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>
19 #include <linux/fsverity.h>
22 #include "transaction.h"
23 #include "btrfs_inode.h"
24 #include "print-tree.h"
29 #include "compression.h"
30 #include "delalloc-space.h"
34 static struct kmem_cache *btrfs_inode_defrag_cachep;
36 * when auto defrag is enabled we
37 * queue up these defrag structs to remember which
38 * inodes need defragging passes
41 struct rb_node rb_node;
45 * transid where the defrag was added, we search for
46 * extents newer than this
54 * The extent size threshold for autodefrag.
56 * This value is different for compressed/non-compressed extents,
57 * thus needs to be passed from higher layer.
58 * (aka, inode_should_defrag())
63 static int __compare_inode_defrag(struct inode_defrag *defrag1,
64 struct inode_defrag *defrag2)
66 if (defrag1->root > defrag2->root)
68 else if (defrag1->root < defrag2->root)
70 else if (defrag1->ino > defrag2->ino)
72 else if (defrag1->ino < defrag2->ino)
78 /* pop a record for an inode into the defrag tree. The lock
79 * must be held already
81 * If you're inserting a record for an older transid than an
82 * existing record, the transid already in the tree is lowered
84 * If an existing record is found the defrag item you
87 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
88 struct inode_defrag *defrag)
90 struct btrfs_fs_info *fs_info = inode->root->fs_info;
91 struct inode_defrag *entry;
93 struct rb_node *parent = NULL;
96 p = &fs_info->defrag_inodes.rb_node;
99 entry = rb_entry(parent, struct inode_defrag, rb_node);
101 ret = __compare_inode_defrag(defrag, entry);
103 p = &parent->rb_left;
105 p = &parent->rb_right;
107 /* if we're reinserting an entry for
108 * an old defrag run, make sure to
109 * lower the transid of our existing record
111 if (defrag->transid < entry->transid)
112 entry->transid = defrag->transid;
113 entry->extent_thresh = min(defrag->extent_thresh,
114 entry->extent_thresh);
118 set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
119 rb_link_node(&defrag->rb_node, parent, p);
120 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
124 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
126 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
129 if (btrfs_fs_closing(fs_info))
136 * insert a defrag record for this inode if auto defrag is
139 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
140 struct btrfs_inode *inode, u32 extent_thresh)
142 struct btrfs_root *root = inode->root;
143 struct btrfs_fs_info *fs_info = root->fs_info;
144 struct inode_defrag *defrag;
148 if (!__need_auto_defrag(fs_info))
151 if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
155 transid = trans->transid;
157 transid = inode->root->last_trans;
159 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
163 defrag->ino = btrfs_ino(inode);
164 defrag->transid = transid;
165 defrag->root = root->root_key.objectid;
166 defrag->extent_thresh = extent_thresh;
168 spin_lock(&fs_info->defrag_inodes_lock);
169 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
171 * If we set IN_DEFRAG flag and evict the inode from memory,
172 * and then re-read this inode, this new inode doesn't have
173 * IN_DEFRAG flag. At the case, we may find the existed defrag.
175 ret = __btrfs_add_inode_defrag(inode, defrag);
177 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
179 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
181 spin_unlock(&fs_info->defrag_inodes_lock);
186 * pick the defragable inode that we want, if it doesn't exist, we will get
189 static struct inode_defrag *
190 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
192 struct inode_defrag *entry = NULL;
193 struct inode_defrag tmp;
195 struct rb_node *parent = NULL;
201 spin_lock(&fs_info->defrag_inodes_lock);
202 p = fs_info->defrag_inodes.rb_node;
205 entry = rb_entry(parent, struct inode_defrag, rb_node);
207 ret = __compare_inode_defrag(&tmp, entry);
211 p = parent->rb_right;
216 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
217 parent = rb_next(parent);
219 entry = rb_entry(parent, struct inode_defrag, rb_node);
225 rb_erase(parent, &fs_info->defrag_inodes);
226 spin_unlock(&fs_info->defrag_inodes_lock);
230 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
232 struct inode_defrag *defrag;
233 struct rb_node *node;
235 spin_lock(&fs_info->defrag_inodes_lock);
236 node = rb_first(&fs_info->defrag_inodes);
238 rb_erase(node, &fs_info->defrag_inodes);
239 defrag = rb_entry(node, struct inode_defrag, rb_node);
240 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
242 cond_resched_lock(&fs_info->defrag_inodes_lock);
244 node = rb_first(&fs_info->defrag_inodes);
246 spin_unlock(&fs_info->defrag_inodes_lock);
249 #define BTRFS_DEFRAG_BATCH 1024
251 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
252 struct inode_defrag *defrag)
254 struct btrfs_root *inode_root;
256 struct btrfs_ioctl_defrag_range_args range;
261 if (test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state))
263 if (!__need_auto_defrag(fs_info))
267 inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
268 if (IS_ERR(inode_root)) {
269 ret = PTR_ERR(inode_root);
273 inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
274 btrfs_put_root(inode_root);
276 ret = PTR_ERR(inode);
280 if (cur >= i_size_read(inode)) {
285 /* do a chunk of defrag */
286 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
287 memset(&range, 0, sizeof(range));
290 range.extent_thresh = defrag->extent_thresh;
292 sb_start_write(fs_info->sb);
293 ret = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
295 sb_end_write(fs_info->sb);
301 cur = max(cur + fs_info->sectorsize, range.start);
305 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
310 * run through the list of inodes in the FS that need
313 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
315 struct inode_defrag *defrag;
317 u64 root_objectid = 0;
319 atomic_inc(&fs_info->defrag_running);
321 /* Pause the auto defragger. */
322 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
326 if (!__need_auto_defrag(fs_info))
329 /* find an inode to defrag */
330 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
333 if (root_objectid || first_ino) {
342 first_ino = defrag->ino + 1;
343 root_objectid = defrag->root;
345 __btrfs_run_defrag_inode(fs_info, defrag);
347 atomic_dec(&fs_info->defrag_running);
350 * during unmount, we use the transaction_wait queue to
351 * wait for the defragger to stop
353 wake_up(&fs_info->transaction_wait);
357 /* simple helper to fault in pages and copy. This should go away
358 * and be replaced with calls into generic code.
360 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
361 struct page **prepared_pages,
365 size_t total_copied = 0;
367 int offset = offset_in_page(pos);
369 while (write_bytes > 0) {
370 size_t count = min_t(size_t,
371 PAGE_SIZE - offset, write_bytes);
372 struct page *page = prepared_pages[pg];
374 * Copy data from userspace to the current page
376 copied = copy_page_from_iter_atomic(page, offset, count, i);
378 /* Flush processor's dcache for this page */
379 flush_dcache_page(page);
382 * if we get a partial write, we can end up with
383 * partially up to date pages. These add
384 * a lot of complexity, so make sure they don't
385 * happen by forcing this copy to be retried.
387 * The rest of the btrfs_file_write code will fall
388 * back to page at a time copies after we return 0.
390 if (unlikely(copied < count)) {
391 if (!PageUptodate(page)) {
392 iov_iter_revert(i, copied);
399 write_bytes -= copied;
400 total_copied += copied;
402 if (offset == PAGE_SIZE) {
411 * unlocks pages after btrfs_file_write is done with them
413 static void btrfs_drop_pages(struct btrfs_fs_info *fs_info,
414 struct page **pages, size_t num_pages,
418 u64 block_start = round_down(pos, fs_info->sectorsize);
419 u64 block_len = round_up(pos + copied, fs_info->sectorsize) - block_start;
421 ASSERT(block_len <= U32_MAX);
422 for (i = 0; i < num_pages; i++) {
423 /* page checked is some magic around finding pages that
424 * have been modified without going through btrfs_set_page_dirty
425 * clear it here. There should be no need to mark the pages
426 * accessed as prepare_pages should have marked them accessed
427 * in prepare_pages via find_or_create_page()
429 btrfs_page_clamp_clear_checked(fs_info, pages[i], block_start,
431 unlock_page(pages[i]);
437 * After btrfs_copy_from_user(), update the following things for delalloc:
438 * - Mark newly dirtied pages as DELALLOC in the io tree.
439 * Used to advise which range is to be written back.
440 * - Mark modified pages as Uptodate/Dirty and not needing COW fixup
441 * - Update inode size for past EOF write
443 int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
444 size_t num_pages, loff_t pos, size_t write_bytes,
445 struct extent_state **cached, bool noreserve)
447 struct btrfs_fs_info *fs_info = inode->root->fs_info;
452 u64 end_of_last_block;
453 u64 end_pos = pos + write_bytes;
454 loff_t isize = i_size_read(&inode->vfs_inode);
455 unsigned int extra_bits = 0;
457 if (write_bytes == 0)
461 extra_bits |= EXTENT_NORESERVE;
463 start_pos = round_down(pos, fs_info->sectorsize);
464 num_bytes = round_up(write_bytes + pos - start_pos,
465 fs_info->sectorsize);
466 ASSERT(num_bytes <= U32_MAX);
468 end_of_last_block = start_pos + num_bytes - 1;
471 * The pages may have already been dirty, clear out old accounting so
472 * we can set things up properly
474 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
475 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
478 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
483 for (i = 0; i < num_pages; i++) {
484 struct page *p = pages[i];
486 btrfs_page_clamp_set_uptodate(fs_info, p, start_pos, num_bytes);
487 btrfs_page_clamp_clear_checked(fs_info, p, start_pos, num_bytes);
488 btrfs_page_clamp_set_dirty(fs_info, p, start_pos, num_bytes);
492 * we've only changed i_size in ram, and we haven't updated
493 * the disk i_size. There is no need to log the inode
497 i_size_write(&inode->vfs_inode, end_pos);
502 * this is very complex, but the basic idea is to drop all extents
503 * in the range start - end. hint_block is filled in with a block number
504 * that would be a good hint to the block allocator for this file.
506 * If an extent intersects the range but is not entirely inside the range
507 * it is either truncated or split. Anything entirely inside the range
508 * is deleted from the tree.
510 * Note: the VFS' inode number of bytes is not updated, it's up to the caller
511 * to deal with that. We set the field 'bytes_found' of the arguments structure
512 * with the number of allocated bytes found in the target range, so that the
513 * caller can update the inode's number of bytes in an atomic way when
514 * replacing extents in a range to avoid races with stat(2).
