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, NULL);
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,
485 if (em->block_start != EXTENT_MAP_HOLE)
489 if (em->start < search_start)
490 em_len -= search_start - em->start;
491 if (em_len > search_len)
494 ret = set_extent_bit(&inode->io_tree, search_start,
495 search_start + em_len - 1,
497 NULL, cached_state, GFP_NOFS);
499 search_start = extent_map_end(em);
508 * after copy_from_user, pages need to be dirtied and we need to make
509 * sure holes are created between the current EOF and the start of
510 * any next extents (if required).
512 * this also makes the decision about creating an inline extent vs
513 * doing real data extents, marking pages dirty and delalloc as required.
515 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
516 size_t num_pages, loff_t pos, size_t write_bytes,
517 struct extent_state **cached)
519 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
524 u64 end_of_last_block;
525 u64 end_pos = pos + write_bytes;
526 loff_t isize = i_size_read(inode);
527 unsigned int extra_bits = 0;
529 start_pos = pos & ~((u64) fs_info->sectorsize - 1);
530 num_bytes = round_up(write_bytes + pos - start_pos,
531 fs_info->sectorsize);
533 end_of_last_block = start_pos + num_bytes - 1;
536 * The pages may have already been dirty, clear out old accounting so
537 * we can set things up properly
539 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos, end_of_last_block,
540 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
543 if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
544 if (start_pos >= isize &&
545 !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
547 * There can't be any extents following eof in this case
548 * so just set the delalloc new bit for the range
551 extra_bits |= EXTENT_DELALLOC_NEW;
553 err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
561 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
566 for (i = 0; i < num_pages; i++) {
567 struct page *p = pages[i];
574 * we've only changed i_size in ram, and we haven't updated
575 * the disk i_size. There is no need to log the inode
579 i_size_write(inode, end_pos);
584 * this drops all the extents in the cache that intersect the range
585 * [start, end]. Existing extents are split as required.
587 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
590 struct extent_map *em;
591 struct extent_map *split = NULL;
592 struct extent_map *split2 = NULL;
593 struct extent_map_tree *em_tree = &inode->extent_tree;
594 u64 len = end - start + 1;
602 WARN_ON(end < start);
603 if (end == (u64)-1) {
612 split = alloc_extent_map();
614 split2 = alloc_extent_map();
615 if (!split || !split2)
618 write_lock(&em_tree->lock);
619 em = lookup_extent_mapping(em_tree, start, len);
621 write_unlock(&em_tree->lock);
625 gen = em->generation;
626 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
627 if (testend && em->start + em->len >= start + len) {
629 write_unlock(&em_tree->lock);
632 start = em->start + em->len;
634 len = start + len - (em->start + em->len);
636 write_unlock(&em_tree->lock);
639 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
640 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
641 clear_bit(EXTENT_FLAG_LOGGING, &flags);
642 modified = !list_empty(&em->list);
646 if (em->start < start) {
647 split->start = em->start;
648 split->len = start - em->start;
650 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
651 split->orig_start = em->orig_start;
652 split->block_start = em->block_start;
655 split->block_len = em->block_len;
657 split->block_len = split->len;
658 split->orig_block_len = max(split->block_len,
660 split->ram_bytes = em->ram_bytes;
662 split->orig_start = split->start;
663 split->block_len = 0;
664 split->block_start = em->block_start;
665 split->orig_block_len = 0;
666 split->ram_bytes = split->len;
669 split->generation = gen;
670 split->bdev = em->bdev;
671 split->flags = flags;
672 split->compress_type = em->compress_type;
673 replace_extent_mapping(em_tree, em, split, modified);
674 free_extent_map(split);
678 if (testend && em->start + em->len > start + len) {
679 u64 diff = start + len - em->start;
681 split->start = start + len;
682 split->len = em->start + em->len - (start + len);
683 split->bdev = em->bdev;
684 split->flags = flags;
685 split->compress_type = em->compress_type;
686 split->generation = gen;
688 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
689 split->orig_block_len = max(em->block_len,
692 split->ram_bytes = em->ram_bytes;
694 split->block_len = em->block_len;
695 split->block_start = em->block_start;
696 split->orig_start = em->orig_start;
698 split->block_len = split->len;
699 split->block_start = em->block_start
701 split->orig_start = em->orig_start;
704 split->ram_bytes = split->len;
705 split->orig_start = split->start;
706 split->block_len = 0;
707 split->block_start = em->block_start;
708 split->orig_block_len = 0;
711 if (extent_map_in_tree(em)) {
712 replace_extent_mapping(em_tree, em, split,
715 ret = add_extent_mapping(em_tree, split,
717 ASSERT(ret == 0); /* Logic error */
719 free_extent_map(split);
723 if (extent_map_in_tree(em))
724 remove_extent_mapping(em_tree, em);
725 write_unlock(&em_tree->lock);
729 /* once for the tree*/
733 free_extent_map(split);
735 free_extent_map(split2);
739 * this is very complex, but the basic idea is to drop all extents
740 * in the range start - end. hint_block is filled in with a block number
741 * that would be a good hint to the block allocator for this file.
743 * If an extent intersects the range but is not entirely inside the range
744 * it is either truncated or split. Anything entirely inside the range
745 * is deleted from the tree.
747 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
748 struct btrfs_root *root, struct inode *inode,
749 struct btrfs_path *path, u64 start, u64 end,
750 u64 *drop_end, int drop_cache,
752 u32 extent_item_size,
755 struct btrfs_fs_info *fs_info = root->fs_info;
756 struct extent_buffer *leaf;
757 struct btrfs_file_extent_item *fi;
758 struct btrfs_ref ref = { 0 };
759 struct btrfs_key key;
760 struct btrfs_key new_key;
761 u64 ino = btrfs_ino(BTRFS_I(inode));
762 u64 search_start = start;
765 u64 extent_offset = 0;
767 u64 last_end = start;
773 int modify_tree = -1;
776 int leafs_visited = 0;
779 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
781 if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
784 update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
785 root == fs_info->tree_root);
788 ret = btrfs_lookup_file_extent(trans, root, path, ino,
789 search_start, modify_tree);
792 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
793 leaf = path->nodes[0];
794 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
795 if (key.objectid == ino &&
796 key.type == BTRFS_EXTENT_DATA_KEY)
802 leaf = path->nodes[0];
803 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
805 ret = btrfs_next_leaf(root, path);
813 leaf = path->nodes[0];
817 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
819 if (key.objectid > ino)
821 if (WARN_ON_ONCE(key.objectid < ino) ||
822 key.type < BTRFS_EXTENT_DATA_KEY) {
827 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
830 fi = btrfs_item_ptr(leaf, path->slots[0],
831 struct btrfs_file_extent_item);
832 extent_type = btrfs_file_extent_type(leaf, fi);
834 if (extent_type == BTRFS_FILE_EXTENT_REG ||
835 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
836 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
837 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
838 extent_offset = btrfs_file_extent_offset(leaf, fi);
839 extent_end = key.offset +
840 btrfs_file_extent_num_bytes(leaf, fi);
841 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
842 extent_end = key.offset +
843 btrfs_file_extent_ram_bytes(leaf, fi);
850 * Don't skip extent items representing 0 byte lengths. They
851 * used to be created (bug) if while punching holes we hit
852 * -ENOSPC condition. So if we find one here, just ensure we
853 * delete it, otherwise we would insert a new file extent item
854 * with the same key (offset) as that 0 bytes length file
855 * extent item in the call to setup_items_for_insert() later
858 if (extent_end == key.offset && extent_end >= search_start) {
859 last_end = extent_end;
860 goto delete_extent_item;
863 if (extent_end <= search_start) {
869 search_start = max(key.offset, start);
870 if (recow || !modify_tree) {
872 btrfs_release_path(path);
877 * | - range to drop - |
878 * | -------- extent -------- |
880 if (start > key.offset && end < extent_end) {
882 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
887 memcpy(&new_key, &key, sizeof(new_key));
888 new_key.offset = start;
889 ret = btrfs_duplicate_item(trans, root, path,
891 if (ret == -EAGAIN) {
892 btrfs_release_path(path);
898 leaf = path->nodes[0];
899 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
900 struct btrfs_file_extent_item);
901 btrfs_set_file_extent_num_bytes(leaf, fi,
904 fi = btrfs_item_ptr(leaf, path->slots[0],
905 struct btrfs_file_extent_item);
907 extent_offset += start - key.offset;
908 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
909 btrfs_set_file_extent_num_bytes(leaf, fi,
911 btrfs_mark_buffer_dirty(leaf);
913 if (update_refs && disk_bytenr > 0) {
914 btrfs_init_generic_ref(&ref,
915 BTRFS_ADD_DELAYED_REF,
916 disk_bytenr, num_bytes, 0);
917 btrfs_init_data_ref(&ref,
918 root->root_key.objectid,
920 start - extent_offset);
921 ret = btrfs_inc_extent_ref(trans, &ref);
922 BUG_ON(ret); /* -ENOMEM */
927 * From here on out we will have actually dropped something, so
928 * last_end can be updated.
