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 static noinline int 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,
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 sector_offset = pos & (fs_info->sectorsize - 1);
1640 reserve_bytes = round_up(write_bytes + sector_offset,
1641 fs_info->sectorsize);
1643 extent_changeset_release(data_reserved);
1644 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1647 if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1648 BTRFS_INODE_PREALLOC)) &&
1649 check_can_nocow(BTRFS_I(inode), pos,
1650 &write_bytes) > 0) {
1652 * For nodata cow case, no need to reserve
1655 only_release_metadata = true;
1657 * our prealloc extent may be smaller than
1658 * write_bytes, so scale down.
1660 num_pages = DIV_ROUND_UP(write_bytes + offset,
1662 reserve_bytes = round_up(write_bytes +
1664 fs_info->sectorsize);
1670 WARN_ON(reserve_bytes == 0);
1671 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1674 if (!only_release_metadata)
1675 btrfs_free_reserved_data_space(inode,
1679 btrfs_end_write_no_snapshotting(root);
1683 release_bytes = reserve_bytes;
1686 * This is going to setup the pages array with the number of
1687 * pages we want, so we don't really need to worry about the
1688 * contents of pages from loop to loop
1690 ret = prepare_pages(inode, pages, num_pages,
1692 force_page_uptodate);
1694 btrfs_delalloc_release_extents(BTRFS_I(inode),
1695 reserve_bytes, true);
1699 extents_locked = lock_and_cleanup_extent_if_need(
1700 BTRFS_I(inode), pages,
1701 num_pages, pos, write_bytes, &lockstart,
1702 &lockend, &cached_state);
1703 if (extents_locked < 0) {
1704 if (extents_locked == -EAGAIN)
1706 btrfs_delalloc_release_extents(BTRFS_I(inode),
1707 reserve_bytes, true);
1708 ret = extents_locked;
1712 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1714 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1715 dirty_sectors = round_up(copied + sector_offset,
1716 fs_info->sectorsize);
1717 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1720 * if we have trouble faulting in the pages, fall
1721 * back to one page at a time
1723 if (copied < write_bytes)
1727 force_page_uptodate = true;
1731 force_page_uptodate = false;
1732 dirty_pages = DIV_ROUND_UP(copied + offset,
1736 if (num_sectors > dirty_sectors) {
1737 /* release everything except the sectors we dirtied */
1738 release_bytes -= dirty_sectors <<
1739 fs_info->sb->s_blocksize_bits;
1740 if (only_release_metadata) {
1741 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1742 release_bytes, true);
1746 __pos = round_down(pos,
1747 fs_info->sectorsize) +
1748 (dirty_pages << PAGE_SHIFT);
1749 btrfs_delalloc_release_space(inode,
1750 data_reserved, __pos,
1751 release_bytes, true);
1755 release_bytes = round_up(copied + sector_offset,
1756 fs_info->sectorsize);
1759 ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1760 pos, copied, &cached_state);
1763 * If we have not locked the extent range, because the range's
1764 * start offset is >= i_size, we might still have a non-NULL
1765 * cached extent state, acquired while marking the extent range
1766 * as delalloc through btrfs_dirty_pages(). Therefore free any
1767 * possible cached extent state to avoid a memory leak.
1770 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1771 lockstart, lockend, &cached_state);
1773 free_extent_state(cached_state);
1775 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,
1795 only_release_metadata = false;
1798 btrfs_drop_pages(pages, num_pages);
1802 balance_dirty_pages_ratelimited(inode->i_mapping);
1803 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1804 btrfs_btree_balance_dirty(fs_info);
1807 num_written += copied;
1812 if (release_bytes) {
1813 if (only_release_metadata) {
1814 btrfs_end_write_no_snapshotting(root);
1815 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1816 release_bytes, true);
1818 btrfs_delalloc_release_space(inode, data_reserved,
1819 round_down(pos, fs_info->sectorsize),
1820 release_bytes, true);
1824 extent_changeset_free(data_reserved);
1825 return num_written ? num_written : ret;
1828 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1830 struct file *file = iocb->ki_filp;
1831 struct inode *inode = file_inode(file);
1834 ssize_t written_buffered;
1838 written = generic_file_direct_write(iocb, from);
1840 if (written < 0 || !iov_iter_count(from))
1844 written_buffered = btrfs_buffered_write(iocb, from);
1845 if (written_buffered < 0) {
1846 err = written_buffered;
1850 * Ensure all data is persisted. We want the next direct IO read to be
1851 * able to read what was just written.
1853 endbyte = pos + written_buffered - 1;
1854 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1857 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1860 written += written_buffered;
1861 iocb->ki_pos = pos + written_buffered;
1862 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1863 endbyte >> PAGE_SHIFT);
1865 return written ? written : err;
1868 static void update_time_for_write(struct inode *inode)
1870 struct timespec64 now;
1872 if (IS_NOCMTIME(inode))
1875 now = current_time(inode);
1876 if (!timespec64_equal(&inode->i_mtime, &now))
1877 inode->i_mtime = now;
1879 if (!timespec64_equal(&inode->i_ctime, &now))
1880 inode->i_ctime = now;
1882 if (IS_I_VERSION(inode))
1883 inode_inc_iversion(inode);
1886 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1887 struct iov_iter *from)
1889 struct file *file = iocb->ki_filp;
1890 struct inode *inode = file_inode(file);
1891 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1892 struct btrfs_root *root = BTRFS_I(inode)->root;
1895 ssize_t num_written = 0;
1896 const bool sync = iocb->ki_flags & IOCB_DSYNC;
1903 if (!(iocb->ki_flags & IOCB_DIRECT) &&
1904 (iocb->ki_flags & IOCB_NOWAIT))
1907 if (!inode_trylock(inode)) {
1908 if (iocb->ki_flags & IOCB_NOWAIT)
1913 err = generic_write_checks(iocb, from);
1915 inode_unlock(inode);
1920 count = iov_iter_count(from);
1921 if (iocb->ki_flags & IOCB_NOWAIT) {
1923 * We will allocate space in case nodatacow is not set,
1926 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1927 BTRFS_INODE_PREALLOC)) ||
1928 check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1929 inode_unlock(inode);
1934 current->backing_dev_info = inode_to_bdi(inode);
1935 err = file_remove_privs(file);
1937 inode_unlock(inode);
1942 * If BTRFS flips readonly due to some impossible error
1943 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1944 * although we have opened a file as writable, we have
1945 * to stop this write operation to ensure FS consistency.
