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>
34 #include <linux/iversion.h>
37 #include "transaction.h"
38 #include "btrfs_inode.h"
39 #include "print-tree.h"
44 #include "compression.h"
46 static struct kmem_cache *btrfs_inode_defrag_cachep;
48 * when auto defrag is enabled we
49 * queue up these defrag structs to remember which
50 * inodes need defragging passes
53 struct rb_node rb_node;
57 * transid where the defrag was added, we search for
58 * extents newer than this
65 /* last offset we were able to defrag */
68 /* if we've wrapped around back to zero once already */
72 static int __compare_inode_defrag(struct inode_defrag *defrag1,
73 struct inode_defrag *defrag2)
75 if (defrag1->root > defrag2->root)
77 else if (defrag1->root < defrag2->root)
79 else if (defrag1->ino > defrag2->ino)
81 else if (defrag1->ino < defrag2->ino)
87 /* pop a record for an inode into the defrag tree. The lock
88 * must be held already
90 * If you're inserting a record for an older transid than an
91 * existing record, the transid already in the tree is lowered
93 * If an existing record is found the defrag item you
96 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
97 struct inode_defrag *defrag)
99 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
100 struct inode_defrag *entry;
102 struct rb_node *parent = NULL;
105 p = &fs_info->defrag_inodes.rb_node;
108 entry = rb_entry(parent, struct inode_defrag, rb_node);
110 ret = __compare_inode_defrag(defrag, entry);
112 p = &parent->rb_left;
114 p = &parent->rb_right;
116 /* if we're reinserting an entry for
117 * an old defrag run, make sure to
118 * lower the transid of our existing record
120 if (defrag->transid < entry->transid)
121 entry->transid = defrag->transid;
122 if (defrag->last_offset > entry->last_offset)
123 entry->last_offset = defrag->last_offset;
127 set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
128 rb_link_node(&defrag->rb_node, parent, p);
129 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
133 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
135 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
138 if (btrfs_fs_closing(fs_info))
145 * insert a defrag record for this inode if auto defrag is
148 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
149 struct btrfs_inode *inode)
151 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
152 struct btrfs_root *root = inode->root;
153 struct inode_defrag *defrag;
157 if (!__need_auto_defrag(fs_info))
160 if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
164 transid = trans->transid;
166 transid = inode->root->last_trans;
168 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
172 defrag->ino = btrfs_ino(inode);
173 defrag->transid = transid;
174 defrag->root = root->root_key.objectid;
176 spin_lock(&fs_info->defrag_inodes_lock);
177 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
179 * If we set IN_DEFRAG flag and evict the inode from memory,
180 * and then re-read this inode, this new inode doesn't have
181 * IN_DEFRAG flag. At the case, we may find the existed defrag.
183 ret = __btrfs_add_inode_defrag(inode, defrag);
185 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
187 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
189 spin_unlock(&fs_info->defrag_inodes_lock);
194 * Requeue the defrag object. If there is a defrag object that points to
195 * the same inode in the tree, we will merge them together (by
196 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
198 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
199 struct inode_defrag *defrag)
201 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
204 if (!__need_auto_defrag(fs_info))
208 * Here we don't check the IN_DEFRAG flag, because we need merge
211 spin_lock(&fs_info->defrag_inodes_lock);
212 ret = __btrfs_add_inode_defrag(inode, defrag);
213 spin_unlock(&fs_info->defrag_inodes_lock);
218 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
222 * pick the defragable inode that we want, if it doesn't exist, we will get
225 static struct inode_defrag *
226 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
228 struct inode_defrag *entry = NULL;
229 struct inode_defrag tmp;
231 struct rb_node *parent = NULL;
237 spin_lock(&fs_info->defrag_inodes_lock);
238 p = fs_info->defrag_inodes.rb_node;
241 entry = rb_entry(parent, struct inode_defrag, rb_node);
243 ret = __compare_inode_defrag(&tmp, entry);
247 p = parent->rb_right;
252 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
253 parent = rb_next(parent);
255 entry = rb_entry(parent, struct inode_defrag, rb_node);
261 rb_erase(parent, &fs_info->defrag_inodes);
262 spin_unlock(&fs_info->defrag_inodes_lock);
266 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
268 struct inode_defrag *defrag;
269 struct rb_node *node;
271 spin_lock(&fs_info->defrag_inodes_lock);
272 node = rb_first(&fs_info->defrag_inodes);
274 rb_erase(node, &fs_info->defrag_inodes);
275 defrag = rb_entry(node, struct inode_defrag, rb_node);
276 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
278 cond_resched_lock(&fs_info->defrag_inodes_lock);
280 node = rb_first(&fs_info->defrag_inodes);
282 spin_unlock(&fs_info->defrag_inodes_lock);
285 #define BTRFS_DEFRAG_BATCH 1024
287 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
288 struct inode_defrag *defrag)
290 struct btrfs_root *inode_root;
292 struct btrfs_key key;
293 struct btrfs_ioctl_defrag_range_args range;
299 key.objectid = defrag->root;
300 key.type = BTRFS_ROOT_ITEM_KEY;
301 key.offset = (u64)-1;
303 index = srcu_read_lock(&fs_info->subvol_srcu);
305 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
306 if (IS_ERR(inode_root)) {
307 ret = PTR_ERR(inode_root);
311 key.objectid = defrag->ino;
312 key.type = BTRFS_INODE_ITEM_KEY;
314 inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
316 ret = PTR_ERR(inode);
319 srcu_read_unlock(&fs_info->subvol_srcu, index);
321 /* do a chunk of defrag */
322 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
323 memset(&range, 0, sizeof(range));
325 range.start = defrag->last_offset;
327 sb_start_write(fs_info->sb);
328 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
330 sb_end_write(fs_info->sb);
332 * if we filled the whole defrag batch, there
333 * must be more work to do. Queue this defrag
336 if (num_defrag == BTRFS_DEFRAG_BATCH) {
337 defrag->last_offset = range.start;
338 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
339 } else if (defrag->last_offset && !defrag->cycled) {
341 * we didn't fill our defrag batch, but
342 * we didn't start at zero. Make sure we loop
343 * around to the start of the file.
345 defrag->last_offset = 0;
347 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
349 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
355 srcu_read_unlock(&fs_info->subvol_srcu, index);
356 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
361 * run through the list of inodes in the FS that need
364 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
366 struct inode_defrag *defrag;
368 u64 root_objectid = 0;
370 atomic_inc(&fs_info->defrag_running);
372 /* Pause the auto defragger. */
373 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
377 if (!__need_auto_defrag(fs_info))
380 /* find an inode to defrag */
381 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
384 if (root_objectid || first_ino) {
393 first_ino = defrag->ino + 1;
394 root_objectid = defrag->root;
396 __btrfs_run_defrag_inode(fs_info, defrag);
398 atomic_dec(&fs_info->defrag_running);
401 * during unmount, we use the transaction_wait queue to
402 * wait for the defragger to stop
404 wake_up(&fs_info->transaction_wait);
408 /* simple helper to fault in pages and copy. This should go away
409 * and be replaced with calls into generic code.
411 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
412 struct page **prepared_pages,
416 size_t total_copied = 0;
418 int offset = pos & (PAGE_SIZE - 1);
420 while (write_bytes > 0) {
421 size_t count = min_t(size_t,
422 PAGE_SIZE - offset, write_bytes);
423 struct page *page = prepared_pages[pg];
425 * Copy data from userspace to the current page
427 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
429 /* Flush processor's dcache for this page */
430 flush_dcache_page(page);
433 * if we get a partial write, we can end up with
434 * partially up to date pages. These add
435 * a lot of complexity, so make sure they don't
436 * happen by forcing this copy to be retried.
438 * The rest of the btrfs_file_write code will fall
439 * back to page at a time copies after we return 0.
441 if (!PageUptodate(page) && copied < count)
444 iov_iter_advance(i, copied);
445 write_bytes -= copied;
446 total_copied += copied;
448 /* Return to btrfs_file_write_iter to fault page */
449 if (unlikely(copied == 0))
452 if (copied < PAGE_SIZE - offset) {
463 * unlocks pages after btrfs_file_write is done with them
465 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
468 for (i = 0; i < num_pages; i++) {
469 /* page checked is some magic around finding pages that
470 * have been modified without going through btrfs_set_page_dirty
471 * clear it here. There should be no need to mark the pages
472 * accessed as prepare_pages should have marked them accessed
473 * in prepare_pages via find_or_create_page()
475 ClearPageChecked(pages[i]);
476 unlock_page(pages[i]);
481 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
484 struct extent_state **cached_state)
486 u64 search_start = start;
487 const u64 end = start + len - 1;
489 while (search_start < end) {
490 const u64 search_len = end - search_start + 1;
491 struct extent_map *em;
495 em = btrfs_get_extent(inode, NULL, 0, search_start,
500 if (em->block_start != EXTENT_MAP_HOLE)
504 if (em->start < search_start)
505 em_len -= search_start - em->start;
506 if (em_len > search_len)
509 ret = set_extent_bit(&inode->io_tree, search_start,
510 search_start + em_len - 1,
512 NULL, cached_state, GFP_NOFS);
514 search_start = extent_map_end(em);
523 * after copy_from_user, pages need to be dirtied and we need to make
524 * sure holes are created between the current EOF and the start of
525 * any next extents (if required).
