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]);
481 * after copy_from_user, pages need to be dirtied and we need to make
482 * sure holes are created between the current EOF and the start of
483 * any next extents (if required).
485 * this also makes the decision about creating an inline extent vs
486 * doing real data extents, marking pages dirty and delalloc as required.
488 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
489 size_t num_pages, loff_t pos, size_t write_bytes,
490 struct extent_state **cached)
492 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
497 u64 end_of_last_block;
498 u64 end_pos = pos + write_bytes;
499 loff_t isize = i_size_read(inode);
501 start_pos = pos & ~((u64) fs_info->sectorsize - 1);
502 num_bytes = round_up(write_bytes + pos - start_pos,
503 fs_info->sectorsize);
505 end_of_last_block = start_pos + num_bytes - 1;
506 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
511 for (i = 0; i < num_pages; i++) {
512 struct page *p = pages[i];
519 * we've only changed i_size in ram, and we haven't updated
520 * the disk i_size. There is no need to log the inode
524 i_size_write(inode, end_pos);
529 * this drops all the extents in the cache that intersect the range
530 * [start, end]. Existing extents are split as required.
532 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
535 struct extent_map *em;
536 struct extent_map *split = NULL;
537 struct extent_map *split2 = NULL;
538 struct extent_map_tree *em_tree = &inode->extent_tree;
539 u64 len = end - start + 1;
547 WARN_ON(end < start);
548 if (end == (u64)-1) {
557 split = alloc_extent_map();
559 split2 = alloc_extent_map();
560 if (!split || !split2)
563 write_lock(&em_tree->lock);
564 em = lookup_extent_mapping(em_tree, start, len);
566 write_unlock(&em_tree->lock);
570 gen = em->generation;
571 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
572 if (testend && em->start + em->len >= start + len) {
574 write_unlock(&em_tree->lock);
577 start = em->start + em->len;
579 len = start + len - (em->start + em->len);
581 write_unlock(&em_tree->lock);
584 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
585 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
586 clear_bit(EXTENT_FLAG_LOGGING, &flags);
587 modified = !list_empty(&em->list);
591 if (em->start < start) {
592 split->start = em->start;
593 split->len = start - em->start;
595 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
596 split->orig_start = em->orig_start;
597 split->block_start = em->block_start;
600 split->block_len = em->block_len;
602 split->block_len = split->len;
603 split->orig_block_len = max(split->block_len,
605 split->ram_bytes = em->ram_bytes;
607 split->orig_start = split->start;
608 split->block_len = 0;
609 split->block_start = em->block_start;
610 split->orig_block_len = 0;
611 split->ram_bytes = split->len;
614 split->generation = gen;
615 split->bdev = em->bdev;
616 split->flags = flags;
617 split->compress_type = em->compress_type;
618 replace_extent_mapping(em_tree, em, split, modified);
619 free_extent_map(split);
623 if (testend && em->start + em->len > start + len) {
624 u64 diff = start + len - em->start;
626 split->start = start + len;
627 split->len = em->start + em->len - (start + len);
628 split->bdev = em->bdev;
629 split->flags = flags;
630 split->compress_type = em->compress_type;
631 split->generation = gen;
633 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
634 split->orig_block_len = max(em->block_len,
637 split->ram_bytes = em->ram_bytes;
639 split->block_len = em->block_len;
640 split->block_start = em->block_start;
641 split->orig_start = em->orig_start;
643 split->block_len = split->len;
644 split->block_start = em->block_start
646 split->orig_start = em->orig_start;
649 split->ram_bytes = split->len;
650 split->orig_start = split->start;
651 split->block_len = 0;
652 split->block_start = em->block_start;
653 split->orig_block_len = 0;
656 if (extent_map_in_tree(em)) {
657 replace_extent_mapping(em_tree, em, split,
660 ret = add_extent_mapping(em_tree, split,
662 ASSERT(ret == 0); /* Logic error */
664 free_extent_map(split);
668 if (extent_map_in_tree(em))
669 remove_extent_mapping(em_tree, em);
670 write_unlock(&em_tree->lock);
674 /* once for the tree*/
678 free_extent_map(split);
680 free_extent_map(split2);
684 * this is very complex, but the basic idea is to drop all extents
685 * in the range start - end. hint_block is filled in with a block number
686 * that would be a good hint to the block allocator for this file.
688 * If an extent intersects the range but is not entirely inside the range
689 * it is either truncated or split. Anything entirely inside the range
690 * is deleted from the tree.
692 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
693 struct btrfs_root *root, struct inode *inode,
694 struct btrfs_path *path, u64 start, u64 end,
695 u64 *drop_end, int drop_cache,
697 u32 extent_item_size,
700 struct btrfs_fs_info *fs_info = root->fs_info;
701 struct extent_buffer *leaf;
702 struct btrfs_file_extent_item *fi;
703 struct btrfs_key key;
704 struct btrfs_key new_key;
705 u64 ino = btrfs_ino(BTRFS_I(inode));
706 u64 search_start = start;
709 u64 extent_offset = 0;
711 u64 last_end = start;
717 int modify_tree = -1;
720 int leafs_visited = 0;
723 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
725 if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
728 update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
729 root == fs_info->tree_root);
732 ret = btrfs_lookup_file_extent(trans, root, path, ino,
733 search_start, modify_tree);
736 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
737 leaf = path->nodes[0];
738 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
739 if (key.objectid == ino &&
740 key.type == BTRFS_EXTENT_DATA_KEY)
746 leaf = path->nodes[0];
747 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
749 ret = btrfs_next_leaf(root, path);
757 leaf = path->nodes[0];
761 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
763 if (key.objectid > ino)
765 if (WARN_ON_ONCE(key.objectid < ino) ||
766 key.type < BTRFS_EXTENT_DATA_KEY) {
771 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
774 fi = btrfs_item_ptr(leaf, path->slots[0],
775 struct btrfs_file_extent_item);
776 extent_type = btrfs_file_extent_type(leaf, fi);
778 if (extent_type == BTRFS_FILE_EXTENT_REG ||
779 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
780 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
781 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
782 extent_offset = btrfs_file_extent_offset(leaf, fi);
783 extent_end = key.offset +
784 btrfs_file_extent_num_bytes(leaf, fi);
785 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
786 extent_end = key.offset +
787 btrfs_file_extent_ram_bytes(leaf, fi);
794 * Don't skip extent items representing 0 byte lengths. They
795 * used to be created (bug) if while punching holes we hit
796 * -ENOSPC condition. So if we find one here, just ensure we
797 * delete it, otherwise we would insert a new file extent item
798 * with the same key (offset) as that 0 bytes length file
799 * extent item in the call to setup_items_for_insert() later
802 if (extent_end == key.offset && extent_end >= search_start) {
803 last_end = extent_end;
804 goto delete_extent_item;
807 if (extent_end <= search_start) {
813 search_start = max(key.offset, start);
814 if (recow || !modify_tree) {
816 btrfs_release_path(path);
821 * | - range to drop - |
822 * | -------- extent -------- |
824 if (start > key.offset && end < extent_end) {
826 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
831 memcpy(&new_key, &key, sizeof(new_key));
832 new_key.offset = start;
833 ret = btrfs_duplicate_item(trans, root, path,
835 if (ret == -EAGAIN) {
836 btrfs_release_path(path);
842 leaf = path->nodes[0];
843 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
844 struct btrfs_file_extent_item);
845 btrfs_set_file_extent_num_bytes(leaf, fi,
848 fi = btrfs_item_ptr(leaf, path->slots[0],
849 struct btrfs_file_extent_item);
851 extent_offset += start - key.offset;
852 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
853 btrfs_set_file_extent_num_bytes(leaf, fi,
855 btrfs_mark_buffer_dirty(leaf);
857 if (update_refs && disk_bytenr > 0) {
858 ret = btrfs_inc_extent_ref(trans, fs_info,
859 disk_bytenr, num_bytes, 0,
860 root->root_key.objectid,
862 start - extent_offset);
863 BUG_ON(ret); /* -ENOMEM */
868 * From here on out we will have actually dropped something, so
869 * last_end can be updated.