1947 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1948 inode_unlock(inode);
1954 * We reserve space for updating the inode when we reserve space for the
1955 * extent we are going to write, so we will enospc out there. We don't
1956 * need to start yet another transaction to update the inode as we will
1957 * update the inode when we finish writing whatever data we write.
1959 update_time_for_write(inode);
1961 start_pos = round_down(pos, fs_info->sectorsize);
1962 oldsize = i_size_read(inode);
1963 if (start_pos > oldsize) {
1964 /* Expand hole size to cover write data, preventing empty gap */
1965 end_pos = round_up(pos + count,
1966 fs_info->sectorsize);
1967 err = btrfs_cont_expand(inode, oldsize, end_pos);
1969 inode_unlock(inode);
1972 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1977 atomic_inc(&BTRFS_I(inode)->sync_writers);
1979 if (iocb->ki_flags & IOCB_DIRECT) {
1980 num_written = __btrfs_direct_write(iocb, from);
1982 num_written = btrfs_buffered_write(iocb, from);
1983 if (num_written > 0)
1984 iocb->ki_pos = pos + num_written;
1986 pagecache_isize_extended(inode, oldsize,
1987 i_size_read(inode));
1990 inode_unlock(inode);
1993 * We also have to set last_sub_trans to the current log transid,
1994 * otherwise subsequent syncs to a file that's been synced in this
1995 * transaction will appear to have already occurred.
1997 spin_lock(&BTRFS_I(inode)->lock);
1998 BTRFS_I(inode)->last_sub_trans = root->log_transid;
1999 spin_unlock(&BTRFS_I(inode)->lock);
2000 if (num_written > 0)
2001 num_written = generic_write_sync(iocb, num_written);
2004 atomic_dec(&BTRFS_I(inode)->sync_writers);
2006 current->backing_dev_info = NULL;
2007 return num_written ? num_written : err;
2010 int btrfs_release_file(struct inode *inode, struct file *filp)
2012 struct btrfs_file_private *private = filp->private_data;
2014 if (private && private->filldir_buf)
2015 kfree(private->filldir_buf);
2017 filp->private_data = NULL;
2020 * ordered_data_close is set by setattr when we are about to truncate
2021 * a file from a non-zero size to a zero size. This tries to
2022 * flush down new bytes that may have been written if the
2023 * application were using truncate to replace a file in place.
2025 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2026 &BTRFS_I(inode)->runtime_flags))
2027 filemap_flush(inode->i_mapping);
2031 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2034 struct blk_plug plug;
2037 * This is only called in fsync, which would do synchronous writes, so
2038 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2039 * multiple disks using raid profile, a large IO can be split to
2040 * several segments of stripe length (currently 64K).
2042 blk_start_plug(&plug);
2043 atomic_inc(&BTRFS_I(inode)->sync_writers);
2044 ret = btrfs_fdatawrite_range(inode, start, end);
2045 atomic_dec(&BTRFS_I(inode)->sync_writers);
2046 blk_finish_plug(&plug);
2052 * fsync call for both files and directories. This logs the inode into
2053 * the tree log instead of forcing full commits whenever possible.
2055 * It needs to call filemap_fdatawait so that all ordered extent updates are
2056 * in the metadata btree are up to date for copying to the log.
2058 * It drops the inode mutex before doing the tree log commit. This is an
2059 * important optimization for directories because holding the mutex prevents
2060 * new operations on the dir while we write to disk.
2062 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2064 struct dentry *dentry = file_dentry(file);
2065 struct inode *inode = d_inode(dentry);
2066 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2067 struct btrfs_root *root = BTRFS_I(inode)->root;
2068 struct btrfs_trans_handle *trans;
2069 struct btrfs_log_ctx ctx;
2074 * If the inode needs a full sync, make sure we use a full range to
2075 * avoid log tree corruption, due to hole detection racing with ordered
2076 * extent completion for adjacent ranges, and assertion failures during
2079 if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2080 &BTRFS_I(inode)->runtime_flags)) {
2086 * The range length can be represented by u64, we have to do the typecasts
2087 * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
2089 len = (u64)end - (u64)start + 1;
2090 trace_btrfs_sync_file(file, datasync);
2092 btrfs_init_log_ctx(&ctx, inode);
2095 * We write the dirty pages in the range and wait until they complete
2096 * out of the ->i_mutex. If so, we can flush the dirty pages by
2097 * multi-task, and make the performance up. See
2098 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2100 ret = start_ordered_ops(inode, start, end);
2107 * We take the dio_sem here because the tree log stuff can race with
2108 * lockless dio writes and get an extent map logged for an extent we
2109 * never waited on. We need it this high up for lockdep reasons.
