xref: /linux/fs/btrfs/free-space-tree.c (revision 4ac81c381102bebf09a47946b343d70ed455c646)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2015 Facebook.  All rights reserved.
4  */
5 
6 #include <linux/kernel.h>
7 #include <linux/sched/mm.h>
8 #include "messages.h"
9 #include "ctree.h"
10 #include "disk-io.h"
11 #include "locking.h"
12 #include "free-space-tree.h"
13 #include "transaction.h"
14 #include "block-group.h"
15 #include "fs.h"
16 #include "accessors.h"
17 #include "extent-tree.h"
18 #include "root-tree.h"
19 
20 static int __add_block_group_free_space(struct btrfs_trans_handle *trans,
21 					struct btrfs_block_group *block_group,
22 					struct btrfs_path *path);
23 
24 struct btrfs_root *btrfs_free_space_root(struct btrfs_block_group *block_group)
25 {
26 	struct btrfs_key key = {
27 		.objectid = BTRFS_FREE_SPACE_TREE_OBJECTID,
28 		.type = BTRFS_ROOT_ITEM_KEY,
29 		.offset = 0,
30 	};
31 
32 	if (btrfs_fs_incompat(block_group->fs_info, EXTENT_TREE_V2))
33 		key.offset = block_group->global_root_id;
34 	return btrfs_global_root(block_group->fs_info, &key);
35 }
36 
37 void btrfs_set_free_space_tree_thresholds(struct btrfs_block_group *cache)
38 {
39 	u32 bitmap_range;
40 	size_t bitmap_size;
41 	u64 num_bitmaps, total_bitmap_size;
42 
43 	if (WARN_ON(cache->length == 0))
44 		btrfs_warn(cache->fs_info, "block group %llu length is zero",
45 			   cache->start);
46 
47 	/*
48 	 * We convert to bitmaps when the disk space required for using extents
49 	 * exceeds that required for using bitmaps.
50 	 */
51 	bitmap_range = cache->fs_info->sectorsize * BTRFS_FREE_SPACE_BITMAP_BITS;
52 	num_bitmaps = div_u64(cache->length + bitmap_range - 1, bitmap_range);
53 	bitmap_size = sizeof(struct btrfs_item) + BTRFS_FREE_SPACE_BITMAP_SIZE;
54 	total_bitmap_size = num_bitmaps * bitmap_size;
55 	cache->bitmap_high_thresh = div_u64(total_bitmap_size,
56 					    sizeof(struct btrfs_item));
57 
58 	/*
59 	 * We allow for a small buffer between the high threshold and low
60 	 * threshold to avoid thrashing back and forth between the two formats.
61 	 */
62 	if (cache->bitmap_high_thresh > 100)
63 		cache->bitmap_low_thresh = cache->bitmap_high_thresh - 100;
64 	else
65 		cache->bitmap_low_thresh = 0;
66 }
67 
68 static int add_new_free_space_info(struct btrfs_trans_handle *trans,
69 				   struct btrfs_block_group *block_group,
70 				   struct btrfs_path *path)
71 {
72 	struct btrfs_root *root = btrfs_free_space_root(block_group);
73 	struct btrfs_free_space_info *info;
74 	struct btrfs_key key;
75 	struct extent_buffer *leaf;
76 	int ret;
77 
78 	key.objectid = block_group->start;
79 	key.type = BTRFS_FREE_SPACE_INFO_KEY;
80 	key.offset = block_group->length;
81 
82 	ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*info));
83 	if (ret)
84 		return ret;
85 
86 	leaf = path->nodes[0];
87 	info = btrfs_item_ptr(leaf, path->slots[0],
88 			      struct btrfs_free_space_info);
89 	btrfs_set_free_space_extent_count(leaf, info, 0);
90 	btrfs_set_free_space_flags(leaf, info, 0);
91 	btrfs_release_path(path);
92 	return 0;
93 }
94 
95 struct btrfs_free_space_info *btrfs_search_free_space_info(
96 		struct btrfs_trans_handle *trans,
97 		struct btrfs_block_group *block_group,
98 		struct btrfs_path *path, int cow)
99 {
100 	struct btrfs_fs_info *fs_info = block_group->fs_info;
101 	struct btrfs_root *root = btrfs_free_space_root(block_group);
102 	struct btrfs_key key;
103 	int ret;
104 
105 	key.objectid = block_group->start;
106 	key.type = BTRFS_FREE_SPACE_INFO_KEY;
107 	key.offset = block_group->length;
108 
109 	ret = btrfs_search_slot(trans, root, &key, path, 0, cow);
110 	if (ret < 0)
111 		return ERR_PTR(ret);
112 	if (ret != 0) {
113 		btrfs_warn(fs_info, "missing free space info for %llu",
114 			   block_group->start);
115 		DEBUG_WARN();
116 		return ERR_PTR(-ENOENT);
117 	}
118 
119 	return btrfs_item_ptr(path->nodes[0], path->slots[0],
120 			      struct btrfs_free_space_info);
121 }
122 
123 /*
124  * btrfs_search_slot() but we're looking for the greatest key less than the
125  * passed key.
126  */
127 static int btrfs_search_prev_slot(struct btrfs_trans_handle *trans,
128 				  struct btrfs_root *root,
129 				  struct btrfs_key *key, struct btrfs_path *p,
130 				  int ins_len, int cow)
131 {
132 	int ret;
133 
134 	ret = btrfs_search_slot(trans, root, key, p, ins_len, cow);
135 	if (ret < 0)
136 		return ret;
137 
138 	if (unlikely(ret == 0)) {
139 		DEBUG_WARN();
140 		return -EIO;
141 	}
142 
143 	if (unlikely(p->slots[0] == 0)) {
144 		DEBUG_WARN("no previous slot found");
145 		return -EIO;
146 	}
147 	p->slots[0]--;
148 
149 	return 0;
150 }
151 
152 static inline u32 free_space_bitmap_size(const struct btrfs_fs_info *fs_info,
153 					 u64 size)
154 {
155 	return DIV_ROUND_UP(size >> fs_info->sectorsize_bits, BITS_PER_BYTE);
156 }
157 
158 static unsigned long *alloc_bitmap(u32 bitmap_size)
159 {
160 	unsigned long *ret;
161 	unsigned int nofs_flag;
162 	u32 bitmap_rounded_size = round_up(bitmap_size, sizeof(unsigned long));
163 
164 	/*
165 	 * GFP_NOFS doesn't work with kvmalloc(), but we really can't recurse
166 	 * into the filesystem here. All callers hold a transaction handle
167 	 * open, so if a GFP_KERNEL allocation recurses into the filesystem
168 	 * and triggers a transaction commit, we would deadlock.
