2 * Copyright (C) 2007 Oracle. All rights reserved.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
20 #include <linux/pagemap.h>
21 #include <linux/highmem.h>
22 #include <linux/time.h>
23 #include <linux/init.h>
24 #include <linux/string.h>
25 #include <linux/backing-dev.h>
26 #include <linux/mpage.h>
27 #include <linux/falloc.h>
28 #include <linux/swap.h>
29 #include <linux/writeback.h>
30 #include <linux/compat.h>
31 #include <linux/slab.h>
32 #include <linux/btrfs.h>
33 #include <linux/uio.h>
36 #include "transaction.h"
37 #include "btrfs_inode.h"
38 #include "print-tree.h"
43 #include "compression.h"
45 static struct kmem_cache *btrfs_inode_defrag_cachep;
47 * when auto defrag is enabled we
48 * queue up these defrag structs to remember which
49 * inodes need defragging passes
52 struct rb_node rb_node;
56 * transid where the defrag was added, we search for
57 * extents newer than this
64 /* last offset we were able to defrag */
67 /* if we've wrapped around back to zero once already */
71 static int __compare_inode_defrag(struct inode_defrag *defrag1,
72 struct inode_defrag *defrag2)
74 if (defrag1->root > defrag2->root)
76 else if (defrag1->root < defrag2->root)
78 else if (defrag1->ino > defrag2->ino)
80 else if (defrag1->ino < defrag2->ino)
86 /* pop a record for an inode into the defrag tree. The lock
87 * must be held already
89 * If you're inserting a record for an older transid than an
90 * existing record, the transid already in the tree is lowered
92 * If an existing record is found the defrag item you
95 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
96 struct inode_defrag *defrag)
98 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
99 struct inode_defrag *entry;
101 struct rb_node *parent = NULL;
104 p = &fs_info->defrag_inodes.rb_node;
107 entry = rb_entry(parent, struct inode_defrag, rb_node);
109 ret = __compare_inode_defrag(defrag, entry);
111 p = &parent->rb_left;
113 p = &parent->rb_right;
115 /* if we're reinserting an entry for
116 * an old defrag run, make sure to
117 * lower the transid of our existing record
119 if (defrag->transid < entry->transid)
120 entry->transid = defrag->transid;
121 if (defrag->last_offset > entry->last_offset)
122 entry->last_offset = defrag->last_offset;
126 set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
127 rb_link_node(&defrag->rb_node, parent, p);
128 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
132 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
134 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
137 if (btrfs_fs_closing(fs_info))
144 * insert a defrag record for this inode if auto defrag is
147 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
148 struct btrfs_inode *inode)
150 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
151 struct btrfs_root *root = inode->root;
152 struct inode_defrag *defrag;
156 if (!__need_auto_defrag(fs_info))
159 if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
163 transid = trans->transid;
165 transid = inode->root->last_trans;
167 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
171 defrag->ino = btrfs_ino(inode);
172 defrag->transid = transid;
173 defrag->root = root->root_key.objectid;
175 spin_lock(&fs_info->defrag_inodes_lock);
176 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
178 * If we set IN_DEFRAG flag and evict the inode from memory,
179 * and then re-read this inode, this new inode doesn't have
180 * IN_DEFRAG flag. At the case, we may find the existed defrag.
182 ret = __btrfs_add_inode_defrag(inode, defrag);
184 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
186 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
188 spin_unlock(&fs_info->defrag_inodes_lock);
193 * Requeue the defrag object. If there is a defrag object that points to
194 * the same inode in the tree, we will merge them together (by
195 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
197 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
198 struct inode_defrag *defrag)
200 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
203 if (!__need_auto_defrag(fs_info))
207 * Here we don't check the IN_DEFRAG flag, because we need merge
210 spin_lock(&fs_info->defrag_inodes_lock);
211 ret = __btrfs_add_inode_defrag(inode, defrag);
212 spin_unlock(&fs_info->defrag_inodes_lock);
217 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
221 * pick the defragable inode that we want, if it doesn't exist, we will get
224 static struct inode_defrag *
225 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
227 struct inode_defrag *entry = NULL;
228 struct inode_defrag tmp;
230 struct rb_node *parent = NULL;
236 spin_lock(&fs_info->defrag_inodes_lock);
237 p = fs_info->defrag_inodes.rb_node;
240 entry = rb_entry(parent, struct inode_defrag, rb_node);
242 ret = __compare_inode_defrag(&tmp, entry);
246 p = parent->rb_right;
251 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
252 parent = rb_next(parent);
254 entry = rb_entry(parent, struct inode_defrag, rb_node);
260 rb_erase(parent, &fs_info->defrag_inodes);
261 spin_unlock(&fs_info->defrag_inodes_lock);
265 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
267 struct inode_defrag *defrag;
268 struct rb_node *node;
270 spin_lock(&fs_info->defrag_inodes_lock);
271 node = rb_first(&fs_info->defrag_inodes);
273 rb_erase(node, &fs_info->defrag_inodes);
274 defrag = rb_entry(node, struct inode_defrag, rb_node);
275 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
277 cond_resched_lock(&fs_info->defrag_inodes_lock);
279 node = rb_first(&fs_info->defrag_inodes);
281 spin_unlock(&fs_info->defrag_inodes_lock);
284 #define BTRFS_DEFRAG_BATCH 1024
286 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
287 struct inode_defrag *defrag)
289 struct btrfs_root *inode_root;
291 struct btrfs_key key;
292 struct btrfs_ioctl_defrag_range_args range;
298 key.objectid = defrag->root;
299 key.type = BTRFS_ROOT_ITEM_KEY;
300 key.offset = (u64)-1;
302 index = srcu_read_lock(&fs_info->subvol_srcu);
304 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
305 if (IS_ERR(inode_root)) {
306 ret = PTR_ERR(inode_root);
310 key.objectid = defrag->ino;
311 key.type = BTRFS_INODE_ITEM_KEY;
313 inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
315 ret = PTR_ERR(inode);
318 srcu_read_unlock(&fs_info->subvol_srcu, index);
320 /* do a chunk of defrag */
321 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
322 memset(&range, 0, sizeof(range));
324 range.start = defrag->last_offset;
326 sb_start_write(fs_info->sb);
327 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
329 sb_end_write(fs_info->sb);
331 * if we filled the whole defrag batch, there
332 * must be more work to do. Queue this defrag
335 if (num_defrag == BTRFS_DEFRAG_BATCH) {
336 defrag->last_offset = range.start;
337 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
338 } else if (defrag->last_offset && !defrag->cycled) {
340 * we didn't fill our defrag batch, but
341 * we didn't start at zero. Make sure we loop
342 * around to the start of the file.
344 defrag->last_offset = 0;
346 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
348 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
354 srcu_read_unlock(&fs_info->subvol_srcu, index);
355 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
360 * run through the list of inodes in the FS that need
363 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
365 struct inode_defrag *defrag;
367 u64 root_objectid = 0;
369 atomic_inc(&fs_info->defrag_running);
371 /* Pause the auto defragger. */
372 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
376 if (!__need_auto_defrag(fs_info))
379 /* find an inode to defrag */
380 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
383 if (root_objectid || first_ino) {
392 first_ino = defrag->ino + 1;
393 root_objectid = defrag->root;
395 __btrfs_run_defrag_inode(fs_info, defrag);
397 atomic_dec(&fs_info->defrag_running);
400 * during unmount, we use the transaction_wait queue to
401 * wait for the defragger to stop
403 wake_up(&fs_info->transaction_wait);
407 /* simple helper to fault in pages and copy. This should go away
408 * and be replaced with calls into generic code.
410 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
411 struct page **prepared_pages,
415 size_t total_copied = 0;
417 int offset = pos & (PAGE_SIZE - 1);
419 while (write_bytes > 0) {
420 size_t count = min_t(size_t,
421 PAGE_SIZE - offset, write_bytes);
422 struct page *page = prepared_pages[pg];
424 * Copy data from userspace to the current page
426 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
428 /* Flush processor's dcache for this page */
429 flush_dcache_page(page);
432 * if we get a partial write, we can end up with
433 * partially up to date pages. These add
434 * a lot of complexity, so make sure they don't
435 * happen by forcing this copy to be retried.
437 * The rest of the btrfs_file_write code will fall
438 * back to page at a time copies after we return 0.
