Loading...
1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/bitops.h>
4#include <linux/slab.h>
5#include <linux/bio.h>
6#include <linux/mm.h>
7#include <linux/pagemap.h>
8#include <linux/page-flags.h>
9#include <linux/sched/mm.h>
10#include <linux/spinlock.h>
11#include <linux/blkdev.h>
12#include <linux/swap.h>
13#include <linux/writeback.h>
14#include <linux/pagevec.h>
15#include <linux/prefetch.h>
16#include <linux/fsverity.h>
17#include "misc.h"
18#include "extent_io.h"
19#include "extent-io-tree.h"
20#include "extent_map.h"
21#include "ctree.h"
22#include "btrfs_inode.h"
23#include "bio.h"
24#include "locking.h"
25#include "rcu-string.h"
26#include "backref.h"
27#include "disk-io.h"
28#include "subpage.h"
29#include "zoned.h"
30#include "block-group.h"
31#include "compression.h"
32#include "fs.h"
33#include "accessors.h"
34#include "file-item.h"
35#include "file.h"
36#include "dev-replace.h"
37#include "super.h"
38#include "transaction.h"
39
40static struct kmem_cache *extent_buffer_cache;
41
42#ifdef CONFIG_BTRFS_DEBUG
43static inline void btrfs_leak_debug_add_eb(struct extent_buffer *eb)
44{
45 struct btrfs_fs_info *fs_info = eb->fs_info;
46 unsigned long flags;
47
48 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
49 list_add(&eb->leak_list, &fs_info->allocated_ebs);
50 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
51}
52
53static inline void btrfs_leak_debug_del_eb(struct extent_buffer *eb)
54{
55 struct btrfs_fs_info *fs_info = eb->fs_info;
56 unsigned long flags;
57
58 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
59 list_del(&eb->leak_list);
60 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
61}
62
63void btrfs_extent_buffer_leak_debug_check(struct btrfs_fs_info *fs_info)
64{
65 struct extent_buffer *eb;
66 unsigned long flags;
67
68 /*
69 * If we didn't get into open_ctree our allocated_ebs will not be
70 * initialized, so just skip this.
71 */
72 if (!fs_info->allocated_ebs.next)
73 return;
74
75 WARN_ON(!list_empty(&fs_info->allocated_ebs));
76 spin_lock_irqsave(&fs_info->eb_leak_lock, flags);
77 while (!list_empty(&fs_info->allocated_ebs)) {
78 eb = list_first_entry(&fs_info->allocated_ebs,
79 struct extent_buffer, leak_list);
80 pr_err(
81 "BTRFS: buffer leak start %llu len %lu refs %d bflags %lu owner %llu\n",
82 eb->start, eb->len, atomic_read(&eb->refs), eb->bflags,
83 btrfs_header_owner(eb));
84 list_del(&eb->leak_list);
85 kmem_cache_free(extent_buffer_cache, eb);
86 }
87 spin_unlock_irqrestore(&fs_info->eb_leak_lock, flags);
88}
89#else
90#define btrfs_leak_debug_add_eb(eb) do {} while (0)
91#define btrfs_leak_debug_del_eb(eb) do {} while (0)
92#endif
93
94/*
95 * Structure to record info about the bio being assembled, and other info like
96 * how many bytes are there before stripe/ordered extent boundary.
97 */
98struct btrfs_bio_ctrl {
99 struct btrfs_bio *bbio;
100 enum btrfs_compression_type compress_type;
101 u32 len_to_oe_boundary;
102 blk_opf_t opf;
103 btrfs_bio_end_io_t end_io_func;
104 struct writeback_control *wbc;
105};
106
107static void submit_one_bio(struct btrfs_bio_ctrl *bio_ctrl)
108{
109 struct btrfs_bio *bbio = bio_ctrl->bbio;
110
111 if (!bbio)
112 return;
113
114 /* Caller should ensure the bio has at least some range added */
115 ASSERT(bbio->bio.bi_iter.bi_size);
116
117 if (btrfs_op(&bbio->bio) == BTRFS_MAP_READ &&
118 bio_ctrl->compress_type != BTRFS_COMPRESS_NONE)
119 btrfs_submit_compressed_read(bbio);
120 else
121 btrfs_submit_bio(bbio, 0);
122
123 /* The bbio is owned by the end_io handler now */
124 bio_ctrl->bbio = NULL;
125}
126
127/*
128 * Submit or fail the current bio in the bio_ctrl structure.
129 */
130static void submit_write_bio(struct btrfs_bio_ctrl *bio_ctrl, int ret)
131{
132 struct btrfs_bio *bbio = bio_ctrl->bbio;
133
134 if (!bbio)
135 return;
136
137 if (ret) {
138 ASSERT(ret < 0);
139 btrfs_bio_end_io(bbio, errno_to_blk_status(ret));
140 /* The bio is owned by the end_io handler now */
141 bio_ctrl->bbio = NULL;
142 } else {
143 submit_one_bio(bio_ctrl);
144 }
145}
146
147int __init extent_buffer_init_cachep(void)
148{
149 extent_buffer_cache = kmem_cache_create("btrfs_extent_buffer",
150 sizeof(struct extent_buffer), 0,
151 SLAB_MEM_SPREAD, NULL);
152 if (!extent_buffer_cache)
153 return -ENOMEM;
154
155 return 0;
156}
157
158void __cold extent_buffer_free_cachep(void)
159{
160 /*
161 * Make sure all delayed rcu free are flushed before we
162 * destroy caches.
163 */
164 rcu_barrier();
165 kmem_cache_destroy(extent_buffer_cache);
166}
167
168void extent_range_clear_dirty_for_io(struct inode *inode, u64 start, u64 end)
169{
170 unsigned long index = start >> PAGE_SHIFT;
171 unsigned long end_index = end >> PAGE_SHIFT;
172 struct page *page;
173
174 while (index <= end_index) {
175 page = find_get_page(inode->i_mapping, index);
176 BUG_ON(!page); /* Pages should be in the extent_io_tree */
177 clear_page_dirty_for_io(page);
178 put_page(page);
179 index++;
180 }
181}
182
183static void process_one_page(struct btrfs_fs_info *fs_info,
184 struct page *page, struct page *locked_page,
185 unsigned long page_ops, u64 start, u64 end)
186{
187 struct folio *folio = page_folio(page);
188 u32 len;
189
190 ASSERT(end + 1 - start != 0 && end + 1 - start < U32_MAX);
191 len = end + 1 - start;
192
193 if (page_ops & PAGE_SET_ORDERED)
194 btrfs_folio_clamp_set_ordered(fs_info, folio, start, len);
195 if (page_ops & PAGE_START_WRITEBACK) {
196 btrfs_folio_clamp_clear_dirty(fs_info, folio, start, len);
197 btrfs_folio_clamp_set_writeback(fs_info, folio, start, len);
198 }
199 if (page_ops & PAGE_END_WRITEBACK)
200 btrfs_folio_clamp_clear_writeback(fs_info, folio, start, len);
201
202 if (page != locked_page && (page_ops & PAGE_UNLOCK))
203 btrfs_folio_end_writer_lock(fs_info, folio, start, len);
204}
205
206static void __process_pages_contig(struct address_space *mapping,
207 struct page *locked_page, u64 start, u64 end,
208 unsigned long page_ops)
209{
210 struct btrfs_fs_info *fs_info = btrfs_sb(mapping->host->i_sb);
211 pgoff_t start_index = start >> PAGE_SHIFT;
212 pgoff_t end_index = end >> PAGE_SHIFT;
213 pgoff_t index = start_index;
214 struct folio_batch fbatch;
215 int i;
216
217 folio_batch_init(&fbatch);
218 while (index <= end_index) {
219 int found_folios;
220
221 found_folios = filemap_get_folios_contig(mapping, &index,
222 end_index, &fbatch);
223 for (i = 0; i < found_folios; i++) {
224 struct folio *folio = fbatch.folios[i];
225
226 process_one_page(fs_info, &folio->page, locked_page,
227 page_ops, start, end);
228 }
229 folio_batch_release(&fbatch);
230 cond_resched();
231 }
232}
233
234static noinline void __unlock_for_delalloc(struct inode *inode,
235 struct page *locked_page,
236 u64 start, u64 end)
237{
238 unsigned long index = start >> PAGE_SHIFT;
239 unsigned long end_index = end >> PAGE_SHIFT;
240
241 ASSERT(locked_page);
242 if (index == locked_page->index && end_index == index)
243 return;
244
245 __process_pages_contig(inode->i_mapping, locked_page, start, end,
246 PAGE_UNLOCK);
247}
248
249static noinline int lock_delalloc_pages(struct inode *inode,
250 struct page *locked_page,
251 u64 start,
252 u64 end)
253{
254 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
255 struct address_space *mapping = inode->i_mapping;
256 pgoff_t start_index = start >> PAGE_SHIFT;
257 pgoff_t end_index = end >> PAGE_SHIFT;
258 pgoff_t index = start_index;
259 u64 processed_end = start;
260 struct folio_batch fbatch;
261
262 if (index == locked_page->index && index == end_index)
263 return 0;
264
265 folio_batch_init(&fbatch);
266 while (index <= end_index) {
267 unsigned int found_folios, i;
268
269 found_folios = filemap_get_folios_contig(mapping, &index,
270 end_index, &fbatch);
271 if (found_folios == 0)
272 goto out;
273
274 for (i = 0; i < found_folios; i++) {
275 struct folio *folio = fbatch.folios[i];
276 struct page *page = folio_page(folio, 0);
277 u32 len = end + 1 - start;
278
279 if (page == locked_page)
280 continue;
281
282 if (btrfs_folio_start_writer_lock(fs_info, folio, start,
283 len))
284 goto out;
285
286 if (!PageDirty(page) || page->mapping != mapping) {
287 btrfs_folio_end_writer_lock(fs_info, folio, start,
288 len);
289 goto out;
290 }
291
292 processed_end = page_offset(page) + PAGE_SIZE - 1;
293 }
294 folio_batch_release(&fbatch);
295 cond_resched();
296 }
297
298 return 0;
299out:
300 folio_batch_release(&fbatch);
301 if (processed_end > start)
302 __unlock_for_delalloc(inode, locked_page, start, processed_end);
303 return -EAGAIN;
304}
305
306/*
307 * Find and lock a contiguous range of bytes in the file marked as delalloc, no
308 * more than @max_bytes.
309 *
310 * @start: The original start bytenr to search.
311 * Will store the extent range start bytenr.
312 * @end: The original end bytenr of the search range
313 * Will store the extent range end bytenr.
314 *
315 * Return true if we find a delalloc range which starts inside the original
316 * range, and @start/@end will store the delalloc range start/end.
317 *
318 * Return false if we can't find any delalloc range which starts inside the
319 * original range, and @start/@end will be the non-delalloc range start/end.
320 */
321EXPORT_FOR_TESTS
322noinline_for_stack bool find_lock_delalloc_range(struct inode *inode,
323 struct page *locked_page, u64 *start,
324 u64 *end)
325{
326 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
327 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
328 const u64 orig_start = *start;
329 const u64 orig_end = *end;
330 /* The sanity tests may not set a valid fs_info. */
331 u64 max_bytes = fs_info ? fs_info->max_extent_size : BTRFS_MAX_EXTENT_SIZE;
332 u64 delalloc_start;
333 u64 delalloc_end;
334 bool found;
335 struct extent_state *cached_state = NULL;
336 int ret;
337 int loops = 0;
338
339 /* Caller should pass a valid @end to indicate the search range end */
340 ASSERT(orig_end > orig_start);
341
342 /* The range should at least cover part of the page */
343 ASSERT(!(orig_start >= page_offset(locked_page) + PAGE_SIZE ||
344 orig_end <= page_offset(locked_page)));
345again:
346 /* step one, find a bunch of delalloc bytes starting at start */
347 delalloc_start = *start;
348 delalloc_end = 0;
349 found = btrfs_find_delalloc_range(tree, &delalloc_start, &delalloc_end,
350 max_bytes, &cached_state);
351 if (!found || delalloc_end <= *start || delalloc_start > orig_end) {
352 *start = delalloc_start;
353
354 /* @delalloc_end can be -1, never go beyond @orig_end */
355 *end = min(delalloc_end, orig_end);
356 free_extent_state(cached_state);
357 return false;
358 }
359
360 /*
361 * start comes from the offset of locked_page. We have to lock
362 * pages in order, so we can't process delalloc bytes before
363 * locked_page
364 */
365 if (delalloc_start < *start)
366 delalloc_start = *start;
367
368 /*
369 * make sure to limit the number of pages we try to lock down
370 */
371 if (delalloc_end + 1 - delalloc_start > max_bytes)
372 delalloc_end = delalloc_start + max_bytes - 1;
373
374 /* step two, lock all the pages after the page that has start */
375 ret = lock_delalloc_pages(inode, locked_page,
376 delalloc_start, delalloc_end);
377 ASSERT(!ret || ret == -EAGAIN);
378 if (ret == -EAGAIN) {
379 /* some of the pages are gone, lets avoid looping by
380 * shortening the size of the delalloc range we're searching
381 */
382 free_extent_state(cached_state);
383 cached_state = NULL;
384 if (!loops) {
385 max_bytes = PAGE_SIZE;
386 loops = 1;
387 goto again;
388 } else {
389 found = false;
390 goto out_failed;
391 }
392 }
393
394 /* step three, lock the state bits for the whole range */
395 lock_extent(tree, delalloc_start, delalloc_end, &cached_state);
396
397 /* then test to make sure it is all still delalloc */
398 ret = test_range_bit(tree, delalloc_start, delalloc_end,
399 EXTENT_DELALLOC, cached_state);
400 if (!ret) {
401 unlock_extent(tree, delalloc_start, delalloc_end,
402 &cached_state);
403 __unlock_for_delalloc(inode, locked_page,
404 delalloc_start, delalloc_end);
405 cond_resched();
406 goto again;
407 }
408 free_extent_state(cached_state);
409 *start = delalloc_start;
410 *end = delalloc_end;
411out_failed:
412 return found;
413}
414
415void extent_clear_unlock_delalloc(struct btrfs_inode *inode, u64 start, u64 end,
416 struct page *locked_page,
417 u32 clear_bits, unsigned long page_ops)
418{
419 clear_extent_bit(&inode->io_tree, start, end, clear_bits, NULL);
420
421 __process_pages_contig(inode->vfs_inode.i_mapping, locked_page,
422 start, end, page_ops);
423}
424
425static bool btrfs_verify_page(struct page *page, u64 start)
426{
427 if (!fsverity_active(page->mapping->host) ||
428 PageUptodate(page) ||
429 start >= i_size_read(page->mapping->host))
430 return true;
431 return fsverity_verify_page(page);
432}
433
434static void end_page_read(struct page *page, bool uptodate, u64 start, u32 len)
435{
436 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
437 struct folio *folio = page_folio(page);
438
439 ASSERT(page_offset(page) <= start &&
440 start + len <= page_offset(page) + PAGE_SIZE);
441
442 if (uptodate && btrfs_verify_page(page, start))
443 btrfs_folio_set_uptodate(fs_info, folio, start, len);
444 else
445 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
446
447 if (!btrfs_is_subpage(fs_info, page->mapping))
448 unlock_page(page);
449 else
450 btrfs_subpage_end_reader(fs_info, folio, start, len);
451}
452
453/*
454 * After a write IO is done, we need to:
455 *
456 * - clear the uptodate bits on error
457 * - clear the writeback bits in the extent tree for the range
458 * - filio_end_writeback() if there is no more pending io for the folio
459 *
460 * Scheduling is not allowed, so the extent state tree is expected
461 * to have one and only one object corresponding to this IO.
462 */
463static void end_bbio_data_write(struct btrfs_bio *bbio)
464{
465 struct bio *bio = &bbio->bio;
466 int error = blk_status_to_errno(bio->bi_status);
467 struct folio_iter fi;
468
469 ASSERT(!bio_flagged(bio, BIO_CLONED));
470 bio_for_each_folio_all(fi, bio) {
471 struct folio *folio = fi.folio;
472 struct inode *inode = folio->mapping->host;
473 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
474 const u32 sectorsize = fs_info->sectorsize;
475 u64 start = folio_pos(folio) + fi.offset;
476 u32 len = fi.length;
477
478 /* Only order 0 (single page) folios are allowed for data. */
479 ASSERT(folio_order(folio) == 0);
480
481 /* Our read/write should always be sector aligned. */
482 if (!IS_ALIGNED(fi.offset, sectorsize))
483 btrfs_err(fs_info,
484 "partial page write in btrfs with offset %zu and length %zu",
485 fi.offset, fi.length);
486 else if (!IS_ALIGNED(fi.length, sectorsize))
487 btrfs_info(fs_info,
488 "incomplete page write with offset %zu and length %zu",
489 fi.offset, fi.length);
490
491 btrfs_finish_ordered_extent(bbio->ordered,
492 folio_page(folio, 0), start, len, !error);
493 if (error)
494 mapping_set_error(folio->mapping, error);
495 btrfs_folio_clear_writeback(fs_info, folio, start, len);
496 }
497
498 bio_put(bio);
499}
500
501/*
502 * Record previously processed extent range
503 *
504 * For endio_readpage_release_extent() to handle a full extent range, reducing
505 * the extent io operations.
506 */
507struct processed_extent {
508 struct btrfs_inode *inode;
509 /* Start of the range in @inode */
510 u64 start;
511 /* End of the range in @inode */
512 u64 end;
513 bool uptodate;
514};
515
516/*
517 * Try to release processed extent range
518 *
519 * May not release the extent range right now if the current range is
520 * contiguous to processed extent.
521 *
522 * Will release processed extent when any of @inode, @uptodate, the range is
523 * no longer contiguous to the processed range.
524 *
525 * Passing @inode == NULL will force processed extent to be released.
526 */
527static void endio_readpage_release_extent(struct processed_extent *processed,
528 struct btrfs_inode *inode, u64 start, u64 end,
529 bool uptodate)
530{
531 struct extent_state *cached = NULL;
532 struct extent_io_tree *tree;
533
534 /* The first extent, initialize @processed */
535 if (!processed->inode)
536 goto update;
537
538 /*
539 * Contiguous to processed extent, just uptodate the end.
540 *
541 * Several things to notice:
542 *
543 * - bio can be merged as long as on-disk bytenr is contiguous
544 * This means we can have page belonging to other inodes, thus need to
545 * check if the inode still matches.
546 * - bvec can contain range beyond current page for multi-page bvec
547 * Thus we need to do processed->end + 1 >= start check
548 */
549 if (processed->inode == inode && processed->uptodate == uptodate &&
550 processed->end + 1 >= start && end >= processed->end) {
551 processed->end = end;
552 return;
553 }
554
555 tree = &processed->inode->io_tree;
556 /*
557 * Now we don't have range contiguous to the processed range, release
558 * the processed range now.
559 */
560 unlock_extent(tree, processed->start, processed->end, &cached);
561
562update:
563 /* Update processed to current range */
564 processed->inode = inode;
565 processed->start = start;
566 processed->end = end;
567 processed->uptodate = uptodate;
568}
569
570static void begin_page_read(struct btrfs_fs_info *fs_info, struct page *page)
571{
572 struct folio *folio = page_folio(page);
573
574 ASSERT(folio_test_locked(folio));
575 if (!btrfs_is_subpage(fs_info, folio->mapping))
576 return;
577
578 ASSERT(folio_test_private(folio));
579 btrfs_subpage_start_reader(fs_info, folio, page_offset(page), PAGE_SIZE);
580}
581
582/*
583 * After a data read IO is done, we need to:
584 *
585 * - clear the uptodate bits on error
586 * - set the uptodate bits if things worked
587 * - set the folio up to date if all extents in the tree are uptodate
588 * - clear the lock bit in the extent tree
589 * - unlock the folio if there are no other extents locked for it
590 *
591 * Scheduling is not allowed, so the extent state tree is expected
592 * to have one and only one object corresponding to this IO.
593 */
594static void end_bbio_data_read(struct btrfs_bio *bbio)
595{
596 struct bio *bio = &bbio->bio;
597 struct processed_extent processed = { 0 };
598 struct folio_iter fi;
599 /*
600 * The offset to the beginning of a bio, since one bio can never be
601 * larger than UINT_MAX, u32 here is enough.
602 */
603 u32 bio_offset = 0;
604
605 ASSERT(!bio_flagged(bio, BIO_CLONED));
606 bio_for_each_folio_all(fi, &bbio->bio) {
607 bool uptodate = !bio->bi_status;
608 struct folio *folio = fi.folio;
609 struct inode *inode = folio->mapping->host;
610 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
611 const u32 sectorsize = fs_info->sectorsize;
612 u64 start;
613 u64 end;
614 u32 len;
615
616 /* For now only order 0 folios are supported for data. */
617 ASSERT(folio_order(folio) == 0);
618 btrfs_debug(fs_info,
619 "%s: bi_sector=%llu, err=%d, mirror=%u",
620 __func__, bio->bi_iter.bi_sector, bio->bi_status,
621 bbio->mirror_num);
622
623 /*
624 * We always issue full-sector reads, but if some block in a
625 * folio fails to read, blk_update_request() will advance
626 * bv_offset and adjust bv_len to compensate. Print a warning
627 * for unaligned offsets, and an error if they don't add up to
628 * a full sector.
629 */
630 if (!IS_ALIGNED(fi.offset, sectorsize))
631 btrfs_err(fs_info,
632 "partial page read in btrfs with offset %zu and length %zu",
633 fi.offset, fi.length);
634 else if (!IS_ALIGNED(fi.offset + fi.length, sectorsize))
635 btrfs_info(fs_info,
636 "incomplete page read with offset %zu and length %zu",
637 fi.offset, fi.length);
638
639 start = folio_pos(folio) + fi.offset;
640 end = start + fi.length - 1;
641 len = fi.length;
642
643 if (likely(uptodate)) {
644 loff_t i_size = i_size_read(inode);
645 pgoff_t end_index = i_size >> folio_shift(folio);
646
647 /*
648 * Zero out the remaining part if this range straddles
649 * i_size.
650 *
651 * Here we should only zero the range inside the folio,
652 * not touch anything else.
653 *
654 * NOTE: i_size is exclusive while end is inclusive.
655 */
656 if (folio_index(folio) == end_index && i_size <= end) {
657 u32 zero_start = max(offset_in_folio(folio, i_size),
658 offset_in_folio(folio, start));
659 u32 zero_len = offset_in_folio(folio, end) + 1 -
660 zero_start;
661
662 folio_zero_range(folio, zero_start, zero_len);
663 }
664 }
665
666 /* Update page status and unlock. */
667 end_page_read(folio_page(folio, 0), uptodate, start, len);
668 endio_readpage_release_extent(&processed, BTRFS_I(inode),
669 start, end, uptodate);
670
671 ASSERT(bio_offset + len > bio_offset);
672 bio_offset += len;
673
674 }
675 /* Release the last extent */
676 endio_readpage_release_extent(&processed, NULL, 0, 0, false);
677 bio_put(bio);
678}
679
680/*
681 * Populate every free slot in a provided array with pages.
682 *
683 * @nr_pages: number of pages to allocate
684 * @page_array: the array to fill with pages; any existing non-null entries in
685 * the array will be skipped
686 * @extra_gfp: the extra GFP flags for the allocation.
687 *
688 * Return: 0 if all pages were able to be allocated;
689 * -ENOMEM otherwise, the partially allocated pages would be freed and
690 * the array slots zeroed
691 */
692int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array,
693 gfp_t extra_gfp)
694{
695 unsigned int allocated;
696
697 for (allocated = 0; allocated < nr_pages;) {
698 unsigned int last = allocated;
699
700 allocated = alloc_pages_bulk_array(GFP_NOFS | extra_gfp,
701 nr_pages, page_array);
702
703 if (allocated == nr_pages)
704 return 0;
705
706 /*
707 * During this iteration, no page could be allocated, even
708 * though alloc_pages_bulk_array() falls back to alloc_page()
709 * if it could not bulk-allocate. So we must be out of memory.
710 */
711 if (allocated == last) {
712 for (int i = 0; i < allocated; i++) {
713 __free_page(page_array[i]);
714 page_array[i] = NULL;
715 }
716 return -ENOMEM;
717 }
718
719 memalloc_retry_wait(GFP_NOFS);
720 }
721 return 0;
722}
723
724/*
725 * Populate needed folios for the extent buffer.
726 *
727 * For now, the folios populated are always in order 0 (aka, single page).
728 */
729static int alloc_eb_folio_array(struct extent_buffer *eb, gfp_t extra_gfp)
730{
731 struct page *page_array[INLINE_EXTENT_BUFFER_PAGES] = { 0 };
732 int num_pages = num_extent_pages(eb);
733 int ret;
734
735 ret = btrfs_alloc_page_array(num_pages, page_array, extra_gfp);
736 if (ret < 0)
737 return ret;
738
739 for (int i = 0; i < num_pages; i++)
740 eb->folios[i] = page_folio(page_array[i]);
741 return 0;
742}
743
744static bool btrfs_bio_is_contig(struct btrfs_bio_ctrl *bio_ctrl,
745 struct page *page, u64 disk_bytenr,
746 unsigned int pg_offset)
747{
748 struct bio *bio = &bio_ctrl->bbio->bio;
749 struct bio_vec *bvec = bio_last_bvec_all(bio);
750 const sector_t sector = disk_bytenr >> SECTOR_SHIFT;
751
752 if (bio_ctrl->compress_type != BTRFS_COMPRESS_NONE) {
753 /*
754 * For compression, all IO should have its logical bytenr set
755 * to the starting bytenr of the compressed extent.
756 */
757 return bio->bi_iter.bi_sector == sector;
758 }
759
760 /*
761 * The contig check requires the following conditions to be met:
762 *
763 * 1) The pages are belonging to the same inode
764 * This is implied by the call chain.
765 *
766 * 2) The range has adjacent logical bytenr
767 *
768 * 3) The range has adjacent file offset
769 * This is required for the usage of btrfs_bio->file_offset.
770 */
771 return bio_end_sector(bio) == sector &&
772 page_offset(bvec->bv_page) + bvec->bv_offset + bvec->bv_len ==
773 page_offset(page) + pg_offset;
774}
775
776static void alloc_new_bio(struct btrfs_inode *inode,
777 struct btrfs_bio_ctrl *bio_ctrl,
778 u64 disk_bytenr, u64 file_offset)
779{
780 struct btrfs_fs_info *fs_info = inode->root->fs_info;
781 struct btrfs_bio *bbio;
782
783 bbio = btrfs_bio_alloc(BIO_MAX_VECS, bio_ctrl->opf, fs_info,
784 bio_ctrl->end_io_func, NULL);
785 bbio->bio.bi_iter.bi_sector = disk_bytenr >> SECTOR_SHIFT;
786 bbio->inode = inode;
787 bbio->file_offset = file_offset;
788 bio_ctrl->bbio = bbio;
789 bio_ctrl->len_to_oe_boundary = U32_MAX;
790
791 /* Limit data write bios to the ordered boundary. */
792 if (bio_ctrl->wbc) {
793 struct btrfs_ordered_extent *ordered;
794
795 ordered = btrfs_lookup_ordered_extent(inode, file_offset);
796 if (ordered) {
797 bio_ctrl->len_to_oe_boundary = min_t(u32, U32_MAX,
798 ordered->file_offset +
799 ordered->disk_num_bytes - file_offset);
800 bbio->ordered = ordered;
801 }
802
803 /*
804 * Pick the last added device to support cgroup writeback. For
805 * multi-device file systems this means blk-cgroup policies have
806 * to always be set on the last added/replaced device.
807 * This is a bit odd but has been like that for a long time.
808 */
809 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
810 wbc_init_bio(bio_ctrl->wbc, &bbio->bio);
811 }
812}
813
814/*
815 * @disk_bytenr: logical bytenr where the write will be
816 * @page: page to add to the bio
817 * @size: portion of page that we want to write to
818 * @pg_offset: offset of the new bio or to check whether we are adding
819 * a contiguous page to the previous one
820 *
821 * The will either add the page into the existing @bio_ctrl->bbio, or allocate a
822 * new one in @bio_ctrl->bbio.
823 * The mirror number for this IO should already be initizlied in
824 * @bio_ctrl->mirror_num.
825 */
826static void submit_extent_page(struct btrfs_bio_ctrl *bio_ctrl,
827 u64 disk_bytenr, struct page *page,
828 size_t size, unsigned long pg_offset)
829{
830 struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
831
832 ASSERT(pg_offset + size <= PAGE_SIZE);
833 ASSERT(bio_ctrl->end_io_func);
834
835 if (bio_ctrl->bbio &&
836 !btrfs_bio_is_contig(bio_ctrl, page, disk_bytenr, pg_offset))
837 submit_one_bio(bio_ctrl);
838
839 do {
840 u32 len = size;
841
842 /* Allocate new bio if needed */
843 if (!bio_ctrl->bbio) {
844 alloc_new_bio(inode, bio_ctrl, disk_bytenr,
845 page_offset(page) + pg_offset);
846 }
847
848 /* Cap to the current ordered extent boundary if there is one. */
849 if (len > bio_ctrl->len_to_oe_boundary) {
850 ASSERT(bio_ctrl->compress_type == BTRFS_COMPRESS_NONE);
851 ASSERT(is_data_inode(&inode->vfs_inode));
852 len = bio_ctrl->len_to_oe_boundary;
853 }
854
855 if (bio_add_page(&bio_ctrl->bbio->bio, page, len, pg_offset) != len) {
856 /* bio full: move on to a new one */
857 submit_one_bio(bio_ctrl);
858 continue;
859 }
860
861 if (bio_ctrl->wbc)
862 wbc_account_cgroup_owner(bio_ctrl->wbc, page, len);
863
864 size -= len;
865 pg_offset += len;
866 disk_bytenr += len;
867
868 /*
869 * len_to_oe_boundary defaults to U32_MAX, which isn't page or
870 * sector aligned. alloc_new_bio() then sets it to the end of
871 * our ordered extent for writes into zoned devices.
872 *
873 * When len_to_oe_boundary is tracking an ordered extent, we
874 * trust the ordered extent code to align things properly, and
875 * the check above to cap our write to the ordered extent
876 * boundary is correct.
877 *
878 * When len_to_oe_boundary is U32_MAX, the cap above would
879 * result in a 4095 byte IO for the last page right before
880 * we hit the bio limit of UINT_MAX. bio_add_page() has all
881 * the checks required to make sure we don't overflow the bio,
882 * and we should just ignore len_to_oe_boundary completely
883 * unless we're using it to track an ordered extent.
884 *
885 * It's pretty hard to make a bio sized U32_MAX, but it can
886 * happen when the page cache is able to feed us contiguous
887 * pages for large extents.
888 */
889 if (bio_ctrl->len_to_oe_boundary != U32_MAX)
890 bio_ctrl->len_to_oe_boundary -= len;
891
892 /* Ordered extent boundary: move on to a new bio. */
893 if (bio_ctrl->len_to_oe_boundary == 0)
894 submit_one_bio(bio_ctrl);
895 } while (size);
896}
897
898static int attach_extent_buffer_folio(struct extent_buffer *eb,
899 struct folio *folio,
900 struct btrfs_subpage *prealloc)
901{
902 struct btrfs_fs_info *fs_info = eb->fs_info;
903 int ret = 0;
904
905 /*
906 * If the page is mapped to btree inode, we should hold the private
907 * lock to prevent race.
908 * For cloned or dummy extent buffers, their pages are not mapped and
909 * will not race with any other ebs.
910 */
911 if (folio->mapping)
912 lockdep_assert_held(&folio->mapping->i_private_lock);
913
914 if (fs_info->nodesize >= PAGE_SIZE) {
915 if (!folio_test_private(folio))
916 folio_attach_private(folio, eb);
917 else
918 WARN_ON(folio_get_private(folio) != eb);
919 return 0;
920 }
921
922 /* Already mapped, just free prealloc */
923 if (folio_test_private(folio)) {
924 btrfs_free_subpage(prealloc);
925 return 0;
926 }
927
928 if (prealloc)
929 /* Has preallocated memory for subpage */
930 folio_attach_private(folio, prealloc);
931 else
932 /* Do new allocation to attach subpage */
933 ret = btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_METADATA);
934 return ret;
935}
936
937int set_page_extent_mapped(struct page *page)
938{
939 struct folio *folio = page_folio(page);
940 struct btrfs_fs_info *fs_info;
941
942 ASSERT(page->mapping);
943
944 if (folio_test_private(folio))
945 return 0;
946
947 fs_info = btrfs_sb(page->mapping->host->i_sb);
948
949 if (btrfs_is_subpage(fs_info, page->mapping))
950 return btrfs_attach_subpage(fs_info, folio, BTRFS_SUBPAGE_DATA);
951
952 folio_attach_private(folio, (void *)EXTENT_FOLIO_PRIVATE);
953 return 0;
954}
955
956void clear_page_extent_mapped(struct page *page)
957{
958 struct folio *folio = page_folio(page);
959 struct btrfs_fs_info *fs_info;
960
961 ASSERT(page->mapping);
962
963 if (!folio_test_private(folio))
964 return;
965
966 fs_info = btrfs_sb(page->mapping->host->i_sb);
967 if (btrfs_is_subpage(fs_info, page->mapping))
968 return btrfs_detach_subpage(fs_info, folio);
969
970 folio_detach_private(folio);
971}
972
973static struct extent_map *
974__get_extent_map(struct inode *inode, struct page *page, size_t pg_offset,
975 u64 start, u64 len, struct extent_map **em_cached)
976{
977 struct extent_map *em;
978
979 if (em_cached && *em_cached) {
980 em = *em_cached;
981 if (extent_map_in_tree(em) && start >= em->start &&
982 start < extent_map_end(em)) {
983 refcount_inc(&em->refs);
984 return em;
985 }
986
987 free_extent_map(em);
988 *em_cached = NULL;
989 }
990
991 em = btrfs_get_extent(BTRFS_I(inode), page, pg_offset, start, len);
992 if (em_cached && !IS_ERR(em)) {
993 BUG_ON(*em_cached);
994 refcount_inc(&em->refs);
995 *em_cached = em;
996 }
997 return em;
998}
999/*
1000 * basic readpage implementation. Locked extent state structs are inserted
1001 * into the tree that are removed when the IO is done (by the end_io
1002 * handlers)
1003 * XXX JDM: This needs looking at to ensure proper page locking
1004 * return 0 on success, otherwise return error
1005 */
1006static int btrfs_do_readpage(struct page *page, struct extent_map **em_cached,
1007 struct btrfs_bio_ctrl *bio_ctrl, u64 *prev_em_start)
1008{
1009 struct inode *inode = page->mapping->host;
1010 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1011 u64 start = page_offset(page);
1012 const u64 end = start + PAGE_SIZE - 1;
1013 u64 cur = start;
1014 u64 extent_offset;
1015 u64 last_byte = i_size_read(inode);
1016 u64 block_start;
1017 struct extent_map *em;
1018 int ret = 0;
1019 size_t pg_offset = 0;
1020 size_t iosize;
1021 size_t blocksize = inode->i_sb->s_blocksize;
1022 struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
1023
1024 ret = set_page_extent_mapped(page);
1025 if (ret < 0) {
1026 unlock_extent(tree, start, end, NULL);
1027 unlock_page(page);
1028 return ret;
1029 }
1030
1031 if (page->index == last_byte >> PAGE_SHIFT) {
1032 size_t zero_offset = offset_in_page(last_byte);
1033
1034 if (zero_offset) {
1035 iosize = PAGE_SIZE - zero_offset;
1036 memzero_page(page, zero_offset, iosize);
1037 }
1038 }
1039 bio_ctrl->end_io_func = end_bbio_data_read;
1040 begin_page_read(fs_info, page);
1041 while (cur <= end) {
1042 enum btrfs_compression_type compress_type = BTRFS_COMPRESS_NONE;
1043 bool force_bio_submit = false;
1044 u64 disk_bytenr;
1045
1046 ASSERT(IS_ALIGNED(cur, fs_info->sectorsize));
1047 if (cur >= last_byte) {
1048 iosize = PAGE_SIZE - pg_offset;
1049 memzero_page(page, pg_offset, iosize);
1050 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1051 end_page_read(page, true, cur, iosize);
1052 break;
1053 }
1054 em = __get_extent_map(inode, page, pg_offset, cur,
1055 end - cur + 1, em_cached);
1056 if (IS_ERR(em)) {
1057 unlock_extent(tree, cur, end, NULL);
1058 end_page_read(page, false, cur, end + 1 - cur);
1059 return PTR_ERR(em);
1060 }
1061 extent_offset = cur - em->start;
1062 BUG_ON(extent_map_end(em) <= cur);
1063 BUG_ON(end < cur);
1064
1065 compress_type = extent_map_compression(em);
1066
1067 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1068 iosize = ALIGN(iosize, blocksize);
1069 if (compress_type != BTRFS_COMPRESS_NONE)
1070 disk_bytenr = em->block_start;
1071 else
1072 disk_bytenr = em->block_start + extent_offset;
1073 block_start = em->block_start;
1074 if (em->flags & EXTENT_FLAG_PREALLOC)
1075 block_start = EXTENT_MAP_HOLE;
1076
1077 /*
1078 * If we have a file range that points to a compressed extent
1079 * and it's followed by a consecutive file range that points
1080 * to the same compressed extent (possibly with a different
1081 * offset and/or length, so it either points to the whole extent
1082 * or only part of it), we must make sure we do not submit a
1083 * single bio to populate the pages for the 2 ranges because
1084 * this makes the compressed extent read zero out the pages
1085 * belonging to the 2nd range. Imagine the following scenario:
1086 *
1087 * File layout
1088 * [0 - 8K] [8K - 24K]
1089 * | |
1090 * | |
1091 * points to extent X, points to extent X,
1092 * offset 4K, length of 8K offset 0, length 16K
1093 *
1094 * [extent X, compressed length = 4K uncompressed length = 16K]
1095 *
1096 * If the bio to read the compressed extent covers both ranges,
1097 * it will decompress extent X into the pages belonging to the
1098 * first range and then it will stop, zeroing out the remaining
1099 * pages that belong to the other range that points to extent X.