930 last_end = extent_end;
933 * | ---- range to drop ----- |
934 * | -------- extent -------- |
936 if (start <= key.offset && end < extent_end) {
937 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
942 memcpy(&new_key, &key, sizeof(new_key));
943 new_key.offset = end;
944 btrfs_set_item_key_safe(fs_info, path, &new_key);
946 extent_offset += end - key.offset;
947 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
948 btrfs_set_file_extent_num_bytes(leaf, fi,
950 btrfs_mark_buffer_dirty(leaf);
951 if (update_refs && disk_bytenr > 0)
952 inode_sub_bytes(inode, end - key.offset);
956 search_start = extent_end;
958 * | ---- range to drop ----- |
959 * | -------- extent -------- |
961 if (start > key.offset && end >= extent_end) {
963 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
968 btrfs_set_file_extent_num_bytes(leaf, fi,
970 btrfs_mark_buffer_dirty(leaf);
971 if (update_refs && disk_bytenr > 0)
972 inode_sub_bytes(inode, extent_end - start);
973 if (end == extent_end)
981 * | ---- range to drop ----- |
982 * | ------ extent ------ |
984 if (start <= key.offset && end >= extent_end) {
987 del_slot = path->slots[0];
990 BUG_ON(del_slot + del_nr != path->slots[0]);
995 extent_type == BTRFS_FILE_EXTENT_INLINE) {
996 inode_sub_bytes(inode,
997 extent_end - key.offset);
998 extent_end = ALIGN(extent_end,
999 fs_info->sectorsize);
1000 } else if (update_refs && disk_bytenr > 0) {
1001 btrfs_init_generic_ref(&ref,
1002 BTRFS_DROP_DELAYED_REF,
1003 disk_bytenr, num_bytes, 0);
1004 btrfs_init_data_ref(&ref,
1005 root->root_key.objectid,
1007 key.offset - extent_offset);
1008 ret = btrfs_free_extent(trans, &ref);
1009 BUG_ON(ret); /* -ENOMEM */
1010 inode_sub_bytes(inode,
1011 extent_end - key.offset);
1014 if (end == extent_end)
1017 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1022 ret = btrfs_del_items(trans, root, path, del_slot,
1025 btrfs_abort_transaction(trans, ret);
1032 btrfs_release_path(path);
1039 if (!ret && del_nr > 0) {
1041 * Set path->slots[0] to first slot, so that after the delete
1042 * if items are move off from our leaf to its immediate left or
1043 * right neighbor leafs, we end up with a correct and adjusted
1044 * path->slots[0] for our insertion (if replace_extent != 0).
1046 path->slots[0] = del_slot;
1047 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1049 btrfs_abort_transaction(trans, ret);
1052 leaf = path->nodes[0];
1054 * If btrfs_del_items() was called, it might have deleted a leaf, in
1055 * which case it unlocked our path, so check path->locks[0] matches a
1058 if (!ret && replace_extent && leafs_visited == 1 &&
1059 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1060 path->locks[0] == BTRFS_WRITE_LOCK) &&
1061 btrfs_leaf_free_space(leaf) >=
1062 sizeof(struct btrfs_item) + extent_item_size) {
1065 key.type = BTRFS_EXTENT_DATA_KEY;
1067 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1068 struct btrfs_key slot_key;
1070 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1071 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1074 setup_items_for_insert(root, path, &key,
1077 sizeof(struct btrfs_item) +
1078 extent_item_size, 1);
1082 if (!replace_extent || !(*key_inserted))
1083 btrfs_release_path(path);
1085 *drop_end = found ? min(end, last_end) : end;
1089 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1090 struct btrfs_root *root, struct inode *inode, u64 start,
1091 u64 end, int drop_cache)
1093 struct btrfs_path *path;
1096 path = btrfs_alloc_path();
1099 ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1100 drop_cache, 0, 0, NULL);
1101 btrfs_free_path(path);
1105 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1106 u64 objectid, u64 bytenr, u64 orig_offset,
1107 u64 *start, u64 *end)
1109 struct btrfs_file_extent_item *fi;
1110 struct btrfs_key key;
1113 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1116 btrfs_item_key_to_cpu(leaf, &key, slot);
1117 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1120 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1121 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1122 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1123 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1124 btrfs_file_extent_compression(leaf, fi) ||
1125 btrfs_file_extent_encryption(leaf, fi) ||
1126 btrfs_file_extent_other_encoding(leaf, fi))
1129 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1130 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1133 *start = key.offset;
1139 * Mark extent in the range start - end as written.
1141 * This changes extent type from 'pre-allocated' to 'regular'. If only
1142 * part of extent is marked as written, the extent will be split into
1145 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1146 struct btrfs_inode *inode, u64 start, u64 end)
1148 struct btrfs_fs_info *fs_info = trans->fs_info;
1149 struct btrfs_root *root = inode->root;
1150 struct extent_buffer *leaf;
1151 struct btrfs_path *path;
1152 struct btrfs_file_extent_item *fi;
1153 struct btrfs_ref ref = { 0 };
1154 struct btrfs_key key;
1155 struct btrfs_key new_key;
1167 u64 ino = btrfs_ino(inode);
1169 path = btrfs_alloc_path();
1176 key.type = BTRFS_EXTENT_DATA_KEY;
1179 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1182 if (ret > 0 && path->slots[0] > 0)
1185 leaf = path->nodes[0];
1186 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1187 if (key.objectid != ino ||
1188 key.type != BTRFS_EXTENT_DATA_KEY) {
1190 btrfs_abort_transaction(trans, ret);
1193 fi = btrfs_item_ptr(leaf, path->slots[0],
1194 struct btrfs_file_extent_item);
1195 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1197 btrfs_abort_transaction(trans, ret);
1200 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1201 if (key.offset > start || extent_end < end) {
1203 btrfs_abort_transaction(trans, ret);
1207 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1208 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1209 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1210 memcpy(&new_key, &key, sizeof(new_key));
1212 if (start == key.offset && end < extent_end) {
1215 if (extent_mergeable(leaf, path->slots[0] - 1,
1216 ino, bytenr, orig_offset,
1217 &other_start, &other_end)) {
1218 new_key.offset = end;
1219 btrfs_set_item_key_safe(fs_info, path, &new_key);
1220 fi = btrfs_item_ptr(leaf, path->slots[0],
1221 struct btrfs_file_extent_item);
1222 btrfs_set_file_extent_generation(leaf, fi,
1224 btrfs_set_file_extent_num_bytes(leaf, fi,
1226 btrfs_set_file_extent_offset(leaf, fi,
1228 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1229 struct btrfs_file_extent_item);
1230 btrfs_set_file_extent_generation(leaf, fi,
1232 btrfs_set_file_extent_num_bytes(leaf, fi,
1234 btrfs_mark_buffer_dirty(leaf);
1239 if (start > key.offset && end == extent_end) {
1242 if (extent_mergeable(leaf, path->slots[0] + 1,
1243 ino, bytenr, orig_offset,
1244 &other_start, &other_end)) {
1245 fi = btrfs_item_ptr(leaf, path->slots[0],
1246 struct btrfs_file_extent_item);
1247 btrfs_set_file_extent_num_bytes(leaf, fi,
1248 start - key.offset);
1249 btrfs_set_file_extent_generation(leaf, fi,
1252 new_key.offset = start;
1253 btrfs_set_item_key_safe(fs_info, path, &new_key);
1255 fi = btrfs_item_ptr(leaf, path->slots[0],
1256 struct btrfs_file_extent_item);
1257 btrfs_set_file_extent_generation(leaf, fi,
1259 btrfs_set_file_extent_num_bytes(leaf, fi,
1261 btrfs_set_file_extent_offset(leaf, fi,
1262 start - orig_offset);
1263 btrfs_mark_buffer_dirty(leaf);
1268 while (start > key.offset || end < extent_end) {
1269 if (key.offset == start)
1272 new_key.offset = split;
1273 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1274 if (ret == -EAGAIN) {
1275 btrfs_release_path(path);
1279 btrfs_abort_transaction(trans, ret);
1283 leaf = path->nodes[0];
1284 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1285 struct btrfs_file_extent_item);
1286 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1287 btrfs_set_file_extent_num_bytes(leaf, fi,
1288 split - key.offset);
1290 fi = btrfs_item_ptr(leaf, path->slots[0],
1291 struct btrfs_file_extent_item);
1293 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1294 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1295 btrfs_set_file_extent_num_bytes(leaf, fi,
1296 extent_end - split);
1297 btrfs_mark_buffer_dirty(leaf);
1299 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1301 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1303 ret = btrfs_inc_extent_ref(trans, &ref);
1305 btrfs_abort_transaction(trans, ret);
1309 if (split == start) {
1312 if (start != key.offset) {
1314 btrfs_abort_transaction(trans, ret);
1325 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1327 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1328 if (extent_mergeable(leaf, path->slots[0] + 1,
1329 ino, bytenr, orig_offset,
1330 &other_start, &other_end)) {
1332 btrfs_release_path(path);
1335 extent_end = other_end;
1336 del_slot = path->slots[0] + 1;
1338 ret = btrfs_free_extent(trans, &ref);
1340 btrfs_abort_transaction(trans, ret);
1346 if (extent_mergeable(leaf, path->slots[0] - 1,
1347 ino, bytenr, orig_offset,
1348 &other_start, &other_end)) {
1350 btrfs_release_path(path);
1353 key.offset = other_start;
1354 del_slot = path->slots[0];
1356 ret = btrfs_free_extent(trans, &ref);
1358 btrfs_abort_transaction(trans, ret);
1363 fi = btrfs_item_ptr(leaf, path->slots[0],
1364 struct btrfs_file_extent_item);
1365 btrfs_set_file_extent_type(leaf, fi,
1366 BTRFS_FILE_EXTENT_REG);
1367 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1368 btrfs_mark_buffer_dirty(leaf);
1370 fi = btrfs_item_ptr(leaf, del_slot - 1,
1371 struct btrfs_file_extent_item);
1372 btrfs_set_file_extent_type(leaf, fi,
1373 BTRFS_FILE_EXTENT_REG);
1374 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1375 btrfs_set_file_extent_num_bytes(leaf, fi,
1376 extent_end - key.offset);
1377 btrfs_mark_buffer_dirty(leaf);
1379 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1381 btrfs_abort_transaction(trans, ret);
1386 btrfs_free_path(path);
1391 * on error we return an unlocked page and the error value
1392 * on success we return a locked page and 0
1394 static int prepare_uptodate_page(struct inode *inode,
1395 struct page *page, u64 pos,
1396 bool force_uptodate)
1400 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1401 !PageUptodate(page)) {
1402 ret = btrfs_readpage(NULL, page);
1406 if (!PageUptodate(page)) {
1410 if (page->mapping != inode->i_mapping) {
1419 * this just gets pages into the page cache and locks them down.