440 if (!PageUptodate(page) && copied < count)
443 iov_iter_advance(i, copied);
444 write_bytes -= copied;
445 total_copied += copied;
447 /* Return to btrfs_file_write_iter to fault page */
448 if (unlikely(copied == 0))
451 if (copied < PAGE_SIZE - offset) {
462 * unlocks pages after btrfs_file_write is done with them
464 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
467 for (i = 0; i < num_pages; i++) {
468 /* page checked is some magic around finding pages that
469 * have been modified without going through btrfs_set_page_dirty
470 * clear it here. There should be no need to mark the pages
471 * accessed as prepare_pages should have marked them accessed
472 * in prepare_pages via find_or_create_page()
474 ClearPageChecked(pages[i]);
475 unlock_page(pages[i]);
480 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
483 struct extent_state **cached_state)
485 u64 search_start = start;
486 const u64 end = start + len - 1;
488 while (search_start < end) {
489 const u64 search_len = end - search_start + 1;
490 struct extent_map *em;
494 em = btrfs_get_extent(inode, NULL, 0, search_start,
499 if (em->block_start != EXTENT_MAP_HOLE)
503 if (em->start < search_start)
504 em_len -= search_start - em->start;
505 if (em_len > search_len)
508 ret = set_extent_bit(&inode->io_tree, search_start,
509 search_start + em_len - 1,
511 NULL, cached_state, GFP_NOFS);
513 search_start = extent_map_end(em);
522 * after copy_from_user, pages need to be dirtied and we need to make
523 * sure holes are created between the current EOF and the start of
524 * any next extents (if required).
526 * this also makes the decision about creating an inline extent vs
527 * doing real data extents, marking pages dirty and delalloc as required.
529 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
530 size_t num_pages, loff_t pos, size_t write_bytes,
531 struct extent_state **cached)
533 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
538 u64 end_of_last_block;
539 u64 end_pos = pos + write_bytes;
540 loff_t isize = i_size_read(inode);
541 unsigned int extra_bits = 0;
543 start_pos = pos & ~((u64) fs_info->sectorsize - 1);
544 num_bytes = round_up(write_bytes + pos - start_pos,
545 fs_info->sectorsize);
547 end_of_last_block = start_pos + num_bytes - 1;
549 if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
550 if (start_pos >= isize &&
551 !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
553 * There can't be any extents following eof in this case
554 * so just set the delalloc new bit for the range
557 extra_bits |= EXTENT_DELALLOC_NEW;
559 err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
567 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
568 extra_bits, cached, 0);
572 for (i = 0; i < num_pages; i++) {
573 struct page *p = pages[i];
580 * we've only changed i_size in ram, and we haven't updated
581 * the disk i_size. There is no need to log the inode
585 i_size_write(inode, end_pos);
590 * this drops all the extents in the cache that intersect the range
591 * [start, end]. Existing extents are split as required.
593 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
596 struct extent_map *em;
597 struct extent_map *split = NULL;
598 struct extent_map *split2 = NULL;
599 struct extent_map_tree *em_tree = &inode->extent_tree;
600 u64 len = end - start + 1;
608 WARN_ON(end < start);
609 if (end == (u64)-1) {
618 split = alloc_extent_map();
620 split2 = alloc_extent_map();
621 if (!split || !split2)
624 write_lock(&em_tree->lock);
625 em = lookup_extent_mapping(em_tree, start, len);
627 write_unlock(&em_tree->lock);
631 gen = em->generation;
632 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
633 if (testend && em->start + em->len >= start + len) {
635 write_unlock(&em_tree->lock);
638 start = em->start + em->len;
640 len = start + len - (em->start + em->len);
642 write_unlock(&em_tree->lock);
645 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
646 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
647 clear_bit(EXTENT_FLAG_LOGGING, &flags);
648 modified = !list_empty(&em->list);
652 if (em->start < start) {
653 split->start = em->start;
654 split->len = start - em->start;
656 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
657 split->orig_start = em->orig_start;
658 split->block_start = em->block_start;
661 split->block_len = em->block_len;
663 split->block_len = split->len;
664 split->orig_block_len = max(split->block_len,
666 split->ram_bytes = em->ram_bytes;
668 split->orig_start = split->start;
669 split->block_len = 0;
670 split->block_start = em->block_start;
671 split->orig_block_len = 0;
672 split->ram_bytes = split->len;
675 split->generation = gen;
676 split->bdev = em->bdev;
677 split->flags = flags;
678 split->compress_type = em->compress_type;
679 replace_extent_mapping(em_tree, em, split, modified);
680 free_extent_map(split);
684 if (testend && em->start + em->len > start + len) {
685 u64 diff = start + len - em->start;
687 split->start = start + len;
688 split->len = em->start + em->len - (start + len);
689 split->bdev = em->bdev;
690 split->flags = flags;
691 split->compress_type = em->compress_type;
692 split->generation = gen;
694 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
695 split->orig_block_len = max(em->block_len,
698 split->ram_bytes = em->ram_bytes;
700 split->block_len = em->block_len;
701 split->block_start = em->block_start;
702 split->orig_start = em->orig_start;
704 split->block_len = split->len;
705 split->block_start = em->block_start
707 split->orig_start = em->orig_start;
710 split->ram_bytes = split->len;
711 split->orig_start = split->start;
712 split->block_len = 0;
713 split->block_start = em->block_start;
714 split->orig_block_len = 0;
717 if (extent_map_in_tree(em)) {
718 replace_extent_mapping(em_tree, em, split,
721 ret = add_extent_mapping(em_tree, split,
723 ASSERT(ret == 0); /* Logic error */
725 free_extent_map(split);
729 if (extent_map_in_tree(em))
730 remove_extent_mapping(em_tree, em);
731 write_unlock(&em_tree->lock);
735 /* once for the tree*/
739 free_extent_map(split);
741 free_extent_map(split2);
745 * this is very complex, but the basic idea is to drop all extents
746 * in the range start - end. hint_block is filled in with a block number
747 * that would be a good hint to the block allocator for this file.
749 * If an extent intersects the range but is not entirely inside the range
750 * it is either truncated or split. Anything entirely inside the range
751 * is deleted from the tree.
753 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
754 struct btrfs_root *root, struct inode *inode,
755 struct btrfs_path *path, u64 start, u64 end,
756 u64 *drop_end, int drop_cache,
758 u32 extent_item_size,
761 struct btrfs_fs_info *fs_info = root->fs_info;
762 struct extent_buffer *leaf;
763 struct btrfs_file_extent_item *fi;
764 struct btrfs_key key;
765 struct btrfs_key new_key;
766 u64 ino = btrfs_ino(BTRFS_I(inode));
767 u64 search_start = start;
770 u64 extent_offset = 0;
772 u64 last_end = start;
778 int modify_tree = -1;
781 int leafs_visited = 0;
784 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
786 if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
789 update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
790 root == fs_info->tree_root);
793 ret = btrfs_lookup_file_extent(trans, root, path, ino,
794 search_start, modify_tree);
797 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
798 leaf = path->nodes[0];
799 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
800 if (key.objectid == ino &&
801 key.type == BTRFS_EXTENT_DATA_KEY)
807 leaf = path->nodes[0];
808 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
810 ret = btrfs_next_leaf(root, path);
818 leaf = path->nodes[0];
822 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
824 if (key.objectid > ino)
826 if (WARN_ON_ONCE(key.objectid < ino) ||
827 key.type < BTRFS_EXTENT_DATA_KEY) {
832 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
835 fi = btrfs_item_ptr(leaf, path->slots[0],
836 struct btrfs_file_extent_item);
837 extent_type = btrfs_file_extent_type(leaf, fi);
839 if (extent_type == BTRFS_FILE_EXTENT_REG ||
840 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
841 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
842 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
843 extent_offset = btrfs_file_extent_offset(leaf, fi);
844 extent_end = key.offset +
845 btrfs_file_extent_num_bytes(leaf, fi);
846 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
847 extent_end = key.offset +
848 btrfs_file_extent_inline_len(leaf,
856 * Don't skip extent items representing 0 byte lengths. They
857 * used to be created (bug) if while punching holes we hit
858 * -ENOSPC condition. So if we find one here, just ensure we
859 * delete it, otherwise we would insert a new file extent item
860 * with the same key (offset) as that 0 bytes length file
861 * extent item in the call to setup_items_for_insert() later
864 if (extent_end == key.offset && extent_end >= search_start) {
865 last_end = extent_end;
866 goto delete_extent_item;
869 if (extent_end <= search_start) {
875 search_start = max(key.offset, start);
876 if (recow || !modify_tree) {
878 btrfs_release_path(path);
883 * | - range to drop - |
884 * | -------- extent -------- |
886 if (start > key.offset && end < extent_end) {
888 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
893 memcpy(&new_key, &key, sizeof(new_key));
894 new_key.offset = start;
895 ret = btrfs_duplicate_item(trans, root, path,
897 if (ret == -EAGAIN) {
898 btrfs_release_path(path);
904 leaf = path->nodes[0];
905 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
906 struct btrfs_file_extent_item);
907 btrfs_set_file_extent_num_bytes(leaf, fi,
910 fi = btrfs_item_ptr(leaf, path->slots[0],
911 struct btrfs_file_extent_item);
913 extent_offset += start - key.offset;
914 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
915 btrfs_set_file_extent_num_bytes(leaf, fi,
917 btrfs_mark_buffer_dirty(leaf);
919 if (update_refs && disk_bytenr > 0) {
920 ret = btrfs_inc_extent_ref(trans, root,
921 disk_bytenr, num_bytes, 0,
922 root->root_key.objectid,
924 start - extent_offset);
925 BUG_ON(ret); /* -ENOMEM */
930 * From here on out we will have actually dropped something, so
931 * last_end can be updated.
