2450 int ret2;
2451
2452 /*
2453 * At this point we can safely clear everything except the
2454 * locked bit, the nodatasum bit and the delalloc new bit.
2455 * The delalloc new bit will be cleared by ordered extent
2456 * completion.
2457 */
2458 ret2 = __clear_extent_bit(tree, start, end, clear_bits, NULL, NULL);
2459
2460 /* if clear_extent_bit failed for enomem reasons,
2461 * we can't allow the release to continue.
2462 */
2463 if (ret2 < 0)
2464 ret = false;
2465 else
2466 ret = true;
2467 }
2468 return ret;
2469 }
2470
2471 /*
2472 * a helper for release_folio. As long as there are no locked extents
2473 * in the range corresponding to the page, both state records and extent
2474 * map records are removed
2475 */
2476 bool try_release_extent_mapping(struct page *page, gfp_t mask)
2477 {
2478 u64 start = page_offset(page);
2479 u64 end = start + PAGE_SIZE - 1;
2480 struct btrfs_inode *inode = page_to_inode(page);
2481 struct extent_io_tree *io_tree = &inode->io_tree;
2482
2483 while (start <= end) {
2484 const u64 cur_gen = btrfs_get_fs_generation(inode->root->fs_info);
2485 const u64 len = end - start + 1;
2486 struct extent_map_tree *extent_tree = &inode->extent_tree;
2487 struct extent_map *em;
2488
2489 write_lock(&extent_tree->lock);
2490 em = lookup_extent_mapping(extent_tree, start, len);
2491 if (!em) {
2492 write_unlock(&extent_tree->lock);
2493 break;
2494 }
2495 if ((em->flags & EXTENT_FLAG_PINNED) || em->start != start) {
2496 write_unlock(&extent_tree->lock);
2497 free_extent_map(em);
2498 break;
2499 }
2500 if (test_range_bit_exists(io_tree, em->start,
2501 extent_map_end(em) - 1, EXTENT_LOCKED))
2502 goto next;
2503 /*
2504 * If it's not in the list of modified extents, used by a fast
2505 * fsync, we can remove it. If it's being logged we can safely
2506 * remove it since fsync took an extra reference on the em.
2507 */
2508 if (list_empty(&em->list) || (em->flags & EXTENT_FLAG_LOGGING))
2509 goto remove_em;
2510 /*
2511 * If it's in the list of modified extents, remove it only if
2512 * its generation is older then the current one, in which case
2513 * we don't need it for a fast fsync. Otherwise don't remove it,
2514 * we could be racing with an ongoing fast fsync that could miss
2515 * the new extent.
2516 */
2517 if (em->generation >= cur_gen)
2518 goto next;
2519 remove_em:
2520 /*
2521 * We only remove extent maps that are not in the list of
2522 * modified extents or that are in the list but with a
2523 * generation lower then the current generation, so there is no
2524 * need to set the full fsync flag on the inode (it hurts the
2525 * fsync performance for workloads with a data size that exceeds
2526 * or is close to the system's memory).
2527 */
2528 remove_extent_mapping(inode, em);
2529 /* Once for the inode's extent map tree. */
2530 free_extent_map(em);
2531 next:
2532 start = extent_map_end(em);
2533 write_unlock(&extent_tree->lock);
2534
2535 /* Once for us, for the lookup_extent_mapping() reference. */
2536 free_extent_map(em);
2537
2538 if (need_resched()) {
2539 /*
2540 * If we need to resched but we can't block just exit
2541 * and leave any remaining extent maps.
2542 */
2543 if (!gfpflags_allow_blocking(mask))
2544 break;
2545
2546 cond_resched();
2547 }
2548 }
2549 return try_release_extent_state(io_tree, page, mask);
2550 }
2551
2552 static void __free_extent_buffer(struct extent_buffer *eb)
2553 {
2554 kmem_cache_free(extent_buffer_cache, eb);
2555 }
2556
2557 static int extent_buffer_under_io(const struct extent_buffer *eb)
2558 {
2559 return (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
2560 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
2561 }
2562
2563 static bool folio_range_has_eb(struct btrfs_fs_info *fs_info, struct folio *folio)
2564 {
2565 struct btrfs_subpage *subpage;
2566
2567 lockdep_assert_held(&folio->mapping->i_private_lock);
2568
2569 if (folio_test_private(folio)) {
2570 subpage = folio_get_private(folio);
2571 if (atomic_read(&subpage->eb_refs))
2572 return true;
2573 /*
2574 * Even there is no eb refs here, we may still have
2575 * end_page_read() call relying on page::private.
2576 */
2577 if (atomic_read(&subpage->readers))
2578 return true;
2579 }
2580 return false;
2581 }
2582
2583 static void detach_extent_buffer_folio(const struct extent_buffer *eb, struct folio *folio)
2584 {
2585 struct btrfs_fs_info *fs_info = eb->fs_info;
2586 const bool mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
2587
2588 /*
2589 * For mapped eb, we're going to change the folio private, which should
2590 * be done under the i_private_lock.
2591 */
2592 if (mapped)
2593 spin_lock(&folio->mapping->i_private_lock);
2594
2595 if (!folio_test_private(folio)) {
2596 if (mapped)
2597 spin_unlock(&folio->mapping->i_private_lock);
2598 return;
2599 }
2600
2601 if (fs_info->nodesize >= PAGE_SIZE) {
2602 /*
2603 * We do this since we'll remove the pages after we've
2604 * removed the eb from the radix tree, so we could race
2605 * and have this page now attached to the new eb. So
2606 * only clear folio if it's still connected to
2607 * this eb.
2608 */
2609 if (folio_test_private(folio) && folio_get_private(folio) == eb) {
2610 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
2611 BUG_ON(folio_test_dirty(folio));
2612 BUG_ON(folio_test_writeback(folio));
2613 /* We need to make sure we haven't be attached to a new eb. */
2614 folio_detach_private(folio);
2615 }
2616 if (mapped)
2617 spin_unlock(&folio->mapping->i_private_lock);
2618 return;
2619 }
2620
2621 /*
2622 * For subpage, we can have dummy eb with folio private attached. In
2623 * this case, we can directly detach the private as such folio is only
2624 * attached to one dummy eb, no sharing.
2625 */
2626 if (!mapped) {
2627 btrfs_detach_subpage(fs_info, folio);
2628 return;
2629 }
2630
2631 btrfs_folio_dec_eb_refs(fs_info, folio);
2632
2633 /*
2634 * We can only detach the folio private if there are no other ebs in the
2635 * page range and no unfinished IO.
2636 */
2637 if (!folio_range_has_eb(fs_info, folio))
2638 btrfs_detach_subpage(fs_info, folio);
2639
2640 spin_unlock(&folio->mapping->i_private_lock);
2641 }
2642
2643 /* Release all pages attached to the extent buffer */
2644 static void btrfs_release_extent_buffer_pages(const struct extent_buffer *eb)
2645 {
2646 ASSERT(!extent_buffer_under_io(eb));
2647
2648 for (int i = 0; i < INLINE_EXTENT_BUFFER_PAGES; i++) {
2649 struct folio *folio = eb->folios[i];
2650
2651 if (!folio)
2652 continue;
2653
2654 detach_extent_buffer_folio(eb, folio);
2655
2656 /* One for when we allocated the folio. */
2657 folio_put(folio);
2658 }
2659 }
2660
2661 /*
2662 * Helper for releasing the extent buffer.
2663 */
2664 static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
2665 {
2666 btrfs_release_extent_buffer_pages(eb);
2667 btrfs_leak_debug_del_eb(eb);
2668 __free_extent_buffer(eb);
2669 }
2670
2671 static struct extent_buffer *
2672 __alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
2673 unsigned long len)
2674 {
2675 struct extent_buffer *eb = NULL;
2676
2677 eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
2678 eb->start = start;
2679 eb->len = len;
2680 eb->fs_info = fs_info;
2681 init_rwsem(&eb->lock);
2682
2683 btrfs_leak_debug_add_eb(eb);
2684
2685 spin_lock_init(&eb->refs_lock);
2686 atomic_set(&eb->refs, 1);
2687
2688 ASSERT(len <= BTRFS_MAX_METADATA_BLOCKSIZE);
2689
2690 return eb;
2691 }
2692
2693 struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
2694 {
2695 struct extent_buffer *new;
2696 int num_folios = num_extent_folios(src);
2697 int ret;
2698
2699 new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
2700 if (new == NULL)
2701 return NULL;
2702
2703 /*
2704 * Set UNMAPPED before calling btrfs_release_extent_buffer(), as
2705 * btrfs_release_extent_buffer() have different behavior for
2706 * UNMAPPED subpage extent buffer.