2111 down_write(&BTRFS_I(inode)->dio_sem);
2113 atomic_inc(&root->log_batch);
2116 * Before we acquired the inode's lock, someone may have dirtied more
2117 * pages in the target range. We need to make sure that writeback for
2118 * any such pages does not start while we are logging the inode, because
2119 * if it does, any of the following might happen when we are not doing a
2122 * 1) We log an extent after its writeback finishes but before its
2123 * checksums are added to the csum tree, leading to -EIO errors
2124 * when attempting to read the extent after a log replay.
2126 * 2) We can end up logging an extent before its writeback finishes.
2127 * Therefore after the log replay we will have a file extent item
2128 * pointing to an unwritten extent (and no data checksums as well).
2130 * So trigger writeback for any eventual new dirty pages and then we
2131 * wait for all ordered extents to complete below.
2133 ret = start_ordered_ops(inode, start, end);
2135 inode_unlock(inode);
2140 * We have to do this here to avoid the priority inversion of waiting on
2141 * IO of a lower priority task while holding a transaction open.
2143 ret = btrfs_wait_ordered_range(inode, start, len);
2145 up_write(&BTRFS_I(inode)->dio_sem);
2146 inode_unlock(inode);
2149 atomic_inc(&root->log_batch);
2152 if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2153 BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2155 * We've had everything committed since the last time we were
2156 * modified so clear this flag in case it was set for whatever
2157 * reason, it's no longer relevant.
2159 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2160 &BTRFS_I(inode)->runtime_flags);
2162 * An ordered extent might have started before and completed
2163 * already with io errors, in which case the inode was not
2164 * updated and we end up here. So check the inode's mapping
2165 * for any errors that might have happened since we last
2166 * checked called fsync.
2168 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2169 up_write(&BTRFS_I(inode)->dio_sem);
2170 inode_unlock(inode);
2175 * We use start here because we will need to wait on the IO to complete
2176 * in btrfs_sync_log, which could require joining a transaction (for
2177 * example checking cross references in the nocow path). If we use join
2178 * here we could get into a situation where we're waiting on IO to
2179 * happen that is blocked on a transaction trying to commit. With start
2180 * we inc the extwriter counter, so we wait for all extwriters to exit
2181 * before we start blocking joiners. This comment is to keep somebody
2182 * from thinking they are super smart and changing this to
2183 * btrfs_join_transaction *cough*Josef*cough*.
2185 trans = btrfs_start_transaction(root, 0);
2186 if (IS_ERR(trans)) {
2187 ret = PTR_ERR(trans);
2188 up_write(&BTRFS_I(inode)->dio_sem);
2189 inode_unlock(inode);
2193 ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2195 /* Fallthrough and commit/free transaction. */
2199 /* we've logged all the items and now have a consistent
2200 * version of the file in the log. It is possible that
2201 * someone will come in and modify the file, but that's
2202 * fine because the log is consistent on disk, and we
2203 * have references to all of the file's extents
2205 * It is possible that someone will come in and log the
2206 * file again, but that will end up using the synchronization
2207 * inside btrfs_sync_log to keep things safe.
2209 up_write(&BTRFS_I(inode)->dio_sem);
2210 inode_unlock(inode);
2212 if (ret != BTRFS_NO_LOG_SYNC) {
2214 ret = btrfs_sync_log(trans, root, &ctx);
2216 ret = btrfs_end_transaction(trans);
2220 ret = btrfs_commit_transaction(trans);
2222 ret = btrfs_end_transaction(trans);
2225 ASSERT(list_empty(&ctx.list));
2226 err = file_check_and_advance_wb_err(file);
2229 return ret > 0 ? -EIO : ret;
2232 static const struct vm_operations_struct btrfs_file_vm_ops = {
2233 .fault = filemap_fault,
2234 .map_pages = filemap_map_pages,
2235 .page_mkwrite = btrfs_page_mkwrite,
2238 static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2240 struct address_space *mapping = filp->f_mapping;
2242 if (!mapping->a_ops->readpage)
2245 file_accessed(filp);
2246 vma->vm_ops = &btrfs_file_vm_ops;
2251 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2252 int slot, u64 start, u64 end)
2254 struct btrfs_file_extent_item *fi;
2255 struct btrfs_key key;
2257 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2260 btrfs_item_key_to_cpu(leaf, &key, slot);
2261 if (key.objectid != btrfs_ino(inode) ||
2262 key.type != BTRFS_EXTENT_DATA_KEY)
2265 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2267 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2270 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2273 if (key.offset == end)
2275 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2280 static int fill_holes(struct btrfs_trans_handle *trans,
2281 struct btrfs_inode *inode,
2282 struct btrfs_path *path, u64 offset, u64 end)
2284 struct btrfs_fs_info *fs_info = trans->fs_info;
2285 struct btrfs_root *root = inode->root;
2286 struct extent_buffer *leaf;
2287 struct btrfs_file_extent_item *fi;
2288 struct extent_map *hole_em;
2289 struct extent_map_tree *em_tree = &inode->extent_tree;
2290 struct btrfs_key key;
2293 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2296 key.objectid = btrfs_ino(inode);
2297 key.type = BTRFS_EXTENT_DATA_KEY;
2298 key.offset = offset;
2300 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2303 * We should have dropped this offset, so if we find it then
2304 * something has gone horribly wrong.