933 last_end = extent_end;
936 * | ---- range to drop ----- |
937 * | -------- extent -------- |
939 if (start <= key.offset && end < extent_end) {
940 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
945 memcpy(&new_key, &key, sizeof(new_key));
946 new_key.offset = end;
947 btrfs_set_item_key_safe(fs_info, path, &new_key);
949 extent_offset += end - key.offset;
950 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
951 btrfs_set_file_extent_num_bytes(leaf, fi,
953 btrfs_mark_buffer_dirty(leaf);
954 if (update_refs && disk_bytenr > 0)
955 inode_sub_bytes(inode, end - key.offset);
959 search_start = extent_end;
961 * | ---- range to drop ----- |
962 * | -------- extent -------- |
964 if (start > key.offset && end >= extent_end) {
966 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
971 btrfs_set_file_extent_num_bytes(leaf, fi,
973 btrfs_mark_buffer_dirty(leaf);
974 if (update_refs && disk_bytenr > 0)
975 inode_sub_bytes(inode, extent_end - start);
976 if (end == extent_end)
984 * | ---- range to drop ----- |
985 * | ------ extent ------ |
987 if (start <= key.offset && end >= extent_end) {
990 del_slot = path->slots[0];
993 BUG_ON(del_slot + del_nr != path->slots[0]);
998 extent_type == BTRFS_FILE_EXTENT_INLINE) {
999 inode_sub_bytes(inode,
1000 extent_end - key.offset);
1001 extent_end = ALIGN(extent_end,
1002 fs_info->sectorsize);
1003 } else if (update_refs && disk_bytenr > 0) {
1004 ret = btrfs_free_extent(trans, root,
1005 disk_bytenr, num_bytes, 0,
1006 root->root_key.objectid,
1007 key.objectid, key.offset -
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(fs_info, 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 = btrfs_sb(inode->vfs_inode.i_sb);
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_key key;
1154 struct btrfs_key new_key;
1166 u64 ino = btrfs_ino(inode);
1168 path = btrfs_alloc_path();
1175 key.type = BTRFS_EXTENT_DATA_KEY;
1178 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1181 if (ret > 0 && path->slots[0] > 0)
1184 leaf = path->nodes[0];
1185 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1186 if (key.objectid != ino ||
1187 key.type != BTRFS_EXTENT_DATA_KEY) {
1189 btrfs_abort_transaction(trans, ret);
1192 fi = btrfs_item_ptr(leaf, path->slots[0],
1193 struct btrfs_file_extent_item);
1194 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1196 btrfs_abort_transaction(trans, ret);
1199 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1200 if (key.offset > start || extent_end < end) {
1202 btrfs_abort_transaction(trans, ret);
1206 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1207 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1208 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1209 memcpy(&new_key, &key, sizeof(new_key));
1211 if (start == key.offset && end < extent_end) {
1214 if (extent_mergeable(leaf, path->slots[0] - 1,
1215 ino, bytenr, orig_offset,
1216 &other_start, &other_end)) {
1217 new_key.offset = end;
1218 btrfs_set_item_key_safe(fs_info, path, &new_key);
1219 fi = btrfs_item_ptr(leaf, path->slots[0],
1220 struct btrfs_file_extent_item);
1221 btrfs_set_file_extent_generation(leaf, fi,
1223 btrfs_set_file_extent_num_bytes(leaf, fi,
1225 btrfs_set_file_extent_offset(leaf, fi,
1227 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1228 struct btrfs_file_extent_item);
1229 btrfs_set_file_extent_generation(leaf, fi,
1231 btrfs_set_file_extent_num_bytes(leaf, fi,
1233 btrfs_mark_buffer_dirty(leaf);
1238 if (start > key.offset && end == extent_end) {
1241 if (extent_mergeable(leaf, path->slots[0] + 1,
1242 ino, bytenr, orig_offset,
1243 &other_start, &other_end)) {
1244 fi = btrfs_item_ptr(leaf, path->slots[0],
1245 struct btrfs_file_extent_item);
1246 btrfs_set_file_extent_num_bytes(leaf, fi,
1247 start - key.offset);
1248 btrfs_set_file_extent_generation(leaf, fi,
1251 new_key.offset = start;
1252 btrfs_set_item_key_safe(fs_info, path, &new_key);
1254 fi = btrfs_item_ptr(leaf, path->slots[0],
1255 struct btrfs_file_extent_item);
1256 btrfs_set_file_extent_generation(leaf, fi,
1258 btrfs_set_file_extent_num_bytes(leaf, fi,
1260 btrfs_set_file_extent_offset(leaf, fi,
1261 start - orig_offset);
1262 btrfs_mark_buffer_dirty(leaf);
1267 while (start > key.offset || end < extent_end) {
1268 if (key.offset == start)
1271 new_key.offset = split;
1272 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1273 if (ret == -EAGAIN) {
1274 btrfs_release_path(path);
1278 btrfs_abort_transaction(trans, ret);
1282 leaf = path->nodes[0];
1283 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1284 struct btrfs_file_extent_item);
1285 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1286 btrfs_set_file_extent_num_bytes(leaf, fi,
1287 split - key.offset);
1289 fi = btrfs_item_ptr(leaf, path->slots[0],
1290 struct btrfs_file_extent_item);
1292 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1293 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1294 btrfs_set_file_extent_num_bytes(leaf, fi,
1295 extent_end - split);
1296 btrfs_mark_buffer_dirty(leaf);
1298 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes,
1299 0, root->root_key.objectid,
1302 btrfs_abort_transaction(trans, ret);
1306 if (split == start) {
1309 if (start != key.offset) {
1311 btrfs_abort_transaction(trans, ret);
1322 if (extent_mergeable(leaf, path->slots[0] + 1,
1323 ino, bytenr, orig_offset,
1324 &other_start, &other_end)) {
1326 btrfs_release_path(path);
1329 extent_end = other_end;
1330 del_slot = path->slots[0] + 1;
1332 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1333 0, root->root_key.objectid,
1336 btrfs_abort_transaction(trans, ret);
1342 if (extent_mergeable(leaf, path->slots[0] - 1,
1343 ino, bytenr, orig_offset,
1344 &other_start, &other_end)) {
1346 btrfs_release_path(path);
1349 key.offset = other_start;
1350 del_slot = path->slots[0];
1352 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1353 0, root->root_key.objectid,
1356 btrfs_abort_transaction(trans, ret);
1361 fi = btrfs_item_ptr(leaf, path->slots[0],
1362 struct btrfs_file_extent_item);
1363 btrfs_set_file_extent_type(leaf, fi,
1364 BTRFS_FILE_EXTENT_REG);
1365 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1366 btrfs_mark_buffer_dirty(leaf);
1368 fi = btrfs_item_ptr(leaf, del_slot - 1,
1369 struct btrfs_file_extent_item);
1370 btrfs_set_file_extent_type(leaf, fi,
1371 BTRFS_FILE_EXTENT_REG);
1372 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1373 btrfs_set_file_extent_num_bytes(leaf, fi,
1374 extent_end - key.offset);
1375 btrfs_mark_buffer_dirty(leaf);
1377 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1379 btrfs_abort_transaction(trans, ret);
1384 btrfs_free_path(path);
1389 * on error we return an unlocked page and the error value
1390 * on success we return a locked page and 0
1392 static int prepare_uptodate_page(struct inode *inode,
1393 struct page *page, u64 pos,
1394 bool force_uptodate)
1398 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1399 !PageUptodate(page)) {
1400 ret = btrfs_readpage(NULL, page);
1404 if (!PageUptodate(page)) {
1408 if (page->mapping != inode->i_mapping) {
1417 * this just gets pages into the page cache and locks them down.
1419 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1420 size_t num_pages, loff_t pos,
1421 size_t write_bytes, bool force_uptodate)
1424 unsigned long index = pos >> PAGE_SHIFT;
1425 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1429 for (i = 0; i < num_pages; i++) {
1431 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1432 mask | __GFP_WRITE);
1440 err = prepare_uptodate_page(inode, pages[i], pos,
1442 if (!err && i == num_pages - 1)
1443 err = prepare_uptodate_page(inode, pages[i],
1444 pos + write_bytes, false);
1447 if (err == -EAGAIN) {
1454 wait_on_page_writeback(pages[i]);
1459 while (faili >= 0) {
1460 unlock_page(pages[faili]);
1461 put_page(pages[faili]);
1469 * This function locks the extent and properly waits for data=ordered extents
1470 * to finish before allowing the pages to be modified if need.