1421 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1422 size_t num_pages, loff_t pos,
1423 size_t write_bytes, bool force_uptodate)
1426 unsigned long index = pos >> PAGE_SHIFT;
1427 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1431 for (i = 0; i < num_pages; i++) {
1433 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1434 mask | __GFP_WRITE);
1442 err = prepare_uptodate_page(inode, pages[i], pos,
1444 if (!err && i == num_pages - 1)
1445 err = prepare_uptodate_page(inode, pages[i],
1446 pos + write_bytes, false);
1449 if (err == -EAGAIN) {
1456 wait_on_page_writeback(pages[i]);
1461 while (faili >= 0) {
1462 unlock_page(pages[faili]);
1463 put_page(pages[faili]);
1471 * This function locks the extent and properly waits for data=ordered extents
1472 * to finish before allowing the pages to be modified if need.
1475 * 1 - the extent is locked
1476 * 0 - the extent is not locked, and everything is OK
1477 * -EAGAIN - need re-prepare the pages
1478 * the other < 0 number - Something wrong happens
1481 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1482 size_t num_pages, loff_t pos,
1484 u64 *lockstart, u64 *lockend,
1485 struct extent_state **cached_state)
1487 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1493 start_pos = round_down(pos, fs_info->sectorsize);
1494 last_pos = start_pos
1495 + round_up(pos + write_bytes - start_pos,
1496 fs_info->sectorsize) - 1;
1498 if (start_pos < inode->vfs_inode.i_size) {
1499 struct btrfs_ordered_extent *ordered;
1501 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1503 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1504 last_pos - start_pos + 1);
1506 ordered->file_offset + ordered->len > start_pos &&
1507 ordered->file_offset <= last_pos) {
1508 unlock_extent_cached(&inode->io_tree, start_pos,
1509 last_pos, cached_state);
1510 for (i = 0; i < num_pages; i++) {
1511 unlock_page(pages[i]);
1514 btrfs_start_ordered_extent(&inode->vfs_inode,
1516 btrfs_put_ordered_extent(ordered);
1520 btrfs_put_ordered_extent(ordered);
1522 *lockstart = start_pos;
1523 *lockend = last_pos;
1528 * It's possible the pages are dirty right now, but we don't want
1529 * to clean them yet because copy_from_user may catch a page fault
1530 * and we might have to fall back to one page at a time. If that
1531 * happens, we'll unlock these pages and we'd have a window where
1532 * reclaim could sneak in and drop the once-dirty page on the floor
1533 * without writing it.
1535 * We have the pages locked and the extent range locked, so there's
1536 * no way someone can start IO on any dirty pages in this range.
1538 * We'll call btrfs_dirty_pages() later on, and that will flip around
1539 * delalloc bits and dirty the pages as required.
1541 for (i = 0; i < num_pages; i++) {
1542 set_page_extent_mapped(pages[i]);
1543 WARN_ON(!PageLocked(pages[i]));
1549 int btrfs_check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1550 size_t *write_bytes)
1552 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1553 struct btrfs_root *root = inode->root;
1554 u64 lockstart, lockend;
1558 ret = btrfs_start_write_no_snapshotting(root);
1562 lockstart = round_down(pos, fs_info->sectorsize);
1563 lockend = round_up(pos + *write_bytes,
1564 fs_info->sectorsize) - 1;
1566 btrfs_lock_and_flush_ordered_range(&inode->io_tree, inode, lockstart,
1569 num_bytes = lockend - lockstart + 1;
1570 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1571 NULL, NULL, NULL, false);
1574 btrfs_end_write_no_snapshotting(root);
1576 *write_bytes = min_t(size_t, *write_bytes ,
1577 num_bytes - pos + lockstart);
1580 unlock_extent(&inode->io_tree, lockstart, lockend);
1585 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1588 struct file *file = iocb->ki_filp;
1589 loff_t pos = iocb->ki_pos;
1590 struct inode *inode = file_inode(file);
1591 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1592 struct btrfs_root *root = BTRFS_I(inode)->root;
1593 struct page **pages = NULL;
1594 struct extent_changeset *data_reserved = NULL;
1595 u64 release_bytes = 0;
1598 size_t num_written = 0;
1601 bool only_release_metadata = false;
1602 bool force_page_uptodate = false;
1604 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1605 PAGE_SIZE / (sizeof(struct page *)));
1606 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1607 nrptrs = max(nrptrs, 8);
1608 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1612 while (iov_iter_count(i) > 0) {
1613 struct extent_state *cached_state = NULL;
1614 size_t offset = offset_in_page(pos);
1615 size_t sector_offset;
1616 size_t write_bytes = min(iov_iter_count(i),
1617 nrptrs * (size_t)PAGE_SIZE -
1619 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1621 size_t reserve_bytes;
1624 size_t dirty_sectors;
1628 WARN_ON(num_pages > nrptrs);
1631 * Fault pages before locking them in prepare_pages
1632 * to avoid recursive lock
1634 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1639 only_release_metadata = false;
1640 sector_offset = pos & (fs_info->sectorsize - 1);
1641 reserve_bytes = round_up(write_bytes + sector_offset,
1642 fs_info->sectorsize);
1644 extent_changeset_release(data_reserved);
1645 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1648 if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1649 BTRFS_INODE_PREALLOC)) &&
1650 btrfs_check_can_nocow(BTRFS_I(inode), pos,
1651 &write_bytes) > 0) {
1653 * For nodata cow case, no need to reserve
1656 only_release_metadata = true;
1658 * our prealloc extent may be smaller than
1659 * write_bytes, so scale down.