527 * this also makes the decision about creating an inline extent vs
528 * doing real data extents, marking pages dirty and delalloc as required.
530 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
531 size_t num_pages, loff_t pos, size_t write_bytes,
532 struct extent_state **cached)
534 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
539 u64 end_of_last_block;
540 u64 end_pos = pos + write_bytes;
541 loff_t isize = i_size_read(inode);
542 unsigned int extra_bits = 0;
544 start_pos = pos & ~((u64) fs_info->sectorsize - 1);
545 num_bytes = round_up(write_bytes + pos - start_pos,
546 fs_info->sectorsize);
548 end_of_last_block = start_pos + num_bytes - 1;
550 if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
551 if (start_pos >= isize &&
552 !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
554 * There can't be any extents following eof in this case
555 * so just set the delalloc new bit for the range
558 extra_bits |= EXTENT_DELALLOC_NEW;
560 err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
568 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
569 extra_bits, cached, 0);
573 for (i = 0; i < num_pages; i++) {
574 struct page *p = pages[i];
581 * we've only changed i_size in ram, and we haven't updated
582 * the disk i_size. There is no need to log the inode
586 i_size_write(inode, end_pos);
591 * this drops all the extents in the cache that intersect the range
592 * [start, end]. Existing extents are split as required.
594 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
597 struct extent_map *em;
598 struct extent_map *split = NULL;
599 struct extent_map *split2 = NULL;
600 struct extent_map_tree *em_tree = &inode->extent_tree;
601 u64 len = end - start + 1;
609 WARN_ON(end < start);
610 if (end == (u64)-1) {
619 split = alloc_extent_map();
621 split2 = alloc_extent_map();
622 if (!split || !split2)
625 write_lock(&em_tree->lock);
626 em = lookup_extent_mapping(em_tree, start, len);
628 write_unlock(&em_tree->lock);
632 gen = em->generation;
633 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
634 if (testend && em->start + em->len >= start + len) {
636 write_unlock(&em_tree->lock);
639 start = em->start + em->len;
641 len = start + len - (em->start + em->len);
643 write_unlock(&em_tree->lock);
646 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
647 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
648 clear_bit(EXTENT_FLAG_LOGGING, &flags);
649 modified = !list_empty(&em->list);
653 if (em->start < start) {
654 split->start = em->start;
655 split->len = start - em->start;
657 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
658 split->orig_start = em->orig_start;
659 split->block_start = em->block_start;
662 split->block_len = em->block_len;
664 split->block_len = split->len;
665 split->orig_block_len = max(split->block_len,
667 split->ram_bytes = em->ram_bytes;
669 split->orig_start = split->start;
670 split->block_len = 0;
671 split->block_start = em->block_start;
672 split->orig_block_len = 0;
673 split->ram_bytes = split->len;
676 split->generation = gen;
677 split->bdev = em->bdev;
678 split->flags = flags;
679 split->compress_type = em->compress_type;
680 replace_extent_mapping(em_tree, em, split, modified);
681 free_extent_map(split);
685 if (testend && em->start + em->len > start + len) {
686 u64 diff = start + len - em->start;
688 split->start = start + len;
689 split->len = em->start + em->len - (start + len);
690 split->bdev = em->bdev;
691 split->flags = flags;
692 split->compress_type = em->compress_type;
693 split->generation = gen;
695 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
696 split->orig_block_len = max(em->block_len,
699 split->ram_bytes = em->ram_bytes;
701 split->block_len = em->block_len;
702 split->block_start = em->block_start;
703 split->orig_start = em->orig_start;
705 split->block_len = split->len;
706 split->block_start = em->block_start
708 split->orig_start = em->orig_start;
711 split->ram_bytes = split->len;
712 split->orig_start = split->start;
713 split->block_len = 0;
714 split->block_start = em->block_start;
715 split->orig_block_len = 0;
718 if (extent_map_in_tree(em)) {
719 replace_extent_mapping(em_tree, em, split,
722 ret = add_extent_mapping(em_tree, split,
724 ASSERT(ret == 0); /* Logic error */
726 free_extent_map(split);
730 if (extent_map_in_tree(em))
731 remove_extent_mapping(em_tree, em);
732 write_unlock(&em_tree->lock);
736 /* once for the tree*/
740 free_extent_map(split);
742 free_extent_map(split2);
746 * this is very complex, but the basic idea is to drop all extents
747 * in the range start - end. hint_block is filled in with a block number
748 * that would be a good hint to the block allocator for this file.
750 * If an extent intersects the range but is not entirely inside the range
751 * it is either truncated or split. Anything entirely inside the range
752 * is deleted from the tree.
754 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
755 struct btrfs_root *root, struct inode *inode,
756 struct btrfs_path *path, u64 start, u64 end,
757 u64 *drop_end, int drop_cache,
759 u32 extent_item_size,
762 struct btrfs_fs_info *fs_info = root->fs_info;
763 struct extent_buffer *leaf;
764 struct btrfs_file_extent_item *fi;
765 struct btrfs_key key;
766 struct btrfs_key new_key;
767 u64 ino = btrfs_ino(BTRFS_I(inode));
768 u64 search_start = start;
771 u64 extent_offset = 0;
773 u64 last_end = start;
779 int modify_tree = -1;
782 int leafs_visited = 0;
785 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
787 if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
790 update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
791 root == fs_info->tree_root);
794 ret = btrfs_lookup_file_extent(trans, root, path, ino,
795 search_start, modify_tree);
798 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
799 leaf = path->nodes[0];
800 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
801 if (key.objectid == ino &&
802 key.type == BTRFS_EXTENT_DATA_KEY)
808 leaf = path->nodes[0];
809 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
811 ret = btrfs_next_leaf(root, path);
819 leaf = path->nodes[0];
823 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
825 if (key.objectid > ino)
827 if (WARN_ON_ONCE(key.objectid < ino) ||
828 key.type < BTRFS_EXTENT_DATA_KEY) {
833 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
836 fi = btrfs_item_ptr(leaf, path->slots[0],
837 struct btrfs_file_extent_item);
838 extent_type = btrfs_file_extent_type(leaf, fi);
840 if (extent_type == BTRFS_FILE_EXTENT_REG ||
841 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
842 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
843 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
844 extent_offset = btrfs_file_extent_offset(leaf, fi);
845 extent_end = key.offset +
846 btrfs_file_extent_num_bytes(leaf, fi);
847 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
848 extent_end = key.offset +
849 btrfs_file_extent_inline_len(leaf,
857 * Don't skip extent items representing 0 byte lengths. They
858 * used to be created (bug) if while punching holes we hit
859 * -ENOSPC condition. So if we find one here, just ensure we
860 * delete it, otherwise we would insert a new file extent item
861 * with the same key (offset) as that 0 bytes length file
862 * extent item in the call to setup_items_for_insert() later
865 if (extent_end == key.offset && extent_end >= search_start) {
866 last_end = extent_end;
867 goto delete_extent_item;
870 if (extent_end <= search_start) {
876 search_start = max(key.offset, start);
877 if (recow || !modify_tree) {
879 btrfs_release_path(path);
884 * | - range to drop - |
885 * | -------- extent -------- |
887 if (start > key.offset && end < extent_end) {
889 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
894 memcpy(&new_key, &key, sizeof(new_key));
895 new_key.offset = start;
896 ret = btrfs_duplicate_item(trans, root, path,
898 if (ret == -EAGAIN) {
899 btrfs_release_path(path);
905 leaf = path->nodes[0];
906 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
907 struct btrfs_file_extent_item);
908 btrfs_set_file_extent_num_bytes(leaf, fi,
911 fi = btrfs_item_ptr(leaf, path->slots[0],
912 struct btrfs_file_extent_item);
914 extent_offset += start - key.offset;
915 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
916 btrfs_set_file_extent_num_bytes(leaf, fi,
918 btrfs_mark_buffer_dirty(leaf);
920 if (update_refs && disk_bytenr > 0) {
921 ret = btrfs_inc_extent_ref(trans, root,
922 disk_bytenr, num_bytes, 0,
923 root->root_key.objectid,
925 start - extent_offset);
926 BUG_ON(ret); /* -ENOMEM */
931 * From here on out we will have actually dropped something, so
932 * last_end can be updated.
