1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2007 Oracle. All rights reserved.
7 #include <linux/pagemap.h>
8 #include <linux/time.h>
9 #include <linux/init.h>
10 #include <linux/string.h>
11 #include <linux/backing-dev.h>
12 #include <linux/falloc.h>
13 #include <linux/writeback.h>
14 #include <linux/compat.h>
15 #include <linux/slab.h>
16 #include <linux/btrfs.h>
17 #include <linux/uio.h>
18 #include <linux/iversion.h>
21 #include "transaction.h"
22 #include "btrfs_inode.h"
23 #include "print-tree.h"
28 #include "compression.h"
30 static struct kmem_cache *btrfs_inode_defrag_cachep;
32 * when auto defrag is enabled we
33 * queue up these defrag structs to remember which
34 * inodes need defragging passes
37 struct rb_node rb_node;
41 * transid where the defrag was added, we search for
42 * extents newer than this
49 /* last offset we were able to defrag */
52 /* if we've wrapped around back to zero once already */
56 static int __compare_inode_defrag(struct inode_defrag *defrag1,
57 struct inode_defrag *defrag2)
59 if (defrag1->root > defrag2->root)
61 else if (defrag1->root < defrag2->root)
63 else if (defrag1->ino > defrag2->ino)
65 else if (defrag1->ino < defrag2->ino)
71 /* pop a record for an inode into the defrag tree. The lock
72 * must be held already
74 * If you're inserting a record for an older transid than an
75 * existing record, the transid already in the tree is lowered
77 * If an existing record is found the defrag item you
80 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
81 struct inode_defrag *defrag)
83 struct btrfs_fs_info *fs_info = inode->root->fs_info;
84 struct inode_defrag *entry;
86 struct rb_node *parent = NULL;
89 p = &fs_info->defrag_inodes.rb_node;
92 entry = rb_entry(parent, struct inode_defrag, rb_node);
94 ret = __compare_inode_defrag(defrag, entry);
98 p = &parent->rb_right;
100 /* if we're reinserting an entry for
101 * an old defrag run, make sure to
102 * lower the transid of our existing record
104 if (defrag->transid < entry->transid)
105 entry->transid = defrag->transid;
106 if (defrag->last_offset > entry->last_offset)
107 entry->last_offset = defrag->last_offset;
111 set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
112 rb_link_node(&defrag->rb_node, parent, p);
113 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
117 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
119 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
122 if (btrfs_fs_closing(fs_info))
129 * insert a defrag record for this inode if auto defrag is
132 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
133 struct btrfs_inode *inode)
135 struct btrfs_root *root = inode->root;
136 struct btrfs_fs_info *fs_info = root->fs_info;
137 struct inode_defrag *defrag;
141 if (!__need_auto_defrag(fs_info))
144 if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
148 transid = trans->transid;
150 transid = inode->root->last_trans;
152 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
156 defrag->ino = btrfs_ino(inode);
157 defrag->transid = transid;
158 defrag->root = root->root_key.objectid;
160 spin_lock(&fs_info->defrag_inodes_lock);
161 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
163 * If we set IN_DEFRAG flag and evict the inode from memory,
164 * and then re-read this inode, this new inode doesn't have
165 * IN_DEFRAG flag. At the case, we may find the existed defrag.
167 ret = __btrfs_add_inode_defrag(inode, defrag);
169 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
171 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
173 spin_unlock(&fs_info->defrag_inodes_lock);
178 * Requeue the defrag object. If there is a defrag object that points to
179 * the same inode in the tree, we will merge them together (by
180 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
182 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
183 struct inode_defrag *defrag)
185 struct btrfs_fs_info *fs_info = inode->root->fs_info;
188 if (!__need_auto_defrag(fs_info))
192 * Here we don't check the IN_DEFRAG flag, because we need merge
195 spin_lock(&fs_info->defrag_inodes_lock);
196 ret = __btrfs_add_inode_defrag(inode, defrag);
197 spin_unlock(&fs_info->defrag_inodes_lock);
202 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
206 * pick the defragable inode that we want, if it doesn't exist, we will get
209 static struct inode_defrag *
210 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
212 struct inode_defrag *entry = NULL;
213 struct inode_defrag tmp;
215 struct rb_node *parent = NULL;
221 spin_lock(&fs_info->defrag_inodes_lock);
222 p = fs_info->defrag_inodes.rb_node;
225 entry = rb_entry(parent, struct inode_defrag, rb_node);
227 ret = __compare_inode_defrag(&tmp, entry);
231 p = parent->rb_right;
236 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
237 parent = rb_next(parent);
239 entry = rb_entry(parent, struct inode_defrag, rb_node);
245 rb_erase(parent, &fs_info->defrag_inodes);
246 spin_unlock(&fs_info->defrag_inodes_lock);
250 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
252 struct inode_defrag *defrag;
253 struct rb_node *node;
255 spin_lock(&fs_info->defrag_inodes_lock);
256 node = rb_first(&fs_info->defrag_inodes);
258 rb_erase(node, &fs_info->defrag_inodes);
259 defrag = rb_entry(node, struct inode_defrag, rb_node);
260 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
262 cond_resched_lock(&fs_info->defrag_inodes_lock);
264 node = rb_first(&fs_info->defrag_inodes);
266 spin_unlock(&fs_info->defrag_inodes_lock);
269 #define BTRFS_DEFRAG_BATCH 1024
271 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
272 struct inode_defrag *defrag)
274 struct btrfs_root *inode_root;
276 struct btrfs_key key;
277 struct btrfs_ioctl_defrag_range_args range;
283 key.objectid = defrag->root;
284 key.type = BTRFS_ROOT_ITEM_KEY;
285 key.offset = (u64)-1;
287 index = srcu_read_lock(&fs_info->subvol_srcu);
289 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
290 if (IS_ERR(inode_root)) {
291 ret = PTR_ERR(inode_root);
295 key.objectid = defrag->ino;
296 key.type = BTRFS_INODE_ITEM_KEY;
298 inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
300 ret = PTR_ERR(inode);
303 srcu_read_unlock(&fs_info->subvol_srcu, index);
305 /* do a chunk of defrag */
306 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
307 memset(&range, 0, sizeof(range));
309 range.start = defrag->last_offset;
311 sb_start_write(fs_info->sb);
312 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
314 sb_end_write(fs_info->sb);
316 * if we filled the whole defrag batch, there
317 * must be more work to do. Queue this defrag
320 if (num_defrag == BTRFS_DEFRAG_BATCH) {
321 defrag->last_offset = range.start;
322 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
323 } else if (defrag->last_offset && !defrag->cycled) {
325 * we didn't fill our defrag batch, but
326 * we didn't start at zero. Make sure we loop
327 * around to the start of the file.
329 defrag->last_offset = 0;
331 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
333 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
339 srcu_read_unlock(&fs_info->subvol_srcu, index);
340 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
345 * run through the list of inodes in the FS that need
348 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
350 struct inode_defrag *defrag;
352 u64 root_objectid = 0;
354 atomic_inc(&fs_info->defrag_running);
356 /* Pause the auto defragger. */
357 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
361 if (!__need_auto_defrag(fs_info))
364 /* find an inode to defrag */
365 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
368 if (root_objectid || first_ino) {
377 first_ino = defrag->ino + 1;
378 root_objectid = defrag->root;
380 __btrfs_run_defrag_inode(fs_info, defrag);
382 atomic_dec(&fs_info->defrag_running);
385 * during unmount, we use the transaction_wait queue to
386 * wait for the defragger to stop
388 wake_up(&fs_info->transaction_wait);
392 /* simple helper to fault in pages and copy. This should go away
393 * and be replaced with calls into generic code.
395 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
396 struct page **prepared_pages,
400 size_t total_copied = 0;
402 int offset = pos & (PAGE_SIZE - 1);
404 while (write_bytes > 0) {
405 size_t count = min_t(size_t,
406 PAGE_SIZE - offset, write_bytes);
407 struct page *page = prepared_pages[pg];
409 * Copy data from userspace to the current page
411 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
413 /* Flush processor's dcache for this page */
414 flush_dcache_page(page);
417 * if we get a partial write, we can end up with
418 * partially up to date pages. These add
419 * a lot of complexity, so make sure they don't
420 * happen by forcing this copy to be retried.
422 * The rest of the btrfs_file_write code will fall
423 * back to page at a time copies after we return 0.
425 if (!PageUptodate(page) && copied < count)
428 iov_iter_advance(i, copied);
429 write_bytes -= copied;
430 total_copied += copied;
432 /* Return to btrfs_file_write_iter to fault page */
433 if (unlikely(copied == 0))
436 if (copied < PAGE_SIZE - offset) {
447 * unlocks pages after btrfs_file_write is done with them
449 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
452 for (i = 0; i < num_pages; i++) {
453 /* page checked is some magic around finding pages that
454 * have been modified without going through btrfs_set_page_dirty
455 * clear it here. There should be no need to mark the pages
456 * accessed as prepare_pages should have marked them accessed
457 * in prepare_pages via find_or_create_page()
459 ClearPageChecked(pages[i]);
460 unlock_page(pages[i]);
465 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
468 struct extent_state **cached_state)
470 u64 search_start = start;
471 const u64 end = start + len - 1;
473 while (search_start < end) {
474 const u64 search_len = end - search_start + 1;
475 struct extent_map *em;
479 em = btrfs_get_extent(inode, NULL, 0, search_start,
484 if (em->block_start != EXTENT_MAP_HOLE)
488 if (em->start < search_start)
489 em_len -= search_start - em->start;
490 if (em_len > search_len)
493 ret = set_extent_bit(&inode->io_tree, search_start,
494 search_start + em_len - 1,
496 NULL, cached_state, GFP_NOFS);
498 search_start = extent_map_end(em);
507 * after copy_from_user, pages need to be dirtied and we need to make
508 * sure holes are created between the current EOF and the start of
509 * any next extents (if required).