1661 num_pages = DIV_ROUND_UP(write_bytes + offset,
1663 reserve_bytes = round_up(write_bytes +
1665 fs_info->sectorsize);
1671 WARN_ON(reserve_bytes == 0);
1672 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1675 if (!only_release_metadata)
1676 btrfs_free_reserved_data_space(inode,
1680 btrfs_end_write_no_snapshotting(root);
1684 release_bytes = reserve_bytes;
1687 * This is going to setup the pages array with the number of
1688 * pages we want, so we don't really need to worry about the
1689 * contents of pages from loop to loop
1691 ret = prepare_pages(inode, pages, num_pages,
1693 force_page_uptodate);
1695 btrfs_delalloc_release_extents(BTRFS_I(inode),
1700 extents_locked = lock_and_cleanup_extent_if_need(
1701 BTRFS_I(inode), pages,
1702 num_pages, pos, write_bytes, &lockstart,
1703 &lockend, &cached_state);
1704 if (extents_locked < 0) {
1705 if (extents_locked == -EAGAIN)
1707 btrfs_delalloc_release_extents(BTRFS_I(inode),
1709 ret = extents_locked;
1713 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1715 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1716 dirty_sectors = round_up(copied + sector_offset,
1717 fs_info->sectorsize);
1718 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1721 * if we have trouble faulting in the pages, fall
1722 * back to one page at a time
1724 if (copied < write_bytes)
1728 force_page_uptodate = true;
1732 force_page_uptodate = false;
1733 dirty_pages = DIV_ROUND_UP(copied + offset,
1737 if (num_sectors > dirty_sectors) {
1738 /* release everything except the sectors we dirtied */
1739 release_bytes -= dirty_sectors <<
1740 fs_info->sb->s_blocksize_bits;
1741 if (only_release_metadata) {
1742 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1743 release_bytes, true);
1747 __pos = round_down(pos,
1748 fs_info->sectorsize) +
1749 (dirty_pages << PAGE_SHIFT);
1750 btrfs_delalloc_release_space(inode,
1751 data_reserved, __pos,
1752 release_bytes, true);
1756 release_bytes = round_up(copied + sector_offset,
1757 fs_info->sectorsize);
1760 ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1761 pos, copied, &cached_state);
1764 * If we have not locked the extent range, because the range's
1765 * start offset is >= i_size, we might still have a non-NULL
1766 * cached extent state, acquired while marking the extent range
1767 * as delalloc through btrfs_dirty_pages(). Therefore free any
1768 * possible cached extent state to avoid a memory leak.
1771 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1772 lockstart, lockend, &cached_state);
1774 free_extent_state(cached_state);
1776 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1778 btrfs_drop_pages(pages, num_pages);
1783 if (only_release_metadata)
1784 btrfs_end_write_no_snapshotting(root);
1786 if (only_release_metadata && copied > 0) {
1787 lockstart = round_down(pos,
1788 fs_info->sectorsize);
1789 lockend = round_up(pos + copied,
1790 fs_info->sectorsize) - 1;
1792 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1793 lockend, EXTENT_NORESERVE, NULL,
1797 btrfs_drop_pages(pages, num_pages);
1801 balance_dirty_pages_ratelimited(inode->i_mapping);
1802 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1803 btrfs_btree_balance_dirty(fs_info);
1806 num_written += copied;
1811 if (release_bytes) {
1812 if (only_release_metadata) {
1813 btrfs_end_write_no_snapshotting(root);
1814 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1815 release_bytes, true);
1817 btrfs_delalloc_release_space(inode, data_reserved,
1818 round_down(pos, fs_info->sectorsize),
1819 release_bytes, true);
1823 extent_changeset_free(data_reserved);
1824 return num_written ? num_written : ret;
1827 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1829 struct file *file = iocb->ki_filp;
1830 struct inode *inode = file_inode(file);
1833 ssize_t written_buffered;
1837 written = generic_file_direct_write(iocb, from);
1839 if (written < 0 || !iov_iter_count(from))
1843 written_buffered = btrfs_buffered_write(iocb, from);
1844 if (written_buffered < 0) {
1845 err = written_buffered;
1849 * Ensure all data is persisted. We want the next direct IO read to be
1850 * able to read what was just written.
1852 endbyte = pos + written_buffered - 1;
1853 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1856 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1859 written += written_buffered;
1860 iocb->ki_pos = pos + written_buffered;
1861 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1862 endbyte >> PAGE_SHIFT);
1864 return written ? written : err;
1867 static void update_time_for_write(struct inode *inode)
1869 struct timespec64 now;
1871 if (IS_NOCMTIME(inode))
1874 now = current_time(inode);
1875 if (!timespec64_equal(&inode->i_mtime, &now))
1876 inode->i_mtime = now;
1878 if (!timespec64_equal(&inode->i_ctime, &now))
1879 inode->i_ctime = now;
1881 if (IS_I_VERSION(inode))
1882 inode_inc_iversion(inode);
1885 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1886 struct iov_iter *from)
1888 struct file *file = iocb->ki_filp;
1889 struct inode *inode = file_inode(file);
1890 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1891 struct btrfs_root *root = BTRFS_I(inode)->root;
1894 ssize_t num_written = 0;
1895 const bool sync = iocb->ki_flags & IOCB_DSYNC;
1902 if (!(iocb->ki_flags & IOCB_DIRECT) &&
1903 (iocb->ki_flags & IOCB_NOWAIT))
1906 if (iocb->ki_flags & IOCB_NOWAIT) {
1907 if (!inode_trylock(inode))
1913 err = generic_write_checks(iocb, from);
1915 inode_unlock(inode);
1920 count = iov_iter_count(from);
1921 if (iocb->ki_flags & IOCB_NOWAIT) {
1922 size_t nocow_bytes = count;
1925 * We will allocate space in case nodatacow is not set,
1928 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1929 BTRFS_INODE_PREALLOC)) ||
1930 btrfs_check_can_nocow(BTRFS_I(inode), pos,
1931 &nocow_bytes) <= 0) {
1932 inode_unlock(inode);
1936 /* check_can_nocow() locks the snapshot lock on success */
1937 btrfs_end_write_no_snapshotting(root);
1939 * There are holes in the range or parts of the range that must
1940 * be COWed (shared extents, RO block groups, etc), so just bail
1943 if (nocow_bytes < count) {
1944 inode_unlock(inode);
1949 current->backing_dev_info = inode_to_bdi(inode);
1950 err = file_remove_privs(file);
1952 inode_unlock(inode);
1957 * If BTRFS flips readonly due to some impossible error
1958 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1959 * although we have opened a file as writable, we have
1960 * to stop this write operation to ensure FS consistency.
1962 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1963 inode_unlock(inode);
1969 * We reserve space for updating the inode when we reserve space for the
1970 * extent we are going to write, so we will enospc out there. We don't
1971 * need to start yet another transaction to update the inode as we will
1972 * update the inode when we finish writing whatever data we write.
1974 update_time_for_write(inode);
1976 start_pos = round_down(pos, fs_info->sectorsize);
1977 oldsize = i_size_read(inode);
1978 if (start_pos > oldsize) {
1979 /* Expand hole size to cover write data, preventing empty gap */
1980 end_pos = round_up(pos + count,
1981 fs_info->sectorsize);
1982 err = btrfs_cont_expand(inode, oldsize, end_pos);
1984 inode_unlock(inode);
1987 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1992 atomic_inc(&BTRFS_I(inode)->sync_writers);
1994 if (iocb->ki_flags & IOCB_DIRECT) {
1995 num_written = __btrfs_direct_write(iocb, from);
1997 num_written = btrfs_buffered_write(iocb, from);
1998 if (num_written > 0)
1999 iocb->ki_pos = pos + num_written;
2001 pagecache_isize_extended(inode, oldsize,
2002 i_size_read(inode));
2005 inode_unlock(inode);
2007 btrfs_set_inode_last_sub_trans(BTRFS_I(inode));
2009 if (num_written > 0)
2010 num_written = generic_write_sync(iocb, num_written);
2013 atomic_dec(&BTRFS_I(inode)->sync_writers);
2015 current->backing_dev_info = NULL;
2016 return num_written ? num_written : err;
2019 int btrfs_release_file(struct inode *inode, struct file *filp)
2021 struct btrfs_file_private *private = filp->private_data;
2023 if (private && private->filldir_buf)
2024 kfree(private->filldir_buf);
2026 filp->private_data = NULL;
2029 * ordered_data_close is set by setattr when we are about to truncate
2030 * a file from a non-zero size to a zero size. This tries to
2031 * flush down new bytes that may have been written if the
2032 * application were using truncate to replace a file in place.
2034 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2035 &BTRFS_I(inode)->runtime_flags))
2036 filemap_flush(inode->i_mapping);
2040 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2043 struct blk_plug plug;
2046 * This is only called in fsync, which would do synchronous writes, so
2047 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2048 * multiple disks using raid profile, a large IO can be split to
2049 * several segments of stripe length (currently 64K).
2051 blk_start_plug(&plug);
2052 atomic_inc(&BTRFS_I(inode)->sync_writers);
2053 ret = btrfs_fdatawrite_range(inode, start, end);
2054 atomic_dec(&BTRFS_I(inode)->sync_writers);
2055 blk_finish_plug(&plug);
2061 * fsync call for both files and directories. This logs the inode into
2062 * the tree log instead of forcing full commits whenever possible.
2064 * It needs to call filemap_fdatawait so that all ordered extent updates are
2065 * in the metadata btree are up to date for copying to the log.
