1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2007 Oracle. All rights reserved.
7 #include <linux/pagemap.h>
8 #include <linux/time.h>
9 #include <linux/init.h>
10 #include <linux/string.h>
11 #include <linux/backing-dev.h>
12 #include <linux/falloc.h>
13 #include <linux/writeback.h>
14 #include <linux/compat.h>
15 #include <linux/slab.h>
16 #include <linux/btrfs.h>
17 #include <linux/uio.h>
18 #include <linux/iversion.h>
21 #include "transaction.h"
22 #include "btrfs_inode.h"
23 #include "print-tree.h"
28 #include "compression.h"
29 #include "delalloc-space.h"
32 static struct kmem_cache *btrfs_inode_defrag_cachep;
34 * when auto defrag is enabled we
35 * queue up these defrag structs to remember which
36 * inodes need defragging passes
39 struct rb_node rb_node;
43 * transid where the defrag was added, we search for
44 * extents newer than this
51 /* last offset we were able to defrag */
54 /* if we've wrapped around back to zero once already */
58 static int __compare_inode_defrag(struct inode_defrag *defrag1,
59 struct inode_defrag *defrag2)
61 if (defrag1->root > defrag2->root)
63 else if (defrag1->root < defrag2->root)
65 else if (defrag1->ino > defrag2->ino)
67 else if (defrag1->ino < defrag2->ino)
73 /* pop a record for an inode into the defrag tree. The lock
74 * must be held already
76 * If you're inserting a record for an older transid than an
77 * existing record, the transid already in the tree is lowered
79 * If an existing record is found the defrag item you
82 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
83 struct inode_defrag *defrag)
85 struct btrfs_fs_info *fs_info = inode->root->fs_info;
86 struct inode_defrag *entry;
88 struct rb_node *parent = NULL;
91 p = &fs_info->defrag_inodes.rb_node;
94 entry = rb_entry(parent, struct inode_defrag, rb_node);
96 ret = __compare_inode_defrag(defrag, entry);
100 p = &parent->rb_right;
102 /* if we're reinserting an entry for
103 * an old defrag run, make sure to
104 * lower the transid of our existing record
106 if (defrag->transid < entry->transid)
107 entry->transid = defrag->transid;
108 if (defrag->last_offset > entry->last_offset)
109 entry->last_offset = defrag->last_offset;
113 set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
114 rb_link_node(&defrag->rb_node, parent, p);
115 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
119 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
121 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
124 if (btrfs_fs_closing(fs_info))
131 * insert a defrag record for this inode if auto defrag is
134 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
135 struct btrfs_inode *inode)
137 struct btrfs_root *root = inode->root;
138 struct btrfs_fs_info *fs_info = root->fs_info;
139 struct inode_defrag *defrag;
143 if (!__need_auto_defrag(fs_info))
146 if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
150 transid = trans->transid;
152 transid = inode->root->last_trans;
154 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
158 defrag->ino = btrfs_ino(inode);
159 defrag->transid = transid;
160 defrag->root = root->root_key.objectid;
162 spin_lock(&fs_info->defrag_inodes_lock);
163 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
165 * If we set IN_DEFRAG flag and evict the inode from memory,
166 * and then re-read this inode, this new inode doesn't have
167 * IN_DEFRAG flag. At the case, we may find the existed defrag.
169 ret = __btrfs_add_inode_defrag(inode, defrag);
171 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
173 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
175 spin_unlock(&fs_info->defrag_inodes_lock);
180 * Requeue the defrag object. If there is a defrag object that points to
181 * the same inode in the tree, we will merge them together (by
182 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
184 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
185 struct inode_defrag *defrag)
187 struct btrfs_fs_info *fs_info = inode->root->fs_info;
190 if (!__need_auto_defrag(fs_info))
194 * Here we don't check the IN_DEFRAG flag, because we need merge
197 spin_lock(&fs_info->defrag_inodes_lock);
198 ret = __btrfs_add_inode_defrag(inode, defrag);
199 spin_unlock(&fs_info->defrag_inodes_lock);
204 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
208 * pick the defragable inode that we want, if it doesn't exist, we will get
211 static struct inode_defrag *
212 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
214 struct inode_defrag *entry = NULL;
215 struct inode_defrag tmp;
217 struct rb_node *parent = NULL;
223 spin_lock(&fs_info->defrag_inodes_lock);
224 p = fs_info->defrag_inodes.rb_node;
227 entry = rb_entry(parent, struct inode_defrag, rb_node);
229 ret = __compare_inode_defrag(&tmp, entry);
233 p = parent->rb_right;
238 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
239 parent = rb_next(parent);
241 entry = rb_entry(parent, struct inode_defrag, rb_node);
247 rb_erase(parent, &fs_info->defrag_inodes);
248 spin_unlock(&fs_info->defrag_inodes_lock);
252 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
254 struct inode_defrag *defrag;
255 struct rb_node *node;
257 spin_lock(&fs_info->defrag_inodes_lock);
258 node = rb_first(&fs_info->defrag_inodes);
260 rb_erase(node, &fs_info->defrag_inodes);
261 defrag = rb_entry(node, struct inode_defrag, rb_node);
262 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
264 cond_resched_lock(&fs_info->defrag_inodes_lock);
266 node = rb_first(&fs_info->defrag_inodes);
268 spin_unlock(&fs_info->defrag_inodes_lock);
271 #define BTRFS_DEFRAG_BATCH 1024
273 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
274 struct inode_defrag *defrag)
276 struct btrfs_root *inode_root;
278 struct btrfs_ioctl_defrag_range_args range;
283 inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
284 if (IS_ERR(inode_root)) {
285 ret = PTR_ERR(inode_root);
289 inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
290 btrfs_put_root(inode_root);
292 ret = PTR_ERR(inode);
296 /* do a chunk of defrag */
297 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
298 memset(&range, 0, sizeof(range));
300 range.start = defrag->last_offset;
302 sb_start_write(fs_info->sb);
303 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
305 sb_end_write(fs_info->sb);
307 * if we filled the whole defrag batch, there
308 * must be more work to do. Queue this defrag
311 if (num_defrag == BTRFS_DEFRAG_BATCH) {
312 defrag->last_offset = range.start;
313 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
314 } else if (defrag->last_offset && !defrag->cycled) {
316 * we didn't fill our defrag batch, but
317 * we didn't start at zero. Make sure we loop
318 * around to the start of the file.
320 defrag->last_offset = 0;
322 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
324 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
330 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
335 * run through the list of inodes in the FS that need
338 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
340 struct inode_defrag *defrag;
342 u64 root_objectid = 0;
344 atomic_inc(&fs_info->defrag_running);
346 /* Pause the auto defragger. */
347 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
351 if (!__need_auto_defrag(fs_info))
354 /* find an inode to defrag */
355 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
358 if (root_objectid || first_ino) {
367 first_ino = defrag->ino + 1;
368 root_objectid = defrag->root;
370 __btrfs_run_defrag_inode(fs_info, defrag);
372 atomic_dec(&fs_info->defrag_running);
375 * during unmount, we use the transaction_wait queue to
376 * wait for the defragger to stop
378 wake_up(&fs_info->transaction_wait);
382 /* simple helper to fault in pages and copy. This should go away
383 * and be replaced with calls into generic code.
385 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
386 struct page **prepared_pages,
390 size_t total_copied = 0;
392 int offset = offset_in_page(pos);
394 while (write_bytes > 0) {
395 size_t count = min_t(size_t,
396 PAGE_SIZE - offset, write_bytes);
397 struct page *page = prepared_pages[pg];
399 * Copy data from userspace to the current page
401 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
403 /* Flush processor's dcache for this page */
404 flush_dcache_page(page);
407 * if we get a partial write, we can end up with
408 * partially up to date pages. These add
409 * a lot of complexity, so make sure they don't
410 * happen by forcing this copy to be retried.
412 * The rest of the btrfs_file_write code will fall
413 * back to page at a time copies after we return 0.
415 if (!PageUptodate(page) && copied < count)
418 iov_iter_advance(i, copied);
419 write_bytes -= copied;
420 total_copied += copied;
422 /* Return to btrfs_file_write_iter to fault page */
423 if (unlikely(copied == 0))
426 if (copied < PAGE_SIZE - offset) {
437 * unlocks pages after btrfs_file_write is done with them
439 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
442 for (i = 0; i < num_pages; i++) {
443 /* page checked is some magic around finding pages that
444 * have been modified without going through btrfs_set_page_dirty
445 * clear it here. There should be no need to mark the pages
446 * accessed as prepare_pages should have marked them accessed
447 * in prepare_pages via find_or_create_page()
449 ClearPageChecked(pages[i]);
450 unlock_page(pages[i]);
455 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
458 struct extent_state **cached_state)
460 u64 search_start = start;
461 const u64 end = start + len - 1;
463 while (search_start < end) {
464 const u64 search_len = end - search_start + 1;
465 struct extent_map *em;
469 em = btrfs_get_extent(inode, NULL, 0, search_start, search_len);
473 if (em->block_start != EXTENT_MAP_HOLE)
477 if (em->start < search_start)
478 em_len -= search_start - em->start;
479 if (em_len > search_len)
482 ret = set_extent_bit(&inode->io_tree, search_start,
483 search_start + em_len - 1,
485 NULL, cached_state, GFP_NOFS);
487 search_start = extent_map_end(em);
496 * after copy_from_user, pages need to be dirtied and we need to make
497 * sure holes are created between the current EOF and the start of
498 * any next extents (if required).