169 	 */
170 	nofs_flag = memalloc_nofs_save();
171 	ret = kvzalloc(bitmap_rounded_size, GFP_KERNEL);
172 	memalloc_nofs_restore(nofs_flag);
173 	return ret;
174 }
175 
176 static void le_bitmap_set(unsigned long *map, unsigned int start, int len)
177 {
178 	u8 *p = ((u8 *)map) + BIT_BYTE(start);
179 	const unsigned int size = start + len;
180 	int bits_to_set = BITS_PER_BYTE - (start % BITS_PER_BYTE);
181 	u8 mask_to_set = BITMAP_FIRST_BYTE_MASK(start);
182 
183 	while (len - bits_to_set >= 0) {
184 		*p |= mask_to_set;
185 		len -= bits_to_set;
186 		bits_to_set = BITS_PER_BYTE;
187 		mask_to_set = ~0;
188 		p++;
189 	}
190 	if (len) {
191 		mask_to_set &= BITMAP_LAST_BYTE_MASK(size);
192 		*p |= mask_to_set;
193 	}
194 }
195 
196 EXPORT_FOR_TESTS
197 int btrfs_convert_free_space_to_bitmaps(struct btrfs_trans_handle *trans,
198 					struct btrfs_block_group *block_group,
199 					struct btrfs_path *path)
200 {
201 	struct btrfs_fs_info *fs_info = trans->fs_info;
202 	struct btrfs_root *root = btrfs_free_space_root(block_group);
203 	struct btrfs_free_space_info *info;
204 	struct btrfs_key key, found_key;
205 	struct extent_buffer *leaf;
206 	unsigned long *bitmap;
207 	char *bitmap_cursor;
208 	u64 start, end;
209 	u64 bitmap_range, i;
210 	u32 bitmap_size, flags, expected_extent_count;
211 	u32 extent_count = 0;
212 	int done = 0, nr;
213 	int ret;
214 
215 	bitmap_size = free_space_bitmap_size(fs_info, block_group->length);
216 	bitmap = alloc_bitmap(bitmap_size);
217 	if (unlikely(!bitmap))
218 		return 0;
219 
220 	start = block_group->start;
221 	end = btrfs_block_group_end(block_group);
222 
223 	key.objectid = end - 1;
224 	key.type = (u8)-1;
225 	key.offset = (u64)-1;
226 
227 	while (!done) {
228 		ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
229 		if (unlikely(ret)) {
230 			btrfs_abort_transaction(trans, ret);
231 			goto out;
232 		}
233 
234 		leaf = path->nodes[0];
235 		nr = 0;
236 		path->slots[0]++;
237 		while (path->slots[0] > 0) {
238 			btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0] - 1);
239 
240 			if (found_key.type == BTRFS_FREE_SPACE_INFO_KEY) {
241 				ASSERT(found_key.objectid == block_group->start);
242 				ASSERT(found_key.offset == block_group->length);
243 				done = 1;
244 				break;
245 			} else if (found_key.type == BTRFS_FREE_SPACE_EXTENT_KEY) {
246 				u64 first, last;
247 
248 				ASSERT(found_key.objectid >= start);
249 				ASSERT(found_key.objectid < end);
250 				ASSERT(found_key.objectid + found_key.offset <= end);
251 
252 				first = div_u64(found_key.objectid - start,
253 						fs_info->sectorsize);
254 				last = div_u64(found_key.objectid + found_key.offset - start,
255 					       fs_info->sectorsize);
256 				le_bitmap_set(bitmap, first, last - first);
257 
258 				extent_count++;
259 				nr++;
260 				path->slots[0]--;
261 			} else {
262 				ASSERT(0);
263 			}
264 		}
265 
266 		ret = btrfs_del_items(trans, root, path, path->slots[0], nr);
267 		if (unlikely(ret)) {
268 			btrfs_abort_transaction(trans, ret);
269 			goto out;
270 		}
271 		btrfs_release_path(path);
272 	}
273 
274 	info = btrfs_search_free_space_info(trans, block_group, path, 1);
275 	if (IS_ERR(info)) {
276 		ret = PTR_ERR(info);
277 		btrfs_abort_transaction(trans, ret);
278 		goto out;
279 	}
280 	leaf = path->nodes[0];
281 	flags = btrfs_free_space_flags(leaf, info);
282 	flags |= BTRFS_FREE_SPACE_USING_BITMAPS;
283 	block_group->using_free_space_bitmaps = true;
284 	block_group->using_free_space_bitmaps_cached = true;
285 	btrfs_set_free_space_flags(leaf, info, flags);
286 	expected_extent_count = btrfs_free_space_extent_count(leaf, info);
287 	btrfs_release_path(path);
288 
289 	if (unlikely(extent_count != expected_extent_count)) {
290 		btrfs_err(fs_info,
291 			  "incorrect extent count for %llu; counted %u, expected %u",
292 			  block_group->start, extent_count,
293 			  expected_extent_count);
294 		ret = -EIO;
295 		btrfs_abort_transaction(trans, ret);
296 		goto out;
297 	}
298 
299 	bitmap_cursor = (char *)bitmap;
300 	bitmap_range = fs_info->sectorsize * BTRFS_FREE_SPACE_BITMAP_BITS;
301 	i = start;
302 	while (i < end) {
303 		unsigned long ptr;
304 		u64 extent_size;
305 		u32 data_size;
306 
307 		extent_size = min(end - i, bitmap_range);
308 		data_size = free_space_bitmap_size(fs_info, extent_size);
309 
310 		key.objectid = i;
311 		key.type = BTRFS_FREE_SPACE_BITMAP_KEY;
312 		key.offset = extent_size;
313 
314 		ret = btrfs_insert_empty_item(trans, root, path, &key,
315 					      data_size);
316 		if (unlikely(ret)) {
317 			btrfs_abort_transaction(trans, ret);
318 			goto out;
319 		}
320 
321 		leaf = path->nodes[0];
322 		ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
323 		write_extent_buffer(leaf, bitmap_cursor, ptr,
324 				    data_size);
325 		btrfs_release_path(path);
326 
327 		i += extent_size;
328 		bitmap_cursor += data_size;
329 	}
330 
331 	ret = 0;
332 out:
333 	kvfree(bitmap);
334 	return ret;
335 }
336 
337 EXPORT_FOR_TESTS
338 int btrfs_convert_free_space_to_extents(struct btrfs_trans_handle *trans,
339 					struct btrfs_block_group *block_group,
340 					struct btrfs_path *path)
341 {
342 	struct btrfs_fs_info *fs_info = trans->fs_info;
343 	struct btrfs_root *root = btrfs_free_space_root(block_group);
344 	struct btrfs_free_space_info *info;
345 	struct btrfs_key key, found_key;
346 	struct extent_buffer *leaf;
347 	unsigned long *bitmap;
348 	u64 start, end;
349 	u32 bitmap_size, flags, expected_extent_count;
350 	unsigned long nrbits, start_bit, end_bit;
351 	u32 extent_count = 0;
352 	int done = 0, nr;
353 	int ret;
354 
355 	bitmap_size = free_space_bitmap_size(fs_info, block_group->length);
356 	bitmap = alloc_bitmap(bitmap_size);
357 	if (unlikely(!bitmap))
358 		return 0;
359 
360 	start = block_group->start;
361 	end = btrfs_block_group_end(block_group);
362 
363 	key.objectid = end - 1;
364 	key.type = (u8)-1;
365 	key.offset = (u64)-1;
366 
367 	while (!done) {
368 		ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
369 		if (unlikely(ret)) {
370 			btrfs_abort_transaction(trans, ret);
371 			goto out;
372 		}
373 
374 		leaf = path->nodes[0];
375 		nr = 0;
376 		path->slots[0]++;
377 		while (path->slots[0] > 0) {
378 			btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0] - 1);
379 
380 			if (found_key.type == BTRFS_FREE_SPACE_INFO_KEY) {
381 				ASSERT(found_key.objectid == block_group->start);
382 				ASSERT(found_key.offset == block_group->length);
383 				done = 1;
384 				break;
385 			} else if (found_key.type == BTRFS_FREE_SPACE_BITMAP_KEY) {
386 				unsigned long ptr;
387 				char *bitmap_cursor;
388 				u32 bitmap_pos, data_size;
389 
390 				ASSERT(found_key.objectid >= start);
391 				ASSERT(found_key.objectid < end);
392 				ASSERT(found_key.objectid + found_key.offset <= end);
393 
394 				bitmap_pos = div_u64(found_key.objectid - start,
395 						     fs_info->sectorsize *
396 						     BITS_PER_BYTE);
397 				bitmap_cursor = ((char *)bitmap) + bitmap_pos;
398 				data_size = free_space_bitmap_size(fs_info,
399 								found_key.offset);
400 
401 				path->slots[0]--;
402 				ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
403 				read_extent_buffer(leaf, bitmap_cursor, ptr,
404 						   data_size);
405 
406 				nr++;
407 			} else {
408 				ASSERT(0);
409 			}
410 		}
411 
412 		ret = btrfs_del_items(trans, root, path, path->slots[0], nr);
413 		if (unlikely(ret)) {
414 			btrfs_abort_transaction(trans, ret);
415 			goto out;
416 		}
417 		btrfs_release_path(path);
418 	}
419 
420 	info = btrfs_search_free_space_info(trans, block_group, path, 1);
421 	if (IS_ERR(info)) {
422 		ret = PTR_ERR(info);
423 		btrfs_abort_transaction(trans, ret);
424 		goto out;
425 	}
426 	leaf = path->nodes[0];