516 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
517 struct btrfs_root *root, struct btrfs_inode *inode,
518 struct btrfs_drop_extents_args *args)
520 struct btrfs_fs_info *fs_info = root->fs_info;
521 struct extent_buffer *leaf;
522 struct btrfs_file_extent_item *fi;
523 struct btrfs_ref ref = { 0 };
524 struct btrfs_key key;
525 struct btrfs_key new_key;
526 u64 ino = btrfs_ino(inode);
527 u64 search_start = args->start;
530 u64 extent_offset = 0;
532 u64 last_end = args->start;
538 int modify_tree = -1;
541 struct btrfs_path *path = args->path;
543 args->bytes_found = 0;
544 args->extent_inserted = false;
546 /* Must always have a path if ->replace_extent is true */
547 ASSERT(!(args->replace_extent && !args->path));
550 path = btrfs_alloc_path();
557 if (args->drop_cache)
558 btrfs_drop_extent_map_range(inode, args->start, args->end - 1, false);
560 if (args->start >= inode->disk_i_size && !args->replace_extent)
563 update_refs = (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID);
566 ret = btrfs_lookup_file_extent(trans, root, path, ino,
567 search_start, modify_tree);
570 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
571 leaf = path->nodes[0];
572 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
573 if (key.objectid == ino &&
574 key.type == BTRFS_EXTENT_DATA_KEY)
579 leaf = path->nodes[0];
580 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
582 ret = btrfs_next_leaf(root, path);
589 leaf = path->nodes[0];
593 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
595 if (key.objectid > ino)
597 if (WARN_ON_ONCE(key.objectid < ino) ||
598 key.type < BTRFS_EXTENT_DATA_KEY) {
603 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
606 fi = btrfs_item_ptr(leaf, path->slots[0],
607 struct btrfs_file_extent_item);
608 extent_type = btrfs_file_extent_type(leaf, fi);
610 if (extent_type == BTRFS_FILE_EXTENT_REG ||
611 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
612 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
613 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
614 extent_offset = btrfs_file_extent_offset(leaf, fi);
615 extent_end = key.offset +
616 btrfs_file_extent_num_bytes(leaf, fi);
617 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
618 extent_end = key.offset +
619 btrfs_file_extent_ram_bytes(leaf, fi);
626 * Don't skip extent items representing 0 byte lengths. They
627 * used to be created (bug) if while punching holes we hit
628 * -ENOSPC condition. So if we find one here, just ensure we
629 * delete it, otherwise we would insert a new file extent item
630 * with the same key (offset) as that 0 bytes length file
631 * extent item in the call to setup_items_for_insert() later
634 if (extent_end == key.offset && extent_end >= search_start) {
635 last_end = extent_end;
636 goto delete_extent_item;
639 if (extent_end <= search_start) {
645 search_start = max(key.offset, args->start);
646 if (recow || !modify_tree) {
648 btrfs_release_path(path);
653 * | - range to drop - |
654 * | -------- extent -------- |
656 if (args->start > key.offset && args->end < extent_end) {
658 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
663 memcpy(&new_key, &key, sizeof(new_key));
664 new_key.offset = args->start;
665 ret = btrfs_duplicate_item(trans, root, path,
667 if (ret == -EAGAIN) {
668 btrfs_release_path(path);
674 leaf = path->nodes[0];
675 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
676 struct btrfs_file_extent_item);
677 btrfs_set_file_extent_num_bytes(leaf, fi,
678 args->start - key.offset);
680 fi = btrfs_item_ptr(leaf, path->slots[0],
681 struct btrfs_file_extent_item);
683 extent_offset += args->start - key.offset;
684 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
685 btrfs_set_file_extent_num_bytes(leaf, fi,
686 extent_end - args->start);
687 btrfs_mark_buffer_dirty(leaf);
689 if (update_refs && disk_bytenr > 0) {
690 btrfs_init_generic_ref(&ref,
691 BTRFS_ADD_DELAYED_REF,
692 disk_bytenr, num_bytes, 0);
693 btrfs_init_data_ref(&ref,
694 root->root_key.objectid,
696 args->start - extent_offset,
698 ret = btrfs_inc_extent_ref(trans, &ref);
699 BUG_ON(ret); /* -ENOMEM */
701 key.offset = args->start;
704 * From here on out we will have actually dropped something, so
705 * last_end can be updated.
707 last_end = extent_end;
710 * | ---- range to drop ----- |
711 * | -------- extent -------- |
713 if (args->start <= key.offset && args->end < extent_end) {
714 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
719 memcpy(&new_key, &key, sizeof(new_key));
720 new_key.offset = args->end;
721 btrfs_set_item_key_safe(fs_info, path, &new_key);
723 extent_offset += args->end - key.offset;
724 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
725 btrfs_set_file_extent_num_bytes(leaf, fi,
726 extent_end - args->end);
727 btrfs_mark_buffer_dirty(leaf);
728 if (update_refs && disk_bytenr > 0)
729 args->bytes_found += args->end - key.offset;
733 search_start = extent_end;
735 * | ---- range to drop ----- |
736 * | -------- extent -------- |
738 if (args->start > key.offset && args->end >= extent_end) {
740 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
745 btrfs_set_file_extent_num_bytes(leaf, fi,
746 args->start - key.offset);
747 btrfs_mark_buffer_dirty(leaf);
748 if (update_refs && disk_bytenr > 0)
749 args->bytes_found += extent_end - args->start;
750 if (args->end == extent_end)
758 * | ---- range to drop ----- |
759 * | ------ extent ------ |
761 if (args->start <= key.offset && args->end >= extent_end) {
764 del_slot = path->slots[0];
767 BUG_ON(del_slot + del_nr != path->slots[0]);
772 extent_type == BTRFS_FILE_EXTENT_INLINE) {
773 args->bytes_found += extent_end - key.offset;
774 extent_end = ALIGN(extent_end,
775 fs_info->sectorsize);
776 } else if (update_refs && disk_bytenr > 0) {
777 btrfs_init_generic_ref(&ref,
778 BTRFS_DROP_DELAYED_REF,
779 disk_bytenr, num_bytes, 0);
780 btrfs_init_data_ref(&ref,
781 root->root_key.objectid,
783 key.offset - extent_offset, 0,
785 ret = btrfs_free_extent(trans, &ref);
786 BUG_ON(ret); /* -ENOMEM */
787 args->bytes_found += extent_end - key.offset;
790 if (args->end == extent_end)
793 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
798 ret = btrfs_del_items(trans, root, path, del_slot,
801 btrfs_abort_transaction(trans, ret);
808 btrfs_release_path(path);
815 if (!ret && del_nr > 0) {
817 * Set path->slots[0] to first slot, so that after the delete
818 * if items are move off from our leaf to its immediate left or
819 * right neighbor leafs, we end up with a correct and adjusted
820 * path->slots[0] for our insertion (if args->replace_extent).
822 path->slots[0] = del_slot;
823 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
825 btrfs_abort_transaction(trans, ret);
828 leaf = path->nodes[0];
830 * If btrfs_del_items() was called, it might have deleted a leaf, in
831 * which case it unlocked our path, so check path->locks[0] matches a
834 if (!ret && args->replace_extent &&
835 path->locks[0] == BTRFS_WRITE_LOCK &&
836 btrfs_leaf_free_space(leaf) >=
837 sizeof(struct btrfs_item) + args->extent_item_size) {
840 key.type = BTRFS_EXTENT_DATA_KEY;
841 key.offset = args->start;
842 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
843 struct btrfs_key slot_key;
845 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
846 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
849 btrfs_setup_item_for_insert(root, path, &key, args->extent_item_size);
850 args->extent_inserted = true;
854 btrfs_free_path(path);
855 else if (!args->extent_inserted)
856 btrfs_release_path(path);
858 args->drop_end = found ? min(args->end, last_end) : args->end;
863 static int extent_mergeable(struct extent_buffer *leaf, int slot,
864 u64 objectid, u64 bytenr, u64 orig_offset,
865 u64 *start, u64 *end)
867 struct btrfs_file_extent_item *fi;
868 struct btrfs_key key;
871 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
874 btrfs_item_key_to_cpu(leaf, &key, slot);
875 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
878 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
879 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
880 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
881 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
882 btrfs_file_extent_compression(leaf, fi) ||
883 btrfs_file_extent_encryption(leaf, fi) ||
884 btrfs_file_extent_other_encoding(leaf, fi))
887 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
888 if ((*start && *start != key.offset) || (*end && *end != extent_end))
897 * Mark extent in the range start - end as written.
899 * This changes extent type from 'pre-allocated' to 'regular'. If only
900 * part of extent is marked as written, the extent will be split into
903 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
904 struct btrfs_inode *inode, u64 start, u64 end)
906 struct btrfs_fs_info *fs_info = trans->fs_info;
907 struct btrfs_root *root = inode->root;
908 struct extent_buffer *leaf;
909 struct btrfs_path *path;
910 struct btrfs_file_extent_item *fi;
911 struct btrfs_ref ref = { 0 };
912 struct btrfs_key key;
913 struct btrfs_key new_key;
925 u64 ino = btrfs_ino(inode);
927 path = btrfs_alloc_path();
934 key.type = BTRFS_EXTENT_DATA_KEY;
937 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
940 if (ret > 0 && path->slots[0] > 0)
943 leaf = path->nodes[0];
944 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
945 if (key.objectid != ino ||
946 key.type != BTRFS_EXTENT_DATA_KEY) {
948 btrfs_abort_transaction(trans, ret);
951 fi = btrfs_item_ptr(leaf, path->slots[0],
952 struct btrfs_file_extent_item);
953 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
955 btrfs_abort_transaction(trans, ret);
958 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
959 if (key.offset > start || extent_end < end) {
961 btrfs_abort_transaction(trans, ret);
965 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
966 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
967 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
968 memcpy(&new_key, &key, sizeof(new_key));
970 if (start == key.offset && end < extent_end) {
973 if (extent_mergeable(leaf, path->slots[0] - 1,
974 ino, bytenr, orig_offset,
975 &other_start, &other_end)) {
976 new_key.offset = end;
977 btrfs_set_item_key_safe(fs_info, path, &new_key);
978 fi = btrfs_item_ptr(leaf, path->slots[0],
979 struct btrfs_file_extent_item);
980 btrfs_set_file_extent_generation(leaf, fi,
982 btrfs_set_file_extent_num_bytes(leaf, fi,
984 btrfs_set_file_extent_offset(leaf, fi,
986 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
987 struct btrfs_file_extent_item);
988 btrfs_set_file_extent_generation(leaf, fi,
990 btrfs_set_file_extent_num_bytes(leaf, fi,
992 btrfs_mark_buffer_dirty(leaf);
997 if (start > key.offset && end == extent_end) {
1000 if (extent_mergeable(leaf, path->slots[0] + 1,
1001 ino, bytenr, orig_offset,
1002 &other_start, &other_end)) {
1003 fi = btrfs_item_ptr(leaf, path->slots[0],
1004 struct btrfs_file_extent_item);
1005 btrfs_set_file_extent_num_bytes(leaf, fi,
1006 start - key.offset);
1007 btrfs_set_file_extent_generation(leaf, fi,
1010 new_key.offset = start;
1011 btrfs_set_item_key_safe(fs_info, path, &new_key);
1013 fi = btrfs_item_ptr(leaf, path->slots[0],
1014 struct btrfs_file_extent_item);
1015 btrfs_set_file_extent_generation(leaf, fi,
1017 btrfs_set_file_extent_num_bytes(leaf, fi,
1019 btrfs_set_file_extent_offset(leaf, fi,
1020 start - orig_offset);
1021 btrfs_mark_buffer_dirty(leaf);
1026 while (start > key.offset || end < extent_end) {
1027 if (key.offset == start)
1030 new_key.offset = split;
1031 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1032 if (ret == -EAGAIN) {
1033 btrfs_release_path(path);
1037 btrfs_abort_transaction(trans, ret);
1041 leaf = path->nodes[0];
1042 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1043 struct btrfs_file_extent_item);
1044 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1045 btrfs_set_file_extent_num_bytes(leaf, fi,
1046 split - key.offset);
1048 fi = btrfs_item_ptr(leaf, path->slots[0],
1049 struct btrfs_file_extent_item);
1051 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1052 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1053 btrfs_set_file_extent_num_bytes(leaf, fi,
1054 extent_end - split);
1055 btrfs_mark_buffer_dirty(leaf);
1057 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1059 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1060 orig_offset, 0, false);
1061 ret = btrfs_inc_extent_ref(trans, &ref);
1063 btrfs_abort_transaction(trans, ret);
1067 if (split == start) {
1070 if (start != key.offset) {
1072 btrfs_abort_transaction(trans, ret);
1083 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1085 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset,
1087 if (extent_mergeable(leaf, path->slots[0] + 1,
1088 ino, bytenr, orig_offset,
1089 &other_start, &other_end)) {
1091 btrfs_release_path(path);
1094 extent_end = other_end;
1095 del_slot = path->slots[0] + 1;
1097 ret = btrfs_free_extent(trans, &ref);
1099 btrfs_abort_transaction(trans, ret);
1105 if (extent_mergeable(leaf, path->slots[0] - 1,
1106 ino, bytenr, orig_offset,
1107 &other_start, &other_end)) {
1109 btrfs_release_path(path);
1112 key.offset = other_start;
1113 del_slot = path->slots[0];
1115 ret = btrfs_free_extent(trans, &ref);
1117 btrfs_abort_transaction(trans, ret);
1122 fi = btrfs_item_ptr(leaf, path->slots[0],
1123 struct btrfs_file_extent_item);
1124 btrfs_set_file_extent_type(leaf, fi,
1125 BTRFS_FILE_EXTENT_REG);
1126 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1127 btrfs_mark_buffer_dirty(leaf);
1129 fi = btrfs_item_ptr(leaf, del_slot - 1,
1130 struct btrfs_file_extent_item);
1131 btrfs_set_file_extent_type(leaf, fi,
1132 BTRFS_FILE_EXTENT_REG);
1133 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1134 btrfs_set_file_extent_num_bytes(leaf, fi,
1135 extent_end - key.offset);
1136 btrfs_mark_buffer_dirty(leaf);
1138 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1140 btrfs_abort_transaction(trans, ret);
1145 btrfs_free_path(path);
1150 * on error we return an unlocked page and the error value
1151 * on success we return a locked page and 0
1153 static int prepare_uptodate_page(struct inode *inode,
1154 struct page *page, u64 pos,
1155 bool force_uptodate)
1157 struct folio *folio = page_folio(page);
1160 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1161 !PageUptodate(page)) {
1162 ret = btrfs_read_folio(NULL, folio);
1166 if (!PageUptodate(page)) {
1172 * Since btrfs_read_folio() will unlock the folio before it
1173 * returns, there is a window where btrfs_release_folio() can be
1174 * called to release the page. Here we check both inode
1175 * mapping and PagePrivate() to make sure the page was not
1178 * The private flag check is essential for subpage as we need
1179 * to store extra bitmap using page->private.