2311 leaf = path->nodes[0];
2312 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2316 fi = btrfs_item_ptr(leaf, path->slots[0],
2317 struct btrfs_file_extent_item);
2318 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2320 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2321 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2322 btrfs_set_file_extent_offset(leaf, fi, 0);
2323 btrfs_mark_buffer_dirty(leaf);
2327 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2330 key.offset = offset;
2331 btrfs_set_item_key_safe(fs_info, path, &key);
2332 fi = btrfs_item_ptr(leaf, path->slots[0],
2333 struct btrfs_file_extent_item);
2334 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2336 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2337 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2338 btrfs_set_file_extent_offset(leaf, fi, 0);
2339 btrfs_mark_buffer_dirty(leaf);
2342 btrfs_release_path(path);
2344 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2345 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2350 btrfs_release_path(path);
2352 hole_em = alloc_extent_map();
2354 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2355 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2357 hole_em->start = offset;
2358 hole_em->len = end - offset;
2359 hole_em->ram_bytes = hole_em->len;
2360 hole_em->orig_start = offset;
2362 hole_em->block_start = EXTENT_MAP_HOLE;
2363 hole_em->block_len = 0;
2364 hole_em->orig_block_len = 0;
2365 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2366 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2367 hole_em->generation = trans->transid;
2370 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2371 write_lock(&em_tree->lock);
2372 ret = add_extent_mapping(em_tree, hole_em, 1);
2373 write_unlock(&em_tree->lock);
2374 } while (ret == -EEXIST);
2375 free_extent_map(hole_em);
2377 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2378 &inode->runtime_flags);
2385 * Find a hole extent on given inode and change start/len to the end of hole
2386 * extent.(hole/vacuum extent whose em->start <= start &&
2387 * em->start + em->len > start)
2388 * When a hole extent is found, return 1 and modify start/len.
2390 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2392 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2393 struct extent_map *em;
2396 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2397 round_down(*start, fs_info->sectorsize),
2398 round_up(*len, fs_info->sectorsize), 0);
2402 /* Hole or vacuum extent(only exists in no-hole mode) */
2403 if (em->block_start == EXTENT_MAP_HOLE) {
2405 *len = em->start + em->len > *start + *len ?
2406 0 : *start + *len - em->start - em->len;
2407 *start = em->start + em->len;
2409 free_extent_map(em);
2413 static int btrfs_punch_hole_lock_range(struct inode *inode,
2414 const u64 lockstart,
2416 struct extent_state **cached_state)
2419 struct btrfs_ordered_extent *ordered;
2422 truncate_pagecache_range(inode, lockstart, lockend);
2424 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2426 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2429 * We need to make sure we have no ordered extents in this range
2430 * and nobody raced in and read a page in this range, if we did
2431 * we need to try again.
2434 (ordered->file_offset + ordered->len <= lockstart ||
2435 ordered->file_offset > lockend)) &&
2436 !filemap_range_has_page(inode->i_mapping,
2437 lockstart, lockend)) {
2439 btrfs_put_ordered_extent(ordered);
2443 btrfs_put_ordered_extent(ordered);
2444 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2445 lockend, cached_state);
2446 ret = btrfs_wait_ordered_range(inode, lockstart,
2447 lockend - lockstart + 1);
2454 static int btrfs_insert_clone_extent(struct btrfs_trans_handle *trans,
2455 struct inode *inode,
2456 struct btrfs_path *path,
2457 struct btrfs_clone_extent_info *clone_info,
2458 const u64 clone_len)
2460 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2461 struct btrfs_root *root = BTRFS_I(inode)->root;
2462 struct btrfs_file_extent_item *extent;
2463 struct extent_buffer *leaf;
2464 struct btrfs_key key;
2466 struct btrfs_ref ref = { 0 };
2473 if (clone_info->disk_offset == 0 &&
2474 btrfs_fs_incompat(fs_info, NO_HOLES))
2477 key.objectid = btrfs_ino(BTRFS_I(inode));
2478 key.type = BTRFS_EXTENT_DATA_KEY;
2479 key.offset = clone_info->file_offset;
2480 ret = btrfs_insert_empty_item(trans, root, path, &key,
2481 clone_info->item_size);
2484 leaf = path->nodes[0];
2485 slot = path->slots[0];
2486 write_extent_buffer(leaf, clone_info->extent_buf,
2487 btrfs_item_ptr_offset(leaf, slot),
2488 clone_info->item_size);
2489 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2490 btrfs_set_file_extent_offset(leaf, extent, clone_info->data_offset);
2491 btrfs_set_file_extent_num_bytes(leaf, extent, clone_len);
2492 btrfs_mark_buffer_dirty(leaf);
2493 btrfs_release_path(path);
2495 /* If it's a hole, nothing more needs to be done. */
2496 if (clone_info->disk_offset == 0)
2499 inode_add_bytes(inode, clone_len);
2500 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2501 clone_info->disk_offset,
2502 clone_info->disk_len, 0);
2503 ref_offset = clone_info->file_offset - clone_info->data_offset;
2504 btrfs_init_data_ref(&ref, root->root_key.objectid,
2505 btrfs_ino(BTRFS_I(inode)), ref_offset);
2506 ret = btrfs_inc_extent_ref(trans, &ref);
2512 * The respective range must have been previously locked, as well as the inode.
2513 * The end offset is inclusive (last byte of the range).