511 * this also makes the decision about creating an inline extent vs
512 * doing real data extents, marking pages dirty and delalloc as required.
514 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
515 size_t num_pages, loff_t pos, size_t write_bytes,
516 struct extent_state **cached)
518 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
523 u64 end_of_last_block;
524 u64 end_pos = pos + write_bytes;
525 loff_t isize = i_size_read(inode);
526 unsigned int extra_bits = 0;
528 start_pos = pos & ~((u64) fs_info->sectorsize - 1);
529 num_bytes = round_up(write_bytes + pos - start_pos,
530 fs_info->sectorsize);
532 end_of_last_block = start_pos + num_bytes - 1;
535 * The pages may have already been dirty, clear out old accounting so
536 * we can set things up properly
538 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos, end_of_last_block,
539 EXTENT_DIRTY | EXTENT_DELALLOC |
540 EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 0, 0, cached);
542 if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
543 if (start_pos >= isize &&
544 !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
546 * There can't be any extents following eof in this case
547 * so just set the delalloc new bit for the range
550 extra_bits |= EXTENT_DELALLOC_NEW;
552 err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
560 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
561 extra_bits, cached, 0);
565 for (i = 0; i < num_pages; i++) {
566 struct page *p = pages[i];
573 * we've only changed i_size in ram, and we haven't updated
574 * the disk i_size. There is no need to log the inode
578 i_size_write(inode, end_pos);
583 * this drops all the extents in the cache that intersect the range
584 * [start, end]. Existing extents are split as required.
586 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
589 struct extent_map *em;
590 struct extent_map *split = NULL;
591 struct extent_map *split2 = NULL;
592 struct extent_map_tree *em_tree = &inode->extent_tree;
593 u64 len = end - start + 1;
601 WARN_ON(end < start);
602 if (end == (u64)-1) {
611 split = alloc_extent_map();
613 split2 = alloc_extent_map();
614 if (!split || !split2)
617 write_lock(&em_tree->lock);
618 em = lookup_extent_mapping(em_tree, start, len);
620 write_unlock(&em_tree->lock);
624 gen = em->generation;
625 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
626 if (testend && em->start + em->len >= start + len) {
628 write_unlock(&em_tree->lock);
631 start = em->start + em->len;
633 len = start + len - (em->start + em->len);
635 write_unlock(&em_tree->lock);
638 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
639 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
640 clear_bit(EXTENT_FLAG_LOGGING, &flags);
641 modified = !list_empty(&em->list);
645 if (em->start < start) {
646 split->start = em->start;
647 split->len = start - em->start;
649 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
650 split->orig_start = em->orig_start;
651 split->block_start = em->block_start;
654 split->block_len = em->block_len;
656 split->block_len = split->len;
657 split->orig_block_len = max(split->block_len,
659 split->ram_bytes = em->ram_bytes;
661 split->orig_start = split->start;
662 split->block_len = 0;
663 split->block_start = em->block_start;
664 split->orig_block_len = 0;
665 split->ram_bytes = split->len;
668 split->generation = gen;
669 split->bdev = em->bdev;
670 split->flags = flags;
671 split->compress_type = em->compress_type;
672 replace_extent_mapping(em_tree, em, split, modified);
673 free_extent_map(split);
677 if (testend && em->start + em->len > start + len) {
678 u64 diff = start + len - em->start;
680 split->start = start + len;
681 split->len = em->start + em->len - (start + len);
682 split->bdev = em->bdev;
683 split->flags = flags;
684 split->compress_type = em->compress_type;
685 split->generation = gen;
687 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
688 split->orig_block_len = max(em->block_len,
691 split->ram_bytes = em->ram_bytes;
693 split->block_len = em->block_len;
694 split->block_start = em->block_start;
695 split->orig_start = em->orig_start;
697 split->block_len = split->len;
698 split->block_start = em->block_start
700 split->orig_start = em->orig_start;
703 split->ram_bytes = split->len;
704 split->orig_start = split->start;
705 split->block_len = 0;
706 split->block_start = em->block_start;
707 split->orig_block_len = 0;
710 if (extent_map_in_tree(em)) {
711 replace_extent_mapping(em_tree, em, split,
714 ret = add_extent_mapping(em_tree, split,
716 ASSERT(ret == 0); /* Logic error */
718 free_extent_map(split);
722 if (extent_map_in_tree(em))
723 remove_extent_mapping(em_tree, em);
724 write_unlock(&em_tree->lock);
728 /* once for the tree*/
732 free_extent_map(split);
734 free_extent_map(split2);
738 * this is very complex, but the basic idea is to drop all extents
739 * in the range start - end. hint_block is filled in with a block number
740 * that would be a good hint to the block allocator for this file.
742 * If an extent intersects the range but is not entirely inside the range
743 * it is either truncated or split. Anything entirely inside the range
744 * is deleted from the tree.
746 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
747 struct btrfs_root *root, struct inode *inode,
748 struct btrfs_path *path, u64 start, u64 end,
749 u64 *drop_end, int drop_cache,
751 u32 extent_item_size,
754 struct btrfs_fs_info *fs_info = root->fs_info;
755 struct extent_buffer *leaf;
756 struct btrfs_file_extent_item *fi;
757 struct btrfs_key key;
758 struct btrfs_key new_key;
759 u64 ino = btrfs_ino(BTRFS_I(inode));
760 u64 search_start = start;
763 u64 extent_offset = 0;
765 u64 last_end = start;
771 int modify_tree = -1;
774 int leafs_visited = 0;
777 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
779 if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
782 update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
783 root == fs_info->tree_root);
786 ret = btrfs_lookup_file_extent(trans, root, path, ino,
787 search_start, modify_tree);
790 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
791 leaf = path->nodes[0];
792 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
793 if (key.objectid == ino &&
794 key.type == BTRFS_EXTENT_DATA_KEY)
800 leaf = path->nodes[0];
801 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
803 ret = btrfs_next_leaf(root, path);
811 leaf = path->nodes[0];
815 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
817 if (key.objectid > ino)
819 if (WARN_ON_ONCE(key.objectid < ino) ||
820 key.type < BTRFS_EXTENT_DATA_KEY) {
825 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
828 fi = btrfs_item_ptr(leaf, path->slots[0],
829 struct btrfs_file_extent_item);
830 extent_type = btrfs_file_extent_type(leaf, fi);
832 if (extent_type == BTRFS_FILE_EXTENT_REG ||
833 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
834 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
835 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
836 extent_offset = btrfs_file_extent_offset(leaf, fi);
837 extent_end = key.offset +
838 btrfs_file_extent_num_bytes(leaf, fi);
839 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
840 extent_end = key.offset +
841 btrfs_file_extent_ram_bytes(leaf, fi);
848 * Don't skip extent items representing 0 byte lengths. They
849 * used to be created (bug) if while punching holes we hit
850 * -ENOSPC condition. So if we find one here, just ensure we
851 * delete it, otherwise we would insert a new file extent item
852 * with the same key (offset) as that 0 bytes length file
853 * extent item in the call to setup_items_for_insert() later
856 if (extent_end == key.offset && extent_end >= search_start) {
857 last_end = extent_end;
858 goto delete_extent_item;
861 if (extent_end <= search_start) {
867 search_start = max(key.offset, start);
868 if (recow || !modify_tree) {
870 btrfs_release_path(path);
875 * | - range to drop - |
876 * | -------- extent -------- |
878 if (start > key.offset && end < extent_end) {
880 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
885 memcpy(&new_key, &key, sizeof(new_key));
886 new_key.offset = start;
887 ret = btrfs_duplicate_item(trans, root, path,
889 if (ret == -EAGAIN) {
890 btrfs_release_path(path);
896 leaf = path->nodes[0];
897 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
898 struct btrfs_file_extent_item);
899 btrfs_set_file_extent_num_bytes(leaf, fi,
902 fi = btrfs_item_ptr(leaf, path->slots[0],
903 struct btrfs_file_extent_item);
905 extent_offset += start - key.offset;
906 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
907 btrfs_set_file_extent_num_bytes(leaf, fi,
909 btrfs_mark_buffer_dirty(leaf);
911 if (update_refs && disk_bytenr > 0) {
912 ret = btrfs_inc_extent_ref(trans, root,
913 disk_bytenr, num_bytes, 0,
914 root->root_key.objectid,
916 start - extent_offset);
917 BUG_ON(ret); /* -ENOMEM */
922 * From here on out we will have actually dropped something, so
923 * last_end can be updated.
