xref: /linux/fs/btrfs/reflink.c (revision 90e0d94d369d342e735a75174439482119b6c393)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include <linux/blkdev.h>
4 #include <linux/iversion.h>
5 #include "compression.h"
6 #include "ctree.h"
7 #include "delalloc-space.h"
8 #include "disk-io.h"
9 #include "reflink.h"
10 #include "transaction.h"
11 #include "subpage.h"
12 
13 #define BTRFS_MAX_DEDUPE_LEN	SZ_16M
14 
15 static int clone_finish_inode_update(struct btrfs_trans_handle *trans,
16 				     struct inode *inode,
17 				     u64 endoff,
18 				     const u64 destoff,
19 				     const u64 olen,
20 				     int no_time_update)
21 {
22 	struct btrfs_root *root = BTRFS_I(inode)->root;
23 	int ret;
24 
25 	inode_inc_iversion(inode);
26 	if (!no_time_update) {
27 		inode->i_mtime = current_time(inode);
28 		inode->i_ctime = inode->i_mtime;
29 	}
30 	/*
31 	 * We round up to the block size at eof when determining which
32 	 * extents to clone above, but shouldn't round up the file size.
33 	 */
34 	if (endoff > destoff + olen)
35 		endoff = destoff + olen;
36 	if (endoff > inode->i_size) {
37 		i_size_write(inode, endoff);
38 		btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
39 	}
40 
41 	ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
42 	if (ret) {
43 		btrfs_abort_transaction(trans, ret);
44 		btrfs_end_transaction(trans);
45 		goto out;
46 	}
47 	ret = btrfs_end_transaction(trans);
48 out:
49 	return ret;
50 }
51 
52 static int copy_inline_to_page(struct btrfs_inode *inode,
53 			       const u64 file_offset,
54 			       char *inline_data,
55 			       const u64 size,
56 			       const u64 datal,
57 			       const u8 comp_type)
58 {
59 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
60 	const u32 block_size = fs_info->sectorsize;
61 	const u64 range_end = file_offset + block_size - 1;
62 	const size_t inline_size = size - btrfs_file_extent_calc_inline_size(0);
63 	char *data_start = inline_data + btrfs_file_extent_calc_inline_size(0);
64 	struct extent_changeset *data_reserved = NULL;
65 	struct page *page = NULL;
66 	struct address_space *mapping = inode->vfs_inode.i_mapping;
67 	int ret;
68 
69 	ASSERT(IS_ALIGNED(file_offset, block_size));
70 
71 	/*
72 	 * We have flushed and locked the ranges of the source and destination
73 	 * inodes, we also have locked the inodes, so we are safe to do a
74 	 * reservation here. Also we must not do the reservation while holding
75 	 * a transaction open, otherwise we would deadlock.
76 	 */
77 	ret = btrfs_delalloc_reserve_space(inode, &data_reserved, file_offset,
78 					   block_size);
79 	if (ret)
80 		goto out;
81 
82 	page = find_or_create_page(mapping, file_offset >> PAGE_SHIFT,
83 				   btrfs_alloc_write_mask(mapping));
84 	if (!page) {
85 		ret = -ENOMEM;
86 		goto out_unlock;
87 	}
88 
89 	ret = set_page_extent_mapped(page);
90 	if (ret < 0)
91 		goto out_unlock;
92 
93 	clear_extent_bit(&inode->io_tree, file_offset, range_end,
94 			 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
95 			 NULL);
96 	ret = btrfs_set_extent_delalloc(inode, file_offset, range_end, 0, NULL);
97 	if (ret)
98 		goto out_unlock;
99 
100 	/*
101 	 * After dirtying the page our caller will need to start a transaction,
102 	 * and if we are low on metadata free space, that can cause flushing of
103 	 * delalloc for all inodes in order to get metadata space released.
104 	 * However we are holding the range locked for the whole duration of
105 	 * the clone/dedupe operation, so we may deadlock if that happens and no
106 	 * other task releases enough space. So mark this inode as not being
107 	 * possible to flush to avoid such deadlock. We will clear that flag
108 	 * when we finish cloning all extents, since a transaction is started
109 	 * after finding each extent to clone.
110 	 */
111 	set_bit(BTRFS_INODE_NO_DELALLOC_FLUSH, &inode->runtime_flags);
112 
113 	if (comp_type == BTRFS_COMPRESS_NONE) {
114 		memcpy_to_page(page, offset_in_page(file_offset), data_start,
115 			       datal);
116 	} else {
117 		ret = btrfs_decompress(comp_type, data_start, page,
118 				       offset_in_page(file_offset),
119 				       inline_size, datal);
120 		if (ret)
121 			goto out_unlock;
122 		flush_dcache_page(page);
123 	}
124 
125 	/*
126 	 * If our inline data is smaller then the block/page size, then the
127 	 * remaining of the block/page is equivalent to zeroes. We had something
128 	 * like the following done:
129 	 *
130 	 * $ xfs_io -f -c "pwrite -S 0xab 0 500" file
131 	 * $ sync  # (or fsync)
132 	 * $ xfs_io -c "falloc 0 4K" file
133 	 * $ xfs_io -c "pwrite -S 0xcd 4K 4K"
134 	 *
135 	 * So what's in the range [500, 4095] corresponds to zeroes.
