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>
21 #include "transaction.h"
22 #include "btrfs_inode.h"
23 #include "print-tree.h"
28 #include "compression.h"
29 #include "delalloc-space.h"
31 static struct kmem_cache *btrfs_inode_defrag_cachep;
33 * when auto defrag is enabled we
34 * queue up these defrag structs to remember which
35 * inodes need defragging passes
38 struct rb_node rb_node;
42 * transid where the defrag was added, we search for
43 * extents newer than this
50 /* last offset we were able to defrag */
53 /* if we've wrapped around back to zero once already */
57 static int __compare_inode_defrag(struct inode_defrag *defrag1,
58 struct inode_defrag *defrag2)
60 if (defrag1->root > defrag2->root)
62 else if (defrag1->root < defrag2->root)
64 else if (defrag1->ino > defrag2->ino)
66 else if (defrag1->ino < defrag2->ino)
72 /* pop a record for an inode into the defrag tree. The lock
73 * must be held already
75 * If you're inserting a record for an older transid than an
76 * existing record, the transid already in the tree is lowered
78 * If an existing record is found the defrag item you
81 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
82 struct inode_defrag *defrag)
84 struct btrfs_fs_info *fs_info = inode->root->fs_info;
85 struct inode_defrag *entry;
87 struct rb_node *parent = NULL;
90 p = &fs_info->defrag_inodes.rb_node;
93 entry = rb_entry(parent, struct inode_defrag, rb_node);
95 ret = __compare_inode_defrag(defrag, entry);
99 p = &parent->rb_right;
101 /* if we're reinserting an entry for
102 * an old defrag run, make sure to
103 * lower the transid of our existing record
105 if (defrag->transid < entry->transid)
106 entry->transid = defrag->transid;
107 if (defrag->last_offset > entry->last_offset)
108 entry->last_offset = defrag->last_offset;
112 set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
113 rb_link_node(&defrag->rb_node, parent, p);
114 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
118 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
120 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
123 if (btrfs_fs_closing(fs_info))
130 * insert a defrag record for this inode if auto defrag is
133 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
134 struct btrfs_inode *inode)
136 struct btrfs_root *root = inode->root;
137 struct btrfs_fs_info *fs_info = root->fs_info;
138 struct inode_defrag *defrag;
142 if (!__need_auto_defrag(fs_info))
145 if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
149 transid = trans->transid;
151 transid = inode->root->last_trans;
153 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
157 defrag->ino = btrfs_ino(inode);
158 defrag->transid = transid;
159 defrag->root = root->root_key.objectid;
161 spin_lock(&fs_info->defrag_inodes_lock);
162 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
164 * If we set IN_DEFRAG flag and evict the inode from memory,
165 * and then re-read this inode, this new inode doesn't have
166 * IN_DEFRAG flag. At the case, we may find the existed defrag.
168 ret = __btrfs_add_inode_defrag(inode, defrag);
170 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
172 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
174 spin_unlock(&fs_info->defrag_inodes_lock);
179 * Requeue the defrag object. If there is a defrag object that points to
180 * the same inode in the tree, we will merge them together (by
181 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
183 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
184 struct inode_defrag *defrag)
186 struct btrfs_fs_info *fs_info = inode->root->fs_info;
189 if (!__need_auto_defrag(fs_info))
193 * Here we don't check the IN_DEFRAG flag, because we need merge
196 spin_lock(&fs_info->defrag_inodes_lock);
197 ret = __btrfs_add_inode_defrag(inode, defrag);
198 spin_unlock(&fs_info->defrag_inodes_lock);
203 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
207 * pick the defragable inode that we want, if it doesn't exist, we will get
210 static struct inode_defrag *
211 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
213 struct inode_defrag *entry = NULL;
214 struct inode_defrag tmp;
216 struct rb_node *parent = NULL;
222 spin_lock(&fs_info->defrag_inodes_lock);
223 p = fs_info->defrag_inodes.rb_node;
226 entry = rb_entry(parent, struct inode_defrag, rb_node);
228 ret = __compare_inode_defrag(&tmp, entry);
232 p = parent->rb_right;
237 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
238 parent = rb_next(parent);
240 entry = rb_entry(parent, struct inode_defrag, rb_node);
246 rb_erase(parent, &fs_info->defrag_inodes);
247 spin_unlock(&fs_info->defrag_inodes_lock);
251 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
253 struct inode_defrag *defrag;
254 struct rb_node *node;
256 spin_lock(&fs_info->defrag_inodes_lock);
257 node = rb_first(&fs_info->defrag_inodes);
259 rb_erase(node, &fs_info->defrag_inodes);
260 defrag = rb_entry(node, struct inode_defrag, rb_node);
261 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
263 cond_resched_lock(&fs_info->defrag_inodes_lock);
265 node = rb_first(&fs_info->defrag_inodes);
267 spin_unlock(&fs_info->defrag_inodes_lock);
270 #define BTRFS_DEFRAG_BATCH 1024
272 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
273 struct inode_defrag *defrag)
275 struct btrfs_root *inode_root;
277 struct btrfs_key key;
278 struct btrfs_ioctl_defrag_range_args range;
284 key.objectid = defrag->root;
285 key.type = BTRFS_ROOT_ITEM_KEY;
286 key.offset = (u64)-1;
288 index = srcu_read_lock(&fs_info->subvol_srcu);
290 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
291 if (IS_ERR(inode_root)) {
292 ret = PTR_ERR(inode_root);
296 key.objectid = defrag->ino;
297 key.type = BTRFS_INODE_ITEM_KEY;
299 inode = btrfs_iget(fs_info->sb, &key, inode_root);
301 ret = PTR_ERR(inode);
304 srcu_read_unlock(&fs_info->subvol_srcu, index);
306 /* do a chunk of defrag */
307 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
308 memset(&range, 0, sizeof(range));
310 range.start = defrag->last_offset;
312 sb_start_write(fs_info->sb);
313 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
315 sb_end_write(fs_info->sb);
317 * if we filled the whole defrag batch, there
318 * must be more work to do. Queue this defrag
321 if (num_defrag == BTRFS_DEFRAG_BATCH) {
322 defrag->last_offset = range.start;
323 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
324 } else if (defrag->last_offset && !defrag->cycled) {
326 * we didn't fill our defrag batch, but
327 * we didn't start at zero. Make sure we loop
328 * around to the start of the file.
330 defrag->last_offset = 0;
332 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
334 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
340 srcu_read_unlock(&fs_info->subvol_srcu, index);
341 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
346 * run through the list of inodes in the FS that need
349 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
351 struct inode_defrag *defrag;
353 u64 root_objectid = 0;
355 atomic_inc(&fs_info->defrag_running);
357 /* Pause the auto defragger. */
358 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
362 if (!__need_auto_defrag(fs_info))
365 /* find an inode to defrag */
366 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
369 if (root_objectid || first_ino) {
378 first_ino = defrag->ino + 1;
379 root_objectid = defrag->root;
381 __btrfs_run_defrag_inode(fs_info, defrag);
383 atomic_dec(&fs_info->defrag_running);
386 * during unmount, we use the transaction_wait queue to
387 * wait for the defragger to stop
389 wake_up(&fs_info->transaction_wait);
393 /* simple helper to fault in pages and copy. This should go away
394 * and be replaced with calls into generic code.
396 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
397 struct page **prepared_pages,
401 size_t total_copied = 0;
403 int offset = offset_in_page(pos);
405 while (write_bytes > 0) {
406 size_t count = min_t(size_t,
407 PAGE_SIZE - offset, write_bytes);
408 struct page *page = prepared_pages[pg];
410 * Copy data from userspace to the current page
412 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
414 /* Flush processor's dcache for this page */
415 flush_dcache_page(page);
418 * if we get a partial write, we can end up with
419 * partially up to date pages. These add
420 * a lot of complexity, so make sure they don't
421 * happen by forcing this copy to be retried.
423 * The rest of the btrfs_file_write code will fall
424 * back to page at a time copies after we return 0.
426 if (!PageUptodate(page) && copied < count)
429 iov_iter_advance(i, copied);
430 write_bytes -= copied;
431 total_copied += copied;
433 /* Return to btrfs_file_write_iter to fault page */
434 if (unlikely(copied == 0))
437 if (copied < PAGE_SIZE - offset) {
448 * unlocks pages after btrfs_file_write is done with them
450 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
453 for (i = 0; i < num_pages; i++) {
454 /* page checked is some magic around finding pages that
455 * have been modified without going through btrfs_set_page_dirty
456 * clear it here. There should be no need to mark the pages
457 * accessed as prepare_pages should have marked them accessed
458 * in prepare_pages via find_or_create_page()
460 ClearPageChecked(pages[i]);
461 unlock_page(pages[i]);
466 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
469 struct extent_state **cached_state)
471 u64 search_start = start;
472 const u64 end = start + len - 1;
474 while (search_start < end) {
475 const u64 search_len = end - search_start + 1;
476 struct extent_map *em;
480 em = btrfs_get_extent(inode, NULL, 0, search_start, search_len);
484 if (em->block_start != EXTENT_MAP_HOLE)
488 if (em->start < search_start)
489 em_len -= search_start - em->start;
490 if (em_len > search_len)
493 ret = set_extent_bit(&inode->io_tree, search_start,
494 search_start + em_len - 1,
496 NULL, cached_state, GFP_NOFS);
498 search_start = extent_map_end(em);
507 * after copy_from_user, pages need to be dirtied and we need to make
508 * sure holes are created between the current EOF and the start of
509 * any next extents (if required).
511 * this also makes the decision about creating an inline extent vs
512 * doing real data extents, marking pages dirty and delalloc as required.