925 last_end = extent_end;
928 * | ---- range to drop ----- |
929 * | -------- extent -------- |
931 if (start <= key.offset && end < extent_end) {
932 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
937 memcpy(&new_key, &key, sizeof(new_key));
938 new_key.offset = end;
939 btrfs_set_item_key_safe(fs_info, path, &new_key);
941 extent_offset += end - key.offset;
942 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
943 btrfs_set_file_extent_num_bytes(leaf, fi,
945 btrfs_mark_buffer_dirty(leaf);
946 if (update_refs && disk_bytenr > 0)
947 inode_sub_bytes(inode, end - key.offset);
951 search_start = extent_end;
953 * | ---- range to drop ----- |
954 * | -------- extent -------- |
956 if (start > key.offset && end >= extent_end) {
958 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
963 btrfs_set_file_extent_num_bytes(leaf, fi,
965 btrfs_mark_buffer_dirty(leaf);
966 if (update_refs && disk_bytenr > 0)
967 inode_sub_bytes(inode, extent_end - start);
968 if (end == extent_end)
976 * | ---- range to drop ----- |
977 * | ------ extent ------ |
979 if (start <= key.offset && end >= extent_end) {
982 del_slot = path->slots[0];
985 BUG_ON(del_slot + del_nr != path->slots[0]);
990 extent_type == BTRFS_FILE_EXTENT_INLINE) {
991 inode_sub_bytes(inode,
992 extent_end - key.offset);
993 extent_end = ALIGN(extent_end,
994 fs_info->sectorsize);
995 } else if (update_refs && disk_bytenr > 0) {
996 ret = btrfs_free_extent(trans, root,
997 disk_bytenr, num_bytes, 0,
998 root->root_key.objectid,
999 key.objectid, key.offset -
1001 BUG_ON(ret); /* -ENOMEM */
1002 inode_sub_bytes(inode,
1003 extent_end - key.offset);
1006 if (end == extent_end)
1009 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1014 ret = btrfs_del_items(trans, root, path, del_slot,
1017 btrfs_abort_transaction(trans, ret);
1024 btrfs_release_path(path);
1031 if (!ret && del_nr > 0) {
1033 * Set path->slots[0] to first slot, so that after the delete
1034 * if items are move off from our leaf to its immediate left or
1035 * right neighbor leafs, we end up with a correct and adjusted
1036 * path->slots[0] for our insertion (if replace_extent != 0).
1038 path->slots[0] = del_slot;
1039 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1041 btrfs_abort_transaction(trans, ret);
1044 leaf = path->nodes[0];
1046 * If btrfs_del_items() was called, it might have deleted a leaf, in
1047 * which case it unlocked our path, so check path->locks[0] matches a
1050 if (!ret && replace_extent && leafs_visited == 1 &&
1051 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1052 path->locks[0] == BTRFS_WRITE_LOCK) &&
1053 btrfs_leaf_free_space(fs_info, leaf) >=
1054 sizeof(struct btrfs_item) + extent_item_size) {
1057 key.type = BTRFS_EXTENT_DATA_KEY;
1059 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1060 struct btrfs_key slot_key;
1062 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1063 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1066 setup_items_for_insert(root, path, &key,
1069 sizeof(struct btrfs_item) +
1070 extent_item_size, 1);
1074 if (!replace_extent || !(*key_inserted))
1075 btrfs_release_path(path);
1077 *drop_end = found ? min(end, last_end) : end;
1081 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1082 struct btrfs_root *root, struct inode *inode, u64 start,
1083 u64 end, int drop_cache)
1085 struct btrfs_path *path;
1088 path = btrfs_alloc_path();
1091 ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1092 drop_cache, 0, 0, NULL);
1093 btrfs_free_path(path);
1097 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1098 u64 objectid, u64 bytenr, u64 orig_offset,
1099 u64 *start, u64 *end)
1101 struct btrfs_file_extent_item *fi;
1102 struct btrfs_key key;
1105 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1108 btrfs_item_key_to_cpu(leaf, &key, slot);
1109 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1112 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1113 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1114 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1115 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1116 btrfs_file_extent_compression(leaf, fi) ||
1117 btrfs_file_extent_encryption(leaf, fi) ||
1118 btrfs_file_extent_other_encoding(leaf, fi))
1121 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1122 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1125 *start = key.offset;
1131 * Mark extent in the range start - end as written.
1133 * This changes extent type from 'pre-allocated' to 'regular'. If only
1134 * part of extent is marked as written, the extent will be split into
1137 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1138 struct btrfs_inode *inode, u64 start, u64 end)
1140 struct btrfs_fs_info *fs_info = trans->fs_info;
1141 struct btrfs_root *root = inode->root;
1142 struct extent_buffer *leaf;
1143 struct btrfs_path *path;
1144 struct btrfs_file_extent_item *fi;
1145 struct btrfs_key key;
1146 struct btrfs_key new_key;
1158 u64 ino = btrfs_ino(inode);
1160 path = btrfs_alloc_path();
1167 key.type = BTRFS_EXTENT_DATA_KEY;
1170 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1173 if (ret > 0 && path->slots[0] > 0)
1176 leaf = path->nodes[0];
1177 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1178 if (key.objectid != ino ||
1179 key.type != BTRFS_EXTENT_DATA_KEY) {
1181 btrfs_abort_transaction(trans, ret);
1184 fi = btrfs_item_ptr(leaf, path->slots[0],
1185 struct btrfs_file_extent_item);
1186 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1188 btrfs_abort_transaction(trans, ret);
1191 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1192 if (key.offset > start || extent_end < end) {
1194 btrfs_abort_transaction(trans, ret);
1198 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1199 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1200 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1201 memcpy(&new_key, &key, sizeof(new_key));
1203 if (start == key.offset && end < extent_end) {
1206 if (extent_mergeable(leaf, path->slots[0] - 1,
1207 ino, bytenr, orig_offset,
1208 &other_start, &other_end)) {
1209 new_key.offset = end;
1210 btrfs_set_item_key_safe(fs_info, path, &new_key);
1211 fi = btrfs_item_ptr(leaf, path->slots[0],
1212 struct btrfs_file_extent_item);
1213 btrfs_set_file_extent_generation(leaf, fi,
1215 btrfs_set_file_extent_num_bytes(leaf, fi,
1217 btrfs_set_file_extent_offset(leaf, fi,
1219 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1220 struct btrfs_file_extent_item);
1221 btrfs_set_file_extent_generation(leaf, fi,
1223 btrfs_set_file_extent_num_bytes(leaf, fi,
1225 btrfs_mark_buffer_dirty(leaf);
1230 if (start > key.offset && end == extent_end) {
1233 if (extent_mergeable(leaf, path->slots[0] + 1,
1234 ino, bytenr, orig_offset,
1235 &other_start, &other_end)) {
1236 fi = btrfs_item_ptr(leaf, path->slots[0],
1237 struct btrfs_file_extent_item);
1238 btrfs_set_file_extent_num_bytes(leaf, fi,
1239 start - key.offset);
1240 btrfs_set_file_extent_generation(leaf, fi,
1243 new_key.offset = start;
1244 btrfs_set_item_key_safe(fs_info, path, &new_key);
1246 fi = btrfs_item_ptr(leaf, path->slots[0],
1247 struct btrfs_file_extent_item);
1248 btrfs_set_file_extent_generation(leaf, fi,
1250 btrfs_set_file_extent_num_bytes(leaf, fi,
1252 btrfs_set_file_extent_offset(leaf, fi,
1253 start - orig_offset);
1254 btrfs_mark_buffer_dirty(leaf);
1259 while (start > key.offset || end < extent_end) {
1260 if (key.offset == start)
1263 new_key.offset = split;
1264 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1265 if (ret == -EAGAIN) {
1266 btrfs_release_path(path);
1270 btrfs_abort_transaction(trans, ret);
1274 leaf = path->nodes[0];
1275 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1276 struct btrfs_file_extent_item);
1277 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1278 btrfs_set_file_extent_num_bytes(leaf, fi,
1279 split - key.offset);
1281 fi = btrfs_item_ptr(leaf, path->slots[0],
1282 struct btrfs_file_extent_item);
1284 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1285 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1286 btrfs_set_file_extent_num_bytes(leaf, fi,
1287 extent_end - split);
1288 btrfs_mark_buffer_dirty(leaf);
1290 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes,
1291 0, root->root_key.objectid,
1294 btrfs_abort_transaction(trans, ret);
1298 if (split == start) {
1301 if (start != key.offset) {
1303 btrfs_abort_transaction(trans, ret);
1314 if (extent_mergeable(leaf, path->slots[0] + 1,
1315 ino, bytenr, orig_offset,
1316 &other_start, &other_end)) {
1318 btrfs_release_path(path);
1321 extent_end = other_end;
1322 del_slot = path->slots[0] + 1;
1324 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1325 0, root->root_key.objectid,
1328 btrfs_abort_transaction(trans, ret);
1334 if (extent_mergeable(leaf, path->slots[0] - 1,
1335 ino, bytenr, orig_offset,
1336 &other_start, &other_end)) {
1338 btrfs_release_path(path);
1341 key.offset = other_start;
1342 del_slot = path->slots[0];
1344 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1345 0, root->root_key.objectid,
1348 btrfs_abort_transaction(trans, ret);
1353 fi = btrfs_item_ptr(leaf, path->slots[0],
1354 struct btrfs_file_extent_item);
1355 btrfs_set_file_extent_type(leaf, fi,
1356 BTRFS_FILE_EXTENT_REG);
1357 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1358 btrfs_mark_buffer_dirty(leaf);
1360 fi = btrfs_item_ptr(leaf, del_slot - 1,
1361 struct btrfs_file_extent_item);
1362 btrfs_set_file_extent_type(leaf, fi,
1363 BTRFS_FILE_EXTENT_REG);
1364 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1365 btrfs_set_file_extent_num_bytes(leaf, fi,
1366 extent_end - key.offset);
1367 btrfs_mark_buffer_dirty(leaf);
1369 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1371 btrfs_abort_transaction(trans, ret);
1376 btrfs_free_path(path);
1381 * on error we return an unlocked page and the error value
1382 * on success we return a locked page and 0
1384 static int prepare_uptodate_page(struct inode *inode,
1385 struct page *page, u64 pos,
1386 bool force_uptodate)
1390 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1391 !PageUptodate(page)) {
1392 ret = btrfs_readpage(NULL, page);
1396 if (!PageUptodate(page)) {
1400 if (page->mapping != inode->i_mapping) {
1409 * this just gets pages into the page cache and locks them down.