2514 * @clone_info is NULL for fallocate's hole punching and non-NULL for extent
2516 * When cloning, we don't want to end up in a state where we dropped extents
2517 * without inserting a new one, so we must abort the transaction to avoid a
2520 int btrfs_punch_hole_range(struct inode *inode, struct btrfs_path *path,
2521 const u64 start, const u64 end,
2522 struct btrfs_clone_extent_info *clone_info,
2523 struct btrfs_trans_handle **trans_out)
2525 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2526 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2527 u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2528 struct btrfs_root *root = BTRFS_I(inode)->root;
2529 struct btrfs_trans_handle *trans = NULL;
2530 struct btrfs_block_rsv *rsv;
2531 unsigned int rsv_count;
2534 u64 len = end - start;
2540 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2545 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2549 * 1 - update the inode
2550 * 1 - removing the extents in the range
2551 * 1 - adding the hole extent if no_holes isn't set or if we are cloning
2554 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || clone_info)
2559 trans = btrfs_start_transaction(root, rsv_count);
2560 if (IS_ERR(trans)) {
2561 ret = PTR_ERR(trans);
2566 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2569 trans->block_rsv = rsv;
2572 while (cur_offset < end) {
2573 ret = __btrfs_drop_extents(trans, root, inode, path,
2574 cur_offset, end + 1, &drop_end,
2576 if (ret != -ENOSPC) {
2578 * When cloning we want to avoid transaction aborts when
2579 * nothing was done and we are attempting to clone parts
2580 * of inline extents, in such cases -EOPNOTSUPP is
2581 * returned by __btrfs_drop_extents() without having
2582 * changed anything in the file.
2584 if (clone_info && ret && ret != -EOPNOTSUPP)
2585 btrfs_abort_transaction(trans, ret);
2589 trans->block_rsv = &fs_info->trans_block_rsv;
2591 if (!clone_info && cur_offset < drop_end &&
2592 cur_offset < ino_size) {
2593 ret = fill_holes(trans, BTRFS_I(inode), path,
2594 cur_offset, drop_end);
2597 * If we failed then we didn't insert our hole
2598 * entries for the area we dropped, so now the
2599 * fs is corrupted, so we must abort the
2602 btrfs_abort_transaction(trans, ret);
2608 u64 clone_len = drop_end - cur_offset;
2610 ret = btrfs_insert_clone_extent(trans, inode, path,
2611 clone_info, clone_len);
2613 btrfs_abort_transaction(trans, ret);
2616 clone_info->data_len -= clone_len;
2617 clone_info->data_offset += clone_len;
2618 clone_info->file_offset += clone_len;
2621 cur_offset = drop_end;
2623 ret = btrfs_update_inode(trans, root, inode);
2627 btrfs_end_transaction(trans);
2628 btrfs_btree_balance_dirty(fs_info);
2630 trans = btrfs_start_transaction(root, rsv_count);
2631 if (IS_ERR(trans)) {
2632 ret = PTR_ERR(trans);
2637 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2638 rsv, min_size, false);
2639 BUG_ON(ret); /* shouldn't happen */
2640 trans->block_rsv = rsv;
2643 ret = find_first_non_hole(inode, &cur_offset, &len);
2644 if (unlikely(ret < 0))
2654 * If we were cloning, force the next fsync to be a full one since we
2655 * we replaced (or just dropped in the case of cloning holes when
2656 * NO_HOLES is enabled) extents and extent maps.
2657 * This is for the sake of simplicity, and cloning into files larger
2658 * than 16Mb would force the full fsync any way (when
2659 * try_release_extent_mapping() is invoked during page cache truncation.
2662 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2663 &BTRFS_I(inode)->runtime_flags);
2668 trans->block_rsv = &fs_info->trans_block_rsv;
2670 * If we are using the NO_HOLES feature we might have had already an
2671 * hole that overlaps a part of the region [lockstart, lockend] and
2672 * ends at (or beyond) lockend. Since we have no file extent items to
2673 * represent holes, drop_end can be less than lockend and so we must
2674 * make sure we have an extent map representing the existing hole (the
2675 * call to __btrfs_drop_extents() might have dropped the existing extent
2676 * map representing the existing hole), otherwise the fast fsync path
2677 * will not record the existence of the hole region
2678 * [existing_hole_start, lockend].
2680 if (drop_end <= end)
2683 * Don't insert file hole extent item if it's for a range beyond eof
2684 * (because it's useless) or if it represents a 0 bytes range (when
2685 * cur_offset == drop_end).
2687 if (!clone_info && cur_offset < ino_size && cur_offset < drop_end) {
2688 ret = fill_holes(trans, BTRFS_I(inode), path,
2689 cur_offset, drop_end);
2691 /* Same comment as above. */
2692 btrfs_abort_transaction(trans, ret);
2697 ret = btrfs_insert_clone_extent(trans, inode, path, clone_info,
2698 clone_info->data_len);
2700 btrfs_abort_transaction(trans, ret);
2709 trans->block_rsv = &fs_info->trans_block_rsv;
2711 btrfs_end_transaction(trans);
2715 btrfs_free_block_rsv(fs_info, rsv);
2720 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2722 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2723 struct btrfs_root *root = BTRFS_I(inode)->root;
2724 struct extent_state *cached_state = NULL;
2725 struct btrfs_path *path;
2726 struct btrfs_trans_handle *trans = NULL;
2731 u64 orig_start = offset;
2735 bool truncated_block = false;
2736 bool updated_inode = false;
2738 ret = btrfs_wait_ordered_range(inode, offset, len);
2743 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2744 ret = find_first_non_hole(inode, &offset, &len);
2746 goto out_only_mutex;
2748 /* Already in a large hole */
2750 goto out_only_mutex;
2753 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2754 lockend = round_down(offset + len,
2755 btrfs_inode_sectorsize(inode)) - 1;
2756 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2757 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2759 * We needn't truncate any block which is beyond the end of the file
2760 * because we are sure there is no data there.
