xref: /linux/fs/btrfs/reflink.c (revision 2cd86f02c017bf9733e5cd891381b7d40f6f37ad)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include <linux/blkdev.h>
4 #include <linux/iversion.h>
5 #include "ctree.h"
6 #include "fs.h"
7 #include "messages.h"
8 #include "compression.h"
9 #include "delalloc-space.h"
10 #include "disk-io.h"
11 #include "reflink.h"
12 #include "transaction.h"
13 #include "subpage.h"
14 #include "accessors.h"
15 #include "file-item.h"
16 #include "file.h"
17 #include "super.h"
18 
19 #define BTRFS_MAX_DEDUPE_LEN	SZ_16M
20 
21 static int clone_finish_inode_update(struct btrfs_trans_handle *trans,
22 				     struct inode *inode,
23 				     u64 endoff,
24 				     const u64 destoff,
25 				     const u64 olen,
26 				     int no_time_update)
27 {
28 	int ret;
29 
30 	inode_inc_iversion(inode);
31 	if (!no_time_update) {
32 		inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
33 	}
34 	/*
35 	 * We round up to the block size at eof when determining which
36 	 * extents to clone above, but shouldn't round up the file size.
37 	 */
38 	if (endoff > destoff + olen)
39 		endoff = destoff + olen;
40 	if (endoff > inode->i_size) {
41 		i_size_write(inode, endoff);
42 		btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
43 	}
44 
45 	ret = btrfs_update_inode(trans, BTRFS_I(inode));
46 	if (ret) {
47 		btrfs_abort_transaction(trans, ret);
48 		btrfs_end_transaction(trans);
49 		goto out;
50 	}
51 	ret = btrfs_end_transaction(trans);
52 out:
53 	return ret;
54 }
55 
56 static int copy_inline_to_page(struct btrfs_inode *inode,
57 			       const u64 file_offset,
58 			       char *inline_data,
59 			       const u64 size,
60 			       const u64 datal,
61 			       const u8 comp_type)
62 {
63 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
64 	const u32 block_size = fs_info->sectorsize;
65 	const u64 range_end = file_offset + block_size - 1;
66 	const size_t inline_size = size - btrfs_file_extent_calc_inline_size(0);
67 	char *data_start = inline_data + btrfs_file_extent_calc_inline_size(0);
68 	struct extent_changeset *data_reserved = NULL;
69 	struct folio *folio = NULL;
70 	struct address_space *mapping = inode->vfs_inode.i_mapping;
71 	int ret;
72 
73 	ASSERT(IS_ALIGNED(file_offset, block_size));
74 
75 	/*
76 	 * We have flushed and locked the ranges of the source and destination
77 	 * inodes, we also have locked the inodes, so we are safe to do a
78 	 * reservation here. Also we must not do the reservation while holding
79 	 * a transaction open, otherwise we would deadlock.
80 	 */
81 	ret = btrfs_delalloc_reserve_space(inode, &data_reserved, file_offset,
82 					   block_size);
83 	if (ret)
84 		goto out;
85 
86 	folio = __filemap_get_folio(mapping, file_offset >> PAGE_SHIFT,
87 					FGP_LOCK | FGP_ACCESSED | FGP_CREAT,
88 					btrfs_alloc_write_mask(mapping));
89 	if (IS_ERR(folio)) {
90 		ret = -ENOMEM;
91 		goto out_unlock;
92 	}
93 
94 	ret = set_folio_extent_mapped(folio);
95 	if (ret < 0)
96 		goto out_unlock;
97 
98 	clear_extent_bit(&inode->io_tree, file_offset, range_end,
99 			 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
100 			 NULL);
101 	ret = btrfs_set_extent_delalloc(inode, file_offset, range_end, 0, NULL);
102 	if (ret)
103 		goto out_unlock;
104 
105 	/*
106 	 * After dirtying the page our caller will need to start a transaction,
107 	 * and if we are low on metadata free space, that can cause flushing of
108 	 * delalloc for all inodes in order to get metadata space released.
109 	 * However we are holding the range locked for the whole duration of
110 	 * the clone/dedupe operation, so we may deadlock if that happens and no
111 	 * other task releases enough space. So mark this inode as not being
112 	 * possible to flush to avoid such deadlock. We will clear that flag
113 	 * when we finish cloning all extents, since a transaction is started
114 	 * after finding each extent to clone.