514 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
515 size_t num_pages, loff_t pos, size_t write_bytes,
516 struct extent_state **cached)
518 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
523 u64 end_of_last_block;
524 u64 end_pos = pos + write_bytes;
525 loff_t isize = i_size_read(inode);
526 unsigned int extra_bits = 0;
528 start_pos = pos & ~((u64) fs_info->sectorsize - 1);
529 num_bytes = round_up(write_bytes + pos - start_pos,
530 fs_info->sectorsize);
532 end_of_last_block = start_pos + num_bytes - 1;
535 * The pages may have already been dirty, clear out old accounting so
536 * we can set things up properly
538 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos, end_of_last_block,
539 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
542 if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
543 if (start_pos >= isize &&
544 !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
546 * There can't be any extents following eof in this case
547 * so just set the delalloc new bit for the range
550 extra_bits |= EXTENT_DELALLOC_NEW;
552 err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
560 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
565 for (i = 0; i < num_pages; i++) {
566 struct page *p = pages[i];
573 * we've only changed i_size in ram, and we haven't updated
574 * the disk i_size. There is no need to log the inode
578 i_size_write(inode, end_pos);
583 * this drops all the extents in the cache that intersect the range
584 * [start, end]. Existing extents are split as required.
586 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
589 struct extent_map *em;
590 struct extent_map *split = NULL;
591 struct extent_map *split2 = NULL;
592 struct extent_map_tree *em_tree = &inode->extent_tree;
593 u64 len = end - start + 1;
601 WARN_ON(end < start);
602 if (end == (u64)-1) {
611 split = alloc_extent_map();
613 split2 = alloc_extent_map();
614 if (!split || !split2)
617 write_lock(&em_tree->lock);
618 em = lookup_extent_mapping(em_tree, start, len);
620 write_unlock(&em_tree->lock);
624 gen = em->generation;
625 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
626 if (testend && em->start + em->len >= start + len) {
628 write_unlock(&em_tree->lock);
631 start = em->start + em->len;
633 len = start + len - (em->start + em->len);
635 write_unlock(&em_tree->lock);
638 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
639 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
640 clear_bit(EXTENT_FLAG_LOGGING, &flags);
641 modified = !list_empty(&em->list);
645 if (em->start < start) {
646 split->start = em->start;
647 split->len = start - em->start;
649 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
650 split->orig_start = em->orig_start;
651 split->block_start = em->block_start;
654 split->block_len = em->block_len;
656 split->block_len = split->len;
657 split->orig_block_len = max(split->block_len,
659 split->ram_bytes = em->ram_bytes;
661 split->orig_start = split->start;
662 split->block_len = 0;
663 split->block_start = em->block_start;
664 split->orig_block_len = 0;
665 split->ram_bytes = split->len;
668 split->generation = gen;
669 split->flags = flags;
670 split->compress_type = em->compress_type;
671 replace_extent_mapping(em_tree, em, split, modified);
672 free_extent_map(split);
676 if (testend && em->start + em->len > start + len) {
677 u64 diff = start + len - em->start;
679 split->start = start + len;
680 split->len = em->start + em->len - (start + len);
681 split->flags = flags;
682 split->compress_type = em->compress_type;
683 split->generation = gen;
685 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
686 split->orig_block_len = max(em->block_len,
689 split->ram_bytes = em->ram_bytes;
691 split->block_len = em->block_len;
692 split->block_start = em->block_start;
693 split->orig_start = em->orig_start;
695 split->block_len = split->len;
696 split->block_start = em->block_start
698 split->orig_start = em->orig_start;
701 split->ram_bytes = split->len;
702 split->orig_start = split->start;
703 split->block_len = 0;
704 split->block_start = em->block_start;
705 split->orig_block_len = 0;
708 if (extent_map_in_tree(em)) {
709 replace_extent_mapping(em_tree, em, split,
712 ret = add_extent_mapping(em_tree, split,
714 ASSERT(ret == 0); /* Logic error */
716 free_extent_map(split);
720 if (extent_map_in_tree(em))
721 remove_extent_mapping(em_tree, em);
722 write_unlock(&em_tree->lock);
726 /* once for the tree*/
730 free_extent_map(split);
732 free_extent_map(split2);
736 * this is very complex, but the basic idea is to drop all extents
737 * in the range start - end. hint_block is filled in with a block number
738 * that would be a good hint to the block allocator for this file.
740 * If an extent intersects the range but is not entirely inside the range
741 * it is either truncated or split. Anything entirely inside the range
742 * is deleted from the tree.
744 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
745 struct btrfs_root *root, struct inode *inode,
746 struct btrfs_path *path, u64 start, u64 end,
747 u64 *drop_end, int drop_cache,
749 u32 extent_item_size,
752 struct btrfs_fs_info *fs_info = root->fs_info;
753 struct extent_buffer *leaf;
754 struct btrfs_file_extent_item *fi;
755 struct btrfs_ref ref = { 0 };
756 struct btrfs_key key;
757 struct btrfs_key new_key;
758 u64 ino = btrfs_ino(BTRFS_I(inode));
759 u64 search_start = start;
762 u64 extent_offset = 0;
764 u64 last_end = start;
770 int modify_tree = -1;
773 int leafs_visited = 0;
776 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
778 if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
781 update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
782 root == fs_info->tree_root);
785 ret = btrfs_lookup_file_extent(trans, root, path, ino,
786 search_start, modify_tree);
789 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
790 leaf = path->nodes[0];
791 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
792 if (key.objectid == ino &&
793 key.type == BTRFS_EXTENT_DATA_KEY)
799 leaf = path->nodes[0];
800 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
802 ret = btrfs_next_leaf(root, path);
810 leaf = path->nodes[0];
814 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
816 if (key.objectid > ino)
818 if (WARN_ON_ONCE(key.objectid < ino) ||
819 key.type < BTRFS_EXTENT_DATA_KEY) {
824 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
827 fi = btrfs_item_ptr(leaf, path->slots[0],
828 struct btrfs_file_extent_item);
829 extent_type = btrfs_file_extent_type(leaf, fi);
831 if (extent_type == BTRFS_FILE_EXTENT_REG ||
832 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
833 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
834 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
835 extent_offset = btrfs_file_extent_offset(leaf, fi);
836 extent_end = key.offset +
837 btrfs_file_extent_num_bytes(leaf, fi);
838 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
839 extent_end = key.offset +
840 btrfs_file_extent_ram_bytes(leaf, fi);
847 * Don't skip extent items representing 0 byte lengths. They
848 * used to be created (bug) if while punching holes we hit
849 * -ENOSPC condition. So if we find one here, just ensure we
850 * delete it, otherwise we would insert a new file extent item
851 * with the same key (offset) as that 0 bytes length file
852 * extent item in the call to setup_items_for_insert() later
855 if (extent_end == key.offset && extent_end >= search_start) {
856 last_end = extent_end;
857 goto delete_extent_item;
860 if (extent_end <= search_start) {
866 search_start = max(key.offset, start);
867 if (recow || !modify_tree) {
869 btrfs_release_path(path);
874 * | - range to drop - |
875 * | -------- extent -------- |
877 if (start > key.offset && end < extent_end) {
879 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
884 memcpy(&new_key, &key, sizeof(new_key));
885 new_key.offset = start;
886 ret = btrfs_duplicate_item(trans, root, path,
888 if (ret == -EAGAIN) {
889 btrfs_release_path(path);
895 leaf = path->nodes[0];
896 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
897 struct btrfs_file_extent_item);
898 btrfs_set_file_extent_num_bytes(leaf, fi,
901 fi = btrfs_item_ptr(leaf, path->slots[0],
902 struct btrfs_file_extent_item);
904 extent_offset += start - key.offset;
905 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
906 btrfs_set_file_extent_num_bytes(leaf, fi,
908 btrfs_mark_buffer_dirty(leaf);
910 if (update_refs && disk_bytenr > 0) {
911 btrfs_init_generic_ref(&ref,
912 BTRFS_ADD_DELAYED_REF,
913 disk_bytenr, num_bytes, 0);
914 btrfs_init_data_ref(&ref,
915 root->root_key.objectid,
917 start - extent_offset);
918 ret = btrfs_inc_extent_ref(trans, &ref);
919 BUG_ON(ret); /* -ENOMEM */
924 * From here on out we will have actually dropped something, so
925 * last_end can be updated.
927 last_end = extent_end;
930 * | ---- range to drop ----- |
931 * | -------- extent -------- |
933 if (start <= key.offset && end < extent_end) {
934 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
939 memcpy(&new_key, &key, sizeof(new_key));
940 new_key.offset = end;
941 btrfs_set_item_key_safe(fs_info, path, &new_key);
943 extent_offset += end - key.offset;
944 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
945 btrfs_set_file_extent_num_bytes(leaf, fi,
947 btrfs_mark_buffer_dirty(leaf);
948 if (update_refs && disk_bytenr > 0)
949 inode_sub_bytes(inode, end - key.offset);
953 search_start = extent_end;
955 * | ---- range to drop ----- |
956 * | -------- extent -------- |
958 if (start > key.offset && end >= extent_end) {
960 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
965 btrfs_set_file_extent_num_bytes(leaf, fi,
967 btrfs_mark_buffer_dirty(leaf);
968 if (update_refs && disk_bytenr > 0)
969 inode_sub_bytes(inode, extent_end - start);
970 if (end == extent_end)
978 * | ---- range to drop ----- |
979 * | ------ extent ------ |
981 if (start <= key.offset && end >= extent_end) {
984 del_slot = path->slots[0];
987 BUG_ON(del_slot + del_nr != path->slots[0]);
992 extent_type == BTRFS_FILE_EXTENT_INLINE) {
993 inode_sub_bytes(inode,
994 extent_end - key.offset);
995 extent_end = ALIGN(extent_end,
996 fs_info->sectorsize);
997 } else if (update_refs && disk_bytenr > 0) {
998 btrfs_init_generic_ref(&ref,
999 BTRFS_DROP_DELAYED_REF,
1000 disk_bytenr, num_bytes, 0);
1001 btrfs_init_data_ref(&ref,
1002 root->root_key.objectid,
1004 key.offset - extent_offset);
1005 ret = btrfs_free_extent(trans, &ref);
1006 BUG_ON(ret); /* -ENOMEM */
1007 inode_sub_bytes(inode,
1008 extent_end - key.offset);
1011 if (end == extent_end)
1014 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1019 ret = btrfs_del_items(trans, root, path, del_slot,
1022 btrfs_abort_transaction(trans, ret);
1029 btrfs_release_path(path);
1036 if (!ret && del_nr > 0) {
1038 * Set path->slots[0] to first slot, so that after the delete
1039 * if items are move off from our leaf to its immediate left or
1040 * right neighbor leafs, we end up with a correct and adjusted
1041 * path->slots[0] for our insertion (if replace_extent != 0).