500 * this also makes the decision about creating an inline extent vs
501 * doing real data extents, marking pages dirty and delalloc as required.
503 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
504 size_t num_pages, loff_t pos, size_t write_bytes,
505 struct extent_state **cached)
507 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
512 u64 end_of_last_block;
513 u64 end_pos = pos + write_bytes;
514 loff_t isize = i_size_read(inode);
515 unsigned int extra_bits = 0;
517 start_pos = pos & ~((u64) fs_info->sectorsize - 1);
518 num_bytes = round_up(write_bytes + pos - start_pos,
519 fs_info->sectorsize);
521 end_of_last_block = start_pos + num_bytes - 1;
524 * The pages may have already been dirty, clear out old accounting so
525 * we can set things up properly
527 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos, end_of_last_block,
528 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
531 if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
532 if (start_pos >= isize &&
533 !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
535 * There can't be any extents following eof in this case
536 * so just set the delalloc new bit for the range
539 extra_bits |= EXTENT_DELALLOC_NEW;
541 err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
549 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
554 for (i = 0; i < num_pages; i++) {
555 struct page *p = pages[i];
562 * we've only changed i_size in ram, and we haven't updated
563 * the disk i_size. There is no need to log the inode
567 i_size_write(inode, end_pos);
572 * this drops all the extents in the cache that intersect the range
573 * [start, end]. Existing extents are split as required.
575 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
578 struct extent_map *em;
579 struct extent_map *split = NULL;
580 struct extent_map *split2 = NULL;
581 struct extent_map_tree *em_tree = &inode->extent_tree;
582 u64 len = end - start + 1;
590 WARN_ON(end < start);
591 if (end == (u64)-1) {
600 split = alloc_extent_map();
602 split2 = alloc_extent_map();
603 if (!split || !split2)
606 write_lock(&em_tree->lock);
607 em = lookup_extent_mapping(em_tree, start, len);
609 write_unlock(&em_tree->lock);
613 gen = em->generation;
614 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
615 if (testend && em->start + em->len >= start + len) {
617 write_unlock(&em_tree->lock);
620 start = em->start + em->len;
622 len = start + len - (em->start + em->len);
624 write_unlock(&em_tree->lock);
627 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
628 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
629 clear_bit(EXTENT_FLAG_LOGGING, &flags);
630 modified = !list_empty(&em->list);
634 if (em->start < start) {
635 split->start = em->start;
636 split->len = start - em->start;
638 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
639 split->orig_start = em->orig_start;
640 split->block_start = em->block_start;
643 split->block_len = em->block_len;
645 split->block_len = split->len;
646 split->orig_block_len = max(split->block_len,
648 split->ram_bytes = em->ram_bytes;
650 split->orig_start = split->start;
651 split->block_len = 0;
652 split->block_start = em->block_start;
653 split->orig_block_len = 0;
654 split->ram_bytes = split->len;
657 split->generation = gen;
658 split->flags = flags;
659 split->compress_type = em->compress_type;
660 replace_extent_mapping(em_tree, em, split, modified);
661 free_extent_map(split);
665 if (testend && em->start + em->len > start + len) {
666 u64 diff = start + len - em->start;
668 split->start = start + len;
669 split->len = em->start + em->len - (start + len);
670 split->flags = flags;
671 split->compress_type = em->compress_type;
672 split->generation = gen;
674 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
675 split->orig_block_len = max(em->block_len,
678 split->ram_bytes = em->ram_bytes;
680 split->block_len = em->block_len;
681 split->block_start = em->block_start;
682 split->orig_start = em->orig_start;
684 split->block_len = split->len;
685 split->block_start = em->block_start
687 split->orig_start = em->orig_start;
690 split->ram_bytes = split->len;
691 split->orig_start = split->start;
692 split->block_len = 0;
693 split->block_start = em->block_start;
694 split->orig_block_len = 0;
697 if (extent_map_in_tree(em)) {
698 replace_extent_mapping(em_tree, em, split,
701 ret = add_extent_mapping(em_tree, split,
703 ASSERT(ret == 0); /* Logic error */
705 free_extent_map(split);
709 if (extent_map_in_tree(em))
710 remove_extent_mapping(em_tree, em);
711 write_unlock(&em_tree->lock);
715 /* once for the tree*/
719 free_extent_map(split);
721 free_extent_map(split2);
725 * this is very complex, but the basic idea is to drop all extents
726 * in the range start - end. hint_block is filled in with a block number
727 * that would be a good hint to the block allocator for this file.
729 * If an extent intersects the range but is not entirely inside the range
730 * it is either truncated or split. Anything entirely inside the range
731 * is deleted from the tree.
733 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
734 struct btrfs_root *root, struct inode *inode,
735 struct btrfs_path *path, u64 start, u64 end,
736 u64 *drop_end, int drop_cache,
738 u32 extent_item_size,
741 struct btrfs_fs_info *fs_info = root->fs_info;
742 struct extent_buffer *leaf;
743 struct btrfs_file_extent_item *fi;
744 struct btrfs_ref ref = { 0 };
745 struct btrfs_key key;
746 struct btrfs_key new_key;
747 u64 ino = btrfs_ino(BTRFS_I(inode));
748 u64 search_start = start;
751 u64 extent_offset = 0;
753 u64 last_end = start;
759 int modify_tree = -1;
762 int leafs_visited = 0;
765 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
767 if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
770 update_refs = (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
771 root == fs_info->tree_root);
774 ret = btrfs_lookup_file_extent(trans, root, path, ino,
775 search_start, modify_tree);
778 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
779 leaf = path->nodes[0];
780 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
781 if (key.objectid == ino &&
782 key.type == BTRFS_EXTENT_DATA_KEY)
788 leaf = path->nodes[0];
789 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
791 ret = btrfs_next_leaf(root, path);
799 leaf = path->nodes[0];
803 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
805 if (key.objectid > ino)
807 if (WARN_ON_ONCE(key.objectid < ino) ||
808 key.type < BTRFS_EXTENT_DATA_KEY) {
813 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
816 fi = btrfs_item_ptr(leaf, path->slots[0],
817 struct btrfs_file_extent_item);
818 extent_type = btrfs_file_extent_type(leaf, fi);
820 if (extent_type == BTRFS_FILE_EXTENT_REG ||
821 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
822 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
823 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
824 extent_offset = btrfs_file_extent_offset(leaf, fi);
825 extent_end = key.offset +
826 btrfs_file_extent_num_bytes(leaf, fi);
827 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
828 extent_end = key.offset +
829 btrfs_file_extent_ram_bytes(leaf, fi);
836 * Don't skip extent items representing 0 byte lengths. They
837 * used to be created (bug) if while punching holes we hit
838 * -ENOSPC condition. So if we find one here, just ensure we
839 * delete it, otherwise we would insert a new file extent item
840 * with the same key (offset) as that 0 bytes length file
841 * extent item in the call to setup_items_for_insert() later
844 if (extent_end == key.offset && extent_end >= search_start) {
845 last_end = extent_end;
846 goto delete_extent_item;
849 if (extent_end <= search_start) {
855 search_start = max(key.offset, start);
856 if (recow || !modify_tree) {
858 btrfs_release_path(path);
863 * | - range to drop - |
864 * | -------- extent -------- |
866 if (start > key.offset && end < extent_end) {
868 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
873 memcpy(&new_key, &key, sizeof(new_key));
874 new_key.offset = start;
875 ret = btrfs_duplicate_item(trans, root, path,
877 if (ret == -EAGAIN) {
878 btrfs_release_path(path);
884 leaf = path->nodes[0];
885 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
886 struct btrfs_file_extent_item);
887 btrfs_set_file_extent_num_bytes(leaf, fi,
890 fi = btrfs_item_ptr(leaf, path->slots[0],
891 struct btrfs_file_extent_item);
893 extent_offset += start - key.offset;
894 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
895 btrfs_set_file_extent_num_bytes(leaf, fi,
897 btrfs_mark_buffer_dirty(leaf);
899 if (update_refs && disk_bytenr > 0) {
900 btrfs_init_generic_ref(&ref,
901 BTRFS_ADD_DELAYED_REF,
902 disk_bytenr, num_bytes, 0);
903 btrfs_init_data_ref(&ref,
904 root->root_key.objectid,
906 start - extent_offset);
907 ret = btrfs_inc_extent_ref(trans, &ref);
908 BUG_ON(ret); /* -ENOMEM */
913 * From here on out we will have actually dropped something, so
914 * last_end can be updated.