2707 */
2708 set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
2709
2710 ret = alloc_eb_folio_array(new, false);
2711 if (ret) {
2712 btrfs_release_extent_buffer(new);
2713 return NULL;
2714 }
2715
2716 for (int i = 0; i < num_folios; i++) {
2717 struct folio *folio = new->folios[i];
2718
2719 ret = attach_extent_buffer_folio(new, folio, NULL);
2720 if (ret < 0) {
2721 btrfs_release_extent_buffer(new);
2722 return NULL;
2723 }
2724 WARN_ON(folio_test_dirty(folio));
2725 }
2726 copy_extent_buffer_full(new, src);
2727 set_extent_buffer_uptodate(new);
2728
2729 return new;
2730 }
2731
2732 struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
2733 u64 start, unsigned long len)
2734 {
2735 struct extent_buffer *eb;
2736 int num_folios = 0;
2737 int ret;
2738
2739 eb = __alloc_extent_buffer(fs_info, start, len);
2740 if (!eb)
2741 return NULL;
2742
2743 ret = alloc_eb_folio_array(eb, false);
2744 if (ret)
2745 goto err;
2746
2747 num_folios = num_extent_folios(eb);
2748 for (int i = 0; i < num_folios; i++) {
2749 ret = attach_extent_buffer_folio(eb, eb->folios[i], NULL);
2750 if (ret < 0)
2751 goto err;
2752 }
2753
2754 set_extent_buffer_uptodate(eb);
2755 btrfs_set_header_nritems(eb, 0);
2756 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
2757
2758 return eb;
2759 err:
2760 for (int i = 0; i < num_folios; i++) {
2761 if (eb->folios[i]) {
2762 detach_extent_buffer_folio(eb, eb->folios[i]);
2763 folio_put(eb->folios[i]);
2764 }
2765 }
2766 __free_extent_buffer(eb);
2767 return NULL;
2768 }
2769
2770 struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
2771 u64 start)
2772 {
2773 return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
2774 }
2775
2776 static void check_buffer_tree_ref(struct extent_buffer *eb)
2777 {
2778 int refs;
2779 /*
2780 * The TREE_REF bit is first set when the extent_buffer is added
2781 * to the radix tree. It is also reset, if unset, when a new reference
2782 * is created by find_extent_buffer.
2783 *
2784 * It is only cleared in two cases: freeing the last non-tree
2785 * reference to the extent_buffer when its STALE bit is set or
2786 * calling release_folio when the tree reference is the only reference.
2787 *
2788 * In both cases, care is taken to ensure that the extent_buffer's
2789 * pages are not under io. However, release_folio can be concurrently
2790 * called with creating new references, which is prone to race
2791 * conditions between the calls to check_buffer_tree_ref in those
2792 * codepaths and clearing TREE_REF in try_release_extent_buffer.
2793 *
2794 * The actual lifetime of the extent_buffer in the radix tree is
2795 * adequately protected by the refcount, but the TREE_REF bit and
2796 * its corresponding reference are not. To protect against this
2797 * class of races, we call check_buffer_tree_ref from the codepaths
2798 * which trigger io. Note that once io is initiated, TREE_REF can no
2799 * longer be cleared, so that is the moment at which any such race is
2800 * best fixed.
2801 */
2802 refs = atomic_read(&eb->refs);
2803 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
2804 return;
2805
2806 spin_lock(&eb->refs_lock);
2807 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
2808 atomic_inc(&eb->refs);
2809 spin_unlock(&eb->refs_lock);
2810 }
2811
2812 static void mark_extent_buffer_accessed(struct extent_buffer *eb)
2813 {
2814 int num_folios= num_extent_folios(eb);
2815
2816 check_buffer_tree_ref(eb);
2817
2818 for (int i = 0; i < num_folios; i++)
2819 folio_mark_accessed(eb->folios[i]);
2820 }
2821
2822 struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
2823 u64 start)
2824 {
2825 struct extent_buffer *eb;
2826
2827 eb = find_extent_buffer_nolock(fs_info, start);
2828 if (!eb)
2829 return NULL;
2830 /*
2831 * Lock our eb's refs_lock to avoid races with free_extent_buffer().
2832 * When we get our eb it might be flagged with EXTENT_BUFFER_STALE and
2833 * another task running free_extent_buffer() might have seen that flag
2834 * set, eb->refs == 2, that the buffer isn't under IO (dirty and
2835 * writeback flags not set) and it's still in the tree (flag
2836 * EXTENT_BUFFER_TREE_REF set), therefore being in the process of
2837 * decrementing the extent buffer's reference count twice. So here we
2838 * could race and increment the eb's reference count, clear its stale
2839 * flag, mark it as dirty and drop our reference before the other task
2840 * finishes executing free_extent_buffer, which would later result in
2841 * an attempt to free an extent buffer that is dirty.
2842 */
2843 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
2844 spin_lock(&eb->refs_lock);
2845 spin_unlock(&eb->refs_lock);
2846 }
2847 mark_extent_buffer_accessed(eb);
2848 return eb;
2849 }
2850
2851 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
2852 struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
2853 u64 start)
2854 {
2855 struct extent_buffer *eb, *exists = NULL;
2856 int ret;
2857
2858 eb = find_extent_buffer(fs_info, start);
2859 if (eb)
2860 return eb;
2861 eb = alloc_dummy_extent_buffer(fs_info, start);
2862 if (!eb)
2863 return ERR_PTR(-ENOMEM);
2864 eb->fs_info = fs_info;
2865 again:
2866 ret = radix_tree_preload(GFP_NOFS);
2867 if (ret) {
2868 exists = ERR_PTR(ret);
2869 goto free_eb;
2870 }
2871 spin_lock(&fs_info->buffer_lock);
2872 ret = radix_tree_insert(&fs_info->buffer_radix,
2873 start >> fs_info->sectorsize_bits, eb);
2874 spin_unlock(&fs_info->buffer_lock);
2875 radix_tree_preload_end();
2876 if (ret == -EEXIST) {
2877 exists = find_extent_buffer(fs_info, start);
2878 if (exists)
2879 goto free_eb;
2880 else
2881 goto again;
2882 }
2883 check_buffer_tree_ref(eb);
2884 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
2885
2886 return eb;
2887 free_eb:
2888 btrfs_release_extent_buffer(eb);
2889 return exists;
2890 }
2891 #endif
2892
2893 static struct extent_buffer *grab_extent_buffer(
2894 struct btrfs_fs_info *fs_info, struct page *page)
2895 {
2896 struct folio *folio = page_folio(page);
2897 struct extent_buffer *exists;
2898
2899 lockdep_assert_held(&page->mapping->i_private_lock);
2900
2901 /*
2902 * For subpage case, we completely rely on radix tree to ensure we
2903 * don't try to insert two ebs for the same bytenr. So here we always
2904 * return NULL and just continue.
2905 */
2906 if (fs_info->nodesize < PAGE_SIZE)
2907 return NULL;
2908
2909 /* Page not yet attached to an extent buffer */
2910 if (!folio_test_private(folio))
2911 return NULL;
2912
2913 /*
2914 * We could have already allocated an eb for this page and attached one
2915 * so lets see if we can get a ref on the existing eb, and if we can we
2916 * know it's good and we can just return that one, else we know we can
2917 * just overwrite folio private.
2918 */
2919 exists = folio_get_private(folio);
2920 if (atomic_inc_not_zero(&exists->refs))
2921 return exists;
2922
2923 WARN_ON(PageDirty(page));
2924 folio_detach_private(folio);
2925 return NULL;
2926 }
2927
2928 static int check_eb_alignment(struct btrfs_fs_info *fs_info, u64 start)
2929 {
2930 if (!IS_ALIGNED(start, fs_info->sectorsize)) {
2931 btrfs_err(fs_info, "bad tree block start %llu", start);
2932 return -EINVAL;
2933 }
2934
2935 if (fs_info->nodesize < PAGE_SIZE &&
2936 offset_in_page(start) + fs_info->nodesize > PAGE_SIZE) {
2937 btrfs_err(fs_info,
2938 "tree block crosses page boundary, start %llu nodesize %u",
2939 start, fs_info->nodesize);
2940 return -EINVAL;
2941 }
2942 if (fs_info->nodesize >= PAGE_SIZE &&
2943 !PAGE_ALIGNED(start)) {
2944 btrfs_err(fs_info,
2945 "tree block is not page aligned, start %llu nodesize %u",
2946 start, fs_info->nodesize);
2947 return -EINVAL;
2948 }
2949 if (!IS_ALIGNED(start, fs_info->nodesize) &&
2950 !test_and_set_bit(BTRFS_FS_UNALIGNED_TREE_BLOCK, &fs_info->flags)) {
2951 btrfs_warn(fs_info,
2952 "tree block not nodesize aligned, start %llu nodesize %u, can be resolved by a full metadata balance",
2953 start, fs_info->nodesize);
2954 }
2955 return 0;
2956 }
2957
2958
2959 /*
2960 * Return 0 if eb->folios[i] is attached to btree inode successfully.