136 	 */
137 	if (datal < block_size)
138 		memzero_page(page, datal, block_size - datal);
139 
140 	btrfs_page_set_uptodate(fs_info, page, file_offset, block_size);
141 	btrfs_page_clear_checked(fs_info, page, file_offset, block_size);
142 	btrfs_page_set_dirty(fs_info, page, file_offset, block_size);
143 out_unlock:
144 	if (page) {
145 		unlock_page(page);
146 		put_page(page);
147 	}
148 	if (ret)
149 		btrfs_delalloc_release_space(inode, data_reserved, file_offset,
150 					     block_size, true);
151 	btrfs_delalloc_release_extents(inode, block_size);
152 out:
153 	extent_changeset_free(data_reserved);
154 
155 	return ret;
156 }
157 
158 /*
159  * Deal with cloning of inline extents. We try to copy the inline extent from
160  * the source inode to destination inode when possible. When not possible we
161  * copy the inline extent's data into the respective page of the inode.
162  */
163 static int clone_copy_inline_extent(struct inode *dst,
164 				    struct btrfs_path *path,
165 				    struct btrfs_key *new_key,
166 				    const u64 drop_start,
167 				    const u64 datal,
168 				    const u64 size,
169 				    const u8 comp_type,
170 				    char *inline_data,
171 				    struct btrfs_trans_handle **trans_out)
172 {
173 	struct btrfs_fs_info *fs_info = btrfs_sb(dst->i_sb);
174 	struct btrfs_root *root = BTRFS_I(dst)->root;
175 	const u64 aligned_end = ALIGN(new_key->offset + datal,
176 				      fs_info->sectorsize);
177 	struct btrfs_trans_handle *trans = NULL;
178 	struct btrfs_drop_extents_args drop_args = { 0 };
179 	int ret;
180 	struct btrfs_key key;
181 
182 	if (new_key->offset > 0) {
183 		ret = copy_inline_to_page(BTRFS_I(dst), new_key->offset,
184 					  inline_data, size, datal, comp_type);
185 		goto out;
186 	}
187 
188 	key.objectid = btrfs_ino(BTRFS_I(dst));
189 	key.type = BTRFS_EXTENT_DATA_KEY;
190 	key.offset = 0;
191 	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
192 	if (ret < 0) {
193 		return ret;
194 	} else if (ret > 0) {
195 		if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
196 			ret = btrfs_next_leaf(root, path);
197 			if (ret < 0)
198 				return ret;
199 			else if (ret > 0)
200 				goto copy_inline_extent;
201 		}
202 		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
203 		if (key.objectid == btrfs_ino(BTRFS_I(dst)) &&
204 		    key.type == BTRFS_EXTENT_DATA_KEY) {
205 			/*
206 			 * There's an implicit hole at file offset 0, copy the
207 			 * inline extent's data to the page.
208 			 */
209 			ASSERT(key.offset > 0);
210 			goto copy_to_page;
211 		}
212 	} else if (i_size_read(dst) <= datal) {
213 		struct btrfs_file_extent_item *ei;
214 
215 		ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
216 				    struct btrfs_file_extent_item);
217 		/*
218 		 * If it's an inline extent replace it with the source inline
219 		 * extent, otherwise copy the source inline extent data into
220 		 * the respective page at the destination inode.
221 		 */
222 		if (btrfs_file_extent_type(path->nodes[0], ei) ==
223 		    BTRFS_FILE_EXTENT_INLINE)
224 			goto copy_inline_extent;
225 
226 		goto copy_to_page;
227 	}
228 
229 copy_inline_extent:
230 	/*
231 	 * We have no extent items, or we have an extent at offset 0 which may
232 	 * or may not be inlined. All these cases are dealt the same way.
233 	 */
234 	if (i_size_read(dst) > datal) {
235 		/*
236 		 * At the destination offset 0 we have either a hole, a regular
237 		 * extent or an inline extent larger then the one we want to
238 		 * clone. Deal with all these cases by copying the inline extent
239 		 * data into the respective page at the destination inode.
240 		 */
241 		goto copy_to_page;
242 	}
243 
244 	/*
245 	 * Release path before starting a new transaction so we don't hold locks
246 	 * that would confuse lockdep.