934 last_end = extent_end;
937 * | ---- range to drop ----- |
938 * | -------- extent -------- |
940 if (start <= key.offset && end < extent_end) {
941 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
946 memcpy(&new_key, &key, sizeof(new_key));
947 new_key.offset = end;
948 btrfs_set_item_key_safe(fs_info, path, &new_key);
950 extent_offset += end - key.offset;
951 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
952 btrfs_set_file_extent_num_bytes(leaf, fi,
954 btrfs_mark_buffer_dirty(leaf);
955 if (update_refs && disk_bytenr > 0)
956 inode_sub_bytes(inode, end - key.offset);
960 search_start = extent_end;
962 * | ---- range to drop ----- |
963 * | -------- extent -------- |
965 if (start > key.offset && end >= extent_end) {
967 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
972 btrfs_set_file_extent_num_bytes(leaf, fi,
974 btrfs_mark_buffer_dirty(leaf);
975 if (update_refs && disk_bytenr > 0)
976 inode_sub_bytes(inode, extent_end - start);
977 if (end == extent_end)
985 * | ---- range to drop ----- |
986 * | ------ extent ------ |
988 if (start <= key.offset && end >= extent_end) {
991 del_slot = path->slots[0];
994 BUG_ON(del_slot + del_nr != path->slots[0]);
999 extent_type == BTRFS_FILE_EXTENT_INLINE) {
1000 inode_sub_bytes(inode,
1001 extent_end - key.offset);
1002 extent_end = ALIGN(extent_end,
1003 fs_info->sectorsize);
1004 } else if (update_refs && disk_bytenr > 0) {
1005 ret = btrfs_free_extent(trans, root,
1006 disk_bytenr, num_bytes, 0,
1007 root->root_key.objectid,
1008 key.objectid, key.offset -
1010 BUG_ON(ret); /* -ENOMEM */
1011 inode_sub_bytes(inode,
1012 extent_end - key.offset);
1015 if (end == extent_end)
1018 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1023 ret = btrfs_del_items(trans, root, path, del_slot,
1026 btrfs_abort_transaction(trans, ret);
1033 btrfs_release_path(path);
1040 if (!ret && del_nr > 0) {
1042 * Set path->slots[0] to first slot, so that after the delete
1043 * if items are move off from our leaf to its immediate left or
1044 * right neighbor leafs, we end up with a correct and adjusted
1045 * path->slots[0] for our insertion (if replace_extent != 0).
1047 path->slots[0] = del_slot;
1048 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1050 btrfs_abort_transaction(trans, ret);
1053 leaf = path->nodes[0];
1055 * If btrfs_del_items() was called, it might have deleted a leaf, in
1056 * which case it unlocked our path, so check path->locks[0] matches a
1059 if (!ret && replace_extent && leafs_visited == 1 &&
1060 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1061 path->locks[0] == BTRFS_WRITE_LOCK) &&
1062 btrfs_leaf_free_space(fs_info, leaf) >=
1063 sizeof(struct btrfs_item) + extent_item_size) {
1066 key.type = BTRFS_EXTENT_DATA_KEY;
1068 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1069 struct btrfs_key slot_key;
1071 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1072 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1075 setup_items_for_insert(root, path, &key,
1078 sizeof(struct btrfs_item) +
1079 extent_item_size, 1);
1083 if (!replace_extent || !(*key_inserted))
1084 btrfs_release_path(path);
1086 *drop_end = found ? min(end, last_end) : end;
1090 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1091 struct btrfs_root *root, struct inode *inode, u64 start,
1092 u64 end, int drop_cache)
1094 struct btrfs_path *path;
1097 path = btrfs_alloc_path();
1100 ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1101 drop_cache, 0, 0, NULL);
1102 btrfs_free_path(path);
1106 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1107 u64 objectid, u64 bytenr, u64 orig_offset,
1108 u64 *start, u64 *end)
1110 struct btrfs_file_extent_item *fi;
1111 struct btrfs_key key;
1114 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1117 btrfs_item_key_to_cpu(leaf, &key, slot);
1118 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1121 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1122 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1123 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1124 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1125 btrfs_file_extent_compression(leaf, fi) ||
1126 btrfs_file_extent_encryption(leaf, fi) ||
1127 btrfs_file_extent_other_encoding(leaf, fi))
1130 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1131 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1134 *start = key.offset;
1140 * Mark extent in the range start - end as written.
1142 * This changes extent type from 'pre-allocated' to 'regular'. If only
1143 * part of extent is marked as written, the extent will be split into
1146 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1147 struct btrfs_inode *inode, u64 start, u64 end)
1149 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1150 struct btrfs_root *root = inode->root;
1151 struct extent_buffer *leaf;
1152 struct btrfs_path *path;
1153 struct btrfs_file_extent_item *fi;
1154 struct btrfs_key key;
1155 struct btrfs_key new_key;
1167 u64 ino = btrfs_ino(inode);
1169 path = btrfs_alloc_path();
1176 key.type = BTRFS_EXTENT_DATA_KEY;
1179 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1182 if (ret > 0 && path->slots[0] > 0)
1185 leaf = path->nodes[0];
1186 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1187 if (key.objectid != ino ||
1188 key.type != BTRFS_EXTENT_DATA_KEY) {
1190 btrfs_abort_transaction(trans, ret);
1193 fi = btrfs_item_ptr(leaf, path->slots[0],
1194 struct btrfs_file_extent_item);
1195 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1197 btrfs_abort_transaction(trans, ret);
1200 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1201 if (key.offset > start || extent_end < end) {
1203 btrfs_abort_transaction(trans, ret);
1207 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1208 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1209 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1210 memcpy(&new_key, &key, sizeof(new_key));
1212 if (start == key.offset && end < extent_end) {
1215 if (extent_mergeable(leaf, path->slots[0] - 1,
1216 ino, bytenr, orig_offset,
1217 &other_start, &other_end)) {
1218 new_key.offset = end;
1219 btrfs_set_item_key_safe(fs_info, path, &new_key);
1220 fi = btrfs_item_ptr(leaf, path->slots[0],
1221 struct btrfs_file_extent_item);
1222 btrfs_set_file_extent_generation(leaf, fi,
1224 btrfs_set_file_extent_num_bytes(leaf, fi,
1226 btrfs_set_file_extent_offset(leaf, fi,
1228 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1229 struct btrfs_file_extent_item);
1230 btrfs_set_file_extent_generation(leaf, fi,
1232 btrfs_set_file_extent_num_bytes(leaf, fi,
1234 btrfs_mark_buffer_dirty(leaf);
1239 if (start > key.offset && end == extent_end) {
1242 if (extent_mergeable(leaf, path->slots[0] + 1,
1243 ino, bytenr, orig_offset,
1244 &other_start, &other_end)) {
1245 fi = btrfs_item_ptr(leaf, path->slots[0],
1246 struct btrfs_file_extent_item);
1247 btrfs_set_file_extent_num_bytes(leaf, fi,
1248 start - key.offset);
1249 btrfs_set_file_extent_generation(leaf, fi,
1252 new_key.offset = start;
1253 btrfs_set_item_key_safe(fs_info, path, &new_key);
1255 fi = btrfs_item_ptr(leaf, path->slots[0],
1256 struct btrfs_file_extent_item);
1257 btrfs_set_file_extent_generation(leaf, fi,
1259 btrfs_set_file_extent_num_bytes(leaf, fi,
1261 btrfs_set_file_extent_offset(leaf, fi,
1262 start - orig_offset);
1263 btrfs_mark_buffer_dirty(leaf);
1268 while (start > key.offset || end < extent_end) {
1269 if (key.offset == start)
1272 new_key.offset = split;
1273 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1274 if (ret == -EAGAIN) {
1275 btrfs_release_path(path);
1279 btrfs_abort_transaction(trans, ret);
1283 leaf = path->nodes[0];
1284 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1285 struct btrfs_file_extent_item);
1286 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1287 btrfs_set_file_extent_num_bytes(leaf, fi,
1288 split - key.offset);
1290 fi = btrfs_item_ptr(leaf, path->slots[0],
1291 struct btrfs_file_extent_item);
1293 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1294 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1295 btrfs_set_file_extent_num_bytes(leaf, fi,
1296 extent_end - split);
1297 btrfs_mark_buffer_dirty(leaf);
1299 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes,
1300 0, root->root_key.objectid,
1303 btrfs_abort_transaction(trans, ret);
1307 if (split == start) {
1310 if (start != key.offset) {
1312 btrfs_abort_transaction(trans, ret);
1323 if (extent_mergeable(leaf, path->slots[0] + 1,
1324 ino, bytenr, orig_offset,
1325 &other_start, &other_end)) {
1327 btrfs_release_path(path);
1330 extent_end = other_end;
1331 del_slot = path->slots[0] + 1;
1333 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1334 0, root->root_key.objectid,
1337 btrfs_abort_transaction(trans, ret);
1343 if (extent_mergeable(leaf, path->slots[0] - 1,
1344 ino, bytenr, orig_offset,
1345 &other_start, &other_end)) {
1347 btrfs_release_path(path);
1350 key.offset = other_start;
1351 del_slot = path->slots[0];
1353 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1354 0, root->root_key.objectid,
1357 btrfs_abort_transaction(trans, ret);
1362 fi = btrfs_item_ptr(leaf, path->slots[0],
1363 struct btrfs_file_extent_item);
1364 btrfs_set_file_extent_type(leaf, fi,
1365 BTRFS_FILE_EXTENT_REG);
1366 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1367 btrfs_mark_buffer_dirty(leaf);
1369 fi = btrfs_item_ptr(leaf, del_slot - 1,
1370 struct btrfs_file_extent_item);
1371 btrfs_set_file_extent_type(leaf, fi,
1372 BTRFS_FILE_EXTENT_REG);
1373 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1374 btrfs_set_file_extent_num_bytes(leaf, fi,
1375 extent_end - key.offset);
1376 btrfs_mark_buffer_dirty(leaf);
1378 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1380 btrfs_abort_transaction(trans, ret);
1385 btrfs_free_path(path);
1390 * on error we return an unlocked page and the error value
1391 * on success we return a locked page and 0
1393 static int prepare_uptodate_page(struct inode *inode,
1394 struct page *page, u64 pos,
1395 bool force_uptodate)
1399 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1400 !PageUptodate(page)) {
1401 ret = btrfs_readpage(NULL, page);
1405 if (!PageUptodate(page)) {
1409 if (page->mapping != inode->i_mapping) {
1418 * this just gets pages into the page cache and locks them down.