2961 * Return >0 if there is already another extent buffer for the range,
2962 * and @found_eb_ret would be updated.
2963 * Return -EAGAIN if the filemap has an existing folio but with different size
2964 * than @eb.
2965 * The caller needs to free the existing folios and retry using the same order.
2966 */
2967 static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
2968 struct btrfs_subpage *prealloc,
2969 struct extent_buffer **found_eb_ret)
2970 {
2971
2972 struct btrfs_fs_info *fs_info = eb->fs_info;
2973 struct address_space *mapping = fs_info->btree_inode->i_mapping;
2974 const unsigned long index = eb->start >> PAGE_SHIFT;
2975 struct folio *existing_folio = NULL;
2976 int ret;
2977
2978 ASSERT(found_eb_ret);
2979
2980 /* Caller should ensure the folio exists. */
2981 ASSERT(eb->folios[i]);
2982
2983 retry:
2984 ret = filemap_add_folio(mapping, eb->folios[i], index + i,
2985 GFP_NOFS | __GFP_NOFAIL);
2986 if (!ret)
2987 goto finish;
2988
2989 existing_folio = filemap_lock_folio(mapping, index + i);
2990 /* The page cache only exists for a very short time, just retry. */
2991 if (IS_ERR(existing_folio)) {
2992 existing_folio = NULL;
2993 goto retry;
2994 }
2995
2996 /* For now, we should only have single-page folios for btree inode. */
2997 ASSERT(folio_nr_pages(existing_folio) == 1);
2998
2999 if (folio_size(existing_folio) != eb->folio_size) {
3000 folio_unlock(existing_folio);
3001 folio_put(existing_folio);
3002 return -EAGAIN;
3003 }
3004
3005 finish:
3006 spin_lock(&mapping->i_private_lock);
3007 if (existing_folio && fs_info->nodesize < PAGE_SIZE) {
3008 /* We're going to reuse the existing page, can drop our folio now. */
3009 __free_page(folio_page(eb->folios[i], 0));
3010 eb->folios[i] = existing_folio;
3011 } else if (existing_folio) {
3012 struct extent_buffer *existing_eb;
3013
3014 existing_eb = grab_extent_buffer(fs_info,
3015 folio_page(existing_folio, 0));
3016 if (existing_eb) {
3017 /* The extent buffer still exists, we can use it directly. */
3018 *found_eb_ret = existing_eb;
3019 spin_unlock(&mapping->i_private_lock);
3020 folio_unlock(existing_folio);
3021 folio_put(existing_folio);
3022 return 1;
3023 }
3024 /* The extent buffer no longer exists, we can reuse the folio. */
3025 __free_page(folio_page(eb->folios[i], 0));
3026 eb->folios[i] = existing_folio;
3027 }
3028 eb->folio_size = folio_size(eb->folios[i]);
3029 eb->folio_shift = folio_shift(eb->folios[i]);
3030 /* Should not fail, as we have preallocated the memory. */
3031 ret = attach_extent_buffer_folio(eb, eb->folios[i], prealloc);
3032 ASSERT(!ret);
3033 /*
3034 * To inform we have an extra eb under allocation, so that
3035 * detach_extent_buffer_page() won't release the folio private when the
3036 * eb hasn't been inserted into radix tree yet.
3037 *
3038 * The ref will be decreased when the eb releases the page, in
3039 * detach_extent_buffer_page(). Thus needs no special handling in the
3040 * error path.
3041 */
3042 btrfs_folio_inc_eb_refs(fs_info, eb->folios[i]);
3043 spin_unlock(&mapping->i_private_lock);
3044 return 0;
3045 }
3046
3047 struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
3048 u64 start, u64 owner_root, int level)
3049 {
3050 unsigned long len = fs_info->nodesize;
3051 int num_folios;
3052 int attached = 0;
3053 struct extent_buffer *eb;
3054 struct extent_buffer *existing_eb = NULL;
3055 struct btrfs_subpage *prealloc = NULL;
3056 u64 lockdep_owner = owner_root;
3057 bool page_contig = true;
3058 int uptodate = 1;
3059 int ret;
3060
3061 if (check_eb_alignment(fs_info, start))
3062 return ERR_PTR(-EINVAL);
3063
3064 #if BITS_PER_LONG == 32
3065 if (start >= MAX_LFS_FILESIZE) {
3066 btrfs_err_rl(fs_info,
3067 "extent buffer %llu is beyond 32bit page cache limit", start);
3068 btrfs_err_32bit_limit(fs_info);
3069 return ERR_PTR(-EOVERFLOW);
3070 }
3071 if (start >= BTRFS_32BIT_EARLY_WARN_THRESHOLD)
3072 btrfs_warn_32bit_limit(fs_info);
3073 #endif
3074
3075 eb = find_extent_buffer(fs_info, start);
3076 if (eb)
3077 return eb;
3078
3079 eb = __alloc_extent_buffer(fs_info, start, len);
3080 if (!eb)
3081 return ERR_PTR(-ENOMEM);
3082
3083 /*
3084 * The reloc trees are just snapshots, so we need them to appear to be
3085 * just like any other fs tree WRT lockdep.
3086 */
3087 if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID)
3088 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
3089
3090 btrfs_set_buffer_lockdep_class(lockdep_owner, eb, level);
3091
3092 /*
3093 * Preallocate folio private for subpage case, so that we won't
3094 * allocate memory with i_private_lock nor page lock hold.
3095 *
3096 * The memory will be freed by attach_extent_buffer_page() or freed
3097 * manually if we exit earlier.
3098 */
3099 if (fs_info->nodesize < PAGE_SIZE) {
3100 prealloc = btrfs_alloc_subpage(fs_info, BTRFS_SUBPAGE_METADATA);
3101 if (IS_ERR(prealloc)) {
3102 ret = PTR_ERR(prealloc);
3103 goto out;
3104 }
3105 }
3106
3107 reallocate:
3108 /* Allocate all pages first. */
3109 ret = alloc_eb_folio_array(eb, true);
3110 if (ret < 0) {
3111 btrfs_free_subpage(prealloc);
3112 goto out;
3113 }
3114
3115 num_folios = num_extent_folios(eb);
3116 /* Attach all pages to the filemap. */
3117 for (int i = 0; i < num_folios; i++) {
3118 struct folio *folio;
3119
3120 ret = attach_eb_folio_to_filemap(eb, i, prealloc, &existing_eb);
3121 if (ret > 0) {
3122 ASSERT(existing_eb);
3123 goto out;
3124 }
3125
3126 /*
3127 * TODO: Special handling for a corner case where the order of
3128 * folios mismatch between the new eb and filemap.
3129 *
3130 * This happens when:
3131 *
3132 * - the new eb is using higher order folio
3133 *
3134 * - the filemap is still using 0-order folios for the range
3135 * This can happen at the previous eb allocation, and we don't
3136 * have higher order folio for the call.
3137 *
3138 * - the existing eb has already been freed
3139 *
3140 * In this case, we have to free the existing folios first, and
3141 * re-allocate using the same order.
3142 * Thankfully this is not going to happen yet, as we're still
3143 * using 0-order folios.
3144 */
3145 if (unlikely(ret == -EAGAIN)) {
3146 ASSERT(0);
3147 goto reallocate;
3148 }
3149 attached++;
3150
3151 /*
3152 * Only after attach_eb_folio_to_filemap(), eb->folios[] is
3153 * reliable, as we may choose to reuse the existing page cache
3154 * and free the allocated page.
3155 */
3156 folio = eb->folios[i];
3157 WARN_ON(btrfs_folio_test_dirty(fs_info, folio, eb->start, eb->len));
3158
3159 /*
3160 * Check if the current page is physically contiguous with previous eb
3161 * page.
3162 * At this stage, either we allocated a large folio, thus @i
3163 * would only be 0, or we fall back to per-page allocation.