115 	 */
116 	set_bit(BTRFS_INODE_NO_DELALLOC_FLUSH, &inode->runtime_flags);
117 
118 	if (comp_type == BTRFS_COMPRESS_NONE) {
119 		memcpy_to_folio(folio, offset_in_folio(folio, file_offset), data_start,
120 					datal);
121 	} else {
122 		ret = btrfs_decompress(comp_type, data_start, folio,
123 				       offset_in_folio(folio, file_offset),
124 				       inline_size, datal);
125 		if (ret)
126 			goto out_unlock;
127 		flush_dcache_folio(folio);
128 	}
129 
130 	/*
131 	 * If our inline data is smaller then the block/page size, then the
132 	 * remaining of the block/page is equivalent to zeroes. We had something
133 	 * like the following done:
134 	 *
135 	 * $ xfs_io -f -c "pwrite -S 0xab 0 500" file
136 	 * $ sync  # (or fsync)
137 	 * $ xfs_io -c "falloc 0 4K" file
138 	 * $ xfs_io -c "pwrite -S 0xcd 4K 4K"
139 	 *
140 	 * So what's in the range [500, 4095] corresponds to zeroes.
141 	 */
142 	if (datal < block_size)
143 		folio_zero_range(folio, datal, block_size - datal);
144 
145 	btrfs_folio_set_uptodate(fs_info, folio, file_offset, block_size);
146 	btrfs_folio_clear_checked(fs_info, folio, file_offset, block_size);
147 	btrfs_folio_set_dirty(fs_info, folio, file_offset, block_size);
148 out_unlock:
149 	if (!IS_ERR(folio)) {
150 		folio_unlock(folio);
151 		folio_put(folio);
152 	}
153 	if (ret)
154 		btrfs_delalloc_release_space(inode, data_reserved, file_offset,
155 					     block_size, true);
156 	btrfs_delalloc_release_extents(inode, block_size);
157 out:
158 	extent_changeset_free(data_reserved);
159 
160 	return ret;
161 }
162 
163 /*
164  * Deal with cloning of inline extents. We try to copy the inline extent from
165  * the source inode to destination inode when possible. When not possible we
166  * copy the inline extent's data into the respective page of the inode.
167  */
168 static int clone_copy_inline_extent(struct inode *dst,
169 				    struct btrfs_path *path,
170 				    struct btrfs_key *new_key,
171 				    const u64 drop_start,
172 				    const u64 datal,
173 				    const u64 size,
174 				    const u8 comp_type,
175 				    char *inline_data,
176 				    struct btrfs_trans_handle **trans_out)
177 {
178 	struct btrfs_fs_info *fs_info = inode_to_fs_info(dst);
179 	struct btrfs_root *root = BTRFS_I(dst)->root;
180 	const u64 aligned_end = ALIGN(new_key->offset + datal,
181 				      fs_info->sectorsize);
182 	struct btrfs_trans_handle *trans = NULL;
183 	struct btrfs_drop_extents_args drop_args = { 0 };
184 	int ret;
185 	struct btrfs_key key;
186 
187 	if (new_key->offset > 0) {
188 		ret = copy_inline_to_page(BTRFS_I(dst), new_key->offset,
189 					  inline_data, size, datal, comp_type);
190 		goto out;
191 	}
192 
193 	key.objectid = btrfs_ino(BTRFS_I(dst));
194 	key.type = BTRFS_EXTENT_DATA_KEY;
195 	key.offset = 0;
196 	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
197 	if (ret < 0) {
198 		return ret;
199 	} else if (ret > 0) {
200 		if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
201 			ret = btrfs_next_leaf(root, path);
202 			if (ret < 0)
203 				return ret;
204 			else if (ret > 0)
205 				goto copy_inline_extent;
206 		}
207 		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
208 		if (key.objectid == btrfs_ino(BTRFS_I(dst)) &&
209 		    key.type == BTRFS_EXTENT_DATA_KEY) {
210 			/*
211 			 * There's an implicit hole at file offset 0, copy the
212 			 * inline extent's data to the page.
213 			 */
214 			ASSERT(key.offset > 0);
215 			goto copy_to_page;
216 		}
217 	} else if (i_size_read(dst) <= datal) {
218 		struct btrfs_file_extent_item *ei;
219 
220 		ei = btrfs_item_ptr(path->nodes[0], path->slots[0],
221 				    struct btrfs_file_extent_item);
222 		/*
223 		 * If it's an inline extent replace it with the source inline
224 		 * extent, otherwise copy the source inline extent data into
225 		 * the respective page at the destination inode.
226 		 */
227 		if (btrfs_file_extent_type(path->nodes[0], ei) ==
228 		    BTRFS_FILE_EXTENT_INLINE)
229 			goto copy_inline_extent;
230 
231 		goto copy_to_page;
232 	}
233 
234 copy_inline_extent:
235 	/*
236 	 * We have no extent items, or we have an extent at offset 0 which may
237 	 * or may not be inlined. All these cases are dealt the same way.