1100 * So here we make sure we submit 2 bios, one for the first
1101 * range and another one for the third range. Both will target
1102 * the same physical extent from disk, but we can't currently
1103 * make the compressed bio endio callback populate the pages
1104 * for both ranges because each compressed bio is tightly
1105 * coupled with a single extent map, and each range can have
1106 * an extent map with a different offset value relative to the
1107 * uncompressed data of our extent and different lengths. This
1108 * is a corner case so we prioritize correctness over
1109 * non-optimal behavior (submitting 2 bios for the same extent).
1110 */
1111 if (compress_type != BTRFS_COMPRESS_NONE &&
1112 prev_em_start && *prev_em_start != (u64)-1 &&
1113 *prev_em_start != em->start)
1114 force_bio_submit = true;
1115
1116 if (prev_em_start)
1117 *prev_em_start = em->start;
1118
1119 free_extent_map(em);
1120 em = NULL;
1121
1122 /* we've found a hole, just zero and go on */
1123 if (block_start == EXTENT_MAP_HOLE) {
1124 memzero_page(page, pg_offset, iosize);
1125
1126 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1127 end_page_read(page, true, cur, iosize);
1128 cur = cur + iosize;
1129 pg_offset += iosize;
1130 continue;
1131 }
1132 /* the get_extent function already copied into the page */
1133 if (block_start == EXTENT_MAP_INLINE) {
1134 unlock_extent(tree, cur, cur + iosize - 1, NULL);
1135 end_page_read(page, true, cur, iosize);
1136 cur = cur + iosize;
1137 pg_offset += iosize;
1138 continue;
1139 }
1140
1141 if (bio_ctrl->compress_type != compress_type) {
1142 submit_one_bio(bio_ctrl);
1143 bio_ctrl->compress_type = compress_type;
1144 }
1145
1146 if (force_bio_submit)
1147 submit_one_bio(bio_ctrl);
1148 submit_extent_page(bio_ctrl, disk_bytenr, page, iosize,
1149 pg_offset);
1150 cur = cur + iosize;
1151 pg_offset += iosize;
1152 }
1153
1154 return 0;
1155}
1156
1157int btrfs_read_folio(struct file *file, struct folio *folio)
1158{
1159 struct page *page = &folio->page;
1160 struct btrfs_inode *inode = BTRFS_I(page->mapping->host);
1161 u64 start = page_offset(page);
1162 u64 end = start + PAGE_SIZE - 1;
1163 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ };
1164 int ret;
1165
1166 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
1167
1168 ret = btrfs_do_readpage(page, NULL, &bio_ctrl, NULL);
1169 /*
1170 * If btrfs_do_readpage() failed we will want to submit the assembled
1171 * bio to do the cleanup.
1172 */
1173 submit_one_bio(&bio_ctrl);
1174 return ret;
1175}
1176
1177static inline void contiguous_readpages(struct page *pages[], int nr_pages,
1178 u64 start, u64 end,
1179 struct extent_map **em_cached,
1180 struct btrfs_bio_ctrl *bio_ctrl,
1181 u64 *prev_em_start)
1182{
1183 struct btrfs_inode *inode = BTRFS_I(pages[0]->mapping->host);
1184 int index;
1185
1186 btrfs_lock_and_flush_ordered_range(inode, start, end, NULL);
1187
1188 for (index = 0; index < nr_pages; index++) {
1189 btrfs_do_readpage(pages[index], em_cached, bio_ctrl,
1190 prev_em_start);
1191 put_page(pages[index]);
1192 }
1193}
1194
1195/*
1196 * helper for __extent_writepage, doing all of the delayed allocation setup.
1197 *
1198 * This returns 1 if btrfs_run_delalloc_range function did all the work required
1199 * to write the page (copy into inline extent). In this case the IO has
1200 * been started and the page is already unlocked.
1201 *
1202 * This returns 0 if all went well (page still locked)
1203 * This returns < 0 if there were errors (page still locked)
1204 */
1205static noinline_for_stack int writepage_delalloc(struct btrfs_inode *inode,
1206 struct page *page, struct writeback_control *wbc)
1207{
1208 const u64 page_start = page_offset(page);
1209 const u64 page_end = page_start + PAGE_SIZE - 1;
1210 u64 delalloc_start = page_start;
1211 u64 delalloc_end = page_end;
1212 u64 delalloc_to_write = 0;
1213 int ret = 0;
1214
1215 while (delalloc_start < page_end) {
1216 delalloc_end = page_end;
1217 if (!find_lock_delalloc_range(&inode->vfs_inode, page,
1218 &delalloc_start, &delalloc_end)) {
1219 delalloc_start = delalloc_end + 1;
1220 continue;
1221 }
1222
1223 ret = btrfs_run_delalloc_range(inode, page, delalloc_start,
1224 delalloc_end, wbc);
1225 if (ret < 0)
1226 return ret;
1227
1228 delalloc_start = delalloc_end + 1;
1229 }
1230
1231 /*
1232 * delalloc_end is already one less than the total length, so
1233 * we don't subtract one from PAGE_SIZE
1234 */
1235 delalloc_to_write +=
1236 DIV_ROUND_UP(delalloc_end + 1 - page_start, PAGE_SIZE);
1237
1238 /*
1239 * If btrfs_run_dealloc_range() already started I/O and unlocked
1240 * the pages, we just need to account for them here.
1241 */
1242 if (ret == 1) {
1243 wbc->nr_to_write -= delalloc_to_write;
1244 return 1;
1245 }
1246
1247 if (wbc->nr_to_write < delalloc_to_write) {
1248 int thresh = 8192;
1249
1250 if (delalloc_to_write < thresh * 2)
1251 thresh = delalloc_to_write;
1252 wbc->nr_to_write = min_t(u64, delalloc_to_write,
1253 thresh);
1254 }
1255
1256 return 0;
1257}
1258
1259/*
1260 * Find the first byte we need to write.
1261 *
1262 * For subpage, one page can contain several sectors, and
1263 * __extent_writepage_io() will just grab all extent maps in the page
1264 * range and try to submit all non-inline/non-compressed extents.
1265 *
1266 * This is a big problem for subpage, we shouldn't re-submit already written
1267 * data at all.
1268 * This function will lookup subpage dirty bit to find which range we really
1269 * need to submit.
1270 *
1271 * Return the next dirty range in [@start, @end).
1272 * If no dirty range is found, @start will be page_offset(page) + PAGE_SIZE.
1273 */
1274static void find_next_dirty_byte(struct btrfs_fs_info *fs_info,
1275 struct page *page, u64 *start, u64 *end)
1276{
1277 struct folio *folio = page_folio(page);
1278 struct btrfs_subpage *subpage = folio_get_private(folio);
1279 struct btrfs_subpage_info *spi = fs_info->subpage_info;
1280 u64 orig_start = *start;
1281 /* Declare as unsigned long so we can use bitmap ops */
1282 unsigned long flags;
1283 int range_start_bit;
1284 int range_end_bit;
1285
1286 /*
1287 * For regular sector size == page size case, since one page only
1288 * contains one sector, we return the page offset directly.
1289 */
1290 if (!btrfs_is_subpage(fs_info, page->mapping)) {
1291 *start = page_offset(page);
1292 *end = page_offset(page) + PAGE_SIZE;
1293 return;
1294 }
1295
1296 range_start_bit = spi->dirty_offset +
1297 (offset_in_page(orig_start) >> fs_info->sectorsize_bits);
1298
1299 /* We should have the page locked, but just in case */
1300 spin_lock_irqsave(&subpage->lock, flags);
1301 bitmap_next_set_region(subpage->bitmaps, &range_start_bit, &range_end_bit,
1302 spi->dirty_offset + spi->bitmap_nr_bits);
1303 spin_unlock_irqrestore(&subpage->lock, flags);
1304
1305 range_start_bit -= spi->dirty_offset;
1306 range_end_bit -= spi->dirty_offset;
1307
1308 *start = page_offset(page) + range_start_bit * fs_info->sectorsize;
1309 *end = page_offset(page) + range_end_bit * fs_info->sectorsize;
1310}
1311
1312/*
1313 * helper for __extent_writepage. This calls the writepage start hooks,
1314 * and does the loop to map the page into extents and bios.
1315 *
1316 * We return 1 if the IO is started and the page is unlocked,
1317 * 0 if all went well (page still locked)
1318 * < 0 if there were errors (page still locked)
1319 */
1320static noinline_for_stack int __extent_writepage_io(struct btrfs_inode *inode,
1321 struct page *page,
1322 struct btrfs_bio_ctrl *bio_ctrl,
1323 loff_t i_size,
1324 int *nr_ret)
1325{
1326 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1327 u64 cur = page_offset(page);
1328 u64 end = cur + PAGE_SIZE - 1;
1329 u64 extent_offset;
1330 u64 block_start;
1331 struct extent_map *em;
1332 int ret = 0;
1333 int nr = 0;
1334
1335 ret = btrfs_writepage_cow_fixup(page);
1336 if (ret) {
1337 /* Fixup worker will requeue */
1338 redirty_page_for_writepage(bio_ctrl->wbc, page);
1339 unlock_page(page);
1340 return 1;
1341 }
1342
1343 bio_ctrl->end_io_func = end_bbio_data_write;
1344 while (cur <= end) {
1345 u32 len = end - cur + 1;
1346 u64 disk_bytenr;
1347 u64 em_end;
1348 u64 dirty_range_start = cur;
1349 u64 dirty_range_end;
1350 u32 iosize;
1351
1352 if (cur >= i_size) {
1353 btrfs_mark_ordered_io_finished(inode, page, cur, len,
1354 true);
1355 /*
1356 * This range is beyond i_size, thus we don't need to
1357 * bother writing back.
1358 * But we still need to clear the dirty subpage bit, or
1359 * the next time the page gets dirtied, we will try to
1360 * writeback the sectors with subpage dirty bits,
1361 * causing writeback without ordered extent.
1362 */
1363 btrfs_folio_clear_dirty(fs_info, page_folio(page), cur, len);
1364 break;
1365 }
1366
1367 find_next_dirty_byte(fs_info, page, &dirty_range_start,
1368 &dirty_range_end);
1369 if (cur < dirty_range_start) {
1370 cur = dirty_range_start;
1371 continue;
1372 }
1373
1374 em = btrfs_get_extent(inode, NULL, 0, cur, len);
1375 if (IS_ERR(em)) {
1376 ret = PTR_ERR_OR_ZERO(em);
1377 goto out_error;
1378 }
1379
1380 extent_offset = cur - em->start;
1381 em_end = extent_map_end(em);
1382 ASSERT(cur <= em_end);
1383 ASSERT(cur < end);
1384 ASSERT(IS_ALIGNED(em->start, fs_info->sectorsize));
1385 ASSERT(IS_ALIGNED(em->len, fs_info->sectorsize));
1386
1387 block_start = em->block_start;
1388 disk_bytenr = em->block_start + extent_offset;
1389
1390 ASSERT(!extent_map_is_compressed(em));
1391 ASSERT(block_start != EXTENT_MAP_HOLE);
1392 ASSERT(block_start != EXTENT_MAP_INLINE);
1393
1394 /*
1395 * Note that em_end from extent_map_end() and dirty_range_end from
1396 * find_next_dirty_byte() are all exclusive
1397 */
1398 iosize = min(min(em_end, end + 1), dirty_range_end) - cur;
1399 free_extent_map(em);
1400 em = NULL;
1401
1402 btrfs_set_range_writeback(inode, cur, cur + iosize - 1);
1403 if (!PageWriteback(page)) {
1404 btrfs_err(inode->root->fs_info,
1405 "page %lu not writeback, cur %llu end %llu",
1406 page->index, cur, end);
1407 }
1408
1409 /*
1410 * Although the PageDirty bit is cleared before entering this
1411 * function, subpage dirty bit is not cleared.
1412 * So clear subpage dirty bit here so next time we won't submit
1413 * page for range already written to disk.
1414 */
1415 btrfs_folio_clear_dirty(fs_info, page_folio(page), cur, iosize);
1416
1417 submit_extent_page(bio_ctrl, disk_bytenr, page, iosize,
1418 cur - page_offset(page));
1419 cur += iosize;
1420 nr++;
1421 }
1422
1423 btrfs_folio_assert_not_dirty(fs_info, page_folio(page));
1424 *nr_ret = nr;
1425 return 0;
1426
1427out_error:
1428 /*
1429 * If we finish without problem, we should not only clear page dirty,
1430 * but also empty subpage dirty bits
1431 */
1432 *nr_ret = nr;
1433 return ret;
1434}
1435
1436/*
1437 * the writepage semantics are similar to regular writepage. extent
1438 * records are inserted to lock ranges in the tree, and as dirty areas
1439 * are found, they are marked writeback. Then the lock bits are removed
1440 * and the end_io handler clears the writeback ranges
1441 *
1442 * Return 0 if everything goes well.
1443 * Return <0 for error.
1444 */
1445static int __extent_writepage(struct page *page, struct btrfs_bio_ctrl *bio_ctrl)
1446{
1447 struct folio *folio = page_folio(page);
1448 struct inode *inode = page->mapping->host;
1449 const u64 page_start = page_offset(page);
1450 int ret;
1451 int nr = 0;
1452 size_t pg_offset;
1453 loff_t i_size = i_size_read(inode);
1454 unsigned long end_index = i_size >> PAGE_SHIFT;
1455
1456 trace___extent_writepage(page, inode, bio_ctrl->wbc);
1457
1458 WARN_ON(!PageLocked(page));
1459
1460 pg_offset = offset_in_page(i_size);
1461 if (page->index > end_index ||
1462 (page->index == end_index && !pg_offset)) {
1463 folio_invalidate(folio, 0, folio_size(folio));
1464 folio_unlock(folio);
1465 return 0;
1466 }
1467
1468 if (page->index == end_index)
1469 memzero_page(page, pg_offset, PAGE_SIZE - pg_offset);
1470
1471 ret = set_page_extent_mapped(page);
1472 if (ret < 0)
1473 goto done;
1474
1475 ret = writepage_delalloc(BTRFS_I(inode), page, bio_ctrl->wbc);
1476 if (ret == 1)
1477 return 0;
1478 if (ret)
1479 goto done;
1480
1481 ret = __extent_writepage_io(BTRFS_I(inode), page, bio_ctrl, i_size, &nr);
1482 if (ret == 1)
1483 return 0;
1484
1485 bio_ctrl->wbc->nr_to_write--;
1486
1487done:
1488 if (nr == 0) {
1489 /* make sure the mapping tag for page dirty gets cleared */
1490 set_page_writeback(page);
1491 end_page_writeback(page);
1492 }
1493 if (ret) {
1494 btrfs_mark_ordered_io_finished(BTRFS_I(inode), page, page_start,
1495 PAGE_SIZE, !ret);
1496 mapping_set_error(page->mapping, ret);
1497 }
1498 unlock_page(page);
1499 ASSERT(ret <= 0);
1500 return ret;
1501}
1502
1503void wait_on_extent_buffer_writeback(struct extent_buffer *eb)
1504{
1505 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_WRITEBACK,
1506 TASK_UNINTERRUPTIBLE);
1507}
1508
1509/*
1510 * Lock extent buffer status and pages for writeback.
1511 *
1512 * Return %false if the extent buffer doesn't need to be submitted (e.g. the
1513 * extent buffer is not dirty)
1514 * Return %true is the extent buffer is submitted to bio.
1515 */
1516static noinline_for_stack bool lock_extent_buffer_for_io(struct extent_buffer *eb,
1517 struct writeback_control *wbc)
1518{
1519 struct btrfs_fs_info *fs_info = eb->fs_info;
1520 bool ret = false;
1521
1522 btrfs_tree_lock(eb);
1523 while (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags)) {
1524 btrfs_tree_unlock(eb);
1525 if (wbc->sync_mode != WB_SYNC_ALL)
1526 return false;
1527 wait_on_extent_buffer_writeback(eb);
1528 btrfs_tree_lock(eb);
1529 }
1530
1531 /*
1532 * We need to do this to prevent races in people who check if the eb is
1533 * under IO since we can end up having no IO bits set for a short period
1534 * of time.
1535 */
1536 spin_lock(&eb->refs_lock);
1537 if (test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
1538 set_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1539 spin_unlock(&eb->refs_lock);
1540 btrfs_set_header_flag(eb, BTRFS_HEADER_FLAG_WRITTEN);
1541 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes,
1542 -eb->len,
1543 fs_info->dirty_metadata_batch);
1544 ret = true;
1545 } else {
1546 spin_unlock(&eb->refs_lock);
1547 }
1548 btrfs_tree_unlock(eb);
1549 return ret;
1550}
1551
1552static void set_btree_ioerr(struct extent_buffer *eb)
1553{
1554 struct btrfs_fs_info *fs_info = eb->fs_info;
1555
1556 set_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1557
1558 /*
1559 * A read may stumble upon this buffer later, make sure that it gets an
1560 * error and knows there was an error.
1561 */
1562 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
1563
1564 /*
1565 * We need to set the mapping with the io error as well because a write
1566 * error will flip the file system readonly, and then syncfs() will
1567 * return a 0 because we are readonly if we don't modify the err seq for
1568 * the superblock.
1569 */
1570 mapping_set_error(eb->fs_info->btree_inode->i_mapping, -EIO);
1571
1572 /*
1573 * If writeback for a btree extent that doesn't belong to a log tree
1574 * failed, increment the counter transaction->eb_write_errors.
1575 * We do this because while the transaction is running and before it's
1576 * committing (when we call filemap_fdata[write|wait]_range against
1577 * the btree inode), we might have
1578 * btree_inode->i_mapping->a_ops->writepages() called by the VM - if it
1579 * returns an error or an error happens during writeback, when we're
1580 * committing the transaction we wouldn't know about it, since the pages
1581 * can be no longer dirty nor marked anymore for writeback (if a
1582 * subsequent modification to the extent buffer didn't happen before the
1583 * transaction commit), which makes filemap_fdata[write|wait]_range not
1584 * able to find the pages tagged with SetPageError at transaction
1585 * commit time. So if this happens we must abort the transaction,
1586 * otherwise we commit a super block with btree roots that point to
1587 * btree nodes/leafs whose content on disk is invalid - either garbage
1588 * or the content of some node/leaf from a past generation that got
1589 * cowed or deleted and is no longer valid.
1590 *
1591 * Note: setting AS_EIO/AS_ENOSPC in the btree inode's i_mapping would
1592 * not be enough - we need to distinguish between log tree extents vs
1593 * non-log tree extents, and the next filemap_fdatawait_range() call
1594 * will catch and clear such errors in the mapping - and that call might
1595 * be from a log sync and not from a transaction commit. Also, checking
1596 * for the eb flag EXTENT_BUFFER_WRITE_ERR at transaction commit time is
1597 * not done and would not be reliable - the eb might have been released
1598 * from memory and reading it back again means that flag would not be
1599 * set (since it's a runtime flag, not persisted on disk).
1600 *
1601 * Using the flags below in the btree inode also makes us achieve the
1602 * goal of AS_EIO/AS_ENOSPC when writepages() returns success, started
1603 * writeback for all dirty pages and before filemap_fdatawait_range()
1604 * is called, the writeback for all dirty pages had already finished
1605 * with errors - because we were not using AS_EIO/AS_ENOSPC,
1606 * filemap_fdatawait_range() would return success, as it could not know
1607 * that writeback errors happened (the pages were no longer tagged for
1608 * writeback).
1609 */
1610 switch (eb->log_index) {
1611 case -1:
1612 set_bit(BTRFS_FS_BTREE_ERR, &fs_info->flags);
1613 break;
1614 case 0:
1615 set_bit(BTRFS_FS_LOG1_ERR, &fs_info->flags);
1616 break;
1617 case 1:
1618 set_bit(BTRFS_FS_LOG2_ERR, &fs_info->flags);
1619 break;
1620 default:
1621 BUG(); /* unexpected, logic error */
1622 }
1623}
1624
1625/*
1626 * The endio specific version which won't touch any unsafe spinlock in endio
1627 * context.
1628 */
1629static struct extent_buffer *find_extent_buffer_nolock(
1630 struct btrfs_fs_info *fs_info, u64 start)
1631{
1632 struct extent_buffer *eb;
1633
1634 rcu_read_lock();
1635 eb = radix_tree_lookup(&fs_info->buffer_radix,
1636 start >> fs_info->sectorsize_bits);
1637 if (eb && atomic_inc_not_zero(&eb->refs)) {
1638 rcu_read_unlock();
1639 return eb;
1640 }
1641 rcu_read_unlock();
1642 return NULL;
1643}
1644
1645static void end_bbio_meta_write(struct btrfs_bio *bbio)
1646{
1647 struct extent_buffer *eb = bbio->private;
1648 struct btrfs_fs_info *fs_info = eb->fs_info;
1649 bool uptodate = !bbio->bio.bi_status;
1650 struct folio_iter fi;
1651 u32 bio_offset = 0;
1652
1653 if (!uptodate)
1654 set_btree_ioerr(eb);
1655
1656 bio_for_each_folio_all(fi, &bbio->bio) {
1657 u64 start = eb->start + bio_offset;
1658 struct folio *folio = fi.folio;
1659 u32 len = fi.length;
1660
1661 btrfs_folio_clear_writeback(fs_info, folio, start, len);
1662 bio_offset += len;
1663 }
1664
1665 clear_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags);
1666 smp_mb__after_atomic();
1667 wake_up_bit(&eb->bflags, EXTENT_BUFFER_WRITEBACK);
1668
1669 bio_put(&bbio->bio);
1670}
1671
1672static void prepare_eb_write(struct extent_buffer *eb)
1673{
1674 u32 nritems;
1675 unsigned long start;
1676 unsigned long end;
1677
1678 clear_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags);
1679
1680 /* Set btree blocks beyond nritems with 0 to avoid stale content */
1681 nritems = btrfs_header_nritems(eb);
1682 if (btrfs_header_level(eb) > 0) {
1683 end = btrfs_node_key_ptr_offset(eb, nritems);
1684 memzero_extent_buffer(eb, end, eb->len - end);
1685 } else {
1686 /*
1687 * Leaf:
1688 * header 0 1 2 .. N ... data_N .. data_2 data_1 data_0
1689 */
1690 start = btrfs_item_nr_offset(eb, nritems);
1691 end = btrfs_item_nr_offset(eb, 0);
1692 if (nritems == 0)
1693 end += BTRFS_LEAF_DATA_SIZE(eb->fs_info);
1694 else
1695 end += btrfs_item_offset(eb, nritems - 1);
1696 memzero_extent_buffer(eb, start, end - start);
1697 }
1698}
1699
1700static noinline_for_stack void write_one_eb(struct extent_buffer *eb,
1701 struct writeback_control *wbc)
1702{
1703 struct btrfs_fs_info *fs_info = eb->fs_info;
1704 struct btrfs_bio *bbio;
1705
1706 prepare_eb_write(eb);
1707
1708 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
1709 REQ_OP_WRITE | REQ_META | wbc_to_write_flags(wbc),
1710 eb->fs_info, end_bbio_meta_write, eb);
1711 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
1712 bio_set_dev(&bbio->bio, fs_info->fs_devices->latest_dev->bdev);
1713 wbc_init_bio(wbc, &bbio->bio);
1714 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
1715 bbio->file_offset = eb->start;
1716 if (fs_info->nodesize < PAGE_SIZE) {
1717 struct folio *folio = eb->folios[0];
1718 bool ret;
1719
1720 folio_lock(folio);
1721 btrfs_subpage_set_writeback(fs_info, folio, eb->start, eb->len);
1722 if (btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start,
1723 eb->len)) {
1724 folio_clear_dirty_for_io(folio);
1725 wbc->nr_to_write--;
1726 }
1727 ret = bio_add_folio(&bbio->bio, folio, eb->len,
1728 eb->start - folio_pos(folio));
1729 ASSERT(ret);
1730 wbc_account_cgroup_owner(wbc, folio_page(folio, 0), eb->len);
1731 folio_unlock(folio);
1732 } else {
1733 int num_folios = num_extent_folios(eb);
1734
1735 for (int i = 0; i < num_folios; i++) {
1736 struct folio *folio = eb->folios[i];
1737 bool ret;
1738
1739 folio_lock(folio);
1740 folio_clear_dirty_for_io(folio);
1741 folio_start_writeback(folio);
1742 ret = bio_add_folio(&bbio->bio, folio, folio_size(folio), 0);
1743 ASSERT(ret);
1744 wbc_account_cgroup_owner(wbc, folio_page(folio, 0),
1745 folio_size(folio));
1746 wbc->nr_to_write -= folio_nr_pages(folio);
1747 folio_unlock(folio);
1748 }
1749 }
1750 btrfs_submit_bio(bbio, 0);
1751}
1752
1753/*
1754 * Submit one subpage btree page.
1755 *
1756 * The main difference to submit_eb_page() is:
1757 * - Page locking
1758 * For subpage, we don't rely on page locking at all.
1759 *
1760 * - Flush write bio
1761 * We only flush bio if we may be unable to fit current extent buffers into
1762 * current bio.
1763 *
1764 * Return >=0 for the number of submitted extent buffers.
1765 * Return <0 for fatal error.
1766 */
1767static int submit_eb_subpage(struct page *page, struct writeback_control *wbc)
1768{
1769 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
1770 struct folio *folio = page_folio(page);
1771 int submitted = 0;
1772 u64 page_start = page_offset(page);
1773 int bit_start = 0;
1774 int sectors_per_node = fs_info->nodesize >> fs_info->sectorsize_bits;
1775
1776 /* Lock and write each dirty extent buffers in the range */
1777 while (bit_start < fs_info->subpage_info->bitmap_nr_bits) {
1778 struct btrfs_subpage *subpage = folio_get_private(folio);
1779 struct extent_buffer *eb;
1780 unsigned long flags;
1781 u64 start;
1782
1783 /*
1784 * Take private lock to ensure the subpage won't be detached
1785 * in the meantime.
1786 */
1787 spin_lock(&page->mapping->i_private_lock);
1788 if (!folio_test_private(folio)) {
1789 spin_unlock(&page->mapping->i_private_lock);
1790 break;
1791 }
1792 spin_lock_irqsave(&subpage->lock, flags);
1793 if (!test_bit(bit_start + fs_info->subpage_info->dirty_offset,
1794 subpage->bitmaps)) {
1795 spin_unlock_irqrestore(&subpage->lock, flags);
1796 spin_unlock(&page->mapping->i_private_lock);
1797 bit_start++;
1798 continue;
1799 }
1800
1801 start = page_start + bit_start * fs_info->sectorsize;
1802 bit_start += sectors_per_node;
1803
1804 /*
1805 * Here we just want to grab the eb without touching extra
1806 * spin locks, so call find_extent_buffer_nolock().
1807 */
1808 eb = find_extent_buffer_nolock(fs_info, start);
1809 spin_unlock_irqrestore(&subpage->lock, flags);
1810 spin_unlock(&page->mapping->i_private_lock);
1811
1812 /*
1813 * The eb has already reached 0 refs thus find_extent_buffer()
1814 * doesn't return it. We don't need to write back such eb
1815 * anyway.
1816 */
1817 if (!eb)
1818 continue;
1819
1820 if (lock_extent_buffer_for_io(eb, wbc)) {
1821 write_one_eb(eb, wbc);
1822 submitted++;
1823 }
1824 free_extent_buffer(eb);
1825 }
1826 return submitted;
1827}
1828
1829/*
1830 * Submit all page(s) of one extent buffer.
1831 *
1832 * @page: the page of one extent buffer
1833 * @eb_context: to determine if we need to submit this page, if current page
1834 * belongs to this eb, we don't need to submit
1835 *
1836 * The caller should pass each page in their bytenr order, and here we use
1837 * @eb_context to determine if we have submitted pages of one extent buffer.
1838 *
1839 * If we have, we just skip until we hit a new page that doesn't belong to
1840 * current @eb_context.
1841 *
1842 * If not, we submit all the page(s) of the extent buffer.
1843 *
1844 * Return >0 if we have submitted the extent buffer successfully.
1845 * Return 0 if we don't need to submit the page, as it's already submitted by
1846 * previous call.
1847 * Return <0 for fatal error.
1848 */
1849static int submit_eb_page(struct page *page, struct btrfs_eb_write_context *ctx)
1850{
1851 struct writeback_control *wbc = ctx->wbc;
1852 struct address_space *mapping = page->mapping;
1853 struct folio *folio = page_folio(page);
1854 struct extent_buffer *eb;
1855 int ret;
1856
1857 if (!folio_test_private(folio))
1858 return 0;
1859
1860 if (btrfs_sb(page->mapping->host->i_sb)->nodesize < PAGE_SIZE)
1861 return submit_eb_subpage(page, wbc);
1862
1863 spin_lock(&mapping->i_private_lock);
1864 if (!folio_test_private(folio)) {
1865 spin_unlock(&mapping->i_private_lock);
1866 return 0;
1867 }
1868
1869 eb = folio_get_private(folio);
1870
1871 /*
1872 * Shouldn't happen and normally this would be a BUG_ON but no point
1873 * crashing the machine for something we can survive anyway.
1874 */
1875 if (WARN_ON(!eb)) {
1876 spin_unlock(&mapping->i_private_lock);
1877 return 0;
1878 }
1879
1880 if (eb == ctx->eb) {
1881 spin_unlock(&mapping->i_private_lock);
1882 return 0;
1883 }
1884 ret = atomic_inc_not_zero(&eb->refs);
1885 spin_unlock(&mapping->i_private_lock);
1886 if (!ret)
1887 return 0;
1888
1889 ctx->eb = eb;
1890
1891 ret = btrfs_check_meta_write_pointer(eb->fs_info, ctx);
1892 if (ret) {
1893 if (ret == -EBUSY)
1894 ret = 0;
1895 free_extent_buffer(eb);
1896 return ret;
1897 }
1898
1899 if (!lock_extent_buffer_for_io(eb, wbc)) {
1900 free_extent_buffer(eb);
1901 return 0;
1902 }
1903 /* Implies write in zoned mode. */
1904 if (ctx->zoned_bg) {
1905 /* Mark the last eb in the block group. */
1906 btrfs_schedule_zone_finish_bg(ctx->zoned_bg, eb);
1907 ctx->zoned_bg->meta_write_pointer += eb->len;
1908 }
1909 write_one_eb(eb, wbc);
1910 free_extent_buffer(eb);
1911 return 1;
1912}
1913
1914int btree_write_cache_pages(struct address_space *mapping,
1915 struct writeback_control *wbc)
1916{
1917 struct btrfs_eb_write_context ctx = { .wbc = wbc };
1918 struct btrfs_fs_info *fs_info = BTRFS_I(mapping->host)->root->fs_info;
1919 int ret = 0;
1920 int done = 0;
1921 int nr_to_write_done = 0;
1922 struct folio_batch fbatch;
1923 unsigned int nr_folios;
1924 pgoff_t index;
1925 pgoff_t end; /* Inclusive */
1926 int scanned = 0;
1927 xa_mark_t tag;
1928
1929 folio_batch_init(&fbatch);
1930 if (wbc->range_cyclic) {
1931 index = mapping->writeback_index; /* Start from prev offset */
1932 end = -1;
1933 /*
1934 * Start from the beginning does not need to cycle over the
1935 * range, mark it as scanned.
1936 */
1937 scanned = (index == 0);
1938 } else {
1939 index = wbc->range_start >> PAGE_SHIFT;
1940 end = wbc->range_end >> PAGE_SHIFT;
1941 scanned = 1;
1942 }
1943 if (wbc->sync_mode == WB_SYNC_ALL)
1944 tag = PAGECACHE_TAG_TOWRITE;
1945 else
1946 tag = PAGECACHE_TAG_DIRTY;
1947 btrfs_zoned_meta_io_lock(fs_info);
1948retry:
1949 if (wbc->sync_mode == WB_SYNC_ALL)
1950 tag_pages_for_writeback(mapping, index, end);
1951 while (!done && !nr_to_write_done && (index <= end) &&
1952 (nr_folios = filemap_get_folios_tag(mapping, &index, end,
1953 tag, &fbatch))) {
1954 unsigned i;
1955
1956 for (i = 0; i < nr_folios; i++) {
1957 struct folio *folio = fbatch.folios[i];
1958
1959 ret = submit_eb_page(&folio->page, &ctx);
1960 if (ret == 0)
1961 continue;
1962 if (ret < 0) {
1963 done = 1;
1964 break;
1965 }
1966
1967 /*
1968 * the filesystem may choose to bump up nr_to_write.
1969 * We have to make sure to honor the new nr_to_write
1970 * at any time
1971 */
1972 nr_to_write_done = wbc->nr_to_write <= 0;
1973 }
1974 folio_batch_release(&fbatch);
1975 cond_resched();
1976 }
1977 if (!scanned && !done) {
1978 /*
1979 * We hit the last page and there is more work to be done: wrap
1980 * back to the start of the file
1981 */
1982 scanned = 1;
1983 index = 0;
1984 goto retry;
1985 }
1986 /*
1987 * If something went wrong, don't allow any metadata write bio to be
1988 * submitted.
1989 *
1990 * This would prevent use-after-free if we had dirty pages not
1991 * cleaned up, which can still happen by fuzzed images.
1992 *
1993 * - Bad extent tree
1994 * Allowing existing tree block to be allocated for other trees.
1995 *
1996 * - Log tree operations
1997 * Exiting tree blocks get allocated to log tree, bumps its
1998 * generation, then get cleaned in tree re-balance.
1999 * Such tree block will not be written back, since it's clean,
2000 * thus no WRITTEN flag set.
2001 * And after log writes back, this tree block is not traced by
2002 * any dirty extent_io_tree.
2003 *
2004 * - Offending tree block gets re-dirtied from its original owner
2005 * Since it has bumped generation, no WRITTEN flag, it can be
2006 * reused without COWing. This tree block will not be traced
2007 * by btrfs_transaction::dirty_pages.
2008 *
2009 * Now such dirty tree block will not be cleaned by any dirty
2010 * extent io tree. Thus we don't want to submit such wild eb
2011 * if the fs already has error.
2012 *
2013 * We can get ret > 0 from submit_extent_page() indicating how many ebs
2014 * were submitted. Reset it to 0 to avoid false alerts for the caller.
2015 */
2016 if (ret > 0)
2017 ret = 0;
2018 if (!ret && BTRFS_FS_ERROR(fs_info))
2019 ret = -EROFS;
2020
2021 if (ctx.zoned_bg)
2022 btrfs_put_block_group(ctx.zoned_bg);
2023 btrfs_zoned_meta_io_unlock(fs_info);
2024 return ret;
2025}
2026
2027/*
2028 * Walk the list of dirty pages of the given address space and write all of them.
2029 *
2030 * @mapping: address space structure to write
2031 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2032 * @bio_ctrl: holds context for the write, namely the bio
2033 *
2034 * If a page is already under I/O, write_cache_pages() skips it, even
2035 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2036 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2037 * and msync() need to guarantee that all the data which was dirty at the time
2038 * the call was made get new I/O started against them. If wbc->sync_mode is
2039 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2040 * existing IO to complete.
2041 */
2042static int extent_write_cache_pages(struct address_space *mapping,
2043 struct btrfs_bio_ctrl *bio_ctrl)
2044{
2045 struct writeback_control *wbc = bio_ctrl->wbc;
2046 struct inode *inode = mapping->host;
2047 int ret = 0;
2048 int done = 0;
2049 int nr_to_write_done = 0;
2050 struct folio_batch fbatch;
2051 unsigned int nr_folios;
2052 pgoff_t index;
2053 pgoff_t end; /* Inclusive */
2054 pgoff_t done_index;
2055 int range_whole = 0;
2056 int scanned = 0;
2057 xa_mark_t tag;
2058
2059 /*
2060 * We have to hold onto the inode so that ordered extents can do their
2061 * work when the IO finishes. The alternative to this is failing to add
2062 * an ordered extent if the igrab() fails there and that is a huge pain
2063 * to deal with, so instead just hold onto the inode throughout the
2064 * writepages operation. If it fails here we are freeing up the inode
2065 * anyway and we'd rather not waste our time writing out stuff that is
2066 * going to be truncated anyway.
2067 */
2068 if (!igrab(inode))
2069 return 0;
2070
2071 folio_batch_init(&fbatch);
2072 if (wbc->range_cyclic) {
2073 index = mapping->writeback_index; /* Start from prev offset */
2074 end = -1;
2075 /*
2076 * Start from the beginning does not need to cycle over the
2077 * range, mark it as scanned.
2078 */
2079 scanned = (index == 0);
2080 } else {
2081 index = wbc->range_start >> PAGE_SHIFT;
2082 end = wbc->range_end >> PAGE_SHIFT;
2083 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2084 range_whole = 1;
2085 scanned = 1;
2086 }
2087
2088 /*
2089 * We do the tagged writepage as long as the snapshot flush bit is set
2090 * and we are the first one who do the filemap_flush() on this inode.
2091 *
2092 * The nr_to_write == LONG_MAX is needed to make sure other flushers do
2093 * not race in and drop the bit.
