xref: /linux/fs/btrfs/block-group.c (revision 26fbb4c8c7c3ee9a4c3b4de555a8587b5a19154e)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "misc.h"
4 #include "ctree.h"
5 #include "block-group.h"
6 #include "space-info.h"
7 #include "disk-io.h"
8 #include "free-space-cache.h"
9 #include "free-space-tree.h"
10 #include "volumes.h"
11 #include "transaction.h"
12 #include "ref-verify.h"
13 #include "sysfs.h"
14 #include "tree-log.h"
15 #include "delalloc-space.h"
16 #include "discard.h"
17 #include "raid56.h"
18 
19 /*
20  * Return target flags in extended format or 0 if restripe for this chunk_type
21  * is not in progress
22  *
23  * Should be called with balance_lock held
24  */
25 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
26 {
27 	struct btrfs_balance_control *bctl = fs_info->balance_ctl;
28 	u64 target = 0;
29 
30 	if (!bctl)
31 		return 0;
32 
33 	if (flags & BTRFS_BLOCK_GROUP_DATA &&
34 	    bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
35 		target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
36 	} else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
37 		   bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
38 		target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
39 	} else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
40 		   bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
41 		target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
42 	}
43 
44 	return target;
45 }
46 
47 /*
48  * @flags: available profiles in extended format (see ctree.h)
49  *
50  * Return reduced profile in chunk format.  If profile changing is in progress
51  * (either running or paused) picks the target profile (if it's already
52  * available), otherwise falls back to plain reducing.
53  */
54 static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
55 {
56 	u64 num_devices = fs_info->fs_devices->rw_devices;
57 	u64 target;
58 	u64 raid_type;
59 	u64 allowed = 0;
60 
61 	/*
62 	 * See if restripe for this chunk_type is in progress, if so try to
63 	 * reduce to the target profile
64 	 */
65 	spin_lock(&fs_info->balance_lock);
66 	target = get_restripe_target(fs_info, flags);
67 	if (target) {
68 		spin_unlock(&fs_info->balance_lock);
69 		return extended_to_chunk(target);
70 	}
71 	spin_unlock(&fs_info->balance_lock);
72 
73 	/* First, mask out the RAID levels which aren't possible */
74 	for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
75 		if (num_devices >= btrfs_raid_array[raid_type].devs_min)
76 			allowed |= btrfs_raid_array[raid_type].bg_flag;
77 	}
78 	allowed &= flags;
79 
80 	if (allowed & BTRFS_BLOCK_GROUP_RAID6)
81 		allowed = BTRFS_BLOCK_GROUP_RAID6;
82 	else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
83 		allowed = BTRFS_BLOCK_GROUP_RAID5;
84 	else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
85 		allowed = BTRFS_BLOCK_GROUP_RAID10;
86 	else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
87 		allowed = BTRFS_BLOCK_GROUP_RAID1;
88 	else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
89 		allowed = BTRFS_BLOCK_GROUP_RAID0;
90 
91 	flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
92 
93 	return extended_to_chunk(flags | allowed);
94 }
95 
96 u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
97 {
98 	unsigned seq;
99 	u64 flags;
100 
101 	do {
102 		flags = orig_flags;
103 		seq = read_seqbegin(&fs_info->profiles_lock);
104 
105 		if (flags & BTRFS_BLOCK_GROUP_DATA)
106 			flags |= fs_info->avail_data_alloc_bits;
107 		else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
108 			flags |= fs_info->avail_system_alloc_bits;
109 		else if (flags & BTRFS_BLOCK_GROUP_METADATA)
110 			flags |= fs_info->avail_metadata_alloc_bits;
111 	} while (read_seqretry(&fs_info->profiles_lock, seq));
112 
113 	return btrfs_reduce_alloc_profile(fs_info, flags);
114 }
115 
116 void btrfs_get_block_group(struct btrfs_block_group *cache)
117 {
118 	refcount_inc(&cache->refs);
119 }
120 
121 void btrfs_put_block_group(struct btrfs_block_group *cache)
122 {
123 	if (refcount_dec_and_test(&cache->refs)) {
124 		WARN_ON(cache->pinned > 0);
125 		WARN_ON(cache->reserved > 0);
126 
127 		/*
128 		 * A block_group shouldn't be on the discard_list anymore.
129 		 * Remove the block_group from the discard_list to prevent us
130 		 * from causing a panic due to NULL pointer dereference.
131 		 */
132 		if (WARN_ON(!list_empty(&cache->discard_list)))
133 			btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
134 						  cache);
135 
136 		/*
137 		 * If not empty, someone is still holding mutex of
138 		 * full_stripe_lock, which can only be released by caller.
139 		 * And it will definitely cause use-after-free when caller
140 		 * tries to release full stripe lock.
141 		 *
142 		 * No better way to resolve, but only to warn.
143 		 */
144 		WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
145 		kfree(cache->free_space_ctl);
146 		kfree(cache);
147 	}
148 }
149 
150 /*
151  * This adds the block group to the fs_info rb tree for the block group cache
152  */
153 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
154 				       struct btrfs_block_group *block_group)
155 {
156 	struct rb_node **p;
157 	struct rb_node *parent = NULL;
158 	struct btrfs_block_group *cache;
159 
160 	ASSERT(block_group->length != 0);
161 
162 	spin_lock(&info->block_group_cache_lock);
163 	p = &info->block_group_cache_tree.rb_node;
164 
165 	while (*p) {
166 		parent = *p;
167 		cache = rb_entry(parent, struct btrfs_block_group, cache_node);
168 		if (block_group->start < cache->start) {
169 			p = &(*p)->rb_left;
170 		} else if (block_group->start > cache->start) {
171 			p = &(*p)->rb_right;
172 		} else {
173 			spin_unlock(&info->block_group_cache_lock);
174 			return -EEXIST;
175 		}
176 	}
177 
178 	rb_link_node(&block_group->cache_node, parent, p);
179 	rb_insert_color(&block_group->cache_node,
180 			&info->block_group_cache_tree);
181 
182 	if (info->first_logical_byte > block_group->start)
183 		info->first_logical_byte = block_group->start;
184 
185 	spin_unlock(&info->block_group_cache_lock);
186 
187 	return 0;
188 }
189 
190 /*
191  * This will return the block group at or after bytenr if contains is 0, else
192  * it will return the block group that contains the bytenr
193  */
194 static struct btrfs_block_group *block_group_cache_tree_search(
195 		struct btrfs_fs_info *info, u64 bytenr, int contains)
196 {
197 	struct btrfs_block_group *cache, *ret = NULL;
198 	struct rb_node *n;
199 	u64 end, start;
200 
201 	spin_lock(&info->block_group_cache_lock);
202 	n = info->block_group_cache_tree.rb_node;
203 
204 	while (n) {
205 		cache = rb_entry(n, struct btrfs_block_group, cache_node);
206 		end = cache->start + cache->length - 1;
207 		start = cache->start;
208 
209 		if (bytenr < start) {
210 			if (!contains && (!ret || start < ret->start))
211 				ret = cache;
212 			n = n->rb_left;
213 		} else if (bytenr > start) {
214 			if (contains && bytenr <= end) {
215 				ret = cache;
216 				break;
217 			}
218 			n = n->rb_right;
219 		} else {
220 			ret = cache;
221 			break;
222 		}
223 	}
224 	if (ret) {
225 		btrfs_get_block_group(ret);
226 		if (bytenr == 0 && info->first_logical_byte > ret->start)
227 			info->first_logical_byte = ret->start;
228 	}
229 	spin_unlock(&info->block_group_cache_lock);
230 
231 	return ret;
232 }
233 
234 /*
235  * Return the block group that starts at or after bytenr
236  */
237 struct btrfs_block_group *btrfs_lookup_first_block_group(
238 		struct btrfs_fs_info *info, u64 bytenr)
239 {
240 	return block_group_cache_tree_search(info, bytenr, 0);
241 }
242 
243 /*
244  * Return the block group that contains the given bytenr
245  */
246 struct btrfs_block_group *btrfs_lookup_block_group(
247 		struct btrfs_fs_info *info, u64 bytenr)
248 {
249 	return block_group_cache_tree_search(info, bytenr, 1);
250 }
251 
252 struct btrfs_block_group *btrfs_next_block_group(
253 		struct btrfs_block_group *cache)
254 {
255 	struct btrfs_fs_info *fs_info = cache->fs_info;
256 	struct rb_node *node;
257 
258 	spin_lock(&fs_info->block_group_cache_lock);
259 
260 	/* If our block group was removed, we need a full search. */
261 	if (RB_EMPTY_NODE(&cache->cache_node)) {
262 		const u64 next_bytenr = cache->start + cache->length;
263 
264 		spin_unlock(&fs_info->block_group_cache_lock);
265 		btrfs_put_block_group(cache);
266 		cache = btrfs_lookup_first_block_group(fs_info, next_bytenr); return cache;
267 	}
268 	node = rb_next(&cache->cache_node);
269 	btrfs_put_block_group(cache);
270 	if (node) {
271 		cache = rb_entry(node, struct btrfs_block_group, cache_node);
272 		btrfs_get_block_group(cache);
273 	} else
274 		cache = NULL;
275 	spin_unlock(&fs_info->block_group_cache_lock);
276 	return cache;
277 }
278 
279 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
280 {
281 	struct btrfs_block_group *bg;
282 	bool ret = true;
283 
284 	bg = btrfs_lookup_block_group(fs_info, bytenr);
285 	if (!bg)
286 		return false;
287 
288 	spin_lock(&bg->lock);
289 	if (bg->ro)
290 		ret = false;
291 	else
292 		atomic_inc(&bg->nocow_writers);
293 	spin_unlock(&bg->lock);
294 
295 	/* No put on block group, done by btrfs_dec_nocow_writers */
296 	if (!ret)
297 		btrfs_put_block_group(bg);
298 
299 	return ret;
300 }
301 
302 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
303 {
304 	struct btrfs_block_group *bg;
305 
306 	bg = btrfs_lookup_block_group(fs_info, bytenr);
307 	ASSERT(bg);
308 	if (atomic_dec_and_test(&bg->nocow_writers))
309 		wake_up_var(&bg->nocow_writers);
310 	/*
311 	 * Once for our lookup and once for the lookup done by a previous call
312 	 * to btrfs_inc_nocow_writers()
313 	 */
314 	btrfs_put_block_group(bg);
315 	btrfs_put_block_group(bg);
316 }
317 
318 void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
319 {
320 	wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
321 }
322 
323 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
324 					const u64 start)
325 {
326 	struct btrfs_block_group *bg;
327 
328 	bg = btrfs_lookup_block_group(fs_info, start);
329 	ASSERT(bg);
330 	if (atomic_dec_and_test(&bg->reservations))
331 		wake_up_var(&bg->reservations);
332 	btrfs_put_block_group(bg);
333 }
334 
335 void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg)
336 {
337 	struct btrfs_space_info *space_info = bg->space_info;
338 
339 	ASSERT(bg->ro);
340 
341 	if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
342 		return;
343 
344 	/*
345 	 * Our block group is read only but before we set it to read only,
346 	 * some task might have had allocated an extent from it already, but it
347 	 * has not yet created a respective ordered extent (and added it to a
348 	 * root's list of ordered extents).
349 	 * Therefore wait for any task currently allocating extents, since the
350 	 * block group's reservations counter is incremented while a read lock
351 	 * on the groups' semaphore is held and decremented after releasing
352 	 * the read access on that semaphore and creating the ordered extent.
353 	 */
354 	down_write(&space_info->groups_sem);
355 	up_write(&space_info->groups_sem);
356 
357 	wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
358 }
359 
360 struct btrfs_caching_control *btrfs_get_caching_control(
361 		struct btrfs_block_group *cache)
362 {
363 	struct btrfs_caching_control *ctl;
364 
365 	spin_lock(&cache->lock);
366 	if (!cache->caching_ctl) {
367 		spin_unlock(&cache->lock);
368 		return NULL;
369 	}
370 
371 	ctl = cache->caching_ctl;
372 	refcount_inc(&ctl->count);
373 	spin_unlock(&cache->lock);
374 	return ctl;
375 }
376 
377 void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
378 {
379 	if (refcount_dec_and_test(&ctl->count))
380 		kfree(ctl);
381 }
382 
383 /*
384  * When we wait for progress in the block group caching, its because our
385  * allocation attempt failed at least once.  So, we must sleep and let some
386  * progress happen before we try again.
387  *
388  * This function will sleep at least once waiting for new free space to show
389  * up, and then it will check the block group free space numbers for our min
390  * num_bytes.  Another option is to have it go ahead and look in the rbtree for
391  * a free extent of a given size, but this is a good start.
392  *
393  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
394  * any of the information in this block group.
395  */
396 void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
397 					   u64 num_bytes)
398 {
399 	struct btrfs_caching_control *caching_ctl;
400 
401 	caching_ctl = btrfs_get_caching_control(cache);
402 	if (!caching_ctl)
403 		return;
404 
405 	wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
406 		   (cache->free_space_ctl->free_space >= num_bytes));
407 
408 	btrfs_put_caching_control(caching_ctl);
409 }
410 
411 int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
412 {
413 	struct btrfs_caching_control *caching_ctl;
414 	int ret = 0;
415 
416 	caching_ctl = btrfs_get_caching_control(cache);
417 	if (!caching_ctl)
418 		return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
419 
420 	wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
421 	if (cache->cached == BTRFS_CACHE_ERROR)
422 		ret = -EIO;
423 	btrfs_put_caching_control(caching_ctl);
424 	return ret;
425 }
426 
427 static bool space_cache_v1_done(struct btrfs_block_group *cache)
428 {
429 	bool ret;
430 
431 	spin_lock(&cache->lock);
432 	ret = cache->cached != BTRFS_CACHE_FAST;
433 	spin_unlock(&cache->lock);
434 
435 	return ret;
436 }
437 
438 void btrfs_wait_space_cache_v1_finished(struct btrfs_block_group *cache,
439 				struct btrfs_caching_control *caching_ctl)
440 {
441 	wait_event(caching_ctl->wait, space_cache_v1_done(cache));
442 }
443 
444 #ifdef CONFIG_BTRFS_DEBUG
445 static void fragment_free_space(struct btrfs_block_group *block_group)
446 {
447 	struct btrfs_fs_info *fs_info = block_group->fs_info;
448 	u64 start = block_group->start;
449 	u64 len = block_group->length;
450 	u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
451 		fs_info->nodesize : fs_info->sectorsize;
452 	u64 step = chunk << 1;
453 
454 	while (len > chunk) {
455 		btrfs_remove_free_space(block_group, start, chunk);
456 		start += step;
457 		if (len < step)
458 			len = 0;
459 		else
460 			len -= step;
461 	}
462 }
463 #endif
464 
465 /*
466  * This is only called by btrfs_cache_block_group, since we could have freed
467  * extents we need to check the pinned_extents for any extents that can't be
468  * used yet since their free space will be released as soon as the transaction
469  * commits.