238 	 */
239 	if (i_size_read(dst) > datal) {
240 		/*
241 		 * At the destination offset 0 we have either a hole, a regular
242 		 * extent or an inline extent larger then the one we want to
243 		 * clone. Deal with all these cases by copying the inline extent
244 		 * data into the respective page at the destination inode.
245 		 */
246 		goto copy_to_page;
247 	}
248 
249 	/*
250 	 * Release path before starting a new transaction so we don't hold locks
251 	 * that would confuse lockdep.
252 	 */
253 	btrfs_release_path(path);
254 	/*
255 	 * If we end up here it means were copy the inline extent into a leaf
256 	 * of the destination inode. We know we will drop or adjust at most one
257 	 * extent item in the destination root.
258 	 *
259 	 * 1 unit - adjusting old extent (we may have to split it)
260 	 * 1 unit - add new extent
261 	 * 1 unit - inode update
262 	 */
263 	trans = btrfs_start_transaction(root, 3);
264 	if (IS_ERR(trans)) {
265 		ret = PTR_ERR(trans);
266 		trans = NULL;
267 		goto out;
268 	}
269 	drop_args.path = path;
270 	drop_args.start = drop_start;
271 	drop_args.end = aligned_end;
272 	drop_args.drop_cache = true;
273 	ret = btrfs_drop_extents(trans, root, BTRFS_I(dst), &drop_args);
274 	if (ret)
275 		goto out;
276 	ret = btrfs_insert_empty_item(trans, root, path, new_key, size);
277 	if (ret)
278 		goto out;
279 
280 	write_extent_buffer(path->nodes[0], inline_data,
281 			    btrfs_item_ptr_offset(path->nodes[0],
282 						  path->slots[0]),
283 			    size);
284 	btrfs_update_inode_bytes(BTRFS_I(dst), datal, drop_args.bytes_found);
285 	btrfs_set_inode_full_sync(BTRFS_I(dst));
286 	ret = btrfs_inode_set_file_extent_range(BTRFS_I(dst), 0, aligned_end);
287 out:
288 	if (!ret && !trans) {
289 		/*
290 		 * No transaction here means we copied the inline extent into a
291 		 * page of the destination inode.
292 		 *
293 		 * 1 unit to update inode item
294 		 */
295 		trans = btrfs_start_transaction(root, 1);
296 		if (IS_ERR(trans)) {
297 			ret = PTR_ERR(trans);
298 			trans = NULL;
299 		}
300 	}
301 	if (ret && trans) {
302 		btrfs_abort_transaction(trans, ret);
303 		btrfs_end_transaction(trans);
304 	}
305 	if (!ret)
306 		*trans_out = trans;
307 
308 	return ret;
309 
310 copy_to_page:
311 	/*
312 	 * Release our path because we don't need it anymore and also because
313 	 * copy_inline_to_page() needs to reserve data and metadata, which may
314 	 * need to flush delalloc when we are low on available space and
315 	 * therefore cause a deadlock if writeback of an inline extent needs to
316 	 * write to the same leaf or an ordered extent completion needs to write
317 	 * to the same leaf.
318 	 */
319 	btrfs_release_path(path);
320 
321 	ret = copy_inline_to_page(BTRFS_I(dst), new_key->offset,
322 				  inline_data, size, datal, comp_type);
323 	goto out;
324 }
325 
326 /*
327  * Clone a range from inode file to another.
328  *
329  * @src:             Inode to clone from
330  * @inode:           Inode to clone to
331  * @off:             Offset within source to start clone from
332  * @olen:            Original length, passed by user, of range to clone
333  * @olen_aligned:    Block-aligned value of olen
334  * @destoff:         Offset within @inode to start clone
335  * @no_time_update:  Whether to update mtime/ctime on the target inode
336  */
337 static int btrfs_clone(struct inode *src, struct inode *inode,
338 		       const u64 off, const u64 olen, const u64 olen_aligned,
339 		       const u64 destoff, int no_time_update)
340 {
341 	struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
342 	struct btrfs_path *path = NULL;
343 	struct extent_buffer *leaf;
344 	struct btrfs_trans_handle *trans;
345 	char *buf = NULL;
346 	struct btrfs_key key;
347 	u32 nritems;
348 	int slot;
349 	int ret;
350 	const u64 len = olen_aligned;
351 	u64 last_dest_end = destoff;
352 	u64 prev_extent_end = off;
353 
354 	ret = -ENOMEM;
355 	buf = kvmalloc(fs_info->nodesize, GFP_KERNEL);
356 	if (!buf)
357 		return ret;
358 
359 	path = btrfs_alloc_path();
360 	if (!path) {
361 		kvfree(buf);
362 		return ret;
363 	}
364 
365 	path->reada = READA_FORWARD;
366 	/* Clone data */
367 	key.objectid = btrfs_ino(BTRFS_I(src));
368 	key.type = BTRFS_EXTENT_DATA_KEY;
369 	key.offset = off;
370 
371 	while (1) {
372 		struct btrfs_file_extent_item *extent;
373 		u64 extent_gen;
374 		int type;
375 		u32 size;
376 		struct btrfs_key new_key;
377 		u64 disko = 0, diskl = 0;
378 		u64 datao = 0, datal = 0;
379 		u8 comp;
380 		u64 drop_start;
381 
382 		/* Note the key will change type as we walk through the tree */
383 		ret = btrfs_search_slot(NULL, BTRFS_I(src)->root, &key, path,
384 				0, 0);
385 		if (ret < 0)
386 			goto out;
387 		/*
388 		 * First search, if no extent item that starts at offset off was
389 		 * found but the previous item is an extent item, it's possible
390 		 * it might overlap our target range, therefore process it.