247 	 */
248 	btrfs_release_path(path);
249 	/*
250 	 * If we end up here it means were copy the inline extent into a leaf
251 	 * of the destination inode. We know we will drop or adjust at most one
252 	 * extent item in the destination root.
253 	 *
254 	 * 1 unit - adjusting old extent (we may have to split it)
255 	 * 1 unit - add new extent
256 	 * 1 unit - inode update
257 	 */
258 	trans = btrfs_start_transaction(root, 3);
259 	if (IS_ERR(trans)) {
260 		ret = PTR_ERR(trans);
261 		trans = NULL;
262 		goto out;
263 	}
264 	drop_args.path = path;
265 	drop_args.start = drop_start;
266 	drop_args.end = aligned_end;
267 	drop_args.drop_cache = true;
268 	ret = btrfs_drop_extents(trans, root, BTRFS_I(dst), &drop_args);
269 	if (ret)
270 		goto out;
271 	ret = btrfs_insert_empty_item(trans, root, path, new_key, size);
272 	if (ret)
273 		goto out;
274 
275 	write_extent_buffer(path->nodes[0], inline_data,
276 			    btrfs_item_ptr_offset(path->nodes[0],
277 						  path->slots[0]),
278 			    size);
279 	btrfs_update_inode_bytes(BTRFS_I(dst), datal, drop_args.bytes_found);
280 	btrfs_set_inode_full_sync(BTRFS_I(dst));
281 	ret = btrfs_inode_set_file_extent_range(BTRFS_I(dst), 0, aligned_end);
282 out:
283 	if (!ret && !trans) {
284 		/*
285 		 * No transaction here means we copied the inline extent into a
286 		 * page of the destination inode.
287 		 *
288 		 * 1 unit to update inode item
289 		 */
290 		trans = btrfs_start_transaction(root, 1);
291 		if (IS_ERR(trans)) {
292 			ret = PTR_ERR(trans);
293 			trans = NULL;
294 		}
295 	}
296 	if (ret && trans) {
297 		btrfs_abort_transaction(trans, ret);
298 		btrfs_end_transaction(trans);
299 	}
300 	if (!ret)
301 		*trans_out = trans;
302 
303 	return ret;
304 
305 copy_to_page:
306 	/*
307 	 * Release our path because we don't need it anymore and also because
308 	 * copy_inline_to_page() needs to reserve data and metadata, which may
309 	 * need to flush delalloc when we are low on available space and
310 	 * therefore cause a deadlock if writeback of an inline extent needs to
311 	 * write to the same leaf or an ordered extent completion needs to write
312 	 * to the same leaf.
313 	 */
314 	btrfs_release_path(path);
315 
316 	ret = copy_inline_to_page(BTRFS_I(dst), new_key->offset,
317 				  inline_data, size, datal, comp_type);
318 	goto out;
319 }
320 
321 /**
322  * btrfs_clone() - clone a range from inode file to another
323  *
324  * @src: Inode to clone from
325  * @inode: Inode to clone to
326  * @off: Offset within source to start clone from
327  * @olen: Original length, passed by user, of range to clone
328  * @olen_aligned: Block-aligned value of olen
329  * @destoff: Offset within @inode to start clone
330  * @no_time_update: Whether to update mtime/ctime on the target inode
331  */
332 static int btrfs_clone(struct inode *src, struct inode *inode,
333 		       const u64 off, const u64 olen, const u64 olen_aligned,
334 		       const u64 destoff, int no_time_update)
335 {
336 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
337 	struct btrfs_path *path = NULL;
338 	struct extent_buffer *leaf;
339 	struct btrfs_trans_handle *trans;
340 	char *buf = NULL;
341 	struct btrfs_key key;
342 	u32 nritems;
343 	int slot;
344 	int ret;
345 	const u64 len = olen_aligned;
346 	u64 last_dest_end = destoff;
347 	u64 prev_extent_end = off;
348 
349 	ret = -ENOMEM;
350 	buf = kvmalloc(fs_info->nodesize, GFP_KERNEL);
351 	if (!buf)
352 		return ret;
353 
354 	path = btrfs_alloc_path();
355 	if (!path) {
356 		kvfree(buf);
357 		return ret;
358 	}
359 
360 	path->reada = READA_FORWARD;
361 	/* Clone data */
362 	key.objectid = btrfs_ino(BTRFS_I(src));
363 	key.type = BTRFS_EXTENT_DATA_KEY;
364 	key.offset = off;
365 
366 	while (1) {
367 		struct btrfs_file_extent_item *extent;
368 		u64 extent_gen;
369 		int type;
370 		u32 size;
371 		struct btrfs_key new_key;
372 		u64 disko = 0, diskl = 0;
373 		u64 datao = 0, datal = 0;
374 		u8 comp;
375 		u64 drop_start;
376 
377 		/* Note the key will change type as we walk through the tree */
378 		ret = btrfs_search_slot(NULL, BTRFS_I(src)->root, &key, path,
379 				0, 0);
380 		if (ret < 0)
381 			goto out;
382 		/*
383 		 * First search, if no extent item that starts at offset off was
384 		 * found but the previous item is an extent item, it's possible
385 		 * it might overlap our target range, therefore process it.
