1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2007 Oracle. All rights reserved.
7 #include <linux/pagemap.h>
8 #include <linux/time.h>
9 #include <linux/init.h>
10 #include <linux/string.h>
11 #include <linux/backing-dev.h>
12 #include <linux/falloc.h>
13 #include <linux/writeback.h>
14 #include <linux/compat.h>
15 #include <linux/slab.h>
16 #include <linux/btrfs.h>
17 #include <linux/uio.h>
18 #include <linux/iversion.h>
19 #include <linux/fsverity.h>
20 #include <linux/iomap.h>
23 #include "transaction.h"
24 #include "btrfs_inode.h"
25 #include "print-tree.h"
30 #include "compression.h"
31 #include "delalloc-space.h"
35 #include "accessors.h"
36 #include "extent-tree.h"
37 #include "file-item.h"
42 /* simple helper to fault in pages and copy. This should go away
43 * and be replaced with calls into generic code.
45 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
46 struct page **prepared_pages,
50 size_t total_copied = 0;
52 int offset = offset_in_page(pos);
54 while (write_bytes > 0) {
55 size_t count = min_t(size_t,
56 PAGE_SIZE - offset, write_bytes);
57 struct page *page = prepared_pages[pg];
59 * Copy data from userspace to the current page
61 copied = copy_page_from_iter_atomic(page, offset, count, i);
63 /* Flush processor's dcache for this page */
64 flush_dcache_page(page);
67 * if we get a partial write, we can end up with
68 * partially up to date pages. These add
69 * a lot of complexity, so make sure they don't
70 * happen by forcing this copy to be retried.
72 * The rest of the btrfs_file_write code will fall
73 * back to page at a time copies after we return 0.
75 if (unlikely(copied < count)) {
76 if (!PageUptodate(page)) {
77 iov_iter_revert(i, copied);
84 write_bytes -= copied;
85 total_copied += copied;
87 if (offset == PAGE_SIZE) {
96 * unlocks pages after btrfs_file_write is done with them
98 static void btrfs_drop_pages(struct btrfs_fs_info *fs_info,
99 struct page **pages, size_t num_pages,
103 u64 block_start = round_down(pos, fs_info->sectorsize);
104 u64 block_len = round_up(pos + copied, fs_info->sectorsize) - block_start;
106 ASSERT(block_len <= U32_MAX);
107 for (i = 0; i < num_pages; i++) {
108 /* page checked is some magic around finding pages that
109 * have been modified without going through btrfs_set_page_dirty
110 * clear it here. There should be no need to mark the pages
111 * accessed as prepare_pages should have marked them accessed
112 * in prepare_pages via find_or_create_page()
114 btrfs_folio_clamp_clear_checked(fs_info, page_folio(pages[i]),
115 block_start, block_len);
116 unlock_page(pages[i]);
122 * After btrfs_copy_from_user(), update the following things for delalloc:
123 * - Mark newly dirtied pages as DELALLOC in the io tree.
124 * Used to advise which range is to be written back.
125 * - Mark modified pages as Uptodate/Dirty and not needing COW fixup
126 * - Update inode size for past EOF write
128 int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
129 size_t num_pages, loff_t pos, size_t write_bytes,
130 struct extent_state **cached, bool noreserve)
132 struct btrfs_fs_info *fs_info = inode->root->fs_info;
137 u64 end_of_last_block;
138 u64 end_pos = pos + write_bytes;
139 loff_t isize = i_size_read(&inode->vfs_inode);
140 unsigned int extra_bits = 0;
142 if (write_bytes == 0)
146 extra_bits |= EXTENT_NORESERVE;
148 start_pos = round_down(pos, fs_info->sectorsize);
149 num_bytes = round_up(write_bytes + pos - start_pos,
150 fs_info->sectorsize);
151 ASSERT(num_bytes <= U32_MAX);
153 end_of_last_block = start_pos + num_bytes - 1;
156 * The pages may have already been dirty, clear out old accounting so
157 * we can set things up properly
159 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
160 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
163 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
168 for (i = 0; i < num_pages; i++) {
169 struct page *p = pages[i];
171 btrfs_folio_clamp_set_uptodate(fs_info, page_folio(p),
172 start_pos, num_bytes);
173 btrfs_folio_clamp_clear_checked(fs_info, page_folio(p),
174 start_pos, num_bytes);
175 btrfs_folio_clamp_set_dirty(fs_info, page_folio(p),
176 start_pos, num_bytes);
180 * we've only changed i_size in ram, and we haven't updated
181 * the disk i_size. There is no need to log the inode
185 i_size_write(&inode->vfs_inode, end_pos);
190 * this is very complex, but the basic idea is to drop all extents
191 * in the range start - end. hint_block is filled in with a block number
192 * that would be a good hint to the block allocator for this file.
194 * If an extent intersects the range but is not entirely inside the range
195 * it is either truncated or split. Anything entirely inside the range
196 * is deleted from the tree.
198 * Note: the VFS' inode number of bytes is not updated, it's up to the caller
199 * to deal with that. We set the field 'bytes_found' of the arguments structure
200 * with the number of allocated bytes found in the target range, so that the
201 * caller can update the inode's number of bytes in an atomic way when
202 * replacing extents in a range to avoid races with stat(2).
204 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
205 struct btrfs_root *root, struct btrfs_inode *inode,
206 struct btrfs_drop_extents_args *args)
208 struct btrfs_fs_info *fs_info = root->fs_info;
209 struct extent_buffer *leaf;
210 struct btrfs_file_extent_item *fi;
211 struct btrfs_ref ref = { 0 };
212 struct btrfs_key key;
213 struct btrfs_key new_key;
214 u64 ino = btrfs_ino(inode);
215 u64 search_start = args->start;
218 u64 extent_offset = 0;
220 u64 last_end = args->start;
226 int modify_tree = -1;
229 struct btrfs_path *path = args->path;
231 args->bytes_found = 0;
232 args->extent_inserted = false;
234 /* Must always have a path if ->replace_extent is true */
235 ASSERT(!(args->replace_extent && !args->path));
238 path = btrfs_alloc_path();
245 if (args->drop_cache)
246 btrfs_drop_extent_map_range(inode, args->start, args->end - 1, false);
248 if (args->start >= inode->disk_i_size && !args->replace_extent)
251 update_refs = (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID);
254 ret = btrfs_lookup_file_extent(trans, root, path, ino,
255 search_start, modify_tree);
258 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
259 leaf = path->nodes[0];
260 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
261 if (key.objectid == ino &&
262 key.type == BTRFS_EXTENT_DATA_KEY)
267 leaf = path->nodes[0];
268 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
270 ret = btrfs_next_leaf(root, path);
277 leaf = path->nodes[0];
281 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
283 if (key.objectid > ino)
285 if (WARN_ON_ONCE(key.objectid < ino) ||
286 key.type < BTRFS_EXTENT_DATA_KEY) {
291 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
294 fi = btrfs_item_ptr(leaf, path->slots[0],
295 struct btrfs_file_extent_item);
296 extent_type = btrfs_file_extent_type(leaf, fi);
298 if (extent_type == BTRFS_FILE_EXTENT_REG ||
299 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
300 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
301 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
302 extent_offset = btrfs_file_extent_offset(leaf, fi);
303 extent_end = key.offset +
304 btrfs_file_extent_num_bytes(leaf, fi);
305 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
306 extent_end = key.offset +
307 btrfs_file_extent_ram_bytes(leaf, fi);
314 * Don't skip extent items representing 0 byte lengths. They
315 * used to be created (bug) if while punching holes we hit
316 * -ENOSPC condition. So if we find one here, just ensure we
317 * delete it, otherwise we would insert a new file extent item
318 * with the same key (offset) as that 0 bytes length file
319 * extent item in the call to setup_items_for_insert() later
322 if (extent_end == key.offset && extent_end >= search_start) {
323 last_end = extent_end;
324 goto delete_extent_item;
327 if (extent_end <= search_start) {
333 search_start = max(key.offset, args->start);
334 if (recow || !modify_tree) {
336 btrfs_release_path(path);
341 * | - range to drop - |
342 * | -------- extent -------- |
344 if (args->start > key.offset && args->end < extent_end) {
346 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
351 memcpy(&new_key, &key, sizeof(new_key));
352 new_key.offset = args->start;
353 ret = btrfs_duplicate_item(trans, root, path,
355 if (ret == -EAGAIN) {
356 btrfs_release_path(path);
362 leaf = path->nodes[0];
363 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
364 struct btrfs_file_extent_item);
365 btrfs_set_file_extent_num_bytes(leaf, fi,
366 args->start - key.offset);
368 fi = btrfs_item_ptr(leaf, path->slots[0],
369 struct btrfs_file_extent_item);
371 extent_offset += args->start - key.offset;
372 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
373 btrfs_set_file_extent_num_bytes(leaf, fi,
374 extent_end - args->start);
375 btrfs_mark_buffer_dirty(trans, leaf);
377 if (update_refs && disk_bytenr > 0) {
378 btrfs_init_generic_ref(&ref,
379 BTRFS_ADD_DELAYED_REF,
380 disk_bytenr, num_bytes, 0,
381 root->root_key.objectid);
382 btrfs_init_data_ref(&ref,
383 root->root_key.objectid,
385 args->start - extent_offset,
387 ret = btrfs_inc_extent_ref(trans, &ref);
389 btrfs_abort_transaction(trans, ret);
393 key.offset = args->start;
396 * From here on out we will have actually dropped something, so
397 * last_end can be updated.
399 last_end = extent_end;
402 * | ---- range to drop ----- |
403 * | -------- extent -------- |
405 if (args->start <= key.offset && args->end < extent_end) {
406 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
411 memcpy(&new_key, &key, sizeof(new_key));
412 new_key.offset = args->end;
413 btrfs_set_item_key_safe(trans, path, &new_key);
415 extent_offset += args->end - key.offset;
416 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
417 btrfs_set_file_extent_num_bytes(leaf, fi,
418 extent_end - args->end);
419 btrfs_mark_buffer_dirty(trans, leaf);
420 if (update_refs && disk_bytenr > 0)
421 args->bytes_found += args->end - key.offset;
425 search_start = extent_end;
427 * | ---- range to drop ----- |
428 * | -------- extent -------- |
430 if (args->start > key.offset && args->end >= extent_end) {
432 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
437 btrfs_set_file_extent_num_bytes(leaf, fi,
438 args->start - key.offset);
439 btrfs_mark_buffer_dirty(trans, leaf);
440 if (update_refs && disk_bytenr > 0)
441 args->bytes_found += extent_end - args->start;
442 if (args->end == extent_end)
450 * | ---- range to drop ----- |
451 * | ------ extent ------ |
453 if (args->start <= key.offset && args->end >= extent_end) {
456 del_slot = path->slots[0];
459 BUG_ON(del_slot + del_nr != path->slots[0]);
464 extent_type == BTRFS_FILE_EXTENT_INLINE) {
465 args->bytes_found += extent_end - key.offset;
466 extent_end = ALIGN(extent_end,
467 fs_info->sectorsize);
468 } else if (update_refs && disk_bytenr > 0) {
469 btrfs_init_generic_ref(&ref,
470 BTRFS_DROP_DELAYED_REF,
471 disk_bytenr, num_bytes, 0,
472 root->root_key.objectid);
473 btrfs_init_data_ref(&ref,
474 root->root_key.objectid,
476 key.offset - extent_offset, 0,
478 ret = btrfs_free_extent(trans, &ref);
480 btrfs_abort_transaction(trans, ret);
483 args->bytes_found += extent_end - key.offset;
486 if (args->end == extent_end)
489 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
494 ret = btrfs_del_items(trans, root, path, del_slot,
497 btrfs_abort_transaction(trans, ret);
504 btrfs_release_path(path);
511 if (!ret && del_nr > 0) {
513 * Set path->slots[0] to first slot, so that after the delete
514 * if items are move off from our leaf to its immediate left or
515 * right neighbor leafs, we end up with a correct and adjusted
516 * path->slots[0] for our insertion (if args->replace_extent).