1043 path->slots[0] = del_slot;
1044 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1046 btrfs_abort_transaction(trans, ret);
1049 leaf = path->nodes[0];
1051 * If btrfs_del_items() was called, it might have deleted a leaf, in
1052 * which case it unlocked our path, so check path->locks[0] matches a
1055 if (!ret && replace_extent && leafs_visited == 1 &&
1056 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1057 path->locks[0] == BTRFS_WRITE_LOCK) &&
1058 btrfs_leaf_free_space(leaf) >=
1059 sizeof(struct btrfs_item) + extent_item_size) {
1062 key.type = BTRFS_EXTENT_DATA_KEY;
1064 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1065 struct btrfs_key slot_key;
1067 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1068 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1071 setup_items_for_insert(root, path, &key,
1074 sizeof(struct btrfs_item) +
1075 extent_item_size, 1);
1079 if (!replace_extent || !(*key_inserted))
1080 btrfs_release_path(path);
1082 *drop_end = found ? min(end, last_end) : end;
1086 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1087 struct btrfs_root *root, struct inode *inode, u64 start,
1088 u64 end, int drop_cache)
1090 struct btrfs_path *path;
1093 path = btrfs_alloc_path();
1096 ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1097 drop_cache, 0, 0, NULL);
1098 btrfs_free_path(path);
1102 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1103 u64 objectid, u64 bytenr, u64 orig_offset,
1104 u64 *start, u64 *end)
1106 struct btrfs_file_extent_item *fi;
1107 struct btrfs_key key;
1110 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1113 btrfs_item_key_to_cpu(leaf, &key, slot);
1114 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1117 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1118 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1119 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1120 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1121 btrfs_file_extent_compression(leaf, fi) ||
1122 btrfs_file_extent_encryption(leaf, fi) ||
1123 btrfs_file_extent_other_encoding(leaf, fi))
1126 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1127 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1130 *start = key.offset;
1136 * Mark extent in the range start - end as written.
1138 * This changes extent type from 'pre-allocated' to 'regular'. If only
1139 * part of extent is marked as written, the extent will be split into
1142 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1143 struct btrfs_inode *inode, u64 start, u64 end)
1145 struct btrfs_fs_info *fs_info = trans->fs_info;
1146 struct btrfs_root *root = inode->root;
1147 struct extent_buffer *leaf;
1148 struct btrfs_path *path;
1149 struct btrfs_file_extent_item *fi;
1150 struct btrfs_ref ref = { 0 };
1151 struct btrfs_key key;
1152 struct btrfs_key new_key;
1164 u64 ino = btrfs_ino(inode);
1166 path = btrfs_alloc_path();
1173 key.type = BTRFS_EXTENT_DATA_KEY;
1176 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1179 if (ret > 0 && path->slots[0] > 0)
1182 leaf = path->nodes[0];
1183 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1184 if (key.objectid != ino ||
1185 key.type != BTRFS_EXTENT_DATA_KEY) {
1187 btrfs_abort_transaction(trans, ret);
1190 fi = btrfs_item_ptr(leaf, path->slots[0],
1191 struct btrfs_file_extent_item);
1192 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1194 btrfs_abort_transaction(trans, ret);
1197 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1198 if (key.offset > start || extent_end < end) {
1200 btrfs_abort_transaction(trans, ret);
1204 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1205 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1206 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1207 memcpy(&new_key, &key, sizeof(new_key));
1209 if (start == key.offset && end < extent_end) {
1212 if (extent_mergeable(leaf, path->slots[0] - 1,
1213 ino, bytenr, orig_offset,
1214 &other_start, &other_end)) {
1215 new_key.offset = end;
1216 btrfs_set_item_key_safe(fs_info, path, &new_key);
1217 fi = btrfs_item_ptr(leaf, path->slots[0],
1218 struct btrfs_file_extent_item);
1219 btrfs_set_file_extent_generation(leaf, fi,
1221 btrfs_set_file_extent_num_bytes(leaf, fi,
1223 btrfs_set_file_extent_offset(leaf, fi,
1225 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1226 struct btrfs_file_extent_item);
1227 btrfs_set_file_extent_generation(leaf, fi,
1229 btrfs_set_file_extent_num_bytes(leaf, fi,
1231 btrfs_mark_buffer_dirty(leaf);
1236 if (start > key.offset && end == extent_end) {
1239 if (extent_mergeable(leaf, path->slots[0] + 1,
1240 ino, bytenr, orig_offset,
1241 &other_start, &other_end)) {
1242 fi = btrfs_item_ptr(leaf, path->slots[0],
1243 struct btrfs_file_extent_item);
1244 btrfs_set_file_extent_num_bytes(leaf, fi,
1245 start - key.offset);
1246 btrfs_set_file_extent_generation(leaf, fi,
1249 new_key.offset = start;
1250 btrfs_set_item_key_safe(fs_info, path, &new_key);
1252 fi = btrfs_item_ptr(leaf, path->slots[0],
1253 struct btrfs_file_extent_item);
1254 btrfs_set_file_extent_generation(leaf, fi,
1256 btrfs_set_file_extent_num_bytes(leaf, fi,
1258 btrfs_set_file_extent_offset(leaf, fi,
1259 start - orig_offset);
1260 btrfs_mark_buffer_dirty(leaf);
1265 while (start > key.offset || end < extent_end) {
1266 if (key.offset == start)
1269 new_key.offset = split;
1270 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1271 if (ret == -EAGAIN) {
1272 btrfs_release_path(path);
1276 btrfs_abort_transaction(trans, ret);
1280 leaf = path->nodes[0];
1281 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1282 struct btrfs_file_extent_item);
1283 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1284 btrfs_set_file_extent_num_bytes(leaf, fi,
1285 split - key.offset);
1287 fi = btrfs_item_ptr(leaf, path->slots[0],
1288 struct btrfs_file_extent_item);
1290 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1291 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1292 btrfs_set_file_extent_num_bytes(leaf, fi,
1293 extent_end - split);
1294 btrfs_mark_buffer_dirty(leaf);
1296 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1298 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1300 ret = btrfs_inc_extent_ref(trans, &ref);
1302 btrfs_abort_transaction(trans, ret);
1306 if (split == start) {
1309 if (start != key.offset) {
1311 btrfs_abort_transaction(trans, ret);
1322 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1324 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1325 if (extent_mergeable(leaf, path->slots[0] + 1,
1326 ino, bytenr, orig_offset,
1327 &other_start, &other_end)) {
1329 btrfs_release_path(path);
1332 extent_end = other_end;
1333 del_slot = path->slots[0] + 1;
1335 ret = btrfs_free_extent(trans, &ref);
1337 btrfs_abort_transaction(trans, ret);
1343 if (extent_mergeable(leaf, path->slots[0] - 1,
1344 ino, bytenr, orig_offset,
1345 &other_start, &other_end)) {
1347 btrfs_release_path(path);
1350 key.offset = other_start;
1351 del_slot = path->slots[0];
1353 ret = btrfs_free_extent(trans, &ref);
1355 btrfs_abort_transaction(trans, ret);
1360 fi = btrfs_item_ptr(leaf, path->slots[0],
1361 struct btrfs_file_extent_item);
1362 btrfs_set_file_extent_type(leaf, fi,
1363 BTRFS_FILE_EXTENT_REG);
1364 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1365 btrfs_mark_buffer_dirty(leaf);
1367 fi = btrfs_item_ptr(leaf, del_slot - 1,
1368 struct btrfs_file_extent_item);
1369 btrfs_set_file_extent_type(leaf, fi,
1370 BTRFS_FILE_EXTENT_REG);
1371 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1372 btrfs_set_file_extent_num_bytes(leaf, fi,
1373 extent_end - key.offset);
1374 btrfs_mark_buffer_dirty(leaf);
1376 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1378 btrfs_abort_transaction(trans, ret);
1383 btrfs_free_path(path);
1388 * on error we return an unlocked page and the error value
1389 * on success we return a locked page and 0
1391 static int prepare_uptodate_page(struct inode *inode,
1392 struct page *page, u64 pos,
1393 bool force_uptodate)
1397 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1398 !PageUptodate(page)) {
1399 ret = btrfs_readpage(NULL, page);
1403 if (!PageUptodate(page)) {
1407 if (page->mapping != inode->i_mapping) {
1416 * this just gets pages into the page cache and locks them down.
1418 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1419 size_t num_pages, loff_t pos,
1420 size_t write_bytes, bool force_uptodate)
1423 unsigned long index = pos >> PAGE_SHIFT;
1424 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1428 for (i = 0; i < num_pages; i++) {
1430 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1431 mask | __GFP_WRITE);
1439 err = prepare_uptodate_page(inode, pages[i], pos,
1441 if (!err && i == num_pages - 1)
1442 err = prepare_uptodate_page(inode, pages[i],
1443 pos + write_bytes, false);
1446 if (err == -EAGAIN) {
1453 wait_on_page_writeback(pages[i]);
1458 while (faili >= 0) {
1459 unlock_page(pages[faili]);
1460 put_page(pages[faili]);
1468 * This function locks the extent and properly waits for data=ordered extents
1469 * to finish before allowing the pages to be modified if need.