2094 */
2095 if (range_whole && wbc->nr_to_write == LONG_MAX &&
2096 test_and_clear_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
2097 &BTRFS_I(inode)->runtime_flags))
2098 wbc->tagged_writepages = 1;
2099
2100 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2101 tag = PAGECACHE_TAG_TOWRITE;
2102 else
2103 tag = PAGECACHE_TAG_DIRTY;
2104retry:
2105 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2106 tag_pages_for_writeback(mapping, index, end);
2107 done_index = index;
2108 while (!done && !nr_to_write_done && (index <= end) &&
2109 (nr_folios = filemap_get_folios_tag(mapping, &index,
2110 end, tag, &fbatch))) {
2111 unsigned i;
2112
2113 for (i = 0; i < nr_folios; i++) {
2114 struct folio *folio = fbatch.folios[i];
2115
2116 done_index = folio_next_index(folio);
2117 /*
2118 * At this point we hold neither the i_pages lock nor
2119 * the page lock: the page may be truncated or
2120 * invalidated (changing page->mapping to NULL),
2121 * or even swizzled back from swapper_space to
2122 * tmpfs file mapping
2123 */
2124 if (!folio_trylock(folio)) {
2125 submit_write_bio(bio_ctrl, 0);
2126 folio_lock(folio);
2127 }
2128
2129 if (unlikely(folio->mapping != mapping)) {
2130 folio_unlock(folio);
2131 continue;
2132 }
2133
2134 if (!folio_test_dirty(folio)) {
2135 /* Someone wrote it for us. */
2136 folio_unlock(folio);
2137 continue;
2138 }
2139
2140 if (wbc->sync_mode != WB_SYNC_NONE) {
2141 if (folio_test_writeback(folio))
2142 submit_write_bio(bio_ctrl, 0);
2143 folio_wait_writeback(folio);
2144 }
2145
2146 if (folio_test_writeback(folio) ||
2147 !folio_clear_dirty_for_io(folio)) {
2148 folio_unlock(folio);
2149 continue;
2150 }
2151
2152 ret = __extent_writepage(&folio->page, bio_ctrl);
2153 if (ret < 0) {
2154 done = 1;
2155 break;
2156 }
2157
2158 /*
2159 * The filesystem may choose to bump up nr_to_write.
2160 * We have to make sure to honor the new nr_to_write
2161 * at any time.
2162 */
2163 nr_to_write_done = (wbc->sync_mode == WB_SYNC_NONE &&
2164 wbc->nr_to_write <= 0);
2165 }
2166 folio_batch_release(&fbatch);
2167 cond_resched();
2168 }
2169 if (!scanned && !done) {
2170 /*
2171 * We hit the last page and there is more work to be done: wrap
2172 * back to the start of the file
2173 */
2174 scanned = 1;
2175 index = 0;
2176
2177 /*
2178 * If we're looping we could run into a page that is locked by a
2179 * writer and that writer could be waiting on writeback for a
2180 * page in our current bio, and thus deadlock, so flush the
2181 * write bio here.
2182 */
2183 submit_write_bio(bio_ctrl, 0);
2184 goto retry;
2185 }
2186
2187 if (wbc->range_cyclic || (wbc->nr_to_write > 0 && range_whole))
2188 mapping->writeback_index = done_index;
2189
2190 btrfs_add_delayed_iput(BTRFS_I(inode));
2191 return ret;
2192}
2193
2194/*
2195 * Submit the pages in the range to bio for call sites which delalloc range has
2196 * already been ran (aka, ordered extent inserted) and all pages are still
2197 * locked.
2198 */
2199void extent_write_locked_range(struct inode *inode, struct page *locked_page,
2200 u64 start, u64 end, struct writeback_control *wbc,
2201 bool pages_dirty)
2202{
2203 bool found_error = false;
2204 int ret = 0;
2205 struct address_space *mapping = inode->i_mapping;
2206 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2207 const u32 sectorsize = fs_info->sectorsize;
2208 loff_t i_size = i_size_read(inode);
2209 u64 cur = start;
2210 struct btrfs_bio_ctrl bio_ctrl = {
2211 .wbc = wbc,
2212 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2213 };
2214
2215 if (wbc->no_cgroup_owner)
2216 bio_ctrl.opf |= REQ_BTRFS_CGROUP_PUNT;
2217
2218 ASSERT(IS_ALIGNED(start, sectorsize) && IS_ALIGNED(end + 1, sectorsize));
2219
2220 while (cur <= end) {
2221 u64 cur_end = min(round_down(cur, PAGE_SIZE) + PAGE_SIZE - 1, end);
2222 u32 cur_len = cur_end + 1 - cur;
2223 struct page *page;
2224 int nr = 0;
2225
2226 page = find_get_page(mapping, cur >> PAGE_SHIFT);
2227 ASSERT(PageLocked(page));
2228 if (pages_dirty && page != locked_page) {
2229 ASSERT(PageDirty(page));
2230 clear_page_dirty_for_io(page);
2231 }
2232
2233 ret = __extent_writepage_io(BTRFS_I(inode), page, &bio_ctrl,
2234 i_size, &nr);
2235 if (ret == 1)
2236 goto next_page;
2237
2238 /* Make sure the mapping tag for page dirty gets cleared. */
2239 if (nr == 0) {
2240 set_page_writeback(page);
2241 end_page_writeback(page);
2242 }
2243 if (ret) {
2244 btrfs_mark_ordered_io_finished(BTRFS_I(inode), page,
2245 cur, cur_len, !ret);
2246 mapping_set_error(page->mapping, ret);
2247 }
2248 btrfs_folio_unlock_writer(fs_info, page_folio(page), cur, cur_len);
2249 if (ret < 0)
2250 found_error = true;
2251next_page:
2252 put_page(page);
2253 cur = cur_end + 1;
2254 }
2255
2256 submit_write_bio(&bio_ctrl, found_error ? ret : 0);
2257}
2258
2259int extent_writepages(struct address_space *mapping,
2260 struct writeback_control *wbc)
2261{
2262 struct inode *inode = mapping->host;
2263 int ret = 0;
2264 struct btrfs_bio_ctrl bio_ctrl = {
2265 .wbc = wbc,
2266 .opf = REQ_OP_WRITE | wbc_to_write_flags(wbc),
2267 };
2268
2269 /*
2270 * Allow only a single thread to do the reloc work in zoned mode to
2271 * protect the write pointer updates.
2272 */
2273 btrfs_zoned_data_reloc_lock(BTRFS_I(inode));
2274 ret = extent_write_cache_pages(mapping, &bio_ctrl);
2275 submit_write_bio(&bio_ctrl, ret);
2276 btrfs_zoned_data_reloc_unlock(BTRFS_I(inode));
2277 return ret;
2278}
2279
2280void extent_readahead(struct readahead_control *rac)
2281{
2282 struct btrfs_bio_ctrl bio_ctrl = { .opf = REQ_OP_READ | REQ_RAHEAD };
2283 struct page *pagepool[16];
2284 struct extent_map *em_cached = NULL;
2285 u64 prev_em_start = (u64)-1;
2286 int nr;
2287
2288 while ((nr = readahead_page_batch(rac, pagepool))) {
2289 u64 contig_start = readahead_pos(rac);
2290 u64 contig_end = contig_start + readahead_batch_length(rac) - 1;
2291
2292 contiguous_readpages(pagepool, nr, contig_start, contig_end,
2293 &em_cached, &bio_ctrl, &prev_em_start);
2294 }
2295
2296 if (em_cached)
2297 free_extent_map(em_cached);
2298 submit_one_bio(&bio_ctrl);
2299}
2300
2301/*
2302 * basic invalidate_folio code, this waits on any locked or writeback
2303 * ranges corresponding to the folio, and then deletes any extent state
2304 * records from the tree
2305 */
2306int extent_invalidate_folio(struct extent_io_tree *tree,
2307 struct folio *folio, size_t offset)
2308{
2309 struct extent_state *cached_state = NULL;
2310 u64 start = folio_pos(folio);
2311 u64 end = start + folio_size(folio) - 1;
2312 size_t blocksize = folio->mapping->host->i_sb->s_blocksize;
2313
2314 /* This function is only called for the btree inode */
2315 ASSERT(tree->owner == IO_TREE_BTREE_INODE_IO);
2316
2317 start += ALIGN(offset, blocksize);
2318 if (start > end)
2319 return 0;
2320
2321 lock_extent(tree, start, end, &cached_state);
2322 folio_wait_writeback(folio);
2323
2324 /*
2325 * Currently for btree io tree, only EXTENT_LOCKED is utilized,
2326 * so here we only need to unlock the extent range to free any
2327 * existing extent state.
2328 */
2329 unlock_extent(tree, start, end, &cached_state);
2330 return 0;
2331}
2332
2333/*
2334 * a helper for release_folio, this tests for areas of the page that
2335 * are locked or under IO and drops the related state bits if it is safe
2336 * to drop the page.
2337 */
2338static int try_release_extent_state(struct extent_io_tree *tree,
2339 struct page *page, gfp_t mask)
2340{
2341 u64 start = page_offset(page);
2342 u64 end = start + PAGE_SIZE - 1;
2343 int ret = 1;
2344
2345 if (test_range_bit_exists(tree, start, end, EXTENT_LOCKED)) {
2346 ret = 0;
2347 } else {
2348 u32 clear_bits = ~(EXTENT_LOCKED | EXTENT_NODATASUM |
2349 EXTENT_DELALLOC_NEW | EXTENT_CTLBITS |
2350 EXTENT_QGROUP_RESERVED);
2351
2352 /*
2353 * At this point we can safely clear everything except the
2354 * locked bit, the nodatasum bit and the delalloc new bit.
2355 * The delalloc new bit will be cleared by ordered extent
2356 * completion.
2357 */
2358 ret = __clear_extent_bit(tree, start, end, clear_bits, NULL, NULL);
2359
2360 /* if clear_extent_bit failed for enomem reasons,
2361 * we can't allow the release to continue.
2362 */
2363 if (ret < 0)
2364 ret = 0;
2365 else
2366 ret = 1;
2367 }
2368 return ret;
2369}
2370
2371/*
2372 * a helper for release_folio. As long as there are no locked extents
2373 * in the range corresponding to the page, both state records and extent
2374 * map records are removed
2375 */
2376int try_release_extent_mapping(struct page *page, gfp_t mask)
2377{
2378 struct extent_map *em;
2379 u64 start = page_offset(page);
2380 u64 end = start + PAGE_SIZE - 1;
2381 struct btrfs_inode *btrfs_inode = BTRFS_I(page->mapping->host);
2382 struct extent_io_tree *tree = &btrfs_inode->io_tree;
2383 struct extent_map_tree *map = &btrfs_inode->extent_tree;
2384
2385 if (gfpflags_allow_blocking(mask) &&
2386 page->mapping->host->i_size > SZ_16M) {
2387 u64 len;
2388 while (start <= end) {
2389 struct btrfs_fs_info *fs_info;
2390 u64 cur_gen;
2391
2392 len = end - start + 1;
2393 write_lock(&map->lock);
2394 em = lookup_extent_mapping(map, start, len);
2395 if (!em) {
2396 write_unlock(&map->lock);
2397 break;
2398 }
2399 if ((em->flags & EXTENT_FLAG_PINNED) ||
2400 em->start != start) {
2401 write_unlock(&map->lock);
2402 free_extent_map(em);
2403 break;
2404 }
2405 if (test_range_bit_exists(tree, em->start,
2406 extent_map_end(em) - 1,
2407 EXTENT_LOCKED))
2408 goto next;
2409 /*
2410 * If it's not in the list of modified extents, used
2411 * by a fast fsync, we can remove it. If it's being
2412 * logged we can safely remove it since fsync took an
2413 * extra reference on the em.
2414 */
2415 if (list_empty(&em->list) ||
2416 (em->flags & EXTENT_FLAG_LOGGING))
2417 goto remove_em;
2418 /*
2419 * If it's in the list of modified extents, remove it
2420 * only if its generation is older then the current one,
2421 * in which case we don't need it for a fast fsync.
2422 * Otherwise don't remove it, we could be racing with an
2423 * ongoing fast fsync that could miss the new extent.
2424 */
2425 fs_info = btrfs_inode->root->fs_info;
2426 spin_lock(&fs_info->trans_lock);
2427 cur_gen = fs_info->generation;
2428 spin_unlock(&fs_info->trans_lock);
2429 if (em->generation >= cur_gen)
2430 goto next;
2431remove_em:
2432 /*
2433 * We only remove extent maps that are not in the list of
2434 * modified extents or that are in the list but with a
2435 * generation lower then the current generation, so there
2436 * is no need to set the full fsync flag on the inode (it
2437 * hurts the fsync performance for workloads with a data
2438 * size that exceeds or is close to the system's memory).
2439 */
2440 remove_extent_mapping(map, em);
2441 /* once for the rb tree */
2442 free_extent_map(em);
2443next:
2444 start = extent_map_end(em);
2445 write_unlock(&map->lock);
2446
2447 /* once for us */
2448 free_extent_map(em);
2449
2450 cond_resched(); /* Allow large-extent preemption. */
2451 }
2452 }
2453 return try_release_extent_state(tree, page, mask);
2454}
2455
2456/*
2457 * To cache previous fiemap extent
2458 *
2459 * Will be used for merging fiemap extent
2460 */
2461struct fiemap_cache {
2462 u64 offset;
2463 u64 phys;
2464 u64 len;
2465 u32 flags;
2466 bool cached;
2467};
2468
2469/*
2470 * Helper to submit fiemap extent.
2471 *
2472 * Will try to merge current fiemap extent specified by @offset, @phys,
2473 * @len and @flags with cached one.
2474 * And only when we fails to merge, cached one will be submitted as
2475 * fiemap extent.
2476 *
2477 * Return value is the same as fiemap_fill_next_extent().
2478 */
2479static int emit_fiemap_extent(struct fiemap_extent_info *fieinfo,
2480 struct fiemap_cache *cache,
2481 u64 offset, u64 phys, u64 len, u32 flags)
2482{
2483 u64 cache_end;
2484 int ret = 0;
2485
2486 /* Set at the end of extent_fiemap(). */
2487 ASSERT((flags & FIEMAP_EXTENT_LAST) == 0);
2488
2489 if (!cache->cached)
2490 goto assign;
2491
2492 /*
2493 * When iterating the extents of the inode, at extent_fiemap(), we may
2494 * find an extent that starts at an offset behind the end offset of the
2495 * previous extent we processed. This happens if fiemap is called
2496 * without FIEMAP_FLAG_SYNC and there are ordered extents completing
2497 * while we call btrfs_next_leaf() (through fiemap_next_leaf_item()).
2498 *
2499 * For example we are in leaf X processing its last item, which is the
2500 * file extent item for file range [512K, 1M[, and after
2501 * btrfs_next_leaf() releases the path, there's an ordered extent that
2502 * completes for the file range [768K, 2M[, and that results in trimming
2503 * the file extent item so that it now corresponds to the file range
2504 * [512K, 768K[ and a new file extent item is inserted for the file
2505 * range [768K, 2M[, which may end up as the last item of leaf X or as
2506 * the first item of the next leaf - in either case btrfs_next_leaf()
2507 * will leave us with a path pointing to the new extent item, for the
2508 * file range [768K, 2M[, since that's the first key that follows the
2509 * last one we processed. So in order not to report overlapping extents
2510 * to user space, we trim the length of the previously cached extent and
2511 * emit it.
2512 *
2513 * Upon calling btrfs_next_leaf() we may also find an extent with an
2514 * offset smaller than or equals to cache->offset, and this happens
2515 * when we had a hole or prealloc extent with several delalloc ranges in
2516 * it, but after btrfs_next_leaf() released the path, delalloc was
2517 * flushed and the resulting ordered extents were completed, so we can
2518 * now have found a file extent item for an offset that is smaller than
2519 * or equals to what we have in cache->offset. We deal with this as
2520 * described below.
2521 */
2522 cache_end = cache->offset + cache->len;
2523 if (cache_end > offset) {
2524 if (offset == cache->offset) {
2525 /*
2526 * We cached a dealloc range (found in the io tree) for
2527 * a hole or prealloc extent and we have now found a
2528 * file extent item for the same offset. What we have
2529 * now is more recent and up to date, so discard what
2530 * we had in the cache and use what we have just found.
2531 */
2532 goto assign;
2533 } else if (offset > cache->offset) {
2534 /*
2535 * The extent range we previously found ends after the
2536 * offset of the file extent item we found and that
2537 * offset falls somewhere in the middle of that previous
2538 * extent range. So adjust the range we previously found
2539 * to end at the offset of the file extent item we have
2540 * just found, since this extent is more up to date.
2541 * Emit that adjusted range and cache the file extent
2542 * item we have just found. This corresponds to the case
2543 * where a previously found file extent item was split
2544 * due to an ordered extent completing.
2545 */
2546 cache->len = offset - cache->offset;
2547 goto emit;
2548 } else {
2549 const u64 range_end = offset + len;
2550
2551 /*
2552 * The offset of the file extent item we have just found
2553 * is behind the cached offset. This means we were
2554 * processing a hole or prealloc extent for which we
2555 * have found delalloc ranges (in the io tree), so what
2556 * we have in the cache is the last delalloc range we
2557 * found while the file extent item we found can be
2558 * either for a whole delalloc range we previously
2559 * emmitted or only a part of that range.
2560 *
2561 * We have two cases here:
2562 *
2563 * 1) The file extent item's range ends at or behind the
2564 * cached extent's end. In this case just ignore the
2565 * current file extent item because we don't want to
2566 * overlap with previous ranges that may have been
2567 * emmitted already;
2568 *
2569 * 2) The file extent item starts behind the currently
2570 * cached extent but its end offset goes beyond the
2571 * end offset of the cached extent. We don't want to
2572 * overlap with a previous range that may have been
2573 * emmitted already, so we emit the currently cached
2574 * extent and then partially store the current file
2575 * extent item's range in the cache, for the subrange
2576 * going the cached extent's end to the end of the
2577 * file extent item.
2578 */
2579 if (range_end <= cache_end)
2580 return 0;
2581
2582 if (!(flags & (FIEMAP_EXTENT_ENCODED | FIEMAP_EXTENT_DELALLOC)))
2583 phys += cache_end - offset;
2584
2585 offset = cache_end;
2586 len = range_end - cache_end;
2587 goto emit;
2588 }
2589 }
2590
2591 /*
2592 * Only merges fiemap extents if
2593 * 1) Their logical addresses are continuous
2594 *
2595 * 2) Their physical addresses are continuous
2596 * So truly compressed (physical size smaller than logical size)
2597 * extents won't get merged with each other
2598 *
2599 * 3) Share same flags
2600 */
2601 if (cache->offset + cache->len == offset &&
2602 cache->phys + cache->len == phys &&
2603 cache->flags == flags) {
2604 cache->len += len;
2605 return 0;
2606 }
2607
2608emit:
2609 /* Not mergeable, need to submit cached one */
2610 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
2611 cache->len, cache->flags);
2612 cache->cached = false;
2613 if (ret)
2614 return ret;
2615assign:
2616 cache->cached = true;
2617 cache->offset = offset;
2618 cache->phys = phys;
2619 cache->len = len;
2620 cache->flags = flags;
2621
2622 return 0;
2623}
2624
2625/*
2626 * Emit last fiemap cache
2627 *
2628 * The last fiemap cache may still be cached in the following case:
2629 * 0 4k 8k
2630 * |<- Fiemap range ->|
2631 * |<------------ First extent ----------->|
2632 *
2633 * In this case, the first extent range will be cached but not emitted.
2634 * So we must emit it before ending extent_fiemap().
2635 */
2636static int emit_last_fiemap_cache(struct fiemap_extent_info *fieinfo,
2637 struct fiemap_cache *cache)
2638{
2639 int ret;
2640
2641 if (!cache->cached)
2642 return 0;
2643
2644 ret = fiemap_fill_next_extent(fieinfo, cache->offset, cache->phys,
2645 cache->len, cache->flags);
2646 cache->cached = false;
2647 if (ret > 0)
2648 ret = 0;
2649 return ret;
2650}
2651
2652static int fiemap_next_leaf_item(struct btrfs_inode *inode, struct btrfs_path *path)
2653{
2654 struct extent_buffer *clone;
2655 struct btrfs_key key;
2656 int slot;
2657 int ret;
2658
2659 path->slots[0]++;
2660 if (path->slots[0] < btrfs_header_nritems(path->nodes[0]))
2661 return 0;
2662
2663 ret = btrfs_next_leaf(inode->root, path);
2664 if (ret != 0)
2665 return ret;
2666
2667 /*
2668 * Don't bother with cloning if there are no more file extent items for
2669 * our inode.
2670 */
2671 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2672 if (key.objectid != btrfs_ino(inode) || key.type != BTRFS_EXTENT_DATA_KEY)
2673 return 1;
2674
2675 /* See the comment at fiemap_search_slot() about why we clone. */
2676 clone = btrfs_clone_extent_buffer(path->nodes[0]);
2677 if (!clone)
2678 return -ENOMEM;
2679
2680 slot = path->slots[0];
2681 btrfs_release_path(path);
2682 path->nodes[0] = clone;
2683 path->slots[0] = slot;
2684
2685 return 0;
2686}
2687
2688/*
2689 * Search for the first file extent item that starts at a given file offset or
2690 * the one that starts immediately before that offset.
2691 * Returns: 0 on success, < 0 on error, 1 if not found.
2692 */
2693static int fiemap_search_slot(struct btrfs_inode *inode, struct btrfs_path *path,
2694 u64 file_offset)
2695{
2696 const u64 ino = btrfs_ino(inode);
2697 struct btrfs_root *root = inode->root;
2698 struct extent_buffer *clone;
2699 struct btrfs_key key;
2700 int slot;
2701 int ret;
2702
2703 key.objectid = ino;
2704 key.type = BTRFS_EXTENT_DATA_KEY;
2705 key.offset = file_offset;
2706
2707 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2708 if (ret < 0)
2709 return ret;
2710
2711 if (ret > 0 && path->slots[0] > 0) {
2712 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
2713 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
2714 path->slots[0]--;
2715 }
2716
2717 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
2718 ret = btrfs_next_leaf(root, path);
2719 if (ret != 0)
2720 return ret;
2721
2722 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
2723 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
2724 return 1;
2725 }
2726
2727 /*
2728 * We clone the leaf and use it during fiemap. This is because while
2729 * using the leaf we do expensive things like checking if an extent is
2730 * shared, which can take a long time. In order to prevent blocking
2731 * other tasks for too long, we use a clone of the leaf. We have locked
2732 * the file range in the inode's io tree, so we know none of our file
2733 * extent items can change. This way we avoid blocking other tasks that
2734 * want to insert items for other inodes in the same leaf or b+tree
2735 * rebalance operations (triggered for example when someone is trying
2736 * to push items into this leaf when trying to insert an item in a
2737 * neighbour leaf).
2738 * We also need the private clone because holding a read lock on an
2739 * extent buffer of the subvolume's b+tree will make lockdep unhappy
2740 * when we call fiemap_fill_next_extent(), because that may cause a page
2741 * fault when filling the user space buffer with fiemap data.
2742 */
2743 clone = btrfs_clone_extent_buffer(path->nodes[0]);
2744 if (!clone)
2745 return -ENOMEM;
2746
2747 slot = path->slots[0];
2748 btrfs_release_path(path);
2749 path->nodes[0] = clone;
2750 path->slots[0] = slot;
2751
2752 return 0;
2753}
2754
2755/*
2756 * Process a range which is a hole or a prealloc extent in the inode's subvolume
2757 * btree. If @disk_bytenr is 0, we are dealing with a hole, otherwise a prealloc
2758 * extent. The end offset (@end) is inclusive.
2759 */
2760static int fiemap_process_hole(struct btrfs_inode *inode,
2761 struct fiemap_extent_info *fieinfo,
2762 struct fiemap_cache *cache,
2763 struct extent_state **delalloc_cached_state,
2764 struct btrfs_backref_share_check_ctx *backref_ctx,
2765 u64 disk_bytenr, u64 extent_offset,
2766 u64 extent_gen,
2767 u64 start, u64 end)
2768{
2769 const u64 i_size = i_size_read(&inode->vfs_inode);
2770 u64 cur_offset = start;
2771 u64 last_delalloc_end = 0;
2772 u32 prealloc_flags = FIEMAP_EXTENT_UNWRITTEN;
2773 bool checked_extent_shared = false;
2774 int ret;
2775
2776 /*
2777 * There can be no delalloc past i_size, so don't waste time looking for
2778 * it beyond i_size.
2779 */
2780 while (cur_offset < end && cur_offset < i_size) {
2781 struct extent_state *cached_state = NULL;
2782 u64 delalloc_start;
2783 u64 delalloc_end;
2784 u64 prealloc_start;
2785 u64 lockstart;
2786 u64 lockend;
2787 u64 prealloc_len = 0;
2788 bool delalloc;
2789
2790 lockstart = round_down(cur_offset, inode->root->fs_info->sectorsize);
2791 lockend = round_up(end, inode->root->fs_info->sectorsize);
2792
2793 /*
2794 * We are only locking for the delalloc range because that's the
2795 * only thing that can change here. With fiemap we have a lock
2796 * on the inode, so no buffered or direct writes can happen.
2797 *
2798 * However mmaps and normal page writeback will cause this to
2799 * change arbitrarily. We have to lock the extent lock here to
2800 * make sure that nobody messes with the tree while we're doing
2801 * btrfs_find_delalloc_in_range.
2802 */
2803 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
2804 delalloc = btrfs_find_delalloc_in_range(inode, cur_offset, end,
2805 delalloc_cached_state,
2806 &delalloc_start,
2807 &delalloc_end);
2808 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
2809 if (!delalloc)
2810 break;
2811
2812 /*
2813 * If this is a prealloc extent we have to report every section
2814 * of it that has no delalloc.
2815 */
2816 if (disk_bytenr != 0) {
2817 if (last_delalloc_end == 0) {
2818 prealloc_start = start;
2819 prealloc_len = delalloc_start - start;
2820 } else {
2821 prealloc_start = last_delalloc_end + 1;
2822 prealloc_len = delalloc_start - prealloc_start;
2823 }
2824 }
2825
2826 if (prealloc_len > 0) {
2827 if (!checked_extent_shared && fieinfo->fi_extents_max) {
2828 ret = btrfs_is_data_extent_shared(inode,
2829 disk_bytenr,
2830 extent_gen,
2831 backref_ctx);
2832 if (ret < 0)
2833 return ret;
2834 else if (ret > 0)
2835 prealloc_flags |= FIEMAP_EXTENT_SHARED;
2836
2837 checked_extent_shared = true;
2838 }
2839 ret = emit_fiemap_extent(fieinfo, cache, prealloc_start,
2840 disk_bytenr + extent_offset,
2841 prealloc_len, prealloc_flags);
2842 if (ret)
2843 return ret;
2844 extent_offset += prealloc_len;
2845 }
2846
2847 ret = emit_fiemap_extent(fieinfo, cache, delalloc_start, 0,
2848 delalloc_end + 1 - delalloc_start,
2849 FIEMAP_EXTENT_DELALLOC |
2850 FIEMAP_EXTENT_UNKNOWN);
2851 if (ret)
2852 return ret;
2853
2854 last_delalloc_end = delalloc_end;
2855 cur_offset = delalloc_end + 1;
2856 extent_offset += cur_offset - delalloc_start;
2857 cond_resched();
2858 }
2859
2860 /*
2861 * Either we found no delalloc for the whole prealloc extent or we have
2862 * a prealloc extent that spans i_size or starts at or after i_size.
2863 */
2864 if (disk_bytenr != 0 && last_delalloc_end < end) {
2865 u64 prealloc_start;
2866 u64 prealloc_len;
2867
2868 if (last_delalloc_end == 0) {
2869 prealloc_start = start;
2870 prealloc_len = end + 1 - start;
2871 } else {
2872 prealloc_start = last_delalloc_end + 1;
2873 prealloc_len = end + 1 - prealloc_start;
2874 }
2875
2876 if (!checked_extent_shared && fieinfo->fi_extents_max) {
2877 ret = btrfs_is_data_extent_shared(inode,
2878 disk_bytenr,
2879 extent_gen,
2880 backref_ctx);
2881 if (ret < 0)
2882 return ret;
2883 else if (ret > 0)
2884 prealloc_flags |= FIEMAP_EXTENT_SHARED;
2885 }
2886 ret = emit_fiemap_extent(fieinfo, cache, prealloc_start,
2887 disk_bytenr + extent_offset,
2888 prealloc_len, prealloc_flags);
2889 if (ret)
2890 return ret;
2891 }
2892
2893 return 0;
2894}
2895
2896static int fiemap_find_last_extent_offset(struct btrfs_inode *inode,
2897 struct btrfs_path *path,
2898 u64 *last_extent_end_ret)
2899{
2900 const u64 ino = btrfs_ino(inode);
2901 struct btrfs_root *root = inode->root;
2902 struct extent_buffer *leaf;
2903 struct btrfs_file_extent_item *ei;
2904 struct btrfs_key key;
2905 u64 disk_bytenr;
2906 int ret;
2907
2908 /*
2909 * Lookup the last file extent. We're not using i_size here because
2910 * there might be preallocation past i_size.
2911 */
2912 ret = btrfs_lookup_file_extent(NULL, root, path, ino, (u64)-1, 0);
2913 /* There can't be a file extent item at offset (u64)-1 */
2914 ASSERT(ret != 0);
2915 if (ret < 0)
2916 return ret;
2917
2918 /*
2919 * For a non-existing key, btrfs_search_slot() always leaves us at a
2920 * slot > 0, except if the btree is empty, which is impossible because
2921 * at least it has the inode item for this inode and all the items for
2922 * the root inode 256.
2923 */
2924 ASSERT(path->slots[0] > 0);
2925 path->slots[0]--;
2926 leaf = path->nodes[0];
2927 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2928 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY) {
2929 /* No file extent items in the subvolume tree. */
2930 *last_extent_end_ret = 0;
2931 return 0;
2932 }
2933
2934 /*
2935 * For an inline extent, the disk_bytenr is where inline data starts at,
2936 * so first check if we have an inline extent item before checking if we
2937 * have an implicit hole (disk_bytenr == 0).
2938 */
2939 ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_file_extent_item);
2940 if (btrfs_file_extent_type(leaf, ei) == BTRFS_FILE_EXTENT_INLINE) {
2941 *last_extent_end_ret = btrfs_file_extent_end(path);
2942 return 0;
2943 }
2944
2945 /*
2946 * Find the last file extent item that is not a hole (when NO_HOLES is
2947 * not enabled). This should take at most 2 iterations in the worst
2948 * case: we have one hole file extent item at slot 0 of a leaf and
2949 * another hole file extent item as the last item in the previous leaf.
2950 * This is because we merge file extent items that represent holes.
2951 */
2952 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
2953 while (disk_bytenr == 0) {
2954 ret = btrfs_previous_item(root, path, ino, BTRFS_EXTENT_DATA_KEY);
2955 if (ret < 0) {
2956 return ret;
2957 } else if (ret > 0) {
2958 /* No file extent items that are not holes. */
2959 *last_extent_end_ret = 0;
2960 return 0;
2961 }
2962 leaf = path->nodes[0];
2963 ei = btrfs_item_ptr(leaf, path->slots[0],
2964 struct btrfs_file_extent_item);
2965 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
2966 }
2967
2968 *last_extent_end_ret = btrfs_file_extent_end(path);
2969 return 0;
2970}
2971
2972int extent_fiemap(struct btrfs_inode *inode, struct fiemap_extent_info *fieinfo,
2973 u64 start, u64 len)
2974{
2975 const u64 ino = btrfs_ino(inode);
2976 struct extent_state *delalloc_cached_state = NULL;
2977 struct btrfs_path *path;
2978 struct fiemap_cache cache = { 0 };
2979 struct btrfs_backref_share_check_ctx *backref_ctx;
2980 u64 last_extent_end;
2981 u64 prev_extent_end;
2982 u64 range_start;
2983 u64 range_end;
2984 const u64 sectorsize = inode->root->fs_info->sectorsize;
2985 bool stopped = false;
2986 int ret;
2987
2988 backref_ctx = btrfs_alloc_backref_share_check_ctx();
2989 path = btrfs_alloc_path();
2990 if (!backref_ctx || !path) {
2991 ret = -ENOMEM;
2992 goto out;
2993 }
2994
2995 range_start = round_down(start, sectorsize);
2996 range_end = round_up(start + len, sectorsize);
2997 prev_extent_end = range_start;
2998
2999 ret = fiemap_find_last_extent_offset(inode, path, &last_extent_end);
3000 if (ret < 0)
3001 goto out;
3002 btrfs_release_path(path);
3003
3004 path->reada = READA_FORWARD;
3005 ret = fiemap_search_slot(inode, path, range_start);
3006 if (ret < 0) {
3007 goto out;
3008 } else if (ret > 0) {
3009 /*
3010 * No file extent item found, but we may have delalloc between
3011 * the current offset and i_size. So check for that.
3012 */
3013 ret = 0;
3014 goto check_eof_delalloc;
3015 }
3016
3017 while (prev_extent_end < range_end) {
3018 struct extent_buffer *leaf = path->nodes[0];
3019 struct btrfs_file_extent_item *ei;
3020 struct btrfs_key key;
3021 u64 extent_end;
3022 u64 extent_len;
3023 u64 extent_offset = 0;
3024 u64 extent_gen;
3025 u64 disk_bytenr = 0;
3026 u64 flags = 0;
3027 int extent_type;
3028 u8 compression;
3029
3030 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3031 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3032 break;
3033
3034 extent_end = btrfs_file_extent_end(path);
3035
3036 /*
3037 * The first iteration can leave us at an extent item that ends
3038 * before our range's start. Move to the next item.
3039 */
3040 if (extent_end <= range_start)
3041 goto next_item;
3042
3043 backref_ctx->curr_leaf_bytenr = leaf->start;
3044
3045 /* We have in implicit hole (NO_HOLES feature enabled). */
3046 if (prev_extent_end < key.offset) {
3047 const u64 hole_end = min(key.offset, range_end) - 1;
3048
3049 ret = fiemap_process_hole(inode, fieinfo, &cache,
3050 &delalloc_cached_state,
3051 backref_ctx, 0, 0, 0,
3052 prev_extent_end, hole_end);
3053 if (ret < 0) {
3054 goto out;
3055 } else if (ret > 0) {
3056 /* fiemap_fill_next_extent() told us to stop. */
3057 stopped = true;
3058 break;
3059 }
3060
3061 /* We've reached the end of the fiemap range, stop. */
3062 if (key.offset >= range_end) {
3063 stopped = true;
3064 break;
3065 }
3066 }
3067
3068 extent_len = extent_end - key.offset;
3069 ei = btrfs_item_ptr(leaf, path->slots[0],
3070 struct btrfs_file_extent_item);
3071 compression = btrfs_file_extent_compression(leaf, ei);
3072 extent_type = btrfs_file_extent_type(leaf, ei);
3073 extent_gen = btrfs_file_extent_generation(leaf, ei);
3074
3075 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
3076 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, ei);
3077 if (compression == BTRFS_COMPRESS_NONE)
3078 extent_offset = btrfs_file_extent_offset(leaf, ei);
3079 }
3080
3081 if (compression != BTRFS_COMPRESS_NONE)
3082 flags |= FIEMAP_EXTENT_ENCODED;
3083
3084 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
3085 flags |= FIEMAP_EXTENT_DATA_INLINE;
3086 flags |= FIEMAP_EXTENT_NOT_ALIGNED;
3087 ret = emit_fiemap_extent(fieinfo, &cache, key.offset, 0,
3088 extent_len, flags);
3089 } else if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
3090 ret = fiemap_process_hole(inode, fieinfo, &cache,
3091 &delalloc_cached_state,
3092 backref_ctx,
3093 disk_bytenr, extent_offset,
3094 extent_gen, key.offset,
3095 extent_end - 1);
3096 } else if (disk_bytenr == 0) {
3097 /* We have an explicit hole. */
3098 ret = fiemap_process_hole(inode, fieinfo, &cache,
3099 &delalloc_cached_state,
3100 backref_ctx, 0, 0, 0,
3101 key.offset, extent_end - 1);
3102 } else {
3103 /* We have a regular extent. */
3104 if (fieinfo->fi_extents_max) {
3105 ret = btrfs_is_data_extent_shared(inode,
3106 disk_bytenr,
3107 extent_gen,
3108 backref_ctx);
3109 if (ret < 0)
3110 goto out;
3111 else if (ret > 0)
3112 flags |= FIEMAP_EXTENT_SHARED;
3113 }
3114
3115 ret = emit_fiemap_extent(fieinfo, &cache, key.offset,
3116 disk_bytenr + extent_offset,
3117 extent_len, flags);
3118 }
3119
3120 if (ret < 0) {
3121 goto out;
3122 } else if (ret > 0) {
3123 /* fiemap_fill_next_extent() told us to stop. */
3124 stopped = true;
3125 break;
3126 }
3127
3128 prev_extent_end = extent_end;
3129next_item:
3130 if (fatal_signal_pending(current)) {
3131 ret = -EINTR;
3132 goto out;
3133 }
3134
3135 ret = fiemap_next_leaf_item(inode, path);
3136 if (ret < 0) {
3137 goto out;
3138 } else if (ret > 0) {
3139 /* No more file extent items for this inode. */
3140 break;
3141 }
3142 cond_resched();
3143 }
3144
3145check_eof_delalloc:
3146 /*
3147 * Release (and free) the path before emitting any final entries to
3148 * fiemap_fill_next_extent() to keep lockdep happy. This is because
3149 * once we find no more file extent items exist, we may have a
3150 * non-cloned leaf, and fiemap_fill_next_extent() can trigger page
3151 * faults when copying data to the user space buffer.
3152 */
3153 btrfs_free_path(path);
3154 path = NULL;
3155
3156 if (!stopped && prev_extent_end < range_end) {
3157 ret = fiemap_process_hole(inode, fieinfo, &cache,
3158 &delalloc_cached_state, backref_ctx,
3159 0, 0, 0, prev_extent_end, range_end - 1);
3160 if (ret < 0)
3161 goto out;
3162 prev_extent_end = range_end;
3163 }
3164
3165 if (cache.cached && cache.offset + cache.len >= last_extent_end) {
3166 const u64 i_size = i_size_read(&inode->vfs_inode);
3167
3168 if (prev_extent_end < i_size) {
3169 struct extent_state *cached_state = NULL;
3170 u64 delalloc_start;
3171 u64 delalloc_end;
3172 u64 lockstart;
3173 u64 lockend;
3174 bool delalloc;
3175
3176 lockstart = round_down(prev_extent_end, sectorsize);
3177 lockend = round_up(i_size, sectorsize);
3178
3179 /*
3180 * See the comment in fiemap_process_hole as to why
3181 * we're doing the locking here.