1420 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1421 size_t num_pages, loff_t pos,
1422 size_t write_bytes, bool force_uptodate)
1425 unsigned long index = pos >> PAGE_SHIFT;
1426 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1430 for (i = 0; i < num_pages; i++) {
1432 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1433 mask | __GFP_WRITE);
1441 err = prepare_uptodate_page(inode, pages[i], pos,
1443 if (!err && i == num_pages - 1)
1444 err = prepare_uptodate_page(inode, pages[i],
1445 pos + write_bytes, false);
1448 if (err == -EAGAIN) {
1455 wait_on_page_writeback(pages[i]);
1460 while (faili >= 0) {
1461 unlock_page(pages[faili]);
1462 put_page(pages[faili]);
1470 * This function locks the extent and properly waits for data=ordered extents
1471 * to finish before allowing the pages to be modified if need.
1474 * 1 - the extent is locked
1475 * 0 - the extent is not locked, and everything is OK
1476 * -EAGAIN - need re-prepare the pages
1477 * the other < 0 number - Something wrong happens
1480 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1481 size_t num_pages, loff_t pos,
1483 u64 *lockstart, u64 *lockend,
1484 struct extent_state **cached_state)
1486 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1492 start_pos = round_down(pos, fs_info->sectorsize);
1493 last_pos = start_pos
1494 + round_up(pos + write_bytes - start_pos,
1495 fs_info->sectorsize) - 1;
1497 if (start_pos < inode->vfs_inode.i_size) {
1498 struct btrfs_ordered_extent *ordered;
1500 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1502 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1503 last_pos - start_pos + 1);
1505 ordered->file_offset + ordered->len > start_pos &&
1506 ordered->file_offset <= last_pos) {
1507 unlock_extent_cached(&inode->io_tree, start_pos,
1508 last_pos, cached_state);
1509 for (i = 0; i < num_pages; i++) {
1510 unlock_page(pages[i]);
1513 btrfs_start_ordered_extent(&inode->vfs_inode,
1515 btrfs_put_ordered_extent(ordered);
1519 btrfs_put_ordered_extent(ordered);
1520 clear_extent_bit(&inode->io_tree, start_pos, last_pos,
1521 EXTENT_DIRTY | EXTENT_DELALLOC |
1522 EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1523 0, 0, cached_state);
1524 *lockstart = start_pos;
1525 *lockend = last_pos;
1529 for (i = 0; i < num_pages; i++) {
1530 if (clear_page_dirty_for_io(pages[i]))
1531 account_page_redirty(pages[i]);
1532 set_page_extent_mapped(pages[i]);
1533 WARN_ON(!PageLocked(pages[i]));
1539 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1540 size_t *write_bytes)
1542 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1543 struct btrfs_root *root = inode->root;
1544 struct btrfs_ordered_extent *ordered;
1545 u64 lockstart, lockend;
1549 ret = btrfs_start_write_no_snapshotting(root);
1553 lockstart = round_down(pos, fs_info->sectorsize);
1554 lockend = round_up(pos + *write_bytes,
1555 fs_info->sectorsize) - 1;
1558 lock_extent(&inode->io_tree, lockstart, lockend);
1559 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1560 lockend - lockstart + 1);
1564 unlock_extent(&inode->io_tree, lockstart, lockend);
1565 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1566 btrfs_put_ordered_extent(ordered);
1569 num_bytes = lockend - lockstart + 1;
1570 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1574 btrfs_end_write_no_snapshotting(root);
1576 *write_bytes = min_t(size_t, *write_bytes ,
1577 num_bytes - pos + lockstart);
1580 unlock_extent(&inode->io_tree, lockstart, lockend);
1585 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1589 struct inode *inode = file_inode(file);
1590 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1591 struct btrfs_root *root = BTRFS_I(inode)->root;
1592 struct page **pages = NULL;
1593 struct extent_state *cached_state = NULL;
1594 struct extent_changeset *data_reserved = NULL;
1595 u64 release_bytes = 0;
1598 size_t num_written = 0;
1601 bool only_release_metadata = false;
1602 bool force_page_uptodate = false;
1604 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1605 PAGE_SIZE / (sizeof(struct page *)));
1606 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1607 nrptrs = max(nrptrs, 8);
1608 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1612 while (iov_iter_count(i) > 0) {
1613 size_t offset = pos & (PAGE_SIZE - 1);
1614 size_t sector_offset;
1615 size_t write_bytes = min(iov_iter_count(i),
1616 nrptrs * (size_t)PAGE_SIZE -
1618 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1620 size_t reserve_bytes;
1623 size_t dirty_sectors;
1627 WARN_ON(num_pages > nrptrs);
1630 * Fault pages before locking them in prepare_pages
1631 * to avoid recursive lock
1633 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1638 sector_offset = pos & (fs_info->sectorsize - 1);
1639 reserve_bytes = round_up(write_bytes + sector_offset,
1640 fs_info->sectorsize);
1642 extent_changeset_release(data_reserved);
1643 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1646 if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1647 BTRFS_INODE_PREALLOC)) &&
1648 check_can_nocow(BTRFS_I(inode), pos,
1649 &write_bytes) > 0) {
1651 * For nodata cow case, no need to reserve
1654 only_release_metadata = true;
1656 * our prealloc extent may be smaller than
1657 * write_bytes, so scale down.