427 	flags = btrfs_free_space_flags(leaf, info);
428 	flags &= ~BTRFS_FREE_SPACE_USING_BITMAPS;
429 	block_group->using_free_space_bitmaps = false;
430 	block_group->using_free_space_bitmaps_cached = true;
431 	btrfs_set_free_space_flags(leaf, info, flags);
432 	expected_extent_count = btrfs_free_space_extent_count(leaf, info);
433 	btrfs_release_path(path);
434 
435 	nrbits = block_group->length >> fs_info->sectorsize_bits;
436 	start_bit = find_next_bit_le(bitmap, nrbits, 0);
437 
438 	while (start_bit < nrbits) {
439 		end_bit = find_next_zero_bit_le(bitmap, nrbits, start_bit);
440 		ASSERT(start_bit < end_bit);
441 
442 		key.objectid = start + start_bit * fs_info->sectorsize;
443 		key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
444 		key.offset = (end_bit - start_bit) * fs_info->sectorsize;
445 
446 		ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
447 		if (unlikely(ret)) {
448 			btrfs_abort_transaction(trans, ret);
449 			goto out;
450 		}
451 		btrfs_release_path(path);
452 
453 		extent_count++;
454 
455 		start_bit = find_next_bit_le(bitmap, nrbits, end_bit);
456 	}
457 
458 	if (unlikely(extent_count != expected_extent_count)) {
459 		btrfs_err(fs_info,
460 			  "incorrect extent count for %llu; counted %u, expected %u",
461 			  block_group->start, extent_count,
462 			  expected_extent_count);
463 		ret = -EIO;
464 		btrfs_abort_transaction(trans, ret);
465 		goto out;
466 	}
467 
468 	ret = 0;
469 out:
470 	kvfree(bitmap);
471 	return ret;
472 }
473 
474 static int update_free_space_extent_count(struct btrfs_trans_handle *trans,
475 					  struct btrfs_block_group *block_group,
476 					  struct btrfs_path *path,
477 					  int new_extents)
478 {
479 	struct btrfs_free_space_info *info;
480 	u32 flags;
481 	u32 extent_count;
482 	int ret = 0;
483 
484 	if (new_extents == 0)
485 		return 0;
486 
487 	info = btrfs_search_free_space_info(trans, block_group, path, 1);
488 	if (IS_ERR(info))
489 		return PTR_ERR(info);
490 
491 	flags = btrfs_free_space_flags(path->nodes[0], info);
492 	extent_count = btrfs_free_space_extent_count(path->nodes[0], info);
493 
494 	extent_count += new_extents;
495 	btrfs_set_free_space_extent_count(path->nodes[0], info, extent_count);
496 	btrfs_release_path(path);
497 
498 	if (!(flags & BTRFS_FREE_SPACE_USING_BITMAPS) &&
499 	    extent_count > block_group->bitmap_high_thresh) {
500 		ret = btrfs_convert_free_space_to_bitmaps(trans, block_group, path);
501 	} else if ((flags & BTRFS_FREE_SPACE_USING_BITMAPS) &&
502 		   extent_count < block_group->bitmap_low_thresh) {
503 		ret = btrfs_convert_free_space_to_extents(trans, block_group, path);
504 	}
505 
506 	return ret;
507 }
508 
509 EXPORT_FOR_TESTS
510 bool btrfs_free_space_test_bit(struct btrfs_block_group *block_group,
511 			       struct btrfs_path *path, u64 offset)
512 {
513 	struct extent_buffer *leaf;
514 	struct btrfs_key key;
515 	u64 found_start, found_end;
516 	unsigned long ptr, i;
517 
518 	leaf = path->nodes[0];
519 	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
520 	ASSERT(key.type == BTRFS_FREE_SPACE_BITMAP_KEY);
521 
522 	found_start = key.objectid;
523 	found_end = key.objectid + key.offset;
524 	ASSERT(offset >= found_start && offset < found_end);
525 
526 	ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
527 	i = div_u64(offset - found_start,
528 		    block_group->fs_info->sectorsize);
529 	return extent_buffer_test_bit(leaf, ptr, i);
530 }
531 
532 static void free_space_modify_bits(struct btrfs_trans_handle *trans,
533 				   struct btrfs_block_group *block_group,
534 				   struct btrfs_path *path, u64 *start, u64 *size,
535 				   bool set_bits)
536 {
537 	struct btrfs_fs_info *fs_info = block_group->fs_info;
538 	struct extent_buffer *leaf;
539 	struct btrfs_key key;
540 	u64 end = *start + *size;
541 	u64 found_start, found_end;
542 	unsigned long ptr, first, last;
543 
544 	leaf = path->nodes[0];
545 	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
546 	ASSERT(key.type == BTRFS_FREE_SPACE_BITMAP_KEY);
547 
548 	found_start = key.objectid;
549 	found_end = key.objectid + key.offset;
550 	ASSERT(*start >= found_start && *start < found_end);
551 	ASSERT(end > found_start);
552 
553 	if (end > found_end)
554 		end = found_end;
555 
556 	ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
557 	first = (*start - found_start) >> fs_info->sectorsize_bits;
558 	last = (end - found_start) >> fs_info->sectorsize_bits;
559 	if (set_bits)
560 		extent_buffer_bitmap_set(leaf, ptr, first, last - first);
561 	else
562 		extent_buffer_bitmap_clear(leaf, ptr, first, last - first);
563 	btrfs_mark_buffer_dirty(trans, leaf);
564 
565 	*size -= end - *start;
566 	*start = end;
567 }
568 
569 /*
570  * We can't use btrfs_next_item() in modify_free_space_bitmap() because
571  * btrfs_next_leaf() doesn't get the path for writing. We can forgo the fancy
572  * tree walking in btrfs_next_leaf() anyways because we know exactly what we're
573  * looking for.
574  */
575 static int free_space_next_bitmap(struct btrfs_trans_handle *trans,
576 				  struct btrfs_root *root, struct btrfs_path *p)
577 {
578 	struct btrfs_key key;
579 
580 	if (p->slots[0] + 1 < btrfs_header_nritems(p->nodes[0])) {
581 		p->slots[0]++;
582 		return 0;
583 	}
584 
585 	btrfs_item_key_to_cpu(p->nodes[0], &key, p->slots[0]);
586 	btrfs_release_path(p);
587 
588 	key.objectid += key.offset;
589 	key.type = (u8)-1;
590 	key.offset = (u64)-1;
591 
592 	return btrfs_search_prev_slot(trans, root, &key, p, 0, 1);
593 }
594 
595 /*
596  * If remove is 1, then we are removing free space, thus clearing bits in the
597  * bitmap. If remove is 0, then we are adding free space, thus setting bits in
598  * the bitmap.
599  */
600 static int modify_free_space_bitmap(struct btrfs_trans_handle *trans,
601 				    struct btrfs_block_group *block_group,
602 				    struct btrfs_path *path,
603 				    u64 start, u64 size, bool remove)
604 {
605 	struct btrfs_root *root = btrfs_free_space_root(block_group);
606 	struct btrfs_key key;
607 	u64 end = start + size;
608 	u64 cur_start, cur_size;
609 	bool prev_bit_set = false;
610 	bool next_bit_set = false;
611 	int new_extents;
612 	int ret;
613 
614 	/*
615 	 * Read the bit for the block immediately before the extent of space if
616 	 * that block is within the block group.
617 	 */
618 	if (start > block_group->start) {
619 		u64 prev_block = start - block_group->fs_info->sectorsize;
620 
621 		key.objectid = prev_block;
622 		key.type = (u8)-1;
623 		key.offset = (u64)-1;
624 
625 		ret = btrfs_search_prev_slot(trans, root, &key, path, 0, 1);
626 		if (ret)
627 			return ret;
628 
629 		prev_bit_set = btrfs_free_space_test_bit(block_group, path, prev_block);
630 
631 		/* The previous block may have been in the previous bitmap. */
632 		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
633 		if (start >= key.objectid + key.offset) {
634 			ret = free_space_next_bitmap(trans, root, path);
635 			if (ret)
636 				return ret;
637 		}
638 	} else {
639 		key.objectid = start;
640 		key.type = (u8)-1;
641 		key.offset = (u64)-1;
642 
643 		ret = btrfs_search_prev_slot(trans, root, &key, path, 0, 1);
644 		if (ret)
645 			return ret;
646 	}
647 
648 	/*
649 	 * Iterate over all of the bitmaps overlapped by the extent of space,
650 	 * clearing/setting bits as required.
651 	 */
652 	cur_start = start;
653 	cur_size = size;
654 	while (1) {
655 		free_space_modify_bits(trans, block_group, path, &cur_start,
656 				       &cur_size, !remove);
657 		if (cur_size == 0)
658 			break;
659 		ret = free_space_next_bitmap(trans, root, path);
660 		if (ret)
661 			return ret;
662 	}
663 
664 	/*
665 	 * Read the bit for the block immediately after the extent of space if
666 	 * that block is within the block group.