1411 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1412 size_t num_pages, loff_t pos,
1413 size_t write_bytes, bool force_uptodate)
1416 unsigned long index = pos >> PAGE_SHIFT;
1417 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1421 for (i = 0; i < num_pages; i++) {
1423 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1424 mask | __GFP_WRITE);
1432 err = prepare_uptodate_page(inode, pages[i], pos,
1434 if (!err && i == num_pages - 1)
1435 err = prepare_uptodate_page(inode, pages[i],
1436 pos + write_bytes, false);
1439 if (err == -EAGAIN) {
1446 wait_on_page_writeback(pages[i]);
1451 while (faili >= 0) {
1452 unlock_page(pages[faili]);
1453 put_page(pages[faili]);
1461 * This function locks the extent and properly waits for data=ordered extents
1462 * to finish before allowing the pages to be modified if need.
1465 * 1 - the extent is locked
1466 * 0 - the extent is not locked, and everything is OK
1467 * -EAGAIN - need re-prepare the pages
1468 * the other < 0 number - Something wrong happens
1471 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1472 size_t num_pages, loff_t pos,
1474 u64 *lockstart, u64 *lockend,
1475 struct extent_state **cached_state)
1477 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1483 start_pos = round_down(pos, fs_info->sectorsize);
1484 last_pos = start_pos
1485 + round_up(pos + write_bytes - start_pos,
1486 fs_info->sectorsize) - 1;
1488 if (start_pos < inode->vfs_inode.i_size) {
1489 struct btrfs_ordered_extent *ordered;
1491 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1493 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1494 last_pos - start_pos + 1);
1496 ordered->file_offset + ordered->len > start_pos &&
1497 ordered->file_offset <= last_pos) {
1498 unlock_extent_cached(&inode->io_tree, start_pos,
1499 last_pos, cached_state);
1500 for (i = 0; i < num_pages; i++) {
1501 unlock_page(pages[i]);
1504 btrfs_start_ordered_extent(&inode->vfs_inode,
1506 btrfs_put_ordered_extent(ordered);
1510 btrfs_put_ordered_extent(ordered);
1512 *lockstart = start_pos;
1513 *lockend = last_pos;
1518 * It's possible the pages are dirty right now, but we don't want
1519 * to clean them yet because copy_from_user may catch a page fault
1520 * and we might have to fall back to one page at a time. If that
1521 * happens, we'll unlock these pages and we'd have a window where
1522 * reclaim could sneak in and drop the once-dirty page on the floor
1523 * without writing it.
1525 * We have the pages locked and the extent range locked, so there's
1526 * no way someone can start IO on any dirty pages in this range.
1528 * We'll call btrfs_dirty_pages() later on, and that will flip around
1529 * delalloc bits and dirty the pages as required.
1531 for (i = 0; i < num_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 = inode->root->fs_info;
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 kiocb *iocb,
1588 struct file *file = iocb->ki_filp;
1589 loff_t pos = iocb->ki_pos;
1590 struct inode *inode = file_inode(file);
1591 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1592 struct btrfs_root *root = BTRFS_I(inode)->root;
1593 struct page **pages = NULL;
1594 struct extent_changeset *data_reserved = NULL;
1595 u64 release_bytes = 0;
1598 size_t num_written = 0;
1601 bool only_release_metadata = false;
1602 bool force_page_uptodate = false;
1604 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1605 PAGE_SIZE / (sizeof(struct page *)));
1606 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1607 nrptrs = max(nrptrs, 8);
1608 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1612 while (iov_iter_count(i) > 0) {
1613 size_t offset = pos & (PAGE_SIZE - 1);
1614 struct extent_state *cached_state = NULL;
1615 size_t sector_offset;
1616 size_t write_bytes = min(iov_iter_count(i),
1617 nrptrs * (size_t)PAGE_SIZE -
1619 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1621 size_t reserve_bytes;
1624 size_t dirty_sectors;
1628 WARN_ON(num_pages > nrptrs);
1631 * Fault pages before locking them in prepare_pages
1632 * to avoid recursive lock
1634 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1639 only_release_metadata = false;
1640 sector_offset = pos & (fs_info->sectorsize - 1);
1641 reserve_bytes = round_up(write_bytes + sector_offset,
1642 fs_info->sectorsize);
1644 extent_changeset_release(data_reserved);
1645 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1648 if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1649 BTRFS_INODE_PREALLOC)) &&
1650 check_can_nocow(BTRFS_I(inode), pos,
1651 &write_bytes) > 0) {
1653 * For nodata cow case, no need to reserve
1656 only_release_metadata = true;
1658 * our prealloc extent may be smaller than
1659 * write_bytes, so scale down.
1661 num_pages = DIV_ROUND_UP(write_bytes + offset,
1663 reserve_bytes = round_up(write_bytes +
1665 fs_info->sectorsize);
1671 WARN_ON(reserve_bytes == 0);
1672 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1675 if (!only_release_metadata)
1676 btrfs_free_reserved_data_space(inode,
1680 btrfs_end_write_no_snapshotting(root);
1684 release_bytes = reserve_bytes;
1687 * This is going to setup the pages array with the number of
1688 * pages we want, so we don't really need to worry about the
1689 * contents of pages from loop to loop
1691 ret = prepare_pages(inode, pages, num_pages,
1693 force_page_uptodate);
1695 btrfs_delalloc_release_extents(BTRFS_I(inode),
1700 extents_locked = lock_and_cleanup_extent_if_need(
1701 BTRFS_I(inode), pages,
1702 num_pages, pos, write_bytes, &lockstart,
1703 &lockend, &cached_state);
1704 if (extents_locked < 0) {
1705 if (extents_locked == -EAGAIN)
1707 btrfs_delalloc_release_extents(BTRFS_I(inode),
1709 ret = extents_locked;
1713 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1715 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1716 dirty_sectors = round_up(copied + sector_offset,
1717 fs_info->sectorsize);
1718 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1721 * if we have trouble faulting in the pages, fall
1722 * back to one page at a time
1724 if (copied < write_bytes)
1728 force_page_uptodate = true;
1732 force_page_uptodate = false;
1733 dirty_pages = DIV_ROUND_UP(copied + offset,
1737 if (num_sectors > dirty_sectors) {
1738 /* release everything except the sectors we dirtied */
1739 release_bytes -= dirty_sectors <<
1740 fs_info->sb->s_blocksize_bits;
1741 if (only_release_metadata) {
1742 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1743 release_bytes, true);
1747 __pos = round_down(pos,
1748 fs_info->sectorsize) +
1749 (dirty_pages << PAGE_SHIFT);
1750 btrfs_delalloc_release_space(inode,
1751 data_reserved, __pos,
1752 release_bytes, true);
1756 release_bytes = round_up(copied + sector_offset,
1757 fs_info->sectorsize);
1760 ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1761 pos, copied, &cached_state);
1764 * If we have not locked the extent range, because the range's
1765 * start offset is >= i_size, we might still have a non-NULL
1766 * cached extent state, acquired while marking the extent range
1767 * as delalloc through btrfs_dirty_pages(). Therefore free any
1768 * possible cached extent state to avoid a memory leak.
1771 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1772 lockstart, lockend, &cached_state);
1774 free_extent_state(cached_state);
1776 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1778 btrfs_drop_pages(pages, num_pages);
1783 if (only_release_metadata)
1784 btrfs_end_write_no_snapshotting(root);
1786 if (only_release_metadata && copied > 0) {
1787 lockstart = round_down(pos,
1788 fs_info->sectorsize);
1789 lockend = round_up(pos + copied,
1790 fs_info->sectorsize) - 1;
1792 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1793 lockend, EXTENT_NORESERVE, NULL,
1797 btrfs_drop_pages(pages, num_pages);
1801 balance_dirty_pages_ratelimited(inode->i_mapping);
1802 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1803 btrfs_btree_balance_dirty(fs_info);
1806 num_written += copied;
1811 if (release_bytes) {
1812 if (only_release_metadata) {
1813 btrfs_end_write_no_snapshotting(root);
1814 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1815 release_bytes, true);
1817 btrfs_delalloc_release_space(inode, data_reserved,
1818 round_down(pos, fs_info->sectorsize),
1819 release_bytes, true);
1823 extent_changeset_free(data_reserved);
1824 return num_written ? num_written : ret;
1827 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1829 struct file *file = iocb->ki_filp;
1830 struct inode *inode = file_inode(file);
1833 ssize_t written_buffered;
1837 written = generic_file_direct_write(iocb, from);
1839 if (written < 0 || !iov_iter_count(from))
1843 written_buffered = btrfs_buffered_write(iocb, from);
1844 if (written_buffered < 0) {
1845 err = written_buffered;
1849 * Ensure all data is persisted. We want the next direct IO read to be
1850 * able to read what was just written.