916 last_end = extent_end;
919 * | ---- range to drop ----- |
920 * | -------- extent -------- |
922 if (start <= key.offset && end < extent_end) {
923 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
928 memcpy(&new_key, &key, sizeof(new_key));
929 new_key.offset = end;
930 btrfs_set_item_key_safe(fs_info, path, &new_key);
932 extent_offset += end - key.offset;
933 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
934 btrfs_set_file_extent_num_bytes(leaf, fi,
936 btrfs_mark_buffer_dirty(leaf);
937 if (update_refs && disk_bytenr > 0)
938 inode_sub_bytes(inode, end - key.offset);
942 search_start = extent_end;
944 * | ---- range to drop ----- |
945 * | -------- extent -------- |
947 if (start > key.offset && end >= extent_end) {
949 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
954 btrfs_set_file_extent_num_bytes(leaf, fi,
956 btrfs_mark_buffer_dirty(leaf);
957 if (update_refs && disk_bytenr > 0)
958 inode_sub_bytes(inode, extent_end - start);
959 if (end == extent_end)
967 * | ---- range to drop ----- |
968 * | ------ extent ------ |
970 if (start <= key.offset && end >= extent_end) {
973 del_slot = path->slots[0];
976 BUG_ON(del_slot + del_nr != path->slots[0]);
981 extent_type == BTRFS_FILE_EXTENT_INLINE) {
982 inode_sub_bytes(inode,
983 extent_end - key.offset);
984 extent_end = ALIGN(extent_end,
985 fs_info->sectorsize);
986 } else if (update_refs && disk_bytenr > 0) {
987 btrfs_init_generic_ref(&ref,
988 BTRFS_DROP_DELAYED_REF,
989 disk_bytenr, num_bytes, 0);
990 btrfs_init_data_ref(&ref,
991 root->root_key.objectid,
993 key.offset - extent_offset);
994 ret = btrfs_free_extent(trans, &ref);
995 BUG_ON(ret); /* -ENOMEM */
996 inode_sub_bytes(inode,
997 extent_end - key.offset);
1000 if (end == extent_end)
1003 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1008 ret = btrfs_del_items(trans, root, path, del_slot,
1011 btrfs_abort_transaction(trans, ret);
1018 btrfs_release_path(path);
1025 if (!ret && del_nr > 0) {
1027 * Set path->slots[0] to first slot, so that after the delete
1028 * if items are move off from our leaf to its immediate left or
1029 * right neighbor leafs, we end up with a correct and adjusted
1030 * path->slots[0] for our insertion (if replace_extent != 0).
1032 path->slots[0] = del_slot;
1033 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1035 btrfs_abort_transaction(trans, ret);
1038 leaf = path->nodes[0];
1040 * If btrfs_del_items() was called, it might have deleted a leaf, in
1041 * which case it unlocked our path, so check path->locks[0] matches a
1044 if (!ret && replace_extent && leafs_visited == 1 &&
1045 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1046 path->locks[0] == BTRFS_WRITE_LOCK) &&
1047 btrfs_leaf_free_space(leaf) >=
1048 sizeof(struct btrfs_item) + extent_item_size) {
1051 key.type = BTRFS_EXTENT_DATA_KEY;
1053 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1054 struct btrfs_key slot_key;
1056 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1057 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1060 setup_items_for_insert(root, path, &key,
1063 sizeof(struct btrfs_item) +
1064 extent_item_size, 1);
1068 if (!replace_extent || !(*key_inserted))
1069 btrfs_release_path(path);
1071 *drop_end = found ? min(end, last_end) : end;
1075 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1076 struct btrfs_root *root, struct inode *inode, u64 start,
1077 u64 end, int drop_cache)
1079 struct btrfs_path *path;
1082 path = btrfs_alloc_path();
1085 ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1086 drop_cache, 0, 0, NULL);
1087 btrfs_free_path(path);
1091 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1092 u64 objectid, u64 bytenr, u64 orig_offset,
1093 u64 *start, u64 *end)
1095 struct btrfs_file_extent_item *fi;
1096 struct btrfs_key key;
1099 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1102 btrfs_item_key_to_cpu(leaf, &key, slot);
1103 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1106 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1107 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1108 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1109 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1110 btrfs_file_extent_compression(leaf, fi) ||
1111 btrfs_file_extent_encryption(leaf, fi) ||
1112 btrfs_file_extent_other_encoding(leaf, fi))
1115 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1116 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1119 *start = key.offset;
1125 * Mark extent in the range start - end as written.
1127 * This changes extent type from 'pre-allocated' to 'regular'. If only
1128 * part of extent is marked as written, the extent will be split into
1131 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1132 struct btrfs_inode *inode, u64 start, u64 end)
1134 struct btrfs_fs_info *fs_info = trans->fs_info;
1135 struct btrfs_root *root = inode->root;
1136 struct extent_buffer *leaf;
1137 struct btrfs_path *path;
1138 struct btrfs_file_extent_item *fi;
1139 struct btrfs_ref ref = { 0 };
1140 struct btrfs_key key;
1141 struct btrfs_key new_key;
1153 u64 ino = btrfs_ino(inode);
1155 path = btrfs_alloc_path();
1162 key.type = BTRFS_EXTENT_DATA_KEY;
1165 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1168 if (ret > 0 && path->slots[0] > 0)
1171 leaf = path->nodes[0];
1172 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1173 if (key.objectid != ino ||
1174 key.type != BTRFS_EXTENT_DATA_KEY) {
1176 btrfs_abort_transaction(trans, ret);
1179 fi = btrfs_item_ptr(leaf, path->slots[0],
1180 struct btrfs_file_extent_item);
1181 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1183 btrfs_abort_transaction(trans, ret);
1186 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1187 if (key.offset > start || extent_end < end) {
1189 btrfs_abort_transaction(trans, ret);
1193 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1194 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1195 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1196 memcpy(&new_key, &key, sizeof(new_key));
1198 if (start == key.offset && end < extent_end) {
1201 if (extent_mergeable(leaf, path->slots[0] - 1,
1202 ino, bytenr, orig_offset,
1203 &other_start, &other_end)) {
1204 new_key.offset = end;
1205 btrfs_set_item_key_safe(fs_info, path, &new_key);
1206 fi = btrfs_item_ptr(leaf, path->slots[0],
1207 struct btrfs_file_extent_item);
1208 btrfs_set_file_extent_generation(leaf, fi,
1210 btrfs_set_file_extent_num_bytes(leaf, fi,
1212 btrfs_set_file_extent_offset(leaf, fi,
1214 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1215 struct btrfs_file_extent_item);
1216 btrfs_set_file_extent_generation(leaf, fi,
1218 btrfs_set_file_extent_num_bytes(leaf, fi,
1220 btrfs_mark_buffer_dirty(leaf);
1225 if (start > key.offset && end == extent_end) {
1228 if (extent_mergeable(leaf, path->slots[0] + 1,
1229 ino, bytenr, orig_offset,
1230 &other_start, &other_end)) {
1231 fi = btrfs_item_ptr(leaf, path->slots[0],
1232 struct btrfs_file_extent_item);
1233 btrfs_set_file_extent_num_bytes(leaf, fi,
1234 start - key.offset);
1235 btrfs_set_file_extent_generation(leaf, fi,
1238 new_key.offset = start;
1239 btrfs_set_item_key_safe(fs_info, path, &new_key);
1241 fi = btrfs_item_ptr(leaf, path->slots[0],
1242 struct btrfs_file_extent_item);
1243 btrfs_set_file_extent_generation(leaf, fi,
1245 btrfs_set_file_extent_num_bytes(leaf, fi,
1247 btrfs_set_file_extent_offset(leaf, fi,
1248 start - orig_offset);
1249 btrfs_mark_buffer_dirty(leaf);
1254 while (start > key.offset || end < extent_end) {
1255 if (key.offset == start)
1258 new_key.offset = split;
1259 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1260 if (ret == -EAGAIN) {
1261 btrfs_release_path(path);
1265 btrfs_abort_transaction(trans, ret);
1269 leaf = path->nodes[0];
1270 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1271 struct btrfs_file_extent_item);
1272 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1273 btrfs_set_file_extent_num_bytes(leaf, fi,
1274 split - key.offset);
1276 fi = btrfs_item_ptr(leaf, path->slots[0],
1277 struct btrfs_file_extent_item);
1279 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1280 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1281 btrfs_set_file_extent_num_bytes(leaf, fi,
1282 extent_end - split);
1283 btrfs_mark_buffer_dirty(leaf);
1285 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1287 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1289 ret = btrfs_inc_extent_ref(trans, &ref);
1291 btrfs_abort_transaction(trans, ret);
1295 if (split == start) {
1298 if (start != key.offset) {
1300 btrfs_abort_transaction(trans, ret);
1311 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1313 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
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, &ref);
1326 btrfs_abort_transaction(trans, ret);
1332 if (extent_mergeable(leaf, path->slots[0] - 1,
1333 ino, bytenr, orig_offset,
1334 &other_start, &other_end)) {
1336 btrfs_release_path(path);
1339 key.offset = other_start;
1340 del_slot = path->slots[0];
1342 ret = btrfs_free_extent(trans, &ref);
1344 btrfs_abort_transaction(trans, ret);
1349 fi = btrfs_item_ptr(leaf, path->slots[0],
1350 struct btrfs_file_extent_item);
1351 btrfs_set_file_extent_type(leaf, fi,
1352 BTRFS_FILE_EXTENT_REG);
1353 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1354 btrfs_mark_buffer_dirty(leaf);
1356 fi = btrfs_item_ptr(leaf, del_slot - 1,
1357 struct btrfs_file_extent_item);
1358 btrfs_set_file_extent_type(leaf, fi,
1359 BTRFS_FILE_EXTENT_REG);
1360 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1361 btrfs_set_file_extent_num_bytes(leaf, fi,
1362 extent_end - key.offset);
1363 btrfs_mark_buffer_dirty(leaf);
1365 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1367 btrfs_abort_transaction(trans, ret);
1372 btrfs_free_path(path);
1377 * on error we return an unlocked page and the error value
1378 * on success we return a locked page and 0
1380 static int prepare_uptodate_page(struct inode *inode,
1381 struct page *page, u64 pos,
1382 bool force_uptodate)
1386 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1387 !PageUptodate(page)) {
1388 ret = btrfs_readpage(NULL, page);
1392 if (!PageUptodate(page)) {
1396 if (page->mapping != inode->i_mapping) {
1405 * this just gets pages into the page cache and locks them down.