391 		 */
392 		if (key.offset == off && ret > 0 && path->slots[0] > 0) {
393 			btrfs_item_key_to_cpu(path->nodes[0], &key,
394 					      path->slots[0] - 1);
395 			if (key.type == BTRFS_EXTENT_DATA_KEY)
396 				path->slots[0]--;
397 		}
398 
399 		nritems = btrfs_header_nritems(path->nodes[0]);
400 process_slot:
401 		if (path->slots[0] >= nritems) {
402 			ret = btrfs_next_leaf(BTRFS_I(src)->root, path);
403 			if (ret < 0)
404 				goto out;
405 			if (ret > 0)
406 				break;
407 			nritems = btrfs_header_nritems(path->nodes[0]);
408 		}
409 		leaf = path->nodes[0];
410 		slot = path->slots[0];
411 
412 		btrfs_item_key_to_cpu(leaf, &key, slot);
413 		if (key.type > BTRFS_EXTENT_DATA_KEY ||
414 		    key.objectid != btrfs_ino(BTRFS_I(src)))
415 			break;
416 
417 		ASSERT(key.type == BTRFS_EXTENT_DATA_KEY);
418 
419 		extent = btrfs_item_ptr(leaf, slot,
420 					struct btrfs_file_extent_item);
421 		extent_gen = btrfs_file_extent_generation(leaf, extent);
422 		comp = btrfs_file_extent_compression(leaf, extent);
423 		type = btrfs_file_extent_type(leaf, extent);
424 		if (type == BTRFS_FILE_EXTENT_REG ||
425 		    type == BTRFS_FILE_EXTENT_PREALLOC) {
426 			disko = btrfs_file_extent_disk_bytenr(leaf, extent);
427 			diskl = btrfs_file_extent_disk_num_bytes(leaf, extent);
428 			datao = btrfs_file_extent_offset(leaf, extent);
429 			datal = btrfs_file_extent_num_bytes(leaf, extent);
430 		} else if (type == BTRFS_FILE_EXTENT_INLINE) {
431 			/* Take upper bound, may be compressed */
432 			datal = btrfs_file_extent_ram_bytes(leaf, extent);
433 		}
434 
435 		/*
436 		 * The first search might have left us at an extent item that
437 		 * ends before our target range's start, can happen if we have
438 		 * holes and NO_HOLES feature enabled.
439 		 *
440 		 * Subsequent searches may leave us on a file range we have
441 		 * processed before - this happens due to a race with ordered
442 		 * extent completion for a file range that is outside our source
443 		 * range, but that range was part of a file extent item that
444 		 * also covered a leading part of our source range.
445 		 */
446 		if (key.offset + datal <= prev_extent_end) {
447 			path->slots[0]++;
448 			goto process_slot;
449 		} else if (key.offset >= off + len) {
450 			break;
451 		}
452 
453 		prev_extent_end = key.offset + datal;
454 		size = btrfs_item_size(leaf, slot);
455 		read_extent_buffer(leaf, buf, btrfs_item_ptr_offset(leaf, slot),
456 				   size);
457 
458 		btrfs_release_path(path);
459 
460 		memcpy(&new_key, &key, sizeof(new_key));
461 		new_key.objectid = btrfs_ino(BTRFS_I(inode));
462 		if (off <= key.offset)
463 			new_key.offset = key.offset + destoff - off;
464 		else
465 			new_key.offset = destoff;
466 
467 		/*
468 		 * Deal with a hole that doesn't have an extent item that
469 		 * represents it (NO_HOLES feature enabled).
470 		 * This hole is either in the middle of the cloning range or at
471 		 * the beginning (fully overlaps it or partially overlaps it).