2763 * Only do this if we are in the same block and we aren't doing the
2766 if (same_block && len < fs_info->sectorsize) {
2767 if (offset < ino_size) {
2768 truncated_block = true;
2769 ret = btrfs_truncate_block(inode, offset, len, 0);
2773 goto out_only_mutex;
2776 /* zero back part of the first block */
2777 if (offset < ino_size) {
2778 truncated_block = true;
2779 ret = btrfs_truncate_block(inode, offset, 0, 0);
2781 inode_unlock(inode);
2786 /* Check the aligned pages after the first unaligned page,
2787 * if offset != orig_start, which means the first unaligned page
2788 * including several following pages are already in holes,
2789 * the extra check can be skipped */
2790 if (offset == orig_start) {
2791 /* after truncate page, check hole again */
2792 len = offset + len - lockstart;
2794 ret = find_first_non_hole(inode, &offset, &len);
2796 goto out_only_mutex;
2799 goto out_only_mutex;
2804 /* Check the tail unaligned part is in a hole */
2805 tail_start = lockend + 1;
2806 tail_len = offset + len - tail_start;
2808 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2809 if (unlikely(ret < 0))
2810 goto out_only_mutex;
2812 /* zero the front end of the last page */
2813 if (tail_start + tail_len < ino_size) {
2814 truncated_block = true;
2815 ret = btrfs_truncate_block(inode,
2816 tail_start + tail_len,
2819 goto out_only_mutex;
2824 if (lockend < lockstart) {
2826 goto out_only_mutex;
2829 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2832 goto out_only_mutex;
2834 path = btrfs_alloc_path();
2840 ret = btrfs_punch_hole_range(inode, path, lockstart, lockend, NULL,
2842 btrfs_free_path(path);
2846 ASSERT(trans != NULL);
2847 inode_inc_iversion(inode);
2848 inode->i_mtime = inode->i_ctime = current_time(inode);
2849 ret = btrfs_update_inode(trans, root, inode);
2850 updated_inode = true;
2851 btrfs_end_transaction(trans);
2852 btrfs_btree_balance_dirty(fs_info);
2854 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2857 if (!updated_inode && truncated_block && !ret) {
2859 * If we only end up zeroing part of a page, we still need to
2860 * update the inode item, so that all the time fields are
2861 * updated as well as the necessary btrfs inode in memory fields
2862 * for detecting, at fsync time, if the inode isn't yet in the
2863 * log tree or it's there but not up to date.
2865 struct timespec64 now = current_time(inode);
2867 inode_inc_iversion(inode);
2868 inode->i_mtime = now;
2869 inode->i_ctime = now;
2870 trans = btrfs_start_transaction(root, 1);
2871 if (IS_ERR(trans)) {
2872 ret = PTR_ERR(trans);
2876 ret = btrfs_update_inode(trans, root, inode);
2877 ret2 = btrfs_end_transaction(trans);
2882 inode_unlock(inode);
2886 /* Helper structure to record which range is already reserved */
2887 struct falloc_range {
2888 struct list_head list;
2894 * Helper function to add falloc range
2896 * Caller should have locked the larger range of extent containing
2899 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2901 struct falloc_range *prev = NULL;
2902 struct falloc_range *range = NULL;
2904 if (list_empty(head))
2908 * As fallocate iterate by bytenr order, we only need to check
2911 prev = list_entry(head->prev, struct falloc_range, list);
2912 if (prev->start + prev->len == start) {
2917 range = kmalloc(sizeof(*range), GFP_KERNEL);
2920 range->start = start;
2922 list_add_tail(&range->list, head);
2926 static int btrfs_fallocate_update_isize(struct inode *inode,
2930 struct btrfs_trans_handle *trans;
2931 struct btrfs_root *root = BTRFS_I(inode)->root;
2935 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2938 trans = btrfs_start_transaction(root, 1);
2940 return PTR_ERR(trans);
2942 inode->i_ctime = current_time(inode);
2943 i_size_write(inode, end);
2944 btrfs_ordered_update_i_size(inode, end, NULL);
2945 ret = btrfs_update_inode(trans, root, inode);
2946 ret2 = btrfs_end_transaction(trans);
2948 return ret ? ret : ret2;
2952 RANGE_BOUNDARY_WRITTEN_EXTENT,
2953 RANGE_BOUNDARY_PREALLOC_EXTENT,
2954 RANGE_BOUNDARY_HOLE,
2957 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2960 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2961 struct extent_map *em;
2964 offset = round_down(offset, sectorsize);
2965 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2969 if (em->block_start == EXTENT_MAP_HOLE)
2970 ret = RANGE_BOUNDARY_HOLE;
2971 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2972 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2974 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2976 free_extent_map(em);
2980 static int btrfs_zero_range(struct inode *inode,
2985 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2986 struct extent_map *em;
2987 struct extent_changeset *data_reserved = NULL;
2990 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2991 u64 alloc_start = round_down(offset, sectorsize);
2992 u64 alloc_end = round_up(offset + len, sectorsize);
2993 u64 bytes_to_reserve = 0;
2994 bool space_reserved = false;
2996 inode_dio_wait(inode);
2998 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2999 alloc_start, alloc_end - alloc_start, 0);
3006 * Avoid hole punching and extent allocation for some cases. More cases
3007 * could be considered, but these are unlikely common and we keep things
3008 * as simple as possible for now. Also, intentionally, if the target
3009 * range contains one or more prealloc extents together with regular
3010 * extents and holes, we drop all the existing extents and allocate a
3011 * new prealloc extent, so that we get a larger contiguous disk extent.