470  */
471 u64 add_new_free_space(struct btrfs_block_group *block_group, u64 start, u64 end)
472 {
473 	struct btrfs_fs_info *info = block_group->fs_info;
474 	u64 extent_start, extent_end, size, total_added = 0;
475 	int ret;
476 
477 	while (start < end) {
478 		ret = find_first_extent_bit(&info->excluded_extents, start,
479 					    &extent_start, &extent_end,
480 					    EXTENT_DIRTY | EXTENT_UPTODATE,
481 					    NULL);
482 		if (ret)
483 			break;
484 
485 		if (extent_start <= start) {
486 			start = extent_end + 1;
487 		} else if (extent_start > start && extent_start < end) {
488 			size = extent_start - start;
489 			total_added += size;
490 			ret = btrfs_add_free_space_async_trimmed(block_group,
491 								 start, size);
492 			BUG_ON(ret); /* -ENOMEM or logic error */
493 			start = extent_end + 1;
494 		} else {
495 			break;
496 		}
497 	}
498 
499 	if (start < end) {
500 		size = end - start;
501 		total_added += size;
502 		ret = btrfs_add_free_space_async_trimmed(block_group, start,
503 							 size);
504 		BUG_ON(ret); /* -ENOMEM or logic error */
505 	}
506 
507 	return total_added;
508 }
509 
510 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
511 {
512 	struct btrfs_block_group *block_group = caching_ctl->block_group;
513 	struct btrfs_fs_info *fs_info = block_group->fs_info;
514 	struct btrfs_root *extent_root = fs_info->extent_root;
515 	struct btrfs_path *path;
516 	struct extent_buffer *leaf;
517 	struct btrfs_key key;
518 	u64 total_found = 0;
519 	u64 last = 0;
520 	u32 nritems;
521 	int ret;
522 	bool wakeup = true;
523 
524 	path = btrfs_alloc_path();
525 	if (!path)
526 		return -ENOMEM;
527 
528 	last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
529 
530 #ifdef CONFIG_BTRFS_DEBUG
531 	/*
532 	 * If we're fragmenting we don't want to make anybody think we can
533 	 * allocate from this block group until we've had a chance to fragment
534 	 * the free space.
535 	 */
536 	if (btrfs_should_fragment_free_space(block_group))
537 		wakeup = false;
538 #endif
539 	/*
540 	 * We don't want to deadlock with somebody trying to allocate a new
541 	 * extent for the extent root while also trying to search the extent
542 	 * root to add free space.  So we skip locking and search the commit
543 	 * root, since its read-only
544 	 */
545 	path->skip_locking = 1;
546 	path->search_commit_root = 1;
547 	path->reada = READA_FORWARD;
548 
549 	key.objectid = last;
550 	key.offset = 0;
551 	key.type = BTRFS_EXTENT_ITEM_KEY;
552 
553 next:
554 	ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
555 	if (ret < 0)
556 		goto out;
557 
558 	leaf = path->nodes[0];
559 	nritems = btrfs_header_nritems(leaf);
560 
561 	while (1) {
562 		if (btrfs_fs_closing(fs_info) > 1) {
563 			last = (u64)-1;
564 			break;
565 		}
566 
567 		if (path->slots[0] < nritems) {
568 			btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
569 		} else {
570 			ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
571 			if (ret)
572 				break;
573 
574 			if (need_resched() ||
575 			    rwsem_is_contended(&fs_info->commit_root_sem)) {
576 				if (wakeup)
577 					caching_ctl->progress = last;
578 				btrfs_release_path(path);
579 				up_read(&fs_info->commit_root_sem);
580 				mutex_unlock(&caching_ctl->mutex);
581 				cond_resched();
582 				mutex_lock(&caching_ctl->mutex);
583 				down_read(&fs_info->commit_root_sem);
584 				goto next;
585 			}
586 
587 			ret = btrfs_next_leaf(extent_root, path);
588 			if (ret < 0)
589 				goto out;
590 			if (ret)
591 				break;
592 			leaf = path->nodes[0];
593 			nritems = btrfs_header_nritems(leaf);
594 			continue;
595 		}
596 
597 		if (key.objectid < last) {
598 			key.objectid = last;
599 			key.offset = 0;
600 			key.type = BTRFS_EXTENT_ITEM_KEY;
601 
602 			if (wakeup)
603 				caching_ctl->progress = last;
604 			btrfs_release_path(path);
605 			goto next;
606 		}
607 
608 		if (key.objectid < block_group->start) {
609 			path->slots[0]++;
610 			continue;
611 		}
612 
613 		if (key.objectid >= block_group->start + block_group->length)
614 			break;
615 
616 		if (key.type == BTRFS_EXTENT_ITEM_KEY ||
617 		    key.type == BTRFS_METADATA_ITEM_KEY) {
618 			total_found += add_new_free_space(block_group, last,
619 							  key.objectid);
620 			if (key.type == BTRFS_METADATA_ITEM_KEY)
621 				last = key.objectid +
622 					fs_info->nodesize;
623 			else
624 				last = key.objectid + key.offset;
625 
626 			if (total_found > CACHING_CTL_WAKE_UP) {
627 				total_found = 0;
628 				if (wakeup)
629 					wake_up(&caching_ctl->wait);
630 			}
631 		}
632 		path->slots[0]++;
633 	}
634 	ret = 0;
635 
636 	total_found += add_new_free_space(block_group, last,
637 				block_group->start + block_group->length);
638 	caching_ctl->progress = (u64)-1;
639 
640 out:
641 	btrfs_free_path(path);
642 	return ret;
643 }
644 
645 static noinline void caching_thread(struct btrfs_work *work)
646 {
647 	struct btrfs_block_group *block_group;
648 	struct btrfs_fs_info *fs_info;
649 	struct btrfs_caching_control *caching_ctl;
650 	int ret;
651 
652 	caching_ctl = container_of(work, struct btrfs_caching_control, work);
653 	block_group = caching_ctl->block_group;
654 	fs_info = block_group->fs_info;
655 
656 	mutex_lock(&caching_ctl->mutex);
657 	down_read(&fs_info->commit_root_sem);
658 
659 	if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
660 		ret = load_free_space_cache(block_group);
661 		if (ret == 1) {
662 			ret = 0;
663 			goto done;
664 		}
665 
666 		/*
667 		 * We failed to load the space cache, set ourselves to
668 		 * CACHE_STARTED and carry on.
669 		 */
670 		spin_lock(&block_group->lock);
671 		block_group->cached = BTRFS_CACHE_STARTED;
672 		spin_unlock(&block_group->lock);
673 		wake_up(&caching_ctl->wait);
674 	}
675 
676 	if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
677 		ret = load_free_space_tree(caching_ctl);
678 	else
679 		ret = load_extent_tree_free(caching_ctl);
680 done:
681 	spin_lock(&block_group->lock);
682 	block_group->caching_ctl = NULL;
683 	block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
684 	spin_unlock(&block_group->lock);
685 
686 #ifdef CONFIG_BTRFS_DEBUG
687 	if (btrfs_should_fragment_free_space(block_group)) {
688 		u64 bytes_used;
689 
690 		spin_lock(&block_group->space_info->lock);
691 		spin_lock(&block_group->lock);
692 		bytes_used = block_group->length - block_group->used;
693 		block_group->space_info->bytes_used += bytes_used >> 1;
694 		spin_unlock(&block_group->lock);
695 		spin_unlock(&block_group->space_info->lock);
696 		fragment_free_space(block_group);
697 	}
698 #endif
699 
700 	caching_ctl->progress = (u64)-1;
701 
702 	up_read(&fs_info->commit_root_sem);
703 	btrfs_free_excluded_extents(block_group);
704 	mutex_unlock(&caching_ctl->mutex);
705 
706 	wake_up(&caching_ctl->wait);
707 
708 	btrfs_put_caching_control(caching_ctl);
709 	btrfs_put_block_group(block_group);
710 }
711 
712 int btrfs_cache_block_group(struct btrfs_block_group *cache, int load_cache_only)
713 {
714 	DEFINE_WAIT(wait);
715 	struct btrfs_fs_info *fs_info = cache->fs_info;
716 	struct btrfs_caching_control *caching_ctl = NULL;
717 	int ret = 0;
718 
719 	caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
720 	if (!caching_ctl)
721 		return -ENOMEM;
722 
723 	INIT_LIST_HEAD(&caching_ctl->list);
724 	mutex_init(&caching_ctl->mutex);
725 	init_waitqueue_head(&caching_ctl->wait);
726 	caching_ctl->block_group = cache;
727 	caching_ctl->progress = cache->start;
728 	refcount_set(&caching_ctl->count, 2);
729 	btrfs_init_work(&caching_ctl->work, caching_thread, NULL, NULL);
730 
731 	spin_lock(&cache->lock);
732 	if (cache->cached != BTRFS_CACHE_NO) {
733 		kfree(caching_ctl);
734 
735 		caching_ctl = cache->caching_ctl;
736 		if (caching_ctl)
737 			refcount_inc(&caching_ctl->count);
738 		spin_unlock(&cache->lock);
739 		goto out;
740 	}
741 	WARN_ON(cache->caching_ctl);
742 	cache->caching_ctl = caching_ctl;
743 	if (btrfs_test_opt(fs_info, SPACE_CACHE))
744 		cache->cached = BTRFS_CACHE_FAST;
745 	else
746 		cache->cached = BTRFS_CACHE_STARTED;
747 	cache->has_caching_ctl = 1;
748 	spin_unlock(&cache->lock);
749 
750 	spin_lock(&fs_info->block_group_cache_lock);
751 	refcount_inc(&caching_ctl->count);
752 	list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
753 	spin_unlock(&fs_info->block_group_cache_lock);
754 
755 	btrfs_get_block_group(cache);
756 
757 	btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
758 out:
759 	if (load_cache_only && caching_ctl)
760 		btrfs_wait_space_cache_v1_finished(cache, caching_ctl);
761 	if (caching_ctl)
762 		btrfs_put_caching_control(caching_ctl);
763 
764 	return ret;
765 }
766 
767 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
768 {
769 	u64 extra_flags = chunk_to_extended(flags) &
770 				BTRFS_EXTENDED_PROFILE_MASK;
771 
772 	write_seqlock(&fs_info->profiles_lock);
773 	if (flags & BTRFS_BLOCK_GROUP_DATA)
774 		fs_info->avail_data_alloc_bits &= ~extra_flags;
775 	if (flags & BTRFS_BLOCK_GROUP_METADATA)
776 		fs_info->avail_metadata_alloc_bits &= ~extra_flags;
777 	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
778 		fs_info->avail_system_alloc_bits &= ~extra_flags;
779 	write_sequnlock(&fs_info->profiles_lock);
780 }
781 
782 /*
783  * Clear incompat bits for the following feature(s):
784  *
785  * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
786  *            in the whole filesystem
787  *
788  * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
789  */
790 static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
791 {
792 	bool found_raid56 = false;
793 	bool found_raid1c34 = false;
794 
795 	if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
796 	    (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
797 	    (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
798 		struct list_head *head = &fs_info->space_info;
799 		struct btrfs_space_info *sinfo;
800 
801 		list_for_each_entry_rcu(sinfo, head, list) {
802 			down_read(&sinfo->groups_sem);
803 			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
804 				found_raid56 = true;
805 			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
806 				found_raid56 = true;
807 			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
808 				found_raid1c34 = true;
809 			if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
810 				found_raid1c34 = true;
811 			up_read(&sinfo->groups_sem);
812 		}
813 		if (!found_raid56)
814 			btrfs_clear_fs_incompat(fs_info, RAID56);
815 		if (!found_raid1c34)
816 			btrfs_clear_fs_incompat(fs_info, RAID1C34);
817 	}
818 }
819 
820 static int remove_block_group_item(struct btrfs_trans_handle *trans,
821 				   struct btrfs_path *path,
822 				   struct btrfs_block_group *block_group)
823 {
824 	struct btrfs_fs_info *fs_info = trans->fs_info;
825 	struct btrfs_root *root;
826 	struct btrfs_key key;
827 	int ret;
828 
829 	root = fs_info->extent_root;
830 	key.objectid = block_group->start;
831 	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
832 	key.offset = block_group->length;
833 
834 	ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
835 	if (ret > 0)
836 		ret = -ENOENT;
837 	if (ret < 0)
838 		return ret;
839 
840 	ret = btrfs_del_item(trans, root, path);
841 	return ret;
842 }
843 
844 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
845 			     u64 group_start, struct extent_map *em)
846 {
847 	struct btrfs_fs_info *fs_info = trans->fs_info;
848 	struct btrfs_path *path;
849 	struct btrfs_block_group *block_group;
850 	struct btrfs_free_cluster *cluster;
851 	struct inode *inode;
852 	struct kobject *kobj = NULL;
853 	int ret;
854 	int index;
855 	int factor;
856 	struct btrfs_caching_control *caching_ctl = NULL;
857 	bool remove_em;
858 	bool remove_rsv = false;
859 
860 	block_group = btrfs_lookup_block_group(fs_info, group_start);
861 	BUG_ON(!block_group);
862 	BUG_ON(!block_group->ro);
863 
864 	trace_btrfs_remove_block_group(block_group);
865 	/*
866 	 * Free the reserved super bytes from this block group before
867 	 * remove it.