871 last_end = extent_end;
874 * | ---- range to drop ----- |
875 * | -------- extent -------- |
877 if (start <= key.offset && end < extent_end) {
878 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
883 memcpy(&new_key, &key, sizeof(new_key));
884 new_key.offset = end;
885 btrfs_set_item_key_safe(fs_info, path, &new_key);
887 extent_offset += end - key.offset;
888 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
889 btrfs_set_file_extent_num_bytes(leaf, fi,
891 btrfs_mark_buffer_dirty(leaf);
892 if (update_refs && disk_bytenr > 0)
893 inode_sub_bytes(inode, end - key.offset);
897 search_start = extent_end;
899 * | ---- range to drop ----- |
900 * | -------- extent -------- |
902 if (start > key.offset && end >= extent_end) {
904 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
909 btrfs_set_file_extent_num_bytes(leaf, fi,
911 btrfs_mark_buffer_dirty(leaf);
912 if (update_refs && disk_bytenr > 0)
913 inode_sub_bytes(inode, extent_end - start);
914 if (end == extent_end)
922 * | ---- range to drop ----- |
923 * | ------ extent ------ |
925 if (start <= key.offset && end >= extent_end) {
928 del_slot = path->slots[0];
931 BUG_ON(del_slot + del_nr != path->slots[0]);
936 extent_type == BTRFS_FILE_EXTENT_INLINE) {
937 inode_sub_bytes(inode,
938 extent_end - key.offset);
939 extent_end = ALIGN(extent_end,
940 fs_info->sectorsize);
941 } else if (update_refs && disk_bytenr > 0) {
942 ret = btrfs_free_extent(trans, fs_info,
943 disk_bytenr, num_bytes, 0,
944 root->root_key.objectid,
945 key.objectid, key.offset -
947 BUG_ON(ret); /* -ENOMEM */
948 inode_sub_bytes(inode,
949 extent_end - key.offset);
952 if (end == extent_end)
955 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
960 ret = btrfs_del_items(trans, root, path, del_slot,
963 btrfs_abort_transaction(trans, ret);
970 btrfs_release_path(path);
977 if (!ret && del_nr > 0) {
979 * Set path->slots[0] to first slot, so that after the delete
980 * if items are move off from our leaf to its immediate left or
981 * right neighbor leafs, we end up with a correct and adjusted
982 * path->slots[0] for our insertion (if replace_extent != 0).
984 path->slots[0] = del_slot;
985 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
987 btrfs_abort_transaction(trans, ret);
990 leaf = path->nodes[0];
992 * If btrfs_del_items() was called, it might have deleted a leaf, in
993 * which case it unlocked our path, so check path->locks[0] matches a
996 if (!ret && replace_extent && leafs_visited == 1 &&
997 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
998 path->locks[0] == BTRFS_WRITE_LOCK) &&
999 btrfs_leaf_free_space(fs_info, leaf) >=
1000 sizeof(struct btrfs_item) + extent_item_size) {
1003 key.type = BTRFS_EXTENT_DATA_KEY;
1005 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1006 struct btrfs_key slot_key;
1008 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1009 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1012 setup_items_for_insert(root, path, &key,
1015 sizeof(struct btrfs_item) +
1016 extent_item_size, 1);
1020 if (!replace_extent || !(*key_inserted))
1021 btrfs_release_path(path);
1023 *drop_end = found ? min(end, last_end) : end;
1027 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1028 struct btrfs_root *root, struct inode *inode, u64 start,
1029 u64 end, int drop_cache)
1031 struct btrfs_path *path;
1034 path = btrfs_alloc_path();
1037 ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1038 drop_cache, 0, 0, NULL);
1039 btrfs_free_path(path);
1043 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1044 u64 objectid, u64 bytenr, u64 orig_offset,
1045 u64 *start, u64 *end)
1047 struct btrfs_file_extent_item *fi;
1048 struct btrfs_key key;
1051 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1054 btrfs_item_key_to_cpu(leaf, &key, slot);
1055 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1058 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1059 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1060 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1061 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1062 btrfs_file_extent_compression(leaf, fi) ||
1063 btrfs_file_extent_encryption(leaf, fi) ||
1064 btrfs_file_extent_other_encoding(leaf, fi))
1067 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1068 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1071 *start = key.offset;
1077 * Mark extent in the range start - end as written.
1079 * This changes extent type from 'pre-allocated' to 'regular'. If only
1080 * part of extent is marked as written, the extent will be split into
1083 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1084 struct btrfs_inode *inode, u64 start, u64 end)
1086 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1087 struct btrfs_root *root = inode->root;
1088 struct extent_buffer *leaf;
1089 struct btrfs_path *path;
1090 struct btrfs_file_extent_item *fi;
1091 struct btrfs_key key;
1092 struct btrfs_key new_key;
1104 u64 ino = btrfs_ino(inode);
1106 path = btrfs_alloc_path();
1113 key.type = BTRFS_EXTENT_DATA_KEY;
1116 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1119 if (ret > 0 && path->slots[0] > 0)
1122 leaf = path->nodes[0];
1123 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1124 if (key.objectid != ino ||
1125 key.type != BTRFS_EXTENT_DATA_KEY) {
1127 btrfs_abort_transaction(trans, ret);
1130 fi = btrfs_item_ptr(leaf, path->slots[0],
1131 struct btrfs_file_extent_item);
1132 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1134 btrfs_abort_transaction(trans, ret);
1137 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1138 if (key.offset > start || extent_end < end) {
1140 btrfs_abort_transaction(trans, ret);
1144 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1145 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1146 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1147 memcpy(&new_key, &key, sizeof(new_key));
1149 if (start == key.offset && end < extent_end) {
1152 if (extent_mergeable(leaf, path->slots[0] - 1,
1153 ino, bytenr, orig_offset,
1154 &other_start, &other_end)) {
1155 new_key.offset = end;
1156 btrfs_set_item_key_safe(fs_info, path, &new_key);
1157 fi = btrfs_item_ptr(leaf, path->slots[0],
1158 struct btrfs_file_extent_item);
1159 btrfs_set_file_extent_generation(leaf, fi,
1161 btrfs_set_file_extent_num_bytes(leaf, fi,
1163 btrfs_set_file_extent_offset(leaf, fi,
1165 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1166 struct btrfs_file_extent_item);
1167 btrfs_set_file_extent_generation(leaf, fi,
1169 btrfs_set_file_extent_num_bytes(leaf, fi,
1171 btrfs_mark_buffer_dirty(leaf);
1176 if (start > key.offset && end == extent_end) {
1179 if (extent_mergeable(leaf, path->slots[0] + 1,
1180 ino, bytenr, orig_offset,
1181 &other_start, &other_end)) {
1182 fi = btrfs_item_ptr(leaf, path->slots[0],
1183 struct btrfs_file_extent_item);
1184 btrfs_set_file_extent_num_bytes(leaf, fi,
1185 start - key.offset);
1186 btrfs_set_file_extent_generation(leaf, fi,
1189 new_key.offset = start;
1190 btrfs_set_item_key_safe(fs_info, path, &new_key);
1192 fi = btrfs_item_ptr(leaf, path->slots[0],
1193 struct btrfs_file_extent_item);
1194 btrfs_set_file_extent_generation(leaf, fi,
1196 btrfs_set_file_extent_num_bytes(leaf, fi,
1198 btrfs_set_file_extent_offset(leaf, fi,
1199 start - orig_offset);
1200 btrfs_mark_buffer_dirty(leaf);
1205 while (start > key.offset || end < extent_end) {