472 		 */
473 		if (new_key.offset != last_dest_end)
474 			drop_start = last_dest_end;
475 		else
476 			drop_start = new_key.offset;
477 
478 		if (type == BTRFS_FILE_EXTENT_REG ||
479 		    type == BTRFS_FILE_EXTENT_PREALLOC) {
480 			struct btrfs_replace_extent_info clone_info;
481 
482 			/*
483 			 *    a  | --- range to clone ---|  b
484 			 * | ------------- extent ------------- |
485 			 */
486 
487 			/* Subtract range b */
488 			if (key.offset + datal > off + len)
489 				datal = off + len - key.offset;
490 
491 			/* Subtract range a */
492 			if (off > key.offset) {
493 				datao += off - key.offset;
494 				datal -= off - key.offset;
495 			}
496 
497 			clone_info.disk_offset = disko;
498 			clone_info.disk_len = diskl;
499 			clone_info.data_offset = datao;
500 			clone_info.data_len = datal;
501 			clone_info.file_offset = new_key.offset;
502 			clone_info.extent_buf = buf;
503 			clone_info.is_new_extent = false;
504 			clone_info.update_times = !no_time_update;
505 			ret = btrfs_replace_file_extents(BTRFS_I(inode), path,
506 					drop_start, new_key.offset + datal - 1,
507 					&clone_info, &trans);
508 			if (ret)
509 				goto out;
510 		} else {
511 			ASSERT(type == BTRFS_FILE_EXTENT_INLINE);
512 			/*
513 			 * Inline extents always have to start at file offset 0
514 			 * and can never be bigger then the sector size. We can
515 			 * never clone only parts of an inline extent, since all
516 			 * reflink operations must start at a sector size aligned
517 			 * offset, and the length must be aligned too or end at
518 			 * the i_size (which implies the whole inlined data).
519 			 */
520 			ASSERT(key.offset == 0);
521 			ASSERT(datal <= fs_info->sectorsize);
522 			if (WARN_ON(type != BTRFS_FILE_EXTENT_INLINE) ||
523 			    WARN_ON(key.offset != 0) ||
524 			    WARN_ON(datal > fs_info->sectorsize)) {
525 				ret = -EUCLEAN;
526 				goto out;
527 			}
528 
529 			ret = clone_copy_inline_extent(inode, path, &new_key,
530 						       drop_start, datal, size,
531 						       comp, buf, &trans);
532 			if (ret)
533 				goto out;
534 		}
535 
536 		btrfs_release_path(path);
537 
538 		/*
539 		 * Whenever we share an extent we update the last_reflink_trans
540 		 * of each inode to the current transaction. This is needed to
541 		 * make sure fsync does not log multiple checksum items with
542 		 * overlapping ranges (because some extent items might refer
543 		 * only to sections of the original extent). For the destination
544 		 * inode we do this regardless of the generation of the extents
545 		 * or even if they are inline extents or explicit holes, to make
546 		 * sure a full fsync does not skip them. For the source inode,
547 		 * we only need to update last_reflink_trans in case it's a new
548 		 * extent that is not a hole or an inline extent, to deal with
549 		 * the checksums problem on fsync.
550 		 */
551 		if (extent_gen == trans->transid && disko > 0)
552 			BTRFS_I(src)->last_reflink_trans = trans->transid;
553 
554 		BTRFS_I(inode)->last_reflink_trans = trans->transid;
555 
556 		last_dest_end = ALIGN(new_key.offset + datal,
557 				      fs_info->sectorsize);
558 		ret = clone_finish_inode_update(trans, inode, last_dest_end,
559 						destoff, olen, no_time_update);
560 		if (ret)
561 			goto out;
562 		if (new_key.offset + datal >= destoff + len)
563 			break;
564 
565 		btrfs_release_path(path);
566 		key.offset = prev_extent_end;
567 
568 		if (fatal_signal_pending(current)) {
569 			ret = -EINTR;
570 			goto out;
571 		}
572 
573 		cond_resched();
574 	}
575 	ret = 0;
576 
577 	if (last_dest_end < destoff + len) {
578 		/*
579 		 * We have an implicit hole that fully or partially overlaps our
580 		 * cloning range at its end. This means that we either have the
581 		 * NO_HOLES feature enabled or the implicit hole happened due to
582 		 * mixing buffered and direct IO writes against this file.
583 		 */
584 		btrfs_release_path(path);
585 
586 		/*
587 		 * When using NO_HOLES and we are cloning a range that covers
588 		 * only a hole (no extents) into a range beyond the current
589 		 * i_size, punching a hole in the target range will not create
590 		 * an extent map defining a hole, because the range starts at or
591 		 * beyond current i_size. If the file previously had an i_size
592 		 * greater than the new i_size set by this clone operation, we
593 		 * need to make sure the next fsync is a full fsync, so that it
594 		 * detects and logs a hole covering a range from the current
595 		 * i_size to the new i_size. If the clone range covers extents,
596 		 * besides a hole, then we know the full sync flag was already
597 		 * set by previous calls to btrfs_replace_file_extents() that
598 		 * replaced file extent items.