518 path->slots[0] = del_slot;
519 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
521 btrfs_abort_transaction(trans, ret);
524 leaf = path->nodes[0];
526 * If btrfs_del_items() was called, it might have deleted a leaf, in
527 * which case it unlocked our path, so check path->locks[0] matches a
530 if (!ret && args->replace_extent &&
531 path->locks[0] == BTRFS_WRITE_LOCK &&
532 btrfs_leaf_free_space(leaf) >=
533 sizeof(struct btrfs_item) + args->extent_item_size) {
536 key.type = BTRFS_EXTENT_DATA_KEY;
537 key.offset = args->start;
538 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
539 struct btrfs_key slot_key;
541 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
542 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
545 btrfs_setup_item_for_insert(trans, root, path, &key,
546 args->extent_item_size);
547 args->extent_inserted = true;
551 btrfs_free_path(path);
552 else if (!args->extent_inserted)
553 btrfs_release_path(path);
555 args->drop_end = found ? min(args->end, last_end) : args->end;
560 static int extent_mergeable(struct extent_buffer *leaf, int slot,
561 u64 objectid, u64 bytenr, u64 orig_offset,
562 u64 *start, u64 *end)
564 struct btrfs_file_extent_item *fi;
565 struct btrfs_key key;
568 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
571 btrfs_item_key_to_cpu(leaf, &key, slot);
572 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
575 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
576 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
577 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
578 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
579 btrfs_file_extent_compression(leaf, fi) ||
580 btrfs_file_extent_encryption(leaf, fi) ||
581 btrfs_file_extent_other_encoding(leaf, fi))
584 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
585 if ((*start && *start != key.offset) || (*end && *end != extent_end))
594 * Mark extent in the range start - end as written.
596 * This changes extent type from 'pre-allocated' to 'regular'. If only
597 * part of extent is marked as written, the extent will be split into
600 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
601 struct btrfs_inode *inode, u64 start, u64 end)
603 struct btrfs_root *root = inode->root;
604 struct extent_buffer *leaf;
605 struct btrfs_path *path;
606 struct btrfs_file_extent_item *fi;
607 struct btrfs_ref ref = { 0 };
608 struct btrfs_key key;
609 struct btrfs_key new_key;
621 u64 ino = btrfs_ino(inode);
623 path = btrfs_alloc_path();
630 key.type = BTRFS_EXTENT_DATA_KEY;
633 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
636 if (ret > 0 && path->slots[0] > 0)
639 leaf = path->nodes[0];
640 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
641 if (key.objectid != ino ||
642 key.type != BTRFS_EXTENT_DATA_KEY) {
644 btrfs_abort_transaction(trans, ret);
647 fi = btrfs_item_ptr(leaf, path->slots[0],
648 struct btrfs_file_extent_item);
649 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
651 btrfs_abort_transaction(trans, ret);
654 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
655 if (key.offset > start || extent_end < end) {
657 btrfs_abort_transaction(trans, ret);
661 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
662 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
663 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
664 memcpy(&new_key, &key, sizeof(new_key));
666 if (start == key.offset && end < extent_end) {
669 if (extent_mergeable(leaf, path->slots[0] - 1,
670 ino, bytenr, orig_offset,
671 &other_start, &other_end)) {
672 new_key.offset = end;
673 btrfs_set_item_key_safe(trans, path, &new_key);
674 fi = btrfs_item_ptr(leaf, path->slots[0],
675 struct btrfs_file_extent_item);
676 btrfs_set_file_extent_generation(leaf, fi,
678 btrfs_set_file_extent_num_bytes(leaf, fi,
680 btrfs_set_file_extent_offset(leaf, fi,
682 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
683 struct btrfs_file_extent_item);
684 btrfs_set_file_extent_generation(leaf, fi,
686 btrfs_set_file_extent_num_bytes(leaf, fi,
688 btrfs_mark_buffer_dirty(trans, leaf);
693 if (start > key.offset && end == extent_end) {
696 if (extent_mergeable(leaf, path->slots[0] + 1,
697 ino, bytenr, orig_offset,
698 &other_start, &other_end)) {
699 fi = btrfs_item_ptr(leaf, path->slots[0],
700 struct btrfs_file_extent_item);
701 btrfs_set_file_extent_num_bytes(leaf, fi,
703 btrfs_set_file_extent_generation(leaf, fi,
706 new_key.offset = start;
707 btrfs_set_item_key_safe(trans, path, &new_key);
709 fi = btrfs_item_ptr(leaf, path->slots[0],
710 struct btrfs_file_extent_item);
711 btrfs_set_file_extent_generation(leaf, fi,
713 btrfs_set_file_extent_num_bytes(leaf, fi,
715 btrfs_set_file_extent_offset(leaf, fi,
716 start - orig_offset);
717 btrfs_mark_buffer_dirty(trans, leaf);
722 while (start > key.offset || end < extent_end) {
723 if (key.offset == start)
726 new_key.offset = split;
727 ret = btrfs_duplicate_item(trans, root, path, &new_key);
728 if (ret == -EAGAIN) {
729 btrfs_release_path(path);
733 btrfs_abort_transaction(trans, ret);
737 leaf = path->nodes[0];
738 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
739 struct btrfs_file_extent_item);
740 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
741 btrfs_set_file_extent_num_bytes(leaf, fi,
744 fi = btrfs_item_ptr(leaf, path->slots[0],
745 struct btrfs_file_extent_item);
747 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
748 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
749 btrfs_set_file_extent_num_bytes(leaf, fi,
751 btrfs_mark_buffer_dirty(trans, leaf);
753 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
754 num_bytes, 0, root->root_key.objectid);
755 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
756 orig_offset, 0, false);
757 ret = btrfs_inc_extent_ref(trans, &ref);
759 btrfs_abort_transaction(trans, ret);
763 if (split == start) {
766 if (start != key.offset) {
768 btrfs_abort_transaction(trans, ret);
779 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
780 num_bytes, 0, root->root_key.objectid);
781 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset,
783 if (extent_mergeable(leaf, path->slots[0] + 1,
784 ino, bytenr, orig_offset,
785 &other_start, &other_end)) {
787 btrfs_release_path(path);
790 extent_end = other_end;
791 del_slot = path->slots[0] + 1;
793 ret = btrfs_free_extent(trans, &ref);
795 btrfs_abort_transaction(trans, ret);
801 if (extent_mergeable(leaf, path->slots[0] - 1,
802 ino, bytenr, orig_offset,
803 &other_start, &other_end)) {
805 btrfs_release_path(path);
808 key.offset = other_start;
809 del_slot = path->slots[0];
811 ret = btrfs_free_extent(trans, &ref);
813 btrfs_abort_transaction(trans, ret);
818 fi = btrfs_item_ptr(leaf, path->slots[0],
819 struct btrfs_file_extent_item);
820 btrfs_set_file_extent_type(leaf, fi,
821 BTRFS_FILE_EXTENT_REG);
822 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
823 btrfs_mark_buffer_dirty(trans, leaf);
825 fi = btrfs_item_ptr(leaf, del_slot - 1,
826 struct btrfs_file_extent_item);
827 btrfs_set_file_extent_type(leaf, fi,
828 BTRFS_FILE_EXTENT_REG);
829 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
830 btrfs_set_file_extent_num_bytes(leaf, fi,
831 extent_end - key.offset);
832 btrfs_mark_buffer_dirty(trans, leaf);
834 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
836 btrfs_abort_transaction(trans, ret);
841 btrfs_free_path(path);
846 * on error we return an unlocked page and the error value
847 * on success we return a locked page and 0
849 static int prepare_uptodate_page(struct inode *inode,
850 struct page *page, u64 pos,
853 struct folio *folio = page_folio(page);
856 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
857 !PageUptodate(page)) {
858 ret = btrfs_read_folio(NULL, folio);
862 if (!PageUptodate(page)) {
868 * Since btrfs_read_folio() will unlock the folio before it
869 * returns, there is a window where btrfs_release_folio() can be
870 * called to release the page. Here we check both inode
871 * mapping and PagePrivate() to make sure the page was not
874 * The private flag check is essential for subpage as we need
875 * to store extra bitmap using folio private.
877 if (page->mapping != inode->i_mapping || !folio_test_private(folio)) {
885 static fgf_t get_prepare_fgp_flags(bool nowait)
887 fgf_t fgp_flags = FGP_LOCK | FGP_ACCESSED | FGP_CREAT;
890 fgp_flags |= FGP_NOWAIT;
895 static gfp_t get_prepare_gfp_flags(struct inode *inode, bool nowait)
899 gfp = btrfs_alloc_write_mask(inode->i_mapping);
901 gfp &= ~__GFP_DIRECT_RECLAIM;
909 * this just gets pages into the page cache and locks them down.
911 static noinline int prepare_pages(struct inode *inode, struct page **pages,
912 size_t num_pages, loff_t pos,
913 size_t write_bytes, bool force_uptodate,
917 unsigned long index = pos >> PAGE_SHIFT;
918 gfp_t mask = get_prepare_gfp_flags(inode, nowait);
919 fgf_t fgp_flags = get_prepare_fgp_flags(nowait);
923 for (i = 0; i < num_pages; i++) {
925 pages[i] = pagecache_get_page(inode->i_mapping, index + i,
926 fgp_flags, mask | __GFP_WRITE);
936 err = set_page_extent_mapped(pages[i]);
943 err = prepare_uptodate_page(inode, pages[i], pos,
945 if (!err && i == num_pages - 1)
946 err = prepare_uptodate_page(inode, pages[i],
947 pos + write_bytes, false);
950 if (!nowait && err == -EAGAIN) {
957 wait_on_page_writeback(pages[i]);
963 unlock_page(pages[faili]);
964 put_page(pages[faili]);
972 * This function locks the extent and properly waits for data=ordered extents
973 * to finish before allowing the pages to be modified if need.
976 * 1 - the extent is locked
977 * 0 - the extent is not locked, and everything is OK
978 * -EAGAIN - need re-prepare the pages
979 * the other < 0 number - Something wrong happens
982 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
983 size_t num_pages, loff_t pos,
985 u64 *lockstart, u64 *lockend, bool nowait,
986 struct extent_state **cached_state)
988 struct btrfs_fs_info *fs_info = inode->root->fs_info;
994 start_pos = round_down(pos, fs_info->sectorsize);
995 last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
997 if (start_pos < inode->vfs_inode.i_size) {
998 struct btrfs_ordered_extent *ordered;
1001 if (!try_lock_extent(&inode->io_tree, start_pos, last_pos,
1003 for (i = 0; i < num_pages; i++) {
1004 unlock_page(pages[i]);
1012 lock_extent(&inode->io_tree, start_pos, last_pos, cached_state);
1015 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1016 last_pos - start_pos + 1);
1018 ordered->file_offset + ordered->num_bytes > start_pos &&
1019 ordered->file_offset <= last_pos) {
1020 unlock_extent(&inode->io_tree, start_pos, last_pos,
1022 for (i = 0; i < num_pages; i++) {
1023 unlock_page(pages[i]);
1026 btrfs_start_ordered_extent(ordered);
1027 btrfs_put_ordered_extent(ordered);
1031 btrfs_put_ordered_extent(ordered);
1033 *lockstart = start_pos;
1034 *lockend = last_pos;
1039 * We should be called after prepare_pages() which should have locked
1040 * all pages in the range.
1042 for (i = 0; i < num_pages; i++)
1043 WARN_ON(!PageLocked(pages[i]));
1049 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1051 * @pos: File offset.