1472 * 1 - the extent is locked
1473 * 0 - the extent is not locked, and everything is OK
1474 * -EAGAIN - need re-prepare the pages
1475 * the other < 0 number - Something wrong happens
1478 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1479 size_t num_pages, loff_t pos,
1481 u64 *lockstart, u64 *lockend,
1482 struct extent_state **cached_state)
1484 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1490 start_pos = round_down(pos, fs_info->sectorsize);
1491 last_pos = start_pos
1492 + round_up(pos + write_bytes - start_pos,
1493 fs_info->sectorsize) - 1;
1495 if (start_pos < inode->vfs_inode.i_size) {
1496 struct btrfs_ordered_extent *ordered;
1498 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1500 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1501 last_pos - start_pos + 1);
1503 ordered->file_offset + ordered->num_bytes > start_pos &&
1504 ordered->file_offset <= last_pos) {
1505 unlock_extent_cached(&inode->io_tree, start_pos,
1506 last_pos, cached_state);
1507 for (i = 0; i < num_pages; i++) {
1508 unlock_page(pages[i]);
1511 btrfs_start_ordered_extent(&inode->vfs_inode,
1513 btrfs_put_ordered_extent(ordered);
1517 btrfs_put_ordered_extent(ordered);
1519 *lockstart = start_pos;
1520 *lockend = last_pos;
1525 * It's possible the pages are dirty right now, but we don't want
1526 * to clean them yet because copy_from_user may catch a page fault
1527 * and we might have to fall back to one page at a time. If that
1528 * happens, we'll unlock these pages and we'd have a window where
1529 * reclaim could sneak in and drop the once-dirty page on the floor
1530 * without writing it.
1532 * We have the pages locked and the extent range locked, so there's
1533 * no way someone can start IO on any dirty pages in this range.
1535 * We'll call btrfs_dirty_pages() later on, and that will flip around
1536 * delalloc bits and dirty the pages as required.
1538 for (i = 0; i < num_pages; i++) {
1539 set_page_extent_mapped(pages[i]);
1540 WARN_ON(!PageLocked(pages[i]));
1546 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1547 size_t *write_bytes)
1549 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1550 struct btrfs_root *root = inode->root;
1551 u64 lockstart, lockend;
1555 ret = btrfs_start_write_no_snapshotting(root);
1559 lockstart = round_down(pos, fs_info->sectorsize);
1560 lockend = round_up(pos + *write_bytes,
1561 fs_info->sectorsize) - 1;
1563 btrfs_lock_and_flush_ordered_range(&inode->io_tree, inode, lockstart,
1566 num_bytes = lockend - lockstart + 1;
1567 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1571 btrfs_end_write_no_snapshotting(root);
1573 *write_bytes = min_t(size_t, *write_bytes ,
1574 num_bytes - pos + lockstart);
1577 unlock_extent(&inode->io_tree, lockstart, lockend);
1582 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1585 struct file *file = iocb->ki_filp;
1586 loff_t pos = iocb->ki_pos;
1587 struct inode *inode = file_inode(file);
1588 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1589 struct btrfs_root *root = BTRFS_I(inode)->root;
1590 struct page **pages = NULL;
1591 struct extent_changeset *data_reserved = NULL;
1592 u64 release_bytes = 0;
1595 size_t num_written = 0;
1598 bool only_release_metadata = false;
1599 bool force_page_uptodate = false;
1601 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1602 PAGE_SIZE / (sizeof(struct page *)));
1603 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1604 nrptrs = max(nrptrs, 8);
1605 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1609 while (iov_iter_count(i) > 0) {
1610 struct extent_state *cached_state = NULL;
1611 size_t offset = offset_in_page(pos);
1612 size_t sector_offset;
1613 size_t write_bytes = min(iov_iter_count(i),
1614 nrptrs * (size_t)PAGE_SIZE -
1616 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1618 size_t reserve_bytes;
1621 size_t dirty_sectors;
1625 WARN_ON(num_pages > nrptrs);
1628 * Fault pages before locking them in prepare_pages
1629 * to avoid recursive lock
1631 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1636 only_release_metadata = false;
1637 sector_offset = pos & (fs_info->sectorsize - 1);
1638 reserve_bytes = round_up(write_bytes + sector_offset,
1639 fs_info->sectorsize);
1641 extent_changeset_release(data_reserved);
1642 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1645 if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1646 BTRFS_INODE_PREALLOC)) &&
1647 check_can_nocow(BTRFS_I(inode), pos,
1648 &write_bytes) > 0) {
1650 * For nodata cow case, no need to reserve
1653 only_release_metadata = true;
1655 * our prealloc extent may be smaller than
1656 * write_bytes, so scale down.
1658 num_pages = DIV_ROUND_UP(write_bytes + offset,
1660 reserve_bytes = round_up(write_bytes +
1662 fs_info->sectorsize);
1668 WARN_ON(reserve_bytes == 0);
1669 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1672 if (!only_release_metadata)
1673 btrfs_free_reserved_data_space(inode,
1677 btrfs_end_write_no_snapshotting(root);
1681 release_bytes = reserve_bytes;
1684 * This is going to setup the pages array with the number of
1685 * pages we want, so we don't really need to worry about the
1686 * contents of pages from loop to loop
1688 ret = prepare_pages(inode, pages, num_pages,
1690 force_page_uptodate);
1692 btrfs_delalloc_release_extents(BTRFS_I(inode),
1697 extents_locked = lock_and_cleanup_extent_if_need(
1698 BTRFS_I(inode), pages,
1699 num_pages, pos, write_bytes, &lockstart,
1700 &lockend, &cached_state);
1701 if (extents_locked < 0) {
1702 if (extents_locked == -EAGAIN)
1704 btrfs_delalloc_release_extents(BTRFS_I(inode),
1706 ret = extents_locked;
1710 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1712 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1713 dirty_sectors = round_up(copied + sector_offset,
1714 fs_info->sectorsize);
1715 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1718 * if we have trouble faulting in the pages, fall
1719 * back to one page at a time
1721 if (copied < write_bytes)
1725 force_page_uptodate = true;
1729 force_page_uptodate = false;
1730 dirty_pages = DIV_ROUND_UP(copied + offset,
1734 if (num_sectors > dirty_sectors) {
1735 /* release everything except the sectors we dirtied */
1736 release_bytes -= dirty_sectors <<
1737 fs_info->sb->s_blocksize_bits;
1738 if (only_release_metadata) {
1739 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1740 release_bytes, true);
1744 __pos = round_down(pos,
1745 fs_info->sectorsize) +
1746 (dirty_pages << PAGE_SHIFT);
1747 btrfs_delalloc_release_space(inode,
1748 data_reserved, __pos,
1749 release_bytes, true);
1753 release_bytes = round_up(copied + sector_offset,
1754 fs_info->sectorsize);
1757 ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1758 pos, copied, &cached_state);
1761 * If we have not locked the extent range, because the range's
1762 * start offset is >= i_size, we might still have a non-NULL
1763 * cached extent state, acquired while marking the extent range
1764 * as delalloc through btrfs_dirty_pages(). Therefore free any
1765 * possible cached extent state to avoid a memory leak.
1768 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1769 lockstart, lockend, &cached_state);
1771 free_extent_state(cached_state);
1773 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1775 btrfs_drop_pages(pages, num_pages);
1780 if (only_release_metadata)
1781 btrfs_end_write_no_snapshotting(root);
1783 if (only_release_metadata && copied > 0) {
1784 lockstart = round_down(pos,
1785 fs_info->sectorsize);
1786 lockend = round_up(pos + copied,
1787 fs_info->sectorsize) - 1;
1789 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1790 lockend, EXTENT_NORESERVE, NULL,
1794 btrfs_drop_pages(pages, num_pages);
1798 balance_dirty_pages_ratelimited(inode->i_mapping);
1799 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1800 btrfs_btree_balance_dirty(fs_info);
1803 num_written += copied;
1808 if (release_bytes) {
1809 if (only_release_metadata) {
1810 btrfs_end_write_no_snapshotting(root);
1811 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1812 release_bytes, true);
1814 btrfs_delalloc_release_space(inode, data_reserved,
1815 round_down(pos, fs_info->sectorsize),
1816 release_bytes, true);
1820 extent_changeset_free(data_reserved);
1821 return num_written ? num_written : ret;
1824 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1826 struct file *file = iocb->ki_filp;
1827 struct inode *inode = file_inode(file);
1830 ssize_t written_buffered;
1834 written = generic_file_direct_write(iocb, from);
1836 if (written < 0 || !iov_iter_count(from))
1840 written_buffered = btrfs_buffered_write(iocb, from);
1841 if (written_buffered < 0) {
1842 err = written_buffered;
1846 * Ensure all data is persisted. We want the next direct IO read to be
1847 * able to read what was just written.