868 	 */
869 	btrfs_free_excluded_extents(block_group);
870 	btrfs_free_ref_tree_range(fs_info, block_group->start,
871 				  block_group->length);
872 
873 	index = btrfs_bg_flags_to_raid_index(block_group->flags);
874 	factor = btrfs_bg_type_to_factor(block_group->flags);
875 
876 	/* make sure this block group isn't part of an allocation cluster */
877 	cluster = &fs_info->data_alloc_cluster;
878 	spin_lock(&cluster->refill_lock);
879 	btrfs_return_cluster_to_free_space(block_group, cluster);
880 	spin_unlock(&cluster->refill_lock);
881 
882 	/*
883 	 * make sure this block group isn't part of a metadata
884 	 * allocation cluster
885 	 */
886 	cluster = &fs_info->meta_alloc_cluster;
887 	spin_lock(&cluster->refill_lock);
888 	btrfs_return_cluster_to_free_space(block_group, cluster);
889 	spin_unlock(&cluster->refill_lock);
890 
891 	path = btrfs_alloc_path();
892 	if (!path) {
893 		ret = -ENOMEM;
894 		goto out;
895 	}
896 
897 	/*
898 	 * get the inode first so any iput calls done for the io_list
899 	 * aren't the final iput (no unlinks allowed now)
900 	 */
901 	inode = lookup_free_space_inode(block_group, path);
902 
903 	mutex_lock(&trans->transaction->cache_write_mutex);
904 	/*
905 	 * Make sure our free space cache IO is done before removing the
906 	 * free space inode
907 	 */
908 	spin_lock(&trans->transaction->dirty_bgs_lock);
909 	if (!list_empty(&block_group->io_list)) {
910 		list_del_init(&block_group->io_list);
911 
912 		WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
913 
914 		spin_unlock(&trans->transaction->dirty_bgs_lock);
915 		btrfs_wait_cache_io(trans, block_group, path);
916 		btrfs_put_block_group(block_group);
917 		spin_lock(&trans->transaction->dirty_bgs_lock);
918 	}
919 
920 	if (!list_empty(&block_group->dirty_list)) {
921 		list_del_init(&block_group->dirty_list);
922 		remove_rsv = true;
923 		btrfs_put_block_group(block_group);
924 	}
925 	spin_unlock(&trans->transaction->dirty_bgs_lock);
926 	mutex_unlock(&trans->transaction->cache_write_mutex);
927 
928 	ret = btrfs_remove_free_space_inode(trans, inode, block_group);
929 	if (ret)
930 		goto out;
931 
932 	spin_lock(&fs_info->block_group_cache_lock);
933 	rb_erase(&block_group->cache_node,
934 		 &fs_info->block_group_cache_tree);
935 	RB_CLEAR_NODE(&block_group->cache_node);
936 
937 	/* Once for the block groups rbtree */
938 	btrfs_put_block_group(block_group);
939 
940 	if (fs_info->first_logical_byte == block_group->start)
941 		fs_info->first_logical_byte = (u64)-1;
942 	spin_unlock(&fs_info->block_group_cache_lock);
943 
944 	down_write(&block_group->space_info->groups_sem);
945 	/*
946 	 * we must use list_del_init so people can check to see if they
947 	 * are still on the list after taking the semaphore
948 	 */
949 	list_del_init(&block_group->list);
950 	if (list_empty(&block_group->space_info->block_groups[index])) {
951 		kobj = block_group->space_info->block_group_kobjs[index];
952 		block_group->space_info->block_group_kobjs[index] = NULL;
953 		clear_avail_alloc_bits(fs_info, block_group->flags);
954 	}
955 	up_write(&block_group->space_info->groups_sem);
956 	clear_incompat_bg_bits(fs_info, block_group->flags);
957 	if (kobj) {
958 		kobject_del(kobj);
959 		kobject_put(kobj);
960 	}
961 
962 	if (block_group->has_caching_ctl)
963 		caching_ctl = btrfs_get_caching_control(block_group);
964 	if (block_group->cached == BTRFS_CACHE_STARTED)
965 		btrfs_wait_block_group_cache_done(block_group);
966 	if (block_group->has_caching_ctl) {
967 		spin_lock(&fs_info->block_group_cache_lock);
968 		if (!caching_ctl) {
969 			struct btrfs_caching_control *ctl;
970 
971 			list_for_each_entry(ctl,
972 				    &fs_info->caching_block_groups, list)
973 				if (ctl->block_group == block_group) {
974 					caching_ctl = ctl;
975 					refcount_inc(&caching_ctl->count);
976 					break;
977 				}
978 		}
979 		if (caching_ctl)
980 			list_del_init(&caching_ctl->list);
981 		spin_unlock(&fs_info->block_group_cache_lock);
982 		if (caching_ctl) {
983 			/* Once for the caching bgs list and once for us. */
984 			btrfs_put_caching_control(caching_ctl);
985 			btrfs_put_caching_control(caching_ctl);
986 		}
987 	}
988 
989 	spin_lock(&trans->transaction->dirty_bgs_lock);
990 	WARN_ON(!list_empty(&block_group->dirty_list));
991 	WARN_ON(!list_empty(&block_group->io_list));
992 	spin_unlock(&trans->transaction->dirty_bgs_lock);
993 
994 	btrfs_remove_free_space_cache(block_group);
995 
996 	spin_lock(&block_group->space_info->lock);
997 	list_del_init(&block_group->ro_list);
998 
999 	if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1000 		WARN_ON(block_group->space_info->total_bytes
1001 			< block_group->length);
1002 		WARN_ON(block_group->space_info->bytes_readonly
1003 			< block_group->length);
1004 		WARN_ON(block_group->space_info->disk_total
1005 			< block_group->length * factor);
1006 	}
1007 	block_group->space_info->total_bytes -= block_group->length;
1008 	block_group->space_info->bytes_readonly -= block_group->length;
1009 	block_group->space_info->disk_total -= block_group->length * factor;
1010 
1011 	spin_unlock(&block_group->space_info->lock);
1012 
1013 	/*
1014 	 * Remove the free space for the block group from the free space tree
1015 	 * and the block group's item from the extent tree before marking the
1016 	 * block group as removed. This is to prevent races with tasks that
1017 	 * freeze and unfreeze a block group, this task and another task
1018 	 * allocating a new block group - the unfreeze task ends up removing
1019 	 * the block group's extent map before the task calling this function
1020 	 * deletes the block group item from the extent tree, allowing for
1021 	 * another task to attempt to create another block group with the same
1022 	 * item key (and failing with -EEXIST and a transaction abort).
1023 	 */
1024 	ret = remove_block_group_free_space(trans, block_group);
1025 	if (ret)
1026 		goto out;
1027 
1028 	ret = remove_block_group_item(trans, path, block_group);
1029 	if (ret < 0)
1030 		goto out;
1031 
1032 	spin_lock(&block_group->lock);
1033 	block_group->removed = 1;
1034 	/*
1035 	 * At this point trimming or scrub can't start on this block group,
1036 	 * because we removed the block group from the rbtree
1037 	 * fs_info->block_group_cache_tree so no one can't find it anymore and
1038 	 * even if someone already got this block group before we removed it
1039 	 * from the rbtree, they have already incremented block_group->frozen -
1040 	 * if they didn't, for the trimming case they won't find any free space
1041 	 * entries because we already removed them all when we called
1042 	 * btrfs_remove_free_space_cache().
1043 	 *
1044 	 * And we must not remove the extent map from the fs_info->mapping_tree
1045 	 * to prevent the same logical address range and physical device space
1046 	 * ranges from being reused for a new block group. This is needed to
1047 	 * avoid races with trimming and scrub.
1048 	 *
1049 	 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1050 	 * completely transactionless, so while it is trimming a range the
1051 	 * currently running transaction might finish and a new one start,
1052 	 * allowing for new block groups to be created that can reuse the same
1053 	 * physical device locations unless we take this special care.
1054 	 *
1055 	 * There may also be an implicit trim operation if the file system
1056 	 * is mounted with -odiscard. The same protections must remain
1057 	 * in place until the extents have been discarded completely when
1058 	 * the transaction commit has completed.
1059 	 */
1060 	remove_em = (atomic_read(&block_group->frozen) == 0);
1061 	spin_unlock(&block_group->lock);
1062 
1063 	if (remove_em) {
1064 		struct extent_map_tree *em_tree;
1065 
1066 		em_tree = &fs_info->mapping_tree;
1067 		write_lock(&em_tree->lock);
1068 		remove_extent_mapping(em_tree, em);
1069 		write_unlock(&em_tree->lock);
1070 		/* once for the tree */
1071 		free_extent_map(em);
1072 	}
1073 
1074 out:
1075 	/* Once for the lookup reference */
1076 	btrfs_put_block_group(block_group);
1077 	if (remove_rsv)
1078 		btrfs_delayed_refs_rsv_release(fs_info, 1);
1079 	btrfs_free_path(path);
1080 	return ret;
1081 }
1082 
1083 struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1084 		struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1085 {
1086 	struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1087 	struct extent_map *em;
1088 	struct map_lookup *map;
1089 	unsigned int num_items;
1090 
1091 	read_lock(&em_tree->lock);
1092 	em = lookup_extent_mapping(em_tree, chunk_offset, 1);
1093 	read_unlock(&em_tree->lock);
1094 	ASSERT(em && em->start == chunk_offset);
1095 
1096 	/*
1097 	 * We need to reserve 3 + N units from the metadata space info in order
1098 	 * to remove a block group (done at btrfs_remove_chunk() and at
1099 	 * btrfs_remove_block_group()), which are used for:
1100 	 *
1101 	 * 1 unit for adding the free space inode's orphan (located in the tree
1102 	 * of tree roots).
1103 	 * 1 unit for deleting the block group item (located in the extent
1104 	 * tree).
1105 	 * 1 unit for deleting the free space item (located in tree of tree
1106 	 * roots).
1107 	 * N units for deleting N device extent items corresponding to each
1108 	 * stripe (located in the device tree).
1109 	 *
1110 	 * In order to remove a block group we also need to reserve units in the
1111 	 * system space info in order to update the chunk tree (update one or
1112 	 * more device items and remove one chunk item), but this is done at
1113 	 * btrfs_remove_chunk() through a call to check_system_chunk().
1114 	 */
1115 	map = em->map_lookup;
1116 	num_items = 3 + map->num_stripes;
1117 	free_extent_map(em);
1118 
1119 	return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
1120 							   num_items);
1121 }
1122 
1123 /*
1124  * Mark block group @cache read-only, so later write won't happen to block
1125  * group @cache.
1126  *
1127  * If @force is not set, this function will only mark the block group readonly
1128  * if we have enough free space (1M) in other metadata/system block groups.
1129  * If @force is not set, this function will mark the block group readonly
1130  * without checking free space.
1131  *
1132  * NOTE: This function doesn't care if other block groups can contain all the
1133  * data in this block group. That check should be done by relocation routine,
1134  * not this function.
1135  */
1136 static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1137 {
1138 	struct btrfs_space_info *sinfo = cache->space_info;
1139 	u64 num_bytes;
1140 	int ret = -ENOSPC;
1141 
1142 	spin_lock(&sinfo->lock);
1143 	spin_lock(&cache->lock);
1144 
1145 	if (cache->ro) {
1146 		cache->ro++;
1147 		ret = 0;
1148 		goto out;
1149 	}
1150 
1151 	num_bytes = cache->length - cache->reserved - cache->pinned -
1152 		    cache->bytes_super - cache->used;
1153 
1154 	/*
1155 	 * Data never overcommits, even in mixed mode, so do just the straight
1156 	 * check of left over space in how much we have allocated.
1157 	 */
1158 	if (force) {
1159 		ret = 0;
1160 	} else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1161 		u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1162 
1163 		/*
1164 		 * Here we make sure if we mark this bg RO, we still have enough
1165 		 * free space as buffer.
1166 		 */
1167 		if (sinfo_used + num_bytes <= sinfo->total_bytes)
1168 			ret = 0;
1169 	} else {
1170 		/*
1171 		 * We overcommit metadata, so we need to do the
1172 		 * btrfs_can_overcommit check here, and we need to pass in
1173 		 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1174 		 * leeway to allow us to mark this block group as read only.
1175 		 */
1176 		if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1177 					 BTRFS_RESERVE_NO_FLUSH))
1178 			ret = 0;
1179 	}
1180 
1181 	if (!ret) {
1182 		sinfo->bytes_readonly += num_bytes;
1183 		cache->ro++;
1184 		list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1185 	}
1186 out:
1187 	spin_unlock(&cache->lock);
1188 	spin_unlock(&sinfo->lock);
1189 	if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1190 		btrfs_info(cache->fs_info,
1191 			"unable to make block group %llu ro", cache->start);
1192 		btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1193 	}
1194 	return ret;
1195 }
1196 
1197 static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1198 				 struct btrfs_block_group *bg)
1199 {
1200 	struct btrfs_fs_info *fs_info = bg->fs_info;
1201 	struct btrfs_transaction *prev_trans = NULL;
1202 	const u64 start = bg->start;
1203 	const u64 end = start + bg->length - 1;
1204 	int ret;
1205 
1206 	spin_lock(&fs_info->trans_lock);
1207 	if (trans->transaction->list.prev != &fs_info->trans_list) {
1208 		prev_trans = list_last_entry(&trans->transaction->list,
1209 					     struct btrfs_transaction, list);
1210 		refcount_inc(&prev_trans->use_count);
1211 	}
1212 	spin_unlock(&fs_info->trans_lock);
1213 
1214 	/*
1215 	 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1216 	 * btrfs_finish_extent_commit(). If we are at transaction N, another
1217 	 * task might be running finish_extent_commit() for the previous
1218 	 * transaction N - 1, and have seen a range belonging to the block
1219 	 * group in pinned_extents before we were able to clear the whole block
1220 	 * group range from pinned_extents. This means that task can lookup for
1221 	 * the block group after we unpinned it from pinned_extents and removed
1222 	 * it, leading to a BUG_ON() at unpin_extent_range().
1223 	 */
1224 	mutex_lock(&fs_info->unused_bg_unpin_mutex);
1225 	if (prev_trans) {
1226 		ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1227 					EXTENT_DIRTY);
1228 		if (ret)
1229 			goto out;
1230 	}
1231 
1232 	ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1233 				EXTENT_DIRTY);
1234 out:
1235 	mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1236 	if (prev_trans)
1237 		btrfs_put_transaction(prev_trans);
1238 
1239 	return ret == 0;
1240 }
1241 
1242 /*
1243  * Process the unused_bgs list and remove any that don't have any allocated
1244  * space inside of them.
1245  */
1246 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1247 {
1248 	struct btrfs_block_group *block_group;
1249 	struct btrfs_space_info *space_info;
1250 	struct btrfs_trans_handle *trans;
1251 	const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1252 	int ret = 0;
1253 
1254 	if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1255 		return;
1256 
1257 	spin_lock(&fs_info->unused_bgs_lock);
1258 	while (!list_empty(&fs_info->unused_bgs)) {
1259 		int trimming;
1260 
1261 		block_group = list_first_entry(&fs_info->unused_bgs,
1262 					       struct btrfs_block_group,
1263 					       bg_list);
1264 		list_del_init(&block_group->bg_list);
1265 
1266 		space_info = block_group->space_info;
1267 
1268 		if (ret || btrfs_mixed_space_info(space_info)) {
1269 			btrfs_put_block_group(block_group);
1270 			continue;
1271 		}
1272 		spin_unlock(&fs_info->unused_bgs_lock);
1273 
1274 		btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1275 
1276 		mutex_lock(&fs_info->delete_unused_bgs_mutex);
1277 
1278 		/* Don't want to race with allocators so take the groups_sem */
1279 		down_write(&space_info->groups_sem);
1280 
1281 		/*
1282 		 * Async discard moves the final block group discard to be prior
1283 		 * to the unused_bgs code path.  Therefore, if it's not fully
1284 		 * trimmed, punt it back to the async discard lists.
1285 		 */
1286 		if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1287 		    !btrfs_is_free_space_trimmed(block_group)) {
1288 			trace_btrfs_skip_unused_block_group(block_group);
1289 			up_write(&space_info->groups_sem);
1290 			/* Requeue if we failed because of async discard */
1291 			btrfs_discard_queue_work(&fs_info->discard_ctl,
1292 						 block_group);
1293 			goto next;
1294 		}
1295 
1296 		spin_lock(&block_group->lock);
1297 		if (block_group->reserved || block_group->pinned ||
1298 		    block_group->used || block_group->ro ||
1299 		    list_is_singular(&block_group->list)) {
1300 			/*
1301 			 * We want to bail if we made new allocations or have
1302 			 * outstanding allocations in this block group.  We do
1303 			 * the ro check in case balance is currently acting on
1304 			 * this block group.
1305 			 */
1306 			trace_btrfs_skip_unused_block_group(block_group);
1307 			spin_unlock(&block_group->lock);
1308 			up_write(&space_info->groups_sem);
1309 			goto next;
1310 		}
1311 		spin_unlock(&block_group->lock);
1312 
1313 		/* We don't want to force the issue, only flip if it's ok. */
1314 		ret = inc_block_group_ro(block_group, 0);
1315 		up_write(&space_info->groups_sem);
1316 		if (ret < 0) {
1317 			ret = 0;
1318 			goto next;
1319 		}
1320 
1321 		/*
1322 		 * Want to do this before we do anything else so we can recover
1323 		 * properly if we fail to join the transaction.
1324 		 */
1325 		trans = btrfs_start_trans_remove_block_group(fs_info,
1326 						     block_group->start);
1327 		if (IS_ERR(trans)) {
1328 			btrfs_dec_block_group_ro(block_group);
1329 			ret = PTR_ERR(trans);
1330 			goto next;
1331 		}
1332 
1333 		/*
1334 		 * We could have pending pinned extents for this block group,
1335 		 * just delete them, we don't care about them anymore.
1336 		 */
1337 		if (!clean_pinned_extents(trans, block_group)) {
1338 			btrfs_dec_block_group_ro(block_group);
1339 			goto end_trans;
1340 		}
1341 
1342 		/*
1343 		 * At this point, the block_group is read only and should fail
1344 		 * new allocations.  However, btrfs_finish_extent_commit() can
1345 		 * cause this block_group to be placed back on the discard
1346 		 * lists because now the block_group isn't fully discarded.