386 		 */
387 		if (key.offset == off && ret > 0 && path->slots[0] > 0) {
388 			btrfs_item_key_to_cpu(path->nodes[0], &key,
389 					      path->slots[0] - 1);
390 			if (key.type == BTRFS_EXTENT_DATA_KEY)
391 				path->slots[0]--;
392 		}
393 
394 		nritems = btrfs_header_nritems(path->nodes[0]);
395 process_slot:
396 		if (path->slots[0] >= nritems) {
397 			ret = btrfs_next_leaf(BTRFS_I(src)->root, path);
398 			if (ret < 0)
399 				goto out;
400 			if (ret > 0)
401 				break;
402 			nritems = btrfs_header_nritems(path->nodes[0]);
403 		}
404 		leaf = path->nodes[0];
405 		slot = path->slots[0];
406 
407 		btrfs_item_key_to_cpu(leaf, &key, slot);
408 		if (key.type > BTRFS_EXTENT_DATA_KEY ||
409 		    key.objectid != btrfs_ino(BTRFS_I(src)))
410 			break;
411 
412 		ASSERT(key.type == BTRFS_EXTENT_DATA_KEY);
413 
414 		extent = btrfs_item_ptr(leaf, slot,
415 					struct btrfs_file_extent_item);
416 		extent_gen = btrfs_file_extent_generation(leaf, extent);
417 		comp = btrfs_file_extent_compression(leaf, extent);
418 		type = btrfs_file_extent_type(leaf, extent);
419 		if (type == BTRFS_FILE_EXTENT_REG ||
420 		    type == BTRFS_FILE_EXTENT_PREALLOC) {
421 			disko = btrfs_file_extent_disk_bytenr(leaf, extent);
422 			diskl = btrfs_file_extent_disk_num_bytes(leaf, extent);
423 			datao = btrfs_file_extent_offset(leaf, extent);
424 			datal = btrfs_file_extent_num_bytes(leaf, extent);
425 		} else if (type == BTRFS_FILE_EXTENT_INLINE) {
426 			/* Take upper bound, may be compressed */
427 			datal = btrfs_file_extent_ram_bytes(leaf, extent);
428 		}
429 
430 		/*
431 		 * The first search might have left us at an extent item that
432 		 * ends before our target range's start, can happen if we have
433 		 * holes and NO_HOLES feature enabled.
434 		 *
435 		 * Subsequent searches may leave us on a file range we have
436 		 * processed before - this happens due to a race with ordered
437 		 * extent completion for a file range that is outside our source
438 		 * range, but that range was part of a file extent item that
439 		 * also covered a leading part of our source range.
440 		 */
441 		if (key.offset + datal <= prev_extent_end) {
442 			path->slots[0]++;
443 			goto process_slot;
444 		} else if (key.offset >= off + len) {
445 			break;
446 		}
447 
448 		prev_extent_end = key.offset + datal;
449 		size = btrfs_item_size(leaf, slot);
450 		read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf, slot),
451 				   size);
452 
453 		btrfs_release_path(path);
454 
455 		memcpy(&new_key, &key, sizeof(new_key));
456 		new_key.objectid = btrfs_ino(BTRFS_I(inode));
457 		if (off <= key.offset)
458 			new_key.offset = key.offset + destoff - off;
459 		else
460 			new_key.offset = destoff;
461 
462 		/*
463 		 * Deal with a hole that doesn't have an extent item that
464 		 * represents it (NO_HOLES feature enabled).
465 		 * This hole is either in the middle of the cloning range or at
466 		 * the beginning (fully overlaps it or partially overlaps it).