1206 if (key.offset == start)
1209 new_key.offset = split;
1210 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1211 if (ret == -EAGAIN) {
1212 btrfs_release_path(path);
1216 btrfs_abort_transaction(trans, ret);
1220 leaf = path->nodes[0];
1221 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1222 struct btrfs_file_extent_item);
1223 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1224 btrfs_set_file_extent_num_bytes(leaf, fi,
1225 split - key.offset);
1227 fi = btrfs_item_ptr(leaf, path->slots[0],
1228 struct btrfs_file_extent_item);
1230 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1231 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1232 btrfs_set_file_extent_num_bytes(leaf, fi,
1233 extent_end - split);
1234 btrfs_mark_buffer_dirty(leaf);
1236 ret = btrfs_inc_extent_ref(trans, fs_info, bytenr, num_bytes,
1237 0, root->root_key.objectid,
1240 btrfs_abort_transaction(trans, ret);
1244 if (split == start) {
1247 if (start != key.offset) {
1249 btrfs_abort_transaction(trans, ret);
1260 if (extent_mergeable(leaf, path->slots[0] + 1,
1261 ino, bytenr, orig_offset,
1262 &other_start, &other_end)) {
1264 btrfs_release_path(path);
1267 extent_end = other_end;
1268 del_slot = path->slots[0] + 1;
1270 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1271 0, root->root_key.objectid,
1274 btrfs_abort_transaction(trans, ret);
1280 if (extent_mergeable(leaf, path->slots[0] - 1,
1281 ino, bytenr, orig_offset,
1282 &other_start, &other_end)) {
1284 btrfs_release_path(path);
1287 key.offset = other_start;
1288 del_slot = path->slots[0];
1290 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1291 0, root->root_key.objectid,
1294 btrfs_abort_transaction(trans, ret);
1299 fi = btrfs_item_ptr(leaf, path->slots[0],
1300 struct btrfs_file_extent_item);
1301 btrfs_set_file_extent_type(leaf, fi,
1302 BTRFS_FILE_EXTENT_REG);
1303 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1304 btrfs_mark_buffer_dirty(leaf);
1306 fi = btrfs_item_ptr(leaf, del_slot - 1,
1307 struct btrfs_file_extent_item);
1308 btrfs_set_file_extent_type(leaf, fi,
1309 BTRFS_FILE_EXTENT_REG);
1310 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1311 btrfs_set_file_extent_num_bytes(leaf, fi,
1312 extent_end - key.offset);
1313 btrfs_mark_buffer_dirty(leaf);
1315 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1317 btrfs_abort_transaction(trans, ret);
1322 btrfs_free_path(path);
1327 * on error we return an unlocked page and the error value
1328 * on success we return a locked page and 0
1330 static int prepare_uptodate_page(struct inode *inode,
1331 struct page *page, u64 pos,
1332 bool force_uptodate)
1336 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1337 !PageUptodate(page)) {
1338 ret = btrfs_readpage(NULL, page);
1342 if (!PageUptodate(page)) {
1346 if (page->mapping != inode->i_mapping) {
1355 * this just gets pages into the page cache and locks them down.
1357 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1358 size_t num_pages, loff_t pos,
1359 size_t write_bytes, bool force_uptodate)
1362 unsigned long index = pos >> PAGE_SHIFT;
1363 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1367 for (i = 0; i < num_pages; i++) {
1369 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1370 mask | __GFP_WRITE);
1378 err = prepare_uptodate_page(inode, pages[i], pos,
1380 if (!err && i == num_pages - 1)
1381 err = prepare_uptodate_page(inode, pages[i],
1382 pos + write_bytes, false);
1385 if (err == -EAGAIN) {
1392 wait_on_page_writeback(pages[i]);
1397 while (faili >= 0) {
1398 unlock_page(pages[faili]);
1399 put_page(pages[faili]);
1406 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
1409 struct extent_state **cached_state)
1411 u64 search_start = start;
1412 const u64 end = start + len - 1;
1414 while (search_start < end) {
1415 const u64 search_len = end - search_start + 1;
1416 struct extent_map *em;
1420 em = btrfs_get_extent(inode, NULL, 0, search_start,
1425 if (em->block_start != EXTENT_MAP_HOLE)
1429 if (em->start < search_start)
1430 em_len -= search_start - em->start;
1431 if (em_len > search_len)
1432 em_len = search_len;
1434 ret = set_extent_bit(&inode->io_tree, search_start,
1435 search_start + em_len - 1,
1436 EXTENT_DELALLOC_NEW,
1437 NULL, cached_state, GFP_NOFS);
1439 search_start = extent_map_end(em);
1440 free_extent_map(em);
1448 * This function locks the extent and properly waits for data=ordered extents
1449 * to finish before allowing the pages to be modified if need.
1452 * 1 - the extent is locked
1453 * 0 - the extent is not locked, and everything is OK
1454 * -EAGAIN - need re-prepare the pages
1455 * the other < 0 number - Something wrong happens
1458 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1459 size_t num_pages, loff_t pos,
1461 u64 *lockstart, u64 *lockend,
1462 struct extent_state **cached_state)
1464 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1470 start_pos = round_down(pos, fs_info->sectorsize);
1471 last_pos = start_pos
1472 + round_up(pos + write_bytes - start_pos,
1473 fs_info->sectorsize) - 1;
1475 if (start_pos < inode->vfs_inode.i_size ||
1476 (inode->flags & BTRFS_INODE_PREALLOC)) {
1477 struct btrfs_ordered_extent *ordered;
1478 unsigned int clear_bits;
1480 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1482 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1483 last_pos - start_pos + 1);
1485 ordered->file_offset + ordered->len > start_pos &&
1486 ordered->file_offset <= last_pos) {
1487 unlock_extent_cached(&inode->io_tree, start_pos,
1488 last_pos, cached_state, GFP_NOFS);
1489 for (i = 0; i < num_pages; i++) {
1490 unlock_page(pages[i]);
1493 btrfs_start_ordered_extent(&inode->vfs_inode,
1495 btrfs_put_ordered_extent(ordered);
1499 btrfs_put_ordered_extent(ordered);
1500 ret = btrfs_find_new_delalloc_bytes(inode, start_pos,
1501 last_pos - start_pos + 1,
1503 clear_bits = EXTENT_DIRTY | EXTENT_DELALLOC |
1504 EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG;
1506 clear_bits |= EXTENT_DELALLOC_NEW | EXTENT_LOCKED;
1507 clear_extent_bit(&inode->io_tree, start_pos,
1508 last_pos, clear_bits,
1509 (clear_bits & EXTENT_LOCKED) ? 1 : 0,
1510 0, cached_state, GFP_NOFS);
1513 *lockstart = start_pos;
1514 *lockend = last_pos;
1518 for (i = 0; i < num_pages; i++) {
1519 if (clear_page_dirty_for_io(pages[i]))
1520 account_page_redirty(pages[i]);
1521 set_page_extent_mapped(pages[i]);
1522 WARN_ON(!PageLocked(pages[i]));
1528 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1529 size_t *write_bytes)
1531 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1532 struct btrfs_root *root = inode->root;
1533 struct btrfs_ordered_extent *ordered;
1534 u64 lockstart, lockend;
1538 ret = btrfs_start_write_no_snapshotting(root);
1542 lockstart = round_down(pos, fs_info->sectorsize);
1543 lockend = round_up(pos + *write_bytes,
1544 fs_info->sectorsize) - 1;
1547 lock_extent(&inode->io_tree, lockstart, lockend);
1548 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1549 lockend - lockstart + 1);
1553 unlock_extent(&inode->io_tree, lockstart, lockend);
1554 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1555 btrfs_put_ordered_extent(ordered);
1558 num_bytes = lockend - lockstart + 1;
1559 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1563 btrfs_end_write_no_snapshotting(root);
1565 *write_bytes = min_t(size_t, *write_bytes ,
1566 num_bytes - pos + lockstart);
1569 unlock_extent(&inode->io_tree, lockstart, lockend);