1852 endbyte = pos + written_buffered - 1;
1853 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1856 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1859 written += written_buffered;
1860 iocb->ki_pos = pos + written_buffered;
1861 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1862 endbyte >> PAGE_SHIFT);
1864 return written ? written : err;
1867 static void update_time_for_write(struct inode *inode)
1869 struct timespec64 now;
1871 if (IS_NOCMTIME(inode))
1874 now = current_time(inode);
1875 if (!timespec64_equal(&inode->i_mtime, &now))
1876 inode->i_mtime = now;
1878 if (!timespec64_equal(&inode->i_ctime, &now))
1879 inode->i_ctime = now;
1881 if (IS_I_VERSION(inode))
1882 inode_inc_iversion(inode);
1885 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1886 struct iov_iter *from)
1888 struct file *file = iocb->ki_filp;
1889 struct inode *inode = file_inode(file);
1890 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1891 struct btrfs_root *root = BTRFS_I(inode)->root;
1894 ssize_t num_written = 0;
1895 bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1902 if (!(iocb->ki_flags & IOCB_DIRECT) &&
1903 (iocb->ki_flags & IOCB_NOWAIT))
1906 if (iocb->ki_flags & IOCB_NOWAIT) {
1907 if (!inode_trylock(inode))
1913 err = generic_write_checks(iocb, from);
1915 inode_unlock(inode);
1920 count = iov_iter_count(from);
1921 if (iocb->ki_flags & IOCB_NOWAIT) {
1923 * We will allocate space in case nodatacow is not set,
1926 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1927 BTRFS_INODE_PREALLOC)) ||
1928 check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1929 inode_unlock(inode);
1934 current->backing_dev_info = inode_to_bdi(inode);
1935 err = file_remove_privs(file);
1937 inode_unlock(inode);
1942 * If BTRFS flips readonly due to some impossible error
1943 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1944 * although we have opened a file as writable, we have
1945 * to stop this write operation to ensure FS consistency.
1947 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1948 inode_unlock(inode);
1954 * We reserve space for updating the inode when we reserve space for the
1955 * extent we are going to write, so we will enospc out there. We don't
1956 * need to start yet another transaction to update the inode as we will
1957 * update the inode when we finish writing whatever data we write.
1959 update_time_for_write(inode);
1961 start_pos = round_down(pos, fs_info->sectorsize);
1962 oldsize = i_size_read(inode);
1963 if (start_pos > oldsize) {
1964 /* Expand hole size to cover write data, preventing empty gap */
1965 end_pos = round_up(pos + count,
1966 fs_info->sectorsize);
1967 err = btrfs_cont_expand(inode, oldsize, end_pos);
1969 inode_unlock(inode);
1972 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1977 atomic_inc(&BTRFS_I(inode)->sync_writers);
1979 if (iocb->ki_flags & IOCB_DIRECT) {
1980 num_written = __btrfs_direct_write(iocb, from);
1982 num_written = btrfs_buffered_write(iocb, from);
1983 if (num_written > 0)
1984 iocb->ki_pos = pos + num_written;
1986 pagecache_isize_extended(inode, oldsize,
1987 i_size_read(inode));
1990 inode_unlock(inode);
1993 * We also have to set last_sub_trans to the current log transid,
1994 * otherwise subsequent syncs to a file that's been synced in this
1995 * transaction will appear to have already occurred.
1997 spin_lock(&BTRFS_I(inode)->lock);
1998 BTRFS_I(inode)->last_sub_trans = root->log_transid;
1999 spin_unlock(&BTRFS_I(inode)->lock);
2000 if (num_written > 0)
2001 num_written = generic_write_sync(iocb, num_written);
2004 atomic_dec(&BTRFS_I(inode)->sync_writers);
2006 current->backing_dev_info = NULL;
2007 return num_written ? num_written : err;
2010 int btrfs_release_file(struct inode *inode, struct file *filp)
2012 struct btrfs_file_private *private = filp->private_data;
2014 if (private && private->filldir_buf)
2015 kfree(private->filldir_buf);
2017 filp->private_data = NULL;
2020 * ordered_data_close is set by settattr when we are about to truncate
2021 * a file from a non-zero size to a zero size. This tries to
2022 * flush down new bytes that may have been written if the
2023 * application were using truncate to replace a file in place.
2025 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2026 &BTRFS_I(inode)->runtime_flags))
2027 filemap_flush(inode->i_mapping);
2031 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2034 struct blk_plug plug;
2037 * This is only called in fsync, which would do synchronous writes, so
2038 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2039 * multiple disks using raid profile, a large IO can be split to
2040 * several segments of stripe length (currently 64K).
2042 blk_start_plug(&plug);
2043 atomic_inc(&BTRFS_I(inode)->sync_writers);
2044 ret = btrfs_fdatawrite_range(inode, start, end);
2045 atomic_dec(&BTRFS_I(inode)->sync_writers);
2046 blk_finish_plug(&plug);
2052 * fsync call for both files and directories. This logs the inode into
2053 * the tree log instead of forcing full commits whenever possible.
2055 * It needs to call filemap_fdatawait so that all ordered extent updates are
2056 * in the metadata btree are up to date for copying to the log.
2058 * It drops the inode mutex before doing the tree log commit. This is an
2059 * important optimization for directories because holding the mutex prevents
2060 * new operations on the dir while we write to disk.
2062 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2064 struct dentry *dentry = file_dentry(file);
2065 struct inode *inode = d_inode(dentry);
2066 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2067 struct btrfs_root *root = BTRFS_I(inode)->root;
2068 struct btrfs_trans_handle *trans;
2069 struct btrfs_log_ctx ctx;
2072 trace_btrfs_sync_file(file, datasync);
2074 btrfs_init_log_ctx(&ctx, inode);
2077 * Set the range to full if the NO_HOLES feature is not enabled.
2078 * This is to avoid missing file extent items representing holes after
2079 * replaying the log.
2081 if (!btrfs_fs_incompat(fs_info, NO_HOLES)) {
2087 * We write the dirty pages in the range and wait until they complete
2088 * out of the ->i_mutex. If so, we can flush the dirty pages by
2089 * multi-task, and make the performance up. See
2090 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2092 ret = start_ordered_ops(inode, start, end);
2099 * We take the dio_sem here because the tree log stuff can race with
2100 * lockless dio writes and get an extent map logged for an extent we
2101 * never waited on. We need it this high up for lockdep reasons.
2103 down_write(&BTRFS_I(inode)->dio_sem);
2105 atomic_inc(&root->log_batch);
2108 * If the inode needs a full sync, make sure we use a full range to
2109 * avoid log tree corruption, due to hole detection racing with ordered
2110 * extent completion for adjacent ranges, and assertion failures during
2111 * hole detection. Do this while holding the inode lock, to avoid races
2114 if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2115 &BTRFS_I(inode)->runtime_flags)) {
2121 * Before we acquired the inode's lock, someone may have dirtied more
2122 * pages in the target range. We need to make sure that writeback for
2123 * any such pages does not start while we are logging the inode, because
2124 * if it does, any of the following might happen when we are not doing a
2127 * 1) We log an extent after its writeback finishes but before its
2128 * checksums are added to the csum tree, leading to -EIO errors
2129 * when attempting to read the extent after a log replay.
2131 * 2) We can end up logging an extent before its writeback finishes.
2132 * Therefore after the log replay we will have a file extent item
2133 * pointing to an unwritten extent (and no data checksums as well).
2135 * So trigger writeback for any eventual new dirty pages and then we
2136 * wait for all ordered extents to complete below.
2138 ret = start_ordered_ops(inode, start, end);
2140 up_write(&BTRFS_I(inode)->dio_sem);
2141 inode_unlock(inode);
2146 * We have to do this here to avoid the priority inversion of waiting on
2147 * IO of a lower priority task while holding a transaciton open.
2149 * Also, the range length can be represented by u64, we have to do the
2150 * typecasts to avoid signed overflow if it's [0, LLONG_MAX].
2152 ret = btrfs_wait_ordered_range(inode, start, (u64)end - (u64)start + 1);
2154 up_write(&BTRFS_I(inode)->dio_sem);
2155 inode_unlock(inode);
2158 atomic_inc(&root->log_batch);
2161 if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2162 BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2164 * We've had everything committed since the last time we were
2165 * modified so clear this flag in case it was set for whatever
2166 * reason, it's no longer relevant.
2168 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2169 &BTRFS_I(inode)->runtime_flags);
2171 * An ordered extent might have started before and completed
2172 * already with io errors, in which case the inode was not
2173 * updated and we end up here. So check the inode's mapping
2174 * for any errors that might have happened since we last
2175 * checked called fsync.