3013 if (em->start <= alloc_start &&
3014 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3015 const u64 em_end = em->start + em->len;
3017 if (em_end >= offset + len) {
3019 * The whole range is already a prealloc extent,
3020 * do nothing except updating the inode's i_size if
3023 free_extent_map(em);
3024 ret = btrfs_fallocate_update_isize(inode, offset + len,
3029 * Part of the range is already a prealloc extent, so operate
3030 * only on the remaining part of the range.
3032 alloc_start = em_end;
3033 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3034 len = offset + len - alloc_start;
3035 offset = alloc_start;
3036 alloc_hint = em->block_start + em->len;
3038 free_extent_map(em);
3040 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3041 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3042 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
3043 alloc_start, sectorsize, 0);
3049 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3050 free_extent_map(em);
3051 ret = btrfs_fallocate_update_isize(inode, offset + len,
3055 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3056 free_extent_map(em);
3057 ret = btrfs_truncate_block(inode, offset, len, 0);
3059 ret = btrfs_fallocate_update_isize(inode,
3064 free_extent_map(em);
3065 alloc_start = round_down(offset, sectorsize);
3066 alloc_end = alloc_start + sectorsize;
3070 alloc_start = round_up(offset, sectorsize);
3071 alloc_end = round_down(offset + len, sectorsize);
3074 * For unaligned ranges, check the pages at the boundaries, they might
3075 * map to an extent, in which case we need to partially zero them, or
3076 * they might map to a hole, in which case we need our allocation range
3079 if (!IS_ALIGNED(offset, sectorsize)) {
3080 ret = btrfs_zero_range_check_range_boundary(inode, offset);
3083 if (ret == RANGE_BOUNDARY_HOLE) {
3084 alloc_start = round_down(offset, sectorsize);
3086 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3087 ret = btrfs_truncate_block(inode, offset, 0, 0);
3095 if (!IS_ALIGNED(offset + len, sectorsize)) {
3096 ret = btrfs_zero_range_check_range_boundary(inode,
3100 if (ret == RANGE_BOUNDARY_HOLE) {
3101 alloc_end = round_up(offset + len, sectorsize);
3103 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3104 ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3113 if (alloc_start < alloc_end) {
3114 struct extent_state *cached_state = NULL;
3115 const u64 lockstart = alloc_start;
3116 const u64 lockend = alloc_end - 1;
3118 bytes_to_reserve = alloc_end - alloc_start;
3119 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3123 space_reserved = true;
3124 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3125 alloc_start, bytes_to_reserve);
3128 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3132 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3133 alloc_end - alloc_start,
3135 offset + len, &alloc_hint);
3136 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3137 lockend, &cached_state);
3138 /* btrfs_prealloc_file_range releases reserved space on error */
3140 space_reserved = false;
3144 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3146 if (ret && space_reserved)
3147 btrfs_free_reserved_data_space(inode, data_reserved,
3148 alloc_start, bytes_to_reserve);
3149 extent_changeset_free(data_reserved);
3154 static long btrfs_fallocate(struct file *file, int mode,
3155 loff_t offset, loff_t len)
3157 struct inode *inode = file_inode(file);
3158 struct extent_state *cached_state = NULL;
3159 struct extent_changeset *data_reserved = NULL;
3160 struct falloc_range *range;
3161 struct falloc_range *tmp;
3162 struct list_head reserve_list;
3170 struct extent_map *em;
3171 int blocksize = btrfs_inode_sectorsize(inode);
3174 alloc_start = round_down(offset, blocksize);
3175 alloc_end = round_up(offset + len, blocksize);
3176 cur_offset = alloc_start;
3178 /* Make sure we aren't being give some crap mode */
3179 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3180 FALLOC_FL_ZERO_RANGE))
3183 if (mode & FALLOC_FL_PUNCH_HOLE)
3184 return btrfs_punch_hole(inode, offset, len);
3187 * Only trigger disk allocation, don't trigger qgroup reserve
3189 * For qgroup space, it will be checked later.
3191 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3192 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3193 alloc_end - alloc_start);
3200 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3201 ret = inode_newsize_ok(inode, offset + len);
3207 * TODO: Move these two operations after we have checked
3208 * accurate reserved space, or fallocate can still fail but
3209 * with page truncated or size expanded.
3211 * But that's a minor problem and won't do much harm BTW.
3213 if (alloc_start > inode->i_size) {
3214 ret = btrfs_cont_expand(inode, i_size_read(inode),
3218 } else if (offset + len > inode->i_size) {
3220 * If we are fallocating from the end of the file onward we
3221 * need to zero out the end of the block if i_size lands in the
3222 * middle of a block.
3224 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3230 * wait for ordered IO before we have any locks. We'll loop again
3231 * below with the locks held.