1574 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1578 struct inode *inode = file_inode(file);
1579 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1580 struct btrfs_root *root = BTRFS_I(inode)->root;
1581 struct page **pages = NULL;
1582 struct extent_state *cached_state = NULL;
1583 struct extent_changeset *data_reserved = NULL;
1584 u64 release_bytes = 0;
1587 size_t num_written = 0;
1590 bool only_release_metadata = false;
1591 bool force_page_uptodate = false;
1594 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1595 PAGE_SIZE / (sizeof(struct page *)));
1596 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1597 nrptrs = max(nrptrs, 8);
1598 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1602 while (iov_iter_count(i) > 0) {
1603 size_t offset = pos & (PAGE_SIZE - 1);
1604 size_t sector_offset;
1605 size_t write_bytes = min(iov_iter_count(i),
1606 nrptrs * (size_t)PAGE_SIZE -
1608 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1610 size_t reserve_bytes;
1613 size_t dirty_sectors;
1616 WARN_ON(num_pages > nrptrs);
1619 * Fault pages before locking them in prepare_pages
1620 * to avoid recursive lock
1622 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1627 only_release_metadata = false;
1628 sector_offset = pos & (fs_info->sectorsize - 1);
1629 reserve_bytes = round_up(write_bytes + sector_offset,
1630 fs_info->sectorsize);
1632 extent_changeset_release(data_reserved);
1633 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1636 if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1637 BTRFS_INODE_PREALLOC)) &&
1638 check_can_nocow(BTRFS_I(inode), pos,
1639 &write_bytes) > 0) {
1641 * For nodata cow case, no need to reserve
1644 only_release_metadata = true;
1646 * our prealloc extent may be smaller than
1647 * write_bytes, so scale down.
1649 num_pages = DIV_ROUND_UP(write_bytes + offset,
1651 reserve_bytes = round_up(write_bytes +
1653 fs_info->sectorsize);
1659 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1662 if (!only_release_metadata)
1663 btrfs_free_reserved_data_space(inode,
1667 btrfs_end_write_no_snapshotting(root);
1671 release_bytes = reserve_bytes;
1672 need_unlock = false;
1675 * This is going to setup the pages array with the number of
1676 * pages we want, so we don't really need to worry about the
1677 * contents of pages from loop to loop
1679 ret = prepare_pages(inode, pages, num_pages,
1681 force_page_uptodate);
1685 ret = lock_and_cleanup_extent_if_need(BTRFS_I(inode), pages,
1686 num_pages, pos, write_bytes, &lockstart,
1687 &lockend, &cached_state);
1692 } else if (ret > 0) {
1697 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1699 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1700 dirty_sectors = round_up(copied + sector_offset,
1701 fs_info->sectorsize);
1702 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1705 * if we have trouble faulting in the pages, fall
1706 * back to one page at a time
1708 if (copied < write_bytes)
1712 force_page_uptodate = true;
1716 force_page_uptodate = false;
1717 dirty_pages = DIV_ROUND_UP(copied + offset,
1722 * If we had a short copy we need to release the excess delaloc
1723 * bytes we reserved. We need to increment outstanding_extents
1724 * because btrfs_delalloc_release_space and
1725 * btrfs_delalloc_release_metadata will decrement it, but
1726 * we still have an outstanding extent for the chunk we actually
1729 if (num_sectors > dirty_sectors) {
1730 /* release everything except the sectors we dirtied */
1731 release_bytes -= dirty_sectors <<
1732 fs_info->sb->s_blocksize_bits;
1734 spin_lock(&BTRFS_I(inode)->lock);
1735 BTRFS_I(inode)->outstanding_extents++;
1736 spin_unlock(&BTRFS_I(inode)->lock);
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,
1760 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1761 lockstart, lockend, &cached_state,
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,
1783 btrfs_drop_pages(pages, num_pages);
1787 balance_dirty_pages_ratelimited(inode->i_mapping);
1788 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1789 btrfs_btree_balance_dirty(fs_info);
1792 num_written += copied;
1797 if (release_bytes) {
1798 if (only_release_metadata) {
1799 btrfs_end_write_no_snapshotting(root);
1800 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1803 btrfs_delalloc_release_space(inode, data_reserved,
1804 round_down(pos, fs_info->sectorsize),
1809 extent_changeset_free(data_reserved);
1810 return num_written ? num_written : ret;
1813 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1815 struct file *file = iocb->ki_filp;
1816 struct inode *inode = file_inode(file);
1817 loff_t pos = iocb->ki_pos;
1819 ssize_t written_buffered;
1823 written = generic_file_direct_write(iocb, from);
1825 if (written < 0 || !iov_iter_count(from))
1829 written_buffered = __btrfs_buffered_write(file, from, pos);
1830 if (written_buffered < 0) {
1831 err = written_buffered;
1835 * Ensure all data is persisted. We want the next direct IO read to be
1836 * able to read what was just written.
1838 endbyte = pos + written_buffered - 1;
1839 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1842 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1845 written += written_buffered;
1846 iocb->ki_pos = pos + written_buffered;
1847 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1848 endbyte >> PAGE_SHIFT);
1850 return written ? written : err;
1853 static void update_time_for_write(struct inode *inode)
1855 struct timespec now;
1857 if (IS_NOCMTIME(inode))
1860 now = current_time(inode);
1861 if (!timespec_equal(&inode->i_mtime, &now))
1862 inode->i_mtime = now;
1864 if (!timespec_equal(&inode->i_ctime, &now))
1865 inode->i_ctime = now;
1867 if (IS_I_VERSION(inode))
1868 inode_inc_iversion(inode);
1871 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1872 struct iov_iter *from)
1874 struct file *file = iocb->ki_filp;
1875 struct inode *inode = file_inode(file);
1876 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1877 struct btrfs_root *root = BTRFS_I(inode)->root;
1880 ssize_t num_written = 0;
1881 bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1888 if (!(iocb->ki_flags & IOCB_DIRECT) &&
1889 (iocb->ki_flags & IOCB_NOWAIT))
1892 if (iocb->ki_flags & IOCB_NOWAIT) {
1893 if (!inode_trylock(inode))
1899 err = generic_write_checks(iocb, from);
1901 inode_unlock(inode);
1906 count = iov_iter_count(from);
1907 if (iocb->ki_flags & IOCB_NOWAIT) {
1909 * We will allocate space in case nodatacow is not set,
1912 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1913 BTRFS_INODE_PREALLOC)) ||
1914 check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1915 inode_unlock(inode);
1920 current->backing_dev_info = inode_to_bdi(inode);
1921 err = file_remove_privs(file);
1923 inode_unlock(inode);
1928 * If BTRFS flips readonly due to some impossible error
1929 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1930 * although we have opened a file as writable, we have
1931 * to stop this write operation to ensure FS consistency.
1933 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1934 inode_unlock(inode);
1940 * We reserve space for updating the inode when we reserve space for the
1941 * extent we are going to write, so we will enospc out there. We don't
1942 * need to start yet another transaction to update the inode as we will
1943 * update the inode when we finish writing whatever data we write.