1659 num_pages = DIV_ROUND_UP(write_bytes + offset,
1661 reserve_bytes = round_up(write_bytes +
1663 fs_info->sectorsize);
1669 WARN_ON(reserve_bytes == 0);
1670 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1673 if (!only_release_metadata)
1674 btrfs_free_reserved_data_space(inode,
1678 btrfs_end_write_no_snapshotting(root);
1682 release_bytes = reserve_bytes;
1685 * This is going to setup the pages array with the number of
1686 * pages we want, so we don't really need to worry about the
1687 * contents of pages from loop to loop
1689 ret = prepare_pages(inode, pages, num_pages,
1691 force_page_uptodate);
1693 btrfs_delalloc_release_extents(BTRFS_I(inode),
1694 reserve_bytes, true);
1698 extents_locked = lock_and_cleanup_extent_if_need(
1699 BTRFS_I(inode), pages,
1700 num_pages, pos, write_bytes, &lockstart,
1701 &lockend, &cached_state);
1702 if (extents_locked < 0) {
1703 if (extents_locked == -EAGAIN)
1705 btrfs_delalloc_release_extents(BTRFS_I(inode),
1706 reserve_bytes, true);
1707 ret = extents_locked;
1711 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1713 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1714 dirty_sectors = round_up(copied + sector_offset,
1715 fs_info->sectorsize);
1716 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1719 * if we have trouble faulting in the pages, fall
1720 * back to one page at a time
1722 if (copied < write_bytes)
1726 force_page_uptodate = true;
1730 force_page_uptodate = false;
1731 dirty_pages = DIV_ROUND_UP(copied + offset,
1735 if (num_sectors > dirty_sectors) {
1736 /* release everything except the sectors we dirtied */
1737 release_bytes -= dirty_sectors <<
1738 fs_info->sb->s_blocksize_bits;
1739 if (only_release_metadata) {
1740 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1741 release_bytes, true);
1745 __pos = round_down(pos,
1746 fs_info->sectorsize) +
1747 (dirty_pages << PAGE_SHIFT);
1748 btrfs_delalloc_release_space(inode,
1749 data_reserved, __pos,
1750 release_bytes, true);
1754 release_bytes = round_up(copied + sector_offset,
1755 fs_info->sectorsize);
1758 ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1759 pos, copied, &cached_state);
1761 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1762 lockstart, lockend, &cached_state);
1763 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes,
1766 btrfs_drop_pages(pages, num_pages);
1771 if (only_release_metadata)
1772 btrfs_end_write_no_snapshotting(root);
1774 if (only_release_metadata && copied > 0) {
1775 lockstart = round_down(pos,
1776 fs_info->sectorsize);
1777 lockend = round_up(pos + copied,
1778 fs_info->sectorsize) - 1;
1780 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1781 lockend, EXTENT_NORESERVE, NULL,
1783 only_release_metadata = false;
1786 btrfs_drop_pages(pages, num_pages);
1790 balance_dirty_pages_ratelimited(inode->i_mapping);
1791 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1792 btrfs_btree_balance_dirty(fs_info);
1795 num_written += copied;
1800 if (release_bytes) {
1801 if (only_release_metadata) {
1802 btrfs_end_write_no_snapshotting(root);
1803 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1804 release_bytes, true);
1806 btrfs_delalloc_release_space(inode, data_reserved,
1807 round_down(pos, fs_info->sectorsize),
1808 release_bytes, true);
1812 extent_changeset_free(data_reserved);
1813 return num_written ? num_written : ret;
1816 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1818 struct file *file = iocb->ki_filp;
1819 struct inode *inode = file_inode(file);
1820 loff_t pos = iocb->ki_pos;
1822 ssize_t written_buffered;
1826 written = generic_file_direct_write(iocb, from);
1828 if (written < 0 || !iov_iter_count(from))
1832 written_buffered = __btrfs_buffered_write(file, from, pos);
1833 if (written_buffered < 0) {
1834 err = written_buffered;
1838 * Ensure all data is persisted. We want the next direct IO read to be
1839 * able to read what was just written.
1841 endbyte = pos + written_buffered - 1;
1842 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1845 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1848 written += written_buffered;
1849 iocb->ki_pos = pos + written_buffered;
1850 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1851 endbyte >> PAGE_SHIFT);
1853 return written ? written : err;
1856 static void update_time_for_write(struct inode *inode)
1858 struct timespec now;
1860 if (IS_NOCMTIME(inode))
1863 now = current_time(inode);
1864 if (!timespec_equal(&inode->i_mtime, &now))
1865 inode->i_mtime = now;
1867 if (!timespec_equal(&inode->i_ctime, &now))
1868 inode->i_ctime = now;
1870 if (IS_I_VERSION(inode))
1871 inode_inc_iversion(inode);
1874 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1875 struct iov_iter *from)
1877 struct file *file = iocb->ki_filp;
1878 struct inode *inode = file_inode(file);
1879 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1880 struct btrfs_root *root = BTRFS_I(inode)->root;
1883 ssize_t num_written = 0;
1884 bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1887 size_t count = iov_iter_count(from);
1891 if (!(iocb->ki_flags & IOCB_DIRECT) &&
1892 (iocb->ki_flags & IOCB_NOWAIT))
1895 if (!inode_trylock(inode)) {
1896 if (iocb->ki_flags & IOCB_NOWAIT)
1901 err = generic_write_checks(iocb, from);
1903 inode_unlock(inode);
1908 if (iocb->ki_flags & IOCB_NOWAIT) {
1910 * We will allocate space in case nodatacow is not set,
1913 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1914 BTRFS_INODE_PREALLOC)) ||
1915 check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1916 inode_unlock(inode);
1921 current->backing_dev_info = inode_to_bdi(inode);
1922 err = file_remove_privs(file);
1924 inode_unlock(inode);
1929 * If BTRFS flips readonly due to some impossible error
1930 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1931 * although we have opened a file as writable, we have
1932 * to stop this write operation to ensure FS consistency.
1934 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1935 inode_unlock(inode);
1941 * We reserve space for updating the inode when we reserve space for the
1942 * extent we are going to write, so we will enospc out there. We don't
1943 * need to start yet another transaction to update the inode as we will
1944 * update the inode when we finish writing whatever data we write.
1946 update_time_for_write(inode);
1948 start_pos = round_down(pos, fs_info->sectorsize);
1949 oldsize = i_size_read(inode);
1950 if (start_pos > oldsize) {
1951 /* Expand hole size to cover write data, preventing empty gap */
1952 end_pos = round_up(pos + count,
1953 fs_info->sectorsize);
1954 err = btrfs_cont_expand(inode, oldsize, end_pos);
1956 inode_unlock(inode);
1959 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1964 atomic_inc(&BTRFS_I(inode)->sync_writers);
1966 if (iocb->ki_flags & IOCB_DIRECT) {
1967 num_written = __btrfs_direct_write(iocb, from);
1969 num_written = __btrfs_buffered_write(file, from, pos);
1970 if (num_written > 0)
1971 iocb->ki_pos = pos + num_written;
1973 pagecache_isize_extended(inode, oldsize,
1974 i_size_read(inode));
1977 inode_unlock(inode);
1980 * We also have to set last_sub_trans to the current log transid,
1981 * otherwise subsequent syncs to a file that's been synced in this
1982 * transaction will appear to have already occurred.
1984 spin_lock(&BTRFS_I(inode)->lock);
1985 BTRFS_I(inode)->last_sub_trans = root->log_transid;
1986 spin_unlock(&BTRFS_I(inode)->lock);
1987 if (num_written > 0)
1988 num_written = generic_write_sync(iocb, num_written);
1991 atomic_dec(&BTRFS_I(inode)->sync_writers);
1993 current->backing_dev_info = NULL;
1994 return num_written ? num_written : err;
1997 int btrfs_release_file(struct inode *inode, struct file *filp)
1999 struct btrfs_file_private *private = filp->private_data;
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 * We use start here because we will need to wait on the IO to complete
2193 * in btrfs_sync_log, which could require joining a transaction (for
2194 * example checking cross references in the nocow path). If we use join
2195 * here we could get into a situation where we're waiting on IO to
2196 * happen that is blocked on a transaction trying to commit. With start
2197 * we inc the extwriter counter, so we wait for all extwriters to exit
2198 * before we start blocking join'ers. This comment is to keep somebody
2199 * from thinking they are super smart and changing this to
2200 * btrfs_join_transaction *cough*Josef*cough*.
2202 trans = btrfs_start_transaction(root, 0);
2203 if (IS_ERR(trans)) {
2204 ret = PTR_ERR(trans);
2205 inode_unlock(inode);
2210 ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2212 /* Fallthrough and commit/free transaction. */
2216 /* we've logged all the items and now have a consistent
2217 * version of the file in the log. It is possible that
2218 * someone will come in and modify the file, but that's
2219 * fine because the log is consistent on disk, and we
2220 * have references to all of the file's extents
2222 * It is possible that someone will come in and log the
2223 * file again, but that will end up using the synchronization
2224 * inside btrfs_sync_log to keep things safe.
2226 inode_unlock(inode);
2229 * If any of the ordered extents had an error, just return it to user
2230 * space, so that the application knows some writes didn't succeed and
2231 * can take proper action (retry for e.g.). Blindly committing the
2232 * transaction in this case, would fool userspace that everything was
2233 * successful. And we also want to make sure our log doesn't contain
2234 * file extent items pointing to extents that weren't fully written to -
2235 * just like in the non fast fsync path, where we check for the ordered
2236 * operation's error flag before writing to the log tree and return -EIO
2237 * if any of them had this flag set (btrfs_wait_ordered_range) -
2238 * therefore we need to check for errors in the ordered operations,
2239 * which are indicated by ctx.io_err.