1181 if (page->mapping != inode->i_mapping || !PagePrivate(page)) {
1189 static unsigned int get_prepare_fgp_flags(bool nowait)
1191 unsigned int fgp_flags = FGP_LOCK | FGP_ACCESSED | FGP_CREAT;
1194 fgp_flags |= FGP_NOWAIT;
1199 static gfp_t get_prepare_gfp_flags(struct inode *inode, bool nowait)
1203 gfp = btrfs_alloc_write_mask(inode->i_mapping);
1205 gfp &= ~__GFP_DIRECT_RECLAIM;
1213 * this just gets pages into the page cache and locks them down.
1215 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1216 size_t num_pages, loff_t pos,
1217 size_t write_bytes, bool force_uptodate,
1221 unsigned long index = pos >> PAGE_SHIFT;
1222 gfp_t mask = get_prepare_gfp_flags(inode, nowait);
1223 unsigned int fgp_flags = get_prepare_fgp_flags(nowait);
1227 for (i = 0; i < num_pages; i++) {
1229 pages[i] = pagecache_get_page(inode->i_mapping, index + i,
1230 fgp_flags, mask | __GFP_WRITE);
1240 err = set_page_extent_mapped(pages[i]);
1247 err = prepare_uptodate_page(inode, pages[i], pos,
1249 if (!err && i == num_pages - 1)
1250 err = prepare_uptodate_page(inode, pages[i],
1251 pos + write_bytes, false);
1254 if (!nowait && err == -EAGAIN) {
1261 wait_on_page_writeback(pages[i]);
1266 while (faili >= 0) {
1267 unlock_page(pages[faili]);
1268 put_page(pages[faili]);
1276 * This function locks the extent and properly waits for data=ordered extents
1277 * to finish before allowing the pages to be modified if need.
1280 * 1 - the extent is locked
1281 * 0 - the extent is not locked, and everything is OK
1282 * -EAGAIN - need re-prepare the pages
1283 * the other < 0 number - Something wrong happens
1286 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1287 size_t num_pages, loff_t pos,
1289 u64 *lockstart, u64 *lockend, bool nowait,
1290 struct extent_state **cached_state)
1292 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1298 start_pos = round_down(pos, fs_info->sectorsize);
1299 last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1301 if (start_pos < inode->vfs_inode.i_size) {
1302 struct btrfs_ordered_extent *ordered;
1305 if (!try_lock_extent(&inode->io_tree, start_pos, last_pos)) {
1306 for (i = 0; i < num_pages; i++) {
1307 unlock_page(pages[i]);
1315 lock_extent(&inode->io_tree, start_pos, last_pos, cached_state);
1318 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1319 last_pos - start_pos + 1);
1321 ordered->file_offset + ordered->num_bytes > start_pos &&
1322 ordered->file_offset <= last_pos) {
1323 unlock_extent(&inode->io_tree, start_pos, last_pos,
1325 for (i = 0; i < num_pages; i++) {
1326 unlock_page(pages[i]);
1329 btrfs_start_ordered_extent(ordered, 1);
1330 btrfs_put_ordered_extent(ordered);
1334 btrfs_put_ordered_extent(ordered);
1336 *lockstart = start_pos;
1337 *lockend = last_pos;
1342 * We should be called after prepare_pages() which should have locked
1343 * all pages in the range.
1345 for (i = 0; i < num_pages; i++)
1346 WARN_ON(!PageLocked(pages[i]));
1352 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1354 * @pos: File offset.
1355 * @write_bytes: The length to write, will be updated to the nocow writeable
1358 * This function will flush ordered extents in the range to ensure proper
1362 * > 0 If we can nocow, and updates @write_bytes.
1363 * 0 If we can't do a nocow write.
1364 * -EAGAIN If we can't do a nocow write because snapshoting of the inode's
1365 * root is in progress.
1366 * < 0 If an error happened.
1368 * NOTE: Callers need to call btrfs_check_nocow_unlock() if we return > 0.
1370 int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1371 size_t *write_bytes, bool nowait)
1373 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1374 struct btrfs_root *root = inode->root;
1375 u64 lockstart, lockend;
1379 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1382 if (!btrfs_drew_try_write_lock(&root->snapshot_lock))
1385 lockstart = round_down(pos, fs_info->sectorsize);
1386 lockend = round_up(pos + *write_bytes,
1387 fs_info->sectorsize) - 1;
1388 num_bytes = lockend - lockstart + 1;
1391 if (!btrfs_try_lock_ordered_range(inode, lockstart, lockend)) {
1392 btrfs_drew_write_unlock(&root->snapshot_lock);
1396 btrfs_lock_and_flush_ordered_range(inode, lockstart, lockend, NULL);
1398 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1399 NULL, NULL, NULL, nowait, false);
1401 btrfs_drew_write_unlock(&root->snapshot_lock);
1403 *write_bytes = min_t(size_t, *write_bytes ,
1404 num_bytes - pos + lockstart);
1405 unlock_extent(&inode->io_tree, lockstart, lockend, NULL);
1410 void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1412 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1415 static void update_time_for_write(struct inode *inode)
1417 struct timespec64 now;
1419 if (IS_NOCMTIME(inode))
1422 now = current_time(inode);
1423 if (!timespec64_equal(&inode->i_mtime, &now))
1424 inode->i_mtime = now;
1426 if (!timespec64_equal(&inode->i_ctime, &now))
1427 inode->i_ctime = now;
1429 if (IS_I_VERSION(inode))
1430 inode_inc_iversion(inode);
1433 static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1436 struct file *file = iocb->ki_filp;
1437 struct inode *inode = file_inode(file);
1438 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1439 loff_t pos = iocb->ki_pos;
1445 * Quickly bail out on NOWAIT writes if we don't have the nodatacow or
1446 * prealloc flags, as without those flags we always have to COW. We will
1447 * later check if we can really COW into the target range (using
1448 * can_nocow_extent() at btrfs_get_blocks_direct_write()).
1450 if ((iocb->ki_flags & IOCB_NOWAIT) &&
1451 !(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1454 current->backing_dev_info = inode_to_bdi(inode);
1455 ret = file_remove_privs(file);
1460 * We reserve space for updating the inode when we reserve space for the
1461 * extent we are going to write, so we will enospc out there. We don't
1462 * need to start yet another transaction to update the inode as we will
1463 * update the inode when we finish writing whatever data we write.
1465 update_time_for_write(inode);
1467 start_pos = round_down(pos, fs_info->sectorsize);
1468 oldsize = i_size_read(inode);
1469 if (start_pos > oldsize) {
1470 /* Expand hole size to cover write data, preventing empty gap */
1471 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1473 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1475 current->backing_dev_info = NULL;
1483 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1486 struct file *file = iocb->ki_filp;
1488 struct inode *inode = file_inode(file);
1489 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1490 struct page **pages = NULL;
1491 struct extent_changeset *data_reserved = NULL;
1492 u64 release_bytes = 0;
1495 size_t num_written = 0;
1498 bool only_release_metadata = false;
1499 bool force_page_uptodate = false;
1500 loff_t old_isize = i_size_read(inode);
1501 unsigned int ilock_flags = 0;
1502 const bool nowait = (iocb->ki_flags & IOCB_NOWAIT);
1503 unsigned int bdp_flags = (nowait ? BDP_ASYNC : 0);
1506 ilock_flags |= BTRFS_ILOCK_TRY;
1508 ret = btrfs_inode_lock(inode, ilock_flags);
1512 ret = generic_write_checks(iocb, i);
1516 ret = btrfs_write_check(iocb, i, ret);
1521 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1522 PAGE_SIZE / (sizeof(struct page *)));
1523 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1524 nrptrs = max(nrptrs, 8);
1525 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1531 while (iov_iter_count(i) > 0) {
1532 struct extent_state *cached_state = NULL;
1533 size_t offset = offset_in_page(pos);
1534 size_t sector_offset;
1535 size_t write_bytes = min(iov_iter_count(i),
1536 nrptrs * (size_t)PAGE_SIZE -
1539 size_t reserve_bytes;
1542 size_t dirty_sectors;
1547 * Fault pages before locking them in prepare_pages
1548 * to avoid recursive lock
1550 if (unlikely(fault_in_iov_iter_readable(i, write_bytes))) {
1555 only_release_metadata = false;
1556 sector_offset = pos & (fs_info->sectorsize - 1);
1558 extent_changeset_release(data_reserved);
1559 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1560 &data_reserved, pos,
1561 write_bytes, nowait);
1565 if (nowait && (ret == -ENOSPC || ret == -EAGAIN)) {
1571 * If we don't have to COW at the offset, reserve
1572 * metadata only. write_bytes may get smaller than
1575 can_nocow = btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1576 &write_bytes, nowait);
1583 only_release_metadata = true;
1586 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1587 WARN_ON(num_pages > nrptrs);
1588 reserve_bytes = round_up(write_bytes + sector_offset,
1589 fs_info->sectorsize);
1590 WARN_ON(reserve_bytes == 0);
1591 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1593 reserve_bytes, nowait);
1595 if (!only_release_metadata)
1596 btrfs_free_reserved_data_space(BTRFS_I(inode),
1600 btrfs_check_nocow_unlock(BTRFS_I(inode));
1602 if (nowait && ret == -ENOSPC)
1607 release_bytes = reserve_bytes;
1609 ret = balance_dirty_pages_ratelimited_flags(inode->i_mapping, bdp_flags);
1611 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1616 * This is going to setup the pages array with the number of
1617 * pages we want, so we don't really need to worry about the
1618 * contents of pages from loop to loop
1620 ret = prepare_pages(inode, pages, num_pages,
1621 pos, write_bytes, force_page_uptodate, false);
1623 btrfs_delalloc_release_extents(BTRFS_I(inode),
1628 extents_locked = lock_and_cleanup_extent_if_need(
1629 BTRFS_I(inode), pages,
1630 num_pages, pos, write_bytes, &lockstart,
1631 &lockend, nowait, &cached_state);
1632 if (extents_locked < 0) {
1633 if (!nowait && extents_locked == -EAGAIN)
1636 btrfs_delalloc_release_extents(BTRFS_I(inode),
1638 ret = extents_locked;
1642 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1644 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1645 dirty_sectors = round_up(copied + sector_offset,
1646 fs_info->sectorsize);
1647 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1650 * if we have trouble faulting in the pages, fall
1651 * back to one page at a time
1653 if (copied < write_bytes)
1657 force_page_uptodate = true;
1661 force_page_uptodate = false;
1662 dirty_pages = DIV_ROUND_UP(copied + offset,
1666 if (num_sectors > dirty_sectors) {
1667 /* release everything except the sectors we dirtied */
1668 release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1669 if (only_release_metadata) {
1670 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1671 release_bytes, true);
1675 __pos = round_down(pos,
1676 fs_info->sectorsize) +
1677 (dirty_pages << PAGE_SHIFT);
1678 btrfs_delalloc_release_space(BTRFS_I(inode),
1679 data_reserved, __pos,
1680 release_bytes, true);
1684 release_bytes = round_up(copied + sector_offset,
1685 fs_info->sectorsize);
1687 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1688 dirty_pages, pos, copied,
1689 &cached_state, only_release_metadata);
1692 * If we have not locked the extent range, because the range's
1693 * start offset is >= i_size, we might still have a non-NULL
1694 * cached extent state, acquired while marking the extent range
1695 * as delalloc through btrfs_dirty_pages(). Therefore free any
1696 * possible cached extent state to avoid a memory leak.