3164 */
3165 if (i && folio_page(eb->folios[i - 1], 0) + 1 != folio_page(folio, 0))
3166 page_contig = false;
3167
3168 if (!btrfs_folio_test_uptodate(fs_info, folio, eb->start, eb->len))
3169 uptodate = 0;
3170
3171 /*
3172 * We can't unlock the pages just yet since the extent buffer
3173 * hasn't been properly inserted in the radix tree, this
3174 * opens a race with btree_release_folio which can free a page
3175 * while we are still filling in all pages for the buffer and
3176 * we could crash.
3177 */
3178 }
3179 if (uptodate)
3180 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3181 /* All pages are physically contiguous, can skip cross page handling. */
3182 if (page_contig)
3183 eb->addr = folio_address(eb->folios[0]) + offset_in_page(eb->start);
3184 again:
3185 ret = radix_tree_preload(GFP_NOFS);
3186 if (ret)
3187 goto out;
3188
3189 spin_lock(&fs_info->buffer_lock);
3190 ret = radix_tree_insert(&fs_info->buffer_radix,
3191 start >> fs_info->sectorsize_bits, eb);
3192 spin_unlock(&fs_info->buffer_lock);
3193 radix_tree_preload_end();
3194 if (ret == -EEXIST) {
3195 ret = 0;
3196 existing_eb = find_extent_buffer(fs_info, start);
3197 if (existing_eb)
3198 goto out;
3199 else
3200 goto again;
3201 }
3202 /* add one reference for the tree */
3203 check_buffer_tree_ref(eb);
3204 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
3205
3206 /*
3207 * Now it's safe to unlock the pages because any calls to
3208 * btree_release_folio will correctly detect that a page belongs to a
3209 * live buffer and won't free them prematurely.
3210 */
3211 for (int i = 0; i < num_folios; i++)
3212 unlock_page(folio_page(eb->folios[i], 0));
3213 return eb;
3214
3215 out:
3216 WARN_ON(!atomic_dec_and_test(&eb->refs));
3217
3218 /*
3219 * Any attached folios need to be detached before we unlock them. This
3220 * is because when we're inserting our new folios into the mapping, and
3221 * then attaching our eb to that folio. If we fail to insert our folio
3222 * we'll lookup the folio for that index, and grab that EB. We do not
3223 * want that to grab this eb, as we're getting ready to free it. So we
3224 * have to detach it first and then unlock it.
3225 *
3226 * We have to drop our reference and NULL it out here because in the
3227 * subpage case detaching does a btrfs_folio_dec_eb_refs() for our eb.
3228 * Below when we call btrfs_release_extent_buffer() we will call
3229 * detach_extent_buffer_folio() on our remaining pages in the !subpage
3230 * case. If we left eb->folios[i] populated in the subpage case we'd
3231 * double put our reference and be super sad.
3232 */
3233 for (int i = 0; i < attached; i++) {
3234 ASSERT(eb->folios[i]);
3235 detach_extent_buffer_folio(eb, eb->folios[i]);
3236 unlock_page(folio_page(eb->folios[i], 0));
3237 folio_put(eb->folios[i]);
3238 eb->folios[i] = NULL;
3239 }
3240 /*
3241 * Now all pages of that extent buffer is unmapped, set UNMAPPED flag,
3242 * so it can be cleaned up without utlizing page->mapping.
3243 */
3244 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3245
3246 btrfs_release_extent_buffer(eb);
3247 if (ret < 0)
3248 return ERR_PTR(ret);
3249 ASSERT(existing_eb);
3250 return existing_eb;
3251 }
3252
3253 static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3254 {
3255 struct extent_buffer *eb =
3256 container_of(head, struct extent_buffer, rcu_head);
3257
3258 __free_extent_buffer(eb);
3259 }
3260
3261 static int release_extent_buffer(struct extent_buffer *eb)
3262 __releases(&eb->refs_lock)
3263 {
3264 lockdep_assert_held(&eb->refs_lock);
3265
3266 WARN_ON(atomic_read(&eb->refs) == 0);
3267 if (atomic_dec_and_test(&eb->refs)) {
3268 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
3269 struct btrfs_fs_info *fs_info = eb->fs_info;
3270
3271 spin_unlock(&eb->refs_lock);
3272
3273 spin_lock(&fs_info->buffer_lock);
3274 radix_tree_delete(&fs_info->buffer_radix,
3275 eb->start >> fs_info->sectorsize_bits);
3276 spin_unlock(&fs_info->buffer_lock);
3277 } else {
3278 spin_unlock(&eb->refs_lock);
3279 }
3280
3281 btrfs_leak_debug_del_eb(eb);
3282 /* Should be safe to release our pages at this point */
3283 btrfs_release_extent_buffer_pages(eb);
3284 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3285 if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
3286 __free_extent_buffer(eb);
3287 return 1;
3288 }
3289 #endif
3290 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3291 return 1;
3292 }
3293 spin_unlock(&eb->refs_lock);
3294
3295 return 0;
3296 }
3297
3298 void free_extent_buffer(struct extent_buffer *eb)
3299 {
3300 int refs;
3301 if (!eb)
3302 return;
3303
3304 refs = atomic_read(&eb->refs);
3305 while (1) {
3306 if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
3307 || (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
3308 refs == 1))
3309 break;
3310 if (atomic_try_cmpxchg(&eb->refs, &refs, refs - 1))
3311 return;
3312 }
3313
3314 spin_lock(&eb->refs_lock);
3315 if (atomic_read(&eb->refs) == 2 &&
3316 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
3317 !extent_buffer_under_io(eb) &&
3318 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3319 atomic_dec(&eb->refs);
3320
3321 /*
3322 * I know this is terrible, but it's temporary until we stop tracking
3323 * the uptodate bits and such for the extent buffers.
3324 */
3325 release_extent_buffer(eb);
3326 }
3327
3328 void free_extent_buffer_stale(struct extent_buffer *eb)
3329 {
3330 if (!eb)
3331 return;
3332
3333 spin_lock(&eb->refs_lock);
3334 set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
3335
3336 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
3337 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3338 atomic_dec(&eb->refs);
3339 release_extent_buffer(eb);
3340 }
3341
3342 static void btree_clear_folio_dirty(struct folio *folio)
3343 {
3344 ASSERT(folio_test_dirty(folio));
3345 ASSERT(folio_test_locked(folio));
3346 folio_clear_dirty_for_io(folio);
3347 xa_lock_irq(&folio->mapping->i_pages);
3348 if (!folio_test_dirty(folio))
3349 __xa_clear_mark(&folio->mapping->i_pages,
3350 folio_index(folio), PAGECACHE_TAG_DIRTY);
3351 xa_unlock_irq(&folio->mapping->i_pages);
3352 }
3353
3354 static void clear_subpage_extent_buffer_dirty(const struct extent_buffer *eb)
3355 {
3356 struct btrfs_fs_info *fs_info = eb->fs_info;
3357 struct folio *folio = eb->folios[0];
3358 bool last;
3359
3360 /* btree_clear_folio_dirty() needs page locked. */
3361 folio_lock(folio);
3362 last = btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start, eb->len);
3363 if (last)
3364 btree_clear_folio_dirty(folio);
3365 folio_unlock(folio);
3366 WARN_ON(atomic_read(&eb->refs) == 0);
3367 }
3368
3369 void btrfs_clear_buffer_dirty(struct btrfs_trans_handle *trans,
3370 struct extent_buffer *eb)
3371 {
3372 struct btrfs_fs_info *fs_info = eb->fs_info;
3373 int num_folios;
3374
3375 btrfs_assert_tree_write_locked(eb);
3376
3377 if (trans && btrfs_header_generation(eb) != trans->transid)
3378 return;
3379
3380 /*
3381 * Instead of clearing the dirty flag off of the buffer, mark it as
3382 * EXTENT_BUFFER_ZONED_ZEROOUT. This allows us to preserve
3383 * write-ordering in zoned mode, without the need to later re-dirty
3384 * the extent_buffer.
3385 *
3386 * The actual zeroout of the buffer will happen later in
3387 * btree_csum_one_bio.