667 	 */
668 	if (end < btrfs_block_group_end(block_group)) {
669 		/* The next block may be in the next bitmap. */
670 		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
671 		if (end >= key.objectid + key.offset) {
672 			ret = free_space_next_bitmap(trans, root, path);
673 			if (ret)
674 				return ret;
675 		}
676 
677 		next_bit_set = btrfs_free_space_test_bit(block_group, path, end);
678 	}
679 
680 	if (remove) {
681 		new_extents = -1;
682 		if (prev_bit_set) {
683 			/* Leftover on the left. */
684 			new_extents++;
685 		}
686 		if (next_bit_set) {
687 			/* Leftover on the right. */
688 			new_extents++;
689 		}
690 	} else {
691 		new_extents = 1;
692 		if (prev_bit_set) {
693 			/* Merging with neighbor on the left. */
694 			new_extents--;
695 		}
696 		if (next_bit_set) {
697 			/* Merging with neighbor on the right. */
698 			new_extents--;
699 		}
700 	}
701 
702 	btrfs_release_path(path);
703 	return update_free_space_extent_count(trans, block_group, path, new_extents);
704 }
705 
706 static int remove_free_space_extent(struct btrfs_trans_handle *trans,
707 				    struct btrfs_block_group *block_group,
708 				    struct btrfs_path *path,
709 				    u64 start, u64 size)
710 {
711 	struct btrfs_root *root = btrfs_free_space_root(block_group);
712 	struct btrfs_key key;
713 	u64 found_start, found_end;
714 	u64 end = start + size;
715 	int new_extents = -1;
716 	int ret;
717 
718 	key.objectid = start;
719 	key.type = (u8)-1;
720 	key.offset = (u64)-1;
721 
722 	ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
723 	if (ret)
724 		return ret;
725 
726 	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
727 
728 	ASSERT(key.type == BTRFS_FREE_SPACE_EXTENT_KEY);
729 
730 	found_start = key.objectid;
731 	found_end = key.objectid + key.offset;
732 	ASSERT(start >= found_start && end <= found_end);
733 
734 	/*
735 	 * Okay, now that we've found the free space extent which contains the
736 	 * free space that we are removing, there are four cases:
737 	 *
738 	 * 1. We're using the whole extent: delete the key we found and
739 	 * decrement the free space extent count.
740 	 * 2. We are using part of the extent starting at the beginning: delete
741 	 * the key we found and insert a new key representing the leftover at
742 	 * the end. There is no net change in the number of extents.
743 	 * 3. We are using part of the extent ending at the end: delete the key
744 	 * we found and insert a new key representing the leftover at the
745 	 * beginning. There is no net change in the number of extents.
746 	 * 4. We are using part of the extent in the middle: delete the key we
747 	 * found and insert two new keys representing the leftovers on each
748 	 * side. Where we used to have one extent, we now have two, so increment
749 	 * the extent count. We may need to convert the block group to bitmaps
750 	 * as a result.
751 	 */
752 
753 	/* Delete the existing key (cases 1-4). */
754 	ret = btrfs_del_item(trans, root, path);
755 	if (ret)
756 		return ret;
757 
758 	/* Add a key for leftovers at the beginning (cases 3 and 4). */
759 	if (start > found_start) {
760 		key.objectid = found_start;
761 		key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
762 		key.offset = start - found_start;
763 
764 		btrfs_release_path(path);
765 		ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
766 		if (ret)
767 			return ret;
768 		new_extents++;
769 	}
770 
771 	/* Add a key for leftovers at the end (cases 2 and 4). */
772 	if (end < found_end) {
773 		key.objectid = end;
774 		key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
775 		key.offset = found_end - end;
776 
777 		btrfs_release_path(path);
778 		ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
779 		if (ret)
780 			return ret;
781 		new_extents++;
782 	}
783 
784 	btrfs_release_path(path);
785 	return update_free_space_extent_count(trans, block_group, path, new_extents);
786 }
787 
788 static int using_bitmaps(struct btrfs_block_group *bg, struct btrfs_path *path)
789 {
790 	struct btrfs_free_space_info *info;
791 	u32 flags;
792 
793 	if (bg->using_free_space_bitmaps_cached)
794 		return bg->using_free_space_bitmaps;
795 
796 	info = btrfs_search_free_space_info(NULL, bg, path, 0);
797 	if (IS_ERR(info))
798 		return PTR_ERR(info);
799 	flags = btrfs_free_space_flags(path->nodes[0], info);
800 	btrfs_release_path(path);
801 
802 	bg->using_free_space_bitmaps = (flags & BTRFS_FREE_SPACE_USING_BITMAPS);
803 	bg->using_free_space_bitmaps_cached = true;
804 
805 	return bg->using_free_space_bitmaps;
806 }
807 
808 EXPORT_FOR_TESTS
809 int __btrfs_remove_from_free_space_tree(struct btrfs_trans_handle *trans,
810 					struct btrfs_block_group *block_group,
811 					struct btrfs_path *path, u64 start, u64 size)
812 {
813 	int ret;
814 
815 	ret = __add_block_group_free_space(trans, block_group, path);
816 	if (ret)
817 		return ret;
818 
819 	ret = using_bitmaps(block_group, path);
820 	if (ret < 0)
821 		return ret;
822 
823 	if (ret)
824 		return modify_free_space_bitmap(trans, block_group, path,
825 						start, size, true);
826 
827 	return remove_free_space_extent(trans, block_group, path, start, size);
828 }
829 
830 int btrfs_remove_from_free_space_tree(struct btrfs_trans_handle *trans,
831 				      u64 start, u64 size)
832 {
833 	struct btrfs_block_group *block_group;
834 	BTRFS_PATH_AUTO_FREE(path);
835 	int ret;
836 
837 	if (!btrfs_fs_compat_ro(trans->fs_info, FREE_SPACE_TREE))
838 		return 0;
839 
840 	path = btrfs_alloc_path();
841 	if (unlikely(!path)) {
842 		ret = -ENOMEM;
843 		btrfs_abort_transaction(trans, ret);
844 		return ret;
845 	}
846 
847 	block_group = btrfs_lookup_block_group(trans->fs_info, start);
848 	if (unlikely(!block_group)) {
849 		DEBUG_WARN("no block group found for start=%llu", start);
850 		ret = -ENOENT;
851 		btrfs_abort_transaction(trans, ret);
852 		return ret;
853 	}
854 
855 	mutex_lock(&block_group->free_space_lock);
856 	ret = __btrfs_remove_from_free_space_tree(trans, block_group, path, start, size);
857 	mutex_unlock(&block_group->free_space_lock);
858 	if (ret)
859 		btrfs_abort_transaction(trans, ret);
860 
861 	btrfs_put_block_group(block_group);
862 
863 	return ret;
864 }
865 
866 static int add_free_space_extent(struct btrfs_trans_handle *trans,
867 				 struct btrfs_block_group *block_group,
868 				 struct btrfs_path *path,
869 				 u64 start, u64 size)
870 {
871 	struct btrfs_root *root = btrfs_free_space_root(block_group);
872 	struct btrfs_key key, new_key;
873 	u64 found_start, found_end;
874 	u64 end = start + size;
875 	int new_extents = 1;
876 	int ret;
877 
878 	/*
879 	 * We are adding a new extent of free space, but we need to merge
880 	 * extents. There are four cases here:
881 	 *
882 	 * 1. The new extent does not have any immediate neighbors to merge
883 	 * with: add the new key and increment the free space extent count. We
884 	 * may need to convert the block group to bitmaps as a result.
885 	 * 2. The new extent has an immediate neighbor before it: remove the
886 	 * previous key and insert a new key combining both of them. There is no
887 	 * net change in the number of extents.
888 	 * 3. The new extent has an immediate neighbor after it: remove the next
889 	 * key and insert a new key combining both of them. There is no net
890 	 * change in the number of extents.
891 	 * 4. The new extent has immediate neighbors on both sides: remove both
892 	 * of the keys and insert a new key combining all of them. Where we used
893 	 * to have two extents, we now have one, so decrement the extent count.