2067 * It drops the inode mutex before doing the tree log commit. This is an
2068 * important optimization for directories because holding the mutex prevents
2069 * new operations on the dir while we write to disk.
2071 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2073 struct dentry *dentry = file_dentry(file);
2074 struct inode *inode = d_inode(dentry);
2075 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2076 struct btrfs_root *root = BTRFS_I(inode)->root;
2077 struct btrfs_trans_handle *trans;
2078 struct btrfs_log_ctx ctx;
2081 trace_btrfs_sync_file(file, datasync);
2083 btrfs_init_log_ctx(&ctx, inode);
2086 * Set the range to full if the NO_HOLES feature is not enabled.
2087 * This is to avoid missing file extent items representing holes after
2088 * replaying the log.
2090 if (!btrfs_fs_incompat(fs_info, NO_HOLES)) {
2096 * We write the dirty pages in the range and wait until they complete
2097 * out of the ->i_mutex. If so, we can flush the dirty pages by
2098 * multi-task, and make the performance up. See
2099 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2101 ret = start_ordered_ops(inode, start, end);
2108 * We take the dio_sem here because the tree log stuff can race with
2109 * lockless dio writes and get an extent map logged for an extent we
2110 * never waited on. We need it this high up for lockdep reasons.
2112 down_write(&BTRFS_I(inode)->dio_sem);
2114 atomic_inc(&root->log_batch);
2117 * If the inode needs a full sync, make sure we use a full range to
2118 * avoid log tree corruption, due to hole detection racing with ordered
2119 * extent completion for adjacent ranges, and assertion failures during
2120 * hole detection. Do this while holding the inode lock, to avoid races
2123 if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2124 &BTRFS_I(inode)->runtime_flags)) {
2130 * Before we acquired the inode's lock, someone may have dirtied more
2131 * pages in the target range. We need to make sure that writeback for
2132 * any such pages does not start while we are logging the inode, because
2133 * if it does, any of the following might happen when we are not doing a
2136 * 1) We log an extent after its writeback finishes but before its
2137 * checksums are added to the csum tree, leading to -EIO errors
2138 * when attempting to read the extent after a log replay.
2140 * 2) We can end up logging an extent before its writeback finishes.
2141 * Therefore after the log replay we will have a file extent item
2142 * pointing to an unwritten extent (and no data checksums as well).
2144 * So trigger writeback for any eventual new dirty pages and then we
2145 * wait for all ordered extents to complete below.
2147 ret = start_ordered_ops(inode, start, end);
2149 up_write(&BTRFS_I(inode)->dio_sem);
2150 inode_unlock(inode);
2155 * We have to do this here to avoid the priority inversion of waiting on
2156 * IO of a lower priority task while holding a transaction open.
2158 * Also, the range length can be represented by u64, we have to do the
2159 * typecasts to avoid signed overflow if it's [0, LLONG_MAX].
2161 ret = btrfs_wait_ordered_range(inode, start, (u64)end - (u64)start + 1);
2163 up_write(&BTRFS_I(inode)->dio_sem);
2164 inode_unlock(inode);
2167 atomic_inc(&root->log_batch);
2170 if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2171 BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2173 * We've had everything committed since the last time we were
2174 * modified so clear this flag in case it was set for whatever
2175 * reason, it's no longer relevant.
2177 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2178 &BTRFS_I(inode)->runtime_flags);
2180 * An ordered extent might have started before and completed
2181 * already with io errors, in which case the inode was not
2182 * updated and we end up here. So check the inode's mapping
2183 * for any errors that might have happened since we last
2184 * checked called fsync.
2186 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2187 up_write(&BTRFS_I(inode)->dio_sem);
2188 inode_unlock(inode);
2193 * We use start here because we will need to wait on the IO to complete
2194 * in btrfs_sync_log, which could require joining a transaction (for
2195 * example checking cross references in the nocow path). If we use join
2196 * here we could get into a situation where we're waiting on IO to
2197 * happen that is blocked on a transaction trying to commit. With start
2198 * we inc the extwriter counter, so we wait for all extwriters to exit
2199 * before we start blocking joiners. This comment is to keep somebody
2200 * from thinking they are super smart and changing this to
2201 * btrfs_join_transaction *cough*Josef*cough*.
2203 trans = btrfs_start_transaction(root, 0);
2204 if (IS_ERR(trans)) {
2205 ret = PTR_ERR(trans);
2206 up_write(&BTRFS_I(inode)->dio_sem);
2207 inode_unlock(inode);
2211 ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2213 /* Fallthrough and commit/free transaction. */
2217 /* we've logged all the items and now have a consistent
2218 * version of the file in the log. It is possible that
2219 * someone will come in and modify the file, but that's
2220 * fine because the log is consistent on disk, and we
2221 * have references to all of the file's extents
2223 * It is possible that someone will come in and log the
2224 * file again, but that will end up using the synchronization
2225 * inside btrfs_sync_log to keep things safe.
2227 up_write(&BTRFS_I(inode)->dio_sem);
2228 inode_unlock(inode);
2230 if (ret != BTRFS_NO_LOG_SYNC) {
2232 ret = btrfs_sync_log(trans, root, &ctx);
2234 ret = btrfs_end_transaction(trans);
2238 ret = btrfs_commit_transaction(trans);
2240 ret = btrfs_end_transaction(trans);
2243 ASSERT(list_empty(&ctx.list));
2244 err = file_check_and_advance_wb_err(file);
2247 return ret > 0 ? -EIO : ret;
2250 static const struct vm_operations_struct btrfs_file_vm_ops = {
2251 .fault = filemap_fault,
2252 .map_pages = filemap_map_pages,
2253 .page_mkwrite = btrfs_page_mkwrite,
2256 static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2258 struct address_space *mapping = filp->f_mapping;
2260 if (!mapping->a_ops->readpage)
2263 file_accessed(filp);
2264 vma->vm_ops = &btrfs_file_vm_ops;
2269 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2270 int slot, u64 start, u64 end)
2272 struct btrfs_file_extent_item *fi;
2273 struct btrfs_key key;
2275 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2278 btrfs_item_key_to_cpu(leaf, &key, slot);
2279 if (key.objectid != btrfs_ino(inode) ||
2280 key.type != BTRFS_EXTENT_DATA_KEY)
2283 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2285 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2288 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2291 if (key.offset == end)
2293 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2298 static int fill_holes(struct btrfs_trans_handle *trans,
2299 struct btrfs_inode *inode,
2300 struct btrfs_path *path, u64 offset, u64 end)
2302 struct btrfs_fs_info *fs_info = trans->fs_info;
2303 struct btrfs_root *root = inode->root;
2304 struct extent_buffer *leaf;
2305 struct btrfs_file_extent_item *fi;
2306 struct extent_map *hole_em;
2307 struct extent_map_tree *em_tree = &inode->extent_tree;
2308 struct btrfs_key key;
2311 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2314 key.objectid = btrfs_ino(inode);
2315 key.type = BTRFS_EXTENT_DATA_KEY;
2316 key.offset = offset;
2318 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2321 * We should have dropped this offset, so if we find it then
2322 * something has gone horribly wrong.
2329 leaf = path->nodes[0];
2330 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2334 fi = btrfs_item_ptr(leaf, path->slots[0],
2335 struct btrfs_file_extent_item);
2336 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2338 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2339 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2340 btrfs_set_file_extent_offset(leaf, fi, 0);
2341 btrfs_mark_buffer_dirty(leaf);
2345 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2348 key.offset = offset;
2349 btrfs_set_item_key_safe(fs_info, path, &key);
2350 fi = btrfs_item_ptr(leaf, path->slots[0],
2351 struct btrfs_file_extent_item);
2352 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2354 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2355 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2356 btrfs_set_file_extent_offset(leaf, fi, 0);
2357 btrfs_mark_buffer_dirty(leaf);
2360 btrfs_release_path(path);
2362 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2363 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2368 btrfs_release_path(path);
2370 hole_em = alloc_extent_map();
2372 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2373 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2375 hole_em->start = offset;
2376 hole_em->len = end - offset;
2377 hole_em->ram_bytes = hole_em->len;
2378 hole_em->orig_start = offset;
2380 hole_em->block_start = EXTENT_MAP_HOLE;
2381 hole_em->block_len = 0;
2382 hole_em->orig_block_len = 0;
2383 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2384 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2385 hole_em->generation = trans->transid;
2388 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2389 write_lock(&em_tree->lock);
2390 ret = add_extent_mapping(em_tree, hole_em, 1);
2391 write_unlock(&em_tree->lock);
2392 } while (ret == -EEXIST);
2393 free_extent_map(hole_em);
2395 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2396 &inode->runtime_flags);
2403 * Find a hole extent on given inode and change start/len to the end of hole
2404 * extent.(hole/vacuum extent whose em->start <= start &&
2405 * em->start + em->len > start)
2406 * When a hole extent is found, return 1 and modify start/len.