1347 		 * Bail here and try again later after discarding everything.
1348 		 */
1349 		spin_lock(&fs_info->discard_ctl.lock);
1350 		if (!list_empty(&block_group->discard_list)) {
1351 			spin_unlock(&fs_info->discard_ctl.lock);
1352 			btrfs_dec_block_group_ro(block_group);
1353 			btrfs_discard_queue_work(&fs_info->discard_ctl,
1354 						 block_group);
1355 			goto end_trans;
1356 		}
1357 		spin_unlock(&fs_info->discard_ctl.lock);
1358 
1359 		/* Reset pinned so btrfs_put_block_group doesn't complain */
1360 		spin_lock(&space_info->lock);
1361 		spin_lock(&block_group->lock);
1362 
1363 		btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1364 						     -block_group->pinned);
1365 		space_info->bytes_readonly += block_group->pinned;
1366 		percpu_counter_add_batch(&space_info->total_bytes_pinned,
1367 				   -block_group->pinned,
1368 				   BTRFS_TOTAL_BYTES_PINNED_BATCH);
1369 		block_group->pinned = 0;
1370 
1371 		spin_unlock(&block_group->lock);
1372 		spin_unlock(&space_info->lock);
1373 
1374 		/*
1375 		 * The normal path here is an unused block group is passed here,
1376 		 * then trimming is handled in the transaction commit path.
1377 		 * Async discard interposes before this to do the trimming
1378 		 * before coming down the unused block group path as trimming
1379 		 * will no longer be done later in the transaction commit path.
1380 		 */
1381 		if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1382 			goto flip_async;
1383 
1384 		/* DISCARD can flip during remount */
1385 		trimming = btrfs_test_opt(fs_info, DISCARD_SYNC);
1386 
1387 		/* Implicit trim during transaction commit. */
1388 		if (trimming)
1389 			btrfs_freeze_block_group(block_group);
1390 
1391 		/*
1392 		 * Btrfs_remove_chunk will abort the transaction if things go
1393 		 * horribly wrong.
1394 		 */
1395 		ret = btrfs_remove_chunk(trans, block_group->start);
1396 
1397 		if (ret) {
1398 			if (trimming)
1399 				btrfs_unfreeze_block_group(block_group);
1400 			goto end_trans;
1401 		}
1402 
1403 		/*
1404 		 * If we're not mounted with -odiscard, we can just forget
1405 		 * about this block group. Otherwise we'll need to wait
1406 		 * until transaction commit to do the actual discard.
1407 		 */
1408 		if (trimming) {
1409 			spin_lock(&fs_info->unused_bgs_lock);
1410 			/*
1411 			 * A concurrent scrub might have added us to the list
1412 			 * fs_info->unused_bgs, so use a list_move operation
1413 			 * to add the block group to the deleted_bgs list.
1414 			 */
1415 			list_move(&block_group->bg_list,
1416 				  &trans->transaction->deleted_bgs);
1417 			spin_unlock(&fs_info->unused_bgs_lock);
1418 			btrfs_get_block_group(block_group);
1419 		}
1420 end_trans:
1421 		btrfs_end_transaction(trans);
1422 next:
1423 		mutex_unlock(&fs_info->delete_unused_bgs_mutex);
1424 		btrfs_put_block_group(block_group);
1425 		spin_lock(&fs_info->unused_bgs_lock);
1426 	}
1427 	spin_unlock(&fs_info->unused_bgs_lock);
1428 	return;
1429 
1430 flip_async:
1431 	btrfs_end_transaction(trans);
1432 	mutex_unlock(&fs_info->delete_unused_bgs_mutex);
1433 	btrfs_put_block_group(block_group);
1434 	btrfs_discard_punt_unused_bgs_list(fs_info);
1435 }
1436 
1437 void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1438 {
1439 	struct btrfs_fs_info *fs_info = bg->fs_info;
1440 
1441 	spin_lock(&fs_info->unused_bgs_lock);
1442 	if (list_empty(&bg->bg_list)) {
1443 		btrfs_get_block_group(bg);
1444 		trace_btrfs_add_unused_block_group(bg);
1445 		list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1446 	}
1447 	spin_unlock(&fs_info->unused_bgs_lock);
1448 }
1449 
1450 static int read_bg_from_eb(struct btrfs_fs_info *fs_info, struct btrfs_key *key,
1451 			   struct btrfs_path *path)
1452 {
1453 	struct extent_map_tree *em_tree;
1454 	struct extent_map *em;
1455 	struct btrfs_block_group_item bg;
1456 	struct extent_buffer *leaf;
1457 	int slot;
1458 	u64 flags;
1459 	int ret = 0;
1460 
1461 	slot = path->slots[0];
1462 	leaf = path->nodes[0];
1463 
1464 	em_tree = &fs_info->mapping_tree;
1465 	read_lock(&em_tree->lock);
1466 	em = lookup_extent_mapping(em_tree, key->objectid, key->offset);
1467 	read_unlock(&em_tree->lock);
1468 	if (!em) {
1469 		btrfs_err(fs_info,
1470 			  "logical %llu len %llu found bg but no related chunk",
1471 			  key->objectid, key->offset);
1472 		return -ENOENT;
1473 	}
1474 
1475 	if (em->start != key->objectid || em->len != key->offset) {
1476 		btrfs_err(fs_info,
1477 			"block group %llu len %llu mismatch with chunk %llu len %llu",
1478 			key->objectid, key->offset, em->start, em->len);
1479 		ret = -EUCLEAN;
1480 		goto out_free_em;
1481 	}
1482 
1483 	read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
1484 			   sizeof(bg));
1485 	flags = btrfs_stack_block_group_flags(&bg) &
1486 		BTRFS_BLOCK_GROUP_TYPE_MASK;
1487 
1488 	if (flags != (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1489 		btrfs_err(fs_info,
1490 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
1491 			  key->objectid, key->offset, flags,
1492 			  (BTRFS_BLOCK_GROUP_TYPE_MASK & em->map_lookup->type));
1493 		ret = -EUCLEAN;
1494 	}
1495 
1496 out_free_em:
1497 	free_extent_map(em);
1498 	return ret;
1499 }
1500 
1501 static int find_first_block_group(struct btrfs_fs_info *fs_info,
1502 				  struct btrfs_path *path,
1503 				  struct btrfs_key *key)
1504 {
1505 	struct btrfs_root *root = fs_info->extent_root;
1506 	int ret;
1507 	struct btrfs_key found_key;
1508 	struct extent_buffer *leaf;
1509 	int slot;
1510 
1511 	ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
1512 	if (ret < 0)
1513 		return ret;
1514 
1515 	while (1) {
1516 		slot = path->slots[0];
1517 		leaf = path->nodes[0];
1518 		if (slot >= btrfs_header_nritems(leaf)) {
1519 			ret = btrfs_next_leaf(root, path);
1520 			if (ret == 0)
1521 				continue;
1522 			if (ret < 0)
1523 				goto out;
1524 			break;
1525 		}
1526 		btrfs_item_key_to_cpu(leaf, &found_key, slot);
1527 
1528 		if (found_key.objectid >= key->objectid &&
1529 		    found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
1530 			ret = read_bg_from_eb(fs_info, &found_key, path);
1531 			break;
1532 		}
1533 
1534 		path->slots[0]++;
1535 	}
1536 out:
1537 	return ret;
1538 }
1539 
1540 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
1541 {
1542 	u64 extra_flags = chunk_to_extended(flags) &
1543 				BTRFS_EXTENDED_PROFILE_MASK;
1544 
1545 	write_seqlock(&fs_info->profiles_lock);
1546 	if (flags & BTRFS_BLOCK_GROUP_DATA)
1547 		fs_info->avail_data_alloc_bits |= extra_flags;
1548 	if (flags & BTRFS_BLOCK_GROUP_METADATA)
1549 		fs_info->avail_metadata_alloc_bits |= extra_flags;
1550 	if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
1551 		fs_info->avail_system_alloc_bits |= extra_flags;
1552 	write_sequnlock(&fs_info->profiles_lock);
1553 }
1554 
1555 /**
1556  * btrfs_rmap_block - Map a physical disk address to a list of logical addresses
1557  * @chunk_start:   logical address of block group
1558  * @physical:	   physical address to map to logical addresses
1559  * @logical:	   return array of logical addresses which map to @physical
1560  * @naddrs:	   length of @logical
1561  * @stripe_len:    size of IO stripe for the given block group
1562  *
1563  * Maps a particular @physical disk address to a list of @logical addresses.
1564  * Used primarily to exclude those portions of a block group that contain super
1565  * block copies.
1566  */
1567 EXPORT_FOR_TESTS
1568 int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
1569 		     u64 physical, u64 **logical, int *naddrs, int *stripe_len)
1570 {
1571 	struct extent_map *em;
1572 	struct map_lookup *map;
1573 	u64 *buf;
1574 	u64 bytenr;
1575 	u64 data_stripe_length;
1576 	u64 io_stripe_size;
1577 	int i, nr = 0;
1578 	int ret = 0;
1579 
1580 	em = btrfs_get_chunk_map(fs_info, chunk_start, 1);
1581 	if (IS_ERR(em))
1582 		return -EIO;
1583 
1584 	map = em->map_lookup;
1585 	data_stripe_length = em->orig_block_len;
1586 	io_stripe_size = map->stripe_len;
1587 
1588 	/* For RAID5/6 adjust to a full IO stripe length */
1589 	if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
1590 		io_stripe_size = map->stripe_len * nr_data_stripes(map);
1591 
1592 	buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
1593 	if (!buf) {
1594 		ret = -ENOMEM;
1595 		goto out;
1596 	}
1597 
1598 	for (i = 0; i < map->num_stripes; i++) {
1599 		bool already_inserted = false;
1600 		u64 stripe_nr;
1601 		int j;
1602 
1603 		if (!in_range(physical, map->stripes[i].physical,
1604 			      data_stripe_length))
1605 			continue;
1606 
1607 		stripe_nr = physical - map->stripes[i].physical;
1608 		stripe_nr = div64_u64(stripe_nr, map->stripe_len);
1609 
1610 		if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1611 			stripe_nr = stripe_nr * map->num_stripes + i;
1612 			stripe_nr = div_u64(stripe_nr, map->sub_stripes);
1613 		} else if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1614 			stripe_nr = stripe_nr * map->num_stripes + i;
1615 		}
1616 		/*
1617 		 * The remaining case would be for RAID56, multiply by
1618 		 * nr_data_stripes().  Alternatively, just use rmap_len below
1619 		 * instead of map->stripe_len
1620 		 */
1621 
1622 		bytenr = chunk_start + stripe_nr * io_stripe_size;
1623 
1624 		/* Ensure we don't add duplicate addresses */
1625 		for (j = 0; j < nr; j++) {
1626 			if (buf[j] == bytenr) {
1627 				already_inserted = true;
1628 				break;
1629 			}
1630 		}
1631 
1632 		if (!already_inserted)
1633 			buf[nr++] = bytenr;
1634 	}
1635 
1636 	*logical = buf;
1637 	*naddrs = nr;
1638 	*stripe_len = io_stripe_size;
1639 out:
1640 	free_extent_map(em);
1641 	return ret;
1642 }
1643 
1644 static int exclude_super_stripes(struct btrfs_block_group *cache)
1645 {
1646 	struct btrfs_fs_info *fs_info = cache->fs_info;
1647 	const bool zoned = btrfs_is_zoned(fs_info);
1648 	u64 bytenr;
1649 	u64 *logical;
1650 	int stripe_len;
1651 	int i, nr, ret;
1652 
1653 	if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
1654 		stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
1655 		cache->bytes_super += stripe_len;
1656 		ret = btrfs_add_excluded_extent(fs_info, cache->start,
1657 						stripe_len);
1658 		if (ret)
1659 			return ret;
1660 	}
1661 
1662 	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1663 		bytenr = btrfs_sb_offset(i);
1664 		ret = btrfs_rmap_block(fs_info, cache->start,
1665 				       bytenr, &logical, &nr, &stripe_len);
1666 		if (ret)
1667 			return ret;
1668 
1669 		/* Shouldn't have super stripes in sequential zones */
1670 		if (zoned && nr) {
1671 			btrfs_err(fs_info,
1672 			"zoned: block group %llu must not contain super block",
1673 				  cache->start);
1674 			return -EUCLEAN;
1675 		}
1676 
1677 		while (nr--) {
1678 			u64 len = min_t(u64, stripe_len,
1679 				cache->start + cache->length - logical[nr]);
1680 
1681 			cache->bytes_super += len;
1682 			ret = btrfs_add_excluded_extent(fs_info, logical[nr],
1683 							len);
1684 			if (ret) {
1685 				kfree(logical);
1686 				return ret;
1687 			}
1688 		}
1689 
1690 		kfree(logical);
1691 	}
1692 	return 0;
1693 }
1694 
1695 static void link_block_group(struct btrfs_block_group *cache)
1696 {
1697 	struct btrfs_space_info *space_info = cache->space_info;
1698 	int index = btrfs_bg_flags_to_raid_index(cache->flags);
1699 
1700 	down_write(&space_info->groups_sem);
1701 	list_add_tail(&cache->list, &space_info->block_groups[index]);
1702 	up_write(&space_info->groups_sem);
1703 }
1704 
1705 static struct btrfs_block_group *btrfs_create_block_group_cache(
1706 		struct btrfs_fs_info *fs_info, u64 start)
1707 {
1708 	struct btrfs_block_group *cache;
1709 
1710 	cache = kzalloc(sizeof(*cache), GFP_NOFS);
1711 	if (!cache)
1712 		return NULL;
1713 
1714 	cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
1715 					GFP_NOFS);
1716 	if (!cache->free_space_ctl) {
1717 		kfree(cache);
1718 		return NULL;
1719 	}
1720 
1721 	cache->start = start;
1722 
1723 	cache->fs_info = fs_info;
1724 	cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
1725 
1726 	cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
1727 
1728 	refcount_set(&cache->refs, 1);
1729 	spin_lock_init(&cache->lock);
1730 	init_rwsem(&cache->data_rwsem);
1731 	INIT_LIST_HEAD(&cache->list);
1732 	INIT_LIST_HEAD(&cache->cluster_list);
1733 	INIT_LIST_HEAD(&cache->bg_list);
1734 	INIT_LIST_HEAD(&cache->ro_list);
1735 	INIT_LIST_HEAD(&cache->discard_list);
1736 	INIT_LIST_HEAD(&cache->dirty_list);
1737 	INIT_LIST_HEAD(&cache->io_list);
1738 	btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
1739 	atomic_set(&cache->frozen, 0);
1740 	mutex_init(&cache->free_space_lock);
1741 	btrfs_init_full_stripe_locks_tree(&cache->full_stripe_locks_root);
1742 
1743 	return cache;
1744 }
1745 
1746 /*
1747  * Iterate all chunks and verify that each of them has the corresponding block
1748  * group
1749  */
1750 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
1751 {
1752 	struct extent_map_tree *map_tree = &fs_info->mapping_tree;
1753 	struct extent_map *em;
1754 	struct btrfs_block_group *bg;
1755 	u64 start = 0;
1756 	int ret = 0;
1757 
1758 	while (1) {
1759 		read_lock(&map_tree->lock);
1760 		/*
1761 		 * lookup_extent_mapping will return the first extent map
1762 		 * intersecting the range, so setting @len to 1 is enough to
1763 		 * get the first chunk.