2177 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2178 up_write(&BTRFS_I(inode)->dio_sem);
2179 inode_unlock(inode);
2184 * We use start here because we will need to wait on the IO to complete
2185 * in btrfs_sync_log, which could require joining a transaction (for
2186 * example checking cross references in the nocow path). If we use join
2187 * here we could get into a situation where we're waiting on IO to
2188 * happen that is blocked on a transaction trying to commit. With start
2189 * we inc the extwriter counter, so we wait for all extwriters to exit
2190 * before we start blocking join'ers. This comment is to keep somebody
2191 * from thinking they are super smart and changing this to
2192 * btrfs_join_transaction *cough*Josef*cough*.
2194 trans = btrfs_start_transaction(root, 0);
2195 if (IS_ERR(trans)) {
2196 ret = PTR_ERR(trans);
2197 up_write(&BTRFS_I(inode)->dio_sem);
2198 inode_unlock(inode);
2203 ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2205 /* Fallthrough and commit/free transaction. */
2209 /* we've logged all the items and now have a consistent
2210 * version of the file in the log. It is possible that
2211 * someone will come in and modify the file, but that's
2212 * fine because the log is consistent on disk, and we
2213 * have references to all of the file's extents
2215 * It is possible that someone will come in and log the
2216 * file again, but that will end up using the synchronization
2217 * inside btrfs_sync_log to keep things safe.
2219 up_write(&BTRFS_I(inode)->dio_sem);
2220 inode_unlock(inode);
2223 * If any of the ordered extents had an error, just return it to user
2224 * space, so that the application knows some writes didn't succeed and
2225 * can take proper action (retry for e.g.). Blindly committing the
2226 * transaction in this case, would fool userspace that everything was
2227 * successful. And we also want to make sure our log doesn't contain
2228 * file extent items pointing to extents that weren't fully written to -
2229 * just like in the non fast fsync path, where we check for the ordered
2230 * operation's error flag before writing to the log tree and return -EIO
2231 * if any of them had this flag set (btrfs_wait_ordered_range) -
2232 * therefore we need to check for errors in the ordered operations,
2233 * which are indicated by ctx.io_err.
2236 btrfs_end_transaction(trans);
2241 if (ret != BTRFS_NO_LOG_SYNC) {
2243 ret = btrfs_sync_log(trans, root, &ctx);
2245 ret = btrfs_end_transaction(trans);
2249 ret = btrfs_commit_transaction(trans);
2251 ret = btrfs_end_transaction(trans);
2254 ASSERT(list_empty(&ctx.list));
2255 err = file_check_and_advance_wb_err(file);
2258 return ret > 0 ? -EIO : ret;
2261 static const struct vm_operations_struct btrfs_file_vm_ops = {
2262 .fault = filemap_fault,
2263 .map_pages = filemap_map_pages,
2264 .page_mkwrite = btrfs_page_mkwrite,
2267 static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2269 struct address_space *mapping = filp->f_mapping;
2271 if (!mapping->a_ops->readpage)
2274 file_accessed(filp);
2275 vma->vm_ops = &btrfs_file_vm_ops;
2280 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2281 int slot, u64 start, u64 end)
2283 struct btrfs_file_extent_item *fi;
2284 struct btrfs_key key;
2286 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2289 btrfs_item_key_to_cpu(leaf, &key, slot);
2290 if (key.objectid != btrfs_ino(inode) ||
2291 key.type != BTRFS_EXTENT_DATA_KEY)
2294 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2296 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2299 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2302 if (key.offset == end)
2304 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2309 static int fill_holes(struct btrfs_trans_handle *trans,
2310 struct btrfs_inode *inode,
2311 struct btrfs_path *path, u64 offset, u64 end)
2313 struct btrfs_fs_info *fs_info = trans->fs_info;
2314 struct btrfs_root *root = inode->root;
2315 struct extent_buffer *leaf;
2316 struct btrfs_file_extent_item *fi;
2317 struct extent_map *hole_em;
2318 struct extent_map_tree *em_tree = &inode->extent_tree;
2319 struct btrfs_key key;
2322 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2325 key.objectid = btrfs_ino(inode);
2326 key.type = BTRFS_EXTENT_DATA_KEY;
2327 key.offset = offset;
2329 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2332 * We should have dropped this offset, so if we find it then
2333 * something has gone horribly wrong.
2340 leaf = path->nodes[0];
2341 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2345 fi = btrfs_item_ptr(leaf, path->slots[0],
2346 struct btrfs_file_extent_item);
2347 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2349 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2350 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2351 btrfs_set_file_extent_offset(leaf, fi, 0);
2352 btrfs_mark_buffer_dirty(leaf);
2356 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2359 key.offset = offset;
2360 btrfs_set_item_key_safe(fs_info, path, &key);
2361 fi = btrfs_item_ptr(leaf, path->slots[0],
2362 struct btrfs_file_extent_item);
2363 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2365 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2366 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2367 btrfs_set_file_extent_offset(leaf, fi, 0);
2368 btrfs_mark_buffer_dirty(leaf);
2371 btrfs_release_path(path);
2373 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2374 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2379 btrfs_release_path(path);
2381 hole_em = alloc_extent_map();
2383 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2384 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2386 hole_em->start = offset;
2387 hole_em->len = end - offset;
2388 hole_em->ram_bytes = hole_em->len;
2389 hole_em->orig_start = offset;
2391 hole_em->block_start = EXTENT_MAP_HOLE;
2392 hole_em->block_len = 0;
2393 hole_em->orig_block_len = 0;
2394 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2395 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2396 hole_em->generation = trans->transid;
2399 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2400 write_lock(&em_tree->lock);
2401 ret = add_extent_mapping(em_tree, hole_em, 1);
2402 write_unlock(&em_tree->lock);
2403 } while (ret == -EEXIST);
2404 free_extent_map(hole_em);
2406 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2407 &inode->runtime_flags);
2414 * Find a hole extent on given inode and change start/len to the end of hole
2415 * extent.(hole/vacuum extent whose em->start <= start &&
2416 * em->start + em->len > start)
2417 * When a hole extent is found, return 1 and modify start/len.
2419 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2421 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2422 struct extent_map *em;
2425 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2426 round_down(*start, fs_info->sectorsize),
2427 round_up(*len, fs_info->sectorsize), 0);
2431 /* Hole or vacuum extent(only exists in no-hole mode) */
2432 if (em->block_start == EXTENT_MAP_HOLE) {
2434 *len = em->start + em->len > *start + *len ?
2435 0 : *start + *len - em->start - em->len;
2436 *start = em->start + em->len;
2438 free_extent_map(em);
2442 static int btrfs_punch_hole_lock_range(struct inode *inode,
2443 const u64 lockstart,
2445 struct extent_state **cached_state)
2448 struct btrfs_ordered_extent *ordered;
2451 truncate_pagecache_range(inode, lockstart, lockend);
2453 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2455 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2458 * We need to make sure we have no ordered extents in this range
2459 * and nobody raced in and read a page in this range, if we did
2460 * we need to try again.
2463 (ordered->file_offset + ordered->len <= lockstart ||
2464 ordered->file_offset > lockend)) &&
2465 !filemap_range_has_page(inode->i_mapping,
2466 lockstart, lockend)) {
2468 btrfs_put_ordered_extent(ordered);
2472 btrfs_put_ordered_extent(ordered);
2473 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2474 lockend, cached_state);
2475 ret = btrfs_wait_ordered_range(inode, lockstart,
2476 lockend - lockstart + 1);
2483 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2485 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2486 struct btrfs_root *root = BTRFS_I(inode)->root;
2487 struct extent_state *cached_state = NULL;
2488 struct btrfs_path *path;
2489 struct btrfs_block_rsv *rsv;
2490 struct btrfs_trans_handle *trans;
2495 u64 orig_start = offset;
2497 u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2501 unsigned int rsv_count;
2503 bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2505 bool truncated_block = false;
2506 bool updated_inode = false;
2508 ret = btrfs_wait_ordered_range(inode, offset, len);
2513 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2514 ret = find_first_non_hole(inode, &offset, &len);
2516 goto out_only_mutex;
2518 /* Already in a large hole */
2520 goto out_only_mutex;
2523 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2524 lockend = round_down(offset + len,
2525 btrfs_inode_sectorsize(inode)) - 1;
2526 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2527 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2529 * We needn't truncate any block which is beyond the end of the file
2530 * because we are sure there is no data there.