3388 */
3389 if (btrfs_is_zoned(fs_info) && test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3390 set_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags);
3391 return;
3392 }
3393
3394 if (!test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags))
3395 return;
3396
3397 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes, -eb->len,
3398 fs_info->dirty_metadata_batch);
3399
3400 if (eb->fs_info->nodesize < PAGE_SIZE)
3401 return clear_subpage_extent_buffer_dirty(eb);
3402
3403 num_folios = num_extent_folios(eb);
3404 for (int i = 0; i < num_folios; i++) {
3405 struct folio *folio = eb->folios[i];
3406
3407 if (!folio_test_dirty(folio))
3408 continue;
3409 folio_lock(folio);
3410 btree_clear_folio_dirty(folio);
3411 folio_unlock(folio);
3412 }
3413 WARN_ON(atomic_read(&eb->refs) == 0);
3414 }
3415
3416 void set_extent_buffer_dirty(struct extent_buffer *eb)
3417 {
3418 int num_folios;
3419 bool was_dirty;
3420
3421 check_buffer_tree_ref(eb);
3422
3423 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3424
3425 num_folios = num_extent_folios(eb);
3426 WARN_ON(atomic_read(&eb->refs) == 0);
3427 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
3428 WARN_ON(test_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags));
3429
3430 if (!was_dirty) {
3431 bool subpage = eb->fs_info->nodesize < PAGE_SIZE;
3432
3433 /*
3434 * For subpage case, we can have other extent buffers in the
3435 * same page, and in clear_subpage_extent_buffer_dirty() we
3436 * have to clear page dirty without subpage lock held.
3437 * This can cause race where our page gets dirty cleared after
3438 * we just set it.
3439 *
3440 * Thankfully, clear_subpage_extent_buffer_dirty() has locked
3441 * its page for other reasons, we can use page lock to prevent
3442 * the above race.
3443 */
3444 if (subpage)
3445 lock_page(folio_page(eb->folios[0], 0));
3446 for (int i = 0; i < num_folios; i++)
3447 btrfs_folio_set_dirty(eb->fs_info, eb->folios[i],
3448 eb->start, eb->len);
3449 if (subpage)
3450 unlock_page(folio_page(eb->folios[0], 0));
3451 percpu_counter_add_batch(&eb->fs_info->dirty_metadata_bytes,
3452 eb->len,
3453 eb->fs_info->dirty_metadata_batch);
3454 }
3455 #ifdef CONFIG_BTRFS_DEBUG
3456 for (int i = 0; i < num_folios; i++)
3457 ASSERT(folio_test_dirty(eb->folios[i]));
3458 #endif
3459 }
3460
3461 void clear_extent_buffer_uptodate(struct extent_buffer *eb)
3462 {
3463 struct btrfs_fs_info *fs_info = eb->fs_info;
3464 int num_folios = num_extent_folios(eb);
3465
3466 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3467 for (int i = 0; i < num_folios; i++) {
3468 struct folio *folio = eb->folios[i];
3469
3470 if (!folio)
3471 continue;
3472
3473 /*
3474 * This is special handling for metadata subpage, as regular
3475 * btrfs_is_subpage() can not handle cloned/dummy metadata.
3476 */
3477 if (fs_info->nodesize >= PAGE_SIZE)
3478 folio_clear_uptodate(folio);
3479 else
3480 btrfs_subpage_clear_uptodate(fs_info, folio,
3481 eb->start, eb->len);
3482 }
3483 }
3484
3485 void set_extent_buffer_uptodate(struct extent_buffer *eb)
3486 {
3487 struct btrfs_fs_info *fs_info = eb->fs_info;
3488 int num_folios = num_extent_folios(eb);
3489
3490 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3491 for (int i = 0; i < num_folios; i++) {
3492 struct folio *folio = eb->folios[i];
3493
3494 /*
3495 * This is special handling for metadata subpage, as regular
3496 * btrfs_is_subpage() can not handle cloned/dummy metadata.
3497 */
3498 if (fs_info->nodesize >= PAGE_SIZE)
3499 folio_mark_uptodate(folio);
3500 else
3501 btrfs_subpage_set_uptodate(fs_info, folio,
3502 eb->start, eb->len);
3503 }
3504 }
3505
3506 static void clear_extent_buffer_reading(struct extent_buffer *eb)
3507 {
3508 clear_bit(EXTENT_BUFFER_READING, &eb->bflags);
3509 smp_mb__after_atomic();
3510 wake_up_bit(&eb->bflags, EXTENT_BUFFER_READING);
3511 }
3512
3513 static void end_bbio_meta_read(struct btrfs_bio *bbio)
3514 {
3515 struct extent_buffer *eb = bbio->private;
3516 struct btrfs_fs_info *fs_info = eb->fs_info;
3517 bool uptodate = !bbio->bio.bi_status;
3518 struct folio_iter fi;
3519 u32 bio_offset = 0;
3520
3521 /*
3522 * If the extent buffer is marked UPTODATE before the read operation
3523 * completes, other calls to read_extent_buffer_pages() will return
3524 * early without waiting for the read to finish, causing data races.
3525 */
3526 WARN_ON(test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags));
3527
3528 eb->read_mirror = bbio->mirror_num;
3529
3530 if (uptodate &&
3531 btrfs_validate_extent_buffer(eb, &bbio->parent_check) < 0)
3532 uptodate = false;
3533
3534 if (uptodate) {
3535 set_extent_buffer_uptodate(eb);
3536 } else {
3537 clear_extent_buffer_uptodate(eb);
3538 set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
3539 }
3540
3541 bio_for_each_folio_all(fi, &bbio->bio) {
3542 struct folio *folio = fi.folio;
3543 u64 start = eb->start + bio_offset;
3544 u32 len = fi.length;
3545
3546 if (uptodate)
3547 btrfs_folio_set_uptodate(fs_info, folio, start, len);
3548 else
3549 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
3550
3551 bio_offset += len;
3552 }
3553
3554 clear_extent_buffer_reading(eb);
3555 free_extent_buffer(eb);
3556
3557 bio_put(&bbio->bio);
3558 }
3559
3560 int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num,
3561 const struct btrfs_tree_parent_check *check)
3562 {
3563 struct btrfs_bio *bbio;
3564 bool ret;
3565
3566 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3567 return 0;
3568
3569 /*
3570 * We could have had EXTENT_BUFFER_UPTODATE cleared by the write
3571 * operation, which could potentially still be in flight. In this case
3572 * we simply want to return an error.
3573 */
3574 if (unlikely(test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)))
3575 return -EIO;
3576
3577 /* Someone else is already reading the buffer, just wait for it. */
3578 if (test_and_set_bit(EXTENT_BUFFER_READING, &eb->bflags))
3579 goto done;
3580
3581 /*
3582 * Between the initial test_bit(EXTENT_BUFFER_UPTODATE) and the above
3583 * test_and_set_bit(EXTENT_BUFFER_READING), someone else could have
3584 * started and finished reading the same eb. In this case, UPTODATE
3585 * will now be set, and we shouldn't read it in again.
3586 */
3587 if (unlikely(test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))) {
3588 clear_extent_buffer_reading(eb);
3589 return 0;
3590 }
3591
3592 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
3593 eb->read_mirror = 0;
3594 check_buffer_tree_ref(eb);
3595 atomic_inc(&eb->refs);
3596
3597 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
3598 REQ_OP_READ | REQ_META, eb->fs_info,
3599 end_bbio_meta_read, eb);
3600 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
3601 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
3602 bbio->file_offset = eb->start;
3603 memcpy(&bbio->parent_check, check, sizeof(*check));
3604 if (eb->fs_info->nodesize < PAGE_SIZE) {
3605 ret = bio_add_folio(&bbio->bio, eb->folios[0], eb->len,
3606 eb->start - folio_pos(eb->folios[0]));
3607 ASSERT(ret);
3608 } else {
3609 int num_folios = num_extent_folios(eb);
3610
3611 for (int i = 0; i < num_folios; i++) {
3612 struct folio *folio = eb->folios[i];
3613
3614 ret = bio_add_folio(&bbio->bio, folio, eb->folio_size, 0);
3615 ASSERT(ret);
3616 }
3617 }
3618 btrfs_submit_bio(bbio, mirror_num);
3619
3620 done:
3621 if (wait == WAIT_COMPLETE) {
3622 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_READING, TASK_UNINTERRUPTIBLE);
3623 if (!test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3624 return -EIO;
3625 }
3626
3627 return 0;
3628 }
3629
3630 static bool report_eb_range(const struct extent_buffer *eb, unsigned long start,
3631 unsigned long len)
3632 {
3633 btrfs_warn(eb->fs_info,
3634 "access to eb bytenr %llu len %u out of range start %lu len %lu",
3635 eb->start, eb->len, start, len);
3636 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
3637
3638 return true;
3639 }
3640
3641 /*
3642 * Check if the [start, start + len) range is valid before reading/writing
3643 * the eb.