1764 		 */
1765 		em = lookup_extent_mapping(map_tree, start, 1);
1766 		read_unlock(&map_tree->lock);
1767 		if (!em)
1768 			break;
1769 
1770 		bg = btrfs_lookup_block_group(fs_info, em->start);
1771 		if (!bg) {
1772 			btrfs_err(fs_info,
1773 	"chunk start=%llu len=%llu doesn't have corresponding block group",
1774 				     em->start, em->len);
1775 			ret = -EUCLEAN;
1776 			free_extent_map(em);
1777 			break;
1778 		}
1779 		if (bg->start != em->start || bg->length != em->len ||
1780 		    (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
1781 		    (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1782 			btrfs_err(fs_info,
1783 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
1784 				em->start, em->len,
1785 				em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
1786 				bg->start, bg->length,
1787 				bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
1788 			ret = -EUCLEAN;
1789 			free_extent_map(em);
1790 			btrfs_put_block_group(bg);
1791 			break;
1792 		}
1793 		start = em->start + em->len;
1794 		free_extent_map(em);
1795 		btrfs_put_block_group(bg);
1796 	}
1797 	return ret;
1798 }
1799 
1800 static void read_block_group_item(struct btrfs_block_group *cache,
1801 				 struct btrfs_path *path,
1802 				 const struct btrfs_key *key)
1803 {
1804 	struct extent_buffer *leaf = path->nodes[0];
1805 	struct btrfs_block_group_item bgi;
1806 	int slot = path->slots[0];
1807 
1808 	cache->length = key->offset;
1809 
1810 	read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
1811 			   sizeof(bgi));
1812 	cache->used = btrfs_stack_block_group_used(&bgi);
1813 	cache->flags = btrfs_stack_block_group_flags(&bgi);
1814 }
1815 
1816 static int read_one_block_group(struct btrfs_fs_info *info,
1817 				struct btrfs_path *path,
1818 				const struct btrfs_key *key,
1819 				int need_clear)
1820 {
1821 	struct btrfs_block_group *cache;
1822 	struct btrfs_space_info *space_info;
1823 	const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
1824 	int ret;
1825 
1826 	ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
1827 
1828 	cache = btrfs_create_block_group_cache(info, key->objectid);
1829 	if (!cache)
1830 		return -ENOMEM;
1831 
1832 	read_block_group_item(cache, path, key);
1833 
1834 	set_free_space_tree_thresholds(cache);
1835 
1836 	if (need_clear) {
1837 		/*
1838 		 * When we mount with old space cache, we need to
1839 		 * set BTRFS_DC_CLEAR and set dirty flag.
1840 		 *
1841 		 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
1842 		 *    truncate the old free space cache inode and
1843 		 *    setup a new one.
1844 		 * b) Setting 'dirty flag' makes sure that we flush
1845 		 *    the new space cache info onto disk.
1846 		 */
1847 		if (btrfs_test_opt(info, SPACE_CACHE))
1848 			cache->disk_cache_state = BTRFS_DC_CLEAR;
1849 	}
1850 	if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
1851 	    (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
1852 			btrfs_err(info,
1853 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
1854 				  cache->start);
1855 			ret = -EINVAL;
1856 			goto error;
1857 	}
1858 
1859 	/*
1860 	 * We need to exclude the super stripes now so that the space info has
1861 	 * super bytes accounted for, otherwise we'll think we have more space
1862 	 * than we actually do.
1863 	 */
1864 	ret = exclude_super_stripes(cache);
1865 	if (ret) {
1866 		/* We may have excluded something, so call this just in case. */
1867 		btrfs_free_excluded_extents(cache);
1868 		goto error;
1869 	}
1870 
1871 	/*
1872 	 * Check for two cases, either we are full, and therefore don't need
1873 	 * to bother with the caching work since we won't find any space, or we
1874 	 * are empty, and we can just add all the space in and be done with it.
1875 	 * This saves us _a_lot_ of time, particularly in the full case.
1876 	 */
1877 	if (cache->length == cache->used) {
1878 		cache->last_byte_to_unpin = (u64)-1;
1879 		cache->cached = BTRFS_CACHE_FINISHED;
1880 		btrfs_free_excluded_extents(cache);
1881 	} else if (cache->used == 0) {
1882 		cache->last_byte_to_unpin = (u64)-1;
1883 		cache->cached = BTRFS_CACHE_FINISHED;
1884 		add_new_free_space(cache, cache->start,
1885 				   cache->start + cache->length);
1886 		btrfs_free_excluded_extents(cache);
1887 	}
1888 
1889 	ret = btrfs_add_block_group_cache(info, cache);
1890 	if (ret) {
1891 		btrfs_remove_free_space_cache(cache);
1892 		goto error;
1893 	}
1894 	trace_btrfs_add_block_group(info, cache, 0);
1895 	btrfs_update_space_info(info, cache->flags, cache->length,
1896 				cache->used, cache->bytes_super, &space_info);
1897 
1898 	cache->space_info = space_info;
1899 
1900 	link_block_group(cache);
1901 
1902 	set_avail_alloc_bits(info, cache->flags);
1903 	if (btrfs_chunk_readonly(info, cache->start)) {
1904 		inc_block_group_ro(cache, 1);
1905 	} else if (cache->used == 0) {
1906 		ASSERT(list_empty(&cache->bg_list));
1907 		if (btrfs_test_opt(info, DISCARD_ASYNC))
1908 			btrfs_discard_queue_work(&info->discard_ctl, cache);
1909 		else
1910 			btrfs_mark_bg_unused(cache);
1911 	}
1912 	return 0;
1913 error:
1914 	btrfs_put_block_group(cache);
1915 	return ret;
1916 }
1917 
1918 static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
1919 {
1920 	struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1921 	struct btrfs_space_info *space_info;
1922 	struct rb_node *node;
1923 	int ret = 0;
1924 
1925 	for (node = rb_first_cached(&em_tree->map); node; node = rb_next(node)) {
1926 		struct extent_map *em;
1927 		struct map_lookup *map;
1928 		struct btrfs_block_group *bg;
1929 
1930 		em = rb_entry(node, struct extent_map, rb_node);
1931 		map = em->map_lookup;
1932 		bg = btrfs_create_block_group_cache(fs_info, em->start);
1933 		if (!bg) {
1934 			ret = -ENOMEM;
1935 			break;
1936 		}
1937 
1938 		/* Fill dummy cache as FULL */
1939 		bg->length = em->len;
1940 		bg->flags = map->type;
1941 		bg->last_byte_to_unpin = (u64)-1;
1942 		bg->cached = BTRFS_CACHE_FINISHED;
1943 		bg->used = em->len;
1944 		bg->flags = map->type;
1945 		ret = btrfs_add_block_group_cache(fs_info, bg);
1946 		if (ret) {
1947 			btrfs_remove_free_space_cache(bg);
1948 			btrfs_put_block_group(bg);
1949 			break;
1950 		}
1951 		btrfs_update_space_info(fs_info, bg->flags, em->len, em->len,
1952 					0, &space_info);
1953 		bg->space_info = space_info;
1954 		link_block_group(bg);
1955 
1956 		set_avail_alloc_bits(fs_info, bg->flags);
1957 	}
1958 	if (!ret)
1959 		btrfs_init_global_block_rsv(fs_info);
1960 	return ret;
1961 }
1962 
1963 int btrfs_read_block_groups(struct btrfs_fs_info *info)
1964 {
1965 	struct btrfs_path *path;
1966 	int ret;
1967 	struct btrfs_block_group *cache;
1968 	struct btrfs_space_info *space_info;
1969 	struct btrfs_key key;
1970 	int need_clear = 0;
1971 	u64 cache_gen;
1972 
1973 	if (!info->extent_root)
1974 		return fill_dummy_bgs(info);
1975 
1976 	key.objectid = 0;
1977 	key.offset = 0;
1978 	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
1979 	path = btrfs_alloc_path();
1980 	if (!path)
1981 		return -ENOMEM;
1982 
1983 	cache_gen = btrfs_super_cache_generation(info->super_copy);
1984 	if (btrfs_test_opt(info, SPACE_CACHE) &&
1985 	    btrfs_super_generation(info->super_copy) != cache_gen)
1986 		need_clear = 1;
1987 	if (btrfs_test_opt(info, CLEAR_CACHE))
1988 		need_clear = 1;
1989 
1990 	while (1) {
1991 		ret = find_first_block_group(info, path, &key);
1992 		if (ret > 0)
1993 			break;
1994 		if (ret != 0)
1995 			goto error;
1996 
1997 		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1998 		ret = read_one_block_group(info, path, &key, need_clear);
1999 		if (ret < 0)
2000 			goto error;
2001 		key.objectid += key.offset;
2002 		key.offset = 0;
2003 		btrfs_release_path(path);
2004 	}
2005 	btrfs_release_path(path);
2006 
2007 	list_for_each_entry(space_info, &info->space_info, list) {
2008 		int i;
2009 
2010 		for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2011 			if (list_empty(&space_info->block_groups[i]))
2012 				continue;
2013 			cache = list_first_entry(&space_info->block_groups[i],
2014 						 struct btrfs_block_group,
2015 						 list);
2016 			btrfs_sysfs_add_block_group_type(cache);
2017 		}
2018 
2019 		if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2020 		      (BTRFS_BLOCK_GROUP_RAID10 |
2021 		       BTRFS_BLOCK_GROUP_RAID1_MASK |
2022 		       BTRFS_BLOCK_GROUP_RAID56_MASK |
2023 		       BTRFS_BLOCK_GROUP_DUP)))
2024 			continue;
2025 		/*
2026 		 * Avoid allocating from un-mirrored block group if there are
2027 		 * mirrored block groups.
2028 		 */
2029 		list_for_each_entry(cache,
2030 				&space_info->block_groups[BTRFS_RAID_RAID0],
2031 				list)
2032 			inc_block_group_ro(cache, 1);
2033 		list_for_each_entry(cache,
2034 				&space_info->block_groups[BTRFS_RAID_SINGLE],
2035 				list)
2036 			inc_block_group_ro(cache, 1);
2037 	}
2038 
2039 	btrfs_init_global_block_rsv(info);
2040 	ret = check_chunk_block_group_mappings(info);
2041 error:
2042 	btrfs_free_path(path);
2043 	return ret;
2044 }
2045 
2046 static int insert_block_group_item(struct btrfs_trans_handle *trans,
2047 				   struct btrfs_block_group *block_group)
2048 {
2049 	struct btrfs_fs_info *fs_info = trans->fs_info;
2050 	struct btrfs_block_group_item bgi;
2051 	struct btrfs_root *root;
2052 	struct btrfs_key key;
2053 
2054 	spin_lock(&block_group->lock);
2055 	btrfs_set_stack_block_group_used(&bgi, block_group->used);
2056 	btrfs_set_stack_block_group_chunk_objectid(&bgi,
2057 				BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2058 	btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2059 	key.objectid = block_group->start;
2060 	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2061 	key.offset = block_group->length;
2062 	spin_unlock(&block_group->lock);
2063 
2064 	root = fs_info->extent_root;
2065 	return btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2066 }
2067 
2068 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2069 {
2070 	struct btrfs_fs_info *fs_info = trans->fs_info;
2071 	struct btrfs_block_group *block_group;
2072 	int ret = 0;
2073 
2074 	if (!trans->can_flush_pending_bgs)
2075 		return;
2076 
2077 	while (!list_empty(&trans->new_bgs)) {
2078 		int index;
2079 
2080 		block_group = list_first_entry(&trans->new_bgs,
2081 					       struct btrfs_block_group,
2082 					       bg_list);
2083 		if (ret)
2084 			goto next;
2085 
2086 		index = btrfs_bg_flags_to_raid_index(block_group->flags);
2087 
2088 		ret = insert_block_group_item(trans, block_group);
2089 		if (ret)
2090 			btrfs_abort_transaction(trans, ret);
2091 		ret = btrfs_finish_chunk_alloc(trans, block_group->start,
2092 					block_group->length);
2093 		if (ret)
2094 			btrfs_abort_transaction(trans, ret);
2095 		add_block_group_free_space(trans, block_group);
2096 
2097 		/*
2098 		 * If we restriped during balance, we may have added a new raid
2099 		 * type, so now add the sysfs entries when it is safe to do so.
2100 		 * We don't have to worry about locking here as it's handled in
2101 		 * btrfs_sysfs_add_block_group_type.
2102 		 */
2103 		if (block_group->space_info->block_group_kobjs[index] == NULL)
2104 			btrfs_sysfs_add_block_group_type(block_group);
2105 
2106 		/* Already aborted the transaction if it failed. */
2107 next:
2108 		btrfs_delayed_refs_rsv_release(fs_info, 1);
2109 		list_del_init(&block_group->bg_list);
2110 	}
2111 	btrfs_trans_release_chunk_metadata(trans);
2112 }
2113 
2114 int btrfs_make_block_group(struct btrfs_trans_handle *trans, u64 bytes_used,
2115 			   u64 type, u64 chunk_offset, u64 size)
2116 {
2117 	struct btrfs_fs_info *fs_info = trans->fs_info;
2118 	struct btrfs_block_group *cache;
2119 	int ret;
2120 
2121 	btrfs_set_log_full_commit(trans);
2122 
2123 	cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2124 	if (!cache)
2125 		return -ENOMEM;
2126 
2127 	cache->length = size;
2128 	set_free_space_tree_thresholds(cache);
2129 	cache->used = bytes_used;
2130 	cache->flags = type;
2131 	cache->last_byte_to_unpin = (u64)-1;
2132 	cache->cached = BTRFS_CACHE_FINISHED;
2133 	if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2134 		cache->needs_free_space = 1;
2135 	ret = exclude_super_stripes(cache);
2136 	if (ret) {
2137 		/* We may have excluded something, so call this just in case */
2138 		btrfs_free_excluded_extents(cache);
2139 		btrfs_put_block_group(cache);
2140 		return ret;
2141 	}
2142 
2143 	add_new_free_space(cache, chunk_offset, chunk_offset + size);
2144 
2145 	btrfs_free_excluded_extents(cache);
2146 
2147 #ifdef CONFIG_BTRFS_DEBUG
2148 	if (btrfs_should_fragment_free_space(cache)) {
2149 		u64 new_bytes_used = size - bytes_used;
2150 
2151 		bytes_used += new_bytes_used >> 1;
2152 		fragment_free_space(cache);
2153 	}
2154 #endif
2155 	/*
2156 	 * Ensure the corresponding space_info object is created and
2157 	 * assigned to our block group. We want our bg to be added to the rbtree
2158 	 * with its ->space_info set.
2159 	 */
2160 	cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2161 	ASSERT(cache->space_info);
2162 
2163 	ret = btrfs_add_block_group_cache(fs_info, cache);
2164 	if (ret) {
2165 		btrfs_remove_free_space_cache(cache);
2166 		btrfs_put_block_group(cache);
2167 		return ret;
2168 	}
2169 
2170 	/*
2171 	 * Now that our block group has its ->space_info set and is inserted in
2172 	 * the rbtree, update the space info's counters.