2533 * Only do this if we are in the same block and we aren't doing the
2536 if (same_block && len < fs_info->sectorsize) {
2537 if (offset < ino_size) {
2538 truncated_block = true;
2539 ret = btrfs_truncate_block(inode, offset, len, 0);
2543 goto out_only_mutex;
2546 /* zero back part of the first block */
2547 if (offset < ino_size) {
2548 truncated_block = true;
2549 ret = btrfs_truncate_block(inode, offset, 0, 0);
2551 inode_unlock(inode);
2556 /* Check the aligned pages after the first unaligned page,
2557 * if offset != orig_start, which means the first unaligned page
2558 * including several following pages are already in holes,
2559 * the extra check can be skipped */
2560 if (offset == orig_start) {
2561 /* after truncate page, check hole again */
2562 len = offset + len - lockstart;
2564 ret = find_first_non_hole(inode, &offset, &len);
2566 goto out_only_mutex;
2569 goto out_only_mutex;
2574 /* Check the tail unaligned part is in a hole */
2575 tail_start = lockend + 1;
2576 tail_len = offset + len - tail_start;
2578 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2579 if (unlikely(ret < 0))
2580 goto out_only_mutex;
2582 /* zero the front end of the last page */
2583 if (tail_start + tail_len < ino_size) {
2584 truncated_block = true;
2585 ret = btrfs_truncate_block(inode,
2586 tail_start + tail_len,
2589 goto out_only_mutex;
2594 if (lockend < lockstart) {
2596 goto out_only_mutex;
2599 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2602 goto out_only_mutex;
2604 path = btrfs_alloc_path();
2610 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2615 rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2619 * 1 - update the inode
2620 * 1 - removing the extents in the range
2621 * 1 - adding the hole extent if no_holes isn't set
2623 rsv_count = no_holes ? 2 : 3;
2624 trans = btrfs_start_transaction(root, rsv_count);
2625 if (IS_ERR(trans)) {
2626 err = PTR_ERR(trans);
2630 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2633 trans->block_rsv = rsv;
2635 cur_offset = lockstart;
2636 len = lockend - cur_offset;
2637 while (cur_offset < lockend) {
2638 ret = __btrfs_drop_extents(trans, root, inode, path,
2639 cur_offset, lockend + 1,
2640 &drop_end, 1, 0, 0, NULL);
2644 trans->block_rsv = &fs_info->trans_block_rsv;
2646 if (cur_offset < drop_end && cur_offset < ino_size) {
2647 ret = fill_holes(trans, BTRFS_I(inode), path,
2648 cur_offset, drop_end);
2651 * If we failed then we didn't insert our hole
2652 * entries for the area we dropped, so now the
2653 * fs is corrupted, so we must abort the
2656 btrfs_abort_transaction(trans, ret);
2662 cur_offset = drop_end;
2664 ret = btrfs_update_inode(trans, root, inode);
2670 btrfs_end_transaction(trans);
2671 btrfs_btree_balance_dirty(fs_info);
2673 trans = btrfs_start_transaction(root, rsv_count);
2674 if (IS_ERR(trans)) {
2675 ret = PTR_ERR(trans);
2680 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2682 BUG_ON(ret); /* shouldn't happen */
2683 trans->block_rsv = rsv;
2685 ret = find_first_non_hole(inode, &cur_offset, &len);
2686 if (unlikely(ret < 0))
2699 trans->block_rsv = &fs_info->trans_block_rsv;
2701 * If we are using the NO_HOLES feature we might have had already an
2702 * hole that overlaps a part of the region [lockstart, lockend] and
2703 * ends at (or beyond) lockend. Since we have no file extent items to
2704 * represent holes, drop_end can be less than lockend and so we must
2705 * make sure we have an extent map representing the existing hole (the
2706 * call to __btrfs_drop_extents() might have dropped the existing extent
2707 * map representing the existing hole), otherwise the fast fsync path
2708 * will not record the existence of the hole region
2709 * [existing_hole_start, lockend].
2711 if (drop_end <= lockend)
2712 drop_end = lockend + 1;
2714 * Don't insert file hole extent item if it's for a range beyond eof
2715 * (because it's useless) or if it represents a 0 bytes range (when
2716 * cur_offset == drop_end).
2718 if (cur_offset < ino_size && cur_offset < drop_end) {
2719 ret = fill_holes(trans, BTRFS_I(inode), path,
2720 cur_offset, drop_end);
2722 /* Same comment as above. */
2723 btrfs_abort_transaction(trans, ret);
2733 inode_inc_iversion(inode);
2734 inode->i_mtime = inode->i_ctime = current_time(inode);
2736 trans->block_rsv = &fs_info->trans_block_rsv;
2737 ret = btrfs_update_inode(trans, root, inode);
2738 updated_inode = true;
2739 btrfs_end_transaction(trans);
2740 btrfs_btree_balance_dirty(fs_info);
2742 btrfs_free_path(path);
2743 btrfs_free_block_rsv(fs_info, rsv);
2745 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2748 if (!updated_inode && truncated_block && !ret && !err) {
2750 * If we only end up zeroing part of a page, we still need to
2751 * update the inode item, so that all the time fields are
2752 * updated as well as the necessary btrfs inode in memory fields
2753 * for detecting, at fsync time, if the inode isn't yet in the
2754 * log tree or it's there but not up to date.
2756 struct timespec64 now = current_time(inode);
2758 inode_inc_iversion(inode);
2759 inode->i_mtime = now;
2760 inode->i_ctime = now;
2761 trans = btrfs_start_transaction(root, 1);
2762 if (IS_ERR(trans)) {
2763 err = PTR_ERR(trans);
2765 err = btrfs_update_inode(trans, root, inode);
2766 ret = btrfs_end_transaction(trans);
2769 inode_unlock(inode);
2775 /* Helper structure to record which range is already reserved */
2776 struct falloc_range {
2777 struct list_head list;
2783 * Helper function to add falloc range
2785 * Caller should have locked the larger range of extent containing
2788 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2790 struct falloc_range *prev = NULL;
2791 struct falloc_range *range = NULL;
2793 if (list_empty(head))
2797 * As fallocate iterate by bytenr order, we only need to check
2800 prev = list_entry(head->prev, struct falloc_range, list);
2801 if (prev->start + prev->len == start) {
2806 range = kmalloc(sizeof(*range), GFP_KERNEL);
2809 range->start = start;
2811 list_add_tail(&range->list, head);
2815 static int btrfs_fallocate_update_isize(struct inode *inode,
2819 struct btrfs_trans_handle *trans;
2820 struct btrfs_root *root = BTRFS_I(inode)->root;
2824 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2827 trans = btrfs_start_transaction(root, 1);
2829 return PTR_ERR(trans);
2831 inode->i_ctime = current_time(inode);
2832 i_size_write(inode, end);
2833 btrfs_ordered_update_i_size(inode, end, NULL);
2834 ret = btrfs_update_inode(trans, root, inode);
2835 ret2 = btrfs_end_transaction(trans);
2837 return ret ? ret : ret2;
2841 RANGE_BOUNDARY_WRITTEN_EXTENT = 0,
2842 RANGE_BOUNDARY_PREALLOC_EXTENT = 1,
2843 RANGE_BOUNDARY_HOLE = 2,
2846 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2849 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2850 struct extent_map *em;
2853 offset = round_down(offset, sectorsize);
2854 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2858 if (em->block_start == EXTENT_MAP_HOLE)
2859 ret = RANGE_BOUNDARY_HOLE;
2860 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2861 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2863 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2865 free_extent_map(em);
2869 static int btrfs_zero_range(struct inode *inode,
2874 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2875 struct extent_map *em;
2876 struct extent_changeset *data_reserved = NULL;
2879 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2880 u64 alloc_start = round_down(offset, sectorsize);
2881 u64 alloc_end = round_up(offset + len, sectorsize);
2882 u64 bytes_to_reserve = 0;
2883 bool space_reserved = false;
2885 inode_dio_wait(inode);
2887 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2888 alloc_start, alloc_end - alloc_start, 0);
2895 * Avoid hole punching and extent allocation for some cases. More cases
2896 * could be considered, but these are unlikely common and we keep things
2897 * as simple as possible for now. Also, intentionally, if the target
2898 * range contains one or more prealloc extents together with regular
2899 * extents and holes, we drop all the existing extents and allocate a
2900 * new prealloc extent, so that we get a larger contiguous disk extent.
2902 if (em->start <= alloc_start &&
2903 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2904 const u64 em_end = em->start + em->len;
2906 if (em_end >= offset + len) {
2908 * The whole range is already a prealloc extent,
2909 * do nothing except updating the inode's i_size if
2912 free_extent_map(em);
2913 ret = btrfs_fallocate_update_isize(inode, offset + len,
2918 * Part of the range is already a prealloc extent, so operate
2919 * only on the remaining part of the range.