1052 * @write_bytes: The length to write, will be updated to the nocow writeable
1055 * This function will flush ordered extents in the range to ensure proper
1059 * > 0 If we can nocow, and updates @write_bytes.
1060 * 0 If we can't do a nocow write.
1061 * -EAGAIN If we can't do a nocow write because snapshoting of the inode's
1062 * root is in progress.
1063 * < 0 If an error happened.
1065 * NOTE: Callers need to call btrfs_check_nocow_unlock() if we return > 0.
1067 int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1068 size_t *write_bytes, bool nowait)
1070 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1071 struct btrfs_root *root = inode->root;
1072 struct extent_state *cached_state = NULL;
1073 u64 lockstart, lockend;
1077 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1080 if (!btrfs_drew_try_write_lock(&root->snapshot_lock))
1083 lockstart = round_down(pos, fs_info->sectorsize);
1084 lockend = round_up(pos + *write_bytes,
1085 fs_info->sectorsize) - 1;
1086 num_bytes = lockend - lockstart + 1;
1089 if (!btrfs_try_lock_ordered_range(inode, lockstart, lockend,
1091 btrfs_drew_write_unlock(&root->snapshot_lock);
1095 btrfs_lock_and_flush_ordered_range(inode, lockstart, lockend,
1098 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1099 NULL, NULL, NULL, nowait, false);
1101 btrfs_drew_write_unlock(&root->snapshot_lock);
1103 *write_bytes = min_t(size_t, *write_bytes ,
1104 num_bytes - pos + lockstart);
1105 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
1110 void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1112 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1115 static void update_time_for_write(struct inode *inode)
1117 struct timespec64 now, ts;
1119 if (IS_NOCMTIME(inode))
1122 now = current_time(inode);
1123 ts = inode_get_mtime(inode);
1124 if (!timespec64_equal(&ts, &now))
1125 inode_set_mtime_to_ts(inode, now);
1127 ts = inode_get_ctime(inode);
1128 if (!timespec64_equal(&ts, &now))
1129 inode_set_ctime_to_ts(inode, now);
1131 if (IS_I_VERSION(inode))
1132 inode_inc_iversion(inode);
1135 static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1138 struct file *file = iocb->ki_filp;
1139 struct inode *inode = file_inode(file);
1140 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1141 loff_t pos = iocb->ki_pos;
1147 * Quickly bail out on NOWAIT writes if we don't have the nodatacow or
1148 * prealloc flags, as without those flags we always have to COW. We will
1149 * later check if we can really COW into the target range (using
1150 * can_nocow_extent() at btrfs_get_blocks_direct_write()).
1152 if ((iocb->ki_flags & IOCB_NOWAIT) &&
1153 !(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1156 ret = file_remove_privs(file);
1161 * We reserve space for updating the inode when we reserve space for the
1162 * extent we are going to write, so we will enospc out there. We don't
1163 * need to start yet another transaction to update the inode as we will
1164 * update the inode when we finish writing whatever data we write.
1166 update_time_for_write(inode);
1168 start_pos = round_down(pos, fs_info->sectorsize);
1169 oldsize = i_size_read(inode);
1170 if (start_pos > oldsize) {
1171 /* Expand hole size to cover write data, preventing empty gap */
1172 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1174 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1182 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1185 struct file *file = iocb->ki_filp;
1187 struct inode *inode = file_inode(file);
1188 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1189 struct page **pages = NULL;
1190 struct extent_changeset *data_reserved = NULL;
1191 u64 release_bytes = 0;
1194 size_t num_written = 0;
1197 bool only_release_metadata = false;
1198 bool force_page_uptodate = false;
1199 loff_t old_isize = i_size_read(inode);
1200 unsigned int ilock_flags = 0;
1201 const bool nowait = (iocb->ki_flags & IOCB_NOWAIT);
1202 unsigned int bdp_flags = (nowait ? BDP_ASYNC : 0);
1205 ilock_flags |= BTRFS_ILOCK_TRY;
1207 ret = btrfs_inode_lock(BTRFS_I(inode), ilock_flags);
1211 ret = generic_write_checks(iocb, i);
1215 ret = btrfs_write_check(iocb, i, ret);
1220 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1221 PAGE_SIZE / (sizeof(struct page *)));
1222 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1223 nrptrs = max(nrptrs, 8);
1224 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1230 while (iov_iter_count(i) > 0) {
1231 struct extent_state *cached_state = NULL;
1232 size_t offset = offset_in_page(pos);
1233 size_t sector_offset;
1234 size_t write_bytes = min(iov_iter_count(i),
1235 nrptrs * (size_t)PAGE_SIZE -
1238 size_t reserve_bytes;
1241 size_t dirty_sectors;
1246 * Fault pages before locking them in prepare_pages
1247 * to avoid recursive lock
1249 if (unlikely(fault_in_iov_iter_readable(i, write_bytes))) {
1254 only_release_metadata = false;
1255 sector_offset = pos & (fs_info->sectorsize - 1);
1257 extent_changeset_release(data_reserved);
1258 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1259 &data_reserved, pos,
1260 write_bytes, nowait);
1264 if (nowait && (ret == -ENOSPC || ret == -EAGAIN)) {
1270 * If we don't have to COW at the offset, reserve
1271 * metadata only. write_bytes may get smaller than
1274 can_nocow = btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1275 &write_bytes, nowait);
1282 only_release_metadata = true;
1285 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1286 WARN_ON(num_pages > nrptrs);
1287 reserve_bytes = round_up(write_bytes + sector_offset,
1288 fs_info->sectorsize);
1289 WARN_ON(reserve_bytes == 0);
1290 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1292 reserve_bytes, nowait);
1294 if (!only_release_metadata)
1295 btrfs_free_reserved_data_space(BTRFS_I(inode),
1299 btrfs_check_nocow_unlock(BTRFS_I(inode));
1301 if (nowait && ret == -ENOSPC)
1306 release_bytes = reserve_bytes;
1308 ret = balance_dirty_pages_ratelimited_flags(inode->i_mapping, bdp_flags);
1310 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1315 * This is going to setup the pages array with the number of
1316 * pages we want, so we don't really need to worry about the
1317 * contents of pages from loop to loop
1319 ret = prepare_pages(inode, pages, num_pages,
1320 pos, write_bytes, force_page_uptodate, false);
1322 btrfs_delalloc_release_extents(BTRFS_I(inode),
1327 extents_locked = lock_and_cleanup_extent_if_need(
1328 BTRFS_I(inode), pages,
1329 num_pages, pos, write_bytes, &lockstart,
1330 &lockend, nowait, &cached_state);
1331 if (extents_locked < 0) {
1332 if (!nowait && extents_locked == -EAGAIN)
1335 btrfs_delalloc_release_extents(BTRFS_I(inode),
1337 ret = extents_locked;
1341 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1343 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1344 dirty_sectors = round_up(copied + sector_offset,
1345 fs_info->sectorsize);
1346 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1349 * if we have trouble faulting in the pages, fall
1350 * back to one page at a time
1352 if (copied < write_bytes)
1356 force_page_uptodate = true;
1360 force_page_uptodate = false;
1361 dirty_pages = DIV_ROUND_UP(copied + offset,
1365 if (num_sectors > dirty_sectors) {
1366 /* release everything except the sectors we dirtied */
1367 release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1368 if (only_release_metadata) {
1369 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1370 release_bytes, true);
1374 __pos = round_down(pos,
1375 fs_info->sectorsize) +
1376 (dirty_pages << PAGE_SHIFT);
1377 btrfs_delalloc_release_space(BTRFS_I(inode),
1378 data_reserved, __pos,
1379 release_bytes, true);
1383 release_bytes = round_up(copied + sector_offset,
1384 fs_info->sectorsize);
1386 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1387 dirty_pages, pos, copied,
1388 &cached_state, only_release_metadata);
1391 * If we have not locked the extent range, because the range's
1392 * start offset is >= i_size, we might still have a non-NULL
1393 * cached extent state, acquired while marking the extent range
1394 * as delalloc through btrfs_dirty_pages(). Therefore free any
1395 * possible cached extent state to avoid a memory leak.
1398 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
1399 lockend, &cached_state);
1401 free_extent_state(cached_state);
1403 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1405 btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1410 if (only_release_metadata)
1411 btrfs_check_nocow_unlock(BTRFS_I(inode));
1413 btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1418 num_written += copied;
1423 if (release_bytes) {
1424 if (only_release_metadata) {
1425 btrfs_check_nocow_unlock(BTRFS_I(inode));
1426 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1427 release_bytes, true);
1429 btrfs_delalloc_release_space(BTRFS_I(inode),
1431 round_down(pos, fs_info->sectorsize),
1432 release_bytes, true);
1436 extent_changeset_free(data_reserved);
1437 if (num_written > 0) {
1438 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1439 iocb->ki_pos += num_written;
1442 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1443 return num_written ? num_written : ret;
1446 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1447 const struct iov_iter *iter, loff_t offset)
1449 const u32 blocksize_mask = fs_info->sectorsize - 1;
1451 if (offset & blocksize_mask)
1454 if (iov_iter_alignment(iter) & blocksize_mask)
1460 static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1462 struct file *file = iocb->ki_filp;
1463 struct inode *inode = file_inode(file);
1464 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1466 ssize_t written = 0;
1467 ssize_t written_buffered;
1468 size_t prev_left = 0;
1471 unsigned int ilock_flags = 0;
1472 struct iomap_dio *dio;
1474 if (iocb->ki_flags & IOCB_NOWAIT)
1475 ilock_flags |= BTRFS_ILOCK_TRY;
1478 * If the write DIO is within EOF, use a shared lock and also only if
1479 * security bits will likely not be dropped by file_remove_privs() called
1480 * from btrfs_write_check(). Either will need to be rechecked after the
1481 * lock was acquired.
1483 if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode) && IS_NOSEC(inode))
1484 ilock_flags |= BTRFS_ILOCK_SHARED;
1487 err = btrfs_inode_lock(BTRFS_I(inode), ilock_flags);
1491 /* Shared lock cannot be used with security bits set. */
1492 if ((ilock_flags & BTRFS_ILOCK_SHARED) && !IS_NOSEC(inode)) {
1493 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1494 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1498 err = generic_write_checks(iocb, from);
1500 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1504 err = btrfs_write_check(iocb, from, err);
1506 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1512 * Re-check since file size may have changed just before taking the
1513 * lock or pos may have changed because of O_APPEND in generic_write_check()
1515 if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1516 pos + iov_iter_count(from) > i_size_read(inode)) {
1517 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1518 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1522 if (check_direct_IO(fs_info, from, pos)) {
1523 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1528 * The iov_iter can be mapped to the same file range we are writing to.
1529 * If that's the case, then we will deadlock in the iomap code, because
1530 * it first calls our callback btrfs_dio_iomap_begin(), which will create
1531 * an ordered extent, and after that it will fault in the pages that the
1532 * iov_iter refers to. During the fault in we end up in the readahead
1533 * pages code (starting at btrfs_readahead()), which will lock the range,
1534 * find that ordered extent and then wait for it to complete (at
1535 * btrfs_lock_and_flush_ordered_range()), resulting in a deadlock since
1536 * obviously the ordered extent can never complete as we didn't submit
1537 * yet the respective bio(s). This always happens when the buffer is
1538 * memory mapped to the same file range, since the iomap DIO code always
1539 * invalidates pages in the target file range (after starting and waiting
1540 * for any writeback).
1542 * So here we disable page faults in the iov_iter and then retry if we
1543 * got -EFAULT, faulting in the pages before the retry.
1545 from->nofault = true;
1546 dio = btrfs_dio_write(iocb, from, written);
1547 from->nofault = false;
1550 * iomap_dio_complete() will call btrfs_sync_file() if we have a dsync
1551 * iocb, and that needs to lock the inode. So unlock it before calling
1552 * iomap_dio_complete() to avoid a deadlock.