1945 update_time_for_write(inode);
1947 start_pos = round_down(pos, fs_info->sectorsize);
1948 oldsize = i_size_read(inode);
1949 if (start_pos > oldsize) {
1950 /* Expand hole size to cover write data, preventing empty gap */
1951 end_pos = round_up(pos + count,
1952 fs_info->sectorsize);
1953 err = btrfs_cont_expand(inode, oldsize, end_pos);
1955 inode_unlock(inode);
1958 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1963 atomic_inc(&BTRFS_I(inode)->sync_writers);
1965 if (iocb->ki_flags & IOCB_DIRECT) {
1966 num_written = __btrfs_direct_write(iocb, from);
1968 num_written = __btrfs_buffered_write(file, from, pos);
1969 if (num_written > 0)
1970 iocb->ki_pos = pos + num_written;
1972 pagecache_isize_extended(inode, oldsize,
1973 i_size_read(inode));
1976 inode_unlock(inode);
1979 * We also have to set last_sub_trans to the current log transid,
1980 * otherwise subsequent syncs to a file that's been synced in this
1981 * transaction will appear to have already occurred.
1983 spin_lock(&BTRFS_I(inode)->lock);
1984 BTRFS_I(inode)->last_sub_trans = root->log_transid;
1985 spin_unlock(&BTRFS_I(inode)->lock);
1986 if (num_written > 0)
1987 num_written = generic_write_sync(iocb, num_written);
1990 atomic_dec(&BTRFS_I(inode)->sync_writers);
1992 current->backing_dev_info = NULL;
1993 return num_written ? num_written : err;
1996 int btrfs_release_file(struct inode *inode, struct file *filp)
1998 struct btrfs_file_private *private = filp->private_data;
2000 if (private && private->trans)
2001 btrfs_ioctl_trans_end(filp);
2002 if (private && private->filldir_buf)
2003 kfree(private->filldir_buf);
2005 filp->private_data = NULL;
2008 * ordered_data_close is set by settattr when we are about to truncate
2009 * a file from a non-zero size to a zero size. This tries to
2010 * flush down new bytes that may have been written if the
2011 * application were using truncate to replace a file in place.
2013 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2014 &BTRFS_I(inode)->runtime_flags))
2015 filemap_flush(inode->i_mapping);
2019 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2022 struct blk_plug plug;
2025 * This is only called in fsync, which would do synchronous writes, so
2026 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2027 * multiple disks using raid profile, a large IO can be split to
2028 * several segments of stripe length (currently 64K).
2030 blk_start_plug(&plug);
2031 atomic_inc(&BTRFS_I(inode)->sync_writers);
2032 ret = btrfs_fdatawrite_range(inode, start, end);
2033 atomic_dec(&BTRFS_I(inode)->sync_writers);
2034 blk_finish_plug(&plug);
2040 * fsync call for both files and directories. This logs the inode into
2041 * the tree log instead of forcing full commits whenever possible.
2043 * It needs to call filemap_fdatawait so that all ordered extent updates are
2044 * in the metadata btree are up to date for copying to the log.
2046 * It drops the inode mutex before doing the tree log commit. This is an
2047 * important optimization for directories because holding the mutex prevents
2048 * new operations on the dir while we write to disk.
2050 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2052 struct dentry *dentry = file_dentry(file);
2053 struct inode *inode = d_inode(dentry);
2054 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2055 struct btrfs_root *root = BTRFS_I(inode)->root;
2056 struct btrfs_trans_handle *trans;
2057 struct btrfs_log_ctx ctx;
2063 * If the inode needs a full sync, make sure we use a full range to
2064 * avoid log tree corruption, due to hole detection racing with ordered
2065 * extent completion for adjacent ranges, and assertion failures during
2068 if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2069 &BTRFS_I(inode)->runtime_flags)) {
2075 * The range length can be represented by u64, we have to do the typecasts
2076 * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
2078 len = (u64)end - (u64)start + 1;
2079 trace_btrfs_sync_file(file, datasync);
2081 btrfs_init_log_ctx(&ctx, inode);
2084 * We write the dirty pages in the range and wait until they complete
2085 * out of the ->i_mutex. If so, we can flush the dirty pages by
2086 * multi-task, and make the performance up. See
2087 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2089 ret = start_ordered_ops(inode, start, end);
2096 * We take the dio_sem here because the tree log stuff can race with
2097 * lockless dio writes and get an extent map logged for an extent we
2098 * never waited on. We need it this high up for lockdep reasons.
2100 down_write(&BTRFS_I(inode)->dio_sem);
2102 atomic_inc(&root->log_batch);
2103 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2104 &BTRFS_I(inode)->runtime_flags);
2107 * We have to do this here to avoid the priority inversion of waiting on
2108 * IO of a lower priority task while holding a transaciton open.
2110 ret = btrfs_wait_ordered_range(inode, start, len);
2112 up_write(&BTRFS_I(inode)->dio_sem);
2113 inode_unlock(inode);
2116 atomic_inc(&root->log_batch);
2119 * If the last transaction that changed this file was before the current
2120 * transaction and we have the full sync flag set in our inode, we can
2121 * bail out now without any syncing.
2123 * Note that we can't bail out if the full sync flag isn't set. This is
2124 * because when the full sync flag is set we start all ordered extents
2125 * and wait for them to fully complete - when they complete they update
2126 * the inode's last_trans field through:
2128 * btrfs_finish_ordered_io() ->
2129 * btrfs_update_inode_fallback() ->
2130 * btrfs_update_inode() ->
2131 * btrfs_set_inode_last_trans()
2133 * So we are sure that last_trans is up to date and can do this check to
2134 * bail out safely. For the fast path, when the full sync flag is not
2135 * set in our inode, we can not do it because we start only our ordered
2136 * extents and don't wait for them to complete (that is when
2137 * btrfs_finish_ordered_io runs), so here at this point their last_trans
2138 * value might be less than or equals to fs_info->last_trans_committed,
2139 * and setting a speculative last_trans for an inode when a buffered
2140 * write is made (such as fs_info->generation + 1 for example) would not
2141 * be reliable since after setting the value and before fsync is called
2142 * any number of transactions can start and commit (transaction kthread
2143 * commits the current transaction periodically), and a transaction
2144 * commit does not start nor waits for ordered extents to complete.
2147 if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2148 (full_sync && BTRFS_I(inode)->last_trans <=
2149 fs_info->last_trans_committed) ||
2150 (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2151 BTRFS_I(inode)->last_trans
2152 <= fs_info->last_trans_committed)) {
2154 * We've had everything committed since the last time we were
2155 * modified so clear this flag in case it was set for whatever
2156 * reason, it's no longer relevant.
2158 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2159 &BTRFS_I(inode)->runtime_flags);
2161 * An ordered extent might have started before and completed
2162 * already with io errors, in which case the inode was not
2163 * updated and we end up here. So check the inode's mapping
2164 * for any errors that might have happened since we last
2165 * checked called fsync.
2167 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2168 up_write(&BTRFS_I(inode)->dio_sem);
2169 inode_unlock(inode);
2174 * ok we haven't committed the transaction yet, lets do a commit
2176 if (file->private_data)
2177 btrfs_ioctl_trans_end(file);
2180 * We use start here because we will need to wait on the IO to complete
2181 * in btrfs_sync_log, which could require joining a transaction (for
2182 * example checking cross references in the nocow path). If we use join
2183 * here we could get into a situation where we're waiting on IO to
2184 * happen that is blocked on a transaction trying to commit. With start
2185 * we inc the extwriter counter, so we wait for all extwriters to exit
2186 * before we start blocking join'ers. This comment is to keep somebody
2187 * from thinking they are super smart and changing this to
2188 * btrfs_join_transaction *cough*Josef*cough*.