894 	 */
895 
896 	new_key.objectid = start;
897 	new_key.type = BTRFS_FREE_SPACE_EXTENT_KEY;
898 	new_key.offset = size;
899 
900 	/* Search for a neighbor on the left. */
901 	if (start == block_group->start)
902 		goto right;
903 	key.objectid = start - 1;
904 	key.type = (u8)-1;
905 	key.offset = (u64)-1;
906 
907 	ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
908 	if (ret)
909 		return ret;
910 
911 	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
912 
913 	if (key.type != BTRFS_FREE_SPACE_EXTENT_KEY) {
914 		ASSERT(key.type == BTRFS_FREE_SPACE_INFO_KEY);
915 		btrfs_release_path(path);
916 		goto right;
917 	}
918 
919 	found_start = key.objectid;
920 	found_end = key.objectid + key.offset;
921 	ASSERT(found_start >= block_group->start &&
922 	       found_end > block_group->start);
923 	ASSERT(found_start < start && found_end <= start);
924 
925 	/*
926 	 * Delete the neighbor on the left and absorb it into the new key (cases
927 	 * 2 and 4).
928 	 */
929 	if (found_end == start) {
930 		ret = btrfs_del_item(trans, root, path);
931 		if (ret)
932 			return ret;
933 		new_key.objectid = found_start;
934 		new_key.offset += key.offset;
935 		new_extents--;
936 	}
937 	btrfs_release_path(path);
938 
939 right:
940 	/* Search for a neighbor on the right. */
941 	if (end == btrfs_block_group_end(block_group))
942 		goto insert;
943 	key.objectid = end;
944 	key.type = (u8)-1;
945 	key.offset = (u64)-1;
946 
947 	ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
948 	if (ret)
949 		return ret;
950 
951 	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
952 
953 	if (key.type != BTRFS_FREE_SPACE_EXTENT_KEY) {
954 		ASSERT(key.type == BTRFS_FREE_SPACE_INFO_KEY);
955 		btrfs_release_path(path);
956 		goto insert;
957 	}
958 
959 	found_start = key.objectid;
960 	found_end = key.objectid + key.offset;
961 	ASSERT(found_start >= block_group->start &&
962 	       found_end > block_group->start);
963 	ASSERT((found_start < start && found_end <= start) ||
964 	       (found_start >= end && found_end > end));
965 
966 	/*
967 	 * Delete the neighbor on the right and absorb it into the new key
968 	 * (cases 3 and 4).
969 	 */
970 	if (found_start == end) {
971 		ret = btrfs_del_item(trans, root, path);
972 		if (ret)
973 			return ret;
974 		new_key.offset += key.offset;
975 		new_extents--;
976 	}
977 	btrfs_release_path(path);
978 
979 insert:
980 	/* Insert the new key (cases 1-4). */
981 	ret = btrfs_insert_empty_item(trans, root, path, &new_key, 0);
982 	if (ret)
983 		return ret;
984 
985 	btrfs_release_path(path);
986 	return update_free_space_extent_count(trans, block_group, path, new_extents);
987 }
988 
989 EXPORT_FOR_TESTS
990 int __btrfs_add_to_free_space_tree(struct btrfs_trans_handle *trans,
991 				   struct btrfs_block_group *block_group,
992 				   struct btrfs_path *path, u64 start, u64 size)
993 {
994 	int ret;
995 
996 	ret = __add_block_group_free_space(trans, block_group, path);
997 	if (ret)
998 		return ret;
999 
1000 	ret = using_bitmaps(block_group, path);
1001 	if (ret < 0)
1002 		return ret;
1003 
1004 	if (ret)
1005 		return modify_free_space_bitmap(trans, block_group, path,
1006 						start, size, false);
1007 
1008 	return add_free_space_extent(trans, block_group, path, start, size);
1009 }
1010 
1011 int btrfs_add_to_free_space_tree(struct btrfs_trans_handle *trans,
1012 				 u64 start, u64 size)
1013 {
1014 	struct btrfs_block_group *block_group;
1015 	BTRFS_PATH_AUTO_FREE(path);
1016 	int ret;
1017 
1018 	if (!btrfs_fs_compat_ro(trans->fs_info, FREE_SPACE_TREE))
1019 		return 0;
1020 
1021 	path = btrfs_alloc_path();
1022 	if (unlikely(!path)) {
1023 		ret = -ENOMEM;
1024 		btrfs_abort_transaction(trans, ret);
1025 		return ret;
1026 	}
1027 
1028 	block_group = btrfs_lookup_block_group(trans->fs_info, start);
1029 	if (unlikely(!block_group)) {
1030 		DEBUG_WARN("no block group found for start=%llu", start);
1031 		ret = -ENOENT;
1032 		btrfs_abort_transaction(trans, ret);
1033 		return ret;
1034 	}
1035 
1036 	mutex_lock(&block_group->free_space_lock);
1037 	ret = __btrfs_add_to_free_space_tree(trans, block_group, path, start, size);
1038 	mutex_unlock(&block_group->free_space_lock);
1039 	if (ret)
1040 		btrfs_abort_transaction(trans, ret);
1041 
1042 	btrfs_put_block_group(block_group);
1043 
1044 	return ret;
1045 }
1046 
1047 /*
1048  * Populate the free space tree by walking the extent tree. Operations on the
1049  * extent tree that happen as a result of writes to the free space tree will go
1050  * through the normal add/remove hooks.
1051  */
1052 static int populate_free_space_tree(struct btrfs_trans_handle *trans,
1053 				    struct btrfs_block_group *block_group)
1054 {
1055 	struct btrfs_root *extent_root;
1056 	BTRFS_PATH_AUTO_FREE(path);
1057 	BTRFS_PATH_AUTO_FREE(path2);
1058 	struct btrfs_key key;
1059 	u64 start, end;
1060 	int ret;
1061 
1062 	path = btrfs_alloc_path();
1063 	if (!path)
1064 		return -ENOMEM;
1065 
1066 	path2 = btrfs_alloc_path();
1067 	if (!path2)
1068 		return -ENOMEM;
1069 
1070 	path->reada = READA_FORWARD;
1071 
1072 	ret = add_new_free_space_info(trans, block_group, path2);
1073 	if (ret)
1074 		return ret;
1075 
1076 	mutex_lock(&block_group->free_space_lock);
1077 
1078 	/*
1079 	 * Iterate through all of the extent and metadata items in this block
1080 	 * group, adding the free space between them and the free space at the
1081 	 * end. Note that EXTENT_ITEM and METADATA_ITEM are less than
1082 	 * BLOCK_GROUP_ITEM, so an extent may precede the block group that it's
1083 	 * contained in.
1084 	 */
1085 	key.objectid = block_group->start;
1086 	key.type = BTRFS_EXTENT_ITEM_KEY;
1087 	key.offset = 0;
1088 
1089 	extent_root = btrfs_extent_root(trans->fs_info, key.objectid);
1090 	ret = btrfs_search_slot_for_read(extent_root, &key, path, 1, 0);
1091 	if (ret < 0)
1092 		goto out_locked;
1093 	/*
1094 	 * If ret is 1 (no key found), it means this is an empty block group,
1095 	 * without any extents allocated from it and there's no block group
1096 	 * item (key BTRFS_BLOCK_GROUP_ITEM_KEY) located in the extent tree
1097 	 * because we are using the block group tree feature (so block group
1098 	 * items are stored in the block group tree) or this is a new block
1099 	 * group created in the current transaction and its block group item
1100 	 * was not yet inserted in the extent tree (that happens in
1101 	 * btrfs_create_pending_block_groups() -> insert_block_group_item()).
1102 	 * It also means there are no extents allocated for block groups with a
1103 	 * start offset beyond this block group's end offset (this is the last,
1104 	 * highest, block group).