3644 * NOTE: @start and @len are offset inside the eb, not logical address.
3645 *
3646 * Caller should not touch the dst/src memory if this function returns error.
3647 */
3648 static inline int check_eb_range(const struct extent_buffer *eb,
3649 unsigned long start, unsigned long len)
3650 {
3651 unsigned long offset;
3652
3653 /* start, start + len should not go beyond eb->len nor overflow */
3654 if (unlikely(check_add_overflow(start, len, &offset) || offset > eb->len))
3655 return report_eb_range(eb, start, len);
3656
3657 return false;
3658 }
3659
3660 void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
3661 unsigned long start, unsigned long len)
3662 {
3663 const int unit_size = eb->folio_size;
3664 size_t cur;
3665 size_t offset;
3666 char *dst = (char *)dstv;
3667 unsigned long i = get_eb_folio_index(eb, start);
3668
3669 if (check_eb_range(eb, start, len)) {
3670 /*
3671 * Invalid range hit, reset the memory, so callers won't get
3672 * some random garbage for their uninitialized memory.
3673 */
3674 memset(dstv, 0, len);
3675 return;
3676 }
3677
3678 if (eb->addr) {
3679 memcpy(dstv, eb->addr + start, len);
3680 return;
3681 }
3682
3683 offset = get_eb_offset_in_folio(eb, start);
3684
3685 while (len > 0) {
3686 char *kaddr;
3687
3688 cur = min(len, unit_size - offset);
3689 kaddr = folio_address(eb->folios[i]);
3690 memcpy(dst, kaddr + offset, cur);
3691
3692 dst += cur;
3693 len -= cur;
3694 offset = 0;
3695 i++;
3696 }
3697 }
3698
3699 int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
3700 void __user *dstv,
3701 unsigned long start, unsigned long len)
3702 {
3703 const int unit_size = eb->folio_size;
3704 size_t cur;
3705 size_t offset;
3706 char __user *dst = (char __user *)dstv;
3707 unsigned long i = get_eb_folio_index(eb, start);
3708 int ret = 0;
3709
3710 WARN_ON(start > eb->len);
3711 WARN_ON(start + len > eb->start + eb->len);
3712
3713 if (eb->addr) {
3714 if (copy_to_user_nofault(dstv, eb->addr + start, len))
3715 ret = -EFAULT;
3716 return ret;
3717 }
3718
3719 offset = get_eb_offset_in_folio(eb, start);
3720
3721 while (len > 0) {
3722 char *kaddr;
3723
3724 cur = min(len, unit_size - offset);
3725 kaddr = folio_address(eb->folios[i]);
3726 if (copy_to_user_nofault(dst, kaddr + offset, cur)) {
3727 ret = -EFAULT;
3728 break;
3729 }
3730
3731 dst += cur;
3732 len -= cur;
3733 offset = 0;
3734 i++;
3735 }
3736
3737 return ret;
3738 }
3739
3740 int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
3741 unsigned long start, unsigned long len)
3742 {
3743 const int unit_size = eb->folio_size;
3744 size_t cur;
3745 size_t offset;
3746 char *kaddr;
3747 char *ptr = (char *)ptrv;
3748 unsigned long i = get_eb_folio_index(eb, start);
3749 int ret = 0;
3750
3751 if (check_eb_range(eb, start, len))
3752 return -EINVAL;
3753
3754 if (eb->addr)
3755 return memcmp(ptrv, eb->addr + start, len);
3756
3757 offset = get_eb_offset_in_folio(eb, start);
3758
3759 while (len > 0) {
3760 cur = min(len, unit_size - offset);
3761 kaddr = folio_address(eb->folios[i]);
3762 ret = memcmp(ptr, kaddr + offset, cur);
3763 if (ret)
3764 break;
3765
3766 ptr += cur;
3767 len -= cur;
3768 offset = 0;
3769 i++;
3770 }
3771 return ret;
3772 }
3773
3774 /*
3775 * Check that the extent buffer is uptodate.
3776 *
3777 * For regular sector size == PAGE_SIZE case, check if @page is uptodate.
3778 * For subpage case, check if the range covered by the eb has EXTENT_UPTODATE.
3779 */
3780 static void assert_eb_folio_uptodate(const struct extent_buffer *eb, int i)
3781 {
3782 struct btrfs_fs_info *fs_info = eb->fs_info;
3783 struct folio *folio = eb->folios[i];
3784
3785 ASSERT(folio);
3786
3787 /*
3788 * If we are using the commit root we could potentially clear a page
3789 * Uptodate while we're using the extent buffer that we've previously
3790 * looked up. We don't want to complain in this case, as the page was
3791 * valid before, we just didn't write it out. Instead we want to catch
3792 * the case where we didn't actually read the block properly, which
3793 * would have !PageUptodate and !EXTENT_BUFFER_WRITE_ERR.
3794 */
3795 if (test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
3796 return;
3797
3798 if (fs_info->nodesize < PAGE_SIZE) {
3799 folio = eb->folios[0];
3800 ASSERT(i == 0);
3801 if (WARN_ON(!btrfs_subpage_test_uptodate(fs_info, folio,
3802 eb->start, eb->len)))
3803 btrfs_subpage_dump_bitmap(fs_info, folio, eb->start, eb->len);
3804 } else {
3805 WARN_ON(!folio_test_uptodate(folio));
3806 }
3807 }
3808
3809 static void __write_extent_buffer(const struct extent_buffer *eb,
3810 const void *srcv, unsigned long start,
3811 unsigned long len, bool use_memmove)
3812 {
3813 const int unit_size = eb->folio_size;
3814 size_t cur;
3815 size_t offset;
3816 char *kaddr;
3817 const char *src = (const char *)srcv;
3818 unsigned long i = get_eb_folio_index(eb, start);
3819 /* For unmapped (dummy) ebs, no need to check their uptodate status. */
3820 const bool check_uptodate = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3821
3822 if (check_eb_range(eb, start, len))
3823 return;
3824
3825 if (eb->addr) {
3826 if (use_memmove)
3827 memmove(eb->addr + start, srcv, len);
3828 else
3829 memcpy(eb->addr + start, srcv, len);
3830 return;
3831 }
3832
3833 offset = get_eb_offset_in_folio(eb, start);
3834
3835 while (len > 0) {
3836 if (check_uptodate)
3837 assert_eb_folio_uptodate(eb, i);
3838
3839 cur = min(len, unit_size - offset);
3840 kaddr = folio_address(eb->folios[i]);
3841 if (use_memmove)
3842 memmove(kaddr + offset, src, cur);
3843 else
3844 memcpy(kaddr + offset, src, cur);
3845
3846 src += cur;
3847 len -= cur;
3848 offset = 0;
3849 i++;
3850 }
3851 }
3852
3853 void write_extent_buffer(const struct extent_buffer *eb, const void *srcv,
3854 unsigned long start, unsigned long len)
3855 {
3856 return __write_extent_buffer(eb, srcv, start, len, false);
3857 }
3858
3859 static void memset_extent_buffer(const struct extent_buffer *eb, int c,
3860 unsigned long start, unsigned long len)
3861 {
3862 const int unit_size = eb->folio_size;
3863 unsigned long cur = start;
3864
3865 if (eb->addr) {
3866 memset(eb->addr + start, c, len);
3867 return;
3868 }
3869
3870 while (cur < start + len) {
3871 unsigned long index = get_eb_folio_index(eb, cur);
3872 unsigned int offset = get_eb_offset_in_folio(eb, cur);
3873 unsigned int cur_len = min(start + len - cur, unit_size - offset);
3874
3875 assert_eb_folio_uptodate(eb, index);
3876 memset(folio_address(eb->folios[index]) + offset, c, cur_len);
3877
3878 cur += cur_len;
3879 }
3880 }
3881
3882 void memzero_extent_buffer(const struct extent_buffer *eb, unsigned long start,
3883 unsigned long len)
3884 {
3885 if (check_eb_range(eb, start, len))
3886 return;
3887 return memset_extent_buffer(eb, 0, start, len);
3888 }
3889
3890 void copy_extent_buffer_full(const struct extent_buffer *dst,
3891 const struct extent_buffer *src)
3892 {
3893 const int unit_size = src->folio_size;
3894 unsigned long cur = 0;
3895
3896 ASSERT(dst->len == src->len);
3897
3898 while (cur < src->len) {
3899 unsigned long index = get_eb_folio_index(src, cur);
3900 unsigned long offset = get_eb_offset_in_folio(src, cur);
3901 unsigned long cur_len = min(src->len, unit_size - offset);
3902 void *addr = folio_address(src->folios[index]) + offset;
3903
3904 write_extent_buffer(dst, addr, cur, cur_len);
3905
3906 cur += cur_len;
3907 }
3908 }
3909
3910 void copy_extent_buffer(const struct extent_buffer *dst,
3911 const struct extent_buffer *src,
3912 unsigned long dst_offset, unsigned long src_offset,
3913 unsigned long len)
3914 {
3915 const int unit_size = dst->folio_size;
3916 u64 dst_len = dst->len;
3917 size_t cur;
3918 size_t offset;
3919 char *kaddr;
3920 unsigned long i = get_eb_folio_index(dst, dst_offset);
3921
3922 if (check_eb_range(dst, dst_offset, len) ||
3923 check_eb_range(src, src_offset, len))
3924 return;
3925
3926 WARN_ON(src->len != dst_len);
3927
3928 offset = get_eb_offset_in_folio(dst, dst_offset);
3929
3930 while (len > 0) {
3931 assert_eb_folio_uptodate(dst, i);
3932
3933 cur = min(len, (unsigned long)(unit_size - offset));
3934
3935 kaddr = folio_address(dst->folios[i]);
3936 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3937
3938 src_offset += cur;
3939 len -= cur;
3940 offset = 0;
3941 i++;
3942 }
3943 }
3944
3945 /*
3946 * Calculate the folio and offset of the byte containing the given bit number.