2190 trans = btrfs_start_transaction(root, 0);
2191 if (IS_ERR(trans)) {
2192 ret = PTR_ERR(trans);
2193 up_write(&BTRFS_I(inode)->dio_sem);
2194 inode_unlock(inode);
2199 ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2201 /* Fallthrough and commit/free transaction. */
2205 /* we've logged all the items and now have a consistent
2206 * version of the file in the log. It is possible that
2207 * someone will come in and modify the file, but that's
2208 * fine because the log is consistent on disk, and we
2209 * have references to all of the file's extents
2211 * It is possible that someone will come in and log the
2212 * file again, but that will end up using the synchronization
2213 * inside btrfs_sync_log to keep things safe.
2215 up_write(&BTRFS_I(inode)->dio_sem);
2216 inode_unlock(inode);
2219 * If any of the ordered extents had an error, just return it to user
2220 * space, so that the application knows some writes didn't succeed and
2221 * can take proper action (retry for e.g.). Blindly committing the
2222 * transaction in this case, would fool userspace that everything was
2223 * successful. And we also want to make sure our log doesn't contain
2224 * file extent items pointing to extents that weren't fully written to -
2225 * just like in the non fast fsync path, where we check for the ordered
2226 * operation's error flag before writing to the log tree and return -EIO
2227 * if any of them had this flag set (btrfs_wait_ordered_range) -
2228 * therefore we need to check for errors in the ordered operations,
2229 * which are indicated by ctx.io_err.
2232 btrfs_end_transaction(trans);
2237 if (ret != BTRFS_NO_LOG_SYNC) {
2239 ret = btrfs_sync_log(trans, root, &ctx);
2241 ret = btrfs_end_transaction(trans);
2245 ret = btrfs_commit_transaction(trans);
2247 ret = btrfs_end_transaction(trans);
2250 ASSERT(list_empty(&ctx.list));
2251 err = file_check_and_advance_wb_err(file);
2254 return ret > 0 ? -EIO : ret;
2257 static const struct vm_operations_struct btrfs_file_vm_ops = {
2258 .fault = filemap_fault,
2259 .map_pages = filemap_map_pages,
2260 .page_mkwrite = btrfs_page_mkwrite,
2263 static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2265 struct address_space *mapping = filp->f_mapping;
2267 if (!mapping->a_ops->readpage)
2270 file_accessed(filp);
2271 vma->vm_ops = &btrfs_file_vm_ops;
2276 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2277 int slot, u64 start, u64 end)
2279 struct btrfs_file_extent_item *fi;
2280 struct btrfs_key key;
2282 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2285 btrfs_item_key_to_cpu(leaf, &key, slot);
2286 if (key.objectid != btrfs_ino(inode) ||
2287 key.type != BTRFS_EXTENT_DATA_KEY)
2290 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2292 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2295 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2298 if (key.offset == end)
2300 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2305 static int fill_holes(struct btrfs_trans_handle *trans,
2306 struct btrfs_inode *inode,
2307 struct btrfs_path *path, u64 offset, u64 end)
2309 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
2310 struct btrfs_root *root = inode->root;
2311 struct extent_buffer *leaf;
2312 struct btrfs_file_extent_item *fi;
2313 struct extent_map *hole_em;
2314 struct extent_map_tree *em_tree = &inode->extent_tree;
2315 struct btrfs_key key;
2318 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2321 key.objectid = btrfs_ino(inode);
2322 key.type = BTRFS_EXTENT_DATA_KEY;
2323 key.offset = offset;
2325 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2328 * We should have dropped this offset, so if we find it then
2329 * something has gone horribly wrong.
2336 leaf = path->nodes[0];
2337 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2341 fi = btrfs_item_ptr(leaf, path->slots[0],
2342 struct btrfs_file_extent_item);
2343 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2345 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2346 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2347 btrfs_set_file_extent_offset(leaf, fi, 0);
2348 btrfs_mark_buffer_dirty(leaf);
2352 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2355 key.offset = offset;
2356 btrfs_set_item_key_safe(fs_info, path, &key);
2357 fi = btrfs_item_ptr(leaf, path->slots[0],
2358 struct btrfs_file_extent_item);
2359 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2361 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2362 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2363 btrfs_set_file_extent_offset(leaf, fi, 0);
2364 btrfs_mark_buffer_dirty(leaf);
2367 btrfs_release_path(path);
2369 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2370 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2375 btrfs_release_path(path);
2377 hole_em = alloc_extent_map();
2379 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2380 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2382 hole_em->start = offset;
2383 hole_em->len = end - offset;
2384 hole_em->ram_bytes = hole_em->len;
2385 hole_em->orig_start = offset;
2387 hole_em->block_start = EXTENT_MAP_HOLE;
2388 hole_em->block_len = 0;
2389 hole_em->orig_block_len = 0;
2390 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2391 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2392 hole_em->generation = trans->transid;
2395 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2396 write_lock(&em_tree->lock);
2397 ret = add_extent_mapping(em_tree, hole_em, 1);
2398 write_unlock(&em_tree->lock);
2399 } while (ret == -EEXIST);
2400 free_extent_map(hole_em);
2402 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2403 &inode->runtime_flags);
2410 * Find a hole extent on given inode and change start/len to the end of hole
2411 * extent.(hole/vacuum extent whose em->start <= start &&
2412 * em->start + em->len > start)
2413 * When a hole extent is found, return 1 and modify start/len.
2415 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2417 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2418 struct extent_map *em;
2421 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2422 round_down(*start, fs_info->sectorsize),
2423 round_up(*len, fs_info->sectorsize), 0);
2427 /* Hole or vacuum extent(only exists in no-hole mode) */
2428 if (em->block_start == EXTENT_MAP_HOLE) {
2430 *len = em->start + em->len > *start + *len ?
2431 0 : *start + *len - em->start - em->len;
2432 *start = em->start + em->len;
2434 free_extent_map(em);
2438 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2440 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2441 struct btrfs_root *root = BTRFS_I(inode)->root;
2442 struct extent_state *cached_state = NULL;
2443 struct btrfs_path *path;
2444 struct btrfs_block_rsv *rsv;
2445 struct btrfs_trans_handle *trans;
2450 u64 orig_start = offset;
2452 u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2456 unsigned int rsv_count;
2458 bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2460 bool truncated_block = false;
2461 bool updated_inode = false;
2463 ret = btrfs_wait_ordered_range(inode, offset, len);
2468 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2469 ret = find_first_non_hole(inode, &offset, &len);
2471 goto out_only_mutex;
2473 /* Already in a large hole */
2475 goto out_only_mutex;
2478 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2479 lockend = round_down(offset + len,
2480 btrfs_inode_sectorsize(inode)) - 1;
2481 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2482 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2484 * We needn't truncate any block which is beyond the end of the file
2485 * because we are sure there is no data there.