599 		 */
600 		if (last_dest_end >= i_size_read(inode))
601 			btrfs_set_inode_full_sync(BTRFS_I(inode));
602 
603 		ret = btrfs_replace_file_extents(BTRFS_I(inode), path,
604 				last_dest_end, destoff + len - 1, NULL, &trans);
605 		if (ret)
606 			goto out;
607 
608 		ret = clone_finish_inode_update(trans, inode, destoff + len,
609 						destoff, olen, no_time_update);
610 	}
611 
612 out:
613 	btrfs_free_path(path);
614 	kvfree(buf);
615 	clear_bit(BTRFS_INODE_NO_DELALLOC_FLUSH, &BTRFS_I(inode)->runtime_flags);
616 
617 	return ret;
618 }
619 
620 static void btrfs_double_mmap_lock(struct inode *inode1, struct inode *inode2)
621 {
622 	if (inode1 < inode2)
623 		swap(inode1, inode2);
624 	down_write(&BTRFS_I(inode1)->i_mmap_lock);
625 	down_write_nested(&BTRFS_I(inode2)->i_mmap_lock, SINGLE_DEPTH_NESTING);
626 }
627 
628 static void btrfs_double_mmap_unlock(struct inode *inode1, struct inode *inode2)
629 {
630 	up_write(&BTRFS_I(inode1)->i_mmap_lock);
631 	up_write(&BTRFS_I(inode2)->i_mmap_lock);
632 }
633 
634 static int btrfs_extent_same_range(struct inode *src, u64 loff, u64 len,
635 				   struct inode *dst, u64 dst_loff)
636 {
637 	const u64 end = dst_loff + len - 1;
638 	struct extent_state *cached_state = NULL;
639 	struct btrfs_fs_info *fs_info = BTRFS_I(src)->root->fs_info;
640 	const u64 bs = fs_info->sectorsize;
641 	int ret;
642 
643 	/*
644 	 * Lock destination range to serialize with concurrent readahead(), and
645 	 * we are safe from concurrency with relocation of source extents
646 	 * because we have already locked the inode's i_mmap_lock in exclusive
647 	 * mode.
648 	 */
649 	lock_extent(&BTRFS_I(dst)->io_tree, dst_loff, end, &cached_state);
650 	ret = btrfs_clone(src, dst, loff, len, ALIGN(len, bs), dst_loff, 1);
651 	unlock_extent(&BTRFS_I(dst)->io_tree, dst_loff, end, &cached_state);
652 
653 	btrfs_btree_balance_dirty(fs_info);
654 
655 	return ret;
656 }
657 
658 static int btrfs_extent_same(struct inode *src, u64 loff, u64 olen,
659 			     struct inode *dst, u64 dst_loff)
660 {
661 	int ret = 0;
662 	u64 i, tail_len, chunk_count;
663 	struct btrfs_root *root_dst = BTRFS_I(dst)->root;
664 
665 	spin_lock(&root_dst->root_item_lock);
666 	if (root_dst->send_in_progress) {
667 		btrfs_warn_rl(root_dst->fs_info,
668 "cannot deduplicate to root %llu while send operations are using it (%d in progress)",
669 			      btrfs_root_id(root_dst),
670 			      root_dst->send_in_progress);
671 		spin_unlock(&root_dst->root_item_lock);
672 		return -EAGAIN;
673 	}
674 	root_dst->dedupe_in_progress++;
675 	spin_unlock(&root_dst->root_item_lock);
676 
677 	tail_len = olen % BTRFS_MAX_DEDUPE_LEN;
678 	chunk_count = div_u64(olen, BTRFS_MAX_DEDUPE_LEN);
679 
680 	for (i = 0; i < chunk_count; i++) {
681 		ret = btrfs_extent_same_range(src, loff, BTRFS_MAX_DEDUPE_LEN,
682 					      dst, dst_loff);
683 		if (ret)
684 			goto out;
685 
686 		loff += BTRFS_MAX_DEDUPE_LEN;
687 		dst_loff += BTRFS_MAX_DEDUPE_LEN;
688 	}
689 
690 	if (tail_len > 0)
691 		ret = btrfs_extent_same_range(src, loff, tail_len, dst, dst_loff);
692 out:
693 	spin_lock(&root_dst->root_item_lock);
694 	root_dst->dedupe_in_progress--;
695 	spin_unlock(&root_dst->root_item_lock);
696 
697 	return ret;
698 }
699 
700 static noinline int btrfs_clone_files(struct file *file, struct file *file_src,
701 					u64 off, u64 olen, u64 destoff)
702 {
703 	struct extent_state *cached_state = NULL;
704 	struct inode *inode = file_inode(file);
705 	struct inode *src = file_inode(file_src);
706 	struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
707 	int ret;
708 	int wb_ret;
709 	u64 len = olen;
710 	u64 bs = fs_info->sectorsize;
711 	u64 end;
712 
713 	/*
714 	 * VFS's generic_remap_file_range_prep() protects us from cloning the
715 	 * eof block into the middle of a file, which would result in corruption
716 	 * if the file size is not blocksize aligned. So we don't need to check
717 	 * for that case here.