1473 * 1 - the extent is locked
1474 * 0 - the extent is not locked, and everything is OK
1475 * -EAGAIN - need re-prepare the pages
1476 * the other < 0 number - Something wrong happens
1479 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1480 size_t num_pages, loff_t pos,
1482 u64 *lockstart, u64 *lockend,
1483 struct extent_state **cached_state)
1485 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1491 start_pos = round_down(pos, fs_info->sectorsize);
1492 last_pos = start_pos
1493 + round_up(pos + write_bytes - start_pos,
1494 fs_info->sectorsize) - 1;
1496 if (start_pos < inode->vfs_inode.i_size) {
1497 struct btrfs_ordered_extent *ordered;
1499 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1501 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1502 last_pos - start_pos + 1);
1504 ordered->file_offset + ordered->len > start_pos &&
1505 ordered->file_offset <= last_pos) {
1506 unlock_extent_cached(&inode->io_tree, start_pos,
1507 last_pos, cached_state);
1508 for (i = 0; i < num_pages; i++) {
1509 unlock_page(pages[i]);
1512 btrfs_start_ordered_extent(&inode->vfs_inode,
1514 btrfs_put_ordered_extent(ordered);
1518 btrfs_put_ordered_extent(ordered);
1519 clear_extent_bit(&inode->io_tree, start_pos, last_pos,
1520 EXTENT_DIRTY | EXTENT_DELALLOC |
1521 EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1522 0, 0, cached_state);
1523 *lockstart = start_pos;
1524 *lockend = last_pos;
1528 for (i = 0; i < num_pages; i++) {
1529 if (clear_page_dirty_for_io(pages[i]))
1530 account_page_redirty(pages[i]);
1531 set_page_extent_mapped(pages[i]);
1532 WARN_ON(!PageLocked(pages[i]));
1538 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1539 size_t *write_bytes)
1541 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1542 struct btrfs_root *root = inode->root;
1543 struct btrfs_ordered_extent *ordered;
1544 u64 lockstart, lockend;
1548 ret = btrfs_start_write_no_snapshotting(root);
1552 lockstart = round_down(pos, fs_info->sectorsize);
1553 lockend = round_up(pos + *write_bytes,
1554 fs_info->sectorsize) - 1;
1557 lock_extent(&inode->io_tree, lockstart, lockend);
1558 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1559 lockend - lockstart + 1);
1563 unlock_extent(&inode->io_tree, lockstart, lockend);
1564 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1565 btrfs_put_ordered_extent(ordered);
1568 num_bytes = lockend - lockstart + 1;
1569 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1573 btrfs_end_write_no_snapshotting(root);
1575 *write_bytes = min_t(size_t, *write_bytes ,
1576 num_bytes - pos + lockstart);
1579 unlock_extent(&inode->io_tree, lockstart, lockend);
1584 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1588 struct inode *inode = file_inode(file);
1589 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1590 struct btrfs_root *root = BTRFS_I(inode)->root;
1591 struct page **pages = NULL;
1592 struct extent_state *cached_state = NULL;
1593 struct extent_changeset *data_reserved = NULL;
1594 u64 release_bytes = 0;
1597 size_t num_written = 0;
1600 bool only_release_metadata = false;
1601 bool force_page_uptodate = false;
1603 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1604 PAGE_SIZE / (sizeof(struct page *)));
1605 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1606 nrptrs = max(nrptrs, 8);
1607 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1611 while (iov_iter_count(i) > 0) {
1612 size_t offset = pos & (PAGE_SIZE - 1);
1613 size_t sector_offset;
1614 size_t write_bytes = min(iov_iter_count(i),
1615 nrptrs * (size_t)PAGE_SIZE -
1617 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1619 size_t reserve_bytes;
1622 size_t dirty_sectors;
1626 WARN_ON(num_pages > nrptrs);
1629 * Fault pages before locking them in prepare_pages
1630 * to avoid recursive lock
1632 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1637 sector_offset = pos & (fs_info->sectorsize - 1);
1638 reserve_bytes = round_up(write_bytes + sector_offset,
1639 fs_info->sectorsize);
1641 extent_changeset_release(data_reserved);
1642 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1645 if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1646 BTRFS_INODE_PREALLOC)) &&
1647 check_can_nocow(BTRFS_I(inode), pos,
1648 &write_bytes) > 0) {
1650 * For nodata cow case, no need to reserve
1653 only_release_metadata = true;
1655 * our prealloc extent may be smaller than
1656 * write_bytes, so scale down.
1658 num_pages = DIV_ROUND_UP(write_bytes + offset,
1660 reserve_bytes = round_up(write_bytes +
1662 fs_info->sectorsize);
1668 WARN_ON(reserve_bytes == 0);
1669 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1672 if (!only_release_metadata)
1673 btrfs_free_reserved_data_space(inode,
1677 btrfs_end_write_no_snapshotting(root);
1681 release_bytes = reserve_bytes;
1684 * This is going to setup the pages array with the number of
1685 * pages we want, so we don't really need to worry about the
1686 * contents of pages from loop to loop
1688 ret = prepare_pages(inode, pages, num_pages,
1690 force_page_uptodate);
1692 btrfs_delalloc_release_extents(BTRFS_I(inode),
1697 extents_locked = lock_and_cleanup_extent_if_need(
1698 BTRFS_I(inode), pages,
1699 num_pages, pos, write_bytes, &lockstart,
1700 &lockend, &cached_state);
1701 if (extents_locked < 0) {
1702 if (extents_locked == -EAGAIN)
1704 btrfs_delalloc_release_extents(BTRFS_I(inode),
1706 ret = extents_locked;
1710 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1712 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1713 dirty_sectors = round_up(copied + sector_offset,
1714 fs_info->sectorsize);
1715 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1718 * if we have trouble faulting in the pages, fall
1719 * back to one page at a time
1721 if (copied < write_bytes)
1725 force_page_uptodate = true;
1729 force_page_uptodate = false;
1730 dirty_pages = DIV_ROUND_UP(copied + offset,
1734 if (num_sectors > dirty_sectors) {
1735 /* release everything except the sectors we dirtied */
1736 release_bytes -= dirty_sectors <<
1737 fs_info->sb->s_blocksize_bits;
1738 if (only_release_metadata) {
1739 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1744 __pos = round_down(pos,
1745 fs_info->sectorsize) +
1746 (dirty_pages << PAGE_SHIFT);
1747 btrfs_delalloc_release_space(inode,
1748 data_reserved, __pos,
1753 release_bytes = round_up(copied + sector_offset,
1754 fs_info->sectorsize);
1757 ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1758 pos, copied, &cached_state);
1760 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1761 lockstart, lockend, &cached_state);
1762 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1764 btrfs_drop_pages(pages, num_pages);
1769 if (only_release_metadata)
1770 btrfs_end_write_no_snapshotting(root);
1772 if (only_release_metadata && copied > 0) {
1773 lockstart = round_down(pos,
1774 fs_info->sectorsize);
1775 lockend = round_up(pos + copied,
1776 fs_info->sectorsize) - 1;
1778 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1779 lockend, EXTENT_NORESERVE, NULL,
1781 only_release_metadata = false;
1784 btrfs_drop_pages(pages, num_pages);
1788 balance_dirty_pages_ratelimited(inode->i_mapping);
1789 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1790 btrfs_btree_balance_dirty(fs_info);
1793 num_written += copied;
1798 if (release_bytes) {
1799 if (only_release_metadata) {
1800 btrfs_end_write_no_snapshotting(root);
1801 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1804 btrfs_delalloc_release_space(inode, data_reserved,
1805 round_down(pos, fs_info->sectorsize),
1810 extent_changeset_free(data_reserved);
1811 return num_written ? num_written : ret;
1814 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1816 struct file *file = iocb->ki_filp;
1817 struct inode *inode = file_inode(file);
1818 loff_t pos = iocb->ki_pos;
1820 ssize_t written_buffered;
1824 written = generic_file_direct_write(iocb, from);
1826 if (written < 0 || !iov_iter_count(from))
1830 written_buffered = __btrfs_buffered_write(file, from, pos);
1831 if (written_buffered < 0) {
1832 err = written_buffered;
1836 * Ensure all data is persisted. We want the next direct IO read to be
1837 * able to read what was just written.
1839 endbyte = pos + written_buffered - 1;
1840 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1843 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1846 written += written_buffered;
1847 iocb->ki_pos = pos + written_buffered;
1848 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1849 endbyte >> PAGE_SHIFT);
1851 return written ? written : err;
1854 static void update_time_for_write(struct inode *inode)
1856 struct timespec now;
1858 if (IS_NOCMTIME(inode))
1861 now = current_time(inode);
1862 if (!timespec_equal(&inode->i_mtime, &now))
1863 inode->i_mtime = now;
1865 if (!timespec_equal(&inode->i_ctime, &now))
1866 inode->i_ctime = now;
1868 if (IS_I_VERSION(inode))
1869 inode_inc_iversion(inode);
1872 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1873 struct iov_iter *from)
1875 struct file *file = iocb->ki_filp;
1876 struct inode *inode = file_inode(file);
1877 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1878 struct btrfs_root *root = BTRFS_I(inode)->root;
1881 ssize_t num_written = 0;
1882 bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1885 size_t count = iov_iter_count(from);
1889 if (!(iocb->ki_flags & IOCB_DIRECT) &&
1890 (iocb->ki_flags & IOCB_NOWAIT))
1893 if (!inode_trylock(inode)) {
1894 if (iocb->ki_flags & IOCB_NOWAIT)
1899 err = generic_write_checks(iocb, from);
1901 inode_unlock(inode);
1906 if (iocb->ki_flags & IOCB_NOWAIT) {
1908 * We will allocate space in case nodatacow is not set,
1911 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1912 BTRFS_INODE_PREALLOC)) ||
1913 check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1914 inode_unlock(inode);
1919 current->backing_dev_info = inode_to_bdi(inode);
1920 err = file_remove_privs(file);
1922 inode_unlock(inode);
1927 * If BTRFS flips readonly due to some impossible error
1928 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1929 * although we have opened a file as writable, we have
1930 * to stop this write operation to ensure FS consistency.