2408 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2410 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2411 struct extent_map *em;
2414 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2415 round_down(*start, fs_info->sectorsize),
2416 round_up(*len, fs_info->sectorsize), 0);
2420 /* Hole or vacuum extent(only exists in no-hole mode) */
2421 if (em->block_start == EXTENT_MAP_HOLE) {
2423 *len = em->start + em->len > *start + *len ?
2424 0 : *start + *len - em->start - em->len;
2425 *start = em->start + em->len;
2427 free_extent_map(em);
2431 static int btrfs_punch_hole_lock_range(struct inode *inode,
2432 const u64 lockstart,
2434 struct extent_state **cached_state)
2437 struct btrfs_ordered_extent *ordered;
2440 truncate_pagecache_range(inode, lockstart, lockend);
2442 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2444 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2447 * We need to make sure we have no ordered extents in this range
2448 * and nobody raced in and read a page in this range, if we did
2449 * we need to try again.
2452 (ordered->file_offset + ordered->len <= lockstart ||
2453 ordered->file_offset > lockend)) &&
2454 !filemap_range_has_page(inode->i_mapping,
2455 lockstart, lockend)) {
2457 btrfs_put_ordered_extent(ordered);
2461 btrfs_put_ordered_extent(ordered);
2462 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2463 lockend, cached_state);
2464 ret = btrfs_wait_ordered_range(inode, lockstart,
2465 lockend - lockstart + 1);
2472 static int btrfs_insert_clone_extent(struct btrfs_trans_handle *trans,
2473 struct inode *inode,
2474 struct btrfs_path *path,
2475 struct btrfs_clone_extent_info *clone_info,
2476 const u64 clone_len)
2478 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2479 struct btrfs_root *root = BTRFS_I(inode)->root;
2480 struct btrfs_file_extent_item *extent;
2481 struct extent_buffer *leaf;
2482 struct btrfs_key key;
2484 struct btrfs_ref ref = { 0 };
2491 if (clone_info->disk_offset == 0 &&
2492 btrfs_fs_incompat(fs_info, NO_HOLES))
2495 key.objectid = btrfs_ino(BTRFS_I(inode));
2496 key.type = BTRFS_EXTENT_DATA_KEY;
2497 key.offset = clone_info->file_offset;
2498 ret = btrfs_insert_empty_item(trans, root, path, &key,
2499 clone_info->item_size);
2502 leaf = path->nodes[0];
2503 slot = path->slots[0];
2504 write_extent_buffer(leaf, clone_info->extent_buf,
2505 btrfs_item_ptr_offset(leaf, slot),
2506 clone_info->item_size);
2507 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2508 btrfs_set_file_extent_offset(leaf, extent, clone_info->data_offset);
2509 btrfs_set_file_extent_num_bytes(leaf, extent, clone_len);
2510 btrfs_mark_buffer_dirty(leaf);
2511 btrfs_release_path(path);
2513 /* If it's a hole, nothing more needs to be done. */
2514 if (clone_info->disk_offset == 0)
2517 inode_add_bytes(inode, clone_len);
2518 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2519 clone_info->disk_offset,
2520 clone_info->disk_len, 0);
2521 ref_offset = clone_info->file_offset - clone_info->data_offset;
2522 btrfs_init_data_ref(&ref, root->root_key.objectid,
2523 btrfs_ino(BTRFS_I(inode)), ref_offset);
2524 ret = btrfs_inc_extent_ref(trans, &ref);
2530 * The respective range must have been previously locked, as well as the inode.
2531 * The end offset is inclusive (last byte of the range).
2532 * @clone_info is NULL for fallocate's hole punching and non-NULL for extent
2534 * When cloning, we don't want to end up in a state where we dropped extents
2535 * without inserting a new one, so we must abort the transaction to avoid a
2538 int btrfs_punch_hole_range(struct inode *inode, struct btrfs_path *path,
2539 const u64 start, const u64 end,
2540 struct btrfs_clone_extent_info *clone_info,
2541 struct btrfs_trans_handle **trans_out)
2543 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2544 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2545 u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2546 struct btrfs_root *root = BTRFS_I(inode)->root;
2547 struct btrfs_trans_handle *trans = NULL;
2548 struct btrfs_block_rsv *rsv;
2549 unsigned int rsv_count;
2552 u64 len = end - start;
2558 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2563 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2567 * 1 - update the inode
2568 * 1 - removing the extents in the range
2569 * 1 - adding the hole extent if no_holes isn't set or if we are cloning
2572 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || clone_info)
2577 trans = btrfs_start_transaction(root, rsv_count);
2578 if (IS_ERR(trans)) {
2579 ret = PTR_ERR(trans);
2584 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2587 trans->block_rsv = rsv;
2590 while (cur_offset < end) {
2591 ret = __btrfs_drop_extents(trans, root, inode, path,
2592 cur_offset, end + 1, &drop_end,
2594 if (ret != -ENOSPC) {
2596 * When cloning we want to avoid transaction aborts when
2597 * nothing was done and we are attempting to clone parts
2598 * of inline extents, in such cases -EOPNOTSUPP is
2599 * returned by __btrfs_drop_extents() without having
2600 * changed anything in the file.
2602 if (clone_info && ret && ret != -EOPNOTSUPP)
2603 btrfs_abort_transaction(trans, ret);
2607 trans->block_rsv = &fs_info->trans_block_rsv;
2609 if (!clone_info && cur_offset < drop_end &&
2610 cur_offset < ino_size) {
2611 ret = fill_holes(trans, BTRFS_I(inode), path,
2612 cur_offset, drop_end);
2615 * If we failed then we didn't insert our hole
2616 * entries for the area we dropped, so now the
2617 * fs is corrupted, so we must abort the
2620 btrfs_abort_transaction(trans, ret);
2625 if (clone_info && drop_end > clone_info->file_offset) {
2626 u64 clone_len = drop_end - clone_info->file_offset;
2628 ret = btrfs_insert_clone_extent(trans, inode, path,
2629 clone_info, clone_len);
2631 btrfs_abort_transaction(trans, ret);
2634 clone_info->data_len -= clone_len;
2635 clone_info->data_offset += clone_len;
2636 clone_info->file_offset += clone_len;
2639 cur_offset = drop_end;
2641 ret = btrfs_update_inode(trans, root, inode);
2645 btrfs_end_transaction(trans);
2646 btrfs_btree_balance_dirty(fs_info);
2648 trans = btrfs_start_transaction(root, rsv_count);
2649 if (IS_ERR(trans)) {
2650 ret = PTR_ERR(trans);
2655 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2656 rsv, min_size, false);
2657 BUG_ON(ret); /* shouldn't happen */
2658 trans->block_rsv = rsv;
2661 ret = find_first_non_hole(inode, &cur_offset, &len);
2662 if (unlikely(ret < 0))
2672 * If we were cloning, force the next fsync to be a full one since we
2673 * we replaced (or just dropped in the case of cloning holes when
2674 * NO_HOLES is enabled) extents and extent maps.
2675 * This is for the sake of simplicity, and cloning into files larger
2676 * than 16Mb would force the full fsync any way (when
2677 * try_release_extent_mapping() is invoked during page cache truncation.
2680 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2681 &BTRFS_I(inode)->runtime_flags);
2686 trans->block_rsv = &fs_info->trans_block_rsv;
2688 * If we are using the NO_HOLES feature we might have had already an
2689 * hole that overlaps a part of the region [lockstart, lockend] and
2690 * ends at (or beyond) lockend. Since we have no file extent items to
2691 * represent holes, drop_end can be less than lockend and so we must
2692 * make sure we have an extent map representing the existing hole (the
2693 * call to __btrfs_drop_extents() might have dropped the existing extent
2694 * map representing the existing hole), otherwise the fast fsync path
2695 * will not record the existence of the hole region
2696 * [existing_hole_start, lockend].
2698 if (drop_end <= end)
2701 * Don't insert file hole extent item if it's for a range beyond eof
2702 * (because it's useless) or if it represents a 0 bytes range (when
2703 * cur_offset == drop_end).