467 		 */
468 		if (new_key.offset != last_dest_end)
469 			drop_start = last_dest_end;
470 		else
471 			drop_start = new_key.offset;
472 
473 		if (type == BTRFS_FILE_EXTENT_REG ||
474 		    type == BTRFS_FILE_EXTENT_PREALLOC) {
475 			struct btrfs_replace_extent_info clone_info;
476 
477 			/*
478 			 *    a  | --- range to clone ---|  b
479 			 * | ------------- extent ------------- |
480 			 */
481 
482 			/* Subtract range b */
483 			if (key.offset + datal > off + len)
484 				datal = off + len - key.offset;
485 
486 			/* Subtract range a */
487 			if (off > key.offset) {
488 				datao += off - key.offset;
489 				datal -= off - key.offset;
490 			}
491 
492 			clone_info.disk_offset = disko;
493 			clone_info.disk_len = diskl;
494 			clone_info.data_offset = datao;
495 			clone_info.data_len = datal;
496 			clone_info.file_offset = new_key.offset;
497 			clone_info.extent_buf = buf;
498 			clone_info.is_new_extent = false;
499 			clone_info.update_times = !no_time_update;
500 			ret = btrfs_replace_file_extents(BTRFS_I(inode), path,
501 					drop_start, new_key.offset + datal - 1,
502 					&clone_info, &trans);
503 			if (ret)
504 				goto out;
505 		} else {
506 			ASSERT(type == BTRFS_FILE_EXTENT_INLINE);
507 			/*
508 			 * Inline extents always have to start at file offset 0
509 			 * and can never be bigger then the sector size. We can
510 			 * never clone only parts of an inline extent, since all
511 			 * reflink operations must start at a sector size aligned
512 			 * offset, and the length must be aligned too or end at
513 			 * the i_size (which implies the whole inlined data).
514 			 */
515 			ASSERT(key.offset == 0);
516 			ASSERT(datal <= fs_info->sectorsize);
517 			if (WARN_ON(type != BTRFS_FILE_EXTENT_INLINE) ||
518 			    WARN_ON(key.offset != 0) ||
519 			    WARN_ON(datal > fs_info->sectorsize)) {
520 				ret = -EUCLEAN;
521 				goto out;
522 			}
523 
524 			ret = clone_copy_inline_extent(inode, path, &new_key,
525 						       drop_start, datal, size,
526 						       comp, buf, &trans);
527 			if (ret)
528 				goto out;
529 		}
530 
531 		btrfs_release_path(path);
532 
533 		/*
534 		 * Whenever we share an extent we update the last_reflink_trans
535 		 * of each inode to the current transaction. This is needed to
536 		 * make sure fsync does not log multiple checksum items with
537 		 * overlapping ranges (because some extent items might refer
538 		 * only to sections of the original extent). For the destination
539 		 * inode we do this regardless of the generation of the extents
540 		 * or even if they are inline extents or explicit holes, to make
541 		 * sure a full fsync does not skip them. For the source inode,
542 		 * we only need to update last_reflink_trans in case it's a new
543 		 * extent that is not a hole or an inline extent, to deal with
544 		 * the checksums problem on fsync.
545 		 */
546 		if (extent_gen == trans->transid && disko > 0)
547 			BTRFS_I(src)->last_reflink_trans = trans->transid;
548 
549 		BTRFS_I(inode)->last_reflink_trans = trans->transid;
550 
551 		last_dest_end = ALIGN(new_key.offset + datal,
552 				      fs_info->sectorsize);
553 		ret = clone_finish_inode_update(trans, inode, last_dest_end,
554 						destoff, olen, no_time_update);
555 		if (ret)
556 			goto out;
557 		if (new_key.offset + datal >= destoff + len)
558 			break;
559 
560 		btrfs_release_path(path);
561 		key.offset = prev_extent_end;
562 
563 		if (fatal_signal_pending(current)) {
564 			ret = -EINTR;
565 			goto out;
566 		}
567 
568 		cond_resched();
569 	}
570 	ret = 0;
571 
572 	if (last_dest_end < destoff + len) {
573 		/*
574 		 * We have an implicit hole that fully or partially overlaps our
575 		 * cloning range at its end. This means that we either have the
576 		 * NO_HOLES feature enabled or the implicit hole happened due to
577 		 * mixing buffered and direct IO writes against this file.
578 		 */
579 		btrfs_release_path(path);
580 
581 		/*
582 		 * When using NO_HOLES and we are cloning a range that covers
583 		 * only a hole (no extents) into a range beyond the current
584 		 * i_size, punching a hole in the target range will not create
585 		 * an extent map defining a hole, because the range starts at or
586 		 * beyond current i_size. If the file previously had an i_size
587 		 * greater than the new i_size set by this clone operation, we
588 		 * need to make sure the next fsync is a full fsync, so that it
589 		 * detects and logs a hole covering a range from the current
590 		 * i_size to the new i_size. If the clone range covers extents,
591 		 * besides a hole, then we know the full sync flag was already
592 		 * set by previous calls to btrfs_replace_file_extents() that
593 		 * replaced file extent items.