1554 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1556 if (IS_ERR_OR_NULL(dio))
1557 err = PTR_ERR_OR_ZERO(dio);
1559 err = iomap_dio_complete(dio);
1561 /* No increment (+=) because iomap returns a cumulative value. */
1565 if (iov_iter_count(from) > 0 && (err == -EFAULT || err > 0)) {
1566 const size_t left = iov_iter_count(from);
1568 * We have more data left to write. Try to fault in as many as
1569 * possible of the remainder pages and retry. We do this without
1570 * releasing and locking again the inode, to prevent races with
1573 * Also, in case the iov refers to pages in the file range of the
1574 * file we want to write to (due to a mmap), we could enter an
1575 * infinite loop if we retry after faulting the pages in, since
1576 * iomap will invalidate any pages in the range early on, before
1577 * it tries to fault in the pages of the iov. So we keep track of
1578 * how much was left of iov in the previous EFAULT and fallback
1579 * to buffered IO in case we haven't made any progress.
1581 if (left == prev_left) {
1584 fault_in_iov_iter_readable(from, left);
1591 * If 'err' is -ENOTBLK or we have not written all data, then it means
1592 * we must fallback to buffered IO.
1594 if ((err < 0 && err != -ENOTBLK) || !iov_iter_count(from))
1599 * If we are in a NOWAIT context, then return -EAGAIN to signal the caller
1600 * it must retry the operation in a context where blocking is acceptable,
1601 * because even if we end up not blocking during the buffered IO attempt
1602 * below, we will block when flushing and waiting for the IO.
1604 if (iocb->ki_flags & IOCB_NOWAIT) {
1610 written_buffered = btrfs_buffered_write(iocb, from);
1611 if (written_buffered < 0) {
1612 err = written_buffered;
1616 * Ensure all data is persisted. We want the next direct IO read to be
1617 * able to read what was just written.
1619 endbyte = pos + written_buffered - 1;
1620 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1623 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1626 written += written_buffered;
1627 iocb->ki_pos = pos + written_buffered;
1628 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1629 endbyte >> PAGE_SHIFT);
1631 return err < 0 ? err : written;
1634 static ssize_t btrfs_encoded_write(struct kiocb *iocb, struct iov_iter *from,
1635 const struct btrfs_ioctl_encoded_io_args *encoded)
1637 struct file *file = iocb->ki_filp;
1638 struct inode *inode = file_inode(file);
1642 btrfs_inode_lock(BTRFS_I(inode), 0);
1643 count = encoded->len;
1644 ret = generic_write_checks_count(iocb, &count);
1645 if (ret == 0 && count != encoded->len) {
1647 * The write got truncated by generic_write_checks_count(). We
1648 * can't do a partial encoded write.
1652 if (ret || encoded->len == 0)
1655 ret = btrfs_write_check(iocb, from, encoded->len);
1659 ret = btrfs_do_encoded_write(iocb, from, encoded);
1661 btrfs_inode_unlock(BTRFS_I(inode), 0);
1665 ssize_t btrfs_do_write_iter(struct kiocb *iocb, struct iov_iter *from,
1666 const struct btrfs_ioctl_encoded_io_args *encoded)
1668 struct file *file = iocb->ki_filp;
1669 struct btrfs_inode *inode = BTRFS_I(file_inode(file));
1670 ssize_t num_written, num_sync;
1673 * If the fs flips readonly due to some impossible error, although we
1674 * have opened a file as writable, we have to stop this write operation
1675 * to ensure consistency.
1677 if (BTRFS_FS_ERROR(inode->root->fs_info))
1680 if (encoded && (iocb->ki_flags & IOCB_NOWAIT))
1684 num_written = btrfs_encoded_write(iocb, from, encoded);
1685 num_sync = encoded->len;
1686 } else if (iocb->ki_flags & IOCB_DIRECT) {
1687 num_written = btrfs_direct_write(iocb, from);
1688 num_sync = num_written;
1690 num_written = btrfs_buffered_write(iocb, from);
1691 num_sync = num_written;
1694 btrfs_set_inode_last_sub_trans(inode);
1697 num_sync = generic_write_sync(iocb, num_sync);
1699 num_written = num_sync;
1705 static ssize_t btrfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1707 return btrfs_do_write_iter(iocb, from, NULL);
1710 int btrfs_release_file(struct inode *inode, struct file *filp)
1712 struct btrfs_file_private *private = filp->private_data;
1715 kfree(private->filldir_buf);
1716 free_extent_state(private->llseek_cached_state);
1718 filp->private_data = NULL;
1722 * Set by setattr when we are about to truncate a file from a non-zero
1723 * size to a zero size. This tries to flush down new bytes that may
1724 * have been written if the application were using truncate to replace
1727 if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
1728 &BTRFS_I(inode)->runtime_flags))
1729 filemap_flush(inode->i_mapping);
1733 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
1736 struct blk_plug plug;
1739 * This is only called in fsync, which would do synchronous writes, so
1740 * a plug can merge adjacent IOs as much as possible. Esp. in case of
1741 * multiple disks using raid profile, a large IO can be split to
1742 * several segments of stripe length (currently 64K).
1744 blk_start_plug(&plug);
1745 ret = btrfs_fdatawrite_range(inode, start, end);
1746 blk_finish_plug(&plug);
1751 static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
1753 struct btrfs_inode *inode = BTRFS_I(ctx->inode);
1754 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1756 if (btrfs_inode_in_log(inode, btrfs_get_fs_generation(fs_info)) &&
1757 list_empty(&ctx->ordered_extents))
1761 * If we are doing a fast fsync we can not bail out if the inode's
1762 * last_trans is <= then the last committed transaction, because we only
1763 * update the last_trans of the inode during ordered extent completion,
1764 * and for a fast fsync we don't wait for that, we only wait for the
1765 * writeback to complete.
1767 if (inode->last_trans <= btrfs_get_last_trans_committed(fs_info) &&
1768 (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
1769 list_empty(&ctx->ordered_extents)))
1776 * fsync call for both files and directories. This logs the inode into
1777 * the tree log instead of forcing full commits whenever possible.
1779 * It needs to call filemap_fdatawait so that all ordered extent updates are
1780 * in the metadata btree are up to date for copying to the log.
1782 * It drops the inode mutex before doing the tree log commit. This is an
1783 * important optimization for directories because holding the mutex prevents
1784 * new operations on the dir while we write to disk.
1786 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1788 struct dentry *dentry = file_dentry(file);
1789 struct inode *inode = d_inode(dentry);
1790 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1791 struct btrfs_root *root = BTRFS_I(inode)->root;
1792 struct btrfs_trans_handle *trans;
1793 struct btrfs_log_ctx ctx;
1798 trace_btrfs_sync_file(file, datasync);
1800 btrfs_init_log_ctx(&ctx, inode);
1803 * Always set the range to a full range, otherwise we can get into
1804 * several problems, from missing file extent items to represent holes
1805 * when not using the NO_HOLES feature, to log tree corruption due to
1806 * races between hole detection during logging and completion of ordered
1807 * extents outside the range, to missing checksums due to ordered extents
1808 * for which we flushed only a subset of their pages.
1812 len = (u64)LLONG_MAX + 1;
1815 * We write the dirty pages in the range and wait until they complete
1816 * out of the ->i_mutex. If so, we can flush the dirty pages by
1817 * multi-task, and make the performance up. See
1818 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1820 ret = start_ordered_ops(inode, start, end);
1824 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
1826 atomic_inc(&root->log_batch);
1829 * Before we acquired the inode's lock and the mmap lock, someone may
1830 * have dirtied more pages in the target range. We need to make sure
1831 * that writeback for any such pages does not start while we are logging
1832 * the inode, because if it does, any of the following might happen when
1833 * we are not doing a full inode sync:
1835 * 1) We log an extent after its writeback finishes but before its
1836 * checksums are added to the csum tree, leading to -EIO errors
1837 * when attempting to read the extent after a log replay.
1839 * 2) We can end up logging an extent before its writeback finishes.
1840 * Therefore after the log replay we will have a file extent item
1841 * pointing to an unwritten extent (and no data checksums as well).
1843 * So trigger writeback for any eventual new dirty pages and then we
1844 * wait for all ordered extents to complete below.
1846 ret = start_ordered_ops(inode, start, end);
1848 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
1853 * Always check for the full sync flag while holding the inode's lock,
1854 * to avoid races with other tasks. The flag must be either set all the
1855 * time during logging or always off all the time while logging.
1856 * We check the flag here after starting delalloc above, because when
1857 * running delalloc the full sync flag may be set if we need to drop
1858 * extra extent map ranges due to temporary memory allocation failures.
1860 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1861 &BTRFS_I(inode)->runtime_flags);
1864 * We have to do this here to avoid the priority inversion of waiting on
1865 * IO of a lower priority task while holding a transaction open.
1867 * For a full fsync we wait for the ordered extents to complete while
1868 * for a fast fsync we wait just for writeback to complete, and then
1869 * attach the ordered extents to the transaction so that a transaction
1870 * commit waits for their completion, to avoid data loss if we fsync,
1871 * the current transaction commits before the ordered extents complete
1872 * and a power failure happens right after that.
1874 * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
1875 * logical address recorded in the ordered extent may change. We need
1876 * to wait for the IO to stabilize the logical address.
1878 if (full_sync || btrfs_is_zoned(fs_info)) {
1879 ret = btrfs_wait_ordered_range(inode, start, len);
1882 * Get our ordered extents as soon as possible to avoid doing
1883 * checksum lookups in the csum tree, and use instead the
1884 * checksums attached to the ordered extents.
1886 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
1887 &ctx.ordered_extents);
1888 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
1892 goto out_release_extents;
1894 atomic_inc(&root->log_batch);
1896 if (skip_inode_logging(&ctx)) {
1898 * We've had everything committed since the last time we were
1899 * modified so clear this flag in case it was set for whatever
1900 * reason, it's no longer relevant.
1902 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1903 &BTRFS_I(inode)->runtime_flags);
1905 * An ordered extent might have started before and completed
1906 * already with io errors, in which case the inode was not
1907 * updated and we end up here. So check the inode's mapping
1908 * for any errors that might have happened since we last
1909 * checked called fsync.
1911 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
1912 goto out_release_extents;
1916 * We use start here because we will need to wait on the IO to complete
1917 * in btrfs_sync_log, which could require joining a transaction (for
1918 * example checking cross references in the nocow path). If we use join
1919 * here we could get into a situation where we're waiting on IO to
1920 * happen that is blocked on a transaction trying to commit. With start
1921 * we inc the extwriter counter, so we wait for all extwriters to exit
1922 * before we start blocking joiners. This comment is to keep somebody
1923 * from thinking they are super smart and changing this to
1924 * btrfs_join_transaction *cough*Josef*cough*.
1926 trans = btrfs_start_transaction(root, 0);
1927 if (IS_ERR(trans)) {
1928 ret = PTR_ERR(trans);
1929 goto out_release_extents;
1931 trans->in_fsync = true;
1933 ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
1934 btrfs_release_log_ctx_extents(&ctx);
1936 /* Fallthrough and commit/free transaction. */
1937 ret = BTRFS_LOG_FORCE_COMMIT;
1940 /* we've logged all the items and now have a consistent
1941 * version of the file in the log. It is possible that
1942 * someone will come in and modify the file, but that's
1943 * fine because the log is consistent on disk, and we
1944 * have references to all of the file's extents
1946 * It is possible that someone will come in and log the
1947 * file again, but that will end up using the synchronization
1948 * inside btrfs_sync_log to keep things safe.