1849 endbyte = pos + written_buffered - 1;
1850 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1853 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1856 written += written_buffered;
1857 iocb->ki_pos = pos + written_buffered;
1858 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1859 endbyte >> PAGE_SHIFT);
1861 return written ? written : err;
1864 static void update_time_for_write(struct inode *inode)
1866 struct timespec64 now;
1868 if (IS_NOCMTIME(inode))
1871 now = current_time(inode);
1872 if (!timespec64_equal(&inode->i_mtime, &now))
1873 inode->i_mtime = now;
1875 if (!timespec64_equal(&inode->i_ctime, &now))
1876 inode->i_ctime = now;
1878 if (IS_I_VERSION(inode))
1879 inode_inc_iversion(inode);
1882 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1883 struct iov_iter *from)
1885 struct file *file = iocb->ki_filp;
1886 struct inode *inode = file_inode(file);
1887 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1888 struct btrfs_root *root = BTRFS_I(inode)->root;
1891 ssize_t num_written = 0;
1892 const bool sync = iocb->ki_flags & IOCB_DSYNC;
1899 if (!(iocb->ki_flags & IOCB_DIRECT) &&
1900 (iocb->ki_flags & IOCB_NOWAIT))
1903 if (iocb->ki_flags & IOCB_NOWAIT) {
1904 if (!inode_trylock(inode))
1910 err = generic_write_checks(iocb, from);
1912 inode_unlock(inode);
1917 count = iov_iter_count(from);
1918 if (iocb->ki_flags & IOCB_NOWAIT) {
1920 * We will allocate space in case nodatacow is not set,
1923 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1924 BTRFS_INODE_PREALLOC)) ||
1925 check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1926 inode_unlock(inode);
1931 current->backing_dev_info = inode_to_bdi(inode);
1932 err = file_remove_privs(file);
1934 inode_unlock(inode);
1939 * If BTRFS flips readonly due to some impossible error
1940 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1941 * although we have opened a file as writable, we have
1942 * to stop this write operation to ensure FS consistency.
1944 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1945 inode_unlock(inode);
1951 * We reserve space for updating the inode when we reserve space for the
1952 * extent we are going to write, so we will enospc out there. We don't
1953 * need to start yet another transaction to update the inode as we will
1954 * update the inode when we finish writing whatever data we write.
1956 update_time_for_write(inode);
1958 start_pos = round_down(pos, fs_info->sectorsize);
1959 oldsize = i_size_read(inode);
1960 if (start_pos > oldsize) {
1961 /* Expand hole size to cover write data, preventing empty gap */
1962 end_pos = round_up(pos + count,
1963 fs_info->sectorsize);
1964 err = btrfs_cont_expand(inode, oldsize, end_pos);
1966 inode_unlock(inode);
1969 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1974 atomic_inc(&BTRFS_I(inode)->sync_writers);
1976 if (iocb->ki_flags & IOCB_DIRECT) {
1977 num_written = __btrfs_direct_write(iocb, from);
1979 num_written = btrfs_buffered_write(iocb, from);
1980 if (num_written > 0)
1981 iocb->ki_pos = pos + num_written;
1983 pagecache_isize_extended(inode, oldsize,
1984 i_size_read(inode));
1987 inode_unlock(inode);
1990 * We also have to set last_sub_trans to the current log transid,
1991 * otherwise subsequent syncs to a file that's been synced in this
1992 * transaction will appear to have already occurred.
1994 spin_lock(&BTRFS_I(inode)->lock);
1995 BTRFS_I(inode)->last_sub_trans = root->log_transid;
1996 spin_unlock(&BTRFS_I(inode)->lock);
1997 if (num_written > 0)
1998 num_written = generic_write_sync(iocb, num_written);
2001 atomic_dec(&BTRFS_I(inode)->sync_writers);
2003 current->backing_dev_info = NULL;
2004 return num_written ? num_written : err;
2007 int btrfs_release_file(struct inode *inode, struct file *filp)
2009 struct btrfs_file_private *private = filp->private_data;
2011 if (private && private->filldir_buf)
2012 kfree(private->filldir_buf);
2014 filp->private_data = NULL;
2017 * ordered_data_close is set by setattr when we are about to truncate
2018 * a file from a non-zero size to a zero size. This tries to
2019 * flush down new bytes that may have been written if the
2020 * application were using truncate to replace a file in place.
2022 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2023 &BTRFS_I(inode)->runtime_flags))
2024 filemap_flush(inode->i_mapping);
2028 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2031 struct blk_plug plug;
2034 * This is only called in fsync, which would do synchronous writes, so
2035 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2036 * multiple disks using raid profile, a large IO can be split to
2037 * several segments of stripe length (currently 64K).
2039 blk_start_plug(&plug);
2040 atomic_inc(&BTRFS_I(inode)->sync_writers);
2041 ret = btrfs_fdatawrite_range(inode, start, end);
2042 atomic_dec(&BTRFS_I(inode)->sync_writers);
2043 blk_finish_plug(&plug);
2049 * fsync call for both files and directories. This logs the inode into
2050 * the tree log instead of forcing full commits whenever possible.
2052 * It needs to call filemap_fdatawait so that all ordered extent updates are
2053 * in the metadata btree are up to date for copying to the log.
2055 * It drops the inode mutex before doing the tree log commit. This is an
2056 * important optimization for directories because holding the mutex prevents
2057 * new operations on the dir while we write to disk.
2059 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2061 struct dentry *dentry = file_dentry(file);
2062 struct inode *inode = d_inode(dentry);
2063 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2064 struct btrfs_root *root = BTRFS_I(inode)->root;
2065 struct btrfs_trans_handle *trans;
2066 struct btrfs_log_ctx ctx;
2069 trace_btrfs_sync_file(file, datasync);
2071 btrfs_init_log_ctx(&ctx, inode);
2074 * We write the dirty pages in the range and wait until they complete
2075 * out of the ->i_mutex. If so, we can flush the dirty pages by
2076 * multi-task, and make the performance up. See
2077 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2079 ret = start_ordered_ops(inode, start, end);
2086 * We take the dio_sem here because the tree log stuff can race with
2087 * lockless dio writes and get an extent map logged for an extent we
2088 * never waited on. We need it this high up for lockdep reasons.
2090 down_write(&BTRFS_I(inode)->dio_sem);
2092 atomic_inc(&root->log_batch);
2095 * If the inode needs a full sync, make sure we use a full range to
2096 * avoid log tree corruption, due to hole detection racing with ordered
2097 * extent completion for adjacent ranges, and assertion failures during
2098 * hole detection. Do this while holding the inode lock, to avoid races
2101 if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2102 &BTRFS_I(inode)->runtime_flags)) {
2108 * Before we acquired the inode's lock, someone may have dirtied more
2109 * pages in the target range. We need to make sure that writeback for
2110 * any such pages does not start while we are logging the inode, because
2111 * if it does, any of the following might happen when we are not doing a
2114 * 1) We log an extent after its writeback finishes but before its
2115 * checksums are added to the csum tree, leading to -EIO errors
2116 * when attempting to read the extent after a log replay.
2118 * 2) We can end up logging an extent before its writeback finishes.
2119 * Therefore after the log replay we will have a file extent item
2120 * pointing to an unwritten extent (and no data checksums as well).
2122 * So trigger writeback for any eventual new dirty pages and then we
2123 * wait for all ordered extents to complete below.
2125 ret = start_ordered_ops(inode, start, end);
2127 inode_unlock(inode);
2132 * We have to do this here to avoid the priority inversion of waiting on
2133 * IO of a lower priority task while holding a transaction open.
2135 * Also, the range length can be represented by u64, we have to do the
2136 * typecasts to avoid signed overflow if it's [0, LLONG_MAX].
2138 ret = btrfs_wait_ordered_range(inode, start, (u64)end - (u64)start + 1);
2140 up_write(&BTRFS_I(inode)->dio_sem);
2141 inode_unlock(inode);
2144 atomic_inc(&root->log_batch);
2147 if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2148 BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2150 * We've had everything committed since the last time we were
2151 * modified so clear this flag in case it was set for whatever
2152 * reason, it's no longer relevant.
2154 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2155 &BTRFS_I(inode)->runtime_flags);
2157 * An ordered extent might have started before and completed
2158 * already with io errors, in which case the inode was not
2159 * updated and we end up here. So check the inode's mapping
2160 * for any errors that might have happened since we last
2161 * checked called fsync.
2163 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2164 up_write(&BTRFS_I(inode)->dio_sem);
2165 inode_unlock(inode);
2170 * We use start here because we will need to wait on the IO to complete
2171 * in btrfs_sync_log, which could require joining a transaction (for
2172 * example checking cross references in the nocow path). If we use join
2173 * here we could get into a situation where we're waiting on IO to
2174 * happen that is blocked on a transaction trying to commit. With start
2175 * we inc the extwriter counter, so we wait for all extwriters to exit
2176 * before we start blocking joiners. This comment is to keep somebody
2177 * from thinking they are super smart and changing this to
2178 * btrfs_join_transaction *cough*Josef*cough*.
2180 trans = btrfs_start_transaction(root, 0);
2181 if (IS_ERR(trans)) {
2182 ret = PTR_ERR(trans);
2183 up_write(&BTRFS_I(inode)->dio_sem);
2184 inode_unlock(inode);
2188 ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2190 /* Fallthrough and commit/free transaction. */
2194 /* we've logged all the items and now have a consistent
2195 * version of the file in the log. It is possible that
2196 * someone will come in and modify the file, but that's
2197 * fine because the log is consistent on disk, and we
2198 * have references to all of the file's extents
2200 * It is possible that someone will come in and log the
2201 * file again, but that will end up using the synchronization
2202 * inside btrfs_sync_log to keep things safe.