594 		 */
595 		if (last_dest_end >= i_size_read(inode))
596 			btrfs_set_inode_full_sync(BTRFS_I(inode));
597 
598 		ret = btrfs_replace_file_extents(BTRFS_I(inode), path,
599 				last_dest_end, destoff + len - 1, NULL, &trans);
600 		if (ret)
601 			goto out;
602 
603 		ret = clone_finish_inode_update(trans, inode, destoff + len,
604 						destoff, olen, no_time_update);
605 	}
606 
607 out:
608 	btrfs_free_path(path);
609 	kvfree(buf);
610 	clear_bit(BTRFS_INODE_NO_DELALLOC_FLUSH, &BTRFS_I(inode)->runtime_flags);
611 
612 	return ret;
613 }
614 
615 static void btrfs_double_extent_unlock(struct inode *inode1, u64 loff1,
616 				       struct inode *inode2, u64 loff2, u64 len)
617 {
618 	unlock_extent(&BTRFS_I(inode1)->io_tree, loff1, loff1 + len - 1, NULL);
619 	unlock_extent(&BTRFS_I(inode2)->io_tree, loff2, loff2 + len - 1, NULL);
620 }
621 
622 static void btrfs_double_extent_lock(struct inode *inode1, u64 loff1,
623 				     struct inode *inode2, u64 loff2, u64 len)
624 {
625 	u64 range1_end = loff1 + len - 1;
626 	u64 range2_end = loff2 + len - 1;
627 
628 	if (inode1 < inode2) {
629 		swap(inode1, inode2);
630 		swap(loff1, loff2);
631 		swap(range1_end, range2_end);
632 	} else if (inode1 == inode2 && loff2 < loff1) {
633 		swap(loff1, loff2);
634 		swap(range1_end, range2_end);
635 	}
636 
637 	lock_extent(&BTRFS_I(inode1)->io_tree, loff1, range1_end, NULL);
638 	lock_extent(&BTRFS_I(inode2)->io_tree, loff2, range2_end, NULL);
639 
640 	btrfs_assert_inode_range_clean(BTRFS_I(inode1), loff1, range1_end);
641 	btrfs_assert_inode_range_clean(BTRFS_I(inode2), loff2, range2_end);
642 }
643 
644 static void btrfs_double_mmap_lock(struct inode *inode1, struct inode *inode2)
645 {
646 	if (inode1 < inode2)
647 		swap(inode1, inode2);
648 	down_write(&BTRFS_I(inode1)->i_mmap_lock);
649 	down_write_nested(&BTRFS_I(inode2)->i_mmap_lock, SINGLE_DEPTH_NESTING);
650 }
651 
652 static void btrfs_double_mmap_unlock(struct inode *inode1, struct inode *inode2)
653 {
654 	up_write(&BTRFS_I(inode1)->i_mmap_lock);
655 	up_write(&BTRFS_I(inode2)->i_mmap_lock);
656 }
657 
658 static int btrfs_extent_same_range(struct inode *src, u64 loff, u64 len,
659 				   struct inode *dst, u64 dst_loff)
660 {
661 	struct btrfs_fs_info *fs_info = BTRFS_I(src)->root->fs_info;
662 	const u64 bs = fs_info->sb->s_blocksize;
663 	int ret;
664 
665 	/*
666 	 * Lock destination range to serialize with concurrent readahead() and
667 	 * source range to serialize with relocation.
668 	 */
669 	btrfs_double_extent_lock(src, loff, dst, dst_loff, len);
670 	ret = btrfs_clone(src, dst, loff, len, ALIGN(len, bs), dst_loff, 1);
671 	btrfs_double_extent_unlock(src, loff, dst, dst_loff, len);
672 
673 	btrfs_btree_balance_dirty(fs_info);
674 
675 	return ret;
676 }
677 
678 static int btrfs_extent_same(struct inode *src, u64 loff, u64 olen,
679 			     struct inode *dst, u64 dst_loff)
680 {
681 	int ret = 0;
682 	u64 i, tail_len, chunk_count;
683 	struct btrfs_root *root_dst = BTRFS_I(dst)->root;
684 
685 	spin_lock(&root_dst->root_item_lock);
686 	if (root_dst->send_in_progress) {
687 		btrfs_warn_rl(root_dst->fs_info,
688 "cannot deduplicate to root %llu while send operations are using it (%d in progress)",
689 			      root_dst->root_key.objectid,
690 			      root_dst->send_in_progress);
691 		spin_unlock(&root_dst->root_item_lock);
692 		return -EAGAIN;
693 	}
694 	root_dst->dedupe_in_progress++;
695 	spin_unlock(&root_dst->root_item_lock);
696 
697 	tail_len = olen % BTRFS_MAX_DEDUPE_LEN;
698 	chunk_count = div_u64(olen, BTRFS_MAX_DEDUPE_LEN);
699 
700 	for (i = 0; i < chunk_count; i++) {
701 		ret = btrfs_extent_same_range(src, loff, BTRFS_MAX_DEDUPE_LEN,
702 					      dst, dst_loff);
703 		if (ret)
704 			goto out;
705 
706 		loff += BTRFS_MAX_DEDUPE_LEN;
707 		dst_loff += BTRFS_MAX_DEDUPE_LEN;
708 	}
709 
710 	if (tail_len > 0)
711 		ret = btrfs_extent_same_range(src, loff, tail_len, dst, dst_loff);
712 out:
713 	spin_lock(&root_dst->root_item_lock);
714 	root_dst->dedupe_in_progress--;
715 	spin_unlock(&root_dst->root_item_lock);
716 
717 	return ret;
718 }
719 
720 static noinline int btrfs_clone_files(struct file *file, struct file *file_src,
721 					u64 off, u64 olen, u64 destoff)
722 {
723 	struct inode *inode = file_inode(file);
724 	struct inode *src = file_inode(file_src);
725 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
726 	int ret;
727 	int wb_ret;
728 	u64 len = olen;
729 	u64 bs = fs_info->sb->s_blocksize;
730 
731 	/*
732 	 * VFS's generic_remap_file_range_prep() protects us from cloning the
733 	 * eof block into the middle of a file, which would result in corruption
734 	 * if the file size is not blocksize aligned. So we don't need to check
735 	 * for that case here.