1950 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
1952 if (ret == BTRFS_NO_LOG_SYNC) {
1953 ret = btrfs_end_transaction(trans);
1957 /* We successfully logged the inode, attempt to sync the log. */
1959 ret = btrfs_sync_log(trans, root, &ctx);
1961 ret = btrfs_end_transaction(trans);
1967 * At this point we need to commit the transaction because we had
1968 * btrfs_need_log_full_commit() or some other error.
1970 * If we didn't do a full sync we have to stop the trans handle, wait on
1971 * the ordered extents, start it again and commit the transaction. If
1972 * we attempt to wait on the ordered extents here we could deadlock with
1973 * something like fallocate() that is holding the extent lock trying to
1974 * start a transaction while some other thread is trying to commit the
1975 * transaction while we (fsync) are currently holding the transaction
1979 ret = btrfs_end_transaction(trans);
1982 ret = btrfs_wait_ordered_range(inode, start, len);
1987 * This is safe to use here because we're only interested in
1988 * making sure the transaction that had the ordered extents is
1989 * committed. We aren't waiting on anything past this point,
1990 * we're purely getting the transaction and committing it.
1992 trans = btrfs_attach_transaction_barrier(root);
1993 if (IS_ERR(trans)) {
1994 ret = PTR_ERR(trans);
1997 * We committed the transaction and there's no currently
1998 * running transaction, this means everything we care
1999 * about made it to disk and we are done.
2007 ret = btrfs_commit_transaction(trans);
2009 ASSERT(list_empty(&ctx.list));
2010 ASSERT(list_empty(&ctx.conflict_inodes));
2011 err = file_check_and_advance_wb_err(file);
2014 return ret > 0 ? -EIO : ret;
2016 out_release_extents:
2017 btrfs_release_log_ctx_extents(&ctx);
2018 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2022 static const struct vm_operations_struct btrfs_file_vm_ops = {
2023 .fault = filemap_fault,
2024 .map_pages = filemap_map_pages,
2025 .page_mkwrite = btrfs_page_mkwrite,
2028 static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2030 struct address_space *mapping = filp->f_mapping;
2032 if (!mapping->a_ops->read_folio)
2035 file_accessed(filp);
2036 vma->vm_ops = &btrfs_file_vm_ops;
2041 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2042 int slot, u64 start, u64 end)
2044 struct btrfs_file_extent_item *fi;
2045 struct btrfs_key key;
2047 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2050 btrfs_item_key_to_cpu(leaf, &key, slot);
2051 if (key.objectid != btrfs_ino(inode) ||
2052 key.type != BTRFS_EXTENT_DATA_KEY)
2055 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2057 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2060 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2063 if (key.offset == end)
2065 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2070 static int fill_holes(struct btrfs_trans_handle *trans,
2071 struct btrfs_inode *inode,
2072 struct btrfs_path *path, u64 offset, u64 end)
2074 struct btrfs_fs_info *fs_info = trans->fs_info;
2075 struct btrfs_root *root = inode->root;
2076 struct extent_buffer *leaf;
2077 struct btrfs_file_extent_item *fi;
2078 struct extent_map *hole_em;
2079 struct btrfs_key key;
2082 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2085 key.objectid = btrfs_ino(inode);
2086 key.type = BTRFS_EXTENT_DATA_KEY;
2087 key.offset = offset;
2089 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2092 * We should have dropped this offset, so if we find it then
2093 * something has gone horribly wrong.
2100 leaf = path->nodes[0];
2101 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2105 fi = btrfs_item_ptr(leaf, path->slots[0],
2106 struct btrfs_file_extent_item);
2107 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2109 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2110 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2111 btrfs_set_file_extent_offset(leaf, fi, 0);
2112 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2113 btrfs_mark_buffer_dirty(trans, leaf);
2117 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2120 key.offset = offset;
2121 btrfs_set_item_key_safe(trans, path, &key);
2122 fi = btrfs_item_ptr(leaf, path->slots[0],
2123 struct btrfs_file_extent_item);
2124 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2126 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2127 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2128 btrfs_set_file_extent_offset(leaf, fi, 0);
2129 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2130 btrfs_mark_buffer_dirty(trans, leaf);
2133 btrfs_release_path(path);
2135 ret = btrfs_insert_hole_extent(trans, root, btrfs_ino(inode), offset,
2141 btrfs_release_path(path);
2143 hole_em = alloc_extent_map();
2145 btrfs_drop_extent_map_range(inode, offset, end - 1, false);
2146 btrfs_set_inode_full_sync(inode);
2148 hole_em->start = offset;
2149 hole_em->len = end - offset;
2150 hole_em->ram_bytes = hole_em->len;
2151 hole_em->orig_start = offset;
2153 hole_em->block_start = EXTENT_MAP_HOLE;
2154 hole_em->block_len = 0;
2155 hole_em->orig_block_len = 0;
2156 hole_em->generation = trans->transid;
2158 ret = btrfs_replace_extent_map_range(inode, hole_em, true);
2159 free_extent_map(hole_em);
2161 btrfs_set_inode_full_sync(inode);
2168 * Find a hole extent on given inode and change start/len to the end of hole
2169 * extent.(hole/vacuum extent whose em->start <= start &&
2170 * em->start + em->len > start)
2171 * When a hole extent is found, return 1 and modify start/len.
2173 static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2175 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2176 struct extent_map *em;
2179 em = btrfs_get_extent(inode, NULL, 0,
2180 round_down(*start, fs_info->sectorsize),
2181 round_up(*len, fs_info->sectorsize));
2185 /* Hole or vacuum extent(only exists in no-hole mode) */
2186 if (em->block_start == EXTENT_MAP_HOLE) {
2188 *len = em->start + em->len > *start + *len ?
2189 0 : *start + *len - em->start - em->len;
2190 *start = em->start + em->len;
2192 free_extent_map(em);
2196 static void btrfs_punch_hole_lock_range(struct inode *inode,
2197 const u64 lockstart,
2199 struct extent_state **cached_state)
2202 * For subpage case, if the range is not at page boundary, we could
2203 * have pages at the leading/tailing part of the range.
2204 * This could lead to dead loop since filemap_range_has_page()
2205 * will always return true.
2206 * So here we need to do extra page alignment for
2207 * filemap_range_has_page().
2209 const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2210 const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2213 truncate_pagecache_range(inode, lockstart, lockend);
2215 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2218 * We can't have ordered extents in the range, nor dirty/writeback
2219 * pages, because we have locked the inode's VFS lock in exclusive
2220 * mode, we have locked the inode's i_mmap_lock in exclusive mode,
2221 * we have flushed all delalloc in the range and we have waited
2222 * for any ordered extents in the range to complete.
2223 * We can race with anyone reading pages from this range, so after
2224 * locking the range check if we have pages in the range, and if
2225 * we do, unlock the range and retry.
2227 if (!filemap_range_has_page(inode->i_mapping, page_lockstart,
2231 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2235 btrfs_assert_inode_range_clean(BTRFS_I(inode), lockstart, lockend);
2238 static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2239 struct btrfs_inode *inode,
2240 struct btrfs_path *path,
2241 struct btrfs_replace_extent_info *extent_info,
2242 const u64 replace_len,
2243 const u64 bytes_to_drop)
2245 struct btrfs_fs_info *fs_info = trans->fs_info;
2246 struct btrfs_root *root = inode->root;
2247 struct btrfs_file_extent_item *extent;
2248 struct extent_buffer *leaf;
2249 struct btrfs_key key;
2251 struct btrfs_ref ref = { 0 };
2254 if (replace_len == 0)
2257 if (extent_info->disk_offset == 0 &&
2258 btrfs_fs_incompat(fs_info, NO_HOLES)) {
2259 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2263 key.objectid = btrfs_ino(inode);
2264 key.type = BTRFS_EXTENT_DATA_KEY;
2265 key.offset = extent_info->file_offset;
2266 ret = btrfs_insert_empty_item(trans, root, path, &key,
2267 sizeof(struct btrfs_file_extent_item));
2270 leaf = path->nodes[0];
2271 slot = path->slots[0];
2272 write_extent_buffer(leaf, extent_info->extent_buf,
2273 btrfs_item_ptr_offset(leaf, slot),
2274 sizeof(struct btrfs_file_extent_item));
2275 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2276 ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2277 btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2278 btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2279 if (extent_info->is_new_extent)
2280 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2281 btrfs_mark_buffer_dirty(trans, leaf);
2282 btrfs_release_path(path);
2284 ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2289 /* If it's a hole, nothing more needs to be done. */
2290 if (extent_info->disk_offset == 0) {
2291 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2295 btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2297 if (extent_info->is_new_extent && extent_info->insertions == 0) {
2298 key.objectid = extent_info->disk_offset;
2299 key.type = BTRFS_EXTENT_ITEM_KEY;
2300 key.offset = extent_info->disk_len;
2301 ret = btrfs_alloc_reserved_file_extent(trans, root,
2303 extent_info->file_offset,
2304 extent_info->qgroup_reserved,
2309 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2310 extent_info->disk_offset,
2311 extent_info->disk_len, 0,
2312 root->root_key.objectid);
2313 ref_offset = extent_info->file_offset - extent_info->data_offset;
2314 btrfs_init_data_ref(&ref, root->root_key.objectid,
2315 btrfs_ino(inode), ref_offset, 0, false);
2316 ret = btrfs_inc_extent_ref(trans, &ref);
2319 extent_info->insertions++;
2325 * The respective range must have been previously locked, as well as the inode.
2326 * The end offset is inclusive (last byte of the range).
2327 * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2328 * the file range with an extent.
2329 * When not punching a hole, we don't want to end up in a state where we dropped
2330 * extents without inserting a new one, so we must abort the transaction to avoid
2333 int btrfs_replace_file_extents(struct btrfs_inode *inode,
2334 struct btrfs_path *path, const u64 start,
2336 struct btrfs_replace_extent_info *extent_info,
2337 struct btrfs_trans_handle **trans_out)
2339 struct btrfs_drop_extents_args drop_args = { 0 };
2340 struct btrfs_root *root = inode->root;
2341 struct btrfs_fs_info *fs_info = root->fs_info;
2342 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2343 u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2344 struct btrfs_trans_handle *trans = NULL;
2345 struct btrfs_block_rsv *rsv;
2346 unsigned int rsv_count;
2348 u64 len = end - start;
2354 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2359 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2360 rsv->failfast = true;
2363 * 1 - update the inode
2364 * 1 - removing the extents in the range
2365 * 1 - adding the hole extent if no_holes isn't set or if we are
2366 * replacing the range with a new extent
2368 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2373 trans = btrfs_start_transaction(root, rsv_count);
2374 if (IS_ERR(trans)) {
2375 ret = PTR_ERR(trans);
2380 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2384 trans->block_rsv = rsv;
2387 drop_args.path = path;
2388 drop_args.end = end + 1;
2389 drop_args.drop_cache = true;
2390 while (cur_offset < end) {
2391 drop_args.start = cur_offset;
2392 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2393 /* If we are punching a hole decrement the inode's byte count */
2395 btrfs_update_inode_bytes(inode, 0,
2396 drop_args.bytes_found);
2397 if (ret != -ENOSPC) {
2399 * The only time we don't want to abort is if we are
2400 * attempting to clone a partial inline extent, in which
2401 * case we'll get EOPNOTSUPP. However if we aren't
2402 * clone we need to abort no matter what, because if we
2403 * got EOPNOTSUPP via prealloc then we messed up and
2407 (ret != -EOPNOTSUPP ||
2408 (extent_info && extent_info->is_new_extent)))
2409 btrfs_abort_transaction(trans, ret);
2413 trans->block_rsv = &fs_info->trans_block_rsv;
2415 if (!extent_info && cur_offset < drop_args.drop_end &&
2416 cur_offset < ino_size) {
2417 ret = fill_holes(trans, inode, path, cur_offset,
2418 drop_args.drop_end);
2421 * If we failed then we didn't insert our hole
2422 * entries for the area we dropped, so now the
2423 * fs is corrupted, so we must abort the
2426 btrfs_abort_transaction(trans, ret);
2429 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2431 * We are past the i_size here, but since we didn't
2432 * insert holes we need to clear the mapped area so we
2433 * know to not set disk_i_size in this area until a new
2434 * file extent is inserted here.