1407 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1408 size_t num_pages, loff_t pos,
1409 size_t write_bytes, bool force_uptodate)
1412 unsigned long index = pos >> PAGE_SHIFT;
1413 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1417 for (i = 0; i < num_pages; i++) {
1419 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1420 mask | __GFP_WRITE);
1428 err = prepare_uptodate_page(inode, pages[i], pos,
1430 if (!err && i == num_pages - 1)
1431 err = prepare_uptodate_page(inode, pages[i],
1432 pos + write_bytes, false);
1435 if (err == -EAGAIN) {
1442 wait_on_page_writeback(pages[i]);
1447 while (faili >= 0) {
1448 unlock_page(pages[faili]);
1449 put_page(pages[faili]);
1457 * This function locks the extent and properly waits for data=ordered extents
1458 * to finish before allowing the pages to be modified if need.
1461 * 1 - the extent is locked
1462 * 0 - the extent is not locked, and everything is OK
1463 * -EAGAIN - need re-prepare the pages
1464 * the other < 0 number - Something wrong happens
1467 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1468 size_t num_pages, loff_t pos,
1470 u64 *lockstart, u64 *lockend,
1471 struct extent_state **cached_state)
1473 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1479 start_pos = round_down(pos, fs_info->sectorsize);
1480 last_pos = start_pos
1481 + round_up(pos + write_bytes - start_pos,
1482 fs_info->sectorsize) - 1;
1484 if (start_pos < inode->vfs_inode.i_size) {
1485 struct btrfs_ordered_extent *ordered;
1487 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1489 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1490 last_pos - start_pos + 1);
1492 ordered->file_offset + ordered->num_bytes > start_pos &&
1493 ordered->file_offset <= last_pos) {
1494 unlock_extent_cached(&inode->io_tree, start_pos,
1495 last_pos, cached_state);
1496 for (i = 0; i < num_pages; i++) {
1497 unlock_page(pages[i]);
1500 btrfs_start_ordered_extent(&inode->vfs_inode,
1502 btrfs_put_ordered_extent(ordered);
1506 btrfs_put_ordered_extent(ordered);
1508 *lockstart = start_pos;
1509 *lockend = last_pos;
1514 * It's possible the pages are dirty right now, but we don't want
1515 * to clean them yet because copy_from_user may catch a page fault
1516 * and we might have to fall back to one page at a time. If that
1517 * happens, we'll unlock these pages and we'd have a window where
1518 * reclaim could sneak in and drop the once-dirty page on the floor
1519 * without writing it.
1521 * We have the pages locked and the extent range locked, so there's
1522 * no way someone can start IO on any dirty pages in this range.
1524 * We'll call btrfs_dirty_pages() later on, and that will flip around
1525 * delalloc bits and dirty the pages as required.
1527 for (i = 0; i < num_pages; i++) {
1528 set_page_extent_mapped(pages[i]);
1529 WARN_ON(!PageLocked(pages[i]));
1535 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1536 size_t *write_bytes)
1538 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1539 struct btrfs_root *root = inode->root;
1540 u64 lockstart, lockend;
1544 if (!btrfs_drew_try_write_lock(&root->snapshot_lock))
1547 lockstart = round_down(pos, fs_info->sectorsize);
1548 lockend = round_up(pos + *write_bytes,
1549 fs_info->sectorsize) - 1;
1551 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1554 num_bytes = lockend - lockstart + 1;
1555 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1559 btrfs_drew_write_unlock(&root->snapshot_lock);
1561 *write_bytes = min_t(size_t, *write_bytes ,
1562 num_bytes - pos + lockstart);
1565 unlock_extent(&inode->io_tree, lockstart, lockend);
1570 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1573 struct file *file = iocb->ki_filp;
1574 loff_t pos = iocb->ki_pos;
1575 struct inode *inode = file_inode(file);
1576 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1577 struct btrfs_root *root = BTRFS_I(inode)->root;
1578 struct page **pages = NULL;
1579 struct extent_changeset *data_reserved = NULL;
1580 u64 release_bytes = 0;
1583 size_t num_written = 0;
1586 bool only_release_metadata = false;
1587 bool force_page_uptodate = false;
1589 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1590 PAGE_SIZE / (sizeof(struct page *)));
1591 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1592 nrptrs = max(nrptrs, 8);
1593 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1597 while (iov_iter_count(i) > 0) {
1598 struct extent_state *cached_state = NULL;
1599 size_t offset = offset_in_page(pos);
1600 size_t sector_offset;
1601 size_t write_bytes = min(iov_iter_count(i),
1602 nrptrs * (size_t)PAGE_SIZE -
1604 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1606 size_t reserve_bytes;
1609 size_t dirty_sectors;
1613 WARN_ON(num_pages > nrptrs);
1616 * Fault pages before locking them in prepare_pages
1617 * to avoid recursive lock
1619 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1624 only_release_metadata = false;
1625 sector_offset = pos & (fs_info->sectorsize - 1);
1626 reserve_bytes = round_up(write_bytes + sector_offset,
1627 fs_info->sectorsize);
1629 extent_changeset_release(data_reserved);
1630 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1633 if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1634 BTRFS_INODE_PREALLOC)) &&
1635 check_can_nocow(BTRFS_I(inode), pos,
1636 &write_bytes) > 0) {
1638 * For nodata cow case, no need to reserve
1641 only_release_metadata = true;
1643 * our prealloc extent may be smaller than
1644 * write_bytes, so scale down.
1646 num_pages = DIV_ROUND_UP(write_bytes + offset,
1648 reserve_bytes = round_up(write_bytes +
1650 fs_info->sectorsize);
1656 WARN_ON(reserve_bytes == 0);
1657 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1660 if (!only_release_metadata)
1661 btrfs_free_reserved_data_space(inode,
1665 btrfs_drew_write_unlock(&root->snapshot_lock);
1669 release_bytes = reserve_bytes;
1672 * This is going to setup the pages array with the number of
1673 * pages we want, so we don't really need to worry about the
1674 * contents of pages from loop to loop
1676 ret = prepare_pages(inode, pages, num_pages,
1678 force_page_uptodate);
1680 btrfs_delalloc_release_extents(BTRFS_I(inode),
1685 extents_locked = lock_and_cleanup_extent_if_need(
1686 BTRFS_I(inode), pages,
1687 num_pages, pos, write_bytes, &lockstart,
1688 &lockend, &cached_state);
1689 if (extents_locked < 0) {
1690 if (extents_locked == -EAGAIN)
1692 btrfs_delalloc_release_extents(BTRFS_I(inode),
1694 ret = extents_locked;
1698 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1700 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1701 dirty_sectors = round_up(copied + sector_offset,
1702 fs_info->sectorsize);
1703 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1706 * if we have trouble faulting in the pages, fall
1707 * back to one page at a time
1709 if (copied < write_bytes)
1713 force_page_uptodate = true;
1717 force_page_uptodate = false;
1718 dirty_pages = DIV_ROUND_UP(copied + offset,
1722 if (num_sectors > dirty_sectors) {
1723 /* release everything except the sectors we dirtied */
1724 release_bytes -= dirty_sectors <<
1725 fs_info->sb->s_blocksize_bits;
1726 if (only_release_metadata) {
1727 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1728 release_bytes, true);
1732 __pos = round_down(pos,
1733 fs_info->sectorsize) +
1734 (dirty_pages << PAGE_SHIFT);
1735 btrfs_delalloc_release_space(inode,
1736 data_reserved, __pos,
1737 release_bytes, true);
1741 release_bytes = round_up(copied + sector_offset,
1742 fs_info->sectorsize);
1745 ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1746 pos, copied, &cached_state);
1749 * If we have not locked the extent range, because the range's
1750 * start offset is >= i_size, we might still have a non-NULL
1751 * cached extent state, acquired while marking the extent range
1752 * as delalloc through btrfs_dirty_pages(). Therefore free any
1753 * possible cached extent state to avoid a memory leak.
1756 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1757 lockstart, lockend, &cached_state);
1759 free_extent_state(cached_state);
1761 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1763 btrfs_drop_pages(pages, num_pages);
1768 if (only_release_metadata)
1769 btrfs_drew_write_unlock(&root->snapshot_lock);
1771 if (only_release_metadata && copied > 0) {
1772 lockstart = round_down(pos,
1773 fs_info->sectorsize);
1774 lockend = round_up(pos + copied,
1775 fs_info->sectorsize) - 1;
1777 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1778 lockend, EXTENT_NORESERVE, NULL,
1782 btrfs_drop_pages(pages, num_pages);
1786 balance_dirty_pages_ratelimited(inode->i_mapping);
1787 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1788 btrfs_btree_balance_dirty(fs_info);
1791 num_written += copied;
1796 if (release_bytes) {
1797 if (only_release_metadata) {
1798 btrfs_drew_write_unlock(&root->snapshot_lock);
1799 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1800 release_bytes, true);
1802 btrfs_delalloc_release_space(inode, data_reserved,
1803 round_down(pos, fs_info->sectorsize),
1804 release_bytes, true);
1808 extent_changeset_free(data_reserved);
1809 return num_written ? num_written : ret;
1812 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1814 struct file *file = iocb->ki_filp;
1815 struct inode *inode = file_inode(file);
1818 ssize_t written_buffered;
1822 written = generic_file_direct_write(iocb, from);
1824 if (written < 0 || !iov_iter_count(from))
1828 written_buffered = btrfs_buffered_write(iocb, from);
1829 if (written_buffered < 0) {
1830 err = written_buffered;
1834 * Ensure all data is persisted. We want the next direct IO read to be
1835 * able to read what was just written.