3182 */
3183 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3184 delalloc = btrfs_find_delalloc_in_range(inode,
3185 prev_extent_end,
3186 i_size - 1,
3187 &delalloc_cached_state,
3188 &delalloc_start,
3189 &delalloc_end);
3190 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3191 if (!delalloc)
3192 cache.flags |= FIEMAP_EXTENT_LAST;
3193 } else {
3194 cache.flags |= FIEMAP_EXTENT_LAST;
3195 }
3196 }
3197
3198 ret = emit_last_fiemap_cache(fieinfo, &cache);
3199out:
3200 free_extent_state(delalloc_cached_state);
3201 btrfs_free_backref_share_ctx(backref_ctx);
3202 btrfs_free_path(path);
3203 return ret;
3204}
3205
3206static void __free_extent_buffer(struct extent_buffer *eb)
3207{
3208 kmem_cache_free(extent_buffer_cache, eb);
3209}
3210
3211static int extent_buffer_under_io(const struct extent_buffer *eb)
3212{
3213 return (test_bit(EXTENT_BUFFER_WRITEBACK, &eb->bflags) ||
3214 test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3215}
3216
3217static bool folio_range_has_eb(struct btrfs_fs_info *fs_info, struct folio *folio)
3218{
3219 struct btrfs_subpage *subpage;
3220
3221 lockdep_assert_held(&folio->mapping->i_private_lock);
3222
3223 if (folio_test_private(folio)) {
3224 subpage = folio_get_private(folio);
3225 if (atomic_read(&subpage->eb_refs))
3226 return true;
3227 /*
3228 * Even there is no eb refs here, we may still have
3229 * end_page_read() call relying on page::private.
3230 */
3231 if (atomic_read(&subpage->readers))
3232 return true;
3233 }
3234 return false;
3235}
3236
3237static void detach_extent_buffer_folio(struct extent_buffer *eb, struct folio *folio)
3238{
3239 struct btrfs_fs_info *fs_info = eb->fs_info;
3240 const bool mapped = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3241
3242 /*
3243 * For mapped eb, we're going to change the folio private, which should
3244 * be done under the i_private_lock.
3245 */
3246 if (mapped)
3247 spin_lock(&folio->mapping->i_private_lock);
3248
3249 if (!folio_test_private(folio)) {
3250 if (mapped)
3251 spin_unlock(&folio->mapping->i_private_lock);
3252 return;
3253 }
3254
3255 if (fs_info->nodesize >= PAGE_SIZE) {
3256 /*
3257 * We do this since we'll remove the pages after we've
3258 * removed the eb from the radix tree, so we could race
3259 * and have this page now attached to the new eb. So
3260 * only clear folio if it's still connected to
3261 * this eb.
3262 */
3263 if (folio_test_private(folio) && folio_get_private(folio) == eb) {
3264 BUG_ON(test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags));
3265 BUG_ON(folio_test_dirty(folio));
3266 BUG_ON(folio_test_writeback(folio));
3267 /* We need to make sure we haven't be attached to a new eb. */
3268 folio_detach_private(folio);
3269 }
3270 if (mapped)
3271 spin_unlock(&folio->mapping->i_private_lock);
3272 return;
3273 }
3274
3275 /*
3276 * For subpage, we can have dummy eb with folio private attached. In
3277 * this case, we can directly detach the private as such folio is only
3278 * attached to one dummy eb, no sharing.
3279 */
3280 if (!mapped) {
3281 btrfs_detach_subpage(fs_info, folio);
3282 return;
3283 }
3284
3285 btrfs_folio_dec_eb_refs(fs_info, folio);
3286
3287 /*
3288 * We can only detach the folio private if there are no other ebs in the
3289 * page range and no unfinished IO.
3290 */
3291 if (!folio_range_has_eb(fs_info, folio))
3292 btrfs_detach_subpage(fs_info, folio);
3293
3294 spin_unlock(&folio->mapping->i_private_lock);
3295}
3296
3297/* Release all pages attached to the extent buffer */
3298static void btrfs_release_extent_buffer_pages(struct extent_buffer *eb)
3299{
3300 ASSERT(!extent_buffer_under_io(eb));
3301
3302 for (int i = 0; i < INLINE_EXTENT_BUFFER_PAGES; i++) {
3303 struct folio *folio = eb->folios[i];
3304
3305 if (!folio)
3306 continue;
3307
3308 detach_extent_buffer_folio(eb, folio);
3309
3310 /* One for when we allocated the folio. */
3311 folio_put(folio);
3312 }
3313}
3314
3315/*
3316 * Helper for releasing the extent buffer.
3317 */
3318static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3319{
3320 btrfs_release_extent_buffer_pages(eb);
3321 btrfs_leak_debug_del_eb(eb);
3322 __free_extent_buffer(eb);
3323}
3324
3325static struct extent_buffer *
3326__alloc_extent_buffer(struct btrfs_fs_info *fs_info, u64 start,
3327 unsigned long len)
3328{
3329 struct extent_buffer *eb = NULL;
3330
3331 eb = kmem_cache_zalloc(extent_buffer_cache, GFP_NOFS|__GFP_NOFAIL);
3332 eb->start = start;
3333 eb->len = len;
3334 eb->fs_info = fs_info;
3335 init_rwsem(&eb->lock);
3336
3337 btrfs_leak_debug_add_eb(eb);
3338
3339 spin_lock_init(&eb->refs_lock);
3340 atomic_set(&eb->refs, 1);
3341
3342 ASSERT(len <= BTRFS_MAX_METADATA_BLOCKSIZE);
3343
3344 return eb;
3345}
3346
3347struct extent_buffer *btrfs_clone_extent_buffer(const struct extent_buffer *src)
3348{
3349 struct extent_buffer *new;
3350 int num_folios = num_extent_folios(src);
3351 int ret;
3352
3353 new = __alloc_extent_buffer(src->fs_info, src->start, src->len);
3354 if (new == NULL)
3355 return NULL;
3356
3357 /*
3358 * Set UNMAPPED before calling btrfs_release_extent_buffer(), as
3359 * btrfs_release_extent_buffer() have different behavior for
3360 * UNMAPPED subpage extent buffer.
3361 */
3362 set_bit(EXTENT_BUFFER_UNMAPPED, &new->bflags);
3363
3364 ret = alloc_eb_folio_array(new, 0);
3365 if (ret) {
3366 btrfs_release_extent_buffer(new);
3367 return NULL;
3368 }
3369
3370 for (int i = 0; i < num_folios; i++) {
3371 struct folio *folio = new->folios[i];
3372 int ret;
3373
3374 ret = attach_extent_buffer_folio(new, folio, NULL);
3375 if (ret < 0) {
3376 btrfs_release_extent_buffer(new);
3377 return NULL;
3378 }
3379 WARN_ON(folio_test_dirty(folio));
3380 }
3381 copy_extent_buffer_full(new, src);
3382 set_extent_buffer_uptodate(new);
3383
3384 return new;
3385}
3386
3387struct extent_buffer *__alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
3388 u64 start, unsigned long len)
3389{
3390 struct extent_buffer *eb;
3391 int num_folios = 0;
3392 int ret;
3393
3394 eb = __alloc_extent_buffer(fs_info, start, len);
3395 if (!eb)
3396 return NULL;
3397
3398 ret = alloc_eb_folio_array(eb, 0);
3399 if (ret)
3400 goto err;
3401
3402 num_folios = num_extent_folios(eb);
3403 for (int i = 0; i < num_folios; i++) {
3404 ret = attach_extent_buffer_folio(eb, eb->folios[i], NULL);
3405 if (ret < 0)
3406 goto err;
3407 }
3408
3409 set_extent_buffer_uptodate(eb);
3410 btrfs_set_header_nritems(eb, 0);
3411 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3412
3413 return eb;
3414err:
3415 for (int i = 0; i < num_folios; i++) {
3416 if (eb->folios[i]) {
3417 detach_extent_buffer_folio(eb, eb->folios[i]);
3418 __folio_put(eb->folios[i]);
3419 }
3420 }
3421 __free_extent_buffer(eb);
3422 return NULL;
3423}
3424
3425struct extent_buffer *alloc_dummy_extent_buffer(struct btrfs_fs_info *fs_info,
3426 u64 start)
3427{
3428 return __alloc_dummy_extent_buffer(fs_info, start, fs_info->nodesize);
3429}
3430
3431static void check_buffer_tree_ref(struct extent_buffer *eb)
3432{
3433 int refs;
3434 /*
3435 * The TREE_REF bit is first set when the extent_buffer is added
3436 * to the radix tree. It is also reset, if unset, when a new reference
3437 * is created by find_extent_buffer.
3438 *
3439 * It is only cleared in two cases: freeing the last non-tree
3440 * reference to the extent_buffer when its STALE bit is set or
3441 * calling release_folio when the tree reference is the only reference.
3442 *
3443 * In both cases, care is taken to ensure that the extent_buffer's
3444 * pages are not under io. However, release_folio can be concurrently
3445 * called with creating new references, which is prone to race
3446 * conditions between the calls to check_buffer_tree_ref in those
3447 * codepaths and clearing TREE_REF in try_release_extent_buffer.
3448 *
3449 * The actual lifetime of the extent_buffer in the radix tree is
3450 * adequately protected by the refcount, but the TREE_REF bit and
3451 * its corresponding reference are not. To protect against this
3452 * class of races, we call check_buffer_tree_ref from the codepaths
3453 * which trigger io. Note that once io is initiated, TREE_REF can no
3454 * longer be cleared, so that is the moment at which any such race is
3455 * best fixed.
3456 */
3457 refs = atomic_read(&eb->refs);
3458 if (refs >= 2 && test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3459 return;
3460
3461 spin_lock(&eb->refs_lock);
3462 if (!test_and_set_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3463 atomic_inc(&eb->refs);
3464 spin_unlock(&eb->refs_lock);
3465}
3466
3467static void mark_extent_buffer_accessed(struct extent_buffer *eb)
3468{
3469 int num_folios= num_extent_folios(eb);
3470
3471 check_buffer_tree_ref(eb);
3472
3473 for (int i = 0; i < num_folios; i++)
3474 folio_mark_accessed(eb->folios[i]);
3475}
3476
3477struct extent_buffer *find_extent_buffer(struct btrfs_fs_info *fs_info,
3478 u64 start)
3479{
3480 struct extent_buffer *eb;
3481
3482 eb = find_extent_buffer_nolock(fs_info, start);
3483 if (!eb)
3484 return NULL;
3485 /*
3486 * Lock our eb's refs_lock to avoid races with free_extent_buffer().
3487 * When we get our eb it might be flagged with EXTENT_BUFFER_STALE and
3488 * another task running free_extent_buffer() might have seen that flag
3489 * set, eb->refs == 2, that the buffer isn't under IO (dirty and
3490 * writeback flags not set) and it's still in the tree (flag
3491 * EXTENT_BUFFER_TREE_REF set), therefore being in the process of
3492 * decrementing the extent buffer's reference count twice. So here we
3493 * could race and increment the eb's reference count, clear its stale
3494 * flag, mark it as dirty and drop our reference before the other task
3495 * finishes executing free_extent_buffer, which would later result in
3496 * an attempt to free an extent buffer that is dirty.
3497 */
3498 if (test_bit(EXTENT_BUFFER_STALE, &eb->bflags)) {
3499 spin_lock(&eb->refs_lock);
3500 spin_unlock(&eb->refs_lock);
3501 }
3502 mark_extent_buffer_accessed(eb);
3503 return eb;
3504}
3505
3506#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3507struct extent_buffer *alloc_test_extent_buffer(struct btrfs_fs_info *fs_info,
3508 u64 start)
3509{
3510 struct extent_buffer *eb, *exists = NULL;
3511 int ret;
3512
3513 eb = find_extent_buffer(fs_info, start);
3514 if (eb)
3515 return eb;
3516 eb = alloc_dummy_extent_buffer(fs_info, start);
3517 if (!eb)
3518 return ERR_PTR(-ENOMEM);
3519 eb->fs_info = fs_info;
3520again:
3521 ret = radix_tree_preload(GFP_NOFS);
3522 if (ret) {
3523 exists = ERR_PTR(ret);
3524 goto free_eb;
3525 }
3526 spin_lock(&fs_info->buffer_lock);
3527 ret = radix_tree_insert(&fs_info->buffer_radix,
3528 start >> fs_info->sectorsize_bits, eb);
3529 spin_unlock(&fs_info->buffer_lock);
3530 radix_tree_preload_end();
3531 if (ret == -EEXIST) {
3532 exists = find_extent_buffer(fs_info, start);
3533 if (exists)
3534 goto free_eb;
3535 else
3536 goto again;
3537 }
3538 check_buffer_tree_ref(eb);
3539 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
3540
3541 return eb;
3542free_eb:
3543 btrfs_release_extent_buffer(eb);
3544 return exists;
3545}
3546#endif
3547
3548static struct extent_buffer *grab_extent_buffer(
3549 struct btrfs_fs_info *fs_info, struct page *page)
3550{
3551 struct folio *folio = page_folio(page);
3552 struct extent_buffer *exists;
3553
3554 /*
3555 * For subpage case, we completely rely on radix tree to ensure we
3556 * don't try to insert two ebs for the same bytenr. So here we always
3557 * return NULL and just continue.
3558 */
3559 if (fs_info->nodesize < PAGE_SIZE)
3560 return NULL;
3561
3562 /* Page not yet attached to an extent buffer */
3563 if (!folio_test_private(folio))
3564 return NULL;
3565
3566 /*
3567 * We could have already allocated an eb for this page and attached one
3568 * so lets see if we can get a ref on the existing eb, and if we can we
3569 * know it's good and we can just return that one, else we know we can
3570 * just overwrite folio private.
3571 */
3572 exists = folio_get_private(folio);
3573 if (atomic_inc_not_zero(&exists->refs))
3574 return exists;
3575
3576 WARN_ON(PageDirty(page));
3577 folio_detach_private(folio);
3578 return NULL;
3579}
3580
3581static int check_eb_alignment(struct btrfs_fs_info *fs_info, u64 start)
3582{
3583 if (!IS_ALIGNED(start, fs_info->sectorsize)) {
3584 btrfs_err(fs_info, "bad tree block start %llu", start);
3585 return -EINVAL;
3586 }
3587
3588 if (fs_info->nodesize < PAGE_SIZE &&
3589 offset_in_page(start) + fs_info->nodesize > PAGE_SIZE) {
3590 btrfs_err(fs_info,
3591 "tree block crosses page boundary, start %llu nodesize %u",
3592 start, fs_info->nodesize);
3593 return -EINVAL;
3594 }
3595 if (fs_info->nodesize >= PAGE_SIZE &&
3596 !PAGE_ALIGNED(start)) {
3597 btrfs_err(fs_info,
3598 "tree block is not page aligned, start %llu nodesize %u",
3599 start, fs_info->nodesize);
3600 return -EINVAL;
3601 }
3602 if (!IS_ALIGNED(start, fs_info->nodesize) &&
3603 !test_and_set_bit(BTRFS_FS_UNALIGNED_TREE_BLOCK, &fs_info->flags)) {
3604 btrfs_warn(fs_info,
3605"tree block not nodesize aligned, start %llu nodesize %u, can be resolved by a full metadata balance",
3606 start, fs_info->nodesize);
3607 }
3608 return 0;
3609}
3610
3611
3612/*
3613 * Return 0 if eb->folios[i] is attached to btree inode successfully.
3614 * Return >0 if there is already another extent buffer for the range,
3615 * and @found_eb_ret would be updated.
3616 * Return -EAGAIN if the filemap has an existing folio but with different size
3617 * than @eb.
3618 * The caller needs to free the existing folios and retry using the same order.
3619 */
3620static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i,
3621 struct extent_buffer **found_eb_ret)
3622{
3623
3624 struct btrfs_fs_info *fs_info = eb->fs_info;
3625 struct address_space *mapping = fs_info->btree_inode->i_mapping;
3626 const unsigned long index = eb->start >> PAGE_SHIFT;
3627 struct folio *existing_folio;
3628 int ret;
3629
3630 ASSERT(found_eb_ret);
3631
3632 /* Caller should ensure the folio exists. */
3633 ASSERT(eb->folios[i]);
3634
3635retry:
3636 ret = filemap_add_folio(mapping, eb->folios[i], index + i,
3637 GFP_NOFS | __GFP_NOFAIL);
3638 if (!ret)
3639 return 0;
3640
3641 existing_folio = filemap_lock_folio(mapping, index + i);
3642 /* The page cache only exists for a very short time, just retry. */
3643 if (IS_ERR(existing_folio))
3644 goto retry;
3645
3646 /* For now, we should only have single-page folios for btree inode. */
3647 ASSERT(folio_nr_pages(existing_folio) == 1);
3648
3649 if (folio_size(existing_folio) != folio_size(eb->folios[0])) {
3650 folio_unlock(existing_folio);
3651 folio_put(existing_folio);
3652 return -EAGAIN;
3653 }
3654
3655 if (fs_info->nodesize < PAGE_SIZE) {
3656 /*
3657 * We're going to reuse the existing page, can drop our page
3658 * and subpage structure now.
3659 */
3660 __free_page(folio_page(eb->folios[i], 0));
3661 eb->folios[i] = existing_folio;
3662 } else {
3663 struct extent_buffer *existing_eb;
3664
3665 existing_eb = grab_extent_buffer(fs_info,
3666 folio_page(existing_folio, 0));
3667 if (existing_eb) {
3668 /* The extent buffer still exists, we can use it directly. */
3669 *found_eb_ret = existing_eb;
3670 folio_unlock(existing_folio);
3671 folio_put(existing_folio);
3672 return 1;
3673 }
3674 /* The extent buffer no longer exists, we can reuse the folio. */
3675 __free_page(folio_page(eb->folios[i], 0));
3676 eb->folios[i] = existing_folio;
3677 }
3678 return 0;
3679}
3680
3681struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
3682 u64 start, u64 owner_root, int level)
3683{
3684 unsigned long len = fs_info->nodesize;
3685 int num_folios;
3686 int attached = 0;
3687 struct extent_buffer *eb;
3688 struct extent_buffer *existing_eb = NULL;
3689 struct address_space *mapping = fs_info->btree_inode->i_mapping;
3690 struct btrfs_subpage *prealloc = NULL;
3691 u64 lockdep_owner = owner_root;
3692 bool page_contig = true;
3693 int uptodate = 1;
3694 int ret;
3695
3696 if (check_eb_alignment(fs_info, start))
3697 return ERR_PTR(-EINVAL);
3698
3699#if BITS_PER_LONG == 32
3700 if (start >= MAX_LFS_FILESIZE) {
3701 btrfs_err_rl(fs_info,
3702 "extent buffer %llu is beyond 32bit page cache limit", start);
3703 btrfs_err_32bit_limit(fs_info);
3704 return ERR_PTR(-EOVERFLOW);
3705 }
3706 if (start >= BTRFS_32BIT_EARLY_WARN_THRESHOLD)
3707 btrfs_warn_32bit_limit(fs_info);
3708#endif
3709
3710 eb = find_extent_buffer(fs_info, start);
3711 if (eb)
3712 return eb;
3713
3714 eb = __alloc_extent_buffer(fs_info, start, len);
3715 if (!eb)
3716 return ERR_PTR(-ENOMEM);
3717
3718 /*
3719 * The reloc trees are just snapshots, so we need them to appear to be
3720 * just like any other fs tree WRT lockdep.
3721 */
3722 if (lockdep_owner == BTRFS_TREE_RELOC_OBJECTID)
3723 lockdep_owner = BTRFS_FS_TREE_OBJECTID;
3724
3725 btrfs_set_buffer_lockdep_class(lockdep_owner, eb, level);
3726
3727 /*
3728 * Preallocate folio private for subpage case, so that we won't
3729 * allocate memory with i_private_lock nor page lock hold.
3730 *
3731 * The memory will be freed by attach_extent_buffer_page() or freed
3732 * manually if we exit earlier.
3733 */
3734 if (fs_info->nodesize < PAGE_SIZE) {
3735 prealloc = btrfs_alloc_subpage(fs_info, BTRFS_SUBPAGE_METADATA);
3736 if (IS_ERR(prealloc)) {
3737 ret = PTR_ERR(prealloc);
3738 goto out;
3739 }
3740 }
3741
3742reallocate:
3743 /* Allocate all pages first. */
3744 ret = alloc_eb_folio_array(eb, __GFP_NOFAIL);
3745 if (ret < 0) {
3746 btrfs_free_subpage(prealloc);
3747 goto out;
3748 }
3749
3750 num_folios = num_extent_folios(eb);
3751 /* Attach all pages to the filemap. */
3752 for (int i = 0; i < num_folios; i++) {
3753 struct folio *folio;
3754
3755 ret = attach_eb_folio_to_filemap(eb, i, &existing_eb);
3756 if (ret > 0) {
3757 ASSERT(existing_eb);
3758 goto out;
3759 }
3760
3761 /*
3762 * TODO: Special handling for a corner case where the order of
3763 * folios mismatch between the new eb and filemap.
3764 *
3765 * This happens when:
3766 *
3767 * - the new eb is using higher order folio
3768 *
3769 * - the filemap is still using 0-order folios for the range
3770 * This can happen at the previous eb allocation, and we don't
3771 * have higher order folio for the call.
3772 *
3773 * - the existing eb has already been freed
3774 *
3775 * In this case, we have to free the existing folios first, and
3776 * re-allocate using the same order.
3777 * Thankfully this is not going to happen yet, as we're still
3778 * using 0-order folios.
3779 */
3780 if (unlikely(ret == -EAGAIN)) {
3781 ASSERT(0);
3782 goto reallocate;
3783 }
3784 attached++;
3785
3786 /*
3787 * Only after attach_eb_folio_to_filemap(), eb->folios[] is
3788 * reliable, as we may choose to reuse the existing page cache
3789 * and free the allocated page.
3790 */
3791 folio = eb->folios[i];
3792 spin_lock(&mapping->i_private_lock);
3793 /* Should not fail, as we have preallocated the memory */
3794 ret = attach_extent_buffer_folio(eb, folio, prealloc);
3795 ASSERT(!ret);
3796 /*
3797 * To inform we have extra eb under allocation, so that
3798 * detach_extent_buffer_page() won't release the folio private
3799 * when the eb hasn't yet been inserted into radix tree.
3800 *
3801 * The ref will be decreased when the eb released the page, in
3802 * detach_extent_buffer_page().
3803 * Thus needs no special handling in error path.
3804 */
3805 btrfs_folio_inc_eb_refs(fs_info, folio);
3806 spin_unlock(&mapping->i_private_lock);
3807
3808 WARN_ON(btrfs_folio_test_dirty(fs_info, folio, eb->start, eb->len));
3809
3810 /*
3811 * Check if the current page is physically contiguous with previous eb
3812 * page.
3813 * At this stage, either we allocated a large folio, thus @i
3814 * would only be 0, or we fall back to per-page allocation.
3815 */
3816 if (i && folio_page(eb->folios[i - 1], 0) + 1 != folio_page(folio, 0))
3817 page_contig = false;
3818
3819 if (!btrfs_folio_test_uptodate(fs_info, folio, eb->start, eb->len))
3820 uptodate = 0;
3821
3822 /*
3823 * We can't unlock the pages just yet since the extent buffer
3824 * hasn't been properly inserted in the radix tree, this
3825 * opens a race with btree_release_folio which can free a page
3826 * while we are still filling in all pages for the buffer and
3827 * we could crash.
3828 */
3829 }
3830 if (uptodate)
3831 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3832 /* All pages are physically contiguous, can skip cross page handling. */
3833 if (page_contig)
3834 eb->addr = folio_address(eb->folios[0]) + offset_in_page(eb->start);
3835again:
3836 ret = radix_tree_preload(GFP_NOFS);
3837 if (ret)
3838 goto out;
3839
3840 spin_lock(&fs_info->buffer_lock);
3841 ret = radix_tree_insert(&fs_info->buffer_radix,
3842 start >> fs_info->sectorsize_bits, eb);
3843 spin_unlock(&fs_info->buffer_lock);
3844 radix_tree_preload_end();
3845 if (ret == -EEXIST) {
3846 ret = 0;
3847 existing_eb = find_extent_buffer(fs_info, start);
3848 if (existing_eb)
3849 goto out;
3850 else
3851 goto again;
3852 }
3853 /* add one reference for the tree */
3854 check_buffer_tree_ref(eb);
3855 set_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags);
3856
3857 /*
3858 * Now it's safe to unlock the pages because any calls to
3859 * btree_release_folio will correctly detect that a page belongs to a
3860 * live buffer and won't free them prematurely.
3861 */
3862 for (int i = 0; i < num_folios; i++)
3863 unlock_page(folio_page(eb->folios[i], 0));
3864 return eb;
3865
3866out:
3867 WARN_ON(!atomic_dec_and_test(&eb->refs));
3868
3869 /*
3870 * Any attached folios need to be detached before we unlock them. This
3871 * is because when we're inserting our new folios into the mapping, and
3872 * then attaching our eb to that folio. If we fail to insert our folio
3873 * we'll lookup the folio for that index, and grab that EB. We do not
3874 * want that to grab this eb, as we're getting ready to free it. So we
3875 * have to detach it first and then unlock it.
3876 *
3877 * We have to drop our reference and NULL it out here because in the
3878 * subpage case detaching does a btrfs_folio_dec_eb_refs() for our eb.
3879 * Below when we call btrfs_release_extent_buffer() we will call
3880 * detach_extent_buffer_folio() on our remaining pages in the !subpage
3881 * case. If we left eb->folios[i] populated in the subpage case we'd
3882 * double put our reference and be super sad.
3883 */
3884 for (int i = 0; i < attached; i++) {
3885 ASSERT(eb->folios[i]);
3886 detach_extent_buffer_folio(eb, eb->folios[i]);
3887 unlock_page(folio_page(eb->folios[i], 0));
3888 folio_put(eb->folios[i]);
3889 eb->folios[i] = NULL;
3890 }
3891 /*
3892 * Now all pages of that extent buffer is unmapped, set UNMAPPED flag,
3893 * so it can be cleaned up without utlizing page->mapping.
3894 */
3895 set_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
3896
3897 btrfs_release_extent_buffer(eb);
3898 if (ret < 0)
3899 return ERR_PTR(ret);
3900 ASSERT(existing_eb);
3901 return existing_eb;
3902}
3903
3904static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3905{
3906 struct extent_buffer *eb =
3907 container_of(head, struct extent_buffer, rcu_head);
3908
3909 __free_extent_buffer(eb);
3910}
3911
3912static int release_extent_buffer(struct extent_buffer *eb)
3913 __releases(&eb->refs_lock)
3914{
3915 lockdep_assert_held(&eb->refs_lock);
3916
3917 WARN_ON(atomic_read(&eb->refs) == 0);
3918 if (atomic_dec_and_test(&eb->refs)) {
3919 if (test_and_clear_bit(EXTENT_BUFFER_IN_TREE, &eb->bflags)) {
3920 struct btrfs_fs_info *fs_info = eb->fs_info;
3921
3922 spin_unlock(&eb->refs_lock);
3923
3924 spin_lock(&fs_info->buffer_lock);
3925 radix_tree_delete(&fs_info->buffer_radix,
3926 eb->start >> fs_info->sectorsize_bits);
3927 spin_unlock(&fs_info->buffer_lock);
3928 } else {
3929 spin_unlock(&eb->refs_lock);
3930 }
3931
3932 btrfs_leak_debug_del_eb(eb);
3933 /* Should be safe to release our pages at this point */
3934 btrfs_release_extent_buffer_pages(eb);
3935#ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3936 if (unlikely(test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags))) {
3937 __free_extent_buffer(eb);
3938 return 1;
3939 }
3940#endif
3941 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3942 return 1;
3943 }
3944 spin_unlock(&eb->refs_lock);
3945
3946 return 0;
3947}
3948
3949void free_extent_buffer(struct extent_buffer *eb)
3950{
3951 int refs;
3952 if (!eb)
3953 return;
3954
3955 refs = atomic_read(&eb->refs);
3956 while (1) {
3957 if ((!test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) && refs <= 3)
3958 || (test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags) &&
3959 refs == 1))
3960 break;
3961 if (atomic_try_cmpxchg(&eb->refs, &refs, refs - 1))
3962 return;
3963 }
3964
3965 spin_lock(&eb->refs_lock);
3966 if (atomic_read(&eb->refs) == 2 &&
3967 test_bit(EXTENT_BUFFER_STALE, &eb->bflags) &&
3968 !extent_buffer_under_io(eb) &&
3969 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3970 atomic_dec(&eb->refs);
3971
3972 /*
3973 * I know this is terrible, but it's temporary until we stop tracking
3974 * the uptodate bits and such for the extent buffers.
3975 */
3976 release_extent_buffer(eb);
3977}
3978
3979void free_extent_buffer_stale(struct extent_buffer *eb)
3980{
3981 if (!eb)
3982 return;
3983
3984 spin_lock(&eb->refs_lock);
3985 set_bit(EXTENT_BUFFER_STALE, &eb->bflags);
3986
3987 if (atomic_read(&eb->refs) == 2 && !extent_buffer_under_io(eb) &&
3988 test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags))
3989 atomic_dec(&eb->refs);
3990 release_extent_buffer(eb);
3991}
3992
3993static void btree_clear_folio_dirty(struct folio *folio)
3994{
3995 ASSERT(folio_test_dirty(folio));
3996 ASSERT(folio_test_locked(folio));
3997 folio_clear_dirty_for_io(folio);
3998 xa_lock_irq(&folio->mapping->i_pages);
3999 if (!folio_test_dirty(folio))
4000 __xa_clear_mark(&folio->mapping->i_pages,
4001 folio_index(folio), PAGECACHE_TAG_DIRTY);
4002 xa_unlock_irq(&folio->mapping->i_pages);
4003}
4004
4005static void clear_subpage_extent_buffer_dirty(const struct extent_buffer *eb)
4006{
4007 struct btrfs_fs_info *fs_info = eb->fs_info;
4008 struct folio *folio = eb->folios[0];
4009 bool last;
4010
4011 /* btree_clear_folio_dirty() needs page locked. */
4012 folio_lock(folio);
4013 last = btrfs_subpage_clear_and_test_dirty(fs_info, folio, eb->start, eb->len);
4014 if (last)
4015 btree_clear_folio_dirty(folio);
4016 folio_unlock(folio);
4017 WARN_ON(atomic_read(&eb->refs) == 0);
4018}
4019
4020void btrfs_clear_buffer_dirty(struct btrfs_trans_handle *trans,
4021 struct extent_buffer *eb)
4022{
4023 struct btrfs_fs_info *fs_info = eb->fs_info;
4024 int num_folios;
4025
4026 btrfs_assert_tree_write_locked(eb);
4027
4028 if (trans && btrfs_header_generation(eb) != trans->transid)
4029 return;
4030
4031 /*
4032 * Instead of clearing the dirty flag off of the buffer, mark it as
4033 * EXTENT_BUFFER_ZONED_ZEROOUT. This allows us to preserve
4034 * write-ordering in zoned mode, without the need to later re-dirty
4035 * the extent_buffer.
4036 *
4037 * The actual zeroout of the buffer will happen later in
4038 * btree_csum_one_bio.
4039 */
4040 if (btrfs_is_zoned(fs_info)) {
4041 set_bit(EXTENT_BUFFER_ZONED_ZEROOUT, &eb->bflags);
4042 return;
4043 }
4044
4045 if (!test_and_clear_bit(EXTENT_BUFFER_DIRTY, &eb->bflags))
4046 return;
4047
4048 percpu_counter_add_batch(&fs_info->dirty_metadata_bytes, -eb->len,
4049 fs_info->dirty_metadata_batch);
4050
4051 if (eb->fs_info->nodesize < PAGE_SIZE)
4052 return clear_subpage_extent_buffer_dirty(eb);
4053
4054 num_folios = num_extent_folios(eb);
4055 for (int i = 0; i < num_folios; i++) {
4056 struct folio *folio = eb->folios[i];
4057
4058 if (!folio_test_dirty(folio))
4059 continue;
4060 folio_lock(folio);
4061 btree_clear_folio_dirty(folio);
4062 folio_unlock(folio);
4063 }
4064 WARN_ON(atomic_read(&eb->refs) == 0);
4065}
4066
4067void set_extent_buffer_dirty(struct extent_buffer *eb)
4068{
4069 int num_folios;
4070 bool was_dirty;
4071
4072 check_buffer_tree_ref(eb);
4073
4074 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
4075
4076 num_folios = num_extent_folios(eb);
4077 WARN_ON(atomic_read(&eb->refs) == 0);
4078 WARN_ON(!test_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags));
4079
4080 if (!was_dirty) {
4081 bool subpage = eb->fs_info->nodesize < PAGE_SIZE;
4082
4083 /*
4084 * For subpage case, we can have other extent buffers in the
4085 * same page, and in clear_subpage_extent_buffer_dirty() we
4086 * have to clear page dirty without subpage lock held.
4087 * This can cause race where our page gets dirty cleared after
4088 * we just set it.
4089 *
4090 * Thankfully, clear_subpage_extent_buffer_dirty() has locked
4091 * its page for other reasons, we can use page lock to prevent
4092 * the above race.
4093 */
4094 if (subpage)
4095 lock_page(folio_page(eb->folios[0], 0));
4096 for (int i = 0; i < num_folios; i++)
4097 btrfs_folio_set_dirty(eb->fs_info, eb->folios[i],
4098 eb->start, eb->len);
4099 if (subpage)
4100 unlock_page(folio_page(eb->folios[0], 0));
4101 percpu_counter_add_batch(&eb->fs_info->dirty_metadata_bytes,
4102 eb->len,
4103 eb->fs_info->dirty_metadata_batch);
4104 }
4105#ifdef CONFIG_BTRFS_DEBUG
4106 for (int i = 0; i < num_folios; i++)
4107 ASSERT(folio_test_dirty(eb->folios[i]));
4108#endif
4109}
4110
4111void clear_extent_buffer_uptodate(struct extent_buffer *eb)
4112{
4113 struct btrfs_fs_info *fs_info = eb->fs_info;
4114 int num_folios = num_extent_folios(eb);
4115
4116 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4117 for (int i = 0; i < num_folios; i++) {
4118 struct folio *folio = eb->folios[i];
4119
4120 if (!folio)
4121 continue;
4122
4123 /*
4124 * This is special handling for metadata subpage, as regular
4125 * btrfs_is_subpage() can not handle cloned/dummy metadata.
4126 */
4127 if (fs_info->nodesize >= PAGE_SIZE)
4128 folio_clear_uptodate(folio);
4129 else
4130 btrfs_subpage_clear_uptodate(fs_info, folio,
4131 eb->start, eb->len);
4132 }
4133}
4134
4135void set_extent_buffer_uptodate(struct extent_buffer *eb)
4136{
4137 struct btrfs_fs_info *fs_info = eb->fs_info;
4138 int num_folios = num_extent_folios(eb);
4139
4140 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
4141 for (int i = 0; i < num_folios; i++) {
4142 struct folio *folio = eb->folios[i];
4143
4144 /*
4145 * This is special handling for metadata subpage, as regular
4146 * btrfs_is_subpage() can not handle cloned/dummy metadata.
4147 */
4148 if (fs_info->nodesize >= PAGE_SIZE)
4149 folio_mark_uptodate(folio);
4150 else
4151 btrfs_subpage_set_uptodate(fs_info, folio,
4152 eb->start, eb->len);
4153 }
4154}
4155
4156static void end_bbio_meta_read(struct btrfs_bio *bbio)
4157{
4158 struct extent_buffer *eb = bbio->private;
4159 struct btrfs_fs_info *fs_info = eb->fs_info;
4160 bool uptodate = !bbio->bio.bi_status;
4161 struct folio_iter fi;
4162 u32 bio_offset = 0;
4163
4164 eb->read_mirror = bbio->mirror_num;
4165
4166 if (uptodate &&
4167 btrfs_validate_extent_buffer(eb, &bbio->parent_check) < 0)
4168 uptodate = false;
4169
4170 if (uptodate) {
4171 set_extent_buffer_uptodate(eb);
4172 } else {
4173 clear_extent_buffer_uptodate(eb);
4174 set_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
4175 }
4176
4177 bio_for_each_folio_all(fi, &bbio->bio) {
4178 struct folio *folio = fi.folio;
4179 u64 start = eb->start + bio_offset;
4180 u32 len = fi.length;
4181
4182 if (uptodate)
4183 btrfs_folio_set_uptodate(fs_info, folio, start, len);
4184 else
4185 btrfs_folio_clear_uptodate(fs_info, folio, start, len);
4186
4187 bio_offset += len;
4188 }
4189
4190 clear_bit(EXTENT_BUFFER_READING, &eb->bflags);
4191 smp_mb__after_atomic();
4192 wake_up_bit(&eb->bflags, EXTENT_BUFFER_READING);
4193 free_extent_buffer(eb);
4194
4195 bio_put(&bbio->bio);
4196}
4197
4198int read_extent_buffer_pages(struct extent_buffer *eb, int wait, int mirror_num,
4199 struct btrfs_tree_parent_check *check)
4200{
4201 struct btrfs_bio *bbio;
4202 bool ret;
4203
4204 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4205 return 0;
4206
4207 /*
4208 * We could have had EXTENT_BUFFER_UPTODATE cleared by the write
4209 * operation, which could potentially still be in flight. In this case
4210 * we simply want to return an error.