1105 	 */
1106 	start = block_group->start;
1107 	end = btrfs_block_group_end(block_group);
1108 	while (ret == 0) {
1109 		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1110 
1111 		if (key.type == BTRFS_EXTENT_ITEM_KEY ||
1112 		    key.type == BTRFS_METADATA_ITEM_KEY) {
1113 			if (key.objectid >= end)
1114 				break;
1115 
1116 			if (start < key.objectid) {
1117 				ret = __btrfs_add_to_free_space_tree(trans,
1118 								     block_group,
1119 								     path2, start,
1120 								     key.objectid -
1121 								     start);
1122 				if (ret)
1123 					goto out_locked;
1124 			}
1125 			start = key.objectid;
1126 			if (key.type == BTRFS_METADATA_ITEM_KEY)
1127 				start += trans->fs_info->nodesize;
1128 			else
1129 				start += key.offset;
1130 		} else if (key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
1131 			if (key.objectid != block_group->start)
1132 				break;
1133 		}
1134 
1135 		ret = btrfs_next_item(extent_root, path);
1136 		if (ret < 0)
1137 			goto out_locked;
1138 	}
1139 	if (start < end) {
1140 		ret = __btrfs_add_to_free_space_tree(trans, block_group, path2,
1141 						     start, end - start);
1142 		if (ret)
1143 			goto out_locked;
1144 	}
1145 
1146 	ret = 0;
1147 out_locked:
1148 	mutex_unlock(&block_group->free_space_lock);
1149 
1150 	return ret;
1151 }
1152 
1153 int btrfs_create_free_space_tree(struct btrfs_fs_info *fs_info)
1154 {
1155 	struct btrfs_trans_handle *trans;
1156 	struct btrfs_root *tree_root = fs_info->tree_root;
1157 	struct btrfs_root *free_space_root;
1158 	struct btrfs_block_group *block_group;
1159 	struct rb_node *node;
1160 	int ret;
1161 
1162 	trans = btrfs_start_transaction(tree_root, 0);
1163 	if (IS_ERR(trans))
1164 		return PTR_ERR(trans);
1165 
1166 	set_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
1167 	set_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags);
1168 	free_space_root = btrfs_create_tree(trans,
1169 					    BTRFS_FREE_SPACE_TREE_OBJECTID);
1170 	if (IS_ERR(free_space_root)) {
1171 		ret = PTR_ERR(free_space_root);
1172 		btrfs_abort_transaction(trans, ret);
1173 		btrfs_end_transaction(trans);
1174 		goto out_clear;
1175 	}
1176 	ret = btrfs_global_root_insert(free_space_root);
1177 	if (unlikely(ret)) {
1178 		btrfs_put_root(free_space_root);
1179 		btrfs_abort_transaction(trans, ret);
1180 		btrfs_end_transaction(trans);
1181 		goto out_clear;
1182 	}
1183 
1184 	node = rb_first_cached(&fs_info->block_group_cache_tree);
1185 	while (node) {
1186 		block_group = rb_entry(node, struct btrfs_block_group,
1187 				       cache_node);
1188 		ret = populate_free_space_tree(trans, block_group);
1189 		if (unlikely(ret)) {
1190 			btrfs_abort_transaction(trans, ret);
1191 			btrfs_end_transaction(trans);
1192 			goto out_clear;
1193 		}
1194 		node = rb_next(node);
1195 	}
1196 
1197 	btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE);
1198 	btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE_VALID);
1199 	clear_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
1200 	ret = btrfs_commit_transaction(trans);
1201 
1202 	/*
1203 	 * Now that we've committed the transaction any reading of our commit
1204 	 * root will be safe, so we can cache from the free space tree now.
1205 	 */
1206 	clear_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags);
1207 	return ret;
1208 
1209 out_clear:
1210 	clear_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
1211 	clear_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags);
1212 	return ret;
1213 }
1214 
1215 static int clear_free_space_tree(struct btrfs_trans_handle *trans,
1216 				 struct btrfs_root *root)
1217 {
1218 	BTRFS_PATH_AUTO_FREE(path);
1219 	struct btrfs_key key;
1220 	struct rb_node *node;
1221 	int nr;
1222 	int ret;
1223 
1224 	path = btrfs_alloc_path();
1225 	if (!path)
1226 		return -ENOMEM;
1227 
1228 	key.objectid = 0;
1229 	key.type = 0;
1230 	key.offset = 0;
1231 
1232 	while (1) {
1233 		ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1234 		if (ret < 0)
1235 			return ret;
1236 
1237 		nr = btrfs_header_nritems(path->nodes[0]);
1238 		if (!nr)
1239 			break;
1240 
1241 		path->slots[0] = 0;
1242 		ret = btrfs_del_items(trans, root, path, 0, nr);
1243 		if (ret)
1244 			return ret;
1245 
1246 		btrfs_release_path(path);
1247 	}
1248 
1249 	node = rb_first_cached(&trans->fs_info->block_group_cache_tree);
1250 	while (node) {
1251 		struct btrfs_block_group *bg;
1252 
1253 		bg = rb_entry(node, struct btrfs_block_group, cache_node);
1254 		clear_bit(BLOCK_GROUP_FLAG_FREE_SPACE_ADDED, &bg->runtime_flags);
1255 		node = rb_next(node);
1256 		cond_resched();
1257 	}
1258 
1259 	return 0;
1260 }
1261 
1262 int btrfs_delete_free_space_tree(struct btrfs_fs_info *fs_info)
1263 {
1264 	struct btrfs_trans_handle *trans;
1265 	struct btrfs_root *tree_root = fs_info->tree_root;
1266 	struct btrfs_key key = {
1267 		.objectid = BTRFS_FREE_SPACE_TREE_OBJECTID,
1268 		.type = BTRFS_ROOT_ITEM_KEY,
1269 		.offset = 0,
1270 	};
1271 	struct btrfs_root *free_space_root = btrfs_global_root(fs_info, &key);
1272 	int ret;
1273 
1274 	trans = btrfs_start_transaction(tree_root, 0);
1275 	if (IS_ERR(trans))
1276 		return PTR_ERR(trans);
1277 
1278 	btrfs_clear_fs_compat_ro(fs_info, FREE_SPACE_TREE);
1279 	btrfs_clear_fs_compat_ro(fs_info, FREE_SPACE_TREE_VALID);
1280 
1281 	ret = clear_free_space_tree(trans, free_space_root);
1282 	if (unlikely(ret)) {
1283 		btrfs_abort_transaction(trans, ret);
1284 		btrfs_end_transaction(trans);
1285 		return ret;
1286 	}
1287 
1288 	ret = btrfs_del_root(trans, &free_space_root->root_key);
1289 	if (unlikely(ret)) {
1290 		btrfs_abort_transaction(trans, ret);
1291 		btrfs_end_transaction(trans);
1292 		return ret;
1293 	}
1294 
1295 	btrfs_global_root_delete(free_space_root);
1296 
1297 	spin_lock(&fs_info->trans_lock);
1298 	list_del(&free_space_root->dirty_list);
1299 	spin_unlock(&fs_info->trans_lock);
1300 
1301 	btrfs_tree_lock(free_space_root->node);
1302 	btrfs_clear_buffer_dirty(trans, free_space_root->node);
1303 	btrfs_tree_unlock(free_space_root->node);
1304 	ret = btrfs_free_tree_block(trans, btrfs_root_id(free_space_root),
1305 				    free_space_root->node, 0, 1);
1306 	btrfs_put_root(free_space_root);
1307 	if (unlikely(ret < 0)) {
1308 		btrfs_abort_transaction(trans, ret);
1309 		btrfs_end_transaction(trans);
1310 		return ret;
1311 	}
1312 
1313 	return btrfs_commit_transaction(trans);
1314 }
1315 
1316 int btrfs_rebuild_free_space_tree(struct btrfs_fs_info *fs_info)
1317 {
1318 	struct btrfs_trans_handle *trans;
1319 	struct btrfs_key key = {
1320 		.objectid = BTRFS_FREE_SPACE_TREE_OBJECTID,
1321 		.type = BTRFS_ROOT_ITEM_KEY,
1322 		.offset = 0,
1323 	};
1324 	struct btrfs_root *free_space_root = btrfs_global_root(fs_info, &key);
1325 	struct rb_node *node;
1326 	int ret;
1327 
1328 	trans = btrfs_start_transaction(free_space_root, 1);
1329 	if (IS_ERR(trans))
1330 		return PTR_ERR(trans);
1331 
1332 	set_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
1333 	set_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags);
1334 
1335 	ret = clear_free_space_tree(trans, free_space_root);
1336 	if (unlikely(ret)) {
1337 		btrfs_abort_transaction(trans, ret);
1338 		btrfs_end_transaction(trans);
1339 		return ret;
1340 	}
1341 
1342 	node = rb_first_cached(&fs_info->block_group_cache_tree);
1343 	while (node) {
1344 		struct btrfs_block_group *block_group;
1345 
1346 		block_group = rb_entry(node, struct btrfs_block_group,
1347 				       cache_node);
1348 
1349 		if (test_bit(BLOCK_GROUP_FLAG_FREE_SPACE_ADDED,
1350 			     &block_group->runtime_flags))
1351 			goto next;
1352 
1353 		ret = populate_free_space_tree(trans, block_group);
1354 		if (unlikely(ret)) {
1355 			btrfs_abort_transaction(trans, ret);
1356 			btrfs_end_transaction(trans);
1357 			return ret;
1358 		}
1359 next:
1360 		if (btrfs_should_end_transaction(trans)) {
1361 			btrfs_end_transaction(trans);
1362 			trans = btrfs_start_transaction(free_space_root, 1);
1363 			if (IS_ERR(trans))
1364 				return PTR_ERR(trans);
1365 		}
1366 		node = rb_next(node);
1367 	}
1368 
1369 	btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE);
1370 	btrfs_set_fs_compat_ro(fs_info, FREE_SPACE_TREE_VALID);
1371 	clear_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &fs_info->flags);
1372 
1373 	ret = btrfs_commit_transaction(trans);
1374 	clear_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags);
1375 	return ret;
1376 }
1377 
1378 static int __add_block_group_free_space(struct btrfs_trans_handle *trans,
1379 					struct btrfs_block_group *block_group,
1380 					struct btrfs_path *path)
1381 {
1382 	bool own_path = false;
1383 	int ret;
1384 
1385 	if (!test_and_clear_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE,
1386 				&block_group->runtime_flags))
1387 		return 0;
1388 
1389 	/*
1390 	 * While rebuilding the free space tree we may allocate new metadata
1391 	 * block groups while modifying the free space tree.