2204 up_write(&BTRFS_I(inode)->dio_sem);
2205 inode_unlock(inode);
2207 if (ret != BTRFS_NO_LOG_SYNC) {
2209 ret = btrfs_sync_log(trans, root, &ctx);
2211 ret = btrfs_end_transaction(trans);
2215 ret = btrfs_commit_transaction(trans);
2217 ret = btrfs_end_transaction(trans);
2220 ASSERT(list_empty(&ctx.list));
2221 err = file_check_and_advance_wb_err(file);
2224 return ret > 0 ? -EIO : ret;
2227 static const struct vm_operations_struct btrfs_file_vm_ops = {
2228 .fault = filemap_fault,
2229 .map_pages = filemap_map_pages,
2230 .page_mkwrite = btrfs_page_mkwrite,
2233 static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2235 struct address_space *mapping = filp->f_mapping;
2237 if (!mapping->a_ops->readpage)
2240 file_accessed(filp);
2241 vma->vm_ops = &btrfs_file_vm_ops;
2246 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2247 int slot, u64 start, u64 end)
2249 struct btrfs_file_extent_item *fi;
2250 struct btrfs_key key;
2252 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2255 btrfs_item_key_to_cpu(leaf, &key, slot);
2256 if (key.objectid != btrfs_ino(inode) ||
2257 key.type != BTRFS_EXTENT_DATA_KEY)
2260 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2262 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2265 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2268 if (key.offset == end)
2270 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2275 static int fill_holes(struct btrfs_trans_handle *trans,
2276 struct btrfs_inode *inode,
2277 struct btrfs_path *path, u64 offset, u64 end)
2279 struct btrfs_fs_info *fs_info = trans->fs_info;
2280 struct btrfs_root *root = inode->root;
2281 struct extent_buffer *leaf;
2282 struct btrfs_file_extent_item *fi;
2283 struct extent_map *hole_em;
2284 struct extent_map_tree *em_tree = &inode->extent_tree;
2285 struct btrfs_key key;
2288 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2291 key.objectid = btrfs_ino(inode);
2292 key.type = BTRFS_EXTENT_DATA_KEY;
2293 key.offset = offset;
2295 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2298 * We should have dropped this offset, so if we find it then
2299 * something has gone horribly wrong.
2306 leaf = path->nodes[0];
2307 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2311 fi = btrfs_item_ptr(leaf, path->slots[0],
2312 struct btrfs_file_extent_item);
2313 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2315 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2316 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2317 btrfs_set_file_extent_offset(leaf, fi, 0);
2318 btrfs_mark_buffer_dirty(leaf);
2322 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2325 key.offset = offset;
2326 btrfs_set_item_key_safe(fs_info, path, &key);
2327 fi = btrfs_item_ptr(leaf, path->slots[0],
2328 struct btrfs_file_extent_item);
2329 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2331 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2332 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2333 btrfs_set_file_extent_offset(leaf, fi, 0);
2334 btrfs_mark_buffer_dirty(leaf);
2337 btrfs_release_path(path);
2339 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2340 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2345 btrfs_release_path(path);
2347 hole_em = alloc_extent_map();
2349 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2350 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2352 hole_em->start = offset;
2353 hole_em->len = end - offset;
2354 hole_em->ram_bytes = hole_em->len;
2355 hole_em->orig_start = offset;
2357 hole_em->block_start = EXTENT_MAP_HOLE;
2358 hole_em->block_len = 0;
2359 hole_em->orig_block_len = 0;
2360 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2361 hole_em->generation = trans->transid;
2364 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2365 write_lock(&em_tree->lock);
2366 ret = add_extent_mapping(em_tree, hole_em, 1);
2367 write_unlock(&em_tree->lock);
2368 } while (ret == -EEXIST);
2369 free_extent_map(hole_em);
2371 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2372 &inode->runtime_flags);
2379 * Find a hole extent on given inode and change start/len to the end of hole
2380 * extent.(hole/vacuum extent whose em->start <= start &&
2381 * em->start + em->len > start)
2382 * When a hole extent is found, return 1 and modify start/len.
2384 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2386 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2387 struct extent_map *em;
2390 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2391 round_down(*start, fs_info->sectorsize),
2392 round_up(*len, fs_info->sectorsize));
2396 /* Hole or vacuum extent(only exists in no-hole mode) */
2397 if (em->block_start == EXTENT_MAP_HOLE) {
2399 *len = em->start + em->len > *start + *len ?
2400 0 : *start + *len - em->start - em->len;
2401 *start = em->start + em->len;
2403 free_extent_map(em);
2407 static int btrfs_punch_hole_lock_range(struct inode *inode,
2408 const u64 lockstart,
2410 struct extent_state **cached_state)
2413 struct btrfs_ordered_extent *ordered;
2416 truncate_pagecache_range(inode, lockstart, lockend);
2418 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2420 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2423 * We need to make sure we have no ordered extents in this range
2424 * and nobody raced in and read a page in this range, if we did
2425 * we need to try again.
2428 (ordered->file_offset + ordered->num_bytes <= lockstart ||
2429 ordered->file_offset > lockend)) &&
2430 !filemap_range_has_page(inode->i_mapping,
2431 lockstart, lockend)) {
2433 btrfs_put_ordered_extent(ordered);
2437 btrfs_put_ordered_extent(ordered);
2438 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2439 lockend, cached_state);
2440 ret = btrfs_wait_ordered_range(inode, lockstart,
2441 lockend - lockstart + 1);
2448 static int btrfs_insert_clone_extent(struct btrfs_trans_handle *trans,
2449 struct inode *inode,
2450 struct btrfs_path *path,
2451 struct btrfs_clone_extent_info *clone_info,
2452 const u64 clone_len)
2454 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2455 struct btrfs_root *root = BTRFS_I(inode)->root;
2456 struct btrfs_file_extent_item *extent;
2457 struct extent_buffer *leaf;
2458 struct btrfs_key key;
2460 struct btrfs_ref ref = { 0 };
2467 if (clone_info->disk_offset == 0 &&
2468 btrfs_fs_incompat(fs_info, NO_HOLES))
2471 key.objectid = btrfs_ino(BTRFS_I(inode));
2472 key.type = BTRFS_EXTENT_DATA_KEY;
2473 key.offset = clone_info->file_offset;
2474 ret = btrfs_insert_empty_item(trans, root, path, &key,
2475 clone_info->item_size);
2478 leaf = path->nodes[0];
2479 slot = path->slots[0];
2480 write_extent_buffer(leaf, clone_info->extent_buf,
2481 btrfs_item_ptr_offset(leaf, slot),
2482 clone_info->item_size);
2483 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2484 btrfs_set_file_extent_offset(leaf, extent, clone_info->data_offset);
2485 btrfs_set_file_extent_num_bytes(leaf, extent, clone_len);
2486 btrfs_mark_buffer_dirty(leaf);
2487 btrfs_release_path(path);
2489 /* If it's a hole, nothing more needs to be done. */
2490 if (clone_info->disk_offset == 0)
2493 inode_add_bytes(inode, clone_len);
2494 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2495 clone_info->disk_offset,
2496 clone_info->disk_len, 0);
2497 ref_offset = clone_info->file_offset - clone_info->data_offset;
2498 btrfs_init_data_ref(&ref, root->root_key.objectid,
2499 btrfs_ino(BTRFS_I(inode)), ref_offset);
2500 ret = btrfs_inc_extent_ref(trans, &ref);
2506 * The respective range must have been previously locked, as well as the inode.
2507 * The end offset is inclusive (last byte of the range).
2508 * @clone_info is NULL for fallocate's hole punching and non-NULL for extent
2510 * When cloning, we don't want to end up in a state where we dropped extents
2511 * without inserting a new one, so we must abort the transaction to avoid a
2514 int btrfs_punch_hole_range(struct inode *inode, struct btrfs_path *path,
2515 const u64 start, const u64 end,
2516 struct btrfs_clone_extent_info *clone_info,
2517 struct btrfs_trans_handle **trans_out)
2519 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2520 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2521 u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2522 struct btrfs_root *root = BTRFS_I(inode)->root;
2523 struct btrfs_trans_handle *trans = NULL;
2524 struct btrfs_block_rsv *rsv;
2525 unsigned int rsv_count;
2528 u64 len = end - start;
2534 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2539 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2543 * 1 - update the inode
2544 * 1 - removing the extents in the range
2545 * 1 - adding the hole extent if no_holes isn't set or if we are cloning
2548 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || clone_info)
2553 trans = btrfs_start_transaction(root, rsv_count);
2554 if (IS_ERR(trans)) {
2555 ret = PTR_ERR(trans);
2560 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2563 trans->block_rsv = rsv;
2566 while (cur_offset < end) {
2567 ret = __btrfs_drop_extents(trans, root, inode, path,
2568 cur_offset, end + 1, &drop_end,
2570 if (ret != -ENOSPC) {
2572 * When cloning we want to avoid transaction aborts when
2573 * nothing was done and we are attempting to clone parts
2574 * of inline extents, in such cases -EOPNOTSUPP is
2575 * returned by __btrfs_drop_extents() without having
2576 * changed anything in the file.
2578 if (clone_info && ret && ret != -EOPNOTSUPP)
2579 btrfs_abort_transaction(trans, ret);
2583 trans->block_rsv = &fs_info->trans_block_rsv;
2585 if (!clone_info && cur_offset < drop_end &&
2586 cur_offset < ino_size) {
2587 ret = fill_holes(trans, BTRFS_I(inode), path,
2588 cur_offset, drop_end);
2591 * If we failed then we didn't insert our hole
2592 * entries for the area we dropped, so now the
2593 * fs is corrupted, so we must abort the
2596 btrfs_abort_transaction(trans, ret);
2601 if (clone_info && drop_end > clone_info->file_offset) {
2602 u64 clone_len = drop_end - clone_info->file_offset;
2604 ret = btrfs_insert_clone_extent(trans, inode, path,
2605 clone_info, clone_len);
2607 btrfs_abort_transaction(trans, ret);
2610 clone_info->data_len -= clone_len;
2611 clone_info->data_offset += clone_len;
2612 clone_info->file_offset += clone_len;
2615 cur_offset = drop_end;
2617 ret = btrfs_update_inode(trans, root, inode);
2621 btrfs_end_transaction(trans);
2622 btrfs_btree_balance_dirty(fs_info);
2624 trans = btrfs_start_transaction(root, rsv_count);
2625 if (IS_ERR(trans)) {
2626 ret = PTR_ERR(trans);
2631 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2632 rsv, min_size, false);
2633 BUG_ON(ret); /* shouldn't happen */
2634 trans->block_rsv = rsv;
2637 ret = find_first_non_hole(inode, &cur_offset, &len);
2638 if (unlikely(ret < 0))
2648 * If we were cloning, force the next fsync to be a full one since we
2649 * we replaced (or just dropped in the case of cloning holes when
2650 * NO_HOLES is enabled) extents and extent maps.