4211 */
4212 if (unlikely(test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags)))
4213 return -EIO;
4214
4215 /* Someone else is already reading the buffer, just wait for it. */
4216 if (test_and_set_bit(EXTENT_BUFFER_READING, &eb->bflags))
4217 goto done;
4218
4219 clear_bit(EXTENT_BUFFER_READ_ERR, &eb->bflags);
4220 eb->read_mirror = 0;
4221 check_buffer_tree_ref(eb);
4222 atomic_inc(&eb->refs);
4223
4224 bbio = btrfs_bio_alloc(INLINE_EXTENT_BUFFER_PAGES,
4225 REQ_OP_READ | REQ_META, eb->fs_info,
4226 end_bbio_meta_read, eb);
4227 bbio->bio.bi_iter.bi_sector = eb->start >> SECTOR_SHIFT;
4228 bbio->inode = BTRFS_I(eb->fs_info->btree_inode);
4229 bbio->file_offset = eb->start;
4230 memcpy(&bbio->parent_check, check, sizeof(*check));
4231 if (eb->fs_info->nodesize < PAGE_SIZE) {
4232 ret = bio_add_folio(&bbio->bio, eb->folios[0], eb->len,
4233 eb->start - folio_pos(eb->folios[0]));
4234 ASSERT(ret);
4235 } else {
4236 int num_folios = num_extent_folios(eb);
4237
4238 for (int i = 0; i < num_folios; i++) {
4239 struct folio *folio = eb->folios[i];
4240
4241 ret = bio_add_folio(&bbio->bio, folio, folio_size(folio), 0);
4242 ASSERT(ret);
4243 }
4244 }
4245 btrfs_submit_bio(bbio, mirror_num);
4246
4247done:
4248 if (wait == WAIT_COMPLETE) {
4249 wait_on_bit_io(&eb->bflags, EXTENT_BUFFER_READING, TASK_UNINTERRUPTIBLE);
4250 if (!test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
4251 return -EIO;
4252 }
4253
4254 return 0;
4255}
4256
4257static bool report_eb_range(const struct extent_buffer *eb, unsigned long start,
4258 unsigned long len)
4259{
4260 btrfs_warn(eb->fs_info,
4261 "access to eb bytenr %llu len %lu out of range start %lu len %lu",
4262 eb->start, eb->len, start, len);
4263 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4264
4265 return true;
4266}
4267
4268/*
4269 * Check if the [start, start + len) range is valid before reading/writing
4270 * the eb.
4271 * NOTE: @start and @len are offset inside the eb, not logical address.
4272 *
4273 * Caller should not touch the dst/src memory if this function returns error.
4274 */
4275static inline int check_eb_range(const struct extent_buffer *eb,
4276 unsigned long start, unsigned long len)
4277{
4278 unsigned long offset;
4279
4280 /* start, start + len should not go beyond eb->len nor overflow */
4281 if (unlikely(check_add_overflow(start, len, &offset) || offset > eb->len))
4282 return report_eb_range(eb, start, len);
4283
4284 return false;
4285}
4286
4287void read_extent_buffer(const struct extent_buffer *eb, void *dstv,
4288 unsigned long start, unsigned long len)
4289{
4290 const int unit_size = folio_size(eb->folios[0]);
4291 size_t cur;
4292 size_t offset;
4293 char *dst = (char *)dstv;
4294 unsigned long i = get_eb_folio_index(eb, start);
4295
4296 if (check_eb_range(eb, start, len)) {
4297 /*
4298 * Invalid range hit, reset the memory, so callers won't get
4299 * some random garbage for their uninitialized memory.
4300 */
4301 memset(dstv, 0, len);
4302 return;
4303 }
4304
4305 if (eb->addr) {
4306 memcpy(dstv, eb->addr + start, len);
4307 return;
4308 }
4309
4310 offset = get_eb_offset_in_folio(eb, start);
4311
4312 while (len > 0) {
4313 char *kaddr;
4314
4315 cur = min(len, unit_size - offset);
4316 kaddr = folio_address(eb->folios[i]);
4317 memcpy(dst, kaddr + offset, cur);
4318
4319 dst += cur;
4320 len -= cur;
4321 offset = 0;
4322 i++;
4323 }
4324}
4325
4326int read_extent_buffer_to_user_nofault(const struct extent_buffer *eb,
4327 void __user *dstv,
4328 unsigned long start, unsigned long len)
4329{
4330 const int unit_size = folio_size(eb->folios[0]);
4331 size_t cur;
4332 size_t offset;
4333 char __user *dst = (char __user *)dstv;
4334 unsigned long i = get_eb_folio_index(eb, start);
4335 int ret = 0;
4336
4337 WARN_ON(start > eb->len);
4338 WARN_ON(start + len > eb->start + eb->len);
4339
4340 if (eb->addr) {
4341 if (copy_to_user_nofault(dstv, eb->addr + start, len))
4342 ret = -EFAULT;
4343 return ret;
4344 }
4345
4346 offset = get_eb_offset_in_folio(eb, start);
4347
4348 while (len > 0) {
4349 char *kaddr;
4350
4351 cur = min(len, unit_size - offset);
4352 kaddr = folio_address(eb->folios[i]);
4353 if (copy_to_user_nofault(dst, kaddr + offset, cur)) {
4354 ret = -EFAULT;
4355 break;
4356 }
4357
4358 dst += cur;
4359 len -= cur;
4360 offset = 0;
4361 i++;
4362 }
4363
4364 return ret;
4365}
4366
4367int memcmp_extent_buffer(const struct extent_buffer *eb, const void *ptrv,
4368 unsigned long start, unsigned long len)
4369{
4370 const int unit_size = folio_size(eb->folios[0]);
4371 size_t cur;
4372 size_t offset;
4373 char *kaddr;
4374 char *ptr = (char *)ptrv;
4375 unsigned long i = get_eb_folio_index(eb, start);
4376 int ret = 0;
4377
4378 if (check_eb_range(eb, start, len))
4379 return -EINVAL;
4380
4381 if (eb->addr)
4382 return memcmp(ptrv, eb->addr + start, len);
4383
4384 offset = get_eb_offset_in_folio(eb, start);
4385
4386 while (len > 0) {
4387 cur = min(len, unit_size - offset);
4388 kaddr = folio_address(eb->folios[i]);
4389 ret = memcmp(ptr, kaddr + offset, cur);
4390 if (ret)
4391 break;
4392
4393 ptr += cur;
4394 len -= cur;
4395 offset = 0;
4396 i++;
4397 }
4398 return ret;
4399}
4400
4401/*
4402 * Check that the extent buffer is uptodate.
4403 *
4404 * For regular sector size == PAGE_SIZE case, check if @page is uptodate.
4405 * For subpage case, check if the range covered by the eb has EXTENT_UPTODATE.
4406 */
4407static void assert_eb_folio_uptodate(const struct extent_buffer *eb, int i)
4408{
4409 struct btrfs_fs_info *fs_info = eb->fs_info;
4410 struct folio *folio = eb->folios[i];
4411
4412 ASSERT(folio);
4413
4414 /*
4415 * If we are using the commit root we could potentially clear a page
4416 * Uptodate while we're using the extent buffer that we've previously
4417 * looked up. We don't want to complain in this case, as the page was
4418 * valid before, we just didn't write it out. Instead we want to catch
4419 * the case where we didn't actually read the block properly, which
4420 * would have !PageUptodate and !EXTENT_BUFFER_WRITE_ERR.
4421 */
4422 if (test_bit(EXTENT_BUFFER_WRITE_ERR, &eb->bflags))
4423 return;
4424
4425 if (fs_info->nodesize < PAGE_SIZE) {
4426 struct folio *folio = eb->folios[0];
4427
4428 ASSERT(i == 0);
4429 if (WARN_ON(!btrfs_subpage_test_uptodate(fs_info, folio,
4430 eb->start, eb->len)))
4431 btrfs_subpage_dump_bitmap(fs_info, folio, eb->start, eb->len);
4432 } else {
4433 WARN_ON(!folio_test_uptodate(folio));
4434 }
4435}
4436
4437static void __write_extent_buffer(const struct extent_buffer *eb,
4438 const void *srcv, unsigned long start,
4439 unsigned long len, bool use_memmove)
4440{
4441 const int unit_size = folio_size(eb->folios[0]);
4442 size_t cur;
4443 size_t offset;
4444 char *kaddr;
4445 char *src = (char *)srcv;
4446 unsigned long i = get_eb_folio_index(eb, start);
4447 /* For unmapped (dummy) ebs, no need to check their uptodate status. */
4448 const bool check_uptodate = !test_bit(EXTENT_BUFFER_UNMAPPED, &eb->bflags);
4449
4450 if (check_eb_range(eb, start, len))
4451 return;
4452
4453 if (eb->addr) {
4454 if (use_memmove)
4455 memmove(eb->addr + start, srcv, len);
4456 else
4457 memcpy(eb->addr + start, srcv, len);
4458 return;
4459 }
4460
4461 offset = get_eb_offset_in_folio(eb, start);
4462
4463 while (len > 0) {
4464 if (check_uptodate)
4465 assert_eb_folio_uptodate(eb, i);
4466
4467 cur = min(len, unit_size - offset);
4468 kaddr = folio_address(eb->folios[i]);
4469 if (use_memmove)
4470 memmove(kaddr + offset, src, cur);
4471 else
4472 memcpy(kaddr + offset, src, cur);
4473
4474 src += cur;
4475 len -= cur;
4476 offset = 0;
4477 i++;
4478 }
4479}
4480
4481void write_extent_buffer(const struct extent_buffer *eb, const void *srcv,
4482 unsigned long start, unsigned long len)
4483{
4484 return __write_extent_buffer(eb, srcv, start, len, false);
4485}
4486
4487static void memset_extent_buffer(const struct extent_buffer *eb, int c,
4488 unsigned long start, unsigned long len)
4489{
4490 const int unit_size = folio_size(eb->folios[0]);
4491 unsigned long cur = start;
4492
4493 if (eb->addr) {
4494 memset(eb->addr + start, c, len);
4495 return;
4496 }
4497
4498 while (cur < start + len) {
4499 unsigned long index = get_eb_folio_index(eb, cur);
4500 unsigned int offset = get_eb_offset_in_folio(eb, cur);
4501 unsigned int cur_len = min(start + len - cur, unit_size - offset);
4502
4503 assert_eb_folio_uptodate(eb, index);
4504 memset(folio_address(eb->folios[index]) + offset, c, cur_len);
4505
4506 cur += cur_len;
4507 }
4508}
4509
4510void memzero_extent_buffer(const struct extent_buffer *eb, unsigned long start,
4511 unsigned long len)
4512{
4513 if (check_eb_range(eb, start, len))
4514 return;
4515 return memset_extent_buffer(eb, 0, start, len);
4516}
4517
4518void copy_extent_buffer_full(const struct extent_buffer *dst,
4519 const struct extent_buffer *src)
4520{
4521 const int unit_size = folio_size(src->folios[0]);
4522 unsigned long cur = 0;
4523
4524 ASSERT(dst->len == src->len);
4525
4526 while (cur < src->len) {
4527 unsigned long index = get_eb_folio_index(src, cur);
4528 unsigned long offset = get_eb_offset_in_folio(src, cur);
4529 unsigned long cur_len = min(src->len, unit_size - offset);
4530 void *addr = folio_address(src->folios[index]) + offset;
4531
4532 write_extent_buffer(dst, addr, cur, cur_len);
4533
4534 cur += cur_len;
4535 }
4536}
4537
4538void copy_extent_buffer(const struct extent_buffer *dst,
4539 const struct extent_buffer *src,
4540 unsigned long dst_offset, unsigned long src_offset,
4541 unsigned long len)
4542{
4543 const int unit_size = folio_size(dst->folios[0]);
4544 u64 dst_len = dst->len;
4545 size_t cur;
4546 size_t offset;
4547 char *kaddr;
4548 unsigned long i = get_eb_folio_index(dst, dst_offset);
4549
4550 if (check_eb_range(dst, dst_offset, len) ||
4551 check_eb_range(src, src_offset, len))
4552 return;
4553
4554 WARN_ON(src->len != dst_len);
4555
4556 offset = get_eb_offset_in_folio(dst, dst_offset);
4557
4558 while (len > 0) {
4559 assert_eb_folio_uptodate(dst, i);
4560
4561 cur = min(len, (unsigned long)(unit_size - offset));
4562
4563 kaddr = folio_address(dst->folios[i]);
4564 read_extent_buffer(src, kaddr + offset, src_offset, cur);
4565
4566 src_offset += cur;
4567 len -= cur;
4568 offset = 0;
4569 i++;
4570 }
4571}
4572
4573/*
4574 * Calculate the folio and offset of the byte containing the given bit number.
4575 *
4576 * @eb: the extent buffer
4577 * @start: offset of the bitmap item in the extent buffer
4578 * @nr: bit number
4579 * @folio_index: return index of the folio in the extent buffer that contains
4580 * the given bit number
4581 * @folio_offset: return offset into the folio given by folio_index
4582 *
4583 * This helper hides the ugliness of finding the byte in an extent buffer which
4584 * contains a given bit.
4585 */
4586static inline void eb_bitmap_offset(const struct extent_buffer *eb,
4587 unsigned long start, unsigned long nr,
4588 unsigned long *folio_index,
4589 size_t *folio_offset)
4590{
4591 size_t byte_offset = BIT_BYTE(nr);
4592 size_t offset;
4593
4594 /*
4595 * The byte we want is the offset of the extent buffer + the offset of
4596 * the bitmap item in the extent buffer + the offset of the byte in the
4597 * bitmap item.
4598 */
4599 offset = start + offset_in_folio(eb->folios[0], eb->start) + byte_offset;
4600
4601 *folio_index = offset >> folio_shift(eb->folios[0]);
4602 *folio_offset = offset_in_folio(eb->folios[0], offset);
4603}
4604
4605/*
4606 * Determine whether a bit in a bitmap item is set.
4607 *
4608 * @eb: the extent buffer
4609 * @start: offset of the bitmap item in the extent buffer
4610 * @nr: bit number to test
4611 */
4612int extent_buffer_test_bit(const struct extent_buffer *eb, unsigned long start,
4613 unsigned long nr)
4614{
4615 unsigned long i;
4616 size_t offset;
4617 u8 *kaddr;
4618
4619 eb_bitmap_offset(eb, start, nr, &i, &offset);
4620 assert_eb_folio_uptodate(eb, i);
4621 kaddr = folio_address(eb->folios[i]);
4622 return 1U & (kaddr[offset] >> (nr & (BITS_PER_BYTE - 1)));
4623}
4624
4625static u8 *extent_buffer_get_byte(const struct extent_buffer *eb, unsigned long bytenr)
4626{
4627 unsigned long index = get_eb_folio_index(eb, bytenr);
4628
4629 if (check_eb_range(eb, bytenr, 1))
4630 return NULL;
4631 return folio_address(eb->folios[index]) + get_eb_offset_in_folio(eb, bytenr);
4632}
4633
4634/*
4635 * Set an area of a bitmap to 1.
4636 *
4637 * @eb: the extent buffer
4638 * @start: offset of the bitmap item in the extent buffer
4639 * @pos: bit number of the first bit
4640 * @len: number of bits to set
4641 */
4642void extent_buffer_bitmap_set(const struct extent_buffer *eb, unsigned long start,
4643 unsigned long pos, unsigned long len)
4644{
4645 unsigned int first_byte = start + BIT_BYTE(pos);
4646 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4647 const bool same_byte = (first_byte == last_byte);
4648 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4649 u8 *kaddr;
4650
4651 if (same_byte)
4652 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4653
4654 /* Handle the first byte. */
4655 kaddr = extent_buffer_get_byte(eb, first_byte);
4656 *kaddr |= mask;
4657 if (same_byte)
4658 return;
4659
4660 /* Handle the byte aligned part. */
4661 ASSERT(first_byte + 1 <= last_byte);
4662 memset_extent_buffer(eb, 0xff, first_byte + 1, last_byte - first_byte - 1);
4663
4664 /* Handle the last byte. */
4665 kaddr = extent_buffer_get_byte(eb, last_byte);
4666 *kaddr |= BITMAP_LAST_BYTE_MASK(pos + len);
4667}
4668
4669
4670/*
4671 * Clear an area of a bitmap.
4672 *
4673 * @eb: the extent buffer
4674 * @start: offset of the bitmap item in the extent buffer
4675 * @pos: bit number of the first bit
4676 * @len: number of bits to clear
4677 */
4678void extent_buffer_bitmap_clear(const struct extent_buffer *eb,
4679 unsigned long start, unsigned long pos,
4680 unsigned long len)
4681{
4682 unsigned int first_byte = start + BIT_BYTE(pos);
4683 unsigned int last_byte = start + BIT_BYTE(pos + len - 1);
4684 const bool same_byte = (first_byte == last_byte);
4685 u8 mask = BITMAP_FIRST_BYTE_MASK(pos);
4686 u8 *kaddr;
4687
4688 if (same_byte)
4689 mask &= BITMAP_LAST_BYTE_MASK(pos + len);
4690
4691 /* Handle the first byte. */
4692 kaddr = extent_buffer_get_byte(eb, first_byte);
4693 *kaddr &= ~mask;
4694 if (same_byte)
4695 return;
4696
4697 /* Handle the byte aligned part. */
4698 ASSERT(first_byte + 1 <= last_byte);
4699 memset_extent_buffer(eb, 0, first_byte + 1, last_byte - first_byte - 1);
4700
4701 /* Handle the last byte. */
4702 kaddr = extent_buffer_get_byte(eb, last_byte);
4703 *kaddr &= ~BITMAP_LAST_BYTE_MASK(pos + len);
4704}
4705
4706static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
4707{
4708 unsigned long distance = (src > dst) ? src - dst : dst - src;
4709 return distance < len;
4710}
4711
4712void memcpy_extent_buffer(const struct extent_buffer *dst,
4713 unsigned long dst_offset, unsigned long src_offset,
4714 unsigned long len)
4715{
4716 const int unit_size = folio_size(dst->folios[0]);
4717 unsigned long cur_off = 0;
4718
4719 if (check_eb_range(dst, dst_offset, len) ||
4720 check_eb_range(dst, src_offset, len))
4721 return;
4722
4723 if (dst->addr) {
4724 const bool use_memmove = areas_overlap(src_offset, dst_offset, len);
4725
4726 if (use_memmove)
4727 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4728 else
4729 memcpy(dst->addr + dst_offset, dst->addr + src_offset, len);
4730 return;
4731 }
4732
4733 while (cur_off < len) {
4734 unsigned long cur_src = cur_off + src_offset;
4735 unsigned long folio_index = get_eb_folio_index(dst, cur_src);
4736 unsigned long folio_off = get_eb_offset_in_folio(dst, cur_src);
4737 unsigned long cur_len = min(src_offset + len - cur_src,
4738 unit_size - folio_off);
4739 void *src_addr = folio_address(dst->folios[folio_index]) + folio_off;
4740 const bool use_memmove = areas_overlap(src_offset + cur_off,
4741 dst_offset + cur_off, cur_len);
4742
4743 __write_extent_buffer(dst, src_addr, dst_offset + cur_off, cur_len,
4744 use_memmove);
4745 cur_off += cur_len;
4746 }
4747}
4748
4749void memmove_extent_buffer(const struct extent_buffer *dst,
4750 unsigned long dst_offset, unsigned long src_offset,
4751 unsigned long len)
4752{
4753 unsigned long dst_end = dst_offset + len - 1;
4754 unsigned long src_end = src_offset + len - 1;
4755
4756 if (check_eb_range(dst, dst_offset, len) ||
4757 check_eb_range(dst, src_offset, len))
4758 return;
4759
4760 if (dst_offset < src_offset) {
4761 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
4762 return;
4763 }
4764
4765 if (dst->addr) {
4766 memmove(dst->addr + dst_offset, dst->addr + src_offset, len);
4767 return;
4768 }
4769
4770 while (len > 0) {
4771 unsigned long src_i;
4772 size_t cur;
4773 size_t dst_off_in_folio;
4774 size_t src_off_in_folio;
4775 void *src_addr;
4776 bool use_memmove;
4777
4778 src_i = get_eb_folio_index(dst, src_end);
4779
4780 dst_off_in_folio = get_eb_offset_in_folio(dst, dst_end);
4781 src_off_in_folio = get_eb_offset_in_folio(dst, src_end);
4782
4783 cur = min_t(unsigned long, len, src_off_in_folio + 1);
4784 cur = min(cur, dst_off_in_folio + 1);
4785
4786 src_addr = folio_address(dst->folios[src_i]) + src_off_in_folio -
4787 cur + 1;
4788 use_memmove = areas_overlap(src_end - cur + 1, dst_end - cur + 1,
4789 cur);
4790
4791 __write_extent_buffer(dst, src_addr, dst_end - cur + 1, cur,
4792 use_memmove);
4793
4794 dst_end -= cur;
4795 src_end -= cur;
4796 len -= cur;
4797 }
4798}
4799
4800#define GANG_LOOKUP_SIZE 16
4801static struct extent_buffer *get_next_extent_buffer(
4802 struct btrfs_fs_info *fs_info, struct page *page, u64 bytenr)
4803{
4804 struct extent_buffer *gang[GANG_LOOKUP_SIZE];
4805 struct extent_buffer *found = NULL;
4806 u64 page_start = page_offset(page);
4807 u64 cur = page_start;
4808
4809 ASSERT(in_range(bytenr, page_start, PAGE_SIZE));
4810 lockdep_assert_held(&fs_info->buffer_lock);
4811
4812 while (cur < page_start + PAGE_SIZE) {
4813 int ret;
4814 int i;
4815
4816 ret = radix_tree_gang_lookup(&fs_info->buffer_radix,
4817 (void **)gang, cur >> fs_info->sectorsize_bits,
4818 min_t(unsigned int, GANG_LOOKUP_SIZE,
4819 PAGE_SIZE / fs_info->nodesize));
4820 if (ret == 0)
4821 goto out;
4822 for (i = 0; i < ret; i++) {
4823 /* Already beyond page end */
4824 if (gang[i]->start >= page_start + PAGE_SIZE)
4825 goto out;
4826 /* Found one */
4827 if (gang[i]->start >= bytenr) {
4828 found = gang[i];
4829 goto out;
4830 }
4831 }
4832 cur = gang[ret - 1]->start + gang[ret - 1]->len;
4833 }
4834out:
4835 return found;
4836}
4837
4838static int try_release_subpage_extent_buffer(struct page *page)
4839{
4840 struct btrfs_fs_info *fs_info = btrfs_sb(page->mapping->host->i_sb);
4841 u64 cur = page_offset(page);
4842 const u64 end = page_offset(page) + PAGE_SIZE;
4843 int ret;
4844
4845 while (cur < end) {
4846 struct extent_buffer *eb = NULL;
4847
4848 /*
4849 * Unlike try_release_extent_buffer() which uses folio private
4850 * to grab buffer, for subpage case we rely on radix tree, thus
4851 * we need to ensure radix tree consistency.
4852 *
4853 * We also want an atomic snapshot of the radix tree, thus go
4854 * with spinlock rather than RCU.
4855 */
4856 spin_lock(&fs_info->buffer_lock);
4857 eb = get_next_extent_buffer(fs_info, page, cur);
4858 if (!eb) {
4859 /* No more eb in the page range after or at cur */
4860 spin_unlock(&fs_info->buffer_lock);
4861 break;
4862 }
4863 cur = eb->start + eb->len;
4864
4865 /*
4866 * The same as try_release_extent_buffer(), to ensure the eb
4867 * won't disappear out from under us.
4868 */
4869 spin_lock(&eb->refs_lock);
4870 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4871 spin_unlock(&eb->refs_lock);
4872 spin_unlock(&fs_info->buffer_lock);
4873 break;
4874 }
4875 spin_unlock(&fs_info->buffer_lock);
4876
4877 /*
4878 * If tree ref isn't set then we know the ref on this eb is a
4879 * real ref, so just return, this eb will likely be freed soon
4880 * anyway.
4881 */
4882 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4883 spin_unlock(&eb->refs_lock);
4884 break;
4885 }
4886
4887 /*
4888 * Here we don't care about the return value, we will always
4889 * check the folio private at the end. And
4890 * release_extent_buffer() will release the refs_lock.
4891 */
4892 release_extent_buffer(eb);
4893 }
4894 /*
4895 * Finally to check if we have cleared folio private, as if we have
4896 * released all ebs in the page, the folio private should be cleared now.
4897 */
4898 spin_lock(&page->mapping->i_private_lock);
4899 if (!folio_test_private(page_folio(page)))
4900 ret = 1;
4901 else
4902 ret = 0;
4903 spin_unlock(&page->mapping->i_private_lock);
4904 return ret;
4905
4906}
4907
4908int try_release_extent_buffer(struct page *page)
4909{
4910 struct folio *folio = page_folio(page);
4911 struct extent_buffer *eb;
4912
4913 if (btrfs_sb(page->mapping->host->i_sb)->nodesize < PAGE_SIZE)
4914 return try_release_subpage_extent_buffer(page);
4915
4916 /*
4917 * We need to make sure nobody is changing folio private, as we rely on
4918 * folio private as the pointer to extent buffer.
4919 */
4920 spin_lock(&page->mapping->i_private_lock);
4921 if (!folio_test_private(folio)) {
4922 spin_unlock(&page->mapping->i_private_lock);
4923 return 1;
4924 }
4925
4926 eb = folio_get_private(folio);
4927 BUG_ON(!eb);
4928
4929 /*
4930 * This is a little awful but should be ok, we need to make sure that
4931 * the eb doesn't disappear out from under us while we're looking at
4932 * this page.
4933 */
4934 spin_lock(&eb->refs_lock);
4935 if (atomic_read(&eb->refs) != 1 || extent_buffer_under_io(eb)) {
4936 spin_unlock(&eb->refs_lock);
4937 spin_unlock(&page->mapping->i_private_lock);
4938 return 0;
4939 }
4940 spin_unlock(&page->mapping->i_private_lock);
4941
4942 /*
4943 * If tree ref isn't set then we know the ref on this eb is a real ref,
4944 * so just return, this page will likely be freed soon anyway.
4945 */
4946 if (!test_and_clear_bit(EXTENT_BUFFER_TREE_REF, &eb->bflags)) {
4947 spin_unlock(&eb->refs_lock);
4948 return 0;
4949 }
4950
4951 return release_extent_buffer(eb);
4952}
4953
4954/*
4955 * Attempt to readahead a child block.
4956 *
4957 * @fs_info: the fs_info
4958 * @bytenr: bytenr to read
4959 * @owner_root: objectid of the root that owns this eb
4960 * @gen: generation for the uptodate check, can be 0
4961 * @level: level for the eb
4962 *
4963 * Attempt to readahead a tree block at @bytenr. If @gen is 0 then we do a
4964 * normal uptodate check of the eb, without checking the generation. If we have
4965 * to read the block we will not block on anything.
4966 */
4967void btrfs_readahead_tree_block(struct btrfs_fs_info *fs_info,
4968 u64 bytenr, u64 owner_root, u64 gen, int level)
4969{
4970 struct btrfs_tree_parent_check check = {
4971 .has_first_key = 0,
4972 .level = level,
4973 .transid = gen
4974 };
4975 struct extent_buffer *eb;
4976 int ret;
4977
4978 eb = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
4979 if (IS_ERR(eb))
4980 return;
4981
4982 if (btrfs_buffer_uptodate(eb, gen, 1)) {
4983 free_extent_buffer(eb);
4984 return;
4985 }
4986
4987 ret = read_extent_buffer_pages(eb, WAIT_NONE, 0, &check);
4988 if (ret < 0)
4989 free_extent_buffer_stale(eb);
4990 else
4991 free_extent_buffer(eb);
4992}
4993
4994/*
4995 * Readahead a node's child block.
4996 *
4997 * @node: parent node we're reading from
4998 * @slot: slot in the parent node for the child we want to read
4999 *
5000 * A helper for btrfs_readahead_tree_block, we simply read the bytenr pointed at
5001 * the slot in the node provided.
5002 */
5003void btrfs_readahead_node_child(struct extent_buffer *node, int slot)
5004{
5005 btrfs_readahead_tree_block(node->fs_info,
5006 btrfs_node_blockptr(node, slot),
5007 btrfs_header_owner(node),
5008 btrfs_node_ptr_generation(node, slot),
5009 btrfs_header_level(node) - 1);
5010}
1#include <linux/bitops.h>
2#include <linux/slab.h>
3#include <linux/bio.h>
4#include <linux/mm.h>
5#include <linux/pagemap.h>
6#include <linux/page-flags.h>
7#include <linux/module.h>
8#include <linux/spinlock.h>
9#include <linux/blkdev.h>
10#include <linux/swap.h>
11#include <linux/writeback.h>
12#include <linux/pagevec.h>
13#include <linux/prefetch.h>
14#include <linux/cleancache.h>
15#include "extent_io.h"
16#include "extent_map.h"
17#include "compat.h"
18#include "ctree.h"
19#include "btrfs_inode.h"
20
21static struct kmem_cache *extent_state_cache;
22static struct kmem_cache *extent_buffer_cache;
23
24static LIST_HEAD(buffers);
25static LIST_HEAD(states);
26
27#define LEAK_DEBUG 0
28#if LEAK_DEBUG
29static DEFINE_SPINLOCK(leak_lock);
30#endif
31
32#define BUFFER_LRU_MAX 64
33
34struct tree_entry {
35 u64 start;
36 u64 end;
37 struct rb_node rb_node;
38};
39
40struct extent_page_data {
41 struct bio *bio;
42 struct extent_io_tree *tree;
43 get_extent_t *get_extent;
44
45 /* tells writepage not to lock the state bits for this range
46 * it still does the unlocking
47 */
48 unsigned int extent_locked:1;
49
50 /* tells the submit_bio code to use a WRITE_SYNC */
51 unsigned int sync_io:1;
52};
53
54int __init extent_io_init(void)
55{
56 extent_state_cache = kmem_cache_create("extent_state",
57 sizeof(struct extent_state), 0,
58 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
59 if (!extent_state_cache)
60 return -ENOMEM;
61
62 extent_buffer_cache = kmem_cache_create("extent_buffers",
63 sizeof(struct extent_buffer), 0,
64 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
65 if (!extent_buffer_cache)
66 goto free_state_cache;
67 return 0;
68
69free_state_cache:
70 kmem_cache_destroy(extent_state_cache);
71 return -ENOMEM;
72}
73
74void extent_io_exit(void)
75{
76 struct extent_state *state;
77 struct extent_buffer *eb;
78
79 while (!list_empty(&states)) {
80 state = list_entry(states.next, struct extent_state, leak_list);
81 printk(KERN_ERR "btrfs state leak: start %llu end %llu "
82 "state %lu in tree %p refs %d\n",
83 (unsigned long long)state->start,
84 (unsigned long long)state->end,
85 state->state, state->tree, atomic_read(&state->refs));
86 list_del(&state->leak_list);
87 kmem_cache_free(extent_state_cache, state);
88
89 }
90
91 while (!list_empty(&buffers)) {
92 eb = list_entry(buffers.next, struct extent_buffer, leak_list);
93 printk(KERN_ERR "btrfs buffer leak start %llu len %lu "
94 "refs %d\n", (unsigned long long)eb->start,
95 eb->len, atomic_read(&eb->refs));
96 list_del(&eb->leak_list);
97 kmem_cache_free(extent_buffer_cache, eb);
98 }
99 if (extent_state_cache)
100 kmem_cache_destroy(extent_state_cache);
101 if (extent_buffer_cache)
102 kmem_cache_destroy(extent_buffer_cache);
103}
104
105void extent_io_tree_init(struct extent_io_tree *tree,
106 struct address_space *mapping)
107{
108 tree->state = RB_ROOT;
109 INIT_RADIX_TREE(&tree->buffer, GFP_ATOMIC);
110 tree->ops = NULL;
111 tree->dirty_bytes = 0;
112 spin_lock_init(&tree->lock);
113 spin_lock_init(&tree->buffer_lock);
114 tree->mapping = mapping;
115}
116
117static struct extent_state *alloc_extent_state(gfp_t mask)
118{
119 struct extent_state *state;
120#if LEAK_DEBUG
121 unsigned long flags;
122#endif
123
124 state = kmem_cache_alloc(extent_state_cache, mask);
125 if (!state)
126 return state;
127 state->state = 0;
128 state->private = 0;
129 state->tree = NULL;
130#if LEAK_DEBUG
131 spin_lock_irqsave(&leak_lock, flags);
132 list_add(&state->leak_list, &states);
133 spin_unlock_irqrestore(&leak_lock, flags);
134#endif
135 atomic_set(&state->refs, 1);
136 init_waitqueue_head(&state->wq);
137 return state;
138}
139
140void free_extent_state(struct extent_state *state)
141{
142 if (!state)
143 return;
144 if (atomic_dec_and_test(&state->refs)) {
145#if LEAK_DEBUG
146 unsigned long flags;
147#endif
148 WARN_ON(state->tree);
149#if LEAK_DEBUG
150 spin_lock_irqsave(&leak_lock, flags);
151 list_del(&state->leak_list);
152 spin_unlock_irqrestore(&leak_lock, flags);
153#endif
154 kmem_cache_free(extent_state_cache, state);
155 }
156}
157
158static struct rb_node *tree_insert(struct rb_root *root, u64 offset,
159 struct rb_node *node)
160{
161 struct rb_node **p = &root->rb_node;
162 struct rb_node *parent = NULL;
163 struct tree_entry *entry;
164
165 while (*p) {
166 parent = *p;
167 entry = rb_entry(parent, struct tree_entry, rb_node);
168
169 if (offset < entry->start)
170 p = &(*p)->rb_left;
171 else if (offset > entry->end)
172 p = &(*p)->rb_right;
173 else
174 return parent;
175 }
176
177 entry = rb_entry(node, struct tree_entry, rb_node);
178 rb_link_node(node, parent, p);
179 rb_insert_color(node, root);
180 return NULL;
181}
182
183static struct rb_node *__etree_search(struct extent_io_tree *tree, u64 offset,
184 struct rb_node **prev_ret,
185 struct rb_node **next_ret)
186{
187 struct rb_root *root = &tree->state;
188 struct rb_node *n = root->rb_node;
189 struct rb_node *prev = NULL;
190 struct rb_node *orig_prev = NULL;
191 struct tree_entry *entry;
192 struct tree_entry *prev_entry = NULL;
193
194 while (n) {
195 entry = rb_entry(n, struct tree_entry, rb_node);
196 prev = n;
197 prev_entry = entry;
198
199 if (offset < entry->start)
200 n = n->rb_left;
201 else if (offset > entry->end)
202 n = n->rb_right;
203 else
204 return n;
205 }
206
207 if (prev_ret) {
208 orig_prev = prev;
209 while (prev && offset > prev_entry->end) {
210 prev = rb_next(prev);
211 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
212 }
213 *prev_ret = prev;
214 prev = orig_prev;
215 }
216
217 if (next_ret) {
218 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
219 while (prev && offset < prev_entry->start) {
220 prev = rb_prev(prev);
221 prev_entry = rb_entry(prev, struct tree_entry, rb_node);
222 }
223 *next_ret = prev;
224 }
225 return NULL;
226}
227
228static inline struct rb_node *tree_search(struct extent_io_tree *tree,
229 u64 offset)
230{
231 struct rb_node *prev = NULL;
232 struct rb_node *ret;
233
234 ret = __etree_search(tree, offset, &prev, NULL);
235 if (!ret)
236 return prev;
237 return ret;
238}
239
240static void merge_cb(struct extent_io_tree *tree, struct extent_state *new,
241 struct extent_state *other)
242{
243 if (tree->ops && tree->ops->merge_extent_hook)
244 tree->ops->merge_extent_hook(tree->mapping->host, new,
245 other);
246}
247
248/*
249 * utility function to look for merge candidates inside a given range.
250 * Any extents with matching state are merged together into a single
251 * extent in the tree. Extents with EXTENT_IO in their state field
252 * are not merged because the end_io handlers need to be able to do
253 * operations on them without sleeping (or doing allocations/splits).
254 *
255 * This should be called with the tree lock held.
256 */
257static void merge_state(struct extent_io_tree *tree,
258 struct extent_state *state)
259{
260 struct extent_state *other;
261 struct rb_node *other_node;
262
263 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY))
264 return;
265
266 other_node = rb_prev(&state->rb_node);
267 if (other_node) {
268 other = rb_entry(other_node, struct extent_state, rb_node);
269 if (other->end == state->start - 1 &&
270 other->state == state->state) {
271 merge_cb(tree, state, other);
272 state->start = other->start;
273 other->tree = NULL;
274 rb_erase(&other->rb_node, &tree->state);
275 free_extent_state(other);
276 }
277 }
278 other_node = rb_next(&state->rb_node);
279 if (other_node) {
280 other = rb_entry(other_node, struct extent_state, rb_node);
281 if (other->start == state->end + 1 &&
282 other->state == state->state) {
283 merge_cb(tree, state, other);
284 state->end = other->end;
285 other->tree = NULL;
286 rb_erase(&other->rb_node, &tree->state);
287 free_extent_state(other);
288 }
289 }
290}
291
292static void set_state_cb(struct extent_io_tree *tree,
293 struct extent_state *state, int *bits)
294{
295 if (tree->ops && tree->ops->set_bit_hook)
296 tree->ops->set_bit_hook(tree->mapping->host, state, bits);
297}
298
299static void clear_state_cb(struct extent_io_tree *tree,
300 struct extent_state *state, int *bits)
301{
302 if (tree->ops && tree->ops->clear_bit_hook)
303 tree->ops->clear_bit_hook(tree->mapping->host, state, bits);
304}
305
306static void set_state_bits(struct extent_io_tree *tree,
307 struct extent_state *state, int *bits);
308
309/*
310 * insert an extent_state struct into the tree. 'bits' are set on the
311 * struct before it is inserted.
312 *
313 * This may return -EEXIST if the extent is already there, in which case the
314 * state struct is freed.
315 *
316 * The tree lock is not taken internally. This is a utility function and
317 * probably isn't what you want to call (see set/clear_extent_bit).