1932 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1933 inode_unlock(inode);
1939 * We reserve space for updating the inode when we reserve space for the
1940 * extent we are going to write, so we will enospc out there. We don't
1941 * need to start yet another transaction to update the inode as we will
1942 * update the inode when we finish writing whatever data we write.
1944 update_time_for_write(inode);
1946 start_pos = round_down(pos, fs_info->sectorsize);
1947 oldsize = i_size_read(inode);
1948 if (start_pos > oldsize) {
1949 /* Expand hole size to cover write data, preventing empty gap */
1950 end_pos = round_up(pos + count,
1951 fs_info->sectorsize);
1952 err = btrfs_cont_expand(inode, oldsize, end_pos);
1954 inode_unlock(inode);
1957 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1962 atomic_inc(&BTRFS_I(inode)->sync_writers);
1964 if (iocb->ki_flags & IOCB_DIRECT) {
1965 num_written = __btrfs_direct_write(iocb, from);
1967 num_written = __btrfs_buffered_write(file, from, pos);
1968 if (num_written > 0)
1969 iocb->ki_pos = pos + num_written;
1971 pagecache_isize_extended(inode, oldsize,
1972 i_size_read(inode));
1975 inode_unlock(inode);
1978 * We also have to set last_sub_trans to the current log transid,
1979 * otherwise subsequent syncs to a file that's been synced in this
1980 * transaction will appear to have already occurred.
1982 spin_lock(&BTRFS_I(inode)->lock);
1983 BTRFS_I(inode)->last_sub_trans = root->log_transid;
1984 spin_unlock(&BTRFS_I(inode)->lock);
1985 if (num_written > 0)
1986 num_written = generic_write_sync(iocb, num_written);
1989 atomic_dec(&BTRFS_I(inode)->sync_writers);
1991 current->backing_dev_info = NULL;
1992 return num_written ? num_written : err;
1995 int btrfs_release_file(struct inode *inode, struct file *filp)
1997 struct btrfs_file_private *private = filp->private_data;
1999 if (private && private->trans)
2000 btrfs_ioctl_trans_end(filp);
2001 if (private && private->filldir_buf)
2002 kfree(private->filldir_buf);
2004 filp->private_data = NULL;
2007 * ordered_data_close is set by settattr when we are about to truncate
2008 * a file from a non-zero size to a zero size. This tries to
2009 * flush down new bytes that may have been written if the
2010 * application were using truncate to replace a file in place.
2012 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2013 &BTRFS_I(inode)->runtime_flags))
2014 filemap_flush(inode->i_mapping);
2018 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2021 struct blk_plug plug;
2024 * This is only called in fsync, which would do synchronous writes, so
2025 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2026 * multiple disks using raid profile, a large IO can be split to
2027 * several segments of stripe length (currently 64K).
2029 blk_start_plug(&plug);
2030 atomic_inc(&BTRFS_I(inode)->sync_writers);
2031 ret = btrfs_fdatawrite_range(inode, start, end);
2032 atomic_dec(&BTRFS_I(inode)->sync_writers);
2033 blk_finish_plug(&plug);
2039 * fsync call for both files and directories. This logs the inode into
2040 * the tree log instead of forcing full commits whenever possible.
2042 * It needs to call filemap_fdatawait so that all ordered extent updates are
2043 * in the metadata btree are up to date for copying to the log.
2045 * It drops the inode mutex before doing the tree log commit. This is an
2046 * important optimization for directories because holding the mutex prevents
2047 * new operations on the dir while we write to disk.
2049 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2051 struct dentry *dentry = file_dentry(file);
2052 struct inode *inode = d_inode(dentry);
2053 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2054 struct btrfs_root *root = BTRFS_I(inode)->root;
2055 struct btrfs_trans_handle *trans;
2056 struct btrfs_log_ctx ctx;
2058 bool full_sync = false;
2062 * The range length can be represented by u64, we have to do the typecasts
2063 * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
2065 len = (u64)end - (u64)start + 1;
2066 trace_btrfs_sync_file(file, datasync);
2068 btrfs_init_log_ctx(&ctx, inode);
2071 * We write the dirty pages in the range and wait until they complete
2072 * out of the ->i_mutex. If so, we can flush the dirty pages by
2073 * multi-task, and make the performance up. See
2074 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2076 ret = start_ordered_ops(inode, start, end);
2081 atomic_inc(&root->log_batch);
2082 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2083 &BTRFS_I(inode)->runtime_flags);
2085 * We might have have had more pages made dirty after calling
2086 * start_ordered_ops and before acquiring the inode's i_mutex.
2090 * For a full sync, we need to make sure any ordered operations
2091 * start and finish before we start logging the inode, so that
2092 * all extents are persisted and the respective file extent
2093 * items are in the fs/subvol btree.
2095 ret = btrfs_wait_ordered_range(inode, start, len);
2098 * Start any new ordered operations before starting to log the
2099 * inode. We will wait for them to finish in btrfs_sync_log().
2101 * Right before acquiring the inode's mutex, we might have new
2102 * writes dirtying pages, which won't immediately start the
2103 * respective ordered operations - that is done through the
2104 * fill_delalloc callbacks invoked from the writepage and
2105 * writepages address space operations. So make sure we start
2106 * all ordered operations before starting to log our inode. Not
2107 * doing this means that while logging the inode, writeback
2108 * could start and invoke writepage/writepages, which would call
2109 * the fill_delalloc callbacks (cow_file_range,
2110 * submit_compressed_extents). These callbacks add first an
2111 * extent map to the modified list of extents and then create
2112 * the respective ordered operation, which means in
2113 * tree-log.c:btrfs_log_inode() we might capture all existing
2114 * ordered operations (with btrfs_get_logged_extents()) before
2115 * the fill_delalloc callback adds its ordered operation, and by
2116 * the time we visit the modified list of extent maps (with
2117 * btrfs_log_changed_extents()), we see and process the extent
2118 * map they created. We then use the extent map to construct a
2119 * file extent item for logging without waiting for the
2120 * respective ordered operation to finish - this file extent
2121 * item points to a disk location that might not have yet been
2122 * written to, containing random data - so after a crash a log
2123 * replay will make our inode have file extent items that point
2124 * to disk locations containing invalid data, as we returned
2125 * success to userspace without waiting for the respective
2126 * ordered operation to finish, because it wasn't captured by
2127 * btrfs_get_logged_extents().
2129 ret = start_ordered_ops(inode, start, end);
2132 inode_unlock(inode);
2135 atomic_inc(&root->log_batch);
2138 * If the last transaction that changed this file was before the current
2139 * transaction and we have the full sync flag set in our inode, we can
2140 * bail out now without any syncing.
2142 * Note that we can't bail out if the full sync flag isn't set. This is
2143 * because when the full sync flag is set we start all ordered extents
2144 * and wait for them to fully complete - when they complete they update
2145 * the inode's last_trans field through:
2147 * btrfs_finish_ordered_io() ->
2148 * btrfs_update_inode_fallback() ->
2149 * btrfs_update_inode() ->
2150 * btrfs_set_inode_last_trans()
2152 * So we are sure that last_trans is up to date and can do this check to
2153 * bail out safely. For the fast path, when the full sync flag is not
2154 * set in our inode, we can not do it because we start only our ordered
2155 * extents and don't wait for them to complete (that is when
2156 * btrfs_finish_ordered_io runs), so here at this point their last_trans
2157 * value might be less than or equals to fs_info->last_trans_committed,
2158 * and setting a speculative last_trans for an inode when a buffered
2159 * write is made (such as fs_info->generation + 1 for example) would not
2160 * be reliable since after setting the value and before fsync is called
2161 * any number of transactions can start and commit (transaction kthread
2162 * commits the current transaction periodically), and a transaction
2163 * commit does not start nor waits for ordered extents to complete.
2166 if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2167 (full_sync && BTRFS_I(inode)->last_trans <=
2168 fs_info->last_trans_committed) ||
2169 (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2170 BTRFS_I(inode)->last_trans
2171 <= 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 inode_unlock(inode);
2192 * ok we haven't committed the transaction yet, lets do a commit
2194 if (file->private_data)
2195 btrfs_ioctl_trans_end(file);
2198 * We use start here because we will need to wait on the IO to complete
2199 * in btrfs_sync_log, which could require joining a transaction (for
2200 * example checking cross references in the nocow path). If we use join
2201 * here we could get into a situation where we're waiting on IO to
2202 * happen that is blocked on a transaction trying to commit. With start
2203 * we inc the extwriter counter, so we wait for all extwriters to exit
2204 * before we start blocking join'ers. This comment is to keep somebody
2205 * from thinking they are super smart and changing this to
2206 * btrfs_join_transaction *cough*Josef*cough*.