1837 endbyte = pos + written_buffered - 1;
1838 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1841 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1844 written += written_buffered;
1845 iocb->ki_pos = pos + written_buffered;
1846 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1847 endbyte >> PAGE_SHIFT);
1849 return written ? written : err;
1852 static void update_time_for_write(struct inode *inode)
1854 struct timespec64 now;
1856 if (IS_NOCMTIME(inode))
1859 now = current_time(inode);
1860 if (!timespec64_equal(&inode->i_mtime, &now))
1861 inode->i_mtime = now;
1863 if (!timespec64_equal(&inode->i_ctime, &now))
1864 inode->i_ctime = now;
1866 if (IS_I_VERSION(inode))
1867 inode_inc_iversion(inode);
1870 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1871 struct iov_iter *from)
1873 struct file *file = iocb->ki_filp;
1874 struct inode *inode = file_inode(file);
1875 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1876 struct btrfs_root *root = BTRFS_I(inode)->root;
1879 ssize_t num_written = 0;
1880 const bool sync = iocb->ki_flags & IOCB_DSYNC;
1887 if (!(iocb->ki_flags & IOCB_DIRECT) &&
1888 (iocb->ki_flags & IOCB_NOWAIT))
1891 if (iocb->ki_flags & IOCB_NOWAIT) {
1892 if (!inode_trylock(inode))
1898 err = generic_write_checks(iocb, from);
1900 inode_unlock(inode);
1905 count = iov_iter_count(from);
1906 if (iocb->ki_flags & IOCB_NOWAIT) {
1908 * We will allocate space in case nodatacow is not set,
1911 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1912 BTRFS_INODE_PREALLOC)) ||
1913 check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1914 inode_unlock(inode);
1917 /* check_can_nocow() locks the snapshot lock on success */
1918 btrfs_drew_write_unlock(&root->snapshot_lock);
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(iocb, from);
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 setattr 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;
2059 trace_btrfs_sync_file(file, datasync);
2061 btrfs_init_log_ctx(&ctx, inode);
2064 * Set the range to full if the NO_HOLES feature is not enabled.
2065 * This is to avoid missing file extent items representing holes after
2066 * replaying the log.
2068 if (!btrfs_fs_incompat(fs_info, NO_HOLES)) {
2074 * We write the dirty pages in the range and wait until they complete
2075 * out of the ->i_mutex. If so, we can flush the dirty pages by
2076 * multi-task, and make the performance up. See
2077 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2079 ret = start_ordered_ops(inode, start, end);
2086 * We take the dio_sem here because the tree log stuff can race with
2087 * lockless dio writes and get an extent map logged for an extent we
2088 * never waited on. We need it this high up for lockdep reasons.
2090 down_write(&BTRFS_I(inode)->dio_sem);
2092 atomic_inc(&root->log_batch);
2095 * If the inode needs a full sync, make sure we use a full range to
2096 * avoid log tree corruption, due to hole detection racing with ordered
2097 * extent completion for adjacent ranges and races between logging and
2098 * completion of ordered extents for adjancent ranges - both races
2099 * could lead to file extent items in the log with overlapping ranges.
2100 * Do this while holding the inode lock, to avoid races with other
2103 if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2104 &BTRFS_I(inode)->runtime_flags)) {
2110 * Before we acquired the inode's lock, someone may have dirtied more
2111 * pages in the target range. We need to make sure that writeback for
2112 * any such pages does not start while we are logging the inode, because
2113 * if it does, any of the following might happen when we are not doing a
2116 * 1) We log an extent after its writeback finishes but before its
2117 * checksums are added to the csum tree, leading to -EIO errors
2118 * when attempting to read the extent after a log replay.
2120 * 2) We can end up logging an extent before its writeback finishes.
2121 * Therefore after the log replay we will have a file extent item
2122 * pointing to an unwritten extent (and no data checksums as well).
2124 * So trigger writeback for any eventual new dirty pages and then we
2125 * wait for all ordered extents to complete below.
2127 ret = start_ordered_ops(inode, start, end);
2129 up_write(&BTRFS_I(inode)->dio_sem);
2130 inode_unlock(inode);
2135 * We have to do this here to avoid the priority inversion of waiting on
2136 * IO of a lower priority task while holding a transaction open.
2138 * Also, the range length can be represented by u64, we have to do the
2139 * typecasts to avoid signed overflow if it's [0, LLONG_MAX].
2141 ret = btrfs_wait_ordered_range(inode, start, (u64)end - (u64)start + 1);
2143 up_write(&BTRFS_I(inode)->dio_sem);
2144 inode_unlock(inode);
2147 atomic_inc(&root->log_batch);
2150 if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2151 BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2153 * We've had everything committed since the last time we were
2154 * modified so clear this flag in case it was set for whatever
2155 * reason, it's no longer relevant.
2157 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2158 &BTRFS_I(inode)->runtime_flags);
2160 * An ordered extent might have started before and completed
2161 * already with io errors, in which case the inode was not
2162 * updated and we end up here. So check the inode's mapping
2163 * for any errors that might have happened since we last
2164 * checked called fsync.
2166 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2167 up_write(&BTRFS_I(inode)->dio_sem);
2168 inode_unlock(inode);
2173 * We use start here because we will need to wait on the IO to complete
2174 * in btrfs_sync_log, which could require joining a transaction (for
2175 * example checking cross references in the nocow path). If we use join
2176 * here we could get into a situation where we're waiting on IO to
2177 * happen that is blocked on a transaction trying to commit. With start
2178 * we inc the extwriter counter, so we wait for all extwriters to exit
2179 * before we start blocking joiners. This comment is to keep somebody
2180 * from thinking they are super smart and changing this to
2181 * btrfs_join_transaction *cough*Josef*cough*.
2183 trans = btrfs_start_transaction(root, 0);
2184 if (IS_ERR(trans)) {
2185 ret = PTR_ERR(trans);
2186 up_write(&BTRFS_I(inode)->dio_sem);
2187 inode_unlock(inode);
2191 ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2193 /* Fallthrough and commit/free transaction. */
2197 /* we've logged all the items and now have a consistent
2198 * version of the file in the log. It is possible that
2199 * someone will come in and modify the file, but that's
2200 * fine because the log is consistent on disk, and we
2201 * have references to all of the file's extents
2203 * It is possible that someone will come in and log the
2204 * file again, but that will end up using the synchronization
2205 * inside btrfs_sync_log to keep things safe.
2207 up_write(&BTRFS_I(inode)->dio_sem);
2208 inode_unlock(inode);
2210 if (ret != BTRFS_NO_LOG_SYNC) {
2212 ret = btrfs_sync_log(trans, root, &ctx);
2214 ret = btrfs_end_transaction(trans);
2218 ret = btrfs_commit_transaction(trans);
2220 ret = btrfs_end_transaction(trans);
2223 ASSERT(list_empty(&ctx.list));
2224 err = file_check_and_advance_wb_err(file);
2227 return ret > 0 ? -EIO : ret;
2230 static const struct vm_operations_struct btrfs_file_vm_ops = {
2231 .fault = filemap_fault,
2232 .map_pages = filemap_map_pages,
2233 .page_mkwrite = btrfs_page_mkwrite,
2236 static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2238 struct address_space *mapping = filp->f_mapping;
2240 if (!mapping->a_ops->readpage)
2243 file_accessed(filp);
2244 vma->vm_ops = &btrfs_file_vm_ops;
2249 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2250 int slot, u64 start, u64 end)
2252 struct btrfs_file_extent_item *fi;
2253 struct btrfs_key key;
2255 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2258 btrfs_item_key_to_cpu(leaf, &key, slot);
2259 if (key.objectid != btrfs_ino(inode) ||
2260 key.type != BTRFS_EXTENT_DATA_KEY)
2263 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2265 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2268 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2271 if (key.offset == end)
2273 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2278 static int fill_holes(struct btrfs_trans_handle *trans,
2279 struct btrfs_inode *inode,
2280 struct btrfs_path *path, u64 offset, u64 end)
2282 struct btrfs_fs_info *fs_info = trans->fs_info;
2283 struct btrfs_root *root = inode->root;
2284 struct extent_buffer *leaf;
2285 struct btrfs_file_extent_item *fi;
2286 struct extent_map *hole_em;
2287 struct extent_map_tree *em_tree = &inode->extent_tree;
2288 struct btrfs_key key;
2291 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2294 key.objectid = btrfs_ino(inode);
2295 key.type = BTRFS_EXTENT_DATA_KEY;
2296 key.offset = offset;
2298 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2301 * We should have dropped this offset, so if we find it then
2302 * something has gone horribly wrong.