318 */
319static int insert_state(struct extent_io_tree *tree,
320 struct extent_state *state, u64 start, u64 end,
321 int *bits)
322{
323 struct rb_node *node;
324
325 if (end < start) {
326 printk(KERN_ERR "btrfs end < start %llu %llu\n",
327 (unsigned long long)end,
328 (unsigned long long)start);
329 WARN_ON(1);
330 }
331 state->start = start;
332 state->end = end;
333
334 set_state_bits(tree, state, bits);
335
336 node = tree_insert(&tree->state, end, &state->rb_node);
337 if (node) {
338 struct extent_state *found;
339 found = rb_entry(node, struct extent_state, rb_node);
340 printk(KERN_ERR "btrfs found node %llu %llu on insert of "
341 "%llu %llu\n", (unsigned long long)found->start,
342 (unsigned long long)found->end,
343 (unsigned long long)start, (unsigned long long)end);
344 return -EEXIST;
345 }
346 state->tree = tree;
347 merge_state(tree, state);
348 return 0;
349}
350
351static void split_cb(struct extent_io_tree *tree, struct extent_state *orig,
352 u64 split)
353{
354 if (tree->ops && tree->ops->split_extent_hook)
355 tree->ops->split_extent_hook(tree->mapping->host, orig, split);
356}
357
358/*
359 * split a given extent state struct in two, inserting the preallocated
360 * struct 'prealloc' as the newly created second half. 'split' indicates an
361 * offset inside 'orig' where it should be split.
362 *
363 * Before calling,
364 * the tree has 'orig' at [orig->start, orig->end]. After calling, there
365 * are two extent state structs in the tree:
366 * prealloc: [orig->start, split - 1]
367 * orig: [ split, orig->end ]
368 *
369 * The tree locks are not taken by this function. They need to be held
370 * by the caller.
371 */
372static int split_state(struct extent_io_tree *tree, struct extent_state *orig,
373 struct extent_state *prealloc, u64 split)
374{
375 struct rb_node *node;
376
377 split_cb(tree, orig, split);
378
379 prealloc->start = orig->start;
380 prealloc->end = split - 1;
381 prealloc->state = orig->state;
382 orig->start = split;
383
384 node = tree_insert(&tree->state, prealloc->end, &prealloc->rb_node);
385 if (node) {
386 free_extent_state(prealloc);
387 return -EEXIST;
388 }
389 prealloc->tree = tree;
390 return 0;
391}
392
393/*
394 * utility function to clear some bits in an extent state struct.
395 * it will optionally wake up any one waiting on this state (wake == 1), or
396 * forcibly remove the state from the tree (delete == 1).
397 *
398 * If no bits are set on the state struct after clearing things, the
399 * struct is freed and removed from the tree
400 */
401static int clear_state_bit(struct extent_io_tree *tree,
402 struct extent_state *state,
403 int *bits, int wake)
404{
405 int bits_to_clear = *bits & ~EXTENT_CTLBITS;
406 int ret = state->state & bits_to_clear;
407
408 if ((bits_to_clear & EXTENT_DIRTY) && (state->state & EXTENT_DIRTY)) {
409 u64 range = state->end - state->start + 1;
410 WARN_ON(range > tree->dirty_bytes);
411 tree->dirty_bytes -= range;
412 }
413 clear_state_cb(tree, state, bits);
414 state->state &= ~bits_to_clear;
415 if (wake)
416 wake_up(&state->wq);
417 if (state->state == 0) {
418 if (state->tree) {
419 rb_erase(&state->rb_node, &tree->state);
420 state->tree = NULL;
421 free_extent_state(state);
422 } else {
423 WARN_ON(1);
424 }
425 } else {
426 merge_state(tree, state);
427 }
428 return ret;
429}
430
431static struct extent_state *
432alloc_extent_state_atomic(struct extent_state *prealloc)
433{
434 if (!prealloc)
435 prealloc = alloc_extent_state(GFP_ATOMIC);
436
437 return prealloc;
438}
439
440/*
441 * clear some bits on a range in the tree. This may require splitting
442 * or inserting elements in the tree, so the gfp mask is used to
443 * indicate which allocations or sleeping are allowed.
444 *
445 * pass 'wake' == 1 to kick any sleepers, and 'delete' == 1 to remove
446 * the given range from the tree regardless of state (ie for truncate).
447 *
448 * the range [start, end] is inclusive.
449 *
450 * This takes the tree lock, and returns < 0 on error, > 0 if any of the
451 * bits were already set, or zero if none of the bits were already set.
452 */
453int clear_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
454 int bits, int wake, int delete,
455 struct extent_state **cached_state,
456 gfp_t mask)
457{
458 struct extent_state *state;
459 struct extent_state *cached;
460 struct extent_state *prealloc = NULL;
461 struct rb_node *next_node;
462 struct rb_node *node;
463 u64 last_end;
464 int err;
465 int set = 0;
466 int clear = 0;
467
468 if (delete)
469 bits |= ~EXTENT_CTLBITS;
470 bits |= EXTENT_FIRST_DELALLOC;
471
472 if (bits & (EXTENT_IOBITS | EXTENT_BOUNDARY))
473 clear = 1;
474again:
475 if (!prealloc && (mask & __GFP_WAIT)) {
476 prealloc = alloc_extent_state(mask);
477 if (!prealloc)
478 return -ENOMEM;
479 }
480
481 spin_lock(&tree->lock);
482 if (cached_state) {
483 cached = *cached_state;
484
485 if (clear) {
486 *cached_state = NULL;
487 cached_state = NULL;
488 }
489
490 if (cached && cached->tree && cached->start <= start &&
491 cached->end > start) {
492 if (clear)
493 atomic_dec(&cached->refs);
494 state = cached;
495 goto hit_next;
496 }
497 if (clear)
498 free_extent_state(cached);
499 }
500 /*
501 * this search will find the extents that end after
502 * our range starts
503 */
504 node = tree_search(tree, start);
505 if (!node)
506 goto out;
507 state = rb_entry(node, struct extent_state, rb_node);
508hit_next:
509 if (state->start > end)
510 goto out;
511 WARN_ON(state->end < start);
512 last_end = state->end;
513
514 /*
515 * | ---- desired range ---- |
516 * | state | or
517 * | ------------- state -------------- |
518 *
519 * We need to split the extent we found, and may flip
520 * bits on second half.
521 *
522 * If the extent we found extends past our range, we
523 * just split and search again. It'll get split again
524 * the next time though.
525 *
526 * If the extent we found is inside our range, we clear
527 * the desired bit on it.
528 */
529
530 if (state->start < start) {
531 prealloc = alloc_extent_state_atomic(prealloc);
532 BUG_ON(!prealloc);
533 err = split_state(tree, state, prealloc, start);
534 BUG_ON(err == -EEXIST);
535 prealloc = NULL;
536 if (err)
537 goto out;
538 if (state->end <= end) {
539 set |= clear_state_bit(tree, state, &bits, wake);
540 if (last_end == (u64)-1)
541 goto out;
542 start = last_end + 1;
543 }
544 goto search_again;
545 }
546 /*
547 * | ---- desired range ---- |
548 * | state |
549 * We need to split the extent, and clear the bit
550 * on the first half
551 */
552 if (state->start <= end && state->end > end) {
553 prealloc = alloc_extent_state_atomic(prealloc);
554 BUG_ON(!prealloc);
555 err = split_state(tree, state, prealloc, end + 1);
556 BUG_ON(err == -EEXIST);
557 if (wake)
558 wake_up(&state->wq);
559
560 set |= clear_state_bit(tree, prealloc, &bits, wake);
561
562 prealloc = NULL;
563 goto out;
564 }
565
566 if (state->end < end && prealloc && !need_resched())
567 next_node = rb_next(&state->rb_node);
568 else
569 next_node = NULL;
570
571 set |= clear_state_bit(tree, state, &bits, wake);
572 if (last_end == (u64)-1)
573 goto out;
574 start = last_end + 1;
575 if (start <= end && next_node) {
576 state = rb_entry(next_node, struct extent_state,
577 rb_node);
578 if (state->start == start)
579 goto hit_next;
580 }
581 goto search_again;
582
583out:
584 spin_unlock(&tree->lock);
585 if (prealloc)
586 free_extent_state(prealloc);
587
588 return set;
589
590search_again:
591 if (start > end)
592 goto out;
593 spin_unlock(&tree->lock);
594 if (mask & __GFP_WAIT)
595 cond_resched();
596 goto again;
597}
598
599static int wait_on_state(struct extent_io_tree *tree,
600 struct extent_state *state)
601 __releases(tree->lock)
602 __acquires(tree->lock)
603{
604 DEFINE_WAIT(wait);
605 prepare_to_wait(&state->wq, &wait, TASK_UNINTERRUPTIBLE);
606 spin_unlock(&tree->lock);
607 schedule();
608 spin_lock(&tree->lock);
609 finish_wait(&state->wq, &wait);
610 return 0;
611}
612
613/*
614 * waits for one or more bits to clear on a range in the state tree.
615 * The range [start, end] is inclusive.
616 * The tree lock is taken by this function
617 */
618int wait_extent_bit(struct extent_io_tree *tree, u64 start, u64 end, int bits)
619{
620 struct extent_state *state;
621 struct rb_node *node;
622
623 spin_lock(&tree->lock);
624again:
625 while (1) {
626 /*
627 * this search will find all the extents that end after
628 * our range starts
629 */
630 node = tree_search(tree, start);
631 if (!node)
632 break;
633
634 state = rb_entry(node, struct extent_state, rb_node);
635
636 if (state->start > end)
637 goto out;
638
639 if (state->state & bits) {
640 start = state->start;
641 atomic_inc(&state->refs);
642 wait_on_state(tree, state);
643 free_extent_state(state);
644 goto again;
645 }
646 start = state->end + 1;
647
648 if (start > end)
649 break;
650
651 cond_resched_lock(&tree->lock);
652 }
653out:
654 spin_unlock(&tree->lock);
655 return 0;
656}
657
658static void set_state_bits(struct extent_io_tree *tree,
659 struct extent_state *state,
660 int *bits)
661{
662 int bits_to_set = *bits & ~EXTENT_CTLBITS;
663
664 set_state_cb(tree, state, bits);
665 if ((bits_to_set & EXTENT_DIRTY) && !(state->state & EXTENT_DIRTY)) {
666 u64 range = state->end - state->start + 1;
667 tree->dirty_bytes += range;
668 }
669 state->state |= bits_to_set;
670}
671
672static void cache_state(struct extent_state *state,
673 struct extent_state **cached_ptr)
674{
675 if (cached_ptr && !(*cached_ptr)) {
676 if (state->state & (EXTENT_IOBITS | EXTENT_BOUNDARY)) {
677 *cached_ptr = state;
678 atomic_inc(&state->refs);
679 }
680 }
681}
682
683static void uncache_state(struct extent_state **cached_ptr)
684{
685 if (cached_ptr && (*cached_ptr)) {
686 struct extent_state *state = *cached_ptr;
687 *cached_ptr = NULL;
688 free_extent_state(state);
689 }
690}
691
692/*
693 * set some bits on a range in the tree. This may require allocations or
694 * sleeping, so the gfp mask is used to indicate what is allowed.
695 *
696 * If any of the exclusive bits are set, this will fail with -EEXIST if some
697 * part of the range already has the desired bits set. The start of the
698 * existing range is returned in failed_start in this case.
699 *
700 * [start, end] is inclusive This takes the tree lock.
701 */
702
703int set_extent_bit(struct extent_io_tree *tree, u64 start, u64 end,
704 int bits, int exclusive_bits, u64 *failed_start,
705 struct extent_state **cached_state, gfp_t mask)
706{
707 struct extent_state *state;
708 struct extent_state *prealloc = NULL;
709 struct rb_node *node;
710 int err = 0;
711 u64 last_start;
712 u64 last_end;
713
714 bits |= EXTENT_FIRST_DELALLOC;
715again:
716 if (!prealloc && (mask & __GFP_WAIT)) {
717 prealloc = alloc_extent_state(mask);
718 BUG_ON(!prealloc);
719 }
720
721 spin_lock(&tree->lock);
722 if (cached_state && *cached_state) {
723 state = *cached_state;
724 if (state->start <= start && state->end > start &&
725 state->tree) {
726 node = &state->rb_node;
727 goto hit_next;
728 }
729 }
730 /*
731 * this search will find all the extents that end after
732 * our range starts.
733 */
734 node = tree_search(tree, start);
735 if (!node) {
736 prealloc = alloc_extent_state_atomic(prealloc);
737 BUG_ON(!prealloc);
738 err = insert_state(tree, prealloc, start, end, &bits);
739 prealloc = NULL;
740 BUG_ON(err == -EEXIST);
741 goto out;
742 }
743 state = rb_entry(node, struct extent_state, rb_node);
744hit_next:
745 last_start = state->start;
746 last_end = state->end;
747
748 /*
749 * | ---- desired range ---- |
750 * | state |
751 *
752 * Just lock what we found and keep going
753 */
754 if (state->start == start && state->end <= end) {
755 struct rb_node *next_node;
756 if (state->state & exclusive_bits) {
757 *failed_start = state->start;
758 err = -EEXIST;
759 goto out;
760 }
761
762 set_state_bits(tree, state, &bits);
763
764 cache_state(state, cached_state);
765 merge_state(tree, state);
766 if (last_end == (u64)-1)
767 goto out;
768
769 start = last_end + 1;
770 next_node = rb_next(&state->rb_node);
771 if (next_node && start < end && prealloc && !need_resched()) {
772 state = rb_entry(next_node, struct extent_state,
773 rb_node);
774 if (state->start == start)
775 goto hit_next;
776 }
777 goto search_again;
778 }
779
780 /*
781 * | ---- desired range ---- |
782 * | state |
783 * or
784 * | ------------- state -------------- |
785 *
786 * We need to split the extent we found, and may flip bits on
787 * second half.
788 *
789 * If the extent we found extends past our
790 * range, we just split and search again. It'll get split
791 * again the next time though.
792 *
793 * If the extent we found is inside our range, we set the
794 * desired bit on it.
795 */
796 if (state->start < start) {
797 if (state->state & exclusive_bits) {
798 *failed_start = start;
799 err = -EEXIST;
800 goto out;
801 }
802
803 prealloc = alloc_extent_state_atomic(prealloc);
804 BUG_ON(!prealloc);
805 err = split_state(tree, state, prealloc, start);
806 BUG_ON(err == -EEXIST);
807 prealloc = NULL;
808 if (err)
809 goto out;
810 if (state->end <= end) {
811 set_state_bits(tree, state, &bits);
812 cache_state(state, cached_state);
813 merge_state(tree, state);
814 if (last_end == (u64)-1)
815 goto out;
816 start = last_end + 1;
817 }
818 goto search_again;
819 }
820 /*
821 * | ---- desired range ---- |
822 * | state | or | state |
823 *
824 * There's a hole, we need to insert something in it and
825 * ignore the extent we found.
826 */
827 if (state->start > start) {
828 u64 this_end;
829 if (end < last_start)
830 this_end = end;
831 else
832 this_end = last_start - 1;
833
834 prealloc = alloc_extent_state_atomic(prealloc);
835 BUG_ON(!prealloc);
836
837 /*
838 * Avoid to free 'prealloc' if it can be merged with
839 * the later extent.
840 */
841 err = insert_state(tree, prealloc, start, this_end,
842 &bits);
843 BUG_ON(err == -EEXIST);
844 if (err) {
845 free_extent_state(prealloc);
846 prealloc = NULL;
847 goto out;
848 }
849 cache_state(prealloc, cached_state);
850 prealloc = NULL;
851 start = this_end + 1;
852 goto search_again;
853 }
854 /*
855 * | ---- desired range ---- |
856 * | state |
857 * We need to split the extent, and set the bit
858 * on the first half
859 */
860 if (state->start <= end && state->end > end) {
861 if (state->state & exclusive_bits) {
862 *failed_start = start;
863 err = -EEXIST;
864 goto out;
865 }
866
867 prealloc = alloc_extent_state_atomic(prealloc);
868 BUG_ON(!prealloc);
869 err = split_state(tree, state, prealloc, end + 1);
870 BUG_ON(err == -EEXIST);
871
872 set_state_bits(tree, prealloc, &bits);
873 cache_state(prealloc, cached_state);
874 merge_state(tree, prealloc);
875 prealloc = NULL;
876 goto out;
877 }
878
879 goto search_again;
880
881out:
882 spin_unlock(&tree->lock);
883 if (prealloc)
884 free_extent_state(prealloc);
885
886 return err;
887
888search_again:
889 if (start > end)
890 goto out;
891 spin_unlock(&tree->lock);
892 if (mask & __GFP_WAIT)
893 cond_resched();
894 goto again;
895}
896
897/* wrappers around set/clear extent bit */
898int set_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
899 gfp_t mask)
900{
901 return set_extent_bit(tree, start, end, EXTENT_DIRTY, 0, NULL,
902 NULL, mask);
903}
904
905int set_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
906 int bits, gfp_t mask)
907{
908 return set_extent_bit(tree, start, end, bits, 0, NULL,
909 NULL, mask);
910}
911
912int clear_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
913 int bits, gfp_t mask)
914{
915 return clear_extent_bit(tree, start, end, bits, 0, 0, NULL, mask);
916}
917
918int set_extent_delalloc(struct extent_io_tree *tree, u64 start, u64 end,
919 struct extent_state **cached_state, gfp_t mask)
920{
921 return set_extent_bit(tree, start, end,
922 EXTENT_DELALLOC | EXTENT_DIRTY | EXTENT_UPTODATE,
923 0, NULL, cached_state, mask);
924}
925
926int clear_extent_dirty(struct extent_io_tree *tree, u64 start, u64 end,
927 gfp_t mask)
928{
929 return clear_extent_bit(tree, start, end,
930 EXTENT_DIRTY | EXTENT_DELALLOC |
931 EXTENT_DO_ACCOUNTING, 0, 0, NULL, mask);
932}
933
934int set_extent_new(struct extent_io_tree *tree, u64 start, u64 end,
935 gfp_t mask)
936{
937 return set_extent_bit(tree, start, end, EXTENT_NEW, 0, NULL,
938 NULL, mask);
939}
940
941int set_extent_uptodate(struct extent_io_tree *tree, u64 start, u64 end,
942 struct extent_state **cached_state, gfp_t mask)
943{
944 return set_extent_bit(tree, start, end, EXTENT_UPTODATE, 0,
945 NULL, cached_state, mask);
946}
947
948static int clear_extent_uptodate(struct extent_io_tree *tree, u64 start,
949 u64 end, struct extent_state **cached_state,
950 gfp_t mask)
951{
952 return clear_extent_bit(tree, start, end, EXTENT_UPTODATE, 0, 0,
953 cached_state, mask);
954}
955
956/*
957 * either insert or lock state struct between start and end use mask to tell
958 * us if waiting is desired.
959 */
960int lock_extent_bits(struct extent_io_tree *tree, u64 start, u64 end,
961 int bits, struct extent_state **cached_state, gfp_t mask)
962{
963 int err;
964 u64 failed_start;
965 while (1) {
966 err = set_extent_bit(tree, start, end, EXTENT_LOCKED | bits,
967 EXTENT_LOCKED, &failed_start,
968 cached_state, mask);
969 if (err == -EEXIST && (mask & __GFP_WAIT)) {
970 wait_extent_bit(tree, failed_start, end, EXTENT_LOCKED);
971 start = failed_start;
972 } else {
973 break;
974 }
975 WARN_ON(start > end);
976 }
977 return err;
978}
979
980int lock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
981{
982 return lock_extent_bits(tree, start, end, 0, NULL, mask);
983}
984
985int try_lock_extent(struct extent_io_tree *tree, u64 start, u64 end,
986 gfp_t mask)
987{
988 int err;
989 u64 failed_start;
990
991 err = set_extent_bit(tree, start, end, EXTENT_LOCKED, EXTENT_LOCKED,
992 &failed_start, NULL, mask);
993 if (err == -EEXIST) {
994 if (failed_start > start)
995 clear_extent_bit(tree, start, failed_start - 1,
996 EXTENT_LOCKED, 1, 0, NULL, mask);
997 return 0;
998 }
999 return 1;
1000}
1001
1002int unlock_extent_cached(struct extent_io_tree *tree, u64 start, u64 end,
1003 struct extent_state **cached, gfp_t mask)
1004{
1005 return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, cached,
1006 mask);
1007}
1008
1009int unlock_extent(struct extent_io_tree *tree, u64 start, u64 end, gfp_t mask)
1010{
1011 return clear_extent_bit(tree, start, end, EXTENT_LOCKED, 1, 0, NULL,
1012 mask);
1013}
1014
1015/*
1016 * helper function to set both pages and extents in the tree writeback
1017 */
1018static int set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
1019{
1020 unsigned long index = start >> PAGE_CACHE_SHIFT;
1021 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1022 struct page *page;
1023
1024 while (index <= end_index) {
1025 page = find_get_page(tree->mapping, index);
1026 BUG_ON(!page);
1027 set_page_writeback(page);
1028 page_cache_release(page);
1029 index++;
1030 }
1031 return 0;
1032}
1033
1034/* find the first state struct with 'bits' set after 'start', and
1035 * return it. tree->lock must be held. NULL will returned if
1036 * nothing was found after 'start'
1037 */
1038struct extent_state *find_first_extent_bit_state(struct extent_io_tree *tree,
1039 u64 start, int bits)
1040{
1041 struct rb_node *node;
1042 struct extent_state *state;
1043
1044 /*
1045 * this search will find all the extents that end after
1046 * our range starts.
1047 */
1048 node = tree_search(tree, start);
1049 if (!node)
1050 goto out;
1051
1052 while (1) {
1053 state = rb_entry(node, struct extent_state, rb_node);
1054 if (state->end >= start && (state->state & bits))
1055 return state;
1056
1057 node = rb_next(node);
1058 if (!node)
1059 break;
1060 }
1061out:
1062 return NULL;
1063}
1064
1065/*
1066 * find the first offset in the io tree with 'bits' set. zero is
1067 * returned if we find something, and *start_ret and *end_ret are
1068 * set to reflect the state struct that was found.
1069 *
1070 * If nothing was found, 1 is returned, < 0 on error
1071 */
1072int find_first_extent_bit(struct extent_io_tree *tree, u64 start,
1073 u64 *start_ret, u64 *end_ret, int bits)
1074{
1075 struct extent_state *state;
1076 int ret = 1;
1077
1078 spin_lock(&tree->lock);
1079 state = find_first_extent_bit_state(tree, start, bits);
1080 if (state) {
1081 *start_ret = state->start;
1082 *end_ret = state->end;
1083 ret = 0;
1084 }
1085 spin_unlock(&tree->lock);
1086 return ret;
1087}
1088
1089/*
1090 * find a contiguous range of bytes in the file marked as delalloc, not
1091 * more than 'max_bytes'. start and end are used to return the range,
1092 *
1093 * 1 is returned if we find something, 0 if nothing was in the tree
1094 */
1095static noinline u64 find_delalloc_range(struct extent_io_tree *tree,
1096 u64 *start, u64 *end, u64 max_bytes,
1097 struct extent_state **cached_state)
1098{
1099 struct rb_node *node;
1100 struct extent_state *state;
1101 u64 cur_start = *start;
1102 u64 found = 0;
1103 u64 total_bytes = 0;
1104
1105 spin_lock(&tree->lock);
1106
1107 /*
1108 * this search will find all the extents that end after
1109 * our range starts.
1110 */
1111 node = tree_search(tree, cur_start);
1112 if (!node) {
1113 if (!found)
1114 *end = (u64)-1;
1115 goto out;
1116 }
1117
1118 while (1) {
1119 state = rb_entry(node, struct extent_state, rb_node);
1120 if (found && (state->start != cur_start ||
1121 (state->state & EXTENT_BOUNDARY))) {
1122 goto out;
1123 }
1124 if (!(state->state & EXTENT_DELALLOC)) {
1125 if (!found)
1126 *end = state->end;
1127 goto out;
1128 }
1129 if (!found) {
1130 *start = state->start;
1131 *cached_state = state;
1132 atomic_inc(&state->refs);
1133 }
1134 found++;
1135 *end = state->end;
1136 cur_start = state->end + 1;
1137 node = rb_next(node);
1138 if (!node)
1139 break;
1140 total_bytes += state->end - state->start + 1;
1141 if (total_bytes >= max_bytes)
1142 break;
1143 }
1144out:
1145 spin_unlock(&tree->lock);
1146 return found;
1147}
1148
1149static noinline int __unlock_for_delalloc(struct inode *inode,
1150 struct page *locked_page,
1151 u64 start, u64 end)
1152{
1153 int ret;
1154 struct page *pages[16];
1155 unsigned long index = start >> PAGE_CACHE_SHIFT;
1156 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1157 unsigned long nr_pages = end_index - index + 1;
1158 int i;
1159
1160 if (index == locked_page->index && end_index == index)
1161 return 0;
1162
1163 while (nr_pages > 0) {
1164 ret = find_get_pages_contig(inode->i_mapping, index,
1165 min_t(unsigned long, nr_pages,
1166 ARRAY_SIZE(pages)), pages);
1167 for (i = 0; i < ret; i++) {
1168 if (pages[i] != locked_page)
1169 unlock_page(pages[i]);
1170 page_cache_release(pages[i]);
1171 }
1172 nr_pages -= ret;
1173 index += ret;
1174 cond_resched();
1175 }
1176 return 0;
1177}
1178
1179static noinline int lock_delalloc_pages(struct inode *inode,
1180 struct page *locked_page,
1181 u64 delalloc_start,
1182 u64 delalloc_end)
1183{
1184 unsigned long index = delalloc_start >> PAGE_CACHE_SHIFT;
1185 unsigned long start_index = index;
1186 unsigned long end_index = delalloc_end >> PAGE_CACHE_SHIFT;
1187 unsigned long pages_locked = 0;
1188 struct page *pages[16];
1189 unsigned long nrpages;
1190 int ret;
1191 int i;
1192
1193 /* the caller is responsible for locking the start index */
1194 if (index == locked_page->index && index == end_index)
1195 return 0;
1196
1197 /* skip the page at the start index */
1198 nrpages = end_index - index + 1;
1199 while (nrpages > 0) {
1200 ret = find_get_pages_contig(inode->i_mapping, index,
1201 min_t(unsigned long,
1202 nrpages, ARRAY_SIZE(pages)), pages);
1203 if (ret == 0) {
1204 ret = -EAGAIN;
1205 goto done;
1206 }
1207 /* now we have an array of pages, lock them all */
1208 for (i = 0; i < ret; i++) {
1209 /*
1210 * the caller is taking responsibility for
1211 * locked_page
1212 */
1213 if (pages[i] != locked_page) {
1214 lock_page(pages[i]);
1215 if (!PageDirty(pages[i]) ||
1216 pages[i]->mapping != inode->i_mapping) {
1217 ret = -EAGAIN;
1218 unlock_page(pages[i]);
1219 page_cache_release(pages[i]);
1220 goto done;
1221 }
1222 }
1223 page_cache_release(pages[i]);
1224 pages_locked++;
1225 }
1226 nrpages -= ret;
1227 index += ret;
1228 cond_resched();
1229 }
1230 ret = 0;
1231done:
1232 if (ret && pages_locked) {
1233 __unlock_for_delalloc(inode, locked_page,
1234 delalloc_start,
1235 ((u64)(start_index + pages_locked - 1)) <<
1236 PAGE_CACHE_SHIFT);
1237 }
1238 return ret;
1239}
1240
1241/*
1242 * find a contiguous range of bytes in the file marked as delalloc, not
1243 * more than 'max_bytes'. start and end are used to return the range,
1244 *
1245 * 1 is returned if we find something, 0 if nothing was in the tree
1246 */
1247static noinline u64 find_lock_delalloc_range(struct inode *inode,
1248 struct extent_io_tree *tree,
1249 struct page *locked_page,
1250 u64 *start, u64 *end,
1251 u64 max_bytes)
1252{
1253 u64 delalloc_start;
1254 u64 delalloc_end;
1255 u64 found;
1256 struct extent_state *cached_state = NULL;
1257 int ret;
1258 int loops = 0;
1259
1260again:
1261 /* step one, find a bunch of delalloc bytes starting at start */
1262 delalloc_start = *start;
1263 delalloc_end = 0;
1264 found = find_delalloc_range(tree, &delalloc_start, &delalloc_end,
1265 max_bytes, &cached_state);
1266 if (!found || delalloc_end <= *start) {
1267 *start = delalloc_start;
1268 *end = delalloc_end;
1269 free_extent_state(cached_state);
1270 return found;
1271 }
1272
1273 /*
1274 * start comes from the offset of locked_page. We have to lock
1275 * pages in order, so we can't process delalloc bytes before
1276 * locked_page
1277 */
1278 if (delalloc_start < *start)
1279 delalloc_start = *start;
1280
1281 /*
1282 * make sure to limit the number of pages we try to lock down
1283 * if we're looping.
1284 */
1285 if (delalloc_end + 1 - delalloc_start > max_bytes && loops)
1286 delalloc_end = delalloc_start + PAGE_CACHE_SIZE - 1;
1287
1288 /* step two, lock all the pages after the page that has start */
1289 ret = lock_delalloc_pages(inode, locked_page,
1290 delalloc_start, delalloc_end);
1291 if (ret == -EAGAIN) {
1292 /* some of the pages are gone, lets avoid looping by
1293 * shortening the size of the delalloc range we're searching
1294 */
1295 free_extent_state(cached_state);
1296 if (!loops) {
1297 unsigned long offset = (*start) & (PAGE_CACHE_SIZE - 1);
1298 max_bytes = PAGE_CACHE_SIZE - offset;
1299 loops = 1;
1300 goto again;
1301 } else {
1302 found = 0;
1303 goto out_failed;
1304 }
1305 }
1306 BUG_ON(ret);
1307
1308 /* step three, lock the state bits for the whole range */
1309 lock_extent_bits(tree, delalloc_start, delalloc_end,
1310 0, &cached_state, GFP_NOFS);
1311
1312 /* then test to make sure it is all still delalloc */
1313 ret = test_range_bit(tree, delalloc_start, delalloc_end,
1314 EXTENT_DELALLOC, 1, cached_state);
1315 if (!ret) {
1316 unlock_extent_cached(tree, delalloc_start, delalloc_end,
1317 &cached_state, GFP_NOFS);
1318 __unlock_for_delalloc(inode, locked_page,
1319 delalloc_start, delalloc_end);
1320 cond_resched();
1321 goto again;
1322 }
1323 free_extent_state(cached_state);
1324 *start = delalloc_start;
1325 *end = delalloc_end;
1326out_failed:
1327 return found;
1328}
1329
1330int extent_clear_unlock_delalloc(struct inode *inode,
1331 struct extent_io_tree *tree,
1332 u64 start, u64 end, struct page *locked_page,
1333 unsigned long op)
1334{
1335 int ret;
1336 struct page *pages[16];
1337 unsigned long index = start >> PAGE_CACHE_SHIFT;
1338 unsigned long end_index = end >> PAGE_CACHE_SHIFT;
1339 unsigned long nr_pages = end_index - index + 1;
1340 int i;
1341 int clear_bits = 0;
1342
1343 if (op & EXTENT_CLEAR_UNLOCK)
1344 clear_bits |= EXTENT_LOCKED;
1345 if (op & EXTENT_CLEAR_DIRTY)
1346 clear_bits |= EXTENT_DIRTY;
1347
1348 if (op & EXTENT_CLEAR_DELALLOC)
1349 clear_bits |= EXTENT_DELALLOC;
1350
1351 clear_extent_bit(tree, start, end, clear_bits, 1, 0, NULL, GFP_NOFS);
1352 if (!(op & (EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
1353 EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK |
1354 EXTENT_SET_PRIVATE2)))
1355 return 0;
1356
1357 while (nr_pages > 0) {
1358 ret = find_get_pages_contig(inode->i_mapping, index,
1359 min_t(unsigned long,
1360 nr_pages, ARRAY_SIZE(pages)), pages);
1361 for (i = 0; i < ret; i++) {
1362
1363 if (op & EXTENT_SET_PRIVATE2)
1364 SetPagePrivate2(pages[i]);
1365
1366 if (pages[i] == locked_page) {
1367 page_cache_release(pages[i]);
1368 continue;
1369 }
1370 if (op & EXTENT_CLEAR_DIRTY)
1371 clear_page_dirty_for_io(pages[i]);
1372 if (op & EXTENT_SET_WRITEBACK)
1373 set_page_writeback(pages[i]);
1374 if (op & EXTENT_END_WRITEBACK)
1375 end_page_writeback(pages[i]);
1376 if (op & EXTENT_CLEAR_UNLOCK_PAGE)
1377 unlock_page(pages[i]);
1378 page_cache_release(pages[i]);
1379 }
1380 nr_pages -= ret;
1381 index += ret;
1382 cond_resched();
1383 }
1384 return 0;
1385}
1386
1387/*
1388 * count the number of bytes in the tree that have a given bit(s)
1389 * set. This can be fairly slow, except for EXTENT_DIRTY which is
1390 * cached. The total number found is returned.
1391 */
1392u64 count_range_bits(struct extent_io_tree *tree,
1393 u64 *start, u64 search_end, u64 max_bytes,
1394 unsigned long bits, int contig)
1395{
1396 struct rb_node *node;
1397 struct extent_state *state;
1398 u64 cur_start = *start;
1399 u64 total_bytes = 0;
1400 u64 last = 0;
1401 int found = 0;
1402
1403 if (search_end <= cur_start) {
1404 WARN_ON(1);
1405 return 0;
1406 }
1407
1408 spin_lock(&tree->lock);
1409 if (cur_start == 0 && bits == EXTENT_DIRTY) {
1410 total_bytes = tree->dirty_bytes;
1411 goto out;
1412 }
1413 /*
1414 * this search will find all the extents that end after
1415 * our range starts.
1416 */
1417 node = tree_search(tree, cur_start);
1418 if (!node)
1419 goto out;
1420
1421 while (1) {
1422 state = rb_entry(node, struct extent_state, rb_node);
1423 if (state->start > search_end)
1424 break;
1425 if (contig && found && state->start > last + 1)
1426 break;
1427 if (state->end >= cur_start && (state->state & bits) == bits) {
1428 total_bytes += min(search_end, state->end) + 1 -
1429 max(cur_start, state->start);
1430 if (total_bytes >= max_bytes)
1431 break;
1432 if (!found) {
1433 *start = max(cur_start, state->start);
1434 found = 1;
1435 }
1436 last = state->end;
1437 } else if (contig && found) {
1438 break;
1439 }
1440 node = rb_next(node);
1441 if (!node)
1442 break;
1443 }
1444out:
1445 spin_unlock(&tree->lock);
1446 return total_bytes;
1447}
1448
1449/*
1450 * set the private field for a given byte offset in the tree. If there isn't
1451 * an extent_state there already, this does nothing.
1452 */
1453int set_state_private(struct extent_io_tree *tree, u64 start, u64 private)
1454{
1455 struct rb_node *node;
1456 struct extent_state *state;
1457 int ret = 0;
1458
1459 spin_lock(&tree->lock);
1460 /*
1461 * this search will find all the extents that end after
1462 * our range starts.
1463 */
1464 node = tree_search(tree, start);
1465 if (!node) {
1466 ret = -ENOENT;
1467 goto out;
1468 }
1469 state = rb_entry(node, struct extent_state, rb_node);
1470 if (state->start != start) {
1471 ret = -ENOENT;
1472 goto out;
1473 }
1474 state->private = private;
1475out:
1476 spin_unlock(&tree->lock);
1477 return ret;
1478}
1479
1480int get_state_private(struct extent_io_tree *tree, u64 start, u64 *private)
1481{
1482 struct rb_node *node;
1483 struct extent_state *state;
1484 int ret = 0;
1485
1486 spin_lock(&tree->lock);
1487 /*
1488 * this search will find all the extents that end after
1489 * our range starts.
1490 */
1491 node = tree_search(tree, start);
1492 if (!node) {
1493 ret = -ENOENT;
1494 goto out;
1495 }
1496 state = rb_entry(node, struct extent_state, rb_node);
1497 if (state->start != start) {
1498 ret = -ENOENT;
1499 goto out;
1500 }
1501 *private = state->private;
1502out:
1503 spin_unlock(&tree->lock);
1504 return ret;
1505}
1506
1507/*
1508 * searches a range in the state tree for a given mask.
1509 * If 'filled' == 1, this returns 1 only if every extent in the tree
1510 * has the bits set. Otherwise, 1 is returned if any bit in the
1511 * range is found set.