3947 *
3948 * @eb: the extent buffer
3949 * @start: offset of the bitmap item in the extent buffer
3950 * @nr: bit number
3951 * @folio_index: return index of the folio in the extent buffer that contains
3952 * the given bit number
3953 * @folio_offset: return offset into the folio given by folio_index
3954 *
3955 * This helper hides the ugliness of finding the byte in an extent buffer which
3956 * contains a given bit.
3957 */
3958 static inline void eb_bitmap_offset(const struct extent_buffer *eb,
3959 unsigned long start, unsigned long nr,
3960 unsigned long *folio_index,
3961 size_t *folio_offset)
3962 {
3963 size_t byte_offset = BIT_BYTE(nr);
3964 size_t offset;
3965
3966 /*
3967 * The byte we want is the offset of the extent buffer + the offset of
3968 * the bitmap item in the extent buffer + the offset of the byte in the
3969 * bitmap item.
3970 */
3971 offset = start + offset_in_eb_folio(eb, eb->start) + byte_offset;
3972
3973 *folio_index = offset >> eb->folio_shift;
3974 *folio_offset = offset_in_eb_folio(eb, offset);
3975 }
3976
3977 /*
3978 * Determine whether a bit in a bitmap item is set.
3979 *
3980 * @eb: the extent buffer
3981 * @start: offset of the bitmap item in the extent buffer
3982 * @nr: bit number to test
3983 */
3984 int extent_buffer_test_bit(const struct extent_buffer *eb, unsigned long start,
3985 unsigned long nr)
3986 {
3987 unsigned long i;
3988 size_t offset;
3989 u8 *kaddr;
3990
3991 eb_bitmap_offset(eb, start, nr, &i, &offset);
3992 assert_eb_folio_uptodate(eb, i);
3993 kaddr = folio_address(eb->folios[i]);
3994 return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
3995 }
3996
3997 static u8 *extent_buffer_get_byte(const struct extent_buffer *eb, unsigned long bytenr)
3998 {
3999 unsigned long index = get_eb_folio_index(eb, bytenr);
4000
4001 if (check_eb_range(eb, bytenr, 1))
4002 return NULL;
4003 return folio_address(eb->folios[index]) + get_eb_offset_in_folio(eb, bytenr);
4004 }
4005
4006 /*
4007 * Set an area of a bitmap to 1.
4008 *
4009 * @eb: the extent buffer
4010 * @start: offset of the bitmap item in the extent buffer
4011 * @pos: bit number of the first bit
4012 * @len: number of bits to set
4013 */
4014 void extent_buffer_bitmap_set(const struct extent_buffer *eb, unsigned long start,
4015 unsigned long pos, unsigned long len)
4016 {
4017 unsigned int first_byte = start + BIT_BYTE(pos);
4018 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4019 const bool same_byte = (first_byte == last_byte);
4020 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4021 u8 *kaddr;
4022
4023 if (same_byte)
4024 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4025
4026 /* Handle the first byte. */
4027 kaddr = extent_buffer_get_byte(eb, first_byte);
4028 *kaddr |= mask;
4029 if (same_byte)
4030 return;
4031
4032 /* Handle the byte aligned part. */
4033 ASSERT(first_byte + 1 <= last_byte);
4034 memset_extent_buffer(eb, 0xff, first_byte + 1, last_byte - first_byte - 1);
4035
4036 /* Handle the last byte. */
4037 kaddr = extent_buffer_get_byte(eb, last_byte);
4038 *kaddr |= BITMAP_LAST_BYTE_MASK(pos + len);
4039 }
4040
4041
4042 /*
4043 * Clear an area of a bitmap.
4044 *
4045 * @eb: the extent buffer
4046 * @start: offset of the bitmap item in the extent buffer
4047 * @pos: bit number of the first bit
4048 * @len: number of bits to clear
4049 */
4050 void extent_buffer_bitmap_clear(const struct extent_buffer *eb,
4051 unsigned long start, unsigned long pos,
4052 unsigned long len)
4053 {
4054 unsigned int first_byte = start + BIT_BYTE(pos);
4055 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4056 const bool same_byte = (first_byte == last_byte);
4057 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4058 u8 *kaddr;
4059
4060 if (same_byte)
4061 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4062
4063 /* Handle the first byte. */
4064 kaddr = extent_buffer_get_byte(eb, first_byte);
4065 *kaddr &= ~mask;
4066 if (same_byte)
4067 return;
4068
4069 /* Handle the byte aligned part. */
4070 ASSERT(first_byte + 1 <= last_byte);
4071 memset_extent_buffer(eb, 0, first_byte + 1, last_byte - first_byte - 1);
4072
4073 /* Handle the last byte. */
4074 kaddr = extent_buffer_get_byte(eb, last_byte);
4075 *kaddr &= ~BITMAP_LAST_BYTE_MASK(pos + len);
4076 }
4077
4078 static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4079 {
4080 unsigned long distance = (src > dst) ? src - dst : dst - src;
4081 return distance < len;
4082 }
4083
4084 void memcpy_extent_buffer(const struct extent_buffer *dst,
4085 unsigned long dst_offset, unsigned long src_offset,
4086 unsigned long len)
4087 {
4088 const int unit_size = dst->folio_size;
4089 unsigned long cur_off = 0;
4090
4091 if (check_eb_range(dst, dst_offset, len) ||
4092 check_eb_range(dst, src_offset, len))
4093 return;
4094
4095 if (dst->addr) {
4096 const bool use_memmove = areas_overlap(src_offset, dst_offset, len);
4097
4098 if (use_memmove)
4099 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4100 else
4101 memcpy(dst->addr + dst_offset, dst->addr + src_offset, len);
4102 return;
4103 }
4104
4105 while (cur_off < len) {
4106 unsigned long cur_src = cur_off + src_offset;
4107 unsigned long folio_index = get_eb_folio_index(dst, cur_src);
4108 unsigned long folio_off = get_eb_offset_in_folio(dst, cur_src);
4109 unsigned long cur_len = min(src_offset + len - cur_src,
4110 unit_size - folio_off);
4111 void *src_addr = folio_address(dst->folios[folio_index]) + folio_off;
4112 const bool use_memmove = areas_overlap(src_offset + cur_off,
4113 dst_offset + cur_off, cur_len);
4114
4115 __write_extent_buffer(dst, src_addr, dst_offset + cur_off, cur_len,
4116 use_memmove);
4117 cur_off += cur_len;
4118 }
4119 }
4120
4121 void memmove_extent_buffer(const struct extent_buffer *dst,
4122 unsigned long dst_offset, unsigned long src_offset,
4123 unsigned long len)
4124 {
4125 unsigned long dst_end = dst_offset + len - 1;
4126 unsigned long src_end = src_offset + len - 1;
4127
4128 if (check_eb_range(dst, dst_offset, len) ||
4129 check_eb_range(dst, src_offset, len))
4130 return;
4131
4132 if (dst_offset < src_offset) {
4133 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4134 return;
4135 }
4136
4137 if (dst->addr) {
4138 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4139 return;
4140 }
4141
4142 while (len > 0) {
4143 unsigned long src_i;
4144 size_t cur;
4145 size_t dst_off_in_folio;
4146 size_t src_off_in_folio;
4147 void *src_addr;
4148 bool use_memmove;
4149
4150 src_i = get_eb_folio_index(dst, src_end);
4151
4152 dst_off_in_folio = get_eb_offset_in_folio(dst, dst_end);
4153 src_off_in_folio = get_eb_offset_in_folio(dst, src_end);
4154
4155 cur = min_t(unsigned long, len, src_off_in_folio + 1);
4156 cur = min(cur, dst_off_in_folio + 1);