1699 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
1700 lockend, &cached_state);
1702 free_extent_state(cached_state);
1704 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1706 btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1711 if (only_release_metadata)
1712 btrfs_check_nocow_unlock(BTRFS_I(inode));
1714 btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1719 num_written += copied;
1724 if (release_bytes) {
1725 if (only_release_metadata) {
1726 btrfs_check_nocow_unlock(BTRFS_I(inode));
1727 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1728 release_bytes, true);
1730 btrfs_delalloc_release_space(BTRFS_I(inode),
1732 round_down(pos, fs_info->sectorsize),
1733 release_bytes, true);
1737 extent_changeset_free(data_reserved);
1738 if (num_written > 0) {
1739 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1740 iocb->ki_pos += num_written;
1743 btrfs_inode_unlock(inode, ilock_flags);
1744 return num_written ? num_written : ret;
1747 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1748 const struct iov_iter *iter, loff_t offset)
1750 const u32 blocksize_mask = fs_info->sectorsize - 1;
1752 if (offset & blocksize_mask)
1755 if (iov_iter_alignment(iter) & blocksize_mask)
1761 static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1763 struct file *file = iocb->ki_filp;
1764 struct inode *inode = file_inode(file);
1765 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1767 ssize_t written = 0;
1768 ssize_t written_buffered;
1769 size_t prev_left = 0;
1772 unsigned int ilock_flags = 0;
1773 struct iomap_dio *dio;
1775 if (iocb->ki_flags & IOCB_NOWAIT)
1776 ilock_flags |= BTRFS_ILOCK_TRY;
1778 /* If the write DIO is within EOF, use a shared lock */
1779 if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode))
1780 ilock_flags |= BTRFS_ILOCK_SHARED;
1783 err = btrfs_inode_lock(inode, ilock_flags);
1787 err = generic_write_checks(iocb, from);
1789 btrfs_inode_unlock(inode, ilock_flags);
1793 err = btrfs_write_check(iocb, from, err);
1795 btrfs_inode_unlock(inode, ilock_flags);
1801 * Re-check since file size may have changed just before taking the
1802 * lock or pos may have changed because of O_APPEND in generic_write_check()
1804 if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1805 pos + iov_iter_count(from) > i_size_read(inode)) {
1806 btrfs_inode_unlock(inode, ilock_flags);
1807 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1811 if (check_direct_IO(fs_info, from, pos)) {
1812 btrfs_inode_unlock(inode, ilock_flags);
1817 * The iov_iter can be mapped to the same file range we are writing to.
1818 * If that's the case, then we will deadlock in the iomap code, because
1819 * it first calls our callback btrfs_dio_iomap_begin(), which will create
1820 * an ordered extent, and after that it will fault in the pages that the
1821 * iov_iter refers to. During the fault in we end up in the readahead
1822 * pages code (starting at btrfs_readahead()), which will lock the range,
1823 * find that ordered extent and then wait for it to complete (at
1824 * btrfs_lock_and_flush_ordered_range()), resulting in a deadlock since
1825 * obviously the ordered extent can never complete as we didn't submit
1826 * yet the respective bio(s). This always happens when the buffer is
1827 * memory mapped to the same file range, since the iomap DIO code always
1828 * invalidates pages in the target file range (after starting and waiting
1829 * for any writeback).
1831 * So here we disable page faults in the iov_iter and then retry if we
1832 * got -EFAULT, faulting in the pages before the retry.
1834 from->nofault = true;
1835 dio = btrfs_dio_write(iocb, from, written);
1836 from->nofault = false;
1839 * iomap_dio_complete() will call btrfs_sync_file() if we have a dsync
1840 * iocb, and that needs to lock the inode. So unlock it before calling
1841 * iomap_dio_complete() to avoid a deadlock.
1843 btrfs_inode_unlock(inode, ilock_flags);
1845 if (IS_ERR_OR_NULL(dio))
1846 err = PTR_ERR_OR_ZERO(dio);
1848 err = iomap_dio_complete(dio);
1850 /* No increment (+=) because iomap returns a cumulative value. */
1854 if (iov_iter_count(from) > 0 && (err == -EFAULT || err > 0)) {
1855 const size_t left = iov_iter_count(from);
1857 * We have more data left to write. Try to fault in as many as
1858 * possible of the remainder pages and retry. We do this without
1859 * releasing and locking again the inode, to prevent races with
1862 * Also, in case the iov refers to pages in the file range of the
1863 * file we want to write to (due to a mmap), we could enter an
1864 * infinite loop if we retry after faulting the pages in, since
1865 * iomap will invalidate any pages in the range early on, before
1866 * it tries to fault in the pages of the iov. So we keep track of
1867 * how much was left of iov in the previous EFAULT and fallback
1868 * to buffered IO in case we haven't made any progress.
1870 if (left == prev_left) {
1873 fault_in_iov_iter_readable(from, left);
1880 * If 'err' is -ENOTBLK or we have not written all data, then it means
1881 * we must fallback to buffered IO.
1883 if ((err < 0 && err != -ENOTBLK) || !iov_iter_count(from))
1888 * If we are in a NOWAIT context, then return -EAGAIN to signal the caller
1889 * it must retry the operation in a context where blocking is acceptable,
1890 * since we currently don't have NOWAIT semantics support for buffered IO
1891 * and may block there for many reasons (reserving space for example).
1893 if (iocb->ki_flags & IOCB_NOWAIT) {
1899 written_buffered = btrfs_buffered_write(iocb, from);
1900 if (written_buffered < 0) {
1901 err = written_buffered;
1905 * Ensure all data is persisted. We want the next direct IO read to be
1906 * able to read what was just written.
1908 endbyte = pos + written_buffered - 1;
1909 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1912 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1915 written += written_buffered;
1916 iocb->ki_pos = pos + written_buffered;
1917 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1918 endbyte >> PAGE_SHIFT);
1920 return err < 0 ? err : written;
1923 static ssize_t btrfs_encoded_write(struct kiocb *iocb, struct iov_iter *from,
1924 const struct btrfs_ioctl_encoded_io_args *encoded)
1926 struct file *file = iocb->ki_filp;
1927 struct inode *inode = file_inode(file);
1931 btrfs_inode_lock(inode, 0);
1932 count = encoded->len;
1933 ret = generic_write_checks_count(iocb, &count);
1934 if (ret == 0 && count != encoded->len) {
1936 * The write got truncated by generic_write_checks_count(). We
1937 * can't do a partial encoded write.
1941 if (ret || encoded->len == 0)
1944 ret = btrfs_write_check(iocb, from, encoded->len);
1948 ret = btrfs_do_encoded_write(iocb, from, encoded);
1950 btrfs_inode_unlock(inode, 0);
1954 ssize_t btrfs_do_write_iter(struct kiocb *iocb, struct iov_iter *from,
1955 const struct btrfs_ioctl_encoded_io_args *encoded)
1957 struct file *file = iocb->ki_filp;
1958 struct btrfs_inode *inode = BTRFS_I(file_inode(file));
1959 ssize_t num_written, num_sync;
1960 const bool sync = iocb_is_dsync(iocb);
1963 * If the fs flips readonly due to some impossible error, although we
1964 * have opened a file as writable, we have to stop this write operation
1965 * to ensure consistency.
1967 if (BTRFS_FS_ERROR(inode->root->fs_info))
1970 if (encoded && (iocb->ki_flags & IOCB_NOWAIT))
1974 atomic_inc(&inode->sync_writers);
1977 num_written = btrfs_encoded_write(iocb, from, encoded);
1978 num_sync = encoded->len;
1979 } else if (iocb->ki_flags & IOCB_DIRECT) {
1980 num_written = btrfs_direct_write(iocb, from);
1981 num_sync = num_written;
1983 num_written = btrfs_buffered_write(iocb, from);
1984 num_sync = num_written;
1987 btrfs_set_inode_last_sub_trans(inode);
1990 num_sync = generic_write_sync(iocb, num_sync);
1992 num_written = num_sync;
1996 atomic_dec(&inode->sync_writers);
1998 current->backing_dev_info = NULL;
2002 static ssize_t btrfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
2004 return btrfs_do_write_iter(iocb, from, NULL);
2007 int btrfs_release_file(struct inode *inode, struct file *filp)
2009 struct btrfs_file_private *private = filp->private_data;
2011 if (private && private->filldir_buf)
2012 kfree(private->filldir_buf);
2014 filp->private_data = NULL;
2017 * Set by setattr when we are about to truncate a file from a non-zero
2018 * size to a zero size. This tries to flush down new bytes that may
2019 * have been written if the application were using truncate to replace
2022 if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
2023 &BTRFS_I(inode)->runtime_flags))
2024 filemap_flush(inode->i_mapping);
2028 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2031 struct blk_plug plug;
2034 * This is only called in fsync, which would do synchronous writes, so
2035 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2036 * multiple disks using raid profile, a large IO can be split to
2037 * several segments of stripe length (currently 64K).
2039 blk_start_plug(&plug);
2040 atomic_inc(&BTRFS_I(inode)->sync_writers);
2041 ret = btrfs_fdatawrite_range(inode, start, end);
2042 atomic_dec(&BTRFS_I(inode)->sync_writers);
2043 blk_finish_plug(&plug);
2048 static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
2050 struct btrfs_inode *inode = BTRFS_I(ctx->inode);
2051 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2053 if (btrfs_inode_in_log(inode, fs_info->generation) &&
2054 list_empty(&ctx->ordered_extents))
2058 * If we are doing a fast fsync we can not bail out if the inode's
2059 * last_trans is <= then the last committed transaction, because we only
2060 * update the last_trans of the inode during ordered extent completion,
2061 * and for a fast fsync we don't wait for that, we only wait for the
2062 * writeback to complete.
2064 if (inode->last_trans <= fs_info->last_trans_committed &&
2065 (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
2066 list_empty(&ctx->ordered_extents)))
2073 * fsync call for both files and directories. This logs the inode into
2074 * the tree log instead of forcing full commits whenever possible.
2076 * It needs to call filemap_fdatawait so that all ordered extent updates are
2077 * in the metadata btree are up to date for copying to the log.
2079 * It drops the inode mutex before doing the tree log commit. This is an
2080 * important optimization for directories because holding the mutex prevents
2081 * new operations on the dir while we write to disk.
2083 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2085 struct dentry *dentry = file_dentry(file);
2086 struct inode *inode = d_inode(dentry);
2087 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2088 struct btrfs_root *root = BTRFS_I(inode)->root;
2089 struct btrfs_trans_handle *trans;
2090 struct btrfs_log_ctx ctx;
2095 trace_btrfs_sync_file(file, datasync);
2097 btrfs_init_log_ctx(&ctx, inode);
2100 * Always set the range to a full range, otherwise we can get into
2101 * several problems, from missing file extent items to represent holes
2102 * when not using the NO_HOLES feature, to log tree corruption due to
2103 * races between hole detection during logging and completion of ordered
2104 * extents outside the range, to missing checksums due to ordered extents
2105 * for which we flushed only a subset of their pages.
2109 len = (u64)LLONG_MAX + 1;
2112 * We write the dirty pages in the range and wait until they complete
2113 * out of the ->i_mutex. If so, we can flush the dirty pages by
2114 * multi-task, and make the performance up. See
2115 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2117 ret = start_ordered_ops(inode, start, end);
2121 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
2123 atomic_inc(&root->log_batch);
2126 * Before we acquired the inode's lock and the mmap lock, someone may
2127 * have dirtied more pages in the target range. We need to make sure
2128 * that writeback for any such pages does not start while we are logging
2129 * the inode, because if it does, any of the following might happen when
2130 * we are not doing a full inode sync:
2132 * 1) We log an extent after its writeback finishes but before its
2133 * checksums are added to the csum tree, leading to -EIO errors
2134 * when attempting to read the extent after a log replay.
2136 * 2) We can end up logging an extent before its writeback finishes.
2137 * Therefore after the log replay we will have a file extent item
2138 * pointing to an unwritten extent (and no data checksums as well).
2140 * So trigger writeback for any eventual new dirty pages and then we
2141 * wait for all ordered extents to complete below.
2143 ret = start_ordered_ops(inode, start, end);
2145 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2150 * Always check for the full sync flag while holding the inode's lock,
2151 * to avoid races with other tasks. The flag must be either set all the
2152 * time during logging or always off all the time while logging.
2153 * We check the flag here after starting delalloc above, because when
2154 * running delalloc the full sync flag may be set if we need to drop
2155 * extra extent map ranges due to temporary memory allocation failures.