2242 btrfs_end_transaction(trans);
2247 if (ret != BTRFS_NO_LOG_SYNC) {
2249 ret = btrfs_sync_log(trans, root, &ctx);
2251 ret = btrfs_end_transaction(trans);
2256 ret = btrfs_wait_ordered_range(inode, start, len);
2258 btrfs_end_transaction(trans);
2262 ret = btrfs_commit_transaction(trans);
2264 ret = btrfs_end_transaction(trans);
2267 ASSERT(list_empty(&ctx.list));
2268 err = file_check_and_advance_wb_err(file);
2271 return ret > 0 ? -EIO : ret;
2274 static const struct vm_operations_struct btrfs_file_vm_ops = {
2275 .fault = filemap_fault,
2276 .map_pages = filemap_map_pages,
2277 .page_mkwrite = btrfs_page_mkwrite,
2280 static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2282 struct address_space *mapping = filp->f_mapping;
2284 if (!mapping->a_ops->readpage)
2287 file_accessed(filp);
2288 vma->vm_ops = &btrfs_file_vm_ops;
2293 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2294 int slot, u64 start, u64 end)
2296 struct btrfs_file_extent_item *fi;
2297 struct btrfs_key key;
2299 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2302 btrfs_item_key_to_cpu(leaf, &key, slot);
2303 if (key.objectid != btrfs_ino(inode) ||
2304 key.type != BTRFS_EXTENT_DATA_KEY)
2307 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2309 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2312 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2315 if (key.offset == end)
2317 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2322 static int fill_holes(struct btrfs_trans_handle *trans,
2323 struct btrfs_inode *inode,
2324 struct btrfs_path *path, u64 offset, u64 end)
2326 struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
2327 struct btrfs_root *root = inode->root;
2328 struct extent_buffer *leaf;
2329 struct btrfs_file_extent_item *fi;
2330 struct extent_map *hole_em;
2331 struct extent_map_tree *em_tree = &inode->extent_tree;
2332 struct btrfs_key key;
2335 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2338 key.objectid = btrfs_ino(inode);
2339 key.type = BTRFS_EXTENT_DATA_KEY;
2340 key.offset = offset;
2342 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2345 * We should have dropped this offset, so if we find it then
2346 * something has gone horribly wrong.
2353 leaf = path->nodes[0];
2354 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2358 fi = btrfs_item_ptr(leaf, path->slots[0],
2359 struct btrfs_file_extent_item);
2360 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2362 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2363 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2364 btrfs_set_file_extent_offset(leaf, fi, 0);
2365 btrfs_mark_buffer_dirty(leaf);
2369 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2372 key.offset = offset;
2373 btrfs_set_item_key_safe(fs_info, path, &key);
2374 fi = btrfs_item_ptr(leaf, path->slots[0],
2375 struct btrfs_file_extent_item);
2376 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2378 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2379 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2380 btrfs_set_file_extent_offset(leaf, fi, 0);
2381 btrfs_mark_buffer_dirty(leaf);
2384 btrfs_release_path(path);
2386 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2387 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2392 btrfs_release_path(path);
2394 hole_em = alloc_extent_map();
2396 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2397 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2399 hole_em->start = offset;
2400 hole_em->len = end - offset;
2401 hole_em->ram_bytes = hole_em->len;
2402 hole_em->orig_start = offset;
2404 hole_em->block_start = EXTENT_MAP_HOLE;
2405 hole_em->block_len = 0;
2406 hole_em->orig_block_len = 0;
2407 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2408 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2409 hole_em->generation = trans->transid;
2412 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2413 write_lock(&em_tree->lock);
2414 ret = add_extent_mapping(em_tree, hole_em, 1);
2415 write_unlock(&em_tree->lock);
2416 } while (ret == -EEXIST);
2417 free_extent_map(hole_em);
2419 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2420 &inode->runtime_flags);
2427 * Find a hole extent on given inode and change start/len to the end of hole
2428 * extent.(hole/vacuum extent whose em->start <= start &&
2429 * em->start + em->len > start)
2430 * When a hole extent is found, return 1 and modify start/len.
2432 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2434 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2435 struct extent_map *em;
2438 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2439 round_down(*start, fs_info->sectorsize),
2440 round_up(*len, fs_info->sectorsize), 0);
2444 /* Hole or vacuum extent(only exists in no-hole mode) */
2445 if (em->block_start == EXTENT_MAP_HOLE) {
2447 *len = em->start + em->len > *start + *len ?
2448 0 : *start + *len - em->start - em->len;
2449 *start = em->start + em->len;
2451 free_extent_map(em);
2455 static int btrfs_punch_hole_lock_range(struct inode *inode,
2456 const u64 lockstart,
2458 struct extent_state **cached_state)
2461 struct btrfs_ordered_extent *ordered;
2464 truncate_pagecache_range(inode, lockstart, lockend);
2466 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2468 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2471 * We need to make sure we have no ordered extents in this range
2472 * and nobody raced in and read a page in this range, if we did
2473 * we need to try again.
2476 (ordered->file_offset + ordered->len <= lockstart ||
2477 ordered->file_offset > lockend)) &&
2478 !filemap_range_has_page(inode->i_mapping,
2479 lockstart, lockend)) {
2481 btrfs_put_ordered_extent(ordered);
2485 btrfs_put_ordered_extent(ordered);
2486 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2487 lockend, cached_state);
2488 ret = btrfs_wait_ordered_range(inode, lockstart,
2489 lockend - lockstart + 1);
2496 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2498 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2499 struct btrfs_root *root = BTRFS_I(inode)->root;
2500 struct extent_state *cached_state = NULL;
2501 struct btrfs_path *path;
2502 struct btrfs_block_rsv *rsv;
2503 struct btrfs_trans_handle *trans;
2508 u64 orig_start = offset;
2510 u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2514 unsigned int rsv_count;
2516 bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2518 bool truncated_block = false;
2519 bool updated_inode = false;
2521 ret = btrfs_wait_ordered_range(inode, offset, len);
2526 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2527 ret = find_first_non_hole(inode, &offset, &len);
2529 goto out_only_mutex;
2531 /* Already in a large hole */
2533 goto out_only_mutex;
2536 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2537 lockend = round_down(offset + len,
2538 btrfs_inode_sectorsize(inode)) - 1;
2539 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2540 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2542 * We needn't truncate any block which is beyond the end of the file
2543 * because we are sure there is no data there.
2546 * Only do this if we are in the same block and we aren't doing the
2549 if (same_block && len < fs_info->sectorsize) {
2550 if (offset < ino_size) {
2551 truncated_block = true;
2552 ret = btrfs_truncate_block(inode, offset, len, 0);
2556 goto out_only_mutex;
2559 /* zero back part of the first block */
2560 if (offset < ino_size) {
2561 truncated_block = true;
2562 ret = btrfs_truncate_block(inode, offset, 0, 0);
2564 inode_unlock(inode);
2569 /* Check the aligned pages after the first unaligned page,
2570 * if offset != orig_start, which means the first unaligned page
2571 * including several following pages are already in holes,
2572 * the extra check can be skipped */
2573 if (offset == orig_start) {
2574 /* after truncate page, check hole again */
2575 len = offset + len - lockstart;
2577 ret = find_first_non_hole(inode, &offset, &len);
2579 goto out_only_mutex;
2582 goto out_only_mutex;
2587 /* Check the tail unaligned part is in a hole */
2588 tail_start = lockend + 1;
2589 tail_len = offset + len - tail_start;
2591 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2592 if (unlikely(ret < 0))
2593 goto out_only_mutex;
2595 /* zero the front end of the last page */
2596 if (tail_start + tail_len < ino_size) {
2597 truncated_block = true;
2598 ret = btrfs_truncate_block(inode,
2599 tail_start + tail_len,
2602 goto out_only_mutex;
2607 if (lockend < lockstart) {
2609 goto out_only_mutex;
2612 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2615 inode_unlock(inode);
2616 goto out_only_mutex;
2619 path = btrfs_alloc_path();
2625 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2630 rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2634 * 1 - update the inode
2635 * 1 - removing the extents in the range
2636 * 1 - adding the hole extent if no_holes isn't set
2638 rsv_count = no_holes ? 2 : 3;
2639 trans = btrfs_start_transaction(root, rsv_count);
2640 if (IS_ERR(trans)) {
2641 err = PTR_ERR(trans);
2645 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2648 trans->block_rsv = rsv;
2650 cur_offset = lockstart;
2651 len = lockend - cur_offset;
2652 while (cur_offset < lockend) {
2653 ret = __btrfs_drop_extents(trans, root, inode, path,
2654 cur_offset, lockend + 1,
2655 &drop_end, 1, 0, 0, NULL);
2659 trans->block_rsv = &fs_info->trans_block_rsv;
2661 if (cur_offset < drop_end && cur_offset < ino_size) {
2662 ret = fill_holes(trans, BTRFS_I(inode), path,
2663 cur_offset, drop_end);
2666 * If we failed then we didn't insert our hole
2667 * entries for the area we dropped, so now the
2668 * fs is corrupted, so we must abort the
2671 btrfs_abort_transaction(trans, ret);
2677 cur_offset = drop_end;
2679 ret = btrfs_update_inode(trans, root, inode);
2685 btrfs_end_transaction(trans);
2686 btrfs_btree_balance_dirty(fs_info);
2688 trans = btrfs_start_transaction(root, rsv_count);
2689 if (IS_ERR(trans)) {
2690 ret = PTR_ERR(trans);
2695 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2697 BUG_ON(ret); /* shouldn't happen */
2698 trans->block_rsv = rsv;
2700 ret = find_first_non_hole(inode, &cur_offset, &len);
2701 if (unlikely(ret < 0))
2714 trans->block_rsv = &fs_info->trans_block_rsv;
2716 * If we are using the NO_HOLES feature we might have had already an
2717 * hole that overlaps a part of the region [lockstart, lockend] and
2718 * ends at (or beyond) lockend. Since we have no file extent items to
2719 * represent holes, drop_end can be less than lockend and so we must
2720 * make sure we have an extent map representing the existing hole (the
2721 * call to __btrfs_drop_extents() might have dropped the existing extent
2722 * map representing the existing hole), otherwise the fast fsync path
2723 * will not record the existence of the hole region
2724 * [existing_hole_start, lockend].