2309 leaf = path->nodes[0];
2310 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2314 fi = btrfs_item_ptr(leaf, path->slots[0],
2315 struct btrfs_file_extent_item);
2316 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2318 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2319 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2320 btrfs_set_file_extent_offset(leaf, fi, 0);
2321 btrfs_mark_buffer_dirty(leaf);
2325 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2328 key.offset = offset;
2329 btrfs_set_item_key_safe(fs_info, path, &key);
2330 fi = btrfs_item_ptr(leaf, path->slots[0],
2331 struct btrfs_file_extent_item);
2332 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2334 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2335 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2336 btrfs_set_file_extent_offset(leaf, fi, 0);
2337 btrfs_mark_buffer_dirty(leaf);
2340 btrfs_release_path(path);
2342 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2343 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2348 btrfs_release_path(path);
2350 hole_em = alloc_extent_map();
2352 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2353 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2355 hole_em->start = offset;
2356 hole_em->len = end - offset;
2357 hole_em->ram_bytes = hole_em->len;
2358 hole_em->orig_start = offset;
2360 hole_em->block_start = EXTENT_MAP_HOLE;
2361 hole_em->block_len = 0;
2362 hole_em->orig_block_len = 0;
2363 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2364 hole_em->generation = trans->transid;
2367 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2368 write_lock(&em_tree->lock);
2369 ret = add_extent_mapping(em_tree, hole_em, 1);
2370 write_unlock(&em_tree->lock);
2371 } while (ret == -EEXIST);
2372 free_extent_map(hole_em);
2374 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2375 &inode->runtime_flags);
2382 * Find a hole extent on given inode and change start/len to the end of hole
2383 * extent.(hole/vacuum extent whose em->start <= start &&
2384 * em->start + em->len > start)
2385 * When a hole extent is found, return 1 and modify start/len.
2387 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2389 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2390 struct extent_map *em;
2393 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2394 round_down(*start, fs_info->sectorsize),
2395 round_up(*len, fs_info->sectorsize));
2399 /* Hole or vacuum extent(only exists in no-hole mode) */
2400 if (em->block_start == EXTENT_MAP_HOLE) {
2402 *len = em->start + em->len > *start + *len ?
2403 0 : *start + *len - em->start - em->len;
2404 *start = em->start + em->len;
2406 free_extent_map(em);
2410 static int btrfs_punch_hole_lock_range(struct inode *inode,
2411 const u64 lockstart,
2413 struct extent_state **cached_state)
2416 struct btrfs_ordered_extent *ordered;
2419 truncate_pagecache_range(inode, lockstart, lockend);
2421 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2423 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2426 * We need to make sure we have no ordered extents in this range
2427 * and nobody raced in and read a page in this range, if we did
2428 * we need to try again.
2431 (ordered->file_offset + ordered->num_bytes <= lockstart ||
2432 ordered->file_offset > lockend)) &&
2433 !filemap_range_has_page(inode->i_mapping,
2434 lockstart, lockend)) {
2436 btrfs_put_ordered_extent(ordered);
2440 btrfs_put_ordered_extent(ordered);
2441 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2442 lockend, cached_state);
2443 ret = btrfs_wait_ordered_range(inode, lockstart,
2444 lockend - lockstart + 1);
2451 static int btrfs_insert_clone_extent(struct btrfs_trans_handle *trans,
2452 struct inode *inode,
2453 struct btrfs_path *path,
2454 struct btrfs_clone_extent_info *clone_info,
2455 const u64 clone_len)
2457 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2458 struct btrfs_root *root = BTRFS_I(inode)->root;
2459 struct btrfs_file_extent_item *extent;
2460 struct extent_buffer *leaf;
2461 struct btrfs_key key;
2463 struct btrfs_ref ref = { 0 };
2470 if (clone_info->disk_offset == 0 &&
2471 btrfs_fs_incompat(fs_info, NO_HOLES))
2474 key.objectid = btrfs_ino(BTRFS_I(inode));
2475 key.type = BTRFS_EXTENT_DATA_KEY;
2476 key.offset = clone_info->file_offset;
2477 ret = btrfs_insert_empty_item(trans, root, path, &key,
2478 clone_info->item_size);
2481 leaf = path->nodes[0];
2482 slot = path->slots[0];
2483 write_extent_buffer(leaf, clone_info->extent_buf,
2484 btrfs_item_ptr_offset(leaf, slot),
2485 clone_info->item_size);
2486 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2487 btrfs_set_file_extent_offset(leaf, extent, clone_info->data_offset);
2488 btrfs_set_file_extent_num_bytes(leaf, extent, clone_len);
2489 btrfs_mark_buffer_dirty(leaf);
2490 btrfs_release_path(path);
2492 ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
2493 clone_info->file_offset, clone_len);
2497 /* If it's a hole, nothing more needs to be done. */
2498 if (clone_info->disk_offset == 0)
2501 inode_add_bytes(inode, clone_len);
2502 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2503 clone_info->disk_offset,
2504 clone_info->disk_len, 0);
2505 ref_offset = clone_info->file_offset - clone_info->data_offset;
2506 btrfs_init_data_ref(&ref, root->root_key.objectid,
2507 btrfs_ino(BTRFS_I(inode)), ref_offset);
2508 ret = btrfs_inc_extent_ref(trans, &ref);
2514 * The respective range must have been previously locked, as well as the inode.
2515 * The end offset is inclusive (last byte of the range).
2516 * @clone_info is NULL for fallocate's hole punching and non-NULL for extent
2518 * When cloning, we don't want to end up in a state where we dropped extents
2519 * without inserting a new one, so we must abort the transaction to avoid a
2522 int btrfs_punch_hole_range(struct inode *inode, struct btrfs_path *path,
2523 const u64 start, const u64 end,
2524 struct btrfs_clone_extent_info *clone_info,
2525 struct btrfs_trans_handle **trans_out)
2527 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2528 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2529 u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2530 struct btrfs_root *root = BTRFS_I(inode)->root;
2531 struct btrfs_trans_handle *trans = NULL;
2532 struct btrfs_block_rsv *rsv;
2533 unsigned int rsv_count;
2536 u64 len = end - start;
2542 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2547 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2551 * 1 - update the inode
2552 * 1 - removing the extents in the range
2553 * 1 - adding the hole extent if no_holes isn't set or if we are cloning
2556 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || clone_info)
2561 trans = btrfs_start_transaction(root, rsv_count);
2562 if (IS_ERR(trans)) {
2563 ret = PTR_ERR(trans);
2568 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2571 trans->block_rsv = rsv;
2574 while (cur_offset < end) {
2575 ret = __btrfs_drop_extents(trans, root, inode, path,
2576 cur_offset, end + 1, &drop_end,
2578 if (ret != -ENOSPC) {
2580 * When cloning we want to avoid transaction aborts when
2581 * nothing was done and we are attempting to clone parts
2582 * of inline extents, in such cases -EOPNOTSUPP is
2583 * returned by __btrfs_drop_extents() without having
2584 * changed anything in the file.
2586 if (clone_info && ret && ret != -EOPNOTSUPP)
2587 btrfs_abort_transaction(trans, ret);
2591 trans->block_rsv = &fs_info->trans_block_rsv;
2593 if (!clone_info && cur_offset < drop_end &&
2594 cur_offset < ino_size) {
2595 ret = fill_holes(trans, BTRFS_I(inode), path,
2596 cur_offset, drop_end);
2599 * If we failed then we didn't insert our hole
2600 * entries for the area we dropped, so now the
2601 * fs is corrupted, so we must abort the
2604 btrfs_abort_transaction(trans, ret);
2607 } else if (!clone_info && cur_offset < drop_end) {
2609 * We are past the i_size here, but since we didn't
2610 * insert holes we need to clear the mapped area so we
2611 * know to not set disk_i_size in this area until a new
2612 * file extent is inserted here.
2614 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2615 cur_offset, drop_end - cur_offset);
2618 * We couldn't clear our area, so we could
2619 * presumably adjust up and corrupt the fs, so
2622 btrfs_abort_transaction(trans, ret);
2627 if (clone_info && drop_end > clone_info->file_offset) {
2628 u64 clone_len = drop_end - clone_info->file_offset;
2630 ret = btrfs_insert_clone_extent(trans, inode, path,
2631 clone_info, clone_len);
2633 btrfs_abort_transaction(trans, ret);
2636 clone_info->data_len -= clone_len;
2637 clone_info->data_offset += clone_len;
2638 clone_info->file_offset += clone_len;
2641 cur_offset = drop_end;
2643 ret = btrfs_update_inode(trans, root, inode);
2647 btrfs_end_transaction(trans);
2648 btrfs_btree_balance_dirty(fs_info);
2650 trans = btrfs_start_transaction(root, rsv_count);
2651 if (IS_ERR(trans)) {
2652 ret = PTR_ERR(trans);
2657 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2658 rsv, min_size, false);
2659 BUG_ON(ret); /* shouldn't happen */
2660 trans->block_rsv = rsv;
2663 ret = find_first_non_hole(inode, &cur_offset, &len);
2664 if (unlikely(ret < 0))
2674 * If we were cloning, force the next fsync to be a full one since we
2675 * we replaced (or just dropped in the case of cloning holes when
2676 * NO_HOLES is enabled) extents and extent maps.
2677 * This is for the sake of simplicity, and cloning into files larger
2678 * than 16Mb would force the full fsync any way (when
2679 * try_release_extent_mapping() is invoked during page cache truncation.