2436 ret = btrfs_inode_clear_file_extent_range(inode,
2438 drop_args.drop_end - cur_offset);
2441 * We couldn't clear our area, so we could
2442 * presumably adjust up and corrupt the fs, so
2445 btrfs_abort_transaction(trans, ret);
2451 drop_args.drop_end > extent_info->file_offset) {
2452 u64 replace_len = drop_args.drop_end -
2453 extent_info->file_offset;
2455 ret = btrfs_insert_replace_extent(trans, inode, path,
2456 extent_info, replace_len,
2457 drop_args.bytes_found);
2459 btrfs_abort_transaction(trans, ret);
2462 extent_info->data_len -= replace_len;
2463 extent_info->data_offset += replace_len;
2464 extent_info->file_offset += replace_len;
2468 * We are releasing our handle on the transaction, balance the
2469 * dirty pages of the btree inode and flush delayed items, and
2470 * then get a new transaction handle, which may now point to a
2471 * new transaction in case someone else may have committed the
2472 * transaction we used to replace/drop file extent items. So
2473 * bump the inode's iversion and update mtime and ctime except
2474 * if we are called from a dedupe context. This is because a
2475 * power failure/crash may happen after the transaction is
2476 * committed and before we finish replacing/dropping all the
2477 * file extent items we need.
2479 inode_inc_iversion(&inode->vfs_inode);
2481 if (!extent_info || extent_info->update_times)
2482 inode_set_mtime_to_ts(&inode->vfs_inode,
2483 inode_set_ctime_current(&inode->vfs_inode));
2485 ret = btrfs_update_inode(trans, inode);
2489 btrfs_end_transaction(trans);
2490 btrfs_btree_balance_dirty(fs_info);
2492 trans = btrfs_start_transaction(root, rsv_count);
2493 if (IS_ERR(trans)) {
2494 ret = PTR_ERR(trans);
2499 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2500 rsv, min_size, false);
2503 trans->block_rsv = rsv;
2505 cur_offset = drop_args.drop_end;
2506 len = end - cur_offset;
2507 if (!extent_info && len) {
2508 ret = find_first_non_hole(inode, &cur_offset, &len);
2509 if (unlikely(ret < 0))
2519 * If we were cloning, force the next fsync to be a full one since we
2520 * we replaced (or just dropped in the case of cloning holes when
2521 * NO_HOLES is enabled) file extent items and did not setup new extent
2522 * maps for the replacement extents (or holes).
2524 if (extent_info && !extent_info->is_new_extent)
2525 btrfs_set_inode_full_sync(inode);
2530 trans->block_rsv = &fs_info->trans_block_rsv;
2532 * If we are using the NO_HOLES feature we might have had already an
2533 * hole that overlaps a part of the region [lockstart, lockend] and
2534 * ends at (or beyond) lockend. Since we have no file extent items to
2535 * represent holes, drop_end can be less than lockend and so we must
2536 * make sure we have an extent map representing the existing hole (the
2537 * call to __btrfs_drop_extents() might have dropped the existing extent
2538 * map representing the existing hole), otherwise the fast fsync path
2539 * will not record the existence of the hole region
2540 * [existing_hole_start, lockend].
2542 if (drop_args.drop_end <= end)
2543 drop_args.drop_end = end + 1;
2545 * Don't insert file hole extent item if it's for a range beyond eof
2546 * (because it's useless) or if it represents a 0 bytes range (when
2547 * cur_offset == drop_end).
2549 if (!extent_info && cur_offset < ino_size &&
2550 cur_offset < drop_args.drop_end) {
2551 ret = fill_holes(trans, inode, path, cur_offset,
2552 drop_args.drop_end);
2554 /* Same comment as above. */
2555 btrfs_abort_transaction(trans, ret);
2558 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2559 /* See the comment in the loop above for the reasoning here. */
2560 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2561 drop_args.drop_end - cur_offset);
2563 btrfs_abort_transaction(trans, ret);
2569 ret = btrfs_insert_replace_extent(trans, inode, path,
2570 extent_info, extent_info->data_len,
2571 drop_args.bytes_found);
2573 btrfs_abort_transaction(trans, ret);
2582 trans->block_rsv = &fs_info->trans_block_rsv;
2584 btrfs_end_transaction(trans);
2588 btrfs_free_block_rsv(fs_info, rsv);
2593 static int btrfs_punch_hole(struct file *file, loff_t offset, loff_t len)
2595 struct inode *inode = file_inode(file);
2596 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2597 struct btrfs_root *root = BTRFS_I(inode)->root;
2598 struct extent_state *cached_state = NULL;
2599 struct btrfs_path *path;
2600 struct btrfs_trans_handle *trans = NULL;
2605 u64 orig_start = offset;
2609 bool truncated_block = false;
2610 bool updated_inode = false;
2612 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2614 ret = btrfs_wait_ordered_range(inode, offset, len);
2616 goto out_only_mutex;
2618 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2619 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2621 goto out_only_mutex;
2623 /* Already in a large hole */
2625 goto out_only_mutex;
2628 ret = file_modified(file);
2630 goto out_only_mutex;
2632 lockstart = round_up(offset, fs_info->sectorsize);
2633 lockend = round_down(offset + len, fs_info->sectorsize) - 1;
2634 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2635 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2637 * We needn't truncate any block which is beyond the end of the file
2638 * because we are sure there is no data there.
2641 * Only do this if we are in the same block and we aren't doing the
2644 if (same_block && len < fs_info->sectorsize) {
2645 if (offset < ino_size) {
2646 truncated_block = true;
2647 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2652 goto out_only_mutex;
2655 /* zero back part of the first block */
2656 if (offset < ino_size) {
2657 truncated_block = true;
2658 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2660 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2665 /* Check the aligned pages after the first unaligned page,
2666 * if offset != orig_start, which means the first unaligned page
2667 * including several following pages are already in holes,
2668 * the extra check can be skipped */
2669 if (offset == orig_start) {
2670 /* after truncate page, check hole again */
2671 len = offset + len - lockstart;
2673 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2675 goto out_only_mutex;
2678 goto out_only_mutex;
2683 /* Check the tail unaligned part is in a hole */
2684 tail_start = lockend + 1;
2685 tail_len = offset + len - tail_start;
2687 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2688 if (unlikely(ret < 0))
2689 goto out_only_mutex;
2691 /* zero the front end of the last page */
2692 if (tail_start + tail_len < ino_size) {
2693 truncated_block = true;
2694 ret = btrfs_truncate_block(BTRFS_I(inode),
2695 tail_start + tail_len,
2698 goto out_only_mutex;
2703 if (lockend < lockstart) {
2705 goto out_only_mutex;
2708 btrfs_punch_hole_lock_range(inode, lockstart, lockend, &cached_state);
2710 path = btrfs_alloc_path();
2716 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
2717 lockend, NULL, &trans);
2718 btrfs_free_path(path);
2722 ASSERT(trans != NULL);
2723 inode_inc_iversion(inode);
2724 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
2725 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2726 updated_inode = true;
2727 btrfs_end_transaction(trans);
2728 btrfs_btree_balance_dirty(fs_info);
2730 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2733 if (!updated_inode && truncated_block && !ret) {
2735 * If we only end up zeroing part of a page, we still need to
2736 * update the inode item, so that all the time fields are
2737 * updated as well as the necessary btrfs inode in memory fields
2738 * for detecting, at fsync time, if the inode isn't yet in the
2739 * log tree or it's there but not up to date.
2741 struct timespec64 now = inode_set_ctime_current(inode);
2743 inode_inc_iversion(inode);
2744 inode_set_mtime_to_ts(inode, now);
2745 trans = btrfs_start_transaction(root, 1);
2746 if (IS_ERR(trans)) {
2747 ret = PTR_ERR(trans);
2751 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2752 ret2 = btrfs_end_transaction(trans);
2757 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2761 /* Helper structure to record which range is already reserved */
2762 struct falloc_range {
2763 struct list_head list;
2769 * Helper function to add falloc range
2771 * Caller should have locked the larger range of extent containing
2774 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2776 struct falloc_range *range = NULL;
2778 if (!list_empty(head)) {
2780 * As fallocate iterates by bytenr order, we only need to check
2783 range = list_last_entry(head, struct falloc_range, list);
2784 if (range->start + range->len == start) {
2790 range = kmalloc(sizeof(*range), GFP_KERNEL);
2793 range->start = start;
2795 list_add_tail(&range->list, head);
2799 static int btrfs_fallocate_update_isize(struct inode *inode,
2803 struct btrfs_trans_handle *trans;
2804 struct btrfs_root *root = BTRFS_I(inode)->root;
2808 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2811 trans = btrfs_start_transaction(root, 1);
2813 return PTR_ERR(trans);
2815 inode_set_ctime_current(inode);
2816 i_size_write(inode, end);
2817 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
2818 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2819 ret2 = btrfs_end_transaction(trans);
2821 return ret ? ret : ret2;
2825 RANGE_BOUNDARY_WRITTEN_EXTENT,
2826 RANGE_BOUNDARY_PREALLOC_EXTENT,
2827 RANGE_BOUNDARY_HOLE,
2830 static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
2833 const u64 sectorsize = inode->root->fs_info->sectorsize;
2834 struct extent_map *em;
2837 offset = round_down(offset, sectorsize);
2838 em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
2842 if (em->block_start == EXTENT_MAP_HOLE)
2843 ret = RANGE_BOUNDARY_HOLE;
2844 else if (em->flags & EXTENT_FLAG_PREALLOC)
2845 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2847 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2849 free_extent_map(em);
2853 static int btrfs_zero_range(struct inode *inode,
2858 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2859 struct extent_map *em;
2860 struct extent_changeset *data_reserved = NULL;
2863 const u64 sectorsize = fs_info->sectorsize;
2864 u64 alloc_start = round_down(offset, sectorsize);
2865 u64 alloc_end = round_up(offset + len, sectorsize);
2866 u64 bytes_to_reserve = 0;
2867 bool space_reserved = false;
2869 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
2870 alloc_end - alloc_start);
2877 * Avoid hole punching and extent allocation for some cases. More cases
2878 * could be considered, but these are unlikely common and we keep things
2879 * as simple as possible for now. Also, intentionally, if the target
2880 * range contains one or more prealloc extents together with regular
2881 * extents and holes, we drop all the existing extents and allocate a
2882 * new prealloc extent, so that we get a larger contiguous disk extent.
2884 if (em->start <= alloc_start && (em->flags & EXTENT_FLAG_PREALLOC)) {
2885 const u64 em_end = em->start + em->len;
2887 if (em_end >= offset + len) {
2889 * The whole range is already a prealloc extent,
2890 * do nothing except updating the inode's i_size if
2893 free_extent_map(em);
2894 ret = btrfs_fallocate_update_isize(inode, offset + len,
2899 * Part of the range is already a prealloc extent, so operate
2900 * only on the remaining part of the range.