2157 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2158 &BTRFS_I(inode)->runtime_flags);
2161 * We have to do this here to avoid the priority inversion of waiting on
2162 * IO of a lower priority task while holding a transaction open.
2164 * For a full fsync we wait for the ordered extents to complete while
2165 * for a fast fsync we wait just for writeback to complete, and then
2166 * attach the ordered extents to the transaction so that a transaction
2167 * commit waits for their completion, to avoid data loss if we fsync,
2168 * the current transaction commits before the ordered extents complete
2169 * and a power failure happens right after that.
2171 * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
2172 * logical address recorded in the ordered extent may change. We need
2173 * to wait for the IO to stabilize the logical address.
2175 if (full_sync || btrfs_is_zoned(fs_info)) {
2176 ret = btrfs_wait_ordered_range(inode, start, len);
2179 * Get our ordered extents as soon as possible to avoid doing
2180 * checksum lookups in the csum tree, and use instead the
2181 * checksums attached to the ordered extents.
2183 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2184 &ctx.ordered_extents);
2185 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2189 goto out_release_extents;
2191 atomic_inc(&root->log_batch);
2194 if (skip_inode_logging(&ctx)) {
2196 * We've had everything committed since the last time we were
2197 * modified so clear this flag in case it was set for whatever
2198 * reason, it's no longer relevant.
2200 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2201 &BTRFS_I(inode)->runtime_flags);
2203 * An ordered extent might have started before and completed
2204 * already with io errors, in which case the inode was not
2205 * updated and we end up here. So check the inode's mapping
2206 * for any errors that might have happened since we last
2207 * checked called fsync.
2209 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2210 goto out_release_extents;
2214 * We use start here because we will need to wait on the IO to complete
2215 * in btrfs_sync_log, which could require joining a transaction (for
2216 * example checking cross references in the nocow path). If we use join
2217 * here we could get into a situation where we're waiting on IO to
2218 * happen that is blocked on a transaction trying to commit. With start
2219 * we inc the extwriter counter, so we wait for all extwriters to exit
2220 * before we start blocking joiners. This comment is to keep somebody
2221 * from thinking they are super smart and changing this to
2222 * btrfs_join_transaction *cough*Josef*cough*.
2224 trans = btrfs_start_transaction(root, 0);
2225 if (IS_ERR(trans)) {
2226 ret = PTR_ERR(trans);
2227 goto out_release_extents;
2229 trans->in_fsync = true;
2231 ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2232 btrfs_release_log_ctx_extents(&ctx);
2234 /* Fallthrough and commit/free transaction. */
2235 ret = BTRFS_LOG_FORCE_COMMIT;
2238 /* we've logged all the items and now have a consistent
2239 * version of the file in the log. It is possible that
2240 * someone will come in and modify the file, but that's
2241 * fine because the log is consistent on disk, and we
2242 * have references to all of the file's extents
2244 * It is possible that someone will come in and log the
2245 * file again, but that will end up using the synchronization
2246 * inside btrfs_sync_log to keep things safe.
2248 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2250 if (ret == BTRFS_NO_LOG_SYNC) {
2251 ret = btrfs_end_transaction(trans);
2255 /* We successfully logged the inode, attempt to sync the log. */
2257 ret = btrfs_sync_log(trans, root, &ctx);
2259 ret = btrfs_end_transaction(trans);
2265 * At this point we need to commit the transaction because we had
2266 * btrfs_need_log_full_commit() or some other error.
2268 * If we didn't do a full sync we have to stop the trans handle, wait on
2269 * the ordered extents, start it again and commit the transaction. If
2270 * we attempt to wait on the ordered extents here we could deadlock with
2271 * something like fallocate() that is holding the extent lock trying to
2272 * start a transaction while some other thread is trying to commit the
2273 * transaction while we (fsync) are currently holding the transaction
2277 ret = btrfs_end_transaction(trans);
2280 ret = btrfs_wait_ordered_range(inode, start, len);
2285 * This is safe to use here because we're only interested in
2286 * making sure the transaction that had the ordered extents is
2287 * committed. We aren't waiting on anything past this point,
2288 * we're purely getting the transaction and committing it.
2290 trans = btrfs_attach_transaction_barrier(root);
2291 if (IS_ERR(trans)) {
2292 ret = PTR_ERR(trans);
2295 * We committed the transaction and there's no currently
2296 * running transaction, this means everything we care
2297 * about made it to disk and we are done.
2305 ret = btrfs_commit_transaction(trans);
2307 ASSERT(list_empty(&ctx.list));
2308 ASSERT(list_empty(&ctx.conflict_inodes));
2309 err = file_check_and_advance_wb_err(file);
2312 return ret > 0 ? -EIO : ret;
2314 out_release_extents:
2315 btrfs_release_log_ctx_extents(&ctx);
2316 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2320 static const struct vm_operations_struct btrfs_file_vm_ops = {
2321 .fault = filemap_fault,
2322 .map_pages = filemap_map_pages,
2323 .page_mkwrite = btrfs_page_mkwrite,
2326 static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2328 struct address_space *mapping = filp->f_mapping;
2330 if (!mapping->a_ops->read_folio)
2333 file_accessed(filp);
2334 vma->vm_ops = &btrfs_file_vm_ops;
2339 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2340 int slot, u64 start, u64 end)
2342 struct btrfs_file_extent_item *fi;
2343 struct btrfs_key key;
2345 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2348 btrfs_item_key_to_cpu(leaf, &key, slot);
2349 if (key.objectid != btrfs_ino(inode) ||
2350 key.type != BTRFS_EXTENT_DATA_KEY)
2353 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2355 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2358 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2361 if (key.offset == end)
2363 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2368 static int fill_holes(struct btrfs_trans_handle *trans,
2369 struct btrfs_inode *inode,
2370 struct btrfs_path *path, u64 offset, u64 end)
2372 struct btrfs_fs_info *fs_info = trans->fs_info;
2373 struct btrfs_root *root = inode->root;
2374 struct extent_buffer *leaf;
2375 struct btrfs_file_extent_item *fi;
2376 struct extent_map *hole_em;
2377 struct btrfs_key key;
2380 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2383 key.objectid = btrfs_ino(inode);
2384 key.type = BTRFS_EXTENT_DATA_KEY;
2385 key.offset = offset;
2387 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2390 * We should have dropped this offset, so if we find it then
2391 * something has gone horribly wrong.
2398 leaf = path->nodes[0];
2399 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2403 fi = btrfs_item_ptr(leaf, path->slots[0],
2404 struct btrfs_file_extent_item);
2405 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2407 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2408 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2409 btrfs_set_file_extent_offset(leaf, fi, 0);
2410 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2411 btrfs_mark_buffer_dirty(leaf);
2415 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2418 key.offset = offset;
2419 btrfs_set_item_key_safe(fs_info, path, &key);
2420 fi = btrfs_item_ptr(leaf, path->slots[0],
2421 struct btrfs_file_extent_item);
2422 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2424 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2425 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2426 btrfs_set_file_extent_offset(leaf, fi, 0);
2427 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2428 btrfs_mark_buffer_dirty(leaf);
2431 btrfs_release_path(path);
2433 ret = btrfs_insert_hole_extent(trans, root, btrfs_ino(inode), offset,
2439 btrfs_release_path(path);
2441 hole_em = alloc_extent_map();
2443 btrfs_drop_extent_map_range(inode, offset, end - 1, false);
2444 btrfs_set_inode_full_sync(inode);
2446 hole_em->start = offset;
2447 hole_em->len = end - offset;
2448 hole_em->ram_bytes = hole_em->len;
2449 hole_em->orig_start = offset;
2451 hole_em->block_start = EXTENT_MAP_HOLE;
2452 hole_em->block_len = 0;
2453 hole_em->orig_block_len = 0;
2454 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2455 hole_em->generation = trans->transid;
2457 ret = btrfs_replace_extent_map_range(inode, hole_em, true);
2458 free_extent_map(hole_em);
2460 btrfs_set_inode_full_sync(inode);
2467 * Find a hole extent on given inode and change start/len to the end of hole
2468 * extent.(hole/vacuum extent whose em->start <= start &&
2469 * em->start + em->len > start)
2470 * When a hole extent is found, return 1 and modify start/len.
2472 static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2474 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2475 struct extent_map *em;
2478 em = btrfs_get_extent(inode, NULL, 0,
2479 round_down(*start, fs_info->sectorsize),
2480 round_up(*len, fs_info->sectorsize));
2484 /* Hole or vacuum extent(only exists in no-hole mode) */
2485 if (em->block_start == EXTENT_MAP_HOLE) {
2487 *len = em->start + em->len > *start + *len ?
2488 0 : *start + *len - em->start - em->len;
2489 *start = em->start + em->len;
2491 free_extent_map(em);
2495 static void btrfs_punch_hole_lock_range(struct inode *inode,
2496 const u64 lockstart,
2498 struct extent_state **cached_state)
2501 * For subpage case, if the range is not at page boundary, we could
2502 * have pages at the leading/tailing part of the range.
2503 * This could lead to dead loop since filemap_range_has_page()
2504 * will always return true.
2505 * So here we need to do extra page alignment for
2506 * filemap_range_has_page().
2508 const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2509 const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2512 truncate_pagecache_range(inode, lockstart, lockend);
2514 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2517 * We can't have ordered extents in the range, nor dirty/writeback
2518 * pages, because we have locked the inode's VFS lock in exclusive
2519 * mode, we have locked the inode's i_mmap_lock in exclusive mode,
2520 * we have flushed all delalloc in the range and we have waited
2521 * for any ordered extents in the range to complete.
2522 * We can race with anyone reading pages from this range, so after
2523 * locking the range check if we have pages in the range, and if
2524 * we do, unlock the range and retry.
2526 if (!filemap_range_has_page(inode->i_mapping, page_lockstart,
2530 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2534 btrfs_assert_inode_range_clean(BTRFS_I(inode), lockstart, lockend);
2537 static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2538 struct btrfs_inode *inode,
2539 struct btrfs_path *path,
2540 struct btrfs_replace_extent_info *extent_info,
2541 const u64 replace_len,
2542 const u64 bytes_to_drop)
2544 struct btrfs_fs_info *fs_info = trans->fs_info;
2545 struct btrfs_root *root = inode->root;
2546 struct btrfs_file_extent_item *extent;
2547 struct extent_buffer *leaf;
2548 struct btrfs_key key;
2550 struct btrfs_ref ref = { 0 };
2553 if (replace_len == 0)
2556 if (extent_info->disk_offset == 0 &&
2557 btrfs_fs_incompat(fs_info, NO_HOLES)) {
2558 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2562 key.objectid = btrfs_ino(inode);
2563 key.type = BTRFS_EXTENT_DATA_KEY;
2564 key.offset = extent_info->file_offset;
2565 ret = btrfs_insert_empty_item(trans, root, path, &key,
2566 sizeof(struct btrfs_file_extent_item));
2569 leaf = path->nodes[0];
2570 slot = path->slots[0];
2571 write_extent_buffer(leaf, extent_info->extent_buf,
2572 btrfs_item_ptr_offset(leaf, slot),
2573 sizeof(struct btrfs_file_extent_item));
2574 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2575 ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2576 btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2577 btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2578 if (extent_info->is_new_extent)
2579 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2580 btrfs_mark_buffer_dirty(leaf);
2581 btrfs_release_path(path);
2583 ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2588 /* If it's a hole, nothing more needs to be done. */
2589 if (extent_info->disk_offset == 0) {
2590 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2594 btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2596 if (extent_info->is_new_extent && extent_info->insertions == 0) {
2597 key.objectid = extent_info->disk_offset;
2598 key.type = BTRFS_EXTENT_ITEM_KEY;
2599 key.offset = extent_info->disk_len;
2600 ret = btrfs_alloc_reserved_file_extent(trans, root,
2602 extent_info->file_offset,
2603 extent_info->qgroup_reserved,
2608 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2609 extent_info->disk_offset,
2610 extent_info->disk_len, 0);
2611 ref_offset = extent_info->file_offset - extent_info->data_offset;
2612 btrfs_init_data_ref(&ref, root->root_key.objectid,
2613 btrfs_ino(inode), ref_offset, 0, false);
2614 ret = btrfs_inc_extent_ref(trans, &ref);
2617 extent_info->insertions++;
2623 * The respective range must have been previously locked, as well as the inode.