2682 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2683 &BTRFS_I(inode)->runtime_flags);
2688 trans->block_rsv = &fs_info->trans_block_rsv;
2690 * If we are using the NO_HOLES feature we might have had already an
2691 * hole that overlaps a part of the region [lockstart, lockend] and
2692 * ends at (or beyond) lockend. Since we have no file extent items to
2693 * represent holes, drop_end can be less than lockend and so we must
2694 * make sure we have an extent map representing the existing hole (the
2695 * call to __btrfs_drop_extents() might have dropped the existing extent
2696 * map representing the existing hole), otherwise the fast fsync path
2697 * will not record the existence of the hole region
2698 * [existing_hole_start, lockend].
2700 if (drop_end <= end)
2703 * Don't insert file hole extent item if it's for a range beyond eof
2704 * (because it's useless) or if it represents a 0 bytes range (when
2705 * cur_offset == drop_end).
2707 if (!clone_info && cur_offset < ino_size && cur_offset < drop_end) {
2708 ret = fill_holes(trans, BTRFS_I(inode), path,
2709 cur_offset, drop_end);
2711 /* Same comment as above. */
2712 btrfs_abort_transaction(trans, ret);
2715 } else if (!clone_info && cur_offset < drop_end) {
2716 /* See the comment in the loop above for the reasoning here. */
2717 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2718 cur_offset, drop_end - cur_offset);
2720 btrfs_abort_transaction(trans, ret);
2726 ret = btrfs_insert_clone_extent(trans, inode, path, clone_info,
2727 clone_info->data_len);
2729 btrfs_abort_transaction(trans, ret);
2738 trans->block_rsv = &fs_info->trans_block_rsv;
2740 btrfs_end_transaction(trans);
2744 btrfs_free_block_rsv(fs_info, rsv);
2749 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2751 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2752 struct btrfs_root *root = BTRFS_I(inode)->root;
2753 struct extent_state *cached_state = NULL;
2754 struct btrfs_path *path;
2755 struct btrfs_trans_handle *trans = NULL;
2760 u64 orig_start = offset;
2764 bool truncated_block = false;
2765 bool updated_inode = false;
2767 ret = btrfs_wait_ordered_range(inode, offset, len);
2772 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2773 ret = find_first_non_hole(inode, &offset, &len);
2775 goto out_only_mutex;
2777 /* Already in a large hole */
2779 goto out_only_mutex;
2782 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2783 lockend = round_down(offset + len,
2784 btrfs_inode_sectorsize(inode)) - 1;
2785 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2786 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2788 * We needn't truncate any block which is beyond the end of the file
2789 * because we are sure there is no data there.
2792 * Only do this if we are in the same block and we aren't doing the
2795 if (same_block && len < fs_info->sectorsize) {
2796 if (offset < ino_size) {
2797 truncated_block = true;
2798 ret = btrfs_truncate_block(inode, offset, len, 0);
2802 goto out_only_mutex;
2805 /* zero back part of the first block */
2806 if (offset < ino_size) {
2807 truncated_block = true;
2808 ret = btrfs_truncate_block(inode, offset, 0, 0);
2810 inode_unlock(inode);
2815 /* Check the aligned pages after the first unaligned page,
2816 * if offset != orig_start, which means the first unaligned page
2817 * including several following pages are already in holes,
2818 * the extra check can be skipped */
2819 if (offset == orig_start) {
2820 /* after truncate page, check hole again */
2821 len = offset + len - lockstart;
2823 ret = find_first_non_hole(inode, &offset, &len);
2825 goto out_only_mutex;
2828 goto out_only_mutex;
2833 /* Check the tail unaligned part is in a hole */
2834 tail_start = lockend + 1;
2835 tail_len = offset + len - tail_start;
2837 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2838 if (unlikely(ret < 0))
2839 goto out_only_mutex;
2841 /* zero the front end of the last page */
2842 if (tail_start + tail_len < ino_size) {
2843 truncated_block = true;
2844 ret = btrfs_truncate_block(inode,
2845 tail_start + tail_len,
2848 goto out_only_mutex;
2853 if (lockend < lockstart) {
2855 goto out_only_mutex;
2858 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2861 goto out_only_mutex;
2863 path = btrfs_alloc_path();
2869 ret = btrfs_punch_hole_range(inode, path, lockstart, lockend, NULL,
2871 btrfs_free_path(path);
2875 ASSERT(trans != NULL);
2876 inode_inc_iversion(inode);
2877 inode->i_mtime = inode->i_ctime = current_time(inode);
2878 ret = btrfs_update_inode(trans, root, inode);
2879 updated_inode = true;
2880 btrfs_end_transaction(trans);
2881 btrfs_btree_balance_dirty(fs_info);
2883 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2886 if (!updated_inode && truncated_block && !ret) {
2888 * If we only end up zeroing part of a page, we still need to
2889 * update the inode item, so that all the time fields are
2890 * updated as well as the necessary btrfs inode in memory fields
2891 * for detecting, at fsync time, if the inode isn't yet in the
2892 * log tree or it's there but not up to date.
2894 struct timespec64 now = current_time(inode);
2896 inode_inc_iversion(inode);
2897 inode->i_mtime = now;
2898 inode->i_ctime = now;
2899 trans = btrfs_start_transaction(root, 1);
2900 if (IS_ERR(trans)) {
2901 ret = PTR_ERR(trans);
2905 ret = btrfs_update_inode(trans, root, inode);
2906 ret2 = btrfs_end_transaction(trans);
2911 inode_unlock(inode);
2915 /* Helper structure to record which range is already reserved */
2916 struct falloc_range {
2917 struct list_head list;
2923 * Helper function to add falloc range
2925 * Caller should have locked the larger range of extent containing
2928 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2930 struct falloc_range *prev = NULL;
2931 struct falloc_range *range = NULL;
2933 if (list_empty(head))
2937 * As fallocate iterate by bytenr order, we only need to check
2940 prev = list_entry(head->prev, struct falloc_range, list);
2941 if (prev->start + prev->len == start) {
2946 range = kmalloc(sizeof(*range), GFP_KERNEL);
2949 range->start = start;
2951 list_add_tail(&range->list, head);
2955 static int btrfs_fallocate_update_isize(struct inode *inode,
2959 struct btrfs_trans_handle *trans;
2960 struct btrfs_root *root = BTRFS_I(inode)->root;
2964 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2967 trans = btrfs_start_transaction(root, 1);
2969 return PTR_ERR(trans);
2971 inode->i_ctime = current_time(inode);
2972 i_size_write(inode, end);
2973 btrfs_inode_safe_disk_i_size_write(inode, 0);
2974 ret = btrfs_update_inode(trans, root, inode);
2975 ret2 = btrfs_end_transaction(trans);
2977 return ret ? ret : ret2;
2981 RANGE_BOUNDARY_WRITTEN_EXTENT,
2982 RANGE_BOUNDARY_PREALLOC_EXTENT,
2983 RANGE_BOUNDARY_HOLE,
2986 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2989 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2990 struct extent_map *em;
2993 offset = round_down(offset, sectorsize);
2994 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize);
2998 if (em->block_start == EXTENT_MAP_HOLE)
2999 ret = RANGE_BOUNDARY_HOLE;
3000 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3001 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3003 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3005 free_extent_map(em);
3009 static int btrfs_zero_range(struct inode *inode,
3014 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3015 struct extent_map *em;
3016 struct extent_changeset *data_reserved = NULL;
3019 const u64 sectorsize = btrfs_inode_sectorsize(inode);
3020 u64 alloc_start = round_down(offset, sectorsize);
3021 u64 alloc_end = round_up(offset + len, sectorsize);
3022 u64 bytes_to_reserve = 0;
3023 bool space_reserved = false;
3025 inode_dio_wait(inode);
3027 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3028 alloc_end - alloc_start);
3035 * Avoid hole punching and extent allocation for some cases. More cases
3036 * could be considered, but these are unlikely common and we keep things
3037 * as simple as possible for now. Also, intentionally, if the target
3038 * range contains one or more prealloc extents together with regular
3039 * extents and holes, we drop all the existing extents and allocate a
3040 * new prealloc extent, so that we get a larger contiguous disk extent.
3042 if (em->start <= alloc_start &&
3043 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3044 const u64 em_end = em->start + em->len;
3046 if (em_end >= offset + len) {
3048 * The whole range is already a prealloc extent,
3049 * do nothing except updating the inode's i_size if
3052 free_extent_map(em);
3053 ret = btrfs_fallocate_update_isize(inode, offset + len,
3058 * Part of the range is already a prealloc extent, so operate
3059 * only on the remaining part of the range.