2173 	 */
2174 	trace_btrfs_add_block_group(fs_info, cache, 1);
2175 	btrfs_update_space_info(fs_info, cache->flags, size, bytes_used,
2176 				cache->bytes_super, &cache->space_info);
2177 	btrfs_update_global_block_rsv(fs_info);
2178 
2179 	link_block_group(cache);
2180 
2181 	list_add_tail(&cache->bg_list, &trans->new_bgs);
2182 	trans->delayed_ref_updates++;
2183 	btrfs_update_delayed_refs_rsv(trans);
2184 
2185 	set_avail_alloc_bits(fs_info, type);
2186 	return 0;
2187 }
2188 
2189 /*
2190  * Mark one block group RO, can be called several times for the same block
2191  * group.
2192  *
2193  * @cache:		the destination block group
2194  * @do_chunk_alloc:	whether need to do chunk pre-allocation, this is to
2195  * 			ensure we still have some free space after marking this
2196  * 			block group RO.
2197  */
2198 int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2199 			     bool do_chunk_alloc)
2200 {
2201 	struct btrfs_fs_info *fs_info = cache->fs_info;
2202 	struct btrfs_trans_handle *trans;
2203 	u64 alloc_flags;
2204 	int ret;
2205 
2206 again:
2207 	trans = btrfs_join_transaction(fs_info->extent_root);
2208 	if (IS_ERR(trans))
2209 		return PTR_ERR(trans);
2210 
2211 	/*
2212 	 * we're not allowed to set block groups readonly after the dirty
2213 	 * block groups cache has started writing.  If it already started,
2214 	 * back off and let this transaction commit
2215 	 */
2216 	mutex_lock(&fs_info->ro_block_group_mutex);
2217 	if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2218 		u64 transid = trans->transid;
2219 
2220 		mutex_unlock(&fs_info->ro_block_group_mutex);
2221 		btrfs_end_transaction(trans);
2222 
2223 		ret = btrfs_wait_for_commit(fs_info, transid);
2224 		if (ret)
2225 			return ret;
2226 		goto again;
2227 	}
2228 
2229 	if (do_chunk_alloc) {
2230 		/*
2231 		 * If we are changing raid levels, try to allocate a
2232 		 * corresponding block group with the new raid level.
2233 		 */
2234 		alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2235 		if (alloc_flags != cache->flags) {
2236 			ret = btrfs_chunk_alloc(trans, alloc_flags,
2237 						CHUNK_ALLOC_FORCE);
2238 			/*
2239 			 * ENOSPC is allowed here, we may have enough space
2240 			 * already allocated at the new raid level to carry on
2241 			 */
2242 			if (ret == -ENOSPC)
2243 				ret = 0;
2244 			if (ret < 0)
2245 				goto out;
2246 		}
2247 	}
2248 
2249 	ret = inc_block_group_ro(cache, 0);
2250 	if (!do_chunk_alloc)
2251 		goto unlock_out;
2252 	if (!ret)
2253 		goto out;
2254 	alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2255 	ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2256 	if (ret < 0)
2257 		goto out;
2258 	ret = inc_block_group_ro(cache, 0);
2259 out:
2260 	if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2261 		alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2262 		mutex_lock(&fs_info->chunk_mutex);
2263 		check_system_chunk(trans, alloc_flags);
2264 		mutex_unlock(&fs_info->chunk_mutex);
2265 	}
2266 unlock_out:
2267 	mutex_unlock(&fs_info->ro_block_group_mutex);
2268 
2269 	btrfs_end_transaction(trans);
2270 	return ret;
2271 }
2272 
2273 void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
2274 {
2275 	struct btrfs_space_info *sinfo = cache->space_info;
2276 	u64 num_bytes;
2277 
2278 	BUG_ON(!cache->ro);
2279 
2280 	spin_lock(&sinfo->lock);
2281 	spin_lock(&cache->lock);
2282 	if (!--cache->ro) {
2283 		num_bytes = cache->length - cache->reserved -
2284 			    cache->pinned - cache->bytes_super - cache->used;
2285 		sinfo->bytes_readonly -= num_bytes;
2286 		list_del_init(&cache->ro_list);
2287 	}
2288 	spin_unlock(&cache->lock);
2289 	spin_unlock(&sinfo->lock);
2290 }
2291 
2292 static int update_block_group_item(struct btrfs_trans_handle *trans,
2293 				   struct btrfs_path *path,
2294 				   struct btrfs_block_group *cache)
2295 {
2296 	struct btrfs_fs_info *fs_info = trans->fs_info;
2297 	int ret;
2298 	struct btrfs_root *root = fs_info->extent_root;
2299 	unsigned long bi;
2300 	struct extent_buffer *leaf;
2301 	struct btrfs_block_group_item bgi;
2302 	struct btrfs_key key;
2303 
2304 	key.objectid = cache->start;
2305 	key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2306 	key.offset = cache->length;
2307 
2308 	ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2309 	if (ret) {
2310 		if (ret > 0)
2311 			ret = -ENOENT;
2312 		goto fail;
2313 	}
2314 
2315 	leaf = path->nodes[0];
2316 	bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2317 	btrfs_set_stack_block_group_used(&bgi, cache->used);
2318 	btrfs_set_stack_block_group_chunk_objectid(&bgi,
2319 			BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2320 	btrfs_set_stack_block_group_flags(&bgi, cache->flags);
2321 	write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
2322 	btrfs_mark_buffer_dirty(leaf);
2323 fail:
2324 	btrfs_release_path(path);
2325 	return ret;
2326 
2327 }
2328 
2329 static int cache_save_setup(struct btrfs_block_group *block_group,
2330 			    struct btrfs_trans_handle *trans,
2331 			    struct btrfs_path *path)
2332 {
2333 	struct btrfs_fs_info *fs_info = block_group->fs_info;
2334 	struct btrfs_root *root = fs_info->tree_root;
2335 	struct inode *inode = NULL;
2336 	struct extent_changeset *data_reserved = NULL;
2337 	u64 alloc_hint = 0;
2338 	int dcs = BTRFS_DC_ERROR;
2339 	u64 num_pages = 0;
2340 	int retries = 0;
2341 	int ret = 0;
2342 
2343 	if (!btrfs_test_opt(fs_info, SPACE_CACHE))
2344 		return 0;
2345 
2346 	/*
2347 	 * If this block group is smaller than 100 megs don't bother caching the
2348 	 * block group.
2349 	 */
2350 	if (block_group->length < (100 * SZ_1M)) {
2351 		spin_lock(&block_group->lock);
2352 		block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2353 		spin_unlock(&block_group->lock);
2354 		return 0;
2355 	}
2356 
2357 	if (TRANS_ABORTED(trans))
2358 		return 0;
2359 again:
2360 	inode = lookup_free_space_inode(block_group, path);
2361 	if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2362 		ret = PTR_ERR(inode);
2363 		btrfs_release_path(path);
2364 		goto out;
2365 	}
2366 
2367 	if (IS_ERR(inode)) {
2368 		BUG_ON(retries);
2369 		retries++;
2370 
2371 		if (block_group->ro)
2372 			goto out_free;
2373 
2374 		ret = create_free_space_inode(trans, block_group, path);
2375 		if (ret)
2376 			goto out_free;
2377 		goto again;
2378 	}
2379 
2380 	/*
2381 	 * We want to set the generation to 0, that way if anything goes wrong
2382 	 * from here on out we know not to trust this cache when we load up next
2383 	 * time.
2384 	 */
2385 	BTRFS_I(inode)->generation = 0;
2386 	ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2387 	if (ret) {
2388 		/*
2389 		 * So theoretically we could recover from this, simply set the
2390 		 * super cache generation to 0 so we know to invalidate the
2391 		 * cache, but then we'd have to keep track of the block groups
2392 		 * that fail this way so we know we _have_ to reset this cache
2393 		 * before the next commit or risk reading stale cache.  So to
2394 		 * limit our exposure to horrible edge cases lets just abort the
2395 		 * transaction, this only happens in really bad situations
2396 		 * anyway.
2397 		 */
2398 		btrfs_abort_transaction(trans, ret);
2399 		goto out_put;
2400 	}
2401 	WARN_ON(ret);
2402 
2403 	/* We've already setup this transaction, go ahead and exit */
2404 	if (block_group->cache_generation == trans->transid &&
2405 	    i_size_read(inode)) {
2406 		dcs = BTRFS_DC_SETUP;
2407 		goto out_put;
2408 	}
2409 
2410 	if (i_size_read(inode) > 0) {
2411 		ret = btrfs_check_trunc_cache_free_space(fs_info,
2412 					&fs_info->global_block_rsv);
2413 		if (ret)
2414 			goto out_put;
2415 
2416 		ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
2417 		if (ret)
2418 			goto out_put;
2419 	}
2420 
2421 	spin_lock(&block_group->lock);
2422 	if (block_group->cached != BTRFS_CACHE_FINISHED ||
2423 	    !btrfs_test_opt(fs_info, SPACE_CACHE)) {
2424 		/*
2425 		 * don't bother trying to write stuff out _if_
2426 		 * a) we're not cached,
2427 		 * b) we're with nospace_cache mount option,
2428 		 * c) we're with v2 space_cache (FREE_SPACE_TREE).
2429 		 */
2430 		dcs = BTRFS_DC_WRITTEN;
2431 		spin_unlock(&block_group->lock);
2432 		goto out_put;
2433 	}
2434 	spin_unlock(&block_group->lock);
2435 
2436 	/*
2437 	 * We hit an ENOSPC when setting up the cache in this transaction, just
2438 	 * skip doing the setup, we've already cleared the cache so we're safe.
2439 	 */
2440 	if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
2441 		ret = -ENOSPC;
2442 		goto out_put;
2443 	}
2444 
2445 	/*
2446 	 * Try to preallocate enough space based on how big the block group is.
2447 	 * Keep in mind this has to include any pinned space which could end up
2448 	 * taking up quite a bit since it's not folded into the other space
2449 	 * cache.
2450 	 */
2451 	num_pages = div_u64(block_group->length, SZ_256M);
2452 	if (!num_pages)
2453 		num_pages = 1;
2454 
2455 	num_pages *= 16;
2456 	num_pages *= PAGE_SIZE;
2457 
2458 	ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
2459 					  num_pages);
2460 	if (ret)
2461 		goto out_put;
2462 
2463 	ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2464 					      num_pages, num_pages,
2465 					      &alloc_hint);
2466 	/*
2467 	 * Our cache requires contiguous chunks so that we don't modify a bunch
2468 	 * of metadata or split extents when writing the cache out, which means
2469 	 * we can enospc if we are heavily fragmented in addition to just normal
2470 	 * out of space conditions.  So if we hit this just skip setting up any
2471 	 * other block groups for this transaction, maybe we'll unpin enough
2472 	 * space the next time around.
2473 	 */
2474 	if (!ret)
2475 		dcs = BTRFS_DC_SETUP;
2476 	else if (ret == -ENOSPC)
2477 		set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
2478 
2479 out_put:
2480 	iput(inode);
2481 out_free:
2482 	btrfs_release_path(path);
2483 out:
2484 	spin_lock(&block_group->lock);
2485 	if (!ret && dcs == BTRFS_DC_SETUP)
2486 		block_group->cache_generation = trans->transid;
2487 	block_group->disk_cache_state = dcs;
2488 	spin_unlock(&block_group->lock);
2489 
2490 	extent_changeset_free(data_reserved);
2491 	return ret;
2492 }
2493 
2494 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
2495 {
2496 	struct btrfs_fs_info *fs_info = trans->fs_info;
2497 	struct btrfs_block_group *cache, *tmp;
2498 	struct btrfs_transaction *cur_trans = trans->transaction;
2499 	struct btrfs_path *path;
2500 
2501 	if (list_empty(&cur_trans->dirty_bgs) ||
2502 	    !btrfs_test_opt(fs_info, SPACE_CACHE))
2503 		return 0;
2504 
2505 	path = btrfs_alloc_path();
2506 	if (!path)
2507 		return -ENOMEM;
2508 
2509 	/* Could add new block groups, use _safe just in case */
2510 	list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
2511 				 dirty_list) {
2512 		if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2513 			cache_save_setup(cache, trans, path);
2514 	}
2515 
2516 	btrfs_free_path(path);
2517 	return 0;
2518 }
2519 
2520 /*
2521  * Transaction commit does final block group cache writeback during a critical
2522  * section where nothing is allowed to change the FS.  This is required in
2523  * order for the cache to actually match the block group, but can introduce a
2524  * lot of latency into the commit.
2525  *
2526  * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
2527  * There's a chance we'll have to redo some of it if the block group changes
2528  * again during the commit, but it greatly reduces the commit latency by
2529  * getting rid of the easy block groups while we're still allowing others to
2530  * join the commit.
2531  */
2532 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
2533 {
2534 	struct btrfs_fs_info *fs_info = trans->fs_info;
2535 	struct btrfs_block_group *cache;
2536 	struct btrfs_transaction *cur_trans = trans->transaction;
2537 	int ret = 0;
2538 	int should_put;
2539 	struct btrfs_path *path = NULL;
2540 	LIST_HEAD(dirty);
2541 	struct list_head *io = &cur_trans->io_bgs;
2542 	int num_started = 0;
2543 	int loops = 0;
2544 
2545 	spin_lock(&cur_trans->dirty_bgs_lock);
2546 	if (list_empty(&cur_trans->dirty_bgs)) {
2547 		spin_unlock(&cur_trans->dirty_bgs_lock);
2548 		return 0;
2549 	}
2550 	list_splice_init(&cur_trans->dirty_bgs, &dirty);
2551 	spin_unlock(&cur_trans->dirty_bgs_lock);
2552 
2553 again:
2554 	/* Make sure all the block groups on our dirty list actually exist */
2555 	btrfs_create_pending_block_groups(trans);
2556 
2557 	if (!path) {
2558 		path = btrfs_alloc_path();
2559 		if (!path)
2560 			return -ENOMEM;
2561 	}
2562 
2563 	/*
2564 	 * cache_write_mutex is here only to save us from balance or automatic
2565 	 * removal of empty block groups deleting this block group while we are
2566 	 * writing out the cache
2567 	 */
2568 	mutex_lock(&trans->transaction->cache_write_mutex);
2569 	while (!list_empty(&dirty)) {
2570 		bool drop_reserve = true;
2571 
2572 		cache = list_first_entry(&dirty, struct btrfs_block_group,
2573 					 dirty_list);
2574 		/*
2575 		 * This can happen if something re-dirties a block group that
2576 		 * is already under IO.  Just wait for it to finish and then do
2577 		 * it all again
2578 		 */
2579 		if (!list_empty(&cache->io_list)) {
2580 			list_del_init(&cache->io_list);
2581 			btrfs_wait_cache_io(trans, cache, path);
2582 			btrfs_put_block_group(cache);
2583 		}
2584 
2585 
2586 		/*
2587 		 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
2588 		 * it should update the cache_state.  Don't delete until after
2589 		 * we wait.
2590 		 *
2591 		 * Since we're not running in the commit critical section
2592 		 * we need the dirty_bgs_lock to protect from update_block_group
2593 		 */
2594 		spin_lock(&cur_trans->dirty_bgs_lock);
2595 		list_del_init(&cache->dirty_list);
2596 		spin_unlock(&cur_trans->dirty_bgs_lock);
2597 
2598 		should_put = 1;
2599 
2600 		cache_save_setup(cache, trans, path);
2601 
2602 		if (cache->disk_cache_state == BTRFS_DC_SETUP) {
2603 			cache->io_ctl.inode = NULL;
2604 			ret = btrfs_write_out_cache(trans, cache, path);
2605 			if (ret == 0 && cache->io_ctl.inode) {
2606 				num_started++;
2607 				should_put = 0;
2608 
2609 				/*
2610 				 * The cache_write_mutex is protecting the
2611 				 * io_list, also refer to the definition of
2612 				 * btrfs_transaction::io_bgs for more details
2613 				 */
2614 				list_add_tail(&cache->io_list, io);
2615 			} else {
2616 				/*
2617 				 * If we failed to write the cache, the
2618 				 * generation will be bad and life goes on
2619 				 */
2620 				ret = 0;
2621 			}
2622 		}
2623 		if (!ret) {
2624 			ret = update_block_group_item(trans, path, cache);
2625 			/*
2626 			 * Our block group might still be attached to the list
2627 			 * of new block groups in the transaction handle of some
2628 			 * other task (struct btrfs_trans_handle->new_bgs). This
2629 			 * means its block group item isn't yet in the extent
2630 			 * tree. If this happens ignore the error, as we will
2631 			 * try again later in the critical section of the
2632 			 * transaction commit.