2726 if (drop_end <= lockend)
2727 drop_end = lockend + 1;
2729 * Don't insert file hole extent item if it's for a range beyond eof
2730 * (because it's useless) or if it represents a 0 bytes range (when
2731 * cur_offset == drop_end).
2733 if (cur_offset < ino_size && cur_offset < drop_end) {
2734 ret = fill_holes(trans, BTRFS_I(inode), path,
2735 cur_offset, drop_end);
2737 /* Same comment as above. */
2738 btrfs_abort_transaction(trans, ret);
2748 inode_inc_iversion(inode);
2749 inode->i_mtime = inode->i_ctime = current_time(inode);
2751 trans->block_rsv = &fs_info->trans_block_rsv;
2752 ret = btrfs_update_inode(trans, root, inode);
2753 updated_inode = true;
2754 btrfs_end_transaction(trans);
2755 btrfs_btree_balance_dirty(fs_info);
2757 btrfs_free_path(path);
2758 btrfs_free_block_rsv(fs_info, rsv);
2760 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2763 if (!updated_inode && truncated_block && !ret && !err) {
2765 * If we only end up zeroing part of a page, we still need to
2766 * update the inode item, so that all the time fields are
2767 * updated as well as the necessary btrfs inode in memory fields
2768 * for detecting, at fsync time, if the inode isn't yet in the
2769 * log tree or it's there but not up to date.
2771 trans = btrfs_start_transaction(root, 1);
2772 if (IS_ERR(trans)) {
2773 err = PTR_ERR(trans);
2775 err = btrfs_update_inode(trans, root, inode);
2776 ret = btrfs_end_transaction(trans);
2779 inode_unlock(inode);
2785 /* Helper structure to record which range is already reserved */
2786 struct falloc_range {
2787 struct list_head list;
2793 * Helper function to add falloc range
2795 * Caller should have locked the larger range of extent containing
2798 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2800 struct falloc_range *prev = NULL;
2801 struct falloc_range *range = NULL;
2803 if (list_empty(head))
2807 * As fallocate iterate by bytenr order, we only need to check
2810 prev = list_entry(head->prev, struct falloc_range, list);
2811 if (prev->start + prev->len == start) {
2816 range = kmalloc(sizeof(*range), GFP_KERNEL);
2819 range->start = start;
2821 list_add_tail(&range->list, head);
2825 static int btrfs_fallocate_update_isize(struct inode *inode,
2829 struct btrfs_trans_handle *trans;
2830 struct btrfs_root *root = BTRFS_I(inode)->root;
2834 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2837 trans = btrfs_start_transaction(root, 1);
2839 return PTR_ERR(trans);
2841 inode->i_ctime = current_time(inode);
2842 i_size_write(inode, end);
2843 btrfs_ordered_update_i_size(inode, end, NULL);
2844 ret = btrfs_update_inode(trans, root, inode);
2845 ret2 = btrfs_end_transaction(trans);
2847 return ret ? ret : ret2;
2851 RANGE_BOUNDARY_WRITTEN_EXTENT = 0,
2852 RANGE_BOUNDARY_PREALLOC_EXTENT = 1,
2853 RANGE_BOUNDARY_HOLE = 2,
2856 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2859 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2860 struct extent_map *em;
2863 offset = round_down(offset, sectorsize);
2864 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2868 if (em->block_start == EXTENT_MAP_HOLE)
2869 ret = RANGE_BOUNDARY_HOLE;
2870 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2871 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2873 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2875 free_extent_map(em);
2879 static int btrfs_zero_range(struct inode *inode,
2884 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2885 struct extent_map *em;
2886 struct extent_changeset *data_reserved = NULL;
2889 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2890 u64 alloc_start = round_down(offset, sectorsize);
2891 u64 alloc_end = round_up(offset + len, sectorsize);
2892 u64 bytes_to_reserve = 0;
2893 bool space_reserved = false;
2895 inode_dio_wait(inode);
2897 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2898 alloc_start, alloc_end - alloc_start, 0);
2905 * Avoid hole punching and extent allocation for some cases. More cases
2906 * could be considered, but these are unlikely common and we keep things
2907 * as simple as possible for now. Also, intentionally, if the target
2908 * range contains one or more prealloc extents together with regular
2909 * extents and holes, we drop all the existing extents and allocate a
2910 * new prealloc extent, so that we get a larger contiguous disk extent.
2912 if (em->start <= alloc_start &&
2913 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2914 const u64 em_end = em->start + em->len;
2916 if (em_end >= offset + len) {
2918 * The whole range is already a prealloc extent,
2919 * do nothing except updating the inode's i_size if
2922 free_extent_map(em);
2923 ret = btrfs_fallocate_update_isize(inode, offset + len,
2928 * Part of the range is already a prealloc extent, so operate
2929 * only on the remaining part of the range.
2931 alloc_start = em_end;
2932 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2933 len = offset + len - alloc_start;
2934 offset = alloc_start;
2935 alloc_hint = em->block_start + em->len;
2937 free_extent_map(em);
2939 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2940 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2941 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2942 alloc_start, sectorsize, 0);
2948 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2949 free_extent_map(em);
2950 ret = btrfs_fallocate_update_isize(inode, offset + len,
2954 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2955 free_extent_map(em);
2956 ret = btrfs_truncate_block(inode, offset, len, 0);
2958 ret = btrfs_fallocate_update_isize(inode,
2963 free_extent_map(em);
2964 alloc_start = round_down(offset, sectorsize);
2965 alloc_end = alloc_start + sectorsize;
2969 alloc_start = round_up(offset, sectorsize);
2970 alloc_end = round_down(offset + len, sectorsize);
2973 * For unaligned ranges, check the pages at the boundaries, they might
2974 * map to an extent, in which case we need to partially zero them, or
2975 * they might map to a hole, in which case we need our allocation range
2978 if (!IS_ALIGNED(offset, sectorsize)) {
2979 ret = btrfs_zero_range_check_range_boundary(inode, offset);
2982 if (ret == RANGE_BOUNDARY_HOLE) {
2983 alloc_start = round_down(offset, sectorsize);
2985 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2986 ret = btrfs_truncate_block(inode, offset, 0, 0);
2994 if (!IS_ALIGNED(offset + len, sectorsize)) {
2995 ret = btrfs_zero_range_check_range_boundary(inode,
2999 if (ret == RANGE_BOUNDARY_HOLE) {
3000 alloc_end = round_up(offset + len, sectorsize);
3002 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3003 ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3012 if (alloc_start < alloc_end) {
3013 struct extent_state *cached_state = NULL;
3014 const u64 lockstart = alloc_start;
3015 const u64 lockend = alloc_end - 1;
3017 bytes_to_reserve = alloc_end - alloc_start;
3018 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3022 space_reserved = true;
3023 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3024 alloc_start, bytes_to_reserve);
3027 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3031 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3032 alloc_end - alloc_start,
3034 offset + len, &alloc_hint);
3035 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3036 lockend, &cached_state);
3037 /* btrfs_prealloc_file_range releases reserved space on error */
3039 space_reserved = false;
3043 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3045 if (ret && space_reserved)
3046 btrfs_free_reserved_data_space(inode, data_reserved,
3047 alloc_start, bytes_to_reserve);
3048 extent_changeset_free(data_reserved);
3053 static long btrfs_fallocate(struct file *file, int mode,
3054 loff_t offset, loff_t len)
3056 struct inode *inode = file_inode(file);
3057 struct extent_state *cached_state = NULL;
3058 struct extent_changeset *data_reserved = NULL;
3059 struct falloc_range *range;
3060 struct falloc_range *tmp;
3061 struct list_head reserve_list;
3069 struct extent_map *em;
3070 int blocksize = btrfs_inode_sectorsize(inode);
3073 alloc_start = round_down(offset, blocksize);
3074 alloc_end = round_up(offset + len, blocksize);
3075 cur_offset = alloc_start;
3077 /* Make sure we aren't being give some crap mode */
3078 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3079 FALLOC_FL_ZERO_RANGE))
3082 if (mode & FALLOC_FL_PUNCH_HOLE)
3083 return btrfs_punch_hole(inode, offset, len);
3086 * Only trigger disk allocation, don't trigger qgroup reserve
3088 * For qgroup space, it will be checked later.