3061 alloc_start = em_end;
3062 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3063 len = offset + len - alloc_start;
3064 offset = alloc_start;
3065 alloc_hint = em->block_start + em->len;
3067 free_extent_map(em);
3069 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3070 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3071 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3078 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3079 free_extent_map(em);
3080 ret = btrfs_fallocate_update_isize(inode, offset + len,
3084 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3085 free_extent_map(em);
3086 ret = btrfs_truncate_block(inode, offset, len, 0);
3088 ret = btrfs_fallocate_update_isize(inode,
3093 free_extent_map(em);
3094 alloc_start = round_down(offset, sectorsize);
3095 alloc_end = alloc_start + sectorsize;
3099 alloc_start = round_up(offset, sectorsize);
3100 alloc_end = round_down(offset + len, sectorsize);
3103 * For unaligned ranges, check the pages at the boundaries, they might
3104 * map to an extent, in which case we need to partially zero them, or
3105 * they might map to a hole, in which case we need our allocation range
3108 if (!IS_ALIGNED(offset, sectorsize)) {
3109 ret = btrfs_zero_range_check_range_boundary(inode, offset);
3112 if (ret == RANGE_BOUNDARY_HOLE) {
3113 alloc_start = round_down(offset, sectorsize);
3115 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3116 ret = btrfs_truncate_block(inode, offset, 0, 0);
3124 if (!IS_ALIGNED(offset + len, sectorsize)) {
3125 ret = btrfs_zero_range_check_range_boundary(inode,
3129 if (ret == RANGE_BOUNDARY_HOLE) {
3130 alloc_end = round_up(offset + len, sectorsize);
3132 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3133 ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3142 if (alloc_start < alloc_end) {
3143 struct extent_state *cached_state = NULL;
3144 const u64 lockstart = alloc_start;
3145 const u64 lockend = alloc_end - 1;
3147 bytes_to_reserve = alloc_end - alloc_start;
3148 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3152 space_reserved = true;
3153 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3154 alloc_start, bytes_to_reserve);
3157 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3161 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3162 alloc_end - alloc_start,
3164 offset + len, &alloc_hint);
3165 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3166 lockend, &cached_state);
3167 /* btrfs_prealloc_file_range releases reserved space on error */
3169 space_reserved = false;
3173 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3175 if (ret && space_reserved)
3176 btrfs_free_reserved_data_space(inode, data_reserved,
3177 alloc_start, bytes_to_reserve);
3178 extent_changeset_free(data_reserved);
3183 static long btrfs_fallocate(struct file *file, int mode,
3184 loff_t offset, loff_t len)
3186 struct inode *inode = file_inode(file);
3187 struct extent_state *cached_state = NULL;
3188 struct extent_changeset *data_reserved = NULL;
3189 struct falloc_range *range;
3190 struct falloc_range *tmp;
3191 struct list_head reserve_list;
3199 struct extent_map *em;
3200 int blocksize = btrfs_inode_sectorsize(inode);
3203 alloc_start = round_down(offset, blocksize);
3204 alloc_end = round_up(offset + len, blocksize);
3205 cur_offset = alloc_start;
3207 /* Make sure we aren't being give some crap mode */
3208 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3209 FALLOC_FL_ZERO_RANGE))
3212 if (mode & FALLOC_FL_PUNCH_HOLE)
3213 return btrfs_punch_hole(inode, offset, len);
3216 * Only trigger disk allocation, don't trigger qgroup reserve
3218 * For qgroup space, it will be checked later.
3220 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3221 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3222 alloc_end - alloc_start);
3229 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3230 ret = inode_newsize_ok(inode, offset + len);
3236 * TODO: Move these two operations after we have checked
3237 * accurate reserved space, or fallocate can still fail but
3238 * with page truncated or size expanded.
3240 * But that's a minor problem and won't do much harm BTW.
3242 if (alloc_start > inode->i_size) {
3243 ret = btrfs_cont_expand(inode, i_size_read(inode),
3247 } else if (offset + len > inode->i_size) {
3249 * If we are fallocating from the end of the file onward we
3250 * need to zero out the end of the block if i_size lands in the
3251 * middle of a block.
3253 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3259 * wait for ordered IO before we have any locks. We'll loop again
3260 * below with the locks held.
3262 ret = btrfs_wait_ordered_range(inode, alloc_start,
3263 alloc_end - alloc_start);
3267 if (mode & FALLOC_FL_ZERO_RANGE) {
3268 ret = btrfs_zero_range(inode, offset, len, mode);
3269 inode_unlock(inode);
3273 locked_end = alloc_end - 1;
3275 struct btrfs_ordered_extent *ordered;
3277 /* the extent lock is ordered inside the running
3280 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3281 locked_end, &cached_state);
3282 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3285 ordered->file_offset + ordered->num_bytes > alloc_start &&
3286 ordered->file_offset < alloc_end) {
3287 btrfs_put_ordered_extent(ordered);
3288 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3289 alloc_start, locked_end,
3292 * we can't wait on the range with the transaction
3293 * running or with the extent lock held
3295 ret = btrfs_wait_ordered_range(inode, alloc_start,
3296 alloc_end - alloc_start);
3301 btrfs_put_ordered_extent(ordered);
3306 /* First, check if we exceed the qgroup limit */
3307 INIT_LIST_HEAD(&reserve_list);
3308 while (cur_offset < alloc_end) {
3309 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3310 alloc_end - cur_offset);
3315 last_byte = min(extent_map_end(em), alloc_end);
3316 actual_end = min_t(u64, extent_map_end(em), offset + len);
3317 last_byte = ALIGN(last_byte, blocksize);
3318 if (em->block_start == EXTENT_MAP_HOLE ||
3319 (cur_offset >= inode->i_size &&
3320 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3321 ret = add_falloc_range(&reserve_list, cur_offset,
3322 last_byte - cur_offset);
3324 free_extent_map(em);
3327 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3328 cur_offset, last_byte - cur_offset);
3330 cur_offset = last_byte;
3331 free_extent_map(em);
3336 * Do not need to reserve unwritten extent for this
3337 * range, free reserved data space first, otherwise
3338 * it'll result in false ENOSPC error.
3340 btrfs_free_reserved_data_space(inode, data_reserved,
3341 cur_offset, last_byte - cur_offset);
3343 free_extent_map(em);
3344 cur_offset = last_byte;
3348 * If ret is still 0, means we're OK to fallocate.
3349 * Or just cleanup the list and exit.
3351 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3353 ret = btrfs_prealloc_file_range(inode, mode,
3355 range->len, i_blocksize(inode),
3356 offset + len, &alloc_hint);
3358 btrfs_free_reserved_data_space(inode,
3359 data_reserved, range->start,
3361 list_del(&range->list);
3368 * We didn't need to allocate any more space, but we still extended the
3369 * size of the file so we need to update i_size and the inode item.
3371 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3373 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3376 inode_unlock(inode);
3377 /* Let go of our reservation. */
3378 if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3379 btrfs_free_reserved_data_space(inode, data_reserved,
3380 cur_offset, alloc_end - cur_offset);
3381 extent_changeset_free(data_reserved);
3385 static loff_t find_desired_extent(struct inode *inode, loff_t offset,
3388 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3389 struct extent_map *em = NULL;
3390 struct extent_state *cached_state = NULL;
3391 loff_t i_size = inode->i_size;
3398 if (i_size == 0 || offset >= i_size)
3402 * offset can be negative, in this case we start finding DATA/HOLE from
3403 * the very start of the file.
3405 start = max_t(loff_t, 0, offset);
3407 lockstart = round_down(start, fs_info->sectorsize);
3408 lockend = round_up(i_size, fs_info->sectorsize);
3409 if (lockend <= lockstart)
3410 lockend = lockstart + fs_info->sectorsize;
3412 len = lockend - lockstart + 1;
3414 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3417 while (start < i_size) {
3418 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3425 if (whence == SEEK_HOLE &&
3426 (em->block_start == EXTENT_MAP_HOLE ||
3427 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3429 else if (whence == SEEK_DATA &&
3430 (em->block_start != EXTENT_MAP_HOLE &&
3431 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3434 start = em->start + em->len;
3435 free_extent_map(em);
3439 free_extent_map(em);
3440 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3445 if (whence == SEEK_DATA && start >= i_size)
3448 offset = min_t(loff_t, start, i_size);
3454 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3456 struct inode *inode = file->f_mapping->host;
3460 return generic_file_llseek(file, offset, whence);
3463 inode_lock_shared(inode);
3464 offset = find_desired_extent(inode, offset, whence);
3465 inode_unlock_shared(inode);
3472 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3475 static int btrfs_file_open(struct inode *inode, struct file *filp)
3477 filp->f_mode |= FMODE_NOWAIT;
3478 return generic_file_open(inode, filp);
3481 const struct file_operations btrfs_file_operations = {
3482 .llseek = btrfs_file_llseek,
3483 .read_iter = generic_file_read_iter,
3484 .splice_read = generic_file_splice_read,
3485 .write_iter = btrfs_file_write_iter,
3486 .mmap = btrfs_file_mmap,
3487 .open = btrfs_file_open,
3488 .release = btrfs_release_file,
3489 .fsync = btrfs_sync_file,
3490 .fallocate = btrfs_fallocate,
3491 .unlocked_ioctl = btrfs_ioctl,
3492 #ifdef CONFIG_COMPAT
3493 .compat_ioctl = btrfs_compat_ioctl,
3495 .remap_file_range = btrfs_remap_file_range,
3498 void __cold btrfs_auto_defrag_exit(void)
3500 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3503 int __init btrfs_auto_defrag_init(void)
3505 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3506 sizeof(struct inode_defrag), 0,
3509 if (!btrfs_inode_defrag_cachep)
3515 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3520 * So with compression we will find and lock a dirty page and clear the
3521 * first one as dirty, setup an async extent, and immediately return
3522 * with the entire range locked but with nobody actually marked with
3523 * writeback. So we can't just filemap_write_and_wait_range() and
3524 * expect it to work since it will just kick off a thread to do the
3525 * actual work. So we need to call filemap_fdatawrite_range _again_
3526 * since it will wait on the page lock, which won't be unlocked until
3527 * after the pages have been marked as writeback and so we're good to go
3528 * from there. We have to do this otherwise we'll miss the ordered
3529 * extents and that results in badness. Please Josef, do not think you
3530 * know better and pull this out at some point in the future, it is
3531 * right and you are wrong.
3533 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3534 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3535 &BTRFS_I(inode)->runtime_flags))
3536 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);