2705 if (!clone_info && cur_offset < ino_size && cur_offset < drop_end) {
2706 ret = fill_holes(trans, BTRFS_I(inode), path,
2707 cur_offset, drop_end);
2709 /* Same comment as above. */
2710 btrfs_abort_transaction(trans, ret);
2715 ret = btrfs_insert_clone_extent(trans, inode, path, clone_info,
2716 clone_info->data_len);
2718 btrfs_abort_transaction(trans, ret);
2727 trans->block_rsv = &fs_info->trans_block_rsv;
2729 btrfs_end_transaction(trans);
2733 btrfs_free_block_rsv(fs_info, rsv);
2738 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2740 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2741 struct btrfs_root *root = BTRFS_I(inode)->root;
2742 struct extent_state *cached_state = NULL;
2743 struct btrfs_path *path;
2744 struct btrfs_trans_handle *trans = NULL;
2749 u64 orig_start = offset;
2753 bool truncated_block = false;
2754 bool updated_inode = false;
2756 ret = btrfs_wait_ordered_range(inode, offset, len);
2761 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2762 ret = find_first_non_hole(inode, &offset, &len);
2764 goto out_only_mutex;
2766 /* Already in a large hole */
2768 goto out_only_mutex;
2771 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2772 lockend = round_down(offset + len,
2773 btrfs_inode_sectorsize(inode)) - 1;
2774 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2775 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2777 * We needn't truncate any block which is beyond the end of the file
2778 * because we are sure there is no data there.
2781 * Only do this if we are in the same block and we aren't doing the
2784 if (same_block && len < fs_info->sectorsize) {
2785 if (offset < ino_size) {
2786 truncated_block = true;
2787 ret = btrfs_truncate_block(inode, offset, len, 0);
2791 goto out_only_mutex;
2794 /* zero back part of the first block */
2795 if (offset < ino_size) {
2796 truncated_block = true;
2797 ret = btrfs_truncate_block(inode, offset, 0, 0);
2799 inode_unlock(inode);
2804 /* Check the aligned pages after the first unaligned page,
2805 * if offset != orig_start, which means the first unaligned page
2806 * including several following pages are already in holes,
2807 * the extra check can be skipped */
2808 if (offset == orig_start) {
2809 /* after truncate page, check hole again */
2810 len = offset + len - lockstart;
2812 ret = find_first_non_hole(inode, &offset, &len);
2814 goto out_only_mutex;
2817 goto out_only_mutex;
2822 /* Check the tail unaligned part is in a hole */
2823 tail_start = lockend + 1;
2824 tail_len = offset + len - tail_start;
2826 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2827 if (unlikely(ret < 0))
2828 goto out_only_mutex;
2830 /* zero the front end of the last page */
2831 if (tail_start + tail_len < ino_size) {
2832 truncated_block = true;
2833 ret = btrfs_truncate_block(inode,
2834 tail_start + tail_len,
2837 goto out_only_mutex;
2842 if (lockend < lockstart) {
2844 goto out_only_mutex;
2847 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2850 goto out_only_mutex;
2852 path = btrfs_alloc_path();
2858 ret = btrfs_punch_hole_range(inode, path, lockstart, lockend, NULL,
2860 btrfs_free_path(path);
2864 ASSERT(trans != NULL);
2865 inode_inc_iversion(inode);
2866 inode->i_mtime = inode->i_ctime = current_time(inode);
2867 ret = btrfs_update_inode(trans, root, inode);
2868 updated_inode = true;
2869 btrfs_end_transaction(trans);
2870 btrfs_btree_balance_dirty(fs_info);
2872 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2875 if (!updated_inode && truncated_block && !ret) {
2877 * If we only end up zeroing part of a page, we still need to
2878 * update the inode item, so that all the time fields are
2879 * updated as well as the necessary btrfs inode in memory fields
2880 * for detecting, at fsync time, if the inode isn't yet in the
2881 * log tree or it's there but not up to date.
2883 struct timespec64 now = current_time(inode);
2885 inode_inc_iversion(inode);
2886 inode->i_mtime = now;
2887 inode->i_ctime = now;
2888 trans = btrfs_start_transaction(root, 1);
2889 if (IS_ERR(trans)) {
2890 ret = PTR_ERR(trans);
2894 ret = btrfs_update_inode(trans, root, inode);
2895 ret2 = btrfs_end_transaction(trans);
2900 inode_unlock(inode);
2904 /* Helper structure to record which range is already reserved */
2905 struct falloc_range {
2906 struct list_head list;
2912 * Helper function to add falloc range
2914 * Caller should have locked the larger range of extent containing
2917 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2919 struct falloc_range *prev = NULL;
2920 struct falloc_range *range = NULL;
2922 if (list_empty(head))
2926 * As fallocate iterate by bytenr order, we only need to check
2929 prev = list_entry(head->prev, struct falloc_range, list);
2930 if (prev->start + prev->len == start) {
2935 range = kmalloc(sizeof(*range), GFP_KERNEL);
2938 range->start = start;
2940 list_add_tail(&range->list, head);
2944 static int btrfs_fallocate_update_isize(struct inode *inode,
2948 struct btrfs_trans_handle *trans;
2949 struct btrfs_root *root = BTRFS_I(inode)->root;
2953 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2956 trans = btrfs_start_transaction(root, 1);
2958 return PTR_ERR(trans);
2960 inode->i_ctime = current_time(inode);
2961 i_size_write(inode, end);
2962 btrfs_ordered_update_i_size(inode, end, NULL);
2963 ret = btrfs_update_inode(trans, root, inode);
2964 ret2 = btrfs_end_transaction(trans);
2966 return ret ? ret : ret2;
2970 RANGE_BOUNDARY_WRITTEN_EXTENT,
2971 RANGE_BOUNDARY_PREALLOC_EXTENT,
2972 RANGE_BOUNDARY_HOLE,
2975 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2978 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2979 struct extent_map *em;
2982 offset = round_down(offset, sectorsize);
2983 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2987 if (em->block_start == EXTENT_MAP_HOLE)
2988 ret = RANGE_BOUNDARY_HOLE;
2989 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2990 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2992 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2994 free_extent_map(em);
2998 static int btrfs_zero_range(struct inode *inode,
3003 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3004 struct extent_map *em;
3005 struct extent_changeset *data_reserved = NULL;
3008 const u64 sectorsize = btrfs_inode_sectorsize(inode);
3009 u64 alloc_start = round_down(offset, sectorsize);
3010 u64 alloc_end = round_up(offset + len, sectorsize);
3011 u64 bytes_to_reserve = 0;
3012 bool space_reserved = false;
3014 inode_dio_wait(inode);
3016 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
3017 alloc_start, alloc_end - alloc_start, 0);
3024 * Avoid hole punching and extent allocation for some cases. More cases
3025 * could be considered, but these are unlikely common and we keep things
3026 * as simple as possible for now. Also, intentionally, if the target
3027 * range contains one or more prealloc extents together with regular
3028 * extents and holes, we drop all the existing extents and allocate a
3029 * new prealloc extent, so that we get a larger contiguous disk extent.
3031 if (em->start <= alloc_start &&
3032 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3033 const u64 em_end = em->start + em->len;
3035 if (em_end >= offset + len) {
3037 * The whole range is already a prealloc extent,
3038 * do nothing except updating the inode's i_size if
3041 free_extent_map(em);
3042 ret = btrfs_fallocate_update_isize(inode, offset + len,
3047 * Part of the range is already a prealloc extent, so operate
3048 * only on the remaining part of the range.
3050 alloc_start = em_end;
3051 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3052 len = offset + len - alloc_start;
3053 offset = alloc_start;
3054 alloc_hint = em->block_start + em->len;
3056 free_extent_map(em);
3058 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3059 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3060 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
3061 alloc_start, sectorsize, 0);
3067 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3068 free_extent_map(em);
3069 ret = btrfs_fallocate_update_isize(inode, offset + len,
3073 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3074 free_extent_map(em);
3075 ret = btrfs_truncate_block(inode, offset, len, 0);
3077 ret = btrfs_fallocate_update_isize(inode,
3082 free_extent_map(em);
3083 alloc_start = round_down(offset, sectorsize);
3084 alloc_end = alloc_start + sectorsize;
3088 alloc_start = round_up(offset, sectorsize);
3089 alloc_end = round_down(offset + len, sectorsize);
3092 * For unaligned ranges, check the pages at the boundaries, they might
3093 * map to an extent, in which case we need to partially zero them, or
3094 * they might map to a hole, in which case we need our allocation range
3097 if (!IS_ALIGNED(offset, sectorsize)) {
3098 ret = btrfs_zero_range_check_range_boundary(inode, offset);
3101 if (ret == RANGE_BOUNDARY_HOLE) {
3102 alloc_start = round_down(offset, sectorsize);
3104 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3105 ret = btrfs_truncate_block(inode, offset, 0, 0);
3113 if (!IS_ALIGNED(offset + len, sectorsize)) {
3114 ret = btrfs_zero_range_check_range_boundary(inode,
3118 if (ret == RANGE_BOUNDARY_HOLE) {
3119 alloc_end = round_up(offset + len, sectorsize);
3121 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3122 ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3131 if (alloc_start < alloc_end) {
3132 struct extent_state *cached_state = NULL;
3133 const u64 lockstart = alloc_start;
3134 const u64 lockend = alloc_end - 1;
3136 bytes_to_reserve = alloc_end - alloc_start;
3137 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3141 space_reserved = true;
3142 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3146 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3147 alloc_start, bytes_to_reserve);
3149 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3150 lockend, &cached_state);
3153 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3154 alloc_end - alloc_start,
3156 offset + len, &alloc_hint);
3157 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3158 lockend, &cached_state);
3159 /* btrfs_prealloc_file_range releases reserved space on error */
3161 space_reserved = false;
3165 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3167 if (ret && space_reserved)
3168 btrfs_free_reserved_data_space(inode, data_reserved,
3169 alloc_start, bytes_to_reserve);
3170 extent_changeset_free(data_reserved);
3175 static long btrfs_fallocate(struct file *file, int mode,
3176 loff_t offset, loff_t len)
3178 struct inode *inode = file_inode(file);
3179 struct extent_state *cached_state = NULL;
3180 struct extent_changeset *data_reserved = NULL;
3181 struct falloc_range *range;
3182 struct falloc_range *tmp;
3183 struct list_head reserve_list;
3191 struct extent_map *em;
3192 int blocksize = btrfs_inode_sectorsize(inode);
3195 alloc_start = round_down(offset, blocksize);
3196 alloc_end = round_up(offset + len, blocksize);
3197 cur_offset = alloc_start;
3199 /* Make sure we aren't being give some crap mode */
3200 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3201 FALLOC_FL_ZERO_RANGE))
3204 if (mode & FALLOC_FL_PUNCH_HOLE)
3205 return btrfs_punch_hole(inode, offset, len);
3208 * Only trigger disk allocation, don't trigger qgroup reserve
3210 * For qgroup space, it will be checked later.