2633 			 */
2634 			if (ret == -ENOENT) {
2635 				ret = 0;
2636 				spin_lock(&cur_trans->dirty_bgs_lock);
2637 				if (list_empty(&cache->dirty_list)) {
2638 					list_add_tail(&cache->dirty_list,
2639 						      &cur_trans->dirty_bgs);
2640 					btrfs_get_block_group(cache);
2641 					drop_reserve = false;
2642 				}
2643 				spin_unlock(&cur_trans->dirty_bgs_lock);
2644 			} else if (ret) {
2645 				btrfs_abort_transaction(trans, ret);
2646 			}
2647 		}
2648 
2649 		/* If it's not on the io list, we need to put the block group */
2650 		if (should_put)
2651 			btrfs_put_block_group(cache);
2652 		if (drop_reserve)
2653 			btrfs_delayed_refs_rsv_release(fs_info, 1);
2654 
2655 		if (ret)
2656 			break;
2657 
2658 		/*
2659 		 * Avoid blocking other tasks for too long. It might even save
2660 		 * us from writing caches for block groups that are going to be
2661 		 * removed.
2662 		 */
2663 		mutex_unlock(&trans->transaction->cache_write_mutex);
2664 		mutex_lock(&trans->transaction->cache_write_mutex);
2665 	}
2666 	mutex_unlock(&trans->transaction->cache_write_mutex);
2667 
2668 	/*
2669 	 * Go through delayed refs for all the stuff we've just kicked off
2670 	 * and then loop back (just once)
2671 	 */
2672 	if (!ret)
2673 		ret = btrfs_run_delayed_refs(trans, 0);
2674 	if (!ret && loops == 0) {
2675 		loops++;
2676 		spin_lock(&cur_trans->dirty_bgs_lock);
2677 		list_splice_init(&cur_trans->dirty_bgs, &dirty);
2678 		/*
2679 		 * dirty_bgs_lock protects us from concurrent block group
2680 		 * deletes too (not just cache_write_mutex).
2681 		 */
2682 		if (!list_empty(&dirty)) {
2683 			spin_unlock(&cur_trans->dirty_bgs_lock);
2684 			goto again;
2685 		}
2686 		spin_unlock(&cur_trans->dirty_bgs_lock);
2687 	} else if (ret < 0) {
2688 		btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
2689 	}
2690 
2691 	btrfs_free_path(path);
2692 	return ret;
2693 }
2694 
2695 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
2696 {
2697 	struct btrfs_fs_info *fs_info = trans->fs_info;
2698 	struct btrfs_block_group *cache;
2699 	struct btrfs_transaction *cur_trans = trans->transaction;
2700 	int ret = 0;
2701 	int should_put;
2702 	struct btrfs_path *path;
2703 	struct list_head *io = &cur_trans->io_bgs;
2704 	int num_started = 0;
2705 
2706 	path = btrfs_alloc_path();
2707 	if (!path)
2708 		return -ENOMEM;
2709 
2710 	/*
2711 	 * Even though we are in the critical section of the transaction commit,
2712 	 * we can still have concurrent tasks adding elements to this
2713 	 * transaction's list of dirty block groups. These tasks correspond to
2714 	 * endio free space workers started when writeback finishes for a
2715 	 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
2716 	 * allocate new block groups as a result of COWing nodes of the root
2717 	 * tree when updating the free space inode. The writeback for the space
2718 	 * caches is triggered by an earlier call to
2719 	 * btrfs_start_dirty_block_groups() and iterations of the following
2720 	 * loop.
2721 	 * Also we want to do the cache_save_setup first and then run the
2722 	 * delayed refs to make sure we have the best chance at doing this all
2723 	 * in one shot.
2724 	 */
2725 	spin_lock(&cur_trans->dirty_bgs_lock);
2726 	while (!list_empty(&cur_trans->dirty_bgs)) {
2727 		cache = list_first_entry(&cur_trans->dirty_bgs,
2728 					 struct btrfs_block_group,
2729 					 dirty_list);
2730 
2731 		/*
2732 		 * This can happen if cache_save_setup re-dirties a block group
2733 		 * that is already under IO.  Just wait for it to finish and
2734 		 * then do it all again
2735 		 */
2736 		if (!list_empty(&cache->io_list)) {
2737 			spin_unlock(&cur_trans->dirty_bgs_lock);
2738 			list_del_init(&cache->io_list);
2739 			btrfs_wait_cache_io(trans, cache, path);
2740 			btrfs_put_block_group(cache);
2741 			spin_lock(&cur_trans->dirty_bgs_lock);
2742 		}
2743 
2744 		/*
2745 		 * Don't remove from the dirty list until after we've waited on
2746 		 * any pending IO
2747 		 */
2748 		list_del_init(&cache->dirty_list);
2749 		spin_unlock(&cur_trans->dirty_bgs_lock);
2750 		should_put = 1;
2751 
2752 		cache_save_setup(cache, trans, path);
2753 
2754 		if (!ret)
2755 			ret = btrfs_run_delayed_refs(trans,
2756 						     (unsigned long) -1);
2757 
2758 		if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
2759 			cache->io_ctl.inode = NULL;
2760 			ret = btrfs_write_out_cache(trans, cache, path);
2761 			if (ret == 0 && cache->io_ctl.inode) {
2762 				num_started++;
2763 				should_put = 0;
2764 				list_add_tail(&cache->io_list, io);
2765 			} else {
2766 				/*
2767 				 * If we failed to write the cache, the
2768 				 * generation will be bad and life goes on
2769 				 */
2770 				ret = 0;
2771 			}
2772 		}
2773 		if (!ret) {
2774 			ret = update_block_group_item(trans, path, cache);
2775 			/*
2776 			 * One of the free space endio workers might have
2777 			 * created a new block group while updating a free space
2778 			 * cache's inode (at inode.c:btrfs_finish_ordered_io())
2779 			 * and hasn't released its transaction handle yet, in
2780 			 * which case the new block group is still attached to
2781 			 * its transaction handle and its creation has not
2782 			 * finished yet (no block group item in the extent tree
2783 			 * yet, etc). If this is the case, wait for all free
2784 			 * space endio workers to finish and retry. This is a
2785 			 * very rare case so no need for a more efficient and
2786 			 * complex approach.
2787 			 */
2788 			if (ret == -ENOENT) {
2789 				wait_event(cur_trans->writer_wait,
2790 				   atomic_read(&cur_trans->num_writers) == 1);
2791 				ret = update_block_group_item(trans, path, cache);
2792 			}
2793 			if (ret)
2794 				btrfs_abort_transaction(trans, ret);
2795 		}
2796 
2797 		/* If its not on the io list, we need to put the block group */
2798 		if (should_put)
2799 			btrfs_put_block_group(cache);
2800 		btrfs_delayed_refs_rsv_release(fs_info, 1);
2801 		spin_lock(&cur_trans->dirty_bgs_lock);
2802 	}
2803 	spin_unlock(&cur_trans->dirty_bgs_lock);
2804 
2805 	/*
2806 	 * Refer to the definition of io_bgs member for details why it's safe
2807 	 * to use it without any locking
2808 	 */
2809 	while (!list_empty(io)) {
2810 		cache = list_first_entry(io, struct btrfs_block_group,
2811 					 io_list);
2812 		list_del_init(&cache->io_list);
2813 		btrfs_wait_cache_io(trans, cache, path);
2814 		btrfs_put_block_group(cache);
2815 	}
2816 
2817 	btrfs_free_path(path);
2818 	return ret;
2819 }
2820 
2821 int btrfs_update_block_group(struct btrfs_trans_handle *trans,
2822 			     u64 bytenr, u64 num_bytes, int alloc)
2823 {
2824 	struct btrfs_fs_info *info = trans->fs_info;
2825 	struct btrfs_block_group *cache = NULL;
2826 	u64 total = num_bytes;
2827 	u64 old_val;
2828 	u64 byte_in_group;
2829 	int factor;
2830 	int ret = 0;
2831 
2832 	/* Block accounting for super block */
2833 	spin_lock(&info->delalloc_root_lock);
2834 	old_val = btrfs_super_bytes_used(info->super_copy);
2835 	if (alloc)
2836 		old_val += num_bytes;
2837 	else
2838 		old_val -= num_bytes;
2839 	btrfs_set_super_bytes_used(info->super_copy, old_val);
2840 	spin_unlock(&info->delalloc_root_lock);
2841 
2842 	while (total) {
2843 		cache = btrfs_lookup_block_group(info, bytenr);
2844 		if (!cache) {
2845 			ret = -ENOENT;
2846 			break;
2847 		}
2848 		factor = btrfs_bg_type_to_factor(cache->flags);
2849 
2850 		/*
2851 		 * If this block group has free space cache written out, we
2852 		 * need to make sure to load it if we are removing space.  This
2853 		 * is because we need the unpinning stage to actually add the
2854 		 * space back to the block group, otherwise we will leak space.
2855 		 */
2856 		if (!alloc && !btrfs_block_group_done(cache))
2857 			btrfs_cache_block_group(cache, 1);
2858 
2859 		byte_in_group = bytenr - cache->start;
2860 		WARN_ON(byte_in_group > cache->length);
2861 
2862 		spin_lock(&cache->space_info->lock);
2863 		spin_lock(&cache->lock);
2864 
2865 		if (btrfs_test_opt(info, SPACE_CACHE) &&
2866 		    cache->disk_cache_state < BTRFS_DC_CLEAR)
2867 			cache->disk_cache_state = BTRFS_DC_CLEAR;
2868 
2869 		old_val = cache->used;
2870 		num_bytes = min(total, cache->length - byte_in_group);
2871 		if (alloc) {
2872 			old_val += num_bytes;
2873 			cache->used = old_val;
2874 			cache->reserved -= num_bytes;
2875 			cache->space_info->bytes_reserved -= num_bytes;
2876 			cache->space_info->bytes_used += num_bytes;
2877 			cache->space_info->disk_used += num_bytes * factor;
2878 			spin_unlock(&cache->lock);
2879 			spin_unlock(&cache->space_info->lock);
2880 		} else {
2881 			old_val -= num_bytes;
2882 			cache->used = old_val;
2883 			cache->pinned += num_bytes;
2884 			btrfs_space_info_update_bytes_pinned(info,
2885 					cache->space_info, num_bytes);
2886 			cache->space_info->bytes_used -= num_bytes;
2887 			cache->space_info->disk_used -= num_bytes * factor;
2888 			spin_unlock(&cache->lock);
2889 			spin_unlock(&cache->space_info->lock);
2890 
2891 			percpu_counter_add_batch(
2892 					&cache->space_info->total_bytes_pinned,
2893 					num_bytes,
2894 					BTRFS_TOTAL_BYTES_PINNED_BATCH);
2895 			set_extent_dirty(&trans->transaction->pinned_extents,
2896 					 bytenr, bytenr + num_bytes - 1,
2897 					 GFP_NOFS | __GFP_NOFAIL);
2898 		}
2899 
2900 		spin_lock(&trans->transaction->dirty_bgs_lock);
2901 		if (list_empty(&cache->dirty_list)) {
2902 			list_add_tail(&cache->dirty_list,
2903 				      &trans->transaction->dirty_bgs);
2904 			trans->delayed_ref_updates++;
2905 			btrfs_get_block_group(cache);
2906 		}
2907 		spin_unlock(&trans->transaction->dirty_bgs_lock);
2908 
2909 		/*
2910 		 * No longer have used bytes in this block group, queue it for
2911 		 * deletion. We do this after adding the block group to the
2912 		 * dirty list to avoid races between cleaner kthread and space
2913 		 * cache writeout.
2914 		 */
2915 		if (!alloc && old_val == 0) {
2916 			if (!btrfs_test_opt(info, DISCARD_ASYNC))
2917 				btrfs_mark_bg_unused(cache);
2918 		}
2919 
2920 		btrfs_put_block_group(cache);
2921 		total -= num_bytes;
2922 		bytenr += num_bytes;
2923 	}
2924 
2925 	/* Modified block groups are accounted for in the delayed_refs_rsv. */
2926 	btrfs_update_delayed_refs_rsv(trans);
2927 	return ret;
2928 }
2929 
2930 /**
2931  * btrfs_add_reserved_bytes - update the block_group and space info counters
2932  * @cache:	The cache we are manipulating
2933  * @ram_bytes:  The number of bytes of file content, and will be same to
2934  *              @num_bytes except for the compress path.
2935  * @num_bytes:	The number of bytes in question
2936  * @delalloc:   The blocks are allocated for the delalloc write
2937  *
2938  * This is called by the allocator when it reserves space. If this is a
2939  * reservation and the block group has become read only we cannot make the
2940  * reservation and return -EAGAIN, otherwise this function always succeeds.
2941  */
2942 int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
2943 			     u64 ram_bytes, u64 num_bytes, int delalloc)
2944 {
2945 	struct btrfs_space_info *space_info = cache->space_info;
2946 	int ret = 0;
2947 
2948 	spin_lock(&space_info->lock);
2949 	spin_lock(&cache->lock);
2950 	if (cache->ro) {
2951 		ret = -EAGAIN;
2952 	} else {
2953 		cache->reserved += num_bytes;
2954 		space_info->bytes_reserved += num_bytes;
2955 		trace_btrfs_space_reservation(cache->fs_info, "space_info",
2956 					      space_info->flags, num_bytes, 1);
2957 		btrfs_space_info_update_bytes_may_use(cache->fs_info,
2958 						      space_info, -ram_bytes);
2959 		if (delalloc)
2960 			cache->delalloc_bytes += num_bytes;
2961 
2962 		/*
2963 		 * Compression can use less space than we reserved, so wake
2964 		 * tickets if that happens
2965 		 */
2966 		if (num_bytes < ram_bytes)
2967 			btrfs_try_granting_tickets(cache->fs_info, space_info);
2968 	}
2969 	spin_unlock(&cache->lock);
2970 	spin_unlock(&space_info->lock);
2971 	return ret;
2972 }
2973 
2974 /**
2975  * btrfs_free_reserved_bytes - update the block_group and space info counters
2976  * @cache:      The cache we are manipulating
2977  * @num_bytes:  The number of bytes in question
2978  * @delalloc:   The blocks are allocated for the delalloc write
2979  *
2980  * This is called by somebody who is freeing space that was never actually used
2981  * on disk.  For example if you reserve some space for a new leaf in transaction
2982  * A and before transaction A commits you free that leaf, you call this with
2983  * reserve set to 0 in order to clear the reservation.