718 	 */
719 	if (off + len == src->i_size)
720 		len = ALIGN(src->i_size, bs) - off;
721 
722 	if (destoff > inode->i_size) {
723 		const u64 wb_start = ALIGN_DOWN(inode->i_size, bs);
724 
725 		ret = btrfs_cont_expand(BTRFS_I(inode), inode->i_size, destoff);
726 		if (ret)
727 			return ret;
728 		/*
729 		 * We may have truncated the last block if the inode's size is
730 		 * not sector size aligned, so we need to wait for writeback to
731 		 * complete before proceeding further, otherwise we can race
732 		 * with cloning and attempt to increment a reference to an
733 		 * extent that no longer exists (writeback completed right after
734 		 * we found the previous extent covering eof and before we
735 		 * attempted to increment its reference count).
736 		 */
737 		ret = btrfs_wait_ordered_range(BTRFS_I(inode), wb_start,
738 					       destoff - wb_start);
739 		if (ret)
740 			return ret;
741 	}
742 
743 	/*
744 	 * Lock destination range to serialize with concurrent readahead(), and
745 	 * we are safe from concurrency with relocation of source extents
746 	 * because we have already locked the inode's i_mmap_lock in exclusive
747 	 * mode.
748 	 */
749 	end = destoff + len - 1;
750 	lock_extent(&BTRFS_I(inode)->io_tree, destoff, end, &cached_state);
751 	ret = btrfs_clone(src, inode, off, olen, len, destoff, 0);
752 	unlock_extent(&BTRFS_I(inode)->io_tree, destoff, end, &cached_state);
753 
754 	/*
755 	 * We may have copied an inline extent into a page of the destination
756 	 * range, so wait for writeback to complete before truncating pages
757 	 * from the page cache. This is a rare case.
758 	 */
759 	wb_ret = btrfs_wait_ordered_range(BTRFS_I(inode), destoff, len);
760 	ret = ret ? ret : wb_ret;
761 	/*
762 	 * Truncate page cache pages so that future reads will see the cloned
763 	 * data immediately and not the previous data.
764 	 */
765 	truncate_inode_pages_range(&inode->i_data,
766 				round_down(destoff, PAGE_SIZE),
767 				round_up(destoff + len, PAGE_SIZE) - 1);
768 
769 	btrfs_btree_balance_dirty(fs_info);
770 
771 	return ret;
772 }
773 
774 static int btrfs_remap_file_range_prep(struct file *file_in, loff_t pos_in,
775 				       struct file *file_out, loff_t pos_out,
776 				       loff_t *len, unsigned int remap_flags)
777 {
778 	struct inode *inode_in = file_inode(file_in);
779 	struct inode *inode_out = file_inode(file_out);
780 	u64 bs = BTRFS_I(inode_out)->root->fs_info->sectorsize;
781 	u64 wb_len;
782 	int ret;
783 
784 	if (!(remap_flags & REMAP_FILE_DEDUP)) {
785 		struct btrfs_root *root_out = BTRFS_I(inode_out)->root;
786 
787 		if (btrfs_root_readonly(root_out))
788 			return -EROFS;
789 
790 		ASSERT(inode_in->i_sb == inode_out->i_sb);
791 	}
792 
793 	/* Don't make the dst file partly checksummed */
794 	if ((BTRFS_I(inode_in)->flags & BTRFS_INODE_NODATASUM) !=
795 	    (BTRFS_I(inode_out)->flags & BTRFS_INODE_NODATASUM)) {
796 		return -EINVAL;
797 	}
798 
799 	/*
800 	 * Now that the inodes are locked, we need to start writeback ourselves
801 	 * and can not rely on the writeback from the VFS's generic helper
802 	 * generic_remap_file_range_prep() because:
803 	 *
804 	 * 1) For compression we must call filemap_fdatawrite_range() range
805 	 *    twice (btrfs_fdatawrite_range() does it for us), and the generic
806 	 *    helper only calls it once;
807 	 *
808 	 * 2) filemap_fdatawrite_range(), called by the generic helper only
809 	 *    waits for the writeback to complete, i.e. for IO to be done, and
810 	 *    not for the ordered extents to complete. We need to wait for them
811 	 *    to complete so that new file extent items are in the fs tree.