1512 */
1513int test_range_bit(struct extent_io_tree *tree, u64 start, u64 end,
1514 int bits, int filled, struct extent_state *cached)
1515{
1516 struct extent_state *state = NULL;
1517 struct rb_node *node;
1518 int bitset = 0;
1519
1520 spin_lock(&tree->lock);
1521 if (cached && cached->tree && cached->start <= start &&
1522 cached->end > start)
1523 node = &cached->rb_node;
1524 else
1525 node = tree_search(tree, start);
1526 while (node && start <= end) {
1527 state = rb_entry(node, struct extent_state, rb_node);
1528
1529 if (filled && state->start > start) {
1530 bitset = 0;
1531 break;
1532 }
1533
1534 if (state->start > end)
1535 break;
1536
1537 if (state->state & bits) {
1538 bitset = 1;
1539 if (!filled)
1540 break;
1541 } else if (filled) {
1542 bitset = 0;
1543 break;
1544 }
1545
1546 if (state->end == (u64)-1)
1547 break;
1548
1549 start = state->end + 1;
1550 if (start > end)
1551 break;
1552 node = rb_next(node);
1553 if (!node) {
1554 if (filled)
1555 bitset = 0;
1556 break;
1557 }
1558 }
1559 spin_unlock(&tree->lock);
1560 return bitset;
1561}
1562
1563/*
1564 * helper function to set a given page up to date if all the
1565 * extents in the tree for that page are up to date
1566 */
1567static int check_page_uptodate(struct extent_io_tree *tree,
1568 struct page *page)
1569{
1570 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1571 u64 end = start + PAGE_CACHE_SIZE - 1;
1572 if (test_range_bit(tree, start, end, EXTENT_UPTODATE, 1, NULL))
1573 SetPageUptodate(page);
1574 return 0;
1575}
1576
1577/*
1578 * helper function to unlock a page if all the extents in the tree
1579 * for that page are unlocked
1580 */
1581static int check_page_locked(struct extent_io_tree *tree,
1582 struct page *page)
1583{
1584 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1585 u64 end = start + PAGE_CACHE_SIZE - 1;
1586 if (!test_range_bit(tree, start, end, EXTENT_LOCKED, 0, NULL))
1587 unlock_page(page);
1588 return 0;
1589}
1590
1591/*
1592 * helper function to end page writeback if all the extents
1593 * in the tree for that page are done with writeback
1594 */
1595static int check_page_writeback(struct extent_io_tree *tree,
1596 struct page *page)
1597{
1598 end_page_writeback(page);
1599 return 0;
1600}
1601
1602/* lots and lots of room for performance fixes in the end_bio funcs */
1603
1604/*
1605 * after a writepage IO is done, we need to:
1606 * clear the uptodate bits on error
1607 * clear the writeback bits in the extent tree for this IO
1608 * end_page_writeback if the page has no more pending IO
1609 *
1610 * Scheduling is not allowed, so the extent state tree is expected
1611 * to have one and only one object corresponding to this IO.
1612 */
1613static void end_bio_extent_writepage(struct bio *bio, int err)
1614{
1615 int uptodate = err == 0;
1616 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1617 struct extent_io_tree *tree;
1618 u64 start;
1619 u64 end;
1620 int whole_page;
1621 int ret;
1622
1623 do {
1624 struct page *page = bvec->bv_page;
1625 tree = &BTRFS_I(page->mapping->host)->io_tree;
1626
1627 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1628 bvec->bv_offset;
1629 end = start + bvec->bv_len - 1;
1630
1631 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1632 whole_page = 1;
1633 else
1634 whole_page = 0;
1635
1636 if (--bvec >= bio->bi_io_vec)
1637 prefetchw(&bvec->bv_page->flags);
1638 if (tree->ops && tree->ops->writepage_end_io_hook) {
1639 ret = tree->ops->writepage_end_io_hook(page, start,
1640 end, NULL, uptodate);
1641 if (ret)
1642 uptodate = 0;
1643 }
1644
1645 if (!uptodate && tree->ops &&
1646 tree->ops->writepage_io_failed_hook) {
1647 ret = tree->ops->writepage_io_failed_hook(bio, page,
1648 start, end, NULL);
1649 if (ret == 0) {
1650 uptodate = (err == 0);
1651 continue;
1652 }
1653 }
1654
1655 if (!uptodate) {
1656 clear_extent_uptodate(tree, start, end, NULL, GFP_NOFS);
1657 ClearPageUptodate(page);
1658 SetPageError(page);
1659 }
1660
1661 if (whole_page)
1662 end_page_writeback(page);
1663 else
1664 check_page_writeback(tree, page);
1665 } while (bvec >= bio->bi_io_vec);
1666
1667 bio_put(bio);
1668}
1669
1670/*
1671 * after a readpage IO is done, we need to:
1672 * clear the uptodate bits on error
1673 * set the uptodate bits if things worked
1674 * set the page up to date if all extents in the tree are uptodate
1675 * clear the lock bit in the extent tree
1676 * unlock the page if there are no other extents locked for it
1677 *
1678 * Scheduling is not allowed, so the extent state tree is expected
1679 * to have one and only one object corresponding to this IO.
1680 */
1681static void end_bio_extent_readpage(struct bio *bio, int err)
1682{
1683 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1684 struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
1685 struct bio_vec *bvec = bio->bi_io_vec;
1686 struct extent_io_tree *tree;
1687 u64 start;
1688 u64 end;
1689 int whole_page;
1690 int ret;
1691
1692 if (err)
1693 uptodate = 0;
1694
1695 do {
1696 struct page *page = bvec->bv_page;
1697 struct extent_state *cached = NULL;
1698 struct extent_state *state;
1699
1700 tree = &BTRFS_I(page->mapping->host)->io_tree;
1701
1702 start = ((u64)page->index << PAGE_CACHE_SHIFT) +
1703 bvec->bv_offset;
1704 end = start + bvec->bv_len - 1;
1705
1706 if (bvec->bv_offset == 0 && bvec->bv_len == PAGE_CACHE_SIZE)
1707 whole_page = 1;
1708 else
1709 whole_page = 0;
1710
1711 if (++bvec <= bvec_end)
1712 prefetchw(&bvec->bv_page->flags);
1713
1714 spin_lock(&tree->lock);
1715 state = find_first_extent_bit_state(tree, start, EXTENT_LOCKED);
1716 if (state && state->start == start) {
1717 /*
1718 * take a reference on the state, unlock will drop
1719 * the ref
1720 */
1721 cache_state(state, &cached);
1722 }
1723 spin_unlock(&tree->lock);
1724
1725 if (uptodate && tree->ops && tree->ops->readpage_end_io_hook) {
1726 ret = tree->ops->readpage_end_io_hook(page, start, end,
1727 state);
1728 if (ret)
1729 uptodate = 0;
1730 }
1731 if (!uptodate && tree->ops &&
1732 tree->ops->readpage_io_failed_hook) {
1733 ret = tree->ops->readpage_io_failed_hook(bio, page,
1734 start, end, NULL);
1735 if (ret == 0) {
1736 uptodate =
1737 test_bit(BIO_UPTODATE, &bio->bi_flags);
1738 if (err)
1739 uptodate = 0;
1740 uncache_state(&cached);
1741 continue;
1742 }
1743 }
1744
1745 if (uptodate) {
1746 set_extent_uptodate(tree, start, end, &cached,
1747 GFP_ATOMIC);
1748 }
1749 unlock_extent_cached(tree, start, end, &cached, GFP_ATOMIC);
1750
1751 if (whole_page) {
1752 if (uptodate) {
1753 SetPageUptodate(page);
1754 } else {
1755 ClearPageUptodate(page);
1756 SetPageError(page);
1757 }
1758 unlock_page(page);
1759 } else {
1760 if (uptodate) {
1761 check_page_uptodate(tree, page);
1762 } else {
1763 ClearPageUptodate(page);
1764 SetPageError(page);
1765 }
1766 check_page_locked(tree, page);
1767 }
1768 } while (bvec <= bvec_end);
1769
1770 bio_put(bio);
1771}
1772
1773struct bio *
1774btrfs_bio_alloc(struct block_device *bdev, u64 first_sector, int nr_vecs,
1775 gfp_t gfp_flags)
1776{
1777 struct bio *bio;
1778
1779 bio = bio_alloc(gfp_flags, nr_vecs);
1780
1781 if (bio == NULL && (current->flags & PF_MEMALLOC)) {
1782 while (!bio && (nr_vecs /= 2))
1783 bio = bio_alloc(gfp_flags, nr_vecs);
1784 }
1785
1786 if (bio) {
1787 bio->bi_size = 0;
1788 bio->bi_bdev = bdev;
1789 bio->bi_sector = first_sector;
1790 }
1791 return bio;
1792}
1793
1794static int submit_one_bio(int rw, struct bio *bio, int mirror_num,
1795 unsigned long bio_flags)
1796{
1797 int ret = 0;
1798 struct bio_vec *bvec = bio->bi_io_vec + bio->bi_vcnt - 1;
1799 struct page *page = bvec->bv_page;
1800 struct extent_io_tree *tree = bio->bi_private;
1801 u64 start;
1802
1803 start = ((u64)page->index << PAGE_CACHE_SHIFT) + bvec->bv_offset;
1804
1805 bio->bi_private = NULL;
1806
1807 bio_get(bio);
1808
1809 if (tree->ops && tree->ops->submit_bio_hook)
1810 ret = tree->ops->submit_bio_hook(page->mapping->host, rw, bio,
1811 mirror_num, bio_flags, start);
1812 else
1813 submit_bio(rw, bio);
1814 if (bio_flagged(bio, BIO_EOPNOTSUPP))
1815 ret = -EOPNOTSUPP;
1816 bio_put(bio);
1817 return ret;
1818}
1819
1820static int submit_extent_page(int rw, struct extent_io_tree *tree,
1821 struct page *page, sector_t sector,
1822 size_t size, unsigned long offset,
1823 struct block_device *bdev,
1824 struct bio **bio_ret,
1825 unsigned long max_pages,
1826 bio_end_io_t end_io_func,
1827 int mirror_num,
1828 unsigned long prev_bio_flags,
1829 unsigned long bio_flags)
1830{
1831 int ret = 0;
1832 struct bio *bio;
1833 int nr;
1834 int contig = 0;
1835 int this_compressed = bio_flags & EXTENT_BIO_COMPRESSED;
1836 int old_compressed = prev_bio_flags & EXTENT_BIO_COMPRESSED;
1837 size_t page_size = min_t(size_t, size, PAGE_CACHE_SIZE);
1838
1839 if (bio_ret && *bio_ret) {
1840 bio = *bio_ret;
1841 if (old_compressed)
1842 contig = bio->bi_sector == sector;
1843 else
1844 contig = bio->bi_sector + (bio->bi_size >> 9) ==
1845 sector;
1846
1847 if (prev_bio_flags != bio_flags || !contig ||
1848 (tree->ops && tree->ops->merge_bio_hook &&
1849 tree->ops->merge_bio_hook(page, offset, page_size, bio,
1850 bio_flags)) ||
1851 bio_add_page(bio, page, page_size, offset) < page_size) {
1852 ret = submit_one_bio(rw, bio, mirror_num,
1853 prev_bio_flags);
1854 bio = NULL;
1855 } else {
1856 return 0;
1857 }
1858 }
1859 if (this_compressed)
1860 nr = BIO_MAX_PAGES;
1861 else
1862 nr = bio_get_nr_vecs(bdev);
1863
1864 bio = btrfs_bio_alloc(bdev, sector, nr, GFP_NOFS | __GFP_HIGH);
1865 if (!bio)
1866 return -ENOMEM;
1867
1868 bio_add_page(bio, page, page_size, offset);
1869 bio->bi_end_io = end_io_func;
1870 bio->bi_private = tree;
1871
1872 if (bio_ret)
1873 *bio_ret = bio;
1874 else
1875 ret = submit_one_bio(rw, bio, mirror_num, bio_flags);
1876
1877 return ret;
1878}
1879
1880void set_page_extent_mapped(struct page *page)
1881{
1882 if (!PagePrivate(page)) {
1883 SetPagePrivate(page);
1884 page_cache_get(page);
1885 set_page_private(page, EXTENT_PAGE_PRIVATE);
1886 }
1887}
1888
1889static void set_page_extent_head(struct page *page, unsigned long len)
1890{
1891 WARN_ON(!PagePrivate(page));
1892 set_page_private(page, EXTENT_PAGE_PRIVATE_FIRST_PAGE | len << 2);
1893}
1894
1895/*
1896 * basic readpage implementation. Locked extent state structs are inserted
1897 * into the tree that are removed when the IO is done (by the end_io
1898 * handlers)
1899 */
1900static int __extent_read_full_page(struct extent_io_tree *tree,
1901 struct page *page,
1902 get_extent_t *get_extent,
1903 struct bio **bio, int mirror_num,
1904 unsigned long *bio_flags)
1905{
1906 struct inode *inode = page->mapping->host;
1907 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
1908 u64 page_end = start + PAGE_CACHE_SIZE - 1;
1909 u64 end;
1910 u64 cur = start;
1911 u64 extent_offset;
1912 u64 last_byte = i_size_read(inode);
1913 u64 block_start;
1914 u64 cur_end;
1915 sector_t sector;
1916 struct extent_map *em;
1917 struct block_device *bdev;
1918 struct btrfs_ordered_extent *ordered;
1919 int ret;
1920 int nr = 0;
1921 size_t pg_offset = 0;
1922 size_t iosize;
1923 size_t disk_io_size;
1924 size_t blocksize = inode->i_sb->s_blocksize;
1925 unsigned long this_bio_flag = 0;
1926
1927 set_page_extent_mapped(page);
1928
1929 if (!PageUptodate(page)) {
1930 if (cleancache_get_page(page) == 0) {
1931 BUG_ON(blocksize != PAGE_SIZE);
1932 goto out;
1933 }
1934 }
1935
1936 end = page_end;
1937 while (1) {
1938 lock_extent(tree, start, end, GFP_NOFS);
1939 ordered = btrfs_lookup_ordered_extent(inode, start);
1940 if (!ordered)
1941 break;
1942 unlock_extent(tree, start, end, GFP_NOFS);
1943 btrfs_start_ordered_extent(inode, ordered, 1);
1944 btrfs_put_ordered_extent(ordered);
1945 }
1946
1947 if (page->index == last_byte >> PAGE_CACHE_SHIFT) {
1948 char *userpage;
1949 size_t zero_offset = last_byte & (PAGE_CACHE_SIZE - 1);
1950
1951 if (zero_offset) {
1952 iosize = PAGE_CACHE_SIZE - zero_offset;
1953 userpage = kmap_atomic(page, KM_USER0);
1954 memset(userpage + zero_offset, 0, iosize);
1955 flush_dcache_page(page);
1956 kunmap_atomic(userpage, KM_USER0);
1957 }
1958 }
1959 while (cur <= end) {
1960 if (cur >= last_byte) {
1961 char *userpage;
1962 struct extent_state *cached = NULL;
1963
1964 iosize = PAGE_CACHE_SIZE - pg_offset;
1965 userpage = kmap_atomic(page, KM_USER0);
1966 memset(userpage + pg_offset, 0, iosize);
1967 flush_dcache_page(page);
1968 kunmap_atomic(userpage, KM_USER0);
1969 set_extent_uptodate(tree, cur, cur + iosize - 1,
1970 &cached, GFP_NOFS);
1971 unlock_extent_cached(tree, cur, cur + iosize - 1,
1972 &cached, GFP_NOFS);
1973 break;
1974 }
1975 em = get_extent(inode, page, pg_offset, cur,
1976 end - cur + 1, 0);
1977 if (IS_ERR_OR_NULL(em)) {
1978 SetPageError(page);
1979 unlock_extent(tree, cur, end, GFP_NOFS);
1980 break;
1981 }
1982 extent_offset = cur - em->start;
1983 BUG_ON(extent_map_end(em) <= cur);
1984 BUG_ON(end < cur);
1985
1986 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
1987 this_bio_flag = EXTENT_BIO_COMPRESSED;
1988 extent_set_compress_type(&this_bio_flag,
1989 em->compress_type);
1990 }
1991
1992 iosize = min(extent_map_end(em) - cur, end - cur + 1);
1993 cur_end = min(extent_map_end(em) - 1, end);
1994 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
1995 if (this_bio_flag & EXTENT_BIO_COMPRESSED) {
1996 disk_io_size = em->block_len;
1997 sector = em->block_start >> 9;
1998 } else {
1999 sector = (em->block_start + extent_offset) >> 9;
2000 disk_io_size = iosize;
2001 }
2002 bdev = em->bdev;
2003 block_start = em->block_start;
2004 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2005 block_start = EXTENT_MAP_HOLE;
2006 free_extent_map(em);
2007 em = NULL;
2008
2009 /* we've found a hole, just zero and go on */
2010 if (block_start == EXTENT_MAP_HOLE) {
2011 char *userpage;
2012 struct extent_state *cached = NULL;
2013
2014 userpage = kmap_atomic(page, KM_USER0);
2015 memset(userpage + pg_offset, 0, iosize);
2016 flush_dcache_page(page);
2017 kunmap_atomic(userpage, KM_USER0);
2018
2019 set_extent_uptodate(tree, cur, cur + iosize - 1,
2020 &cached, GFP_NOFS);
2021 unlock_extent_cached(tree, cur, cur + iosize - 1,
2022 &cached, GFP_NOFS);
2023 cur = cur + iosize;
2024 pg_offset += iosize;
2025 continue;
2026 }
2027 /* the get_extent function already copied into the page */
2028 if (test_range_bit(tree, cur, cur_end,
2029 EXTENT_UPTODATE, 1, NULL)) {
2030 check_page_uptodate(tree, page);
2031 unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2032 cur = cur + iosize;
2033 pg_offset += iosize;
2034 continue;
2035 }
2036 /* we have an inline extent but it didn't get marked up
2037 * to date. Error out
2038 */
2039 if (block_start == EXTENT_MAP_INLINE) {
2040 SetPageError(page);
2041 unlock_extent(tree, cur, cur + iosize - 1, GFP_NOFS);
2042 cur = cur + iosize;
2043 pg_offset += iosize;
2044 continue;
2045 }
2046
2047 ret = 0;
2048 if (tree->ops && tree->ops->readpage_io_hook) {
2049 ret = tree->ops->readpage_io_hook(page, cur,
2050 cur + iosize - 1);
2051 }
2052 if (!ret) {
2053 unsigned long pnr = (last_byte >> PAGE_CACHE_SHIFT) + 1;
2054 pnr -= page->index;
2055 ret = submit_extent_page(READ, tree, page,
2056 sector, disk_io_size, pg_offset,
2057 bdev, bio, pnr,
2058 end_bio_extent_readpage, mirror_num,
2059 *bio_flags,
2060 this_bio_flag);
2061 nr++;
2062 *bio_flags = this_bio_flag;
2063 }
2064 if (ret)
2065 SetPageError(page);
2066 cur = cur + iosize;
2067 pg_offset += iosize;
2068 }
2069out:
2070 if (!nr) {
2071 if (!PageError(page))
2072 SetPageUptodate(page);
2073 unlock_page(page);
2074 }
2075 return 0;
2076}
2077
2078int extent_read_full_page(struct extent_io_tree *tree, struct page *page,
2079 get_extent_t *get_extent)
2080{
2081 struct bio *bio = NULL;
2082 unsigned long bio_flags = 0;
2083 int ret;
2084
2085 ret = __extent_read_full_page(tree, page, get_extent, &bio, 0,
2086 &bio_flags);
2087 if (bio)
2088 ret = submit_one_bio(READ, bio, 0, bio_flags);
2089 return ret;
2090}
2091
2092static noinline void update_nr_written(struct page *page,
2093 struct writeback_control *wbc,
2094 unsigned long nr_written)
2095{
2096 wbc->nr_to_write -= nr_written;
2097 if (wbc->range_cyclic || (wbc->nr_to_write > 0 &&
2098 wbc->range_start == 0 && wbc->range_end == LLONG_MAX))
2099 page->mapping->writeback_index = page->index + nr_written;
2100}
2101
2102/*
2103 * the writepage semantics are similar to regular writepage. extent
2104 * records are inserted to lock ranges in the tree, and as dirty areas
2105 * are found, they are marked writeback. Then the lock bits are removed
2106 * and the end_io handler clears the writeback ranges
2107 */
2108static int __extent_writepage(struct page *page, struct writeback_control *wbc,
2109 void *data)
2110{
2111 struct inode *inode = page->mapping->host;
2112 struct extent_page_data *epd = data;
2113 struct extent_io_tree *tree = epd->tree;
2114 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2115 u64 delalloc_start;
2116 u64 page_end = start + PAGE_CACHE_SIZE - 1;
2117 u64 end;
2118 u64 cur = start;
2119 u64 extent_offset;
2120 u64 last_byte = i_size_read(inode);
2121 u64 block_start;
2122 u64 iosize;
2123 sector_t sector;
2124 struct extent_state *cached_state = NULL;
2125 struct extent_map *em;
2126 struct block_device *bdev;
2127 int ret;
2128 int nr = 0;
2129 size_t pg_offset = 0;
2130 size_t blocksize;
2131 loff_t i_size = i_size_read(inode);
2132 unsigned long end_index = i_size >> PAGE_CACHE_SHIFT;
2133 u64 nr_delalloc;
2134 u64 delalloc_end;
2135 int page_started;
2136 int compressed;
2137 int write_flags;
2138 unsigned long nr_written = 0;
2139
2140 if (wbc->sync_mode == WB_SYNC_ALL)
2141 write_flags = WRITE_SYNC;
2142 else
2143 write_flags = WRITE;
2144
2145 trace___extent_writepage(page, inode, wbc);
2146
2147 WARN_ON(!PageLocked(page));
2148 pg_offset = i_size & (PAGE_CACHE_SIZE - 1);
2149 if (page->index > end_index ||
2150 (page->index == end_index && !pg_offset)) {
2151 page->mapping->a_ops->invalidatepage(page, 0);
2152 unlock_page(page);
2153 return 0;
2154 }
2155
2156 if (page->index == end_index) {
2157 char *userpage;
2158
2159 userpage = kmap_atomic(page, KM_USER0);
2160 memset(userpage + pg_offset, 0,
2161 PAGE_CACHE_SIZE - pg_offset);
2162 kunmap_atomic(userpage, KM_USER0);
2163 flush_dcache_page(page);
2164 }
2165 pg_offset = 0;
2166
2167 set_page_extent_mapped(page);
2168
2169 delalloc_start = start;
2170 delalloc_end = 0;
2171 page_started = 0;
2172 if (!epd->extent_locked) {
2173 u64 delalloc_to_write = 0;
2174 /*
2175 * make sure the wbc mapping index is at least updated
2176 * to this page.
2177 */
2178 update_nr_written(page, wbc, 0);
2179
2180 while (delalloc_end < page_end) {
2181 nr_delalloc = find_lock_delalloc_range(inode, tree,
2182 page,
2183 &delalloc_start,
2184 &delalloc_end,
2185 128 * 1024 * 1024);
2186 if (nr_delalloc == 0) {
2187 delalloc_start = delalloc_end + 1;
2188 continue;
2189 }
2190 tree->ops->fill_delalloc(inode, page, delalloc_start,
2191 delalloc_end, &page_started,
2192 &nr_written);
2193 /*
2194 * delalloc_end is already one less than the total
2195 * length, so we don't subtract one from
2196 * PAGE_CACHE_SIZE
2197 */
2198 delalloc_to_write += (delalloc_end - delalloc_start +
2199 PAGE_CACHE_SIZE) >>
2200 PAGE_CACHE_SHIFT;
2201 delalloc_start = delalloc_end + 1;
2202 }
2203 if (wbc->nr_to_write < delalloc_to_write) {
2204 int thresh = 8192;
2205
2206 if (delalloc_to_write < thresh * 2)
2207 thresh = delalloc_to_write;
2208 wbc->nr_to_write = min_t(u64, delalloc_to_write,
2209 thresh);
2210 }
2211
2212 /* did the fill delalloc function already unlock and start
2213 * the IO?
2214 */
2215 if (page_started) {
2216 ret = 0;
2217 /*
2218 * we've unlocked the page, so we can't update
2219 * the mapping's writeback index, just update
2220 * nr_to_write.
2221 */
2222 wbc->nr_to_write -= nr_written;
2223 goto done_unlocked;
2224 }
2225 }
2226 if (tree->ops && tree->ops->writepage_start_hook) {
2227 ret = tree->ops->writepage_start_hook(page, start,
2228 page_end);
2229 if (ret == -EAGAIN) {
2230 redirty_page_for_writepage(wbc, page);
2231 update_nr_written(page, wbc, nr_written);
2232 unlock_page(page);
2233 ret = 0;
2234 goto done_unlocked;
2235 }
2236 }
2237
2238 /*
2239 * we don't want to touch the inode after unlocking the page,
2240 * so we update the mapping writeback index now
2241 */
2242 update_nr_written(page, wbc, nr_written + 1);
2243
2244 end = page_end;
2245 if (last_byte <= start) {
2246 if (tree->ops && tree->ops->writepage_end_io_hook)
2247 tree->ops->writepage_end_io_hook(page, start,
2248 page_end, NULL, 1);
2249 goto done;
2250 }
2251
2252 blocksize = inode->i_sb->s_blocksize;
2253
2254 while (cur <= end) {
2255 if (cur >= last_byte) {
2256 if (tree->ops && tree->ops->writepage_end_io_hook)
2257 tree->ops->writepage_end_io_hook(page, cur,
2258 page_end, NULL, 1);
2259 break;
2260 }
2261 em = epd->get_extent(inode, page, pg_offset, cur,
2262 end - cur + 1, 1);
2263 if (IS_ERR_OR_NULL(em)) {
2264 SetPageError(page);
2265 break;
2266 }
2267
2268 extent_offset = cur - em->start;
2269 BUG_ON(extent_map_end(em) <= cur);
2270 BUG_ON(end < cur);
2271 iosize = min(extent_map_end(em) - cur, end - cur + 1);
2272 iosize = (iosize + blocksize - 1) & ~((u64)blocksize - 1);
2273 sector = (em->block_start + extent_offset) >> 9;
2274 bdev = em->bdev;
2275 block_start = em->block_start;
2276 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
2277 free_extent_map(em);
2278 em = NULL;
2279
2280 /*
2281 * compressed and inline extents are written through other
2282 * paths in the FS
2283 */
2284 if (compressed || block_start == EXTENT_MAP_HOLE ||
2285 block_start == EXTENT_MAP_INLINE) {
2286 /*
2287 * end_io notification does not happen here for
2288 * compressed extents
2289 */
2290 if (!compressed && tree->ops &&
2291 tree->ops->writepage_end_io_hook)
2292 tree->ops->writepage_end_io_hook(page, cur,
2293 cur + iosize - 1,
2294 NULL, 1);
2295 else if (compressed) {
2296 /* we don't want to end_page_writeback on
2297 * a compressed extent. this happens
2298 * elsewhere
2299 */
2300 nr++;
2301 }
2302
2303 cur += iosize;
2304 pg_offset += iosize;
2305 continue;
2306 }
2307 /* leave this out until we have a page_mkwrite call */
2308 if (0 && !test_range_bit(tree, cur, cur + iosize - 1,
2309 EXTENT_DIRTY, 0, NULL)) {
2310 cur = cur + iosize;
2311 pg_offset += iosize;
2312 continue;
2313 }
2314
2315 if (tree->ops && tree->ops->writepage_io_hook) {
2316 ret = tree->ops->writepage_io_hook(page, cur,
2317 cur + iosize - 1);
2318 } else {
2319 ret = 0;
2320 }
2321 if (ret) {
2322 SetPageError(page);
2323 } else {
2324 unsigned long max_nr = end_index + 1;
2325
2326 set_range_writeback(tree, cur, cur + iosize - 1);
2327 if (!PageWriteback(page)) {
2328 printk(KERN_ERR "btrfs warning page %lu not "
2329 "writeback, cur %llu end %llu\n",
2330 page->index, (unsigned long long)cur,
2331 (unsigned long long)end);
2332 }
2333
2334 ret = submit_extent_page(write_flags, tree, page,
2335 sector, iosize, pg_offset,
2336 bdev, &epd->bio, max_nr,
2337 end_bio_extent_writepage,
2338 0, 0, 0);
2339 if (ret)
2340 SetPageError(page);
2341 }
2342 cur = cur + iosize;
2343 pg_offset += iosize;
2344 nr++;
2345 }
2346done:
2347 if (nr == 0) {
2348 /* make sure the mapping tag for page dirty gets cleared */
2349 set_page_writeback(page);
2350 end_page_writeback(page);
2351 }
2352 unlock_page(page);
2353
2354done_unlocked:
2355
2356 /* drop our reference on any cached states */
2357 free_extent_state(cached_state);
2358 return 0;
2359}
2360
2361/**
2362 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
2363 * @mapping: address space structure to write
2364 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2365 * @writepage: function called for each page
2366 * @data: data passed to writepage function
2367 *
2368 * If a page is already under I/O, write_cache_pages() skips it, even
2369 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2370 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2371 * and msync() need to guarantee that all the data which was dirty at the time
2372 * the call was made get new I/O started against them. If wbc->sync_mode is
2373 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2374 * existing IO to complete.
2375 */
2376static int extent_write_cache_pages(struct extent_io_tree *tree,
2377 struct address_space *mapping,
2378 struct writeback_control *wbc,
2379 writepage_t writepage, void *data,
2380 void (*flush_fn)(void *))
2381{
2382 int ret = 0;
2383 int done = 0;
2384 int nr_to_write_done = 0;
2385 struct pagevec pvec;
2386 int nr_pages;
2387 pgoff_t index;
2388 pgoff_t end; /* Inclusive */
2389 int scanned = 0;
2390 int tag;
2391
2392 pagevec_init(&pvec, 0);
2393 if (wbc->range_cyclic) {
2394 index = mapping->writeback_index; /* Start from prev offset */
2395 end = -1;
2396 } else {
2397 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2398 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2399 scanned = 1;
2400 }
2401 if (wbc->sync_mode == WB_SYNC_ALL)
2402 tag = PAGECACHE_TAG_TOWRITE;
2403 else
2404 tag = PAGECACHE_TAG_DIRTY;
2405retry:
2406 if (wbc->sync_mode == WB_SYNC_ALL)
2407 tag_pages_for_writeback(mapping, index, end);
2408 while (!done && !nr_to_write_done && (index <= end) &&
2409 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2410 min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
2411 unsigned i;
2412
2413 scanned = 1;
2414 for (i = 0; i < nr_pages; i++) {
2415 struct page *page = pvec.pages[i];
2416
2417 /*
2418 * At this point we hold neither mapping->tree_lock nor
2419 * lock on the page itself: the page may be truncated or
2420 * invalidated (changing page->mapping to NULL), or even
2421 * swizzled back from swapper_space to tmpfs file
2422 * mapping
2423 */
2424 if (tree->ops && tree->ops->write_cache_pages_lock_hook)
2425 tree->ops->write_cache_pages_lock_hook(page);
2426 else
2427 lock_page(page);
2428
2429 if (unlikely(page->mapping != mapping)) {
2430 unlock_page(page);
2431 continue;
2432 }
2433
2434 if (!wbc->range_cyclic && page->index > end) {
2435 done = 1;
2436 unlock_page(page);
2437 continue;
2438 }
2439
2440 if (wbc->sync_mode != WB_SYNC_NONE) {
2441 if (PageWriteback(page))
2442 flush_fn(data);
2443 wait_on_page_writeback(page);
2444 }
2445
2446 if (PageWriteback(page) ||
2447 !clear_page_dirty_for_io(page)) {
2448 unlock_page(page);
2449 continue;
2450 }
2451
2452 ret = (*writepage)(page, wbc, data);
2453
2454 if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
2455 unlock_page(page);
2456 ret = 0;
2457 }
2458 if (ret)
2459 done = 1;
2460
2461 /*
2462 * the filesystem may choose to bump up nr_to_write.
2463 * We have to make sure to honor the new nr_to_write
2464 * at any time
2465 */
2466 nr_to_write_done = wbc->nr_to_write <= 0;
2467 }
2468 pagevec_release(&pvec);
2469 cond_resched();
2470 }
2471 if (!scanned && !done) {
2472 /*
2473 * We hit the last page and there is more work to be done: wrap
2474 * back to the start of the file
2475 */
2476 scanned = 1;
2477 index = 0;
2478 goto retry;
2479 }
2480 return ret;
2481}
2482
2483static void flush_epd_write_bio(struct extent_page_data *epd)
2484{
2485 if (epd->bio) {
2486 if (epd->sync_io)
2487 submit_one_bio(WRITE_SYNC, epd->bio, 0, 0);
2488 else
2489 submit_one_bio(WRITE, epd->bio, 0, 0);
2490 epd->bio = NULL;
2491 }
2492}
2493
2494static noinline void flush_write_bio(void *data)
2495{
2496 struct extent_page_data *epd = data;
2497 flush_epd_write_bio(epd);
2498}
2499
2500int extent_write_full_page(struct extent_io_tree *tree, struct page *page,
2501 get_extent_t *get_extent,
2502 struct writeback_control *wbc)
2503{
2504 int ret;
2505 struct extent_page_data epd = {
2506 .bio = NULL,
2507 .tree = tree,
2508 .get_extent = get_extent,
2509 .extent_locked = 0,
2510 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2511 };
2512
2513 ret = __extent_writepage(page, wbc, &epd);
2514
2515 flush_epd_write_bio(&epd);
2516 return ret;
2517}
2518
2519int extent_write_locked_range(struct extent_io_tree *tree, struct inode *inode,
2520 u64 start, u64 end, get_extent_t *get_extent,
2521 int mode)
2522{
2523 int ret = 0;
2524 struct address_space *mapping = inode->i_mapping;
2525 struct page *page;
2526 unsigned long nr_pages = (end - start + PAGE_CACHE_SIZE) >>
2527 PAGE_CACHE_SHIFT;
2528
2529 struct extent_page_data epd = {
2530 .bio = NULL,
2531 .tree = tree,
2532 .get_extent = get_extent,
2533 .extent_locked = 1,
2534 .sync_io = mode == WB_SYNC_ALL,
2535 };
2536 struct writeback_control wbc_writepages = {
2537 .sync_mode = mode,
2538 .nr_to_write = nr_pages * 2,
2539 .range_start = start,
2540 .range_end = end + 1,
2541 };
2542
2543 while (start <= end) {
2544 page = find_get_page(mapping, start >> PAGE_CACHE_SHIFT);
2545 if (clear_page_dirty_for_io(page))
2546 ret = __extent_writepage(page, &wbc_writepages, &epd);
2547 else {
2548 if (tree->ops && tree->ops->writepage_end_io_hook)
2549 tree->ops->writepage_end_io_hook(page, start,
2550 start + PAGE_CACHE_SIZE - 1,
2551 NULL, 1);
2552 unlock_page(page);
2553 }
2554 page_cache_release(page);
2555 start += PAGE_CACHE_SIZE;
2556 }
2557
2558 flush_epd_write_bio(&epd);
2559 return ret;
2560}
2561
2562int extent_writepages(struct extent_io_tree *tree,
2563 struct address_space *mapping,
2564 get_extent_t *get_extent,
2565 struct writeback_control *wbc)
2566{
2567 int ret = 0;
2568 struct extent_page_data epd = {
2569 .bio = NULL,
2570 .tree = tree,
2571 .get_extent = get_extent,
2572 .extent_locked = 0,
2573 .sync_io = wbc->sync_mode == WB_SYNC_ALL,
2574 };
2575
2576 ret = extent_write_cache_pages(tree, mapping, wbc,
2577 __extent_writepage, &epd,
2578 flush_write_bio);
2579 flush_epd_write_bio(&epd);
2580 return ret;
2581}
2582
2583int extent_readpages(struct extent_io_tree *tree,
2584 struct address_space *mapping,
2585 struct list_head *pages, unsigned nr_pages,
2586 get_extent_t get_extent)
2587{
2588 struct bio *bio = NULL;
2589 unsigned page_idx;
2590 unsigned long bio_flags = 0;
2591
2592 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
2593 struct page *page = list_entry(pages->prev, struct page, lru);
2594
2595 prefetchw(&page->flags);
2596 list_del(&page->lru);
2597 if (!add_to_page_cache_lru(page, mapping,
2598 page->index, GFP_NOFS)) {
2599 __extent_read_full_page(tree, page, get_extent,
2600 &bio, 0, &bio_flags);
2601 }
2602 page_cache_release(page);
2603 }
2604 BUG_ON(!list_empty(pages));
2605 if (bio)
2606 submit_one_bio(READ, bio, 0, bio_flags);
2607 return 0;
2608}
2609
2610/*
2611 * basic invalidatepage code, this waits on any locked or writeback
2612 * ranges corresponding to the page, and then deletes any extent state
2613 * records from the tree
2614 */
2615int extent_invalidatepage(struct extent_io_tree *tree,
2616 struct page *page, unsigned long offset)
2617{
2618 struct extent_state *cached_state = NULL;
2619 u64 start = ((u64)page->index << PAGE_CACHE_SHIFT);
2620 u64 end = start + PAGE_CACHE_SIZE - 1;
2621 size_t blocksize = page->mapping->host->i_sb->s_blocksize;
2622
2623 start += (offset + blocksize - 1) & ~(blocksize - 1);
2624 if (start > end)
2625 return 0;
2626
2627 lock_extent_bits(tree, start, end, 0, &cached_state, GFP_NOFS);
2628 wait_on_page_writeback(page);
2629 clear_extent_bit(tree, start, end,
2630 EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
2631 EXTENT_DO_ACCOUNTING,
2632 1, 1, &cached_state, GFP_NOFS);
2633 return 0;
2634}
2635
2636/*
2637 * a helper for releasepage, this tests for areas of the page that
2638 * are locked or under IO and drops the related state bits if it is safe
2639 * to drop the page.
2640 */
2641int try_release_extent_state(struct extent_map_tree *map,
2642 struct extent_io_tree *tree, struct page *page,
2643 gfp_t mask)
2644{
2645 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2646 u64 end = start + PAGE_CACHE_SIZE - 1;
2647 int ret = 1;
2648
2649 if (test_range_bit(tree, start, end,
2650 EXTENT_IOBITS, 0, NULL))
2651 ret = 0;
2652 else {
2653 if ((mask & GFP_NOFS) == GFP_NOFS)
2654 mask = GFP_NOFS;
2655 /*
2656 * at this point we can safely clear everything except the
2657 * locked bit and the nodatasum bit
2658 */
2659 ret = clear_extent_bit(tree, start, end,
2660 ~(EXTENT_LOCKED | EXTENT_NODATASUM),
2661 0, 0, NULL, mask);
2662
2663 /* if clear_extent_bit failed for enomem reasons,
2664 * we can't allow the release to continue.