1392 	 *
1393 	 * Because during the rebuild (at btrfs_rebuild_free_space_tree()) we
1394 	 * can use multiple transactions, every time btrfs_end_transaction() is
1395 	 * called at btrfs_rebuild_free_space_tree() we finish the creation of
1396 	 * new block groups by calling btrfs_create_pending_block_groups(), and
1397 	 * that in turn calls us, through btrfs_add_block_group_free_space(),
1398 	 * to add a free space info item and a free space extent item for the
1399 	 * block group.
1400 	 *
1401 	 * Then later btrfs_rebuild_free_space_tree() may find such new block
1402 	 * groups and processes them with populate_free_space_tree(), which can
1403 	 * fail with EEXIST since there are already items for the block group in
1404 	 * the free space tree. Notice that we say "may find" because a new
1405 	 * block group may be added to the block groups rbtree in a node before
1406 	 * or after the block group currently being processed by the rebuild
1407 	 * process. So signal the rebuild process to skip such new block groups
1408 	 * if it finds them.
1409 	 */
1410 	set_bit(BLOCK_GROUP_FLAG_FREE_SPACE_ADDED, &block_group->runtime_flags);
1411 
1412 	if (!path) {
1413 		path = btrfs_alloc_path();
1414 		if (unlikely(!path)) {
1415 			btrfs_abort_transaction(trans, -ENOMEM);
1416 			return -ENOMEM;
1417 		}
1418 		own_path = true;
1419 	}
1420 
1421 	ret = add_new_free_space_info(trans, block_group, path);
1422 	if (unlikely(ret)) {
1423 		btrfs_abort_transaction(trans, ret);
1424 		goto out;
1425 	}
1426 
1427 	ret = __btrfs_add_to_free_space_tree(trans, block_group, path,
1428 					     block_group->start, block_group->length);
1429 	if (ret)
1430 		btrfs_abort_transaction(trans, ret);
1431 
1432 out:
1433 	if (own_path)
1434 		btrfs_free_path(path);
1435 
1436 	return ret;
1437 }
1438 
1439 int btrfs_add_block_group_free_space(struct btrfs_trans_handle *trans,
1440 				     struct btrfs_block_group *block_group)
1441 {
1442 	int ret;
1443 
1444 	if (!btrfs_fs_compat_ro(trans->fs_info, FREE_SPACE_TREE))
1445 		return 0;
1446 
1447 	mutex_lock(&block_group->free_space_lock);
1448 	ret = __add_block_group_free_space(trans, block_group, NULL);
1449 	mutex_unlock(&block_group->free_space_lock);
1450 	return ret;
1451 }
1452 
1453 int btrfs_remove_block_group_free_space(struct btrfs_trans_handle *trans,
1454 					struct btrfs_block_group *block_group)
1455 {
1456 	struct btrfs_root *root = btrfs_free_space_root(block_group);
1457 	BTRFS_PATH_AUTO_FREE(path);
1458 	struct btrfs_key key, found_key;
1459 	struct extent_buffer *leaf;
1460 	u64 start, end;
1461 	int done = 0, nr;
1462 	int ret;
1463 
1464 	if (!btrfs_fs_compat_ro(trans->fs_info, FREE_SPACE_TREE))
1465 		return 0;
1466 
1467 	if (test_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE, &block_group->runtime_flags)) {
1468 		/* We never added this block group to the free space tree. */
1469 		return 0;
1470 	}
1471 
1472 	path = btrfs_alloc_path();
1473 	if (unlikely(!path)) {
1474 		ret = -ENOMEM;
1475 		btrfs_abort_transaction(trans, ret);
1476 		return ret;
1477 	}
1478 
1479 	start = block_group->start;
1480 	end = btrfs_block_group_end(block_group);
1481 
1482 	key.objectid = end - 1;
1483 	key.type = (u8)-1;
1484 	key.offset = (u64)-1;
1485 
1486 	while (!done) {
1487 		ret = btrfs_search_prev_slot(trans, root, &key, path, -1, 1);
1488 		if (unlikely(ret)) {
1489 			btrfs_abort_transaction(trans, ret);
1490 			return ret;
1491 		}
1492 
1493 		leaf = path->nodes[0];
1494 		nr = 0;
1495 		path->slots[0]++;
1496 		while (path->slots[0] > 0) {
1497 			btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0] - 1);
1498 
1499 			if (found_key.type == BTRFS_FREE_SPACE_INFO_KEY) {
1500 				ASSERT(found_key.objectid == block_group->start);
1501 				ASSERT(found_key.offset == block_group->length);
1502 				done = 1;
1503 				nr++;
1504 				path->slots[0]--;
1505 				break;
1506 			} else if (found_key.type == BTRFS_FREE_SPACE_EXTENT_KEY ||
1507 				   found_key.type == BTRFS_FREE_SPACE_BITMAP_KEY) {
1508 				ASSERT(found_key.objectid >= start);
1509 				ASSERT(found_key.objectid < end);
1510 				ASSERT(found_key.objectid + found_key.offset <= end);
1511 				nr++;
1512 				path->slots[0]--;
1513 			} else {
1514 				ASSERT(0);
1515 			}
1516 		}
1517 
1518 		ret = btrfs_del_items(trans, root, path, path->slots[0], nr);
1519 		if (unlikely(ret)) {
1520 			btrfs_abort_transaction(trans, ret);
1521 			return ret;
1522 		}
1523 		btrfs_release_path(path);
1524 	}
1525 
1526 	return 0;
1527 }
1528 
1529 static int load_free_space_bitmaps(struct btrfs_caching_control *caching_ctl,
1530 				   struct btrfs_path *path,
1531 				   u32 expected_extent_count)
1532 {
1533 	struct btrfs_block_group *block_group = caching_ctl->block_group;
1534 	struct btrfs_fs_info *fs_info = block_group->fs_info;
1535 	struct btrfs_root *root;
1536 	struct btrfs_key key;
1537 	bool prev_bit_set = false;
1538 	/* Initialize to silence GCC. */
1539 	u64 extent_start = 0;
1540 	const u64 end = btrfs_block_group_end(block_group);
1541 	u64 offset;
1542 	u64 total_found = 0;
1543 	u32 extent_count = 0;
1544 	int ret;
1545 
1546 	root = btrfs_free_space_root(block_group);
1547 
1548 	while (1) {
1549 		ret = btrfs_next_item(root, path);
1550 		if (ret < 0)
1551 			return ret;
1552 		if (ret)
1553 			break;
1554 
1555 		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1556 
1557 		if (key.type == BTRFS_FREE_SPACE_INFO_KEY)
1558 			break;
1559 
1560 		ASSERT(key.type == BTRFS_FREE_SPACE_BITMAP_KEY);
1561 		ASSERT(key.objectid < end && key.objectid + key.offset <= end);
1562 
1563 		offset = key.objectid;
1564 		while (offset < key.objectid + key.offset) {
1565 			bool bit_set;
1566 
1567 			bit_set = btrfs_free_space_test_bit(block_group, path, offset);
1568 			if (!prev_bit_set && bit_set) {
1569 				extent_start = offset;
1570 			} else if (prev_bit_set && !bit_set) {
1571 				u64 space_added;
1572 
1573 				ret = btrfs_add_new_free_space(block_group,
1574 							       extent_start,
1575 							       offset,
1576 							       &space_added);
1577 				if (ret)
1578 					return ret;
1579 				total_found += space_added;
1580 				if (total_found > CACHING_CTL_WAKE_UP) {
1581 					total_found = 0;
1582 					wake_up(&caching_ctl->wait);
1583 				}
1584 				extent_count++;
1585 			}
1586 			prev_bit_set = bit_set;
1587 			offset += fs_info->sectorsize;
1588 		}
1589 	}
1590 	if (prev_bit_set) {
1591 		ret = btrfs_add_new_free_space(block_group, extent_start, end, NULL);