812 	 */
813 	if (*len == 0 && !(remap_flags & REMAP_FILE_DEDUP))
814 		wb_len = ALIGN(inode_in->i_size, bs) - ALIGN_DOWN(pos_in, bs);
815 	else
816 		wb_len = ALIGN(*len, bs);
817 
818 	/*
819 	 * Workaround to make sure NOCOW buffered write reach disk as NOCOW.
820 	 *
821 	 * Btrfs' back references do not have a block level granularity, they
822 	 * work at the whole extent level.
823 	 * NOCOW buffered write without data space reserved may not be able
824 	 * to fall back to CoW due to lack of data space, thus could cause
825 	 * data loss.
826 	 *
827 	 * Here we take a shortcut by flushing the whole inode, so that all
828 	 * nocow write should reach disk as nocow before we increase the
829 	 * reference of the extent. We could do better by only flushing NOCOW
830 	 * data, but that needs extra accounting.
831 	 *
832 	 * Also we don't need to check ASYNC_EXTENT, as async extent will be
833 	 * CoWed anyway, not affecting nocow part.
834 	 */
835 	ret = filemap_flush(inode_in->i_mapping);
836 	if (ret < 0)
837 		return ret;
838 
839 	ret = btrfs_wait_ordered_range(BTRFS_I(inode_in), ALIGN_DOWN(pos_in, bs),
840 				       wb_len);
841 	if (ret < 0)
842 		return ret;
843 	ret = btrfs_wait_ordered_range(BTRFS_I(inode_out), ALIGN_DOWN(pos_out, bs),
844 				       wb_len);
845 	if (ret < 0)
846 		return ret;
847 
848 	return generic_remap_file_range_prep(file_in, pos_in, file_out, pos_out,
849 					    len, remap_flags);
850 }
851 
852 static bool file_sync_write(const struct file *file)
853 {
854 	if (file->f_flags & (__O_SYNC | O_DSYNC))
855 		return true;
856 	if (IS_SYNC(file_inode(file)))
857 		return true;
858 
859 	return false;
860 }
861 
862 loff_t btrfs_remap_file_range(struct file *src_file, loff_t off,
863 		struct file *dst_file, loff_t destoff, loff_t len,
864 		unsigned int remap_flags)
865 {
866 	struct inode *src_inode = file_inode(src_file);
867 	struct inode *dst_inode = file_inode(dst_file);
868 	bool same_inode = dst_inode == src_inode;
869 	int ret;
870 
871 	if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
872 		return -EINVAL;
873 
874 	if (same_inode) {
875 		btrfs_inode_lock(BTRFS_I(src_inode), BTRFS_ILOCK_MMAP);
876 	} else {
877 		lock_two_nondirectories(src_inode, dst_inode);
878 		btrfs_double_mmap_lock(src_inode, dst_inode);
879 	}
880 
881 	ret = btrfs_remap_file_range_prep(src_file, off, dst_file, destoff,
882 					  &len, remap_flags);
883 	if (ret < 0 || len == 0)
884 		goto out_unlock;
885 
886 	if (remap_flags & REMAP_FILE_DEDUP)
887 		ret = btrfs_extent_same(src_inode, off, len, dst_inode, destoff);
888 	else
889 		ret = btrfs_clone_files(dst_file, src_file, off, len, destoff);
890 
891 out_unlock:
892 	if (same_inode) {
893 		btrfs_inode_unlock(BTRFS_I(src_inode), BTRFS_ILOCK_MMAP);
894 	} else {
895 		btrfs_double_mmap_unlock(src_inode, dst_inode);
896 		unlock_two_nondirectories(src_inode, dst_inode);
897 	}
898 
899 	/*
900 	 * If either the source or the destination file was opened with O_SYNC,
901 	 * O_DSYNC or has the S_SYNC attribute, fsync both the destination and
902 	 * source files/ranges, so that after a successful return (0) followed
903 	 * by a power failure results in the reflinked data to be readable from
904 	 * both files/ranges.
905 	 */
906 	if (ret == 0 && len > 0 &&
907 	    (file_sync_write(src_file) || file_sync_write(dst_file))) {
908 		ret = btrfs_sync_file(src_file, off, off + len - 1, 0);
909 		if (ret == 0)
910 			ret = btrfs_sync_file(dst_file, destoff,
911 					      destoff + len - 1, 0);
912 	}
913 
914 	return ret < 0 ? ret : len;
915 }
916