4157
4158 src_addr = folio_address(dst->folios[src_i]) + src_off_in_folio -
4159 cur + 1;
4160 use_memmove = areas_overlap(src_end - cur + 1, dst_end - cur + 1,
4161 cur);
4162
4163 __write_extent_buffer(dst, src_addr, dst_end - cur + 1, cur,
4164 use_memmove);
4165
4166 dst_end -= cur;
4167 src_end -= cur;
4168 len -= cur;
4169 }
4170 }
4171
4172 #define GANG_LOOKUP_SIZE 16
4173 static struct extent_buffer *get_next_extent_buffer(
4174 const struct btrfs_fs_info *fs_info, struct page *page, u64 bytenr)
4175 {
4176 struct extent_buffer *gang[GANG_LOOKUP_SIZE];
4177 struct extent_buffer *found = NULL;
4178 u64 page_start = page_offset(page);
4179 u64 cur = page_start;
4180
4181 ASSERT(in_range(bytenr, page_start, PAGE_SIZE));
4182 lockdep_assert_held(&fs_info->buffer_lock);
4183
4184 while (cur < page_start + PAGE_SIZE) {
4185 int ret;
4186 int i;
4187
4188 ret = radix_tree_gang_lookup(&fs_info->buffer_radix,
4189 (void **)gang, cur >> fs_info->sectorsize_bits,
4190 min_t(unsigned int, GANG_LOOKUP_SIZE,
4191 PAGE_SIZE / fs_info->nodesize));
4192 if (ret == 0)
4193 goto out;
4194 for (i = 0; i < ret; i++) {
4195 /* Already beyond page end */
4196 if (gang[i]->start >= page_start + PAGE_SIZE)
4197 goto out;
4198 /* Found one */
4199 if (gang[i]->start >= bytenr) {
4200 found = gang[i];
4201 goto out;
4202 }
4203 }
4204 cur = gang[ret - 1]->start + gang[ret - 1]->len;
4205 }
4206 out:
4207 return found;
4208 }
4209
4210 static int try_release_subpage_extent_buffer(struct page *page)
4211 {
4212 struct btrfs_fs_info *fs_info = page_to_fs_info(page);
4213 u64 cur = page_offset(page);
4214 const u64 end = page_offset(page) + PAGE_SIZE;
4215 int ret;
4216
4217 while (cur < end) {
4218 struct extent_buffer *eb = NULL;
4219
4220 /*
4221 * Unlike try_release_extent_buffer() which uses folio private
4222 * to grab buffer, for subpage case we rely on radix tree, thus
4223 * we need to ensure radix tree consistency.
4224 *
4225 * We also want an atomic snapshot of the radix tree, thus go
4226 * with spinlock rather than RCU.
4227 */
4228 spin_lock(&fs_info->buffer_lock);
4229 eb = get_next_extent_buffer(fs_info, page, cur);
4230 if (!eb) {
4231 /* No more eb in the page range after or at cur */
4232 spin_unlock(&fs_info->buffer_lock);
4233 break;
4234 }
4235 cur = eb->start + eb->len;
4236
4237 /*
4238 * The same as try_release_extent_buffer(), to ensure the eb
4239 * won't disappear out from under us.
4240 */
4241 spin_lock(&eb->refs_lock);
4242 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4243 spin_unlock(&eb->refs_lock);
4244 spin_unlock(&fs_info->buffer_lock);
4245 break;
4246 }
4247 spin_unlock(&fs_info->buffer_lock);
4248
4249 /*
4250 * If tree ref isn't set then we know the ref on this eb is a
4251 * real ref, so just return, this eb will likely be freed soon
4252 * anyway.
4253 */
4254 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4255 spin_unlock(&eb->refs_lock);
4256 break;
4257 }
4258
4259 /*
4260 * Here we don't care about the return value, we will always
4261 * check the folio private at the end. And
4262 * release_extent_buffer() will release the refs_lock.
4263 */
4264 release_extent_buffer(eb);
4265 }
4266 /*
4267 * Finally to check if we have cleared folio private, as if we have
4268 * released all ebs in the page, the folio private should be cleared now.
4269 */
4270 spin_lock(&page->mapping->i_private_lock);
4271 if (!folio_test_private(page_folio(page)))
4272 ret = 1;
4273 else
4274 ret = 0;
4275 spin_unlock(&page->mapping->i_private_lock);
4276 return ret;
4277
4278 }
4279
4280 int try_release_extent_buffer(struct page *page)
4281 {
4282 struct folio *folio = page_folio(page);
4283 struct extent_buffer *eb;
4284
4285 if (page_to_fs_info(page)->nodesize < PAGE_SIZE)
4286 return try_release_subpage_extent_buffer(page);
4287
4288 /*
4289 * We need to make sure nobody is changing folio private, as we rely on
4290 * folio private as the pointer to extent buffer.
4291 */
4292 spin_lock(&page->mapping->i_private_lock);
4293 if (!folio_test_private(folio)) {
4294 spin_unlock(&page->mapping->i_private_lock);
4295 return 1;
4296 }
4297
4298 eb = folio_get_private(folio);
4299 BUG_ON(!eb);
4300
4301 /*
4302 * This is a little awful but should be ok, we need to make sure that
4303 * the eb doesn't disappear out from under us while we're looking at
4304 * this page.
4305 */
4306 spin_lock(&eb->refs_lock);
4307 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4308 spin_unlock(&eb->refs_lock);
4309 spin_unlock(&page->mapping->i_private_lock);
4310 return 0;
4311 }
4312 spin_unlock(&page->mapping->i_private_lock);
4313
4314 /*
4315 * If tree ref isn't set then we know the ref on this eb is a real ref,
4316 * so just return, this page will likely be freed soon anyway.
4317 */
4318 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4319 spin_unlock(&eb->refs_lock);
4320 return 0;
4321 }
4322
4323 return release_extent_buffer(eb);
4324 }
4325
4326 /*
4327 * Attempt to readahead a child block.
4328 *
4329 * @fs_info: the fs_info
4330 * @bytenr: bytenr to read
4331 * @owner_root: objectid of the root that owns this eb
4332 * @gen: generation for the uptodate check, can be 0
4333 * @level: level for the eb
4334 *
4335 * Attempt to readahead a tree block at @bytenr. If @gen is 0 then we do a
4336 * normal uptodate check of the eb, without checking the generation. If we have
4337 * to read the block we will not block on anything.
4338 */
4339 void btrfs_readahead_tree_block(struct btrfs_fs_info *fs_info,
4340 u64 bytenr, u64 owner_root, u64 gen, int level)
4341 {
4342 struct btrfs_tree_parent_check check = {
4343 .has_first_key = 0,
4344 .level = level,
4345 .transid = gen
4346 };
4347 struct extent_buffer *eb;
4348 int ret;
4349
4350 eb = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
4351 if (IS_ERR(eb))
4352 return;
4353
4354 if (btrfs_buffer_uptodate(eb, gen, 1)) {
4355 free_extent_buffer(eb);
4356 return;
4357 }
4358
4359 ret = read_extent_buffer_pages(eb, WAIT_NONE, 0, &check);
4360 if (ret < 0)
4361 free_extent_buffer_stale(eb);
4362 else
4363 free_extent_buffer(eb);
4364 }
4365
4366 /*
4367 * Readahead a node's child block.
4368 *
4369 * @node: parent node we're reading from
4370 * @slot: slot in the parent node for the child we want to read
4371 *
4372 * A helper for btrfs_readahead_tree_block, we simply read the bytenr pointed at
4373 * the slot in the node provided.
4374 */
4375 void btrfs_readahead_node_child(struct extent_buffer *node, int slot)
4376 {
4377 btrfs_readahead_tree_block(node->fs_info,
4378 btrfs_node_blockptr(node, slot),
4379 btrfs_header_owner(node),
4380 btrfs_node_ptr_generation(node, slot),
4381 btrfs_header_level(node) - 1);
4382 }
4383
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.