2902 alloc_start = em_end;
2903 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2904 len = offset + len - alloc_start;
2905 offset = alloc_start;
2906 alloc_hint = em->block_start + em->len;
2908 free_extent_map(em);
2910 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2911 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2912 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
2919 if (em->flags & EXTENT_FLAG_PREALLOC) {
2920 free_extent_map(em);
2921 ret = btrfs_fallocate_update_isize(inode, offset + len,
2925 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2926 free_extent_map(em);
2927 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2930 ret = btrfs_fallocate_update_isize(inode,
2935 free_extent_map(em);
2936 alloc_start = round_down(offset, sectorsize);
2937 alloc_end = alloc_start + sectorsize;
2941 alloc_start = round_up(offset, sectorsize);
2942 alloc_end = round_down(offset + len, sectorsize);
2945 * For unaligned ranges, check the pages at the boundaries, they might
2946 * map to an extent, in which case we need to partially zero them, or
2947 * they might map to a hole, in which case we need our allocation range
2950 if (!IS_ALIGNED(offset, sectorsize)) {
2951 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2955 if (ret == RANGE_BOUNDARY_HOLE) {
2956 alloc_start = round_down(offset, sectorsize);
2958 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2959 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2967 if (!IS_ALIGNED(offset + len, sectorsize)) {
2968 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2972 if (ret == RANGE_BOUNDARY_HOLE) {
2973 alloc_end = round_up(offset + len, sectorsize);
2975 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2976 ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
2986 if (alloc_start < alloc_end) {
2987 struct extent_state *cached_state = NULL;
2988 const u64 lockstart = alloc_start;
2989 const u64 lockend = alloc_end - 1;
2991 bytes_to_reserve = alloc_end - alloc_start;
2992 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2996 space_reserved = true;
2997 btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2999 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3000 alloc_start, bytes_to_reserve);
3002 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
3003 lockend, &cached_state);
3006 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3007 alloc_end - alloc_start,
3009 offset + len, &alloc_hint);
3010 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3012 /* btrfs_prealloc_file_range releases reserved space on error */
3014 space_reserved = false;
3018 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3020 if (ret && space_reserved)
3021 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3022 alloc_start, bytes_to_reserve);
3023 extent_changeset_free(data_reserved);
3028 static long btrfs_fallocate(struct file *file, int mode,
3029 loff_t offset, loff_t len)
3031 struct inode *inode = file_inode(file);
3032 struct extent_state *cached_state = NULL;
3033 struct extent_changeset *data_reserved = NULL;
3034 struct falloc_range *range;
3035 struct falloc_range *tmp;
3036 LIST_HEAD(reserve_list);
3044 u64 data_space_needed = 0;
3045 u64 data_space_reserved = 0;
3046 u64 qgroup_reserved = 0;
3047 struct extent_map *em;
3048 int blocksize = BTRFS_I(inode)->root->fs_info->sectorsize;
3051 /* Do not allow fallocate in ZONED mode */
3052 if (btrfs_is_zoned(btrfs_sb(inode->i_sb)))
3055 alloc_start = round_down(offset, blocksize);
3056 alloc_end = round_up(offset + len, blocksize);
3057 cur_offset = alloc_start;
3059 /* Make sure we aren't being give some crap mode */
3060 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3061 FALLOC_FL_ZERO_RANGE))
3064 if (mode & FALLOC_FL_PUNCH_HOLE)
3065 return btrfs_punch_hole(file, offset, len);
3067 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3069 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3070 ret = inode_newsize_ok(inode, offset + len);
3075 ret = file_modified(file);
3080 * TODO: Move these two operations after we have checked
3081 * accurate reserved space, or fallocate can still fail but
3082 * with page truncated or size expanded.
3084 * But that's a minor problem and won't do much harm BTW.
3086 if (alloc_start > inode->i_size) {
3087 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3091 } else if (offset + len > inode->i_size) {
3093 * If we are fallocating from the end of the file onward we
3094 * need to zero out the end of the block if i_size lands in the
3095 * middle of a block.
3097 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3103 * We have locked the inode at the VFS level (in exclusive mode) and we
3104 * have locked the i_mmap_lock lock (in exclusive mode). Now before
3105 * locking the file range, flush all dealloc in the range and wait for
3106 * all ordered extents in the range to complete. After this we can lock
3107 * the file range and, due to the previous locking we did, we know there
3108 * can't be more delalloc or ordered extents in the range.
3110 ret = btrfs_wait_ordered_range(inode, alloc_start,
3111 alloc_end - alloc_start);
3115 if (mode & FALLOC_FL_ZERO_RANGE) {
3116 ret = btrfs_zero_range(inode, offset, len, mode);
3117 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3121 locked_end = alloc_end - 1;
3122 lock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3125 btrfs_assert_inode_range_clean(BTRFS_I(inode), alloc_start, locked_end);
3127 /* First, check if we exceed the qgroup limit */
3128 while (cur_offset < alloc_end) {
3129 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3130 alloc_end - cur_offset);
3135 last_byte = min(extent_map_end(em), alloc_end);
3136 actual_end = min_t(u64, extent_map_end(em), offset + len);
3137 last_byte = ALIGN(last_byte, blocksize);
3138 if (em->block_start == EXTENT_MAP_HOLE ||
3139 (cur_offset >= inode->i_size &&
3140 !(em->flags & EXTENT_FLAG_PREALLOC))) {
3141 const u64 range_len = last_byte - cur_offset;
3143 ret = add_falloc_range(&reserve_list, cur_offset, range_len);
3145 free_extent_map(em);
3148 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3149 &data_reserved, cur_offset, range_len);
3151 free_extent_map(em);
3154 qgroup_reserved += range_len;
3155 data_space_needed += range_len;
3157 free_extent_map(em);
3158 cur_offset = last_byte;
3161 if (!ret && data_space_needed > 0) {
3163 * We are safe to reserve space here as we can't have delalloc
3164 * in the range, see above.
3166 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3169 data_space_reserved = data_space_needed;
3173 * If ret is still 0, means we're OK to fallocate.
3174 * Or just cleanup the list and exit.
3176 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3178 ret = btrfs_prealloc_file_range(inode, mode,
3180 range->len, i_blocksize(inode),
3181 offset + len, &alloc_hint);
3183 * btrfs_prealloc_file_range() releases space even
3184 * if it returns an error.
3186 data_space_reserved -= range->len;
3187 qgroup_reserved -= range->len;
3188 } else if (data_space_reserved > 0) {
3189 btrfs_free_reserved_data_space(BTRFS_I(inode),
3190 data_reserved, range->start,
3192 data_space_reserved -= range->len;
3193 qgroup_reserved -= range->len;
3194 } else if (qgroup_reserved > 0) {
3195 btrfs_qgroup_free_data(BTRFS_I(inode), data_reserved,
3196 range->start, range->len, NULL);
3197 qgroup_reserved -= range->len;
3199 list_del(&range->list);
3206 * We didn't need to allocate any more space, but we still extended the
3207 * size of the file so we need to update i_size and the inode item.
3209 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3211 unlock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3214 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3215 extent_changeset_free(data_reserved);
3220 * Helper for btrfs_find_delalloc_in_range(). Find a subrange in a given range
3221 * that has unflushed and/or flushing delalloc. There might be other adjacent
3222 * subranges after the one it found, so btrfs_find_delalloc_in_range() keeps
3223 * looping while it gets adjacent subranges, and merging them together.
3225 static bool find_delalloc_subrange(struct btrfs_inode *inode, u64 start, u64 end,
3226 struct extent_state **cached_state,
3227 bool *search_io_tree,
3228 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3230 u64 len = end + 1 - start;
3231 u64 delalloc_len = 0;
3232 struct btrfs_ordered_extent *oe;
3237 * Search the io tree first for EXTENT_DELALLOC. If we find any, it
3238 * means we have delalloc (dirty pages) for which writeback has not
3241 if (*search_io_tree) {
3242 spin_lock(&inode->lock);
3243 if (inode->delalloc_bytes > 0) {
3244 spin_unlock(&inode->lock);
3245 *delalloc_start_ret = start;
3246 delalloc_len = count_range_bits(&inode->io_tree,
3247 delalloc_start_ret, end,
3248 len, EXTENT_DELALLOC, 1,
3251 spin_unlock(&inode->lock);
3255 if (delalloc_len > 0) {
3257 * If delalloc was found then *delalloc_start_ret has a sector size
3258 * aligned value (rounded down).
3260 *delalloc_end_ret = *delalloc_start_ret + delalloc_len - 1;
3262 if (*delalloc_start_ret == start) {
3263 /* Delalloc for the whole range, nothing more to do. */
3264 if (*delalloc_end_ret == end)
3266 /* Else trim our search range for ordered extents. */
3267 start = *delalloc_end_ret + 1;
3268 len = end + 1 - start;
3271 /* No delalloc, future calls don't need to search again. */
3272 *search_io_tree = false;
3276 * Now also check if there's any ordered extent in the range.
3277 * We do this because:
3279 * 1) When delalloc is flushed, the file range is locked, we clear the
3280 * EXTENT_DELALLOC bit from the io tree and create an extent map and
3281 * an ordered extent for the write. So we might just have been called
3282 * after delalloc is flushed and before the ordered extent completes
3283 * and inserts the new file extent item in the subvolume's btree;
3285 * 2) We may have an ordered extent created by flushing delalloc for a
3286 * subrange that starts before the subrange we found marked with
3287 * EXTENT_DELALLOC in the io tree.
3289 * We could also use the extent map tree to find such delalloc that is
3290 * being flushed, but using the ordered extents tree is more efficient
3291 * because it's usually much smaller as ordered extents are removed from
3292 * the tree once they complete. With the extent maps, we mau have them
3293 * in the extent map tree for a very long time, and they were either
3294 * created by previous writes or loaded by read operations.
3296 oe = btrfs_lookup_first_ordered_range(inode, start, len);
3298 return (delalloc_len > 0);
3300 /* The ordered extent may span beyond our search range. */
3301 oe_start = max(oe->file_offset, start);
3302 oe_end = min(oe->file_offset + oe->num_bytes - 1, end);
3304 btrfs_put_ordered_extent(oe);
3306 /* Don't have unflushed delalloc, return the ordered extent range. */
3307 if (delalloc_len == 0) {
3308 *delalloc_start_ret = oe_start;
3309 *delalloc_end_ret = oe_end;
3314 * We have both unflushed delalloc (io_tree) and an ordered extent.
3315 * If the ranges are adjacent returned a combined range, otherwise
3316 * return the leftmost range.
3318 if (oe_start < *delalloc_start_ret) {
3319 if (oe_end < *delalloc_start_ret)
3320 *delalloc_end_ret = oe_end;
3321 *delalloc_start_ret = oe_start;
3322 } else if (*delalloc_end_ret + 1 == oe_start) {
3323 *delalloc_end_ret = oe_end;
3330 * Check if there's delalloc in a given range.
3332 * @inode: The inode.
3333 * @start: The start offset of the range. It does not need to be
3334 * sector size aligned.
3335 * @end: The end offset (inclusive value) of the search range.
3336 * It does not need to be sector size aligned.
3337 * @cached_state: Extent state record used for speeding up delalloc
3338 * searches in the inode's io_tree. Can be NULL.
3339 * @delalloc_start_ret: Output argument, set to the start offset of the
3340 * subrange found with delalloc (may not be sector size
3342 * @delalloc_end_ret: Output argument, set to he end offset (inclusive value)
3343 * of the subrange found with delalloc.
3345 * Returns true if a subrange with delalloc is found within the given range, and
3346 * if so it sets @delalloc_start_ret and @delalloc_end_ret with the start and
3347 * end offsets of the subrange.
3349 bool btrfs_find_delalloc_in_range(struct btrfs_inode *inode, u64 start, u64 end,
3350 struct extent_state **cached_state,
3351 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3353 u64 cur_offset = round_down(start, inode->root->fs_info->sectorsize);
3354 u64 prev_delalloc_end = 0;
3355 bool search_io_tree = true;
3358 while (cur_offset <= end) {
3363 delalloc = find_delalloc_subrange(inode, cur_offset, end,
3364 cached_state, &search_io_tree,
3370 if (prev_delalloc_end == 0) {
3371 /* First subrange found. */
3372 *delalloc_start_ret = max(delalloc_start, start);
3373 *delalloc_end_ret = delalloc_end;
3375 } else if (delalloc_start == prev_delalloc_end + 1) {
3376 /* Subrange adjacent to the previous one, merge them. */
3377 *delalloc_end_ret = delalloc_end;
3379 /* Subrange not adjacent to the previous one, exit. */
3383 prev_delalloc_end = delalloc_end;
3384 cur_offset = delalloc_end + 1;
3392 * Check if there's a hole or delalloc range in a range representing a hole (or
3393 * prealloc extent) found in the inode's subvolume btree.