2208 trans = btrfs_start_transaction(root, 0);
2209 if (IS_ERR(trans)) {
2210 ret = PTR_ERR(trans);
2211 inode_unlock(inode);
2216 ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2218 /* Fallthrough and commit/free transaction. */
2222 /* we've logged all the items and now have a consistent
2223 * version of the file in the log. It is possible that
2224 * someone will come in and modify the file, but that's
2225 * fine because the log is consistent on disk, and we
2226 * have references to all of the file's extents
2228 * It is possible that someone will come in and log the
2229 * file again, but that will end up using the synchronization
2230 * inside btrfs_sync_log to keep things safe.
2232 inode_unlock(inode);
2235 * If any of the ordered extents had an error, just return it to user
2236 * space, so that the application knows some writes didn't succeed and
2237 * can take proper action (retry for e.g.). Blindly committing the
2238 * transaction in this case, would fool userspace that everything was
2239 * successful. And we also want to make sure our log doesn't contain
2240 * file extent items pointing to extents that weren't fully written to -
2241 * just like in the non fast fsync path, where we check for the ordered
2242 * operation's error flag before writing to the log tree and return -EIO
2243 * if any of them had this flag set (btrfs_wait_ordered_range) -
2244 * therefore we need to check for errors in the ordered operations,
2245 * which are indicated by ctx.io_err.
2248 btrfs_end_transaction(trans);
2253 if (ret != BTRFS_NO_LOG_SYNC) {
2255 ret = btrfs_sync_log(trans, root, &ctx);
2257 ret = btrfs_end_transaction(trans);
2262 ret = btrfs_wait_ordered_range(inode, start, len);
2264 btrfs_end_transaction(trans);
2268 ret = btrfs_commit_transaction(trans);
2270 ret = btrfs_end_transaction(trans);
2273 ASSERT(list_empty(&ctx.list));
2274 err = file_check_and_advance_wb_err(file);
2277 return ret > 0 ? -EIO : ret;
2280 static const struct vm_operations_struct btrfs_file_vm_ops = {
2281 .fault = filemap_fault,
2282 .map_pages = filemap_map_pages,
2283 .page_mkwrite = btrfs_page_mkwrite,
2286 static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2288 struct address_space *mapping = filp->f_mapping;
2290 if (!mapping->a_ops->readpage)
2293 file_accessed(filp);
2294 vma->vm_ops = &btrfs_file_vm_ops;
2299 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2300 int slot, u64 start, u64 end)
2302 struct btrfs_file_extent_item *fi;
2303 struct btrfs_key key;
2305 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2308 btrfs_item_key_to_cpu(leaf, &key, slot);
2309 if (key.objectid != btrfs_ino(inode) ||
2310 key.type != BTRFS_EXTENT_DATA_KEY)
2313 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2315 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2318 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2321 if (key.offset == end)
2323 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2328 static int fill_holes(struct btrfs_trans_handle *trans,
2329 struct btrfs_inode *inode,
2330 struct btrfs_path *path, u64 offset, u64 end)
2332 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
2333 struct btrfs_root *root = inode->root;
2334 struct extent_buffer *leaf;
2335 struct btrfs_file_extent_item *fi;
2336 struct extent_map *hole_em;
2337 struct extent_map_tree *em_tree = &inode->extent_tree;
2338 struct btrfs_key key;
2341 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2344 key.objectid = btrfs_ino(inode);
2345 key.type = BTRFS_EXTENT_DATA_KEY;
2346 key.offset = offset;
2348 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2351 * We should have dropped this offset, so if we find it then
2352 * something has gone horribly wrong.
2359 leaf = path->nodes[0];
2360 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2364 fi = btrfs_item_ptr(leaf, path->slots[0],
2365 struct btrfs_file_extent_item);
2366 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2368 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2369 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2370 btrfs_set_file_extent_offset(leaf, fi, 0);
2371 btrfs_mark_buffer_dirty(leaf);
2375 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2378 key.offset = offset;
2379 btrfs_set_item_key_safe(fs_info, path, &key);
2380 fi = btrfs_item_ptr(leaf, path->slots[0],
2381 struct btrfs_file_extent_item);
2382 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2384 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2385 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2386 btrfs_set_file_extent_offset(leaf, fi, 0);
2387 btrfs_mark_buffer_dirty(leaf);
2390 btrfs_release_path(path);
2392 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2393 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2398 btrfs_release_path(path);
2400 hole_em = alloc_extent_map();
2402 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2403 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2405 hole_em->start = offset;
2406 hole_em->len = end - offset;
2407 hole_em->ram_bytes = hole_em->len;
2408 hole_em->orig_start = offset;
2410 hole_em->block_start = EXTENT_MAP_HOLE;
2411 hole_em->block_len = 0;
2412 hole_em->orig_block_len = 0;
2413 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2414 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2415 hole_em->generation = trans->transid;
2418 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2419 write_lock(&em_tree->lock);
2420 ret = add_extent_mapping(em_tree, hole_em, 1);
2421 write_unlock(&em_tree->lock);
2422 } while (ret == -EEXIST);
2423 free_extent_map(hole_em);
2425 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2426 &inode->runtime_flags);
2433 * Find a hole extent on given inode and change start/len to the end of hole
2434 * extent.(hole/vacuum extent whose em->start <= start &&
2435 * em->start + em->len > start)
2436 * When a hole extent is found, return 1 and modify start/len.
2438 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2440 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2441 struct extent_map *em;
2444 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2445 round_down(*start, fs_info->sectorsize),
2446 round_up(*len, fs_info->sectorsize), 0);
2450 /* Hole or vacuum extent(only exists in no-hole mode) */
2451 if (em->block_start == EXTENT_MAP_HOLE) {
2453 *len = em->start + em->len > *start + *len ?
2454 0 : *start + *len - em->start - em->len;
2455 *start = em->start + em->len;
2457 free_extent_map(em);
2461 static int btrfs_punch_hole_lock_range(struct inode *inode,
2462 const u64 lockstart,
2464 struct extent_state **cached_state)
2467 struct btrfs_ordered_extent *ordered;
2470 truncate_pagecache_range(inode, lockstart, lockend);
2472 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2474 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2477 * We need to make sure we have no ordered extents in this range
2478 * and nobody raced in and read a page in this range, if we did
2479 * we need to try again.
2482 (ordered->file_offset + ordered->len <= lockstart ||
2483 ordered->file_offset > lockend)) &&
2484 !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2486 btrfs_put_ordered_extent(ordered);
2490 btrfs_put_ordered_extent(ordered);
2491 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2492 lockend, cached_state);
2493 ret = btrfs_wait_ordered_range(inode, lockstart,
2494 lockend - lockstart + 1);
2501 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2503 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2504 struct btrfs_root *root = BTRFS_I(inode)->root;
2505 struct extent_state *cached_state = NULL;
2506 struct btrfs_path *path;
2507 struct btrfs_block_rsv *rsv;
2508 struct btrfs_trans_handle *trans;
2513 u64 orig_start = offset;
2515 u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2519 unsigned int rsv_count;
2521 bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2523 bool truncated_block = false;
2524 bool updated_inode = false;
2526 ret = btrfs_wait_ordered_range(inode, offset, len);
2531 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2532 ret = find_first_non_hole(inode, &offset, &len);
2534 goto out_only_mutex;
2536 /* Already in a large hole */
2538 goto out_only_mutex;
2541 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2542 lockend = round_down(offset + len,
2543 btrfs_inode_sectorsize(inode)) - 1;
2544 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2545 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2547 * We needn't truncate any block which is beyond the end of the file
2548 * because we are sure there is no data there.