2488 * Only do this if we are in the same block and we aren't doing the
2491 if (same_block && len < fs_info->sectorsize) {
2492 if (offset < ino_size) {
2493 truncated_block = true;
2494 ret = btrfs_truncate_block(inode, offset, len, 0);
2498 goto out_only_mutex;
2501 /* zero back part of the first block */
2502 if (offset < ino_size) {
2503 truncated_block = true;
2504 ret = btrfs_truncate_block(inode, offset, 0, 0);
2506 inode_unlock(inode);
2511 /* Check the aligned pages after the first unaligned page,
2512 * if offset != orig_start, which means the first unaligned page
2513 * including several following pages are already in holes,
2514 * the extra check can be skipped */
2515 if (offset == orig_start) {
2516 /* after truncate page, check hole again */
2517 len = offset + len - lockstart;
2519 ret = find_first_non_hole(inode, &offset, &len);
2521 goto out_only_mutex;
2524 goto out_only_mutex;
2529 /* Check the tail unaligned part is in a hole */
2530 tail_start = lockend + 1;
2531 tail_len = offset + len - tail_start;
2533 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2534 if (unlikely(ret < 0))
2535 goto out_only_mutex;
2537 /* zero the front end of the last page */
2538 if (tail_start + tail_len < ino_size) {
2539 truncated_block = true;
2540 ret = btrfs_truncate_block(inode,
2541 tail_start + tail_len,
2544 goto out_only_mutex;
2549 if (lockend < lockstart) {
2551 goto out_only_mutex;
2555 struct btrfs_ordered_extent *ordered;
2557 truncate_pagecache_range(inode, lockstart, lockend);
2559 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2561 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2564 * We need to make sure we have no ordered extents in this range
2565 * and nobody raced in and read a page in this range, if we did
2566 * we need to try again.
2569 (ordered->file_offset + ordered->len <= lockstart ||
2570 ordered->file_offset > lockend)) &&
2571 !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2573 btrfs_put_ordered_extent(ordered);
2577 btrfs_put_ordered_extent(ordered);
2578 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2579 lockend, &cached_state, GFP_NOFS);
2580 ret = btrfs_wait_ordered_range(inode, lockstart,
2581 lockend - lockstart + 1);
2583 inode_unlock(inode);
2588 path = btrfs_alloc_path();
2594 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2599 rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2603 * 1 - update the inode
2604 * 1 - removing the extents in the range
2605 * 1 - adding the hole extent if no_holes isn't set
2607 rsv_count = no_holes ? 2 : 3;
2608 trans = btrfs_start_transaction(root, rsv_count);
2609 if (IS_ERR(trans)) {
2610 err = PTR_ERR(trans);
2614 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2617 trans->block_rsv = rsv;
2619 cur_offset = lockstart;
2620 len = lockend - cur_offset;
2621 while (cur_offset < lockend) {
2622 ret = __btrfs_drop_extents(trans, root, inode, path,
2623 cur_offset, lockend + 1,
2624 &drop_end, 1, 0, 0, NULL);
2628 trans->block_rsv = &fs_info->trans_block_rsv;
2630 if (cur_offset < drop_end && cur_offset < ino_size) {
2631 ret = fill_holes(trans, BTRFS_I(inode), path,
2632 cur_offset, drop_end);
2635 * If we failed then we didn't insert our hole
2636 * entries for the area we dropped, so now the
2637 * fs is corrupted, so we must abort the
2640 btrfs_abort_transaction(trans, ret);
2646 cur_offset = drop_end;
2648 ret = btrfs_update_inode(trans, root, inode);
2654 btrfs_end_transaction(trans);
2655 btrfs_btree_balance_dirty(fs_info);
2657 trans = btrfs_start_transaction(root, rsv_count);
2658 if (IS_ERR(trans)) {
2659 ret = PTR_ERR(trans);
2664 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2666 BUG_ON(ret); /* shouldn't happen */
2667 trans->block_rsv = rsv;
2669 ret = find_first_non_hole(inode, &cur_offset, &len);
2670 if (unlikely(ret < 0))
2683 trans->block_rsv = &fs_info->trans_block_rsv;
2685 * If we are using the NO_HOLES feature we might have had already an
2686 * hole that overlaps a part of the region [lockstart, lockend] and
2687 * ends at (or beyond) lockend. Since we have no file extent items to
2688 * represent holes, drop_end can be less than lockend and so we must
2689 * make sure we have an extent map representing the existing hole (the
2690 * call to __btrfs_drop_extents() might have dropped the existing extent
2691 * map representing the existing hole), otherwise the fast fsync path
2692 * will not record the existence of the hole region
2693 * [existing_hole_start, lockend].
2695 if (drop_end <= lockend)
2696 drop_end = lockend + 1;
2698 * Don't insert file hole extent item if it's for a range beyond eof
2699 * (because it's useless) or if it represents a 0 bytes range (when
2700 * cur_offset == drop_end).
2702 if (cur_offset < ino_size && cur_offset < drop_end) {
2703 ret = fill_holes(trans, BTRFS_I(inode), path,
2704 cur_offset, drop_end);
2706 /* Same comment as above. */
2707 btrfs_abort_transaction(trans, ret);
2717 inode_inc_iversion(inode);
2718 inode->i_mtime = inode->i_ctime = current_time(inode);
2720 trans->block_rsv = &fs_info->trans_block_rsv;
2721 ret = btrfs_update_inode(trans, root, inode);
2722 updated_inode = true;
2723 btrfs_end_transaction(trans);
2724 btrfs_btree_balance_dirty(fs_info);
2726 btrfs_free_path(path);
2727 btrfs_free_block_rsv(fs_info, rsv);
2729 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2730 &cached_state, GFP_NOFS);
2732 if (!updated_inode && truncated_block && !ret && !err) {
2734 * If we only end up zeroing part of a page, we still need to
2735 * update the inode item, so that all the time fields are
2736 * updated as well as the necessary btrfs inode in memory fields
2737 * for detecting, at fsync time, if the inode isn't yet in the
2738 * log tree or it's there but not up to date.
2740 struct timespec now = current_time(inode);
2742 inode_inc_iversion(inode);
2743 inode->i_mtime = now;
2744 inode->i_ctime = now;
2745 trans = btrfs_start_transaction(root, 1);
2746 if (IS_ERR(trans)) {
2747 err = PTR_ERR(trans);
2749 err = btrfs_update_inode(trans, root, inode);
2750 ret = btrfs_end_transaction(trans);
2753 inode_unlock(inode);
2759 /* Helper structure to record which range is already reserved */
2760 struct falloc_range {
2761 struct list_head list;
2767 * Helper function to add falloc range
2769 * Caller should have locked the larger range of extent containing
2772 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2774 struct falloc_range *prev = NULL;
2775 struct falloc_range *range = NULL;
2777 if (list_empty(head))
2781 * As fallocate iterate by bytenr order, we only need to check
2784 prev = list_entry(head->prev, struct falloc_range, list);
2785 if (prev->start + prev->len == start) {
2790 range = kmalloc(sizeof(*range), GFP_KERNEL);
2793 range->start = start;
2795 list_add_tail(&range->list, head);
2799 static long btrfs_fallocate(struct file *file, int mode,
2800 loff_t offset, loff_t len)
2802 struct inode *inode = file_inode(file);
2803 struct extent_state *cached_state = NULL;
2804 struct extent_changeset *data_reserved = NULL;
2805 struct falloc_range *range;
2806 struct falloc_range *tmp;
2807 struct list_head reserve_list;
2815 struct extent_map *em;
2816 int blocksize = btrfs_inode_sectorsize(inode);
2819 alloc_start = round_down(offset, blocksize);
2820 alloc_end = round_up(offset + len, blocksize);
2821 cur_offset = alloc_start;
2823 /* Make sure we aren't being give some crap mode */
2824 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2827 if (mode & FALLOC_FL_PUNCH_HOLE)
2828 return btrfs_punch_hole(inode, offset, len);
2831 * Only trigger disk allocation, don't trigger qgroup reserve
2833 * For qgroup space, it will be checked later.
2835 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2836 alloc_end - alloc_start);
2842 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2843 ret = inode_newsize_ok(inode, offset + len);
2849 * TODO: Move these two operations after we have checked
2850 * accurate reserved space, or fallocate can still fail but
2851 * with page truncated or size expanded.