2984  */
2985 void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
2986 			       u64 num_bytes, int delalloc)
2987 {
2988 	struct btrfs_space_info *space_info = cache->space_info;
2989 
2990 	spin_lock(&space_info->lock);
2991 	spin_lock(&cache->lock);
2992 	if (cache->ro)
2993 		space_info->bytes_readonly += num_bytes;
2994 	cache->reserved -= num_bytes;
2995 	space_info->bytes_reserved -= num_bytes;
2996 	space_info->max_extent_size = 0;
2997 
2998 	if (delalloc)
2999 		cache->delalloc_bytes -= num_bytes;
3000 	spin_unlock(&cache->lock);
3001 
3002 	btrfs_try_granting_tickets(cache->fs_info, space_info);
3003 	spin_unlock(&space_info->lock);
3004 }
3005 
3006 static void force_metadata_allocation(struct btrfs_fs_info *info)
3007 {
3008 	struct list_head *head = &info->space_info;
3009 	struct btrfs_space_info *found;
3010 
3011 	list_for_each_entry(found, head, list) {
3012 		if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3013 			found->force_alloc = CHUNK_ALLOC_FORCE;
3014 	}
3015 }
3016 
3017 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3018 			      struct btrfs_space_info *sinfo, int force)
3019 {
3020 	u64 bytes_used = btrfs_space_info_used(sinfo, false);
3021 	u64 thresh;
3022 
3023 	if (force == CHUNK_ALLOC_FORCE)
3024 		return 1;
3025 
3026 	/*
3027 	 * in limited mode, we want to have some free space up to
3028 	 * about 1% of the FS size.
3029 	 */
3030 	if (force == CHUNK_ALLOC_LIMITED) {
3031 		thresh = btrfs_super_total_bytes(fs_info->super_copy);
3032 		thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
3033 
3034 		if (sinfo->total_bytes - bytes_used < thresh)
3035 			return 1;
3036 	}
3037 
3038 	if (bytes_used + SZ_2M < div_factor(sinfo->total_bytes, 8))
3039 		return 0;
3040 	return 1;
3041 }
3042 
3043 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3044 {
3045 	u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3046 
3047 	return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3048 }
3049 
3050 /*
3051  * If force is CHUNK_ALLOC_FORCE:
3052  *    - return 1 if it successfully allocates a chunk,
3053  *    - return errors including -ENOSPC otherwise.
3054  * If force is NOT CHUNK_ALLOC_FORCE:
3055  *    - return 0 if it doesn't need to allocate a new chunk,
3056  *    - return 1 if it successfully allocates a chunk,
3057  *    - return errors including -ENOSPC otherwise.
3058  */
3059 int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
3060 		      enum btrfs_chunk_alloc_enum force)
3061 {
3062 	struct btrfs_fs_info *fs_info = trans->fs_info;
3063 	struct btrfs_space_info *space_info;
3064 	bool wait_for_alloc = false;
3065 	bool should_alloc = false;
3066 	int ret = 0;
3067 
3068 	/* Don't re-enter if we're already allocating a chunk */
3069 	if (trans->allocating_chunk)
3070 		return -ENOSPC;
3071 
3072 	space_info = btrfs_find_space_info(fs_info, flags);
3073 	ASSERT(space_info);
3074 
3075 	do {
3076 		spin_lock(&space_info->lock);
3077 		if (force < space_info->force_alloc)
3078 			force = space_info->force_alloc;
3079 		should_alloc = should_alloc_chunk(fs_info, space_info, force);
3080 		if (space_info->full) {
3081 			/* No more free physical space */
3082 			if (should_alloc)
3083 				ret = -ENOSPC;
3084 			else
3085 				ret = 0;
3086 			spin_unlock(&space_info->lock);
3087 			return ret;
3088 		} else if (!should_alloc) {
3089 			spin_unlock(&space_info->lock);
3090 			return 0;
3091 		} else if (space_info->chunk_alloc) {
3092 			/*
3093 			 * Someone is already allocating, so we need to block
3094 			 * until this someone is finished and then loop to
3095 			 * recheck if we should continue with our allocation
3096 			 * attempt.
3097 			 */
3098 			wait_for_alloc = true;
3099 			spin_unlock(&space_info->lock);
3100 			mutex_lock(&fs_info->chunk_mutex);
3101 			mutex_unlock(&fs_info->chunk_mutex);
3102 		} else {
3103 			/* Proceed with allocation */
3104 			space_info->chunk_alloc = 1;
3105 			wait_for_alloc = false;
3106 			spin_unlock(&space_info->lock);
3107 		}
3108 
3109 		cond_resched();
3110 	} while (wait_for_alloc);
3111 
3112 	mutex_lock(&fs_info->chunk_mutex);
3113 	trans->allocating_chunk = true;
3114 
3115 	/*
3116 	 * If we have mixed data/metadata chunks we want to make sure we keep
3117 	 * allocating mixed chunks instead of individual chunks.
3118 	 */
3119 	if (btrfs_mixed_space_info(space_info))
3120 		flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3121 
3122 	/*
3123 	 * if we're doing a data chunk, go ahead and make sure that
3124 	 * we keep a reasonable number of metadata chunks allocated in the
3125 	 * FS as well.
3126 	 */
3127 	if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3128 		fs_info->data_chunk_allocations++;
3129 		if (!(fs_info->data_chunk_allocations %
3130 		      fs_info->metadata_ratio))
3131 			force_metadata_allocation(fs_info);
3132 	}
3133 
3134 	/*
3135 	 * Check if we have enough space in SYSTEM chunk because we may need
3136 	 * to update devices.
3137 	 */
3138 	check_system_chunk(trans, flags);
3139 
3140 	ret = btrfs_alloc_chunk(trans, flags);
3141 	trans->allocating_chunk = false;
3142 
3143 	spin_lock(&space_info->lock);
3144 	if (ret < 0) {
3145 		if (ret == -ENOSPC)
3146 			space_info->full = 1;
3147 		else
3148 			goto out;
3149 	} else {
3150 		ret = 1;
3151 		space_info->max_extent_size = 0;
3152 	}
3153 
3154 	space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3155 out:
3156 	space_info->chunk_alloc = 0;
3157 	spin_unlock(&space_info->lock);
3158 	mutex_unlock(&fs_info->chunk_mutex);
3159 	/*
3160 	 * When we allocate a new chunk we reserve space in the chunk block
3161 	 * reserve to make sure we can COW nodes/leafs in the chunk tree or
3162 	 * add new nodes/leafs to it if we end up needing to do it when
3163 	 * inserting the chunk item and updating device items as part of the
3164 	 * second phase of chunk allocation, performed by
3165 	 * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
3166 	 * large number of new block groups to create in our transaction
3167 	 * handle's new_bgs list to avoid exhausting the chunk block reserve
3168 	 * in extreme cases - like having a single transaction create many new
3169 	 * block groups when starting to write out the free space caches of all
3170 	 * the block groups that were made dirty during the lifetime of the
3171 	 * transaction.
3172 	 */
3173 	if (trans->chunk_bytes_reserved >= (u64)SZ_2M)
3174 		btrfs_create_pending_block_groups(trans);
3175 
3176 	return ret;
3177 }
3178 
3179 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
3180 {
3181 	u64 num_dev;
3182 
3183 	num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
3184 	if (!num_dev)
3185 		num_dev = fs_info->fs_devices->rw_devices;
3186 
3187 	return num_dev;
3188 }
3189 
3190 /*
3191  * Reserve space in the system space for allocating or removing a chunk
3192  */
3193 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
3194 {
3195 	struct btrfs_fs_info *fs_info = trans->fs_info;
3196 	struct btrfs_space_info *info;
3197 	u64 left;
3198 	u64 thresh;
3199 	int ret = 0;
3200 	u64 num_devs;
3201 
3202 	/*
3203 	 * Needed because we can end up allocating a system chunk and for an
3204 	 * atomic and race free space reservation in the chunk block reserve.
3205 	 */
3206 	lockdep_assert_held(&fs_info->chunk_mutex);
3207 
3208 	info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3209 	spin_lock(&info->lock);
3210 	left = info->total_bytes - btrfs_space_info_used(info, true);
3211 	spin_unlock(&info->lock);
3212 
3213 	num_devs = get_profile_num_devs(fs_info, type);
3214 
3215 	/* num_devs device items to update and 1 chunk item to add or remove */
3216 	thresh = btrfs_calc_metadata_size(fs_info, num_devs) +
3217 		btrfs_calc_insert_metadata_size(fs_info, 1);
3218 
3219 	if (left < thresh && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
3220 		btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
3221 			   left, thresh, type);
3222 		btrfs_dump_space_info(fs_info, info, 0, 0);
3223 	}
3224 
3225 	if (left < thresh) {
3226 		u64 flags = btrfs_system_alloc_profile(fs_info);
3227 
3228 		/*
3229 		 * Ignore failure to create system chunk. We might end up not
3230 		 * needing it, as we might not need to COW all nodes/leafs from
3231 		 * the paths we visit in the chunk tree (they were already COWed
3232 		 * or created in the current transaction for example).
3233 		 */
3234 		ret = btrfs_alloc_chunk(trans, flags);
3235 	}
3236 
3237 	if (!ret) {
3238 		ret = btrfs_block_rsv_add(fs_info->chunk_root,
3239 					  &fs_info->chunk_block_rsv,
3240 					  thresh, BTRFS_RESERVE_NO_FLUSH);
3241 		if (!ret)
3242 			trans->chunk_bytes_reserved += thresh;
3243 	}
3244 }
3245 
3246 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
3247 {
3248 	struct btrfs_block_group *block_group;
3249 	u64 last = 0;
3250 
3251 	while (1) {
3252 		struct inode *inode;
3253 
3254 		block_group = btrfs_lookup_first_block_group(info, last);
3255 		while (block_group) {
3256 			btrfs_wait_block_group_cache_done(block_group);
3257 			spin_lock(&block_group->lock);
3258 			if (block_group->iref)
3259 				break;
3260 			spin_unlock(&block_group->lock);
3261 			block_group = btrfs_next_block_group(block_group);
3262 		}
3263 		if (!block_group) {
3264 			if (last == 0)
3265 				break;
3266 			last = 0;
3267 			continue;
3268 		}
3269 
3270 		inode = block_group->inode;
3271 		block_group->iref = 0;
3272 		block_group->inode = NULL;
3273 		spin_unlock(&block_group->lock);
3274 		ASSERT(block_group->io_ctl.inode == NULL);
3275 		iput(inode);
3276 		last = block_group->start + block_group->length;
3277 		btrfs_put_block_group(block_group);
3278 	}
3279 }
3280 
3281 /*
3282  * Must be called only after stopping all workers, since we could have block
3283  * group caching kthreads running, and therefore they could race with us if we
3284  * freed the block groups before stopping them.
3285  */
3286 int btrfs_free_block_groups(struct btrfs_fs_info *info)
3287 {
3288 	struct btrfs_block_group *block_group;
3289 	struct btrfs_space_info *space_info;
3290 	struct btrfs_caching_control *caching_ctl;
3291 	struct rb_node *n;
3292 
3293 	spin_lock(&info->block_group_cache_lock);
3294 	while (!list_empty(&info->caching_block_groups)) {
3295 		caching_ctl = list_entry(info->caching_block_groups.next,
3296 					 struct btrfs_caching_control, list);
3297 		list_del(&caching_ctl->list);
3298 		btrfs_put_caching_control(caching_ctl);
3299 	}
3300 	spin_unlock(&info->block_group_cache_lock);
3301 
3302 	spin_lock(&info->unused_bgs_lock);
3303 	while (!list_empty(&info->unused_bgs)) {
3304 		block_group = list_first_entry(&info->unused_bgs,
3305 					       struct btrfs_block_group,
3306 					       bg_list);
3307 		list_del_init(&block_group->bg_list);
3308 		btrfs_put_block_group(block_group);
3309 	}
3310 	spin_unlock(&info->unused_bgs_lock);
3311 
3312 	spin_lock(&info->block_group_cache_lock);
3313 	while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
3314 		block_group = rb_entry(n, struct btrfs_block_group,
3315 				       cache_node);
3316 		rb_erase(&block_group->cache_node,
3317 			 &info->block_group_cache_tree);
3318 		RB_CLEAR_NODE(&block_group->cache_node);
3319 		spin_unlock(&info->block_group_cache_lock);
3320 
3321 		down_write(&block_group->space_info->groups_sem);
3322 		list_del(&block_group->list);
3323 		up_write(&block_group->space_info->groups_sem);
3324 
3325 		/*
3326 		 * We haven't cached this block group, which means we could
3327 		 * possibly have excluded extents on this block group.
3328 		 */
3329 		if (block_group->cached == BTRFS_CACHE_NO ||
3330 		    block_group->cached == BTRFS_CACHE_ERROR)
3331 			btrfs_free_excluded_extents(block_group);
3332 
3333 		btrfs_remove_free_space_cache(block_group);
3334 		ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
3335 		ASSERT(list_empty(&block_group->dirty_list));
3336 		ASSERT(list_empty(&block_group->io_list));
3337 		ASSERT(list_empty(&block_group->bg_list));
3338 		ASSERT(refcount_read(&block_group->refs) == 1);
3339 		btrfs_put_block_group(block_group);
3340 
3341 		spin_lock(&info->block_group_cache_lock);
3342 	}
3343 	spin_unlock(&info->block_group_cache_lock);
3344 
3345 	btrfs_release_global_block_rsv(info);
3346 
3347 	while (!list_empty(&info->space_info)) {
3348 		space_info = list_entry(info->space_info.next,
3349 					struct btrfs_space_info,
3350 					list);
3351 
3352 		/*
3353 		 * Do not hide this behind enospc_debug, this is actually
3354 		 * important and indicates a real bug if this happens.
3355 		 */
3356 		if (WARN_ON(space_info->bytes_pinned > 0 ||
3357 			    space_info->bytes_reserved > 0 ||
3358 			    space_info->bytes_may_use > 0))
3359 			btrfs_dump_space_info(info, space_info, 0, 0);
3360 		WARN_ON(space_info->reclaim_size > 0);
3361 		list_del(&space_info->list);
3362 		btrfs_sysfs_remove_space_info(space_info);
3363 	}
3364 	return 0;
3365 }
3366 
3367 void btrfs_freeze_block_group(struct btrfs_block_group *cache)
3368 {
3369 	atomic_inc(&cache->frozen);
3370 }
3371 
3372 void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
3373 {
3374 	struct btrfs_fs_info *fs_info = block_group->fs_info;
3375 	struct extent_map_tree *em_tree;
3376 	struct extent_map *em;
3377 	bool cleanup;
3378 
3379 	spin_lock(&block_group->lock);
3380 	cleanup = (atomic_dec_and_test(&block_group->frozen) &&
3381 		   block_group->removed);
3382 	spin_unlock(&block_group->lock);
3383 
3384 	if (cleanup) {
3385 		em_tree = &fs_info->mapping_tree;
3386 		write_lock(&em_tree->lock);
3387 		em = lookup_extent_mapping(em_tree, block_group->start,
3388 					   1);
3389 		BUG_ON(!em); /* logic error, can't happen */
3390 		remove_extent_mapping(em_tree, em);
3391 		write_unlock(&em_tree->lock);
3392 
3393 		/* once for us and once for the tree */
3394 		free_extent_map(em);
3395 		free_extent_map(em);
3396 
3397 		/*
3398 		 * We may have left one free space entry and other possible
3399 		 * tasks trimming this block group have left 1 entry each one.
3400 		 * Free them if any.
3401 		 */
3402 		__btrfs_remove_free_space_cache(block_group->free_space_ctl);
3403 	}
3404 }
3405