3212 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3213 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3214 alloc_end - alloc_start);
3221 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3222 ret = inode_newsize_ok(inode, offset + len);
3228 * TODO: Move these two operations after we have checked
3229 * accurate reserved space, or fallocate can still fail but
3230 * with page truncated or size expanded.
3232 * But that's a minor problem and won't do much harm BTW.
3234 if (alloc_start > inode->i_size) {
3235 ret = btrfs_cont_expand(inode, i_size_read(inode),
3239 } else if (offset + len > inode->i_size) {
3241 * If we are fallocating from the end of the file onward we
3242 * need to zero out the end of the block if i_size lands in the
3243 * middle of a block.
3245 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3251 * wait for ordered IO before we have any locks. We'll loop again
3252 * below with the locks held.
3254 ret = btrfs_wait_ordered_range(inode, alloc_start,
3255 alloc_end - alloc_start);
3259 if (mode & FALLOC_FL_ZERO_RANGE) {
3260 ret = btrfs_zero_range(inode, offset, len, mode);
3261 inode_unlock(inode);
3265 locked_end = alloc_end - 1;
3267 struct btrfs_ordered_extent *ordered;
3269 /* the extent lock is ordered inside the running
3272 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3273 locked_end, &cached_state);
3274 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3277 ordered->file_offset + ordered->len > alloc_start &&
3278 ordered->file_offset < alloc_end) {
3279 btrfs_put_ordered_extent(ordered);
3280 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3281 alloc_start, locked_end,
3284 * we can't wait on the range with the transaction
3285 * running or with the extent lock held
3287 ret = btrfs_wait_ordered_range(inode, alloc_start,
3288 alloc_end - alloc_start);
3293 btrfs_put_ordered_extent(ordered);
3298 /* First, check if we exceed the qgroup limit */
3299 INIT_LIST_HEAD(&reserve_list);
3300 while (cur_offset < alloc_end) {
3301 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3302 alloc_end - cur_offset, 0);
3307 last_byte = min(extent_map_end(em), alloc_end);
3308 actual_end = min_t(u64, extent_map_end(em), offset + len);
3309 last_byte = ALIGN(last_byte, blocksize);
3310 if (em->block_start == EXTENT_MAP_HOLE ||
3311 (cur_offset >= inode->i_size &&
3312 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3313 ret = add_falloc_range(&reserve_list, cur_offset,
3314 last_byte - cur_offset);
3316 free_extent_map(em);
3319 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3320 &data_reserved, cur_offset,
3321 last_byte - cur_offset);
3323 cur_offset = last_byte;
3324 free_extent_map(em);
3329 * Do not need to reserve unwritten extent for this
3330 * range, free reserved data space first, otherwise
3331 * it'll result in false ENOSPC error.
3333 btrfs_free_reserved_data_space(inode, data_reserved,
3334 cur_offset, last_byte - cur_offset);
3336 free_extent_map(em);
3337 cur_offset = last_byte;
3341 * If ret is still 0, means we're OK to fallocate.
3342 * Or just cleanup the list and exit.
3344 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3346 ret = btrfs_prealloc_file_range(inode, mode,
3348 range->len, i_blocksize(inode),
3349 offset + len, &alloc_hint);
3351 btrfs_free_reserved_data_space(inode,
3352 data_reserved, range->start,
3354 list_del(&range->list);
3361 * We didn't need to allocate any more space, but we still extended the
3362 * size of the file so we need to update i_size and the inode item.
3364 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3366 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3369 inode_unlock(inode);
3370 /* Let go of our reservation. */
3371 if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3372 btrfs_free_reserved_data_space(inode, data_reserved,
3373 cur_offset, alloc_end - cur_offset);
3374 extent_changeset_free(data_reserved);
3378 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3380 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3381 struct extent_map *em = NULL;
3382 struct extent_state *cached_state = NULL;
3389 if (inode->i_size == 0)
3393 * *offset can be negative, in this case we start finding DATA/HOLE from
3394 * the very start of the file.
3396 start = max_t(loff_t, 0, *offset);
3398 lockstart = round_down(start, fs_info->sectorsize);
3399 lockend = round_up(i_size_read(inode),
3400 fs_info->sectorsize);
3401 if (lockend <= lockstart)
3402 lockend = lockstart + fs_info->sectorsize;
3404 len = lockend - lockstart + 1;
3406 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3409 while (start < inode->i_size) {
3410 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3417 if (whence == SEEK_HOLE &&
3418 (em->block_start == EXTENT_MAP_HOLE ||
3419 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3421 else if (whence == SEEK_DATA &&
3422 (em->block_start != EXTENT_MAP_HOLE &&
3423 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3426 start = em->start + em->len;
3427 free_extent_map(em);
3431 free_extent_map(em);
3433 if (whence == SEEK_DATA && start >= inode->i_size)
3436 *offset = min_t(loff_t, start, inode->i_size);
3438 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3443 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3445 struct inode *inode = file->f_mapping->host;
3452 offset = generic_file_llseek(file, offset, whence);
3456 if (offset >= i_size_read(inode)) {
3457 inode_unlock(inode);
3461 ret = find_desired_extent(inode, &offset, whence);
3463 inode_unlock(inode);
3468 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3470 inode_unlock(inode);
3474 static int btrfs_file_open(struct inode *inode, struct file *filp)
3476 filp->f_mode |= FMODE_NOWAIT;
3477 return generic_file_open(inode, filp);
3480 const struct file_operations btrfs_file_operations = {
3481 .llseek = btrfs_file_llseek,
3482 .read_iter = generic_file_read_iter,
3483 .splice_read = generic_file_splice_read,
3484 .write_iter = btrfs_file_write_iter,
3485 .mmap = btrfs_file_mmap,
3486 .open = btrfs_file_open,
3487 .release = btrfs_release_file,
3488 .fsync = btrfs_sync_file,
3489 .fallocate = btrfs_fallocate,
3490 .unlocked_ioctl = btrfs_ioctl,
3491 #ifdef CONFIG_COMPAT
3492 .compat_ioctl = btrfs_compat_ioctl,
3494 .remap_file_range = btrfs_remap_file_range,
3497 void __cold btrfs_auto_defrag_exit(void)
3499 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3502 int __init btrfs_auto_defrag_init(void)
3504 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3505 sizeof(struct inode_defrag), 0,
3508 if (!btrfs_inode_defrag_cachep)
3514 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3519 * So with compression we will find and lock a dirty page and clear the
3520 * first one as dirty, setup an async extent, and immediately return
3521 * with the entire range locked but with nobody actually marked with
3522 * writeback. So we can't just filemap_write_and_wait_range() and
3523 * expect it to work since it will just kick off a thread to do the
3524 * actual work. So we need to call filemap_fdatawrite_range _again_
3525 * since it will wait on the page lock, which won't be unlocked until
3526 * after the pages have been marked as writeback and so we're good to go
3527 * from there. We have to do this otherwise we'll miss the ordered
3528 * extents and that results in badness. Please Josef, do not think you
3529 * know better and pull this out at some point in the future, it is
3530 * right and you are wrong.
3532 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3533 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3534 &BTRFS_I(inode)->runtime_flags))
3535 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);