2624 * The end offset is inclusive (last byte of the range).
2625 * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2626 * the file range with an extent.
2627 * When not punching a hole, we don't want to end up in a state where we dropped
2628 * extents without inserting a new one, so we must abort the transaction to avoid
2631 int btrfs_replace_file_extents(struct btrfs_inode *inode,
2632 struct btrfs_path *path, const u64 start,
2634 struct btrfs_replace_extent_info *extent_info,
2635 struct btrfs_trans_handle **trans_out)
2637 struct btrfs_drop_extents_args drop_args = { 0 };
2638 struct btrfs_root *root = inode->root;
2639 struct btrfs_fs_info *fs_info = root->fs_info;
2640 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2641 u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2642 struct btrfs_trans_handle *trans = NULL;
2643 struct btrfs_block_rsv *rsv;
2644 unsigned int rsv_count;
2646 u64 len = end - start;
2652 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2657 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2658 rsv->failfast = true;
2661 * 1 - update the inode
2662 * 1 - removing the extents in the range
2663 * 1 - adding the hole extent if no_holes isn't set or if we are
2664 * replacing the range with a new extent
2666 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2671 trans = btrfs_start_transaction(root, rsv_count);
2672 if (IS_ERR(trans)) {
2673 ret = PTR_ERR(trans);
2678 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2682 trans->block_rsv = rsv;
2685 drop_args.path = path;
2686 drop_args.end = end + 1;
2687 drop_args.drop_cache = true;
2688 while (cur_offset < end) {
2689 drop_args.start = cur_offset;
2690 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2691 /* If we are punching a hole decrement the inode's byte count */
2693 btrfs_update_inode_bytes(inode, 0,
2694 drop_args.bytes_found);
2695 if (ret != -ENOSPC) {
2697 * The only time we don't want to abort is if we are
2698 * attempting to clone a partial inline extent, in which
2699 * case we'll get EOPNOTSUPP. However if we aren't
2700 * clone we need to abort no matter what, because if we
2701 * got EOPNOTSUPP via prealloc then we messed up and
2705 (ret != -EOPNOTSUPP ||
2706 (extent_info && extent_info->is_new_extent)))
2707 btrfs_abort_transaction(trans, ret);
2711 trans->block_rsv = &fs_info->trans_block_rsv;
2713 if (!extent_info && cur_offset < drop_args.drop_end &&
2714 cur_offset < ino_size) {
2715 ret = fill_holes(trans, inode, path, cur_offset,
2716 drop_args.drop_end);
2719 * If we failed then we didn't insert our hole
2720 * entries for the area we dropped, so now the
2721 * fs is corrupted, so we must abort the
2724 btrfs_abort_transaction(trans, ret);
2727 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2729 * We are past the i_size here, but since we didn't
2730 * insert holes we need to clear the mapped area so we
2731 * know to not set disk_i_size in this area until a new
2732 * file extent is inserted here.
2734 ret = btrfs_inode_clear_file_extent_range(inode,
2736 drop_args.drop_end - cur_offset);
2739 * We couldn't clear our area, so we could
2740 * presumably adjust up and corrupt the fs, so
2743 btrfs_abort_transaction(trans, ret);
2749 drop_args.drop_end > extent_info->file_offset) {
2750 u64 replace_len = drop_args.drop_end -
2751 extent_info->file_offset;
2753 ret = btrfs_insert_replace_extent(trans, inode, path,
2754 extent_info, replace_len,
2755 drop_args.bytes_found);
2757 btrfs_abort_transaction(trans, ret);
2760 extent_info->data_len -= replace_len;
2761 extent_info->data_offset += replace_len;
2762 extent_info->file_offset += replace_len;
2766 * We are releasing our handle on the transaction, balance the
2767 * dirty pages of the btree inode and flush delayed items, and
2768 * then get a new transaction handle, which may now point to a
2769 * new transaction in case someone else may have committed the
2770 * transaction we used to replace/drop file extent items. So
2771 * bump the inode's iversion and update mtime and ctime except
2772 * if we are called from a dedupe context. This is because a
2773 * power failure/crash may happen after the transaction is
2774 * committed and before we finish replacing/dropping all the
2775 * file extent items we need.
2777 inode_inc_iversion(&inode->vfs_inode);
2779 if (!extent_info || extent_info->update_times) {
2780 inode->vfs_inode.i_mtime = current_time(&inode->vfs_inode);
2781 inode->vfs_inode.i_ctime = inode->vfs_inode.i_mtime;
2784 ret = btrfs_update_inode(trans, root, inode);
2788 btrfs_end_transaction(trans);
2789 btrfs_btree_balance_dirty(fs_info);
2791 trans = btrfs_start_transaction(root, rsv_count);
2792 if (IS_ERR(trans)) {
2793 ret = PTR_ERR(trans);
2798 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2799 rsv, min_size, false);
2802 trans->block_rsv = rsv;
2804 cur_offset = drop_args.drop_end;
2805 len = end - cur_offset;
2806 if (!extent_info && len) {
2807 ret = find_first_non_hole(inode, &cur_offset, &len);
2808 if (unlikely(ret < 0))
2818 * If we were cloning, force the next fsync to be a full one since we
2819 * we replaced (or just dropped in the case of cloning holes when
2820 * NO_HOLES is enabled) file extent items and did not setup new extent
2821 * maps for the replacement extents (or holes).
2823 if (extent_info && !extent_info->is_new_extent)
2824 btrfs_set_inode_full_sync(inode);
2829 trans->block_rsv = &fs_info->trans_block_rsv;
2831 * If we are using the NO_HOLES feature we might have had already an
2832 * hole that overlaps a part of the region [lockstart, lockend] and
2833 * ends at (or beyond) lockend. Since we have no file extent items to
2834 * represent holes, drop_end can be less than lockend and so we must
2835 * make sure we have an extent map representing the existing hole (the
2836 * call to __btrfs_drop_extents() might have dropped the existing extent
2837 * map representing the existing hole), otherwise the fast fsync path
2838 * will not record the existence of the hole region
2839 * [existing_hole_start, lockend].
2841 if (drop_args.drop_end <= end)
2842 drop_args.drop_end = end + 1;
2844 * Don't insert file hole extent item if it's for a range beyond eof
2845 * (because it's useless) or if it represents a 0 bytes range (when
2846 * cur_offset == drop_end).
2848 if (!extent_info && cur_offset < ino_size &&
2849 cur_offset < drop_args.drop_end) {
2850 ret = fill_holes(trans, inode, path, cur_offset,
2851 drop_args.drop_end);
2853 /* Same comment as above. */
2854 btrfs_abort_transaction(trans, ret);
2857 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2858 /* See the comment in the loop above for the reasoning here. */
2859 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2860 drop_args.drop_end - cur_offset);
2862 btrfs_abort_transaction(trans, ret);
2868 ret = btrfs_insert_replace_extent(trans, inode, path,
2869 extent_info, extent_info->data_len,
2870 drop_args.bytes_found);
2872 btrfs_abort_transaction(trans, ret);
2881 trans->block_rsv = &fs_info->trans_block_rsv;
2883 btrfs_end_transaction(trans);
2887 btrfs_free_block_rsv(fs_info, rsv);
2892 static int btrfs_punch_hole(struct file *file, loff_t offset, loff_t len)
2894 struct inode *inode = file_inode(file);
2895 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2896 struct btrfs_root *root = BTRFS_I(inode)->root;
2897 struct extent_state *cached_state = NULL;
2898 struct btrfs_path *path;
2899 struct btrfs_trans_handle *trans = NULL;
2904 u64 orig_start = offset;
2908 bool truncated_block = false;
2909 bool updated_inode = false;
2911 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
2913 ret = btrfs_wait_ordered_range(inode, offset, len);
2915 goto out_only_mutex;
2917 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2918 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2920 goto out_only_mutex;
2922 /* Already in a large hole */
2924 goto out_only_mutex;
2927 ret = file_modified(file);
2929 goto out_only_mutex;
2931 lockstart = round_up(offset, fs_info->sectorsize);
2932 lockend = round_down(offset + len, fs_info->sectorsize) - 1;
2933 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2934 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2936 * We needn't truncate any block which is beyond the end of the file
2937 * because we are sure there is no data there.
2940 * Only do this if we are in the same block and we aren't doing the
2943 if (same_block && len < fs_info->sectorsize) {
2944 if (offset < ino_size) {
2945 truncated_block = true;
2946 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2951 goto out_only_mutex;
2954 /* zero back part of the first block */
2955 if (offset < ino_size) {
2956 truncated_block = true;
2957 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2959 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2964 /* Check the aligned pages after the first unaligned page,
2965 * if offset != orig_start, which means the first unaligned page
2966 * including several following pages are already in holes,
2967 * the extra check can be skipped */
2968 if (offset == orig_start) {
2969 /* after truncate page, check hole again */
2970 len = offset + len - lockstart;
2972 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2974 goto out_only_mutex;
2977 goto out_only_mutex;
2982 /* Check the tail unaligned part is in a hole */
2983 tail_start = lockend + 1;
2984 tail_len = offset + len - tail_start;
2986 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2987 if (unlikely(ret < 0))
2988 goto out_only_mutex;
2990 /* zero the front end of the last page */
2991 if (tail_start + tail_len < ino_size) {
2992 truncated_block = true;
2993 ret = btrfs_truncate_block(BTRFS_I(inode),
2994 tail_start + tail_len,
2997 goto out_only_mutex;
3002 if (lockend < lockstart) {
3004 goto out_only_mutex;
3007 btrfs_punch_hole_lock_range(inode, lockstart, lockend, &cached_state);
3009 path = btrfs_alloc_path();
3015 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
3016 lockend, NULL, &trans);
3017 btrfs_free_path(path);
3021 ASSERT(trans != NULL);
3022 inode_inc_iversion(inode);
3023 inode->i_mtime = current_time(inode);
3024 inode->i_ctime = inode->i_mtime;
3025 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3026 updated_inode = true;
3027 btrfs_end_transaction(trans);
3028 btrfs_btree_balance_dirty(fs_info);
3030 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3033 if (!updated_inode && truncated_block && !ret) {
3035 * If we only end up zeroing part of a page, we still need to
3036 * update the inode item, so that all the time fields are
3037 * updated as well as the necessary btrfs inode in memory fields
3038 * for detecting, at fsync time, if the inode isn't yet in the
3039 * log tree or it's there but not up to date.
3041 struct timespec64 now = current_time(inode);
3043 inode_inc_iversion(inode);
3044 inode->i_mtime = now;
3045 inode->i_ctime = now;
3046 trans = btrfs_start_transaction(root, 1);
3047 if (IS_ERR(trans)) {
3048 ret = PTR_ERR(trans);
3052 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3053 ret2 = btrfs_end_transaction(trans);
3058 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3062 /* Helper structure to record which range is already reserved */
3063 struct falloc_range {
3064 struct list_head list;
3070 * Helper function to add falloc range
3072 * Caller should have locked the larger range of extent containing
3075 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3077 struct falloc_range *range = NULL;
3079 if (!list_empty(head)) {
3081 * As fallocate iterates by bytenr order, we only need to check
3084 range = list_last_entry(head, struct falloc_range, list);
3085 if (range->start + range->len == start) {
3091 range = kmalloc(sizeof(*range), GFP_KERNEL);
3094 range->start = start;
3096 list_add_tail(&range->list, head);
3100 static int btrfs_fallocate_update_isize(struct inode *inode,
3104 struct btrfs_trans_handle *trans;
3105 struct btrfs_root *root = BTRFS_I(inode)->root;
3109 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3112 trans = btrfs_start_transaction(root, 1);
3114 return PTR_ERR(trans);
3116 inode->i_ctime = current_time(inode);
3117 i_size_write(inode, end);
3118 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
3119 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3120 ret2 = btrfs_end_transaction(trans);
3122 return ret ? ret : ret2;
3126 RANGE_BOUNDARY_WRITTEN_EXTENT,
3127 RANGE_BOUNDARY_PREALLOC_EXTENT,
3128 RANGE_BOUNDARY_HOLE,
3131 static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3134 const u64 sectorsize = inode->root->fs_info->sectorsize;
3135 struct extent_map *em;
3138 offset = round_down(offset, sectorsize);
3139 em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3143 if (em->block_start == EXTENT_MAP_HOLE)
3144 ret = RANGE_BOUNDARY_HOLE;
3145 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3146 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3148 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3150 free_extent_map(em);
3154 static int btrfs_zero_range(struct inode *inode,
3159 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3160 struct extent_map *em;
3161 struct extent_changeset *data_reserved = NULL;
3164 const u64 sectorsize = fs_info->sectorsize;
3165 u64 alloc_start = round_down(offset, sectorsize);
3166 u64 alloc_end = round_up(offset + len, sectorsize);
3167 u64 bytes_to_reserve = 0;
3168 bool space_reserved = false;
3170 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3171 alloc_end - alloc_start);
3178 * Avoid hole punching and extent allocation for some cases. More cases
3179 * could be considered, but these are unlikely common and we keep things
3180 * as simple as possible for now. Also, intentionally, if the target
3181 * range contains one or more prealloc extents together with regular
3182 * extents and holes, we drop all the existing extents and allocate a
3183 * new prealloc extent, so that we get a larger contiguous disk extent.