2551 * Only do this if we are in the same block and we aren't doing the
2554 if (same_block && len < fs_info->sectorsize) {
2555 if (offset < ino_size) {
2556 truncated_block = true;
2557 ret = btrfs_truncate_block(inode, offset, len, 0);
2561 goto out_only_mutex;
2564 /* zero back part of the first block */
2565 if (offset < ino_size) {
2566 truncated_block = true;
2567 ret = btrfs_truncate_block(inode, offset, 0, 0);
2569 inode_unlock(inode);
2574 /* Check the aligned pages after the first unaligned page,
2575 * if offset != orig_start, which means the first unaligned page
2576 * including several following pages are already in holes,
2577 * the extra check can be skipped */
2578 if (offset == orig_start) {
2579 /* after truncate page, check hole again */
2580 len = offset + len - lockstart;
2582 ret = find_first_non_hole(inode, &offset, &len);
2584 goto out_only_mutex;
2587 goto out_only_mutex;
2592 /* Check the tail unaligned part is in a hole */
2593 tail_start = lockend + 1;
2594 tail_len = offset + len - tail_start;
2596 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2597 if (unlikely(ret < 0))
2598 goto out_only_mutex;
2600 /* zero the front end of the last page */
2601 if (tail_start + tail_len < ino_size) {
2602 truncated_block = true;
2603 ret = btrfs_truncate_block(inode,
2604 tail_start + tail_len,
2607 goto out_only_mutex;
2612 if (lockend < lockstart) {
2614 goto out_only_mutex;
2617 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2620 inode_unlock(inode);
2621 goto out_only_mutex;
2624 path = btrfs_alloc_path();
2630 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2635 rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2639 * 1 - update the inode
2640 * 1 - removing the extents in the range
2641 * 1 - adding the hole extent if no_holes isn't set
2643 rsv_count = no_holes ? 2 : 3;
2644 trans = btrfs_start_transaction(root, rsv_count);
2645 if (IS_ERR(trans)) {
2646 err = PTR_ERR(trans);
2650 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2653 trans->block_rsv = rsv;
2655 cur_offset = lockstart;
2656 len = lockend - cur_offset;
2657 while (cur_offset < lockend) {
2658 ret = __btrfs_drop_extents(trans, root, inode, path,
2659 cur_offset, lockend + 1,
2660 &drop_end, 1, 0, 0, NULL);
2664 trans->block_rsv = &fs_info->trans_block_rsv;
2666 if (cur_offset < drop_end && cur_offset < ino_size) {
2667 ret = fill_holes(trans, BTRFS_I(inode), path,
2668 cur_offset, drop_end);
2671 * If we failed then we didn't insert our hole
2672 * entries for the area we dropped, so now the
2673 * fs is corrupted, so we must abort the
2676 btrfs_abort_transaction(trans, ret);
2682 cur_offset = drop_end;
2684 ret = btrfs_update_inode(trans, root, inode);
2690 btrfs_end_transaction(trans);
2691 btrfs_btree_balance_dirty(fs_info);
2693 trans = btrfs_start_transaction(root, rsv_count);
2694 if (IS_ERR(trans)) {
2695 ret = PTR_ERR(trans);
2700 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2702 BUG_ON(ret); /* shouldn't happen */
2703 trans->block_rsv = rsv;
2705 ret = find_first_non_hole(inode, &cur_offset, &len);
2706 if (unlikely(ret < 0))
2719 trans->block_rsv = &fs_info->trans_block_rsv;
2721 * If we are using the NO_HOLES feature we might have had already an
2722 * hole that overlaps a part of the region [lockstart, lockend] and
2723 * ends at (or beyond) lockend. Since we have no file extent items to
2724 * represent holes, drop_end can be less than lockend and so we must
2725 * make sure we have an extent map representing the existing hole (the
2726 * call to __btrfs_drop_extents() might have dropped the existing extent
2727 * map representing the existing hole), otherwise the fast fsync path
2728 * will not record the existence of the hole region
2729 * [existing_hole_start, lockend].
2731 if (drop_end <= lockend)
2732 drop_end = lockend + 1;
2734 * Don't insert file hole extent item if it's for a range beyond eof
2735 * (because it's useless) or if it represents a 0 bytes range (when
2736 * cur_offset == drop_end).
2738 if (cur_offset < ino_size && cur_offset < drop_end) {
2739 ret = fill_holes(trans, BTRFS_I(inode), path,
2740 cur_offset, drop_end);
2742 /* Same comment as above. */
2743 btrfs_abort_transaction(trans, ret);
2753 inode_inc_iversion(inode);
2754 inode->i_mtime = inode->i_ctime = current_time(inode);
2756 trans->block_rsv = &fs_info->trans_block_rsv;
2757 ret = btrfs_update_inode(trans, root, inode);
2758 updated_inode = true;
2759 btrfs_end_transaction(trans);
2760 btrfs_btree_balance_dirty(fs_info);
2762 btrfs_free_path(path);
2763 btrfs_free_block_rsv(fs_info, rsv);
2765 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2768 if (!updated_inode && truncated_block && !ret && !err) {
2770 * If we only end up zeroing part of a page, we still need to
2771 * update the inode item, so that all the time fields are
2772 * updated as well as the necessary btrfs inode in memory fields
2773 * for detecting, at fsync time, if the inode isn't yet in the
2774 * log tree or it's there but not up to date.
2776 trans = btrfs_start_transaction(root, 1);
2777 if (IS_ERR(trans)) {
2778 err = PTR_ERR(trans);
2780 err = btrfs_update_inode(trans, root, inode);
2781 ret = btrfs_end_transaction(trans);
2784 inode_unlock(inode);
2790 /* Helper structure to record which range is already reserved */
2791 struct falloc_range {
2792 struct list_head list;
2798 * Helper function to add falloc range
2800 * Caller should have locked the larger range of extent containing
2803 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2805 struct falloc_range *prev = NULL;
2806 struct falloc_range *range = NULL;
2808 if (list_empty(head))
2812 * As fallocate iterate by bytenr order, we only need to check
2815 prev = list_entry(head->prev, struct falloc_range, list);
2816 if (prev->start + prev->len == start) {
2821 range = kmalloc(sizeof(*range), GFP_KERNEL);
2824 range->start = start;
2826 list_add_tail(&range->list, head);
2830 static int btrfs_fallocate_update_isize(struct inode *inode,
2834 struct btrfs_trans_handle *trans;
2835 struct btrfs_root *root = BTRFS_I(inode)->root;
2839 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2842 trans = btrfs_start_transaction(root, 1);
2844 return PTR_ERR(trans);
2846 inode->i_ctime = current_time(inode);
2847 i_size_write(inode, end);
2848 btrfs_ordered_update_i_size(inode, end, NULL);
2849 ret = btrfs_update_inode(trans, root, inode);
2850 ret2 = btrfs_end_transaction(trans);
2852 return ret ? ret : ret2;
2856 RANGE_BOUNDARY_WRITTEN_EXTENT = 0,
2857 RANGE_BOUNDARY_PREALLOC_EXTENT = 1,
2858 RANGE_BOUNDARY_HOLE = 2,
2861 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2864 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2865 struct extent_map *em;
2868 offset = round_down(offset, sectorsize);
2869 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2873 if (em->block_start == EXTENT_MAP_HOLE)
2874 ret = RANGE_BOUNDARY_HOLE;
2875 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2876 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2878 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2880 free_extent_map(em);
2884 static int btrfs_zero_range(struct inode *inode,
2889 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2890 struct extent_map *em;
2891 struct extent_changeset *data_reserved = NULL;
2894 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2895 u64 alloc_start = round_down(offset, sectorsize);
2896 u64 alloc_end = round_up(offset + len, sectorsize);
2897 u64 bytes_to_reserve = 0;
2898 bool space_reserved = false;
2900 inode_dio_wait(inode);
2902 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2903 alloc_start, alloc_end - alloc_start, 0);
2910 * Avoid hole punching and extent allocation for some cases. More cases
2911 * could be considered, but these are unlikely common and we keep things
2912 * as simple as possible for now. Also, intentionally, if the target
2913 * range contains one or more prealloc extents together with regular
2914 * extents and holes, we drop all the existing extents and allocate a
2915 * new prealloc extent, so that we get a larger contiguous disk extent.
2917 if (em->start <= alloc_start &&
2918 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2919 const u64 em_end = em->start + em->len;
2921 if (em_end >= offset + len) {
2923 * The whole range is already a prealloc extent,
2924 * do nothing except updating the inode's i_size if
2927 free_extent_map(em);
2928 ret = btrfs_fallocate_update_isize(inode, offset + len,
2933 * Part of the range is already a prealloc extent, so operate
2934 * only on the remaining part of the range.