2853 * But that's a minor problem and won't do much harm BTW.
2855 if (alloc_start > inode->i_size) {
2856 ret = btrfs_cont_expand(inode, i_size_read(inode),
2860 } else if (offset + len > inode->i_size) {
2862 * If we are fallocating from the end of the file onward we
2863 * need to zero out the end of the block if i_size lands in the
2864 * middle of a block.
2866 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
2872 * wait for ordered IO before we have any locks. We'll loop again
2873 * below with the locks held.
2875 ret = btrfs_wait_ordered_range(inode, alloc_start,
2876 alloc_end - alloc_start);
2880 locked_end = alloc_end - 1;
2882 struct btrfs_ordered_extent *ordered;
2884 /* the extent lock is ordered inside the running
2887 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2888 locked_end, &cached_state);
2889 ordered = btrfs_lookup_first_ordered_extent(inode,
2892 ordered->file_offset + ordered->len > alloc_start &&
2893 ordered->file_offset < alloc_end) {
2894 btrfs_put_ordered_extent(ordered);
2895 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2896 alloc_start, locked_end,
2897 &cached_state, GFP_KERNEL);
2899 * we can't wait on the range with the transaction
2900 * running or with the extent lock held
2902 ret = btrfs_wait_ordered_range(inode, alloc_start,
2903 alloc_end - alloc_start);
2908 btrfs_put_ordered_extent(ordered);
2913 /* First, check if we exceed the qgroup limit */
2914 INIT_LIST_HEAD(&reserve_list);
2916 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
2917 alloc_end - cur_offset, 0);
2922 last_byte = min(extent_map_end(em), alloc_end);
2923 actual_end = min_t(u64, extent_map_end(em), offset + len);
2924 last_byte = ALIGN(last_byte, blocksize);
2925 if (em->block_start == EXTENT_MAP_HOLE ||
2926 (cur_offset >= inode->i_size &&
2927 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2928 ret = add_falloc_range(&reserve_list, cur_offset,
2929 last_byte - cur_offset);
2931 free_extent_map(em);
2934 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
2935 cur_offset, last_byte - cur_offset);
2937 cur_offset = last_byte;
2938 free_extent_map(em);
2943 * Do not need to reserve unwritten extent for this
2944 * range, free reserved data space first, otherwise
2945 * it'll result in false ENOSPC error.
2947 btrfs_free_reserved_data_space(inode, data_reserved,
2948 cur_offset, last_byte - cur_offset);
2950 free_extent_map(em);
2951 cur_offset = last_byte;
2952 if (cur_offset >= alloc_end)
2957 * If ret is still 0, means we're OK to fallocate.
2958 * Or just cleanup the list and exit.
2960 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
2962 ret = btrfs_prealloc_file_range(inode, mode,
2964 range->len, i_blocksize(inode),
2965 offset + len, &alloc_hint);
2967 btrfs_free_reserved_data_space(inode,
2968 data_reserved, range->start,
2970 list_del(&range->list);
2976 if (actual_end > inode->i_size &&
2977 !(mode & FALLOC_FL_KEEP_SIZE)) {
2978 struct btrfs_trans_handle *trans;
2979 struct btrfs_root *root = BTRFS_I(inode)->root;
2982 * We didn't need to allocate any more space, but we
2983 * still extended the size of the file so we need to
2984 * update i_size and the inode item.
2986 trans = btrfs_start_transaction(root, 1);
2987 if (IS_ERR(trans)) {
2988 ret = PTR_ERR(trans);
2990 inode->i_ctime = current_time(inode);
2991 i_size_write(inode, actual_end);
2992 btrfs_ordered_update_i_size(inode, actual_end, NULL);
2993 ret = btrfs_update_inode(trans, root, inode);
2995 btrfs_end_transaction(trans);
2997 ret = btrfs_end_transaction(trans);
3001 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3002 &cached_state, GFP_KERNEL);
3004 inode_unlock(inode);
3005 /* Let go of our reservation. */
3007 btrfs_free_reserved_data_space(inode, data_reserved,
3008 cur_offset, alloc_end - cur_offset);
3009 extent_changeset_free(data_reserved);
3013 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3015 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3016 struct extent_map *em = NULL;
3017 struct extent_state *cached_state = NULL;
3024 if (inode->i_size == 0)
3028 * *offset can be negative, in this case we start finding DATA/HOLE from
3029 * the very start of the file.
3031 start = max_t(loff_t, 0, *offset);
3033 lockstart = round_down(start, fs_info->sectorsize);
3034 lockend = round_up(i_size_read(inode),
3035 fs_info->sectorsize);
3036 if (lockend <= lockstart)
3037 lockend = lockstart + fs_info->sectorsize;
3039 len = lockend - lockstart + 1;
3041 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3044 while (start < inode->i_size) {
3045 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0,
3053 if (whence == SEEK_HOLE &&
3054 (em->block_start == EXTENT_MAP_HOLE ||
3055 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3057 else if (whence == SEEK_DATA &&
3058 (em->block_start != EXTENT_MAP_HOLE &&
3059 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3062 start = em->start + em->len;
3063 free_extent_map(em);
3067 free_extent_map(em);
3069 if (whence == SEEK_DATA && start >= inode->i_size)
3072 *offset = min_t(loff_t, start, inode->i_size);
3074 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3075 &cached_state, GFP_NOFS);
3079 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3081 struct inode *inode = file->f_mapping->host;
3088 offset = generic_file_llseek(file, offset, whence);
3092 if (offset >= i_size_read(inode)) {
3093 inode_unlock(inode);
3097 ret = find_desired_extent(inode, &offset, whence);
3099 inode_unlock(inode);
3104 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3106 inode_unlock(inode);
3110 static int btrfs_file_open(struct inode *inode, struct file *filp)
3112 filp->f_mode |= FMODE_NOWAIT;
3113 return generic_file_open(inode, filp);
3116 const struct file_operations btrfs_file_operations = {
3117 .llseek = btrfs_file_llseek,
3118 .read_iter = generic_file_read_iter,
3119 .splice_read = generic_file_splice_read,
3120 .write_iter = btrfs_file_write_iter,
3121 .mmap = btrfs_file_mmap,
3122 .open = btrfs_file_open,
3123 .release = btrfs_release_file,
3124 .fsync = btrfs_sync_file,
3125 .fallocate = btrfs_fallocate,
3126 .unlocked_ioctl = btrfs_ioctl,
3127 #ifdef CONFIG_COMPAT
3128 .compat_ioctl = btrfs_compat_ioctl,
3130 .clone_file_range = btrfs_clone_file_range,
3131 .dedupe_file_range = btrfs_dedupe_file_range,
3134 void btrfs_auto_defrag_exit(void)
3136 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3139 int btrfs_auto_defrag_init(void)
3141 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3142 sizeof(struct inode_defrag), 0,
3145 if (!btrfs_inode_defrag_cachep)
3151 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3156 * So with compression we will find and lock a dirty page and clear the
3157 * first one as dirty, setup an async extent, and immediately return
3158 * with the entire range locked but with nobody actually marked with
3159 * writeback. So we can't just filemap_write_and_wait_range() and
3160 * expect it to work since it will just kick off a thread to do the
3161 * actual work. So we need to call filemap_fdatawrite_range _again_
3162 * since it will wait on the page lock, which won't be unlocked until
3163 * after the pages have been marked as writeback and so we're good to go
3164 * from there. We have to do this otherwise we'll miss the ordered
3165 * extents and that results in badness. Please Josef, do not think you
3166 * know better and pull this out at some point in the future, it is
3167 * right and you are wrong.
3169 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3170 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3171 &BTRFS_I(inode)->runtime_flags))
3172 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);