736 	 */
737 	if (off + len == src->i_size)
738 		len = ALIGN(src->i_size, bs) - off;
739 
740 	if (destoff > inode->i_size) {
741 		const u64 wb_start = ALIGN_DOWN(inode->i_size, bs);
742 
743 		ret = btrfs_cont_expand(BTRFS_I(inode), inode->i_size, destoff);
744 		if (ret)
745 			return ret;
746 		/*
747 		 * We may have truncated the last block if the inode's size is
748 		 * not sector size aligned, so we need to wait for writeback to
749 		 * complete before proceeding further, otherwise we can race
750 		 * with cloning and attempt to increment a reference to an
751 		 * extent that no longer exists (writeback completed right after
752 		 * we found the previous extent covering eof and before we
753 		 * attempted to increment its reference count).
754 		 */
755 		ret = btrfs_wait_ordered_range(inode, wb_start,
756 					       destoff - wb_start);
757 		if (ret)
758 			return ret;
759 	}
760 
761 	/*
762 	 * Lock destination range to serialize with concurrent readahead() and
763 	 * source range to serialize with relocation.
764 	 */
765 	btrfs_double_extent_lock(src, off, inode, destoff, len);
766 	ret = btrfs_clone(src, inode, off, olen, len, destoff, 0);
767 	btrfs_double_extent_unlock(src, off, inode, destoff, len);
768 
769 	/*
770 	 * We may have copied an inline extent into a page of the destination
771 	 * range, so wait for writeback to complete before truncating pages
772 	 * from the page cache. This is a rare case.
773 	 */
774 	wb_ret = btrfs_wait_ordered_range(inode, destoff, len);
775 	ret = ret ? ret : wb_ret;
776 	/*
777 	 * Truncate page cache pages so that future reads will see the cloned
778 	 * data immediately and not the previous data.
779 	 */
780 	truncate_inode_pages_range(&inode->i_data,
781 				round_down(destoff, PAGE_SIZE),
782 				round_up(destoff + len, PAGE_SIZE) - 1);
783 
784 	btrfs_btree_balance_dirty(fs_info);
785 
786 	return ret;
787 }
788 
789 static int btrfs_remap_file_range_prep(struct file *file_in, loff_t pos_in,
790 				       struct file *file_out, loff_t pos_out,
791 				       loff_t *len, unsigned int remap_flags)
792 {
793 	struct inode *inode_in = file_inode(file_in);
794 	struct inode *inode_out = file_inode(file_out);
795 	u64 bs = BTRFS_I(inode_out)->root->fs_info->sb->s_blocksize;
796 	u64 wb_len;
797 	int ret;
798 
799 	if (!(remap_flags & REMAP_FILE_DEDUP)) {
800 		struct btrfs_root *root_out = BTRFS_I(inode_out)->root;
801 
802 		if (btrfs_root_readonly(root_out))
803 			return -EROFS;
804 
805 		ASSERT(inode_in->i_sb == inode_out->i_sb);
806 	}
807 
808 	/* Don't make the dst file partly checksummed */
809 	if ((BTRFS_I(inode_in)->flags & BTRFS_INODE_NODATASUM) !=
810 	    (BTRFS_I(inode_out)->flags & BTRFS_INODE_NODATASUM)) {
811 		return -EINVAL;
812 	}
813 
814 	/*
815 	 * Now that the inodes are locked, we need to start writeback ourselves
816 	 * and can not rely on the writeback from the VFS's generic helper
817 	 * generic_remap_file_range_prep() because:
818 	 *
819 	 * 1) For compression we must call filemap_fdatawrite_range() range
820 	 *    twice (btrfs_fdatawrite_range() does it for us), and the generic
821 	 *    helper only calls it once;
822 	 *
823 	 * 2) filemap_fdatawrite_range(), called by the generic helper only
824 	 *    waits for the writeback to complete, i.e. for IO to be done, and
825 	 *    not for the ordered extents to complete. We need to wait for them
826 	 *    to complete so that new file extent items are in the fs tree.