3090 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3091 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3092 alloc_end - alloc_start);
3099 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3100 ret = inode_newsize_ok(inode, offset + len);
3106 * TODO: Move these two operations after we have checked
3107 * accurate reserved space, or fallocate can still fail but
3108 * with page truncated or size expanded.
3110 * But that's a minor problem and won't do much harm BTW.
3112 if (alloc_start > inode->i_size) {
3113 ret = btrfs_cont_expand(inode, i_size_read(inode),
3117 } else if (offset + len > inode->i_size) {
3119 * If we are fallocating from the end of the file onward we
3120 * need to zero out the end of the block if i_size lands in the
3121 * middle of a block.
3123 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3129 * wait for ordered IO before we have any locks. We'll loop again
3130 * below with the locks held.
3132 ret = btrfs_wait_ordered_range(inode, alloc_start,
3133 alloc_end - alloc_start);
3137 if (mode & FALLOC_FL_ZERO_RANGE) {
3138 ret = btrfs_zero_range(inode, offset, len, mode);
3139 inode_unlock(inode);
3143 locked_end = alloc_end - 1;
3145 struct btrfs_ordered_extent *ordered;
3147 /* the extent lock is ordered inside the running
3150 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3151 locked_end, &cached_state);
3152 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3155 ordered->file_offset + ordered->len > alloc_start &&
3156 ordered->file_offset < alloc_end) {
3157 btrfs_put_ordered_extent(ordered);
3158 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3159 alloc_start, locked_end,
3162 * we can't wait on the range with the transaction
3163 * running or with the extent lock held
3165 ret = btrfs_wait_ordered_range(inode, alloc_start,
3166 alloc_end - alloc_start);
3171 btrfs_put_ordered_extent(ordered);
3176 /* First, check if we exceed the qgroup limit */
3177 INIT_LIST_HEAD(&reserve_list);
3178 while (cur_offset < alloc_end) {
3179 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3180 alloc_end - cur_offset, 0);
3185 last_byte = min(extent_map_end(em), alloc_end);
3186 actual_end = min_t(u64, extent_map_end(em), offset + len);
3187 last_byte = ALIGN(last_byte, blocksize);
3188 if (em->block_start == EXTENT_MAP_HOLE ||
3189 (cur_offset >= inode->i_size &&
3190 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3191 ret = add_falloc_range(&reserve_list, cur_offset,
3192 last_byte - cur_offset);
3194 free_extent_map(em);
3197 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3198 cur_offset, last_byte - cur_offset);
3200 free_extent_map(em);
3205 * Do not need to reserve unwritten extent for this
3206 * range, free reserved data space first, otherwise
3207 * it'll result in false ENOSPC error.
3209 btrfs_free_reserved_data_space(inode, data_reserved,
3210 cur_offset, last_byte - cur_offset);
3212 free_extent_map(em);
3213 cur_offset = last_byte;
3217 * If ret is still 0, means we're OK to fallocate.
3218 * Or just cleanup the list and exit.
3220 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3222 ret = btrfs_prealloc_file_range(inode, mode,
3224 range->len, i_blocksize(inode),
3225 offset + len, &alloc_hint);
3227 btrfs_free_reserved_data_space(inode,
3228 data_reserved, range->start,
3230 list_del(&range->list);
3237 * We didn't need to allocate any more space, but we still extended the
3238 * size of the file so we need to update i_size and the inode item.
3240 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3242 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3245 inode_unlock(inode);
3246 /* Let go of our reservation. */
3247 if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3248 btrfs_free_reserved_data_space(inode, data_reserved,
3249 alloc_start, alloc_end - cur_offset);
3250 extent_changeset_free(data_reserved);
3254 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3256 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3257 struct extent_map *em = NULL;
3258 struct extent_state *cached_state = NULL;
3265 if (inode->i_size == 0)
3269 * *offset can be negative, in this case we start finding DATA/HOLE from
3270 * the very start of the file.
3272 start = max_t(loff_t, 0, *offset);
3274 lockstart = round_down(start, fs_info->sectorsize);
3275 lockend = round_up(i_size_read(inode),
3276 fs_info->sectorsize);
3277 if (lockend <= lockstart)
3278 lockend = lockstart + fs_info->sectorsize;
3280 len = lockend - lockstart + 1;
3282 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3285 while (start < inode->i_size) {
3286 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0,
3294 if (whence == SEEK_HOLE &&
3295 (em->block_start == EXTENT_MAP_HOLE ||
3296 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3298 else if (whence == SEEK_DATA &&
3299 (em->block_start != EXTENT_MAP_HOLE &&
3300 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3303 start = em->start + em->len;
3304 free_extent_map(em);
3308 free_extent_map(em);
3310 if (whence == SEEK_DATA && start >= inode->i_size)
3313 *offset = min_t(loff_t, start, inode->i_size);
3315 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3320 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3322 struct inode *inode = file->f_mapping->host;
3329 offset = generic_file_llseek(file, offset, whence);
3333 if (offset >= i_size_read(inode)) {
3334 inode_unlock(inode);
3338 ret = find_desired_extent(inode, &offset, whence);
3340 inode_unlock(inode);
3345 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3347 inode_unlock(inode);
3351 static int btrfs_file_open(struct inode *inode, struct file *filp)
3353 filp->f_mode |= FMODE_NOWAIT;
3354 return generic_file_open(inode, filp);
3357 const struct file_operations btrfs_file_operations = {
3358 .llseek = btrfs_file_llseek,
3359 .read_iter = generic_file_read_iter,
3360 .splice_read = generic_file_splice_read,
3361 .write_iter = btrfs_file_write_iter,
3362 .mmap = btrfs_file_mmap,
3363 .open = btrfs_file_open,
3364 .release = btrfs_release_file,
3365 .fsync = btrfs_sync_file,
3366 .fallocate = btrfs_fallocate,
3367 .unlocked_ioctl = btrfs_ioctl,
3368 #ifdef CONFIG_COMPAT
3369 .compat_ioctl = btrfs_compat_ioctl,
3371 .clone_file_range = btrfs_clone_file_range,
3372 .dedupe_file_range = btrfs_dedupe_file_range,
3375 void __cold btrfs_auto_defrag_exit(void)
3377 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3380 int __init btrfs_auto_defrag_init(void)
3382 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3383 sizeof(struct inode_defrag), 0,
3386 if (!btrfs_inode_defrag_cachep)
3392 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3397 * So with compression we will find and lock a dirty page and clear the
3398 * first one as dirty, setup an async extent, and immediately return
3399 * with the entire range locked but with nobody actually marked with
3400 * writeback. So we can't just filemap_write_and_wait_range() and
3401 * expect it to work since it will just kick off a thread to do the
3402 * actual work. So we need to call filemap_fdatawrite_range _again_
3403 * since it will wait on the page lock, which won't be unlocked until
3404 * after the pages have been marked as writeback and so we're good to go
3405 * from there. We have to do this otherwise we'll miss the ordered
3406 * extents and that results in badness. Please Josef, do not think you
3407 * know better and pull this out at some point in the future, it is
3408 * right and you are wrong.
3410 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3411 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3412 &BTRFS_I(inode)->runtime_flags))
3413 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);