2651 * This is for the sake of simplicity, and cloning into files larger
2652 * than 16Mb would force the full fsync any way (when
2653 * try_release_extent_mapping() is invoked during page cache truncation.
2656 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2657 &BTRFS_I(inode)->runtime_flags);
2662 trans->block_rsv = &fs_info->trans_block_rsv;
2664 * If we are using the NO_HOLES feature we might have had already an
2665 * hole that overlaps a part of the region [lockstart, lockend] and
2666 * ends at (or beyond) lockend. Since we have no file extent items to
2667 * represent holes, drop_end can be less than lockend and so we must
2668 * make sure we have an extent map representing the existing hole (the
2669 * call to __btrfs_drop_extents() might have dropped the existing extent
2670 * map representing the existing hole), otherwise the fast fsync path
2671 * will not record the existence of the hole region
2672 * [existing_hole_start, lockend].
2674 if (drop_end <= end)
2677 * Don't insert file hole extent item if it's for a range beyond eof
2678 * (because it's useless) or if it represents a 0 bytes range (when
2679 * cur_offset == drop_end).
2681 if (!clone_info && cur_offset < ino_size && cur_offset < drop_end) {
2682 ret = fill_holes(trans, BTRFS_I(inode), path,
2683 cur_offset, drop_end);
2685 /* Same comment as above. */
2686 btrfs_abort_transaction(trans, ret);
2691 ret = btrfs_insert_clone_extent(trans, inode, path, clone_info,
2692 clone_info->data_len);
2694 btrfs_abort_transaction(trans, ret);
2703 trans->block_rsv = &fs_info->trans_block_rsv;
2705 btrfs_end_transaction(trans);
2709 btrfs_free_block_rsv(fs_info, rsv);
2714 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2716 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2717 struct btrfs_root *root = BTRFS_I(inode)->root;
2718 struct extent_state *cached_state = NULL;
2719 struct btrfs_path *path;
2720 struct btrfs_trans_handle *trans = NULL;
2725 u64 orig_start = offset;
2729 bool truncated_block = false;
2730 bool updated_inode = false;
2732 ret = btrfs_wait_ordered_range(inode, offset, len);
2737 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2738 ret = find_first_non_hole(inode, &offset, &len);
2740 goto out_only_mutex;
2742 /* Already in a large hole */
2744 goto out_only_mutex;
2747 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2748 lockend = round_down(offset + len,
2749 btrfs_inode_sectorsize(inode)) - 1;
2750 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2751 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2753 * We needn't truncate any block which is beyond the end of the file
2754 * because we are sure there is no data there.
2757 * Only do this if we are in the same block and we aren't doing the
2760 if (same_block && len < fs_info->sectorsize) {
2761 if (offset < ino_size) {
2762 truncated_block = true;
2763 ret = btrfs_truncate_block(inode, offset, len, 0);
2767 goto out_only_mutex;
2770 /* zero back part of the first block */
2771 if (offset < ino_size) {
2772 truncated_block = true;
2773 ret = btrfs_truncate_block(inode, offset, 0, 0);
2775 inode_unlock(inode);
2780 /* Check the aligned pages after the first unaligned page,
2781 * if offset != orig_start, which means the first unaligned page
2782 * including several following pages are already in holes,
2783 * the extra check can be skipped */
2784 if (offset == orig_start) {
2785 /* after truncate page, check hole again */
2786 len = offset + len - lockstart;
2788 ret = find_first_non_hole(inode, &offset, &len);
2790 goto out_only_mutex;
2793 goto out_only_mutex;
2798 /* Check the tail unaligned part is in a hole */
2799 tail_start = lockend + 1;
2800 tail_len = offset + len - tail_start;
2802 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2803 if (unlikely(ret < 0))
2804 goto out_only_mutex;
2806 /* zero the front end of the last page */
2807 if (tail_start + tail_len < ino_size) {
2808 truncated_block = true;
2809 ret = btrfs_truncate_block(inode,
2810 tail_start + tail_len,
2813 goto out_only_mutex;
2818 if (lockend < lockstart) {
2820 goto out_only_mutex;
2823 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2826 goto out_only_mutex;
2828 path = btrfs_alloc_path();
2834 ret = btrfs_punch_hole_range(inode, path, lockstart, lockend, NULL,
2836 btrfs_free_path(path);
2840 ASSERT(trans != NULL);
2841 inode_inc_iversion(inode);
2842 inode->i_mtime = inode->i_ctime = current_time(inode);
2843 ret = btrfs_update_inode(trans, root, inode);
2844 updated_inode = true;
2845 btrfs_end_transaction(trans);
2846 btrfs_btree_balance_dirty(fs_info);
2848 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2851 if (!updated_inode && truncated_block && !ret) {
2853 * If we only end up zeroing part of a page, we still need to
2854 * update the inode item, so that all the time fields are
2855 * updated as well as the necessary btrfs inode in memory fields
2856 * for detecting, at fsync time, if the inode isn't yet in the
2857 * log tree or it's there but not up to date.
2859 struct timespec64 now = current_time(inode);
2861 inode_inc_iversion(inode);
2862 inode->i_mtime = now;
2863 inode->i_ctime = now;
2864 trans = btrfs_start_transaction(root, 1);
2865 if (IS_ERR(trans)) {
2866 ret = PTR_ERR(trans);
2870 ret = btrfs_update_inode(trans, root, inode);
2871 ret2 = btrfs_end_transaction(trans);
2876 inode_unlock(inode);
2880 /* Helper structure to record which range is already reserved */
2881 struct falloc_range {
2882 struct list_head list;
2888 * Helper function to add falloc range
2890 * Caller should have locked the larger range of extent containing
2893 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2895 struct falloc_range *prev = NULL;
2896 struct falloc_range *range = NULL;
2898 if (list_empty(head))
2902 * As fallocate iterate by bytenr order, we only need to check
2905 prev = list_entry(head->prev, struct falloc_range, list);
2906 if (prev->start + prev->len == start) {
2911 range = kmalloc(sizeof(*range), GFP_KERNEL);
2914 range->start = start;
2916 list_add_tail(&range->list, head);
2920 static int btrfs_fallocate_update_isize(struct inode *inode,
2924 struct btrfs_trans_handle *trans;
2925 struct btrfs_root *root = BTRFS_I(inode)->root;
2929 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2932 trans = btrfs_start_transaction(root, 1);
2934 return PTR_ERR(trans);
2936 inode->i_ctime = current_time(inode);
2937 i_size_write(inode, end);
2938 btrfs_ordered_update_i_size(inode, end, NULL);
2939 ret = btrfs_update_inode(trans, root, inode);
2940 ret2 = btrfs_end_transaction(trans);
2942 return ret ? ret : ret2;
2946 RANGE_BOUNDARY_WRITTEN_EXTENT,
2947 RANGE_BOUNDARY_PREALLOC_EXTENT,
2948 RANGE_BOUNDARY_HOLE,
2951 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2954 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2955 struct extent_map *em;
2958 offset = round_down(offset, sectorsize);
2959 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize);
2963 if (em->block_start == EXTENT_MAP_HOLE)
2964 ret = RANGE_BOUNDARY_HOLE;
2965 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2966 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2968 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2970 free_extent_map(em);
2974 static int btrfs_zero_range(struct inode *inode,
2979 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2980 struct extent_map *em;
2981 struct extent_changeset *data_reserved = NULL;
2984 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2985 u64 alloc_start = round_down(offset, sectorsize);
2986 u64 alloc_end = round_up(offset + len, sectorsize);
2987 u64 bytes_to_reserve = 0;
2988 bool space_reserved = false;
2990 inode_dio_wait(inode);
2992 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
2993 alloc_end - alloc_start);
3000 * Avoid hole punching and extent allocation for some cases. More cases
3001 * could be considered, but these are unlikely common and we keep things
3002 * as simple as possible for now. Also, intentionally, if the target
3003 * range contains one or more prealloc extents together with regular
3004 * extents and holes, we drop all the existing extents and allocate a
3005 * new prealloc extent, so that we get a larger contiguous disk extent.
3007 if (em->start <= alloc_start &&
3008 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3009 const u64 em_end = em->start + em->len;
3011 if (em_end >= offset + len) {
3013 * The whole range is already a prealloc extent,
3014 * do nothing except updating the inode's i_size if
3017 free_extent_map(em);
3018 ret = btrfs_fallocate_update_isize(inode, offset + len,
3023 * Part of the range is already a prealloc extent, so operate
3024 * only on the remaining part of the range.