3185 if (em->start <= alloc_start &&
3186 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3187 const u64 em_end = em->start + em->len;
3189 if (em_end >= offset + len) {
3191 * The whole range is already a prealloc extent,
3192 * do nothing except updating the inode's i_size if
3195 free_extent_map(em);
3196 ret = btrfs_fallocate_update_isize(inode, offset + len,
3201 * Part of the range is already a prealloc extent, so operate
3202 * only on the remaining part of the range.
3204 alloc_start = em_end;
3205 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3206 len = offset + len - alloc_start;
3207 offset = alloc_start;
3208 alloc_hint = em->block_start + em->len;
3210 free_extent_map(em);
3212 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3213 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3214 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3221 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3222 free_extent_map(em);
3223 ret = btrfs_fallocate_update_isize(inode, offset + len,
3227 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3228 free_extent_map(em);
3229 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
3232 ret = btrfs_fallocate_update_isize(inode,
3237 free_extent_map(em);
3238 alloc_start = round_down(offset, sectorsize);
3239 alloc_end = alloc_start + sectorsize;
3243 alloc_start = round_up(offset, sectorsize);
3244 alloc_end = round_down(offset + len, sectorsize);
3247 * For unaligned ranges, check the pages at the boundaries, they might
3248 * map to an extent, in which case we need to partially zero them, or
3249 * they might map to a hole, in which case we need our allocation range
3252 if (!IS_ALIGNED(offset, sectorsize)) {
3253 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3257 if (ret == RANGE_BOUNDARY_HOLE) {
3258 alloc_start = round_down(offset, sectorsize);
3260 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3261 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
3269 if (!IS_ALIGNED(offset + len, sectorsize)) {
3270 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3274 if (ret == RANGE_BOUNDARY_HOLE) {
3275 alloc_end = round_up(offset + len, sectorsize);
3277 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3278 ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
3288 if (alloc_start < alloc_end) {
3289 struct extent_state *cached_state = NULL;
3290 const u64 lockstart = alloc_start;
3291 const u64 lockend = alloc_end - 1;
3293 bytes_to_reserve = alloc_end - alloc_start;
3294 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3298 space_reserved = true;
3299 btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3301 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3302 alloc_start, bytes_to_reserve);
3304 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
3305 lockend, &cached_state);
3308 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3309 alloc_end - alloc_start,
3311 offset + len, &alloc_hint);
3312 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3314 /* btrfs_prealloc_file_range releases reserved space on error */
3316 space_reserved = false;
3320 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3322 if (ret && space_reserved)
3323 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3324 alloc_start, bytes_to_reserve);
3325 extent_changeset_free(data_reserved);
3330 static long btrfs_fallocate(struct file *file, int mode,
3331 loff_t offset, loff_t len)
3333 struct inode *inode = file_inode(file);
3334 struct extent_state *cached_state = NULL;
3335 struct extent_changeset *data_reserved = NULL;
3336 struct falloc_range *range;
3337 struct falloc_range *tmp;
3338 struct list_head reserve_list;
3346 u64 data_space_needed = 0;
3347 u64 data_space_reserved = 0;
3348 u64 qgroup_reserved = 0;
3349 struct extent_map *em;
3350 int blocksize = BTRFS_I(inode)->root->fs_info->sectorsize;
3353 /* Do not allow fallocate in ZONED mode */
3354 if (btrfs_is_zoned(btrfs_sb(inode->i_sb)))
3357 alloc_start = round_down(offset, blocksize);
3358 alloc_end = round_up(offset + len, blocksize);
3359 cur_offset = alloc_start;
3361 /* Make sure we aren't being give some crap mode */
3362 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3363 FALLOC_FL_ZERO_RANGE))
3366 if (mode & FALLOC_FL_PUNCH_HOLE)
3367 return btrfs_punch_hole(file, offset, len);
3369 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
3371 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3372 ret = inode_newsize_ok(inode, offset + len);
3377 ret = file_modified(file);
3382 * TODO: Move these two operations after we have checked
3383 * accurate reserved space, or fallocate can still fail but
3384 * with page truncated or size expanded.
3386 * But that's a minor problem and won't do much harm BTW.
3388 if (alloc_start > inode->i_size) {
3389 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3393 } else if (offset + len > inode->i_size) {
3395 * If we are fallocating from the end of the file onward we
3396 * need to zero out the end of the block if i_size lands in the
3397 * middle of a block.
3399 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3405 * We have locked the inode at the VFS level (in exclusive mode) and we
3406 * have locked the i_mmap_lock lock (in exclusive mode). Now before
3407 * locking the file range, flush all dealloc in the range and wait for
3408 * all ordered extents in the range to complete. After this we can lock
3409 * the file range and, due to the previous locking we did, we know there
3410 * can't be more delalloc or ordered extents in the range.
3412 ret = btrfs_wait_ordered_range(inode, alloc_start,
3413 alloc_end - alloc_start);
3417 if (mode & FALLOC_FL_ZERO_RANGE) {
3418 ret = btrfs_zero_range(inode, offset, len, mode);
3419 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3423 locked_end = alloc_end - 1;
3424 lock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3427 btrfs_assert_inode_range_clean(BTRFS_I(inode), alloc_start, locked_end);
3429 /* First, check if we exceed the qgroup limit */
3430 INIT_LIST_HEAD(&reserve_list);
3431 while (cur_offset < alloc_end) {
3432 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3433 alloc_end - cur_offset);
3438 last_byte = min(extent_map_end(em), alloc_end);
3439 actual_end = min_t(u64, extent_map_end(em), offset + len);
3440 last_byte = ALIGN(last_byte, blocksize);
3441 if (em->block_start == EXTENT_MAP_HOLE ||
3442 (cur_offset >= inode->i_size &&
3443 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3444 const u64 range_len = last_byte - cur_offset;
3446 ret = add_falloc_range(&reserve_list, cur_offset, range_len);
3448 free_extent_map(em);
3451 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3452 &data_reserved, cur_offset, range_len);
3454 free_extent_map(em);
3457 qgroup_reserved += range_len;
3458 data_space_needed += range_len;
3460 free_extent_map(em);
3461 cur_offset = last_byte;
3464 if (!ret && data_space_needed > 0) {
3466 * We are safe to reserve space here as we can't have delalloc
3467 * in the range, see above.
3469 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3472 data_space_reserved = data_space_needed;
3476 * If ret is still 0, means we're OK to fallocate.
3477 * Or just cleanup the list and exit.
3479 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3481 ret = btrfs_prealloc_file_range(inode, mode,
3483 range->len, i_blocksize(inode),
3484 offset + len, &alloc_hint);
3486 * btrfs_prealloc_file_range() releases space even
3487 * if it returns an error.
3489 data_space_reserved -= range->len;
3490 qgroup_reserved -= range->len;
3491 } else if (data_space_reserved > 0) {
3492 btrfs_free_reserved_data_space(BTRFS_I(inode),
3493 data_reserved, range->start,
3495 data_space_reserved -= range->len;
3496 qgroup_reserved -= range->len;
3497 } else if (qgroup_reserved > 0) {
3498 btrfs_qgroup_free_data(BTRFS_I(inode), data_reserved,
3499 range->start, range->len);
3500 qgroup_reserved -= range->len;
3502 list_del(&range->list);
3509 * We didn't need to allocate any more space, but we still extended the
3510 * size of the file so we need to update i_size and the inode item.
3512 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3514 unlock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3517 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3518 extent_changeset_free(data_reserved);
3523 * Helper for btrfs_find_delalloc_in_range(). Find a subrange in a given range
3524 * that has unflushed and/or flushing delalloc. There might be other adjacent
3525 * subranges after the one it found, so btrfs_find_delalloc_in_range() keeps
3526 * looping while it gets adjacent subranges, and merging them together.
3528 static bool find_delalloc_subrange(struct btrfs_inode *inode, u64 start, u64 end,
3529 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3531 const u64 len = end + 1 - start;
3532 struct extent_map_tree *em_tree = &inode->extent_tree;
3533 struct extent_map *em;
3538 * Search the io tree first for EXTENT_DELALLOC. If we find any, it
3539 * means we have delalloc (dirty pages) for which writeback has not
3542 *delalloc_start_ret = start;
3543 delalloc_len = count_range_bits(&inode->io_tree, delalloc_start_ret, end,
3544 len, EXTENT_DELALLOC, 1);
3546 * If delalloc was found then *delalloc_start_ret has a sector size
3547 * aligned value (rounded down).
3549 if (delalloc_len > 0)
3550 *delalloc_end_ret = *delalloc_start_ret + delalloc_len - 1;
3553 * Now also check if there's any extent map in the range that does not
3554 * map to a hole or prealloc extent. We do this because:
3556 * 1) When delalloc is flushed, the file range is locked, we clear the
3557 * EXTENT_DELALLOC bit from the io tree and create an extent map for
3558 * an allocated extent. So we might just have been called after
3559 * delalloc is flushed and before the ordered extent completes and
3560 * inserts the new file extent item in the subvolume's btree;
3562 * 2) We may have an extent map created by flushing delalloc for a
3563 * subrange that starts before the subrange we found marked with
3564 * EXTENT_DELALLOC in the io tree.
3566 read_lock(&em_tree->lock);
3567 em = lookup_extent_mapping(em_tree, start, len);
3568 read_unlock(&em_tree->lock);
3570 /* extent_map_end() returns a non-inclusive end offset. */
3571 em_end = em ? extent_map_end(em) : 0;
3574 * If we have a hole/prealloc extent map, check the next one if this one
3575 * ends before our range's end.
3577 if (em && (em->block_start == EXTENT_MAP_HOLE ||
3578 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) && em_end < end) {
3579 struct extent_map *next_em;
3581 read_lock(&em_tree->lock);
3582 next_em = lookup_extent_mapping(em_tree, em_end, len - em_end);
3583 read_unlock(&em_tree->lock);
3585 free_extent_map(em);
3586 em_end = next_em ? extent_map_end(next_em) : 0;
3590 if (em && (em->block_start == EXTENT_MAP_HOLE ||
3591 test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3592 free_extent_map(em);
3597 * No extent map or one for a hole or prealloc extent. Use the delalloc
3598 * range we found in the io tree if we have one.
3601 return (delalloc_len > 0);
3604 * We don't have any range as EXTENT_DELALLOC in the io tree, so the
3605 * extent map is the only subrange representing delalloc.
3607 if (delalloc_len == 0) {
3608 *delalloc_start_ret = em->start;
3609 *delalloc_end_ret = min(end, em_end - 1);
3610 free_extent_map(em);
3615 * The extent map represents a delalloc range that starts before the
3616 * delalloc range we found in the io tree.
3618 if (em->start < *delalloc_start_ret) {
3619 *delalloc_start_ret = em->start;
3621 * If the ranges are adjacent, return a combined range.
3622 * Otherwise return the extent map's range.
3624 if (em_end < *delalloc_start_ret)
3625 *delalloc_end_ret = min(end, em_end - 1);
3627 free_extent_map(em);
3632 * The extent map starts after the delalloc range we found in the io
3633 * tree. If it's adjacent, return a combined range, otherwise return
3634 * the range found in the io tree.
3636 if (*delalloc_end_ret + 1 == em->start)
3637 *delalloc_end_ret = min(end, em_end - 1);
3639 free_extent_map(em);
3644 * Check if there's delalloc in a given range.