2921 alloc_start = em_end;
2922 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2923 len = offset + len - alloc_start;
2924 offset = alloc_start;
2925 alloc_hint = em->block_start + em->len;
2927 free_extent_map(em);
2929 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2930 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2931 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2932 alloc_start, sectorsize, 0);
2938 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2939 free_extent_map(em);
2940 ret = btrfs_fallocate_update_isize(inode, offset + len,
2944 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2945 free_extent_map(em);
2946 ret = btrfs_truncate_block(inode, offset, len, 0);
2948 ret = btrfs_fallocate_update_isize(inode,
2953 free_extent_map(em);
2954 alloc_start = round_down(offset, sectorsize);
2955 alloc_end = alloc_start + sectorsize;
2959 alloc_start = round_up(offset, sectorsize);
2960 alloc_end = round_down(offset + len, sectorsize);
2963 * For unaligned ranges, check the pages at the boundaries, they might
2964 * map to an extent, in which case we need to partially zero them, or
2965 * they might map to a hole, in which case we need our allocation range
2968 if (!IS_ALIGNED(offset, sectorsize)) {
2969 ret = btrfs_zero_range_check_range_boundary(inode, offset);
2972 if (ret == RANGE_BOUNDARY_HOLE) {
2973 alloc_start = round_down(offset, sectorsize);
2975 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2976 ret = btrfs_truncate_block(inode, offset, 0, 0);
2984 if (!IS_ALIGNED(offset + len, sectorsize)) {
2985 ret = btrfs_zero_range_check_range_boundary(inode,
2989 if (ret == RANGE_BOUNDARY_HOLE) {
2990 alloc_end = round_up(offset + len, sectorsize);
2992 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2993 ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3002 if (alloc_start < alloc_end) {
3003 struct extent_state *cached_state = NULL;
3004 const u64 lockstart = alloc_start;
3005 const u64 lockend = alloc_end - 1;
3007 bytes_to_reserve = alloc_end - alloc_start;
3008 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3012 space_reserved = true;
3013 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3017 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3018 alloc_start, bytes_to_reserve);
3020 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3021 lockend, &cached_state);
3024 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3025 alloc_end - alloc_start,
3027 offset + len, &alloc_hint);
3028 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3029 lockend, &cached_state);
3030 /* btrfs_prealloc_file_range releases reserved space on error */
3032 space_reserved = false;
3036 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3038 if (ret && space_reserved)
3039 btrfs_free_reserved_data_space(inode, data_reserved,
3040 alloc_start, bytes_to_reserve);
3041 extent_changeset_free(data_reserved);
3046 static long btrfs_fallocate(struct file *file, int mode,
3047 loff_t offset, loff_t len)
3049 struct inode *inode = file_inode(file);
3050 struct extent_state *cached_state = NULL;
3051 struct extent_changeset *data_reserved = NULL;
3052 struct falloc_range *range;
3053 struct falloc_range *tmp;
3054 struct list_head reserve_list;
3062 struct extent_map *em;
3063 int blocksize = btrfs_inode_sectorsize(inode);
3066 alloc_start = round_down(offset, blocksize);
3067 alloc_end = round_up(offset + len, blocksize);
3068 cur_offset = alloc_start;
3070 /* Make sure we aren't being give some crap mode */
3071 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3072 FALLOC_FL_ZERO_RANGE))
3075 if (mode & FALLOC_FL_PUNCH_HOLE)
3076 return btrfs_punch_hole(inode, offset, len);
3079 * Only trigger disk allocation, don't trigger qgroup reserve
3081 * For qgroup space, it will be checked later.
3083 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3084 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3085 alloc_end - alloc_start);
3092 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3093 ret = inode_newsize_ok(inode, offset + len);
3099 * TODO: Move these two operations after we have checked
3100 * accurate reserved space, or fallocate can still fail but
3101 * with page truncated or size expanded.
3103 * But that's a minor problem and won't do much harm BTW.
3105 if (alloc_start > inode->i_size) {
3106 ret = btrfs_cont_expand(inode, i_size_read(inode),
3110 } else if (offset + len > inode->i_size) {
3112 * If we are fallocating from the end of the file onward we
3113 * need to zero out the end of the block if i_size lands in the
3114 * middle of a block.
3116 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3122 * wait for ordered IO before we have any locks. We'll loop again
3123 * below with the locks held.
3125 ret = btrfs_wait_ordered_range(inode, alloc_start,
3126 alloc_end - alloc_start);
3130 if (mode & FALLOC_FL_ZERO_RANGE) {
3131 ret = btrfs_zero_range(inode, offset, len, mode);
3132 inode_unlock(inode);
3136 locked_end = alloc_end - 1;
3138 struct btrfs_ordered_extent *ordered;
3140 /* the extent lock is ordered inside the running
3143 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3144 locked_end, &cached_state);
3145 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3148 ordered->file_offset + ordered->len > alloc_start &&
3149 ordered->file_offset < alloc_end) {
3150 btrfs_put_ordered_extent(ordered);
3151 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3152 alloc_start, locked_end,
3155 * we can't wait on the range with the transaction
3156 * running or with the extent lock held
3158 ret = btrfs_wait_ordered_range(inode, alloc_start,
3159 alloc_end - alloc_start);
3164 btrfs_put_ordered_extent(ordered);
3169 /* First, check if we exceed the qgroup limit */
3170 INIT_LIST_HEAD(&reserve_list);
3171 while (cur_offset < alloc_end) {
3172 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3173 alloc_end - cur_offset, 0);
3178 last_byte = min(extent_map_end(em), alloc_end);
3179 actual_end = min_t(u64, extent_map_end(em), offset + len);
3180 last_byte = ALIGN(last_byte, blocksize);
3181 if (em->block_start == EXTENT_MAP_HOLE ||
3182 (cur_offset >= inode->i_size &&
3183 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3184 ret = add_falloc_range(&reserve_list, cur_offset,
3185 last_byte - cur_offset);
3187 free_extent_map(em);
3190 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3191 cur_offset, last_byte - cur_offset);
3193 cur_offset = last_byte;
3194 free_extent_map(em);
3199 * Do not need to reserve unwritten extent for this
3200 * range, free reserved data space first, otherwise
3201 * it'll result in false ENOSPC error.
3203 btrfs_free_reserved_data_space(inode, data_reserved,
3204 cur_offset, last_byte - cur_offset);
3206 free_extent_map(em);
3207 cur_offset = last_byte;
3211 * If ret is still 0, means we're OK to fallocate.
3212 * Or just cleanup the list and exit.
3214 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3216 ret = btrfs_prealloc_file_range(inode, mode,
3218 range->len, i_blocksize(inode),
3219 offset + len, &alloc_hint);
3221 btrfs_free_reserved_data_space(inode,
3222 data_reserved, range->start,
3224 list_del(&range->list);
3231 * We didn't need to allocate any more space, but we still extended the
3232 * size of the file so we need to update i_size and the inode item.
3234 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3236 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3239 inode_unlock(inode);
3240 /* Let go of our reservation. */
3241 if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3242 btrfs_free_reserved_data_space(inode, data_reserved,
3243 cur_offset, alloc_end - cur_offset);
3244 extent_changeset_free(data_reserved);
3248 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3250 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3251 struct extent_map *em = NULL;
3252 struct extent_state *cached_state = NULL;
3259 if (inode->i_size == 0)
3263 * *offset can be negative, in this case we start finding DATA/HOLE from
3264 * the very start of the file.
3266 start = max_t(loff_t, 0, *offset);
3268 lockstart = round_down(start, fs_info->sectorsize);
3269 lockend = round_up(i_size_read(inode),
3270 fs_info->sectorsize);
3271 if (lockend <= lockstart)
3272 lockend = lockstart + fs_info->sectorsize;
3274 len = lockend - lockstart + 1;
3276 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3279 while (start < inode->i_size) {
3280 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0,
3288 if (whence == SEEK_HOLE &&
3289 (em->block_start == EXTENT_MAP_HOLE ||
3290 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3292 else if (whence == SEEK_DATA &&
3293 (em->block_start != EXTENT_MAP_HOLE &&
3294 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3297 start = em->start + em->len;
3298 free_extent_map(em);
3302 free_extent_map(em);
3304 if (whence == SEEK_DATA && start >= inode->i_size)
3307 *offset = min_t(loff_t, start, inode->i_size);
3309 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3314 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3316 struct inode *inode = file->f_mapping->host;
3323 offset = generic_file_llseek(file, offset, whence);
3327 if (offset >= i_size_read(inode)) {
3328 inode_unlock(inode);
3332 ret = find_desired_extent(inode, &offset, whence);
3334 inode_unlock(inode);
3339 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3341 inode_unlock(inode);
3345 static int btrfs_file_open(struct inode *inode, struct file *filp)
3347 filp->f_mode |= FMODE_NOWAIT;
3348 return generic_file_open(inode, filp);
3351 const struct file_operations btrfs_file_operations = {
3352 .llseek = btrfs_file_llseek,
3353 .read_iter = generic_file_read_iter,
3354 .splice_read = generic_file_splice_read,
3355 .write_iter = btrfs_file_write_iter,
3356 .mmap = btrfs_file_mmap,
3357 .open = btrfs_file_open,
3358 .release = btrfs_release_file,
3359 .fsync = btrfs_sync_file,
3360 .fallocate = btrfs_fallocate,
3361 .unlocked_ioctl = btrfs_ioctl,
3362 #ifdef CONFIG_COMPAT
3363 .compat_ioctl = btrfs_compat_ioctl,
3365 .clone_file_range = btrfs_clone_file_range,
3366 .dedupe_file_range = btrfs_dedupe_file_range,
3369 void __cold btrfs_auto_defrag_exit(void)
3371 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3374 int __init btrfs_auto_defrag_init(void)
3376 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3377 sizeof(struct inode_defrag), 0,
3380 if (!btrfs_inode_defrag_cachep)
3386 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3391 * So with compression we will find and lock a dirty page and clear the
3392 * first one as dirty, setup an async extent, and immediately return
3393 * with the entire range locked but with nobody actually marked with
3394 * writeback. So we can't just filemap_write_and_wait_range() and
3395 * expect it to work since it will just kick off a thread to do the
3396 * actual work. So we need to call filemap_fdatawrite_range _again_
3397 * since it will wait on the page lock, which won't be unlocked until
3398 * after the pages have been marked as writeback and so we're good to go
3399 * from there. We have to do this otherwise we'll miss the ordered
3400 * extents and that results in badness. Please Josef, do not think you
3401 * know better and pull this out at some point in the future, it is
3402 * right and you are wrong.
3404 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3405 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3406 &BTRFS_I(inode)->runtime_flags))
3407 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);