3395 * @inode: The inode.
3396 * @whence: Seek mode (SEEK_DATA or SEEK_HOLE).
3397 * @start: Start offset of the hole region. It does not need to be sector
3399 * @end: End offset (inclusive value) of the hole region. It does not
3400 * need to be sector size aligned.
3401 * @start_ret: Return parameter, used to set the start of the subrange in the
3402 * hole that matches the search criteria (seek mode), if such
3403 * subrange is found (return value of the function is true).
3404 * The value returned here may not be sector size aligned.
3406 * Returns true if a subrange matching the given seek mode is found, and if one
3407 * is found, it updates @start_ret with the start of the subrange.
3409 static bool find_desired_extent_in_hole(struct btrfs_inode *inode, int whence,
3410 struct extent_state **cached_state,
3411 u64 start, u64 end, u64 *start_ret)
3417 delalloc = btrfs_find_delalloc_in_range(inode, start, end, cached_state,
3418 &delalloc_start, &delalloc_end);
3419 if (delalloc && whence == SEEK_DATA) {
3420 *start_ret = delalloc_start;
3424 if (delalloc && whence == SEEK_HOLE) {
3426 * We found delalloc but it starts after out start offset. So we
3427 * have a hole between our start offset and the delalloc start.
3429 if (start < delalloc_start) {
3434 * Delalloc range starts at our start offset.
3435 * If the delalloc range's length is smaller than our range,
3436 * then it means we have a hole that starts where the delalloc
3439 if (delalloc_end < end) {
3440 *start_ret = delalloc_end + 1;
3444 /* There's delalloc for the whole range. */
3448 if (!delalloc && whence == SEEK_HOLE) {
3454 * No delalloc in the range and we are seeking for data. The caller has
3455 * to iterate to the next extent item in the subvolume btree.
3460 static loff_t find_desired_extent(struct file *file, loff_t offset, int whence)
3462 struct btrfs_inode *inode = BTRFS_I(file->f_mapping->host);
3463 struct btrfs_file_private *private = file->private_data;
3464 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3465 struct extent_state *cached_state = NULL;
3466 struct extent_state **delalloc_cached_state;
3467 const loff_t i_size = i_size_read(&inode->vfs_inode);
3468 const u64 ino = btrfs_ino(inode);
3469 struct btrfs_root *root = inode->root;
3470 struct btrfs_path *path;
3471 struct btrfs_key key;
3472 u64 last_extent_end;
3479 if (i_size == 0 || offset >= i_size)
3483 * Quick path. If the inode has no prealloc extents and its number of
3484 * bytes used matches its i_size, then it can not have holes.
3486 if (whence == SEEK_HOLE &&
3487 !(inode->flags & BTRFS_INODE_PREALLOC) &&
3488 inode_get_bytes(&inode->vfs_inode) == i_size)
3492 private = kzalloc(sizeof(*private), GFP_KERNEL);
3494 * No worries if memory allocation failed.
3495 * The private structure is used only for speeding up multiple
3496 * lseek SEEK_HOLE/DATA calls to a file when there's delalloc,
3497 * so everything will still be correct.
3499 file->private_data = private;
3503 delalloc_cached_state = &private->llseek_cached_state;
3505 delalloc_cached_state = NULL;
3508 * offset can be negative, in this case we start finding DATA/HOLE from
3509 * the very start of the file.
3511 start = max_t(loff_t, 0, offset);
3513 lockstart = round_down(start, fs_info->sectorsize);
3514 lockend = round_up(i_size, fs_info->sectorsize);
3515 if (lockend <= lockstart)
3516 lockend = lockstart + fs_info->sectorsize;
3519 path = btrfs_alloc_path();
3522 path->reada = READA_FORWARD;
3525 key.type = BTRFS_EXTENT_DATA_KEY;
3528 last_extent_end = lockstart;
3530 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3532 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3535 } else if (ret > 0 && path->slots[0] > 0) {
3536 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
3537 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
3541 while (start < i_size) {
3542 struct extent_buffer *leaf = path->nodes[0];
3543 struct btrfs_file_extent_item *extent;
3547 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3548 ret = btrfs_next_leaf(root, path);
3554 leaf = path->nodes[0];
3557 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3558 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3561 extent_end = btrfs_file_extent_end(path);
3564 * In the first iteration we may have a slot that points to an
3565 * extent that ends before our start offset, so skip it.
3567 if (extent_end <= start) {
3572 /* We have an implicit hole, NO_HOLES feature is likely set. */
3573 if (last_extent_end < key.offset) {
3574 u64 search_start = last_extent_end;
3578 * First iteration, @start matches @offset and it's
3581 if (start == offset)
3582 search_start = offset;
3584 found = find_desired_extent_in_hole(inode, whence,
3585 delalloc_cached_state,
3590 start = found_start;
3594 * Didn't find data or a hole (due to delalloc) in the
3595 * implicit hole range, so need to analyze the extent.
3599 extent = btrfs_item_ptr(leaf, path->slots[0],
3600 struct btrfs_file_extent_item);
3601 type = btrfs_file_extent_type(leaf, extent);
3604 * Can't access the extent's disk_bytenr field if this is an
3605 * inline extent, since at that offset, it's where the extent
3608 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
3609 (type == BTRFS_FILE_EXTENT_REG &&
3610 btrfs_file_extent_disk_bytenr(leaf, extent) == 0)) {
3612 * Explicit hole or prealloc extent, search for delalloc.
3613 * A prealloc extent is treated like a hole.
3615 u64 search_start = key.offset;
3619 * First iteration, @start matches @offset and it's
3622 if (start == offset)
3623 search_start = offset;
3625 found = find_desired_extent_in_hole(inode, whence,
3626 delalloc_cached_state,
3631 start = found_start;
3635 * Didn't find data or a hole (due to delalloc) in the
3636 * implicit hole range, so need to analyze the next
3641 * Found a regular or inline extent.
3642 * If we are seeking for data, adjust the start offset
3643 * and stop, we're done.
3645 if (whence == SEEK_DATA) {
3646 start = max_t(u64, key.offset, offset);
3651 * Else, we are seeking for a hole, check the next file
3657 last_extent_end = extent_end;
3659 if (fatal_signal_pending(current)) {
3666 /* We have an implicit hole from the last extent found up to i_size. */
3667 if (!found && start < i_size) {
3668 found = find_desired_extent_in_hole(inode, whence,
3669 delalloc_cached_state, start,
3670 i_size - 1, &start);
3676 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3677 btrfs_free_path(path);
3682 if (whence == SEEK_DATA && start >= i_size)
3685 return min_t(loff_t, start, i_size);
3688 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3690 struct inode *inode = file->f_mapping->host;
3694 return generic_file_llseek(file, offset, whence);
3697 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3698 offset = find_desired_extent(file, offset, whence);
3699 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3706 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3709 static int btrfs_file_open(struct inode *inode, struct file *filp)
3713 filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC | FMODE_BUF_WASYNC |
3716 ret = fsverity_file_open(inode, filp);
3719 return generic_file_open(inode, filp);
3722 static int check_direct_read(struct btrfs_fs_info *fs_info,
3723 const struct iov_iter *iter, loff_t offset)
3728 ret = check_direct_IO(fs_info, iter, offset);
3732 if (!iter_is_iovec(iter))
3735 for (seg = 0; seg < iter->nr_segs; seg++) {
3736 for (i = seg + 1; i < iter->nr_segs; i++) {
3737 const struct iovec *iov1 = iter_iov(iter) + seg;
3738 const struct iovec *iov2 = iter_iov(iter) + i;
3740 if (iov1->iov_base == iov2->iov_base)
3747 static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3749 struct inode *inode = file_inode(iocb->ki_filp);
3750 size_t prev_left = 0;
3754 if (fsverity_active(inode))
3757 if (check_direct_read(btrfs_sb(inode->i_sb), to, iocb->ki_pos))
3760 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3763 * This is similar to what we do for direct IO writes, see the comment
3764 * at btrfs_direct_write(), but we also disable page faults in addition
3765 * to disabling them only at the iov_iter level. This is because when
3766 * reading from a hole or prealloc extent, iomap calls iov_iter_zero(),
3767 * which can still trigger page fault ins despite having set ->nofault
3768 * to true of our 'to' iov_iter.
3770 * The difference to direct IO writes is that we deadlock when trying
3771 * to lock the extent range in the inode's tree during he page reads
3772 * triggered by the fault in (while for writes it is due to waiting for
3773 * our own ordered extent). This is because for direct IO reads,
3774 * btrfs_dio_iomap_begin() returns with the extent range locked, which
3775 * is only unlocked in the endio callback (end_bio_extent_readpage()).
3777 pagefault_disable();
3779 ret = btrfs_dio_read(iocb, to, read);
3780 to->nofault = false;
3783 /* No increment (+=) because iomap returns a cumulative value. */
3787 if (iov_iter_count(to) > 0 && (ret == -EFAULT || ret > 0)) {
3788 const size_t left = iov_iter_count(to);
3790 if (left == prev_left) {
3792 * We didn't make any progress since the last attempt,
3793 * fallback to a buffered read for the remainder of the
3794 * range. This is just to avoid any possibility of looping
3800 * We made some progress since the last retry or this is
3801 * the first time we are retrying. Fault in as many pages
3802 * as possible and retry.
3804 fault_in_iov_iter_writeable(to, left);
3809 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3810 return ret < 0 ? ret : read;
3813 static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3817 if (iocb->ki_flags & IOCB_DIRECT) {
3818 ret = btrfs_direct_read(iocb, to);
3819 if (ret < 0 || !iov_iter_count(to) ||
3820 iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3824 return filemap_read(iocb, to, ret);
3827 const struct file_operations btrfs_file_operations = {
3828 .llseek = btrfs_file_llseek,
3829 .read_iter = btrfs_file_read_iter,
3830 .splice_read = filemap_splice_read,
3831 .write_iter = btrfs_file_write_iter,
3832 .splice_write = iter_file_splice_write,
3833 .mmap = btrfs_file_mmap,
3834 .open = btrfs_file_open,
3835 .release = btrfs_release_file,
3836 .get_unmapped_area = thp_get_unmapped_area,
3837 .fsync = btrfs_sync_file,
3838 .fallocate = btrfs_fallocate,
3839 .unlocked_ioctl = btrfs_ioctl,
3840 #ifdef CONFIG_COMPAT
3841 .compat_ioctl = btrfs_compat_ioctl,
3843 .remap_file_range = btrfs_remap_file_range,
3846 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3851 * So with compression we will find and lock a dirty page and clear the
3852 * first one as dirty, setup an async extent, and immediately return
3853 * with the entire range locked but with nobody actually marked with
3854 * writeback. So we can't just filemap_write_and_wait_range() and
3855 * expect it to work since it will just kick off a thread to do the
3856 * actual work. So we need to call filemap_fdatawrite_range _again_
3857 * since it will wait on the page lock, which won't be unlocked until
3858 * after the pages have been marked as writeback and so we're good to go
3859 * from there. We have to do this otherwise we'll miss the ordered
3860 * extents and that results in badness. Please Josef, do not think you
3861 * know better and pull this out at some point in the future, it is
3862 * right and you are wrong.
3864 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3865 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3866 &BTRFS_I(inode)->runtime_flags))
3867 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);