827 	 */
828 	if (*len == 0 && !(remap_flags & REMAP_FILE_DEDUP))
829 		wb_len = ALIGN(inode_in->i_size, bs) - ALIGN_DOWN(pos_in, bs);
830 	else
831 		wb_len = ALIGN(*len, bs);
832 
833 	/*
834 	 * Workaround to make sure NOCOW buffered write reach disk as NOCOW.
835 	 *
836 	 * Btrfs' back references do not have a block level granularity, they
837 	 * work at the whole extent level.
838 	 * NOCOW buffered write without data space reserved may not be able
839 	 * to fall back to CoW due to lack of data space, thus could cause
840 	 * data loss.
841 	 *
842 	 * Here we take a shortcut by flushing the whole inode, so that all
843 	 * nocow write should reach disk as nocow before we increase the
844 	 * reference of the extent. We could do better by only flushing NOCOW
845 	 * data, but that needs extra accounting.
846 	 *
847 	 * Also we don't need to check ASYNC_EXTENT, as async extent will be
848 	 * CoWed anyway, not affecting nocow part.
849 	 */
850 	ret = filemap_flush(inode_in->i_mapping);
851 	if (ret < 0)
852 		return ret;
853 
854 	ret = btrfs_wait_ordered_range(inode_in, ALIGN_DOWN(pos_in, bs),
855 				       wb_len);
856 	if (ret < 0)
857 		return ret;
858 	ret = btrfs_wait_ordered_range(inode_out, ALIGN_DOWN(pos_out, bs),
859 				       wb_len);
860 	if (ret < 0)
861 		return ret;
862 
863 	return generic_remap_file_range_prep(file_in, pos_in, file_out, pos_out,
864 					    len, remap_flags);
865 }
866 
867 static bool file_sync_write(const struct file *file)
868 {
869 	if (file->f_flags & (__O_SYNC | O_DSYNC))
870 		return true;
871 	if (IS_SYNC(file_inode(file)))
872 		return true;
873 
874 	return false;
875 }
876 
877 loff_t btrfs_remap_file_range(struct file *src_file, loff_t off,
878 		struct file *dst_file, loff_t destoff, loff_t len,
879 		unsigned int remap_flags)
880 {
881 	struct inode *src_inode = file_inode(src_file);
882 	struct inode *dst_inode = file_inode(dst_file);
883 	bool same_inode = dst_inode == src_inode;
884 	int ret;
885 
886 	if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
887 		return -EINVAL;
888 
889 	if (same_inode) {
890 		btrfs_inode_lock(src_inode, BTRFS_ILOCK_MMAP);
891 	} else {
892 		lock_two_nondirectories(src_inode, dst_inode);
893 		btrfs_double_mmap_lock(src_inode, dst_inode);
894 	}
895 
896 	ret = btrfs_remap_file_range_prep(src_file, off, dst_file, destoff,
897 					  &len, remap_flags);
898 	if (ret < 0 || len == 0)
899 		goto out_unlock;
900 
901 	if (remap_flags & REMAP_FILE_DEDUP)
902 		ret = btrfs_extent_same(src_inode, off, len, dst_inode, destoff);
903 	else
904 		ret = btrfs_clone_files(dst_file, src_file, off, len, destoff);
905 
906 out_unlock:
907 	if (same_inode) {
908 		btrfs_inode_unlock(src_inode, BTRFS_ILOCK_MMAP);
909 	} else {
910 		btrfs_double_mmap_unlock(src_inode, dst_inode);
911 		unlock_two_nondirectories(src_inode, dst_inode);
912 	}
913 
914 	/*
915 	 * If either the source or the destination file was opened with O_SYNC,
916 	 * O_DSYNC or has the S_SYNC attribute, fsync both the destination and
917 	 * source files/ranges, so that after a successful return (0) followed
918 	 * by a power failure results in the reflinked data to be readable from
919 	 * both files/ranges.
920 	 */
921 	if (ret == 0 && len > 0 &&
922 	    (file_sync_write(src_file) || file_sync_write(dst_file))) {
923 		ret = btrfs_sync_file(src_file, off, off + len - 1, 0);
924 		if (ret == 0)
925 			ret = btrfs_sync_file(dst_file, destoff,
926 					      destoff + len - 1, 0);
927 	}
928 
929 	return ret < 0 ? ret : len;
930 }
931