3233 ret = btrfs_wait_ordered_range(inode, alloc_start,
3234 alloc_end - alloc_start);
3238 if (mode & FALLOC_FL_ZERO_RANGE) {
3239 ret = btrfs_zero_range(inode, offset, len, mode);
3240 inode_unlock(inode);
3244 locked_end = alloc_end - 1;
3246 struct btrfs_ordered_extent *ordered;
3248 /* the extent lock is ordered inside the running
3251 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3252 locked_end, &cached_state);
3253 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3256 ordered->file_offset + ordered->len > alloc_start &&
3257 ordered->file_offset < alloc_end) {
3258 btrfs_put_ordered_extent(ordered);
3259 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3260 alloc_start, locked_end,
3263 * we can't wait on the range with the transaction
3264 * running or with the extent lock held
3266 ret = btrfs_wait_ordered_range(inode, alloc_start,
3267 alloc_end - alloc_start);
3272 btrfs_put_ordered_extent(ordered);
3277 /* First, check if we exceed the qgroup limit */
3278 INIT_LIST_HEAD(&reserve_list);
3279 while (cur_offset < alloc_end) {
3280 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3281 alloc_end - cur_offset, 0);
3286 last_byte = min(extent_map_end(em), alloc_end);
3287 actual_end = min_t(u64, extent_map_end(em), offset + len);
3288 last_byte = ALIGN(last_byte, blocksize);
3289 if (em->block_start == EXTENT_MAP_HOLE ||
3290 (cur_offset >= inode->i_size &&
3291 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3292 ret = add_falloc_range(&reserve_list, cur_offset,
3293 last_byte - cur_offset);
3295 free_extent_map(em);
3298 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3299 cur_offset, last_byte - cur_offset);
3301 cur_offset = last_byte;
3302 free_extent_map(em);
3307 * Do not need to reserve unwritten extent for this
3308 * range, free reserved data space first, otherwise
3309 * it'll result in false ENOSPC error.
3311 btrfs_free_reserved_data_space(inode, data_reserved,
3312 cur_offset, last_byte - cur_offset);
3314 free_extent_map(em);
3315 cur_offset = last_byte;
3319 * If ret is still 0, means we're OK to fallocate.
3320 * Or just cleanup the list and exit.
3322 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3324 ret = btrfs_prealloc_file_range(inode, mode,
3326 range->len, i_blocksize(inode),
3327 offset + len, &alloc_hint);
3329 btrfs_free_reserved_data_space(inode,
3330 data_reserved, range->start,
3332 list_del(&range->list);
3339 * We didn't need to allocate any more space, but we still extended the
3340 * size of the file so we need to update i_size and the inode item.
3342 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3344 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3347 inode_unlock(inode);
3348 /* Let go of our reservation. */
3349 if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3350 btrfs_free_reserved_data_space(inode, data_reserved,
3351 cur_offset, alloc_end - cur_offset);
3352 extent_changeset_free(data_reserved);
3356 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3358 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3359 struct extent_map *em = NULL;
3360 struct extent_state *cached_state = NULL;
3367 if (inode->i_size == 0)
3371 * *offset can be negative, in this case we start finding DATA/HOLE from
3372 * the very start of the file.
3374 start = max_t(loff_t, 0, *offset);
3376 lockstart = round_down(start, fs_info->sectorsize);
3377 lockend = round_up(i_size_read(inode),
3378 fs_info->sectorsize);
3379 if (lockend <= lockstart)
3380 lockend = lockstart + fs_info->sectorsize;
3382 len = lockend - lockstart + 1;
3384 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3387 while (start < inode->i_size) {
3388 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3395 if (whence == SEEK_HOLE &&
3396 (em->block_start == EXTENT_MAP_HOLE ||
3397 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3399 else if (whence == SEEK_DATA &&
3400 (em->block_start != EXTENT_MAP_HOLE &&
3401 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3404 start = em->start + em->len;
3405 free_extent_map(em);
3409 free_extent_map(em);
3411 if (whence == SEEK_DATA && start >= inode->i_size)
3414 *offset = min_t(loff_t, start, inode->i_size);
3416 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3421 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3423 struct inode *inode = file->f_mapping->host;
3430 offset = generic_file_llseek(file, offset, whence);
3434 if (offset >= i_size_read(inode)) {
3435 inode_unlock(inode);
3439 ret = find_desired_extent(inode, &offset, whence);
3441 inode_unlock(inode);
3446 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3448 inode_unlock(inode);
3452 static int btrfs_file_open(struct inode *inode, struct file *filp)
3454 filp->f_mode |= FMODE_NOWAIT;
3455 return generic_file_open(inode, filp);
3458 const struct file_operations btrfs_file_operations = {
3459 .llseek = btrfs_file_llseek,
3460 .read_iter = generic_file_read_iter,
3461 .splice_read = generic_file_splice_read,
3462 .write_iter = btrfs_file_write_iter,
3463 .mmap = btrfs_file_mmap,
3464 .open = btrfs_file_open,
3465 .release = btrfs_release_file,
3466 .fsync = btrfs_sync_file,
3467 .fallocate = btrfs_fallocate,
3468 .unlocked_ioctl = btrfs_ioctl,
3469 #ifdef CONFIG_COMPAT
3470 .compat_ioctl = btrfs_compat_ioctl,
3472 .remap_file_range = btrfs_remap_file_range,
3475 void __cold btrfs_auto_defrag_exit(void)
3477 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3480 int __init btrfs_auto_defrag_init(void)
3482 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3483 sizeof(struct inode_defrag), 0,
3486 if (!btrfs_inode_defrag_cachep)
3492 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3497 * So with compression we will find and lock a dirty page and clear the
3498 * first one as dirty, setup an async extent, and immediately return
3499 * with the entire range locked but with nobody actually marked with
3500 * writeback. So we can't just filemap_write_and_wait_range() and
3501 * expect it to work since it will just kick off a thread to do the
3502 * actual work. So we need to call filemap_fdatawrite_range _again_
3503 * since it will wait on the page lock, which won't be unlocked until
3504 * after the pages have been marked as writeback and so we're good to go
3505 * from there. We have to do this otherwise we'll miss the ordered
3506 * extents and that results in badness. Please Josef, do not think you
3507 * know better and pull this out at some point in the future, it is
3508 * right and you are wrong.
3510 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3511 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3512 &BTRFS_I(inode)->runtime_flags))
3513 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);