2936 alloc_start = em_end;
2937 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2938 len = offset + len - alloc_start;
2939 offset = alloc_start;
2940 alloc_hint = em->block_start + em->len;
2942 free_extent_map(em);
2944 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2945 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2946 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2947 alloc_start, sectorsize, 0);
2953 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2954 free_extent_map(em);
2955 ret = btrfs_fallocate_update_isize(inode, offset + len,
2959 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2960 free_extent_map(em);
2961 ret = btrfs_truncate_block(inode, offset, len, 0);
2963 ret = btrfs_fallocate_update_isize(inode,
2968 free_extent_map(em);
2969 alloc_start = round_down(offset, sectorsize);
2970 alloc_end = alloc_start + sectorsize;
2974 alloc_start = round_up(offset, sectorsize);
2975 alloc_end = round_down(offset + len, sectorsize);
2978 * For unaligned ranges, check the pages at the boundaries, they might
2979 * map to an extent, in which case we need to partially zero them, or
2980 * they might map to a hole, in which case we need our allocation range
2983 if (!IS_ALIGNED(offset, sectorsize)) {
2984 ret = btrfs_zero_range_check_range_boundary(inode, offset);
2987 if (ret == RANGE_BOUNDARY_HOLE) {
2988 alloc_start = round_down(offset, sectorsize);
2990 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2991 ret = btrfs_truncate_block(inode, offset, 0, 0);
2999 if (!IS_ALIGNED(offset + len, sectorsize)) {
3000 ret = btrfs_zero_range_check_range_boundary(inode,
3004 if (ret == RANGE_BOUNDARY_HOLE) {
3005 alloc_end = round_up(offset + len, sectorsize);
3007 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3008 ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3017 if (alloc_start < alloc_end) {
3018 struct extent_state *cached_state = NULL;
3019 const u64 lockstart = alloc_start;
3020 const u64 lockend = alloc_end - 1;
3022 bytes_to_reserve = alloc_end - alloc_start;
3023 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3027 space_reserved = true;
3028 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3029 alloc_start, bytes_to_reserve);
3032 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3036 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3037 alloc_end - alloc_start,
3039 offset + len, &alloc_hint);
3040 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3041 lockend, &cached_state);
3042 /* btrfs_prealloc_file_range releases reserved space on error */
3044 space_reserved = false;
3048 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3050 if (ret && space_reserved)
3051 btrfs_free_reserved_data_space(inode, data_reserved,
3052 alloc_start, bytes_to_reserve);
3053 extent_changeset_free(data_reserved);
3058 static long btrfs_fallocate(struct file *file, int mode,
3059 loff_t offset, loff_t len)
3061 struct inode *inode = file_inode(file);
3062 struct extent_state *cached_state = NULL;
3063 struct extent_changeset *data_reserved = NULL;
3064 struct falloc_range *range;
3065 struct falloc_range *tmp;
3066 struct list_head reserve_list;
3074 struct extent_map *em;
3075 int blocksize = btrfs_inode_sectorsize(inode);
3078 alloc_start = round_down(offset, blocksize);
3079 alloc_end = round_up(offset + len, blocksize);
3080 cur_offset = alloc_start;
3082 /* Make sure we aren't being give some crap mode */
3083 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3084 FALLOC_FL_ZERO_RANGE))
3087 if (mode & FALLOC_FL_PUNCH_HOLE)
3088 return btrfs_punch_hole(inode, offset, len);
3091 * Only trigger disk allocation, don't trigger qgroup reserve
3093 * For qgroup space, it will be checked later.
3095 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3096 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3097 alloc_end - alloc_start);
3104 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3105 ret = inode_newsize_ok(inode, offset + len);
3111 * TODO: Move these two operations after we have checked
3112 * accurate reserved space, or fallocate can still fail but
3113 * with page truncated or size expanded.
3115 * But that's a minor problem and won't do much harm BTW.
3117 if (alloc_start > inode->i_size) {
3118 ret = btrfs_cont_expand(inode, i_size_read(inode),
3122 } else if (offset + len > inode->i_size) {
3124 * If we are fallocating from the end of the file onward we
3125 * need to zero out the end of the block if i_size lands in the
3126 * middle of a block.
3128 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3134 * wait for ordered IO before we have any locks. We'll loop again
3135 * below with the locks held.
3137 ret = btrfs_wait_ordered_range(inode, alloc_start,
3138 alloc_end - alloc_start);
3142 if (mode & FALLOC_FL_ZERO_RANGE) {
3143 ret = btrfs_zero_range(inode, offset, len, mode);
3144 inode_unlock(inode);
3148 locked_end = alloc_end - 1;
3150 struct btrfs_ordered_extent *ordered;
3152 /* the extent lock is ordered inside the running
3155 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3156 locked_end, &cached_state);
3157 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3160 ordered->file_offset + ordered->len > alloc_start &&
3161 ordered->file_offset < alloc_end) {
3162 btrfs_put_ordered_extent(ordered);
3163 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3164 alloc_start, locked_end,
3167 * we can't wait on the range with the transaction
3168 * running or with the extent lock held
3170 ret = btrfs_wait_ordered_range(inode, alloc_start,
3171 alloc_end - alloc_start);
3176 btrfs_put_ordered_extent(ordered);
3181 /* First, check if we exceed the qgroup limit */
3182 INIT_LIST_HEAD(&reserve_list);
3183 while (cur_offset < alloc_end) {
3184 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3185 alloc_end - cur_offset, 0);
3190 last_byte = min(extent_map_end(em), alloc_end);
3191 actual_end = min_t(u64, extent_map_end(em), offset + len);
3192 last_byte = ALIGN(last_byte, blocksize);
3193 if (em->block_start == EXTENT_MAP_HOLE ||
3194 (cur_offset >= inode->i_size &&
3195 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3196 ret = add_falloc_range(&reserve_list, cur_offset,
3197 last_byte - cur_offset);
3199 free_extent_map(em);
3202 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3203 cur_offset, last_byte - cur_offset);
3205 free_extent_map(em);
3210 * Do not need to reserve unwritten extent for this
3211 * range, free reserved data space first, otherwise
3212 * it'll result in false ENOSPC error.
3214 btrfs_free_reserved_data_space(inode, data_reserved,
3215 cur_offset, last_byte - cur_offset);
3217 free_extent_map(em);
3218 cur_offset = last_byte;
3222 * If ret is still 0, means we're OK to fallocate.
3223 * Or just cleanup the list and exit.
3225 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3227 ret = btrfs_prealloc_file_range(inode, mode,
3229 range->len, i_blocksize(inode),
3230 offset + len, &alloc_hint);
3232 btrfs_free_reserved_data_space(inode,
3233 data_reserved, range->start,
3235 list_del(&range->list);
3242 * We didn't need to allocate any more space, but we still extended the
3243 * size of the file so we need to update i_size and the inode item.
3245 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3247 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3250 inode_unlock(inode);
3251 /* Let go of our reservation. */
3252 if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3253 btrfs_free_reserved_data_space(inode, data_reserved,
3254 alloc_start, alloc_end - cur_offset);
3255 extent_changeset_free(data_reserved);
3259 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3261 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3262 struct extent_map *em = NULL;
3263 struct extent_state *cached_state = NULL;
3270 if (inode->i_size == 0)
3274 * *offset can be negative, in this case we start finding DATA/HOLE from
3275 * the very start of the file.
3277 start = max_t(loff_t, 0, *offset);
3279 lockstart = round_down(start, fs_info->sectorsize);
3280 lockend = round_up(i_size_read(inode),
3281 fs_info->sectorsize);
3282 if (lockend <= lockstart)
3283 lockend = lockstart + fs_info->sectorsize;
3285 len = lockend - lockstart + 1;
3287 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3290 while (start < inode->i_size) {
3291 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0,
3299 if (whence == SEEK_HOLE &&
3300 (em->block_start == EXTENT_MAP_HOLE ||
3301 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3303 else if (whence == SEEK_DATA &&
3304 (em->block_start != EXTENT_MAP_HOLE &&
3305 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3308 start = em->start + em->len;
3309 free_extent_map(em);
3313 free_extent_map(em);
3315 if (whence == SEEK_DATA && start >= inode->i_size)
3318 *offset = min_t(loff_t, start, inode->i_size);
3320 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3325 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3327 struct inode *inode = file->f_mapping->host;
3334 offset = generic_file_llseek(file, offset, whence);
3338 if (offset >= i_size_read(inode)) {
3339 inode_unlock(inode);
3343 ret = find_desired_extent(inode, &offset, whence);
3345 inode_unlock(inode);
3350 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3352 inode_unlock(inode);
3356 static int btrfs_file_open(struct inode *inode, struct file *filp)
3358 filp->f_mode |= FMODE_NOWAIT;
3359 return generic_file_open(inode, filp);
3362 const struct file_operations btrfs_file_operations = {
3363 .llseek = btrfs_file_llseek,
3364 .read_iter = generic_file_read_iter,
3365 .splice_read = generic_file_splice_read,
3366 .write_iter = btrfs_file_write_iter,
3367 .mmap = btrfs_file_mmap,
3368 .open = btrfs_file_open,
3369 .release = btrfs_release_file,
3370 .fsync = btrfs_sync_file,
3371 .fallocate = btrfs_fallocate,
3372 .unlocked_ioctl = btrfs_ioctl,
3373 #ifdef CONFIG_COMPAT
3374 .compat_ioctl = btrfs_compat_ioctl,
3376 .clone_file_range = btrfs_clone_file_range,
3377 .dedupe_file_range = btrfs_dedupe_file_range,
3380 void btrfs_auto_defrag_exit(void)
3382 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3385 int __init btrfs_auto_defrag_init(void)
3387 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3388 sizeof(struct inode_defrag), 0,
3391 if (!btrfs_inode_defrag_cachep)
3397 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3402 * So with compression we will find and lock a dirty page and clear the
3403 * first one as dirty, setup an async extent, and immediately return
3404 * with the entire range locked but with nobody actually marked with
3405 * writeback. So we can't just filemap_write_and_wait_range() and
3406 * expect it to work since it will just kick off a thread to do the
3407 * actual work. So we need to call filemap_fdatawrite_range _again_
3408 * since it will wait on the page lock, which won't be unlocked until
3409 * after the pages have been marked as writeback and so we're good to go
3410 * from there. We have to do this otherwise we'll miss the ordered
3411 * extents and that results in badness. Please Josef, do not think you
3412 * know better and pull this out at some point in the future, it is
3413 * right and you are wrong.
3415 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3416 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3417 &BTRFS_I(inode)->runtime_flags))
3418 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);