1592 		if (ret)
1593 			return ret;
1594 		extent_count++;
1595 	}
1596 
1597 	if (unlikely(extent_count != expected_extent_count)) {
1598 		btrfs_err(fs_info,
1599 			  "incorrect extent count for %llu; counted %u, expected %u",
1600 			  block_group->start, extent_count,
1601 			  expected_extent_count);
1602 		DEBUG_WARN();
1603 		return -EIO;
1604 	}
1605 
1606 	return 0;
1607 }
1608 
1609 static int load_free_space_extents(struct btrfs_caching_control *caching_ctl,
1610 				   struct btrfs_path *path,
1611 				   u32 expected_extent_count)
1612 {
1613 	struct btrfs_block_group *block_group = caching_ctl->block_group;
1614 	struct btrfs_fs_info *fs_info = block_group->fs_info;
1615 	struct btrfs_root *root;
1616 	struct btrfs_key key;
1617 	const u64 end = btrfs_block_group_end(block_group);
1618 	u64 total_found = 0;
1619 	u32 extent_count = 0;
1620 	int ret;
1621 
1622 	root = btrfs_free_space_root(block_group);
1623 
1624 	while (1) {
1625 		u64 space_added;
1626 
1627 		ret = btrfs_next_item(root, path);
1628 		if (ret < 0)
1629 			return ret;
1630 		if (ret)
1631 			break;
1632 
1633 		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1634 
1635 		if (key.type == BTRFS_FREE_SPACE_INFO_KEY)
1636 			break;
1637 
1638 		ASSERT(key.type == BTRFS_FREE_SPACE_EXTENT_KEY);
1639 		ASSERT(key.objectid < end && key.objectid + key.offset <= end);
1640 
1641 		ret = btrfs_add_new_free_space(block_group, key.objectid,
1642 					       key.objectid + key.offset,
1643 					       &space_added);
1644 		if (ret)
1645 			return ret;
1646 		total_found += space_added;
1647 		if (total_found > CACHING_CTL_WAKE_UP) {
1648 			total_found = 0;
1649 			wake_up(&caching_ctl->wait);
1650 		}
1651 		extent_count++;
1652 	}
1653 
1654 	if (unlikely(extent_count != expected_extent_count)) {
1655 		btrfs_err(fs_info,
1656 			  "incorrect extent count for %llu; counted %u, expected %u",
1657 			  block_group->start, extent_count,
1658 			  expected_extent_count);
1659 		DEBUG_WARN();
1660 		return -EIO;
1661 	}
1662 
1663 	return 0;
1664 }
1665 
1666 int btrfs_load_free_space_tree(struct btrfs_caching_control *caching_ctl)
1667 {
1668 	struct btrfs_block_group *block_group;
1669 	struct btrfs_free_space_info *info;
1670 	BTRFS_PATH_AUTO_FREE(path);
1671 	u32 extent_count, flags;
1672 
1673 	block_group = caching_ctl->block_group;
1674 
1675 	path = btrfs_alloc_path();
1676 	if (!path)
1677 		return -ENOMEM;
1678 
1679 	/*
1680 	 * Just like caching_thread() doesn't want to deadlock on the extent
1681 	 * tree, we don't want to deadlock on the free space tree.
1682 	 */
1683 	path->skip_locking = true;
1684 	path->search_commit_root = true;
1685 	path->reada = READA_FORWARD;
1686 
1687 	info = btrfs_search_free_space_info(NULL, block_group, path, 0);
1688 	if (IS_ERR(info))
1689 		return PTR_ERR(info);
1690 
1691 	extent_count = btrfs_free_space_extent_count(path->nodes[0], info);
1692 	flags = btrfs_free_space_flags(path->nodes[0], info);
1693 
1694 	/*
1695 	 * We left path pointing to the free space info item, so now
1696 	 * load_free_space_foo can just iterate through the free space tree from
1697 	 * there.
1698 	 */
1699 	if (flags & BTRFS_FREE_SPACE_USING_BITMAPS)
1700 		return load_free_space_bitmaps(caching_ctl, path, extent_count);
1701 	else
1702 		return load_free_space_extents(caching_ctl, path, extent_count);
1703 }
1704 
1705 static int delete_orphan_free_space_entries(struct btrfs_root *fst_root,
1706 					    struct btrfs_path *path,
1707 					    u64 first_bg_bytenr)
1708 {
1709 	struct btrfs_trans_handle *trans;
1710 	int ret;
1711 
1712 	trans = btrfs_start_transaction(fst_root, 1);
1713 	if (IS_ERR(trans))
1714 		return PTR_ERR(trans);
1715 
1716 	while (true) {
1717 		struct btrfs_key key = { 0 };
1718 		int i;
1719 
1720 		ret = btrfs_search_slot(trans, fst_root, &key, path, -1, 1);
1721 		if (ret < 0)
1722 			break;
1723 		ASSERT(ret > 0);
1724 		ret = 0;
1725 		for (i = 0; i < btrfs_header_nritems(path->nodes[0]); i++) {
1726 			btrfs_item_key_to_cpu(path->nodes[0], &key, i);
1727 			if (key.objectid >= first_bg_bytenr) {
1728 				/*
1729 				 * Only break the for() loop and continue to
1730 				 * delete items.
1731 				 */
1732 				break;
1733 			}
1734 		}
1735 		/* No items to delete, finished. */
1736 		if (i == 0)
1737 			break;
1738 
1739 		ret = btrfs_del_items(trans, fst_root, path, 0, i);
1740 		if (ret < 0)
1741 			break;
1742 		btrfs_release_path(path);
1743 	}
1744 	btrfs_release_path(path);
1745 	btrfs_end_transaction(trans);
1746 	if (ret == 0)
1747 		btrfs_info(fst_root->fs_info, "deleted orphan free space tree entries");
1748 	return ret;
1749 }
1750 
1751 /* Remove any free space entry before the first block group. */
1752 int btrfs_delete_orphan_free_space_entries(struct btrfs_fs_info *fs_info)
1753 {
1754 	BTRFS_PATH_AUTO_RELEASE(path);
1755 	struct btrfs_key key = {
1756 		.objectid = BTRFS_FREE_SPACE_TREE_OBJECTID,
1757 		.type = BTRFS_ROOT_ITEM_KEY,
1758 		.offset = 0,
1759 	};
1760 	struct btrfs_root *root;
1761 	struct btrfs_block_group *bg;
1762 	u64 first_bg_bytenr;
1763 	int ret;
1764 
1765 	/*
1766 	 * Extent tree v2 has multiple global roots based on the block group.
1767 	 * This means we cannot easily grab the global free space tree and locate
1768 	 * orphan items.  Furthermore this is still experimental, all users
1769 	 * should use the latest btrfs-progs anyway.
1770 	 */
1771 	if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2))
1772 		return 0;
1773 	if (!btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
1774 		return 0;
1775 	root = btrfs_global_root(fs_info, &key);
1776 	if (!root)
1777 		return 0;
1778 
1779 	key.objectid = 0;
1780 	key.type = 0;
1781 	key.offset = 0;
1782 
1783 	bg = btrfs_lookup_first_block_group(fs_info, 0);
1784 	if (unlikely(!bg)) {
1785 		btrfs_err(fs_info, "no block group found");
1786 		return -EUCLEAN;
1787 	}
1788 	first_bg_bytenr = bg->start;
1789 	btrfs_put_block_group(bg);
1790 
1791 	ret = btrfs_search_slot(NULL, root, &key, &path, 0, 0);
1792 	if (ret < 0)
1793 		return ret;
1794 	/* There should not be an all-zero key in fst. */
1795 	ASSERT(ret > 0);
1796 
1797 	/* Empty free space tree. */
1798 	if (path.slots[0] >= btrfs_header_nritems(path.nodes[0]))
1799 		return 0;
1800 
1801 	btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
1802 	if (key.objectid >= first_bg_bytenr)
1803 		return 0;
1804 	btrfs_release_path(&path);
1805 	return delete_orphan_free_space_entries(root, &path, first_bg_bytenr);
1806 }
1807