3646 * @inode: The inode.
3647 * @start: The start offset of the range. It does not need to be
3648 * sector size aligned.
3649 * @end: The end offset (inclusive value) of the search range.
3650 * It does not need to be sector size aligned.
3651 * @delalloc_start_ret: Output argument, set to the start offset of the
3652 * subrange found with delalloc (may not be sector size
3654 * @delalloc_end_ret: Output argument, set to he end offset (inclusive value)
3655 * of the subrange found with delalloc.
3657 * Returns true if a subrange with delalloc is found within the given range, and
3658 * if so it sets @delalloc_start_ret and @delalloc_end_ret with the start and
3659 * end offsets of the subrange.
3661 bool btrfs_find_delalloc_in_range(struct btrfs_inode *inode, u64 start, u64 end,
3662 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3664 u64 cur_offset = round_down(start, inode->root->fs_info->sectorsize);
3665 u64 prev_delalloc_end = 0;
3668 while (cur_offset < end) {
3673 delalloc = find_delalloc_subrange(inode, cur_offset, end,
3679 if (prev_delalloc_end == 0) {
3680 /* First subrange found. */
3681 *delalloc_start_ret = max(delalloc_start, start);
3682 *delalloc_end_ret = delalloc_end;
3684 } else if (delalloc_start == prev_delalloc_end + 1) {
3685 /* Subrange adjacent to the previous one, merge them. */
3686 *delalloc_end_ret = delalloc_end;
3688 /* Subrange not adjacent to the previous one, exit. */
3692 prev_delalloc_end = delalloc_end;
3693 cur_offset = delalloc_end + 1;
3701 * Check if there's a hole or delalloc range in a range representing a hole (or
3702 * prealloc extent) found in the inode's subvolume btree.
3704 * @inode: The inode.
3705 * @whence: Seek mode (SEEK_DATA or SEEK_HOLE).
3706 * @start: Start offset of the hole region. It does not need to be sector
3708 * @end: End offset (inclusive value) of the hole region. It does not
3709 * need to be sector size aligned.
3710 * @start_ret: Return parameter, used to set the start of the subrange in the
3711 * hole that matches the search criteria (seek mode), if such
3712 * subrange is found (return value of the function is true).
3713 * The value returned here may not be sector size aligned.
3715 * Returns true if a subrange matching the given seek mode is found, and if one
3716 * is found, it updates @start_ret with the start of the subrange.
3718 static bool find_desired_extent_in_hole(struct btrfs_inode *inode, int whence,
3719 u64 start, u64 end, u64 *start_ret)
3725 delalloc = btrfs_find_delalloc_in_range(inode, start, end,
3726 &delalloc_start, &delalloc_end);
3727 if (delalloc && whence == SEEK_DATA) {
3728 *start_ret = delalloc_start;
3732 if (delalloc && whence == SEEK_HOLE) {
3734 * We found delalloc but it starts after out start offset. So we
3735 * have a hole between our start offset and the delalloc start.
3737 if (start < delalloc_start) {
3742 * Delalloc range starts at our start offset.
3743 * If the delalloc range's length is smaller than our range,
3744 * then it means we have a hole that starts where the delalloc
3747 if (delalloc_end < end) {
3748 *start_ret = delalloc_end + 1;
3752 /* There's delalloc for the whole range. */
3756 if (!delalloc && whence == SEEK_HOLE) {
3762 * No delalloc in the range and we are seeking for data. The caller has
3763 * to iterate to the next extent item in the subvolume btree.
3768 static loff_t find_desired_extent(struct btrfs_inode *inode, loff_t offset,
3771 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3772 struct extent_state *cached_state = NULL;
3773 const loff_t i_size = i_size_read(&inode->vfs_inode);
3774 const u64 ino = btrfs_ino(inode);
3775 struct btrfs_root *root = inode->root;
3776 struct btrfs_path *path;
3777 struct btrfs_key key;
3778 u64 last_extent_end;
3785 if (i_size == 0 || offset >= i_size)
3789 * Quick path. If the inode has no prealloc extents and its number of
3790 * bytes used matches its i_size, then it can not have holes.
3792 if (whence == SEEK_HOLE &&
3793 !(inode->flags & BTRFS_INODE_PREALLOC) &&
3794 inode_get_bytes(&inode->vfs_inode) == i_size)
3798 * offset can be negative, in this case we start finding DATA/HOLE from
3799 * the very start of the file.
3801 start = max_t(loff_t, 0, offset);
3803 lockstart = round_down(start, fs_info->sectorsize);
3804 lockend = round_up(i_size, fs_info->sectorsize);
3805 if (lockend <= lockstart)
3806 lockend = lockstart + fs_info->sectorsize;
3809 path = btrfs_alloc_path();
3812 path->reada = READA_FORWARD;
3815 key.type = BTRFS_EXTENT_DATA_KEY;
3818 last_extent_end = lockstart;
3820 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3822 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3825 } else if (ret > 0 && path->slots[0] > 0) {
3826 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
3827 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
3831 while (start < i_size) {
3832 struct extent_buffer *leaf = path->nodes[0];
3833 struct btrfs_file_extent_item *extent;
3836 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3837 ret = btrfs_next_leaf(root, path);
3843 leaf = path->nodes[0];
3846 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3847 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3850 extent_end = btrfs_file_extent_end(path);
3853 * In the first iteration we may have a slot that points to an
3854 * extent that ends before our start offset, so skip it.
3856 if (extent_end <= start) {
3861 /* We have an implicit hole, NO_HOLES feature is likely set. */
3862 if (last_extent_end < key.offset) {
3863 u64 search_start = last_extent_end;
3867 * First iteration, @start matches @offset and it's
3870 if (start == offset)
3871 search_start = offset;
3873 found = find_desired_extent_in_hole(inode, whence,
3878 start = found_start;
3882 * Didn't find data or a hole (due to delalloc) in the
3883 * implicit hole range, so need to analyze the extent.
3887 extent = btrfs_item_ptr(leaf, path->slots[0],
3888 struct btrfs_file_extent_item);
3890 if (btrfs_file_extent_disk_bytenr(leaf, extent) == 0 ||
3891 btrfs_file_extent_type(leaf, extent) ==
3892 BTRFS_FILE_EXTENT_PREALLOC) {
3894 * Explicit hole or prealloc extent, search for delalloc.
3895 * A prealloc extent is treated like a hole.
3897 u64 search_start = key.offset;
3901 * First iteration, @start matches @offset and it's
3904 if (start == offset)
3905 search_start = offset;
3907 found = find_desired_extent_in_hole(inode, whence,
3912 start = found_start;
3916 * Didn't find data or a hole (due to delalloc) in the
3917 * implicit hole range, so need to analyze the next
3922 * Found a regular or inline extent.
3923 * If we are seeking for data, adjust the start offset
3924 * and stop, we're done.
3926 if (whence == SEEK_DATA) {
3927 start = max_t(u64, key.offset, offset);
3932 * Else, we are seeking for a hole, check the next file
3938 last_extent_end = extent_end;
3940 if (fatal_signal_pending(current)) {
3947 /* We have an implicit hole from the last extent found up to i_size. */
3948 if (!found && start < i_size) {
3949 found = find_desired_extent_in_hole(inode, whence, start,
3950 i_size - 1, &start);
3956 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3957 btrfs_free_path(path);
3962 if (whence == SEEK_DATA && start >= i_size)
3965 return min_t(loff_t, start, i_size);
3968 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3970 struct inode *inode = file->f_mapping->host;
3974 return generic_file_llseek(file, offset, whence);
3977 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3978 offset = find_desired_extent(BTRFS_I(inode), offset, whence);
3979 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3986 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3989 static int btrfs_file_open(struct inode *inode, struct file *filp)
3993 filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC | FMODE_BUF_WASYNC;
3995 ret = fsverity_file_open(inode, filp);
3998 return generic_file_open(inode, filp);
4001 static int check_direct_read(struct btrfs_fs_info *fs_info,
4002 const struct iov_iter *iter, loff_t offset)
4007 ret = check_direct_IO(fs_info, iter, offset);
4011 if (!iter_is_iovec(iter))
4014 for (seg = 0; seg < iter->nr_segs; seg++)
4015 for (i = seg + 1; i < iter->nr_segs; i++)
4016 if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
4021 static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
4023 struct inode *inode = file_inode(iocb->ki_filp);
4024 size_t prev_left = 0;
4028 if (fsverity_active(inode))
4031 if (check_direct_read(btrfs_sb(inode->i_sb), to, iocb->ki_pos))
4034 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
4037 * This is similar to what we do for direct IO writes, see the comment
4038 * at btrfs_direct_write(), but we also disable page faults in addition
4039 * to disabling them only at the iov_iter level. This is because when
4040 * reading from a hole or prealloc extent, iomap calls iov_iter_zero(),
4041 * which can still trigger page fault ins despite having set ->nofault
4042 * to true of our 'to' iov_iter.
4044 * The difference to direct IO writes is that we deadlock when trying
4045 * to lock the extent range in the inode's tree during he page reads
4046 * triggered by the fault in (while for writes it is due to waiting for
4047 * our own ordered extent). This is because for direct IO reads,
4048 * btrfs_dio_iomap_begin() returns with the extent range locked, which
4049 * is only unlocked in the endio callback (end_bio_extent_readpage()).
4051 pagefault_disable();
4053 ret = btrfs_dio_read(iocb, to, read);
4054 to->nofault = false;
4057 /* No increment (+=) because iomap returns a cumulative value. */
4061 if (iov_iter_count(to) > 0 && (ret == -EFAULT || ret > 0)) {
4062 const size_t left = iov_iter_count(to);
4064 if (left == prev_left) {
4066 * We didn't make any progress since the last attempt,
4067 * fallback to a buffered read for the remainder of the
4068 * range. This is just to avoid any possibility of looping
4074 * We made some progress since the last retry or this is
4075 * the first time we are retrying. Fault in as many pages
4076 * as possible and retry.
4078 fault_in_iov_iter_writeable(to, left);
4083 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
4084 return ret < 0 ? ret : read;
4087 static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
4091 if (iocb->ki_flags & IOCB_DIRECT) {
4092 ret = btrfs_direct_read(iocb, to);
4093 if (ret < 0 || !iov_iter_count(to) ||
4094 iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
4098 return filemap_read(iocb, to, ret);
4101 const struct file_operations btrfs_file_operations = {
4102 .llseek = btrfs_file_llseek,
4103 .read_iter = btrfs_file_read_iter,
4104 .splice_read = generic_file_splice_read,
4105 .write_iter = btrfs_file_write_iter,
4106 .splice_write = iter_file_splice_write,
4107 .mmap = btrfs_file_mmap,
4108 .open = btrfs_file_open,
4109 .release = btrfs_release_file,
4110 .get_unmapped_area = thp_get_unmapped_area,
4111 .fsync = btrfs_sync_file,
4112 .fallocate = btrfs_fallocate,
4113 .unlocked_ioctl = btrfs_ioctl,
4114 #ifdef CONFIG_COMPAT
4115 .compat_ioctl = btrfs_compat_ioctl,
4117 .remap_file_range = btrfs_remap_file_range,
4120 void __cold btrfs_auto_defrag_exit(void)
4122 kmem_cache_destroy(btrfs_inode_defrag_cachep);
4125 int __init btrfs_auto_defrag_init(void)
4127 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
4128 sizeof(struct inode_defrag), 0,
4131 if (!btrfs_inode_defrag_cachep)
4137 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
4142 * So with compression we will find and lock a dirty page and clear the
4143 * first one as dirty, setup an async extent, and immediately return
4144 * with the entire range locked but with nobody actually marked with
4145 * writeback. So we can't just filemap_write_and_wait_range() and
4146 * expect it to work since it will just kick off a thread to do the
4147 * actual work. So we need to call filemap_fdatawrite_range _again_
4148 * since it will wait on the page lock, which won't be unlocked until
4149 * after the pages have been marked as writeback and so we're good to go
4150 * from there. We have to do this otherwise we'll miss the ordered
4151 * extents and that results in badness. Please Josef, do not think you
4152 * know better and pull this out at some point in the future, it is
4153 * right and you are wrong.
4155 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
4156 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
4157 &BTRFS_I(inode)->runtime_flags))
4158 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);