2665 */
2666 if (ret < 0)
2667 ret = 0;
2668 else
2669 ret = 1;
2670 }
2671 return ret;
2672}
2673
2674/*
2675 * a helper for releasepage. As long as there are no locked extents
2676 * in the range corresponding to the page, both state records and extent
2677 * map records are removed
2678 */
2679int try_release_extent_mapping(struct extent_map_tree *map,
2680 struct extent_io_tree *tree, struct page *page,
2681 gfp_t mask)
2682{
2683 struct extent_map *em;
2684 u64 start = (u64)page->index << PAGE_CACHE_SHIFT;
2685 u64 end = start + PAGE_CACHE_SIZE - 1;
2686
2687 if ((mask & __GFP_WAIT) &&
2688 page->mapping->host->i_size > 16 * 1024 * 1024) {
2689 u64 len;
2690 while (start <= end) {
2691 len = end - start + 1;
2692 write_lock(&map->lock);
2693 em = lookup_extent_mapping(map, start, len);
2694 if (IS_ERR_OR_NULL(em)) {
2695 write_unlock(&map->lock);
2696 break;
2697 }
2698 if (test_bit(EXTENT_FLAG_PINNED, &em->flags) ||
2699 em->start != start) {
2700 write_unlock(&map->lock);
2701 free_extent_map(em);
2702 break;
2703 }
2704 if (!test_range_bit(tree, em->start,
2705 extent_map_end(em) - 1,
2706 EXTENT_LOCKED | EXTENT_WRITEBACK,
2707 0, NULL)) {
2708 remove_extent_mapping(map, em);
2709 /* once for the rb tree */
2710 free_extent_map(em);
2711 }
2712 start = extent_map_end(em);
2713 write_unlock(&map->lock);
2714
2715 /* once for us */
2716 free_extent_map(em);
2717 }
2718 }
2719 return try_release_extent_state(map, tree, page, mask);
2720}
2721
2722/*
2723 * helper function for fiemap, which doesn't want to see any holes.
2724 * This maps until we find something past 'last'
2725 */
2726static struct extent_map *get_extent_skip_holes(struct inode *inode,
2727 u64 offset,
2728 u64 last,
2729 get_extent_t *get_extent)
2730{
2731 u64 sectorsize = BTRFS_I(inode)->root->sectorsize;
2732 struct extent_map *em;
2733 u64 len;
2734
2735 if (offset >= last)
2736 return NULL;
2737
2738 while(1) {
2739 len = last - offset;
2740 if (len == 0)
2741 break;
2742 len = (len + sectorsize - 1) & ~(sectorsize - 1);
2743 em = get_extent(inode, NULL, 0, offset, len, 0);
2744 if (IS_ERR_OR_NULL(em))
2745 return em;
2746
2747 /* if this isn't a hole return it */
2748 if (!test_bit(EXTENT_FLAG_VACANCY, &em->flags) &&
2749 em->block_start != EXTENT_MAP_HOLE) {
2750 return em;
2751 }
2752
2753 /* this is a hole, advance to the next extent */
2754 offset = extent_map_end(em);
2755 free_extent_map(em);
2756 if (offset >= last)
2757 break;
2758 }
2759 return NULL;
2760}
2761
2762int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
2763 __u64 start, __u64 len, get_extent_t *get_extent)
2764{
2765 int ret = 0;
2766 u64 off = start;
2767 u64 max = start + len;
2768 u32 flags = 0;
2769 u32 found_type;
2770 u64 last;
2771 u64 last_for_get_extent = 0;
2772 u64 disko = 0;
2773 u64 isize = i_size_read(inode);
2774 struct btrfs_key found_key;
2775 struct extent_map *em = NULL;
2776 struct extent_state *cached_state = NULL;
2777 struct btrfs_path *path;
2778 struct btrfs_file_extent_item *item;
2779 int end = 0;
2780 u64 em_start = 0;
2781 u64 em_len = 0;
2782 u64 em_end = 0;
2783 unsigned long emflags;
2784
2785 if (len == 0)
2786 return -EINVAL;
2787
2788 path = btrfs_alloc_path();
2789 if (!path)
2790 return -ENOMEM;
2791 path->leave_spinning = 1;
2792
2793 /*
2794 * lookup the last file extent. We're not using i_size here
2795 * because there might be preallocation past i_size
2796 */
2797 ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
2798 path, btrfs_ino(inode), -1, 0);
2799 if (ret < 0) {
2800 btrfs_free_path(path);
2801 return ret;
2802 }
2803 WARN_ON(!ret);
2804 path->slots[0]--;
2805 item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2806 struct btrfs_file_extent_item);
2807 btrfs_item_key_to_cpu(path->nodes[0], &found_key, path->slots[0]);
2808 found_type = btrfs_key_type(&found_key);
2809
2810 /* No extents, but there might be delalloc bits */
2811 if (found_key.objectid != btrfs_ino(inode) ||
2812 found_type != BTRFS_EXTENT_DATA_KEY) {
2813 /* have to trust i_size as the end */
2814 last = (u64)-1;
2815 last_for_get_extent = isize;
2816 } else {
2817 /*
2818 * remember the start of the last extent. There are a
2819 * bunch of different factors that go into the length of the
2820 * extent, so its much less complex to remember where it started
2821 */
2822 last = found_key.offset;
2823 last_for_get_extent = last + 1;
2824 }
2825 btrfs_free_path(path);
2826
2827 /*
2828 * we might have some extents allocated but more delalloc past those
2829 * extents. so, we trust isize unless the start of the last extent is
2830 * beyond isize
2831 */
2832 if (last < isize) {
2833 last = (u64)-1;
2834 last_for_get_extent = isize;
2835 }
2836
2837 lock_extent_bits(&BTRFS_I(inode)->io_tree, start, start + len, 0,
2838 &cached_state, GFP_NOFS);
2839
2840 em = get_extent_skip_holes(inode, off, last_for_get_extent,
2841 get_extent);
2842 if (!em)
2843 goto out;
2844 if (IS_ERR(em)) {
2845 ret = PTR_ERR(em);
2846 goto out;
2847 }
2848
2849 while (!end) {
2850 u64 offset_in_extent;
2851
2852 /* break if the extent we found is outside the range */
2853 if (em->start >= max || extent_map_end(em) < off)
2854 break;
2855
2856 /*
2857 * get_extent may return an extent that starts before our
2858 * requested range. We have to make sure the ranges
2859 * we return to fiemap always move forward and don't
2860 * overlap, so adjust the offsets here
2861 */
2862 em_start = max(em->start, off);
2863
2864 /*
2865 * record the offset from the start of the extent
2866 * for adjusting the disk offset below
2867 */
2868 offset_in_extent = em_start - em->start;
2869 em_end = extent_map_end(em);
2870 em_len = em_end - em_start;
2871 emflags = em->flags;
2872 disko = 0;
2873 flags = 0;
2874
2875 /*
2876 * bump off for our next call to get_extent
2877 */
2878 off = extent_map_end(em);
2879 if (off >= max)
2880 end = 1;
2881
2882 if (em->block_start == EXTENT_MAP_LAST_BYTE) {
2883 end = 1;
2884 flags |= FIEMAP_EXTENT_LAST;
2885 } else if (em->block_start == EXTENT_MAP_INLINE) {
2886 flags |= (FIEMAP_EXTENT_DATA_INLINE |
2887 FIEMAP_EXTENT_NOT_ALIGNED);
2888 } else if (em->block_start == EXTENT_MAP_DELALLOC) {
2889 flags |= (FIEMAP_EXTENT_DELALLOC |
2890 FIEMAP_EXTENT_UNKNOWN);
2891 } else {
2892 disko = em->block_start + offset_in_extent;
2893 }
2894 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
2895 flags |= FIEMAP_EXTENT_ENCODED;
2896
2897 free_extent_map(em);
2898 em = NULL;
2899 if ((em_start >= last) || em_len == (u64)-1 ||
2900 (last == (u64)-1 && isize <= em_end)) {
2901 flags |= FIEMAP_EXTENT_LAST;
2902 end = 1;
2903 }
2904
2905 /* now scan forward to see if this is really the last extent. */
2906 em = get_extent_skip_holes(inode, off, last_for_get_extent,
2907 get_extent);
2908 if (IS_ERR(em)) {
2909 ret = PTR_ERR(em);
2910 goto out;
2911 }
2912 if (!em) {
2913 flags |= FIEMAP_EXTENT_LAST;
2914 end = 1;
2915 }
2916 ret = fiemap_fill_next_extent(fieinfo, em_start, disko,
2917 em_len, flags);
2918 if (ret)
2919 goto out_free;
2920 }
2921out_free:
2922 free_extent_map(em);
2923out:
2924 unlock_extent_cached(&BTRFS_I(inode)->io_tree, start, start + len,
2925 &cached_state, GFP_NOFS);
2926 return ret;
2927}
2928
2929static inline struct page *extent_buffer_page(struct extent_buffer *eb,
2930 unsigned long i)
2931{
2932 struct page *p;
2933 struct address_space *mapping;
2934
2935 if (i == 0)
2936 return eb->first_page;
2937 i += eb->start >> PAGE_CACHE_SHIFT;
2938 mapping = eb->first_page->mapping;
2939 if (!mapping)
2940 return NULL;
2941
2942 /*
2943 * extent_buffer_page is only called after pinning the page
2944 * by increasing the reference count. So we know the page must
2945 * be in the radix tree.
2946 */
2947 rcu_read_lock();
2948 p = radix_tree_lookup(&mapping->page_tree, i);
2949 rcu_read_unlock();
2950
2951 return p;
2952}
2953
2954static inline unsigned long num_extent_pages(u64 start, u64 len)
2955{
2956 return ((start + len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
2957 (start >> PAGE_CACHE_SHIFT);
2958}
2959
2960static struct extent_buffer *__alloc_extent_buffer(struct extent_io_tree *tree,
2961 u64 start,
2962 unsigned long len,
2963 gfp_t mask)
2964{
2965 struct extent_buffer *eb = NULL;
2966#if LEAK_DEBUG
2967 unsigned long flags;
2968#endif
2969
2970 eb = kmem_cache_zalloc(extent_buffer_cache, mask);
2971 if (eb == NULL)
2972 return NULL;
2973 eb->start = start;
2974 eb->len = len;
2975 rwlock_init(&eb->lock);
2976 atomic_set(&eb->write_locks, 0);
2977 atomic_set(&eb->read_locks, 0);
2978 atomic_set(&eb->blocking_readers, 0);
2979 atomic_set(&eb->blocking_writers, 0);
2980 atomic_set(&eb->spinning_readers, 0);
2981 atomic_set(&eb->spinning_writers, 0);
2982 init_waitqueue_head(&eb->write_lock_wq);
2983 init_waitqueue_head(&eb->read_lock_wq);
2984
2985#if LEAK_DEBUG
2986 spin_lock_irqsave(&leak_lock, flags);
2987 list_add(&eb->leak_list, &buffers);
2988 spin_unlock_irqrestore(&leak_lock, flags);
2989#endif
2990 atomic_set(&eb->refs, 1);
2991
2992 return eb;
2993}
2994
2995static void __free_extent_buffer(struct extent_buffer *eb)
2996{
2997#if LEAK_DEBUG
2998 unsigned long flags;
2999 spin_lock_irqsave(&leak_lock, flags);
3000 list_del(&eb->leak_list);
3001 spin_unlock_irqrestore(&leak_lock, flags);
3002#endif
3003 kmem_cache_free(extent_buffer_cache, eb);
3004}
3005
3006/*
3007 * Helper for releasing extent buffer page.
3008 */
3009static void btrfs_release_extent_buffer_page(struct extent_buffer *eb,
3010 unsigned long start_idx)
3011{
3012 unsigned long index;
3013 struct page *page;
3014
3015 if (!eb->first_page)
3016 return;
3017
3018 index = num_extent_pages(eb->start, eb->len);
3019 if (start_idx >= index)
3020 return;
3021
3022 do {
3023 index--;
3024 page = extent_buffer_page(eb, index);
3025 if (page)
3026 page_cache_release(page);
3027 } while (index != start_idx);
3028}
3029
3030/*
3031 * Helper for releasing the extent buffer.
3032 */
3033static inline void btrfs_release_extent_buffer(struct extent_buffer *eb)
3034{
3035 btrfs_release_extent_buffer_page(eb, 0);
3036 __free_extent_buffer(eb);
3037}
3038
3039struct extent_buffer *alloc_extent_buffer(struct extent_io_tree *tree,
3040 u64 start, unsigned long len,
3041 struct page *page0)
3042{
3043 unsigned long num_pages = num_extent_pages(start, len);
3044 unsigned long i;
3045 unsigned long index = start >> PAGE_CACHE_SHIFT;
3046 struct extent_buffer *eb;
3047 struct extent_buffer *exists = NULL;
3048 struct page *p;
3049 struct address_space *mapping = tree->mapping;
3050 int uptodate = 1;
3051 int ret;
3052
3053 rcu_read_lock();
3054 eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3055 if (eb && atomic_inc_not_zero(&eb->refs)) {
3056 rcu_read_unlock();
3057 mark_page_accessed(eb->first_page);
3058 return eb;
3059 }
3060 rcu_read_unlock();
3061
3062 eb = __alloc_extent_buffer(tree, start, len, GFP_NOFS);
3063 if (!eb)
3064 return NULL;
3065
3066 if (page0) {
3067 eb->first_page = page0;
3068 i = 1;
3069 index++;
3070 page_cache_get(page0);
3071 mark_page_accessed(page0);
3072 set_page_extent_mapped(page0);
3073 set_page_extent_head(page0, len);
3074 uptodate = PageUptodate(page0);
3075 } else {
3076 i = 0;
3077 }
3078 for (; i < num_pages; i++, index++) {
3079 p = find_or_create_page(mapping, index, GFP_NOFS);
3080 if (!p) {
3081 WARN_ON(1);
3082 goto free_eb;
3083 }
3084 set_page_extent_mapped(p);
3085 mark_page_accessed(p);
3086 if (i == 0) {
3087 eb->first_page = p;
3088 set_page_extent_head(p, len);
3089 } else {
3090 set_page_private(p, EXTENT_PAGE_PRIVATE);
3091 }
3092 if (!PageUptodate(p))
3093 uptodate = 0;
3094
3095 /*
3096 * see below about how we avoid a nasty race with release page
3097 * and why we unlock later
3098 */
3099 if (i != 0)
3100 unlock_page(p);
3101 }
3102 if (uptodate)
3103 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3104
3105 ret = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
3106 if (ret)
3107 goto free_eb;
3108
3109 spin_lock(&tree->buffer_lock);
3110 ret = radix_tree_insert(&tree->buffer, start >> PAGE_CACHE_SHIFT, eb);
3111 if (ret == -EEXIST) {
3112 exists = radix_tree_lookup(&tree->buffer,
3113 start >> PAGE_CACHE_SHIFT);
3114 /* add one reference for the caller */
3115 atomic_inc(&exists->refs);
3116 spin_unlock(&tree->buffer_lock);
3117 radix_tree_preload_end();
3118 goto free_eb;
3119 }
3120 /* add one reference for the tree */
3121 atomic_inc(&eb->refs);
3122 spin_unlock(&tree->buffer_lock);
3123 radix_tree_preload_end();
3124
3125 /*
3126 * there is a race where release page may have
3127 * tried to find this extent buffer in the radix
3128 * but failed. It will tell the VM it is safe to
3129 * reclaim the, and it will clear the page private bit.
3130 * We must make sure to set the page private bit properly
3131 * after the extent buffer is in the radix tree so
3132 * it doesn't get lost
3133 */
3134 set_page_extent_mapped(eb->first_page);
3135 set_page_extent_head(eb->first_page, eb->len);
3136 if (!page0)
3137 unlock_page(eb->first_page);
3138 return eb;
3139
3140free_eb:
3141 if (eb->first_page && !page0)
3142 unlock_page(eb->first_page);
3143
3144 if (!atomic_dec_and_test(&eb->refs))
3145 return exists;
3146 btrfs_release_extent_buffer(eb);
3147 return exists;
3148}
3149
3150struct extent_buffer *find_extent_buffer(struct extent_io_tree *tree,
3151 u64 start, unsigned long len)
3152{
3153 struct extent_buffer *eb;
3154
3155 rcu_read_lock();
3156 eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3157 if (eb && atomic_inc_not_zero(&eb->refs)) {
3158 rcu_read_unlock();
3159 mark_page_accessed(eb->first_page);
3160 return eb;
3161 }
3162 rcu_read_unlock();
3163
3164 return NULL;
3165}
3166
3167void free_extent_buffer(struct extent_buffer *eb)
3168{
3169 if (!eb)
3170 return;
3171
3172 if (!atomic_dec_and_test(&eb->refs))
3173 return;
3174
3175 WARN_ON(1);
3176}
3177
3178int clear_extent_buffer_dirty(struct extent_io_tree *tree,
3179 struct extent_buffer *eb)
3180{
3181 unsigned long i;
3182 unsigned long num_pages;
3183 struct page *page;
3184
3185 num_pages = num_extent_pages(eb->start, eb->len);
3186
3187 for (i = 0; i < num_pages; i++) {
3188 page = extent_buffer_page(eb, i);
3189 if (!PageDirty(page))
3190 continue;
3191
3192 lock_page(page);
3193 WARN_ON(!PagePrivate(page));
3194
3195 set_page_extent_mapped(page);
3196 if (i == 0)
3197 set_page_extent_head(page, eb->len);
3198
3199 clear_page_dirty_for_io(page);
3200 spin_lock_irq(&page->mapping->tree_lock);
3201 if (!PageDirty(page)) {
3202 radix_tree_tag_clear(&page->mapping->page_tree,
3203 page_index(page),
3204 PAGECACHE_TAG_DIRTY);
3205 }
3206 spin_unlock_irq(&page->mapping->tree_lock);
3207 unlock_page(page);
3208 }
3209 return 0;
3210}
3211
3212int set_extent_buffer_dirty(struct extent_io_tree *tree,
3213 struct extent_buffer *eb)
3214{
3215 unsigned long i;
3216 unsigned long num_pages;
3217 int was_dirty = 0;
3218
3219 was_dirty = test_and_set_bit(EXTENT_BUFFER_DIRTY, &eb->bflags);
3220 num_pages = num_extent_pages(eb->start, eb->len);
3221 for (i = 0; i < num_pages; i++)
3222 __set_page_dirty_nobuffers(extent_buffer_page(eb, i));
3223 return was_dirty;
3224}
3225
3226static int __eb_straddles_pages(u64 start, u64 len)
3227{
3228 if (len < PAGE_CACHE_SIZE)
3229 return 1;
3230 if (start & (PAGE_CACHE_SIZE - 1))
3231 return 1;
3232 if ((start + len) & (PAGE_CACHE_SIZE - 1))
3233 return 1;
3234 return 0;
3235}
3236
3237static int eb_straddles_pages(struct extent_buffer *eb)
3238{
3239 return __eb_straddles_pages(eb->start, eb->len);
3240}
3241
3242int clear_extent_buffer_uptodate(struct extent_io_tree *tree,
3243 struct extent_buffer *eb,
3244 struct extent_state **cached_state)
3245{
3246 unsigned long i;
3247 struct page *page;
3248 unsigned long num_pages;
3249
3250 num_pages = num_extent_pages(eb->start, eb->len);
3251 clear_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3252
3253 if (eb_straddles_pages(eb)) {
3254 clear_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3255 cached_state, GFP_NOFS);
3256 }
3257 for (i = 0; i < num_pages; i++) {
3258 page = extent_buffer_page(eb, i);
3259 if (page)
3260 ClearPageUptodate(page);
3261 }
3262 return 0;
3263}
3264
3265int set_extent_buffer_uptodate(struct extent_io_tree *tree,
3266 struct extent_buffer *eb)
3267{
3268 unsigned long i;
3269 struct page *page;
3270 unsigned long num_pages;
3271
3272 num_pages = num_extent_pages(eb->start, eb->len);
3273
3274 if (eb_straddles_pages(eb)) {
3275 set_extent_uptodate(tree, eb->start, eb->start + eb->len - 1,
3276 NULL, GFP_NOFS);
3277 }
3278 for (i = 0; i < num_pages; i++) {
3279 page = extent_buffer_page(eb, i);
3280 if ((i == 0 && (eb->start & (PAGE_CACHE_SIZE - 1))) ||
3281 ((i == num_pages - 1) &&
3282 ((eb->start + eb->len) & (PAGE_CACHE_SIZE - 1)))) {
3283 check_page_uptodate(tree, page);
3284 continue;
3285 }
3286 SetPageUptodate(page);
3287 }
3288 return 0;
3289}
3290
3291int extent_range_uptodate(struct extent_io_tree *tree,
3292 u64 start, u64 end)
3293{
3294 struct page *page;
3295 int ret;
3296 int pg_uptodate = 1;
3297 int uptodate;
3298 unsigned long index;
3299
3300 if (__eb_straddles_pages(start, end - start + 1)) {
3301 ret = test_range_bit(tree, start, end,
3302 EXTENT_UPTODATE, 1, NULL);
3303 if (ret)
3304 return 1;
3305 }
3306 while (start <= end) {
3307 index = start >> PAGE_CACHE_SHIFT;
3308 page = find_get_page(tree->mapping, index);
3309 uptodate = PageUptodate(page);
3310 page_cache_release(page);
3311 if (!uptodate) {
3312 pg_uptodate = 0;
3313 break;
3314 }
3315 start += PAGE_CACHE_SIZE;
3316 }
3317 return pg_uptodate;
3318}
3319
3320int extent_buffer_uptodate(struct extent_io_tree *tree,
3321 struct extent_buffer *eb,
3322 struct extent_state *cached_state)
3323{
3324 int ret = 0;
3325 unsigned long num_pages;
3326 unsigned long i;
3327 struct page *page;
3328 int pg_uptodate = 1;
3329
3330 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3331 return 1;
3332
3333 if (eb_straddles_pages(eb)) {
3334 ret = test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3335 EXTENT_UPTODATE, 1, cached_state);
3336 if (ret)
3337 return ret;
3338 }
3339
3340 num_pages = num_extent_pages(eb->start, eb->len);
3341 for (i = 0; i < num_pages; i++) {
3342 page = extent_buffer_page(eb, i);
3343 if (!PageUptodate(page)) {
3344 pg_uptodate = 0;
3345 break;
3346 }
3347 }
3348 return pg_uptodate;
3349}
3350
3351int read_extent_buffer_pages(struct extent_io_tree *tree,
3352 struct extent_buffer *eb,
3353 u64 start, int wait,
3354 get_extent_t *get_extent, int mirror_num)
3355{
3356 unsigned long i;
3357 unsigned long start_i;
3358 struct page *page;
3359 int err;
3360 int ret = 0;
3361 int locked_pages = 0;
3362 int all_uptodate = 1;
3363 int inc_all_pages = 0;
3364 unsigned long num_pages;
3365 struct bio *bio = NULL;
3366 unsigned long bio_flags = 0;
3367
3368 if (test_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags))
3369 return 0;
3370
3371 if (eb_straddles_pages(eb)) {
3372 if (test_range_bit(tree, eb->start, eb->start + eb->len - 1,
3373 EXTENT_UPTODATE, 1, NULL)) {
3374 return 0;
3375 }
3376 }
3377
3378 if (start) {
3379 WARN_ON(start < eb->start);
3380 start_i = (start >> PAGE_CACHE_SHIFT) -
3381 (eb->start >> PAGE_CACHE_SHIFT);
3382 } else {
3383 start_i = 0;
3384 }
3385
3386 num_pages = num_extent_pages(eb->start, eb->len);
3387 for (i = start_i; i < num_pages; i++) {
3388 page = extent_buffer_page(eb, i);
3389 if (!wait) {
3390 if (!trylock_page(page))
3391 goto unlock_exit;
3392 } else {
3393 lock_page(page);
3394 }
3395 locked_pages++;
3396 if (!PageUptodate(page))
3397 all_uptodate = 0;
3398 }
3399 if (all_uptodate) {
3400 if (start_i == 0)
3401 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3402 goto unlock_exit;
3403 }
3404
3405 for (i = start_i; i < num_pages; i++) {
3406 page = extent_buffer_page(eb, i);
3407
3408 WARN_ON(!PagePrivate(page));
3409
3410 set_page_extent_mapped(page);
3411 if (i == 0)
3412 set_page_extent_head(page, eb->len);
3413
3414 if (inc_all_pages)
3415 page_cache_get(page);
3416 if (!PageUptodate(page)) {
3417 if (start_i == 0)
3418 inc_all_pages = 1;
3419 ClearPageError(page);
3420 err = __extent_read_full_page(tree, page,
3421 get_extent, &bio,
3422 mirror_num, &bio_flags);
3423 if (err)
3424 ret = err;
3425 } else {
3426 unlock_page(page);
3427 }
3428 }
3429
3430 if (bio)
3431 submit_one_bio(READ, bio, mirror_num, bio_flags);
3432
3433 if (ret || !wait)
3434 return ret;
3435
3436 for (i = start_i; i < num_pages; i++) {
3437 page = extent_buffer_page(eb, i);
3438 wait_on_page_locked(page);
3439 if (!PageUptodate(page))
3440 ret = -EIO;
3441 }
3442
3443 if (!ret)
3444 set_bit(EXTENT_BUFFER_UPTODATE, &eb->bflags);
3445 return ret;
3446
3447unlock_exit:
3448 i = start_i;
3449 while (locked_pages > 0) {
3450 page = extent_buffer_page(eb, i);
3451 i++;
3452 unlock_page(page);
3453 locked_pages--;
3454 }
3455 return ret;
3456}
3457
3458void read_extent_buffer(struct extent_buffer *eb, void *dstv,
3459 unsigned long start,
3460 unsigned long len)
3461{
3462 size_t cur;
3463 size_t offset;
3464 struct page *page;
3465 char *kaddr;
3466 char *dst = (char *)dstv;
3467 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3468 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3469
3470 WARN_ON(start > eb->len);
3471 WARN_ON(start + len > eb->start + eb->len);
3472
3473 offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3474
3475 while (len > 0) {
3476 page = extent_buffer_page(eb, i);
3477
3478 cur = min(len, (PAGE_CACHE_SIZE - offset));
3479 kaddr = page_address(page);
3480 memcpy(dst, kaddr + offset, cur);
3481
3482 dst += cur;
3483 len -= cur;
3484 offset = 0;
3485 i++;
3486 }
3487}
3488
3489int map_private_extent_buffer(struct extent_buffer *eb, unsigned long start,
3490 unsigned long min_len, char **map,
3491 unsigned long *map_start,
3492 unsigned long *map_len)
3493{
3494 size_t offset = start & (PAGE_CACHE_SIZE - 1);
3495 char *kaddr;
3496 struct page *p;
3497 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3498 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3499 unsigned long end_i = (start_offset + start + min_len - 1) >>
3500 PAGE_CACHE_SHIFT;
3501
3502 if (i != end_i)
3503 return -EINVAL;
3504
3505 if (i == 0) {
3506 offset = start_offset;
3507 *map_start = 0;
3508 } else {
3509 offset = 0;
3510 *map_start = ((u64)i << PAGE_CACHE_SHIFT) - start_offset;
3511 }
3512
3513 if (start + min_len > eb->len) {
3514 printk(KERN_ERR "btrfs bad mapping eb start %llu len %lu, "
3515 "wanted %lu %lu\n", (unsigned long long)eb->start,
3516 eb->len, start, min_len);
3517 WARN_ON(1);
3518 return -EINVAL;
3519 }
3520
3521 p = extent_buffer_page(eb, i);
3522 kaddr = page_address(p);
3523 *map = kaddr + offset;
3524 *map_len = PAGE_CACHE_SIZE - offset;
3525 return 0;
3526}
3527
3528int memcmp_extent_buffer(struct extent_buffer *eb, const void *ptrv,
3529 unsigned long start,
3530 unsigned long len)
3531{
3532 size_t cur;
3533 size_t offset;
3534 struct page *page;
3535 char *kaddr;
3536 char *ptr = (char *)ptrv;
3537 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3538 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3539 int ret = 0;
3540
3541 WARN_ON(start > eb->len);
3542 WARN_ON(start + len > eb->start + eb->len);
3543
3544 offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3545
3546 while (len > 0) {
3547 page = extent_buffer_page(eb, i);
3548
3549 cur = min(len, (PAGE_CACHE_SIZE - offset));
3550
3551 kaddr = page_address(page);
3552 ret = memcmp(ptr, kaddr + offset, cur);
3553 if (ret)
3554 break;
3555
3556 ptr += cur;
3557 len -= cur;
3558 offset = 0;
3559 i++;
3560 }
3561 return ret;
3562}
3563
3564void write_extent_buffer(struct extent_buffer *eb, const void *srcv,
3565 unsigned long start, unsigned long len)
3566{
3567 size_t cur;
3568 size_t offset;
3569 struct page *page;
3570 char *kaddr;
3571 char *src = (char *)srcv;
3572 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3573 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3574
3575 WARN_ON(start > eb->len);
3576 WARN_ON(start + len > eb->start + eb->len);
3577
3578 offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3579
3580 while (len > 0) {
3581 page = extent_buffer_page(eb, i);
3582 WARN_ON(!PageUptodate(page));
3583
3584 cur = min(len, PAGE_CACHE_SIZE - offset);
3585 kaddr = page_address(page);
3586 memcpy(kaddr + offset, src, cur);
3587
3588 src += cur;
3589 len -= cur;
3590 offset = 0;
3591 i++;
3592 }
3593}
3594
3595void memset_extent_buffer(struct extent_buffer *eb, char c,
3596 unsigned long start, unsigned long len)
3597{
3598 size_t cur;
3599 size_t offset;
3600 struct page *page;
3601 char *kaddr;
3602 size_t start_offset = eb->start & ((u64)PAGE_CACHE_SIZE - 1);
3603 unsigned long i = (start_offset + start) >> PAGE_CACHE_SHIFT;
3604
3605 WARN_ON(start > eb->len);
3606 WARN_ON(start + len > eb->start + eb->len);
3607
3608 offset = (start_offset + start) & ((unsigned long)PAGE_CACHE_SIZE - 1);
3609
3610 while (len > 0) {
3611 page = extent_buffer_page(eb, i);
3612 WARN_ON(!PageUptodate(page));
3613
3614 cur = min(len, PAGE_CACHE_SIZE - offset);
3615 kaddr = page_address(page);
3616 memset(kaddr + offset, c, cur);
3617
3618 len -= cur;
3619 offset = 0;
3620 i++;
3621 }
3622}
3623
3624void copy_extent_buffer(struct extent_buffer *dst, struct extent_buffer *src,
3625 unsigned long dst_offset, unsigned long src_offset,
3626 unsigned long len)
3627{
3628 u64 dst_len = dst->len;
3629 size_t cur;
3630 size_t offset;
3631 struct page *page;
3632 char *kaddr;
3633 size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3634 unsigned long i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3635
3636 WARN_ON(src->len != dst_len);
3637
3638 offset = (start_offset + dst_offset) &
3639 ((unsigned long)PAGE_CACHE_SIZE - 1);
3640
3641 while (len > 0) {
3642 page = extent_buffer_page(dst, i);
3643 WARN_ON(!PageUptodate(page));
3644
3645 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE - offset));
3646
3647 kaddr = page_address(page);
3648 read_extent_buffer(src, kaddr + offset, src_offset, cur);
3649
3650 src_offset += cur;
3651 len -= cur;
3652 offset = 0;
3653 i++;
3654 }
3655}
3656
3657static void move_pages(struct page *dst_page, struct page *src_page,
3658 unsigned long dst_off, unsigned long src_off,
3659 unsigned long len)
3660{
3661 char *dst_kaddr = page_address(dst_page);
3662 if (dst_page == src_page) {
3663 memmove(dst_kaddr + dst_off, dst_kaddr + src_off, len);
3664 } else {
3665 char *src_kaddr = page_address(src_page);
3666 char *p = dst_kaddr + dst_off + len;
3667 char *s = src_kaddr + src_off + len;
3668
3669 while (len--)
3670 *--p = *--s;
3671 }
3672}
3673
3674static inline bool areas_overlap(unsigned long src, unsigned long dst, unsigned long len)
3675{
3676 unsigned long distance = (src > dst) ? src - dst : dst - src;
3677 return distance < len;
3678}
3679
3680static void copy_pages(struct page *dst_page, struct page *src_page,
3681 unsigned long dst_off, unsigned long src_off,
3682 unsigned long len)
3683{
3684 char *dst_kaddr = page_address(dst_page);
3685 char *src_kaddr;
3686
3687 if (dst_page != src_page) {
3688 src_kaddr = page_address(src_page);
3689 } else {
3690 src_kaddr = dst_kaddr;
3691 BUG_ON(areas_overlap(src_off, dst_off, len));
3692 }
3693
3694 memcpy(dst_kaddr + dst_off, src_kaddr + src_off, len);
3695}
3696
3697void memcpy_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3698 unsigned long src_offset, unsigned long len)
3699{
3700 size_t cur;
3701 size_t dst_off_in_page;
3702 size_t src_off_in_page;
3703 size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3704 unsigned long dst_i;
3705 unsigned long src_i;
3706
3707 if (src_offset + len > dst->len) {
3708 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3709 "len %lu dst len %lu\n", src_offset, len, dst->len);
3710 BUG_ON(1);
3711 }
3712 if (dst_offset + len > dst->len) {
3713 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3714 "len %lu dst len %lu\n", dst_offset, len, dst->len);
3715 BUG_ON(1);
3716 }
3717
3718 while (len > 0) {
3719 dst_off_in_page = (start_offset + dst_offset) &
3720 ((unsigned long)PAGE_CACHE_SIZE - 1);
3721 src_off_in_page = (start_offset + src_offset) &
3722 ((unsigned long)PAGE_CACHE_SIZE - 1);
3723
3724 dst_i = (start_offset + dst_offset) >> PAGE_CACHE_SHIFT;
3725 src_i = (start_offset + src_offset) >> PAGE_CACHE_SHIFT;
3726
3727 cur = min(len, (unsigned long)(PAGE_CACHE_SIZE -
3728 src_off_in_page));
3729 cur = min_t(unsigned long, cur,
3730 (unsigned long)(PAGE_CACHE_SIZE - dst_off_in_page));
3731
3732 copy_pages(extent_buffer_page(dst, dst_i),
3733 extent_buffer_page(dst, src_i),
3734 dst_off_in_page, src_off_in_page, cur);
3735
3736 src_offset += cur;
3737 dst_offset += cur;
3738 len -= cur;
3739 }
3740}
3741
3742void memmove_extent_buffer(struct extent_buffer *dst, unsigned long dst_offset,
3743 unsigned long src_offset, unsigned long len)
3744{
3745 size_t cur;
3746 size_t dst_off_in_page;
3747 size_t src_off_in_page;
3748 unsigned long dst_end = dst_offset + len - 1;
3749 unsigned long src_end = src_offset + len - 1;
3750 size_t start_offset = dst->start & ((u64)PAGE_CACHE_SIZE - 1);
3751 unsigned long dst_i;
3752 unsigned long src_i;
3753
3754 if (src_offset + len > dst->len) {
3755 printk(KERN_ERR "btrfs memmove bogus src_offset %lu move "
3756 "len %lu len %lu\n", src_offset, len, dst->len);
3757 BUG_ON(1);
3758 }
3759 if (dst_offset + len > dst->len) {
3760 printk(KERN_ERR "btrfs memmove bogus dst_offset %lu move "
3761 "len %lu len %lu\n", dst_offset, len, dst->len);
3762 BUG_ON(1);
3763 }
3764 if (!areas_overlap(src_offset, dst_offset, len)) {
3765 memcpy_extent_buffer(dst, dst_offset, src_offset, len);
3766 return;
3767 }
3768 while (len > 0) {
3769 dst_i = (start_offset + dst_end) >> PAGE_CACHE_SHIFT;
3770 src_i = (start_offset + src_end) >> PAGE_CACHE_SHIFT;
3771
3772 dst_off_in_page = (start_offset + dst_end) &
3773 ((unsigned long)PAGE_CACHE_SIZE - 1);
3774 src_off_in_page = (start_offset + src_end) &
3775 ((unsigned long)PAGE_CACHE_SIZE - 1);
3776
3777 cur = min_t(unsigned long, len, src_off_in_page + 1);
3778 cur = min(cur, dst_off_in_page + 1);
3779 move_pages(extent_buffer_page(dst, dst_i),
3780 extent_buffer_page(dst, src_i),
3781 dst_off_in_page - cur + 1,
3782 src_off_in_page - cur + 1, cur);
3783
3784 dst_end -= cur;
3785 src_end -= cur;
3786 len -= cur;
3787 }
3788}
3789
3790static inline void btrfs_release_extent_buffer_rcu(struct rcu_head *head)
3791{
3792 struct extent_buffer *eb =
3793 container_of(head, struct extent_buffer, rcu_head);
3794
3795 btrfs_release_extent_buffer(eb);
3796}
3797
3798int try_release_extent_buffer(struct extent_io_tree *tree, struct page *page)
3799{
3800 u64 start = page_offset(page);
3801 struct extent_buffer *eb;
3802 int ret = 1;
3803
3804 spin_lock(&tree->buffer_lock);
3805 eb = radix_tree_lookup(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3806 if (!eb) {
3807 spin_unlock(&tree->buffer_lock);
3808 return ret;
3809 }
3810
3811 if (test_bit(EXTENT_BUFFER_DIRTY, &eb->bflags)) {
3812 ret = 0;
3813 goto out;
3814 }
3815
3816 /*
3817 * set @eb->refs to 0 if it is already 1, and then release the @eb.
3818 * Or go back.
3819 */
3820 if (atomic_cmpxchg(&eb->refs, 1, 0) != 1) {
3821 ret = 0;
3822 goto out;
3823 }
3824
3825 radix_tree_delete(&tree->buffer, start >> PAGE_CACHE_SHIFT);
3826out:
3827 spin_unlock(&tree->buffer_lock);
3828
3829 /* at this point we can safely release the extent buffer */
3830 if (atomic_read(&eb->refs) == 0)
3831 call_rcu(&eb->rcu_head, btrfs_release_extent_buffer_rcu);
3832 return ret;
3833}