3026 alloc_start = em_end;
3027 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3028 len = offset + len - alloc_start;
3029 offset = alloc_start;
3030 alloc_hint = em->block_start + em->len;
3032 free_extent_map(em);
3034 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3035 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3036 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3043 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3044 free_extent_map(em);
3045 ret = btrfs_fallocate_update_isize(inode, offset + len,
3049 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3050 free_extent_map(em);
3051 ret = btrfs_truncate_block(inode, offset, len, 0);
3053 ret = btrfs_fallocate_update_isize(inode,
3058 free_extent_map(em);
3059 alloc_start = round_down(offset, sectorsize);
3060 alloc_end = alloc_start + sectorsize;
3064 alloc_start = round_up(offset, sectorsize);
3065 alloc_end = round_down(offset + len, sectorsize);
3068 * For unaligned ranges, check the pages at the boundaries, they might
3069 * map to an extent, in which case we need to partially zero them, or
3070 * they might map to a hole, in which case we need our allocation range
3073 if (!IS_ALIGNED(offset, sectorsize)) {
3074 ret = btrfs_zero_range_check_range_boundary(inode, offset);
3077 if (ret == RANGE_BOUNDARY_HOLE) {
3078 alloc_start = round_down(offset, sectorsize);
3080 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3081 ret = btrfs_truncate_block(inode, offset, 0, 0);
3089 if (!IS_ALIGNED(offset + len, sectorsize)) {
3090 ret = btrfs_zero_range_check_range_boundary(inode,
3094 if (ret == RANGE_BOUNDARY_HOLE) {
3095 alloc_end = round_up(offset + len, sectorsize);
3097 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3098 ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3107 if (alloc_start < alloc_end) {
3108 struct extent_state *cached_state = NULL;
3109 const u64 lockstart = alloc_start;
3110 const u64 lockend = alloc_end - 1;
3112 bytes_to_reserve = alloc_end - alloc_start;
3113 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3117 space_reserved = true;
3118 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3119 alloc_start, bytes_to_reserve);
3122 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3126 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3127 alloc_end - alloc_start,
3129 offset + len, &alloc_hint);
3130 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3131 lockend, &cached_state);
3132 /* btrfs_prealloc_file_range releases reserved space on error */
3134 space_reserved = false;
3138 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3140 if (ret && space_reserved)
3141 btrfs_free_reserved_data_space(inode, data_reserved,
3142 alloc_start, bytes_to_reserve);
3143 extent_changeset_free(data_reserved);
3148 static long btrfs_fallocate(struct file *file, int mode,
3149 loff_t offset, loff_t len)
3151 struct inode *inode = file_inode(file);
3152 struct extent_state *cached_state = NULL;
3153 struct extent_changeset *data_reserved = NULL;
3154 struct falloc_range *range;
3155 struct falloc_range *tmp;
3156 struct list_head reserve_list;
3164 struct extent_map *em;
3165 int blocksize = btrfs_inode_sectorsize(inode);
3168 alloc_start = round_down(offset, blocksize);
3169 alloc_end = round_up(offset + len, blocksize);
3170 cur_offset = alloc_start;
3172 /* Make sure we aren't being give some crap mode */
3173 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3174 FALLOC_FL_ZERO_RANGE))
3177 if (mode & FALLOC_FL_PUNCH_HOLE)
3178 return btrfs_punch_hole(inode, offset, len);
3181 * Only trigger disk allocation, don't trigger qgroup reserve
3183 * For qgroup space, it will be checked later.
3185 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3186 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3187 alloc_end - alloc_start);
3194 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3195 ret = inode_newsize_ok(inode, offset + len);
3201 * TODO: Move these two operations after we have checked
3202 * accurate reserved space, or fallocate can still fail but
3203 * with page truncated or size expanded.
3205 * But that's a minor problem and won't do much harm BTW.
3207 if (alloc_start > inode->i_size) {
3208 ret = btrfs_cont_expand(inode, i_size_read(inode),
3212 } else if (offset + len > inode->i_size) {
3214 * If we are fallocating from the end of the file onward we
3215 * need to zero out the end of the block if i_size lands in the
3216 * middle of a block.
3218 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3224 * wait for ordered IO before we have any locks. We'll loop again
3225 * below with the locks held.
3227 ret = btrfs_wait_ordered_range(inode, alloc_start,
3228 alloc_end - alloc_start);
3232 if (mode & FALLOC_FL_ZERO_RANGE) {
3233 ret = btrfs_zero_range(inode, offset, len, mode);
3234 inode_unlock(inode);
3238 locked_end = alloc_end - 1;
3240 struct btrfs_ordered_extent *ordered;
3242 /* the extent lock is ordered inside the running
3245 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3246 locked_end, &cached_state);
3247 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3250 ordered->file_offset + ordered->num_bytes > alloc_start &&
3251 ordered->file_offset < alloc_end) {
3252 btrfs_put_ordered_extent(ordered);
3253 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3254 alloc_start, locked_end,
3257 * we can't wait on the range with the transaction
3258 * running or with the extent lock held
3260 ret = btrfs_wait_ordered_range(inode, alloc_start,
3261 alloc_end - alloc_start);
3266 btrfs_put_ordered_extent(ordered);
3271 /* First, check if we exceed the qgroup limit */
3272 INIT_LIST_HEAD(&reserve_list);
3273 while (cur_offset < alloc_end) {
3274 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3275 alloc_end - cur_offset);
3280 last_byte = min(extent_map_end(em), alloc_end);
3281 actual_end = min_t(u64, extent_map_end(em), offset + len);
3282 last_byte = ALIGN(last_byte, blocksize);
3283 if (em->block_start == EXTENT_MAP_HOLE ||
3284 (cur_offset >= inode->i_size &&
3285 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3286 ret = add_falloc_range(&reserve_list, cur_offset,
3287 last_byte - cur_offset);
3289 free_extent_map(em);
3292 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3293 cur_offset, last_byte - cur_offset);
3295 cur_offset = last_byte;
3296 free_extent_map(em);
3301 * Do not need to reserve unwritten extent for this
3302 * range, free reserved data space first, otherwise
3303 * it'll result in false ENOSPC error.
3305 btrfs_free_reserved_data_space(inode, data_reserved,
3306 cur_offset, last_byte - cur_offset);
3308 free_extent_map(em);
3309 cur_offset = last_byte;
3313 * If ret is still 0, means we're OK to fallocate.
3314 * Or just cleanup the list and exit.
3316 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3318 ret = btrfs_prealloc_file_range(inode, mode,
3320 range->len, i_blocksize(inode),
3321 offset + len, &alloc_hint);
3323 btrfs_free_reserved_data_space(inode,
3324 data_reserved, range->start,
3326 list_del(&range->list);
3333 * We didn't need to allocate any more space, but we still extended the
3334 * size of the file so we need to update i_size and the inode item.
3336 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3338 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3341 inode_unlock(inode);
3342 /* Let go of our reservation. */
3343 if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3344 btrfs_free_reserved_data_space(inode, data_reserved,
3345 cur_offset, alloc_end - cur_offset);
3346 extent_changeset_free(data_reserved);
3350 static loff_t find_desired_extent(struct inode *inode, loff_t offset,
3353 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3354 struct extent_map *em = NULL;
3355 struct extent_state *cached_state = NULL;
3356 loff_t i_size = inode->i_size;
3363 if (i_size == 0 || offset >= i_size)
3367 * offset can be negative, in this case we start finding DATA/HOLE from
3368 * the very start of the file.
3370 start = max_t(loff_t, 0, offset);
3372 lockstart = round_down(start, fs_info->sectorsize);
3373 lockend = round_up(i_size, fs_info->sectorsize);
3374 if (lockend <= lockstart)
3375 lockend = lockstart + fs_info->sectorsize;
3377 len = lockend - lockstart + 1;
3379 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3382 while (start < i_size) {
3383 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3390 if (whence == SEEK_HOLE &&
3391 (em->block_start == EXTENT_MAP_HOLE ||
3392 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3394 else if (whence == SEEK_DATA &&
3395 (em->block_start != EXTENT_MAP_HOLE &&
3396 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3399 start = em->start + em->len;
3400 free_extent_map(em);
3404 free_extent_map(em);
3405 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3410 if (whence == SEEK_DATA && start >= i_size)
3413 offset = min_t(loff_t, start, i_size);
3419 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3421 struct inode *inode = file->f_mapping->host;
3425 return generic_file_llseek(file, offset, whence);
3428 inode_lock_shared(inode);
3429 offset = find_desired_extent(inode, offset, whence);
3430 inode_unlock_shared(inode);
3437 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3440 static int btrfs_file_open(struct inode *inode, struct file *filp)
3442 filp->f_mode |= FMODE_NOWAIT;
3443 return generic_file_open(inode, filp);
3446 const struct file_operations btrfs_file_operations = {
3447 .llseek = btrfs_file_llseek,
3448 .read_iter = generic_file_read_iter,
3449 .splice_read = generic_file_splice_read,
3450 .write_iter = btrfs_file_write_iter,
3451 .mmap = btrfs_file_mmap,
3452 .open = btrfs_file_open,
3453 .release = btrfs_release_file,
3454 .fsync = btrfs_sync_file,
3455 .fallocate = btrfs_fallocate,
3456 .unlocked_ioctl = btrfs_ioctl,
3457 #ifdef CONFIG_COMPAT
3458 .compat_ioctl = btrfs_compat_ioctl,
3460 .remap_file_range = btrfs_remap_file_range,
3463 void __cold btrfs_auto_defrag_exit(void)
3465 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3468 int __init btrfs_auto_defrag_init(void)
3470 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3471 sizeof(struct inode_defrag), 0,
3474 if (!btrfs_inode_defrag_cachep)
3480 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3485 * So with compression we will find and lock a dirty page and clear the
3486 * first one as dirty, setup an async extent, and immediately return
3487 * with the entire range locked but with nobody actually marked with
3488 * writeback. So we can't just filemap_write_and_wait_range() and
3489 * expect it to work since it will just kick off a thread to do the
3490 * actual work. So we need to call filemap_fdatawrite_range _again_
3491 * since it will wait on the page lock, which won't be unlocked until
3492 * after the pages have been marked as writeback and so we're good to go
3493 * from there. We have to do this otherwise we'll miss the ordered
3494 * extents and that results in badness. Please Josef, do not think you
3495 * know better and pull this out at some point in the future, it is
3496 * right and you are wrong.
3498 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3499 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3500 &BTRFS_I(inode)->runtime_flags))
3501 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);